diff --git a/.github/ISSUE_TEMPLATE/custom.md b/.github/ISSUE_TEMPLATE/bug.md similarity index 96% rename from .github/ISSUE_TEMPLATE/custom.md rename to .github/ISSUE_TEMPLATE/bug.md index 8fd955356..c003fe7c1 100644 --- a/.github/ISSUE_TEMPLATE/custom.md +++ b/.github/ISSUE_TEMPLATE/bug.md @@ -1,8 +1,7 @@ --- -name: Issue and enhancement template -about: Used to report issues and request enhancements for llama.cpp -title: "[User] Insert summary of your issue or enhancement.." -labels: '' +name: Bug template +about: Used to report bugs in llama.cpp +labels: ["bug-unconfirmed"] assignees: '' --- @@ -46,7 +45,7 @@ $ g++ --version # Failure Information (for bugs) -Please help provide information about the failure if this is a bug. If it is not a bug, please remove the rest of this template. +Please help provide information about the failure / bug. # Steps to Reproduce diff --git a/.github/ISSUE_TEMPLATE/enhancement.md b/.github/ISSUE_TEMPLATE/enhancement.md new file mode 100644 index 000000000..dcffda750 --- /dev/null +++ b/.github/ISSUE_TEMPLATE/enhancement.md @@ -0,0 +1,28 @@ +--- +name: Enhancement template +about: Used to request enhancements for llama.cpp +labels: ["enhancement"] +assignees: '' + +--- + +# Prerequisites + +Please answer the following questions for yourself before submitting an issue. + +- [ ] I am running the latest code. Development is very rapid so there are no tagged versions as of now. +- [ ] I carefully followed the [README.md](https://github.com/ggerganov/llama.cpp/blob/master/README.md). +- [ ] I [searched using keywords relevant to my issue](https://docs.github.com/en/issues/tracking-your-work-with-issues/filtering-and-searching-issues-and-pull-requests) to make sure that I am creating a new issue that is not already open (or closed). +- [ ] I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new bug or useful enhancement to share. + +# Feature Description + +Please provide a detailed written description of what you were trying to do, and what you expected `llama.cpp` to do as an enhancement. + +# Motivation + +Please provide a detailed written description of reasons why this feature is necessary and how it is useful to `llama.cpp` users. + +# Possible Implementation + +If you have an idea as to how it can be implemented, please write a detailed description. Feel free to give links to external sources or share visuals that might be helpful to understand the details better. diff --git a/.gitignore b/.gitignore index 471cf90d5..5d7c5479e 100644 --- a/.gitignore +++ b/.gitignore @@ -10,10 +10,12 @@ *.gcno *.gcda *.dot +*.bat *.metallib .DS_Store .build/ .cache/ +.ccls-cache/ .direnv/ .envrc .swiftpm diff --git a/CMakeLists.txt b/CMakeLists.txt index 6af42a6c2..3659279e2 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -82,6 +82,7 @@ set(LLAMA_BLAS_VENDOR "Generic" CACHE STRING "llama: BLAS library vendor") option(LLAMA_CUBLAS "llama: use CUDA" OFF) #option(LLAMA_CUDA_CUBLAS "llama: use cuBLAS for prompt processing" OFF) option(LLAMA_CUDA_FORCE_DMMV "llama: use dmmv instead of mmvq CUDA kernels" OFF) +option(LLAMA_CUDA_FORCE_MMQ "llama: use mmq kernels instead of cuBLAS" OFF) set(LLAMA_CUDA_DMMV_X "32" CACHE STRING "llama: x stride for dmmv CUDA kernels") set(LLAMA_CUDA_MMV_Y "1" CACHE STRING "llama: y block size for mmv CUDA kernels") option(LLAMA_CUDA_F16 "llama: use 16 bit floats for some calculations" OFF) @@ -93,7 +94,6 @@ option(LLAMA_CLBLAST "llama: use CLBlast" option(LLAMA_METAL "llama: use Metal" ${LLAMA_METAL_DEFAULT}) option(LLAMA_METAL_NDEBUG "llama: disable Metal debugging" OFF) option(LLAMA_MPI "llama: use MPI" OFF) -option(LLAMA_K_QUANTS "llama: use k-quants" ON) option(LLAMA_QKK_64 "llama: use super-block size of 64 for k-quants" OFF) option(LLAMA_BUILD_TESTS "llama: build tests" ${LLAMA_STANDALONE}) @@ -277,13 +277,8 @@ if (LLAMA_BLAS) endif() endif() -if (LLAMA_K_QUANTS) - set(GGML_HEADERS_EXTRA k_quants.h) - set(GGML_SOURCES_EXTRA k_quants.c) - add_compile_definitions(GGML_USE_K_QUANTS) - if (LLAMA_QKK_64) - add_compile_definitions(GGML_QKK_64) - endif() +if (LLAMA_QKK_64) + add_compile_definitions(GGML_QKK_64) endif() if (LLAMA_CUBLAS) @@ -305,6 +300,9 @@ if (LLAMA_CUBLAS) if (LLAMA_CUDA_FORCE_DMMV) add_compile_definitions(GGML_CUDA_FORCE_DMMV) endif() + if (LLAMA_CUDA_FORCE_MMQ) + add_compile_definitions(GGML_CUDA_FORCE_MMQ) + endif() add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X}) add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y}) if (DEFINED LLAMA_CUDA_DMMV_Y) @@ -331,6 +329,7 @@ if (LLAMA_CUBLAS) set(CMAKE_CUDA_ARCHITECTURES "60;61;70") # needed for f16 CUDA intrinsics else() set(CMAKE_CUDA_ARCHITECTURES "52;61;70") # lowest CUDA 12 standard + lowest for integer intrinsics + #set(CMAKE_CUDA_ARCHITECTURES "") # use this to compile much faster, but only F16 models work endif() endif() message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}") @@ -404,6 +403,9 @@ if (LLAMA_HIPBLAS) if (LLAMA_CUDA_FORCE_DMMV) target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_DMMV) endif() + if (LLAMA_CUDA_FORCE_MMQ) + target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_MMQ) + endif() target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X}) target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y}) target_compile_definitions(ggml-rocm PRIVATE K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER}) @@ -665,6 +667,8 @@ add_library(ggml OBJECT ggml-alloc.h ggml-backend.c ggml-backend.h + ggml-quants.c + ggml-quants.h ${GGML_SOURCES_CUDA} ${GGML_HEADERS_CUDA} ${GGML_SOURCES_OPENCL} ${GGML_HEADERS_OPENCL} ${GGML_SOURCES_METAL} ${GGML_HEADERS_METAL} diff --git a/Makefile b/Makefile index 325ae747b..c53c1e726 100644 --- a/Makefile +++ b/Makefile @@ -342,13 +342,9 @@ else MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d endif -ifndef LLAMA_NO_K_QUANTS - MK_CPPFLAGS += -DGGML_USE_K_QUANTS - OBJS += k_quants.o ifdef LLAMA_QKK_64 MK_CPPFLAGS += -DGGML_QKK_64 endif -endif ifndef LLAMA_NO_ACCELERATE # Mac OS - include Accelerate framework. @@ -365,7 +361,7 @@ ifdef LLAMA_MPI MK_CPPFLAGS += -DGGML_USE_MPI MK_CFLAGS += -Wno-cast-qual MK_CXXFLAGS += -Wno-cast-qual - OBJS += ggml-mpi.o + OBJS += ggml-mpi.o endif # LLAMA_MPI ifdef LLAMA_OPENBLAS @@ -382,7 +378,7 @@ endif # LLAMA_BLIS ifdef LLAMA_CUBLAS MK_CPPFLAGS += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include MK_LDFLAGS += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib - OBJS += ggml-cuda.o + OBJS += ggml-cuda.o NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math ifdef LLAMA_CUDA_NVCC NVCC = $(LLAMA_CUDA_NVCC) @@ -397,6 +393,9 @@ endif # CUDA_DOCKER_ARCH ifdef LLAMA_CUDA_FORCE_DMMV NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV endif # LLAMA_CUDA_FORCE_DMMV +ifdef LLAMA_CUDA_FORCE_MMQ + NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ +endif # LLAMA_CUDA_FORCE_MMQ ifdef LLAMA_CUDA_DMMV_X NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X) else @@ -494,11 +493,6 @@ ggml-mpi.o: ggml-mpi.c ggml-mpi.h $(CC) $(CFLAGS) -c $< -o $@ endif # LLAMA_MPI -ifndef LLAMA_NO_K_QUANTS -k_quants.o: k_quants.c k_quants.h - $(CC) $(CFLAGS) -c $< -o $@ -endif # LLAMA_NO_K_QUANTS - # combine build flags with cmdline overrides override CFLAGS := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(CFLAGS) override CXXFLAGS := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS) @@ -539,15 +533,18 @@ ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h ggml-backend.o: ggml-backend.c ggml.h ggml-backend.h $(CC) $(CFLAGS) -c $< -o $@ -OBJS += ggml-alloc.o ggml-backend.o +ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h + $(CC) $(CFLAGS) -c $< -o $@ + +OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o llama.o: llama.cpp ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h $(CXX) $(CXXFLAGS) -c $< -o $@ -COMMON_H_DEPS = common/common.h common/sampling.h build-info.h common/log.h -COMMON_DEPS = $(COMMON_H_DEPS) common.o sampling.o grammar-parser.o +COMMON_H_DEPS = common/common.h common/sampling.h common/log.h +COMMON_DEPS = common.o sampling.o grammar-parser.o -common.o: common/common.cpp $(COMMON_H_DEPS) +common.o: common/common.cpp build-info.h $(COMMON_H_DEPS) $(CXX) $(CXXFLAGS) -c $< -o $@ sampling.o: common/sampling.cpp $(COMMON_H_DEPS) @@ -605,8 +602,8 @@ embedding: examples/embedding/embedding.cpp build-info.h ggml. save-load-state: examples/save-load-state/save-load-state.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS) - $(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2) +server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS) + $(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2) -Wno-cast-qual gguf: examples/gguf/gguf.cpp ggml.o llama.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) diff --git a/Package.swift b/Package.swift index 4ab055b19..5b3bd72ca 100644 --- a/Package.swift +++ b/Package.swift @@ -42,13 +42,12 @@ let package = Package( "llama.cpp", "ggml-alloc.c", "ggml-backend.c", - "k_quants.c", + "ggml-quants.c", ] + additionalSources, resources: resources, publicHeadersPath: "spm-headers", cSettings: [ .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]), - .define("GGML_USE_K_QUANTS"), .define("GGML_USE_ACCELERATE") // NOTE: NEW_LAPACK will required iOS version 16.4+ // We should consider add this in the future when we drop support for iOS 14 diff --git a/README.md b/README.md index 49bb556a8..b56ecaec7 100644 --- a/README.md +++ b/README.md @@ -101,7 +101,7 @@ as the main playground for developing new features for the [ggml](https://github - Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python) - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp) -- Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp), [hlhr202/llama-node](https://github.com/hlhr202/llama-node) +- Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp) - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb) - Rust: [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp) - C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp) diff --git a/build.zig b/build.zig index 0b74cee48..9b58b74ca 100644 --- a/build.zig +++ b/build.zig @@ -116,30 +116,26 @@ pub fn build(b: *std.build.Builder) !void { var make = try Maker.init(b); make.enable_lto = b.option(bool, "lto", "Enable LTO optimization, (default: false)") orelse false; - if (b.option(bool, "k-quants", "Enable K-quants, (default: true)") orelse true) { - try make.addFlag("-DGGML_USE_K_QUANTS"); - const k_quants = make.obj("k_quants", "k_quants.c"); - try make.objs.append(k_quants); - } - const ggml = make.obj("ggml", "ggml.c"); const ggml_alloc = make.obj("ggml-alloc", "ggml-alloc.c"); const ggml_backend = make.obj("ggml-backend", "ggml-backend.c"); + const ggml_quants = make.obj("ggml-quants", "ggml-quants.c"); const llama = make.obj("llama", "llama.cpp"); const common = make.obj("common", "common/common.cpp"); const console = make.obj("console", "common/console.cpp"); const sampling = make.obj("sampling", "common/sampling.cpp"); const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.cpp"); const train = make.obj("train", "common/train.cpp"); + const clip = make.obj("clip", "examples/llava/clip.cpp"); - _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, sampling, console, grammar_parser }); - _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common }); - _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common }); - _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common }); - _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, train }); - _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, train }); + _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, sampling, console, grammar_parser }); + _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common }); + _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common }); + _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common }); + _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, train }); + _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, train }); - const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, sampling, grammar_parser }); + const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, sampling, grammar_parser, clip }); if (server.target.isWindows()) { server.linkSystemLibrary("ws2_32"); } diff --git a/common/common.cpp b/common/common.cpp index 6f979c56d..817710eed 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -233,12 +233,19 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { break; } sparams.top_p = std::stof(argv[i]); + } else if (arg == "--min-p") { + if (++i >= argc) { + invalid_param = true; + break; + } + sparams.min_p = std::stof(argv[i]); } else if (arg == "--temp") { if (++i >= argc) { invalid_param = true; break; } sparams.temp = std::stof(argv[i]); + sparams.temp = std::max(sparams.temp, 0.0f); } else if (arg == "--tfs") { if (++i >= argc) { invalid_param = true; @@ -639,6 +646,7 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { process_escapes(params.prompt); process_escapes(params.input_prefix); process_escapes(params.input_suffix); + process_escapes(sparams.cfg_negative_prompt); for (auto & antiprompt : params.antiprompt) { process_escapes(antiprompt); } @@ -685,6 +693,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" -b N, --batch-size N batch size for prompt processing (default: %d)\n", params.n_batch); printf(" --top-k N top-k sampling (default: %d, 0 = disabled)\n", sparams.top_k); printf(" --top-p N top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)sparams.top_p); + printf(" --min-p N min-p sampling (default: %.1f, 0.0 = disabled)\n", (double)sparams.min_p); printf(" --tfs N tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)sparams.tfs_z); printf(" --typical N locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)sparams.typical_p); printf(" --repeat-last-n N last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", sparams.penalty_last_n); @@ -750,7 +759,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { #endif // GGML_USE_CUBLAS #endif printf(" --verbose-prompt print prompt before generation\n"); - fprintf(stderr, " --simple-io use basic IO for better compatibility in subprocesses and limited consoles\n"); + printf(" --simple-io use basic IO for better compatibility in subprocesses and limited consoles\n"); printf(" --lora FNAME apply LoRA adapter (implies --no-mmap)\n"); printf(" --lora-scaled FNAME S apply LoRA adapter with user defined scaling S (implies --no-mmap)\n"); printf(" --lora-base FNAME optional model to use as a base for the layers modified by the LoRA adapter\n"); @@ -890,15 +899,15 @@ std::tuple llama_init_from_gpt_par } if (params.ignore_eos) { - params.sparams.logit_bias[llama_token_eos(lctx)] = -INFINITY; + params.sparams.logit_bias[llama_token_eos(model)] = -INFINITY; } { LOG("warming up the model with an empty run\n"); - std::vector tmp = { llama_token_bos(lctx), llama_token_eos(lctx), }; + std::vector tmp = { llama_token_bos(model), llama_token_eos(model), }; llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0)); - llama_kv_cache_tokens_rm(lctx, -1, -1); + llama_kv_cache_clear(lctx); llama_reset_timings(lctx); } @@ -951,7 +960,7 @@ std::string llama_token_to_piece(const struct llama_context * ctx, llama_token t } std::string llama_detokenize_spm(llama_context * ctx, const std::vector & tokens) { - const llama_token bos_id = llama_token_bos(ctx); + const llama_token bos_id = llama_token_bos(llama_get_model(ctx)); std::string piece; std::string result; @@ -1196,7 +1205,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l fprintf(stream, "hellaswag: %s # default: false\n", params.hellaswag ? "true" : "false"); fprintf(stream, "hellaswag_tasks: %zu # default: 400\n", params.hellaswag_tasks); - const auto logit_bias_eos = sparams.logit_bias.find(llama_token_eos(lctx)); + const auto logit_bias_eos = sparams.logit_bias.find(llama_token_eos(llama_get_model(lctx))); const bool ignore_eos = logit_bias_eos != sparams.logit_bias.end() && logit_bias_eos->second == -INFINITY; fprintf(stream, "ignore_eos: %s # default: false\n", ignore_eos ? "true" : "false"); @@ -1284,6 +1293,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l fprintf(stream, "threads: %d # default: %d\n", params.n_threads, std::thread::hardware_concurrency()); fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k); fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p); + fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p); fprintf(stream, "typical_p: %f # default: 1.0\n", sparams.typical_p); fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false"); } diff --git a/common/log.h b/common/log.h index 70e7e4ca2..c0e814861 100644 --- a/common/log.h +++ b/common/log.h @@ -97,37 +97,56 @@ #define LOG_TEE_TARGET stderr #endif +// Utility for synchronizing log configuration state +// since std::optional was introduced only in c++17 +enum LogTriState +{ + LogTriStateSame, + LogTriStateFalse, + LogTriStateTrue +}; + // Utility to obtain "pid" like unique process id and use it when creating log files. inline std::string log_get_pid() { - static std::string pid; - if (pid.empty()) - { - // std::this_thread::get_id() is the most portable way of obtaining a "process id" - // it's not the same as "pid" but is unique enough to solve multiple instances - // trying to write to the same log. - std::stringstream ss; - ss << std::this_thread::get_id(); - pid = ss.str(); - } + static std::string pid; + if (pid.empty()) + { + // std::this_thread::get_id() is the most portable way of obtaining a "process id" + // it's not the same as "pid" but is unique enough to solve multiple instances + // trying to write to the same log. + std::stringstream ss; + ss << std::this_thread::get_id(); + pid = ss.str(); + } - return pid; + return pid; } // Utility function for generating log file names with unique id based on thread id. // invocation with log_filename_generator( "llama", "log" ) creates a string "llama..log" // where the number is a runtime id of the current thread. -#define log_filename_generator(log_file_basename, log_file_extension) log_filename_generator_impl(log_file_basename, log_file_extension) +#define log_filename_generator(log_file_basename, log_file_extension) log_filename_generator_impl(LogTriStateSame, log_file_basename, log_file_extension) // INTERNAL, DO NOT USE -inline std::string log_filename_generator_impl(const std::string & log_file_basename, const std::string & log_file_extension) +inline std::string log_filename_generator_impl(LogTriState multilog, const std::string & log_file_basename, const std::string & log_file_extension) { + static bool _multilog = false; + + if (multilog != LogTriStateSame) + { + _multilog = multilog == LogTriStateTrue; + } + std::stringstream buf; buf << log_file_basename; - buf << "."; - buf << log_get_pid(); + if (_multilog) + { + buf << "."; + buf << log_get_pid(); + } buf << "."; buf << log_file_extension; @@ -212,15 +231,6 @@ inline std::string log_filename_generator_impl(const std::string & log_file_base #define LOG_TEE_FLF_VAL ,"" #endif -// Utility for synchronizing log configuration state -// since std::optional was introduced only in c++17 -enum LogTriState -{ - LogTriStateSame, - LogTriStateFalse, - LogTriStateTrue -}; - // INTERNAL, DO NOT USE // USE LOG() INSTEAD // @@ -314,16 +324,23 @@ enum LogTriState #endif // INTERNAL, DO NOT USE -inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTriStateSame, const std::string & filename = LOG_DEFAULT_FILE_NAME, FILE *target = nullptr) +inline FILE *log_handler1_impl(bool change = false, LogTriState append = LogTriStateSame, LogTriState disable = LogTriStateSame, const std::string & filename = LOG_DEFAULT_FILE_NAME, FILE *target = nullptr) { - static bool _initialized{false}; - static bool _disabled{(filename.empty() && target == nullptr)}; + static bool _initialized = false; + static bool _append = false; + static bool _disabled = filename.empty() && target == nullptr; static std::string log_current_filename{filename}; static FILE *log_current_target{target}; static FILE *logfile = nullptr; if (change) { + if (append != LogTriStateSame) + { + _append = append == LogTriStateTrue; + return logfile; + } + if (disable == LogTriStateTrue) { // Disable primary target @@ -376,7 +393,7 @@ inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTri } } - logfile = fopen(filename.c_str(), "w"); + logfile = fopen(filename.c_str(), _append ? "a" : "w"); } if (!logfile) @@ -397,9 +414,9 @@ inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTri } // INTERNAL, DO NOT USE -inline FILE *log_handler2_impl(bool change = false, LogTriState disable = LogTriStateSame, FILE *target = nullptr, const std::string & filename = LOG_DEFAULT_FILE_NAME) +inline FILE *log_handler2_impl(bool change = false, LogTriState append = LogTriStateSame, LogTriState disable = LogTriStateSame, FILE *target = nullptr, const std::string & filename = LOG_DEFAULT_FILE_NAME) { - return log_handler1_impl(change, disable, filename, target); + return log_handler1_impl(change, append, disable, filename, target); } // Disables logs entirely at runtime. @@ -410,7 +427,7 @@ inline FILE *log_handler2_impl(bool change = false, LogTriState disable = LogTri // INTERNAL, DO NOT USE inline FILE *log_disable_impl() { - return log_handler1_impl(true, LogTriStateTrue); + return log_handler1_impl(true, LogTriStateSame, LogTriStateTrue); } // Enables logs at runtime. @@ -419,19 +436,31 @@ inline FILE *log_disable_impl() // INTERNAL, DO NOT USE inline FILE *log_enable_impl() { - return log_handler1_impl(true, LogTriStateFalse); + return log_handler1_impl(true, LogTriStateSame, LogTriStateFalse); } // Sets target fir logs, either by a file name or FILE* pointer (stdout, stderr, or any valid FILE*) #define log_set_target(target) log_set_target_impl(target) // INTERNAL, DO NOT USE -inline FILE *log_set_target_impl(const std::string & filename) { return log_handler1_impl(true, LogTriStateSame, filename); } -inline FILE *log_set_target_impl(FILE *target) { return log_handler2_impl(true, LogTriStateSame, target); } +inline FILE *log_set_target_impl(const std::string & filename) { return log_handler1_impl(true, LogTriStateSame, LogTriStateSame, filename); } +inline FILE *log_set_target_impl(FILE *target) { return log_handler2_impl(true, LogTriStateSame, LogTriStateSame, target); } // INTERNAL, DO NOT USE inline FILE *log_handler() { return log_handler1_impl(); } +// Enable or disable creating separate log files for each run. +// can ONLY be invoked BEFORE first log use. +#define log_multilog(enable) log_filename_generator_impl((enable) ? LogTriStateTrue : LogTriStateFalse, "", "") +// Enable or disable append mode for log file. +// can ONLY be invoked BEFORE first log use. +#define log_append(enable) log_append_impl(enable) +// INTERNAL, DO NOT USE +inline FILE *log_append_impl(bool enable) +{ + return log_handler1_impl(true, enable ? LogTriStateTrue : LogTriStateFalse, LogTriStateSame); +} + inline void log_test() { log_disable(); @@ -493,6 +522,18 @@ inline bool log_param_single_parse(const std::string & param) return true; } + if (param == "--log-new") + { + log_multilog(true); + return true; + } + + if (param == "--log-append") + { + log_append(true); + return true; + } + return false; } @@ -522,7 +563,9 @@ inline void log_print_usage() printf(" --log-disable Disable trace logs\n"); printf(" --log-enable Enable trace logs\n"); printf(" --log-file Specify a log filename (without extension)\n"); - printf(" Log file will be tagged with unique ID and written as \"..log\"\n"); /* */ + printf(" --log-new Create a separate new log file on start. " + "Each log file will have unique name: \"..log\"\n"); + printf(" --log-append Don't truncate the old log file.\n"); } #define log_dump_cmdline(argc, argv) log_dump_cmdline_impl(argc, argv) diff --git a/common/sampling.cpp b/common/sampling.cpp index 6f0af3c4a..1317024c2 100644 --- a/common/sampling.cpp +++ b/common/sampling.cpp @@ -39,6 +39,7 @@ void llama_sampling_free(struct llama_sampling_context * ctx) { void llama_sampling_reset(llama_sampling_context * ctx) { if (ctx->grammar != NULL) { llama_grammar_free(ctx->grammar); + ctx->grammar = NULL; } if (!ctx->parsed_grammar.rules.empty()) { @@ -89,10 +90,10 @@ std::string llama_sampling_print(const llama_sampling_params & params) { snprintf(result, sizeof(result), "\trepeat_last_n = %d, repeat_penalty = %.3f, frequency_penalty = %.3f, presence_penalty = %.3f\n" - "\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, typical_p = %.3f, temp = %.3f\n" + "\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, min_p = %.3f, typical_p = %.3f, temp = %.3f\n" "\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f", params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present, - params.top_k, params.tfs_z, params.top_p, params.typical_p, params.temp, + params.top_k, params.tfs_z, params.top_p, params.min_p, params.typical_p, params.temp, params.mirostat, params.mirostat_eta, params.mirostat_tau); return std::string(result); @@ -110,6 +111,7 @@ llama_token llama_sampling_sample( const float temp = params.temp; const int32_t top_k = params.top_k <= 0 ? n_vocab : params.top_k; const float top_p = params.top_p; + const float min_p = params.min_p; const float tfs_z = params.tfs_z; const float typical_p = params.typical_p; const int32_t penalty_last_n = params.penalty_last_n < 0 ? params.n_prev : params.penalty_last_n; @@ -147,7 +149,7 @@ llama_token llama_sampling_sample( // apply penalties if (!prev.empty()) { - const float nl_logit = logits[llama_token_nl(ctx_main)]; + const float nl_logit = logits[llama_token_nl(llama_get_model(ctx_main))]; llama_sample_repetition_penalties(ctx_main, &cur_p, prev.data() + prev.size() - penalty_last_n, @@ -155,7 +157,7 @@ llama_token llama_sampling_sample( if (!penalize_nl) { for (size_t idx = 0; idx < cur_p.size; idx++) { - if (cur_p.data[idx].id == llama_token_nl(ctx_main)) { + if (cur_p.data[idx].id == llama_token_nl(llama_get_model(ctx_main))) { cur_p.data[idx].logit = nl_logit; break; } @@ -167,8 +169,12 @@ llama_token llama_sampling_sample( llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar); } - if (temp <= 0) { - // greedy sampling + if (temp < 0.0) { + // greedy sampling, with probs + llama_sample_softmax(ctx_main, &cur_p); + id = cur_p.data[0].id; + } else if (temp == 0.0) { + // greedy sampling, no probs id = llama_sample_token_greedy(ctx_main, &cur_p); } else { if (mirostat == 1) { @@ -186,6 +192,7 @@ llama_token llama_sampling_sample( llama_sample_tail_free(ctx_main, &cur_p, tfs_z, min_keep); llama_sample_typical (ctx_main, &cur_p, typical_p, min_keep); llama_sample_top_p (ctx_main, &cur_p, top_p, min_keep); + llama_sample_min_p (ctx_main, &cur_p, min_p, min_keep); llama_sample_temp (ctx_main, &cur_p, temp); id = llama_sample_token(ctx_main, &cur_p); diff --git a/common/sampling.h b/common/sampling.h index 62ea6d4cf..7c9b8dcf2 100644 --- a/common/sampling.h +++ b/common/sampling.h @@ -14,6 +14,7 @@ typedef struct llama_sampling_params { int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens. int32_t top_k = 40; // <= 0 to use vocab size float top_p = 0.95f; // 1.0 = disabled + float min_p = 0.05f; // 0.0 = disabled float tfs_z = 1.00f; // 1.0 = disabled float typical_p = 1.00f; // 1.0 = disabled float temp = 0.80f; // 1.0 = disabled diff --git a/common/train.cpp b/common/train.cpp index 154ca56e5..bc15b7a03 100644 --- a/common/train.cpp +++ b/common/train.cpp @@ -236,8 +236,8 @@ int64_t get_example_targets_batch( int64_t used_samples = 0; ggml_set_f32(target_probs, 0.0f); - llama_token bos = llama_token_bos(lctx); - llama_token eos = llama_token_eos(lctx); + llama_token bos = llama_token_bos(llama_get_model(lctx)); + llama_token eos = llama_token_eos(llama_get_model(lctx)); // printf("%s: example_id=%d n_batch=%d n_train_samples=%zu\n", __func__, example_id, n_batch, n_train_samples); for (int k=0; k= vocab_size} + expected_new_ids = list(range(vocab_size, vocab_size + len(new_tokens))) + actual_new_ids = sorted(new_tokens.keys()) + + if expected_new_ids != actual_new_ids: + raise ValueError(f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}") + + # Token pieces that were added to the base vocabulary. + self.added_tokens_list = [new_tokens[id] for id in actual_new_ids] + self.vocab_size_base = vocab_size + self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) + self.fname_tokenizer = fname_tokenizer self.fname_added_tokens = fname_added_tokens def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: @@ -1163,10 +1166,13 @@ def main(args_in: list[str] | None = None) -> None: vocab: Vocab if args.vocab_only: - assert args.outfile, "need --outfile if using --vocab-only" + if not args.outfile: + raise ValueError("need --outfile if using --vocab-only") # FIXME: Try to respect vocab_dir somehow? vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype) - special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, load_merges = args.vocabtype == 'bpe') + special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, + load_merges = args.vocabtype == 'bpe', + n_vocab = vocab.vocab_size) outfile = args.outfile OutputFile.write_vocab_only(outfile, params, vocab, special_vocab) print(f"Wrote {outfile}") @@ -1178,7 +1184,9 @@ def main(args_in: list[str] | None = None) -> None: vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent vocab = load_vocab(vocab_dir, args.vocabtype) # FIXME: Try to respect vocab_dir somehow? - special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, load_merges = args.vocabtype == 'bpe') + special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, + load_merges = args.vocabtype == 'bpe', + n_vocab = vocab.vocab_size) model = model_plus.model model = convert_model_names(model, params) diff --git a/examples/batched-bench/batched-bench.cpp b/examples/batched-bench/batched-bench.cpp index c552eaa73..533c55c17 100644 --- a/examples/batched-bench/batched-bench.cpp +++ b/examples/batched-bench/batched-bench.cpp @@ -154,6 +154,10 @@ int main(int argc, char ** argv) { } } + LOG_TEE("\n"); + LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, mmq = %d\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq); + LOG_TEE("\n"); + LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP", "TG", "B", "N_KV", "T_PP s", "S_PP t/s", "T_TG s", "S_TG t/s", "T s", "S t/s"); LOG_TEE("|%6s-|-%6s-|-%4s-|-%6s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|\n", "------", "------", "----", "------", "--------", "--------", "--------", "--------", "--------", "--------"); @@ -181,7 +185,7 @@ int main(int argc, char ** argv) { const auto t_pp_start = ggml_time_us(); - llama_kv_cache_tokens_rm(ctx, -1, -1); + llama_kv_cache_clear(ctx); if (!decode_helper(ctx, batch, ctx_params.n_batch)) { LOG_TEE("%s: llama_decode() failed\n", __func__); diff --git a/examples/batched/batched.cpp b/examples/batched/batched.cpp index 155212165..22a4265df 100644 --- a/examples/batched/batched.cpp +++ b/examples/batched/batched.cpp @@ -11,12 +11,19 @@ int main(int argc, char ** argv) { gpt_params params; if (argc == 1 || argv[1][0] == '-') { - printf("usage: %s MODEL_PATH [PROMPT] [PARALLEL]\n" , argv[0]); + printf("usage: %s MODEL_PATH [PROMPT] [PARALLEL] [LEN] [NGL]\n" , argv[0]); return 1 ; } + // number of parallel batches int n_parallel = 1; + // total length of the sequences including the prompt + int n_len = 32; + + // number of layers to offload to the GPU + int n_gpu_layers = 0; + if (argc >= 2) { params.model = argv[1]; } @@ -29,13 +36,18 @@ int main(int argc, char ** argv) { n_parallel = std::atoi(argv[3]); } + if (argc >= 5) { + n_len = std::atoi(argv[4]); + } + + if (argc >= 6) { + n_gpu_layers = std::atoi(argv[5]); + } + if (params.prompt.empty()) { params.prompt = "Hello my name is"; } - // total length of the sequences including the prompt - const int n_len = 32; - // init LLM llama_backend_init(params.numa); @@ -44,7 +56,7 @@ int main(int argc, char ** argv) { llama_model_params model_params = llama_model_default_params(); - // model_params.n_gpu_layers = 99; // offload all layers to the GPU + model_params.n_gpu_layers = n_gpu_layers; llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params); @@ -175,7 +187,7 @@ int main(int argc, char ** argv) { //const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p); // is it an end of stream? -> mark the stream as finished - if (new_token_id == llama_token_eos(ctx) || n_cur == n_len) { + if (new_token_id == llama_token_eos(model) || n_cur == n_len) { i_batch[i] = -1; LOG_TEE("\n"); if (n_parallel > 1) { diff --git a/examples/beam-search/beam-search.cpp b/examples/beam-search/beam-search.cpp index f078ab8a8..679b382e1 100644 --- a/examples/beam-search/beam-search.cpp +++ b/examples/beam-search/beam-search.cpp @@ -47,7 +47,7 @@ struct beam_search_callback_data { // In this case, end-of-beam (eob) is equivalent to end-of-sentence (eos) but this need not always be the same. // For example, eob can be flagged due to maximum token length, stop words, etc. static bool is_at_eob(const beam_search_callback_data & callback_data, const llama_token * tokens, size_t n_tokens) { - return n_tokens && tokens[n_tokens-1] == llama_token_eos(callback_data.ctx); + return n_tokens && tokens[n_tokens-1] == llama_token_eos(llama_get_model(callback_data.ctx)); } // Function matching type llama_beam_search_callback_fn_t. diff --git a/examples/finetune/finetune.cpp b/examples/finetune/finetune.cpp index 35824cd2d..60c7faa79 100644 --- a/examples/finetune/finetune.cpp +++ b/examples/finetune/finetune.cpp @@ -652,7 +652,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs( GGML_ASSERT(tokens_input->type == GGML_TYPE_I32); auto add_to_f32 = [] (struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) { - if (ggml_is_quantized(a->type)) { + if (ggml_is_quantized(a->type) || a->type == GGML_TYPE_F16) { return ggml_add_cast(ctx, a, b, GGML_TYPE_F32); } else if (a->type == GGML_TYPE_F32) { return ggml_add(ctx, a, b); @@ -1459,6 +1459,17 @@ static bool train_params_parse(int argc, char ** argv, struct train_params * par } params->n_rank_w3 = std::stoi(argv[i]); params->custom_n_rank_w3 = true; + } else if (arg == "--gpu-layers" || arg == "-ngl" || arg == "--n-gpu-layers") { + if (++i >= argc) { + invalid_param = true; + break; + } +#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD + params->common.n_gpu_layers = std::stoi(argv[i]); +#else + fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n"); + fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n"); +#endif } else { fprintf(stderr, "error: unknown argument: %s\n", arg.c_str()); train_print_usage(argc, argv, &default_params); @@ -1545,6 +1556,7 @@ int main(int argc, char ** argv) { srand(params.common.seed); struct llama_model_params llama_mparams = llama_model_default_params(); + llama_mparams.n_gpu_layers = params.common.n_gpu_layers; llama_mparams.vocab_only = false; printf("%s: model base = '%s'\n", __func__, params.fn_model_base); diff --git a/examples/finetune/finetune.sh b/examples/finetune/finetune.sh new file mode 100644 index 000000000..079bfa113 --- /dev/null +++ b/examples/finetune/finetune.sh @@ -0,0 +1,34 @@ +#!/bin/bash +cd `dirname $0` +cd ../.. + +EXE="./finetune" + +if [[ ! $LLAMA_MODEL_DIR ]]; then LLAMA_MODEL_DIR="./models"; fi +if [[ ! $LLAMA_TRAINING_DIR ]]; then LLAMA_TRAINING_DIR="."; fi + +# MODEL="$LLAMA_MODEL_DIR/openllama-3b-v2-q8_0.gguf" # This is the model the readme uses. +MODEL="$LLAMA_MODEL_DIR/openllama-3b-v2.gguf" # An f16 model. Note in this case with "-g", you get an f32-format .BIN file that isn't yet supported if you use it with "main --lora" with GPU inferencing. + +while getopts "dg" opt; do + case $opt in + d) + DEBUGGER="gdb --args" + ;; + g) + EXE="./build/bin/Release/finetune" + GPUARG="--gpu-layers 25" + ;; + esac +done + +$DEBUGGER $EXE \ + --model-base $MODEL \ + $GPUARG \ + --checkpoint-in chk-ol3b-shakespeare-LATEST.gguf \ + --checkpoint-out chk-ol3b-shakespeare-ITERATION.gguf \ + --lora-out lora-ol3b-shakespeare-ITERATION.bin \ + --train-data "$LLAMA_TRAINING_DIR\shakespeare.txt" \ + --save-every 10 \ + --threads 10 --adam-iter 30 --batch 4 --ctx 64 \ + --use-checkpointing diff --git a/examples/infill/infill.cpp b/examples/infill/infill.cpp index 6331335e3..9c52b7bba 100644 --- a/examples/infill/infill.cpp +++ b/examples/infill/infill.cpp @@ -246,14 +246,14 @@ int main(int argc, char ** argv) { if (suff_rm_leading_spc && inp_sfx[0] == space_token) { inp_sfx.erase(inp_sfx.begin()); } - inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(ctx)); + inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model)); if (add_bos) { - inp_pfx.insert(inp_pfx.begin(), llama_token_bos(ctx)); + inp_pfx.insert(inp_pfx.begin(), llama_token_bos(model)); } - inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(ctx)); + inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model)); embd_inp = inp_pfx; embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end()); - embd_inp.push_back(llama_token_middle(ctx)); + embd_inp.push_back(llama_token_middle(model)); LOG("prefix: \"%s\"\n", log_tostr(params.input_prefix)); LOG("suffix: \"%s\"\n", log_tostr(params.input_suffix)); @@ -261,7 +261,7 @@ int main(int argc, char ** argv) { // Should not run without any tokens if (embd_inp.empty()) { - embd_inp.push_back(llama_token_bos(ctx)); + embd_inp.push_back(llama_token_bos(model)); LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str()); } @@ -577,10 +577,10 @@ int main(int argc, char ** argv) { if ((int) embd_inp.size() <= n_consumed) { // deal with eot token in infill mode - if ((llama_sampling_last(ctx_sampling) == llama_token_eot(ctx) || is_interacting) && params.interactive){ + if ((llama_sampling_last(ctx_sampling) == llama_token_eot(model) || is_interacting) && params.interactive){ if(is_interacting && !params.interactive_first) { // print an eot token - printf("%s", llama_token_to_piece(ctx, llama_token_eot(ctx)).c_str()); + printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str()); } fflush(stdout); printf("\n"); @@ -627,14 +627,14 @@ int main(int argc, char ** argv) { if (suff_rm_leading_spc && inp_sfx[0] == space_token) { inp_sfx.erase(inp_sfx.begin()); } - inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(ctx)); + inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model)); if (add_bos) { - inp_pfx.insert(inp_pfx.begin(), llama_token_bos(ctx)); + inp_pfx.insert(inp_pfx.begin(), llama_token_bos(model)); } - inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(ctx)); + inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model)); embd_inp = inp_pfx; embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end()); - embd_inp.push_back(llama_token_middle(ctx)); + embd_inp.push_back(llama_token_middle(model)); embd.clear(); embd_guidance.clear(); n_remain = params.n_predict; @@ -644,7 +644,7 @@ int main(int argc, char ** argv) { is_interacting = false; } // deal with end of text token in interactive mode - else if (llama_sampling_last(ctx_sampling) == llama_token_eos(ctx)) { + else if (llama_sampling_last(ctx_sampling) == llama_token_eos(model)) { LOG("found EOS token\n"); if (params.interactive) { @@ -661,7 +661,7 @@ int main(int argc, char ** argv) { if (params.input_prefix_bos) { LOG("adding input prefix BOS token\n"); - embd_inp.push_back(llama_token_bos(ctx)); + embd_inp.push_back(llama_token_bos(model)); } std::string buffer; @@ -724,7 +724,7 @@ int main(int argc, char ** argv) { } // end of text token - if (!embd.empty() && embd.back() == llama_token_eos(ctx) && !params.interactive) { + if (!embd.empty() && embd.back() == llama_token_eos(model) && !params.interactive) { break; } @@ -736,7 +736,7 @@ int main(int argc, char ** argv) { } } if (!params.interactive && n_remain <= 0) { - printf("%s", llama_token_to_piece(ctx, llama_token_eot(ctx)).c_str()); + printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str()); fflush(stdout); } diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp index a04115c96..780398184 100644 --- a/examples/llama-bench/llama-bench.cpp +++ b/examples/llama-bench/llama-bench.cpp @@ -933,7 +933,7 @@ struct sql_printer : public printer { }; static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_batch, int n_threads) { - std::vector tokens(n_batch, llama_token_bos(ctx)); + std::vector tokens(n_batch, llama_token_bos(llama_get_model(ctx))); int n_processed = 0; llama_set_n_threads(ctx, n_threads, n_threads); @@ -946,7 +946,7 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_bat } static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) { - llama_token token = llama_token_bos(ctx); + llama_token token = llama_token_bos(llama_get_model(ctx)); llama_set_n_threads(ctx, n_threads, n_threads); @@ -1037,7 +1037,7 @@ int main(int argc, char ** argv) { test t(inst, lmodel, ctx); - llama_kv_cache_tokens_rm(ctx, -1, -1); + llama_kv_cache_clear(ctx); // warmup run if (t.n_prompt > 0) { @@ -1048,7 +1048,7 @@ int main(int argc, char ** argv) { } for (int i = 0; i < params.reps; i++) { - llama_kv_cache_tokens_rm(ctx, -1, -1); + llama_kv_cache_clear(ctx); uint64_t t_start = get_time_ns(); if (t.n_prompt > 0) { diff --git a/examples/llava/CMakeLists.txt b/examples/llava/CMakeLists.txt index d02e6ab46..2d7979ecd 100644 --- a/examples/llava/CMakeLists.txt +++ b/examples/llava/CMakeLists.txt @@ -1,7 +1,7 @@ set(TARGET clip) add_library(${TARGET} clip.cpp clip.h) install(TARGETS ${TARGET} LIBRARY) -target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) +target_link_libraries(${TARGET} PRIVATE common ggml ${CMAKE_THREAD_LIBS_INIT}) target_compile_features(${TARGET} PRIVATE cxx_std_11) if (NOT MSVC) target_compile_options(${TARGET} PRIVATE -Wno-cast-qual) # stb_image.h diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp index 1ae9077b4..61932e659 100644 --- a/examples/llava/clip.cpp +++ b/examples/llava/clip.cpp @@ -610,8 +610,8 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { int idx_mean = get_key_idx(ctx, KEY_IMAGE_MEAN); int idx_std = get_key_idx(ctx, KEY_IMAGE_STD); for (int i = 0; i < 3; ++i) { - new_clip->image_mean[i] = *((float *)gguf_get_arr_data(ctx, idx_mean)); - new_clip->image_std[i] = *((float *)gguf_get_arr_data(ctx, idx_std)); + new_clip->image_mean[i] = *((const float *)gguf_get_arr_data(ctx, idx_mean)); + new_clip->image_std[i] = *((const float *)gguf_get_arr_data(ctx, idx_std)); } if (verbosity >= 2) { diff --git a/examples/llava/llava-utils.h b/examples/llava/llava-utils.h index 45b2b1ad3..320c71967 100644 --- a/examples/llava/llava-utils.h +++ b/examples/llava/llava-utils.h @@ -137,7 +137,7 @@ inline llama_token sample_id(llama_context * ctx_llama, gpt_params & params) { inline const char * sample(struct llama_context * ctx_llama, gpt_params & params, int * n_past) { int id = sample_id(ctx_llama, params); static std::string ret; - if (id == llama_token_eos(ctx_llama)) { + if (id == llama_token_eos(llama_get_model(ctx_llama))) { ret = ""; } else { ret = llama_token_to_piece(ctx_llama, id); diff --git a/examples/main-cmake-pkg/CMakeLists.txt b/examples/main-cmake-pkg/CMakeLists.txt index 908131884..cb00edbbb 100644 --- a/examples/main-cmake-pkg/CMakeLists.txt +++ b/examples/main-cmake-pkg/CMakeLists.txt @@ -16,6 +16,8 @@ add_library(common OBJECT ${_common_path}/console.cpp ${_common_path}/grammar-parser.h ${_common_path}/grammar-parser.cpp + ${_common_path}/sampling.h + ${_common_path}/sampling.cpp ) # WARNING: because build-info.h is auto-generated, it will only diff --git a/examples/main/README.md b/examples/main/README.md index a9561c383..a3428b487 100644 --- a/examples/main/README.md +++ b/examples/main/README.md @@ -208,6 +208,14 @@ Top-p sampling, also known as nucleus sampling, is another text generation metho Example usage: `--top-p 0.95` +### Min P Sampling + +- `--min-p N`: Sets a minimum base probability threshold for token selection (default: 0.05). + +The Min-P sampling method was designed as an alternative to Top-P, and aims to ensure a balance of quality and variety. The parameter *p* represents the minimum probability for a token to be considered, relative to the probability of the most likely token. For example, with *p*=0.05 and the most likely token having a probability of 0.9, logits with a value less than 0.045 are filtered out. + +Example usage: `--min-p 0.05` + ### Tail Free Sampling (TFS) - `--tfs N`: Enable tail free sampling with parameter z (default: 1.0, 1.0 = disabled). diff --git a/examples/main/main.cpp b/examples/main/main.cpp index db5309afe..8a43b6ab8 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -248,7 +248,7 @@ int main(int argc, char ** argv) { // Should not run without any tokens if (embd_inp.empty()) { - embd_inp.push_back(llama_token_bos(ctx)); + embd_inp.push_back(llama_token_bos(model)); LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str()); } @@ -298,7 +298,7 @@ int main(int argc, char ** argv) { } // remove any "future" tokens that we might have inherited from the previous session - llama_kv_cache_tokens_rm(ctx, n_matching_session_tokens, -1); + llama_kv_cache_seq_rm(ctx, -1, n_matching_session_tokens, -1); } LOGLN( @@ -693,7 +693,7 @@ int main(int argc, char ** argv) { } // deal with end of text token in interactive mode - if (llama_sampling_last(ctx_sampling) == llama_token_eos(ctx)) { + if (llama_sampling_last(ctx_sampling) == llama_token_eos(model)) { LOG("found EOS token\n"); if (params.interactive) { @@ -720,7 +720,7 @@ int main(int argc, char ** argv) { if (params.input_prefix_bos) { LOG("adding input prefix BOS token\n"); - embd_inp.push_back(llama_token_bos(ctx)); + embd_inp.push_back(llama_token_bos(model)); } std::string buffer; @@ -761,6 +761,9 @@ int main(int argc, char ** argv) { n_consumed = embd_inp.size(); embd_inp.insert(embd_inp.end(), inp_pfx.begin(), inp_pfx.end()); } + if (params.escape) { + process_escapes(buffer); + } const auto line_pfx = ::llama_tokenize(ctx, params.input_prefix, false, true); const auto line_inp = ::llama_tokenize(ctx, buffer, false, false); @@ -801,7 +804,7 @@ int main(int argc, char ** argv) { } // end of text token - if (!embd.empty() && embd.back() == llama_token_eos(ctx) && !(params.instruct || params.interactive)) { + if (!embd.empty() && embd.back() == llama_token_eos(model) && !(params.instruct || params.interactive)) { LOG_TEE(" [end of text]\n"); break; } diff --git a/examples/parallel/parallel.cpp b/examples/parallel/parallel.cpp index eb64adef8..9a0b9c183 100644 --- a/examples/parallel/parallel.cpp +++ b/examples/parallel/parallel.cpp @@ -347,7 +347,7 @@ int main(int argc, char ** argv) { // client.id, client.seq_id, id, client.n_decoded, client.i_batch, token_str.c_str()); if (client.n_decoded > 2 && - (id == llama_token_eos(ctx) || + (id == llama_token_eos(model) || (params.n_predict > 0 && client.n_decoded + client.n_prompt >= params.n_predict) || client.response.find("User:") != std::string::npos || client.response.find('\n') != std::string::npos)) { diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index 7d0038bd4..bd2c73d87 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -210,7 +210,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & const auto t_start = std::chrono::high_resolution_clock::now(); // clear the KV cache - llama_kv_cache_tokens_rm(ctx, -1, -1); + llama_kv_cache_clear(ctx); for (int j = 0; j < num_batches; ++j) { const int batch_start = start + j * n_batch; @@ -227,7 +227,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & // add BOS token for the first batch of each chunk if (add_bos && j == 0) { - tokens[batch_start] = llama_token_bos(ctx); + tokens[batch_start] = llama_token_bos(llama_get_model(ctx)); } const auto batch_logits = llama_get_logits(ctx); @@ -339,7 +339,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par const auto t_start = std::chrono::high_resolution_clock::now(); // clear the KV cache - llama_kv_cache_tokens_rm(ctx, -1, -1); + llama_kv_cache_clear(ctx); for (int j = 0; j < num_batches; ++j) { const int batch_start = start + j * n_batch; @@ -350,7 +350,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par // add BOS token for the first batch of each chunk if (add_bos && j == 0) { - tokens[batch_start] = llama_token_bos(ctx); + tokens[batch_start] = llama_token_bos(llama_get_model(ctx)); } if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) { @@ -573,7 +573,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) { } // clear the KV cache - llama_kv_cache_tokens_rm(ctx, -1, -1); + llama_kv_cache_clear(ctx); auto logits = hellaswag_evaluate_tokens(ctx, query_embd, 0, params.n_batch, n_vocab); if (logits.empty()) { diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index c7dd0d894..be0b2fe1e 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -18,7 +18,6 @@ static const std::vector QUANT_OPTIONS = { { "Q4_1", LLAMA_FTYPE_MOSTLY_Q4_1, " 3.90G, +0.1585 ppl @ LLaMA-v1-7B", }, { "Q5_0", LLAMA_FTYPE_MOSTLY_Q5_0, " 4.33G, +0.0683 ppl @ LLaMA-v1-7B", }, { "Q5_1", LLAMA_FTYPE_MOSTLY_Q5_1, " 4.70G, +0.0349 ppl @ LLaMA-v1-7B", }, -#ifdef GGML_USE_K_QUANTS { "Q2_K", LLAMA_FTYPE_MOSTLY_Q2_K, " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", }, { "Q3_K", LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" }, { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", }, @@ -31,7 +30,6 @@ static const std::vector QUANT_OPTIONS = { { "Q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S, " 4.33G, +0.0400 ppl @ LLaMA-v1-7B", }, { "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 4.45G, +0.0122 ppl @ LLaMA-v1-7B", }, { "Q6_K", LLAMA_FTYPE_MOSTLY_Q6_K, " 5.15G, -0.0008 ppl @ LLaMA-v1-7B", }, -#endif { "Q8_0", LLAMA_FTYPE_MOSTLY_Q8_0, " 6.70G, +0.0004 ppl @ LLaMA-v1-7B", }, { "F16", LLAMA_FTYPE_MOSTLY_F16, "13.00G @ 7B", }, { "F32", LLAMA_FTYPE_ALL_F32, "26.00G @ 7B", }, @@ -70,13 +68,14 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp } // usage: -// ./quantize [--allow-requantize] [--leave-output-tensor] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads] +// ./quantize [--allow-requantize] [--leave-output-tensor] [--pure] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads] // [[noreturn]] static void usage(const char * executable) { - printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable); + printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable); printf(" --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n"); printf(" --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n"); + printf(" --pure: Disable k-quant mixtures and quantize all tensors to the same type\n"); printf("\nAllowed quantization types:\n"); for (auto & it : QUANT_OPTIONS) { if (it.name != "COPY") { @@ -103,6 +102,8 @@ int main(int argc, char ** argv) { params.quantize_output_tensor = false; } else if (strcmp(argv[arg_idx], "--allow-requantize") == 0) { params.allow_requantize = true; + } else if (strcmp(argv[arg_idx], "--pure") == 0) { + params.pure = true; } else { usage(argv[0]); } diff --git a/examples/server/CMakeLists.txt b/examples/server/CMakeLists.txt index 3782f9b80..a23ddcc55 100644 --- a/examples/server/CMakeLists.txt +++ b/examples/server/CMakeLists.txt @@ -6,7 +6,7 @@ install(TARGETS ${TARGET} RUNTIME) target_compile_definitions(${TARGET} PRIVATE SERVER_VERBOSE=$ ) -target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) +target_link_libraries(${TARGET} PRIVATE common llama clip ${CMAKE_THREAD_LIBS_INIT}) if (WIN32) TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32) endif() diff --git a/examples/server/README.md b/examples/server/README.md index 9f0ace3d7..715007735 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -24,6 +24,10 @@ Command line options: - `--port`: Set the port to listen. Default: `8080`. - `--path`: path from which to serve static files (default examples/server/public) - `--embedding`: Enable embedding extraction, Default: disabled. +- `-np N`, `--parallel N`: Set the number of slots for process requests (default: 1) +- `-cb`, `--cont-batching`: enable continuous batching (a.k.a dynamic batching) (default: disabled) +- `-spf FNAME`, `--system-prompt-file FNAME` Set a file to load "a system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime) +- `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA. ## Build @@ -158,6 +162,8 @@ node index.js `n_probs`: If greater than 0, the response also contains the probabilities of top N tokens for each generated token (default: 0) + `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `prompt`. You can determine the place of the image in the prompt as in the following: `USER:[img-12]Describe the image in detail.\nASSISTANT:` In this case, `[img-12]` will be replaced by the embeddings of the image id 12 in the following `image_data` array: `{..., "image_data": [{"data": "", "id": 12}]}`. Use `image_data` only with multimodal models, e.g., LLaVA. + *Result JSON:* Note: When using streaming mode (`stream`) only `content` and `stop` will be returned until end of completion. @@ -188,6 +194,12 @@ node index.js `truncated`: Boolean indicating if the context size was exceeded during generation, i.e. the number of tokens provided in the prompt (`tokens_evaluated`) plus tokens generated (`tokens predicted`) exceeded the context size (`n_ctx`) + `slot_id`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot (default: -1) + + `cache_prompt`: Save the prompt and generation for avoid reprocess entire prompt if a part of this isn't change (default: false) + + `system_prompt`: Change the system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime) + - **POST** `/tokenize`: Tokenize a given text. *Options:* @@ -218,8 +230,32 @@ node index.js It also accepts all the options of `/completion` except `stream` and `prompt`. +- **GET** `/props`: Return the required assistant name and anti-prompt to generate the prompt in case you have specified a system prompt for all slots. + ## More examples +### Change system prompt on runtime + +To use the server example to serve multiple chat-type clients while keeping the same system prompt, you can utilize the option `system_prompt` to achieve that. This only needs to be done once to establish it. + +`prompt`: Specify a context that you want all connecting clients to respect. + +`anti_prompt`: Specify the word you want to use to instruct the model to stop. This must be sent to each client through the `/props` endpoint. + +`assistant_name`: The bot's name is necessary for each customer to generate the prompt. This must be sent to each client through the `/props` endpoint. + +```json +{ + "system_prompt": { + "prompt": "Transcript of a never ending dialog, where the User interacts with an Assistant.\nThe Assistant is helpful, kind, honest, good at writing, and never fails to answer the User's requests immediately and with precision.\nUser: Recommend a nice restaurant in the area.\nAssistant: I recommend the restaurant \"The Golden Duck\". It is a 5 star restaurant with a great view of the city. The food is delicious and the service is excellent. The prices are reasonable and the portions are generous. The restaurant is located at 123 Main Street, New York, NY 10001. The phone number is (212) 555-1234. The hours are Monday through Friday from 11:00 am to 10:00 pm. The restaurant is closed on Saturdays and Sundays.\nUser: Who is Richard Feynman?\nAssistant: Richard Feynman was an American physicist who is best known for his work in quantum mechanics and particle physics. He was awarded the Nobel Prize in Physics in 1965 for his contributions to the development of quantum electrodynamics. He was a popular lecturer and author, and he wrote several books, including \"Surely You're Joking, Mr. Feynman!\" and \"What Do You Care What Other People Think?\".\nUser:", + "anti_prompt": "User:", + "assistant_name": "Assistant:" + } +} +``` + +**NOTE**: You can do this automatically when starting the server by simply creating a .json file with these options and using the CLI option `-spf FNAME` or `--system-prompt-file FNAME`. + ### Interactive mode Check the sample in [chat.mjs](chat.mjs). diff --git a/examples/server/api_like_OAI.py b/examples/server/api_like_OAI.py index 14d2dcf65..313e1a965 100755 --- a/examples/server/api_like_OAI.py +++ b/examples/server/api_like_OAI.py @@ -8,6 +8,7 @@ import json app = Flask(__name__) +slot_id = -1 parser = argparse.ArgumentParser(description="An example of using server.cpp with a similar API to OAI. It must be used together with server.cpp.") parser.add_argument("--chat-prompt", type=str, help="the top prompt in chat completions(default: 'A chat between a curious user and an artificial intelligence assistant. The assistant follows the given rules no matter what.\\n')", default='A chat between a curious user and an artificial intelligence assistant. The assistant follows the given rules no matter what.\\n') @@ -77,7 +78,8 @@ def make_postData(body, chat=False, stream=False): if(is_present(body, "stop")): postData["stop"] += body["stop"] postData["n_keep"] = -1 postData["stream"] = stream - + postData["cache_prompt"] = True + postData["slot_id"] = slot_id return postData def make_resData(data, chat=False, promptToken=[]): @@ -128,6 +130,7 @@ def make_resData_stream(data, chat=False, time_now = 0, start=False): } ] } + slot_id = data["slot_id"] if (chat): if (start): resData["choices"][0]["delta"] = { diff --git a/examples/server/chat.mjs b/examples/server/chat.mjs index 87f4d2926..219ebb51a 100644 --- a/examples/server/chat.mjs +++ b/examples/server/chat.mjs @@ -7,6 +7,11 @@ const args = process.argv.slice(2); const grammarJsonSchemaFile = args.find( (_, index) => args[index - 1] === "--grammar-json-schema" ); + +const no_cached_prompt = args.find( + (_, index) => args[index - 1] === "--no-cache-prompt" +) ?? "false"; + const grammarFile = args.find((_, index) => args[index - 1] === "--grammar"); // Example usage: function,arguments @@ -30,6 +35,9 @@ if (grammarFile) { grammar = readFileSync(grammarFile, 'utf-8') } +// for cached prompt +let slot_id = -1; + const API_URL = 'http://127.0.0.1:8080' const chat = [ @@ -76,6 +84,8 @@ async function chat_completion(question) { top_p: 0.9, n_keep: n_keep, n_predict: 256, + cache_prompt: no_cached_prompt === "false", + slot_id: slot_id, stop: ["\n### Human:"], // stop completion after generating this grammar, stream: true, @@ -92,6 +102,7 @@ async function chat_completion(question) { const t = Buffer.from(chunk).toString('utf8') if (t.startsWith('data: ')) { const message = JSON.parse(t.substring(6)) + slot_id = message.slot_id answer += message.content process.stdout.write(message.content) if (message.stop) { diff --git a/examples/server/index.html.hpp b/examples/server/index.html.hpp index 58e3387d1..5d3bdfbdd 100644 --- a/examples/server/index.html.hpp +++ b/examples/server/index.html.hpp @@ -186,351 +186,369 @@ unsigned char index_html[] = { 0x64, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x23, 0x32, 0x32, 0x32, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x23, 0x64, 0x64, 0x64, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x64, 0x65, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x6e, - 0x74, 0x2d, 0x66, 0x61, 0x6d, 0x69, 0x6c, 0x79, 0x3a, 0x20, 0x6d, 0x6f, - 0x6e, 0x6f, 0x73, 0x70, 0x61, 0x63, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x70, 0x61, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x3a, 0x20, - 0x30, 0x2e, 0x31, 0x65, 0x6d, 0x20, 0x30, 0x2e, 0x33, 0x65, 0x6d, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x6f, 0x72, 0x64, 0x65, - 0x72, 0x2d, 0x72, 0x61, 0x64, 0x69, 0x75, 0x73, 0x3a, 0x20, 0x33, 0x70, - 0x78, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x6c, - 0x61, 0x62, 0x65, 0x6c, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x6d, 0x61, 0x72, 0x67, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x35, - 0x65, 0x6d, 0x20, 0x30, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x64, 0x69, 0x73, 0x70, 0x6c, 0x61, 0x79, 0x3a, 0x20, 0x62, 0x6c, 0x6f, - 0x63, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x64, + 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, + 0x6e, 0x74, 0x2d, 0x66, 0x61, 0x6d, 0x69, 0x6c, 0x79, 0x3a, 0x20, 0x6d, + 0x6f, 0x6e, 0x6f, 0x73, 0x70, 0x61, 0x63, 0x65, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x3a, + 0x20, 0x30, 0x2e, 0x31, 0x65, 0x6d, 0x20, 0x30, 0x2e, 0x33, 0x65, 0x6d, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x6f, 0x72, 0x64, + 0x65, 0x72, 0x2d, 0x72, 0x61, 0x64, 0x69, 0x75, 0x73, 0x3a, 0x20, 0x33, + 0x70, 0x78, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, - 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x2e, 0x73, 0x6c, 0x69, 0x6d, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x61, 0x72, 0x67, 0x69, - 0x6e, 0x3a, 0x20, 0x30, 0x20, 0x30, 0x2e, 0x35, 0x65, 0x6d, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x69, 0x73, 0x70, 0x6c, 0x61, - 0x79, 0x3a, 0x20, 0x69, 0x6e, 0x6c, 0x69, 0x6e, 0x65, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x68, 0x65, - 0x61, 0x64, 0x65, 0x72, 0x2c, 0x20, 0x66, 0x6f, 0x6f, 0x74, 0x65, 0x72, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, 0x78, - 0x74, 0x2d, 0x61, 0x6c, 0x69, 0x67, 0x6e, 0x3a, 0x20, 0x63, 0x65, 0x6e, - 0x74, 0x65, 0x72, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x6f, 0x74, 0x65, 0x72, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x6e, 0x74, 0x2d, - 0x73, 0x69, 0x7a, 0x65, 0x3a, 0x20, 0x38, 0x30, 0x25, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, - 0x23, 0x38, 0x38, 0x38, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x6d, 0x61, 0x72, 0x67, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, + 0x35, 0x65, 0x6d, 0x20, 0x30, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x64, 0x69, 0x73, 0x70, 0x6c, 0x61, 0x79, 0x3a, 0x20, 0x62, 0x6c, + 0x6f, 0x63, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, + 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x2e, 0x73, 0x6c, 0x69, 0x6d, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x61, 0x72, 0x67, + 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x20, 0x30, 0x2e, 0x35, 0x65, 0x6d, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x69, 0x73, 0x70, 0x6c, + 0x61, 0x79, 0x3a, 0x20, 0x69, 0x6e, 0x6c, 0x69, 0x6e, 0x65, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x68, + 0x65, 0x61, 0x64, 0x65, 0x72, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, + 0x6f, 0x6f, 0x74, 0x65, 0x72, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x74, 0x65, 0x78, 0x74, 0x2d, 0x61, 0x6c, 0x69, 0x67, 0x6e, + 0x3a, 0x20, 0x63, 0x65, 0x6e, 0x74, 0x65, 0x72, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x6f, + 0x74, 0x65, 0x72, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x66, 0x6f, 0x6e, 0x74, 0x2d, 0x73, 0x69, 0x7a, 0x65, 0x3a, 0x20, 0x38, + 0x30, 0x25, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x23, 0x38, 0x38, 0x38, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x6d, + 0x6f, 0x64, 0x65, 0x2d, 0x63, 0x68, 0x61, 0x74, 0x20, 0x74, 0x65, 0x78, + 0x74, 0x61, 0x72, 0x65, 0x61, 0x5b, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x70, + 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x5d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x68, 0x65, 0x69, 0x67, 0x68, 0x74, 0x3a, 0x20, 0x34, + 0x2e, 0x35, 0x65, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x6d, 0x6f, 0x64, 0x65, 0x2d, 0x63, - 0x68, 0x61, 0x74, 0x20, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, - 0x5b, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, - 0x5d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x65, - 0x69, 0x67, 0x68, 0x74, 0x3a, 0x20, 0x34, 0x2e, 0x35, 0x65, 0x6d, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x2e, 0x6d, 0x6f, 0x64, 0x65, 0x2d, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, - 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, - 0x61, 0x5b, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x70, 0x72, 0x6f, 0x6d, 0x70, - 0x74, 0x5d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, - 0x65, 0x69, 0x67, 0x68, 0x74, 0x3a, 0x20, 0x31, 0x30, 0x65, 0x6d, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x40, 0x6b, 0x65, 0x79, 0x66, 0x72, 0x61, 0x6d, 0x65, 0x73, 0x20, - 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x2d, 0x62, 0x67, 0x2d, 0x77, - 0x69, 0x70, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x74, 0x65, + 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x5b, 0x6e, 0x61, 0x6d, 0x65, 0x3d, + 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x5d, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x68, 0x65, 0x69, 0x67, 0x68, 0x74, 0x3a, 0x20, + 0x31, 0x30, 0x65, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x40, 0x6b, 0x65, 0x79, 0x66, 0x72, + 0x61, 0x6d, 0x65, 0x73, 0x20, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, + 0x2d, 0x62, 0x67, 0x2d, 0x77, 0x69, 0x70, 0x65, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x30, 0x25, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, + 0x6f, 0x75, 0x6e, 0x64, 0x2d, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, + 0x6e, 0x3a, 0x20, 0x30, 0x25, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x31, 0x30, 0x30, 0x25, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64, 0x2d, - 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x3a, 0x20, 0x30, 0x25, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x31, 0x30, 0x30, 0x25, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, - 0x72, 0x6f, 0x75, 0x6e, 0x64, 0x2d, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, - 0x6f, 0x6e, 0x3a, 0x20, 0x31, 0x30, 0x30, 0x25, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, - 0x67, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2d, 0x2d, + 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x3a, 0x20, 0x31, 0x30, + 0x30, 0x25, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2e, + 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, + 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, 0x31, 0x3a, 0x20, 0x23, + 0x65, 0x65, 0x65, 0x65, 0x65, 0x65, 0x30, 0x30, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, + 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, 0x32, 0x3a, 0x20, 0x23, + 0x65, 0x65, 0x65, 0x65, 0x65, 0x65, 0x66, 0x66, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, + 0x6e, 0x64, 0x2d, 0x73, 0x69, 0x7a, 0x65, 0x3a, 0x20, 0x35, 0x30, 0x25, + 0x20, 0x31, 0x30, 0x30, 0x25, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64, 0x2d, + 0x69, 0x6d, 0x61, 0x67, 0x65, 0x3a, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x61, + 0x72, 0x2d, 0x67, 0x72, 0x61, 0x64, 0x69, 0x65, 0x6e, 0x74, 0x28, 0x39, + 0x30, 0x64, 0x65, 0x67, 0x2c, 0x20, 0x76, 0x61, 0x72, 0x28, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, - 0x72, 0x2d, 0x31, 0x3a, 0x20, 0x23, 0x65, 0x65, 0x65, 0x65, 0x65, 0x65, - 0x30, 0x30, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2d, 0x2d, + 0x72, 0x2d, 0x31, 0x29, 0x2c, 0x20, 0x76, 0x61, 0x72, 0x28, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, - 0x72, 0x2d, 0x32, 0x3a, 0x20, 0x23, 0x65, 0x65, 0x65, 0x65, 0x65, 0x65, - 0x66, 0x66, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, - 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64, 0x2d, 0x73, 0x69, 0x7a, - 0x65, 0x3a, 0x20, 0x35, 0x30, 0x25, 0x20, 0x31, 0x30, 0x30, 0x25, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, - 0x72, 0x6f, 0x75, 0x6e, 0x64, 0x2d, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x3a, - 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x61, 0x72, 0x2d, 0x67, 0x72, 0x61, 0x64, - 0x69, 0x65, 0x6e, 0x74, 0x28, 0x39, 0x30, 0x64, 0x65, 0x67, 0x2c, 0x20, - 0x76, 0x61, 0x72, 0x28, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, - 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, 0x31, 0x29, 0x2c, 0x20, - 0x76, 0x61, 0x72, 0x28, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, - 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, 0x32, 0x29, 0x2c, 0x20, - 0x76, 0x61, 0x72, 0x28, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, - 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, 0x31, 0x29, 0x29, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x61, 0x6e, 0x69, 0x6d, 0x61, - 0x74, 0x69, 0x6f, 0x6e, 0x3a, 0x20, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, - 0x67, 0x2d, 0x62, 0x67, 0x2d, 0x77, 0x69, 0x70, 0x65, 0x20, 0x32, 0x73, - 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x61, 0x72, 0x20, 0x69, 0x6e, 0x66, 0x69, - 0x6e, 0x69, 0x74, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x40, 0x6d, 0x65, 0x64, 0x69, 0x61, 0x20, - 0x28, 0x70, 0x72, 0x65, 0x66, 0x65, 0x72, 0x73, 0x2d, 0x63, 0x6f, 0x6c, - 0x6f, 0x72, 0x2d, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x65, 0x3a, 0x20, 0x64, - 0x61, 0x72, 0x6b, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x2e, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2d, 0x2d, 0x6c, 0x6f, - 0x61, 0x64, 0x69, 0x6e, 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, - 0x31, 0x3a, 0x20, 0x23, 0x32, 0x32, 0x32, 0x32, 0x32, 0x32, 0x30, 0x30, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2d, 0x2d, + 0x72, 0x2d, 0x32, 0x29, 0x2c, 0x20, 0x76, 0x61, 0x72, 0x28, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, - 0x72, 0x2d, 0x32, 0x3a, 0x20, 0x23, 0x32, 0x32, 0x32, 0x32, 0x32, 0x32, - 0x66, 0x66, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x70, 0x6f, 0x70, 0x6f, 0x76, - 0x65, 0x72, 0x2d, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, - 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, - 0x72, 0x3a, 0x20, 0x62, 0x6c, 0x61, 0x63, 0x6b, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3e, 0x0a, 0x0a, - 0x20, 0x20, 0x3c, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x20, 0x74, 0x79, - 0x70, 0x65, 0x3d, 0x22, 0x6d, 0x6f, 0x64, 0x75, 0x6c, 0x65, 0x22, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x20, - 0x7b, 0x43, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x2c, 0x20, - 0x63, 0x6f, 0x6d, 0x70, 0x75, 0x74, 0x65, 0x64, 0x2c, 0x20, 0x68, 0x2c, - 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x2c, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, - 0x72, 0x2c, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x20, 0x75, - 0x73, 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x2c, 0x20, 0x75, 0x73, - 0x65, 0x52, 0x65, 0x66, 0x2c, 0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, - 0x6e, 0x61, 0x6c, 0x7d, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, - 0x69, 0x6e, 0x64, 0x65, 0x78, 0x2e, 0x6a, 0x73, 0x27, 0x3b, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x7b, - 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x7d, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, - 0x27, 0x2f, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, - 0x2e, 0x6a, 0x73, 0x27, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, - 0x70, 0x6f, 0x72, 0x74, 0x20, 0x7b, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, - 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x7d, 0x20, 0x66, - 0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x6a, 0x73, 0x6f, 0x6e, 0x2d, 0x73, - 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2d, 0x74, 0x6f, 0x2d, 0x67, 0x72, 0x61, - 0x6d, 0x6d, 0x61, 0x72, 0x2e, 0x6d, 0x6a, 0x73, 0x27, 0x3b, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x65, - 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, - 0x61, 0x6c, 0x28, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, - 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x3a, 0x20, 0x22, 0x54, 0x68, 0x69, 0x73, - 0x20, 0x69, 0x73, 0x20, 0x61, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, - 0x73, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x62, 0x65, 0x74, 0x77, 0x65, - 0x65, 0x6e, 0x20, 0x55, 0x73, 0x65, 0x72, 0x20, 0x61, 0x6e, 0x64, 0x20, - 0x4c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x61, 0x20, 0x66, 0x72, 0x69, - 0x65, 0x6e, 0x64, 0x6c, 0x79, 0x20, 0x63, 0x68, 0x61, 0x74, 0x62, 0x6f, - 0x74, 0x2e, 0x20, 0x4c, 0x6c, 0x61, 0x6d, 0x61, 0x20, 0x69, 0x73, 0x20, - 0x68, 0x65, 0x6c, 0x70, 0x66, 0x75, 0x6c, 0x2c, 0x20, 0x6b, 0x69, 0x6e, - 0x64, 0x2c, 0x20, 0x68, 0x6f, 0x6e, 0x65, 0x73, 0x74, 0x2c, 0x20, 0x67, - 0x6f, 0x6f, 0x64, 0x20, 0x61, 0x74, 0x20, 0x77, 0x72, 0x69, 0x74, 0x69, - 0x6e, 0x67, 0x2c, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x6e, 0x65, 0x76, 0x65, - 0x72, 0x20, 0x66, 0x61, 0x69, 0x6c, 0x73, 0x20, 0x74, 0x6f, 0x20, 0x61, - 0x6e, 0x73, 0x77, 0x65, 0x72, 0x20, 0x61, 0x6e, 0x79, 0x20, 0x72, 0x65, - 0x71, 0x75, 0x65, 0x73, 0x74, 0x73, 0x20, 0x69, 0x6d, 0x6d, 0x65, 0x64, - 0x69, 0x61, 0x74, 0x65, 0x6c, 0x79, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x77, - 0x69, 0x74, 0x68, 0x20, 0x70, 0x72, 0x65, 0x63, 0x69, 0x73, 0x69, 0x6f, - 0x6e, 0x2e, 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3a, 0x20, 0x22, 0x7b, 0x7b, - 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x7d, 0x7d, 0x5c, 0x6e, 0x5c, 0x6e, - 0x7b, 0x7b, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x7d, 0x7d, 0x5c, - 0x6e, 0x7b, 0x7b, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x7d, 0x3a, 0x22, 0x2c, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, - 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3a, 0x20, - 0x22, 0x7b, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x7d, 0x3a, 0x20, 0x7b, - 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x7d, 0x22, 0x2c, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, - 0x63, 0x72, 0x69, 0x70, 0x74, 0x3a, 0x20, 0x5b, 0x5d, 0x2c, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3a, 0x20, 0x22, - 0x63, 0x68, 0x61, 0x74, 0x22, 0x2c, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x22, - 0x63, 0x68, 0x61, 0x74, 0x22, 0x20, 0x7c, 0x20, 0x22, 0x63, 0x6f, 0x6d, - 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x68, 0x61, 0x72, 0x3a, 0x20, 0x22, 0x4c, 0x6c, - 0x61, 0x6d, 0x61, 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x75, 0x73, 0x65, 0x72, 0x3a, 0x20, 0x22, 0x55, 0x73, 0x65, 0x72, 0x22, - 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x61, 0x72, 0x61, - 0x6d, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6e, 0x5f, 0x70, 0x72, - 0x65, 0x64, 0x69, 0x63, 0x74, 0x3a, 0x20, 0x34, 0x30, 0x30, 0x2c, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x65, 0x72, - 0x61, 0x74, 0x75, 0x72, 0x65, 0x3a, 0x20, 0x30, 0x2e, 0x37, 0x2c, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, - 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x3a, 0x20, 0x32, 0x35, 0x36, - 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, - 0x61, 0x62, 0x6c, 0x65, 0x20, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, - 0x2c, 0x20, 0x2d, 0x31, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, - 0x78, 0x74, 0x20, 0x73, 0x69, 0x7a, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x70, 0x65, 0x6e, - 0x61, 0x6c, 0x74, 0x79, 0x3a, 0x20, 0x31, 0x2e, 0x31, 0x38, 0x2c, 0x20, - 0x2f, 0x2f, 0x20, 0x31, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, + 0x72, 0x2d, 0x31, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x61, 0x6e, 0x69, 0x6d, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x3a, 0x20, + 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x2d, 0x62, 0x67, 0x2d, 0x77, + 0x69, 0x70, 0x65, 0x20, 0x32, 0x73, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x61, + 0x72, 0x20, 0x69, 0x6e, 0x66, 0x69, 0x6e, 0x69, 0x74, 0x65, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x40, + 0x6d, 0x65, 0x64, 0x69, 0x61, 0x20, 0x28, 0x70, 0x72, 0x65, 0x66, 0x65, + 0x72, 0x73, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, 0x73, 0x63, 0x68, + 0x65, 0x6d, 0x65, 0x3a, 0x20, 0x64, 0x61, 0x72, 0x6b, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x6c, 0x6f, 0x61, 0x64, + 0x69, 0x6e, 0x67, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x2d, + 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, 0x31, 0x3a, 0x20, 0x23, 0x32, 0x32, + 0x32, 0x32, 0x32, 0x32, 0x30, 0x30, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, + 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, 0x32, 0x3a, 0x20, 0x23, + 0x32, 0x32, 0x32, 0x32, 0x32, 0x32, 0x66, 0x66, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x2e, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x2d, 0x63, 0x6f, + 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, + 0x6e, 0x64, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x62, 0x6c, + 0x61, 0x63, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x3c, 0x2f, 0x73, + 0x74, 0x79, 0x6c, 0x65, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x3c, 0x73, 0x63, + 0x72, 0x69, 0x70, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x6d, + 0x6f, 0x64, 0x75, 0x6c, 0x65, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x2c, 0x20, 0x68, 0x2c, 0x20, + 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x20, 0x65, 0x66, 0x66, 0x65, + 0x63, 0x74, 0x2c, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x75, 0x74, 0x65, 0x64, + 0x2c, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x2c, 0x20, 0x75, 0x73, + 0x65, 0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x20, 0x75, 0x73, 0x65, + 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x2c, 0x20, 0x75, 0x73, 0x65, 0x52, + 0x65, 0x66, 0x2c, 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, + 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x66, 0x72, 0x6f, 0x6d, + 0x20, 0x27, 0x2f, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x2e, 0x6a, 0x73, 0x27, + 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, + 0x74, 0x20, 0x7b, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x20, 0x7d, 0x20, + 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x63, 0x6f, 0x6d, 0x70, 0x6c, + 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x6a, 0x73, 0x27, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x7b, 0x20, + 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, + 0x74, 0x65, 0x72, 0x20, 0x7d, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x27, + 0x2f, 0x6a, 0x73, 0x6f, 0x6e, 0x2d, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, + 0x2d, 0x74, 0x6f, 0x2d, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x2e, + 0x6d, 0x6a, 0x73, 0x27, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, + 0x74, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x5f, 0x69, + 0x6d, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x76, 0x61, 0x72, 0x20, 0x73, 0x6c, + 0x6f, 0x74, 0x5f, 0x69, 0x64, 0x20, 0x3d, 0x20, 0x2d, 0x31, 0x3b, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, + 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, + 0x6e, 0x61, 0x6c, 0x28, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x3a, 0x20, 0x22, 0x54, 0x68, 0x69, + 0x73, 0x20, 0x69, 0x73, 0x20, 0x61, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, + 0x72, 0x73, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x62, 0x65, 0x74, 0x77, + 0x65, 0x65, 0x6e, 0x20, 0x55, 0x73, 0x65, 0x72, 0x20, 0x61, 0x6e, 0x64, + 0x20, 0x4c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x61, 0x20, 0x66, 0x72, + 0x69, 0x65, 0x6e, 0x64, 0x6c, 0x79, 0x20, 0x63, 0x68, 0x61, 0x74, 0x62, + 0x6f, 0x74, 0x2e, 0x20, 0x4c, 0x6c, 0x61, 0x6d, 0x61, 0x20, 0x69, 0x73, + 0x20, 0x68, 0x65, 0x6c, 0x70, 0x66, 0x75, 0x6c, 0x2c, 0x20, 0x6b, 0x69, + 0x6e, 0x64, 0x2c, 0x20, 0x68, 0x6f, 0x6e, 0x65, 0x73, 0x74, 0x2c, 0x20, + 0x67, 0x6f, 0x6f, 0x64, 0x20, 0x61, 0x74, 0x20, 0x77, 0x72, 0x69, 0x74, + 0x69, 0x6e, 0x67, 0x2c, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x6e, 0x65, 0x76, + 0x65, 0x72, 0x20, 0x66, 0x61, 0x69, 0x6c, 0x73, 0x20, 0x74, 0x6f, 0x20, + 0x61, 0x6e, 0x73, 0x77, 0x65, 0x72, 0x20, 0x61, 0x6e, 0x79, 0x20, 0x72, + 0x65, 0x71, 0x75, 0x65, 0x73, 0x74, 0x73, 0x20, 0x69, 0x6d, 0x6d, 0x65, + 0x64, 0x69, 0x61, 0x74, 0x65, 0x6c, 0x79, 0x20, 0x61, 0x6e, 0x64, 0x20, + 0x77, 0x69, 0x74, 0x68, 0x20, 0x70, 0x72, 0x65, 0x63, 0x69, 0x73, 0x69, + 0x6f, 0x6e, 0x2e, 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3a, 0x20, 0x22, 0x7b, + 0x7b, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x7d, 0x7d, 0x5c, 0x6e, 0x5c, + 0x6e, 0x7b, 0x7b, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x7d, 0x7d, + 0x5c, 0x6e, 0x7b, 0x7b, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x7d, 0x3a, 0x22, + 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x69, 0x73, 0x74, + 0x6f, 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3a, + 0x20, 0x22, 0x7b, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x7d, 0x3a, 0x20, + 0x7b, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x7d, 0x22, + 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, + 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x3a, 0x20, 0x5b, 0x5d, 0x2c, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3a, 0x20, + 0x22, 0x63, 0x68, 0x61, 0x74, 0x22, 0x2c, 0x20, 0x20, 0x2f, 0x2f, 0x20, + 0x22, 0x63, 0x68, 0x61, 0x74, 0x22, 0x20, 0x7c, 0x20, 0x22, 0x63, 0x6f, + 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x68, 0x61, 0x72, 0x3a, 0x20, 0x22, 0x4c, + 0x6c, 0x61, 0x6d, 0x61, 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x75, 0x73, 0x65, 0x72, 0x3a, 0x20, 0x22, 0x55, 0x73, 0x65, 0x72, + 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x61, + 0x67, 0x65, 0x5f, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x3a, + 0x20, 0x27, 0x27, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x61, + 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, + 0x6c, 0x28, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6e, 0x5f, + 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x3a, 0x20, 0x34, 0x30, 0x30, + 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, 0x6d, 0x70, + 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x3a, 0x20, 0x30, 0x2e, 0x37, + 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x70, 0x65, + 0x61, 0x74, 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x3a, 0x20, 0x32, + 0x35, 0x36, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x30, 0x20, 0x3d, 0x20, 0x64, + 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x20, 0x70, 0x65, 0x6e, 0x61, 0x6c, + 0x74, 0x79, 0x2c, 0x20, 0x2d, 0x31, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6e, + 0x74, 0x65, 0x78, 0x74, 0x20, 0x73, 0x69, 0x7a, 0x65, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x70, + 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x3a, 0x20, 0x31, 0x2e, 0x31, 0x38, + 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x31, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, + 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x74, 0x6f, 0x70, 0x5f, 0x6b, 0x3a, 0x20, 0x34, 0x30, 0x2c, + 0x20, 0x2f, 0x2f, 0x20, 0x3c, 0x3d, 0x20, 0x30, 0x20, 0x74, 0x6f, 0x20, + 0x75, 0x73, 0x65, 0x20, 0x76, 0x6f, 0x63, 0x61, 0x62, 0x20, 0x73, 0x69, + 0x7a, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x6f, 0x70, + 0x5f, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x35, 0x2c, 0x20, 0x2f, 0x2f, 0x20, + 0x31, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, + 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x66, 0x73, + 0x5f, 0x7a, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, + 0x31, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, + 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x79, 0x70, + 0x69, 0x63, 0x61, 0x6c, 0x5f, 0x70, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, + 0x20, 0x2f, 0x2f, 0x20, 0x31, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, + 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x70, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, 0x65, + 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, + 0x2f, 0x2f, 0x20, 0x30, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x74, 0x6f, 0x70, 0x5f, 0x6b, 0x3a, 0x20, 0x34, 0x30, 0x2c, 0x20, 0x2f, - 0x2f, 0x20, 0x3c, 0x3d, 0x20, 0x30, 0x20, 0x74, 0x6f, 0x20, 0x75, 0x73, - 0x65, 0x20, 0x76, 0x6f, 0x63, 0x61, 0x62, 0x20, 0x73, 0x69, 0x7a, 0x65, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x6f, 0x70, 0x5f, 0x70, - 0x3a, 0x20, 0x30, 0x2e, 0x35, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x31, 0x2e, - 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x66, 0x73, 0x5f, 0x7a, - 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x31, 0x2e, - 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x79, 0x70, 0x69, 0x63, - 0x61, 0x6c, 0x5f, 0x70, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x2f, - 0x2f, 0x20, 0x31, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, - 0x62, 0x6c, 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, - 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, 0x65, 0x6e, 0x61, - 0x6c, 0x74, 0x79, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, - 0x20, 0x30, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, - 0x6c, 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x72, - 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, 0x5f, 0x70, 0x65, 0x6e, 0x61, - 0x6c, 0x74, 0x79, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, - 0x20, 0x30, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, - 0x6c, 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x69, - 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x2f, - 0x2f, 0x20, 0x30, 0x2f, 0x31, 0x2f, 0x32, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x74, - 0x61, 0x75, 0x3a, 0x20, 0x35, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x74, 0x61, - 0x72, 0x67, 0x65, 0x74, 0x20, 0x65, 0x6e, 0x74, 0x72, 0x6f, 0x70, 0x79, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x69, 0x72, 0x6f, 0x73, - 0x74, 0x61, 0x74, 0x5f, 0x65, 0x74, 0x61, 0x3a, 0x20, 0x30, 0x2e, 0x31, - 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x6c, 0x65, 0x61, 0x72, 0x6e, 0x69, 0x6e, - 0x67, 0x20, 0x72, 0x61, 0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x3a, 0x20, 0x27, 0x27, - 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6e, 0x5f, 0x70, 0x72, - 0x6f, 0x62, 0x73, 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x6e, - 0x6f, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, - 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, - 0x65, 0x73, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x2f, 0x2a, 0x20, 0x53, 0x54, 0x41, 0x52, 0x54, 0x3a, - 0x20, 0x53, 0x75, 0x70, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x66, 0x6f, 0x72, - 0x20, 0x73, 0x74, 0x6f, 0x72, 0x69, 0x6e, 0x67, 0x20, 0x70, 0x72, 0x6f, - 0x6d, 0x70, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, - 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x65, - 0x74, 0x65, 0x72, 0x73, 0x20, 0x69, 0x6e, 0x20, 0x62, 0x6f, 0x72, 0x77, - 0x73, 0x65, 0x72, 0x20, 0x4c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, - 0x72, 0x61, 0x67, 0x65, 0x20, 0x2a, 0x2f, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, - 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x74, 0x6f, - 0x72, 0x61, 0x67, 0x65, 0x4b, 0x65, 0x79, 0x20, 0x3d, 0x20, 0x22, 0x6c, - 0x6c, 0x61, 0x6d, 0x61, 0x63, 0x70, 0x70, 0x5f, 0x73, 0x65, 0x72, 0x76, - 0x65, 0x72, 0x5f, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, - 0x72, 0x61, 0x67, 0x65, 0x22, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6c, 0x6f, 0x63, + 0x66, 0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, 0x5f, 0x70, 0x65, + 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, + 0x2f, 0x2f, 0x20, 0x30, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, + 0x61, 0x62, 0x6c, 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x3a, 0x20, 0x30, 0x2c, + 0x20, 0x2f, 0x2f, 0x20, 0x30, 0x2f, 0x31, 0x2f, 0x32, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, + 0x5f, 0x74, 0x61, 0x75, 0x3a, 0x20, 0x35, 0x2c, 0x20, 0x2f, 0x2f, 0x20, + 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x65, 0x6e, 0x74, 0x72, 0x6f, + 0x70, 0x79, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x69, 0x72, + 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x65, 0x74, 0x61, 0x3a, 0x20, 0x30, + 0x2e, 0x31, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x6c, 0x65, 0x61, 0x72, 0x6e, + 0x69, 0x6e, 0x67, 0x20, 0x72, 0x61, 0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x3a, 0x20, + 0x27, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6e, 0x5f, + 0x70, 0x72, 0x6f, 0x62, 0x73, 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x2f, 0x2f, + 0x20, 0x6e, 0x6f, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, + 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, + 0x74, 0x69, 0x65, 0x73, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x64, 0x61, 0x74, 0x61, 0x3a, 0x20, + 0x5b, 0x5d, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x61, + 0x63, 0x68, 0x65, 0x5f, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x3a, 0x20, + 0x74, 0x72, 0x75, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2a, 0x20, 0x53, 0x54, 0x41, 0x52, + 0x54, 0x3a, 0x20, 0x53, 0x75, 0x70, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x66, + 0x6f, 0x72, 0x20, 0x73, 0x74, 0x6f, 0x72, 0x69, 0x6e, 0x67, 0x20, 0x70, + 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, + 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x70, 0x61, 0x72, 0x61, + 0x6d, 0x65, 0x74, 0x65, 0x72, 0x73, 0x20, 0x69, 0x6e, 0x20, 0x62, 0x6f, + 0x72, 0x77, 0x73, 0x65, 0x72, 0x20, 0x4c, 0x6f, 0x63, 0x61, 0x6c, 0x53, + 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x20, 0x2a, 0x2f, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, - 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d, 0x4f, 0x62, - 0x6a, 0x65, 0x63, 0x74, 0x28, 0x74, 0x61, 0x67, 0x2c, 0x20, 0x63, 0x6f, - 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, - 0x61, 0x67, 0x65, 0x2e, 0x73, 0x65, 0x74, 0x49, 0x74, 0x65, 0x6d, 0x28, + 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x4b, 0x65, 0x79, 0x20, 0x3d, 0x20, + 0x22, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x63, 0x70, 0x70, 0x5f, 0x73, 0x65, + 0x72, 0x76, 0x65, 0x72, 0x5f, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, + 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x22, 0x3b, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6c, + 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, + 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d, + 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x74, 0x61, 0x67, 0x2c, 0x20, + 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, + 0x6f, 0x72, 0x61, 0x67, 0x65, 0x2e, 0x73, 0x65, 0x74, 0x49, 0x74, 0x65, + 0x6d, 0x28, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, + 0x61, 0x67, 0x65, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x4b, + 0x65, 0x79, 0x20, 0x2b, 0x20, 0x27, 0x2f, 0x27, 0x20, 0x2b, 0x20, 0x74, + 0x61, 0x67, 0x2c, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x73, 0x74, 0x72, + 0x69, 0x6e, 0x67, 0x69, 0x66, 0x79, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, + 0x6e, 0x74, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, + 0x6e, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, + 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, + 0x72, 0x6f, 0x6d, 0x52, 0x61, 0x77, 0x54, 0x65, 0x78, 0x74, 0x28, 0x74, + 0x61, 0x67, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, + 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x2e, 0x73, 0x65, + 0x74, 0x49, 0x74, 0x65, 0x6d, 0x28, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, + 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x74, 0x6f, 0x72, + 0x61, 0x67, 0x65, 0x4b, 0x65, 0x79, 0x20, 0x2b, 0x20, 0x27, 0x2f, 0x27, + 0x20, 0x2b, 0x20, 0x74, 0x61, 0x67, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, + 0x65, 0x6e, 0x74, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, + 0x6e, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, + 0x61, 0x67, 0x65, 0x5f, 0x67, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x41, + 0x73, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x74, 0x61, 0x67, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x69, 0x74, 0x65, 0x6d, 0x20, 0x3d, 0x20, 0x6c, 0x6f, + 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x2e, 0x67, + 0x65, 0x74, 0x49, 0x74, 0x65, 0x6d, 0x28, 0x6c, 0x6f, 0x63, 0x61, 0x6c, + 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x74, 0x6f, + 0x72, 0x61, 0x67, 0x65, 0x4b, 0x65, 0x79, 0x20, 0x2b, 0x20, 0x27, 0x2f, + 0x27, 0x20, 0x2b, 0x20, 0x74, 0x61, 0x67, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x69, 0x74, 0x65, + 0x6d, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x75, 0x6c, 0x6c, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, + 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x4a, 0x53, 0x4f, 0x4e, + 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, 0x28, 0x69, 0x74, 0x65, 0x6d, 0x29, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, + 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, + 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x67, 0x65, 0x74, 0x44, + 0x61, 0x74, 0x61, 0x41, 0x73, 0x52, 0x61, 0x77, 0x54, 0x65, 0x78, 0x74, + 0x28, 0x74, 0x61, 0x67, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x74, 0x65, 0x6d, + 0x20, 0x3d, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, + 0x61, 0x67, 0x65, 0x2e, 0x67, 0x65, 0x74, 0x49, 0x74, 0x65, 0x6d, 0x28, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x4b, 0x65, 0x79, 0x20, 0x2b, 0x20, 0x27, 0x2f, 0x27, 0x20, 0x2b, 0x20, 0x74, 0x61, 0x67, - 0x2c, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x73, 0x74, 0x72, 0x69, 0x6e, - 0x67, 0x69, 0x66, 0x79, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, - 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, - 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, - 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, - 0x6d, 0x52, 0x61, 0x77, 0x54, 0x65, 0x78, 0x74, 0x28, 0x74, 0x61, 0x67, - 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, - 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x2e, 0x73, 0x65, 0x74, 0x49, - 0x74, 0x65, 0x6d, 0x28, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, - 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, - 0x65, 0x4b, 0x65, 0x79, 0x20, 0x2b, 0x20, 0x27, 0x2f, 0x27, 0x20, 0x2b, - 0x20, 0x74, 0x61, 0x67, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, - 0x74, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, - 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, - 0x65, 0x5f, 0x67, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x41, 0x73, 0x4f, - 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x74, 0x61, 0x67, 0x29, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x69, 0x74, 0x65, 0x6d, 0x20, 0x3d, 0x20, 0x6c, 0x6f, 0x63, 0x61, - 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x2e, 0x67, 0x65, 0x74, - 0x49, 0x74, 0x65, 0x6d, 0x28, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, - 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, - 0x67, 0x65, 0x4b, 0x65, 0x79, 0x20, 0x2b, 0x20, 0x27, 0x2f, 0x27, 0x20, - 0x2b, 0x20, 0x74, 0x61, 0x67, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x69, 0x74, 0x65, 0x6d, 0x29, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, - 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, - 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x70, - 0x61, 0x72, 0x73, 0x65, 0x28, 0x69, 0x74, 0x65, 0x6d, 0x29, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, - 0x69, 0x6f, 0x6e, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, - 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x67, 0x65, 0x74, 0x44, 0x61, 0x74, - 0x61, 0x41, 0x73, 0x52, 0x61, 0x77, 0x54, 0x65, 0x78, 0x74, 0x28, 0x74, - 0x61, 0x67, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x74, 0x65, 0x6d, 0x20, 0x3d, - 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, - 0x65, 0x2e, 0x67, 0x65, 0x74, 0x49, 0x74, 0x65, 0x6d, 0x28, 0x6c, 0x6f, - 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, - 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x4b, 0x65, 0x79, 0x20, 0x2b, - 0x20, 0x27, 0x2f, 0x27, 0x20, 0x2b, 0x20, 0x74, 0x61, 0x67, 0x29, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, - 0x69, 0x74, 0x65, 0x6d, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, - 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x69, - 0x74, 0x65, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x2f, 0x2f, 0x20, 0x63, 0x72, 0x65, 0x61, 0x74, 0x65, 0x20, 0x61, 0x20, - 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x20, 0x66, 0x6f, - 0x72, 0x20, 0x75, 0x73, 0x65, 0x72, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x73, 0x65, 0x74, - 0x74, 0x69, 0x6e, 0x67, 0x73, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, - 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, - 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x7b, 0x7d, 0x29, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, - 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x73, 0x69, - 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x7b, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, - 0x20, 0x27, 0x27, 0x2c, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x3a, 0x20, 0x7b, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, - 0x3a, 0x20, 0x7b, 0x7d, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, - 0x3a, 0x20, 0x7b, 0x7d, 0x20, 0x7d, 0x20, 0x7d, 0x29, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x6c, 0x65, 0x74, 0x27, 0x73, 0x20, - 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, - 0x6c, 0x79, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, + 0x28, 0x21, 0x69, 0x74, 0x65, 0x6d, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, + 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, + 0x20, 0x69, 0x74, 0x65, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x63, 0x72, 0x65, 0x61, 0x74, 0x65, 0x20, + 0x61, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x20, + 0x66, 0x6f, 0x72, 0x20, 0x75, 0x73, 0x65, 0x72, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x73, - 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x69, 0x66, 0x20, 0x74, - 0x68, 0x65, 0x72, 0x65, 0x20, 0x61, 0x72, 0x65, 0x20, 0x61, 0x6e, 0x79, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x75, 0x73, 0x65, 0x72, - 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, - 0x6e, 0x64, 0x20, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, - 0x61, 0x72, 0x65, 0x20, 0x73, 0x74, 0x6f, 0x72, 0x65, 0x64, 0x20, 0x69, - 0x6e, 0x20, 0x6f, 0x6e, 0x65, 0x20, 0x6f, 0x62, 0x6a, 0x65, 0x63, 0x74, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x69, 0x6e, 0x20, 0x66, - 0x6f, 0x72, 0x6d, 0x20, 0x6f, 0x66, 0x20, 0x7b, 0x20, 0x22, 0x74, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x6e, 0x61, 0x6d, 0x65, 0x22, 0x3a, - 0x20, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x64, 0x61, - 0x74, 0x61, 0x22, 0x20, 0x7d, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x7b, 0x20, - 0x22, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x74, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x6e, 0x61, 0x6d, 0x65, 0x22, 0x3a, 0x22, - 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x64, 0x61, 0x74, 0x61, - 0x22, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, - 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x49, 0x6d, - 0x70, 0x6f, 0x72, 0x74, 0x69, 0x6e, 0x67, 0x20, 0x73, 0x61, 0x76, 0x65, - 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x27, - 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x69, - 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, - 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x67, 0x65, 0x74, - 0x44, 0x61, 0x74, 0x61, 0x41, 0x73, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, - 0x28, 0x27, 0x75, 0x73, 0x65, 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x73, 0x27, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x69, 0x66, 0x20, 0x28, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, - 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x29, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x61, + 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, + 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x73, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x7b, + 0x7d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, + 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, + 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x7b, 0x20, 0x6e, 0x61, 0x6d, + 0x65, 0x3a, 0x20, 0x27, 0x27, 0x2c, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, + 0x61, 0x74, 0x65, 0x3a, 0x20, 0x7b, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, + 0x6f, 0x6e, 0x3a, 0x20, 0x7b, 0x7d, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, + 0x6d, 0x73, 0x3a, 0x20, 0x7b, 0x7d, 0x20, 0x7d, 0x20, 0x7d, 0x29, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x6c, 0x65, 0x74, 0x27, + 0x73, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x6c, 0x6f, 0x63, + 0x61, 0x6c, 0x6c, 0x79, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, + 0x20, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x69, 0x66, + 0x20, 0x74, 0x68, 0x65, 0x72, 0x65, 0x20, 0x61, 0x72, 0x65, 0x20, 0x61, + 0x6e, 0x79, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x75, 0x73, + 0x65, 0x72, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, + 0x20, 0x61, 0x6e, 0x64, 0x20, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, + 0x73, 0x20, 0x61, 0x72, 0x65, 0x20, 0x73, 0x74, 0x6f, 0x72, 0x65, 0x64, + 0x20, 0x69, 0x6e, 0x20, 0x6f, 0x6e, 0x65, 0x20, 0x6f, 0x62, 0x6a, 0x65, + 0x63, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x69, 0x6e, + 0x20, 0x66, 0x6f, 0x72, 0x6d, 0x20, 0x6f, 0x66, 0x20, 0x7b, 0x20, 0x22, + 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x6e, 0x61, 0x6d, 0x65, + 0x22, 0x3a, 0x20, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x64, 0x61, 0x74, 0x61, 0x22, 0x20, 0x7d, 0x20, 0x61, 0x6e, 0x64, 0x20, + 0x7b, 0x20, 0x22, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x74, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x6e, 0x61, 0x6d, 0x65, 0x22, + 0x3a, 0x22, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x64, 0x61, + 0x74, 0x61, 0x22, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, + 0x49, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x69, 0x6e, 0x67, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, - 0x73, 0x20, 0x77, 0x65, 0x72, 0x65, 0x20, 0x73, 0x75, 0x63, 0x63, 0x65, - 0x73, 0x73, 0x66, 0x75, 0x6c, 0x79, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, - 0x74, 0x65, 0x64, 0x2e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, - 0x27, 0x50, 0x72, 0x6f, 0x63, 0x65, 0x73, 0x73, 0x69, 0x6e, 0x67, 0x20, + 0x73, 0x27, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, + 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x54, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x6c, 0x6f, 0x63, + 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x67, + 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x41, 0x73, 0x4f, 0x62, 0x6a, 0x65, + 0x63, 0x74, 0x28, 0x27, 0x75, 0x73, 0x65, 0x72, 0x5f, 0x74, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x27, 0x29, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, + 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x75, 0x70, 0x64, 0x61, - 0x74, 0x69, 0x6e, 0x67, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, - 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x27, 0x29, 0x0a, + 0x74, 0x65, 0x73, 0x20, 0x77, 0x65, 0x72, 0x65, 0x20, 0x73, 0x75, 0x63, + 0x63, 0x65, 0x73, 0x73, 0x66, 0x75, 0x6c, 0x79, 0x20, 0x69, 0x6d, 0x70, + 0x6f, 0x72, 0x74, 0x65, 0x64, 0x2e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, + 0x67, 0x28, 0x27, 0x50, 0x72, 0x6f, 0x63, 0x65, 0x73, 0x73, 0x69, 0x6e, + 0x67, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, + 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x75, 0x70, + 0x64, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, + 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x27, + 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x72, 0x61, + 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, + 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, + 0x64, 0x61, 0x74, 0x61, 0x3a, 0x20, 0x5b, 0x5d, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, @@ -608,515 +626,639 @@ unsigned char index_html[] = { 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, + 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, + 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, + 0x65, 0x64, 0x3a, 0x20, 0x27, 0x27, 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x74, 0x2e, 0x64, 0x61, 0x74, + 0x61, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x20, 0x3d, 0x20, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x70, 0x61, - 0x72, 0x61, 0x6d, 0x73, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, - 0x6e, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, - 0x61, 0x75, 0x6c, 0x74, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, - 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, - 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, - 0x65, 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, - 0x74, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, - 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x41, 0x70, - 0x70, 0x6c, 0x79, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, - 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x2c, 0x20, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x64, 0x61, 0x74, 0x61, + 0x3a, 0x20, 0x5b, 0x5d, 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x4c, 0x6f, 0x61, 0x64, 0x41, 0x6e, 0x64, 0x41, - 0x70, 0x70, 0x6c, 0x79, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, - 0x64, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x2f, 0x2f, 0x20, 0x67, 0x65, 0x74, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, - 0x61, 0x76, 0x65, 0x64, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x75, 0x73, - 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6c, 0x61, - 0x73, 0x74, 0x55, 0x73, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x20, 0x3d, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, - 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x67, 0x65, 0x74, 0x44, 0x61, - 0x74, 0x61, 0x41, 0x73, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, - 0x75, 0x73, 0x65, 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x73, 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x27, 0x29, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x6c, 0x61, 0x73, + 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, + 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, + 0x6c, 0x79, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, + 0x75, 0x6c, 0x74, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x41, 0x70, 0x70, 0x6c, 0x79, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, + 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, + 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, + 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4c, 0x6f, 0x61, 0x64, 0x41, 0x6e, + 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, + 0x76, 0x65, 0x64, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x67, 0x65, 0x74, 0x20, 0x61, 0x75, 0x74, + 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, + 0x75, 0x73, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, + 0x6c, 0x61, 0x73, 0x74, 0x55, 0x73, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, + 0x6c, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, + 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x67, 0x65, 0x74, + 0x44, 0x61, 0x74, 0x61, 0x41, 0x73, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, + 0x28, 0x27, 0x75, 0x73, 0x65, 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, + 0x61, 0x74, 0x65, 0x73, 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x27, 0x29, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x6c, + 0x61, 0x73, 0x74, 0x55, 0x73, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, + 0x61, 0x74, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, + 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, + 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, + 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x2c, 0x20, 0x72, 0x65, 0x73, 0x74, 0x6f, + 0x72, 0x69, 0x6e, 0x67, 0x27, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, + 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x55, 0x73, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, - 0x67, 0x28, 0x27, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, - 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x66, 0x6f, - 0x75, 0x6e, 0x64, 0x2c, 0x20, 0x72, 0x65, 0x73, 0x74, 0x6f, 0x72, 0x69, - 0x6e, 0x67, 0x27, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, - 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x55, - 0x73, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, - 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x4e, 0x6f, 0x20, 0x61, 0x75, - 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x20, 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x2c, 0x20, - 0x75, 0x73, 0x69, 0x6e, 0x67, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, - 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x27, 0x29, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, - 0x6e, 0x6f, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, - 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x75, 0x73, 0x65, 0x64, 0x20, 0x74, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x77, 0x61, 0x73, 0x20, - 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x2c, 0x20, 0x73, 0x6f, 0x20, 0x6c, 0x6f, - 0x61, 0x64, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x64, 0x65, 0x66, 0x61, - 0x75, 0x6c, 0x74, 0x2e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, - 0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, - 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x41, - 0x70, 0x70, 0x6c, 0x79, 0x69, 0x6e, 0x67, 0x20, 0x74, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x2f, 0x2f, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x75, 0x70, 0x64, 0x61, - 0x74, 0x65, 0x20, 0x69, 0x6e, 0x74, 0x65, 0x72, 0x6e, 0x61, 0x6c, 0x20, - 0x64, 0x61, 0x74, 0x61, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x74, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x28, 0x73, 0x65, 0x6c, - 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x2f, 0x2f, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, - 0x67, 0x28, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x63, 0x6f, - 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x73, 0x65, - 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, - 0x69, 0x6f, 0x6e, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, - 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, - 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x41, 0x75, 0x74, - 0x6f, 0x73, 0x61, 0x76, 0x65, 0x2e, 0x2e, 0x2e, 0x27, 0x29, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x73, 0x65, 0x6c, - 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, - 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x3d, 0x20, 0x27, 0x64, 0x65, 0x66, - 0x61, 0x75, 0x6c, 0x74, 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x77, 0x65, 0x20, 0x64, - 0x6f, 0x6e, 0x27, 0x74, 0x20, 0x77, 0x61, 0x6e, 0x74, 0x20, 0x74, 0x6f, - 0x20, 0x73, 0x61, 0x76, 0x65, 0x20, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x64, - 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x2c, 0x20, 0x73, 0x6f, 0x20, 0x6c, 0x65, 0x74, 0x27, - 0x73, 0x20, 0x63, 0x72, 0x65, 0x61, 0x74, 0x65, 0x20, 0x61, 0x20, 0x6e, - 0x65, 0x77, 0x20, 0x6f, 0x6e, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, + 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x4e, 0x6f, 0x20, + 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x66, 0x6f, 0x75, 0x6e, 0x64, + 0x2c, 0x20, 0x75, 0x73, 0x69, 0x6e, 0x67, 0x20, 0x64, 0x65, 0x66, 0x61, + 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, + 0x2f, 0x20, 0x6e, 0x6f, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, + 0x65, 0x64, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x75, 0x73, 0x65, 0x64, + 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x77, 0x61, + 0x73, 0x20, 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x2c, 0x20, 0x73, 0x6f, 0x20, + 0x6c, 0x6f, 0x61, 0x64, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x64, 0x65, + 0x66, 0x61, 0x75, 0x6c, 0x74, 0x2e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, + 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, + 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, + 0x27, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x69, 0x6e, 0x67, 0x20, 0x74, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x75, 0x70, + 0x64, 0x61, 0x74, 0x65, 0x20, 0x69, 0x6e, 0x74, 0x65, 0x72, 0x6e, 0x61, + 0x6c, 0x20, 0x64, 0x61, 0x74, 0x61, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, + 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x28, 0x73, + 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x2f, 0x2f, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, + 0x6c, 0x6f, 0x67, 0x28, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, + 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, + 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, + 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, + 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, + 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, + 0x76, 0x65, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, + 0x28, 0x27, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x41, + 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x2e, 0x2e, 0x2e, 0x27, 0x29, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x73, + 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x3d, 0x20, 0x27, 0x64, + 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x77, 0x65, + 0x20, 0x64, 0x6f, 0x6e, 0x27, 0x74, 0x20, 0x77, 0x61, 0x6e, 0x74, 0x20, + 0x74, 0x6f, 0x20, 0x73, 0x61, 0x76, 0x65, 0x20, 0x6f, 0x76, 0x65, 0x72, + 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2c, 0x20, 0x73, 0x6f, 0x20, 0x6c, 0x65, + 0x74, 0x27, 0x73, 0x20, 0x63, 0x72, 0x65, 0x61, 0x74, 0x65, 0x20, 0x61, + 0x20, 0x6e, 0x65, 0x77, 0x20, 0x6f, 0x6e, 0x65, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6e, 0x65, 0x77, + 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4e, 0x61, 0x6d, 0x65, + 0x20, 0x3d, 0x20, 0x27, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, + 0x6c, 0x61, 0x74, 0x65, 0x2d, 0x27, 0x20, 0x2b, 0x20, 0x44, 0x61, 0x74, + 0x65, 0x2e, 0x6e, 0x6f, 0x77, 0x28, 0x29, 0x2e, 0x74, 0x6f, 0x53, 0x74, + 0x72, 0x69, 0x6e, 0x67, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x20, 0x3d, - 0x20, 0x27, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x2d, 0x27, 0x20, 0x2b, 0x20, 0x44, 0x61, 0x74, 0x65, 0x2e, - 0x6e, 0x6f, 0x77, 0x28, 0x29, 0x2e, 0x74, 0x6f, 0x53, 0x74, 0x72, 0x69, - 0x6e, 0x67, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x27, 0x6e, 0x61, - 0x6d, 0x65, 0x27, 0x3a, 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x27, 0x64, - 0x61, 0x74, 0x61, 0x27, 0x3a, 0x20, 0x7b, 0x20, 0x27, 0x73, 0x65, 0x73, - 0x73, 0x69, 0x6f, 0x6e, 0x27, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, - 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x27, 0x70, - 0x61, 0x72, 0x61, 0x6d, 0x73, 0x27, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, - 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x20, 0x7d, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, - 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x53, - 0x61, 0x76, 0x69, 0x6e, 0x67, 0x20, 0x61, 0x73, 0x20, 0x27, 0x20, 0x2b, - 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, - 0x4e, 0x61, 0x6d, 0x65, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x61, 0x76, 0x65, 0x20, 0x69, - 0x6e, 0x20, 0x74, 0x68, 0x65, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, - 0x76, 0x65, 0x20, 0x73, 0x6c, 0x6f, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, - 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, - 0x61, 0x46, 0x72, 0x6f, 0x6d, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, - 0x27, 0x75, 0x73, 0x65, 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x73, 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x27, 0x2c, 0x20, 0x6e, - 0x65, 0x77, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x29, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, - 0x61, 0x6e, 0x64, 0x20, 0x6c, 0x6f, 0x61, 0x64, 0x20, 0x69, 0x74, 0x20, - 0x62, 0x61, 0x63, 0x6b, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x61, 0x70, 0x70, - 0x6c, 0x79, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, - 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4c, - 0x6f, 0x61, 0x64, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x41, - 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x28, 0x29, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6f, - 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, - 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d, 0x4f, - 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, 0x75, 0x73, 0x65, 0x72, 0x5f, - 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x5f, 0x6c, 0x61, - 0x73, 0x74, 0x27, 0x2c, 0x20, 0x7b, 0x20, 0x27, 0x6e, 0x61, 0x6d, 0x65, - 0x27, 0x3a, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, - 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x27, + 0x6e, 0x61, 0x6d, 0x65, 0x27, 0x3a, 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x27, 0x64, 0x61, 0x74, 0x61, 0x27, 0x3a, 0x20, 0x7b, 0x20, 0x27, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x27, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x27, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x27, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, - 0x20, 0x7d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, - 0x43, 0x68, 0x65, 0x63, 0x6b, 0x69, 0x6e, 0x67, 0x20, 0x66, 0x6f, 0x72, - 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x6c, - 0x61, 0x73, 0x74, 0x20, 0x75, 0x73, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, - 0x4c, 0x6f, 0x61, 0x64, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, - 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x28, 0x29, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2a, 0x20, 0x45, 0x4e, 0x44, 0x3a, - 0x20, 0x53, 0x75, 0x70, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x66, 0x6f, 0x72, - 0x20, 0x73, 0x74, 0x6f, 0x72, 0x69, 0x6e, 0x67, 0x20, 0x70, 0x72, 0x6f, - 0x6d, 0x70, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, - 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x65, - 0x74, 0x65, 0x72, 0x73, 0x20, 0x69, 0x6e, 0x20, 0x62, 0x72, 0x6f, 0x77, - 0x73, 0x65, 0x72, 0x73, 0x20, 0x4c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, - 0x6f, 0x72, 0x61, 0x67, 0x65, 0x20, 0x2a, 0x2f, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, - 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, - 0x6e, 0x61, 0x6c, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, - 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, - 0x6e, 0x61, 0x6c, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, - 0x74, 0x6c, 0x79, 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, - 0x6e, 0x67, 0x20, 0x61, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, - 0x69, 0x6f, 0x6e, 0x3f, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, - 0x73, 0x74, 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6e, - 0x67, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x75, 0x74, 0x65, 0x64, - 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, - 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, - 0x21, 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x68, 0x61, 0x73, 0x20, 0x74, 0x68, 0x65, - 0x20, 0x75, 0x73, 0x65, 0x72, 0x20, 0x73, 0x74, 0x61, 0x72, 0x74, 0x65, - 0x64, 0x20, 0x61, 0x20, 0x63, 0x68, 0x61, 0x74, 0x3f, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x61, 0x74, - 0x53, 0x74, 0x61, 0x72, 0x74, 0x65, 0x64, 0x20, 0x3d, 0x20, 0x63, 0x6f, - 0x6d, 0x70, 0x75, 0x74, 0x65, 0x64, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, - 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, - 0x74, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x3e, 0x20, 0x30, - 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, - 0x70, 0x64, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x74, 0x72, 0x61, - 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, - 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, - 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x69, 0x6d, - 0x70, 0x6c, 0x65, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, - 0x20, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x73, 0x74, 0x72, 0x2c, 0x20, - 0x65, 0x78, 0x74, 0x72, 0x61, 0x53, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, - 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, - 0x67, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, 0x78, 0x74, 0x72, 0x61, 0x53, - 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x74, 0x74, 0x69, - 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x73, - 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x2c, 0x20, 0x2e, 0x2e, 0x2e, - 0x65, 0x78, 0x74, 0x72, 0x61, 0x53, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, - 0x73, 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, - 0x6e, 0x20, 0x53, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x28, 0x73, 0x74, 0x72, - 0x29, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x41, 0x6c, 0x6c, - 0x28, 0x2f, 0x5c, 0x7b, 0x5c, 0x7b, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5c, - 0x7d, 0x5c, 0x7d, 0x2f, 0x67, 0x2c, 0x20, 0x28, 0x5f, 0x2c, 0x20, 0x6b, - 0x65, 0x79, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x28, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, - 0x5b, 0x6b, 0x65, 0x79, 0x5d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x61, 0x73, 0x79, 0x6e, - 0x63, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x72, - 0x75, 0x6e, 0x4c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, - 0x70, 0x74, 0x2c, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x61, 0x72, - 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63, 0x68, 0x61, 0x72, 0x29, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, - 0x61, 0x67, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x5b, 0x5d, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x68, - 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, - 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, - 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6e, + 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, + 0x27, 0x53, 0x61, 0x76, 0x69, 0x6e, 0x67, 0x20, 0x61, 0x73, 0x20, 0x27, + 0x20, 0x2b, 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, + 0x74, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x61, 0x76, 0x65, + 0x20, 0x69, 0x6e, 0x20, 0x74, 0x68, 0x65, 0x20, 0x61, 0x75, 0x74, 0x6f, + 0x73, 0x61, 0x76, 0x65, 0x20, 0x73, 0x6c, 0x6f, 0x74, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, + 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, + 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d, 0x4f, 0x62, 0x6a, 0x65, 0x63, + 0x74, 0x28, 0x27, 0x75, 0x73, 0x65, 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, + 0x6c, 0x61, 0x74, 0x65, 0x73, 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x27, 0x2c, + 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, + 0x2f, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x6c, 0x6f, 0x61, 0x64, 0x20, 0x69, + 0x74, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x61, + 0x70, 0x70, 0x6c, 0x79, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x4c, 0x6f, 0x61, 0x64, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, + 0x79, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x28, 0x29, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, + 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, + 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, + 0x6d, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, 0x75, 0x73, 0x65, + 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x5f, + 0x6c, 0x61, 0x73, 0x74, 0x27, 0x2c, 0x20, 0x7b, 0x20, 0x27, 0x6e, 0x61, + 0x6d, 0x65, 0x27, 0x3a, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, + 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65, + 0x2c, 0x20, 0x27, 0x64, 0x61, 0x74, 0x61, 0x27, 0x3a, 0x20, 0x7b, 0x20, + 0x27, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x27, 0x3a, 0x20, 0x73, + 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x2c, 0x20, 0x27, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x27, 0x3a, 0x20, + 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x20, 0x7d, 0x20, 0x7d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, + 0x28, 0x27, 0x43, 0x68, 0x65, 0x63, 0x6b, 0x69, 0x6e, 0x67, 0x20, 0x66, + 0x6f, 0x72, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, + 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x75, 0x73, 0x65, 0x64, 0x20, 0x74, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x27, 0x29, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, + 0x74, 0x65, 0x4c, 0x6f, 0x61, 0x64, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, + 0x6c, 0x79, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x28, + 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2a, 0x20, 0x45, 0x4e, + 0x44, 0x3a, 0x20, 0x53, 0x75, 0x70, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x66, + 0x6f, 0x72, 0x20, 0x73, 0x74, 0x6f, 0x72, 0x69, 0x6e, 0x67, 0x20, 0x70, + 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, + 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x70, 0x61, 0x72, 0x61, + 0x6d, 0x65, 0x74, 0x65, 0x72, 0x73, 0x20, 0x69, 0x6e, 0x20, 0x62, 0x72, + 0x6f, 0x77, 0x73, 0x65, 0x72, 0x73, 0x20, 0x4c, 0x6f, 0x63, 0x61, 0x6c, + 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x20, 0x2a, 0x2f, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, + 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x20, 0x3d, 0x20, 0x73, + 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, + 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x73, + 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x63, 0x75, 0x72, 0x72, + 0x65, 0x6e, 0x74, 0x6c, 0x79, 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, + 0x74, 0x69, 0x6e, 0x67, 0x20, 0x61, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, + 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x3f, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, + 0x69, 0x6e, 0x67, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x75, 0x74, + 0x65, 0x64, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x45, - 0x72, 0x72, 0x6f, 0x72, 0x28, 0x22, 0x61, 0x6c, 0x72, 0x65, 0x61, 0x64, - 0x79, 0x20, 0x72, 0x75, 0x6e, 0x6e, 0x69, 0x6e, 0x67, 0x22, 0x29, 0x3b, + 0x65, 0x20, 0x21, 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x68, 0x61, 0x73, 0x20, 0x74, + 0x68, 0x65, 0x20, 0x75, 0x73, 0x65, 0x72, 0x20, 0x73, 0x74, 0x61, 0x72, + 0x74, 0x65, 0x64, 0x20, 0x61, 0x20, 0x63, 0x68, 0x61, 0x74, 0x3f, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, + 0x61, 0x74, 0x53, 0x74, 0x61, 0x72, 0x74, 0x65, 0x64, 0x20, 0x3d, 0x20, + 0x63, 0x6f, 0x6d, 0x70, 0x75, 0x74, 0x65, 0x64, 0x28, 0x28, 0x29, 0x20, + 0x3d, 0x3e, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, + 0x69, 0x70, 0x74, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x3e, + 0x20, 0x30, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, + 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x74, + 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x29, 0x20, 0x3d, + 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, + 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, + 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, - 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x6e, 0x65, - 0x77, 0x20, 0x41, 0x62, 0x6f, 0x72, 0x74, 0x43, 0x6f, 0x6e, 0x74, 0x72, - 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, - 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, - 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, - 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, - 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x7b, 0x63, 0x6f, 0x6e, - 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x3a, 0x20, 0x63, 0x6f, 0x6e, - 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x7d, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x64, 0x61, 0x74, - 0x61, 0x20, 0x3d, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, - 0x74, 0x61, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x69, 0x66, 0x20, 0x28, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x73, 0x74, - 0x6f, 0x70, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x77, 0x68, 0x69, 0x6c, 0x65, 0x20, 0x28, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, - 0x67, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x3e, - 0x20, 0x30, 0x20, 0x26, 0x26, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, - 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x5b, 0x63, 0x75, + 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, + 0x69, 0x6d, 0x70, 0x6c, 0x65, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, + 0x74, 0x65, 0x20, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x73, 0x74, 0x72, + 0x2c, 0x20, 0x65, 0x78, 0x74, 0x72, 0x61, 0x53, 0x65, 0x74, 0x74, 0x69, + 0x6e, 0x67, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x73, 0x65, 0x74, 0x74, + 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, + 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, 0x78, 0x74, 0x72, + 0x61, 0x53, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x74, + 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, + 0x2e, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x2c, 0x20, 0x2e, + 0x2e, 0x2e, 0x65, 0x78, 0x74, 0x72, 0x61, 0x53, 0x65, 0x74, 0x74, 0x69, + 0x6e, 0x67, 0x73, 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, + 0x75, 0x72, 0x6e, 0x20, 0x53, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x28, 0x73, + 0x74, 0x72, 0x29, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x41, + 0x6c, 0x6c, 0x28, 0x2f, 0x5c, 0x7b, 0x5c, 0x7b, 0x28, 0x2e, 0x2a, 0x3f, + 0x29, 0x5c, 0x7d, 0x5c, 0x7d, 0x2f, 0x67, 0x2c, 0x20, 0x28, 0x5f, 0x2c, + 0x20, 0x6b, 0x65, 0x79, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x74, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, + 0x67, 0x73, 0x5b, 0x6b, 0x65, 0x79, 0x5d, 0x29, 0x29, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x61, 0x73, + 0x79, 0x6e, 0x63, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, + 0x20, 0x72, 0x75, 0x6e, 0x4c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, + 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, + 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63, 0x68, 0x61, 0x72, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, + 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x5b, 0x5d, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x20, 0x3d, 0x20, 0x73, + 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, + 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x6e, 0x65, 0x77, + 0x20, 0x45, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x22, 0x61, 0x6c, 0x72, 0x65, + 0x61, 0x64, 0x79, 0x20, 0x72, 0x75, 0x6e, 0x6e, 0x69, 0x6e, 0x67, 0x22, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, + 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, + 0x6e, 0x65, 0x77, 0x20, 0x41, 0x62, 0x6f, 0x72, 0x74, 0x43, 0x6f, 0x6e, + 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x28, 0x29, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, + 0x69, 0x74, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, + 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, + 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x6c, 0x6c, 0x61, + 0x6d, 0x61, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x7b, 0x20, + 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x3a, 0x20, + 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x64, 0x61, 0x74, 0x61, 0x20, 0x3d, 0x20, 0x63, 0x68, 0x75, 0x6e, + 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x64, 0x61, 0x74, + 0x61, 0x2e, 0x73, 0x74, 0x6f, 0x70, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x77, 0x68, 0x69, 0x6c, + 0x65, 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, + 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, + 0x74, 0x68, 0x20, 0x3e, 0x20, 0x30, 0x20, 0x26, 0x26, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, - 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x2d, 0x20, 0x31, - 0x5d, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x2e, 0x6d, 0x61, - 0x74, 0x63, 0x68, 0x28, 0x2f, 0x5c, 0x6e, 0x24, 0x2f, 0x29, 0x20, 0x21, - 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, - 0x65, 0x73, 0x2e, 0x70, 0x6f, 0x70, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, - 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, - 0x28, 0x5b, 0x2e, 0x2e, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, - 0x2c, 0x20, 0x5b, 0x63, 0x68, 0x61, 0x72, 0x2c, 0x20, 0x63, 0x75, 0x72, - 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, - 0x5d, 0x5d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, - 0x67, 0x28, 0x22, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, - 0x6e, 0x20, 0x66, 0x69, 0x6e, 0x69, 0x73, 0x68, 0x65, 0x64, 0x3a, 0x20, - 0x27, 0x22, 0x2c, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, - 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x6d, 0x61, 0x70, 0x28, - 0x6d, 0x73, 0x67, 0x20, 0x3d, 0x3e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, - 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, - 0x28, 0x27, 0x27, 0x29, 0x2c, 0x20, 0x22, 0x27, 0x2c, 0x20, 0x73, 0x75, - 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x3a, 0x20, 0x22, 0x2c, 0x20, 0x64, 0x61, - 0x74, 0x61, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x75, 0x72, 0x72, - 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, - 0x70, 0x75, 0x73, 0x68, 0x28, 0x64, 0x61, 0x74, 0x61, 0x29, 0x3b, 0x0a, + 0x73, 0x5b, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, + 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, + 0x20, 0x2d, 0x20, 0x31, 0x5d, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, + 0x74, 0x2e, 0x6d, 0x61, 0x74, 0x63, 0x68, 0x28, 0x2f, 0x5c, 0x6e, 0x24, + 0x2f, 0x29, 0x20, 0x21, 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, + 0x61, 0x67, 0x65, 0x73, 0x2e, 0x70, 0x6f, 0x70, 0x28, 0x29, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, 0x2e, 0x2e, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x2c, 0x20, 0x5b, 0x63, 0x68, 0x61, 0x72, 0x2c, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x5d, 0x5d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x69, 0x66, 0x20, 0x28, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, - 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, - 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, - 0x3d, 0x20, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, 0x6d, 0x69, 0x6e, - 0x67, 0x73, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, - 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, - 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x65, 0x6e, 0x64, - 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x74, 0x6f, 0x20, - 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x61, 0x74, 0x20, 0x3d, 0x20, - 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x6d, 0x73, 0x67, 0x29, 0x20, - 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, - 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, - 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, - 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x61, 0x6c, 0x72, 0x65, - 0x61, 0x64, 0x79, 0x20, 0x72, 0x75, 0x6e, 0x6e, 0x69, 0x6e, 0x67, 0x2e, - 0x2e, 0x2e, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, - 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, 0x2e, 0x2e, 0x2e, 0x73, 0x65, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, + 0x6c, 0x6f, 0x67, 0x28, 0x22, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, + 0x69, 0x6f, 0x6e, 0x20, 0x66, 0x69, 0x6e, 0x69, 0x73, 0x68, 0x65, 0x64, + 0x3a, 0x20, 0x27, 0x22, 0x2c, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, + 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x6d, 0x61, + 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20, 0x3d, 0x3e, 0x20, 0x6d, 0x73, 0x67, + 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x2e, 0x6a, 0x6f, + 0x69, 0x6e, 0x28, 0x27, 0x27, 0x29, 0x2c, 0x20, 0x22, 0x27, 0x2c, 0x20, + 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x3a, 0x20, 0x22, 0x2c, 0x20, + 0x64, 0x61, 0x74, 0x61, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x75, + 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, + 0x73, 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, 0x64, 0x61, 0x74, 0x61, 0x29, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x73, 0x6c, 0x6f, 0x74, 0x5f, 0x69, 0x64, 0x20, 0x3d, 0x20, 0x64, 0x61, + 0x74, 0x61, 0x2e, 0x73, 0x6c, 0x6f, 0x74, 0x5f, 0x69, 0x64, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, + 0x20, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x5f, 0x69, + 0x6d, 0x61, 0x67, 0x65, 0x20, 0x26, 0x26, 0x20, 0x21, 0x64, 0x61, 0x74, + 0x61, 0x2e, 0x6d, 0x75, 0x6c, 0x74, 0x69, 0x6d, 0x6f, 0x64, 0x61, 0x6c, + 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x61, 0x6c, 0x65, 0x72, 0x74, 0x28, 0x22, 0x54, + 0x68, 0x65, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x20, 0x77, 0x61, + 0x73, 0x20, 0x6e, 0x6f, 0x74, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x69, 0x6c, + 0x65, 0x64, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x6d, 0x75, 0x6c, 0x74, 0x69, + 0x6d, 0x6f, 0x64, 0x61, 0x6c, 0x20, 0x6f, 0x72, 0x20, 0x74, 0x68, 0x65, + 0x20, 0x6d, 0x6f, 0x64, 0x65, 0x6c, 0x20, 0x70, 0x72, 0x6f, 0x6a, 0x65, + 0x63, 0x74, 0x6f, 0x72, 0x20, 0x63, 0x61, 0x6e, 0x27, 0x74, 0x20, 0x62, + 0x65, 0x20, 0x6c, 0x6f, 0x61, 0x64, 0x65, 0x64, 0x2e, 0x22, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, + 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, + 0x5b, 0x2e, 0x2e, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x2c, + 0x20, 0x5b, 0x63, 0x68, 0x61, 0x72, 0x2c, 0x20, 0x63, 0x75, 0x72, 0x72, + 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x5d, + 0x5d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, + 0x20, 0x28, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, 0x6d, 0x69, 0x6e, + 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, + 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x64, + 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x6e, 0x75, 0x6c, + 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x65, 0x6e, 0x64, 0x20, 0x6d, 0x65, + 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x74, 0x6f, 0x20, 0x73, 0x65, 0x72, + 0x76, 0x65, 0x72, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, + 0x74, 0x20, 0x63, 0x68, 0x61, 0x74, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, + 0x6e, 0x63, 0x20, 0x28, 0x6d, 0x73, 0x67, 0x29, 0x20, 0x3d, 0x3e, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, + 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, + 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x61, 0x6c, 0x72, 0x65, 0x61, 0x64, 0x79, + 0x20, 0x72, 0x75, 0x6e, 0x6e, 0x69, 0x6e, 0x67, 0x2e, 0x2e, 0x2e, 0x27, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, + 0x65, 0x74, 0x75, 0x72, 0x6e, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, + 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, + 0x74, 0x65, 0x28, 0x5b, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, + 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, + 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x2c, 0x20, 0x5b, 0x22, 0x7b, + 0x7b, 0x75, 0x73, 0x65, 0x72, 0x7d, 0x7d, 0x22, 0x2c, 0x20, 0x6d, 0x73, + 0x67, 0x5d, 0x5d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x6c, 0x65, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x3d, + 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, - 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x2c, 0x20, - 0x5b, 0x22, 0x7b, 0x7b, 0x75, 0x73, 0x65, 0x72, 0x7d, 0x7d, 0x22, 0x2c, - 0x20, 0x6d, 0x73, 0x67, 0x5d, 0x5d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x72, 0x6f, - 0x6d, 0x70, 0x74, 0x20, 0x3d, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x28, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x2c, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x3a, 0x20, 0x6d, 0x73, - 0x67, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, - 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, - 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, - 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x2e, 0x66, 0x6c, 0x61, - 0x74, 0x4d, 0x61, 0x70, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x28, 0x5b, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, - 0x64, 0x61, 0x74, 0x61, 0x5d, 0x29, 0x20, 0x3d, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, - 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, - 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2c, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, + 0x61, 0x67, 0x65, 0x3a, 0x20, 0x6d, 0x73, 0x67, 0x2c, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, + 0x79, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, + 0x69, 0x70, 0x74, 0x2e, 0x66, 0x6c, 0x61, 0x74, 0x4d, 0x61, 0x70, 0x28, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x28, + 0x5b, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x64, 0x61, 0x74, 0x61, 0x5d, + 0x29, 0x20, 0x3d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f, + 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, - 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x3a, 0x20, 0x41, 0x72, 0x72, 0x61, - 0x79, 0x2e, 0x69, 0x73, 0x41, 0x72, 0x72, 0x61, 0x79, 0x28, 0x64, 0x61, - 0x74, 0x61, 0x29, 0x20, 0x3f, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6e, 0x61, 0x6d, 0x65, + 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, + 0x65, 0x3a, 0x20, 0x41, 0x72, 0x72, 0x61, 0x79, 0x2e, 0x69, 0x73, 0x41, + 0x72, 0x72, 0x61, 0x79, 0x28, 0x64, 0x61, 0x74, 0x61, 0x29, 0x20, 0x3f, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x61, 0x74, 0x61, 0x2e, + 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20, 0x3d, 0x3e, 0x20, 0x6d, + 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x2e, + 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, 0x27, 0x29, 0x2e, 0x72, 0x65, 0x70, + 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5e, 0x5c, 0x73, 0x2f, 0x2c, 0x20, + 0x27, 0x27, 0x29, 0x20, 0x3a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x64, 0x61, 0x74, 0x61, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, 0x67, - 0x20, 0x3d, 0x3e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, - 0x65, 0x6e, 0x74, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, 0x27, - 0x29, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5e, - 0x5c, 0x73, 0x2f, 0x2c, 0x20, 0x27, 0x27, 0x29, 0x20, 0x3a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x61, 0x74, 0x61, 0x2c, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x22, 0x5c, 0x6e, 0x22, - 0x29, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x3b, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x61, 0x77, 0x61, 0x69, + 0x64, 0x61, 0x74, 0x61, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x2e, 0x6a, 0x6f, + 0x69, 0x6e, 0x28, 0x22, 0x5c, 0x6e, 0x22, 0x29, 0x2c, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, + 0x65, 0x64, 0x5f, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x29, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x72, 0x6f, 0x6d, + 0x70, 0x74, 0x20, 0x3d, 0x20, 0x60, 0x41, 0x20, 0x63, 0x68, 0x61, 0x74, + 0x20, 0x62, 0x65, 0x74, 0x77, 0x65, 0x65, 0x6e, 0x20, 0x61, 0x20, 0x63, + 0x75, 0x72, 0x69, 0x6f, 0x75, 0x73, 0x20, 0x68, 0x75, 0x6d, 0x61, 0x6e, + 0x20, 0x61, 0x6e, 0x64, 0x20, 0x61, 0x6e, 0x20, 0x61, 0x72, 0x74, 0x69, + 0x66, 0x69, 0x63, 0x69, 0x61, 0x6c, 0x20, 0x69, 0x6e, 0x74, 0x65, 0x6c, + 0x6c, 0x69, 0x67, 0x65, 0x6e, 0x63, 0x65, 0x20, 0x61, 0x73, 0x73, 0x69, + 0x73, 0x74, 0x61, 0x6e, 0x74, 0x2e, 0x20, 0x54, 0x68, 0x65, 0x20, 0x61, + 0x73, 0x73, 0x69, 0x73, 0x74, 0x61, 0x6e, 0x74, 0x20, 0x67, 0x69, 0x76, + 0x65, 0x73, 0x20, 0x68, 0x65, 0x6c, 0x70, 0x66, 0x75, 0x6c, 0x2c, 0x20, + 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x65, 0x64, 0x2c, 0x20, 0x61, 0x6e, + 0x64, 0x20, 0x70, 0x6f, 0x6c, 0x69, 0x74, 0x65, 0x20, 0x61, 0x6e, 0x73, + 0x77, 0x65, 0x72, 0x73, 0x20, 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, + 0x68, 0x75, 0x6d, 0x61, 0x6e, 0x27, 0x73, 0x20, 0x71, 0x75, 0x65, 0x73, + 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x2e, 0x5c, 0x6e, 0x55, 0x53, 0x45, 0x52, + 0x3a, 0x5b, 0x69, 0x6d, 0x67, 0x2d, 0x31, 0x30, 0x5d, 0x24, 0x7b, 0x6d, + 0x73, 0x67, 0x7d, 0x5c, 0x6e, 0x41, 0x53, 0x53, 0x49, 0x53, 0x54, 0x41, + 0x4e, 0x54, 0x3a, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x72, 0x75, 0x6e, 0x4c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70, 0x3a, 0x20, - 0x5b, 0x22, 0x3c, 0x2f, 0x73, 0x3e, 0x22, 0x2c, 0x20, 0x74, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x22, 0x7b, 0x7b, 0x63, 0x68, 0x61, - 0x72, 0x7d, 0x7d, 0x3a, 0x22, 0x29, 0x2c, 0x20, 0x74, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x28, 0x22, 0x7b, 0x7b, 0x75, 0x73, 0x65, 0x72, - 0x7d, 0x7d, 0x3a, 0x22, 0x29, 0x5d, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x22, 0x7b, 0x7b, 0x63, 0x68, 0x61, 0x72, - 0x7d, 0x7d, 0x22, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, - 0x75, 0x6e, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, - 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x29, 0x20, - 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, - 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, - 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, - 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x61, 0x6c, 0x72, 0x65, - 0x61, 0x64, 0x79, 0x20, 0x72, 0x75, 0x6e, 0x6e, 0x69, 0x6e, 0x67, 0x2e, - 0x2e, 0x2e, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x7b, 0x70, 0x72, 0x6f, 0x6d, 0x70, - 0x74, 0x7d, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, - 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, 0x2e, 0x2e, 0x2e, 0x73, - 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x2c, - 0x20, 0x5b, 0x22, 0x22, 0x2c, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, - 0x5d, 0x5d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x61, - 0x77, 0x61, 0x69, 0x74, 0x20, 0x72, 0x75, 0x6e, 0x4c, 0x6c, 0x61, 0x6d, - 0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x70, - 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, - 0x70, 0x3a, 0x20, 0x5b, 0x5d, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x2c, 0x20, 0x22, 0x22, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, - 0x74, 0x20, 0x73, 0x74, 0x6f, 0x70, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, - 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x65, 0x2e, 0x70, 0x72, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x44, 0x65, 0x66, - 0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, - 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, - 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x2e, 0x61, 0x62, 0x6f, 0x72, 0x74, 0x28, 0x29, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, - 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x20, 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, - 0x73, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, - 0x70, 0x28, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, - 0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, 0x5d, 0x29, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, - 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, - 0x65, 0x49, 0x6e, 0x70, 0x75, 0x74, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, - 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, - 0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x22, 0x22, 0x29, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, - 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, - 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70, 0x28, 0x65, 0x29, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x68, 0x61, 0x74, 0x28, - 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x65, 0x6e, 0x74, 0x65, 0x72, 0x53, 0x75, - 0x62, 0x6d, 0x69, 0x74, 0x73, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x76, 0x65, - 0x6e, 0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, 0x76, 0x65, - 0x6e, 0x74, 0x2e, 0x77, 0x68, 0x69, 0x63, 0x68, 0x20, 0x3d, 0x3d, 0x3d, - 0x20, 0x31, 0x33, 0x20, 0x26, 0x26, 0x20, 0x21, 0x65, 0x76, 0x65, 0x6e, - 0x74, 0x2e, 0x73, 0x68, 0x69, 0x66, 0x74, 0x4b, 0x65, 0x79, 0x29, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x6c, 0x6f, 0x74, 0x5f, 0x69, + 0x64, 0x3a, 0x20, 0x73, 0x6c, 0x6f, 0x74, 0x5f, 0x69, 0x64, 0x2c, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70, + 0x3a, 0x20, 0x5b, 0x22, 0x3c, 0x2f, 0x73, 0x3e, 0x22, 0x2c, 0x20, 0x74, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x22, 0x7b, 0x7b, 0x63, + 0x68, 0x61, 0x72, 0x7d, 0x7d, 0x3a, 0x22, 0x29, 0x2c, 0x20, 0x74, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x22, 0x7b, 0x7b, 0x75, 0x73, + 0x65, 0x72, 0x7d, 0x7d, 0x3a, 0x22, 0x29, 0x5d, 0x2c, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x22, 0x7b, 0x7b, 0x63, 0x68, + 0x61, 0x72, 0x7d, 0x7d, 0x22, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x72, 0x75, 0x6e, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, + 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, + 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, + 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x61, 0x6c, + 0x72, 0x65, 0x61, 0x64, 0x79, 0x20, 0x72, 0x75, 0x6e, 0x6e, 0x69, 0x6e, + 0x67, 0x2e, 0x2e, 0x2e, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x7b, 0x20, 0x70, 0x72, + 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x7d, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, + 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, + 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, + 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, + 0x69, 0x70, 0x74, 0x2c, 0x20, 0x5b, 0x22, 0x22, 0x2c, 0x20, 0x70, 0x72, + 0x6f, 0x6d, 0x70, 0x74, 0x5d, 0x5d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x72, 0x75, 0x6e, + 0x4c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, + 0x2c, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x73, 0x6c, 0x6f, 0x74, 0x5f, 0x69, 0x64, 0x3a, 0x20, 0x73, 0x6c, + 0x6f, 0x74, 0x5f, 0x69, 0x64, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x5b, 0x5d, 0x2c, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x22, 0x22, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x74, 0x6f, 0x70, + 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x2e, 0x70, 0x72, 0x65, 0x76, + 0x65, 0x6e, 0x74, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, + 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, + 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x61, 0x62, 0x6f, + 0x72, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x6e, 0x75, 0x6c, + 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x73, 0x65, 0x74, 0x20, 0x3d, 0x20, + 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70, 0x28, 0x65, 0x29, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, + 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, + 0x5d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, 0x6c, + 0x6f, 0x61, 0x64, 0x49, 0x6d, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, 0x28, + 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x65, 0x2e, 0x70, 0x72, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x44, + 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, + 0x2e, 0x67, 0x65, 0x74, 0x45, 0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x42, + 0x79, 0x49, 0x64, 0x28, 0x22, 0x66, 0x69, 0x6c, 0x65, 0x49, 0x6e, 0x70, + 0x75, 0x74, 0x22, 0x29, 0x2e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x28, 0x29, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, 0x63, 0x75, + 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x67, 0x65, 0x74, 0x45, 0x6c, 0x65, 0x6d, + 0x65, 0x6e, 0x74, 0x42, 0x79, 0x49, 0x64, 0x28, 0x22, 0x66, 0x69, 0x6c, + 0x65, 0x49, 0x6e, 0x70, 0x75, 0x74, 0x22, 0x29, 0x2e, 0x61, 0x64, 0x64, + 0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65, 0x6e, 0x65, + 0x72, 0x28, 0x22, 0x63, 0x68, 0x61, 0x6e, 0x67, 0x65, 0x22, 0x2c, 0x20, + 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x28, 0x65, 0x76, + 0x65, 0x6e, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x65, 0x6c, + 0x65, 0x63, 0x74, 0x65, 0x64, 0x46, 0x69, 0x6c, 0x65, 0x20, 0x3d, 0x20, + 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, + 0x2e, 0x66, 0x69, 0x6c, 0x65, 0x73, 0x5b, 0x30, 0x5d, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x73, + 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x46, 0x69, 0x6c, 0x65, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x61, 0x64, 0x65, + 0x72, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x46, 0x69, 0x6c, 0x65, + 0x52, 0x65, 0x61, 0x64, 0x65, 0x72, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x61, 0x64, + 0x65, 0x72, 0x2e, 0x6f, 0x6e, 0x6c, 0x6f, 0x61, 0x64, 0x20, 0x3d, 0x20, + 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, - 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, - 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x66, 0x6f, 0x72, 0x6d, 0x20, 0x6f, 0x6e, 0x73, 0x75, 0x62, - 0x6d, 0x69, 0x74, 0x3d, 0x24, 0x7b, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, - 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, - 0x61, 0x72, 0x65, 0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, - 0x73, 0x4e, 0x61, 0x6d, 0x65, 0x3d, 0x24, 0x7b, 0x67, 0x65, 0x6e, 0x65, - 0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x20, 0x3f, 0x20, 0x22, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x22, - 0x20, 0x3a, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6d, 0x61, 0x67, + 0x65, 0x5f, 0x64, 0x61, 0x74, 0x61, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x61, + 0x64, 0x65, 0x72, 0x2e, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x28, 0x65, - 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x65, 0x2e, 0x74, - 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x6f, 0x6e, 0x6b, 0x65, 0x79, 0x70, 0x72, 0x65, - 0x73, 0x73, 0x3d, 0x24, 0x7b, 0x65, 0x6e, 0x74, 0x65, 0x72, 0x53, 0x75, - 0x62, 0x6d, 0x69, 0x74, 0x73, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x6c, - 0x61, 0x63, 0x65, 0x68, 0x6f, 0x6c, 0x64, 0x65, 0x72, 0x3d, 0x22, 0x53, - 0x61, 0x79, 0x20, 0x73, 0x6f, 0x6d, 0x65, 0x74, 0x68, 0x69, 0x6e, 0x67, - 0x2e, 0x2e, 0x2e, 0x22, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x6f, 0x77, 0x73, - 0x3d, 0x32, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, - 0x74, 0x65, 0x78, 0x74, 0x22, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, - 0x65, 0x7d, 0x22, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, - 0x69, 0x76, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x72, 0x69, - 0x67, 0x68, 0x74, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, - 0x6e, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x73, 0x75, 0x62, 0x6d, - 0x69, 0x74, 0x22, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, + 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, + 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, + 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, + 0x65, 0x64, 0x3a, 0x20, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x64, 0x61, + 0x74, 0x61, 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, + 0x64, 0x61, 0x74, 0x61, 0x3a, 0x20, 0x5b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7b, 0x20, 0x64, 0x61, 0x74, 0x61, 0x3a, 0x20, 0x69, 0x6d, 0x61, 0x67, + 0x65, 0x5f, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, + 0x63, 0x65, 0x28, 0x2f, 0x64, 0x61, 0x74, 0x61, 0x3a, 0x69, 0x6d, 0x61, + 0x67, 0x65, 0x5c, 0x2f, 0x5b, 0x5e, 0x3b, 0x5d, 0x2b, 0x3b, 0x62, 0x61, + 0x73, 0x65, 0x36, 0x34, 0x2c, 0x2f, 0x2c, 0x20, 0x27, 0x27, 0x29, 0x2c, + 0x20, 0x69, 0x64, 0x3a, 0x20, 0x31, 0x30, 0x20, 0x7d, 0x5d, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x5f, 0x69, 0x6d, 0x61, + 0x67, 0x65, 0x20, 0x3d, 0x20, 0x74, 0x72, 0x75, 0x65, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x61, + 0x64, 0x65, 0x72, 0x2e, 0x72, 0x65, 0x61, 0x64, 0x41, 0x73, 0x44, 0x61, + 0x74, 0x61, 0x55, 0x52, 0x4c, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, + 0x65, 0x64, 0x46, 0x69, 0x6c, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, + 0x20, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x49, 0x6e, 0x70, 0x75, + 0x74, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, + 0x65, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, 0x6e, 0x61, + 0x6c, 0x28, 0x22, 0x22, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x75, 0x62, 0x6d, 0x69, + 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, + 0x70, 0x28, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x68, 0x61, 0x74, 0x28, 0x6d, 0x65, 0x73, 0x73, 0x61, + 0x67, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, + 0x67, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x22, + 0x22, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x65, 0x6e, 0x74, 0x65, 0x72, 0x53, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x73, + 0x20, 0x3d, 0x20, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x3d, + 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x68, + 0x69, 0x63, 0x68, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x31, 0x33, 0x20, 0x26, + 0x26, 0x20, 0x21, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x68, 0x69, + 0x66, 0x74, 0x4b, 0x65, 0x79, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x75, 0x62, 0x6d, 0x69, + 0x74, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, + 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x6f, 0x72, + 0x6d, 0x20, 0x6f, 0x6e, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x3d, 0x24, + 0x7b, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x7d, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x4e, 0x61, 0x6d, 0x65, 0x3d, 0x24, 0x7b, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6e, - 0x67, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x3e, 0x53, 0x65, 0x6e, - 0x64, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, + 0x67, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3f, 0x20, 0x22, 0x6c, + 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x22, 0x20, 0x3a, 0x20, 0x6e, 0x75, + 0x6c, 0x6c, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, + 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, + 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x20, 0x3d, 0x20, 0x65, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6f, + 0x6e, 0x6b, 0x65, 0x79, 0x70, 0x72, 0x65, 0x73, 0x73, 0x3d, 0x24, 0x7b, + 0x65, 0x6e, 0x74, 0x65, 0x72, 0x53, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x73, + 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x68, 0x6f, + 0x6c, 0x64, 0x65, 0x72, 0x3d, 0x22, 0x53, 0x61, 0x79, 0x20, 0x73, 0x6f, + 0x6d, 0x65, 0x74, 0x68, 0x69, 0x6e, 0x67, 0x2e, 0x2e, 0x2e, 0x22, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x3d, 0x32, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, + 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x22, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, 0x63, 0x6c, + 0x61, 0x73, 0x73, 0x3d, 0x22, 0x72, 0x69, 0x67, 0x68, 0x74, 0x22, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x74, 0x79, 0x70, + 0x65, 0x3d, 0x22, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x22, 0x20, 0x64, + 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x67, 0x65, + 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x7d, 0x3e, 0x53, 0x65, 0x6e, 0x64, 0x3c, 0x2f, 0x62, 0x75, + 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, + 0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, + 0x75, 0x70, 0x6c, 0x6f, 0x61, 0x64, 0x49, 0x6d, 0x61, 0x67, 0x65, 0x7d, + 0x3e, 0x55, 0x70, 0x6c, 0x6f, 0x61, 0x64, 0x20, 0x49, 0x6d, 0x61, 0x67, + 0x65, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x73, 0x74, 0x6f, 0x70, 0x7d, 0x20, 0x64, @@ -1238,28 +1380,39 @@ unsigned char index_html[] = { 0x74, 0x65, 0x78, 0x74, 0x3d, 0x24, 0x7b, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x74, 0x65, 0x78, 0x74, 0x29, 0x7d, 0x20, 0x2f, 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x28, - 0x75, 0x73, 0x65, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, + 0x28, 0x75, 0x73, 0x65, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, + 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x70, 0x20, 0x6b, 0x65, + 0x79, 0x3d, 0x24, 0x7b, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x7d, 0x3e, 0x3c, + 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x7b, 0x74, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x75, 0x73, 0x65, 0x72, 0x29, 0x7d, + 0x3a, 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x20, 0x24, + 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x3c, 0x2f, 0x70, + 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x70, 0x20, 0x6b, 0x65, 0x79, - 0x3d, 0x24, 0x7b, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x7d, 0x3e, 0x3c, 0x73, - 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x7b, 0x74, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x28, 0x75, 0x73, 0x65, 0x72, 0x29, 0x7d, 0x3a, - 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x20, 0x24, 0x7b, + 0x3d, 0x24, 0x7b, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x7d, 0x3e, 0x24, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x3c, 0x2f, 0x70, 0x3e, - 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, - 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, - 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x70, 0x20, 0x6b, 0x65, 0x79, 0x3d, - 0x24, 0x7b, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x7d, 0x3e, 0x24, 0x7b, 0x6d, - 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x3c, 0x2f, 0x70, 0x3e, 0x60, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, - 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x73, 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x69, 0x64, 0x3d, - 0x22, 0x63, 0x68, 0x61, 0x74, 0x22, 0x20, 0x72, 0x65, 0x66, 0x3d, 0x24, - 0x7b, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x7d, 0x3e, + 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, + 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x73, 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x69, 0x64, + 0x3d, 0x22, 0x63, 0x68, 0x61, 0x74, 0x22, 0x20, 0x72, 0x65, 0x66, 0x3d, + 0x24, 0x7b, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x7d, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x69, 0x6d, 0x67, 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x22, + 0x77, 0x69, 0x64, 0x74, 0x68, 0x3a, 0x20, 0x36, 0x30, 0x25, 0x3b, 0x24, + 0x7b, 0x21, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2e, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, + 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x20, 0x3f, 0x20, 0x60, 0x64, 0x69, + 0x73, 0x70, 0x6c, 0x61, 0x79, 0x3a, 0x20, 0x6e, 0x6f, 0x6e, 0x65, 0x3b, + 0x60, 0x20, 0x3a, 0x20, 0x60, 0x60, 0x7d, 0x22, 0x20, 0x73, 0x72, 0x63, + 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, + 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x7d, 0x22, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x66, 0x6c, 0x61, 0x74, 0x4d, 0x61, 0x70, 0x28, 0x63, 0x68, 0x61, 0x74, 0x4c, 0x69, @@ -1344,703 +1497,704 @@ unsigned char index_html[] = { 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x64, 0x75, 0x63, 0x65, 0x28, 0x28, 0x61, 0x63, 0x63, 0x2c, 0x20, 0x63, 0x75, 0x72, 0x2c, 0x20, 0x69, 0x29, 0x20, 0x3d, - 0x3e, 0x20, 0x28, 0x7b, 0x2e, 0x2e, 0x2e, 0x61, 0x63, 0x63, 0x2c, 0x20, - 0x5b, 0x63, 0x75, 0x72, 0x2e, 0x74, 0x72, 0x69, 0x6d, 0x28, 0x29, 0x5d, - 0x3a, 0x20, 0x69, 0x7d, 0x29, 0x2c, 0x20, 0x7b, 0x7d, 0x29, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, - 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x2e, 0x76, 0x69, 0x73, 0x69, 0x74, - 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2c, 0x20, 0x27, 0x27, 0x29, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, - 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, - 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, - 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, 0x72, 0x61, - 0x6d, 0x6d, 0x61, 0x72, 0x3a, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, - 0x74, 0x65, 0x72, 0x2e, 0x66, 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x47, 0x72, - 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x28, 0x29, 0x2c, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x63, 0x61, 0x74, 0x63, 0x68, - 0x20, 0x28, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x61, 0x6c, 0x65, 0x72, 0x74, 0x28, 0x60, - 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x20, 0x66, 0x61, 0x69, 0x6c, - 0x65, 0x64, 0x3a, 0x20, 0x24, 0x7b, 0x65, 0x2e, 0x6d, 0x65, 0x73, 0x73, - 0x61, 0x67, 0x65, 0x7d, 0x60, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, - 0x74, 0x20, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, - 0x20, 0x3d, 0x20, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x2c, 0x20, - 0x6d, 0x61, 0x78, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x2c, 0x20, 0x6e, 0x61, - 0x6d, 0x65, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x2c, 0x20, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, - 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, - 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x3e, 0x24, 0x7b, 0x6c, - 0x61, 0x62, 0x65, 0x6c, 0x7d, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, - 0x65, 0x3d, 0x22, 0x72, 0x61, 0x6e, 0x67, 0x65, 0x22, 0x20, 0x69, 0x64, - 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x6d, - 0x69, 0x6e, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x69, 0x6e, 0x7d, 0x22, 0x20, - 0x6d, 0x61, 0x78, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x61, 0x78, 0x7d, 0x22, - 0x20, 0x73, 0x74, 0x65, 0x70, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x74, 0x65, - 0x70, 0x7d, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x24, 0x7b, - 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x3d, 0x22, 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x22, 0x20, - 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, - 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x46, 0x6c, - 0x6f, 0x61, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, - 0x6e, 0x3e, 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x3c, 0x2f, - 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x49, 0x6e, 0x74, 0x46, 0x69, - 0x65, 0x6c, 0x64, 0x20, 0x3d, 0x20, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, + 0x3e, 0x20, 0x28, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x61, 0x63, 0x63, 0x2c, + 0x20, 0x5b, 0x63, 0x75, 0x72, 0x2e, 0x74, 0x72, 0x69, 0x6d, 0x28, 0x29, + 0x5d, 0x3a, 0x20, 0x69, 0x20, 0x7d, 0x29, 0x2c, 0x20, 0x7b, 0x7d, 0x29, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x2e, 0x76, 0x69, 0x73, + 0x69, 0x74, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2c, 0x20, 0x27, + 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, + 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x3a, 0x20, 0x63, 0x6f, 0x6e, 0x76, + 0x65, 0x72, 0x74, 0x65, 0x72, 0x2e, 0x66, 0x6f, 0x72, 0x6d, 0x61, 0x74, + 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x28, 0x29, 0x2c, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x63, 0x61, 0x74, + 0x63, 0x68, 0x20, 0x28, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x61, 0x6c, 0x65, 0x72, 0x74, + 0x28, 0x60, 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x20, 0x66, 0x61, + 0x69, 0x6c, 0x65, 0x64, 0x3a, 0x20, 0x24, 0x7b, 0x65, 0x2e, 0x6d, 0x65, + 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x60, 0x29, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, + 0x6c, 0x64, 0x20, 0x3d, 0x20, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x2c, - 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, - 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, - 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x24, 0x7b, 0x6e, - 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x3e, 0x24, 0x7b, 0x6c, 0x61, 0x62, 0x65, - 0x6c, 0x7d, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, - 0x72, 0x61, 0x6e, 0x67, 0x65, 0x22, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x24, - 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x6d, 0x69, 0x6e, 0x3d, - 0x22, 0x24, 0x7b, 0x6d, 0x69, 0x6e, 0x7d, 0x22, 0x20, 0x6d, 0x61, 0x78, - 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x61, 0x78, 0x7d, 0x22, 0x20, 0x6e, 0x61, - 0x6d, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, - 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, - 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, - 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, + 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x2c, + 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x29, 0x20, 0x3d, 0x3e, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, + 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, + 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, + 0x6f, 0x72, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, + 0x3e, 0x24, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x7d, 0x3c, 0x2f, 0x6c, + 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, + 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x6e, 0x67, 0x65, + 0x22, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, + 0x7d, 0x22, 0x20, 0x6d, 0x69, 0x6e, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x69, + 0x6e, 0x7d, 0x22, 0x20, 0x6d, 0x61, 0x78, 0x3d, 0x22, 0x24, 0x7b, 0x6d, + 0x61, 0x78, 0x7d, 0x22, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3d, 0x22, 0x24, + 0x7b, 0x73, 0x74, 0x65, 0x70, 0x7d, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, + 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, + 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, + 0x6d, 0x73, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, - 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, - 0x65, 0x73, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, - 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x65, 0x2e, 0x70, 0x72, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x44, 0x65, 0x66, - 0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, - 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, - 0x6c, 0x79, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, - 0x74, 0x20, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x42, 0x75, 0x74, 0x74, 0x6f, - 0x6e, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, - 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, - 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x3d, 0x20, 0x27, - 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x27, 0x29, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x49, + 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x20, 0x3d, 0x20, 0x28, 0x7b, + 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x2c, + 0x20, 0x6d, 0x69, 0x6e, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, - 0x6c, 0x65, 0x64, 0x3e, 0x55, 0x73, 0x69, 0x6e, 0x67, 0x20, 0x64, 0x65, - 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, - 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, - 0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, - 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, - 0x52, 0x65, 0x73, 0x65, 0x74, 0x7d, 0x3e, 0x52, 0x65, 0x73, 0x65, 0x74, - 0x20, 0x61, 0x6c, 0x6c, 0x20, 0x74, 0x6f, 0x20, 0x64, 0x65, 0x66, 0x61, - 0x75, 0x6c, 0x74, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, - 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x61, 0x75, 0x74, 0x6f, - 0x73, 0x61, 0x76, 0x65, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x20, 0x6f, 0x6e, 0x20, 0x65, 0x76, 0x65, 0x72, 0x79, 0x20, 0x63, - 0x68, 0x61, 0x6e, 0x67, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x28, 0x29, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x5b, 0x73, - 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x5d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, - 0x72, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x20, 0x3d, 0x20, 0x28, - 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, - 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x3e, 0x47, - 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, - 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, - 0x20, 0x69, 0x64, 0x3d, 0x22, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, - 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x67, 0x72, 0x61, 0x6d, - 0x6d, 0x61, 0x72, 0x22, 0x20, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x68, 0x6f, - 0x6c, 0x64, 0x65, 0x72, 0x3d, 0x22, 0x55, 0x73, 0x65, 0x20, 0x67, 0x62, - 0x6e, 0x66, 0x20, 0x6f, 0x72, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x20, 0x53, - 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2b, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, - 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, - 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x2e, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x7d, 0x22, 0x20, 0x72, - 0x6f, 0x77, 0x73, 0x3d, 0x34, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, - 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, - 0x72, 0x61, 0x6d, 0x73, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, - 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, - 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x70, 0x72, 0x6f, - 0x70, 0x2d, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x22, 0x20, 0x70, 0x6c, 0x61, - 0x63, 0x65, 0x68, 0x6f, 0x6c, 0x64, 0x65, 0x72, 0x3d, 0x22, 0x6f, 0x72, - 0x64, 0x65, 0x72, 0x3a, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x31, 0x2c, 0x70, - 0x72, 0x6f, 0x70, 0x32, 0x2c, 0x70, 0x72, 0x6f, 0x70, 0x33, 0x22, 0x20, - 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, - 0x64, 0x61, 0x74, 0x65, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, - 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, - 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, - 0x22, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x22, 0x20, 0x6f, 0x6e, 0x63, - 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x63, 0x6f, 0x6e, 0x76, 0x65, - 0x72, 0x74, 0x4a, 0x53, 0x4f, 0x4e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, - 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x7d, 0x3e, 0x43, 0x6f, 0x6e, - 0x76, 0x65, 0x72, 0x74, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x20, 0x53, 0x63, - 0x68, 0x65, 0x6d, 0x61, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x43, - 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x53, - 0x65, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x28, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, - 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x68, 0x74, 0x6d, - 0x6c, 0x46, 0x6f, 0x72, 0x3d, 0x22, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, - 0x22, 0x3e, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x3c, 0x2f, 0x6c, 0x61, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, + 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, + 0x72, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x3e, + 0x24, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x7d, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, - 0x65, 0x61, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, - 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x70, 0x72, 0x6f, - 0x6d, 0x70, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, - 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x2e, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x7d, 0x22, - 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, - 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, - 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, - 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x3b, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x43, 0x68, 0x61, 0x74, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, - 0x6f, 0x72, 0x6d, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, - 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, - 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x24, 0x7b, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x43, 0x6f, - 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x53, 0x65, - 0x74, 0x28, 0x29, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, - 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, - 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x75, - 0x73, 0x65, 0x72, 0x22, 0x3e, 0x55, 0x73, 0x65, 0x72, 0x20, 0x6e, 0x61, - 0x6d, 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, - 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, - 0x3d, 0x22, 0x75, 0x73, 0x65, 0x72, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x75, 0x73, 0x65, 0x72, 0x7d, - 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, - 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, - 0x6e, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, - 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x62, 0x6f, - 0x74, 0x22, 0x3e, 0x42, 0x6f, 0x74, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3c, - 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, - 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, - 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x63, - 0x68, 0x61, 0x72, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, - 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x2e, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x22, 0x20, 0x6f, - 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, - 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x20, - 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, - 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, - 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, - 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x3e, - 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, - 0x69, 0x64, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, - 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, - 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x7d, 0x22, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x3d, 0x34, 0x20, 0x6f, - 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, - 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, - 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, - 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x22, 0x3e, 0x43, 0x68, 0x61, 0x74, 0x20, 0x68, 0x69, 0x73, - 0x74, 0x6f, 0x72, 0x79, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x69, 0x64, - 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, - 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, - 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, - 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x68, 0x69, - 0x73, 0x74, 0x6f, 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x7d, 0x22, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x3d, 0x31, 0x20, 0x6f, - 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, - 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x47, - 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, - 0x6c, 0x28, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, - 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6c, - 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, - 0x6f, 0x72, 0x6d, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, - 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, - 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x24, 0x7b, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x43, 0x6f, - 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x53, 0x65, - 0x74, 0x28, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, - 0x3e, 0x24, 0x7b, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x43, 0x6f, - 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x28, 0x29, 0x7d, 0x3c, 0x2f, 0x66, 0x69, - 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, - 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x6f, 0x72, 0x6d, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, - 0x61, 0x73, 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x24, 0x7b, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x42, 0x75, 0x74, 0x74, 0x6f, - 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x63, 0x6c, 0x61, 0x73, - 0x73, 0x3d, 0x22, 0x73, 0x6c, 0x69, 0x6d, 0x22, 0x3e, 0x3c, 0x69, 0x6e, - 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, - 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x74, - 0x79, 0x70, 0x65, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, - 0x63, 0x68, 0x61, 0x74, 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, - 0x64, 0x3d, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x20, 0x3d, - 0x3d, 0x3d, 0x20, 0x22, 0x63, 0x68, 0x61, 0x74, 0x22, 0x7d, 0x20, 0x6f, - 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, - 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x20, - 0x2f, 0x3e, 0x20, 0x43, 0x68, 0x61, 0x74, 0x3c, 0x2f, 0x6c, 0x61, 0x62, - 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, - 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x73, 0x6c, 0x69, 0x6d, - 0x22, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, - 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e, 0x61, - 0x6d, 0x65, 0x3d, 0x22, 0x74, 0x79, 0x70, 0x65, 0x22, 0x20, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, - 0x69, 0x6f, 0x6e, 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, - 0x3d, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, - 0x3d, 0x20, 0x22, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, - 0x6e, 0x22, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, - 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, - 0x69, 0x6f, 0x6e, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x43, 0x6f, 0x6d, 0x70, - 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, - 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, - 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x73, 0x65, 0x73, - 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, - 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x63, 0x68, 0x61, - 0x74, 0x27, 0x20, 0x3f, 0x20, 0x43, 0x68, 0x61, 0x74, 0x43, 0x6f, 0x6e, - 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x28, 0x29, 0x20, 0x3a, 0x20, - 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x43, 0x6f, - 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x28, 0x29, 0x7d, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, - 0x73, 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, - 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, - 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x50, 0x72, 0x65, 0x64, 0x69, 0x63, - 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, - 0x20, 0x32, 0x30, 0x34, 0x38, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, - 0x2d, 0x31, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x6e, - 0x5f, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x22, 0x2c, 0x20, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x5f, 0x70, 0x72, 0x65, - 0x64, 0x69, 0x63, 0x74, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, - 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, - 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x65, 0x6d, 0x70, 0x65, 0x72, - 0x61, 0x74, 0x75, 0x72, 0x65, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, - 0x20, 0x31, 0x2e, 0x35, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, - 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, - 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x22, 0x2c, - 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, - 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, - 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x65, 0x6d, - 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x7d, 0x29, 0x7d, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, - 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x50, 0x65, - 0x6e, 0x61, 0x6c, 0x69, 0x7a, 0x65, 0x20, 0x72, 0x65, 0x70, 0x65, 0x61, - 0x74, 0x20, 0x73, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x65, 0x22, 0x2c, - 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x32, 0x2e, 0x30, 0x2c, 0x20, 0x6d, - 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, - 0x65, 0x3a, 0x20, 0x22, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x70, - 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, - 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, - 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x7d, 0x29, 0x7d, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, - 0x7b, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, - 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x43, 0x6f, 0x6e, 0x73, 0x69, - 0x64, 0x65, 0x72, 0x20, 0x4e, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x73, - 0x20, 0x66, 0x6f, 0x72, 0x20, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x69, 0x7a, - 0x65, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x32, 0x30, 0x34, - 0x38, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x6e, - 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, - 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x22, 0x2c, 0x20, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, - 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x7d, 0x29, 0x7d, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, - 0x7b, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, - 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x6f, 0x70, 0x2d, 0x4b, - 0x20, 0x73, 0x61, 0x6d, 0x70, 0x6c, 0x69, 0x6e, 0x67, 0x22, 0x2c, 0x20, - 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x30, 0x30, 0x2c, 0x20, 0x6d, 0x69, - 0x6e, 0x3a, 0x20, 0x2d, 0x31, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, - 0x20, 0x22, 0x74, 0x6f, 0x70, 0x5f, 0x6b, 0x22, 0x2c, 0x20, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x6f, 0x70, 0x5f, 0x6b, 0x7d, - 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, - 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, - 0x22, 0x54, 0x6f, 0x70, 0x2d, 0x50, 0x20, 0x73, 0x61, 0x6d, 0x70, 0x6c, - 0x69, 0x6e, 0x67, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, - 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, - 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x6f, 0x70, - 0x5f, 0x70, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, - 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, - 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x2e, 0x74, 0x6f, 0x70, 0x5f, 0x70, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, - 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x65, 0x74, 0x61, 0x69, - 0x6c, 0x73, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, - 0x3e, 0x4d, 0x6f, 0x72, 0x65, 0x20, 0x6f, 0x70, 0x74, 0x69, 0x6f, 0x6e, - 0x73, 0x3c, 0x2f, 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, - 0x61, 0x73, 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, - 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, - 0x46, 0x53, 0x2d, 0x5a, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, - 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, - 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x66, - 0x73, 0x5f, 0x7a, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, - 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, - 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x2e, 0x74, 0x66, 0x73, 0x5f, 0x7a, 0x7d, 0x29, 0x7d, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, - 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, - 0x79, 0x70, 0x69, 0x63, 0x61, 0x6c, 0x20, 0x50, 0x22, 0x2c, 0x20, 0x6d, - 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, - 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, - 0x20, 0x22, 0x74, 0x79, 0x70, 0x69, 0x63, 0x61, 0x6c, 0x5f, 0x70, 0x22, - 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, - 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, - 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, - 0x70, 0x69, 0x63, 0x61, 0x6c, 0x5f, 0x70, 0x7d, 0x29, 0x7d, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, - 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x50, - 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x20, 0x70, 0x65, 0x6e, 0x61, - 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, - 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, - 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x70, 0x72, 0x65, - 0x73, 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, - 0x79, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, - 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, - 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, - 0x70, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, 0x65, 0x6e, - 0x61, 0x6c, 0x74, 0x79, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, - 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, - 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x46, 0x72, 0x65, 0x71, - 0x75, 0x65, 0x6e, 0x63, 0x79, 0x20, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, - 0x79, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, - 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, - 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x66, 0x72, 0x65, 0x71, 0x75, - 0x65, 0x6e, 0x63, 0x79, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, - 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, - 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, - 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x66, - 0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, 0x5f, 0x70, 0x65, 0x6e, - 0x61, 0x6c, 0x74, 0x79, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, - 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x68, 0x72, 0x20, - 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, - 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x74, 0x68, 0x72, 0x65, - 0x65, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x3c, - 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, - 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, - 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x22, 0x20, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x30, 0x22, 0x20, 0x63, 0x68, 0x65, - 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, - 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, - 0x73, 0x74, 0x61, 0x74, 0x20, 0x3d, 0x3d, 0x20, 0x30, 0x7d, 0x20, 0x6f, - 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, - 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, - 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x6e, 0x6f, 0x20, 0x4d, 0x69, 0x72, 0x6f, - 0x73, 0x74, 0x61, 0x74, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, - 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, - 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, - 0x3d, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x22, 0x20, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x31, 0x22, 0x20, 0x63, 0x68, - 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, - 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, - 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x3d, 0x3d, 0x20, 0x31, 0x7d, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, + 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x6e, 0x67, 0x65, 0x22, + 0x20, 0x69, 0x64, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, + 0x22, 0x20, 0x6d, 0x69, 0x6e, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x69, 0x6e, + 0x7d, 0x22, 0x20, 0x6d, 0x61, 0x78, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x61, + 0x78, 0x7d, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x24, 0x7b, + 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x3d, 0x22, 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, - 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, - 0x61, 0x74, 0x20, 0x76, 0x31, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, - 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, - 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e, 0x61, 0x6d, - 0x65, 0x3d, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x22, - 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x32, 0x22, 0x20, 0x63, - 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x70, 0x61, 0x72, - 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, - 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x3d, 0x3d, 0x20, 0x32, 0x7d, - 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, - 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, - 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x4d, 0x69, 0x72, 0x6f, 0x73, - 0x74, 0x61, 0x74, 0x20, 0x76, 0x32, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, - 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, - 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, - 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x74, 0x61, 0x75, - 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x30, 0x2e, 0x30, - 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, - 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, - 0x74, 0x61, 0x74, 0x5f, 0x74, 0x61, 0x75, 0x22, 0x2c, 0x20, 0x73, 0x74, - 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, - 0x61, 0x74, 0x5f, 0x74, 0x61, 0x75, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, - 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x4d, 0x69, - 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x65, 0x74, 0x61, 0x22, 0x2c, - 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, - 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, - 0x65, 0x3a, 0x20, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, - 0x5f, 0x65, 0x74, 0x61, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, - 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, - 0x65, 0x74, 0x61, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, - 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, - 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, - 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, - 0x6c, 0x3a, 0x20, 0x22, 0x53, 0x68, 0x6f, 0x77, 0x20, 0x50, 0x72, 0x6f, - 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x22, 0x2c, - 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x30, 0x2c, 0x20, 0x6d, 0x69, - 0x6e, 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, - 0x22, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x22, 0x2c, 0x20, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x5f, 0x70, 0x72, 0x6f, - 0x62, 0x73, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, - 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, - 0x73, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x2f, 0x66, 0x6f, 0x72, 0x6d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x62, - 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x29, 0x20, - 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x20, 0x3d, 0x20, 0x4d, 0x61, 0x74, - 0x68, 0x2e, 0x66, 0x6c, 0x6f, 0x6f, 0x72, 0x28, 0x31, 0x39, 0x32, 0x20, - 0x2a, 0x20, 0x28, 0x31, 0x20, 0x2d, 0x20, 0x70, 0x29, 0x29, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, - 0x67, 0x20, 0x3d, 0x20, 0x4d, 0x61, 0x74, 0x68, 0x2e, 0x66, 0x6c, 0x6f, - 0x6f, 0x72, 0x28, 0x31, 0x39, 0x32, 0x20, 0x2a, 0x20, 0x70, 0x29, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, - 0x6e, 0x20, 0x60, 0x72, 0x67, 0x62, 0x61, 0x28, 0x24, 0x7b, 0x72, 0x7d, - 0x2c, 0x24, 0x7b, 0x67, 0x7d, 0x2c, 0x30, 0x2c, 0x30, 0x2e, 0x33, 0x29, - 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x50, 0x72, 0x6f, 0x62, - 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x20, 0x3d, 0x20, - 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, - 0x72, 0x6e, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x64, 0x61, - 0x74, 0x61, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20, 0x3d, - 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6d, 0x70, - 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, - 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x20, 0x7d, 0x20, 0x3d, - 0x20, 0x6d, 0x73, 0x67, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x21, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, - 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, - 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x20, 0x7c, 0x7c, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, - 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, - 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, - 0x67, 0x74, 0x68, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x30, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x20, 0x72, 0x65, 0x74, 0x75, - 0x72, 0x6e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, - 0x6e, 0x74, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, - 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, - 0x74, 0x69, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, - 0x3e, 0x20, 0x31, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x4e, 0x6f, 0x74, 0x20, - 0x66, 0x6f, 0x72, 0x20, 0x62, 0x79, 0x74, 0x65, 0x20, 0x70, 0x61, 0x69, - 0x72, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, - 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, - 0x74, 0x69, 0x65, 0x73, 0x5b, 0x30, 0x5d, 0x2e, 0x63, 0x6f, 0x6e, 0x74, - 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x74, 0x61, 0x72, 0x74, 0x73, 0x57, 0x69, - 0x74, 0x68, 0x28, 0x27, 0x62, 0x79, 0x74, 0x65, 0x3a, 0x20, 0x5c, 0x5c, - 0x27, 0x29, 0x29, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6d, - 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, - 0x6e, 0x73, 0x74, 0x20, 0x73, 0x70, 0x6c, 0x69, 0x74, 0x44, 0x61, 0x74, - 0x61, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, - 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, - 0x74, 0x69, 0x65, 0x73, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x70, 0x72, 0x6f, - 0x62, 0x20, 0x3d, 0x3e, 0x20, 0x28, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, - 0x65, 0x6e, 0x74, 0x3a, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x2e, 0x63, 0x6f, - 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, - 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, - 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x3a, 0x20, 0x5b, 0x70, 0x72, - 0x6f, 0x62, 0x5d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x29, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, + 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x3c, 0x2f, 0x73, 0x70, + 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x20, 0x3d, + 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x2e, 0x70, 0x72, 0x65, 0x76, + 0x65, 0x6e, 0x74, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, + 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, + 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, + 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x28, 0x29, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x55, 0x73, 0x65, 0x72, + 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, + 0x74, 0x42, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x28, 0x29, + 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, + 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, + 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x61, 0x6d, + 0x65, 0x20, 0x3d, 0x3d, 0x20, 0x27, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, + 0x74, 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, - 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x24, 0x7b, 0x50, 0x72, 0x6f, 0x62, 0x61, - 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x7d, 0x20, 0x64, 0x61, - 0x74, 0x61, 0x3d, 0x24, 0x7b, 0x73, 0x70, 0x6c, 0x69, 0x74, 0x44, 0x61, - 0x74, 0x61, 0x7d, 0x20, 0x2f, 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x7b, 0x20, 0x70, - 0x72, 0x6f, 0x62, 0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, - 0x74, 0x20, 0x7d, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, - 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, - 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x5b, 0x30, 0x5d, 0x0a, 0x20, 0x20, + 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, + 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x3e, 0x55, 0x73, + 0x69, 0x6e, 0x67, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, + 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3c, 0x2f, 0x62, 0x75, + 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, + 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, + 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x7d, + 0x3e, 0x52, 0x65, 0x73, 0x65, 0x74, 0x20, 0x61, 0x6c, 0x6c, 0x20, 0x74, + 0x6f, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x3c, 0x2f, 0x62, + 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, + 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, + 0x2f, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x20, 0x74, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x6f, 0x6e, 0x20, 0x65, + 0x76, 0x65, 0x72, 0x79, 0x20, 0x63, 0x68, 0x61, 0x6e, 0x67, 0x65, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, + 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x41, 0x75, 0x74, 0x6f, + 0x73, 0x61, 0x76, 0x65, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x2c, 0x20, 0x5b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, + 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x5d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, - 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x20, 0x3d, 0x20, 0x70, 0x72, 0x6f, 0x62, - 0x73, 0x2e, 0x66, 0x69, 0x6e, 0x64, 0x28, 0x70, 0x20, 0x3d, 0x3e, 0x20, - 0x70, 0x2e, 0x74, 0x6f, 0x6b, 0x5f, 0x73, 0x74, 0x72, 0x20, 0x3d, 0x3d, - 0x3d, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, - 0x74, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x20, - 0x3d, 0x20, 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x20, 0x3f, 0x20, 0x70, 0x72, - 0x6f, 0x62, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x28, 0x66, 0x6f, 0x75, 0x6e, - 0x64, 0x2e, 0x70, 0x72, 0x6f, 0x62, 0x29, 0x20, 0x3a, 0x20, 0x27, 0x74, - 0x72, 0x61, 0x6e, 0x73, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x27, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, - 0x73, 0x74, 0x20, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x43, 0x68, - 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x20, 0x3d, 0x20, 0x68, 0x74, 0x6d, + 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x43, 0x6f, 0x6e, 0x74, 0x72, + 0x6f, 0x6c, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x28, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, - 0x22, 0x70, 0x72, 0x6f, 0x62, 0x2d, 0x73, 0x65, 0x74, 0x22, 0x3e, 0x0a, + 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, + 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, + 0x61, 0x74, 0x65, 0x22, 0x3e, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, + 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, + 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x67, + 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, + 0x3d, 0x22, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x22, 0x20, 0x70, + 0x6c, 0x61, 0x63, 0x65, 0x68, 0x6f, 0x6c, 0x64, 0x65, 0x72, 0x3d, 0x22, + 0x55, 0x73, 0x65, 0x20, 0x67, 0x62, 0x6e, 0x66, 0x20, 0x6f, 0x72, 0x20, + 0x4a, 0x53, 0x4f, 0x4e, 0x20, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2b, + 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x67, 0x72, 0x61, 0x6d, 0x6d, + 0x61, 0x72, 0x7d, 0x22, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x3d, 0x34, 0x20, + 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, + 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x7d, 0x2f, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, + 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, + 0x65, 0x3d, 0x22, 0x70, 0x72, 0x6f, 0x70, 0x2d, 0x6f, 0x72, 0x64, 0x65, + 0x72, 0x22, 0x20, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x68, 0x6f, 0x6c, 0x64, + 0x65, 0x72, 0x3d, 0x22, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x3a, 0x20, 0x70, + 0x72, 0x6f, 0x70, 0x31, 0x2c, 0x70, 0x72, 0x6f, 0x70, 0x32, 0x2c, 0x70, + 0x72, 0x6f, 0x70, 0x33, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, + 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x47, 0x72, + 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, + 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, + 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, + 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x62, 0x75, 0x74, 0x74, 0x6f, + 0x6e, 0x22, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, + 0x7b, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x4a, 0x53, 0x4f, 0x4e, + 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, + 0x72, 0x7d, 0x3e, 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x20, 0x4a, + 0x53, 0x4f, 0x4e, 0x20, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x3c, 0x2f, + 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x3b, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x50, + 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, + 0x46, 0x69, 0x65, 0x6c, 0x64, 0x53, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x28, + 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, + 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, + 0x65, 0x6c, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x46, 0x6f, 0x72, 0x3d, 0x22, + 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x22, 0x3e, 0x50, 0x72, 0x6f, 0x6d, + 0x70, 0x74, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x74, 0x79, 0x70, + 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, + 0x65, 0x3d, 0x22, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x22, 0x20, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, + 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x70, 0x72, + 0x6f, 0x6d, 0x70, 0x74, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, + 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, + 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, + 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x43, 0x68, 0x61, 0x74, 0x43, + 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x20, 0x3d, 0x20, + 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x50, 0x72, + 0x6f, 0x6d, 0x70, 0x74, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x46, + 0x69, 0x65, 0x6c, 0x64, 0x53, 0x65, 0x74, 0x28, 0x29, 0x7d, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, + 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, + 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, + 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, + 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x75, 0x73, 0x65, 0x72, 0x22, 0x3e, 0x55, + 0x73, 0x65, 0x72, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3c, 0x2f, 0x6c, 0x61, + 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, + 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, + 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x75, 0x73, 0x65, 0x72, + 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, + 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x2e, 0x75, 0x73, 0x65, 0x72, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, + 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, + 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x2e, 0x6d, 0x61, 0x70, 0x28, - 0x28, 0x70, 0x2c, 0x20, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x29, 0x20, 0x3d, - 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, + 0x6f, 0x72, 0x3d, 0x22, 0x62, 0x6f, 0x74, 0x22, 0x3e, 0x42, 0x6f, 0x74, + 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, + 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, + 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x63, 0x68, 0x61, 0x72, 0x22, 0x20, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, + 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x63, 0x68, + 0x61, 0x72, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, + 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, + 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, + 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, + 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, + 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, + 0x6c, 0x61, 0x74, 0x65, 0x22, 0x3e, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, + 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3c, 0x2f, 0x6c, + 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, + 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x74, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, + 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, + 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x7d, 0x22, 0x20, 0x72, 0x6f, + 0x77, 0x73, 0x3d, 0x34, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, + 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, + 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, + 0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, + 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x3e, 0x43, 0x68, + 0x61, 0x74, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x20, 0x74, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, + 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, + 0x72, 0x65, 0x61, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, + 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, + 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, + 0x61, 0x74, 0x65, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, + 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x54, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x7d, 0x22, 0x20, 0x72, 0x6f, + 0x77, 0x73, 0x3d, 0x31, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, + 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, + 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, + 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x24, 0x7b, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, + 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x28, 0x29, 0x7d, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, + 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, + 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, + 0x6e, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x20, + 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x28, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, + 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, + 0x6c, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x53, 0x65, 0x74, 0x28, 0x29, 0x7d, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x24, 0x7b, 0x47, + 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, + 0x6c, 0x28, 0x29, 0x7d, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, + 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x3c, 0x66, 0x6f, 0x72, 0x6d, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, + 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, + 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x55, 0x73, + 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, + 0x73, 0x65, 0x74, 0x42, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, + 0x62, 0x65, 0x6c, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x73, + 0x6c, 0x69, 0x6d, 0x22, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, + 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, + 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x74, 0x79, 0x70, 0x65, 0x22, + 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x63, 0x68, 0x61, 0x74, + 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b, + 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x22, + 0x63, 0x68, 0x61, 0x74, 0x22, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, + 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, + 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x43, + 0x68, 0x61, 0x74, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x63, 0x6c, 0x61, + 0x73, 0x73, 0x3d, 0x22, 0x73, 0x6c, 0x69, 0x6d, 0x22, 0x3e, 0x3c, 0x69, + 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, + 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, + 0x74, 0x79, 0x70, 0x65, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, + 0x22, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x22, + 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x73, + 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x2e, 0x74, 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x22, 0x63, + 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x7d, 0x20, + 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, + 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, + 0x20, 0x2f, 0x3e, 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, + 0x6f, 0x6e, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x64, 0x69, 0x76, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6b, 0x65, - 0x79, 0x3d, 0x24, 0x7b, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x7d, 0x0a, 0x20, + 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, + 0x65, 0x74, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x20, + 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x63, 0x68, 0x61, 0x74, 0x27, 0x20, 0x3f, + 0x20, 0x43, 0x68, 0x61, 0x74, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, + 0x6f, 0x72, 0x6d, 0x28, 0x29, 0x20, 0x3a, 0x20, 0x43, 0x6f, 0x6d, 0x70, + 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, + 0x46, 0x6f, 0x72, 0x6d, 0x28, 0x29, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, + 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, + 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, 0x46, + 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, + 0x3a, 0x20, 0x22, 0x50, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x69, 0x6f, + 0x6e, 0x73, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x32, 0x30, + 0x34, 0x38, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x2d, 0x31, 0x2c, + 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x6e, 0x5f, 0x70, 0x72, + 0x65, 0x64, 0x69, 0x63, 0x74, 0x22, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x5f, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, + 0x74, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, + 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, + 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, + 0x74, 0x75, 0x72, 0x65, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, + 0x31, 0x2e, 0x35, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, + 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x65, + 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x22, 0x2c, 0x20, + 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, + 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x65, 0x6d, 0x70, + 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x69, 0x74, 0x6c, 0x65, 0x3d, 0x24, - 0x7b, 0x60, 0x70, 0x72, 0x6f, 0x62, 0x3a, 0x20, 0x24, 0x7b, 0x70, 0x2e, - 0x70, 0x72, 0x6f, 0x62, 0x7d, 0x60, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, + 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x50, + 0x65, 0x6e, 0x61, 0x6c, 0x69, 0x7a, 0x65, 0x20, 0x72, 0x65, 0x70, 0x65, + 0x61, 0x74, 0x20, 0x73, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x65, 0x22, + 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x32, 0x2e, 0x30, 0x2c, 0x20, + 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, + 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, + 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x73, 0x74, + 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, + 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x20, 0x7d, 0x29, 0x7d, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, + 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x43, 0x6f, + 0x6e, 0x73, 0x69, 0x64, 0x65, 0x72, 0x20, 0x4e, 0x20, 0x74, 0x6f, 0x6b, + 0x65, 0x6e, 0x73, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x70, 0x65, 0x6e, 0x61, + 0x6c, 0x69, 0x7a, 0x65, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, + 0x32, 0x30, 0x34, 0x38, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, + 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x72, 0x65, 0x70, + 0x65, 0x61, 0x74, 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x22, 0x2c, + 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, + 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x72, 0x65, 0x70, + 0x65, 0x61, 0x74, 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x20, 0x7d, + 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, + 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, + 0x54, 0x6f, 0x70, 0x2d, 0x4b, 0x20, 0x73, 0x61, 0x6d, 0x70, 0x6c, 0x69, + 0x6e, 0x67, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x30, + 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x2d, 0x31, 0x2c, 0x20, + 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x6f, 0x70, 0x5f, 0x6b, + 0x22, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, + 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, + 0x6f, 0x70, 0x5f, 0x6b, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, + 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, + 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x6f, 0x70, 0x2d, + 0x50, 0x20, 0x73, 0x61, 0x6d, 0x70, 0x6c, 0x69, 0x6e, 0x67, 0x22, 0x2c, + 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, + 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, + 0x65, 0x3a, 0x20, 0x22, 0x74, 0x6f, 0x70, 0x5f, 0x70, 0x22, 0x2c, 0x20, + 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, + 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x6f, 0x70, 0x5f, + 0x70, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, + 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x73, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x7b, 0x0a, + 0x3c, 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x3e, 0x4d, 0x6f, 0x72, + 0x65, 0x20, 0x6f, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x3c, 0x2f, 0x73, + 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, + 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, + 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, + 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, + 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x46, 0x53, 0x2d, + 0x5a, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, + 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, + 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x66, 0x73, 0x5f, 0x7a, + 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, + 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, + 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, + 0x66, 0x73, 0x5f, 0x7a, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, + 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, + 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x79, + 0x70, 0x69, 0x63, 0x61, 0x6c, 0x20, 0x50, 0x22, 0x2c, 0x20, 0x6d, 0x61, + 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, + 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, + 0x22, 0x74, 0x79, 0x70, 0x69, 0x63, 0x61, 0x6c, 0x5f, 0x70, 0x22, 0x2c, + 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, + 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, + 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, 0x70, + 0x69, 0x63, 0x61, 0x6c, 0x5f, 0x70, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, + 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, + 0x50, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x20, 0x70, 0x65, 0x6e, + 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, + 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, + 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x70, 0x72, + 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, + 0x74, 0x79, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, + 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, + 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x2e, 0x70, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, 0x65, + 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, + 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x46, + 0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, 0x20, 0x70, 0x65, 0x6e, + 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, + 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, + 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x66, 0x72, + 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, 0x5f, 0x70, 0x65, 0x6e, 0x61, + 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, + 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, + 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x66, 0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, 0x5f, + 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x20, 0x7d, 0x29, 0x7d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x68, 0x72, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, + 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, + 0x74, 0x68, 0x72, 0x65, 0x65, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, + 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, + 0x65, 0x6c, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, + 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e, + 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, + 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x30, 0x22, + 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x70, + 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, + 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x3d, 0x3d, 0x20, + 0x30, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, + 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, + 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x6e, 0x6f, 0x20, + 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x3c, 0x2f, 0x6c, 0x61, + 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, + 0x62, 0x65, 0x6c, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, + 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20, + 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, + 0x61, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x31, + 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b, + 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x3d, 0x3d, + 0x20, 0x31, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, + 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, + 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x4d, 0x69, + 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x76, 0x31, 0x3c, 0x2f, 0x6c, + 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, + 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, + 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, + 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, + 0x74, 0x61, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, + 0x32, 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, + 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x3d, + 0x3d, 0x20, 0x32, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, + 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x4d, + 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x76, 0x32, 0x3c, 0x2f, + 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, + 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, + 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, + 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, + 0x74, 0x20, 0x74, 0x61, 0x75, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, + 0x20, 0x31, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, + 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, + 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x74, 0x61, 0x75, + 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, + 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, + 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, + 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x74, 0x61, 0x75, 0x20, + 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, + 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, + 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, + 0x74, 0x20, 0x65, 0x74, 0x61, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, + 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, + 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x6d, + 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x65, 0x74, 0x61, 0x22, + 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, + 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, + 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x65, 0x74, 0x61, 0x20, 0x7d, + 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, + 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, + 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, + 0x22, 0x53, 0x68, 0x6f, 0x77, 0x20, 0x50, 0x72, 0x6f, 0x62, 0x61, 0x62, + 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x22, 0x2c, 0x20, 0x6d, 0x61, + 0x78, 0x3a, 0x20, 0x31, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, + 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x6e, 0x5f, + 0x70, 0x72, 0x6f, 0x62, 0x73, 0x22, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x20, + 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, + 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x73, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, + 0x6f, 0x72, 0x6d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x43, 0x6f, + 0x6c, 0x6f, 0x72, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x29, 0x20, 0x3d, 0x3e, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x72, 0x20, 0x3d, 0x20, 0x4d, 0x61, 0x74, 0x68, 0x2e, + 0x66, 0x6c, 0x6f, 0x6f, 0x72, 0x28, 0x31, 0x39, 0x32, 0x20, 0x2a, 0x20, + 0x28, 0x31, 0x20, 0x2d, 0x20, 0x70, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x67, 0x20, + 0x3d, 0x20, 0x4d, 0x61, 0x74, 0x68, 0x2e, 0x66, 0x6c, 0x6f, 0x6f, 0x72, + 0x28, 0x31, 0x39, 0x32, 0x20, 0x2a, 0x20, 0x70, 0x29, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, + 0x60, 0x72, 0x67, 0x62, 0x61, 0x28, 0x24, 0x7b, 0x72, 0x7d, 0x2c, 0x24, + 0x7b, 0x67, 0x7d, 0x2c, 0x30, 0x2c, 0x30, 0x2e, 0x33, 0x29, 0x60, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x50, 0x72, 0x6f, 0x62, 0x61, 0x62, + 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x28, 0x70, + 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, + 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x64, 0x61, 0x74, 0x61, + 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20, 0x3d, 0x3e, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, + 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, + 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x20, 0x7d, 0x20, 0x3d, 0x20, 0x6d, + 0x73, 0x67, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x21, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, + 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, + 0x74, 0x69, 0x65, 0x73, 0x20, 0x7c, 0x7c, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, + 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, + 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, + 0x68, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x30, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x29, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, + 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, + 0x20, 0x28, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, + 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, + 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x3e, 0x20, + 0x31, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x4e, 0x6f, 0x74, 0x20, 0x66, 0x6f, + 0x72, 0x20, 0x62, 0x79, 0x74, 0x65, 0x20, 0x70, 0x61, 0x69, 0x72, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, + 0x20, 0x28, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, + 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, + 0x65, 0x73, 0x5b, 0x30, 0x5d, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, + 0x74, 0x2e, 0x73, 0x74, 0x61, 0x72, 0x74, 0x73, 0x57, 0x69, 0x74, 0x68, + 0x28, 0x27, 0x62, 0x79, 0x74, 0x65, 0x3a, 0x20, 0x5c, 0x5c, 0x27, 0x29, + 0x29, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6d, 0x73, 0x67, + 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, + 0x74, 0x20, 0x73, 0x70, 0x6c, 0x69, 0x74, 0x44, 0x61, 0x74, 0x61, 0x20, + 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, + 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, + 0x65, 0x73, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x70, 0x72, 0x6f, 0x62, 0x20, + 0x3d, 0x3e, 0x20, 0x28, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, + 0x74, 0x3a, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x2e, 0x63, 0x6f, 0x6e, 0x74, + 0x65, 0x6e, 0x74, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, + 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, + 0x69, 0x74, 0x69, 0x65, 0x73, 0x3a, 0x20, 0x5b, 0x70, 0x72, 0x6f, 0x62, + 0x5d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x29, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, + 0x6c, 0x60, 0x3c, 0x24, 0x7b, 0x50, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, + 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x7d, 0x20, 0x64, 0x61, 0x74, 0x61, + 0x3d, 0x24, 0x7b, 0x73, 0x70, 0x6c, 0x69, 0x74, 0x44, 0x61, 0x74, 0x61, + 0x7d, 0x20, 0x2f, 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x7b, 0x20, 0x70, 0x72, 0x6f, + 0x62, 0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, + 0x7d, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, + 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, + 0x74, 0x69, 0x65, 0x73, 0x5b, 0x30, 0x5d, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x66, 0x6f, + 0x75, 0x6e, 0x64, 0x20, 0x3d, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x2e, + 0x66, 0x69, 0x6e, 0x64, 0x28, 0x70, 0x20, 0x3d, 0x3e, 0x20, 0x70, 0x2e, + 0x74, 0x6f, 0x6b, 0x5f, 0x73, 0x74, 0x72, 0x20, 0x3d, 0x3d, 0x3d, 0x20, + 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x70, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x20, 0x3d, 0x20, + 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x20, 0x3f, 0x20, 0x70, 0x72, 0x6f, 0x62, + 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x28, 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x2e, + 0x70, 0x72, 0x6f, 0x62, 0x29, 0x20, 0x3a, 0x20, 0x27, 0x74, 0x72, 0x61, + 0x6e, 0x73, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x27, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x43, 0x68, 0x69, 0x6c, + 0x64, 0x72, 0x65, 0x6e, 0x20, 0x3d, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x64, 0x69, 0x76, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x70, + 0x72, 0x6f, 0x62, 0x2d, 0x73, 0x65, 0x74, 0x22, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, + 0x70, 0x72, 0x6f, 0x62, 0x73, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x28, 0x70, + 0x2c, 0x20, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x29, 0x20, 0x3d, 0x3e, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x6b, 0x65, 0x79, 0x3d, 0x24, 0x7b, 0x69, 0x6e, + 0x64, 0x65, 0x78, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, + 0x69, 0x74, 0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x60, 0x70, 0x72, 0x6f, 0x62, + 0x3a, 0x20, 0x24, 0x7b, 0x70, 0x2e, 0x70, 0x72, 0x6f, 0x62, 0x7d, 0x60, + 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x79, 0x6c, + 0x65, 0x3d, 0x24, 0x7b, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x3a, 0x20, 0x27, 0x30, 0x2e, 0x33, 0x65, 0x6d, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, - 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64, 0x43, 0x6f, 0x6c, 0x6f, - 0x72, 0x3a, 0x20, 0x70, 0x2e, 0x74, 0x6f, 0x6b, 0x5f, 0x73, 0x74, 0x72, - 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, - 0x20, 0x3f, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x43, 0x6f, 0x6c, 0x6f, 0x72, - 0x28, 0x70, 0x2e, 0x70, 0x72, 0x6f, 0x62, 0x29, 0x20, 0x3a, 0x20, 0x27, - 0x74, 0x72, 0x61, 0x6e, 0x73, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x27, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, + 0x6e, 0x64, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x70, 0x2e, 0x74, + 0x6f, 0x6b, 0x5f, 0x73, 0x74, 0x72, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x63, + 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x3f, 0x20, 0x70, 0x72, 0x6f, + 0x62, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x28, 0x70, 0x2e, 0x70, 0x72, 0x6f, + 0x62, 0x29, 0x20, 0x3a, 0x20, 0x27, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x70, + 0x61, 0x72, 0x65, 0x6e, 0x74, 0x27, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, @@ -2056,464 +2210,470 @@ unsigned char index_html[] = { 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, - 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x50, - 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x7d, 0x20, 0x73, 0x74, 0x79, 0x6c, - 0x65, 0x3d, 0x24, 0x7b, 0x7b, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, - 0x6f, 0x75, 0x6e, 0x64, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x70, - 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x20, 0x7d, 0x7d, 0x20, 0x70, 0x6f, 0x70, - 0x6f, 0x76, 0x65, 0x72, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, - 0x3d, 0x24, 0x7b, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x43, 0x68, - 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x6d, - 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x2e, 0x6d, - 0x61, 0x74, 0x63, 0x68, 0x28, 0x2f, 0x5c, 0x6e, 0x2f, 0x67, 0x69, 0x6d, - 0x29, 0x20, 0x3f, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x62, 0x72, - 0x20, 0x2f, 0x3e, 0x60, 0x20, 0x3a, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, - 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x70, 0x6f, 0x6f, 0x72, - 0x20, 0x6d, 0x61, 0x6e, 0x73, 0x20, 0x6d, 0x61, 0x72, 0x6b, 0x64, 0x6f, - 0x77, 0x6e, 0x20, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x6d, 0x65, - 0x6e, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x4d, 0x61, 0x72, 0x6b, 0x64, 0x6f, 0x77, 0x6e, 0x69, 0x73, 0x68, - 0x20, 0x3d, 0x20, 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x20, - 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, 0x64, 0x20, 0x3d, 0x20, 0x70, 0x61, - 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x74, 0x65, 0x78, 0x74, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, - 0x63, 0x65, 0x28, 0x2f, 0x26, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x26, 0x61, - 0x6d, 0x70, 0x3b, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, + 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x70, 0x6f, 0x76, + 0x65, 0x72, 0x7d, 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x24, 0x7b, + 0x7b, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64, + 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x70, 0x43, 0x6f, 0x6c, 0x6f, + 0x72, 0x20, 0x7d, 0x7d, 0x20, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, + 0x43, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x3d, 0x24, 0x7b, 0x70, + 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x72, + 0x65, 0x6e, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x6d, 0x73, 0x67, 0x2e, 0x63, + 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x2e, 0x6d, 0x61, 0x74, 0x63, 0x68, + 0x28, 0x2f, 0x5c, 0x6e, 0x2f, 0x67, 0x69, 0x6d, 0x29, 0x20, 0x3f, 0x20, + 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x62, 0x72, 0x20, 0x2f, 0x3e, 0x60, + 0x20, 0x3a, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, + 0x6e, 0x74, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x2f, 0x2f, 0x20, 0x70, 0x6f, 0x6f, 0x72, 0x20, 0x6d, 0x61, 0x6e, + 0x73, 0x20, 0x6d, 0x61, 0x72, 0x6b, 0x64, 0x6f, 0x77, 0x6e, 0x20, 0x72, + 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x4d, 0x61, 0x72, + 0x6b, 0x64, 0x6f, 0x77, 0x6e, 0x69, 0x73, 0x68, 0x20, 0x3d, 0x20, 0x28, + 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x6d, 0x64, 0x20, 0x3d, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, + 0x2e, 0x74, 0x65, 0x78, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, - 0x3c, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x26, 0x6c, 0x74, 0x3b, 0x27, 0x29, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, - 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x3e, 0x2f, 0x67, 0x2c, 0x20, - 0x27, 0x26, 0x67, 0x74, 0x3b, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, - 0x28, 0x2f, 0x5e, 0x23, 0x7b, 0x31, 0x2c, 0x36, 0x7d, 0x20, 0x28, 0x2e, - 0x2a, 0x29, 0x24, 0x2f, 0x67, 0x69, 0x6d, 0x2c, 0x20, 0x27, 0x3c, 0x68, - 0x33, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x68, 0x33, 0x3e, 0x27, 0x29, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, - 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5c, 0x2a, 0x5c, 0x2a, 0x28, 0x2e, - 0x2a, 0x3f, 0x29, 0x5c, 0x2a, 0x5c, 0x2a, 0x2f, 0x67, 0x2c, 0x20, 0x27, - 0x3c, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x31, 0x3c, 0x2f, - 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, - 0x63, 0x65, 0x28, 0x2f, 0x5f, 0x5f, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5f, - 0x5f, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x73, 0x74, 0x72, 0x6f, 0x6e, - 0x67, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, - 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5c, 0x2a, - 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5c, 0x2a, 0x2f, 0x67, 0x2c, 0x20, 0x27, - 0x3c, 0x65, 0x6d, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x65, 0x6d, 0x3e, 0x27, + 0x26, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x26, 0x61, 0x6d, 0x70, 0x3b, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, - 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5f, 0x28, 0x2e, 0x2a, - 0x3f, 0x29, 0x5f, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x65, 0x6d, 0x3e, + 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x3c, 0x2f, 0x67, 0x2c, + 0x20, 0x27, 0x26, 0x6c, 0x74, 0x3b, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, + 0x65, 0x28, 0x2f, 0x3e, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x26, 0x67, 0x74, + 0x3b, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5e, 0x23, + 0x7b, 0x31, 0x2c, 0x36, 0x7d, 0x20, 0x28, 0x2e, 0x2a, 0x29, 0x24, 0x2f, + 0x67, 0x69, 0x6d, 0x2c, 0x20, 0x27, 0x3c, 0x68, 0x33, 0x3e, 0x24, 0x31, + 0x3c, 0x2f, 0x68, 0x33, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, + 0x28, 0x2f, 0x5c, 0x2a, 0x5c, 0x2a, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5c, + 0x2a, 0x5c, 0x2a, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x73, 0x74, 0x72, + 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f, + 0x6e, 0x67, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, + 0x5f, 0x5f, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5f, 0x5f, 0x2f, 0x67, 0x2c, + 0x20, 0x27, 0x3c, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x31, + 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x27, 0x29, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, + 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5c, 0x2a, 0x28, 0x2e, 0x2a, 0x3f, + 0x29, 0x5c, 0x2a, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x65, 0x6d, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x65, 0x6d, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, - 0x63, 0x65, 0x28, 0x2f, 0x60, 0x60, 0x60, 0x2e, 0x2a, 0x3f, 0x5c, 0x6e, - 0x28, 0x5b, 0x5c, 0x73, 0x5c, 0x53, 0x5d, 0x2a, 0x3f, 0x29, 0x60, 0x60, - 0x60, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x70, 0x72, 0x65, 0x3e, 0x3c, - 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x63, 0x6f, 0x64, - 0x65, 0x3e, 0x3c, 0x2f, 0x70, 0x72, 0x65, 0x3e, 0x27, 0x29, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, - 0x61, 0x63, 0x65, 0x28, 0x2f, 0x60, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x60, - 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x24, - 0x31, 0x3c, 0x2f, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x27, 0x29, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, - 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5c, 0x6e, 0x2f, 0x67, 0x69, 0x6d, 0x2c, - 0x20, 0x27, 0x3c, 0x62, 0x72, 0x20, 0x2f, 0x3e, 0x27, 0x29, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, - 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x20, - 0x64, 0x61, 0x6e, 0x67, 0x65, 0x72, 0x6f, 0x75, 0x73, 0x6c, 0x79, 0x53, - 0x65, 0x74, 0x49, 0x6e, 0x6e, 0x65, 0x72, 0x48, 0x54, 0x4d, 0x4c, 0x3d, - 0x24, 0x7b, 0x7b, 0x20, 0x5f, 0x5f, 0x68, 0x74, 0x6d, 0x6c, 0x3a, 0x20, - 0x6d, 0x64, 0x20, 0x7d, 0x7d, 0x20, 0x2f, 0x3e, 0x60, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x47, 0x65, - 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x49, 0x6e, 0x66, 0x6f, - 0x20, 0x3d, 0x20, 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x20, - 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, - 0x66, 0x20, 0x28, 0x21, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, - 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, - 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x73, 0x70, 0x61, - 0x6e, 0x2f, 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, - 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x6c, - 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x2e, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x65, - 0x64, 0x5f, 0x70, 0x65, 0x72, 0x5f, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x5f, - 0x6d, 0x73, 0x2e, 0x74, 0x6f, 0x46, 0x69, 0x78, 0x65, 0x64, 0x28, 0x29, - 0x7d, 0x6d, 0x73, 0x20, 0x70, 0x65, 0x72, 0x20, 0x74, 0x6f, 0x6b, 0x65, - 0x6e, 0x2c, 0x20, 0x24, 0x7b, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, - 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x70, 0x72, - 0x65, 0x64, 0x69, 0x63, 0x74, 0x65, 0x64, 0x5f, 0x70, 0x65, 0x72, 0x5f, - 0x73, 0x65, 0x63, 0x6f, 0x6e, 0x64, 0x2e, 0x74, 0x6f, 0x46, 0x69, 0x78, - 0x65, 0x64, 0x28, 0x32, 0x29, 0x7d, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, - 0x73, 0x20, 0x70, 0x65, 0x72, 0x20, 0x73, 0x65, 0x63, 0x6f, 0x6e, 0x64, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x73, - 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x2f, 0x2f, 0x20, 0x73, 0x69, 0x6d, 0x70, 0x6c, 0x65, 0x20, 0x70, 0x6f, - 0x70, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x69, 0x6d, 0x70, 0x6c, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x50, 0x6f, 0x70, - 0x6f, 0x76, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x70, - 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x73, 0x4f, 0x70, - 0x65, 0x6e, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, 0x6e, - 0x61, 0x6c, 0x28, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x29, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, - 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x75, 0x73, - 0x65, 0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x7b, 0x20, 0x74, 0x6f, - 0x70, 0x3a, 0x20, 0x27, 0x30, 0x70, 0x78, 0x27, 0x2c, 0x20, 0x6c, 0x65, - 0x66, 0x74, 0x3a, 0x20, 0x27, 0x30, 0x70, 0x78, 0x27, 0x20, 0x7d, 0x29, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, - 0x74, 0x20, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, 0x66, 0x20, - 0x3d, 0x20, 0x75, 0x73, 0x65, 0x52, 0x65, 0x66, 0x28, 0x6e, 0x75, 0x6c, - 0x6c, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, - 0x6e, 0x73, 0x74, 0x20, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x52, - 0x65, 0x66, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, 0x52, 0x65, 0x66, 0x28, - 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, 0x6f, 0x67, 0x67, - 0x6c, 0x65, 0x50, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x3d, 0x20, - 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x62, 0x75, 0x74, 0x74, - 0x6f, 0x6e, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, - 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x63, - 0x74, 0x20, 0x3d, 0x20, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, - 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x67, 0x65, - 0x74, 0x42, 0x6f, 0x75, 0x6e, 0x64, 0x69, 0x6e, 0x67, 0x43, 0x6c, 0x69, - 0x65, 0x6e, 0x74, 0x52, 0x65, 0x63, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x6f, 0x73, - 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, - 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x60, 0x24, 0x7b, - 0x72, 0x65, 0x63, 0x74, 0x2e, 0x62, 0x6f, 0x74, 0x74, 0x6f, 0x6d, 0x20, - 0x2b, 0x20, 0x77, 0x69, 0x6e, 0x64, 0x6f, 0x77, 0x2e, 0x73, 0x63, 0x72, - 0x6f, 0x6c, 0x6c, 0x59, 0x7d, 0x70, 0x78, 0x60, 0x2c, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, - 0x66, 0x74, 0x3a, 0x20, 0x60, 0x24, 0x7b, 0x72, 0x65, 0x63, 0x74, 0x2e, - 0x6c, 0x65, 0x66, 0x74, 0x20, 0x2b, 0x20, 0x77, 0x69, 0x6e, 0x64, 0x6f, - 0x77, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x58, 0x7d, 0x70, 0x78, - 0x60, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x73, - 0x4f, 0x70, 0x65, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, - 0x20, 0x21, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, - 0x74, 0x20, 0x68, 0x61, 0x6e, 0x64, 0x6c, 0x65, 0x43, 0x6c, 0x69, 0x63, - 0x6b, 0x4f, 0x75, 0x74, 0x73, 0x69, 0x64, 0x65, 0x20, 0x3d, 0x20, 0x28, - 0x65, 0x76, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, - 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x52, 0x65, 0x66, 0x2e, 0x63, - 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x20, 0x26, 0x26, 0x20, 0x21, 0x70, - 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, - 0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, - 0x6e, 0x73, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x74, 0x61, 0x72, - 0x67, 0x65, 0x74, 0x29, 0x20, 0x26, 0x26, 0x20, 0x21, 0x62, 0x75, 0x74, - 0x74, 0x6f, 0x6e, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, - 0x6e, 0x74, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x73, 0x28, - 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, - 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, - 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, - 0x74, 0x2e, 0x61, 0x64, 0x64, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, - 0x73, 0x74, 0x65, 0x6e, 0x65, 0x72, 0x28, 0x27, 0x6d, 0x6f, 0x75, 0x73, - 0x65, 0x64, 0x6f, 0x77, 0x6e, 0x27, 0x2c, 0x20, 0x68, 0x61, 0x6e, 0x64, - 0x6c, 0x65, 0x43, 0x6c, 0x69, 0x63, 0x6b, 0x4f, 0x75, 0x74, 0x73, 0x69, - 0x64, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x28, 0x29, 0x20, 0x3d, - 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x72, - 0x65, 0x6d, 0x6f, 0x76, 0x65, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, - 0x73, 0x74, 0x65, 0x6e, 0x65, 0x72, 0x28, 0x27, 0x6d, 0x6f, 0x75, 0x73, - 0x65, 0x64, 0x6f, 0x77, 0x6e, 0x27, 0x2c, 0x20, 0x68, 0x61, 0x6e, 0x64, - 0x6c, 0x65, 0x43, 0x6c, 0x69, 0x63, 0x6b, 0x4f, 0x75, 0x74, 0x73, 0x69, - 0x64, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, - 0x20, 0x5b, 0x5d, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, - 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, - 0x70, 0x61, 0x6e, 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x24, 0x7b, - 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x7d, - 0x20, 0x72, 0x65, 0x66, 0x3d, 0x24, 0x7b, 0x62, 0x75, 0x74, 0x74, 0x6f, - 0x6e, 0x52, 0x65, 0x66, 0x7d, 0x20, 0x6f, 0x6e, 0x43, 0x6c, 0x69, 0x63, - 0x6b, 0x3d, 0x24, 0x7b, 0x74, 0x6f, 0x67, 0x67, 0x6c, 0x65, 0x50, 0x6f, - 0x70, 0x6f, 0x76, 0x65, 0x72, 0x7d, 0x3e, 0x24, 0x7b, 0x70, 0x72, 0x6f, - 0x70, 0x73, 0x2e, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x7d, - 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x26, 0x26, 0x20, 0x68, 0x74, + 0x63, 0x65, 0x28, 0x2f, 0x5f, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5f, 0x2f, + 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x65, 0x6d, 0x3e, 0x24, 0x31, 0x3c, 0x2f, + 0x65, 0x6d, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, + 0x60, 0x60, 0x60, 0x2e, 0x2a, 0x3f, 0x5c, 0x6e, 0x28, 0x5b, 0x5c, 0x73, + 0x5c, 0x53, 0x5d, 0x2a, 0x3f, 0x29, 0x60, 0x60, 0x60, 0x2f, 0x67, 0x2c, + 0x20, 0x27, 0x3c, 0x70, 0x72, 0x65, 0x3e, 0x3c, 0x63, 0x6f, 0x64, 0x65, + 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x3c, 0x2f, + 0x70, 0x72, 0x65, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, + 0x2f, 0x60, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x60, 0x2f, 0x67, 0x2c, 0x20, + 0x27, 0x3c, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x63, + 0x6f, 0x64, 0x65, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, + 0x2f, 0x5c, 0x6e, 0x2f, 0x67, 0x69, 0x6d, 0x2c, 0x20, 0x27, 0x3c, 0x62, + 0x72, 0x20, 0x2f, 0x3e, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, + 0x6c, 0x60, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x20, 0x64, 0x61, 0x6e, 0x67, + 0x65, 0x72, 0x6f, 0x75, 0x73, 0x6c, 0x79, 0x53, 0x65, 0x74, 0x49, 0x6e, + 0x6e, 0x65, 0x72, 0x48, 0x54, 0x4d, 0x4c, 0x3d, 0x24, 0x7b, 0x7b, 0x20, + 0x5f, 0x5f, 0x68, 0x74, 0x6d, 0x6c, 0x3a, 0x20, 0x6d, 0x64, 0x20, 0x7d, + 0x7d, 0x20, 0x2f, 0x3e, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x47, 0x65, 0x6e, 0x65, 0x72, 0x61, + 0x74, 0x69, 0x6f, 0x6e, 0x49, 0x6e, 0x66, 0x6f, 0x20, 0x3d, 0x20, 0x28, + 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, + 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, + 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x2f, 0x3e, 0x60, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x7d, - 0x20, 0x69, 0x6e, 0x74, 0x6f, 0x3d, 0x22, 0x23, 0x70, 0x6f, 0x72, 0x74, - 0x61, 0x6c, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x0a, 0x20, 0x20, + 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x6c, 0x6c, 0x61, 0x6d, 0x61, + 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, + 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x65, 0x64, 0x5f, 0x70, 0x65, + 0x72, 0x5f, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x5f, 0x6d, 0x73, 0x2e, 0x74, + 0x6f, 0x46, 0x69, 0x78, 0x65, 0x64, 0x28, 0x29, 0x7d, 0x6d, 0x73, 0x20, + 0x70, 0x65, 0x72, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x2c, 0x20, 0x24, + 0x7b, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, + 0x74, 0x65, 0x64, 0x5f, 0x70, 0x65, 0x72, 0x5f, 0x73, 0x65, 0x63, 0x6f, + 0x6e, 0x64, 0x2e, 0x74, 0x6f, 0x46, 0x69, 0x78, 0x65, 0x64, 0x28, 0x32, + 0x29, 0x7d, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x73, 0x20, 0x70, 0x65, + 0x72, 0x20, 0x73, 0x65, 0x63, 0x6f, 0x6e, 0x64, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, + 0x69, 0x6d, 0x70, 0x6c, 0x65, 0x20, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, + 0x72, 0x20, 0x69, 0x6d, 0x70, 0x6c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x50, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, + 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x3d, + 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, 0x20, 0x3d, + 0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x66, + 0x61, 0x6c, 0x73, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x6f, 0x73, 0x69, 0x74, + 0x69, 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, + 0x6e, 0x61, 0x6c, 0x28, 0x7b, 0x20, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x27, + 0x30, 0x70, 0x78, 0x27, 0x2c, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x3a, 0x20, + 0x27, 0x30, 0x70, 0x78, 0x27, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x62, 0x75, + 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, 0x66, 0x20, 0x3d, 0x20, 0x75, 0x73, + 0x65, 0x52, 0x65, 0x66, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x52, 0x65, 0x66, 0x20, 0x3d, + 0x20, 0x75, 0x73, 0x65, 0x52, 0x65, 0x66, 0x28, 0x6e, 0x75, 0x6c, 0x6c, + 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x74, 0x6f, 0x67, 0x67, 0x6c, 0x65, 0x50, 0x6f, + 0x70, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, + 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, + 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x63, 0x74, 0x20, 0x3d, 0x20, + 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, + 0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x67, 0x65, 0x74, 0x42, 0x6f, 0x75, + 0x6e, 0x64, 0x69, 0x6e, 0x67, 0x43, 0x6c, 0x69, 0x65, 0x6e, 0x74, 0x52, + 0x65, 0x63, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, + 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x72, 0x65, 0x66, 0x3d, 0x24, 0x7b, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, - 0x72, 0x52, 0x65, 0x66, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, - 0x73, 0x3d, 0x22, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x2d, 0x63, - 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x22, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, - 0x79, 0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, - 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x6f, 0x70, 0x2c, + 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x60, 0x24, 0x7b, 0x72, 0x65, 0x63, 0x74, + 0x2e, 0x62, 0x6f, 0x74, 0x74, 0x6f, 0x6d, 0x20, 0x2b, 0x20, 0x77, 0x69, + 0x6e, 0x64, 0x6f, 0x77, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x59, + 0x7d, 0x70, 0x78, 0x60, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x3a, 0x20, + 0x60, 0x24, 0x7b, 0x72, 0x65, 0x63, 0x74, 0x2e, 0x6c, 0x65, 0x66, 0x74, + 0x20, 0x2b, 0x20, 0x77, 0x69, 0x6e, 0x64, 0x6f, 0x77, 0x2e, 0x73, 0x63, + 0x72, 0x6f, 0x6c, 0x6c, 0x58, 0x7d, 0x70, 0x78, 0x60, 0x2c, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x21, 0x69, 0x73, + 0x4f, 0x70, 0x65, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x68, 0x61, + 0x6e, 0x64, 0x6c, 0x65, 0x43, 0x6c, 0x69, 0x63, 0x6b, 0x4f, 0x75, 0x74, + 0x73, 0x69, 0x64, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x76, 0x65, 0x6e, + 0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x70, 0x6f, 0x70, 0x6f, + 0x76, 0x65, 0x72, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, + 0x6e, 0x74, 0x20, 0x26, 0x26, 0x20, 0x21, 0x70, 0x6f, 0x70, 0x6f, 0x76, + 0x65, 0x72, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, + 0x74, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x73, 0x28, 0x65, + 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x29, + 0x20, 0x26, 0x26, 0x20, 0x21, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, + 0x65, 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x63, + 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x73, 0x28, 0x65, 0x76, 0x65, 0x6e, + 0x74, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x29, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, + 0x73, 0x4f, 0x70, 0x65, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, + 0x3d, 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, + 0x73, 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x28, 0x28, 0x29, 0x20, + 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x61, 0x64, + 0x64, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65, 0x6e, + 0x65, 0x72, 0x28, 0x27, 0x6d, 0x6f, 0x75, 0x73, 0x65, 0x64, 0x6f, 0x77, + 0x6e, 0x27, 0x2c, 0x20, 0x68, 0x61, 0x6e, 0x64, 0x6c, 0x65, 0x43, 0x6c, + 0x69, 0x63, 0x6b, 0x4f, 0x75, 0x74, 0x73, 0x69, 0x64, 0x65, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, + 0x75, 0x72, 0x6e, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, + 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x76, + 0x65, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65, 0x6e, + 0x65, 0x72, 0x28, 0x27, 0x6d, 0x6f, 0x75, 0x73, 0x65, 0x64, 0x6f, 0x77, + 0x6e, 0x27, 0x2c, 0x20, 0x68, 0x61, 0x6e, 0x64, 0x6c, 0x65, 0x43, 0x6c, + 0x69, 0x63, 0x6b, 0x4f, 0x75, 0x74, 0x73, 0x69, 0x64, 0x65, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x5b, 0x5d, 0x29, + 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, + 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x20, + 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70, + 0x73, 0x2e, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x7d, 0x20, 0x72, 0x65, 0x66, + 0x3d, 0x24, 0x7b, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, 0x66, + 0x7d, 0x20, 0x6f, 0x6e, 0x43, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, + 0x74, 0x6f, 0x67, 0x67, 0x6c, 0x65, 0x50, 0x6f, 0x70, 0x6f, 0x76, 0x65, + 0x72, 0x7d, 0x3e, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x63, + 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x7d, 0x3c, 0x2f, 0x73, 0x70, + 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x24, 0x7b, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x20, 0x26, 0x26, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, + 0x7b, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x7d, 0x20, 0x69, 0x6e, 0x74, + 0x6f, 0x3d, 0x22, 0x23, 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x64, 0x69, 0x76, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x66, 0x3d, + 0x24, 0x7b, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x52, 0x65, 0x66, + 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x70, + 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x2d, 0x63, 0x6f, 0x6e, 0x74, 0x65, + 0x6e, 0x74, 0x22, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, + 0x24, 0x7b, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x70, 0x6f, 0x73, + 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, + 0x74, 0x6f, 0x70, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x3a, 0x20, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6c, 0x65, 0x66, 0x74, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x7d, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70, 0x73, - 0x2e, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x43, 0x68, 0x69, 0x6c, - 0x64, 0x72, 0x65, 0x6e, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x2f, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x7d, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x7d, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x53, - 0x6f, 0x75, 0x72, 0x63, 0x65, 0x3a, 0x20, 0x70, 0x72, 0x65, 0x61, 0x63, - 0x74, 0x2d, 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x20, 0x28, 0x68, 0x74, - 0x74, 0x70, 0x73, 0x3a, 0x2f, 0x2f, 0x67, 0x69, 0x74, 0x68, 0x75, 0x62, - 0x2e, 0x63, 0x6f, 0x6d, 0x2f, 0x64, 0x65, 0x76, 0x65, 0x6c, 0x6f, 0x70, - 0x69, 0x74, 0x2f, 0x70, 0x72, 0x65, 0x61, 0x63, 0x74, 0x2d, 0x70, 0x6f, - 0x72, 0x74, 0x61, 0x6c, 0x2f, 0x62, 0x6c, 0x6f, 0x62, 0x2f, 0x6d, 0x61, - 0x73, 0x74, 0x65, 0x72, 0x2f, 0x73, 0x72, 0x63, 0x2f, 0x70, 0x72, 0x65, - 0x61, 0x63, 0x74, 0x2d, 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x2e, 0x6a, - 0x73, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2a, 0x2a, 0x20, 0x52, - 0x65, 0x64, 0x69, 0x72, 0x65, 0x63, 0x74, 0x20, 0x72, 0x65, 0x6e, 0x64, - 0x65, 0x72, 0x69, 0x6e, 0x67, 0x20, 0x6f, 0x66, 0x20, 0x64, 0x65, 0x73, - 0x63, 0x65, 0x6e, 0x64, 0x61, 0x6e, 0x74, 0x73, 0x20, 0x69, 0x6e, 0x74, - 0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x67, 0x69, 0x76, 0x65, 0x6e, 0x20, - 0x43, 0x53, 0x53, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x6f, 0x72, - 0x20, 0x2a, 0x2f, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, - 0x73, 0x20, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x20, 0x65, 0x78, 0x74, - 0x65, 0x6e, 0x64, 0x73, 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, - 0x6e, 0x74, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x44, 0x69, 0x64, 0x55, - 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, - 0x6f, 0x72, 0x20, 0x28, 0x6c, 0x65, 0x74, 0x20, 0x69, 0x20, 0x69, 0x6e, - 0x20, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, - 0x70, 0x72, 0x6f, 0x70, 0x73, 0x5b, 0x69, 0x5d, 0x20, 0x21, 0x3d, 0x3d, - 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x5b, - 0x69, 0x5d, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, - 0x20, 0x73, 0x65, 0x74, 0x54, 0x69, 0x6d, 0x65, 0x6f, 0x75, 0x74, 0x28, - 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, - 0x61, 0x79, 0x65, 0x72, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, - 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x44, 0x69, 0x64, 0x4d, 0x6f, 0x75, 0x6e, - 0x74, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x73, 0x4d, 0x6f, 0x75, - 0x6e, 0x74, 0x65, 0x64, 0x20, 0x3d, 0x20, 0x74, 0x72, 0x75, 0x65, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, - 0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, - 0x72, 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6e, - 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72, 0x2e, 0x62, 0x69, 0x6e, - 0x64, 0x28, 0x74, 0x68, 0x69, 0x73, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, - 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72, 0x28, 0x29, 0x3b, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x70, 0x6f, 0x70, + 0x6f, 0x76, 0x65, 0x72, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, + 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x24, 0x7b, 0x50, + 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x60, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x53, 0x6f, 0x75, 0x72, 0x63, + 0x65, 0x3a, 0x20, 0x70, 0x72, 0x65, 0x61, 0x63, 0x74, 0x2d, 0x70, 0x6f, + 0x72, 0x74, 0x61, 0x6c, 0x20, 0x28, 0x68, 0x74, 0x74, 0x70, 0x73, 0x3a, + 0x2f, 0x2f, 0x67, 0x69, 0x74, 0x68, 0x75, 0x62, 0x2e, 0x63, 0x6f, 0x6d, + 0x2f, 0x64, 0x65, 0x76, 0x65, 0x6c, 0x6f, 0x70, 0x69, 0x74, 0x2f, 0x70, + 0x72, 0x65, 0x61, 0x63, 0x74, 0x2d, 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, + 0x2f, 0x62, 0x6c, 0x6f, 0x62, 0x2f, 0x6d, 0x61, 0x73, 0x74, 0x65, 0x72, + 0x2f, 0x73, 0x72, 0x63, 0x2f, 0x70, 0x72, 0x65, 0x61, 0x63, 0x74, 0x2d, + 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x2e, 0x6a, 0x73, 0x29, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x2f, 0x2a, 0x2a, 0x20, 0x52, 0x65, 0x64, 0x69, 0x72, + 0x65, 0x63, 0x74, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x69, 0x6e, + 0x67, 0x20, 0x6f, 0x66, 0x20, 0x64, 0x65, 0x73, 0x63, 0x65, 0x6e, 0x64, + 0x61, 0x6e, 0x74, 0x73, 0x20, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x74, 0x68, + 0x65, 0x20, 0x67, 0x69, 0x76, 0x65, 0x6e, 0x20, 0x43, 0x53, 0x53, 0x20, + 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x6f, 0x72, 0x20, 0x2a, 0x2f, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x20, 0x50, 0x6f, + 0x72, 0x74, 0x61, 0x6c, 0x20, 0x65, 0x78, 0x74, 0x65, 0x6e, 0x64, 0x73, + 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6f, + 0x6e, 0x65, 0x6e, 0x74, 0x44, 0x69, 0x64, 0x55, 0x70, 0x64, 0x61, 0x74, + 0x65, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x28, + 0x6c, 0x65, 0x74, 0x20, 0x69, 0x20, 0x69, 0x6e, 0x20, 0x70, 0x72, 0x6f, + 0x70, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x70, + 0x73, 0x5b, 0x69, 0x5d, 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x5b, 0x69, 0x5d, 0x29, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x73, 0x65, 0x74, + 0x54, 0x69, 0x6d, 0x65, 0x6f, 0x75, 0x74, 0x28, 0x74, 0x68, 0x69, 0x73, + 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, - 0x74, 0x57, 0x69, 0x6c, 0x6c, 0x55, 0x6e, 0x6d, 0x6f, 0x75, 0x6e, 0x74, - 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, - 0x4c, 0x61, 0x79, 0x65, 0x72, 0x28, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x29, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, + 0x74, 0x44, 0x69, 0x64, 0x4d, 0x6f, 0x75, 0x6e, 0x74, 0x28, 0x29, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x73, 0x4d, 0x6f, 0x75, 0x6e, 0x74, 0x65, 0x64, - 0x20, 0x3d, 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, 0x68, - 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x20, 0x26, 0x26, - 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, - 0x2e, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x4e, 0x6f, 0x64, 0x65, 0x29, - 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, - 0x2e, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x4e, 0x6f, 0x64, 0x65, 0x2e, - 0x72, 0x65, 0x6d, 0x6f, 0x76, 0x65, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x28, - 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x29, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x69, 0x6e, 0x64, 0x4e, 0x6f, 0x64, - 0x65, 0x28, 0x6e, 0x6f, 0x64, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, - 0x20, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x6e, 0x6f, 0x64, 0x65, - 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, - 0x27, 0x20, 0x3f, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, - 0x2e, 0x71, 0x75, 0x65, 0x72, 0x79, 0x53, 0x65, 0x6c, 0x65, 0x63, 0x74, - 0x6f, 0x72, 0x28, 0x6e, 0x6f, 0x64, 0x65, 0x29, 0x20, 0x3a, 0x20, 0x6e, - 0x6f, 0x64, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6e, 0x64, - 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72, 0x28, 0x73, 0x68, 0x6f, 0x77, - 0x20, 0x3d, 0x20, 0x74, 0x72, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, - 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x73, 0x4d, 0x6f, 0x75, 0x6e, 0x74, - 0x65, 0x64, 0x29, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x3b, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, - 0x63, 0x6c, 0x65, 0x61, 0x6e, 0x20, 0x75, 0x70, 0x20, 0x6f, 0x6c, 0x64, - 0x20, 0x6e, 0x6f, 0x64, 0x65, 0x20, 0x69, 0x66, 0x20, 0x6d, 0x6f, 0x76, - 0x69, 0x6e, 0x67, 0x20, 0x62, 0x61, 0x73, 0x65, 0x73, 0x3a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, - 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x69, 0x6e, - 0x74, 0x6f, 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, - 0x69, 0x6e, 0x74, 0x6f, 0x50, 0x6f, 0x69, 0x6e, 0x74, 0x65, 0x72, 0x29, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x50, 0x6f, - 0x69, 0x6e, 0x74, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, - 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, - 0x66, 0x20, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, - 0x20, 0x26, 0x26, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, - 0x6f, 0x74, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, - 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x6e, - 0x64, 0x65, 0x72, 0x28, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x24, 0x7b, - 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x78, 0x79, 0x7d, - 0x20, 0x2f, 0x3e, 0x60, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, - 0x6e, 0x74, 0x6f, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, - 0x6d, 0x6f, 0x74, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, - 0x74, 0x6f, 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x66, 0x69, - 0x6e, 0x64, 0x4e, 0x6f, 0x64, 0x65, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, - 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x29, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, - 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x72, 0x65, - 0x6e, 0x64, 0x65, 0x72, 0x28, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, - 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x78, 0x79, 0x7d, - 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x3d, 0x24, 0x7b, 0x74, - 0x68, 0x69, 0x73, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x7d, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x24, 0x7b, 0x73, 0x68, 0x6f, 0x77, 0x20, 0x26, 0x26, 0x20, - 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x63, - 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x20, 0x7c, 0x7c, 0x20, 0x6e, - 0x75, 0x6c, 0x6c, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, 0x61, - 0x6c, 0x50, 0x72, 0x6f, 0x78, 0x79, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, - 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, - 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x3d, 0x20, 0x74, 0x72, 0x75, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, + 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72, 0x20, 0x3d, 0x20, + 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, + 0x61, 0x79, 0x65, 0x72, 0x2e, 0x62, 0x69, 0x6e, 0x64, 0x28, 0x74, 0x68, + 0x69, 0x73, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, + 0x4c, 0x61, 0x79, 0x65, 0x72, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, - 0x6e, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x68, 0x69, 0x67, 0x68, 0x2d, 0x6f, 0x72, - 0x64, 0x65, 0x72, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, - 0x74, 0x20, 0x74, 0x68, 0x61, 0x74, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, - 0x72, 0x73, 0x20, 0x69, 0x74, 0x73, 0x20, 0x66, 0x69, 0x72, 0x73, 0x74, - 0x20, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x20, 0x69, 0x66, 0x20, 0x69, 0x74, - 0x20, 0x65, 0x78, 0x69, 0x73, 0x74, 0x73, 0x2e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x2f, 0x2f, 0x20, 0x75, 0x73, 0x65, 0x64, 0x20, 0x61, 0x73, 0x20, - 0x61, 0x20, 0x63, 0x6f, 0x6e, 0x64, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x61, - 0x6c, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x69, 0x6e, 0x67, 0x20, - 0x70, 0x72, 0x6f, 0x78, 0x79, 0x2e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6c, 0x61, 0x73, 0x73, 0x20, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x50, - 0x72, 0x6f, 0x78, 0x79, 0x20, 0x65, 0x78, 0x74, 0x65, 0x6e, 0x64, 0x73, - 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, 0x65, 0x74, 0x43, 0x68, - 0x69, 0x6c, 0x64, 0x43, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x28, 0x29, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, - 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, - 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x7b, - 0x20, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x20, 0x7d, 0x29, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, - 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x72, - 0x65, 0x6e, 0x20, 0x7c, 0x7c, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, - 0x69, 0x6f, 0x6e, 0x20, 0x41, 0x70, 0x70, 0x28, 0x70, 0x72, 0x6f, 0x70, - 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, - 0x76, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x6d, 0x6f, 0x64, - 0x65, 0x2d, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x7d, 0x22, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x68, 0x65, 0x61, 0x64, 0x65, 0x72, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x68, 0x31, - 0x3e, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70, 0x70, 0x3c, 0x2f, - 0x68, 0x31, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x2f, 0x68, 0x65, 0x61, 0x64, 0x65, 0x72, 0x3e, 0x0a, + 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x57, 0x69, 0x6c, + 0x6c, 0x55, 0x6e, 0x6d, 0x6f, 0x75, 0x6e, 0x74, 0x28, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, + 0x72, 0x28, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, + 0x73, 0x4d, 0x6f, 0x75, 0x6e, 0x74, 0x65, 0x64, 0x20, 0x3d, 0x20, 0x66, + 0x61, 0x6c, 0x73, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, + 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x20, 0x26, 0x26, 0x20, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x2e, 0x70, 0x61, 0x72, + 0x65, 0x6e, 0x74, 0x4e, 0x6f, 0x64, 0x65, 0x29, 0x20, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x2e, 0x70, 0x61, 0x72, + 0x65, 0x6e, 0x74, 0x4e, 0x6f, 0x64, 0x65, 0x2e, 0x72, 0x65, 0x6d, 0x6f, + 0x76, 0x65, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x28, 0x74, 0x68, 0x69, 0x73, + 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x66, 0x69, 0x6e, 0x64, 0x4e, 0x6f, 0x64, 0x65, 0x28, 0x6e, 0x6f, + 0x64, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x79, 0x70, + 0x65, 0x6f, 0x66, 0x20, 0x6e, 0x6f, 0x64, 0x65, 0x20, 0x3d, 0x3d, 0x3d, + 0x20, 0x27, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x27, 0x20, 0x3f, 0x20, + 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x71, 0x75, 0x65, + 0x72, 0x79, 0x53, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x6f, 0x72, 0x28, 0x6e, + 0x6f, 0x64, 0x65, 0x29, 0x20, 0x3a, 0x20, 0x6e, 0x6f, 0x64, 0x65, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, + 0x79, 0x65, 0x72, 0x28, 0x73, 0x68, 0x6f, 0x77, 0x20, 0x3d, 0x20, 0x74, + 0x72, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x74, 0x68, 0x69, 0x73, + 0x2e, 0x69, 0x73, 0x4d, 0x6f, 0x75, 0x6e, 0x74, 0x65, 0x64, 0x29, 0x20, + 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x63, 0x6c, 0x65, 0x61, + 0x6e, 0x20, 0x75, 0x70, 0x20, 0x6f, 0x6c, 0x64, 0x20, 0x6e, 0x6f, 0x64, + 0x65, 0x20, 0x69, 0x66, 0x20, 0x6d, 0x6f, 0x76, 0x69, 0x6e, 0x67, 0x20, + 0x62, 0x61, 0x73, 0x65, 0x73, 0x3a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, + 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x21, + 0x3d, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, + 0x50, 0x6f, 0x69, 0x6e, 0x74, 0x65, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x50, 0x6f, 0x69, 0x6e, 0x74, 0x65, + 0x72, 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, + 0x70, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, + 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x26, 0x26, 0x20, + 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, + 0x74, 0x65, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, + 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, + 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x78, 0x79, 0x7d, 0x20, 0x2f, 0x3e, 0x60, + 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x2c, + 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x3d, + 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x66, 0x69, 0x6e, 0x64, 0x4e, 0x6f, + 0x64, 0x65, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, + 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, + 0x6f, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, + 0x28, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, + 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x78, 0x79, 0x7d, 0x20, 0x63, 0x6f, 0x6e, + 0x74, 0x65, 0x78, 0x74, 0x3d, 0x24, 0x7b, 0x74, 0x68, 0x69, 0x73, 0x2e, + 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x7d, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, + 0x73, 0x68, 0x6f, 0x77, 0x20, 0x26, 0x26, 0x20, 0x74, 0x68, 0x69, 0x73, + 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x63, 0x68, 0x69, 0x6c, 0x64, + 0x72, 0x65, 0x6e, 0x20, 0x7c, 0x7c, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x6d, 0x61, 0x69, 0x6e, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x63, 0x6f, 0x6e, - 0x74, 0x65, 0x6e, 0x74, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x63, 0x68, - 0x61, 0x74, 0x53, 0x74, 0x61, 0x72, 0x74, 0x65, 0x64, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x20, 0x3f, 0x20, 0x43, 0x68, 0x61, 0x74, 0x4c, 0x6f, - 0x67, 0x20, 0x3a, 0x20, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, - 0x72, 0x6d, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x6d, 0x61, 0x69, 0x6e, 0x3e, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x73, 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x69, 0x64, 0x3d, - 0x22, 0x77, 0x72, 0x69, 0x74, 0x65, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, + 0x2f, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x50, 0x72, 0x6f, + 0x78, 0x79, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x60, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, + 0x6f, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, + 0x74, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6e, 0x64, + 0x65, 0x72, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x75, + 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, + 0x20, 0x68, 0x69, 0x67, 0x68, 0x2d, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x20, + 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x20, 0x74, 0x68, + 0x61, 0x74, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x73, 0x20, 0x69, + 0x74, 0x73, 0x20, 0x66, 0x69, 0x72, 0x73, 0x74, 0x20, 0x63, 0x68, 0x69, + 0x6c, 0x64, 0x20, 0x69, 0x66, 0x20, 0x69, 0x74, 0x20, 0x65, 0x78, 0x69, + 0x73, 0x74, 0x73, 0x2e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, + 0x75, 0x73, 0x65, 0x64, 0x20, 0x61, 0x73, 0x20, 0x61, 0x20, 0x63, 0x6f, + 0x6e, 0x64, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x61, 0x6c, 0x20, 0x72, 0x65, + 0x6e, 0x64, 0x65, 0x72, 0x69, 0x6e, 0x67, 0x20, 0x70, 0x72, 0x6f, 0x78, + 0x79, 0x2e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, + 0x20, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x78, 0x79, + 0x20, 0x65, 0x78, 0x74, 0x65, 0x6e, 0x64, 0x73, 0x20, 0x43, 0x6f, 0x6d, + 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x67, 0x65, 0x74, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x43, + 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, + 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, + 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x7b, 0x20, 0x63, 0x68, 0x69, + 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x20, 0x7d, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, + 0x6e, 0x20, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x20, 0x7c, + 0x7c, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, + 0x41, 0x70, 0x70, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x7b, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, + 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, 0x63, 0x6c, + 0x61, 0x73, 0x73, 0x3d, 0x22, 0x6d, 0x6f, 0x64, 0x65, 0x2d, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, - 0x63, 0x68, 0x61, 0x74, 0x27, 0x20, 0x3f, 0x20, 0x4d, 0x65, 0x73, 0x73, - 0x61, 0x67, 0x65, 0x49, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x3a, 0x20, 0x43, - 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x43, 0x6f, 0x6e, - 0x74, 0x72, 0x6f, 0x6c, 0x73, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x65, - 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x6f, 0x6f, 0x74, 0x65, - 0x72, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x70, 0x3e, 0x3c, 0x24, 0x7b, 0x4d, 0x6f, 0x64, - 0x65, 0x6c, 0x47, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, - 0x49, 0x6e, 0x66, 0x6f, 0x7d, 0x20, 0x2f, 0x3e, 0x3c, 0x2f, 0x70, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x70, 0x3e, 0x50, 0x6f, 0x77, 0x65, 0x72, 0x65, 0x64, 0x20, - 0x62, 0x79, 0x20, 0x3c, 0x61, 0x20, 0x68, 0x72, 0x65, 0x66, 0x3d, 0x22, - 0x68, 0x74, 0x74, 0x70, 0x73, 0x3a, 0x2f, 0x2f, 0x67, 0x69, 0x74, 0x68, - 0x75, 0x62, 0x2e, 0x63, 0x6f, 0x6d, 0x2f, 0x67, 0x67, 0x65, 0x72, 0x67, - 0x61, 0x6e, 0x6f, 0x76, 0x2f, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63, - 0x70, 0x70, 0x22, 0x3e, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70, - 0x70, 0x3c, 0x2f, 0x61, 0x3e, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x3c, 0x61, - 0x20, 0x68, 0x72, 0x65, 0x66, 0x3d, 0x22, 0x68, 0x74, 0x74, 0x70, 0x73, - 0x3a, 0x2f, 0x2f, 0x67, 0x67, 0x6d, 0x6c, 0x2e, 0x61, 0x69, 0x22, 0x3e, - 0x67, 0x67, 0x6d, 0x6c, 0x2e, 0x61, 0x69, 0x3c, 0x2f, 0x61, 0x3e, 0x2e, - 0x3c, 0x2f, 0x70, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x6f, 0x6f, 0x74, 0x65, 0x72, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, - 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x68, 0x28, 0x41, 0x70, 0x70, - 0x29, 0x2c, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, - 0x71, 0x75, 0x65, 0x72, 0x79, 0x53, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x6f, - 0x72, 0x28, 0x27, 0x23, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, - 0x72, 0x27, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x63, - 0x72, 0x69, 0x70, 0x74, 0x3e, 0x0a, 0x3c, 0x2f, 0x68, 0x65, 0x61, 0x64, - 0x3e, 0x0a, 0x0a, 0x3c, 0x62, 0x6f, 0x64, 0x79, 0x3e, 0x0a, 0x20, 0x20, - 0x3c, 0x64, 0x69, 0x76, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x63, 0x6f, 0x6e, - 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x22, 0x3e, 0x3c, 0x2f, 0x64, 0x69, - 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, 0x69, 0x64, - 0x3d, 0x22, 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x22, 0x3e, 0x3c, 0x2f, - 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x3c, 0x2f, 0x62, 0x6f, 0x64, 0x79, 0x3e, - 0x0a, 0x0a, 0x3c, 0x2f, 0x68, 0x74, 0x6d, 0x6c, 0x3e, 0x0a + 0x65, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x7d, 0x22, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x68, 0x65, 0x61, + 0x64, 0x65, 0x72, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x68, 0x31, 0x3e, 0x6c, 0x6c, 0x61, + 0x6d, 0x61, 0x2e, 0x63, 0x70, 0x70, 0x3c, 0x2f, 0x68, 0x31, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, + 0x68, 0x65, 0x61, 0x64, 0x65, 0x72, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6d, 0x61, 0x69, 0x6e, + 0x20, 0x69, 0x64, 0x3d, 0x22, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, + 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x63, 0x68, 0x61, 0x74, 0x53, 0x74, + 0x61, 0x72, 0x74, 0x65, 0x64, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, + 0x3f, 0x20, 0x43, 0x68, 0x61, 0x74, 0x4c, 0x6f, 0x67, 0x20, 0x3a, 0x20, + 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x7d, 0x20, + 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x2f, 0x6d, 0x61, 0x69, 0x6e, 0x3e, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x65, 0x63, + 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x77, 0x72, 0x69, + 0x74, 0x65, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, + 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, + 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x63, 0x68, 0x61, 0x74, + 0x27, 0x20, 0x3f, 0x20, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x49, + 0x6e, 0x70, 0x75, 0x74, 0x20, 0x3a, 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6c, + 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, + 0x73, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x65, 0x63, 0x74, 0x69, 0x6f, + 0x6e, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x66, 0x6f, 0x6f, 0x74, 0x65, 0x72, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x70, 0x3e, 0x3c, 0x24, 0x7b, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x47, 0x65, + 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x49, 0x6e, 0x66, 0x6f, + 0x7d, 0x20, 0x2f, 0x3e, 0x3c, 0x2f, 0x70, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x70, 0x3e, + 0x50, 0x6f, 0x77, 0x65, 0x72, 0x65, 0x64, 0x20, 0x62, 0x79, 0x20, 0x3c, + 0x61, 0x20, 0x68, 0x72, 0x65, 0x66, 0x3d, 0x22, 0x68, 0x74, 0x74, 0x70, + 0x73, 0x3a, 0x2f, 0x2f, 0x67, 0x69, 0x74, 0x68, 0x75, 0x62, 0x2e, 0x63, + 0x6f, 0x6d, 0x2f, 0x67, 0x67, 0x65, 0x72, 0x67, 0x61, 0x6e, 0x6f, 0x76, + 0x2f, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70, 0x70, 0x22, 0x3e, + 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70, 0x70, 0x3c, 0x2f, 0x61, + 0x3e, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x3c, 0x61, 0x20, 0x68, 0x72, 0x65, + 0x66, 0x3d, 0x22, 0x68, 0x74, 0x74, 0x70, 0x73, 0x3a, 0x2f, 0x2f, 0x67, + 0x67, 0x6d, 0x6c, 0x2e, 0x61, 0x69, 0x22, 0x3e, 0x67, 0x67, 0x6d, 0x6c, + 0x2e, 0x61, 0x69, 0x3c, 0x2f, 0x61, 0x3e, 0x2e, 0x3c, 0x2f, 0x70, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x2f, 0x66, 0x6f, 0x6f, 0x74, 0x65, 0x72, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6e, 0x64, + 0x65, 0x72, 0x28, 0x68, 0x28, 0x41, 0x70, 0x70, 0x29, 0x2c, 0x20, 0x64, + 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x71, 0x75, 0x65, 0x72, + 0x79, 0x53, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x6f, 0x72, 0x28, 0x27, 0x23, + 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x27, 0x29, 0x29, + 0x3b, 0x0a, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, + 0x3e, 0x0a, 0x3c, 0x2f, 0x68, 0x65, 0x61, 0x64, 0x3e, 0x0a, 0x0a, 0x3c, + 0x62, 0x6f, 0x64, 0x79, 0x3e, 0x0a, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, + 0x20, 0x69, 0x64, 0x3d, 0x22, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, + 0x65, 0x72, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, + 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x66, 0x69, + 0x6c, 0x65, 0x22, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x66, 0x69, 0x6c, 0x65, + 0x49, 0x6e, 0x70, 0x75, 0x74, 0x22, 0x20, 0x61, 0x63, 0x63, 0x65, 0x70, + 0x74, 0x3d, 0x22, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x2f, 0x2a, 0x22, 0x20, + 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x22, 0x64, 0x69, 0x73, 0x70, 0x6c, + 0x61, 0x79, 0x3a, 0x20, 0x6e, 0x6f, 0x6e, 0x65, 0x3b, 0x22, 0x3e, 0x0a, + 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x3c, + 0x64, 0x69, 0x76, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x70, 0x6f, 0x72, 0x74, + 0x61, 0x6c, 0x22, 0x3e, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x3c, + 0x2f, 0x62, 0x6f, 0x64, 0x79, 0x3e, 0x0a, 0x0a, 0x3c, 0x2f, 0x68, 0x74, + 0x6d, 0x6c, 0x3e, 0x0a, 0x0a }; -unsigned int index_html_len = 30190; +unsigned int index_html_len = 32105; diff --git a/examples/server/public/index.html b/examples/server/public/index.html index f539884e2..39d7bb93d 100644 --- a/examples/server/public/index.html +++ b/examples/server/public/index.html @@ -125,6 +125,7 @@ background-color: #222; color: #ddd; } + code { font-family: monospace; padding: 0.1em 0.3em; @@ -141,7 +142,8 @@ display: inline; } - header, footer { + header, + footer { text-align: center; } @@ -163,6 +165,7 @@ 0% { background-position: 0%; } + 100% { background-position: 100%; } @@ -181,6 +184,7 @@ --loading-color-1: #22222200; --loading-color-2: #222222ff; } + .popover-content { background-color: black; } @@ -194,6 +198,8 @@ import { llama } from '/completion.js'; import { SchemaConverter } from '/json-schema-to-grammar.mjs'; + let selected_image = false; + var slot_id = -1; const session = signal({ prompt: "This is a conversation between User and Llama, a friendly chatbot. Llama is helpful, kind, honest, good at writing, and never fails to answer any requests immediately and with precision.", @@ -203,6 +209,7 @@ type: "chat", // "chat" | "completion" char: "Llama", user: "User", + image_selected: '' }) const params = signal({ @@ -220,7 +227,9 @@ mirostat_tau: 5, // target entropy mirostat_eta: 0.1, // learning rate grammar: '', - n_probs: 0, // no completion_probabilities + n_probs: 0, // no completion_probabilities, + image_data: [], + cache_prompt: true }) /* START: Support for storing prompt templates and parameters in borwser LocalStorage */ @@ -270,6 +279,7 @@ // saved templates were successfuly imported. console.log('Processing saved templates and updating default template') + params.value = { ...params.value, image_data: [] }; //console.log(importedTemplates); savedUserTemplates.value = importedTemplates; @@ -294,7 +304,9 @@ function userTemplateApply(t) { session.value = t.data.session; + session.value = { ...session.value, image_selected: '' }; params.value = t.data.params; + params.value = { ...params.value, image_data: [] }; } function userTemplateResetToDefaultAndApply() { @@ -385,20 +397,25 @@ throw new Error("already running"); } controller.value = new AbortController(); - for await (const chunk of llama(prompt, llamaParams, {controller: controller.value})) { + for await (const chunk of llama(prompt, llamaParams, { controller: controller.value })) { const data = chunk.data; if (data.stop) { while ( currentMessages.length > 0 && currentMessages[currentMessages.length - 1].content.match(/\n$/) != null - ) { + ) { currentMessages.pop(); } transcriptUpdate([...history, [char, currentMessages]]) console.log("Completion finished: '", currentMessages.map(msg => msg.content).join(''), "', summary: ", data); } else { currentMessages.push(data); + slot_id = data.slot_id; + if (selected_image && !data.multimodal) { + alert("The server was not compiled for multimodal or the model projector can't be loaded."); + return; + } transcriptUpdate([...history, [char, currentMessages]]) } @@ -419,7 +436,7 @@ transcriptUpdate([...session.value.transcript, ["{{user}}", msg]]) - const prompt = template(session.value.template, { + let prompt = template(session.value.template, { message: msg, history: session.value.transcript.flatMap( ([name, data]) => @@ -434,9 +451,12 @@ ) ).join("\n"), }); - + if (selected_image) { + prompt = `A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:[img-10]${msg}\nASSISTANT:`; + } await runLlama(prompt, { ...params.value, + slot_id: slot_id, stop: ["", template("{{char}}:"), template("{{user}}:")], }, "{{char}}"); } @@ -446,10 +466,11 @@ console.log('already running...'); return; } - const {prompt} = session.value; + const { prompt } = session.value; transcriptUpdate([...session.value.transcript, ["", prompt]]); await runLlama(prompt, { ...params.value, + slot_id: slot_id, stop: [], }, ""); } @@ -467,6 +488,27 @@ transcriptUpdate([]); } + const uploadImage = (e) => { + e.preventDefault(); + document.getElementById("fileInput").click(); + document.getElementById("fileInput").addEventListener("change", function (event) { + const selectedFile = event.target.files[0]; + if (selectedFile) { + const reader = new FileReader(); + reader.onload = function () { + const image_data = reader.result; + session.value = { ...session.value, image_selected: image_data }; + params.value = { + ...params.value, image_data: [ + { data: image_data.replace(/data:image\/[^;]+;base64,/, ''), id: 10 }] + } + }; + selected_image = true; + reader.readAsDataURL(selectedFile); + } + }); + } + function MessageInput() { const message = useSignal("") @@ -497,6 +539,7 @@
+
@@ -540,7 +583,7 @@ data; message = html`<${Markdownish} text=${template(text)} />` } - if(user) { + if (user) { return html`

${template(user)}: ${message}

` } else { return html`

${message}

` @@ -549,6 +592,7 @@ return html`
+ ${messages.flatMap(chatLine)}
`; }; @@ -567,7 +611,7 @@ const converter = new SchemaConverter( grammarJsonSchemaPropOrder.value .split(',') - .reduce((acc, cur, i) => ({...acc, [cur.trim()]: i}), {}) + .reduce((acc, cur, i) => ({ ...acc, [cur.trim()]: i }), {}) ) converter.visit(schema, '') params.value = { @@ -579,7 +623,7 @@ } } - const FloatField = ({label, max, min, name, step, value}) => { + const FloatField = ({ label, max, min, name, step, value }) => { return html`
@@ -589,7 +633,7 @@ ` }; - const IntField = ({label, max, min, name, value}) => { + const IntField = ({ label, max, min, name, value }) => { return html`
@@ -672,7 +716,7 @@ ${GrammarControl()} ` - ); + ); const CompletionConfigForm = () => ( html` @@ -694,20 +738,20 @@ ${session.value.type === 'chat' ? ChatConfigForm() : CompletionConfigForm()}
- ${IntField({label: "Predictions", max: 2048, min: -1, name: "n_predict", value: params.value.n_predict})} - ${FloatField({label: "Temperature", max: 1.5, min: 0.0, name: "temperature", step: 0.01, value: params.value.temperature})} - ${FloatField({label: "Penalize repeat sequence", max: 2.0, min: 0.0, name: "repeat_penalty", step: 0.01, value: params.value.repeat_penalty})} - ${IntField({label: "Consider N tokens for penalize", max: 2048, min: 0, name: "repeat_last_n", value: params.value.repeat_last_n})} - ${IntField({label: "Top-K sampling", max: 100, min: -1, name: "top_k", value: params.value.top_k})} - ${FloatField({label: "Top-P sampling", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p})} + ${IntField({ label: "Predictions", max: 2048, min: -1, name: "n_predict", value: params.value.n_predict })} + ${FloatField({ label: "Temperature", max: 1.5, min: 0.0, name: "temperature", step: 0.01, value: params.value.temperature })} + ${FloatField({ label: "Penalize repeat sequence", max: 2.0, min: 0.0, name: "repeat_penalty", step: 0.01, value: params.value.repeat_penalty })} + ${IntField({ label: "Consider N tokens for penalize", max: 2048, min: 0, name: "repeat_last_n", value: params.value.repeat_last_n })} + ${IntField({ label: "Top-K sampling", max: 100, min: -1, name: "top_k", value: params.value.top_k })} + ${FloatField({ label: "Top-P sampling", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p })}
More options
- ${FloatField({label: "TFS-Z", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z})} - ${FloatField({label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p})} - ${FloatField({label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty})} - ${FloatField({label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty})} + ${FloatField({ label: "TFS-Z", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })} + ${FloatField({ label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p })} + ${FloatField({ label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })} + ${FloatField({ label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })}
@@ -716,11 +760,11 @@
- ${FloatField({label: "Mirostat tau", max: 10.0, min: 0.0, name: "mirostat_tau", step: 0.01, value: params.value.mirostat_tau})} - ${FloatField({label: "Mirostat eta", max: 1.0, min: 0.0, name: "mirostat_eta", step: 0.01, value: params.value.mirostat_eta})} + ${FloatField({ label: "Mirostat tau", max: 10.0, min: 0.0, name: "mirostat_tau", step: 0.01, value: params.value.mirostat_tau })} + ${FloatField({ label: "Mirostat eta", max: 1.0, min: 0.0, name: "mirostat_eta", step: 0.01, value: params.value.mirostat_eta })}
- ${IntField({label: "Show Probabilities", max: 10, min: 0, name: "n_probs", value: params.value.n_probs})} + ${IntField({ label: "Show Probabilities", max: 10, min: 0, name: "n_probs", value: params.value.n_probs })}
@@ -759,20 +803,20 @@ const popoverChildren = html`
${probs.map((p, index) => { - return html` + return html`
${p.tok_str}: ${Math.floor(p.prob * 100)}%
` - })} + })}
` @@ -851,9 +895,9 @@ ref=${popoverRef} class="popover-content" style=${{ - top: position.value.top, - left: position.value.left, - }} + top: position.value.top, + left: position.value.left, + }} > ${props.popoverChildren}
@@ -952,8 +996,11 @@ -
+
+ +
+ diff --git a/examples/server/server.cpp b/examples/server/server.cpp index b5ad3cc99..47ae0d558 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -1,6 +1,11 @@ #include "common.h" #include "llama.h" #include "build-info.h" +#include "grammar-parser.h" + +#include "../llava/clip.h" + +#include "stb_image.h" #ifndef NDEBUG // crash the server in debug mode, otherwise send an http 500 error @@ -17,12 +22,14 @@ #include "json-schema-to-grammar.mjs.hpp" #include +#include +#include +#include #ifndef SERVER_VERBOSE #define SERVER_VERBOSE 1 #endif -using namespace httplib; using json = nlohmann::json; struct server_params @@ -34,6 +41,166 @@ struct server_params int32_t write_timeout = 600; }; +static bool server_verbose = false; + +#if SERVER_VERBOSE != 1 +#define LOG_VERBOSE(MSG, ...) +#else +#define LOG_VERBOSE(MSG, ...) \ + do \ + { \ + if (server_verbose) \ + { \ + server_log("VERBOSE", __func__, __LINE__, MSG, __VA_ARGS__); \ + } \ + } while (0) +#endif + +#define LOG_ERROR( MSG, ...) server_log("ERROR", __func__, __LINE__, MSG, __VA_ARGS__) +#define LOG_WARNING(MSG, ...) server_log("WARNING", __func__, __LINE__, MSG, __VA_ARGS__) +#define LOG_INFO( MSG, ...) server_log("INFO", __func__, __LINE__, MSG, __VA_ARGS__) + +// +// base64 utils (TODO: move to common in the future) +// + +static const std::string base64_chars = + "ABCDEFGHIJKLMNOPQRSTUVWXYZ" + "abcdefghijklmnopqrstuvwxyz" + "0123456789+/"; + +static inline bool is_base64(uint8_t c) +{ + return (isalnum(c) || (c == '+') || (c == '/')); +} + +static std::vector base64_decode(std::string const &encoded_string) +{ + int i = 0; + int j = 0; + int in_ = 0; + + int in_len = encoded_string.size(); + + uint8_t char_array_4[4]; + uint8_t char_array_3[3]; + + std::vector ret; + + while (in_len-- && (encoded_string[in_] != '=') && is_base64(encoded_string[in_])) + { + char_array_4[i++] = encoded_string[in_]; in_++; + if (i == 4) + { + for (i = 0; i <4; i++) + { + char_array_4[i] = base64_chars.find(char_array_4[i]); + } + + char_array_3[0] = ((char_array_4[0] ) << 2) + ((char_array_4[1] & 0x30) >> 4); + char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2); + char_array_3[2] = ((char_array_4[2] & 0x3) << 6) + char_array_4[3]; + + for (i = 0; (i < 3); i++) + { + ret.push_back(char_array_3[i]); + } + i = 0; + } + } + + if (i) + { + for (j = i; j <4; j++) + { + char_array_4[j] = 0; + } + + for (j = 0; j <4; j++) + { + char_array_4[j] = base64_chars.find(char_array_4[j]); + } + + char_array_3[0] = ((char_array_4[0] ) << 2) + ((char_array_4[1] & 0x30) >> 4); + char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2); + char_array_3[2] = ((char_array_4[2] & 0x3) << 6) + char_array_4[3]; + + for (j = 0; (j < i - 1); j++) + { + ret.push_back(char_array_3[j]); + } + } + + return ret; +} + +// +// parallel +// + +enum task_type { + COMPLETION_TASK, + CANCEL_TASK +}; + +struct task_server { + int id; + int target_id; + task_type type; + json data; + bool infill_mode = false; + bool embedding_mode = false; +}; + +struct task_result { + int id; + bool stop; + bool error; + json result_json; +}; + +// TODO: can become bool if we can't find use of more states +enum slot_state +{ + IDLE, + PROCESSING, +}; + +enum slot_command +{ + NONE, + LOAD_PROMPT, + RELEASE, +}; + +struct slot_params +{ + bool stream = true; + bool cache_prompt = false; // remember the prompt to avoid reprocessing all prompt + + uint32_t seed = -1; // RNG seed + int32_t n_keep = 0; // number of tokens to keep from initial prompt + int32_t n_predict = -1; // new tokens to predict + + std::vector antiprompt; + + json input_prefix; + json input_suffix; +}; + +struct slot_image +{ + int32_t id; + + bool request_encode_image = false; + float* image_embedding = nullptr; + int32_t image_tokens = 0; + + clip_image_u8 img_data; + + std::string prefix_prompt; // before of this image +}; + // completion token output with probabilities struct completion_token_output { @@ -45,6 +212,7 @@ struct completion_token_output std::vector probs; llama_token tok; + std::string text_to_send; }; static size_t common_part(const std::vector &a, const std::vector &b) @@ -89,6 +257,7 @@ static size_t find_partial_stop_string(const std::string &stop, return std::string::npos; } +// TODO: reuse llama_detokenize template static std::string tokens_to_str(llama_context *ctx, Iter begin, Iter end) { @@ -103,12 +272,13 @@ static std::string tokens_to_str(llama_context *ctx, Iter begin, Iter end) static void server_log(const char *level, const char *function, int line, const char *message, const nlohmann::ordered_json &extra) { - nlohmann::ordered_json log{ + nlohmann::ordered_json log + { {"timestamp", time(nullptr)}, - {"level", level}, - {"function", function}, - {"line", line}, - {"message", message}, + {"level", level}, + {"function", function}, + {"line", line}, + {"message", message}, }; if (!extra.empty()) @@ -138,7 +308,7 @@ static std::string tokens_to_output_formatted_string(const llama_context *ctx, c } // convert a vector of completion_token_output to json -static json probs_vector_to_json(const llama_context *ctx, const std::vector & probs) +static json probs_vector_to_json(const llama_context *ctx, const std::vector &probs) { json out = json::array(); for (const auto &prob : probs) @@ -147,76 +317,212 @@ static json probs_vector_to_json(const llama_context *ctx, const std::vector +static T json_value(const json &body, const std::string &key, const T &default_value) { - bool stream = false; - bool has_next_token = false; - std::string generated_text; - std::vector generated_token_probs; + // Fallback null to default value + return body.contains(key) && !body.at(key).is_null() + ? body.value(key, default_value) + : default_value; +} - size_t num_prompt_tokens = 0; - size_t num_tokens_predicted = 0; - size_t n_past = 0; - size_t n_remain = 0; +struct llama_client_slot +{ + int id; + int task_id = -1; + + struct slot_params params; + + slot_state state = IDLE; + slot_command command = NONE; + + // used to determine the slot that has been used the longest + int64_t t_last_used = -1; + + // generation props + int32_t n_ctx = 0; // context size per slot + int32_t n_past = 0; + int32_t n_decoded = 0; + int32_t n_remaining = -1; + int32_t i_batch = -1; + + int32_t num_prompt_tokens = 0; + int32_t num_prompt_tokens_processed = 0; + int32_t multibyte_pending = 0; json prompt; - std::vector embd; - - gpt_params params; - - llama_model *model = nullptr; - llama_context *ctx = nullptr; - llama_sampling_context *ctx_sampling = nullptr; - - int n_ctx; + std::string generated_text; + llama_token sampled; + std::vector cache_tokens; + std::vector generated_token_probs; + bool infill = false; + bool embedding = false; + bool has_next_token = true; bool truncated = false; bool stopped_eos = false; bool stopped_word = false; bool stopped_limit = false; + std::string stopping_word; - int32_t multibyte_pending = 0; - std::mutex mutex; + // sampling + struct llama_sampling_params sparams; + llama_sampling_context *ctx_sampling = nullptr; - std::unique_lock lock() - { - return std::unique_lock(mutex); + // multimodal + std::vector images; + + // stats + size_t sent_count = 0; + size_t sent_token_probs_index = 0; + + int64_t t_start_process_prompt; + int64_t t_start_genereration; + + double t_prompt_processing; // ms + double t_token_generation; // ms + + void reset() { + num_prompt_tokens = 0; + generated_text = ""; + truncated = false; + stopped_eos = false; + stopped_word = false; + stopped_limit = false; + stopping_word = ""; + multibyte_pending = 0; + n_past = 0; + sent_count = 0; + sent_token_probs_index = 0; + infill = false; + + generated_token_probs.clear(); + + for (slot_image &img : images) + { + free(img.image_embedding); + delete[] img.img_data.data; + img.prefix_prompt = ""; + } + + images.clear(); + // llama_set_rng_seed(ctx, params.seed); in batched the seed matter??????? } + bool has_budget(gpt_params &global_params) { + n_remaining = -1; + if(params.n_predict != -1) + { + n_remaining = params.n_predict - n_decoded; + } + else if (global_params.n_predict != -1) + { + n_remaining = global_params.n_predict - n_decoded; + } + return n_remaining > 0 || n_remaining == -1; // no budget || limitless + } + + bool available() const { + return state == IDLE && command == NONE; + } + + bool is_processing() const { + return (state == IDLE && command == LOAD_PROMPT) || state == PROCESSING; + } + + void add_token_string(const completion_token_output &token) { + if (command == RELEASE) + { + return; + } + cache_tokens.push_back(token.tok); + generated_token_probs.push_back(token); + } + + void release() { + if (state == IDLE || state == PROCESSING) + { + t_token_generation = (ggml_time_us() - t_start_genereration) / 1e3; + command = RELEASE; + } + } + + json get_formated_timings() { + return json + { + {"prompt_n", num_prompt_tokens_processed}, + {"prompt_ms", t_prompt_processing}, + {"prompt_per_token_ms", t_prompt_processing / num_prompt_tokens_processed}, + {"prompt_per_second", 1e3 / t_prompt_processing * num_prompt_tokens_processed}, + + {"predicted_n", n_decoded}, + {"predicted_ms", t_token_generation}, + {"predicted_per_token_ms", t_token_generation / n_decoded}, + {"predicted_per_second", 1e3 / t_token_generation * n_decoded}, + }; + } + + void print_timings() { + LOG_TEE("\n"); + LOG_TEE("%s: prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n", + __func__, t_prompt_processing, num_prompt_tokens_processed, t_prompt_processing / num_prompt_tokens_processed, 1e3 / t_prompt_processing * num_prompt_tokens_processed); + LOG_TEE("%s: eval time = %10.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)\n", + __func__, t_token_generation, n_decoded,t_token_generation / n_decoded, 1e3 / t_token_generation * n_decoded); + LOG_TEE("%s: total time = %10.2f ms\n", __func__, t_prompt_processing + t_token_generation); + } +}; + +struct llama_server_context +{ + llama_model *model = nullptr; + llama_context *ctx = nullptr; + + clip_ctx *clp_ctx = nullptr; + + gpt_params params; + + llama_batch batch; + + bool multimodal = false; + bool clean_kv_cache = true; + bool all_slots_are_idle = false; + + int32_t id_gen; + int32_t n_ctx; // total context for all clients / slots + + // system prompt + bool system_need_update = false; + + std::string system_prompt; + std::vector system_tokens; + + std::string name_user; // this should be the antiprompt + std::string name_assistant; + + // slots / clients + std::vector slots; + + std::vector queue_tasks; + std::vector queue_results; + std::mutex mutex_tasks; + std::mutex mutex_results; + ~llama_server_context() { if (ctx) @@ -231,46 +537,74 @@ struct llama_server_context } } - void rewind() - { - params.antiprompt.clear(); - params.sparams.grammar.clear(); - num_prompt_tokens = 0; - num_tokens_predicted = 0; - generated_text = ""; - generated_text.reserve(n_ctx); - generated_token_probs.clear(); - truncated = false; - stopped_eos = false; - stopped_word = false; - stopped_limit = false; - stopping_word = ""; - multibyte_pending = 0; - n_remain = 0; - n_past = 0; - params.sparams.n_prev = n_ctx; - } - - void initSampling() { - if (ctx_sampling != nullptr) { - llama_sampling_free(ctx_sampling); - } - ctx_sampling = llama_sampling_init(params.sparams); - } - - bool loadModel(const gpt_params ¶ms_) + bool load_model(const gpt_params ¶ms_) { params = params_; + if (!params.mmproj.empty()) { + multimodal = true; + LOG_TEE("Multi Modal Mode Enabled"); + clp_ctx = clip_model_load(params.mmproj.c_str(), /*verbosity=*/ 1); + if(clp_ctx == nullptr) { + LOG_ERROR("unable to load clip model", {{"model", params.mmproj}}); + return false; + } + + if (params.n_ctx < 2048) { // request larger context for the image embedding + params.n_ctx = 2048; + } + } + std::tie(model, ctx) = llama_init_from_gpt_params(params); if (model == nullptr) { - LOG_ERROR("unable to load model", {{"model", params_.model}}); + LOG_ERROR("unable to load model", {{"model", params.model}}); return false; } + + if (multimodal) { + const int n_embd_clip = clip_n_mmproj_embd(clp_ctx); + const int n_embd_llm = llama_n_embd(model); + if (n_embd_clip != n_embd_llm) { + LOG_TEE("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_embd_clip, n_embd_llm); + llama_free(ctx); + llama_free_model(model); + return false; + } + } + n_ctx = llama_n_ctx(ctx); + return true; } + void initialize() { + id_gen = 0; + + // create slots + all_slots_are_idle = true; + + const int32_t n_ctx_slot = n_ctx / params.n_parallel; + + LOG_TEE("Available slots:\n"); + for (int i = 0; i < params.n_parallel; i++) + { + llama_client_slot slot; + + slot.id = i; + slot.n_ctx = n_ctx_slot; + slot.reset(); + + LOG_TEE(" -> Slot %i - max context: %i\n", slot.id, n_ctx_slot); + slots.push_back(slot); + } + + batch = llama_batch_init(n_ctx, 0, params.n_parallel); + + // empty system prompt + system_prompt = ""; + system_tokens.clear(); + } + std::vector tokenize(const json & json_prompt, bool add_bos) const { // If `add_bos` is true, we only add BOS, when json_prompt is a string, @@ -316,260 +650,274 @@ struct llama_server_context return prompt_tokens; } - void truncatePrompt(std::vector &prompt_tokens) { - const int n_left = n_ctx - params.n_keep; - const int n_block_size = n_left / 2; - const int erased_blocks = (prompt_tokens.size() - params.n_keep - n_block_size) / n_block_size; + llama_client_slot* get_slot(int id) { + int64_t t_last = ggml_time_us(); + llama_client_slot *last_used = nullptr; - // Keep n_keep tokens at start of prompt (at most n_ctx - 4) - std::vector new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + params.n_keep); + for (llama_client_slot & slot : slots) + { + if (slot.id == id && slot.available()) + { + return &slot; + } - new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + params.n_keep + erased_blocks * n_block_size, prompt_tokens.end()); + if (slot.available() && slot.t_last_used < t_last) + { + last_used = &slot; + t_last = slot.t_last_used; + } + } - LOG_VERBOSE("input truncated", { - {"n_ctx", n_ctx}, - {"n_keep", params.n_keep}, - {"n_left", n_left}, - {"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())}, - {"num_prompt_tokens", new_tokens.size()} - }); - - truncated = true; - prompt_tokens = new_tokens; + return last_used; } - void loadInfill() - { - bool suff_rm_leading_spc = true; - if (params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) { - params.input_suffix.erase(0, 1); - suff_rm_leading_spc = false; - } + bool launch_slot_with_data(llama_client_slot* &slot, json data) { + slot_params default_params; + llama_sampling_params default_sparams; - auto prefix_tokens = tokenize(params.input_prefix, false); - auto suffix_tokens = tokenize(params.input_suffix, false); - const int space_token = 29871; - if (suff_rm_leading_spc && suffix_tokens[0] == space_token) { - suffix_tokens.erase(suffix_tokens.begin()); - } - prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(ctx)); - prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(ctx)); // always add BOS - prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(ctx)); - prefix_tokens.insert(prefix_tokens.end(), suffix_tokens.begin(), suffix_tokens.end()); - prefix_tokens.push_back(llama_token_middle(ctx)); + slot->params.stream = json_value(data, "stream", false); + slot->params.cache_prompt = json_value(data, "cache_prompt", false); + slot->params.n_predict = json_value(data, "n_predict", default_params.n_predict); + slot->sparams.top_k = json_value(data, "top_k", default_sparams.top_k); + slot->sparams.top_p = json_value(data, "top_p", default_sparams.top_p); + slot->sparams.tfs_z = json_value(data, "tfs_z", default_sparams.tfs_z); + slot->sparams.typical_p = json_value(data, "typical_p", default_sparams.typical_p); + slot->sparams.temp = json_value(data, "temperature", default_sparams.temp); + slot->sparams.penalty_last_n = json_value(data, "repeat_last_n", default_sparams.penalty_last_n); + slot->sparams.penalty_repeat = json_value(data, "repeat_penalty", default_sparams.penalty_repeat); + slot->sparams.penalty_freq = json_value(data, "frequency_penalty", default_sparams.penalty_freq); + slot->sparams.penalty_present = json_value(data, "presence_penalty", default_sparams.penalty_present); + slot->sparams.mirostat = json_value(data, "mirostat", default_sparams.mirostat); + slot->sparams.mirostat_tau = json_value(data, "mirostat_tau", default_sparams.mirostat_tau); + slot->sparams.mirostat_eta = json_value(data, "mirostat_eta", default_sparams.mirostat_eta); + slot->sparams.penalize_nl = json_value(data, "penalize_nl", default_sparams.penalize_nl); + slot->params.n_keep = json_value(data, "n_keep", slot->params.n_keep); + slot->params.seed = json_value(data, "seed", default_params.seed); + slot->sparams.grammar = json_value(data, "grammar", default_sparams.grammar); + slot->sparams.n_probs = json_value(data, "n_probs", default_sparams.n_probs); - auto prompt_tokens = prefix_tokens; - - num_prompt_tokens = prompt_tokens.size(); - - if (params.n_keep < 0) + // infill + if (data.count("input_prefix") != 0) { - params.n_keep = (int)num_prompt_tokens; + slot->params.input_prefix = data["input_prefix"]; } - params.n_keep = std::min(params.n_ctx - 4, params.n_keep); - - // if input prompt is too big, truncate like normal - if (num_prompt_tokens >= (size_t) n_ctx) + else { - truncatePrompt(prompt_tokens); - num_prompt_tokens = prompt_tokens.size(); - - GGML_ASSERT(num_prompt_tokens < (size_t)n_ctx); + slot->params.input_prefix = ""; } - // push the prompt into the sampling context (do not apply grammar) - for (auto & token : prompt_tokens) + if (data.count("input_suffix") != 0) { - llama_sampling_accept(ctx_sampling, ctx, token, false); + slot->params.input_suffix = data["input_suffix"]; } - - // compare the evaluated prompt with the new prompt - n_past = common_part(embd, prompt_tokens); - embd = prompt_tokens; - - if (n_past == num_prompt_tokens) + else { - // we have to evaluate at least 1 token to generate logits. - printf("we have to evaluate at least 1 token to generate logits\n"); - n_past--; + slot->params.input_suffix = ""; } - // since #3228 we now have to manually manage the KV cache - llama_kv_cache_seq_rm(ctx, 0, n_past, -1); - - LOG_VERBOSE("prompt ingested", { - {"n_past", n_past}, - {"cached", tokens_to_str(ctx, embd.cbegin(), embd.cbegin() + n_past)}, - {"to_eval", tokens_to_str(ctx, embd.cbegin() + n_past, embd.cend())}, - }); - - has_next_token = true; - } - void loadPrompt() - { - auto prompt_tokens = tokenize(prompt, true); // always add BOS - - num_prompt_tokens = prompt_tokens.size(); - - if (params.n_keep < 0) + if (data.count("prompt") != 0) { - params.n_keep = (int)num_prompt_tokens; + slot->prompt = data["prompt"]; } - params.n_keep = std::min(n_ctx - 4, params.n_keep); - - // if input prompt is too big, truncate like normal - if (num_prompt_tokens >= (size_t) n_ctx) + else { - truncatePrompt(prompt_tokens); - num_prompt_tokens = prompt_tokens.size(); - - GGML_ASSERT(num_prompt_tokens < (size_t)n_ctx); + slot->prompt = ""; } - // push the prompt into the sampling context (do not apply grammar) - for (auto & token : prompt_tokens) + slot->sparams.logit_bias.clear(); + + if (json_value(data, "ignore_eos", false)) { - llama_sampling_accept(ctx_sampling, ctx, token, false); + slot->sparams.logit_bias[llama_token_eos(model)] = -INFINITY; } - // compare the evaluated prompt with the new prompt - n_past = common_part(embd, prompt_tokens); - - embd = prompt_tokens; - if (n_past == num_prompt_tokens) + const auto &logit_bias = data.find("logit_bias"); + if (logit_bias != data.end() && logit_bias->is_array()) { - // we have to evaluate at least 1 token to generate logits. - n_past--; + const int n_vocab = llama_n_vocab(model); + for (const auto &el : *logit_bias) + { + if (el.is_array() && el.size() == 2 && el[0].is_number_integer()) + { + llama_token tok = el[0].get(); + if (tok >= 0 && tok < n_vocab) + { + if (el[1].is_number()) + { + slot->sparams.logit_bias[tok] = el[1].get(); + } + else if (el[1].is_boolean() && !el[1].get()) + { + slot->sparams.logit_bias[tok] = -INFINITY; + } + } + } + } } - // since #3228 we now have to manually manage the KV cache - llama_kv_cache_seq_rm(ctx, 0, n_past, -1); + slot->params.antiprompt.clear(); - LOG_VERBOSE("prompt ingested", { - {"n_past", n_past}, - {"cached", tokens_to_str(ctx, embd.cbegin(), embd.cbegin() + n_past)}, - {"to_eval", tokens_to_str(ctx, embd.cbegin() + n_past, embd.cend())}, - }); + const auto &stop = data.find("stop"); + if (stop != data.end() && stop->is_array()) + { + for (const auto &word : *stop) + { + if (!word.empty()) + { + slot->params.antiprompt.push_back(word); + } + } + } - has_next_token = true; + if (multimodal) + { + const auto &images_data = data.find("image_data"); + if (images_data != data.end() && images_data->is_array()) + { + for (const auto &img : *images_data) + { + std::string data_b64 = img["data"].get(); + slot_image img_sl; + img_sl.id = img.count("id") != 0 ? img["id"].get() : slot->images.size(); + int width, height, channels; + std::vector image_buffer = base64_decode(data_b64); + data_b64.clear(); + auto data = stbi_load_from_memory(image_buffer.data(), image_buffer.size(), &width, &height, &channels, 3); + if (!data) { + LOG_TEE("slot %i - failed to load image [id: %i]\n", slot->id, img_sl.id); + return false; + } + LOG_TEE("slot %i - image loaded [id: %i] resolution (%i x %i)\n", slot->id, img_sl.id, width, height); + img_sl.img_data.nx = width; + img_sl.img_data.ny = height; + img_sl.img_data.size = width * height * 3; + img_sl.img_data.data = new uint8_t[width * height * 3](); + memcpy(img_sl.img_data.data, data, width * height * 3); + stbi_image_free(data); + img_sl.request_encode_image = true; + slot->images.push_back(img_sl); + } + // process prompt + // example: system prompt [img-102] user [img-103] describe [img-134] -> [{id: 102, prefix: 'system prompt '}, {id: 103, prefix: ' user '}, {id: 134, prefix: ' describe '}]} + if (slot->images.size() > 0 && !slot->prompt.is_array()) + { + std::string prompt = slot->prompt.get(); + size_t pos = 0, begin_prefix = 0; + std::string pattern = "[img-"; + while ((pos = prompt.find(pattern, pos)) != std::string::npos) { + size_t end_prefix = pos; + pos += pattern.length(); + size_t end_pos = prompt.find("]", pos); + if (end_pos != std::string::npos) + { + std::string image_id = prompt.substr(pos, end_pos - pos); + try + { + int img_id = std::stoi(image_id); + bool found = false; + for (slot_image &img : slot->images) + { + if (img.id == img_id) { + found = true; + img.prefix_prompt = prompt.substr(begin_prefix, end_prefix - begin_prefix); + begin_prefix = end_pos + 1; + break; + } + } + if (!found) { + LOG_TEE("ERROR: Image with id: %i, not found.\n", img_id); + slot->images.clear(); + return false; + } + } catch (const std::invalid_argument& e) { + LOG_TEE("Invalid image number id in prompt\n"); + slot->images.clear(); + return false; + } + } + } + slot->prompt = ""; + slot->params.input_suffix = prompt.substr(begin_prefix); + slot->params.cache_prompt = false; // multimodal doesn't support cache prompt + } + } + } + + if (slot->ctx_sampling != nullptr) + { + llama_sampling_free(slot->ctx_sampling); + } + slot->ctx_sampling = llama_sampling_init(slot->sparams); + slot->command = LOAD_PROMPT; + + all_slots_are_idle = false; + + LOG_TEE("slot %i is processing [task id: %i]\n", slot->id, slot->task_id); + + return true; } - void beginCompletion() - { - // number of tokens to keep when resetting context - n_remain = params.n_predict; - llama_set_rng_seed(ctx, params.seed); + void kv_cache_clear() { + // clear the entire KV cache + llama_kv_cache_clear(ctx); + clean_kv_cache = false; } - completion_token_output nextToken() - { - completion_token_output result; - result.tok = -1; + void update_system_prompt() { + system_tokens = ::llama_tokenize(ctx, system_prompt, true); - if (embd.size() >= (size_t)n_ctx) + llama_batch_clear(batch); + + kv_cache_clear(); + + for (int i = 0; i < (int) system_tokens.size(); ++i) { - // Shift context - - const int n_left = n_past - params.n_keep - 1; - const int n_discard = n_left/2; - - llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1 , params.n_keep + n_discard + 1); - llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard); - - for (size_t i = params.n_keep + 1 + n_discard; i < embd.size(); i++) - { - embd[i - n_discard] = embd[i]; - } - embd.resize(embd.size() - n_discard); - - n_past -= n_discard; - - truncated = true; - LOG_VERBOSE("input truncated", { - {"n_ctx", n_ctx}, - {"n_keep", params.n_keep}, - {"n_left", n_left}, - }); + llama_batch_add(batch, system_tokens[i], i, { 0 }, false); } - bool tg = true; - while (n_past < embd.size()) + if (llama_decode(ctx, batch) != 0) { - int n_eval = (int)embd.size() - n_past; - tg = n_eval == 1; - if (n_eval > params.n_batch) - { - n_eval = params.n_batch; - } - - if (llama_decode(ctx, llama_batch_get_one(&embd[n_past], n_eval, n_past, 0))) - { - LOG_ERROR("failed to eval", { - {"n_eval", n_eval}, - {"n_past", n_past}, - {"embd", tokens_to_str(ctx, embd.cbegin() + n_past, embd.cend())}, - }); - has_next_token = false; - return result; - } - n_past += n_eval; + LOG_TEE("%s: llama_decode() failed\n", __func__); + return; } - if (params.n_predict == 0) + // assign the system KV cache to all parallel sequences + for (int32_t i = 1; i < params.n_parallel; ++i) { - has_next_token = false; - result.tok = llama_token_eos(ctx); - return result; + llama_kv_cache_seq_cp(ctx, 0, i, 0, system_tokens.size()); } - { - // out of user input, sample next token - result.tok = llama_sampling_sample(ctx_sampling, ctx, NULL); - - llama_token_data_array cur_p = { ctx_sampling->cur.data(), ctx_sampling->cur.size(), false }; - - const int32_t n_probs = params.sparams.n_probs; - if (params.sparams.temp <= 0 && n_probs > 0) - { - // For llama_sample_token_greedy we need to sort candidates - llama_sample_softmax(ctx, &cur_p); - } - - for (size_t i = 0; i < std::min(cur_p.size, (size_t)n_probs); ++i) - { - result.probs.push_back({cur_p.data[i].id, cur_p.data[i].p}); - } - - llama_sampling_accept(ctx_sampling, ctx, result.tok, true); - - if (tg) { - num_tokens_predicted++; - } - } - - // add it to the context - embd.push_back(result.tok); - // decrement remaining sampling budget - --n_remain; - - if (!embd.empty() && embd.back() == llama_token_eos(ctx)) - { - // stopping_word = llama_token_to_piece(ctx, embd.back()); - has_next_token = false; - stopped_eos = true; - LOG_VERBOSE("eos token found", {}); - return result; - } - - has_next_token = params.n_predict == -1 || n_remain != 0; - return result; + LOG_TEE("system prompt updated\n"); + system_need_update = false; } - size_t findStoppingStrings(const std::string &text, const size_t last_token_size, - const stop_type type) + void notify_system_prompt_changed() { + // release all slots + for (llama_client_slot &slot : slots) + { + slot.release(); + } + + system_need_update = true; + } + + void process_system_prompt_data(const json &sys_props) { + system_prompt = sys_props.value("prompt", ""); + name_user = sys_props.value("anti_prompt", ""); + name_assistant = sys_props.value("assistant_name", ""); + + if (slots.size() > 0) + { + notify_system_prompt_changed(); + } + } + + static size_t find_stopping_strings(const std::string &text, const size_t last_token_size, + const stop_type type, llama_client_slot &slot) { size_t stop_pos = std::string::npos; - for (const std::string &word : params.antiprompt) + + for (const std::string &word : slot.params.antiprompt) { size_t pos; if (type == STOP_FULL) @@ -587,95 +935,815 @@ struct llama_server_context { if (type == STOP_FULL) { - stopping_word = word; - stopped_word = true; - has_next_token = false; + slot.stopped_word = true; + slot.stopping_word = word; + slot.has_next_token = false; } stop_pos = pos; } } + return stop_pos; } - completion_token_output doCompletion() - { - auto token_with_probs = nextToken(); + bool process_token(completion_token_output &result, llama_client_slot &slot) { + // remember which tokens were sampled - used for repetition penalties during sampling + const std::string token_str = llama_token_to_piece(ctx, result.tok); + slot.sampled = result.tok; - const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_piece(ctx, token_with_probs.tok); - generated_text += token_text; + // search stop word and delete it + slot.generated_text += token_str; + slot.has_next_token = true; - if (params.sparams.n_probs > 0) + if (slot.multibyte_pending > 0) { - generated_token_probs.push_back(token_with_probs); + slot.multibyte_pending -= token_str.size(); } - - if (multibyte_pending > 0) + else if (token_str.size() == 1) { - multibyte_pending -= token_text.size(); - } - else if (token_text.size() == 1) - { - const char c = token_text[0]; + const char c = token_str[0]; // 2-byte characters: 110xxxxx 10xxxxxx if ((c & 0xE0) == 0xC0) { - multibyte_pending = 1; + slot.multibyte_pending = 1; // 3-byte characters: 1110xxxx 10xxxxxx 10xxxxxx } else if ((c & 0xF0) == 0xE0) { - multibyte_pending = 2; + slot.multibyte_pending = 2; // 4-byte characters: 11110xxx 10xxxxxx 10xxxxxx 10xxxxxx } else if ((c & 0xF8) == 0xF0) { - multibyte_pending = 3; + slot.multibyte_pending = 3; } else { - multibyte_pending = 0; + slot.multibyte_pending = 0; } } - if (multibyte_pending > 0 && !has_next_token) + if (slot.multibyte_pending == 0) { - has_next_token = true; - n_remain++; + size_t pos = std::min(slot.sent_count, slot.generated_text.size()); + const std::string str_test = slot.generated_text.substr(pos); + bool is_stop_full = false; + size_t stop_pos = find_stopping_strings(str_test, token_str.size(), STOP_FULL, slot); + if (stop_pos != std::string::npos) + { + is_stop_full = true; + slot.generated_text.erase( + slot.generated_text.begin() + pos + stop_pos, + slot.generated_text.end()); + pos = std::min(slot.sent_count, slot.generated_text.size()); + } + else + { + is_stop_full = false; + stop_pos = find_stopping_strings(str_test, token_str.size(), STOP_PARTIAL, slot); + } + + // check if there is any token to predict + if (stop_pos == std::string::npos || (!slot.has_next_token && !is_stop_full && stop_pos > 0)) + { + // no send the stop word in the response + result.text_to_send = slot.generated_text.substr(pos, std::string::npos); + slot.sent_count += result.text_to_send.size(); + // add the token to slot queue and cache + } + slot.add_token_string(result); + if (slot.params.stream) + { + send_partial_response(slot, result); + } } - if (!has_next_token && n_remain == 0) + if (slot.multibyte_pending > 0 && !slot.has_next_token) { - stopped_limit = true; + slot.has_next_token = true; + } + + // check the limits + if (slot.n_decoded > 2 && slot.has_next_token && !slot.has_budget(params)) + { + slot.stopped_limit = true; + slot.has_next_token = false; + } + + if (!slot.cache_tokens.empty() && result.tok == llama_token_eos(model)) + { + slot.stopped_eos = true; + slot.has_next_token = false; + LOG_VERBOSE("eos token found", {}); } LOG_VERBOSE("next token", { - {"token", token_with_probs.tok}, - {"token_text", tokens_to_output_formatted_string(ctx, token_with_probs.tok)}, - {"has_next_token", has_next_token}, - {"n_remain", n_remain}, - {"num_tokens_predicted", num_tokens_predicted}, - {"stopped_eos", stopped_eos}, - {"stopped_word", stopped_word}, - {"stopped_limit", stopped_limit}, - {"stopping_word", stopping_word}, + {"token", result.tok}, + {"token_text", tokens_to_output_formatted_string(ctx, result.tok)}, + {"has_next_token", slot.has_next_token}, + {"n_remain", slot.n_remaining}, + {"num_tokens_predicted", slot.n_decoded}, + {"stopped_eos", slot.stopped_eos}, + {"stopped_word", slot.stopped_word}, + {"stopped_limit", slot.stopped_limit}, + {"stopping_word", slot.stopping_word}, }); - return token_with_probs; + return slot.has_next_token; // continue } - std::vector getEmbedding() + bool process_images(llama_client_slot &slot) const { - static const int n_embd = llama_n_embd(model); + for (slot_image &img : slot.images) + { + if (!img.request_encode_image) + { + continue; + } + clip_image_f32 img_res; + if (!clip_image_preprocess(clp_ctx, &img.img_data, &img_res, /*pad2square =*/ true)) + { + LOG_TEE("Error processing the given image"); + clip_free(clp_ctx); + return false; + } + img.image_tokens = clip_n_patches(clp_ctx); + img.image_embedding = (float *)malloc(clip_embd_nbytes(clp_ctx)); + if (!img.image_embedding) + { + LOG_TEE("Unable to allocate memory for image embeddings\n"); + clip_free(clp_ctx); + return false; + } + LOG_TEE("slot %i - encoding image [id: %i]\n", slot.id, img.id); + if (!clip_image_encode(clp_ctx, params.n_threads, &img_res, img.image_embedding)) + { + LOG_TEE("Unable to encode image\n"); + return false; + } + img.request_encode_image = false; + } + + return slot.images.size() > 0; + } + + void send_error(int id, std::string error) + { + std::lock_guard lock(mutex_results); + task_result res; + res.id = id; + res.error = true; + res.result_json = { { "content", error } }; + queue_results.push_back(res); + } + + json get_model_props() + { + return get_formated_generation(slots[0]); + } + + json get_formated_generation(llama_client_slot &slot) + { + const auto eos_bias = slot.sparams.logit_bias.find(llama_token_eos(model)); + const bool ignore_eos = eos_bias != slot.sparams.logit_bias.end() && + eos_bias->second < 0.0f && std::isinf(eos_bias->second); + return json { + {"n_ctx", slot.n_ctx}, + {"model", params.model_alias}, + {"seed", slot.params.seed}, + {"temp", slot.sparams.temp}, + {"top_k", slot.sparams.top_k}, + {"top_p", slot.sparams.top_p}, + {"tfs_z", slot.sparams.tfs_z}, + {"typical_p", slot.sparams.typical_p}, + {"repeat_last_n", slot.sparams.penalty_last_n}, + {"repeat_penalty", slot.sparams.penalty_repeat}, + {"presence_penalty", slot.sparams.penalty_present}, + {"frequency_penalty", slot.sparams.penalty_freq}, + {"mirostat", slot.sparams.mirostat}, + {"mirostat_tau", slot.sparams.mirostat_tau}, + {"mirostat_eta", slot.sparams.mirostat_eta}, + {"penalize_nl", slot.sparams.penalize_nl}, + {"stop", slot.params.antiprompt}, + {"n_predict", slot.params.n_predict}, + {"n_keep", params.n_keep}, + {"ignore_eos", ignore_eos}, + {"stream", slot.params.stream}, + {"logit_bias", slot.sparams.logit_bias}, + {"n_probs", slot.sparams.n_probs}, + {"grammar", slot.sparams.grammar}, + }; + } + + void send_partial_response(llama_client_slot &slot, completion_token_output tkn) + { + std::lock_guard lock(mutex_results); + task_result res; + res.id = slot.task_id; + res.error = false; + res.stop = false; + + res.result_json = json + { + {"content", tkn.text_to_send}, + {"stop", false}, + {"slot_id", slot.id}, + {"multimodal", multimodal} + }; + + if (slot.sparams.n_probs > 0) + { + std::vector probs_output = {}; + const std::vector to_send_toks = llama_tokenize(ctx, tkn.text_to_send, false); + size_t probs_pos = std::min(slot.sent_token_probs_index, slot.generated_token_probs.size()); + size_t probs_stop_pos = std::min(slot.sent_token_probs_index + to_send_toks.size(), slot.generated_token_probs.size()); + if (probs_pos < probs_stop_pos) + { + probs_output = std::vector(slot.generated_token_probs.begin() + probs_pos, slot.generated_token_probs.begin() + probs_stop_pos); + } + slot.sent_token_probs_index = probs_stop_pos; + res.result_json["completion_probabilities"] = probs_vector_to_json(ctx, probs_output); + } + + queue_results.push_back(res); + } + + void send_final_response(llama_client_slot &slot) + { + std::lock_guard lock(mutex_results); + task_result res; + res.id = slot.task_id; + res.error = false; + res.stop = true; + + res.result_json = json + { + {"content", !slot.params.stream ? slot.generated_text : ""}, + {"slot_id", slot.id}, + {"stop", true}, + {"model", params.model_alias}, + {"tokens_predicted", slot.n_decoded}, + {"tokens_evaluated", slot.num_prompt_tokens}, + {"generation_settings", get_formated_generation(slot)}, + {"prompt", slot.prompt}, + {"truncated", slot.truncated}, + {"stopped_eos", slot.stopped_eos}, + {"stopped_word", slot.stopped_word}, + {"stopped_limit", slot.stopped_limit}, + {"stopping_word", slot.stopping_word}, + {"tokens_cached", slot.n_past}, + {"timings", slot.get_formated_timings()} + }; + + if (slot.sparams.n_probs > 0) + { + std::vector probs = {}; + if (!slot.params.stream && slot.stopped_word) + { + const std::vector stop_word_toks = llama_tokenize(ctx, slot.stopping_word, false); + probs = std::vector(slot.generated_token_probs.begin(), slot.generated_token_probs.end() - stop_word_toks.size()); + } + else + { + probs = std::vector( + slot.generated_token_probs.begin(), + slot.generated_token_probs.begin() + slot.sent_token_probs_index); + } + res.result_json["completion_probabilities"] = probs_vector_to_json(ctx, probs); + } + + queue_results.push_back(res); + } + + void send_embedding(llama_client_slot &slot) + { + std::lock_guard lock(mutex_results); + task_result res; + res.id = slot.task_id; + res.error = false; + res.stop = true; + + const int n_embd = llama_n_embd(model); if (!params.embedding) { LOG_WARNING("embedding disabled", { {"params.embedding", params.embedding}, }); - return std::vector(n_embd, 0.0f); + res.result_json = json + { + {"embedding", std::vector(n_embd, 0.0f)}, + }; } - const float *data = llama_get_embeddings(ctx); - std::vector embedding(data, data + n_embd); - return embedding; + else + { + const float *data = llama_get_embeddings(ctx); + std::vector embedding(data, data + n_embd); + res.result_json = json + { + {"embedding", embedding }, + }; + } + queue_results.push_back(res); + } + + int request_completion(json data, bool infill, bool embedding) + { + std::lock_guard lock(mutex_tasks); + task_server task; + task.id = id_gen++; + task.data = data; + task.infill_mode = infill; + task.embedding_mode = embedding; + task.type = COMPLETION_TASK; + queue_tasks.push_back(task); + return task.id; + } + + task_result next_result(int task_id) + { + while (true) + { + std::this_thread::sleep_for(std::chrono::microseconds(5)); + std::lock_guard lock(mutex_results); + + if (queue_results.empty()) + { + continue; + } + + for (int i = 0; i < (int) queue_results.size(); i++) + { + if (queue_results[i].id == task_id) + { + task_result res = queue_results[i]; + queue_results.erase(queue_results.begin() + i); + return res; + } + } + } + + // never reached + //return task_result{-1, false, false, {}}; + } + + // for multiple images processing + bool ingest_images(llama_client_slot &slot, int n_batch) + { + int image_idx = 0; + + while (image_idx < (int) slot.images.size()) + { + slot_image &img = slot.images[image_idx]; + + // process prefix prompt + for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) + { + const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i)); + llama_batch batch_view = { + n_tokens, + batch.token + i, + nullptr, + batch.pos + i, + batch.n_seq_id + i, + batch.seq_id + i, + batch.logits + i, + 0, 0, 0, // unused + }; + if (llama_decode(ctx, batch_view)) + { + LOG_TEE("%s : failed to eval\n", __func__); + return false; + } + } + + // process image with llm + for (int i = 0; i < img.image_tokens; i += n_batch) + { + int n_eval = img.image_tokens - i; + if (n_eval > n_batch) + { + n_eval = n_batch; + } + + const int n_embd = llama_n_embd(model); + llama_batch batch_img = { n_eval, nullptr, (img.image_embedding + i * n_embd), nullptr, nullptr, nullptr, nullptr, slot.n_past, 1, 0, }; + if (llama_decode(ctx, batch_img)) + { + LOG_TEE("%s : failed to eval image\n", __func__); + return false; + } + slot.n_past += n_eval; + } + image_idx++; + + llama_batch_clear(batch); + + // append prefix of next image + const auto json_prompt = (image_idx >= (int) slot.images.size()) ? + slot.params.input_suffix : // no more images, then process suffix prompt + (json)(slot.images[image_idx].prefix_prompt); + + std::vector append_tokens = tokenize(json_prompt, false); // has next image + for (int i = 0; i < (int) append_tokens.size(); ++i) + { + llama_batch_add(batch, append_tokens[i], slot.n_past, { slot.id }, true); + slot.n_past += 1; + } + } + + return true; + } + + void request_cancel(int task_id) + { + std::lock_guard lock(mutex_tasks); + task_server task; + task.id = id_gen++; + task.type = CANCEL_TASK; + task.target_id = task_id; + queue_tasks.push_back(task); + } + + void process_tasks() + { + std::lock_guard lock(mutex_tasks); + while (!queue_tasks.empty()) + { + task_server task = queue_tasks.front(); + queue_tasks.erase(queue_tasks.begin()); + switch (task.type) + { + case COMPLETION_TASK: { + llama_client_slot *slot = get_slot(json_value(task.data, "slot_id", -1)); + if (slot == nullptr) + { + LOG_TEE("slot unavailable\n"); + // send error result + send_error(task.id, "slot unavailable"); + return; + } + + if (task.data.contains("system_prompt")) + { + process_system_prompt_data(task.data["system_prompt"]); + } + + slot->reset(); + + slot->infill = task.infill_mode; + slot->embedding = task.embedding_mode; + slot->task_id = task.id; + + if (!launch_slot_with_data(slot, task.data)) + { + // send error result + send_error(task.id, "internal_error"); + break; + } + } break; + case CANCEL_TASK: { // release slot linked with the task id + for (auto & slot : slots) + { + if (slot.task_id == task.target_id) + { + slot.release(); + break; + } + } + } break; + } + } + } + + bool update_slots() { + // attend tasks + process_tasks(); + + // update the system prompt wait until all slots are idle state + if (system_need_update && all_slots_are_idle) + { + LOG_TEE("updating system prompt\n"); + update_system_prompt(); + } + + llama_batch_clear(batch); + + if (all_slots_are_idle) + { + if (system_prompt.empty() && clean_kv_cache) + { + LOG_TEE("all slots are idle and system prompt is empty, clear the KV cache\n"); + kv_cache_clear(); + } + // avoid 100% usage of cpu all time + std::this_thread::sleep_for(std::chrono::milliseconds(5)); + } + + for (llama_client_slot &slot : slots) + { + if (slot.is_processing() && slot.cache_tokens.size() >= (size_t) slot.n_ctx) + { + // Shift context + const int n_left = slot.n_past - slot.params.n_keep - 1; + const int n_discard = n_left / 2; + + LOG_TEE("slot %d: context shift - n_keep = %d, n_left = %d, n_discard = %d\n", slot.id, slot.params.n_keep, n_left, n_discard); + llama_kv_cache_seq_rm (ctx, slot.id, slot.params.n_keep + 1 , slot.params.n_keep + n_discard + 1); + llama_kv_cache_seq_shift(ctx, slot.id, slot.params.n_keep + 1 + n_discard, slot.n_past, -n_discard); + + for (size_t i = slot.params.n_keep + 1 + n_discard; i < slot.cache_tokens.size(); i++) + { + slot.cache_tokens[i - n_discard] = slot.cache_tokens[i]; + } + + slot.cache_tokens.resize(slot.cache_tokens.size() - n_discard); + + slot.n_past -= n_discard; + + slot.truncated = true; + + LOG_VERBOSE("context shift", { + {"n_ctx", n_ctx}, + {"n_keep", params.n_keep}, + {"n_left", n_left}, + }); + } + } + + // decode any currently ongoing sequences + for (auto & slot : slots) + { + // release the slot + if (slot.command == RELEASE) + { + slot.state = IDLE; + slot.command = NONE; + slot.t_last_used = ggml_time_us(); + + LOG_TEE("slot %d released (%d tokens in cache)\n", slot.id, (int) slot.cache_tokens.size()); + + continue; + } + + if (slot.state == IDLE) + { + continue; + } + + slot.i_batch = batch.n_tokens; + + llama_batch_add(batch, slot.sampled, system_tokens.size() + slot.n_past, { slot.id }, true); + + slot.n_decoded += 1; + slot.n_past += 1; + } + + // process in chunks of params.n_batch + int32_t n_batch = params.n_batch; + + // assign workload to the slots + if (params.cont_batching || batch.n_tokens == 0) + { + for (auto & slot : slots) + { + const bool has_prompt = slot.prompt.is_array() || (slot.prompt.is_string() && !slot.prompt.get().empty()) || !slot.images.empty(); + + // empty prompt passed -> release the slot and send empty response + if (slot.state == IDLE && slot.command == LOAD_PROMPT && !has_prompt) + { + slot.release(); + slot.print_timings(); + send_final_response(slot); + continue; + } + + // need process the prompt + if (slot.state == IDLE && slot.command == LOAD_PROMPT) + { + slot.state = PROCESSING; + slot.command = NONE; + std::vector prompt_tokens; + slot.t_start_process_prompt = ggml_time_us(); + slot.t_start_genereration = 0; + + if (slot.infill) + { + bool suff_rm_leading_spc = true; + if (params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) + { + params.input_suffix.erase(0, 1); + suff_rm_leading_spc = false; + } + auto prefix_tokens = tokenize(slot.params.input_prefix, false); + auto suffix_tokens = tokenize(slot.params.input_suffix, false); + + const int space_token = 29871; // TODO: this should not be hardcoded + if (suff_rm_leading_spc && !suffix_tokens.empty() && suffix_tokens[0] == space_token) { + suffix_tokens.erase(suffix_tokens.begin()); + } + + prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(model)); + prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(model)); // always add BOS + prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(model)); + prefix_tokens.insert(prefix_tokens.end(), suffix_tokens.begin(), suffix_tokens.end()); + prefix_tokens.push_back(llama_token_middle(model)); + prompt_tokens = prefix_tokens; + } + else + { + prompt_tokens = tokenize(slot.prompt, system_prompt.empty()); // add BOS if there isn't system prompt + } + + slot.num_prompt_tokens = prompt_tokens.size(); + + if (!slot.params.cache_prompt) + { + llama_sampling_reset(slot.ctx_sampling); + + slot.n_past = 0; + slot.num_prompt_tokens_processed = slot.num_prompt_tokens; + } + else + { + if (slot.params.n_keep < 0) + { + slot.params.n_keep = slot.num_prompt_tokens; + } + slot.params.n_keep = std::min(slot.n_ctx - 4, slot.params.n_keep); + + // if input prompt is too big, truncate it + if (slot.num_prompt_tokens >= slot.n_ctx) + { + const int n_left = slot.n_ctx - slot.params.n_keep; + const int n_block_size = n_left / 2; + const int erased_blocks = (slot.num_prompt_tokens - slot.params.n_keep - n_block_size) / n_block_size; + + std::vector new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + slot.params.n_keep); + new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + slot.params.n_keep + erased_blocks * n_block_size, prompt_tokens.end()); + + LOG_VERBOSE("input truncated", { + {"n_ctx", slot.n_ctx}, + {"n_keep", slot.params.n_keep}, + {"n_left", n_left}, + {"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())}, + }); + slot.truncated = true; + prompt_tokens = new_tokens; + + slot.num_prompt_tokens = prompt_tokens.size(); + GGML_ASSERT(slot.num_prompt_tokens < slot.n_ctx); + } + + // push the prompt into the sampling context (do not apply grammar) + for (auto &token : prompt_tokens) + { + llama_sampling_accept(slot.ctx_sampling, ctx, token, false); + } + + slot.n_past = common_part(slot.cache_tokens, prompt_tokens); + slot.num_prompt_tokens_processed = slot.num_prompt_tokens - slot.n_past; + + LOG_TEE("slot %d : in cache: %i tokens | to process: %i tokens\n", slot.id, slot.n_past, slot.num_prompt_tokens_processed); + } + + LOG_TEE("slot %d : kv cache rm - [%d, end)\n", slot.id, (int) system_tokens.size() + slot.n_past); + + llama_kv_cache_seq_rm(ctx, slot.id, system_tokens.size() + slot.n_past, -1); + + slot.cache_tokens = prompt_tokens; + + if (slot.n_past == slot.num_prompt_tokens) + { + // we have to evaluate at least 1 token to generate logits. + LOG_TEE("slot %d : we have to evaluate at least 1 token to generate logits\n", slot.id); + slot.n_past--; + } + + LOG_VERBOSE("prompt ingested", { + {"n_past", slot.n_past}, + {"cached", tokens_to_str(ctx, slot.cache_tokens.cbegin(), slot.cache_tokens.cbegin() + slot.n_past)}, + {"to_eval", tokens_to_str(ctx, slot.cache_tokens.cbegin() + slot.n_past, slot.cache_tokens.cend())}, + }); + + const bool has_images = process_images(slot); + + // process the prefix of first image + std::vector prefix_tokens = has_images ? tokenize(slot.images[0].prefix_prompt, true) : prompt_tokens; + for (; slot.n_past < (int) prefix_tokens.size(); ++slot.n_past) + { + llama_batch_add(batch, prefix_tokens[slot.n_past], system_tokens.size() + slot.n_past, { slot.id }, false); + } + + if (has_images && !ingest_images(slot, n_batch)) + { + LOG_TEE("failed processing images\n"); + return false; + } + + // extract the logits only for the last token + if (batch.n_tokens > 0) + { + batch.logits[batch.n_tokens - 1] = true; + } + + slot.n_decoded = 0; + slot.i_batch = batch.n_tokens - 1; + } + } + } + + if (batch.n_tokens == 0) + { + all_slots_are_idle = true; + return true; + } + + for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) + { + const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i)); + llama_batch batch_view = + { + n_tokens, + batch.token + i, + nullptr, + batch.pos + i, + batch.n_seq_id + i, + batch.seq_id + i, + batch.logits + i, + 0, 0, 0, // unused + }; + + const int ret = llama_decode(ctx, batch_view); + if (ret != 0) + { + if (n_batch == 1 || ret < 0) + { + // if you get here, it means the KV cache is full - try increasing it via the context size + LOG_TEE("%s : failed to decode the batch, n_batch = %d, ret = %d\n", __func__, n_batch, ret); + return false; + } + + LOG_TEE("%s : failed to find free space in the KV cache, retrying with smaller n_batch = %d\n", __func__, n_batch / 2); + + // retry with half the batch size to try to find a free slot in the KV cache + n_batch /= 2; + i -= n_batch; + continue; + } + + for (auto & slot : slots) + { + if (slot.i_batch < (int) i || slot.i_batch >= (int) (i + n_tokens)) + { + continue; + } + + // prompt evaluated for embedding + if (slot.embedding) + { + send_embedding(slot); + slot.release(); + slot.i_batch = -1; + return true; + } + + completion_token_output result; + const llama_token id = llama_sampling_sample(slot.ctx_sampling, ctx, NULL, slot.i_batch - i); + + llama_sampling_accept(slot.ctx_sampling, ctx, id, true); + + if (slot.n_decoded == 1) + { + slot.t_start_genereration = ggml_time_us(); + slot.t_prompt_processing = (slot.t_start_genereration - slot.t_start_process_prompt) / 1e3; + } + + llama_token_data_array cur_p = { slot.ctx_sampling->cur.data(), slot.ctx_sampling->cur.size(), false }; + result.tok = id; + + const int32_t n_probs = slot.sparams.n_probs; + if (slot.sparams.temp <= 0 && n_probs > 0) + { + // for llama_sample_token_greedy we need to sort candidates + llama_sample_softmax(ctx, &cur_p); + } + + for (size_t i = 0; i < std::min(cur_p.size, (size_t)n_probs); ++i) + { + result.probs.push_back({cur_p.data[i].id, cur_p.data[i].p}); + } + + if (!process_token(result, slot)) + { + slot.release(); + slot.print_timings(); + send_final_response(slot); + } + + slot.i_batch = -1; + } + } + return true; } }; @@ -725,11 +1793,16 @@ static void server_print_usage(const char *argv0, const gpt_params ¶ms, printf(" --path PUBLIC_PATH path from which to serve static files (default %s)\n", sparams.public_path.c_str()); printf(" -to N, --timeout N server read/write timeout in seconds (default: %d)\n", sparams.read_timeout); printf(" --embedding enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled"); + printf(" -np N, --parallel N number of slots for process requests (default: %d)\n", params.n_parallel); + printf(" -cb, --cont-batching enable continuous batching (a.k.a dynamic batching) (default: disabled)\n"); + printf(" -spf FNAME, --system-prompt-file FNAME\n"); + printf(" Set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications.\n"); + printf(" --mmproj MMPROJ_FILE path to a multimodal projector file for LLaVA.\n"); printf("\n"); } static void server_params_parse(int argc, char **argv, server_params &sparams, - gpt_params ¶ms) + gpt_params ¶ms, llama_server_context& llama) { gpt_params default_params; server_params default_sparams; @@ -984,6 +2057,56 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, { params.embedding = true; } + else if (arg == "-cb" || arg == "--cont-batching") + { + params.cont_batching = true; + } + else if (arg == "-np" || arg == "--parallel") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + params.n_parallel = std::stoi(argv[i]); + } else if (arg == "-n" || arg == "--n-predict") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + params.n_predict = std::stoi(argv[i]); + } else if (arg == "-spf" || arg == "--system-prompt-file") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + std::ifstream file(argv[i]); + if (!file) { + fprintf(stderr, "error: failed to open file '%s'\n", argv[i]); + invalid_param = true; + break; + } + std::string systm_content; + std::copy( + std::istreambuf_iterator(file), + std::istreambuf_iterator(), + std::back_inserter(systm_content) + ); + llama.process_system_prompt_data(json::parse(systm_content)); + } + else if(arg == "--mmproj") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + params.mmproj = argv[i]; + } else { fprintf(stderr, "error: unknown argument: %s\n", arg.c_str()); @@ -1000,102 +2123,18 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, } } -static json format_generation_settings(llama_server_context &llama) -{ - const auto & sparams = llama.params.sparams; - const auto eos_bias = sparams.logit_bias.find(llama_token_eos(llama.ctx)); - const bool ignore_eos = eos_bias != sparams.logit_bias.end() && - eos_bias->second < 0.0f && std::isinf(eos_bias->second); - - return json{ - {"n_ctx", llama.n_ctx}, - {"model", llama.params.model_alias}, - {"seed", llama.params.seed}, - {"temp", sparams.temp}, - {"top_k", sparams.top_k}, - {"top_p", sparams.top_p}, - {"tfs_z", sparams.tfs_z}, - {"typical_p", sparams.typical_p}, - {"repeat_last_n", sparams.penalty_last_n}, - {"repeat_penalty", sparams.penalty_repeat}, - {"frequency_penalty", sparams.penalty_freq}, - {"presence_penalty", sparams.penalty_present}, - {"mirostat", sparams.mirostat}, - {"mirostat_tau", sparams.mirostat_tau}, - {"mirostat_eta", sparams.mirostat_eta}, - {"penalize_nl", sparams.penalize_nl}, - {"stop", llama.params.antiprompt}, - {"n_predict", llama.params.n_predict}, - {"n_keep", llama.params.n_keep}, - {"ignore_eos", ignore_eos}, - {"stream", llama.stream}, - {"logit_bias", sparams.logit_bias}, - {"n_probs", sparams.n_probs}, - {"grammar", llama.params.sparams.grammar}, - }; -} - -static json format_embedding_response(llama_server_context &llama) -{ - return json{ - {"embedding", llama.getEmbedding()}, - }; -} - -static json format_timings(llama_server_context &llama) -{ - const auto timings = llama_get_timings(llama.ctx); - - return json{ - {"prompt_n", timings.n_p_eval}, - {"prompt_ms", timings.t_p_eval_ms}, - {"prompt_per_token_ms", timings.t_p_eval_ms / timings.n_p_eval}, - {"prompt_per_second", 1e3 / timings.t_p_eval_ms * timings.n_p_eval}, - - {"predicted_n", timings.n_eval}, - {"predicted_ms", timings.t_eval_ms}, - {"predicted_per_token_ms", timings.t_eval_ms / timings.n_eval}, - {"predicted_per_second", 1e3 / timings.t_eval_ms * timings.n_eval}, - }; -} - -static json format_final_response(llama_server_context &llama, const std::string &content, const std::vector &probs) -{ - - json res = json{ - {"content", content}, - {"stop", true}, - {"model", llama.params.model_alias}, - {"tokens_predicted", llama.num_tokens_predicted}, - {"tokens_evaluated", llama.num_prompt_tokens}, - {"generation_settings", format_generation_settings(llama)}, - {"prompt", llama.prompt}, - {"truncated", llama.truncated}, - {"stopped_eos", llama.stopped_eos}, - {"stopped_word", llama.stopped_word}, - {"stopped_limit", llama.stopped_limit}, - {"stopping_word", llama.stopping_word}, - {"tokens_cached", llama.n_past}, - {"timings", format_timings(llama)}, - }; - - if (llama.params.sparams.n_probs > 0) - { - res["completion_probabilities"] = probs_vector_to_json(llama.ctx, probs); - } - - return res; -} - static json format_partial_response( - llama_server_context &llama, const std::string &content, const std::vector &probs + llama_server_context &llama, llama_client_slot *slot, const std::string &content, const std::vector &probs ) { - json res = json{ - {"content", content}, - {"stop", false}, + json res = json + { + {"content", content }, + {"stop", false}, + {"slot_id", slot->id }, + {"multimodal", llama.multimodal } }; - if (llama.params.sparams.n_probs > 0) + if (slot->sparams.n_probs > 0) { res["completion_probabilities"] = probs_vector_to_json(llama.ctx, probs); } @@ -1115,120 +2154,8 @@ static json format_detokenized_response(std::string content) {"content", content}}; } -template -static T json_value(const json &body, const std::string &key, const T &default_value) -{ - // Fallback null to default value - return body.contains(key) && !body.at(key).is_null() - ? body.value(key, default_value) - : default_value; -} -static void parse_options_completion(const json &body, llama_server_context &llama) -{ - gpt_params default_params; - const auto & default_sparams = default_params.sparams; - - auto & params = llama.params; - auto & sparams = llama.params.sparams; - - llama.stream = json_value(body, "stream", false); - params.n_predict = json_value(body, "n_predict", default_params.n_predict); - sparams.top_k = json_value(body, "top_k", default_sparams.top_k); - sparams.top_p = json_value(body, "top_p", default_sparams.top_p); - sparams.tfs_z = json_value(body, "tfs_z", default_sparams.tfs_z); - sparams.typical_p = json_value(body, "typical_p", default_sparams.typical_p); - sparams.temp = json_value(body, "temperature", default_sparams.temp); - sparams.penalty_last_n = json_value(body, "repeat_last_n", default_sparams.penalty_last_n); - sparams.penalty_repeat = json_value(body, "repeat_penalty", default_sparams.penalty_repeat); - sparams.penalty_freq = json_value(body, "frequency_penalty", default_sparams.penalty_freq); - sparams.penalty_present = json_value(body, "presence_penalty", default_sparams.penalty_present); - sparams.mirostat = json_value(body, "mirostat", default_sparams.mirostat); - sparams.mirostat_tau = json_value(body, "mirostat_tau", default_sparams.mirostat_tau); - sparams.mirostat_eta = json_value(body, "mirostat_eta", default_sparams.mirostat_eta); - sparams.penalize_nl = json_value(body, "penalize_nl", default_sparams.penalize_nl); - params.n_keep = json_value(body, "n_keep", default_params.n_keep); - params.seed = json_value(body, "seed", default_params.seed); - sparams.grammar = json_value(body, "grammar", default_sparams.grammar); - sparams.n_probs = json_value(body, "n_probs", default_sparams.n_probs); - - if (body.count("prompt") != 0) - { - llama.prompt = body["prompt"]; - } - else - { - llama.prompt = ""; - } - - sparams.logit_bias.clear(); - if (json_value(body, "ignore_eos", false)) - { - sparams.logit_bias[llama_token_eos(llama.ctx)] = -INFINITY; - } - - const auto &logit_bias = body.find("logit_bias"); - if (logit_bias != body.end() && logit_bias->is_array()) - { - const int n_vocab = llama_n_vocab(llama.model); - for (const auto &el : *logit_bias) - { - if (el.is_array() && el.size() == 2 && el[0].is_number_integer()) - { - llama_token tok = el[0].get(); - if (tok >= 0 && tok < n_vocab) - { - if (el[1].is_number()) - { - sparams.logit_bias[tok] = el[1].get(); - } - else if (el[1].is_boolean() && !el[1].get()) - { - sparams.logit_bias[tok] = -INFINITY; - } - } - } - } - } - - llama.params.antiprompt.clear(); - const auto &stop = body.find("stop"); - if (stop != body.end() && stop->is_array()) - { - for (const auto &word : *stop) - { - if (!word.empty()) - { - llama.params.antiprompt.push_back(word); - } - } - } - - LOG_VERBOSE("completion parameters parsed", format_generation_settings(llama)); -} - -static void parse_options_infill(const json &body, llama_server_context &llama) -{ - if (body.count("input_prefix") != 0) - { - llama.params.input_prefix = body["input_prefix"]; - } - else - { - llama.params.input_prefix = ""; - } - if (body.count("input_suffix") != 0) - { - llama.params.input_suffix = body["input_suffix"]; - } - else - { - llama.params.input_suffix = ""; - } - parse_options_completion(body, llama); -} - -static void log_server_request(const Request &req, const Response &res) +static void log_server_request(const httplib::Request &req, const httplib::Response &res) { LOG_INFO("request", { {"remote_addr", req.remote_addr}, @@ -1245,60 +2172,26 @@ static void log_server_request(const Request &req, const Response &res) }); } -static bool is_at_eob(llama_server_context &server_context, const llama_token *tokens, const size_t n_tokens) { - return n_tokens && tokens[n_tokens-1] == llama_token_eos(server_context.ctx); -} - -// Function matching type llama_beam_search_callback_fn_t. -// Custom callback example is called each time the beams lengths increase: -// * Show progress by printing ',' following by number of convergent beam tokens if any. -// * When all beams converge to a common prefix, they are made available in beams_state.beams[0]. -// This is also called when the stop condition is met. -// Collect tokens into std::vector response which is pointed to by callback_data. -static void beam_search_callback(void *callback_data, llama_beams_state beams_state) { - auto & llama = *static_cast(callback_data); - // Mark beams as EOS as needed. - for (size_t i = 0 ; i < beams_state.n_beams ; ++i) { - llama_beam_view& beam_view = beams_state.beam_views[i]; - if (!beam_view.eob && is_at_eob(llama, beam_view.tokens, beam_view.n_tokens)) { - beam_view.eob = true; - } - } - printf(","); // Show progress - if (const size_t n = beams_state.common_prefix_length) { - llama.generated_token_probs.resize(llama.generated_token_probs.size() + n); - assert(0u < beams_state.n_beams); - const llama_token * tokens = beams_state.beam_views[0].tokens; - const auto map = [](llama_token tok) { return completion_token_output{{},tok}; }; - std::transform(tokens, tokens + n, llama.generated_token_probs.end() - n, map); - printf("%zu", n); - } - fflush(stdout); -#if 0 // DEBUG: print current beams for this iteration - std::cout << "\n\nCurrent beams:\n"; - for (size_t i=0 ; i < beams_state.n_beams ; ++i) { - std::cout << "beams["<generated_token_probs; auto translator = token_translator{llama.ctx}; auto add_strlen = [=](size_t sum, const completion_token_output & cto) { return sum + translator(cto).size(); }; const size_t len = std::accumulate(gtps.begin(), gtps.end(), size_t(0), add_strlen); - if (llama.generated_text.capacity() < llama.generated_text.size() + len) { - llama.generated_text.reserve(llama.generated_text.size() + len); + if (slot->generated_text.capacity() < slot->generated_text.size() + len) + { + slot->generated_text.reserve(slot->generated_text.size() + len); } - for (const completion_token_output & cto : gtps) { - llama.generated_text += translator(cto); + for (const completion_token_output & cto : gtps) + { + slot->generated_text += translator(cto); } } @@ -1311,7 +2204,7 @@ int main(int argc, char **argv) // struct that contains llama context and inference llama_server_context llama; - server_params_parse(argc, argv, sparams, params); + server_params_parse(argc, argv, sparams, params, llama); if (params.model_alias == "unknown") { @@ -1322,6 +2215,7 @@ int main(int argc, char **argv) LOG_INFO("build info", {{"build", BUILD_NUMBER}, {"commit", BUILD_COMMIT}}); + LOG_INFO("system info", { {"n_threads", params.n_threads}, {"n_threads_batch", params.n_threads_batch}, @@ -1330,405 +2224,261 @@ int main(int argc, char **argv) }); // load the model - if (!llama.loadModel(params)) + if (!llama.load_model(params)) { return 1; } - Server svr; + llama.initialize(); + + httplib::Server svr; svr.set_default_headers({{"Server", "llama.cpp"}, {"Access-Control-Allow-Origin", "*"}, {"Access-Control-Allow-Headers", "content-type"}}); // this is only called if no index.html is found in the public --path - svr.Get("/", [](const Request &, Response &res) + svr.Get("/", [](const httplib::Request &, httplib::Response &res) { - res.set_content(reinterpret_cast(&index_html), index_html_len, "text/html"); - return false; }); + res.set_content(reinterpret_cast(&index_html), index_html_len, "text/html"); + return false; + }); // this is only called if no index.js is found in the public --path - svr.Get("/index.js", [](const Request &, Response &res) + svr.Get("/index.js", [](const httplib::Request &, httplib::Response &res) { - res.set_content(reinterpret_cast(&index_js), index_js_len, "text/javascript"); - return false; }); + res.set_content(reinterpret_cast(&index_js), index_js_len, "text/javascript"); + return false; + }); // this is only called if no index.html is found in the public --path - svr.Get("/completion.js", [](const Request &, Response &res) + svr.Get("/completion.js", [](const httplib::Request &, httplib::Response &res) { - res.set_content(reinterpret_cast(&completion_js), completion_js_len, "application/javascript"); - return false; }); + res.set_content(reinterpret_cast(&completion_js), completion_js_len, "application/javascript"); + return false; + }); // this is only called if no index.html is found in the public --path - svr.Get("/json-schema-to-grammar.mjs", [](const Request &, Response &res) + svr.Get("/json-schema-to-grammar.mjs", [](const httplib::Request &, httplib::Response &res) { - res.set_content(reinterpret_cast(&json_schema_to_grammar_mjs), json_schema_to_grammar_mjs_len, "application/javascript"); - return false; }); + res.set_content(reinterpret_cast(&json_schema_to_grammar_mjs), json_schema_to_grammar_mjs_len, "application/javascript"); + return false; + }); - svr.Post("/completion", [&llama](const Request &req, Response &res) - { - auto lock = llama.lock(); + svr.Get("/props", [&llama](const httplib::Request & /*req*/, httplib::Response &res) + { + res.set_header("Access-Control-Allow-Origin", "*"); + json data = { + { "user_name", llama.name_user.c_str() }, + { "assistant_name", llama.name_assistant.c_str() } + }; + res.set_content(data.dump(), "application/json"); + }); - llama.rewind(); - - llama_reset_timings(llama.ctx); - parse_options_completion(json::parse(req.body), llama); - - llama.initSampling(); - llama.loadPrompt(); - llama.beginCompletion(); - - if (!llama.stream) { - if (llama.params.n_beams) { - // Fill llama.generated_token_probs vector with final beam. - llama_beam_search(llama.ctx, beam_search_callback, &llama, llama.params.n_beams, - llama.n_past, llama.n_remain); - // Translate llama.generated_token_probs to llama.generated_text. - append_to_generated_text_from_generated_token_probs(llama); - } else { - size_t stop_pos = std::string::npos; - - while (llama.has_next_token) { - const completion_token_output token_with_probs = llama.doCompletion(); - const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_piece(llama.ctx, token_with_probs.tok); - - stop_pos = llama.findStoppingStrings(llama.generated_text, - token_text.size(), STOP_FULL); - } - - if (stop_pos == std::string::npos) { - stop_pos = llama.findStoppingStrings(llama.generated_text, 0, STOP_PARTIAL); - } - if (stop_pos != std::string::npos) { - llama.generated_text.erase(llama.generated_text.begin() + stop_pos, - llama.generated_text.end()); - } - } - - auto probs = llama.generated_token_probs; - if (llama.params.sparams.n_probs > 0 && llama.stopped_word) { - const std::vector stop_word_toks = llama_tokenize(llama.ctx, llama.stopping_word, false); - probs = std::vector(llama.generated_token_probs.begin(), llama.generated_token_probs.end() - stop_word_toks.size()); - } - - const json data = format_final_response(llama, llama.generated_text, probs); - - llama_print_timings(llama.ctx); - - res.set_content(data.dump(-1, ' ', false, json::error_handler_t::replace), - "application/json"); - } else { - const auto chunked_content_provider = [&](size_t, DataSink & sink) { - size_t sent_count = 0; - size_t sent_token_probs_index = 0; - - while (llama.has_next_token) { - const completion_token_output token_with_probs = llama.doCompletion(); - if (token_with_probs.tok == -1 || llama.multibyte_pending > 0) { - continue; + svr.Post("/completion", [&llama](const httplib::Request &req, httplib::Response &res) + { + json data = json::parse(req.body); + const int task_id = llama.request_completion(data, false, false); + if (!json_value(data, "stream", false)) { + std::string completion_text; + task_result result = llama.next_result(task_id); + if (!result.error && result.stop) { + res.set_content(result.result_json.dump(-1, ' ', false, json::error_handler_t::replace), "application/json"); } - const std::string token_text = llama_token_to_piece(llama.ctx, token_with_probs.tok); - - size_t pos = std::min(sent_count, llama.generated_text.size()); - - const std::string str_test = llama.generated_text.substr(pos); - bool is_stop_full = false; - size_t stop_pos = - llama.findStoppingStrings(str_test, token_text.size(), STOP_FULL); - if (stop_pos != std::string::npos) { - is_stop_full = true; - llama.generated_text.erase( - llama.generated_text.begin() + pos + stop_pos, - llama.generated_text.end()); - pos = std::min(sent_count, llama.generated_text.size()); - } else { - is_stop_full = false; - stop_pos = llama.findStoppingStrings(str_test, token_text.size(), - STOP_PARTIAL); + else + { + res.status = 404; + res.set_content(result.result_json["content"], "text/plain"); + return; } - - if ( - stop_pos == std::string::npos || - // Send rest of the text if we are at the end of the generation - (!llama.has_next_token && !is_stop_full && stop_pos > 0) - ) { - const std::string to_send = llama.generated_text.substr(pos, std::string::npos); - - sent_count += to_send.size(); - - std::vector probs_output = {}; - - if (llama.params.sparams.n_probs > 0) { - const std::vector to_send_toks = llama_tokenize(llama.ctx, to_send, false); - size_t probs_pos = std::min(sent_token_probs_index, llama.generated_token_probs.size()); - size_t probs_stop_pos = std::min(sent_token_probs_index + to_send_toks.size(), llama.generated_token_probs.size()); - if (probs_pos < probs_stop_pos) { - probs_output = std::vector(llama.generated_token_probs.begin() + probs_pos, llama.generated_token_probs.begin() + probs_stop_pos); - } - sent_token_probs_index = probs_stop_pos; - } - - const json data = format_partial_response(llama, to_send, probs_output); - - const std::string str = - "data: " + - data.dump(-1, ' ', false, json::error_handler_t::replace) + - "\n\n"; - - LOG_VERBOSE("data stream", { - { "to_send", str } - }); - - if (!sink.write(str.data(), str.size())) { - LOG_VERBOSE("stream closed", {}); - llama_print_timings(llama.ctx); - return false; - } - } - - if (!llama.has_next_token) { - // Generation is done, send extra information. - const json data = format_final_response( - llama, - "", - std::vector(llama.generated_token_probs.begin(), llama.generated_token_probs.begin() + sent_token_probs_index) - ); - - const std::string str = - "data: " + - data.dump(-1, ' ', false, json::error_handler_t::replace) + - "\n\n"; - - LOG_VERBOSE("data stream", { - { "to_send", str } - }); - - if (!sink.write(str.data(), str.size())) { - LOG_VERBOSE("stream closed", {}); - llama_print_timings(llama.ctx); - return false; - } - } - } - - llama_print_timings(llama.ctx); - sink.done(); - return true; - }; - const auto on_complete = [&](bool) { - llama.mutex.unlock(); - }; - lock.release(); - res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete); - } }); - - svr.Post("/infill", [&llama](const Request &req, Response &res) - { - auto lock = llama.lock(); - - llama.rewind(); - - llama_reset_timings(llama.ctx); - parse_options_infill(json::parse(req.body), llama); - - llama.initSampling(); - llama.loadInfill(); - llama.beginCompletion(); - const auto chunked_content_provider = [&](size_t, DataSink & sink) { - size_t sent_count = 0; - size_t sent_token_probs_index = 0; - - while (llama.has_next_token) { - const completion_token_output token_with_probs = llama.doCompletion(); - if (token_with_probs.tok == -1 || llama.multibyte_pending > 0) { - continue; - } - const std::string token_text = llama_token_to_piece(llama.ctx, token_with_probs.tok); - - size_t pos = std::min(sent_count, llama.generated_text.size()); - - const std::string str_test = llama.generated_text.substr(pos); - bool is_stop_full = false; - size_t stop_pos = - llama.findStoppingStrings(str_test, token_text.size(), STOP_FULL); - if (stop_pos != std::string::npos) { - is_stop_full = true; - llama.generated_text.erase( - llama.generated_text.begin() + pos + stop_pos, - llama.generated_text.end()); - pos = std::min(sent_count, llama.generated_text.size()); } else { - is_stop_full = false; - stop_pos = llama.findStoppingStrings(str_test, token_text.size(), - STOP_PARTIAL); - } - - if ( - stop_pos == std::string::npos || - // Send rest of the text if we are at the end of the generation - (!llama.has_next_token && !is_stop_full && stop_pos > 0) - ) { - const std::string to_send = llama.generated_text.substr(pos, std::string::npos); - - sent_count += to_send.size(); - - std::vector probs_output = {}; - - if (llama.params.sparams.n_probs > 0) { - const std::vector to_send_toks = llama_tokenize(llama.ctx, to_send, false); - size_t probs_pos = std::min(sent_token_probs_index, llama.generated_token_probs.size()); - size_t probs_stop_pos = std::min(sent_token_probs_index + to_send_toks.size(), llama.generated_token_probs.size()); - if (probs_pos < probs_stop_pos) { - probs_output = std::vector(llama.generated_token_probs.begin() + probs_pos, llama.generated_token_probs.begin() + probs_stop_pos); + const auto chunked_content_provider = [task_id, &llama](size_t, httplib::DataSink & sink) + { + while (true) + { + task_result result = llama.next_result(task_id); + if (!result.error) { + const std::string str = + "data: " + + result.result_json.dump(-1, ' ', false, json::error_handler_t::replace) + + "\n\n"; + LOG_VERBOSE("data stream", { + { "to_send", str } + }); + if (!sink.write(str.c_str(), str.size())) + { + return false; + } + if (result.stop) { + break; + } + } else { + break; + } } - sent_token_probs_index = probs_stop_pos; - } + sink.done(); + return true; + }; - const json data = format_partial_response(llama, to_send, probs_output); + auto on_complete = [task_id, &llama] (bool) + { + // cancel + llama.request_cancel(task_id); + }; - const std::string str = - "data: " + - data.dump(-1, ' ', false, json::error_handler_t::replace) + - "\n\n"; - - LOG_VERBOSE("data stream", { - { "to_send", str } - }); - - if (!sink.write(str.data(), str.size())) { - LOG_VERBOSE("stream closed", {}); - llama_print_timings(llama.ctx); - return false; - } + res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete); } + }); - if (!llama.has_next_token) { - // Generation is done, send extra information. - const json data = format_final_response( - llama, - "", - std::vector(llama.generated_token_probs.begin(), llama.generated_token_probs.begin() + sent_token_probs_index) - ); - - const std::string str = - "data: " + - data.dump(-1, ' ', false, json::error_handler_t::replace) + - "\n\n"; - - LOG_VERBOSE("data stream", { - { "to_send", str } - }); - - if (!sink.write(str.data(), str.size())) { - LOG_VERBOSE("stream closed", {}); - llama_print_timings(llama.ctx); - return false; - } - } - } - - llama_print_timings(llama.ctx); - sink.done(); - return true; - }; - const auto on_complete = [&](bool) { - llama.mutex.unlock(); - }; - lock.release(); - res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete); - }); - - svr.Get("/model.json", [&llama](const Request &, Response &res) + svr.Post("/infill", [&llama](const httplib::Request &req, httplib::Response &res) { - const json data = format_generation_settings(llama); - return res.set_content(data.dump(), "application/json"); }); + json data = json::parse(req.body); + const int task_id = llama.request_completion(data, true, false); + if (!json_value(data, "stream", false)) { + std::string completion_text; + task_result result = llama.next_result(task_id); + if (!result.error && result.stop) + { + res.set_content(result.result_json.dump(-1, ' ', false, json::error_handler_t::replace), "application/json"); + } + else + { + res.status = 404; + res.set_content(result.result_json["content"], "text/plain"); + return; + } + } else { + const auto chunked_content_provider = [task_id, &llama](size_t, httplib::DataSink & sink) { + while (true) + { + task_result result = llama.next_result(task_id); + if (!result.error) { + const std::string str = + "data: " + + result.result_json.dump(-1, ' ', false, json::error_handler_t::replace) + + "\n\n"; + LOG_VERBOSE("data stream", { + { "to_send", str } + }); + if (!sink.write(str.c_str(), str.size())) + { + return false; + } + if (result.stop) + { + break; + } + } + else + { + break; + } + } - svr.Options(R"(/.*)", [](const Request &, Response &res) + sink.done(); + + return true; + }; + + auto on_complete = [task_id, &llama] (bool) + { + // cancel + llama.request_cancel(task_id); + }; + + res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete); + } + }); + + svr.Get("/model.json", [&llama](const httplib::Request &, httplib::Response &res) + { + const json data = llama.get_model_props(); + return res.set_content(data.dump(), "application/json"); + }); + + svr.Options(R"(/.*)", [](const httplib::Request &, httplib::Response &res) { return res.set_content("", "application/json"); }); - svr.Post("/tokenize", [&llama](const Request &req, Response &res) - { - auto lock = llama.lock(); + svr.Post("/tokenize", [&llama](const httplib::Request &req, httplib::Response &res) + { + const json body = json::parse(req.body); + std::vector tokens; + if (body.count("content") != 0) + { + tokens = llama.tokenize(body["content"], false); + } + const json data = format_tokenizer_response(tokens); + return res.set_content(data.dump(), "application/json"); + }); - const json body = json::parse(req.body); - std::vector tokens; - if (body.count("content") != 0) - { - tokens = llama.tokenize(body["content"], false); - } - const json data = format_tokenizer_response(tokens); - return res.set_content(data.dump(), "application/json"); }); + svr.Post("/detokenize", [&llama](const httplib::Request &req, httplib::Response &res) + { + const json body = json::parse(req.body); + std::string content; + if (body.count("tokens") != 0) + { + const std::vector tokens = body["tokens"]; + content = tokens_to_str(llama.ctx, tokens.cbegin(), tokens.cend()); + } - svr.Post("/detokenize", [&llama](const Request &req, Response &res) - { - auto lock = llama.lock(); + const json data = format_detokenized_response(content); + return res.set_content(data.dump(), "application/json"); + }); - const json body = json::parse(req.body); - std::string content; - if (body.count("tokens") != 0) - { - const std::vector tokens = body["tokens"]; - content = tokens_to_str(llama.ctx, tokens.cbegin(), tokens.cend()); - } - - const json data = format_detokenized_response(content); - return res.set_content(data.dump(), "application/json"); }); - - svr.Post("/embedding", [&llama](const Request &req, Response &res) - { - auto lock = llama.lock(); - - const json body = json::parse(req.body); - - llama.rewind(); - - llama_reset_timings(llama.ctx); - - if (body.count("content") != 0) - { - llama.prompt = body["content"]; - } - else - { - llama.prompt = ""; - } - llama.params.n_predict = 0; - - llama.initSampling(); - llama.loadPrompt(); - llama.beginCompletion(); - llama.doCompletion(); - - const json data = format_embedding_response(llama); - return res.set_content(data.dump(), "application/json"); }); + svr.Post("/embedding", [&llama](const httplib::Request &req, httplib::Response &res) + { + const json body = json::parse(req.body); + json prompt; + if (body.count("content") != 0) + { + prompt = body["content"]; + } + else + { + prompt = ""; + } + const int task_id = llama.request_completion({ {"prompt", prompt}, { "n_predict", 0} }, false, true); + task_result result = llama.next_result(task_id); + return res.set_content(result.result_json.dump(), "application/json"); + }); svr.set_logger(log_server_request); - svr.set_exception_handler([](const Request &, Response &res, std::exception_ptr ep) - { - const char fmt[] = "500 Internal Server Error\n%s"; - char buf[BUFSIZ]; - try { - std::rethrow_exception(std::move(ep)); - } catch (std::exception & e) { - snprintf(buf, sizeof(buf), fmt, e.what()); - } catch (...) { - snprintf(buf, sizeof(buf), fmt, "Unknown Exception"); - } - res.set_content(buf, "text/plain"); - res.status = 500; }); + svr.set_exception_handler([](const httplib::Request &, httplib::Response &res, std::exception_ptr ep) + { + const char fmt[] = "500 Internal Server Error\n%s"; + char buf[BUFSIZ]; + try + { + std::rethrow_exception(std::move(ep)); + } + catch (std::exception &e) + { + snprintf(buf, sizeof(buf), fmt, e.what()); + } + catch (...) + { + snprintf(buf, sizeof(buf), fmt, "Unknown Exception"); + } + res.set_content(buf, "text/plain"); + res.status = 500; + }); - svr.set_error_handler([](const Request &, Response &res) - { - if (res.status == 400) { - res.set_content("Invalid request", "text/plain"); - } else if (res.status != 500) { - res.set_content("File Not Found", "text/plain"); - res.status = 404; - } }); + svr.set_error_handler([](const httplib::Request &, httplib::Response &res) + { + if (res.status == 400) + { + res.set_content("Invalid request", "text/plain"); + } + else if (res.status != 500) + { + res.set_content("File Not Found", "text/plain"); + res.status = 404; + } + }); // set timeouts and change hostname and port - svr.set_read_timeout(sparams.read_timeout); + svr.set_read_timeout (sparams.read_timeout); svr.set_write_timeout(sparams.write_timeout); if (!svr.bind_to_port(sparams.hostname, sparams.port)) @@ -1741,20 +2491,38 @@ int main(int argc, char **argv) svr.set_base_dir(sparams.public_path); // to make it ctrl+clickable: - printf("\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port); + LOG_TEE("\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port); LOG_INFO("HTTP server listening", { {"hostname", sparams.hostname}, {"port", sparams.port}, }); - if (!svr.listen_after_bind()) + // run the HTTP server in a thread - see comment below + std::thread t([&]() + { + if (!svr.listen_after_bind()) + { + return 1; + } + + return 0; + }); + + // GG: if I put the main loop inside a thread, it crashes on the first request when build in Debug!? + // "Bus error: 10" - this is on macOS, it does not crash on Linux + //std::thread t2([&]() { - return 1; + bool running = true; + while (running) + { + running = llama.update_slots(); + } } + //); + + t.join(); - llama_sampling_free(llama.ctx_sampling); llama_backend_free(); - return 0; } diff --git a/examples/simple/simple.cpp b/examples/simple/simple.cpp index 55385f566..374aef6f1 100644 --- a/examples/simple/simple.cpp +++ b/examples/simple/simple.cpp @@ -95,13 +95,8 @@ int main(int argc, char ** argv) { llama_batch batch = llama_batch_init(512, 0, 1); // evaluate the initial prompt - batch.n_tokens = tokens_list.size(); - - for (int32_t i = 0; i < batch.n_tokens; i++) { - batch.token[i] = tokens_list[i]; - batch.pos[i] = i; - batch.seq_id[i] = 0; - batch.logits[i] = false; + for (size_t i = 0; i < tokens_list.size(); i++) { + llama_batch_add(batch, tokens_list[i], i, { 0 }, false); } // llama_decode will output logits only for the last token of the prompt @@ -138,7 +133,7 @@ int main(int argc, char ** argv) { const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p); // is it an end of stream? - if (new_token_id == llama_token_eos(ctx) || n_cur == n_len) { + if (new_token_id == llama_token_eos(model) || n_cur == n_len) { LOG_TEE("\n"); break; @@ -148,15 +143,10 @@ int main(int argc, char ** argv) { fflush(stdout); // prepare the next batch - batch.n_tokens = 0; + llama_batch_clear(batch); // push this new token for next evaluation - batch.token [batch.n_tokens] = new_token_id; - batch.pos [batch.n_tokens] = n_cur; - batch.seq_id[batch.n_tokens] = 0; - batch.logits[batch.n_tokens] = true; - - batch.n_tokens += 1; + llama_batch_add(batch, new_token_id, n_cur, { 0 }, true); n_decode += 1; } diff --git a/examples/speculative/speculative.cpp b/examples/speculative/speculative.cpp index 894321ce9..323c74652 100644 --- a/examples/speculative/speculative.cpp +++ b/examples/speculative/speculative.cpp @@ -8,6 +8,9 @@ #include #include +#define SPEC_VOCAB_MAX_SIZE_DIFFERENCE 100 +#define SPEC_VOCAB_CHECK_START_TOKEN_ID 5 + struct seq_draft { bool active = false; bool drafting = false; @@ -64,6 +67,33 @@ int main(int argc, char ** argv) { params.n_gpu_layers = params.n_gpu_layers_draft; std::tie(model_dft, ctx_dft) = llama_init_from_gpt_params(params); + { + const int n_vocab_tgt = llama_n_vocab(model_tgt); + const int n_vocab_dft = llama_n_vocab(model_dft); + const int vocab_diff = n_vocab_tgt > n_vocab_dft + ? n_vocab_tgt - n_vocab_dft + : n_vocab_dft - n_vocab_tgt; + + if (vocab_diff > SPEC_VOCAB_MAX_SIZE_DIFFERENCE) { + fprintf(stderr, "%s: error: draft model vocab must closely match target model to use speculation but ", __func__); + fprintf(stderr, "target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n", + n_vocab_tgt, llama_n_vocab(model_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE); + return 1; + } + + for (int i = SPEC_VOCAB_CHECK_START_TOKEN_ID; i < std::min(n_vocab_tgt, n_vocab_dft); ++i) { + const char * token_text_tgt = llama_token_get_text(model_tgt, i); + const char * token_text_dft = llama_token_get_text(model_dft, i); + if (std::strcmp(token_text_tgt, token_text_dft) != 0) { + fprintf(stderr, "%s: error: draft model vocab must match target model to use speculation but ", __func__); + fprintf(stderr, "token %d content differs - target '%s', draft '%s'\n", i, + llama_token_to_piece(ctx_tgt, i).c_str(), + llama_token_to_piece(ctx_dft, i).c_str()); + return 1; + } + } + } + // tokenize the prompt std::vector inp; inp = ::llama_tokenize(ctx_tgt, params.prompt, true); @@ -118,7 +148,7 @@ int main(int argc, char ** argv) { std::vector drafts(n_seq_dft); params.sparams.grammar.clear(); // the draft samplers will copy the target sampler's grammar - params.sparams.temp = std::max(0.01f, params.sparams.temp); + params.sparams.temp = -1.0f; // force greedy sampling with probs for the draft model for (int s = 0; s < n_seq_dft; ++s) { drafts[s].ctx_sampling = llama_sampling_init(params.sparams); @@ -163,7 +193,7 @@ int main(int argc, char ** argv) { printf("%s", token_str.c_str()); fflush(stdout); - if (id == llama_token_eos(ctx_tgt)) { + if (id == llama_token_eos(model_tgt)) { has_eos = true; } @@ -227,6 +257,7 @@ int main(int argc, char ** argv) { llama_batch_add (batch_dft, id, n_past_dft, { 0 }, true); llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1); + // LOG("dft batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_dft, batch_dft).c_str()); llama_decode (ctx_dft, batch_dft); ++n_past_dft; @@ -370,7 +401,7 @@ int main(int argc, char ** argv) { llama_kv_cache_seq_cp(ctx_tgt, 0, s, -1, -1); } - //LOG("target batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_tgt, batch_tgt)); + // LOG("target batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_tgt, batch_tgt).c_str()); llama_decode(ctx_tgt, batch_tgt); ++n_past_tgt; } diff --git a/flake.lock b/flake.lock index a7777d05d..0455f6561 100644 --- a/flake.lock +++ b/flake.lock @@ -5,11 +5,11 @@ "systems": "systems" }, "locked": { - "lastModified": 1692799911, - "narHash": "sha256-3eihraek4qL744EvQXsK1Ha6C3CR7nnT8X2qWap4RNk=", + "lastModified": 1694529238, + "narHash": "sha256-zsNZZGTGnMOf9YpHKJqMSsa0dXbfmxeoJ7xHlrt+xmY=", "owner": "numtide", "repo": "flake-utils", - "rev": "f9e7cf818399d17d347f847525c5a5a8032e4e44", + "rev": "ff7b65b44d01cf9ba6a71320833626af21126384", "type": "github" }, "original": { @@ -20,11 +20,11 @@ }, "nixpkgs": { "locked": { - "lastModified": 1692913444, - "narHash": "sha256-1SvMQm2DwofNxXVtNWWtIcTh7GctEVrS/Xel/mdc6iY=", + "lastModified": 1698318101, + "narHash": "sha256-gUihHt3yPD7bVqg+k/UVHgngyaJ3DMEBchbymBMvK1E=", "owner": "NixOS", "repo": "nixpkgs", - "rev": "18324978d632ffc55ef1d928e81630c620f4f447", + "rev": "63678e9f3d3afecfeafa0acead6239cdb447574c", "type": "github" }, "original": { diff --git a/flake.nix b/flake.nix index cfc4776a4..4cf28d5c1 100644 --- a/flake.nix +++ b/flake.nix @@ -11,8 +11,7 @@ meta.mainProgram = "llama"; inherit (pkgs.stdenv) isAarch32 isAarch64 isDarwin; buildInputs = with pkgs; [ openmpi ]; - osSpecific = with pkgs; buildInputs ++ - ( + osSpecific = with pkgs; buildInputs ++ ( if isAarch64 && isDarwin then with pkgs.darwin.apple_sdk_11_0.frameworks; [ Accelerate @@ -51,6 +50,9 @@ }; llama-python = pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece ]); + # TODO(Green-Sky): find a better way to opt-into the heavy ml python runtime + llama-python-extra = + pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece torchWithoutCuda transformers ]); postPatch = '' substituteInPlace ./ggml-metal.m \ --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";" @@ -93,12 +95,15 @@ }; packages.rocm = pkgs.stdenv.mkDerivation { inherit name src meta postPatch nativeBuildInputs postInstall; - buildInputs = with pkgs; buildInputs ++ [ hip hipblas rocblas ]; + buildInputs = with pkgs.rocmPackages; buildInputs ++ [ clr hipblas rocblas ]; cmakeFlags = cmakeFlags ++ [ "-DLLAMA_HIPBLAS=1" "-DCMAKE_C_COMPILER=hipcc" "-DCMAKE_CXX_COMPILER=hipcc" - "-DCMAKE_POSITION_INDEPENDENT_CODE=ON" + # Build all targets supported by rocBLAS. When updating search for TARGET_LIST_ROCM + # in github.com/ROCmSoftwarePlatform/rocBLAS/blob/develop/CMakeLists.txt + # and select the line that matches the current nixpkgs version of rocBLAS. + "-DAMDGPU_TARGETS=gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102" ]; }; apps.llama-server = { @@ -126,5 +131,9 @@ buildInputs = [ llama-python ]; packages = nativeBuildInputs ++ osSpecific; }; + devShells.extra = pkgs.mkShell { + buildInputs = [ llama-python-extra ]; + packages = nativeBuildInputs ++ osSpecific; + }; }); } diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 654d3632f..4e6e7cd94 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -29,6 +29,8 @@ #define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width) #define cublasCreate hipblasCreate #define cublasGemmEx hipblasGemmEx +#define cublasGemmBatchedEx hipblasGemmBatchedEx +#define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx #define cublasHandle_t hipblasHandle_t #define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS #define cublasSetStream hipblasSetStream @@ -85,6 +87,24 @@ #define CC_OFFSET_AMD 1000000 #define CC_RDNA2 (CC_OFFSET_AMD + 1030) +// define this if you want to always fallback to MMQ kernels and not use cuBLAS for matrix multiplication +// on modern hardware, using cuBLAS is recommended as it utilizes F16 tensor cores which are very performant +// for large computational tasks. the drawback is that this requires some extra amount of VRAM: +// - 7B quantum model: +100-200 MB +// - 13B quantum model: +200-400 MB +// +//#define GGML_CUDA_FORCE_MMQ + +// TODO: improve this to be correct for more hardware +// for example, currently fails for GeForce GTX 1660 which is TURING arch (> VOLTA) but does not have tensor cores +// probably other such cases, and not sure what happens on AMD hardware +#if !defined(GGML_CUDA_FORCE_MMQ) +#define CUDA_USE_TENSOR_CORES +#endif + +// max batch size to use MMQ kernels when tensor cores are available +#define MMQ_MAX_BATCH_SIZE 32 + #if defined(GGML_USE_HIPBLAS) #define __CUDA_ARCH__ 1300 @@ -468,7 +488,6 @@ static int g_device_count = -1; static int g_main_device = 0; static int g_compute_capabilities[GGML_CUDA_MAX_DEVICES]; static float g_tensor_split[GGML_CUDA_MAX_DEVICES] = {0}; -static bool g_mul_mat_q = true; static void * g_scratch_buffer = nullptr; static size_t g_scratch_size = 0; // disabled by default @@ -494,6 +513,15 @@ static __global__ void add_f16_f32_f16(const half * x, const float * y, half * d dst[i] = __hadd(x[i], __float2half(y[i])); } +static __global__ void add_f16_f32_f32(const half * x, const float * y, float * dst, const int k) { + const int i = blockDim.x*blockIdx.x + threadIdx.x; + + if (i >= k) { + return; + } + dst[i] = __half2float(x[i]) + y[i]; +} + static __global__ void mul_f32(const float * x, const float * y, float * dst, const int kx, const int ky) { const int i = blockDim.x*blockIdx.x + threadIdx.x; @@ -3552,9 +3580,15 @@ static __device__ __forceinline__ void mul_mat_q( #define MMQ_X_Q4_0_RDNA1 64 #define MMQ_Y_Q4_0_RDNA1 64 #define NWARPS_Q4_0_RDNA1 8 +#if defined(CUDA_USE_TENSOR_CORES) +#define MMQ_X_Q4_0_AMPERE 4 +#define MMQ_Y_Q4_0_AMPERE 32 +#define NWARPS_Q4_0_AMPERE 4 +#else #define MMQ_X_Q4_0_AMPERE 64 #define MMQ_Y_Q4_0_AMPERE 128 #define NWARPS_Q4_0_AMPERE 4 +#endif #define MMQ_X_Q4_0_PASCAL 64 #define MMQ_Y_Q4_0_PASCAL 64 #define NWARPS_Q4_0_PASCAL 8 @@ -3613,9 +3647,15 @@ template static __global__ void #define MMQ_X_Q4_1_RDNA1 64 #define MMQ_Y_Q4_1_RDNA1 64 #define NWARPS_Q4_1_RDNA1 8 +#if defined(CUDA_USE_TENSOR_CORES) +#define MMQ_X_Q4_1_AMPERE 4 +#define MMQ_Y_Q4_1_AMPERE 32 +#define NWARPS_Q4_1_AMPERE 4 +#else #define MMQ_X_Q4_1_AMPERE 64 #define MMQ_Y_Q4_1_AMPERE 128 #define NWARPS_Q4_1_AMPERE 4 +#endif #define MMQ_X_Q4_1_PASCAL 64 #define MMQ_Y_Q4_1_PASCAL 64 #define NWARPS_Q4_1_PASCAL 8 @@ -3676,9 +3716,15 @@ template static __global__ void #define MMQ_X_Q5_0_RDNA1 64 #define MMQ_Y_Q5_0_RDNA1 64 #define NWARPS_Q5_0_RDNA1 8 +#if defined(CUDA_USE_TENSOR_CORES) +#define MMQ_X_Q5_0_AMPERE 4 +#define MMQ_Y_Q5_0_AMPERE 32 +#define NWARPS_Q5_0_AMPERE 4 +#else #define MMQ_X_Q5_0_AMPERE 128 #define MMQ_Y_Q5_0_AMPERE 64 #define NWARPS_Q5_0_AMPERE 4 +#endif #define MMQ_X_Q5_0_PASCAL 64 #define MMQ_Y_Q5_0_PASCAL 64 #define NWARPS_Q5_0_PASCAL 8 @@ -3737,9 +3783,15 @@ template static __global__ void #define MMQ_X_Q5_1_RDNA1 64 #define MMQ_Y_Q5_1_RDNA1 64 #define NWARPS_Q5_1_RDNA1 8 +#if defined(CUDA_USE_TENSOR_CORES) +#define MMQ_X_Q5_1_AMPERE 4 +#define MMQ_Y_Q5_1_AMPERE 32 +#define NWARPS_Q5_1_AMPERE 4 +#else #define MMQ_X_Q5_1_AMPERE 128 #define MMQ_Y_Q5_1_AMPERE 64 #define NWARPS_Q5_1_AMPERE 4 +#endif #define MMQ_X_Q5_1_PASCAL 64 #define MMQ_Y_Q5_1_PASCAL 64 #define NWARPS_Q5_1_PASCAL 8 @@ -3798,9 +3850,15 @@ mul_mat_q5_1( #define MMQ_X_Q8_0_RDNA1 64 #define MMQ_Y_Q8_0_RDNA1 64 #define NWARPS_Q8_0_RDNA1 8 +#if defined(CUDA_USE_TENSOR_CORES) +#define MMQ_X_Q8_0_AMPERE 4 +#define MMQ_Y_Q8_0_AMPERE 32 +#define NWARPS_Q8_0_AMPERE 4 +#else #define MMQ_X_Q8_0_AMPERE 128 #define MMQ_Y_Q8_0_AMPERE 64 #define NWARPS_Q8_0_AMPERE 4 +#endif #define MMQ_X_Q8_0_PASCAL 64 #define MMQ_Y_Q8_0_PASCAL 64 #define NWARPS_Q8_0_PASCAL 8 @@ -3859,9 +3917,15 @@ template static __global__ void #define MMQ_X_Q2_K_RDNA1 128 #define MMQ_Y_Q2_K_RDNA1 32 #define NWARPS_Q2_K_RDNA1 8 +#if defined(CUDA_USE_TENSOR_CORES) +#define MMQ_X_Q2_K_AMPERE 4 +#define MMQ_Y_Q2_K_AMPERE 32 +#define NWARPS_Q2_K_AMPERE 4 +#else #define MMQ_X_Q2_K_AMPERE 64 #define MMQ_Y_Q2_K_AMPERE 128 #define NWARPS_Q2_K_AMPERE 4 +#endif #define MMQ_X_Q2_K_PASCAL 64 #define MMQ_Y_Q2_K_PASCAL 64 #define NWARPS_Q2_K_PASCAL 8 @@ -3920,9 +3984,15 @@ mul_mat_q2_K( #define MMQ_X_Q3_K_RDNA1 32 #define MMQ_Y_Q3_K_RDNA1 128 #define NWARPS_Q3_K_RDNA1 8 +#if defined(CUDA_USE_TENSOR_CORES) +#define MMQ_X_Q3_K_AMPERE 4 +#define MMQ_Y_Q3_K_AMPERE 32 +#define NWARPS_Q3_K_AMPERE 4 +#else #define MMQ_X_Q3_K_AMPERE 128 #define MMQ_Y_Q3_K_AMPERE 128 #define NWARPS_Q3_K_AMPERE 4 +#endif #define MMQ_X_Q3_K_PASCAL 64 #define MMQ_Y_Q3_K_PASCAL 64 #define NWARPS_Q3_K_PASCAL 8 @@ -3983,9 +4053,15 @@ template static __global__ void #define MMQ_X_Q4_K_RDNA1 32 #define MMQ_Y_Q4_K_RDNA1 64 #define NWARPS_Q4_K_RDNA1 8 +#if defined(CUDA_USE_TENSOR_CORES) +#define MMQ_X_Q4_K_AMPERE 4 +#define MMQ_Y_Q4_K_AMPERE 32 +#define NWARPS_Q4_K_AMPERE 4 +#else #define MMQ_X_Q4_K_AMPERE 64 #define MMQ_Y_Q4_K_AMPERE 128 #define NWARPS_Q4_K_AMPERE 4 +#endif #define MMQ_X_Q4_K_PASCAL 64 #define MMQ_Y_Q4_K_PASCAL 64 #define NWARPS_Q4_K_PASCAL 8 @@ -4046,9 +4122,15 @@ template static __global__ void #define MMQ_X_Q5_K_RDNA1 32 #define MMQ_Y_Q5_K_RDNA1 64 #define NWARPS_Q5_K_RDNA1 8 +#if defined(CUDA_USE_TENSOR_CORES) +#define MMQ_X_Q5_K_AMPERE 4 +#define MMQ_Y_Q5_K_AMPERE 32 +#define NWARPS_Q5_K_AMPERE 4 +#else #define MMQ_X_Q5_K_AMPERE 64 #define MMQ_Y_Q5_K_AMPERE 128 #define NWARPS_Q5_K_AMPERE 4 +#endif #define MMQ_X_Q5_K_PASCAL 64 #define MMQ_Y_Q5_K_PASCAL 64 #define NWARPS_Q5_K_PASCAL 8 @@ -4107,9 +4189,15 @@ mul_mat_q5_K( #define MMQ_X_Q6_K_RDNA1 32 #define MMQ_Y_Q6_K_RDNA1 64 #define NWARPS_Q6_K_RDNA1 8 +#if defined(CUDA_USE_TENSOR_CORES) +#define MMQ_X_Q6_K_AMPERE 4 +#define MMQ_Y_Q6_K_AMPERE 32 +#define NWARPS_Q6_K_AMPERE 4 +#else #define MMQ_X_Q6_K_AMPERE 64 #define MMQ_Y_Q6_K_AMPERE 64 #define NWARPS_Q6_K_AMPERE 4 +#endif #define MMQ_X_Q6_K_PASCAL 64 #define MMQ_Y_Q6_K_PASCAL 64 #define NWARPS_Q6_K_PASCAL 8 @@ -4326,13 +4414,13 @@ static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous const half * x = (const half *) vx; - const int row_x = blockDim.y*blockIdx.y + threadIdx.y; - const int channel = blockDim.z*blockIdx.z + threadIdx.z; + const int row_x = blockDim.y*blockIdx.y + threadIdx.y; + const int channel = blockDim.z*blockIdx.z + threadIdx.z; const int channel_x = channel / channel_x_divisor; - const int nrows_y = ncols_x; + const int nrows_y = ncols_x; const int nrows_dst = nrows_x; - const int row_dst = row_x; + const int row_dst = row_x; const int idst = channel*nrows_dst + row_dst; @@ -4345,13 +4433,13 @@ static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous break; } - const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x; - const float xi = __half2float(x[ix]); - const int row_y = col_x; + const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x; const int iy = channel*nrows_y + row_y; + const float xi = __half2float(x[ix]); + tmp += xi * y[iy]; } @@ -4614,6 +4702,11 @@ static void add_f16_f32_f16_cuda(const half * x, const float * y, half * dst, co add_f16_f32_f16<<>>(x, y, dst, k); } +static void add_f16_f32_f32_cuda(const half * x, const float * y, float * dst, const int k, cudaStream_t stream) { + const int num_blocks = (k + CUDA_ADD_BLOCK_SIZE - 1) / CUDA_ADD_BLOCK_SIZE; + add_f16_f32_f32<<>>(x, y, dst, k); +} + static void mul_f32_cuda(const float * x, const float * y, float * dst, const int kx, const int ky, cudaStream_t stream) { const int num_blocks = (kx + CUDA_MUL_BLOCK_SIZE - 1) / CUDA_MUL_BLOCK_SIZE; mul_f32<<>>(x, y, dst, kx, ky); @@ -5661,11 +5754,21 @@ void ggml_init_cublas() { CUDA_CHECK(cudaGetDeviceCount(&g_device_count)); GGML_ASSERT(g_device_count <= GGML_CUDA_MAX_DEVICES); int64_t total_vram = 0; +#if defined(GGML_CUDA_FORCE_MMQ) + fprintf(stderr, "%s: GGML_CUDA_FORCE_MMQ: yes\n", __func__); +#else + fprintf(stderr, "%s: GGML_CUDA_FORCE_MMQ: no\n", __func__); +#endif +#if defined(CUDA_USE_TENSOR_CORES) + fprintf(stderr, "%s: CUDA_USE_TENSOR_CORES: yes\n", __func__); +#else + fprintf(stderr, "%s: CUDA_USE_TENSOR_CORES: no\n", __func__); +#endif fprintf(stderr, "%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, g_device_count); - for (int64_t id = 0; id < g_device_count; ++id) { + for (int id = 0; id < g_device_count; ++id) { cudaDeviceProp prop; CUDA_CHECK(cudaGetDeviceProperties(&prop, id)); - fprintf(stderr, " Device %ld: %s, compute capability %d.%d\n", id, prop.name, prop.major, prop.minor); + fprintf(stderr, " Device %d: %s, compute capability %d.%d\n", id, prop.name, prop.major, prop.minor); g_tensor_split[id] = total_vram; total_vram += prop.totalGlobalMem; @@ -5675,15 +5778,15 @@ void ggml_init_cublas() { g_compute_capabilities[id] = 100*prop.major + 10*prop.minor; #endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) } - for (int64_t id = 0; id < g_device_count; ++id) { + for (int id = 0; id < g_device_count; ++id) { g_tensor_split[id] /= total_vram; } - for (int64_t id = 0; id < g_device_count; ++id) { + for (int id = 0; id < g_device_count; ++id) { CUDA_CHECK(ggml_cuda_set_device(id)); // create cuda streams - for (int64_t is = 0; is < MAX_STREAMS; ++is) { + for (int is = 0; is < MAX_STREAMS; ++is) { CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams[id][is], cudaStreamNonBlocking)); } @@ -5907,7 +6010,10 @@ inline void ggml_cuda_op_add( add_f32_cuda(src0_dd, src1_dd, dst_dd, ggml_nelements(src0), ne10*ne11, main_stream); } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { add_f16_f32_f16_cuda((const half *) src0_dd, src1_dd, (half *) dst_dd, ggml_nelements(src0), main_stream); + } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) { + add_f16_f32_f32_cuda((const half *) src0_dd, src1_dd, dst_dd, ggml_nelements(src0), main_stream); } else { + fprintf(stderr, "src0->type: %d dst->type: %d\n", src0->type, dst->type); GGML_ASSERT(false); } @@ -6252,16 +6358,15 @@ inline void ggml_cuda_op_mul_mat_cublas( const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols, const int64_t src1_padded_row_size, const cudaStream_t & stream) { - GGML_ASSERT(src0_dd_i != nullptr); + GGML_ASSERT(src0_dd_i != nullptr); GGML_ASSERT(src1_ddf_i != nullptr); - GGML_ASSERT(dst_dd_i != nullptr); - + GGML_ASSERT(dst_dd_i != nullptr); const int64_t ne00 = src0->ne[0]; - const int64_t ne10 = src1->ne[0]; const int64_t ne0 = dst->ne[0]; + const int64_t row_diff = row_high - row_low; int id; @@ -6346,7 +6451,7 @@ inline void ggml_cuda_op_mul_mat_cublas( cublasSgemm(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N, row_diff, src1_ncols, ne10, &alpha, src0_ddf_i, ne00, - src1_ddf_i, ne10, + src1_ddf_i, ne10, &beta, dst_dd_i, ldc)); if (src0_as != 0) { @@ -7013,7 +7118,8 @@ static void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tens } static void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){ - GGML_ASSERT(!ggml_is_contiguous(src0) && ggml_is_contiguous(src1)); + GGML_ASSERT(!ggml_is_transposed(src0)); + GGML_ASSERT(!ggml_is_transposed(src1)); GGML_ASSERT(!ggml_is_permuted(src0)); GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT); GGML_ASSERT(src0->type == GGML_TYPE_F16); @@ -7023,11 +7129,11 @@ static void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor const int64_t ne01 = src0->ne[1]; const int64_t ne02 = src0->ne[2]; - const int64_t ne12 = src1->ne[2]; - const int64_t nb01 = src0->nb[1]; const int64_t nb02 = src0->nb[2]; + const int64_t ne12 = src1->ne[2]; + CUDA_CHECK(ggml_cuda_set_device(g_main_device)); cudaStream_t main_stream = g_cudaStreams[g_main_device][0]; @@ -7046,27 +7152,200 @@ static void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, main_stream); } +static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_ASSERT(!ggml_is_transposed(src0)); + GGML_ASSERT(!ggml_is_transposed(src1)); + + GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT); + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + + const int64_t ne00 = src0->ne[0]; GGML_UNUSED(ne00); + const int64_t ne01 = src0->ne[1]; + const int64_t ne02 = src0->ne[2]; + const int64_t ne03 = src0->ne[3]; + + const int64_t nb01 = src0->nb[1]; + const int64_t nb02 = src0->nb[2]; GGML_UNUSED(nb02); + const int64_t nb03 = src0->nb[3]; GGML_UNUSED(nb03); + + const int64_t ne10 = src1->ne[0]; + const int64_t ne11 = src1->ne[1]; + const int64_t ne12 = src1->ne[2]; + const int64_t ne13 = src1->ne[3]; + + const int64_t nb11 = src1->nb[1]; + const int64_t nb12 = src1->nb[2]; GGML_UNUSED(nb12); + const int64_t nb13 = src1->nb[3]; GGML_UNUSED(nb13); + + const int64_t ne1 = ggml_nelements(src1); + const int64_t ne = ggml_nelements(dst); + + CUDA_CHECK(ggml_cuda_set_device(g_main_device)); + cudaStream_t main_stream = g_cudaStreams[g_main_device][0]; + + int id; + CUDA_CHECK(cudaGetDevice(&id)); + CUBLAS_CHECK(cublasSetStream(g_cublas_handles[id], main_stream)); + + ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; + void * src0_ddq = src0_extra->data_device[g_main_device]; + half * src0_as_f16 = (half *) src0_ddq; + + ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra; + float * src1_ddf = (float *) src1_extra->data_device[g_main_device]; + + ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; + float * dst_ddf = (float *) dst_extra->data_device[g_main_device]; + + // convert src1 to fp16 + const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type); + GGML_ASSERT(to_fp16_cuda != nullptr); + + size_t src1_as = 0; + half * src1_as_f16 = (half *) ggml_cuda_pool_malloc(ne1 * sizeof(half), &src1_as); + to_fp16_cuda(src1_ddf, src1_as_f16, ne1, main_stream); + + size_t dst_as = 0; + half * dst_f16 = (half *) ggml_cuda_pool_malloc(ne * sizeof(half), &dst_as); + + GGML_ASSERT(ne12 % ne02 == 0); + GGML_ASSERT(ne13 % ne03 == 0); + + // broadcast factors + const int64_t r2 = ne12/ne02; + const int64_t r3 = ne13/ne03; + + const half alpha_f16 = 1.0f; + const half beta_f16 = 0.0f; + +#if 0 + // use cublasGemmEx + { + for (int i13 = 0; i13 < ne13; ++i13) { + for (int i12 = 0; i12 < ne12; ++i12) { + int i03 = i13 / r3; + int i02 = i12 / r2; + + CUBLAS_CHECK( + cublasGemmEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N, + ne01, ne11, ne10, + &alpha_f16, (const char *) src0_as_f16 + i02*src0->nb[2] + i03*src0->nb[3] , CUDA_R_16F, nb01/sizeof(half), + (const char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, CUDA_R_16F, nb11/sizeof(float), + &beta_f16, ( char *) dst_f16 + i12* dst->nb[2]/2 + i13* dst->nb[3]/2, CUDA_R_16F, ne01, + CUBLAS_COMPUTE_16F, + CUBLAS_GEMM_DEFAULT_TENSOR_OP)); + } + } + } +#else + if (r2 == 1 && r3 == 1 && src0->nb[2]*src0->ne[2] == src0->nb[3] && src1->nb[2]*src1->ne[2] == src1->nb[3]) { + // there is no broadcast and src0, src1 are contiguous across dims 2, 3 + // use cublasGemmStridedBatchedEx + CUBLAS_CHECK( + cublasGemmStridedBatchedEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N, + ne01, ne11, ne10, + &alpha_f16, (const char *) src0_as_f16, CUDA_R_16F, nb01/sizeof(half), src0->nb[2]/sizeof(half), // strideA + (const char *) src1_as_f16, CUDA_R_16F, nb11/sizeof(float), src1->nb[2]/sizeof(float), // strideB + &beta_f16, ( char *) dst_f16, CUDA_R_16F, ne01, dst->nb[2]/sizeof(float), // strideC + ne12*ne13, + CUBLAS_COMPUTE_16F, + CUBLAS_GEMM_DEFAULT_TENSOR_OP)); + } else { + // use cublasGemmBatchedEx + // TODO: https://github.com/ggerganov/llama.cpp/pull/3749#discussion_r1369997000 + const int ne23 = ne12*ne13; + + // TODO: avoid this alloc + void ** ptrs = (void **) malloc(3*ne23*sizeof(void *)); + + for (int i13 = 0; i13 < ne13; ++i13) { + for (int i12 = 0; i12 < ne12; ++i12) { + int i03 = i13 / r3; + int i02 = i12 / r2; + + ptrs[0*ne23 + i12 + i13*ne12] = (char *) src0_as_f16 + i02*src0->nb[2] + i03*src0->nb[3]; + ptrs[1*ne23 + i12 + i13*ne12] = (char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2; + ptrs[2*ne23 + i12 + i13*ne12] = (char *) dst_f16 + i12* dst->nb[2]/2 + i13* dst->nb[3]/2; + } + } + + // allocate device memory for pointers + void ** ptrs_as = nullptr; + CUDA_CHECK(cudaMalloc(&ptrs_as, 3*ne23*sizeof(void *))); + + // TODO: this does not work for some reason -- not sure why? + //size_t ptrs_s = 0; + //ptrs_as = (void **) ggml_cuda_pool_malloc(3*ne23*sizeof(void *), &ptrs_s); + + // copy pointers to device + CUDA_CHECK(cudaMemcpy(ptrs_as, ptrs, 3*ne23*sizeof(void *), cudaMemcpyHostToDevice)); + + free(ptrs); + + CUBLAS_CHECK( + cublasGemmBatchedEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N, + ne01, ne11, ne10, + &alpha_f16, (const void **) (ptrs_as + 0*ne23), CUDA_R_16F, nb01/sizeof(half), + (const void **) (ptrs_as + 1*ne23), CUDA_R_16F, nb11/sizeof(float), + &beta_f16, ( void **) (ptrs_as + 2*ne23), CUDA_R_16F, ne01, + ne23, + CUBLAS_COMPUTE_16F, + CUBLAS_GEMM_DEFAULT_TENSOR_OP)); + + // free device memory for pointers + CUDA_CHECK(cudaFree(ptrs_as)); + //ggml_cuda_pool_free(ptrs_as, ptrs_s); + } +#endif + + const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16); + to_fp32_cuda(dst_f16, dst_ddf, ne, main_stream); + + ggml_cuda_pool_free(src1_as_f16, src1_as); + ggml_cuda_pool_free(dst_f16, dst_as); +} + static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - bool all_on_device = (src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT) && - src1->backend == GGML_BACKEND_GPU && dst->backend == GGML_BACKEND_GPU; + const bool all_on_device = + (src0->backend == GGML_BACKEND_GPU) && + (src1->backend == GGML_BACKEND_GPU) && + ( dst->backend == GGML_BACKEND_GPU); int64_t min_compute_capability = INT_MAX; for (int64_t id = 0; id < g_device_count; ++id) { - if (min_compute_capability > g_compute_capabilities[id] - && g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) { + if (min_compute_capability > g_compute_capabilities[id] && g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) { min_compute_capability = g_compute_capabilities[id]; } } - if (all_on_device && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) { +#ifdef CUDA_USE_TENSOR_CORES + const bool use_tensor_cores = true; +#else + const bool use_tensor_cores = false; +#endif + + // debug helpers + //printf("src0: %8d %8d %8d %8d\n", src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]); + //printf(" %8d %8d %8d %8d\n", src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]); + //printf("src1: %8d %8d %8d %8d\n", src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3]); + //printf(" %8d %8d %8d %8d\n", src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3]); + //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name); + //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name); + + if (all_on_device && !use_tensor_cores && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) { + // KQ single-batch ggml_cuda_mul_mat_vec_p021(src0, src1, dst); - } else if (all_on_device && !ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && src1->ne[1] == 1) { + } else if (all_on_device && !use_tensor_cores && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) { + // KQV single-batch ggml_cuda_mul_mat_vec_nc(src0, src1, dst); + } else if (all_on_device && src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1)) { + // KQ + KQV multi-batch + ggml_cuda_mul_mat_mat_batched_cublas(src0, src1, dst); } else if (src0->type == GGML_TYPE_F32) { ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false); } else if (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16) { if (src1->ne[1] == 1 && src0->ne[0] % GGML_CUDA_DMMV_X == 0) { - #ifdef GGML_CUDA_FORCE_DMMV const bool use_mul_mat_vec_q = false; #else @@ -7079,7 +7358,15 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1 ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, false); } } else { - if (g_mul_mat_q && ggml_is_quantized(src0->type) && min_compute_capability >= MIN_CC_DP4A) { + bool use_mul_mat_q = min_compute_capability >= MIN_CC_DP4A && ggml_is_quantized(src0->type); + + // when tensor cores are available, use them for large batch size + // ref: https://github.com/ggerganov/llama.cpp/pull/3776 + if (use_tensor_cores && min_compute_capability >= CC_VOLTA && src1->ne[1] > MMQ_MAX_BATCH_SIZE) { + use_mul_mat_q = false; + } + + if (use_mul_mat_q) { ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_q, true); } else { ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false); @@ -7433,10 +7720,6 @@ void ggml_cuda_set_main_device(const int main_device) { } } -void ggml_cuda_set_mul_mat_q(const bool mul_mat_q) { - g_mul_mat_q = mul_mat_q; -} - void ggml_cuda_set_scratch_size(const size_t scratch_size) { // this is a hack to not completely break llama.cpp when using multiple models or contexts simultaneously // it still won't always work as expected, but it's better than nothing diff --git a/ggml-impl.h b/ggml-impl.h new file mode 100644 index 000000000..5ec18a50c --- /dev/null +++ b/ggml-impl.h @@ -0,0 +1,237 @@ +#pragma once + +#include "ggml.h" + +// GGML internal header + +#include +#include +#include +#include // memcpy +#include // fabsf + +#ifdef __cplusplus +extern "C" { +#endif + +// static_assert should be a #define, but if it's not, +// fall back to the _Static_assert C11 keyword. +// if C99 - static_assert is noop +// ref: https://stackoverflow.com/a/53923785/4039976 +#ifndef static_assert +#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L) +#define static_assert(cond, msg) _Static_assert(cond, msg) +#else +#define static_assert(cond, msg) struct global_scope_noop_trick +#endif +#endif + +// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512 +#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__)) +#ifndef __FMA__ +#define __FMA__ +#endif +#ifndef __F16C__ +#define __F16C__ +#endif +#ifndef __SSE3__ +#define __SSE3__ +#endif +#endif + +#undef MIN +#undef MAX + +#define MIN(a, b) ((a) < (b) ? (a) : (b)) +#define MAX(a, b) ((a) > (b) ? (a) : (b)) + +// 16-bit float +// on Arm, we use __fp16 +// on x86, we use uint16_t +#if defined(__ARM_NEON) && !defined(_MSC_VER) + +// if YCM cannot find , make a symbolic link to it, for example: +// +// $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/ +// +#include + +#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x)) +#define GGML_COMPUTE_FP32_TO_FP16(x) (x) + +#define GGML_FP16_TO_FP32(x) ((float) (x)) +#define GGML_FP32_TO_FP16(x) (x) + +#else + +#ifdef __wasm_simd128__ +#include +#else +#ifdef __POWER9_VECTOR__ +#include +#undef bool +#define bool _Bool +#else +#if defined(_MSC_VER) || defined(__MINGW32__) +#include +#else +#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) +#if !defined(__riscv) +#include +#endif +#endif +#endif +#endif +#endif + +#ifdef __riscv_v_intrinsic +#include +#endif + +#ifdef __F16C__ + +#ifdef _MSC_VER +#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x))) +#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0) +#else +#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x) +#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0) +#endif + +#elif defined(__POWER9_VECTOR__) + +#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) +#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) +/* the inline asm below is about 12% faster than the lookup method */ +#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x) +#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) + +static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { + register float f; + register double d; + __asm__( + "mtfprd %0,%2\n" + "xscvhpdp %0,%0\n" + "frsp %1,%0\n" : + /* temp */ "=d"(d), + /* out */ "=f"(f): + /* in */ "r"(h)); + return f; +} + +static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { + register double d; + register ggml_fp16_t r; + __asm__( /* xscvdphp can work on double or single precision */ + "xscvdphp %0,%2\n" + "mffprd %1,%0\n" : + /* temp */ "=d"(d), + /* out */ "=r"(r): + /* in */ "f"(f)); + return r; +} + +#else + +// FP16 <-> FP32 +// ref: https://github.com/Maratyszcza/FP16 + +static inline float fp32_from_bits(uint32_t w) { + union { + uint32_t as_bits; + float as_value; + } fp32; + fp32.as_bits = w; + return fp32.as_value; +} + +static inline uint32_t fp32_to_bits(float f) { + union { + float as_value; + uint32_t as_bits; + } fp32; + fp32.as_value = f; + return fp32.as_bits; +} + +static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { + const uint32_t w = (uint32_t) h << 16; + const uint32_t sign = w & UINT32_C(0x80000000); + const uint32_t two_w = w + w; + + const uint32_t exp_offset = UINT32_C(0xE0) << 23; +#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__) + const float exp_scale = 0x1.0p-112f; +#else + const float exp_scale = fp32_from_bits(UINT32_C(0x7800000)); +#endif + const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale; + + const uint32_t magic_mask = UINT32_C(126) << 23; + const float magic_bias = 0.5f; + const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias; + + const uint32_t denormalized_cutoff = UINT32_C(1) << 27; + const uint32_t result = sign | + (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value)); + return fp32_from_bits(result); +} + +static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { +#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__) + const float scale_to_inf = 0x1.0p+112f; + const float scale_to_zero = 0x1.0p-110f; +#else + const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000)); + const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000)); +#endif + float base = (fabsf(f) * scale_to_inf) * scale_to_zero; + + const uint32_t w = fp32_to_bits(f); + const uint32_t shl1_w = w + w; + const uint32_t sign = w & UINT32_C(0x80000000); + uint32_t bias = shl1_w & UINT32_C(0xFF000000); + if (bias < UINT32_C(0x71000000)) { + bias = UINT32_C(0x71000000); + } + + base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base; + const uint32_t bits = fp32_to_bits(base); + const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00); + const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF); + const uint32_t nonsign = exp_bits + mantissa_bits; + return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign); +} + +#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) +#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) + +#endif // __F16C__ + +#endif // __ARM_NEON + +// precomputed f32 table for f16 (256 KB) +// defined in ggml.c, initialized in ggml_init() +extern float ggml_table_f32_f16[1 << 16]; + +// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32, +// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON. +// This is also true for POWER9. +#if !defined(GGML_FP16_TO_FP32) || !defined(GGML_FP32_TO_FP16) + +inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) { + uint16_t s; + memcpy(&s, &f, sizeof(uint16_t)); + return ggml_table_f32_f16[s]; +} + +#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x) +#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) + +#endif + + // TODO: backend v2 PR + +#ifdef __cplusplus +} +#endif diff --git a/ggml-metal.m b/ggml-metal.m index c908106be..bc881395a 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -62,6 +62,7 @@ struct ggml_metal_context { GGML_METAL_DECL_KERNEL(mul); GGML_METAL_DECL_KERNEL(mul_row); // TODO: avoid this extra kernel, instead extend the "mul" kernel to support broadcast GGML_METAL_DECL_KERNEL(scale); + GGML_METAL_DECL_KERNEL(scale_4); GGML_METAL_DECL_KERNEL(silu); GGML_METAL_DECL_KERNEL(relu); GGML_METAL_DECL_KERNEL(gelu); @@ -209,6 +210,10 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__); NSString * sourcePath = [bundle pathForResource:@"ggml-metal" ofType:@"metal"]; + if (sourcePath == nil) { + GGML_METAL_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__); + sourcePath = @"ggml-metal.metal"; + } GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [sourcePath UTF8String]); NSString * src = [NSString stringWithContentsOfFile:sourcePath encoding:NSUTF8StringEncoding error:&error]; if (error) { @@ -233,14 +238,17 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { // load kernels { NSError * error = nil; -#define GGML_METAL_ADD_KERNEL(name) \ - ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \ - ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:&error]; \ + + /* GGML_METAL_LOG_INFO("%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \ (int) ctx->pipeline_##name.maxTotalThreadsPerThreadgroup, \ (int) ctx->pipeline_##name.threadExecutionWidth); \ + */ +#define GGML_METAL_ADD_KERNEL(name) \ + ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \ + ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:&error]; \ if (error) { \ - GGML_METAL_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \ + GGML_METAL_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \ return NULL; \ } @@ -249,6 +257,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(mul); GGML_METAL_ADD_KERNEL(mul_row); GGML_METAL_ADD_KERNEL(scale); + GGML_METAL_ADD_KERNEL(scale_4); GGML_METAL_ADD_KERNEL(silu); GGML_METAL_ADD_KERNEL(relu); GGML_METAL_ADD_KERNEL(gelu); @@ -347,6 +356,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) { GGML_METAL_DEL_KERNEL(mul); GGML_METAL_DEL_KERNEL(mul_row); GGML_METAL_DEL_KERNEL(scale); + GGML_METAL_DEL_KERNEL(scale_4); GGML_METAL_DEL_KERNEL(silu); GGML_METAL_DEL_KERNEL(relu); GGML_METAL_DEL_KERNEL(gelu); @@ -923,15 +933,20 @@ void ggml_metal_graph_compute( const float scale = *(const float *) src1->data; - [encoder setComputePipelineState:ctx->pipeline_scale]; + int64_t n = ggml_nelements(dst); + + if (n % 4 == 0) { + n /= 4; + [encoder setComputePipelineState:ctx->pipeline_scale_4]; + } else { + [encoder setComputePipelineState:ctx->pipeline_scale]; + } + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; [encoder setBytes:&scale length:sizeof(scale) atIndex:2]; - const int64_t n = ggml_nelements(dst); - GGML_ASSERT(n % 4 == 0); - - [encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; } break; case GGML_OP_UNARY: switch (ggml_get_unary_op(gf->nodes[i])) { diff --git a/ggml-metal.metal b/ggml-metal.metal index 69fc71362..f4b460564 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -125,9 +125,17 @@ kernel void kernel_mul_row( } kernel void kernel_scale( + device const float * src0, + device float * dst, + constant float & scale, + uint tpig[[thread_position_in_grid]]) { + dst[tpig] = src0[tpig] * scale; +} + +kernel void kernel_scale_4( device const float4 * src0, device float4 * dst, - constant float & scale, + constant float & scale, uint tpig[[thread_position_in_grid]]) { dst[tpig] = src0[tpig] * scale; } diff --git a/k_quants.c b/ggml-quants.c similarity index 69% rename from k_quants.c rename to ggml-quants.c index 801941fbe..740be6dc5 100644 --- a/k_quants.c +++ b/ggml-quants.c @@ -1,9 +1,10 @@ -#include "k_quants.h" -#include "ggml.h" +#include "ggml-quants.h" +#include "ggml-impl.h" #include #include #include +#include #ifdef __ARM_NEON @@ -65,6 +66,1026 @@ inline static int32_t vaddvq_s32(int32x4_t v) { #define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1) +#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) +// multiply int8_t, add results pairwise twice +static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) { + // Get absolute values of x vectors + const __m128i ax = _mm_sign_epi8(x, x); + // Sign the values of the y vectors + const __m128i sy = _mm_sign_epi8(y, x); + // Perform multiplication and create 16-bit values + const __m128i dot = _mm_maddubs_epi16(ax, sy); + const __m128i ones = _mm_set1_epi16(1); + return _mm_madd_epi16(ones, dot); +} + +#if __AVX__ || __AVX2__ || __AVX512F__ +// horizontally add 8 floats +static inline float hsum_float_8(const __m256 x) { + __m128 res = _mm256_extractf128_ps(x, 1); + res = _mm_add_ps(res, _mm256_castps256_ps128(x)); + res = _mm_add_ps(res, _mm_movehl_ps(res, res)); + res = _mm_add_ss(res, _mm_movehdup_ps(res)); + return _mm_cvtss_f32(res); +} + +// horizontally add 8 int32_t +static inline int hsum_i32_8(const __m256i a) { + const __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(a), _mm256_extractf128_si256(a, 1)); + const __m128i hi64 = _mm_unpackhi_epi64(sum128, sum128); + const __m128i sum64 = _mm_add_epi32(hi64, sum128); + const __m128i hi32 = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1)); + return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32)); +} + +// horizontally add 4 int32_t +static inline int hsum_i32_4(const __m128i a) { + const __m128i hi64 = _mm_unpackhi_epi64(a, a); + const __m128i sum64 = _mm_add_epi32(hi64, a); + const __m128i hi32 = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1)); + return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32)); +} + +#if defined(__AVX2__) || defined(__AVX512F__) +// spread 32 bits to 32 bytes { 0x00, 0xFF } +static inline __m256i bytes_from_bits_32(const uint8_t * x) { + uint32_t x32; + memcpy(&x32, x, sizeof(uint32_t)); + const __m256i shuf_mask = _mm256_set_epi64x( + 0x0303030303030303, 0x0202020202020202, + 0x0101010101010101, 0x0000000000000000); + __m256i bytes = _mm256_shuffle_epi8(_mm256_set1_epi32(x32), shuf_mask); + const __m256i bit_mask = _mm256_set1_epi64x(0x7fbfdfeff7fbfdfe); + bytes = _mm256_or_si256(bytes, bit_mask); + return _mm256_cmpeq_epi8(bytes, _mm256_set1_epi64x(-1)); +} + +// Unpack 32 4-bit fields into 32 bytes +// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval +static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi) +{ + const __m128i tmp = _mm_loadu_si128((const __m128i *)rsi); + const __m256i bytes = MM256_SET_M128I(_mm_srli_epi16(tmp, 4), tmp); + const __m256i lowMask = _mm256_set1_epi8( 0xF ); + return _mm256_and_si256(lowMask, bytes); +} + +// add int16_t pairwise and return as float vector +static inline __m256 sum_i16_pairs_float(const __m256i x) { + const __m256i ones = _mm256_set1_epi16(1); + const __m256i summed_pairs = _mm256_madd_epi16(ones, x); + return _mm256_cvtepi32_ps(summed_pairs); +} + +static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) { +#if __AVXVNNI__ + const __m256i zero = _mm256_setzero_si256(); + const __m256i summed_pairs = _mm256_dpbusd_epi32(zero, ax, sy); + return _mm256_cvtepi32_ps(summed_pairs); +#else + // Perform multiplication and create 16-bit values + const __m256i dot = _mm256_maddubs_epi16(ax, sy); + return sum_i16_pairs_float(dot); +#endif +} + +// multiply int8_t, add results pairwise twice and return as float vector +static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) { +#if __AVXVNNIINT8__ + const __m256i zero = _mm256_setzero_si256(); + const __m256i summed_pairs = _mm256_dpbssd_epi32(zero, x, y); + return _mm256_cvtepi32_ps(summed_pairs); +#else + // Get absolute values of x vectors + const __m256i ax = _mm256_sign_epi8(x, x); + // Sign the values of the y vectors + const __m256i sy = _mm256_sign_epi8(y, x); + return mul_sum_us8_pairs_float(ax, sy); +#endif +} + +static inline __m128i packNibbles( __m256i bytes ) +{ + // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh +#if __AVX512F__ + const __m256i bytes_srli_4 = _mm256_srli_epi16(bytes, 4); // 0000_0000_abcd_0000 + bytes = _mm256_or_si256(bytes, bytes_srli_4); // 0000_abcd_abcd_efgh + return _mm256_cvtepi16_epi8(bytes); // abcd_efgh +#else + const __m256i lowByte = _mm256_set1_epi16( 0xFF ); + __m256i high = _mm256_andnot_si256( lowByte, bytes ); + __m256i low = _mm256_and_si256( lowByte, bytes ); + high = _mm256_srli_epi16( high, 4 ); + bytes = _mm256_or_si256( low, high ); + + // Compress uint16_t lanes into bytes + __m128i r0 = _mm256_castsi256_si128( bytes ); + __m128i r1 = _mm256_extracti128_si256( bytes, 1 ); + return _mm_packus_epi16( r0, r1 ); +#endif +} +#elif defined(__AVX__) +// spread 32 bits to 32 bytes { 0x00, 0xFF } +static inline __m256i bytes_from_bits_32(const uint8_t * x) { + uint32_t x32; + memcpy(&x32, x, sizeof(uint32_t)); + const __m128i shuf_maskl = _mm_set_epi64x(0x0101010101010101, 0x0000000000000000); + const __m128i shuf_maskh = _mm_set_epi64x(0x0303030303030303, 0x0202020202020202); + __m128i bytesl = _mm_shuffle_epi8(_mm_set1_epi32(x32), shuf_maskl); + __m128i bytesh = _mm_shuffle_epi8(_mm_set1_epi32(x32), shuf_maskh); + const __m128i bit_mask = _mm_set1_epi64x(0x7fbfdfeff7fbfdfe); + bytesl = _mm_or_si128(bytesl, bit_mask); + bytesh = _mm_or_si128(bytesh, bit_mask); + bytesl = _mm_cmpeq_epi8(bytesl, _mm_set1_epi64x(-1)); + bytesh = _mm_cmpeq_epi8(bytesh, _mm_set1_epi64x(-1)); + return MM256_SET_M128I(bytesh, bytesl); +} + +// Unpack 32 4-bit fields into 32 bytes +// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval +static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi) +{ + // Load 16 bytes from memory + __m128i tmpl = _mm_loadu_si128((const __m128i *)rsi); + __m128i tmph = _mm_srli_epi16(tmpl, 4); + const __m128i lowMask = _mm_set1_epi8(0xF); + tmpl = _mm_and_si128(lowMask, tmpl); + tmph = _mm_and_si128(lowMask, tmph); + return MM256_SET_M128I(tmph, tmpl); +} + +// add int16_t pairwise and return as float vector +static inline __m256 sum_i16_pairs_float(const __m128i xh, const __m128i xl) { + const __m128i ones = _mm_set1_epi16(1); + const __m128i summed_pairsl = _mm_madd_epi16(ones, xl); + const __m128i summed_pairsh = _mm_madd_epi16(ones, xh); + const __m256i summed_pairs = MM256_SET_M128I(summed_pairsh, summed_pairsl); + return _mm256_cvtepi32_ps(summed_pairs); +} + +static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) { + const __m128i axl = _mm256_castsi256_si128(ax); + const __m128i axh = _mm256_extractf128_si256(ax, 1); + const __m128i syl = _mm256_castsi256_si128(sy); + const __m128i syh = _mm256_extractf128_si256(sy, 1); + // Perform multiplication and create 16-bit values + const __m128i dotl = _mm_maddubs_epi16(axl, syl); + const __m128i doth = _mm_maddubs_epi16(axh, syh); + return sum_i16_pairs_float(doth, dotl); +} + +// multiply int8_t, add results pairwise twice and return as float vector +static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) { + const __m128i xl = _mm256_castsi256_si128(x); + const __m128i xh = _mm256_extractf128_si256(x, 1); + const __m128i yl = _mm256_castsi256_si128(y); + const __m128i yh = _mm256_extractf128_si256(y, 1); + // Get absolute values of x vectors + const __m128i axl = _mm_sign_epi8(xl, xl); + const __m128i axh = _mm_sign_epi8(xh, xh); + // Sign the values of the y vectors + const __m128i syl = _mm_sign_epi8(yl, xl); + const __m128i syh = _mm_sign_epi8(yh, xh); + // Perform multiplication and create 16-bit values + const __m128i dotl = _mm_maddubs_epi16(axl, syl); + const __m128i doth = _mm_maddubs_epi16(axh, syh); + return sum_i16_pairs_float(doth, dotl); +} + +static inline __m128i packNibbles( __m128i bytes1, __m128i bytes2 ) +{ + // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh + const __m128i lowByte = _mm_set1_epi16( 0xFF ); + __m128i high = _mm_andnot_si128( lowByte, bytes1 ); + __m128i low = _mm_and_si128( lowByte, bytes1 ); + high = _mm_srli_epi16( high, 4 ); + bytes1 = _mm_or_si128( low, high ); + high = _mm_andnot_si128( lowByte, bytes2 ); + low = _mm_and_si128( lowByte, bytes2 ); + high = _mm_srli_epi16( high, 4 ); + bytes2 = _mm_or_si128( low, high ); + + return _mm_packus_epi16( bytes1, bytes2); +} +#endif +#elif defined(__SSSE3__) +// horizontally add 4x4 floats +static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 c, const __m128 d) { + __m128 res_0 =_mm_hadd_ps(a, b); + __m128 res_1 =_mm_hadd_ps(c, d); + __m128 res =_mm_hadd_ps(res_0, res_1); + res =_mm_hadd_ps(res, res); + res =_mm_hadd_ps(res, res); + + return _mm_cvtss_f32(res); +} +#endif // __AVX__ || __AVX2__ || __AVX512F__ +#endif // defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) + +#if defined(__ARM_NEON) + +#if !defined(__aarch64__) + +inline static int32_t vaddvq_s32(int32x4_t v) { + return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3); +} + +inline static float vaddvq_f32(float32x4_t v) { + return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3); +} + +inline static float vmaxvq_f32(float32x4_t v) { + return + MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)), + MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3))); +} + +inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) { + int32x4_t res; + + res[0] = roundf(vgetq_lane_f32(v, 0)); + res[1] = roundf(vgetq_lane_f32(v, 1)); + res[2] = roundf(vgetq_lane_f32(v, 2)); + res[3] = roundf(vgetq_lane_f32(v, 3)); + + return res; +} + +#endif +#endif + +#if defined(__ARM_NEON) || defined(__wasm_simd128__) +#define B1(c,s,n) 0x ## n ## c , 0x ## n ## s +#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s) +#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s) +#define B4(c,s,n) B3(c,s,n ## c), B3(c,s,n ## s) +#define B5(c,s,n) B4(c,s,n ## c), B4(c,s,n ## s) +#define B6(c,s,n) B5(c,s,n ## c), B5(c,s,n ## s) +#define B7(c,s,n) B6(c,s,n ## c), B6(c,s,n ## s) +#define B8(c,s ) B7(c,s, c), B7(c,s, s) + +// precomputed tables for expanding 8bits to 8 bytes: +static const uint64_t table_b2b_0[1 << 8] = { B8(00, 10) }; // ( b) << 4 +static const uint64_t table_b2b_1[1 << 8] = { B8(10, 00) }; // (!b) << 4 +#endif + +// reference implementation for deterministic creation of model files +void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k) { + static const int qk = QK4_0; + + assert(k % qk == 0); + + const int nb = k / qk; + + for (int i = 0; i < nb; i++) { + float amax = 0.0f; // absolute max + float max = 0.0f; + + for (int j = 0; j < qk; j++) { + const float v = x[i*qk + j]; + if (amax < fabsf(v)) { + amax = fabsf(v); + max = v; + } + } + + const float d = max / -8; + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = GGML_FP32_TO_FP16(d); + + for (int j = 0; j < qk/2; ++j) { + const float x0 = x[i*qk + 0 + j]*id; + const float x1 = x[i*qk + qk/2 + j]*id; + + const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f)); + const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f)); + + y[i].qs[j] = xi0; + y[i].qs[j] |= xi1 << 4; + } + } +} + +void quantize_row_q4_0(const float * restrict x, void * restrict y, int k) { + quantize_row_q4_0_reference(x, y, k); +} + +void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k) { + const int qk = QK4_1; + + assert(k % qk == 0); + + const int nb = k / qk; + + for (int i = 0; i < nb; i++) { + float min = FLT_MAX; + float max = -FLT_MAX; + + for (int j = 0; j < qk; j++) { + const float v = x[i*qk + j]; + + if (v < min) min = v; + if (v > max) max = v; + } + + const float d = (max - min) / ((1 << 4) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = GGML_FP32_TO_FP16(d); + y[i].m = GGML_FP32_TO_FP16(min); + + for (int j = 0; j < qk/2; ++j) { + const float x0 = (x[i*qk + 0 + j] - min)*id; + const float x1 = (x[i*qk + qk/2 + j] - min)*id; + + const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f)); + const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f)); + + y[i].qs[j] = xi0; + y[i].qs[j] |= xi1 << 4; + } + } +} + +void quantize_row_q4_1(const float * restrict x, void * restrict y, int k) { + quantize_row_q4_1_reference(x, y, k); +} + +void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k) { + static const int qk = QK5_0; + + assert(k % qk == 0); + + const int nb = k / qk; + + for (int i = 0; i < nb; i++) { + float amax = 0.0f; // absolute max + float max = 0.0f; + + for (int j = 0; j < qk; j++) { + const float v = x[i*qk + j]; + if (amax < fabsf(v)) { + amax = fabsf(v); + max = v; + } + } + + const float d = max / -16; + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = GGML_FP32_TO_FP16(d); + + uint32_t qh = 0; + + for (int j = 0; j < qk/2; ++j) { + const float x0 = x[i*qk + 0 + j]*id; + const float x1 = x[i*qk + qk/2 + j]*id; + + const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f)); + const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f)); + + y[i].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4); + + // get the 5-th bit and store it in qh at the right position + qh |= ((xi0 & 0x10u) >> 4) << (j + 0); + qh |= ((xi1 & 0x10u) >> 4) << (j + qk/2); + } + + memcpy(&y[i].qh, &qh, sizeof(qh)); + } +} + +void quantize_row_q5_0(const float * restrict x, void * restrict y, int k) { + quantize_row_q5_0_reference(x, y, k); +} + +void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k) { + const int qk = QK5_1; + + assert(k % qk == 0); + + const int nb = k / qk; + + for (int i = 0; i < nb; i++) { + float min = FLT_MAX; + float max = -FLT_MAX; + + for (int j = 0; j < qk; j++) { + const float v = x[i*qk + j]; + + if (v < min) min = v; + if (v > max) max = v; + } + + const float d = (max - min) / ((1 << 5) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = GGML_FP32_TO_FP16(d); + y[i].m = GGML_FP32_TO_FP16(min); + + uint32_t qh = 0; + + for (int j = 0; j < qk/2; ++j) { + const float x0 = (x[i*qk + 0 + j] - min)*id; + const float x1 = (x[i*qk + qk/2 + j] - min)*id; + + const uint8_t xi0 = (uint8_t)(x0 + 0.5f); + const uint8_t xi1 = (uint8_t)(x1 + 0.5f); + + y[i].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4); + + // get the 5-th bit and store it in qh at the right position + qh |= ((xi0 & 0x10u) >> 4) << (j + 0); + qh |= ((xi1 & 0x10u) >> 4) << (j + qk/2); + } + + memcpy(&y[i].qh, &qh, sizeof(y[i].qh)); + } +} + +void quantize_row_q5_1(const float * restrict x, void * restrict y, int k) { + quantize_row_q5_1_reference(x, y, k); +} + +// reference implementation for deterministic creation of model files +void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k) { + assert(k % QK8_0 == 0); + const int nb = k / QK8_0; + + for (int i = 0; i < nb; i++) { + float amax = 0.0f; // absolute max + + for (int j = 0; j < QK8_0; j++) { + const float v = x[i*QK8_0 + j]; + amax = MAX(amax, fabsf(v)); + } + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = GGML_FP32_TO_FP16(d); + + for (int j = 0; j < QK8_0; ++j) { + const float x0 = x[i*QK8_0 + j]*id; + + y[i].qs[j] = roundf(x0); + } + } +} + +void quantize_row_q8_0(const float * restrict x, void * restrict vy, int k) { + assert(QK8_0 == 32); + assert(k % QK8_0 == 0); + const int nb = k / QK8_0; + + block_q8_0 * restrict y = vy; + +#if defined(__ARM_NEON) + for (int i = 0; i < nb; i++) { + float32x4_t srcv [8]; + float32x4_t asrcv[8]; + float32x4_t amaxv[8]; + + for (int j = 0; j < 8; j++) srcv[j] = vld1q_f32(x + i*32 + 4*j); + for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]); + + for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]); + for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]); + for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]); + + const float amax = vmaxvq_f32(amaxv[0]); + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = GGML_FP32_TO_FP16(d); + + for (int j = 0; j < 8; j++) { + const float32x4_t v = vmulq_n_f32(srcv[j], id); + const int32x4_t vi = vcvtnq_s32_f32(v); + + y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0); + y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1); + y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2); + y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3); + } + } +#elif defined(__wasm_simd128__) + for (int i = 0; i < nb; i++) { + v128_t srcv [8]; + v128_t asrcv[8]; + v128_t amaxv[8]; + + for (int j = 0; j < 8; j++) srcv[j] = wasm_v128_load(x + i*32 + 4*j); + for (int j = 0; j < 8; j++) asrcv[j] = wasm_f32x4_abs(srcv[j]); + + for (int j = 0; j < 4; j++) amaxv[2*j] = wasm_f32x4_max(asrcv[2*j], asrcv[2*j+1]); + for (int j = 0; j < 2; j++) amaxv[4*j] = wasm_f32x4_max(amaxv[4*j], amaxv[4*j+2]); + for (int j = 0; j < 1; j++) amaxv[8*j] = wasm_f32x4_max(amaxv[8*j], amaxv[8*j+4]); + + const float amax = MAX(MAX(wasm_f32x4_extract_lane(amaxv[0], 0), + wasm_f32x4_extract_lane(amaxv[0], 1)), + MAX(wasm_f32x4_extract_lane(amaxv[0], 2), + wasm_f32x4_extract_lane(amaxv[0], 3))); + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = GGML_FP32_TO_FP16(d); + + for (int j = 0; j < 8; j++) { + const v128_t v = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id)); + const v128_t vi = wasm_i32x4_trunc_sat_f32x4(v); + + y[i].qs[4*j + 0] = wasm_i32x4_extract_lane(vi, 0); + y[i].qs[4*j + 1] = wasm_i32x4_extract_lane(vi, 1); + y[i].qs[4*j + 2] = wasm_i32x4_extract_lane(vi, 2); + y[i].qs[4*j + 3] = wasm_i32x4_extract_lane(vi, 3); + } + } +#elif defined(__AVX2__) || defined(__AVX__) + for (int i = 0; i < nb; i++) { + // Load elements into 4 AVX vectors + __m256 v0 = _mm256_loadu_ps( x ); + __m256 v1 = _mm256_loadu_ps( x + 8 ); + __m256 v2 = _mm256_loadu_ps( x + 16 ); + __m256 v3 = _mm256_loadu_ps( x + 24 ); + x += 32; + + // Compute max(abs(e)) for the block + const __m256 signBit = _mm256_set1_ps( -0.0f ); + __m256 maxAbs = _mm256_andnot_ps( signBit, v0 ); + maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) ); + maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) ); + maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) ); + + __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) ); + max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) ); + max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) ); + const float maxScalar = _mm_cvtss_f32( max4 ); + + // Quantize these floats + const float d = maxScalar / 127.f; + y[i].d = GGML_FP32_TO_FP16(d); + const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; + const __m256 mul = _mm256_set1_ps( id ); + + // Apply the multiplier + v0 = _mm256_mul_ps( v0, mul ); + v1 = _mm256_mul_ps( v1, mul ); + v2 = _mm256_mul_ps( v2, mul ); + v3 = _mm256_mul_ps( v3, mul ); + + // Round to nearest integer + v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST ); + v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST ); + v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST ); + v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST ); + + // Convert floats to integers + __m256i i0 = _mm256_cvtps_epi32( v0 ); + __m256i i1 = _mm256_cvtps_epi32( v1 ); + __m256i i2 = _mm256_cvtps_epi32( v2 ); + __m256i i3 = _mm256_cvtps_epi32( v3 ); + +#if defined(__AVX2__) + // Convert int32 to int16 + i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 + i2 = _mm256_packs_epi32( i2, i3 ); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31 + // Convert int16 to int8 + i0 = _mm256_packs_epi16( i0, i2 ); // 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27, 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31 + + // We got our precious signed bytes, but the order is now wrong + // These AVX2 pack instructions process 16-byte pieces independently + // The following instruction is fixing the order + const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 ); + i0 = _mm256_permutevar8x32_epi32( i0, perm ); + + _mm256_storeu_si256((__m256i *)y[i].qs, i0); +#else + // Since we don't have in AVX some necessary functions, + // we split the registers in half and call AVX2 analogs from SSE + __m128i ni0 = _mm256_castsi256_si128( i0 ); + __m128i ni1 = _mm256_extractf128_si256( i0, 1); + __m128i ni2 = _mm256_castsi256_si128( i1 ); + __m128i ni3 = _mm256_extractf128_si256( i1, 1); + __m128i ni4 = _mm256_castsi256_si128( i2 ); + __m128i ni5 = _mm256_extractf128_si256( i2, 1); + __m128i ni6 = _mm256_castsi256_si128( i3 ); + __m128i ni7 = _mm256_extractf128_si256( i3, 1); + + // Convert int32 to int16 + ni0 = _mm_packs_epi32( ni0, ni1 ); + ni2 = _mm_packs_epi32( ni2, ni3 ); + ni4 = _mm_packs_epi32( ni4, ni5 ); + ni6 = _mm_packs_epi32( ni6, ni7 ); + // Convert int16 to int8 + ni0 = _mm_packs_epi16( ni0, ni2 ); + ni4 = _mm_packs_epi16( ni4, ni6 ); + + _mm_storeu_si128((__m128i *)(y[i].qs + 0), ni0); + _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4); +#endif + } +#elif defined(__riscv_v_intrinsic) + + size_t vl = __riscv_vsetvl_e32m4(QK8_0); + + for (int i = 0; i < nb; i++) { + // load elements + vfloat32m4_t v_x = __riscv_vle32_v_f32m4(x+i*QK8_0, vl); + + vfloat32m4_t vfabs = __riscv_vfabs_v_f32m4(v_x, vl); + vfloat32m1_t tmp = __riscv_vfmv_v_f_f32m1(0.0f, vl); + vfloat32m1_t vmax = __riscv_vfredmax_vs_f32m4_f32m1(vfabs, tmp, vl); + float amax = __riscv_vfmv_f_s_f32m1_f32(vmax); + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = GGML_FP32_TO_FP16(d); + + vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl); + + // convert to integer + vint16m2_t vi = __riscv_vfncvt_x_f_w_i16m2(x0, vl); + vint8m1_t vs = __riscv_vncvt_x_x_w_i8m1(vi, vl); + + // store result + __riscv_vse8_v_i8m1(y[i].qs , vs, vl); + } +#else + GGML_UNUSED(nb); + // scalar + quantize_row_q8_0_reference(x, y, k); +#endif +} + +// reference implementation for deterministic creation of model files +void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int k) { + assert(QK8_1 == 32); + assert(k % QK8_1 == 0); + const int nb = k / QK8_1; + + for (int i = 0; i < nb; i++) { + float amax = 0.0f; // absolute max + + for (int j = 0; j < QK8_1; j++) { + const float v = x[i*QK8_1 + j]; + amax = MAX(amax, fabsf(v)); + } + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = d; + + int sum = 0; + + for (int j = 0; j < QK8_1/2; ++j) { + const float v0 = x[i*QK8_1 + j]*id; + const float v1 = x[i*QK8_1 + QK8_1/2 + j]*id; + + y[i].qs[ j] = roundf(v0); + y[i].qs[QK8_1/2 + j] = roundf(v1); + + sum += y[i].qs[ j]; + sum += y[i].qs[QK8_1/2 + j]; + } + + y[i].s = sum*d; + } +} + +void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) { + assert(k % QK8_1 == 0); + const int nb = k / QK8_1; + + block_q8_1 * restrict y = vy; + +#if defined(__ARM_NEON) + for (int i = 0; i < nb; i++) { + float32x4_t srcv [8]; + float32x4_t asrcv[8]; + float32x4_t amaxv[8]; + + for (int j = 0; j < 8; j++) srcv[j] = vld1q_f32(x + i*32 + 4*j); + for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]); + + for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]); + for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]); + for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]); + + const float amax = vmaxvq_f32(amaxv[0]); + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = d; + + int32x4_t accv = vdupq_n_s32(0); + + for (int j = 0; j < 8; j++) { + const float32x4_t v = vmulq_n_f32(srcv[j], id); + const int32x4_t vi = vcvtnq_s32_f32(v); + + y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0); + y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1); + y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2); + y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3); + + accv = vaddq_s32(accv, vi); + } + + y[i].s = d * vaddvq_s32(accv); + } +#elif defined(__wasm_simd128__) + for (int i = 0; i < nb; i++) { + v128_t srcv [8]; + v128_t asrcv[8]; + v128_t amaxv[8]; + + for (int j = 0; j < 8; j++) srcv[j] = wasm_v128_load(x + i*32 + 4*j); + for (int j = 0; j < 8; j++) asrcv[j] = wasm_f32x4_abs(srcv[j]); + + for (int j = 0; j < 4; j++) amaxv[2*j] = wasm_f32x4_max(asrcv[2*j], asrcv[2*j+1]); + for (int j = 0; j < 2; j++) amaxv[4*j] = wasm_f32x4_max(amaxv[4*j], amaxv[4*j+2]); + for (int j = 0; j < 1; j++) amaxv[8*j] = wasm_f32x4_max(amaxv[8*j], amaxv[8*j+4]); + + const float amax = MAX(MAX(wasm_f32x4_extract_lane(amaxv[0], 0), + wasm_f32x4_extract_lane(amaxv[0], 1)), + MAX(wasm_f32x4_extract_lane(amaxv[0], 2), + wasm_f32x4_extract_lane(amaxv[0], 3))); + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = d; + + v128_t accv = wasm_i32x4_splat(0); + + for (int j = 0; j < 8; j++) { + const v128_t v = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id)); + const v128_t vi = wasm_i32x4_trunc_sat_f32x4(v); + + y[i].qs[4*j + 0] = wasm_i32x4_extract_lane(vi, 0); + y[i].qs[4*j + 1] = wasm_i32x4_extract_lane(vi, 1); + y[i].qs[4*j + 2] = wasm_i32x4_extract_lane(vi, 2); + y[i].qs[4*j + 3] = wasm_i32x4_extract_lane(vi, 3); + + accv = wasm_i32x4_add(accv, vi); + } + + y[i].s = d * (wasm_i32x4_extract_lane(accv, 0) + + wasm_i32x4_extract_lane(accv, 1) + + wasm_i32x4_extract_lane(accv, 2) + + wasm_i32x4_extract_lane(accv, 3)); + } +#elif defined(__AVX2__) || defined(__AVX__) + for (int i = 0; i < nb; i++) { + // Load elements into 4 AVX vectors + __m256 v0 = _mm256_loadu_ps( x ); + __m256 v1 = _mm256_loadu_ps( x + 8 ); + __m256 v2 = _mm256_loadu_ps( x + 16 ); + __m256 v3 = _mm256_loadu_ps( x + 24 ); + x += 32; + + // Compute max(abs(e)) for the block + const __m256 signBit = _mm256_set1_ps( -0.0f ); + __m256 maxAbs = _mm256_andnot_ps( signBit, v0 ); + maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) ); + maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) ); + maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) ); + + __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) ); + max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) ); + max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) ); + const float maxScalar = _mm_cvtss_f32( max4 ); + + // Quantize these floats + const float d = maxScalar / 127.f; + y[i].d = d; + const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; + const __m256 mul = _mm256_set1_ps( id ); + + // Apply the multiplier + v0 = _mm256_mul_ps( v0, mul ); + v1 = _mm256_mul_ps( v1, mul ); + v2 = _mm256_mul_ps( v2, mul ); + v3 = _mm256_mul_ps( v3, mul ); + + // Round to nearest integer + v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST ); + v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST ); + v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST ); + v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST ); + + // Convert floats to integers + __m256i i0 = _mm256_cvtps_epi32( v0 ); + __m256i i1 = _mm256_cvtps_epi32( v1 ); + __m256i i2 = _mm256_cvtps_epi32( v2 ); + __m256i i3 = _mm256_cvtps_epi32( v3 ); + +#if defined(__AVX2__) + // Compute the sum of the quants and set y[i].s + y[i].s = d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))); + + // Convert int32 to int16 + i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 + i2 = _mm256_packs_epi32( i2, i3 ); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31 + // Convert int16 to int8 + i0 = _mm256_packs_epi16( i0, i2 ); // 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27, 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31 + + // We got our precious signed bytes, but the order is now wrong + // These AVX2 pack instructions process 16-byte pieces independently + // The following instruction is fixing the order + const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 ); + i0 = _mm256_permutevar8x32_epi32( i0, perm ); + + _mm256_storeu_si256((__m256i *)y[i].qs, i0); +#else + // Since we don't have in AVX some necessary functions, + // we split the registers in half and call AVX2 analogs from SSE + __m128i ni0 = _mm256_castsi256_si128( i0 ); + __m128i ni1 = _mm256_extractf128_si256( i0, 1); + __m128i ni2 = _mm256_castsi256_si128( i1 ); + __m128i ni3 = _mm256_extractf128_si256( i1, 1); + __m128i ni4 = _mm256_castsi256_si128( i2 ); + __m128i ni5 = _mm256_extractf128_si256( i2, 1); + __m128i ni6 = _mm256_castsi256_si128( i3 ); + __m128i ni7 = _mm256_extractf128_si256( i3, 1); + + // Compute the sum of the quants and set y[i].s + const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3)); + const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7)); + y[i].s = d * hsum_i32_4(_mm_add_epi32(s0, s1)); + + // Convert int32 to int16 + ni0 = _mm_packs_epi32( ni0, ni1 ); + ni2 = _mm_packs_epi32( ni2, ni3 ); + ni4 = _mm_packs_epi32( ni4, ni5 ); + ni6 = _mm_packs_epi32( ni6, ni7 ); + // Convert int16 to int8 + ni0 = _mm_packs_epi16( ni0, ni2 ); + ni4 = _mm_packs_epi16( ni4, ni6 ); + + _mm_storeu_si128((__m128i *)(y[i].qs + 0), ni0); + _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4); +#endif + } +#elif defined(__riscv_v_intrinsic) + + size_t vl = __riscv_vsetvl_e32m4(QK8_1); + + for (int i = 0; i < nb; i++) { + // load elements + vfloat32m4_t v_x = __riscv_vle32_v_f32m4(x+i*QK8_1, vl); + + vfloat32m4_t vfabs = __riscv_vfabs_v_f32m4(v_x, vl); + vfloat32m1_t tmp = __riscv_vfmv_v_f_f32m1(0.0, vl); + vfloat32m1_t vmax = __riscv_vfredmax_vs_f32m4_f32m1(vfabs, tmp, vl); + float amax = __riscv_vfmv_f_s_f32m1_f32(vmax); + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = d; + + vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl); + + // convert to integer + vint16m2_t vi = __riscv_vfncvt_x_f_w_i16m2(x0, vl); + vint8m1_t vs = __riscv_vncvt_x_x_w_i8m1(vi, vl); + + // store result + __riscv_vse8_v_i8m1(y[i].qs , vs, vl); + + // compute sum for y[i].s + vint16m1_t tmp2 = __riscv_vmv_v_x_i16m1(0, vl); + vint16m1_t vwrs = __riscv_vwredsum_vs_i8m1_i16m1(vs, tmp2, vl); + + // set y[i].s + int sum = __riscv_vmv_x_s_i16m1_i16(vwrs); + y[i].s = sum*d; + } +#else + GGML_UNUSED(nb); + // scalar + quantize_row_q8_1_reference(x, y, k); +#endif +} + +void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k) { + static const int qk = QK4_0; + + assert(k % qk == 0); + + const int nb = k / qk; + + for (int i = 0; i < nb; i++) { + const float d = GGML_FP16_TO_FP32(x[i].d); + + for (int j = 0; j < qk/2; ++j) { + const int x0 = (x[i].qs[j] & 0x0F) - 8; + const int x1 = (x[i].qs[j] >> 4) - 8; + + y[i*qk + j + 0 ] = x0*d; + y[i*qk + j + qk/2] = x1*d; + } + } +} + +void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int k) { + static const int qk = QK4_1; + + assert(k % qk == 0); + + const int nb = k / qk; + + for (int i = 0; i < nb; i++) { + const float d = GGML_FP16_TO_FP32(x[i].d); + const float m = GGML_FP16_TO_FP32(x[i].m); + + for (int j = 0; j < qk/2; ++j) { + const int x0 = (x[i].qs[j] & 0x0F); + const int x1 = (x[i].qs[j] >> 4); + + y[i*qk + j + 0 ] = x0*d + m; + y[i*qk + j + qk/2] = x1*d + m; + } + } +} + +void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int k) { + static const int qk = QK5_0; + + assert(k % qk == 0); + + const int nb = k / qk; + + for (int i = 0; i < nb; i++) { + const float d = GGML_FP16_TO_FP32(x[i].d); + + uint32_t qh; + memcpy(&qh, x[i].qh, sizeof(qh)); + + for (int j = 0; j < qk/2; ++j) { + const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; + + const int32_t x0 = ((x[i].qs[j] & 0x0F) | xh_0) - 16; + const int32_t x1 = ((x[i].qs[j] >> 4) | xh_1) - 16; + + y[i*qk + j + 0 ] = x0*d; + y[i*qk + j + qk/2] = x1*d; + } + } +} + +void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int k) { + static const int qk = QK5_1; + + assert(k % qk == 0); + + const int nb = k / qk; + + for (int i = 0; i < nb; i++) { + const float d = GGML_FP16_TO_FP32(x[i].d); + const float m = GGML_FP16_TO_FP32(x[i].m); + + uint32_t qh; + memcpy(&qh, x[i].qh, sizeof(qh)); + + for (int j = 0; j < qk/2; ++j) { + const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; + + const int x0 = (x[i].qs[j] & 0x0F) | xh_0; + const int x1 = (x[i].qs[j] >> 4) | xh_1; + + y[i*qk + j + 0 ] = x0*d + m; + y[i*qk + j + qk/2] = x1*d + m; + } + } +} + +void dequantize_row_q8_0(const block_q8_0 * restrict x, float * restrict y, int k) { + static const int qk = QK8_0; + + assert(k % qk == 0); + + const int nb = k / qk; + + for (int i = 0; i < nb; i++) { + const float d = GGML_FP16_TO_FP32(x[i].d); + + for (int j = 0; j < qk; ++j) { + y[i*qk + j] = x[i].qs[j]*d; + } + } +} + // // 2-6 bit quantization in super-blocks // @@ -368,10 +1389,10 @@ void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict int l = nearest_int(iscale*scales[j]); y[i].scales[j] = l; } - y[i].d = ggml_fp32_to_fp16(max_scale/q4scale); + y[i].d = GGML_FP32_TO_FP16(max_scale/q4scale); } else { for (int j = 0; j < QK_K/16; ++j) y[i].scales[j] = 0; - y[i].d = ggml_fp32_to_fp16(0.f); + y[i].d = GGML_FP32_TO_FP16(0.f); } if (max_min > 0) { float iscale = q4scale/max_min; @@ -379,14 +1400,14 @@ void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict int l = nearest_int(iscale*mins[j]); y[i].scales[j] |= (l << 4); } - y[i].dmin = ggml_fp32_to_fp16(max_min/q4scale); + y[i].dmin = GGML_FP32_TO_FP16(max_min/q4scale); } else { - y[i].dmin = ggml_fp32_to_fp16(0.f); + y[i].dmin = GGML_FP32_TO_FP16(0.f); } for (int j = 0; j < QK_K/16; ++j) { - const float d = ggml_fp16_to_fp32(y[i].d) * (y[i].scales[j] & 0xF); + const float d = GGML_FP16_TO_FP32(y[i].d) * (y[i].scales[j] & 0xF); if (!d) continue; - const float dm = ggml_fp16_to_fp32(y[i].dmin) * (y[i].scales[j] >> 4); + const float dm = GGML_FP16_TO_FP32(y[i].dmin) * (y[i].scales[j] >> 4); for (int ii = 0; ii < 16; ++ii) { int l = nearest_int((x[16*j + ii] + dm)/d); l = MAX(0, MIN(3, l)); @@ -417,8 +1438,8 @@ void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int for (int i = 0; i < nb; i++) { - const float d = ggml_fp16_to_fp32(x[i].d); - const float min = ggml_fp16_to_fp32(x[i].dmin); + const float d = GGML_FP16_TO_FP32(x[i].d); + const float min = GGML_FP16_TO_FP32(x[i].dmin); const uint8_t * q = x[i].qs; @@ -507,16 +1528,16 @@ void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict l >>= 4; y[i].scales[j%4 + 8] |= (l << (2*(j/4))); } - y[i].d = ggml_fp32_to_fp16(1/iscale); + y[i].d = GGML_FP32_TO_FP16(1/iscale); } else { - y[i].d = ggml_fp32_to_fp16(0.f); + y[i].d = GGML_FP32_TO_FP16(0.f); } int8_t sc; for (int j = 0; j < QK_K/16; ++j) { sc = j < 8 ? y[i].scales[j] & 0xF : y[i].scales[j-8] >> 4; sc = (sc | (((y[i].scales[8 + j%4] >> (2*(j/4))) & 3) << 4)) - 32; - float d = ggml_fp16_to_fp32(y[i].d) * sc; + float d = GGML_FP16_TO_FP32(y[i].d) * sc; if (!d) { continue; } @@ -536,16 +1557,16 @@ void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict l2 = 8 + MAX(-8, MIN(7, l2)); y[i].scales[j/2] = l1 | (l2 << 4); } - y[i].d = ggml_fp32_to_fp16(1/iscale); + y[i].d = GGML_FP32_TO_FP16(1/iscale); } else { for (int j = 0; j < QK_K/16; j+=2) { y[i].scales[j/2] = 0; } - y[i].d = ggml_fp32_to_fp16(0.f); + y[i].d = GGML_FP32_TO_FP16(0.f); } for (int j = 0; j < QK_K/16; ++j) { int s = j%2 == 0 ? y[i].scales[j/2] & 0xF : y[i].scales[j/2] >> 4; - float d = ggml_fp16_to_fp32(y[i].d) * (s - 8); + float d = GGML_FP16_TO_FP32(y[i].d) * (s - 8); if (!d) { continue; } @@ -599,7 +1620,7 @@ void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int for (int i = 0; i < nb; i++) { - const float d_all = ggml_fp16_to_fp32(x[i].d); + const float d_all = GGML_FP16_TO_FP32(x[i].d); const uint8_t * restrict q = x[i].qs; const uint8_t * restrict hm = x[i].hmask; @@ -644,7 +1665,7 @@ void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int for (int i = 0; i < nb; i++) { - const float d_all = ggml_fp16_to_fp32(x[i].d); + const float d_all = GGML_FP16_TO_FP32(x[i].d); const uint8_t * restrict q = x[i].qs; const uint8_t * restrict hm = x[i].hmask; @@ -734,15 +1755,15 @@ void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y[i].scales[j-0] |= ((lm >> 4) << 6); } } - y[i].d = ggml_fp32_to_fp16(max_scale/63.f); - y[i].dmin = ggml_fp32_to_fp16(max_min/63.f); + y[i].d = GGML_FP32_TO_FP16(max_scale/63.f); + y[i].dmin = GGML_FP32_TO_FP16(max_min/63.f); uint8_t sc, m; for (int j = 0; j < QK_K/32; ++j) { get_scale_min_k4(j, y[i].scales, &sc, &m); - const float d = ggml_fp16_to_fp32(y[i].d) * sc; + const float d = GGML_FP16_TO_FP32(y[i].d) * sc; if (!d) continue; - const float dm = ggml_fp16_to_fp32(y[i].dmin) * m; + const float dm = GGML_FP16_TO_FP32(y[i].dmin) * m; for (int ii = 0; ii < 32; ++ii) { int l = nearest_int((x[32*j + ii] + dm)/d); l = MAX(0, MIN(15, l)); @@ -759,17 +1780,17 @@ void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict int m2 = nearest_int(inv_min*mins[1]); y[i].scales[0] = d1 | (m1 << 4); y[i].scales[1] = d2 | (m2 << 4); - y[i].d[0] = ggml_fp32_to_fp16(max_scale/s_factor); - y[i].d[1] = ggml_fp32_to_fp16(max_min/s_factor); + y[i].d[0] = GGML_FP32_TO_FP16(max_scale/s_factor); + y[i].d[1] = GGML_FP32_TO_FP16(max_min/s_factor); float sumlx = 0; int suml2 = 0; for (int j = 0; j < QK_K/32; ++j) { const uint8_t sd = y[i].scales[j] & 0xF; const uint8_t sm = y[i].scales[j] >> 4; - const float d = ggml_fp16_to_fp32(y[i].d[0]) * sd; + const float d = GGML_FP16_TO_FP32(y[i].d[0]) * sd; if (!d) continue; - const float m = ggml_fp16_to_fp32(y[i].d[1]) * sm; + const float m = GGML_FP16_TO_FP32(y[i].d[1]) * sm; for (int ii = 0; ii < 32; ++ii) { int l = nearest_int((x[32*j + ii] + m)/d); l = MAX(0, MIN(15, l)); @@ -779,7 +1800,7 @@ void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict } } if (suml2) { - y[i].d[0] = ggml_fp32_to_fp16(sumlx/suml2); + y[i].d[0] = GGML_FP32_TO_FP16(sumlx/suml2); } #endif uint8_t * q = y[i].qs; @@ -803,8 +1824,8 @@ void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int #if QK_K == 256 - const float d = ggml_fp16_to_fp32(x[i].d); - const float min = ggml_fp16_to_fp32(x[i].dmin); + const float d = GGML_FP16_TO_FP32(x[i].d); + const float min = GGML_FP16_TO_FP32(x[i].dmin); int is = 0; uint8_t sc, m; @@ -818,8 +1839,8 @@ void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int q += 32; is += 2; } #else - const float dall = ggml_fp16_to_fp32(x[i].d[0]); - const float mall = ggml_fp16_to_fp32(x[i].d[1]); + const float dall = GGML_FP16_TO_FP32(x[i].d[0]); + const float mall = GGML_FP16_TO_FP32(x[i].d[1]); const float d1 = dall * (x[i].scales[0] & 0xF), m1 = mall * (x[i].scales[0] >> 4); const float d2 = dall * (x[i].scales[1] & 0xF), m2 = mall * (x[i].scales[1] >> 4); for (int l = 0; l < 32; ++l) { @@ -905,15 +1926,15 @@ void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y[i].scales[j-0] |= ((lm >> 4) << 6); } } - y[i].d = ggml_fp32_to_fp16(max_scale/63.f); - y[i].dmin = ggml_fp32_to_fp16(max_min/63.f); + y[i].d = GGML_FP32_TO_FP16(max_scale/63.f); + y[i].dmin = GGML_FP32_TO_FP16(max_min/63.f); uint8_t sc, m; for (int j = 0; j < QK_K/32; ++j) { get_scale_min_k4(j, y[i].scales, &sc, &m); - const float d = ggml_fp16_to_fp32(y[i].d) * sc; + const float d = GGML_FP16_TO_FP32(y[i].d) * sc; if (!d) continue; - const float dm = ggml_fp16_to_fp32(y[i].dmin) * m; + const float dm = GGML_FP16_TO_FP32(y[i].dmin) * m; for (int ii = 0; ii < 32; ++ii) { int l = nearest_int((x[32*j + ii] + dm)/d); l = MAX(0, MIN(31, l)); @@ -957,10 +1978,10 @@ void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict int l = nearest_int(iscale*scales[j]); y[i].scales[j] = MAX(-128, MIN(127, l)); } - y[i].d = ggml_fp32_to_fp16(1/iscale); + y[i].d = GGML_FP32_TO_FP16(1/iscale); for (int j = 0; j < QK_K/16; ++j) { - const float d = ggml_fp16_to_fp32(y[i].d) * y[i].scales[j]; + const float d = GGML_FP16_TO_FP32(y[i].d) * y[i].scales[j]; if (!d) continue; for (int ii = 0; ii < 16; ++ii) { int l = nearest_int(x[16*j + ii]/d); @@ -1004,8 +2025,8 @@ void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int #if QK_K == 256 - const float d = ggml_fp16_to_fp32(x[i].d); - const float min = ggml_fp16_to_fp32(x[i].dmin); + const float d = GGML_FP16_TO_FP32(x[i].d); + const float min = GGML_FP16_TO_FP32(x[i].dmin); int is = 0; uint8_t sc, m; @@ -1021,7 +2042,7 @@ void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int u1 <<= 2; u2 <<= 2; } #else - float d = ggml_fp16_to_fp32(x[i].d); + float d = GGML_FP16_TO_FP32(x[i].d); const int8_t * restrict s = x[i].scales; for (int l = 0; l < 8; ++l) { y[l+ 0] = d * s[0] * ((ql[l+ 0] & 0xF) - (qh[l] & 0x01 ? 0 : 16)); @@ -1084,19 +2105,19 @@ void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict if (!max_abs_scale) { memset(&y[i], 0, sizeof(block_q6_K)); - y[i].d = ggml_fp32_to_fp16(0.f); + y[i].d = GGML_FP32_TO_FP16(0.f); x += QK_K; continue; } float iscale = -128.f/max_scale; - y[i].d = ggml_fp32_to_fp16(1/iscale); + y[i].d = GGML_FP32_TO_FP16(1/iscale); for (int ib = 0; ib < QK_K/16; ++ib) { y[i].scales[ib] = MIN(127, nearest_int(iscale*scales[ib])); } for (int j = 0; j < QK_K/16; ++j) { - float d = ggml_fp16_to_fp32(y[i].d) * y[i].scales[j]; + float d = GGML_FP16_TO_FP32(y[i].d) * y[i].scales[j]; if (!d) { continue; } @@ -1145,7 +2166,7 @@ void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int for (int i = 0; i < nb; i++) { - const float d = ggml_fp16_to_fp32(x[i].d); + const float d = GGML_FP16_TO_FP32(x[i].d); const uint8_t * restrict ql = x[i].ql; const uint8_t * restrict qh = x[i].qh; @@ -1264,15 +2285,6 @@ void quantize_row_q8_K(const float * restrict x, void * restrict y, int k) { // #if __AVX__ || __AVX2__ || __AVX512F__ -// horizontally add 8 floats -static inline float hsum_float_8(const __m256 x) { - __m128 res = _mm256_extractf128_ps(x, 1); - res = _mm_add_ps(res, _mm256_castps256_ps128(x)); - res = _mm_add_ps(res, _mm_movehl_ps(res, res)); - res = _mm_add_ss(res, _mm_movehdup_ps(res)); - return _mm_cvtss_f32(res); -} - // shuffles to pick the required scales in dot products static inline __m256i get_scale_shuffle_q3k(int i) { static const uint8_t k_shuffle[128] = { @@ -1311,6 +2323,1224 @@ static inline __m128i get_scale_shuffle(int i) { } #endif +void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + const int qk = QK8_0; + const int nb = n / qk; + + assert(n % qk == 0); + + const block_q4_0 * restrict x = vx; + const block_q8_0 * restrict y = vy; + +#if defined(__ARM_NEON) + float32x4_t sumv0 = vdupq_n_f32(0.0f); + float32x4_t sumv1 = vdupq_n_f32(0.0f); + + assert(nb % 2 == 0); // TODO: handle odd nb + + for (int i = 0; i < nb; i += 2) { + const block_q4_0 * restrict x0 = &x[i + 0]; + const block_q4_0 * restrict x1 = &x[i + 1]; + const block_q8_0 * restrict y0 = &y[i + 0]; + const block_q8_0 * restrict y1 = &y[i + 1]; + + const uint8x16_t m4b = vdupq_n_u8(0x0F); + const int8x16_t s8b = vdupq_n_s8(0x8); + + const uint8x16_t v0_0 = vld1q_u8(x0->qs); + const uint8x16_t v0_1 = vld1q_u8(x1->qs); + + // 4-bit -> 8-bit + const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); + const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); + const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); + const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); + + // sub 8 + const int8x16_t v0_0ls = vsubq_s8(v0_0l, s8b); + const int8x16_t v0_0hs = vsubq_s8(v0_0h, s8b); + const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b); + const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b); + + // load y + const int8x16_t v1_0l = vld1q_s8(y0->qs); + const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); + const int8x16_t v1_1l = vld1q_s8(y1->qs); + const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); + +#if defined(__ARM_FEATURE_DOTPROD) + // dot product into int32x4_t + const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h); + const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h); + + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); +#else + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hs), vget_low_s8 (v1_0h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hs), vget_high_s8(v1_0h)); + + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1ls), vget_low_s8 (v1_1l)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1ls), vget_high_s8(v1_1l)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hs), vget_low_s8 (v1_1h)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hs), vget_high_s8(v1_1h)); + + const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); + const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); + const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); + const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); + + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); +#endif + } + + *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); +#elif defined(__AVX2__) + // Initialize accumulator with zeros + __m256 acc = _mm256_setzero_ps(); + + // Main loop + for (int i = 0; i < nb; ++i) { + /* Compute combined scale for the block */ + const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) ); + + __m256i bx = bytes_from_nibbles_32(x[i].qs); + + // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval. + const __m256i off = _mm256_set1_epi8( 8 ); + bx = _mm256_sub_epi8( bx, off ); + + __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); + + const __m256 q = mul_sum_i8_pairs_float(bx, by); + + /* Multiply q with scale and accumulate */ + acc = _mm256_fmadd_ps( d, q, acc ); + } + + *s = hsum_float_8(acc); +#elif defined(__AVX__) + // Initialize accumulator with zeros + __m256 acc = _mm256_setzero_ps(); + + // Main loop + for (int i = 0; i < nb; ++i) { + // Compute combined scale for the block + const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) ); + + const __m128i lowMask = _mm_set1_epi8(0xF); + const __m128i off = _mm_set1_epi8(8); + + const __m128i tmp = _mm_loadu_si128((const __m128i *)x[i].qs); + + __m128i bx = _mm_and_si128(lowMask, tmp); + __m128i by = _mm_loadu_si128((const __m128i *)y[i].qs); + bx = _mm_sub_epi8(bx, off); + const __m128i i32_0 = mul_sum_i8_pairs(bx, by); + + bx = _mm_and_si128(lowMask, _mm_srli_epi64(tmp, 4)); + by = _mm_loadu_si128((const __m128i *)(y[i].qs + 16)); + bx = _mm_sub_epi8(bx, off); + const __m128i i32_1 = mul_sum_i8_pairs(bx, by); + + // Convert int32_t to float + __m256 p = _mm256_cvtepi32_ps(MM256_SET_M128I(i32_0, i32_1)); + + // Apply the scale, and accumulate + acc = _mm256_add_ps(_mm256_mul_ps( d, p ), acc); + } + + *s = hsum_float_8(acc); +#elif defined(__SSSE3__) + // set constants + const __m128i lowMask = _mm_set1_epi8(0xF); + const __m128i off = _mm_set1_epi8(8); + + // Initialize accumulator with zeros + __m128 acc_0 = _mm_setzero_ps(); + __m128 acc_1 = _mm_setzero_ps(); + __m128 acc_2 = _mm_setzero_ps(); + __m128 acc_3 = _mm_setzero_ps(); + + // First round without accumulation + { + _mm_prefetch(&x[0] + sizeof(block_q4_0), _MM_HINT_T0); + _mm_prefetch(&y[0] + sizeof(block_q8_0), _MM_HINT_T0); + + // Compute combined scale for the block 0 and 1 + const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[0].d) * GGML_FP16_TO_FP32(y[0].d) ); + + const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[0].qs); + + __m128i bx_0 = _mm_and_si128(lowMask, tmp_0_1); + __m128i by_0 = _mm_loadu_si128((const __m128i *)y[0].qs); + bx_0 = _mm_sub_epi8(bx_0, off); + const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0); + + __m128i bx_1 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_0_1, 4)); + __m128i by_1 = _mm_loadu_si128((const __m128i *)(y[0].qs + 16)); + bx_1 = _mm_sub_epi8(bx_1, off); + const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1); + + _mm_prefetch(&x[1] + sizeof(block_q4_0), _MM_HINT_T0); + _mm_prefetch(&y[1] + sizeof(block_q8_0), _MM_HINT_T0); + + // Compute combined scale for the block 2 and 3 + const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[1].d) * GGML_FP16_TO_FP32(y[1].d) ); + + const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[1].qs); + + __m128i bx_2 = _mm_and_si128(lowMask, tmp_2_3); + __m128i by_2 = _mm_loadu_si128((const __m128i *)y[1].qs); + bx_2 = _mm_sub_epi8(bx_2, off); + const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2); + + __m128i bx_3 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_2_3, 4)); + __m128i by_3 = _mm_loadu_si128((const __m128i *)(y[1].qs + 16)); + bx_3 = _mm_sub_epi8(bx_3, off); + const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3); + + // Convert int32_t to float + __m128 p0 = _mm_cvtepi32_ps(i32_0); + __m128 p1 = _mm_cvtepi32_ps(i32_1); + __m128 p2 = _mm_cvtepi32_ps(i32_2); + __m128 p3 = _mm_cvtepi32_ps(i32_3); + + // Apply the scale + acc_0 = _mm_mul_ps( d_0_1, p0 ); + acc_1 = _mm_mul_ps( d_0_1, p1 ); + acc_2 = _mm_mul_ps( d_2_3, p2 ); + acc_3 = _mm_mul_ps( d_2_3, p3 ); + } + + assert(nb % 2 == 0); // TODO: handle odd nb + + // Main loop + for (int i = 2; i < nb; i+=2) { + _mm_prefetch(&x[i] + sizeof(block_q4_0), _MM_HINT_T0); + _mm_prefetch(&y[i] + sizeof(block_q8_0), _MM_HINT_T0); + + // Compute combined scale for the block 0 and 1 + const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) ); + + const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[i].qs); + + __m128i bx_0 = _mm_and_si128(lowMask, tmp_0_1); + __m128i by_0 = _mm_loadu_si128((const __m128i *)y[i].qs); + bx_0 = _mm_sub_epi8(bx_0, off); + const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0); + + __m128i bx_1 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_0_1, 4)); + __m128i by_1 = _mm_loadu_si128((const __m128i *)(y[i].qs + 16)); + bx_1 = _mm_sub_epi8(bx_1, off); + const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1); + + _mm_prefetch(&x[i] + 2 * sizeof(block_q4_0), _MM_HINT_T0); + _mm_prefetch(&y[i] + 2 * sizeof(block_q8_0), _MM_HINT_T0); + + // Compute combined scale for the block 2 and 3 + const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[i + 1].d) * GGML_FP16_TO_FP32(y[i + 1].d) ); + + const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[i + 1].qs); + + __m128i bx_2 = _mm_and_si128(lowMask, tmp_2_3); + __m128i by_2 = _mm_loadu_si128((const __m128i *)y[i + 1].qs); + bx_2 = _mm_sub_epi8(bx_2, off); + const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2); + + __m128i bx_3 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_2_3, 4)); + __m128i by_3 = _mm_loadu_si128((const __m128i *)(y[i + 1].qs + 16)); + bx_3 = _mm_sub_epi8(bx_3, off); + const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3); + + // Convert int32_t to float + __m128 p0 = _mm_cvtepi32_ps(i32_0); + __m128 p1 = _mm_cvtepi32_ps(i32_1); + __m128 p2 = _mm_cvtepi32_ps(i32_2); + __m128 p3 = _mm_cvtepi32_ps(i32_3); + + // Apply the scale + __m128 p0_d = _mm_mul_ps( d_0_1, p0 ); + __m128 p1_d = _mm_mul_ps( d_0_1, p1 ); + __m128 p2_d = _mm_mul_ps( d_2_3, p2 ); + __m128 p3_d = _mm_mul_ps( d_2_3, p3 ); + + // Acummulate + acc_0 = _mm_add_ps(p0_d, acc_0); + acc_1 = _mm_add_ps(p1_d, acc_1); + acc_2 = _mm_add_ps(p2_d, acc_2); + acc_3 = _mm_add_ps(p3_d, acc_3); + } + + *s = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3); +#elif defined(__riscv_v_intrinsic) + float sumf = 0.0; + + size_t vl = __riscv_vsetvl_e8m1(qk/2); + + for (int i = 0; i < nb; i++) { + // load elements + vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl); + + vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl); + vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl); + + // mask and store lower part of x, and then upper part + vuint8mf2_t x_a = __riscv_vand_vx_u8mf2(tx, 0x0F, vl); + vuint8mf2_t x_l = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl); + + vint8mf2_t x_ai = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a); + vint8mf2_t x_li = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l); + + // subtract offset + vint8mf2_t v0 = __riscv_vsub_vx_i8mf2(x_ai, 8, vl); + vint8mf2_t v1 = __riscv_vsub_vx_i8mf2(x_li, 8, vl); + + vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl); + vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl); + + vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl); + + vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl); + vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl); + + int sumi = __riscv_vmv_x_s_i32m1_i32(vs2); + + sumf += sumi*GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d); + } + + *s = sumf; +#else + // scalar + float sumf = 0.0; + + for (int i = 0; i < nb; i++) { + int sumi = 0; + + for (int j = 0; j < qk/2; ++j) { + const int v0 = (x[i].qs[j] & 0x0F) - 8; + const int v1 = (x[i].qs[j] >> 4) - 8; + + sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]); + } + + sumf += sumi*GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d); + } + + *s = sumf; +#endif +} + +void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + const int qk = QK8_1; + const int nb = n / qk; + + assert(n % qk == 0); + + const block_q4_1 * restrict x = vx; + const block_q8_1 * restrict y = vy; + + // TODO: add WASM SIMD +#if defined(__ARM_NEON) + float32x4_t sumv0 = vdupq_n_f32(0.0f); + float32x4_t sumv1 = vdupq_n_f32(0.0f); + + float summs = 0; + + assert(nb % 2 == 0); // TODO: handle odd nb + + for (int i = 0; i < nb; i += 2) { + const block_q4_1 * restrict x0 = &x[i + 0]; + const block_q4_1 * restrict x1 = &x[i + 1]; + const block_q8_1 * restrict y0 = &y[i + 0]; + const block_q8_1 * restrict y1 = &y[i + 1]; + + summs += GGML_FP16_TO_FP32(x0->m) * y0->s + GGML_FP16_TO_FP32(x1->m) * y1->s; + + const uint8x16_t m4b = vdupq_n_u8(0x0F); + + const uint8x16_t v0_0 = vld1q_u8(x0->qs); + const uint8x16_t v0_1 = vld1q_u8(x1->qs); + + // 4-bit -> 8-bit + const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); + const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); + const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); + const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); + + // load y + const int8x16_t v1_0l = vld1q_s8(y0->qs); + const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); + const int8x16_t v1_1l = vld1q_s8(y1->qs); + const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); + +#if defined(__ARM_FEATURE_DOTPROD) + // dot product into int32x4_t + const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h); + const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h); + + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*y1->d); +#else + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0l), vget_low_s8 (v1_0l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0l), vget_high_s8(v1_0l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0h), vget_low_s8 (v1_0h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0h), vget_high_s8(v1_0h)); + + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1l), vget_low_s8 (v1_1l)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1l), vget_high_s8(v1_1l)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1h), vget_low_s8 (v1_1h)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1h), vget_high_s8(v1_1h)); + + const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); + const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); + const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); + const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); + + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*y1->d); +#endif + } + + *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs; +#elif defined(__AVX2__) || defined(__AVX__) + // Initialize accumulator with zeros + __m256 acc = _mm256_setzero_ps(); + + float summs = 0; + + // Main loop + for (int i = 0; i < nb; ++i) { + const float d0 = GGML_FP16_TO_FP32(x[i].d); + const float d1 = y[i].d; + + summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s; + + const __m256 d0v = _mm256_set1_ps( d0 ); + const __m256 d1v = _mm256_set1_ps( d1 ); + + // Compute combined scales + const __m256 d0d1 = _mm256_mul_ps( d0v, d1v ); + + // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes + const __m256i bx = bytes_from_nibbles_32(x[i].qs); + const __m256i by = _mm256_loadu_si256( (const __m256i *)y[i].qs ); + + const __m256 xy = mul_sum_us8_pairs_float(bx, by); + + // Accumulate d0*d1*x*y +#if defined(__AVX2__) + acc = _mm256_fmadd_ps( d0d1, xy, acc ); +#else + acc = _mm256_add_ps( _mm256_mul_ps( d0d1, xy ), acc ); +#endif + } + + *s = hsum_float_8(acc) + summs; +#elif defined(__riscv_v_intrinsic) + float sumf = 0.0; + + size_t vl = __riscv_vsetvl_e8m1(qk/2); + + for (int i = 0; i < nb; i++) { + // load elements + vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl); + + vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl); + vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl); + + // mask and store lower part of x, and then upper part + vuint8mf2_t x_a = __riscv_vand_vx_u8mf2(tx, 0x0F, vl); + vuint8mf2_t x_l = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl); + + vint8mf2_t v0 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a); + vint8mf2_t v1 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l); + + vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl); + vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl); + + vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl); + + vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl); + vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl); + + int sumi = __riscv_vmv_x_s_i32m1_i32(vs2); + + sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s; + } + + *s = sumf; +#else + // scalar + float sumf = 0.0; + + for (int i = 0; i < nb; i++) { + int sumi = 0; + + for (int j = 0; j < qk/2; ++j) { + const int v0 = (x[i].qs[j] & 0x0F); + const int v1 = (x[i].qs[j] >> 4); + + sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]); + } + + sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s; + } + + *s = sumf; +#endif +} + +void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + const int qk = QK8_0; + const int nb = n / qk; + + assert(n % qk == 0); + assert(qk == QK5_0); + + const block_q5_0 * restrict x = vx; + const block_q8_0 * restrict y = vy; + +#if defined(__ARM_NEON) + float32x4_t sumv0 = vdupq_n_f32(0.0f); + float32x4_t sumv1 = vdupq_n_f32(0.0f); + + uint32_t qh0; + uint32_t qh1; + + uint64_t tmp0[4]; + uint64_t tmp1[4]; + + assert(nb % 2 == 0); // TODO: handle odd nb + + for (int i = 0; i < nb; i += 2) { + const block_q5_0 * restrict x0 = &x[i]; + const block_q5_0 * restrict x1 = &x[i + 1]; + const block_q8_0 * restrict y0 = &y[i]; + const block_q8_0 * restrict y1 = &y[i + 1]; + + const uint8x16_t m4b = vdupq_n_u8(0x0F); + + // extract the 5th bit via lookup table ((!b) << 4) + memcpy(&qh0, x0->qh, sizeof(qh0)); + memcpy(&qh1, x1->qh, sizeof(qh1)); + + tmp0[0] = table_b2b_1[(qh0 >> 0) & 0xFF]; + tmp0[1] = table_b2b_1[(qh0 >> 8) & 0xFF]; + tmp0[2] = table_b2b_1[(qh0 >> 16) & 0xFF]; + tmp0[3] = table_b2b_1[(qh0 >> 24) ]; + + tmp1[0] = table_b2b_1[(qh1 >> 0) & 0xFF]; + tmp1[1] = table_b2b_1[(qh1 >> 8) & 0xFF]; + tmp1[2] = table_b2b_1[(qh1 >> 16) & 0xFF]; + tmp1[3] = table_b2b_1[(qh1 >> 24) ]; + + const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0)); + const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2)); + const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0)); + const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2)); + + const uint8x16_t v0_0 = vld1q_u8(x0->qs); + const uint8x16_t v0_1 = vld1q_u8(x1->qs); + + // 4-bit -> 8-bit + int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); + int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); + int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); + int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); + + // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero) + const int8x16_t v0_0lf = vsubq_s8(v0_0l, qhl0); + const int8x16_t v0_0hf = vsubq_s8(v0_0h, qhh0); + const int8x16_t v0_1lf = vsubq_s8(v0_1l, qhl1); + const int8x16_t v0_1hf = vsubq_s8(v0_1h, qhh1); + + // load y + const int8x16_t v1_0l = vld1q_s8(y0->qs); + const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); + const int8x16_t v1_1l = vld1q_s8(y1->qs); + const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); + +#if defined(__ARM_FEATURE_DOTPROD) + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l), + vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l), + vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); +#else + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lf), vget_low_s8 (v1_0l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lf), vget_high_s8(v1_0l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hf), vget_low_s8 (v1_0h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hf), vget_high_s8(v1_0h)); + + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lf), vget_low_s8 (v1_1l)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lf), vget_high_s8(v1_1l)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hf), vget_low_s8 (v1_1h)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hf), vget_high_s8(v1_1h)); + + const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); + const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); + const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); + const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); + + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); +#endif + } + + *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); +#elif defined(__wasm_simd128__) + v128_t sumv = wasm_f32x4_splat(0.0f); + + uint32_t qh; + uint64_t tmp[4]; + + // TODO: check if unrolling this is better + for (int i = 0; i < nb; ++i) { + const block_q5_0 * restrict x0 = &x[i]; + const block_q8_0 * restrict y0 = &y[i]; + + const v128_t m4b = wasm_i8x16_splat(0x0F); + + // extract the 5th bit + memcpy(&qh, x0->qh, sizeof(qh)); + + tmp[0] = table_b2b_1[(qh >> 0) & 0xFF]; + tmp[1] = table_b2b_1[(qh >> 8) & 0xFF]; + tmp[2] = table_b2b_1[(qh >> 16) & 0xFF]; + tmp[3] = table_b2b_1[(qh >> 24) ]; + + const v128_t qhl = wasm_v128_load(tmp + 0); + const v128_t qhh = wasm_v128_load(tmp + 2); + + const v128_t v0 = wasm_v128_load(x0->qs); + + // 4-bit -> 8-bit + const v128_t v0l = wasm_v128_and (v0, m4b); + const v128_t v0h = wasm_u8x16_shr(v0, 4); + + // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero) + const v128_t v0lf = wasm_i8x16_sub(v0l, qhl); + const v128_t v0hf = wasm_i8x16_sub(v0h, qhh); + + // load y + const v128_t v1l = wasm_v128_load(y0->qs); + const v128_t v1h = wasm_v128_load(y0->qs + 16); + + // int8x16 -> int16x8 + const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf); + const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf); + const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf); + const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf); + + const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l); + const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l); + const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h); + const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h); + + // dot product + sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4( + wasm_i32x4_add( + wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll), + wasm_i32x4_dot_i16x8(v0lfh, v1lh)), + wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl), + wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), + wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d)))); + } + + *s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) + + wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3); +#elif defined(__AVX2__) + // Initialize accumulator with zeros + __m256 acc = _mm256_setzero_ps(); + + // Main loop + for (int i = 0; i < nb; i++) { + /* Compute combined scale for the block */ + const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d)); + + __m256i bx = bytes_from_nibbles_32(x[i].qs); + __m256i bxhi = bytes_from_bits_32(x[i].qh); + bxhi = _mm256_andnot_si256(bxhi, _mm256_set1_epi8((char)0xF0)); + bx = _mm256_or_si256(bx, bxhi); + + __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); + + const __m256 q = mul_sum_i8_pairs_float(bx, by); + + /* Multiply q with scale and accumulate */ + acc = _mm256_fmadd_ps(d, q, acc); + } + + *s = hsum_float_8(acc); +#elif defined(__AVX__) + // Initialize accumulator with zeros + __m256 acc = _mm256_setzero_ps(); + __m128i mask = _mm_set1_epi8((char)0xF0); + + // Main loop + for (int i = 0; i < nb; i++) { + /* Compute combined scale for the block */ + const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d)); + + __m256i bx = bytes_from_nibbles_32(x[i].qs); + const __m256i bxhi = bytes_from_bits_32(x[i].qh); + __m128i bxhil = _mm256_castsi256_si128(bxhi); + __m128i bxhih = _mm256_extractf128_si256(bxhi, 1); + bxhil = _mm_andnot_si128(bxhil, mask); + bxhih = _mm_andnot_si128(bxhih, mask); + __m128i bxl = _mm256_castsi256_si128(bx); + __m128i bxh = _mm256_extractf128_si256(bx, 1); + bxl = _mm_or_si128(bxl, bxhil); + bxh = _mm_or_si128(bxh, bxhih); + bx = MM256_SET_M128I(bxh, bxl); + + const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); + + const __m256 q = mul_sum_i8_pairs_float(bx, by); + + /* Multiply q with scale and accumulate */ + acc = _mm256_add_ps(_mm256_mul_ps(d, q), acc); + } + + *s = hsum_float_8(acc); +#elif defined(__riscv_v_intrinsic) + float sumf = 0.0; + + uint32_t qh; + + size_t vl = __riscv_vsetvl_e8m1(qk/2); + + // These tempory registers are for masking and shift operations + vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl); + vuint32m2_t vt_2 = __riscv_vsll_vv_u32m2(__riscv_vmv_v_x_u32m2(1, vl), vt_1, vl); + + vuint32m2_t vt_3 = __riscv_vsll_vx_u32m2(vt_2, 16, vl); + vuint32m2_t vt_4 = __riscv_vadd_vx_u32m2(vt_1, 12, vl); + + for (int i = 0; i < nb; i++) { + memcpy(&qh, x[i].qh, sizeof(uint32_t)); + + // ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4; + vuint32m2_t xha_0 = __riscv_vand_vx_u32m2(vt_2, qh, vl); + vuint32m2_t xhr_0 = __riscv_vsrl_vv_u32m2(xha_0, vt_1, vl); + vuint32m2_t xhl_0 = __riscv_vsll_vx_u32m2(xhr_0, 4, vl); + + // ((qh & (1u << (j + 16))) >> (j + 12)); + vuint32m2_t xha_1 = __riscv_vand_vx_u32m2(vt_3, qh, vl); + vuint32m2_t xhl_1 = __riscv_vsrl_vv_u32m2(xha_1, vt_4, vl); + + // narrowing + vuint16m1_t xhc_0 = __riscv_vncvt_x_x_w_u16m1(xhl_0, vl); + vuint8mf2_t xh_0 = __riscv_vncvt_x_x_w_u8mf2(xhc_0, vl); + + vuint16m1_t xhc_1 = __riscv_vncvt_x_x_w_u16m1(xhl_1, vl); + vuint8mf2_t xh_1 = __riscv_vncvt_x_x_w_u8mf2(xhc_1, vl); + + // load + vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl); + + vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl); + vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl); + + vuint8mf2_t x_at = __riscv_vand_vx_u8mf2(tx, 0x0F, vl); + vuint8mf2_t x_lt = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl); + + vuint8mf2_t x_a = __riscv_vor_vv_u8mf2(x_at, xh_0, vl); + vuint8mf2_t x_l = __riscv_vor_vv_u8mf2(x_lt, xh_1, vl); + + vint8mf2_t x_ai = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a); + vint8mf2_t x_li = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l); + + vint8mf2_t v0 = __riscv_vsub_vx_i8mf2(x_ai, 16, vl); + vint8mf2_t v1 = __riscv_vsub_vx_i8mf2(x_li, 16, vl); + + vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl); + vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl); + + vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl); + + vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl); + vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl); + + int sumi = __riscv_vmv_x_s_i32m1_i32(vs2); + + sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d)) * sumi; + } + + *s = sumf; +#else + // scalar + float sumf = 0.0; + + for (int i = 0; i < nb; i++) { + uint32_t qh; + memcpy(&qh, x[i].qh, sizeof(qh)); + + int sumi = 0; + + for (int j = 0; j < qk/2; ++j) { + const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4; + const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12)); + + const int32_t x0 = ((x[i].qs[j] & 0x0F) | xh_0) - 16; + const int32_t x1 = ((x[i].qs[j] >> 4) | xh_1) - 16; + + sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]); + } + + sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d)) * sumi; + } + + *s = sumf; +#endif +} + +void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + const int qk = QK8_1; + const int nb = n / qk; + + assert(n % qk == 0); + assert(qk == QK5_1); + + const block_q5_1 * restrict x = vx; + const block_q8_1 * restrict y = vy; + +#if defined(__ARM_NEON) + float32x4_t sumv0 = vdupq_n_f32(0.0f); + float32x4_t sumv1 = vdupq_n_f32(0.0f); + + float summs0 = 0.0f; + float summs1 = 0.0f; + + uint32_t qh0; + uint32_t qh1; + + uint64_t tmp0[4]; + uint64_t tmp1[4]; + + assert(nb % 2 == 0); // TODO: handle odd nb + + for (int i = 0; i < nb; i += 2) { + const block_q5_1 * restrict x0 = &x[i]; + const block_q5_1 * restrict x1 = &x[i + 1]; + const block_q8_1 * restrict y0 = &y[i]; + const block_q8_1 * restrict y1 = &y[i + 1]; + + const uint8x16_t m4b = vdupq_n_u8(0x0F); + + summs0 += GGML_FP16_TO_FP32(x0->m) * y0->s; + summs1 += GGML_FP16_TO_FP32(x1->m) * y1->s; + + // extract the 5th bit via lookup table ((b) << 4) + memcpy(&qh0, x0->qh, sizeof(qh0)); + memcpy(&qh1, x1->qh, sizeof(qh1)); + + tmp0[0] = table_b2b_0[(qh0 >> 0) & 0xFF]; + tmp0[1] = table_b2b_0[(qh0 >> 8) & 0xFF]; + tmp0[2] = table_b2b_0[(qh0 >> 16) & 0xFF]; + tmp0[3] = table_b2b_0[(qh0 >> 24) ]; + + tmp1[0] = table_b2b_0[(qh1 >> 0) & 0xFF]; + tmp1[1] = table_b2b_0[(qh1 >> 8) & 0xFF]; + tmp1[2] = table_b2b_0[(qh1 >> 16) & 0xFF]; + tmp1[3] = table_b2b_0[(qh1 >> 24) ]; + + const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0)); + const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2)); + const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0)); + const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2)); + + const uint8x16_t v0_0 = vld1q_u8(x0->qs); + const uint8x16_t v0_1 = vld1q_u8(x1->qs); + + // 4-bit -> 8-bit + const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); + const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); + const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); + const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); + + // add high bit + const int8x16_t v0_0lf = vorrq_s8(v0_0l, qhl0); + const int8x16_t v0_0hf = vorrq_s8(v0_0h, qhh0); + const int8x16_t v0_1lf = vorrq_s8(v0_1l, qhl1); + const int8x16_t v0_1hf = vorrq_s8(v0_1h, qhh1); + + // load y + const int8x16_t v1_0l = vld1q_s8(y0->qs); + const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); + const int8x16_t v1_1l = vld1q_s8(y1->qs); + const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); + +#if defined(__ARM_FEATURE_DOTPROD) + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l), + vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l), + vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*y1->d); +#else + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lf), vget_low_s8 (v1_0l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lf), vget_high_s8(v1_0l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hf), vget_low_s8 (v1_0h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hf), vget_high_s8(v1_0h)); + + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lf), vget_low_s8 (v1_1l)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lf), vget_high_s8(v1_1l)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hf), vget_low_s8 (v1_1h)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hf), vget_high_s8(v1_1h)); + + const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); + const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); + const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); + const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); + + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*y1->d); +#endif + } + + *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1; +#elif defined(__wasm_simd128__) + v128_t sumv = wasm_f32x4_splat(0.0f); + + float summs = 0.0f; + + uint32_t qh; + uint64_t tmp[4]; + + // TODO: check if unrolling this is better + for (int i = 0; i < nb; ++i) { + const block_q5_1 * restrict x0 = &x[i]; + const block_q8_1 * restrict y0 = &y[i]; + + summs += GGML_FP16_TO_FP32(x0->m) * y0->s; + + const v128_t m4b = wasm_i8x16_splat(0x0F); + + // extract the 5th bit + memcpy(&qh, x0->qh, sizeof(qh)); + + tmp[0] = table_b2b_0[(qh >> 0) & 0xFF]; + tmp[1] = table_b2b_0[(qh >> 8) & 0xFF]; + tmp[2] = table_b2b_0[(qh >> 16) & 0xFF]; + tmp[3] = table_b2b_0[(qh >> 24) ]; + + const v128_t qhl = wasm_v128_load(tmp + 0); + const v128_t qhh = wasm_v128_load(tmp + 2); + + const v128_t v0 = wasm_v128_load(x0->qs); + + // 4-bit -> 8-bit + const v128_t v0l = wasm_v128_and (v0, m4b); + const v128_t v0h = wasm_u8x16_shr(v0, 4); + + // add high bit + const v128_t v0lf = wasm_v128_or(v0l, qhl); + const v128_t v0hf = wasm_v128_or(v0h, qhh); + + // load y + const v128_t v1l = wasm_v128_load(y0->qs); + const v128_t v1h = wasm_v128_load(y0->qs + 16); + + // int8x16 -> int16x8 + const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf); + const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf); + const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf); + const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf); + + const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l); + const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l); + const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h); + const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h); + + // dot product + sumv = wasm_f32x4_add(sumv, + wasm_f32x4_mul(wasm_f32x4_convert_i32x4(wasm_i32x4_add( + wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll), + wasm_i32x4_dot_i16x8(v0lfh, v1lh)), + wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl), + wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), + wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * y0->d))); + } + + *s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) + + wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3) + summs; +#elif defined(__AVX2__) + // Initialize accumulator with zeros + __m256 acc = _mm256_setzero_ps(); + + float summs = 0.0f; + + // Main loop + for (int i = 0; i < nb; i++) { + const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d)); + + summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s; + + __m256i bx = bytes_from_nibbles_32(x[i].qs); + __m256i bxhi = bytes_from_bits_32(x[i].qh); + bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10)); + bx = _mm256_or_si256(bx, bxhi); + + const __m256 dy = _mm256_set1_ps(y[i].d); + const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); + + const __m256 q = mul_sum_us8_pairs_float(bx, by); + + acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc); + } + + *s = hsum_float_8(acc) + summs; +#elif defined(__AVX__) + // Initialize accumulator with zeros + __m256 acc = _mm256_setzero_ps(); + __m128i mask = _mm_set1_epi8(0x10); + + float summs = 0.0f; + + // Main loop + for (int i = 0; i < nb; i++) { + const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d)); + + summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s; + + __m256i bx = bytes_from_nibbles_32(x[i].qs); + const __m256i bxhi = bytes_from_bits_32(x[i].qh); + __m128i bxhil = _mm256_castsi256_si128(bxhi); + __m128i bxhih = _mm256_extractf128_si256(bxhi, 1); + bxhil = _mm_and_si128(bxhil, mask); + bxhih = _mm_and_si128(bxhih, mask); + __m128i bxl = _mm256_castsi256_si128(bx); + __m128i bxh = _mm256_extractf128_si256(bx, 1); + bxl = _mm_or_si128(bxl, bxhil); + bxh = _mm_or_si128(bxh, bxhih); + bx = MM256_SET_M128I(bxh, bxl); + + const __m256 dy = _mm256_set1_ps(y[i].d); + const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); + + const __m256 q = mul_sum_us8_pairs_float(bx, by); + + acc = _mm256_add_ps(_mm256_mul_ps(q, _mm256_mul_ps(dx, dy)), acc); + } + + *s = hsum_float_8(acc) + summs; +#elif defined(__riscv_v_intrinsic) + float sumf = 0.0; + + uint32_t qh; + + size_t vl = __riscv_vsetvl_e8m1(qk/2); + + // temporary registers for shift operations + vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl); + vuint32m2_t vt_2 = __riscv_vadd_vx_u32m2(vt_1, 12, vl); + + for (int i = 0; i < nb; i++) { + memcpy(&qh, x[i].qh, sizeof(uint32_t)); + + // load qh + vuint32m2_t vqh = __riscv_vmv_v_x_u32m2(qh, vl); + + // ((qh >> (j + 0)) << 4) & 0x10; + vuint32m2_t xhr_0 = __riscv_vsrl_vv_u32m2(vqh, vt_1, vl); + vuint32m2_t xhl_0 = __riscv_vsll_vx_u32m2(xhr_0, 4, vl); + vuint32m2_t xha_0 = __riscv_vand_vx_u32m2(xhl_0, 0x10, vl); + + // ((qh >> (j + 12)) ) & 0x10; + vuint32m2_t xhr_1 = __riscv_vsrl_vv_u32m2(vqh, vt_2, vl); + vuint32m2_t xha_1 = __riscv_vand_vx_u32m2(xhr_1, 0x10, vl); + + // narrowing + vuint16m1_t xhc_0 = __riscv_vncvt_x_x_w_u16m1(xha_0, vl); + vuint8mf2_t xh_0 = __riscv_vncvt_x_x_w_u8mf2(xhc_0, vl); + + vuint16m1_t xhc_1 = __riscv_vncvt_x_x_w_u16m1(xha_1, vl); + vuint8mf2_t xh_1 = __riscv_vncvt_x_x_w_u8mf2(xhc_1, vl); + + // load + vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl); + + vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl); + vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl); + + vuint8mf2_t x_at = __riscv_vand_vx_u8mf2(tx, 0x0F, vl); + vuint8mf2_t x_lt = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl); + + vuint8mf2_t x_a = __riscv_vor_vv_u8mf2(x_at, xh_0, vl); + vuint8mf2_t x_l = __riscv_vor_vv_u8mf2(x_lt, xh_1, vl); + + vint8mf2_t v0 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a); + vint8mf2_t v1 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l); + + vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl); + vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl); + + vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl); + + vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl); + vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl); + + int sumi = __riscv_vmv_x_s_i32m1_i32(vs2); + + sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s; + } + + *s = sumf; +#else + // scalar + float sumf = 0.0; + + for (int i = 0; i < nb; i++) { + uint32_t qh; + memcpy(&qh, x[i].qh, sizeof(qh)); + + int sumi = 0; + + for (int j = 0; j < qk/2; ++j) { + const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; + + const int32_t x0 = (x[i].qs[j] & 0xF) | xh_0; + const int32_t x1 = (x[i].qs[j] >> 4) | xh_1; + + sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]); + } + + sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s; + } + + *s = sumf; +#endif +} + +void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + const int qk = QK8_0; + const int nb = n / qk; + + assert(n % qk == 0); + + const block_q8_0 * restrict x = vx; + const block_q8_0 * restrict y = vy; + +#if defined(__ARM_NEON) + float32x4_t sumv0 = vdupq_n_f32(0.0f); + float32x4_t sumv1 = vdupq_n_f32(0.0f); + + assert(nb % 2 == 0); // TODO: handle odd nb + + for (int i = 0; i < nb; i += 2) { + const block_q8_0 * restrict x0 = &x[i + 0]; + const block_q8_0 * restrict x1 = &x[i + 1]; + const block_q8_0 * restrict y0 = &y[i + 0]; + const block_q8_0 * restrict y1 = &y[i + 1]; + + const int8x16_t x0_0 = vld1q_s8(x0->qs); + const int8x16_t x0_1 = vld1q_s8(x0->qs + 16); + const int8x16_t x1_0 = vld1q_s8(x1->qs); + const int8x16_t x1_1 = vld1q_s8(x1->qs + 16); + + // load y + const int8x16_t y0_0 = vld1q_s8(y0->qs); + const int8x16_t y0_1 = vld1q_s8(y0->qs + 16); + const int8x16_t y1_0 = vld1q_s8(y1->qs); + const int8x16_t y1_1 = vld1q_s8(y1->qs + 16); + +#if defined(__ARM_FEATURE_DOTPROD) + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), x0_0, y0_0), + vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), x1_0, y1_0), + vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); + +#else + const int16x8_t p0_0 = vmull_s8(vget_low_s8 (x0_0), vget_low_s8 (y0_0)); + const int16x8_t p0_1 = vmull_s8(vget_high_s8(x0_0), vget_high_s8(y0_0)); + const int16x8_t p0_2 = vmull_s8(vget_low_s8 (x0_1), vget_low_s8 (y0_1)); + const int16x8_t p0_3 = vmull_s8(vget_high_s8(x0_1), vget_high_s8(y0_1)); + + const int16x8_t p1_0 = vmull_s8(vget_low_s8 (x1_0), vget_low_s8 (y1_0)); + const int16x8_t p1_1 = vmull_s8(vget_high_s8(x1_0), vget_high_s8(y1_0)); + const int16x8_t p1_2 = vmull_s8(vget_low_s8 (x1_1), vget_low_s8 (y1_1)); + const int16x8_t p1_3 = vmull_s8(vget_high_s8(x1_1), vget_high_s8(y1_1)); + + const int32x4_t p0 = vaddq_s32(vpaddlq_s16(p0_0), vpaddlq_s16(p0_1)); + const int32x4_t p1 = vaddq_s32(vpaddlq_s16(p0_2), vpaddlq_s16(p0_3)); + const int32x4_t p2 = vaddq_s32(vpaddlq_s16(p1_0), vpaddlq_s16(p1_1)); + const int32x4_t p3 = vaddq_s32(vpaddlq_s16(p1_2), vpaddlq_s16(p1_3)); + + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(p0, p1)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(p2, p3)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); +#endif + } + + *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); +#elif defined(__AVX2__) || defined(__AVX__) + // Initialize accumulator with zeros + __m256 acc = _mm256_setzero_ps(); + + // Main loop + for (int i = 0; i < nb; ++i) { + // Compute combined scale for the block + const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d)); + __m256i bx = _mm256_loadu_si256((const __m256i *)x[i].qs); + __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); + + const __m256 q = mul_sum_i8_pairs_float(bx, by); + + // Multiply q with scale and accumulate +#if defined(__AVX2__) + acc = _mm256_fmadd_ps( d, q, acc ); +#else + acc = _mm256_add_ps( _mm256_mul_ps( d, q ), acc ); +#endif + } + + *s = hsum_float_8(acc); +#elif defined(__riscv_v_intrinsic) + float sumf = 0.0; + size_t vl = __riscv_vsetvl_e8m1(qk); + + for (int i = 0; i < nb; i++) { + // load elements + vint8m1_t bx = __riscv_vle8_v_i8m1(x[i].qs, vl); + vint8m1_t by = __riscv_vle8_v_i8m1(y[i].qs, vl); + + vint16m2_t vw_mul = __riscv_vwmul_vv_i16m2(bx, by, vl); + + vint32m1_t v_zero = __riscv_vmv_v_x_i32m1(0, vl); + vint32m1_t v_sum = __riscv_vwredsum_vs_i16m2_i32m1(vw_mul, v_zero, vl); + + int sumi = __riscv_vmv_x_s_i32m1_i32(v_sum); + + sumf += sumi*(GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d)); + } + + *s = sumf; +#else + // scalar + float sumf = 0.0; + + for (int i = 0; i < nb; i++) { + int sumi = 0; + + for (int j = 0; j < qk; j++) { + sumi += x[i].qs[j]*y[i].qs[j]; + } + + sumf += sumi*(GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d)); + } + + *s = sumf; +#endif +} + #if QK_K == 256 void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { @@ -1334,8 +3564,8 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); - const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const uint8_t * restrict q2 = x[i].qs; const int8_t * restrict q8 = y[i].qs; @@ -1413,8 +3643,8 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); - const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const uint8_t * restrict q2 = x[i].qs; const int8_t * restrict q8 = y[i].qs; @@ -1480,8 +3710,8 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float dall = y[i].d * ggml_fp16_to_fp32(x[i].d); - const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin); + const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const uint8_t * restrict q2 = x[i].qs; const int8_t * restrict q8 = y[i].qs; @@ -1588,8 +3818,8 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri const int8_t * q8 = y[i].qs; const uint8_t * sc = x[i].scales; - const float dall = y[i].d * ggml_fp16_to_fp32(x[i].d); - const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin); + const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); size_t vl = 16; @@ -1675,8 +3905,8 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri summs += y[i].bsums[j] * (sc[j] >> 4); } - const float dall = y[i].d * ggml_fp16_to_fp32(x[i].d); - const float dmin = y[i].d * ggml_fp16_to_fp32(x[i].dmin); + const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); int isum = 0; int is = 0; @@ -1793,8 +4023,8 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); - const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const uint8_t * restrict q2 = x[i].qs; const int8_t * restrict q8 = y[i].qs; @@ -1845,8 +4075,8 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); - const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const uint8_t * restrict q2 = x[i].qs; const int8_t * restrict q8 = y[i].qs; @@ -1960,8 +4190,8 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri summs += y[i].bsums[j] * (sc[j] >> 4); } - const float dall = y[i].d * ggml_fp16_to_fp32(x[i].d); - const float dmin = y[i].d * ggml_fp16_to_fp32(x[i].dmin); + const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); isum[0] = isum[1] = isum[2] = isum[3] = 0; for (int l = 0; l < 16; ++l) { @@ -2014,7 +4244,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const uint8_t * restrict q3 = x[i].qs; const uint8_t * restrict qh = x[i].hmask; @@ -2122,7 +4352,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const uint8_t * restrict q3 = x[i].qs; const int8_t * restrict q8 = y[i].qs; @@ -2227,7 +4457,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const uint8_t * restrict q3 = x[i].qs; const int8_t * restrict q8 = y[i].qs; @@ -2448,7 +4678,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri } - const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d; + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; sumf += d*sum_t; @@ -2513,7 +4743,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; q8 += 8; a += 8; } - const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d; + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -2615,7 +4845,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const uint8_t * restrict q3 = x[i].qs; const int8_t * restrict q8 = y[i].qs; @@ -2686,7 +4916,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const uint8_t * restrict q3 = x[i].qs; const int8_t * restrict q8 = y[i].qs; @@ -2871,7 +5101,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri q8 += 8; a += 8; for (int l = 0; l < 8; ++l) aux32[l] += scales[j] * aux16[l]; } - const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d; + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -2911,8 +5141,8 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); - const float dmin = y[i].d * ggml_fp16_to_fp32(x[i].dmin); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8)); @@ -2994,8 +5224,8 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); - const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); memcpy(utmp, x[i].scales, 12); utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); @@ -3060,8 +5290,8 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); - const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const uint8_t * restrict q4 = x[i].qs; const int8_t * restrict q8 = y[i].qs; @@ -3143,8 +5373,8 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri size_t vl = 8; - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); - const float dmin = y[i].d * ggml_fp16_to_fp32(x[i].dmin); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl); vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl); @@ -3254,9 +5484,9 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d; + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = ggml_fp16_to_fp32(x[i].dmin) * y[i].d; + const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; sumf -= dmin * sumi; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -3358,8 +5588,8 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = ggml_fp16_to_fp32(x[i].d[0]) * y[i].d; - const float m = ggml_fp16_to_fp32(x[i].d[1]) * y[i].d; + const float d = GGML_FP16_TO_FP32(x[i].d[0]) * y[i].d; + const float m = GGML_FP16_TO_FP32(x[i].d[1]) * y[i].d; const __m256 vd = _mm256_set1_ps(d); const uint16_t * a = (const uint16_t *)x[i].scales; @@ -3404,8 +5634,8 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = ggml_fp16_to_fp32(x[i].d[0]) * y[i].d; - const float m = ggml_fp16_to_fp32(x[i].d[1]) * y[i].d; + const float d = GGML_FP16_TO_FP32(x[i].d[0]) * y[i].d; + const float m = GGML_FP16_TO_FP32(x[i].d[1]) * y[i].d; const __m256 vd = _mm256_set1_ps(d); const uint16_t * a = (const uint16_t *)x[i].scales; @@ -3461,8 +5691,8 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri s16[0] = b[0] & 0x0f0f; s16[1] = (b[0] >> 4) & 0x0f0f; - sumf -= y[i].d * ggml_fp16_to_fp32(x[i].d[1]) * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3])); - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d[0]); + sumf -= y[i].d * GGML_FP16_TO_FP32(x[i].d[1]) * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3])); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d[0]); size_t vl = 32; @@ -3511,9 +5741,9 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri s16[0] = b[0] & 0x0f0f; s16[1] = (b[0] >> 4) & 0x0f0f; - sumf -= y[i].d * ggml_fp16_to_fp32(x[i].d[1]) * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3])); + sumf -= y[i].d * GGML_FP16_TO_FP32(x[i].d[1]) * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3])); - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d[0]); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d[0]); for (int j = 0; j < QK_K/32; ++j) { for (int l = 0; l < 16; ++l) aux16[l] = q8[l] * a[l]; @@ -3561,8 +5791,8 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); - const float dmin = y[i].d * ggml_fp16_to_fp32(x[i].dmin); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8)); @@ -3650,8 +5880,8 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri const int8_t * restrict q8 = y[i].qs; #if QK_K == 256 - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); - const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); memcpy(utmp, x[i].scales, 12); utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); @@ -3732,8 +5962,8 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); - const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const uint8_t * restrict q5 = x[i].qs; const int8_t * restrict q8 = y[i].qs; @@ -3837,8 +6067,8 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri const uint8_t * restrict hm = x[i].qh; const int8_t * restrict q8 = y[i].qs; - const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d; - const float dmin = ggml_fp16_to_fp32(x[i].dmin) * y[i].d; + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl); vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl); @@ -3960,9 +6190,9 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d; + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = ggml_fp16_to_fp32(x[i].dmin) * y[i].d; + const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; sumf -= dmin * sumi; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -4060,7 +6290,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri const uint8_t * restrict q5 = x[i].qs; const int8_t * restrict q8 = y[i].qs; - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5); @@ -4106,7 +6336,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri const uint8_t * restrict q5 = x[i].qs; const int8_t * restrict q8 = y[i].qs; - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5); @@ -4243,7 +6473,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri for (int l = 0; l < 8; ++l) a[8*is + l] -= (hm[l] & m ? 0 : 16); } - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const int8_t * restrict sc = x[i].scales; for (int j = 0; j < QK_K/16; ++j) { @@ -4286,7 +6516,7 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d_all = ggml_fp16_to_fp32(x[i].d); + const float d_all = GGML_FP16_TO_FP32(x[i].d); const uint8_t * restrict q6 = x[i].ql; const uint8_t * restrict qh = x[i].qh; @@ -4418,7 +6648,7 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const uint8_t * restrict q4 = x[i].ql; const uint8_t * restrict qh = x[i].qh; @@ -4498,7 +6728,7 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const uint8_t * restrict q4 = x[i].ql; const uint8_t * restrict qh = x[i].qh; @@ -4610,7 +6840,7 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri float sumf = 0; for (int i = 0; i < nb; ++i) { - const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d; + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * restrict q6 = x[i].ql; const uint8_t * restrict qh = x[i].qh; @@ -4727,7 +6957,7 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d; + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -4825,7 +7055,7 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const uint8_t * restrict q4 = x[i].ql; const uint8_t * restrict qh = x[i].qh; @@ -4882,7 +7112,7 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const uint8_t * restrict q4 = x[i].ql; const uint8_t * restrict qh = x[i].qh; @@ -5041,7 +7271,7 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d; + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; } for (int l = 0; l < 8; ++l) sumf += sums[l]; diff --git a/k_quants.h b/ggml-quants.h similarity index 63% rename from k_quants.h rename to ggml-quants.h index 9de089e7a..70c12c274 100644 --- a/k_quants.h +++ b/ggml-quants.h @@ -1,11 +1,63 @@ #pragma once -#include "ggml.h" +#include "ggml-impl.h" + +// GGML internal header #include -#include #include +#define QK4_0 32 +typedef struct { + ggml_fp16_t d; // delta + uint8_t qs[QK4_0 / 2]; // nibbles / quants +} block_q4_0; +static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding"); + +#define QK4_1 32 +typedef struct { + ggml_fp16_t d; // delta + ggml_fp16_t m; // min + uint8_t qs[QK4_1 / 2]; // nibbles / quants +} block_q4_1; +static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_fp16_t) + QK4_1 / 2, "wrong q4_1 block size/padding"); + +#define QK5_0 32 +typedef struct { + ggml_fp16_t d; // delta + uint8_t qh[4]; // 5-th bit of quants + uint8_t qs[QK5_0 / 2]; // nibbles / quants +} block_q5_0; +static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding"); + +#define QK5_1 32 +typedef struct { + ggml_fp16_t d; // delta + ggml_fp16_t m; // min + uint8_t qh[4]; // 5-th bit of quants + uint8_t qs[QK5_1 / 2]; // nibbles / quants +} block_q5_1; +static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding"); + +#define QK8_0 32 +typedef struct { + ggml_fp16_t d; // delta + int8_t qs[QK8_0]; // quants +} block_q8_0; +static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding"); + +#define QK8_1 32 +typedef struct { + float d; // delta + float s; // d * sum(qs[i]) + int8_t qs[QK8_1]; // quants +} block_q8_1; +static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding"); + +// +// Super-block quantization structures +// + // Super-block size #ifdef GGML_QKK_64 #define QK_K 64 @@ -15,18 +67,6 @@ #define K_SCALE_SIZE 12 #endif -#ifndef static_assert -#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L) -#define static_assert(cond, msg) _Static_assert(cond, msg) -#else -#define static_assert(cond, msg) struct global_scope_noop_trick -#endif -#endif - -// -// Super-block quantization structures -// - // 2-bit quantization // weight is represented as x = a * q + b // 16 blocks of 16 elements each @@ -127,6 +167,13 @@ static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_ // Quantization +void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k); +void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k); +void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k); +void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k); +void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k); +void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int k); + void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k); void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k); void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k); @@ -134,6 +181,13 @@ void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k); void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k); +void quantize_row_q4_0(const float * restrict x, void * restrict y, int k); +void quantize_row_q4_1(const float * restrict x, void * restrict y, int k); +void quantize_row_q5_0(const float * restrict x, void * restrict y, int k); +void quantize_row_q5_1(const float * restrict x, void * restrict y, int k); +void quantize_row_q8_0(const float * restrict x, void * restrict y, int k); +void quantize_row_q8_1(const float * restrict x, void * restrict y, int k); + void quantize_row_q2_K(const float * restrict x, void * restrict y, int k); void quantize_row_q3_K(const float * restrict x, void * restrict y, int k); void quantize_row_q4_K(const float * restrict x, void * restrict y, int k); @@ -142,6 +196,13 @@ void quantize_row_q6_K(const float * restrict x, void * restrict y, int k); void quantize_row_q8_K(const float * restrict x, void * restrict y, int k); // Dequantization +void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k); +void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int k); +void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int k); +void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int k); +void dequantize_row_q8_0(const block_q8_0 * restrict x, float * restrict y, int k); +//void dequantize_row_q8_1(const block_q8_1 * restrict x, float * restrict y, int k); + void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k); void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k); void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k); @@ -150,16 +211,14 @@ void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k); // Dot product +void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy); +void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy); +void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy); +void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy); +void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy); + void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy); void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy); void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy); void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy); void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy); - -// Quantization with histogram collection -size_t ggml_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist); -size_t ggml_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist); -size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist); -size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist); -size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist); - diff --git a/ggml.c b/ggml.c index 49f3b7aba..80d682255 100644 --- a/ggml.c +++ b/ggml.c @@ -1,10 +1,7 @@ #define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnigns on Windows -#include "ggml.h" - -#ifdef GGML_USE_K_QUANTS -#include "k_quants.h" -#endif +#include "ggml-impl.h" +#include "ggml-quants.h" #if defined(_MSC_VER) || defined(__MINGW32__) #include // using malloc.h with MSC/MINGW @@ -30,18 +27,6 @@ #include #endif -// static_assert should be a #define, but if it's not, -// fall back to the _Static_assert C11 keyword. -// if C99 - static_assert is noop -// ref: https://stackoverflow.com/a/53923785/4039976 -#ifndef static_assert -#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L) -#define static_assert(cond, msg) _Static_assert(cond, msg) -#else -#define static_assert(cond, msg) struct global_scope_noop_trick -#endif -#endif - #if defined(_MSC_VER) // disable "possible loss of data" to avoid hundreds of casts // we should just be careful :) @@ -109,23 +94,11 @@ typedef void * thread_ret_t; #include #endif + #ifdef GGML_USE_CPU_HBM #include #endif -// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512 -#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__)) -#ifndef __FMA__ -#define __FMA__ -#endif -#ifndef __F16C__ -#define __F16C__ -#endif -#ifndef __SSE3__ -#define __SSE3__ -#endif -#endif - /*#define GGML_PERF*/ #define GGML_DEBUG 0 #define GGML_GELU_FP16 @@ -251,228 +224,27 @@ inline static void * ggml_aligned_malloc(size_t size) { #include "ggml-opencl.h" #endif -#undef MIN -#undef MAX -#define MIN(a, b) ((a) < (b) ? (a) : (b)) -#define MAX(a, b) ((a) > (b) ? (a) : (b)) - // floating point type used to accumulate sums typedef double ggml_float; -// 16-bit float -// on Arm, we use __fp16 -// on x86, we use uint16_t -#if defined(__ARM_NEON) && !defined(_MSC_VER) - -// if YCM cannot find , make a symbolic link to it, for example: -// -// $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/ -// -#include - -#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x)) -#define GGML_COMPUTE_FP32_TO_FP16(x) (x) - -#define GGML_FP16_TO_FP32(x) ((float) (x)) -#define GGML_FP32_TO_FP16(x) (x) - -#else - -#ifdef __wasm_simd128__ -#include -#else -#ifdef __POWER9_VECTOR__ -#include -#undef bool -#define bool _Bool -#else -#if defined(_MSC_VER) || defined(__MINGW32__) -#include -#else -#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) -#if !defined(__riscv) -#include -#endif -#endif -#endif -#endif -#endif - -#ifdef __riscv_v_intrinsic -#include -#endif - -#ifdef __F16C__ - -#ifdef _MSC_VER -#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x))) -#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0) -#else -#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x) -#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0) -#endif - -#elif defined(__POWER9_VECTOR__) - -#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) -#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) -/* the inline asm below is about 12% faster than the lookup method */ -#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x) -#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) - -static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { - register float f; - register double d; - __asm__( - "mtfprd %0,%2\n" - "xscvhpdp %0,%0\n" - "frsp %1,%0\n" : - /* temp */ "=d"(d), - /* out */ "=f"(f): - /* in */ "r"(h)); - return f; -} - -static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { - register double d; - register ggml_fp16_t r; - __asm__( /* xscvdphp can work on double or single precision */ - "xscvdphp %0,%2\n" - "mffprd %1,%0\n" : - /* temp */ "=d"(d), - /* out */ "=r"(r): - /* in */ "f"(f)); - return r; -} - -#else - -// FP16 <-> FP32 -// ref: https://github.com/Maratyszcza/FP16 - -static inline float fp32_from_bits(uint32_t w) { - union { - uint32_t as_bits; - float as_value; - } fp32; - fp32.as_bits = w; - return fp32.as_value; -} - -static inline uint32_t fp32_to_bits(float f) { - union { - float as_value; - uint32_t as_bits; - } fp32; - fp32.as_value = f; - return fp32.as_bits; -} - -static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { - const uint32_t w = (uint32_t) h << 16; - const uint32_t sign = w & UINT32_C(0x80000000); - const uint32_t two_w = w + w; - - const uint32_t exp_offset = UINT32_C(0xE0) << 23; -#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__) - const float exp_scale = 0x1.0p-112f; -#else - const float exp_scale = fp32_from_bits(UINT32_C(0x7800000)); -#endif - const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale; - - const uint32_t magic_mask = UINT32_C(126) << 23; - const float magic_bias = 0.5f; - const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias; - - const uint32_t denormalized_cutoff = UINT32_C(1) << 27; - const uint32_t result = sign | - (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value)); - return fp32_from_bits(result); -} - -static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { -#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__) - const float scale_to_inf = 0x1.0p+112f; - const float scale_to_zero = 0x1.0p-110f; -#else - const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000)); - const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000)); -#endif - float base = (fabsf(f) * scale_to_inf) * scale_to_zero; - - const uint32_t w = fp32_to_bits(f); - const uint32_t shl1_w = w + w; - const uint32_t sign = w & UINT32_C(0x80000000); - uint32_t bias = shl1_w & UINT32_C(0xFF000000); - if (bias < UINT32_C(0x71000000)) { - bias = UINT32_C(0x71000000); - } - - base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base; - const uint32_t bits = fp32_to_bits(base); - const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00); - const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF); - const uint32_t nonsign = exp_bits + mantissa_bits; - return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign); -} - -#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) -#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) - -#endif // __F16C__ - -#endif // __ARM_NEON - // // global data // // precomputed gelu table for f16 (128 KB) -static ggml_fp16_t table_gelu_f16[1 << 16]; +static ggml_fp16_t ggml_table_gelu_f16[1 << 16]; // precomputed quick gelu table for f16 (128 KB) -static ggml_fp16_t table_gelu_quick_f16[1 << 16]; +static ggml_fp16_t ggml_table_gelu_quick_f16[1 << 16]; // precomputed silu table for f16 (128 KB) -static ggml_fp16_t table_silu_f16[1 << 16]; +static ggml_fp16_t ggml_table_silu_f16[1 << 16]; // precomputed exp table for f16 (128 KB) -static ggml_fp16_t table_exp_f16[1 << 16]; +static ggml_fp16_t ggml_table_exp_f16[1 << 16]; -// precomputed f32 table for f16 (256 KB) -static float table_f32_f16[1 << 16]; - -#if defined(__ARM_NEON) || defined(__wasm_simd128__) -#define B1(c,s,n) 0x ## n ## c , 0x ## n ## s -#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s) -#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s) -#define B4(c,s,n) B3(c,s,n ## c), B3(c,s,n ## s) -#define B5(c,s,n) B4(c,s,n ## c), B4(c,s,n ## s) -#define B6(c,s,n) B5(c,s,n ## c), B5(c,s,n ## s) -#define B7(c,s,n) B6(c,s,n ## c), B6(c,s,n ## s) -#define B8(c,s ) B7(c,s, c), B7(c,s, s) - -// precomputed tables for expanding 8bits to 8 bytes: -static const uint64_t table_b2b_0[1 << 8] = { B8(00, 10) }; // ( b) << 4 -static const uint64_t table_b2b_1[1 << 8] = { B8(10, 00) }; // (!b) << 4 -#endif - -// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32, -// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON. -// This is also true for POWER9. -#if !defined(GGML_FP16_TO_FP32) || !defined(GGML_FP32_TO_FP16) - -inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) { - uint16_t s; - memcpy(&s, &f, sizeof(uint16_t)); - return table_f32_f16[s]; -} - -#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x) -#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) - -#endif +// precomputed f32 table for f16 (256 KB) (ggml-impl.h) +float ggml_table_f32_f16[1 << 16]; // note: do not use these inside ggml.c // these are meant to be used via the ggml.h API @@ -571,7 +343,6 @@ int64_t ggml_cycles_per_ms(void) { #define ggml_perf_cycles_per_ms() 0 #endif - // // cache line // @@ -588,1071 +359,8 @@ int64_t ggml_cycles_per_ms(void) { static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float); -// -// quantization -// - -#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1) - -#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) -// multiply int8_t, add results pairwise twice -static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) { - // Get absolute values of x vectors - const __m128i ax = _mm_sign_epi8(x, x); - // Sign the values of the y vectors - const __m128i sy = _mm_sign_epi8(y, x); - // Perform multiplication and create 16-bit values - const __m128i dot = _mm_maddubs_epi16(ax, sy); - const __m128i ones = _mm_set1_epi16(1); - return _mm_madd_epi16(ones, dot); -} - -#if __AVX__ || __AVX2__ || __AVX512F__ -// horizontally add 8 floats -static inline float hsum_float_8(const __m256 x) { - __m128 res = _mm256_extractf128_ps(x, 1); - res = _mm_add_ps(res, _mm256_castps256_ps128(x)); - res = _mm_add_ps(res, _mm_movehl_ps(res, res)); - res = _mm_add_ss(res, _mm_movehdup_ps(res)); - return _mm_cvtss_f32(res); -} - -// horizontally add 8 int32_t -static inline int hsum_i32_8(const __m256i a) { - const __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(a), _mm256_extractf128_si256(a, 1)); - const __m128i hi64 = _mm_unpackhi_epi64(sum128, sum128); - const __m128i sum64 = _mm_add_epi32(hi64, sum128); - const __m128i hi32 = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1)); - return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32)); -} - -// horizontally add 4 int32_t -static inline int hsum_i32_4(const __m128i a) { - const __m128i hi64 = _mm_unpackhi_epi64(a, a); - const __m128i sum64 = _mm_add_epi32(hi64, a); - const __m128i hi32 = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1)); - return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32)); -} - -#if defined(__AVX2__) || defined(__AVX512F__) -// spread 32 bits to 32 bytes { 0x00, 0xFF } -static inline __m256i bytes_from_bits_32(const uint8_t * x) { - uint32_t x32; - memcpy(&x32, x, sizeof(uint32_t)); - const __m256i shuf_mask = _mm256_set_epi64x( - 0x0303030303030303, 0x0202020202020202, - 0x0101010101010101, 0x0000000000000000); - __m256i bytes = _mm256_shuffle_epi8(_mm256_set1_epi32(x32), shuf_mask); - const __m256i bit_mask = _mm256_set1_epi64x(0x7fbfdfeff7fbfdfe); - bytes = _mm256_or_si256(bytes, bit_mask); - return _mm256_cmpeq_epi8(bytes, _mm256_set1_epi64x(-1)); -} - -// Unpack 32 4-bit fields into 32 bytes -// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval -static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi) -{ - const __m128i tmp = _mm_loadu_si128((const __m128i *)rsi); - const __m256i bytes = MM256_SET_M128I(_mm_srli_epi16(tmp, 4), tmp); - const __m256i lowMask = _mm256_set1_epi8( 0xF ); - return _mm256_and_si256(lowMask, bytes); -} - -// add int16_t pairwise and return as float vector -static inline __m256 sum_i16_pairs_float(const __m256i x) { - const __m256i ones = _mm256_set1_epi16(1); - const __m256i summed_pairs = _mm256_madd_epi16(ones, x); - return _mm256_cvtepi32_ps(summed_pairs); -} - -static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) { -#if __AVXVNNI__ - const __m256i zero = _mm256_setzero_si256(); - const __m256i summed_pairs = _mm256_dpbusd_epi32(zero, ax, sy); - return _mm256_cvtepi32_ps(summed_pairs); -#else - // Perform multiplication and create 16-bit values - const __m256i dot = _mm256_maddubs_epi16(ax, sy); - return sum_i16_pairs_float(dot); -#endif -} - -// multiply int8_t, add results pairwise twice and return as float vector -static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) { -#if __AVXVNNIINT8__ - const __m256i zero = _mm256_setzero_si256(); - const __m256i summed_pairs = _mm256_dpbssd_epi32(zero, x, y); - return _mm256_cvtepi32_ps(summed_pairs); -#else - // Get absolute values of x vectors - const __m256i ax = _mm256_sign_epi8(x, x); - // Sign the values of the y vectors - const __m256i sy = _mm256_sign_epi8(y, x); - return mul_sum_us8_pairs_float(ax, sy); -#endif -} - -static inline __m128i packNibbles( __m256i bytes ) -{ - // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh -#if __AVX512F__ - const __m256i bytes_srli_4 = _mm256_srli_epi16(bytes, 4); // 0000_0000_abcd_0000 - bytes = _mm256_or_si256(bytes, bytes_srli_4); // 0000_abcd_abcd_efgh - return _mm256_cvtepi16_epi8(bytes); // abcd_efgh -#else - const __m256i lowByte = _mm256_set1_epi16( 0xFF ); - __m256i high = _mm256_andnot_si256( lowByte, bytes ); - __m256i low = _mm256_and_si256( lowByte, bytes ); - high = _mm256_srli_epi16( high, 4 ); - bytes = _mm256_or_si256( low, high ); - - // Compress uint16_t lanes into bytes - __m128i r0 = _mm256_castsi256_si128( bytes ); - __m128i r1 = _mm256_extracti128_si256( bytes, 1 ); - return _mm_packus_epi16( r0, r1 ); -#endif -} -#elif defined(__AVX__) -// spread 32 bits to 32 bytes { 0x00, 0xFF } -static inline __m256i bytes_from_bits_32(const uint8_t * x) { - uint32_t x32; - memcpy(&x32, x, sizeof(uint32_t)); - const __m128i shuf_maskl = _mm_set_epi64x(0x0101010101010101, 0x0000000000000000); - const __m128i shuf_maskh = _mm_set_epi64x(0x0303030303030303, 0x0202020202020202); - __m128i bytesl = _mm_shuffle_epi8(_mm_set1_epi32(x32), shuf_maskl); - __m128i bytesh = _mm_shuffle_epi8(_mm_set1_epi32(x32), shuf_maskh); - const __m128i bit_mask = _mm_set1_epi64x(0x7fbfdfeff7fbfdfe); - bytesl = _mm_or_si128(bytesl, bit_mask); - bytesh = _mm_or_si128(bytesh, bit_mask); - bytesl = _mm_cmpeq_epi8(bytesl, _mm_set1_epi64x(-1)); - bytesh = _mm_cmpeq_epi8(bytesh, _mm_set1_epi64x(-1)); - return MM256_SET_M128I(bytesh, bytesl); -} - -// Unpack 32 4-bit fields into 32 bytes -// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval -static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi) -{ - // Load 16 bytes from memory - __m128i tmpl = _mm_loadu_si128((const __m128i *)rsi); - __m128i tmph = _mm_srli_epi16(tmpl, 4); - const __m128i lowMask = _mm_set1_epi8(0xF); - tmpl = _mm_and_si128(lowMask, tmpl); - tmph = _mm_and_si128(lowMask, tmph); - return MM256_SET_M128I(tmph, tmpl); -} - -// add int16_t pairwise and return as float vector -static inline __m256 sum_i16_pairs_float(const __m128i xh, const __m128i xl) { - const __m128i ones = _mm_set1_epi16(1); - const __m128i summed_pairsl = _mm_madd_epi16(ones, xl); - const __m128i summed_pairsh = _mm_madd_epi16(ones, xh); - const __m256i summed_pairs = MM256_SET_M128I(summed_pairsh, summed_pairsl); - return _mm256_cvtepi32_ps(summed_pairs); -} - -static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) { - const __m128i axl = _mm256_castsi256_si128(ax); - const __m128i axh = _mm256_extractf128_si256(ax, 1); - const __m128i syl = _mm256_castsi256_si128(sy); - const __m128i syh = _mm256_extractf128_si256(sy, 1); - // Perform multiplication and create 16-bit values - const __m128i dotl = _mm_maddubs_epi16(axl, syl); - const __m128i doth = _mm_maddubs_epi16(axh, syh); - return sum_i16_pairs_float(doth, dotl); -} - -// multiply int8_t, add results pairwise twice and return as float vector -static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) { - const __m128i xl = _mm256_castsi256_si128(x); - const __m128i xh = _mm256_extractf128_si256(x, 1); - const __m128i yl = _mm256_castsi256_si128(y); - const __m128i yh = _mm256_extractf128_si256(y, 1); - // Get absolute values of x vectors - const __m128i axl = _mm_sign_epi8(xl, xl); - const __m128i axh = _mm_sign_epi8(xh, xh); - // Sign the values of the y vectors - const __m128i syl = _mm_sign_epi8(yl, xl); - const __m128i syh = _mm_sign_epi8(yh, xh); - // Perform multiplication and create 16-bit values - const __m128i dotl = _mm_maddubs_epi16(axl, syl); - const __m128i doth = _mm_maddubs_epi16(axh, syh); - return sum_i16_pairs_float(doth, dotl); -} - -static inline __m128i packNibbles( __m128i bytes1, __m128i bytes2 ) -{ - // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh - const __m128i lowByte = _mm_set1_epi16( 0xFF ); - __m128i high = _mm_andnot_si128( lowByte, bytes1 ); - __m128i low = _mm_and_si128( lowByte, bytes1 ); - high = _mm_srli_epi16( high, 4 ); - bytes1 = _mm_or_si128( low, high ); - high = _mm_andnot_si128( lowByte, bytes2 ); - low = _mm_and_si128( lowByte, bytes2 ); - high = _mm_srli_epi16( high, 4 ); - bytes2 = _mm_or_si128( low, high ); - - return _mm_packus_epi16( bytes1, bytes2); -} -#endif -#elif defined(__SSSE3__) -// horizontally add 4x4 floats -static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 c, const __m128 d) { - __m128 res_0 =_mm_hadd_ps(a, b); - __m128 res_1 =_mm_hadd_ps(c, d); - __m128 res =_mm_hadd_ps(res_0, res_1); - res =_mm_hadd_ps(res, res); - res =_mm_hadd_ps(res, res); - - return _mm_cvtss_f32(res); -} -#endif // __AVX__ || __AVX2__ || __AVX512F__ -#endif // defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) - -#if defined(__ARM_NEON) - -#if !defined(__aarch64__) - -inline static int32_t vaddvq_s32(int32x4_t v) { - return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3); -} - -inline static float vaddvq_f32(float32x4_t v) { - return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3); -} - -inline static float vmaxvq_f32(float32x4_t v) { - return - MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)), - MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3))); -} - -inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) { - int32x4_t res; - - res[0] = roundf(vgetq_lane_f32(v, 0)); - res[1] = roundf(vgetq_lane_f32(v, 1)); - res[2] = roundf(vgetq_lane_f32(v, 2)); - res[3] = roundf(vgetq_lane_f32(v, 3)); - - return res; -} - -#endif -#endif - -#define QK4_0 32 -typedef struct { - ggml_fp16_t d; // delta - uint8_t qs[QK4_0 / 2]; // nibbles / quants -} block_q4_0; -static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding"); - -#define QK4_1 32 -typedef struct { - ggml_fp16_t d; // delta - ggml_fp16_t m; // min - uint8_t qs[QK4_1 / 2]; // nibbles / quants -} block_q4_1; -static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_fp16_t) + QK4_1 / 2, "wrong q4_1 block size/padding"); - -#define QK5_0 32 -typedef struct { - ggml_fp16_t d; // delta - uint8_t qh[4]; // 5-th bit of quants - uint8_t qs[QK5_0 / 2]; // nibbles / quants -} block_q5_0; -static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding"); - -#define QK5_1 32 -typedef struct { - ggml_fp16_t d; // delta - ggml_fp16_t m; // min - uint8_t qh[4]; // 5-th bit of quants - uint8_t qs[QK5_1 / 2]; // nibbles / quants -} block_q5_1; -static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding"); - -#define QK8_0 32 -typedef struct { - ggml_fp16_t d; // delta - int8_t qs[QK8_0]; // quants -} block_q8_0; -static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding"); - -#define QK8_1 32 -typedef struct { - float d; // delta - float s; // d * sum(qs[i]) - int8_t qs[QK8_1]; // quants -} block_q8_1; -static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding"); - -// reference implementation for deterministic creation of model files -static void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k) { - static const int qk = QK4_0; - - assert(k % qk == 0); - - const int nb = k / qk; - - for (int i = 0; i < nb; i++) { - float amax = 0.0f; // absolute max - float max = 0.0f; - - for (int j = 0; j < qk; j++) { - const float v = x[i*qk + j]; - if (amax < fabsf(v)) { - amax = fabsf(v); - max = v; - } - } - - const float d = max / -8; - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = GGML_FP32_TO_FP16(d); - - for (int j = 0; j < qk/2; ++j) { - const float x0 = x[i*qk + 0 + j]*id; - const float x1 = x[i*qk + qk/2 + j]*id; - - const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f)); - const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f)); - - y[i].qs[j] = xi0; - y[i].qs[j] |= xi1 << 4; - } - } -} - -static void quantize_row_q4_0(const float * restrict x, void * restrict y, int k) { - quantize_row_q4_0_reference(x, y, k); -} - -static void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k) { - const int qk = QK4_1; - - assert(k % qk == 0); - - const int nb = k / qk; - - for (int i = 0; i < nb; i++) { - float min = FLT_MAX; - float max = -FLT_MAX; - - for (int j = 0; j < qk; j++) { - const float v = x[i*qk + j]; - - if (v < min) min = v; - if (v > max) max = v; - } - - const float d = (max - min) / ((1 << 4) - 1); - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = GGML_FP32_TO_FP16(d); - y[i].m = GGML_FP32_TO_FP16(min); - - for (int j = 0; j < qk/2; ++j) { - const float x0 = (x[i*qk + 0 + j] - min)*id; - const float x1 = (x[i*qk + qk/2 + j] - min)*id; - - const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f)); - const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f)); - - y[i].qs[j] = xi0; - y[i].qs[j] |= xi1 << 4; - } - } -} - -static void quantize_row_q4_1(const float * restrict x, void * restrict y, int k) { - quantize_row_q4_1_reference(x, y, k); -} - -static void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k) { - static const int qk = QK5_0; - - assert(k % qk == 0); - - const int nb = k / qk; - - for (int i = 0; i < nb; i++) { - float amax = 0.0f; // absolute max - float max = 0.0f; - - for (int j = 0; j < qk; j++) { - const float v = x[i*qk + j]; - if (amax < fabsf(v)) { - amax = fabsf(v); - max = v; - } - } - - const float d = max / -16; - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = GGML_FP32_TO_FP16(d); - - uint32_t qh = 0; - - for (int j = 0; j < qk/2; ++j) { - const float x0 = x[i*qk + 0 + j]*id; - const float x1 = x[i*qk + qk/2 + j]*id; - - const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f)); - const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f)); - - y[i].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4); - - // get the 5-th bit and store it in qh at the right position - qh |= ((xi0 & 0x10u) >> 4) << (j + 0); - qh |= ((xi1 & 0x10u) >> 4) << (j + qk/2); - } - - memcpy(&y[i].qh, &qh, sizeof(qh)); - } -} - -static void quantize_row_q5_0(const float * restrict x, void * restrict y, int k) { - quantize_row_q5_0_reference(x, y, k); -} - -static void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k) { - const int qk = QK5_1; - - assert(k % qk == 0); - - const int nb = k / qk; - - for (int i = 0; i < nb; i++) { - float min = FLT_MAX; - float max = -FLT_MAX; - - for (int j = 0; j < qk; j++) { - const float v = x[i*qk + j]; - - if (v < min) min = v; - if (v > max) max = v; - } - - const float d = (max - min) / ((1 << 5) - 1); - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = GGML_FP32_TO_FP16(d); - y[i].m = GGML_FP32_TO_FP16(min); - - uint32_t qh = 0; - - for (int j = 0; j < qk/2; ++j) { - const float x0 = (x[i*qk + 0 + j] - min)*id; - const float x1 = (x[i*qk + qk/2 + j] - min)*id; - - const uint8_t xi0 = (uint8_t)(x0 + 0.5f); - const uint8_t xi1 = (uint8_t)(x1 + 0.5f); - - y[i].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4); - - // get the 5-th bit and store it in qh at the right position - qh |= ((xi0 & 0x10u) >> 4) << (j + 0); - qh |= ((xi1 & 0x10u) >> 4) << (j + qk/2); - } - - memcpy(&y[i].qh, &qh, sizeof(y[i].qh)); - } -} - -static void quantize_row_q5_1(const float * restrict x, void * restrict y, int k) { - quantize_row_q5_1_reference(x, y, k); -} - -// reference implementation for deterministic creation of model files -static void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k) { - assert(k % QK8_0 == 0); - const int nb = k / QK8_0; - - for (int i = 0; i < nb; i++) { - float amax = 0.0f; // absolute max - - for (int j = 0; j < QK8_0; j++) { - const float v = x[i*QK8_0 + j]; - amax = MAX(amax, fabsf(v)); - } - - const float d = amax / ((1 << 7) - 1); - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = GGML_FP32_TO_FP16(d); - - for (int j = 0; j < QK8_0; ++j) { - const float x0 = x[i*QK8_0 + j]*id; - - y[i].qs[j] = roundf(x0); - } - } -} - -static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int k) { - assert(QK8_0 == 32); - assert(k % QK8_0 == 0); - const int nb = k / QK8_0; - - block_q8_0 * restrict y = vy; - -#if defined(__ARM_NEON) - for (int i = 0; i < nb; i++) { - float32x4_t srcv [8]; - float32x4_t asrcv[8]; - float32x4_t amaxv[8]; - - for (int j = 0; j < 8; j++) srcv[j] = vld1q_f32(x + i*32 + 4*j); - for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]); - - for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]); - for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]); - for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]); - - const float amax = vmaxvq_f32(amaxv[0]); - - const float d = amax / ((1 << 7) - 1); - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = GGML_FP32_TO_FP16(d); - - for (int j = 0; j < 8; j++) { - const float32x4_t v = vmulq_n_f32(srcv[j], id); - const int32x4_t vi = vcvtnq_s32_f32(v); - - y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0); - y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1); - y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2); - y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3); - } - } -#elif defined(__wasm_simd128__) - for (int i = 0; i < nb; i++) { - v128_t srcv [8]; - v128_t asrcv[8]; - v128_t amaxv[8]; - - for (int j = 0; j < 8; j++) srcv[j] = wasm_v128_load(x + i*32 + 4*j); - for (int j = 0; j < 8; j++) asrcv[j] = wasm_f32x4_abs(srcv[j]); - - for (int j = 0; j < 4; j++) amaxv[2*j] = wasm_f32x4_max(asrcv[2*j], asrcv[2*j+1]); - for (int j = 0; j < 2; j++) amaxv[4*j] = wasm_f32x4_max(amaxv[4*j], amaxv[4*j+2]); - for (int j = 0; j < 1; j++) amaxv[8*j] = wasm_f32x4_max(amaxv[8*j], amaxv[8*j+4]); - - const float amax = MAX(MAX(wasm_f32x4_extract_lane(amaxv[0], 0), - wasm_f32x4_extract_lane(amaxv[0], 1)), - MAX(wasm_f32x4_extract_lane(amaxv[0], 2), - wasm_f32x4_extract_lane(amaxv[0], 3))); - - const float d = amax / ((1 << 7) - 1); - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = GGML_FP32_TO_FP16(d); - - for (int j = 0; j < 8; j++) { - const v128_t v = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id)); - const v128_t vi = wasm_i32x4_trunc_sat_f32x4(v); - - y[i].qs[4*j + 0] = wasm_i32x4_extract_lane(vi, 0); - y[i].qs[4*j + 1] = wasm_i32x4_extract_lane(vi, 1); - y[i].qs[4*j + 2] = wasm_i32x4_extract_lane(vi, 2); - y[i].qs[4*j + 3] = wasm_i32x4_extract_lane(vi, 3); - } - } -#elif defined(__AVX2__) || defined(__AVX__) - for (int i = 0; i < nb; i++) { - // Load elements into 4 AVX vectors - __m256 v0 = _mm256_loadu_ps( x ); - __m256 v1 = _mm256_loadu_ps( x + 8 ); - __m256 v2 = _mm256_loadu_ps( x + 16 ); - __m256 v3 = _mm256_loadu_ps( x + 24 ); - x += 32; - - // Compute max(abs(e)) for the block - const __m256 signBit = _mm256_set1_ps( -0.0f ); - __m256 maxAbs = _mm256_andnot_ps( signBit, v0 ); - maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) ); - maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) ); - maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) ); - - __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) ); - max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) ); - max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) ); - const float maxScalar = _mm_cvtss_f32( max4 ); - - // Quantize these floats - const float d = maxScalar / 127.f; - y[i].d = GGML_FP32_TO_FP16(d); - const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; - const __m256 mul = _mm256_set1_ps( id ); - - // Apply the multiplier - v0 = _mm256_mul_ps( v0, mul ); - v1 = _mm256_mul_ps( v1, mul ); - v2 = _mm256_mul_ps( v2, mul ); - v3 = _mm256_mul_ps( v3, mul ); - - // Round to nearest integer - v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST ); - v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST ); - v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST ); - v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST ); - - // Convert floats to integers - __m256i i0 = _mm256_cvtps_epi32( v0 ); - __m256i i1 = _mm256_cvtps_epi32( v1 ); - __m256i i2 = _mm256_cvtps_epi32( v2 ); - __m256i i3 = _mm256_cvtps_epi32( v3 ); - -#if defined(__AVX2__) - // Convert int32 to int16 - i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 - i2 = _mm256_packs_epi32( i2, i3 ); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31 - // Convert int16 to int8 - i0 = _mm256_packs_epi16( i0, i2 ); // 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27, 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31 - - // We got our precious signed bytes, but the order is now wrong - // These AVX2 pack instructions process 16-byte pieces independently - // The following instruction is fixing the order - const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 ); - i0 = _mm256_permutevar8x32_epi32( i0, perm ); - - _mm256_storeu_si256((__m256i *)y[i].qs, i0); -#else - // Since we don't have in AVX some necessary functions, - // we split the registers in half and call AVX2 analogs from SSE - __m128i ni0 = _mm256_castsi256_si128( i0 ); - __m128i ni1 = _mm256_extractf128_si256( i0, 1); - __m128i ni2 = _mm256_castsi256_si128( i1 ); - __m128i ni3 = _mm256_extractf128_si256( i1, 1); - __m128i ni4 = _mm256_castsi256_si128( i2 ); - __m128i ni5 = _mm256_extractf128_si256( i2, 1); - __m128i ni6 = _mm256_castsi256_si128( i3 ); - __m128i ni7 = _mm256_extractf128_si256( i3, 1); - - // Convert int32 to int16 - ni0 = _mm_packs_epi32( ni0, ni1 ); - ni2 = _mm_packs_epi32( ni2, ni3 ); - ni4 = _mm_packs_epi32( ni4, ni5 ); - ni6 = _mm_packs_epi32( ni6, ni7 ); - // Convert int16 to int8 - ni0 = _mm_packs_epi16( ni0, ni2 ); - ni4 = _mm_packs_epi16( ni4, ni6 ); - - _mm_storeu_si128((__m128i *)(y[i].qs + 0), ni0); - _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4); -#endif - } -#elif defined(__riscv_v_intrinsic) - - size_t vl = __riscv_vsetvl_e32m4(QK8_0); - - for (int i = 0; i < nb; i++) { - // load elements - vfloat32m4_t v_x = __riscv_vle32_v_f32m4(x+i*QK8_0, vl); - - vfloat32m4_t vfabs = __riscv_vfabs_v_f32m4(v_x, vl); - vfloat32m1_t tmp = __riscv_vfmv_v_f_f32m1(0.0f, vl); - vfloat32m1_t vmax = __riscv_vfredmax_vs_f32m4_f32m1(vfabs, tmp, vl); - float amax = __riscv_vfmv_f_s_f32m1_f32(vmax); - - const float d = amax / ((1 << 7) - 1); - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = GGML_FP32_TO_FP16(d); - - vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl); - - // convert to integer - vint16m2_t vi = __riscv_vfncvt_x_f_w_i16m2(x0, vl); - vint8m1_t vs = __riscv_vncvt_x_x_w_i8m1(vi, vl); - - // store result - __riscv_vse8_v_i8m1(y[i].qs , vs, vl); - } -#else - // scalar - quantize_row_q8_0_reference(x, y, k); -#endif -} - -// reference implementation for deterministic creation of model files -static void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int k) { - assert(QK8_1 == 32); - assert(k % QK8_1 == 0); - const int nb = k / QK8_1; - - for (int i = 0; i < nb; i++) { - float amax = 0.0f; // absolute max - - for (int j = 0; j < QK8_1; j++) { - const float v = x[i*QK8_1 + j]; - amax = MAX(amax, fabsf(v)); - } - - const float d = amax / ((1 << 7) - 1); - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = d; - - int sum = 0; - - for (int j = 0; j < QK8_1/2; ++j) { - const float v0 = x[i*QK8_1 + j]*id; - const float v1 = x[i*QK8_1 + QK8_1/2 + j]*id; - - y[i].qs[ j] = roundf(v0); - y[i].qs[QK8_1/2 + j] = roundf(v1); - - sum += y[i].qs[ j]; - sum += y[i].qs[QK8_1/2 + j]; - } - - y[i].s = sum*d; - } -} - -static void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) { - assert(k % QK8_1 == 0); - const int nb = k / QK8_1; - - block_q8_1 * restrict y = vy; - -#if defined(__ARM_NEON) - for (int i = 0; i < nb; i++) { - float32x4_t srcv [8]; - float32x4_t asrcv[8]; - float32x4_t amaxv[8]; - - for (int j = 0; j < 8; j++) srcv[j] = vld1q_f32(x + i*32 + 4*j); - for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]); - - for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]); - for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]); - for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]); - - const float amax = vmaxvq_f32(amaxv[0]); - - const float d = amax / ((1 << 7) - 1); - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = d; - - int32x4_t accv = vdupq_n_s32(0); - - for (int j = 0; j < 8; j++) { - const float32x4_t v = vmulq_n_f32(srcv[j], id); - const int32x4_t vi = vcvtnq_s32_f32(v); - - y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0); - y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1); - y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2); - y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3); - - accv = vaddq_s32(accv, vi); - } - - y[i].s = d * vaddvq_s32(accv); - } -#elif defined(__wasm_simd128__) - for (int i = 0; i < nb; i++) { - v128_t srcv [8]; - v128_t asrcv[8]; - v128_t amaxv[8]; - - for (int j = 0; j < 8; j++) srcv[j] = wasm_v128_load(x + i*32 + 4*j); - for (int j = 0; j < 8; j++) asrcv[j] = wasm_f32x4_abs(srcv[j]); - - for (int j = 0; j < 4; j++) amaxv[2*j] = wasm_f32x4_max(asrcv[2*j], asrcv[2*j+1]); - for (int j = 0; j < 2; j++) amaxv[4*j] = wasm_f32x4_max(amaxv[4*j], amaxv[4*j+2]); - for (int j = 0; j < 1; j++) amaxv[8*j] = wasm_f32x4_max(amaxv[8*j], amaxv[8*j+4]); - - const float amax = MAX(MAX(wasm_f32x4_extract_lane(amaxv[0], 0), - wasm_f32x4_extract_lane(amaxv[0], 1)), - MAX(wasm_f32x4_extract_lane(amaxv[0], 2), - wasm_f32x4_extract_lane(amaxv[0], 3))); - - const float d = amax / ((1 << 7) - 1); - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = d; - - v128_t accv = wasm_i32x4_splat(0); - - for (int j = 0; j < 8; j++) { - const v128_t v = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id)); - const v128_t vi = wasm_i32x4_trunc_sat_f32x4(v); - - y[i].qs[4*j + 0] = wasm_i32x4_extract_lane(vi, 0); - y[i].qs[4*j + 1] = wasm_i32x4_extract_lane(vi, 1); - y[i].qs[4*j + 2] = wasm_i32x4_extract_lane(vi, 2); - y[i].qs[4*j + 3] = wasm_i32x4_extract_lane(vi, 3); - - accv = wasm_i32x4_add(accv, vi); - } - - y[i].s = d * (wasm_i32x4_extract_lane(accv, 0) + - wasm_i32x4_extract_lane(accv, 1) + - wasm_i32x4_extract_lane(accv, 2) + - wasm_i32x4_extract_lane(accv, 3)); - } -#elif defined(__AVX2__) || defined(__AVX__) - for (int i = 0; i < nb; i++) { - // Load elements into 4 AVX vectors - __m256 v0 = _mm256_loadu_ps( x ); - __m256 v1 = _mm256_loadu_ps( x + 8 ); - __m256 v2 = _mm256_loadu_ps( x + 16 ); - __m256 v3 = _mm256_loadu_ps( x + 24 ); - x += 32; - - // Compute max(abs(e)) for the block - const __m256 signBit = _mm256_set1_ps( -0.0f ); - __m256 maxAbs = _mm256_andnot_ps( signBit, v0 ); - maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) ); - maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) ); - maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) ); - - __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) ); - max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) ); - max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) ); - const float maxScalar = _mm_cvtss_f32( max4 ); - - // Quantize these floats - const float d = maxScalar / 127.f; - y[i].d = d; - const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; - const __m256 mul = _mm256_set1_ps( id ); - - // Apply the multiplier - v0 = _mm256_mul_ps( v0, mul ); - v1 = _mm256_mul_ps( v1, mul ); - v2 = _mm256_mul_ps( v2, mul ); - v3 = _mm256_mul_ps( v3, mul ); - - // Round to nearest integer - v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST ); - v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST ); - v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST ); - v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST ); - - // Convert floats to integers - __m256i i0 = _mm256_cvtps_epi32( v0 ); - __m256i i1 = _mm256_cvtps_epi32( v1 ); - __m256i i2 = _mm256_cvtps_epi32( v2 ); - __m256i i3 = _mm256_cvtps_epi32( v3 ); - -#if defined(__AVX2__) - // Compute the sum of the quants and set y[i].s - y[i].s = d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))); - - // Convert int32 to int16 - i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 - i2 = _mm256_packs_epi32( i2, i3 ); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31 - // Convert int16 to int8 - i0 = _mm256_packs_epi16( i0, i2 ); // 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27, 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31 - - // We got our precious signed bytes, but the order is now wrong - // These AVX2 pack instructions process 16-byte pieces independently - // The following instruction is fixing the order - const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 ); - i0 = _mm256_permutevar8x32_epi32( i0, perm ); - - _mm256_storeu_si256((__m256i *)y[i].qs, i0); -#else - // Since we don't have in AVX some necessary functions, - // we split the registers in half and call AVX2 analogs from SSE - __m128i ni0 = _mm256_castsi256_si128( i0 ); - __m128i ni1 = _mm256_extractf128_si256( i0, 1); - __m128i ni2 = _mm256_castsi256_si128( i1 ); - __m128i ni3 = _mm256_extractf128_si256( i1, 1); - __m128i ni4 = _mm256_castsi256_si128( i2 ); - __m128i ni5 = _mm256_extractf128_si256( i2, 1); - __m128i ni6 = _mm256_castsi256_si128( i3 ); - __m128i ni7 = _mm256_extractf128_si256( i3, 1); - - // Compute the sum of the quants and set y[i].s - const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3)); - const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7)); - y[i].s = d * hsum_i32_4(_mm_add_epi32(s0, s1)); - - // Convert int32 to int16 - ni0 = _mm_packs_epi32( ni0, ni1 ); - ni2 = _mm_packs_epi32( ni2, ni3 ); - ni4 = _mm_packs_epi32( ni4, ni5 ); - ni6 = _mm_packs_epi32( ni6, ni7 ); - // Convert int16 to int8 - ni0 = _mm_packs_epi16( ni0, ni2 ); - ni4 = _mm_packs_epi16( ni4, ni6 ); - - _mm_storeu_si128((__m128i *)(y[i].qs + 0), ni0); - _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4); -#endif - } -#elif defined(__riscv_v_intrinsic) - - size_t vl = __riscv_vsetvl_e32m4(QK8_1); - - for (int i = 0; i < nb; i++) { - // load elements - vfloat32m4_t v_x = __riscv_vle32_v_f32m4(x+i*QK8_1, vl); - - vfloat32m4_t vfabs = __riscv_vfabs_v_f32m4(v_x, vl); - vfloat32m1_t tmp = __riscv_vfmv_v_f_f32m1(0.0, vl); - vfloat32m1_t vmax = __riscv_vfredmax_vs_f32m4_f32m1(vfabs, tmp, vl); - float amax = __riscv_vfmv_f_s_f32m1_f32(vmax); - - const float d = amax / ((1 << 7) - 1); - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = d; - - vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl); - - // convert to integer - vint16m2_t vi = __riscv_vfncvt_x_f_w_i16m2(x0, vl); - vint8m1_t vs = __riscv_vncvt_x_x_w_i8m1(vi, vl); - - // store result - __riscv_vse8_v_i8m1(y[i].qs , vs, vl); - - // compute sum for y[i].s - vint16m1_t tmp2 = __riscv_vmv_v_x_i16m1(0, vl); - vint16m1_t vwrs = __riscv_vwredsum_vs_i8m1_i16m1(vs, tmp2, vl); - - // set y[i].s - int sum = __riscv_vmv_x_s_i16m1_i16(vwrs); - y[i].s = sum*d; - } -#else - // scalar - quantize_row_q8_1_reference(x, y, k); -#endif -} - -static void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k) { - static const int qk = QK4_0; - - assert(k % qk == 0); - - const int nb = k / qk; - - for (int i = 0; i < nb; i++) { - const float d = GGML_FP16_TO_FP32(x[i].d); - - for (int j = 0; j < qk/2; ++j) { - const int x0 = (x[i].qs[j] & 0x0F) - 8; - const int x1 = (x[i].qs[j] >> 4) - 8; - - y[i*qk + j + 0 ] = x0*d; - y[i*qk + j + qk/2] = x1*d; - } - } -} - -static void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int k) { - static const int qk = QK4_1; - - assert(k % qk == 0); - - const int nb = k / qk; - - for (int i = 0; i < nb; i++) { - const float d = GGML_FP16_TO_FP32(x[i].d); - const float m = GGML_FP16_TO_FP32(x[i].m); - - for (int j = 0; j < qk/2; ++j) { - const int x0 = (x[i].qs[j] & 0x0F); - const int x1 = (x[i].qs[j] >> 4); - - y[i*qk + j + 0 ] = x0*d + m; - y[i*qk + j + qk/2] = x1*d + m; - } - } -} - -static void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int k) { - static const int qk = QK5_0; - - assert(k % qk == 0); - - const int nb = k / qk; - - for (int i = 0; i < nb; i++) { - const float d = GGML_FP16_TO_FP32(x[i].d); - - uint32_t qh; - memcpy(&qh, x[i].qh, sizeof(qh)); - - for (int j = 0; j < qk/2; ++j) { - const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; - const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; - - const int32_t x0 = ((x[i].qs[j] & 0x0F) | xh_0) - 16; - const int32_t x1 = ((x[i].qs[j] >> 4) | xh_1) - 16; - - y[i*qk + j + 0 ] = x0*d; - y[i*qk + j + qk/2] = x1*d; - } - } -} - -static void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int k) { - static const int qk = QK5_1; - - assert(k % qk == 0); - - const int nb = k / qk; - - for (int i = 0; i < nb; i++) { - const float d = GGML_FP16_TO_FP32(x[i].d); - const float m = GGML_FP16_TO_FP32(x[i].m); - - uint32_t qh; - memcpy(&qh, x[i].qh, sizeof(qh)); - - for (int j = 0; j < qk/2; ++j) { - const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; - const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; - - const int x0 = (x[i].qs[j] & 0x0F) | xh_0; - const int x1 = (x[i].qs[j] >> 4) | xh_1; - - y[i*qk + j + 0 ] = x0*d + m; - y[i*qk + j + qk/2] = x1*d + m; - } - } -} - -static void dequantize_row_q8_0(const void * restrict vx, float * restrict y, int k) { - static const int qk = QK8_0; - - assert(k % qk == 0); - - const int nb = k / qk; - - const block_q8_0 * restrict x = vx; - - for (int i = 0; i < nb; i++) { - const float d = GGML_FP16_TO_FP32(x[i].d); - - for (int j = 0; j < qk; ++j) { - y[i*qk + j] = x[i].qs[j]*d; - } - } -} - static void ggml_vec_dot_f32(const int n, float * restrict s, const float * restrict x, const float * restrict y); static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t * restrict x, ggml_fp16_t * restrict y); -static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); -static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); -static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); -static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); -static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { [GGML_TYPE_I8] = { @@ -1714,6 +422,28 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .vec_dot = ggml_vec_dot_q4_1_q8_1, .vec_dot_type = GGML_TYPE_Q8_1, }, + [4] = { // GGML_TYPE_Q4_2 + .type_name = "DEPRECATED", + .blck_size = 0, + .type_size = 0, + .is_quantized = false, + .to_float = NULL, + .from_float = NULL, + .from_float_reference = NULL, + .vec_dot = NULL, + .vec_dot_type = GGML_TYPE_COUNT, + }, + [5] = { // GGML_TYPE_Q4_3 + .type_name = "DEPRECATED", + .blck_size = 0, + .type_size = 0, + .is_quantized = false, + .to_float = NULL, + .from_float = NULL, + .from_float_reference = NULL, + .vec_dot = NULL, + .vec_dot_type = GGML_TYPE_COUNT, + }, [GGML_TYPE_Q5_0] = { .type_name = "q5_0", .blck_size = QK5_0, @@ -1741,7 +471,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .blck_size = QK8_0, .type_size = sizeof(block_q8_0), .is_quantized = true, - .to_float = dequantize_row_q8_0, + .to_float = (ggml_to_float_t) dequantize_row_q8_0, .from_float = quantize_row_q8_0, .from_float_reference = (ggml_from_float_t) quantize_row_q8_0_reference, .vec_dot = ggml_vec_dot_q8_0_q8_0, @@ -1756,7 +486,6 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .from_float_reference = (ggml_from_float_t) quantize_row_q8_1_reference, .vec_dot_type = GGML_TYPE_Q8_1, }, -#ifdef GGML_USE_K_QUANTS [GGML_TYPE_Q2_K] = { .type_name = "q2_K", .blck_size = QK_K, @@ -1819,7 +548,6 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .is_quantized = true, .from_float = quantize_row_q8_K, } -#endif }; // For internal test use @@ -1828,7 +556,6 @@ ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type) { return type_traits[type]; } - // // simd mappings // @@ -2444,1218 +1171,6 @@ static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t * rest *s = sumf; } -static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int qk = QK8_0; - const int nb = n / qk; - - assert(n % qk == 0); - - const block_q4_0 * restrict x = vx; - const block_q8_0 * restrict y = vy; - -#if defined(__ARM_NEON) - float32x4_t sumv0 = vdupq_n_f32(0.0f); - float32x4_t sumv1 = vdupq_n_f32(0.0f); - - GGML_ASSERT(nb % 2 == 0); // TODO: handle odd nb - for (int i = 0; i < nb; i += 2) { - const block_q4_0 * restrict x0 = &x[i + 0]; - const block_q4_0 * restrict x1 = &x[i + 1]; - const block_q8_0 * restrict y0 = &y[i + 0]; - const block_q8_0 * restrict y1 = &y[i + 1]; - - const uint8x16_t m4b = vdupq_n_u8(0x0F); - const int8x16_t s8b = vdupq_n_s8(0x8); - - const uint8x16_t v0_0 = vld1q_u8(x0->qs); - const uint8x16_t v0_1 = vld1q_u8(x1->qs); - - // 4-bit -> 8-bit - const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); - const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); - const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); - const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); - - // sub 8 - const int8x16_t v0_0ls = vsubq_s8(v0_0l, s8b); - const int8x16_t v0_0hs = vsubq_s8(v0_0h, s8b); - const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b); - const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b); - - // load y - const int8x16_t v1_0l = vld1q_s8(y0->qs); - const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); - const int8x16_t v1_1l = vld1q_s8(y1->qs); - const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); - -#if defined(__ARM_FEATURE_DOTPROD) - // dot product into int32x4_t - const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h); - const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h); - - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); -#else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0l)); - const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0l)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hs), vget_low_s8 (v1_0h)); - const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hs), vget_high_s8(v1_0h)); - - const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1ls), vget_low_s8 (v1_1l)); - const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1ls), vget_high_s8(v1_1l)); - const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hs), vget_low_s8 (v1_1h)); - const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hs), vget_high_s8(v1_1h)); - - const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); - const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); - const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); - const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); - - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); -#endif - } - - *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); -#elif defined(__AVX2__) - // Initialize accumulator with zeros - __m256 acc = _mm256_setzero_ps(); - - // Main loop - for (int i = 0; i < nb; ++i) { - /* Compute combined scale for the block */ - const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) ); - - __m256i bx = bytes_from_nibbles_32(x[i].qs); - - // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval. - const __m256i off = _mm256_set1_epi8( 8 ); - bx = _mm256_sub_epi8( bx, off ); - - __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); - - const __m256 q = mul_sum_i8_pairs_float(bx, by); - - /* Multiply q with scale and accumulate */ - acc = _mm256_fmadd_ps( d, q, acc ); - } - - *s = hsum_float_8(acc); -#elif defined(__AVX__) - // Initialize accumulator with zeros - __m256 acc = _mm256_setzero_ps(); - - // Main loop - for (int i = 0; i < nb; ++i) { - // Compute combined scale for the block - const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) ); - - const __m128i lowMask = _mm_set1_epi8(0xF); - const __m128i off = _mm_set1_epi8(8); - - const __m128i tmp = _mm_loadu_si128((const __m128i *)x[i].qs); - - __m128i bx = _mm_and_si128(lowMask, tmp); - __m128i by = _mm_loadu_si128((const __m128i *)y[i].qs); - bx = _mm_sub_epi8(bx, off); - const __m128i i32_0 = mul_sum_i8_pairs(bx, by); - - bx = _mm_and_si128(lowMask, _mm_srli_epi64(tmp, 4)); - by = _mm_loadu_si128((const __m128i *)(y[i].qs + 16)); - bx = _mm_sub_epi8(bx, off); - const __m128i i32_1 = mul_sum_i8_pairs(bx, by); - - // Convert int32_t to float - __m256 p = _mm256_cvtepi32_ps(MM256_SET_M128I(i32_0, i32_1)); - - // Apply the scale, and accumulate - acc = _mm256_add_ps(_mm256_mul_ps( d, p ), acc); - } - - *s = hsum_float_8(acc); -#elif defined(__SSSE3__) - // set constants - const __m128i lowMask = _mm_set1_epi8(0xF); - const __m128i off = _mm_set1_epi8(8); - - // Initialize accumulator with zeros - __m128 acc_0 = _mm_setzero_ps(); - __m128 acc_1 = _mm_setzero_ps(); - __m128 acc_2 = _mm_setzero_ps(); - __m128 acc_3 = _mm_setzero_ps(); - - // First round without accumulation - { - _mm_prefetch(&x[0] + sizeof(block_q4_0), _MM_HINT_T0); - _mm_prefetch(&y[0] + sizeof(block_q8_0), _MM_HINT_T0); - - // Compute combined scale for the block 0 and 1 - const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[0].d) * GGML_FP16_TO_FP32(y[0].d) ); - - const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[0].qs); - - __m128i bx_0 = _mm_and_si128(lowMask, tmp_0_1); - __m128i by_0 = _mm_loadu_si128((const __m128i *)y[0].qs); - bx_0 = _mm_sub_epi8(bx_0, off); - const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0); - - __m128i bx_1 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_0_1, 4)); - __m128i by_1 = _mm_loadu_si128((const __m128i *)(y[0].qs + 16)); - bx_1 = _mm_sub_epi8(bx_1, off); - const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1); - - _mm_prefetch(&x[1] + sizeof(block_q4_0), _MM_HINT_T0); - _mm_prefetch(&y[1] + sizeof(block_q8_0), _MM_HINT_T0); - - // Compute combined scale for the block 2 and 3 - const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[1].d) * GGML_FP16_TO_FP32(y[1].d) ); - - const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[1].qs); - - __m128i bx_2 = _mm_and_si128(lowMask, tmp_2_3); - __m128i by_2 = _mm_loadu_si128((const __m128i *)y[1].qs); - bx_2 = _mm_sub_epi8(bx_2, off); - const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2); - - __m128i bx_3 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_2_3, 4)); - __m128i by_3 = _mm_loadu_si128((const __m128i *)(y[1].qs + 16)); - bx_3 = _mm_sub_epi8(bx_3, off); - const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3); - - // Convert int32_t to float - __m128 p0 = _mm_cvtepi32_ps(i32_0); - __m128 p1 = _mm_cvtepi32_ps(i32_1); - __m128 p2 = _mm_cvtepi32_ps(i32_2); - __m128 p3 = _mm_cvtepi32_ps(i32_3); - - // Apply the scale - acc_0 = _mm_mul_ps( d_0_1, p0 ); - acc_1 = _mm_mul_ps( d_0_1, p1 ); - acc_2 = _mm_mul_ps( d_2_3, p2 ); - acc_3 = _mm_mul_ps( d_2_3, p3 ); - } - - // Main loop - GGML_ASSERT(nb % 2 == 0); // TODO: handle odd nb - for (int i = 2; i < nb; i+=2) { - _mm_prefetch(&x[i] + sizeof(block_q4_0), _MM_HINT_T0); - _mm_prefetch(&y[i] + sizeof(block_q8_0), _MM_HINT_T0); - - // Compute combined scale for the block 0 and 1 - const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) ); - - const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[i].qs); - - __m128i bx_0 = _mm_and_si128(lowMask, tmp_0_1); - __m128i by_0 = _mm_loadu_si128((const __m128i *)y[i].qs); - bx_0 = _mm_sub_epi8(bx_0, off); - const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0); - - __m128i bx_1 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_0_1, 4)); - __m128i by_1 = _mm_loadu_si128((const __m128i *)(y[i].qs + 16)); - bx_1 = _mm_sub_epi8(bx_1, off); - const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1); - - _mm_prefetch(&x[i] + 2 * sizeof(block_q4_0), _MM_HINT_T0); - _mm_prefetch(&y[i] + 2 * sizeof(block_q8_0), _MM_HINT_T0); - - // Compute combined scale for the block 2 and 3 - const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[i + 1].d) * GGML_FP16_TO_FP32(y[i + 1].d) ); - - const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[i + 1].qs); - - __m128i bx_2 = _mm_and_si128(lowMask, tmp_2_3); - __m128i by_2 = _mm_loadu_si128((const __m128i *)y[i + 1].qs); - bx_2 = _mm_sub_epi8(bx_2, off); - const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2); - - __m128i bx_3 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_2_3, 4)); - __m128i by_3 = _mm_loadu_si128((const __m128i *)(y[i + 1].qs + 16)); - bx_3 = _mm_sub_epi8(bx_3, off); - const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3); - - // Convert int32_t to float - __m128 p0 = _mm_cvtepi32_ps(i32_0); - __m128 p1 = _mm_cvtepi32_ps(i32_1); - __m128 p2 = _mm_cvtepi32_ps(i32_2); - __m128 p3 = _mm_cvtepi32_ps(i32_3); - - // Apply the scale - __m128 p0_d = _mm_mul_ps( d_0_1, p0 ); - __m128 p1_d = _mm_mul_ps( d_0_1, p1 ); - __m128 p2_d = _mm_mul_ps( d_2_3, p2 ); - __m128 p3_d = _mm_mul_ps( d_2_3, p3 ); - - // Acummulate - acc_0 = _mm_add_ps(p0_d, acc_0); - acc_1 = _mm_add_ps(p1_d, acc_1); - acc_2 = _mm_add_ps(p2_d, acc_2); - acc_3 = _mm_add_ps(p3_d, acc_3); - } - - *s = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3); -#elif defined(__riscv_v_intrinsic) - float sumf = 0.0; - - size_t vl = __riscv_vsetvl_e8m1(qk/2); - - for (int i = 0; i < nb; i++) { - // load elements - vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl); - - vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl); - vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl); - - // mask and store lower part of x, and then upper part - vuint8mf2_t x_a = __riscv_vand_vx_u8mf2(tx, 0x0F, vl); - vuint8mf2_t x_l = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl); - - vint8mf2_t x_ai = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a); - vint8mf2_t x_li = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l); - - // subtract offset - vint8mf2_t v0 = __riscv_vsub_vx_i8mf2(x_ai, 8, vl); - vint8mf2_t v1 = __riscv_vsub_vx_i8mf2(x_li, 8, vl); - - vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl); - vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl); - - vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl); - - vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl); - vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl); - - int sumi = __riscv_vmv_x_s_i32m1_i32(vs2); - - sumf += sumi*GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d); - } - - *s = sumf; -#else - // scalar - float sumf = 0.0; - - for (int i = 0; i < nb; i++) { - int sumi = 0; - - for (int j = 0; j < qk/2; ++j) { - const int v0 = (x[i].qs[j] & 0x0F) - 8; - const int v1 = (x[i].qs[j] >> 4) - 8; - - sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]); - } - - sumf += sumi*GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d); - } - - *s = sumf; -#endif -} - -static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int qk = QK8_1; - const int nb = n / qk; - - assert(n % qk == 0); - - const block_q4_1 * restrict x = vx; - const block_q8_1 * restrict y = vy; - - // TODO: add WASM SIMD -#if defined(__ARM_NEON) - float32x4_t sumv0 = vdupq_n_f32(0.0f); - float32x4_t sumv1 = vdupq_n_f32(0.0f); - - float summs = 0; - - GGML_ASSERT(nb % 2 == 0); // TODO: handle odd nb - for (int i = 0; i < nb; i += 2) { - const block_q4_1 * restrict x0 = &x[i + 0]; - const block_q4_1 * restrict x1 = &x[i + 1]; - const block_q8_1 * restrict y0 = &y[i + 0]; - const block_q8_1 * restrict y1 = &y[i + 1]; - - summs += GGML_FP16_TO_FP32(x0->m) * y0->s + GGML_FP16_TO_FP32(x1->m) * y1->s; - - const uint8x16_t m4b = vdupq_n_u8(0x0F); - - const uint8x16_t v0_0 = vld1q_u8(x0->qs); - const uint8x16_t v0_1 = vld1q_u8(x1->qs); - - // 4-bit -> 8-bit - const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); - const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); - const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); - const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); - - // load y - const int8x16_t v1_0l = vld1q_s8(y0->qs); - const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); - const int8x16_t v1_1l = vld1q_s8(y1->qs); - const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); - -#if defined(__ARM_FEATURE_DOTPROD) - // dot product into int32x4_t - const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h); - const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h); - - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*y0->d); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*y1->d); -#else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0l), vget_low_s8 (v1_0l)); - const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0l), vget_high_s8(v1_0l)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0h), vget_low_s8 (v1_0h)); - const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0h), vget_high_s8(v1_0h)); - - const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1l), vget_low_s8 (v1_1l)); - const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1l), vget_high_s8(v1_1l)); - const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1h), vget_low_s8 (v1_1h)); - const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1h), vget_high_s8(v1_1h)); - - const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); - const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); - const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); - const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); - - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*y0->d); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*y1->d); -#endif - } - - *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs; -#elif defined(__AVX2__) || defined(__AVX__) - // Initialize accumulator with zeros - __m256 acc = _mm256_setzero_ps(); - - float summs = 0; - - // Main loop - for (int i = 0; i < nb; ++i) { - const float d0 = GGML_FP16_TO_FP32(x[i].d); - const float d1 = y[i].d; - - summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s; - - const __m256 d0v = _mm256_set1_ps( d0 ); - const __m256 d1v = _mm256_set1_ps( d1 ); - - // Compute combined scales - const __m256 d0d1 = _mm256_mul_ps( d0v, d1v ); - - // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes - const __m256i bx = bytes_from_nibbles_32(x[i].qs); - const __m256i by = _mm256_loadu_si256( (const __m256i *)y[i].qs ); - - const __m256 xy = mul_sum_us8_pairs_float(bx, by); - - // Accumulate d0*d1*x*y -#if defined(__AVX2__) - acc = _mm256_fmadd_ps( d0d1, xy, acc ); -#else - acc = _mm256_add_ps( _mm256_mul_ps( d0d1, xy ), acc ); -#endif - } - - *s = hsum_float_8(acc) + summs; -#elif defined(__riscv_v_intrinsic) - float sumf = 0.0; - - size_t vl = __riscv_vsetvl_e8m1(qk/2); - - for (int i = 0; i < nb; i++) { - // load elements - vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl); - - vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl); - vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl); - - // mask and store lower part of x, and then upper part - vuint8mf2_t x_a = __riscv_vand_vx_u8mf2(tx, 0x0F, vl); - vuint8mf2_t x_l = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl); - - vint8mf2_t v0 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a); - vint8mf2_t v1 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l); - - vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl); - vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl); - - vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl); - - vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl); - vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl); - - int sumi = __riscv_vmv_x_s_i32m1_i32(vs2); - - sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s; - } - - *s = sumf; -#else - // scalar - float sumf = 0.0; - - for (int i = 0; i < nb; i++) { - int sumi = 0; - - for (int j = 0; j < qk/2; ++j) { - const int v0 = (x[i].qs[j] & 0x0F); - const int v1 = (x[i].qs[j] >> 4); - - sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]); - } - - sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s; - } - - *s = sumf; -#endif -} - -static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int qk = QK8_0; - const int nb = n / qk; - - assert(n % qk == 0); - assert(qk == QK5_0); - - const block_q5_0 * restrict x = vx; - const block_q8_0 * restrict y = vy; - -#if defined(__ARM_NEON) - float32x4_t sumv0 = vdupq_n_f32(0.0f); - float32x4_t sumv1 = vdupq_n_f32(0.0f); - - uint32_t qh0; - uint32_t qh1; - - uint64_t tmp0[4]; - uint64_t tmp1[4]; - - GGML_ASSERT(nb % 2 == 0); // TODO: handle odd nb - for (int i = 0; i < nb; i += 2) { - const block_q5_0 * restrict x0 = &x[i]; - const block_q5_0 * restrict x1 = &x[i + 1]; - const block_q8_0 * restrict y0 = &y[i]; - const block_q8_0 * restrict y1 = &y[i + 1]; - - const uint8x16_t m4b = vdupq_n_u8(0x0F); - - // extract the 5th bit via lookup table ((!b) << 4) - memcpy(&qh0, x0->qh, sizeof(qh0)); - memcpy(&qh1, x1->qh, sizeof(qh1)); - - tmp0[0] = table_b2b_1[(qh0 >> 0) & 0xFF]; - tmp0[1] = table_b2b_1[(qh0 >> 8) & 0xFF]; - tmp0[2] = table_b2b_1[(qh0 >> 16) & 0xFF]; - tmp0[3] = table_b2b_1[(qh0 >> 24) ]; - - tmp1[0] = table_b2b_1[(qh1 >> 0) & 0xFF]; - tmp1[1] = table_b2b_1[(qh1 >> 8) & 0xFF]; - tmp1[2] = table_b2b_1[(qh1 >> 16) & 0xFF]; - tmp1[3] = table_b2b_1[(qh1 >> 24) ]; - - const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0)); - const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2)); - const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0)); - const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2)); - - const uint8x16_t v0_0 = vld1q_u8(x0->qs); - const uint8x16_t v0_1 = vld1q_u8(x1->qs); - - // 4-bit -> 8-bit - int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); - int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); - int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); - int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); - - // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero) - const int8x16_t v0_0lf = vsubq_s8(v0_0l, qhl0); - const int8x16_t v0_0hf = vsubq_s8(v0_0h, qhh0); - const int8x16_t v0_1lf = vsubq_s8(v0_1l, qhl1); - const int8x16_t v0_1hf = vsubq_s8(v0_1h, qhh1); - - // load y - const int8x16_t v1_0l = vld1q_s8(y0->qs); - const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); - const int8x16_t v1_1l = vld1q_s8(y1->qs); - const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); - -#if defined(__ARM_FEATURE_DOTPROD) - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( - vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l), - vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( - vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l), - vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); -#else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lf), vget_low_s8 (v1_0l)); - const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lf), vget_high_s8(v1_0l)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hf), vget_low_s8 (v1_0h)); - const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hf), vget_high_s8(v1_0h)); - - const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lf), vget_low_s8 (v1_1l)); - const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lf), vget_high_s8(v1_1l)); - const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hf), vget_low_s8 (v1_1h)); - const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hf), vget_high_s8(v1_1h)); - - const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); - const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); - const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); - const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); - - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); -#endif - } - - *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); -#elif defined(__wasm_simd128__) - v128_t sumv = wasm_f32x4_splat(0.0f); - - uint32_t qh; - uint64_t tmp[4]; - - // TODO: check if unrolling this is better - for (int i = 0; i < nb; ++i) { - const block_q5_0 * restrict x0 = &x[i]; - const block_q8_0 * restrict y0 = &y[i]; - - const v128_t m4b = wasm_i8x16_splat(0x0F); - - // extract the 5th bit - memcpy(&qh, x0->qh, sizeof(qh)); - - tmp[0] = table_b2b_1[(qh >> 0) & 0xFF]; - tmp[1] = table_b2b_1[(qh >> 8) & 0xFF]; - tmp[2] = table_b2b_1[(qh >> 16) & 0xFF]; - tmp[3] = table_b2b_1[(qh >> 24) ]; - - const v128_t qhl = wasm_v128_load(tmp + 0); - const v128_t qhh = wasm_v128_load(tmp + 2); - - const v128_t v0 = wasm_v128_load(x0->qs); - - // 4-bit -> 8-bit - const v128_t v0l = wasm_v128_and (v0, m4b); - const v128_t v0h = wasm_u8x16_shr(v0, 4); - - // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero) - const v128_t v0lf = wasm_i8x16_sub(v0l, qhl); - const v128_t v0hf = wasm_i8x16_sub(v0h, qhh); - - // load y - const v128_t v1l = wasm_v128_load(y0->qs); - const v128_t v1h = wasm_v128_load(y0->qs + 16); - - // int8x16 -> int16x8 - const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf); - const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf); - const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf); - const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf); - - const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l); - const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l); - const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h); - const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h); - - // dot product - sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4( - wasm_i32x4_add( - wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll), - wasm_i32x4_dot_i16x8(v0lfh, v1lh)), - wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl), - wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), - wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d)))); - } - - *s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) + - wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3); -#elif defined(__AVX2__) - // Initialize accumulator with zeros - __m256 acc = _mm256_setzero_ps(); - - // Main loop - for (int i = 0; i < nb; i++) { - /* Compute combined scale for the block */ - const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d)); - - __m256i bx = bytes_from_nibbles_32(x[i].qs); - __m256i bxhi = bytes_from_bits_32(x[i].qh); - bxhi = _mm256_andnot_si256(bxhi, _mm256_set1_epi8((char)0xF0)); - bx = _mm256_or_si256(bx, bxhi); - - __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); - - const __m256 q = mul_sum_i8_pairs_float(bx, by); - - /* Multiply q with scale and accumulate */ - acc = _mm256_fmadd_ps(d, q, acc); - } - - *s = hsum_float_8(acc); -#elif defined(__AVX__) - // Initialize accumulator with zeros - __m256 acc = _mm256_setzero_ps(); - __m128i mask = _mm_set1_epi8((char)0xF0); - - // Main loop - for (int i = 0; i < nb; i++) { - /* Compute combined scale for the block */ - const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d)); - - __m256i bx = bytes_from_nibbles_32(x[i].qs); - const __m256i bxhi = bytes_from_bits_32(x[i].qh); - __m128i bxhil = _mm256_castsi256_si128(bxhi); - __m128i bxhih = _mm256_extractf128_si256(bxhi, 1); - bxhil = _mm_andnot_si128(bxhil, mask); - bxhih = _mm_andnot_si128(bxhih, mask); - __m128i bxl = _mm256_castsi256_si128(bx); - __m128i bxh = _mm256_extractf128_si256(bx, 1); - bxl = _mm_or_si128(bxl, bxhil); - bxh = _mm_or_si128(bxh, bxhih); - bx = MM256_SET_M128I(bxh, bxl); - - const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); - - const __m256 q = mul_sum_i8_pairs_float(bx, by); - - /* Multiply q with scale and accumulate */ - acc = _mm256_add_ps(_mm256_mul_ps(d, q), acc); - } - - *s = hsum_float_8(acc); -#elif defined(__riscv_v_intrinsic) - float sumf = 0.0; - - uint32_t qh; - - size_t vl = __riscv_vsetvl_e8m1(qk/2); - - // These tempory registers are for masking and shift operations - vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl); - vuint32m2_t vt_2 = __riscv_vsll_vv_u32m2(__riscv_vmv_v_x_u32m2(1, vl), vt_1, vl); - - vuint32m2_t vt_3 = __riscv_vsll_vx_u32m2(vt_2, 16, vl); - vuint32m2_t vt_4 = __riscv_vadd_vx_u32m2(vt_1, 12, vl); - - for (int i = 0; i < nb; i++) { - memcpy(&qh, x[i].qh, sizeof(uint32_t)); - - // ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4; - vuint32m2_t xha_0 = __riscv_vand_vx_u32m2(vt_2, qh, vl); - vuint32m2_t xhr_0 = __riscv_vsrl_vv_u32m2(xha_0, vt_1, vl); - vuint32m2_t xhl_0 = __riscv_vsll_vx_u32m2(xhr_0, 4, vl); - - // ((qh & (1u << (j + 16))) >> (j + 12)); - vuint32m2_t xha_1 = __riscv_vand_vx_u32m2(vt_3, qh, vl); - vuint32m2_t xhl_1 = __riscv_vsrl_vv_u32m2(xha_1, vt_4, vl); - - // narrowing - vuint16m1_t xhc_0 = __riscv_vncvt_x_x_w_u16m1(xhl_0, vl); - vuint8mf2_t xh_0 = __riscv_vncvt_x_x_w_u8mf2(xhc_0, vl); - - vuint16m1_t xhc_1 = __riscv_vncvt_x_x_w_u16m1(xhl_1, vl); - vuint8mf2_t xh_1 = __riscv_vncvt_x_x_w_u8mf2(xhc_1, vl); - - // load - vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl); - - vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl); - vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl); - - vuint8mf2_t x_at = __riscv_vand_vx_u8mf2(tx, 0x0F, vl); - vuint8mf2_t x_lt = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl); - - vuint8mf2_t x_a = __riscv_vor_vv_u8mf2(x_at, xh_0, vl); - vuint8mf2_t x_l = __riscv_vor_vv_u8mf2(x_lt, xh_1, vl); - - vint8mf2_t x_ai = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a); - vint8mf2_t x_li = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l); - - vint8mf2_t v0 = __riscv_vsub_vx_i8mf2(x_ai, 16, vl); - vint8mf2_t v1 = __riscv_vsub_vx_i8mf2(x_li, 16, vl); - - vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl); - vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl); - - vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl); - - vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl); - vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl); - - int sumi = __riscv_vmv_x_s_i32m1_i32(vs2); - - sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d)) * sumi; - } - - *s = sumf; -#else - // scalar - float sumf = 0.0; - - for (int i = 0; i < nb; i++) { - uint32_t qh; - memcpy(&qh, x[i].qh, sizeof(qh)); - - int sumi = 0; - - for (int j = 0; j < qk/2; ++j) { - const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4; - const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12)); - - const int32_t x0 = ((x[i].qs[j] & 0x0F) | xh_0) - 16; - const int32_t x1 = ((x[i].qs[j] >> 4) | xh_1) - 16; - - sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]); - } - - sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d)) * sumi; - } - - *s = sumf; -#endif -} - -static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int qk = QK8_1; - const int nb = n / qk; - - assert(n % qk == 0); - assert(qk == QK5_1); - - const block_q5_1 * restrict x = vx; - const block_q8_1 * restrict y = vy; - -#if defined(__ARM_NEON) - float32x4_t sumv0 = vdupq_n_f32(0.0f); - float32x4_t sumv1 = vdupq_n_f32(0.0f); - - float summs0 = 0.0f; - float summs1 = 0.0f; - - uint32_t qh0; - uint32_t qh1; - - uint64_t tmp0[4]; - uint64_t tmp1[4]; - - GGML_ASSERT(nb % 2 == 0); // TODO: handle odd nb - for (int i = 0; i < nb; i += 2) { - const block_q5_1 * restrict x0 = &x[i]; - const block_q5_1 * restrict x1 = &x[i + 1]; - const block_q8_1 * restrict y0 = &y[i]; - const block_q8_1 * restrict y1 = &y[i + 1]; - - const uint8x16_t m4b = vdupq_n_u8(0x0F); - - summs0 += GGML_FP16_TO_FP32(x0->m) * y0->s; - summs1 += GGML_FP16_TO_FP32(x1->m) * y1->s; - - // extract the 5th bit via lookup table ((b) << 4) - memcpy(&qh0, x0->qh, sizeof(qh0)); - memcpy(&qh1, x1->qh, sizeof(qh1)); - - tmp0[0] = table_b2b_0[(qh0 >> 0) & 0xFF]; - tmp0[1] = table_b2b_0[(qh0 >> 8) & 0xFF]; - tmp0[2] = table_b2b_0[(qh0 >> 16) & 0xFF]; - tmp0[3] = table_b2b_0[(qh0 >> 24) ]; - - tmp1[0] = table_b2b_0[(qh1 >> 0) & 0xFF]; - tmp1[1] = table_b2b_0[(qh1 >> 8) & 0xFF]; - tmp1[2] = table_b2b_0[(qh1 >> 16) & 0xFF]; - tmp1[3] = table_b2b_0[(qh1 >> 24) ]; - - const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0)); - const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2)); - const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0)); - const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2)); - - const uint8x16_t v0_0 = vld1q_u8(x0->qs); - const uint8x16_t v0_1 = vld1q_u8(x1->qs); - - // 4-bit -> 8-bit - const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); - const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); - const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); - const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); - - // add high bit - const int8x16_t v0_0lf = vorrq_s8(v0_0l, qhl0); - const int8x16_t v0_0hf = vorrq_s8(v0_0h, qhh0); - const int8x16_t v0_1lf = vorrq_s8(v0_1l, qhl1); - const int8x16_t v0_1hf = vorrq_s8(v0_1h, qhh1); - - // load y - const int8x16_t v1_0l = vld1q_s8(y0->qs); - const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); - const int8x16_t v1_1l = vld1q_s8(y1->qs); - const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); - -#if defined(__ARM_FEATURE_DOTPROD) - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( - vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l), - vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*y0->d); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( - vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l), - vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*y1->d); -#else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lf), vget_low_s8 (v1_0l)); - const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lf), vget_high_s8(v1_0l)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hf), vget_low_s8 (v1_0h)); - const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hf), vget_high_s8(v1_0h)); - - const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lf), vget_low_s8 (v1_1l)); - const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lf), vget_high_s8(v1_1l)); - const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hf), vget_low_s8 (v1_1h)); - const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hf), vget_high_s8(v1_1h)); - - const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); - const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); - const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); - const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); - - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*y0->d); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*y1->d); -#endif - } - - *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1; -#elif defined(__wasm_simd128__) - v128_t sumv = wasm_f32x4_splat(0.0f); - - float summs = 0.0f; - - uint32_t qh; - uint64_t tmp[4]; - - // TODO: check if unrolling this is better - for (int i = 0; i < nb; ++i) { - const block_q5_1 * restrict x0 = &x[i]; - const block_q8_1 * restrict y0 = &y[i]; - - summs += GGML_FP16_TO_FP32(x0->m) * y0->s; - - const v128_t m4b = wasm_i8x16_splat(0x0F); - - // extract the 5th bit - memcpy(&qh, x0->qh, sizeof(qh)); - - tmp[0] = table_b2b_0[(qh >> 0) & 0xFF]; - tmp[1] = table_b2b_0[(qh >> 8) & 0xFF]; - tmp[2] = table_b2b_0[(qh >> 16) & 0xFF]; - tmp[3] = table_b2b_0[(qh >> 24) ]; - - const v128_t qhl = wasm_v128_load(tmp + 0); - const v128_t qhh = wasm_v128_load(tmp + 2); - - const v128_t v0 = wasm_v128_load(x0->qs); - - // 4-bit -> 8-bit - const v128_t v0l = wasm_v128_and (v0, m4b); - const v128_t v0h = wasm_u8x16_shr(v0, 4); - - // add high bit - const v128_t v0lf = wasm_v128_or(v0l, qhl); - const v128_t v0hf = wasm_v128_or(v0h, qhh); - - // load y - const v128_t v1l = wasm_v128_load(y0->qs); - const v128_t v1h = wasm_v128_load(y0->qs + 16); - - // int8x16 -> int16x8 - const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf); - const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf); - const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf); - const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf); - - const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l); - const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l); - const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h); - const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h); - - // dot product - sumv = wasm_f32x4_add(sumv, - wasm_f32x4_mul(wasm_f32x4_convert_i32x4(wasm_i32x4_add( - wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll), - wasm_i32x4_dot_i16x8(v0lfh, v1lh)), - wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl), - wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), - wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * y0->d))); - } - - *s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) + - wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3) + summs; -#elif defined(__AVX2__) - // Initialize accumulator with zeros - __m256 acc = _mm256_setzero_ps(); - - float summs = 0.0f; - - // Main loop - for (int i = 0; i < nb; i++) { - const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d)); - - summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s; - - __m256i bx = bytes_from_nibbles_32(x[i].qs); - __m256i bxhi = bytes_from_bits_32(x[i].qh); - bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10)); - bx = _mm256_or_si256(bx, bxhi); - - const __m256 dy = _mm256_set1_ps(y[i].d); - const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); - - const __m256 q = mul_sum_us8_pairs_float(bx, by); - - acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc); - } - - *s = hsum_float_8(acc) + summs; -#elif defined(__AVX__) - // Initialize accumulator with zeros - __m256 acc = _mm256_setzero_ps(); - __m128i mask = _mm_set1_epi8(0x10); - - float summs = 0.0f; - - // Main loop - for (int i = 0; i < nb; i++) { - const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d)); - - summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s; - - __m256i bx = bytes_from_nibbles_32(x[i].qs); - const __m256i bxhi = bytes_from_bits_32(x[i].qh); - __m128i bxhil = _mm256_castsi256_si128(bxhi); - __m128i bxhih = _mm256_extractf128_si256(bxhi, 1); - bxhil = _mm_and_si128(bxhil, mask); - bxhih = _mm_and_si128(bxhih, mask); - __m128i bxl = _mm256_castsi256_si128(bx); - __m128i bxh = _mm256_extractf128_si256(bx, 1); - bxl = _mm_or_si128(bxl, bxhil); - bxh = _mm_or_si128(bxh, bxhih); - bx = MM256_SET_M128I(bxh, bxl); - - const __m256 dy = _mm256_set1_ps(y[i].d); - const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); - - const __m256 q = mul_sum_us8_pairs_float(bx, by); - - acc = _mm256_add_ps(_mm256_mul_ps(q, _mm256_mul_ps(dx, dy)), acc); - } - - *s = hsum_float_8(acc) + summs; -#elif defined(__riscv_v_intrinsic) - float sumf = 0.0; - - uint32_t qh; - - size_t vl = __riscv_vsetvl_e8m1(qk/2); - - // temporary registers for shift operations - vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl); - vuint32m2_t vt_2 = __riscv_vadd_vx_u32m2(vt_1, 12, vl); - - for (int i = 0; i < nb; i++) { - memcpy(&qh, x[i].qh, sizeof(uint32_t)); - - // load qh - vuint32m2_t vqh = __riscv_vmv_v_x_u32m2(qh, vl); - - // ((qh >> (j + 0)) << 4) & 0x10; - vuint32m2_t xhr_0 = __riscv_vsrl_vv_u32m2(vqh, vt_1, vl); - vuint32m2_t xhl_0 = __riscv_vsll_vx_u32m2(xhr_0, 4, vl); - vuint32m2_t xha_0 = __riscv_vand_vx_u32m2(xhl_0, 0x10, vl); - - // ((qh >> (j + 12)) ) & 0x10; - vuint32m2_t xhr_1 = __riscv_vsrl_vv_u32m2(vqh, vt_2, vl); - vuint32m2_t xha_1 = __riscv_vand_vx_u32m2(xhr_1, 0x10, vl); - - // narrowing - vuint16m1_t xhc_0 = __riscv_vncvt_x_x_w_u16m1(xha_0, vl); - vuint8mf2_t xh_0 = __riscv_vncvt_x_x_w_u8mf2(xhc_0, vl); - - vuint16m1_t xhc_1 = __riscv_vncvt_x_x_w_u16m1(xha_1, vl); - vuint8mf2_t xh_1 = __riscv_vncvt_x_x_w_u8mf2(xhc_1, vl); - - // load - vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl); - - vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl); - vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl); - - vuint8mf2_t x_at = __riscv_vand_vx_u8mf2(tx, 0x0F, vl); - vuint8mf2_t x_lt = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl); - - vuint8mf2_t x_a = __riscv_vor_vv_u8mf2(x_at, xh_0, vl); - vuint8mf2_t x_l = __riscv_vor_vv_u8mf2(x_lt, xh_1, vl); - - vint8mf2_t v0 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a); - vint8mf2_t v1 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l); - - vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl); - vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl); - - vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl); - - vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl); - vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl); - - int sumi = __riscv_vmv_x_s_i32m1_i32(vs2); - - sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s; - } - - *s = sumf; -#else - // scalar - float sumf = 0.0; - - for (int i = 0; i < nb; i++) { - uint32_t qh; - memcpy(&qh, x[i].qh, sizeof(qh)); - - int sumi = 0; - - for (int j = 0; j < qk/2; ++j) { - const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; - const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; - - const int32_t x0 = (x[i].qs[j] & 0xF) | xh_0; - const int32_t x1 = (x[i].qs[j] >> 4) | xh_1; - - sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]); - } - - sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s; - } - - *s = sumf; -#endif -} - -static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int qk = QK8_0; - const int nb = n / qk; - - assert(n % qk == 0); - - const block_q8_0 * restrict x = vx; - const block_q8_0 * restrict y = vy; - -#if defined(__ARM_NEON) - float32x4_t sumv0 = vdupq_n_f32(0.0f); - float32x4_t sumv1 = vdupq_n_f32(0.0f); - - GGML_ASSERT(nb % 2 == 0); // TODO: handle odd nb - for (int i = 0; i < nb; i += 2) { - const block_q8_0 * restrict x0 = &x[i + 0]; - const block_q8_0 * restrict x1 = &x[i + 1]; - const block_q8_0 * restrict y0 = &y[i + 0]; - const block_q8_0 * restrict y1 = &y[i + 1]; - - const int8x16_t x0_0 = vld1q_s8(x0->qs); - const int8x16_t x0_1 = vld1q_s8(x0->qs + 16); - const int8x16_t x1_0 = vld1q_s8(x1->qs); - const int8x16_t x1_1 = vld1q_s8(x1->qs + 16); - - // load y - const int8x16_t y0_0 = vld1q_s8(y0->qs); - const int8x16_t y0_1 = vld1q_s8(y0->qs + 16); - const int8x16_t y1_0 = vld1q_s8(y1->qs); - const int8x16_t y1_1 = vld1q_s8(y1->qs + 16); - -#if defined(__ARM_FEATURE_DOTPROD) - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( - vdotq_s32(vdupq_n_s32(0), x0_0, y0_0), - vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); - - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( - vdotq_s32(vdupq_n_s32(0), x1_0, y1_0), - vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); - -#else - const int16x8_t p0_0 = vmull_s8(vget_low_s8 (x0_0), vget_low_s8 (y0_0)); - const int16x8_t p0_1 = vmull_s8(vget_high_s8(x0_0), vget_high_s8(y0_0)); - const int16x8_t p0_2 = vmull_s8(vget_low_s8 (x0_1), vget_low_s8 (y0_1)); - const int16x8_t p0_3 = vmull_s8(vget_high_s8(x0_1), vget_high_s8(y0_1)); - - const int16x8_t p1_0 = vmull_s8(vget_low_s8 (x1_0), vget_low_s8 (y1_0)); - const int16x8_t p1_1 = vmull_s8(vget_high_s8(x1_0), vget_high_s8(y1_0)); - const int16x8_t p1_2 = vmull_s8(vget_low_s8 (x1_1), vget_low_s8 (y1_1)); - const int16x8_t p1_3 = vmull_s8(vget_high_s8(x1_1), vget_high_s8(y1_1)); - - const int32x4_t p0 = vaddq_s32(vpaddlq_s16(p0_0), vpaddlq_s16(p0_1)); - const int32x4_t p1 = vaddq_s32(vpaddlq_s16(p0_2), vpaddlq_s16(p0_3)); - const int32x4_t p2 = vaddq_s32(vpaddlq_s16(p1_0), vpaddlq_s16(p1_1)); - const int32x4_t p3 = vaddq_s32(vpaddlq_s16(p1_2), vpaddlq_s16(p1_3)); - - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(p0, p1)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(p2, p3)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); -#endif - } - - *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); -#elif defined(__AVX2__) || defined(__AVX__) - // Initialize accumulator with zeros - __m256 acc = _mm256_setzero_ps(); - - // Main loop - for (int i = 0; i < nb; ++i) { - // Compute combined scale for the block - const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d)); - __m256i bx = _mm256_loadu_si256((const __m256i *)x[i].qs); - __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); - - const __m256 q = mul_sum_i8_pairs_float(bx, by); - - // Multiply q with scale and accumulate -#if defined(__AVX2__) - acc = _mm256_fmadd_ps( d, q, acc ); -#else - acc = _mm256_add_ps( _mm256_mul_ps( d, q ), acc ); -#endif - } - - *s = hsum_float_8(acc); -#elif defined(__riscv_v_intrinsic) - float sumf = 0.0; - size_t vl = __riscv_vsetvl_e8m1(qk); - - for (int i = 0; i < nb; i++) { - // load elements - vint8m1_t bx = __riscv_vle8_v_i8m1(x[i].qs, vl); - vint8m1_t by = __riscv_vle8_v_i8m1(y[i].qs, vl); - - vint16m2_t vw_mul = __riscv_vwmul_vv_i16m2(bx, by, vl); - - vint32m1_t v_zero = __riscv_vmv_v_x_i32m1(0, vl); - vint32m1_t v_sum = __riscv_vwredsum_vs_i16m2_i32m1(vw_mul, v_zero, vl); - - int sumi = __riscv_vmv_x_s_i32m1_i32(v_sum); - - sumf += sumi*(GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d)); - } - - *s = sumf; -#else - // scalar - float sumf = 0.0; - - for (int i = 0; i < nb; i++) { - int sumi = 0; - - for (int j = 0; j < qk; j++) { - sumi += x[i].qs[j]*y[i].qs[j]; - } - - sumf += sumi*(GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d)); - } - - *s = sumf; -#endif -} - // compute GGML_VEC_DOT_UNROLL dot products at once // xs - x row stride in bytes inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * restrict s, void * restrict xv, ggml_fp16_t * restrict y) { @@ -3848,7 +1363,7 @@ inline static float ggml_gelu_f32(float x) { inline static void ggml_vec_gelu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { const uint16_t * i16 = (const uint16_t *) x; for (int i = 0; i < n; ++i) { - y[i] = table_gelu_f16[i16[i]]; + y[i] = ggml_table_gelu_f16[i16[i]]; } } @@ -3858,7 +1373,7 @@ inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) { ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]); memcpy(&t, &fp16, sizeof(uint16_t)); - y[i] = GGML_FP16_TO_FP32(table_gelu_f16[t]); + y[i] = GGML_FP16_TO_FP32(ggml_table_gelu_f16[t]); } } #else @@ -3876,7 +1391,7 @@ inline static float ggml_gelu_quick_f32(float x) { //inline static void ggml_vec_gelu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { // const uint16_t * i16 = (const uint16_t *) x; // for (int i = 0; i < n; ++i) { -// y[i] = table_gelu_quick_f16[i16[i]]; +// y[i] = ggml_table_gelu_quick_f16[i16[i]]; // } //} @@ -3886,7 +1401,7 @@ inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float * for (int i = 0; i < n; ++i) { ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]); memcpy(&t, &fp16, sizeof(uint16_t)); - y[i] = GGML_FP16_TO_FP32(table_gelu_quick_f16[t]); + y[i] = GGML_FP16_TO_FP32(ggml_table_gelu_quick_f16[t]); } } #else @@ -3905,7 +1420,7 @@ inline static float ggml_silu_f32(float x) { //inline static void ggml_vec_silu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { // const uint16_t * i16 = (const uint16_t *) x; // for (int i = 0; i < n; ++i) { -// y[i] = table_silu_f16[i16[i]]; +// y[i] = ggml_table_silu_f16[i16[i]]; // } //} @@ -3915,7 +1430,7 @@ inline static void ggml_vec_silu_f32(const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) { ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]); memcpy(&t, &fp16, sizeof(uint16_t)); - y[i] = GGML_FP16_TO_FP32(table_silu_f16[t]); + y[i] = GGML_FP16_TO_FP32(ggml_table_silu_f16[t]); } } #else @@ -4057,16 +1572,17 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "ALIBI", "CLAMP", "CONV_1D", + "CONV_1D_STAGE_0", + "CONV_1D_STAGE_1", "CONV_TRANSPOSE_1D", "CONV_2D", + "CONV_2D_STAGE_0", + "CONV_2D_STAGE_1", "CONV_TRANSPOSE_2D", "POOL_1D", "POOL_2D", "UPSCALE", - "CONV_1D_STAGE_0", - "CONV_1D_STAGE_1", - "FLASH_ATTN", "FLASH_FF", "FLASH_ATTN_BACK", @@ -4092,7 +1608,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "CROSS_ENTROPY_LOSS_BACK", }; -static_assert(GGML_OP_COUNT == 71, "GGML_OP_COUNT != 71"); +static_assert(GGML_OP_COUNT == 73, "GGML_OP_COUNT != 73"); static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "none", @@ -4143,16 +1659,17 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "alibi(x)", "clamp(x)", "conv_1d(x)", + "conv_1d_stage_0(x)", + "conv_1d_stage_1(x)", "conv_transpose_1d(x)", "conv_2d(x)", + "conv_2d_stage_0(x)", + "conv_2d_stage_1(x)", "conv_transpose_2d(x)", "pool_1d(x)", "pool_2d(x)", "upscale(x)", - "conv_1d_stage_0(x)", - "conv_1d_stage_1(x)", - "flash_attn(x)", "flash_ff(x)", "flash_attn_back(x)", @@ -4178,7 +1695,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "cross_entropy_loss_back(x,y)", }; -static_assert(GGML_OP_COUNT == 71, "GGML_OP_COUNT != 71"); +static_assert(GGML_OP_COUNT == 73, "GGML_OP_COUNT != 73"); static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2"); @@ -4209,8 +1726,10 @@ static void ggml_setup_op_has_task_pass(void) { p[GGML_OP_CONV_1D ] = true; p[GGML_OP_CONV_1D_STAGE_0 ] = true; p[GGML_OP_CONV_1D_STAGE_1 ] = true; - p[GGML_OP_CONV_2D ] = true; p[GGML_OP_CONV_TRANSPOSE_1D ] = true; + p[GGML_OP_CONV_2D ] = true; + p[GGML_OP_CONV_2D_STAGE_0 ] = true; + p[GGML_OP_CONV_2D_STAGE_1 ] = true; p[GGML_OP_CONV_TRANSPOSE_2D ] = true; p[GGML_OP_FLASH_ATTN_BACK ] = true; p[GGML_OP_CROSS_ENTROPY_LOSS ] = true; @@ -4627,11 +2146,11 @@ struct ggml_context * ggml_init(struct ggml_init_params params) { for (int i = 0; i < (1 << 16); ++i) { uint16_t ui = i; memcpy(&ii, &ui, sizeof(ii)); - const float f = table_f32_f16[i] = GGML_COMPUTE_FP16_TO_FP32(ii); - table_gelu_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_f32(f)); - table_gelu_quick_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_quick_f32(f)); - table_silu_f16[i] = GGML_FP32_TO_FP16(ggml_silu_f32(f)); - table_exp_f16[i] = GGML_FP32_TO_FP16(expf(f)); + const float f = ggml_table_f32_f16[i] = GGML_COMPUTE_FP16_TO_FP32(ii); + ggml_table_gelu_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_f32(f)); + ggml_table_gelu_quick_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_quick_f32(f)); + ggml_table_silu_f16[i] = GGML_FP32_TO_FP16(ggml_silu_f32(f)); + ggml_table_exp_f16[i] = GGML_FP32_TO_FP16(expf(f)); } const uint64_t t_end = ggml_time_us(); UNUSED(t_end); @@ -5634,7 +3153,7 @@ static struct ggml_tensor * ggml_add_cast_impl( // TODO: support less-strict constraint // GGML_ASSERT(ggml_can_repeat(b, a)); GGML_ASSERT(ggml_can_repeat_rows(b, a)); - GGML_ASSERT(ggml_is_quantized(a->type)); // currently only supported for quantized input + GGML_ASSERT(ggml_is_quantized(a->type) || a->type == GGML_TYPE_F16); // currently only supported for quantized input and f16 bool is_node = false; @@ -5954,7 +3473,6 @@ struct ggml_tensor * ggml_sqrt_inplace( return ggml_sqrt_impl(ctx, a, true); } - // ggml_log static struct ggml_tensor * ggml_log_impl( @@ -6008,7 +3526,6 @@ struct ggml_tensor * ggml_sum( return result; } - // ggml_sum_rows struct ggml_tensor * ggml_sum_rows( @@ -6640,7 +4157,6 @@ struct ggml_tensor * ggml_set_2d_inplace( return ggml_set_impl(ctx, a, b, nb1, a->nb[2], a->nb[3], offset, false); } - // ggml_cpy static struct ggml_tensor * ggml_cpy_impl( @@ -6720,7 +4236,6 @@ struct ggml_tensor * ggml_cont_inplace( return ggml_cont_impl(ctx, a, true); } - // make contiguous, with new shape GGML_API struct ggml_tensor * ggml_cont_1d( struct ggml_context * ctx, @@ -7173,7 +4688,6 @@ struct ggml_tensor * ggml_diag( return result; } - // ggml_diag_mask_inf static struct ggml_tensor * ggml_diag_mask_inf_impl( @@ -7285,7 +4799,6 @@ struct ggml_tensor * ggml_soft_max_inplace( return ggml_soft_max_impl(ctx, a, true); } - // ggml_soft_max_back static struct ggml_tensor * ggml_soft_max_back_impl( @@ -7702,7 +5215,11 @@ GGML_API struct ggml_tensor * ggml_conv_transpose_1d( // ggml_conv_2d -struct ggml_tensor * ggml_conv_2d( +// im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW] +// a: [OC，IC, KH, KW] +// b: [N, IC, IH, IW] +// result: [N, OH, OW, IC*KH*KW] +static struct ggml_tensor * ggml_conv_2d_stage_0( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, @@ -7721,17 +5238,21 @@ struct ggml_tensor * ggml_conv_2d( is_node = true; } + const int64_t OH = ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1); + const int64_t OW = ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0); + const int64_t ne[4] = { - ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0), - ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1), - a->ne[3], b->ne[3], + a->ne[2] * a->ne[1] * a->ne[0], + OW, + OH, + b->ne[3], }; - struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); + struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F16, 4, ne); int32_t params[] = { s0, s1, p0, p1, d0, d1 }; ggml_set_op_params(result, params, sizeof(params)); - result->op = GGML_OP_CONV_2D; + result->op = GGML_OP_CONV_2D_STAGE_0; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; result->src[1] = b; @@ -7740,8 +5261,61 @@ struct ggml_tensor * ggml_conv_2d( } -// ggml_conv_2d_sk_p0 +// gemm: [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW] +// a: [OC, IC, KH, KW] +// b: [N, OH, OW, IC * KH * KW] +// result: [N, OC, OH, OW] +static struct ggml_tensor * ggml_conv_2d_stage_1( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b) { + bool is_node = false; + + if (a->grad || b->grad) { + GGML_ASSERT(false); // TODO: implement backward + is_node = true; + } + + const int64_t ne[4] = { + b->ne[1], + b->ne[2], + a->ne[3], + b->ne[3], + }; + struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); + + result->op = GGML_OP_CONV_2D_STAGE_1; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src[0] = a; + result->src[1] = b; + + return result; + +} + +// a: [OC，IC, KH, KW] +// b: [N, IC, IH, IW] +// result: [N, OC, OH, OW] +struct ggml_tensor * ggml_conv_2d( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int s0, + int s1, + int p0, + int p1, + int d0, + int d1) { + + struct ggml_tensor * result = ggml_conv_2d_stage_0(ctx, a, b, s0, s1, p0, p1, d0, d1); // [N, OH, OW, IC * KH * KW] + result = ggml_conv_2d_stage_1(ctx, a, result); + + return result; + +} + +// ggml_conv_2d_sk_p0 struct ggml_tensor * ggml_conv_2d_sk_p0( struct ggml_context * ctx, struct ggml_tensor * a, @@ -8180,7 +5754,6 @@ static struct ggml_tensor * ggml_add_rel_pos_impl( return result; } - struct ggml_tensor * ggml_add_rel_pos( struct ggml_context * ctx, struct ggml_tensor * a, @@ -8625,8 +6198,6 @@ struct ggml_tensor * ggml_map_custom3_inplace( return ggml_map_custom3_impl(ctx, a, b, c, fun, n_tasks, userdata, true); } - - // ggml_cross_entropy_loss struct ggml_tensor * ggml_cross_entropy_loss( @@ -9356,9 +6927,15 @@ static void ggml_compute_forward_add_f16_f32( GGML_ASSERT(src0->type == GGML_TYPE_F16); GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT(dst->type == GGML_TYPE_F16); - GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); + if (dst->type == GGML_TYPE_F32) { + GGML_ASSERT( nb0 == sizeof(float)); + } + else { + GGML_ASSERT(dst->type == GGML_TYPE_F16); + GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); + } + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); // rows per thread @@ -9369,18 +6946,35 @@ static void ggml_compute_forward_add_f16_f32( const int ir1 = MIN(ir0 + dr, nr); if (nb10 == sizeof(float)) { - for (int ir = ir0; ir < ir1; ++ir) { - // src0, src1 and dst are same shape => same indices - const int i3 = ir/(ne2*ne1); - const int i2 = (ir - i3*ne2*ne1)/ne1; - const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + if (dst->type == GGML_TYPE_F16) { + for (int ir = ir0; ir < ir1; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); - ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); - ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); - float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11); + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); + ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11); - for (int i = 0; i < ne0; i++) { - dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + src1_ptr[i]); + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + src1_ptr[i]); + } + } + } else { + for (int ir = ir0; ir < ir1; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + float * dst_ptr = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); + ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11); + + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP16_TO_FP32(src0_ptr[i]) + src1_ptr[i]; + } } } } @@ -9828,7 +7422,6 @@ static void ggml_compute_forward_add1( } } - // ggml_compute_forward_acc static void ggml_compute_forward_acc_f32( @@ -9968,7 +7561,6 @@ static void ggml_compute_forward_sub_f32( const int i2 = (ir - i3*ne2*ne1)/ne1; const int i1 = (ir - i3*ne2*ne1 - i2*ne1); - #ifdef GGML_USE_ACCELERATE vDSP_vsub( (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1, @@ -10149,7 +7741,6 @@ static void ggml_compute_forward_div_f32( const int i2 = (ir - i3*ne2*ne1)/ne1; const int i1 = (ir - i3*ne2*ne1 - i2*ne1); - #ifdef GGML_USE_ACCELERATE UNUSED(ggml_vec_div_f32); @@ -10287,7 +7878,6 @@ static void ggml_compute_forward_sqrt( } } - // ggml_compute_forward_log static void ggml_compute_forward_log_f32( @@ -12120,7 +9710,6 @@ static void ggml_compute_forward_out_prod_f32( } } - //int64_t t1 = ggml_perf_time_us(); //static int64_t acc = 0; //acc += t1 - t0; @@ -12316,7 +9905,6 @@ static void ggml_compute_forward_scale_f32( const size_t nb1 = dst->nb[1]; - for (int i1 = ir0; i1 < ir1; i1++) { if (dst->data != src0->data) { // src0 is same shape as dst => same indices @@ -12714,7 +10302,6 @@ static void ggml_compute_forward_get_rows_back_f32( } } - static void ggml_compute_forward_get_rows_back( const struct ggml_compute_params * params, const struct ggml_tensor * src0, @@ -12949,7 +10536,7 @@ static void ggml_compute_forward_soft_max_f32( // const float val = (sp[i] == -INFINITY) ? 0.0 : exp(sp[i] - max); ggml_fp16_t s = GGML_FP32_TO_FP16(sp[i] - max); memcpy(&scvt, &s, sizeof(scvt)); - const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt]); + const float val = GGML_FP16_TO_FP32(ggml_table_exp_f16[scvt]); sum += (ggml_float)val; dp[i] = val; } @@ -13997,6 +11584,7 @@ static void ggml_compute_forward_conv_1d_f32( } } +// TODO: reuse ggml_mul_mat or implement ggml_im2col and remove stage_0 and stage_1 static void gemm_f16_out_f32(int64_t m, int64_t n, int64_t k, ggml_fp16_t * A, ggml_fp16_t * B, @@ -14298,6 +11886,9 @@ static void ggml_compute_forward_conv_transpose_1d_f16_f32( } } + // need to zero dst since we are accumulating into it + memset(dst->data, 0, ggml_nbytes(dst)); + return; } @@ -14370,7 +11961,7 @@ static void ggml_compute_forward_conv_transpose_1d_f32( const float * const src = (float *)((char *) src0->data + i02*nb02 + i01*nb01); float * dst_data = wdata + i01*ne00*ne02; for (int64_t i00 = 0; i00 < ne00; i00++) { - dst_data[i01*ne00*ne02 + i00*ne02 + i02] = src[i00]; + dst_data[i00*ne02 + i02] = src[i00]; } } } @@ -14389,6 +11980,9 @@ static void ggml_compute_forward_conv_transpose_1d_f32( } } + // need to zero dst since we are accumulating into it + memset(dst->data, 0, ggml_nbytes(dst)); + return; } @@ -14450,6 +12044,144 @@ static void ggml_compute_forward_conv_transpose_1d( // ggml_compute_forward_conv_2d +// src0: kernel [OC, IC, KH, KW] +// src1: image [N, IC, IH, IW] +// dst: result [N, OH, OW, IC*KH*KW] +static void ggml_compute_forward_conv_2d_stage_0_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F16); + + int64_t t0 = ggml_perf_time_us(); + UNUSED(t0); + + GGML_TENSOR_BINARY_OP_LOCALS; + + const int64_t N = ne13; + const int64_t IC = ne12; + const int64_t IH = ne11; + const int64_t IW = ne10; + + // const int64_t OC = ne03; + // const int64_t IC = ne02; + const int64_t KH = ne01; + const int64_t KW = ne00; + + const int64_t OH = ne2; + const int64_t OW = ne1; + + const int ith = params->ith; + const int nth = params->nth; + + const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; + const int32_t s1 = ((const int32_t*)(dst->op_params))[1]; + const int32_t p0 = ((const int32_t*)(dst->op_params))[2]; + const int32_t p1 = ((const int32_t*)(dst->op_params))[3]; + const int32_t d0 = ((const int32_t*)(dst->op_params))[4]; + const int32_t d1 = ((const int32_t*)(dst->op_params))[5]; + + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb10 == sizeof(float)); + + if (params->type == GGML_TASK_INIT) { + memset(dst->data, 0, ggml_nbytes(dst)); + return; + } + + if (params->type == GGML_TASK_FINALIZE) { + return; + } + + // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW] + { + ggml_fp16_t * const wdata = (ggml_fp16_t *) dst->data; + + for (int64_t in = 0; in < N; in++) { + for (int64_t ioh = 0; ioh < OH; ioh++) { + for (int64_t iow = 0; iow < OW; iow++) { + for (int64_t iic = ith; iic < IC; iic+=nth) { + + // micro kernel + ggml_fp16_t * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW] + const float * const src_data = (float *)((char *) src1->data + in*nb13 + iic*nb12); // [IH, IW] + + for (int64_t ikh = 0; ikh < KH; ikh++) { + for (int64_t ikw = 0; ikw < KW; ikw++) { + const int64_t iiw = iow*s0 + ikw*d0 - p0; + const int64_t iih = ioh*s1 + ikh*d1 - p1; + + if (!(iih < 0 || iih >= IH || iiw < 0 || iiw >= IW)) { + dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_FP32_TO_FP16(src_data[iih*IW + iiw]); + } + } + } + } + } + } + } + } +} + +// gemm: [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW] +// src0: [OC, IC, KH, KW] +// src1: [N, OH, OW, IC * KH * KW] +// result: [N, OC, OH, OW] +static void ggml_compute_forward_conv_2d_stage_1_f16( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F16); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + int64_t t0 = ggml_perf_time_us(); + UNUSED(t0); + + if (params->type == GGML_TASK_INIT) { + return; + } + + if (params->type == GGML_TASK_FINALIZE) { + return; + } + + GGML_TENSOR_BINARY_OP_LOCALS; + + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb10 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb0 == sizeof(float)); + + const int N = ne13; + const int OH = ne12; + const int OW = ne11; + + const int OC = ne03; + const int IC = ne02; + const int KH = ne01; + const int KW = ne00; + + const int ith = params->ith; + const int nth = params->nth; + + int64_t m = OC; + int64_t n = OH * OW; + int64_t k = IC * KH * KW; + + // [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW] + for (int i = 0; i < N; i++) { + ggml_fp16_t * A = (ggml_fp16_t *)src0->data; // [m, k] + ggml_fp16_t * B = (ggml_fp16_t *)src1->data + i * m * k; // [n, k] + float * C = (float *)dst->data + i * m * n; // [m, n] + + gemm_f16_out_f32(m, n, k, A, B, C, ith, nth); + } +} + static void ggml_compute_forward_conv_2d_f16_f32( const struct ggml_compute_params * params, const struct ggml_tensor * src0, @@ -14462,16 +12194,40 @@ static void ggml_compute_forward_conv_2d_f16_f32( int64_t t0 = ggml_perf_time_us(); UNUSED(t0); - GGML_TENSOR_BINARY_OP_LOCALS; + GGML_TENSOR_BINARY_OP_LOCALS + + // src1: image [N, IC, IH, IW] + // src0: kernel [OC, IC, KH, KW] + // dst: result [N, OC, OH, OW] + // ne12: IC + // ne0: OW + // ne1: OH + // nk0: KW + // nk1: KH + // ne13: N + + const int N = ne13; + const int IC = ne12; + const int IH = ne11; + const int IW = ne10; + + const int OC = ne03; + // const int IC = ne02; + const int KH = ne01; + const int KW = ne00; + + const int OH = ne1; + const int OW = ne0; const int ith = params->ith; const int nth = params->nth; - const int nk0 = ne00; - const int nk1 = ne01; + // const int nk0 = ne00; + // const int nk1 = ne01; // size of the convolution row - the kernel size unrolled across all channels - const int ew0 = nk0*nk1*ne02; + // const int ew0 = nk0*nk1*ne02; + // ew0: IC*KH*KW const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; const int32_t s1 = ((const int32_t*)(dst->op_params))[1]; @@ -14487,24 +12243,27 @@ static void ggml_compute_forward_conv_2d_f16_f32( memset(params->wdata, 0, params->wsize); // prepare source data (src1) + // im2col: [N, IC, IH, IW] => [N*OH*OW, IC*KH*KW] + { ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; - for (int i13 = 0; i13 < ne13; i13++) { - for (int i12 = 0; i12 < ne12; i12++) { - const float * const src = (float *)((char *) src1->data + i13*nb13 + i12*nb12); - ggml_fp16_t * dst_data = wdata + i13*(ne1*ne0*ew0); + for (int in = 0; in < N; in++) { + for (int iic = 0; iic < IC; iic++) { + for (int ioh = 0; ioh < OH; ioh++) { + for (int iow = 0; iow < OW; iow++) { - for (int i1 = 0; i1 < ne1; i1++) { - for (int i0 = 0; i0 < ne0; i0++) { - for (int ik1 = 0; ik1 < nk1; ik1++) { - for (int ik0 = 0; ik0 < nk0; ik0++) { - const int idx0 = i0*s0 + ik0*d0 - p0; - const int idx1 = i1*s1 + ik1*d1 - p1; + // micro kernel + ggml_fp16_t * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW] + const float * const src_data = (float *)((char *) src1->data + in*nb13 + iic*nb12); // [IH, IW] - if (!(idx1 < 0 || idx1 >= ne11 || idx0 < 0 || idx0 >= ne10)) { - dst_data[(i1*ne0 + i0)*ew0 + i12*(nk0*nk1) + ik1*nk0 + ik0] = - GGML_FP32_TO_FP16(src[idx1*ne10 + idx0]); + for (int ikh = 0; ikh < KH; ikh++) { + for (int ikw = 0; ikw < KW; ikw++) { + const int iiw = iow*s0 + ikw*d0 - p0; + const int iih = ioh*s1 + ikh*d1 - p1; + + if (!(iih < 0 || iih >= IH || iiw < 0 || iiw >= IW)) { + dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_FP32_TO_FP16(src_data[iih*IW + iiw]); } } } @@ -14521,30 +12280,22 @@ static void ggml_compute_forward_conv_2d_f16_f32( return; } - // total patches in dst - const int np = ne2; - - // patches per thread - const int dp = (np + nth - 1)/nth; - - // patch range for this thread - const int ip0 = dp*ith; - const int ip1 = MIN(ip0 + dp, np); - ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; + // wdata: [N*OH*OW, IC*KH*KW] + // dst: result [N, OC, OH, OW] + // src0: kernel [OC, IC, KH, KW] - for (int i3 = 0; i3 < ne3; i3++) { - for (int i2 = ip0; i2 < ip1; i2++) { - float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2); + int64_t m = OC; + int64_t n = OH * OW; + int64_t k = IC * KH * KW; - for (int i1 = 0; i1 < ne1; ++i1) { - for (int i0 = 0; i0 < ne0; ++i0) { - ggml_vec_dot_f16(ew0, dst_data + i1*ne0 + i0, - (ggml_fp16_t *) ((char *) src0->data + i2*nb03), - (ggml_fp16_t *) wdata + i3*nb3 + (i1*ne0 + i0)*ew0); - } - } - } + // [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW] + for (int i = 0; i < N; i++) { + ggml_fp16_t * A = (ggml_fp16_t *)src0->data; // [m, k] + ggml_fp16_t * B = (ggml_fp16_t *)wdata + i * m * k; // [n, k] + float * C = (float *)dst->data + i * m * n; // [m * k] + + gemm_f16_out_f32(m, n, k, A, B, C, ith, nth); } } @@ -14570,6 +12321,48 @@ static void ggml_compute_forward_conv_2d( } } +static void ggml_compute_forward_conv_2d_stage_0( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F16: + { + ggml_compute_forward_conv_2d_stage_0_f32(params, src0, src1, dst); + } break; + case GGML_TYPE_F32: + { + GGML_ASSERT(false); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + +static void ggml_compute_forward_conv_2d_stage_1( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F16: + { + ggml_compute_forward_conv_2d_stage_1_f16(params, src0, src1, dst); + } break; + case GGML_TYPE_F32: + { + GGML_ASSERT(false); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + // ggml_compute_forward_conv_transpose_2d static void ggml_compute_forward_conv_transpose_2d( @@ -14628,6 +12421,8 @@ static void ggml_compute_forward_conv_transpose_2d( } } + memset(dst->data, 0, ggml_nbytes(dst)); + return; } @@ -15030,7 +12825,7 @@ static void ggml_compute_forward_flash_attn_f32( #else ggml_fp16_t s = GGML_FP32_TO_FP16(SS[j] - max); memcpy(&scvt[j], &s, sizeof(uint16_t)); - const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt[j]]); + const float val = GGML_FP16_TO_FP32(ggml_table_exp_f16[scvt[j]]); #endif sump[j] += (ggml_float)val; SS[j] = val; @@ -15232,7 +13027,7 @@ static void ggml_compute_forward_flash_attn_f16( } else { ggml_fp16_t s = GGML_FP32_TO_FP16(SS[j] - max); memcpy(&scvt[j], &s, sizeof(uint16_t)); - const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt[j]]); + const float val = GGML_FP16_TO_FP32(ggml_table_exp_f16[scvt[j]]); sump[j] += (ggml_float)val; SS[j] = val; } @@ -15683,7 +13478,7 @@ static void ggml_compute_forward_flash_attn_back_f32( #else ggml_fp16_t s = GGML_FP32_TO_FP16(SR[j] - max); memcpy(&scvt[j], &s, sizeof(uint16_t)); - const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt[j]]); + const float val = GGML_FP16_TO_FP32(ggml_table_exp_f16[scvt[j]]); #endif sump[j] += (ggml_float)val; SW[j] = val; @@ -16126,7 +13921,6 @@ static void ggml_compute_forward_add_rel_pos_f32( const int ip0 = dp*ith; const int ip1 = MIN(ip0 + dp, np); - for (int64_t i13 = ip0; i13 < ip1; ++i13) { for (int64_t i12 = 0; i12 < ne12; ++i12) { for (int64_t i11 = 0; i11 < ne11; ++i11) { @@ -16193,7 +13987,6 @@ static void ggml_compute_forward_map_unary_f32( } } - static void ggml_compute_forward_map_unary( const struct ggml_compute_params * params, const struct ggml_tensor * src0, @@ -16241,7 +14034,6 @@ static void ggml_compute_forward_map_binary_f32( } } - static void ggml_compute_forward_map_binary( const struct ggml_compute_params * params, const struct ggml_tensor * src0, @@ -16293,7 +14085,6 @@ static void ggml_compute_forward_map_custom2_f32( fun(dst, a, b); } - // ggml_compute_forward_map_custom3 static void ggml_compute_forward_map_custom3_f32( @@ -16437,7 +14228,7 @@ static void ggml_compute_forward_cross_entropy_loss_f32( #else ggml_fp16_t s = GGML_FP32_TO_FP16(s0[i] - max); memcpy(&scvt, &s, sizeof(scvt)); - const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt]); + const float val = GGML_FP16_TO_FP32(ggml_table_exp_f16[scvt]); #endif sum += (ggml_float)val; st[i] = val; @@ -16551,7 +14342,7 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32( #else ggml_fp16_t s = GGML_FP32_TO_FP16(s0[i] - max); memcpy(&scvt, &s, sizeof(scvt)); - const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt]); + const float val = GGML_FP16_TO_FP32(ggml_table_exp_f16[scvt]); #endif sum += (ggml_float)val; ds0[i] = val; @@ -16568,7 +14359,6 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32( ggml_vec_sub_f32(nc, ds0, ds0, s1); ggml_vec_scale_f32(nc, ds0, d[0] / (float) nr); - #ifndef NDEBUG for (int i = 0; i < nc; ++i) { assert(!isnan(ds0[i])); @@ -16596,12 +14386,15 @@ static void ggml_compute_forward_cross_entropy_loss_back( } } - ///////////////////////////////// static void ggml_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) { GGML_ASSERT(params); + if (tensor->op == GGML_OP_NONE) { + return; + } + #ifdef GGML_USE_CUBLAS bool skip_cpu = ggml_cuda_compute_forward(params, tensor); if (skip_cpu) { @@ -16804,6 +14597,14 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_conv_2d(params, tensor->src[0], tensor->src[1], tensor); } break; + case GGML_OP_CONV_2D_STAGE_0: + { + ggml_compute_forward_conv_2d_stage_0(params, tensor->src[0], tensor->src[1], tensor); + } break; + case GGML_OP_CONV_2D_STAGE_1: + { + ggml_compute_forward_conv_2d_stage_1(params, tensor->src[0], tensor->src[1], tensor); + } break; case GGML_OP_CONV_TRANSPOSE_2D: { ggml_compute_forward_conv_transpose_2d(params, tensor->src[0], tensor->src[1], tensor); @@ -17733,11 +15534,19 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor { GGML_ASSERT(false); // TODO: not implemented } break; + case GGML_OP_CONV_TRANSPOSE_1D: + { + GGML_ASSERT(false); // TODO: not implemented + } break; case GGML_OP_CONV_2D: { GGML_ASSERT(false); // TODO: not implemented } break; - case GGML_OP_CONV_TRANSPOSE_1D: + case GGML_OP_CONV_2D_STAGE_0: + { + GGML_ASSERT(false); // TODO: not implemented + } break; + case GGML_OP_CONV_2D_STAGE_1: { GGML_ASSERT(false); // TODO: not implemented } break; @@ -18666,6 +16475,7 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) { const int64_t ne0 = node->ne[0]; const int64_t ne1 = node->ne[1]; const int64_t ne2 = node->ne[2]; + const int64_t ne3 = node->ne[3]; const int64_t nk = ne00*ne01; const int64_t ew0 = nk * ne02; @@ -18676,7 +16486,8 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) { if (node->src[0]->type == GGML_TYPE_F16 && node->src[1]->type == GGML_TYPE_F32) { - cur = sizeof(ggml_fp16_t)*(ne0*ne1*ew0); + // im2col: [N*OH*OW, IC*KH*KW] + cur = sizeof(ggml_fp16_t)*(ne3*ne0*ne1*ew0); } else if (node->src[0]->type == GGML_TYPE_F32 && node->src[1]->type == GGML_TYPE_F32) { cur = sizeof(float)* (ne10*ne11*ne12); @@ -18686,6 +16497,14 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) { work_size = MAX(work_size, cur); } break; + case GGML_OP_CONV_2D_STAGE_0: + { + n_tasks = n_threads; + } break; + case GGML_OP_CONV_2D_STAGE_1: + { + n_tasks = n_threads; + } break; case GGML_OP_CONV_TRANSPOSE_2D: { n_tasks = n_threads; @@ -19874,7 +17693,6 @@ static enum ggml_opt_result ggml_opt_adam( opt->loss_after = fx; - // check convergence if (fabsf(fx - fx_prev[0])/fx < params.adam.eps_f) { GGML_PRINT_DEBUG("converged\n"); @@ -20723,7 +18541,6 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i block_q8_0 * block = (block_q8_0*)dst + start / QK8_0; result = ggml_quantize_q8_0(src + start, block, n, n, hist); } break; -#ifdef GGML_USE_K_QUANTS case GGML_TYPE_Q2_K: { GGML_ASSERT(start % QK_K == 0); @@ -20754,7 +18571,6 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i block_q6_K * block = (block_q6_K*)dst + start / QK_K; result = ggml_quantize_q6_K(src + start, block, n, n, hist); } break; -#endif case GGML_TYPE_F16: { int elemsize = sizeof(ggml_fp16_t); diff --git a/ggml.h b/ggml.h index 16aaf169e..9d16c5a72 100644 --- a/ggml.h +++ b/ggml.h @@ -401,15 +401,16 @@ extern "C" { GGML_OP_ALIBI, GGML_OP_CLAMP, GGML_OP_CONV_1D, - GGML_OP_CONV_2D, + GGML_OP_CONV_1D_STAGE_0, // internal + GGML_OP_CONV_1D_STAGE_1, // internal GGML_OP_CONV_TRANSPOSE_1D, + GGML_OP_CONV_2D, + GGML_OP_CONV_2D_STAGE_0, // internal + GGML_OP_CONV_2D_STAGE_1, // internal GGML_OP_CONV_TRANSPOSE_2D, GGML_OP_POOL_1D, GGML_OP_POOL_2D, - GGML_OP_CONV_1D_STAGE_0, // internal - GGML_OP_CONV_1D_STAGE_1, // internal - GGML_OP_UPSCALE, // nearest interpolate GGML_OP_FLASH_ATTN, @@ -708,7 +709,7 @@ extern "C" { // Context tensor enumeration and lookup GGML_API struct ggml_tensor * ggml_get_first_tensor(struct ggml_context * ctx); GGML_API struct ggml_tensor * ggml_get_next_tensor (struct ggml_context * ctx, struct ggml_tensor * tensor); - GGML_API struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name); + GGML_API struct ggml_tensor * ggml_get_tensor (struct ggml_context * ctx, const char * name); GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor); GGML_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value); @@ -1020,9 +1021,9 @@ extern "C" { struct ggml_tensor * b, float eps); - // A: n columns, m rows - // B: n columns, p rows (i.e. we transpose it internally) - // result is m columns, p rows + // A: k columns, n rows => [ne03, ne02, n, k] + // B: k columns, m rows (i.e. we transpose it internally) => [ne03 * x, ne02 * y, m, k] + // result is n columns, m rows => [ne03 * x, ne02 * y, m, n] GGML_API struct ggml_tensor * ggml_mul_mat( struct ggml_context * ctx, struct ggml_tensor * a, @@ -1929,12 +1930,19 @@ extern "C" { // quantization // + // TODO: these would probably get removed in favor of the more general ggml_quantize_chunk GGML_API size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist); GGML_API size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist); GGML_API size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist); GGML_API size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * hist); GGML_API size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * hist); + GGML_API size_t ggml_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist); + GGML_API size_t ggml_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist); + GGML_API size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist); + GGML_API size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist); + GGML_API size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist); + GGML_API size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist); // diff --git a/gguf-py/gguf/gguf.py b/gguf-py/gguf/gguf.py index 072c839c4..6b7d65429 100644 --- a/gguf-py/gguf/gguf.py +++ b/gguf-py/gguf/gguf.py @@ -987,12 +987,15 @@ class SpecialVocab: merges: list[str] = [] special_token_types: tuple[str, ...] = ('bos', 'eos', 'unk', 'sep', 'pad') special_token_ids: dict[str, int] = {} + n_vocab: int | None = None def __init__( self, path: str | os.PathLike[str], load_merges: bool = False, special_token_types: tuple[str, ...] | None = None, + n_vocab: int | None = None, ): self.special_token_ids = {} + self.n_vocab = n_vocab self.load_merges = load_merges if special_token_types is not None: self.special_token_types = special_token_types @@ -1002,6 +1005,16 @@ class SpecialVocab: if not self._try_load_from_tokenizer_json(path): self._try_load_from_config_json(path) + def _set_special_token(self, typ: str, tid: Any): + if not isinstance(tid, int) or tid < 0: + return + if self.n_vocab is None or tid < self.n_vocab: + self.special_token_ids[typ] = tid + return + print(f'gguf: WARNING: Special token type {typ}, id {tid} out of range, must be under {self.n_vocab} - skipping', + file = sys.stderr) + + def _try_load_from_tokenizer_json(self, path: Path) -> bool: tokenizer_file = path / 'tokenizer.json' if not tokenizer_file.is_file(): @@ -1029,10 +1042,11 @@ class SpecialVocab: tc_content = entry_content else: continue - for maybe_token_id in (atok.get('id') for atok in added_tokens if atok.get('content') == tc_content): - if isinstance(maybe_token_id, int) and maybe_token_id >= 0: - self.special_token_ids[typ] = maybe_token_id - break + # We only need the first match here. + maybe_token_id = next(( + atok.get('id') for atok in added_tokens + if atok.get('content') == tc_content), None) + self._set_special_token(typ, maybe_token_id) return True def _try_load_from_config_json(self, path: Path) -> bool: @@ -1042,21 +1056,21 @@ class SpecialVocab: with open(config_file, encoding = 'utf-8') as f: config = json.load(f) for typ in self.special_token_types: - maybe_token_id = config.get(f'{typ}_token_id') - if isinstance(maybe_token_id, int) and maybe_token_id >= 0: - self.special_token_ids[typ] = maybe_token_id + self._set_special_token(typ, config.get(f'{typ}_token_id')) return True - def add_to_gguf(self, gw: GGUFWriter) -> None: + def add_to_gguf(self, gw: GGUFWriter, quiet: bool = False) -> None: if len(self.merges) > 0: - print(f'gguf: Adding {len(self.merges)} merge(s).') + if not quiet: + print(f'gguf: Adding {len(self.merges)} merge(s).') gw.add_token_merges(self.merges) for typ, tokid in self.special_token_ids.items(): handler: Callable[[int], None] | None = getattr(gw, f'add_{typ}_token_id', None) if handler is None: - print(f'gguf: WARNING: No handler for special token type {typ} with id {tokid} - skipping') + print(f'gguf: WARNING: No handler for special token type {typ} with id {tokid} - skipping', file = sys.stderr) continue - print(f'gguf: Setting special token type {typ} to {tokid}') + if not quiet: + print(f'gguf: Setting special token type {typ} to {tokid}') handler(tokid) def __repr__(self) -> str: diff --git a/llama.cpp b/llama.cpp index 365349335..42cedc7a1 100644 --- a/llama.cpp +++ b/llama.cpp @@ -19,13 +19,11 @@ #ifdef GGML_USE_MPI # include "ggml-mpi.h" #endif -#ifdef GGML_USE_K_QUANTS -# ifndef QK_K -# ifdef GGML_QKK_64 -# define QK_K 64 -# else -# define QK_K 256 -# endif +#ifndef QK_K +# ifdef GGML_QKK_64 +# define QK_K 64 +# else +# define QK_K 256 # endif #endif @@ -62,7 +60,9 @@ #include #include #include +#include #include +#include #include #include #include @@ -71,11 +71,10 @@ #include #include #include +#include #include #include #include -#include -#include #if defined(_MSC_VER) #pragma warning(disable: 4244 4267) // possible loss of data @@ -971,18 +970,19 @@ struct llama_mlock { typedef void (*offload_func_t)(struct ggml_tensor * tensor); -static void llama_nop(struct ggml_tensor * tensor) { // don't offload by default +static void ggml_offload_nop(struct ggml_tensor * tensor) { (void) tensor; } -static std::string llama_token_to_str(const struct llama_context * ctx, llama_token token) { +static std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) { std::vector result(8, 0); const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size()); if (n_tokens < 0) { result.resize(-n_tokens); int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size()); GGML_ASSERT(check == -n_tokens); - } else { + } + else { result.resize(n_tokens); } @@ -1114,13 +1114,13 @@ struct llama_layer { struct ggml_tensor * ffn_norm_b; // ff - struct ggml_tensor * w1; // ffn_gate - struct ggml_tensor * w2; // ffn_down - struct ggml_tensor * w3; // ffn_up + struct ggml_tensor * ffn_gate; // w1 + struct ggml_tensor * ffn_down; // w2 + struct ggml_tensor * ffn_up; // w3 // ff bias - struct ggml_tensor * b2; // ffn_down - struct ggml_tensor * b3; // ffn_up + struct ggml_tensor * ffn_down_b; // b2 + struct ggml_tensor * ffn_up_b; // b3 }; struct llama_kv_cell { @@ -1202,10 +1202,10 @@ struct llama_vocab { id special_eot_id = 32010; int find_bpe_rank(std::string token_left, std::string token_right) const { - replace_all(token_left, " ", "\u0120"); - replace_all(token_left, "\n", "\u010A"); - replace_all(token_right, " ", "\u0120"); - replace_all(token_right, "\n", "\u010A"); + GGML_ASSERT(token_left.find(" ") == std::string::npos); + GGML_ASSERT(token_left.find("\n") == std::string::npos); + GGML_ASSERT(token_right.find(" ") == std::string::npos); + GGML_ASSERT(token_right.find("\n") == std::string::npos); auto it = bpe_ranks.find(std::make_pair(token_left, token_right)); if (it == bpe_ranks.end()) { @@ -1226,8 +1226,8 @@ struct llama_model { llama_hparams hparams = {}; llama_vocab vocab; - struct ggml_tensor * tok_embeddings; - struct ggml_tensor * pos_embeddings; + struct ggml_tensor * tok_embd; + struct ggml_tensor * pos_embd; struct ggml_tensor * tok_norm; struct ggml_tensor * tok_norm_b; @@ -1467,17 +1467,12 @@ static int32_t llama_kv_cache_cell_max(const struct llama_kv_cache & cache) { return 0; } -static void llama_kv_cache_tokens_rm(struct llama_kv_cache & cache, int32_t c0, int32_t c1) { - if (c0 < 0) c0 = 0; - if (c1 < 0) c1 = cache.size; - - for (int32_t i = c0; i < c1; ++i) { +static void llama_kv_cache_clear(struct llama_kv_cache & cache) { + for (int32_t i = 0; i < (int32_t) cache.size; ++i) { cache.cells[i].pos = -1; cache.cells[i].seq_id.clear(); } - - // Searching for a free slot can start here since we know it will be empty. - cache.head = uint32_t(c0); + cache.head = 0; } static void llama_kv_cache_seq_rm( @@ -1491,8 +1486,14 @@ static void llama_kv_cache_seq_rm( if (p1 < 0) p1 = std::numeric_limits::max(); for (uint32_t i = 0; i < cache.size; ++i) { - if (cache.cells[i].has_seq_id(seq_id) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) { - cache.cells[i].seq_id.erase(seq_id); + if (cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) { + if (seq_id < 0) { + cache.cells[i].seq_id.clear(); + } else if (cache.cells[i].has_seq_id(seq_id)) { + cache.cells[i].seq_id.erase(seq_id); + } else { + continue; + } if (cache.cells[i].seq_id.empty()) { cache.cells[i].pos = -1; if (new_head == cache.size) new_head = i; @@ -1553,14 +1554,14 @@ static void llama_kv_cache_seq_shift( for (uint32_t i = 0; i < cache.size; ++i) { if (cache.cells[i].has_seq_id(seq_id) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) { - cache.cells[i].pos += delta; + cache.has_shift = true; + cache.cells[i].pos += delta; + cache.cells[i].delta += delta; + if (cache.cells[i].pos < 0) { cache.cells[i].pos = -1; cache.cells[i].seq_id.clear(); if (new_head == cache.size) new_head = i; - } else { - cache.has_shift = true; - cache.cells[i].delta = delta; } } } @@ -1577,12 +1578,14 @@ static void llama_kv_cache_seq_shift( enum llama_fver { GGUF_FILE_VERSION_V1 = 1, GGUF_FILE_VERSION_V2 = 2, + GGUF_FILE_VERSION_V3 = 3, }; static const char * llama_file_version_name(llama_fver version) { switch (version) { case GGUF_FILE_VERSION_V1: return "GGUF V1 (support until nov 2023)"; - case GGUF_FILE_VERSION_V2: return "GGUF V2 (latest)"; + case GGUF_FILE_VERSION_V2: return "GGUF V2"; + case GGUF_FILE_VERSION_V3: return "GGUF V3 (latest)"; } return "unknown"; @@ -2238,15 +2241,35 @@ static void llm_load_vocab( if (vocab.type == LLAMA_VOCAB_TYPE_SPM) { vocab.linefeed_id = llama_byte_to_token(vocab, '\n'); } else { - vocab.linefeed_id = llama_tokenize_internal(vocab, "\u010A", false)[0]; + const std::vector ids = llama_tokenize_internal(vocab, "\u010A", false); + GGML_ASSERT(!ids.empty() && "model vocab missing newline token"); + vocab.linefeed_id = ids[0]; } // special tokens - GGUF_GET_KEY(ctx, vocab.special_bos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_BOS_ID)); - GGUF_GET_KEY(ctx, vocab.special_eos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_EOS_ID)); - GGUF_GET_KEY(ctx, vocab.special_unk_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_UNK_ID)); - GGUF_GET_KEY(ctx, vocab.special_sep_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_SEP_ID)); - GGUF_GET_KEY(ctx, vocab.special_pad_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_PAD_ID)); + { + const std::vector> special_token_types = { + { LLM_KV_TOKENIZER_BOS_ID, vocab.special_bos_id }, + { LLM_KV_TOKENIZER_EOS_ID, vocab.special_eos_id }, + { LLM_KV_TOKENIZER_UNK_ID, vocab.special_unk_id }, + { LLM_KV_TOKENIZER_SEP_ID, vocab.special_sep_id }, + { LLM_KV_TOKENIZER_PAD_ID, vocab.special_pad_id }, + }; + for (const auto & it : special_token_types) { + const std::string & key = kv(std::get<0>(it)); + int32_t & id = std::get<1>(it), old_id = id; + + GGUF_GET_KEY(ctx, id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, key); + // Must be >= -1 and < vocab size. Since the key is unsigned, -1 + // can only come from the default value, so there's no point in + // validating that. + if (size_t(id + 1) > vocab.id_to_token.size()) { + LLAMA_LOG_WARN("%s: bad special token: '%s' = %d, using default id %d\n", + __func__, key.c_str(), id, old_id); + id = old_id; + } + } + } // build special tokens cache { @@ -2460,7 +2483,7 @@ static void llm_load_tensors( case LLM_ARCH_LLAMA: case LLM_ARCH_REFACT: { - model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); // output { @@ -2514,21 +2537,21 @@ static void llm_load_tensors( layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); - layer.w1 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split); - layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); - layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); if (backend == GGML_BACKEND_GPU) { vram_weights += - ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk) + - ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) + - ggml_nbytes(layer.w1) + ggml_nbytes(layer.w2) + ggml_nbytes(layer.w3); + ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk) + + ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) + + ggml_nbytes(layer.ffn_gate) + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up); } } } break; case LLM_ARCH_BAICHUAN: { - model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); { ggml_backend_type backend_norm; ggml_backend_type backend_output; @@ -2580,15 +2603,15 @@ static void llm_load_tensors( layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); - layer.w1 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split); - layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); - layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); if (backend == GGML_BACKEND_GPU) { vram_weights += - ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk) + - ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) + - ggml_nbytes(layer.w1) + ggml_nbytes(layer.w2) + ggml_nbytes(layer.w3); + ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk) + + ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) + + ggml_nbytes(layer.ffn_gate) + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up); } } } break; @@ -2596,7 +2619,7 @@ static void llm_load_tensors( { // TODO: CPU-only for now - model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); // output { @@ -2659,21 +2682,21 @@ static void llm_load_tensors( layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); - layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); - layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); if (backend == GGML_BACKEND_GPU) { vram_weights += ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.attn_norm_b) + ggml_nbytes(layer.wqkv) + ggml_nbytes(layer.wo) + - ggml_nbytes(layer.w2) + ggml_nbytes(layer.w3); + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up); } } } break; case LLM_ARCH_STARCODER: { - model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); - model.pos_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, hparams.n_ctx_train}, GGML_BACKEND_CPU); + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + model.pos_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, hparams.n_ctx_train}, GGML_BACKEND_CPU); // output { @@ -2724,19 +2747,19 @@ static void llm_load_tensors( layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend); layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); - layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, backend_split); + layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, backend); layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); - layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend_split); + layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend); layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, backend); - layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split); - layer.b2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, backend_split); + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split); + layer.ffn_down_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, backend); - layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); - layer.b3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend_split); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend); if (backend == GGML_BACKEND_GPU) { vram_weights += @@ -2744,14 +2767,14 @@ static void llm_load_tensors( ggml_nbytes(layer.wqkv) + ggml_nbytes(layer.bqkv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.bo) + ggml_nbytes(layer.ffn_norm) + ggml_nbytes(layer.ffn_norm_b) + - ggml_nbytes(layer.w2) + ggml_nbytes(layer.b2) + - ggml_nbytes(layer.w3) + ggml_nbytes(layer.b3); + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_down_b) + + ggml_nbytes(layer.ffn_up) + ggml_nbytes(layer.ffn_up_b); } } } break; case LLM_ARCH_PERSIMMON: { - model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); { ggml_backend_type backend_norm; @@ -2792,31 +2815,31 @@ static void llm_load_tensors( const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; auto & layer = model.layers[i]; - layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); - layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend); - layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); - layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, backend_split); - layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); - layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend_split); - layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split); - layer.b2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, backend_split); - layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); - layer.b3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend_split); - layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); - layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, backend); + layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); + layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend); + layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); + layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, backend_split); + layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); + layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend_split); + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split); + layer.ffn_down_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, backend_split); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend_split); + layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); + layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, backend); layer.attn_q_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {64}, backend); - layer.attn_q_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q_NORM, "bias", i), {64}, backend); + layer.attn_q_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q_NORM, "bias", i), {64}, backend); layer.attn_k_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {64}, backend); - layer.attn_k_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K_NORM, "bias", i), {64}, backend); + layer.attn_k_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K_NORM, "bias", i), {64}, backend); } } break; case LLM_ARCH_BLOOM: { // TODO: CPU-only for now - model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); - model.tok_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, GGML_BACKEND_CPU); - model.tok_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd}, GGML_BACKEND_CPU); + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + model.tok_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, GGML_BACKEND_CPU); + model.tok_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd}, GGML_BACKEND_CPU); // output { @@ -2875,11 +2898,11 @@ static void llm_load_tensors( layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, backend); - layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split); - layer.b2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, backend_split); + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split); + layer.ffn_down_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, backend_split); - layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); - layer.b3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend_split); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend_split); if (backend == GGML_BACKEND_GPU) { vram_weights += @@ -2887,14 +2910,14 @@ static void llm_load_tensors( ggml_nbytes(layer.wqkv) + ggml_nbytes(layer.bqkv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.bo) + ggml_nbytes(layer.ffn_norm) + ggml_nbytes(layer.ffn_norm_b) + - ggml_nbytes(layer.w3) + ggml_nbytes(layer.b3) + - ggml_nbytes(layer.w2) + ggml_nbytes(layer.b2); + ggml_nbytes(layer.ffn_up) + ggml_nbytes(layer.ffn_up_b) + + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_down_b); } } } break; case LLM_ARCH_MPT: { - model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); // output { @@ -2945,8 +2968,8 @@ static void llm_load_tensors( layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); - layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); - layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); if (backend == GGML_BACKEND_GPU) { vram_weights += @@ -2954,8 +2977,8 @@ static void llm_load_tensors( ggml_nbytes(layer.wqkv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) + - ggml_nbytes(layer.w2) + - ggml_nbytes(layer.w3); + ggml_nbytes(layer.ffn_down) + + ggml_nbytes(layer.ffn_up); } } } break; @@ -2985,10 +3008,10 @@ static void llm_load_tensors( #ifdef GGML_USE_CUBLAS const int max_backend_supported_layers = hparams.n_layer + 3; - const int max_offloadable_layers = hparams.n_layer + 3; -#elif defined(GGML_USE_CLBLAST) + const int max_offloadable_layers = hparams.n_layer + 3; +#elif GGML_USE_CLBLAST const int max_backend_supported_layers = hparams.n_layer + 1; - const int max_offloadable_layers = hparams.n_layer + 1; + const int max_offloadable_layers = hparams.n_layer + 1; #endif // GGML_USE_CUBLAS LLAMA_LOG_INFO("%s: offloaded %d/%d layers to GPU\n", __func__, std::min(n_gpu_layers, max_offloadable_layers), max_backend_supported_layers); @@ -3067,9 +3090,359 @@ static bool llama_model_load( return true; } +using llm_build_cb = std::function; + +enum llm_rope_type { + LLM_ROPE, + LLM_ROPE_NEOX, + LLM_ROPE_GLM, +}; + +static struct ggml_tensor * llm_build_inp_embd( + struct ggml_context * ctx, + const llama_batch & batch, + struct ggml_tensor * tok_embd, + int64_t n_embd, + int32_t n_tokens, + const llm_build_cb & cb) { + struct ggml_tensor * inpL; + + if (batch.token) { + struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, n_tokens); + cb(inp_tokens, "inp_tokens", -1); + + inpL = ggml_get_rows(ctx, tok_embd, inp_tokens); + } else { +#ifdef GGML_USE_MPI + GGML_ASSERT(false && "not implemented"); +#endif + + inpL = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_tokens); + } + + return inpL; +} + +// Persimmon: n_rot = n_embd_head/2 +// Other: n_rot = n_embd_head +static void llm_build_k_shift( + const llama_context & lctx, + struct ggml_context * ctx, + struct ggml_cgraph * graph, + int64_t n_rot, + llm_rope_type type, + const llm_build_cb & cb) { + const auto & model = lctx.model; + const auto & kv_self = lctx.kv_self; + const auto & cparams = lctx.cparams; + + const auto & hparams = model.hparams; + + const int64_t n_layer = hparams.n_layer; + const int64_t n_head_kv = hparams.n_head_kv; + const int64_t n_embd_gqa = hparams.n_embd_gqa(); + const int64_t n_embd_head = hparams.n_embd_head(); + + const int64_t n_ctx = lctx.cparams.n_ctx; + + const float freq_base = cparams.rope_freq_base; + const float freq_scale = cparams.rope_freq_scale; + + GGML_ASSERT(n_embd_head % n_rot == 0); + + struct ggml_tensor * K_shift = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, n_ctx); + cb(K_shift, "K_shift", -1); + + int rope_type = 0; + + switch (type) { + case LLM_ROPE: rope_type = 0; break; + case LLM_ROPE_NEOX: rope_type = 2; break; + case LLM_ROPE_GLM: rope_type = 4; break; + } + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * tmp = + // we rotate only the first n_rot dimensions + ggml_rope_custom_inplace(ctx, + ggml_view_3d(ctx, kv_self.k, + n_rot, n_head_kv, n_ctx, + ggml_element_size(kv_self.k)*n_embd_head, + ggml_element_size(kv_self.k)*n_embd_gqa, + ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il), + K_shift, n_rot, rope_type, 0, freq_base, freq_scale); + cb(tmp, "K_shifted", il); + ggml_build_forward_expand(graph, tmp); + } +} + +static void llm_build_kv_store( + const llama_context & lctx, + struct ggml_context * ctx, + struct ggml_cgraph * graph, + struct ggml_tensor * k_cur, + struct ggml_tensor * v_cur, + int32_t n_tokens, + int32_t kv_head, + const llm_build_cb & cb, + int64_t il) { + const auto & model = lctx.model; + const auto & kv_self = lctx.kv_self; + const auto & cparams = lctx.cparams; + + const auto & hparams = model.hparams; + + const int64_t n_ctx = cparams.n_ctx; + const int64_t n_embd_gqa = hparams.n_embd_gqa(); + + // compute the transposed [n_tokens, n_embd] V matrix + struct ggml_tensor * v_cur_t = ggml_transpose(ctx, ggml_reshape_2d(ctx, v_cur, n_embd_gqa, n_tokens)); + //struct ggml_tensor * v_cur_t = ggml_transpose(ctx, v_cur); // TODO: reshape above is likely not needed + cb(v_cur_t, "v_cur_t", il); + + struct ggml_tensor * k_cache_view = ggml_view_1d(ctx, kv_self.k, n_tokens*n_embd_gqa, + (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head)); + cb(k_cache_view, "k_cache_view", il); + + struct ggml_tensor * v_cache_view = ggml_view_2d(ctx, kv_self.v, n_tokens, n_embd_gqa, + ( n_ctx)*ggml_element_size(kv_self.v), + (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v)); + cb(v_cache_view, "v_cache_view", il); + + // important: storing RoPE-ed version of K in the KV cache! + ggml_build_forward_expand(graph, ggml_cpy(ctx, k_cur, k_cache_view)); + ggml_build_forward_expand(graph, ggml_cpy(ctx, v_cur_t, v_cache_view)); +} + +enum llm_norm_type { + LLM_NORM, + LLM_NORM_RMS, +}; + +static struct ggml_tensor * llm_build_norm( + struct ggml_context * ctx, + struct ggml_tensor * cur, + struct ggml_tensor * mw, + struct ggml_tensor * mb, + llm_norm_type type, + float eps, + const llm_build_cb & cb, + int il) { + switch (type) { + case LLM_NORM: cur = ggml_norm (ctx, cur, eps); break; + case LLM_NORM_RMS: cur = ggml_rms_norm(ctx, cur, eps); break; + } + + if (mw || mb) { + cb(cur, "norm", il); + } + + if (mw) { + cur = ggml_mul(ctx, cur, mw); + if (mb) { + cb(cur, "norm_w", il); + } + } + + if (mb) { + cur = ggml_add(ctx, cur, mb); + } + + return cur; +} + +enum llm_ffn_op_type { + LLM_FFN_SILU, + LLM_FFN_GELU, + LLM_FFN_RELU, + LLM_FFN_RELU_SQR, +}; + +enum llm_ffn_gate_type { + LLM_FFN_SEQ, + LLM_FFN_PAR, // ffn_gate is parallel to ffn_up +}; + +static struct ggml_tensor * llm_build_ffn( + struct ggml_context * ctx, + struct ggml_tensor * cur, + struct ggml_tensor * up, + struct ggml_tensor * up_b, + struct ggml_tensor * gate, + struct ggml_tensor * gate_b, + struct ggml_tensor * down, + struct ggml_tensor * down_b, + llm_ffn_op_type type_op, + llm_ffn_gate_type type_gate, + const llm_build_cb & cb, + int il) { + struct ggml_tensor * tmp = ggml_mul_mat(ctx, up, cur); + cb(tmp, "ffn_up", il); + + if (up_b) { + tmp = ggml_add(ctx, tmp, up_b); + cb(tmp, "ffn_up_b", il); + } + + if (gate) { + switch (type_gate) { + case LLM_FFN_SEQ: + { + cur = ggml_mul_mat(ctx, gate, tmp); + cb(cur, "ffn_gate", il); + } break; + case LLM_FFN_PAR: + { + cur = ggml_mul_mat(ctx, gate, cur); + cb(cur, "ffn_gate", il); + } break; + } + + if (gate_b) { + cur = ggml_add(ctx, cur, gate_b); + cb(cur, "ffn_gate_b", il); + } + } else { + cur = tmp; + } + + switch (type_op) { + case LLM_FFN_SILU: + { + cur = ggml_silu(ctx, cur); + cb(cur, "ffn_silu", il); + } break; + case LLM_FFN_GELU: + { + cur = ggml_gelu(ctx, cur); + cb(cur, "ffn_gelu", il); + } break; + case LLM_FFN_RELU: + { + cur = ggml_relu(ctx, cur); + cb(cur, "ffn_relu", il); + } break; + case LLM_FFN_RELU_SQR: + { + cur = ggml_relu(ctx, cur); + cb(cur, "ffn_relu", il); + + cur = ggml_sqr(ctx, cur); + cb(cur, "ffn_sqr(relu)", il); + } break; + } + + if (type_gate == LLM_FFN_PAR) { + cur = ggml_mul(ctx, cur, tmp); + cb(cur, "ffn_gate_par", il); + } + + cur = ggml_mul_mat(ctx, down, cur); + if (down_b) { + cb(cur, "ffn_down", il); + } + + if (down_b) { + cur = ggml_add(ctx, cur, down_b); + } + + return cur; +} + +// if max_alibi_bias > 0 then apply ALiBi +static struct ggml_tensor * llm_build_kqv( + const llama_context & lctx, + struct ggml_context * ctx, + struct ggml_tensor * cur, + struct ggml_tensor * wo, + struct ggml_tensor * wo_b, + struct ggml_tensor * q_cur, + struct ggml_tensor * kq_scale, + struct ggml_tensor * kq_mask, + int32_t n_tokens, + int32_t n_kv, + float alibi_bias_max, + const llm_build_cb & cb, + int il) { + const auto & model = lctx.model; + const auto & kv_self = lctx.kv_self; + const auto & cparams = lctx.cparams; + + const auto & hparams = model.hparams; + + const int64_t n_ctx = cparams.n_ctx; + const int64_t n_embd = hparams.n_embd; + const int64_t n_head = hparams.n_head; + const int64_t n_head_kv = hparams.n_head_kv; + const int64_t n_embd_head = hparams.n_embd_head(); + const int64_t n_embd_gqa = hparams.n_embd_gqa(); + + struct ggml_tensor * q = ggml_permute(ctx, q_cur, 0, 2, 1, 3); + cb(q, "q", il); + + struct ggml_tensor * k = + ggml_view_3d(ctx, kv_self.k, + n_embd_head, n_kv, n_head_kv, + ggml_element_size(kv_self.k)*n_embd_gqa, + ggml_element_size(kv_self.k)*n_embd_head, + ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il); + cb(k, "k", il); + + struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q); + cb(kq, "kq", il); + + kq = ggml_scale(ctx, kq, kq_scale); + cb(kq, "kq_scaled", il); + + if (alibi_bias_max > 0.0f) { + // TODO: n_head or n_head_kv + // TODO: K-shift is likely not working + // TODO: change to ggml_add + kq = ggml_alibi(ctx, kq, /*n_past*/ 0, n_head, alibi_bias_max); + cb(kq, "kq_scaled_alibi", il); + } + + kq = ggml_add(ctx, kq, kq_mask); + cb(kq, "kq_masked", il); + + kq = ggml_soft_max(ctx, kq); + cb(kq, "kq_soft_max", il); + + // split cached v into n_head heads + struct ggml_tensor * v = + ggml_view_3d(ctx, kv_self.v, + n_kv, n_embd_head, n_head_kv, + ggml_element_size(kv_self.v)*n_ctx, + ggml_element_size(kv_self.v)*n_ctx*n_embd_head, + ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il); + cb(v, "v", il); + + struct ggml_tensor * kqv = ggml_mul_mat(ctx, v, kq); + cb(kqv, "kqv", il); + + struct ggml_tensor * kqv_merged = ggml_permute(ctx, kqv, 0, 2, 1, 3); + cb(kqv_merged, "kqv_merged", il); + + cur = ggml_cont_2d(ctx, kqv_merged, n_embd, n_tokens); + cb(cur, "kqv_merged_cont", il); + + cur = ggml_mul_mat(ctx, wo, cur); + if (wo_b) { + cb(cur, "kqv_wo", il); + } + + if (wo_b) { + cur = ggml_add(ctx, cur, wo_b); + } + + return cur; +} + static struct ggml_cgraph * llm_build_llama( - llama_context & lctx, - const llama_batch & batch) { + llama_context & lctx, + const llama_batch & batch, + const llm_build_cb & cb, + bool worst_case) { const auto & model = lctx.model; const auto & hparams = model.hparams; const auto & cparams = lctx.cparams; @@ -3084,7 +3457,6 @@ static struct ggml_cgraph * llm_build_llama( const int64_t n_head = hparams.n_head; const int64_t n_head_kv = hparams.n_head_kv; const int64_t n_embd_head = hparams.n_embd_head(); - const int64_t n_embd_gqa = hparams.n_embd_gqa(); GGML_ASSERT(n_embd_head == hparams.n_rot); @@ -3092,13 +3464,11 @@ static struct ggml_cgraph * llm_build_llama( const float freq_scale = cparams.rope_freq_scale; const float norm_rms_eps = hparams.f_norm_rms_eps; - const int n_gpu_layers = model.n_gpu_layers; - const int32_t n_tokens = batch.n_tokens; - const int32_t n_kv = ggml_allocr_is_measure(lctx.alloc) ? n_ctx : kv_self.n; - const int32_t kv_head = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head; + const int32_t n_kv = worst_case ? n_ctx : kv_self.n; + const int32_t kv_head = worst_case ? n_ctx - n_tokens : kv_self.head; - const bool do_rope_shift = ggml_allocr_is_measure(lctx.alloc) || kv_self.has_shift; + const bool do_rope_shift = worst_case || kv_self.has_shift; //printf("n_kv = %d\n", n_kv); @@ -3117,314 +3487,81 @@ static struct ggml_cgraph * llm_build_llama( struct ggml_tensor * cur; struct ggml_tensor * inpL; - if (batch.token) { - struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + inpL = llm_build_inp_embd(ctx0, batch, model.tok_embd, n_embd, n_tokens, cb); + cb(inpL, "inp_embd", -1); - ggml_allocr_alloc(lctx.alloc, inp_tokens); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens)); - } - ggml_set_name(inp_tokens, "inp_tokens"); - - inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens); - } else { -#ifdef GGML_USE_MPI - GGML_ASSERT(false && "not implemented"); -#endif - - inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens); - - ggml_allocr_alloc(lctx.alloc, inpL); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inpL->data, batch.embd, n_tokens * n_embd * ggml_element_size(inpL)); - } - } - - const int i_gpu_start = n_layer - n_gpu_layers; - (void) i_gpu_start; - - // offload functions set the tensor output backend to GPU - // tensors are GPU-accelerated if any input or the output has been offloaded - offload_func_t offload_func_nr = llama_nop; // nr = non-repeating - offload_func_t offload_func_kq = llama_nop; - offload_func_t offload_func_v = llama_nop; - -#ifdef GGML_USE_CUBLAS - if (n_gpu_layers > n_layer) { - offload_func_nr = ggml_cuda_assign_buffers_no_alloc; - } - if (n_gpu_layers > n_layer + 1) { - offload_func_v = ggml_cuda_assign_buffers_no_alloc; - } - if (n_gpu_layers > n_layer + 2) { - offload_func_kq = ggml_cuda_assign_buffers_no_alloc; - } -#endif // GGML_USE_CUBLAS + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + cb(inp_pos, "inp_pos", -1); // KQ_scale struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)"); - ggml_allocr_alloc(lctx.alloc, KQ_scale); - if (!ggml_allocr_is_measure(lctx.alloc)) { - ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd_head))); - } + cb(KQ_scale, "KQ_scale", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); - offload_func_kq(KQ_mask); - ggml_set_name(KQ_mask, "KQ_mask"); - ggml_allocr_alloc(lctx.alloc, KQ_mask); - if (!ggml_allocr_is_measure(lctx.alloc)) { - float * data = (float *) KQ_mask->data; - memset(data, 0, ggml_nbytes(KQ_mask)); - - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - const llama_pos pos = batch.pos[j]; - const llama_seq_id seq_id = batch.seq_id[j][0]; - - for (int i = 0; i < n_kv; ++i) { - if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) { - data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY; - } - } - } - } - } - - // KQ_pos - contains the positions - struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); - offload_func_kq(KQ_pos); - ggml_set_name(KQ_pos, "KQ_pos"); - ggml_allocr_alloc(lctx.alloc, KQ_pos); - if (!ggml_allocr_is_measure(lctx.alloc)) { - int * data = (int *) KQ_pos->data; - for (int i = 0; i < n_tokens; ++i) { - data[i] = batch.pos[i]; - } - } + cb(KQ_mask, "KQ_mask", -1); // shift the entire K-cache if needed if (do_rope_shift) { - struct ggml_tensor * K_shift = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_ctx); - offload_func_kq(K_shift); - ggml_set_name(K_shift, "K_shift"); - ggml_allocr_alloc(lctx.alloc, K_shift); - if (!ggml_allocr_is_measure(lctx.alloc)) { - int * data = (int *) K_shift->data; - for (int i = 0; i < n_ctx; ++i) { - data[i] = kv_self.cells[i].delta; - } - } - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * tmp = - ggml_rope_custom_inplace(ctx0, - ggml_view_3d(ctx0, kv_self.k, - n_embd_head, n_head_kv, n_ctx, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il), - K_shift, n_embd_head, 0, 0, freq_base, freq_scale); - offload_func_kq(tmp); - ggml_build_forward_expand(gf, tmp); - } + llm_build_k_shift(lctx, ctx0, gf, n_embd_head, LLM_ROPE, cb); } for (int il = 0; il < n_layer; ++il) { - ggml_format_name(inpL, "layer_inp_%d", il); - - offload_func_t offload_func = llama_nop; - -#ifdef GGML_USE_CUBLAS - if (il >= i_gpu_start) { - offload_func = ggml_cuda_assign_buffers_no_alloc; - } -#endif // GGML_USE_CUBLAS - struct ggml_tensor * inpSA = inpL; // norm - { - cur = ggml_rms_norm(ctx0, inpL, norm_rms_eps); - offload_func(cur); - ggml_set_name(cur, "rms_norm_0"); - - // cur = cur*attn_norm(broadcasted) - cur = ggml_mul(ctx0, cur, model.layers[il].attn_norm); - offload_func(cur); - ggml_set_name(cur, "attention_norm_0"); - } + cur = llm_build_norm(ctx0, inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, norm_rms_eps, cb, il); + cb(cur, "attn_norm", il); // self-attention { // compute Q and K and RoPE them - struct ggml_tensor * tmpk = ggml_mul_mat(ctx0, model.layers[il].wk, cur); - offload_func_kq(tmpk); - ggml_set_name(tmpk, "tmpk"); + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); - struct ggml_tensor * tmpq = ggml_mul_mat(ctx0, model.layers[il].wq, cur); - offload_func_kq(tmpq); - ggml_set_name(tmpq, "tmpq"); + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); - struct ggml_tensor * Kcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens), KQ_pos, n_embd_head, 0, 0, freq_base, freq_scale); - offload_func_kq(Kcur); - ggml_set_name(Kcur, "Kcur"); + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); - struct ggml_tensor * Qcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, n_tokens), KQ_pos, n_embd_head, 0, 0, freq_base, freq_scale); - offload_func_kq(Qcur); - ggml_set_name(Qcur, "Qcur"); + Qcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, n_embd_head, 0, 0, freq_base, freq_scale); + cb(Qcur, "Qcur", il); - // store key and value to memory - { - // compute the transposed [n_tokens, n_embd] V matrix + Kcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, n_embd_head, 0, 0, freq_base, freq_scale); + cb(Kcur, "Kcur", il); - struct ggml_tensor * tmpv = ggml_mul_mat(ctx0, model.layers[il].wv, cur); - offload_func_v(tmpv); - ggml_set_name(tmpv, "tmpv"); + llm_build_kv_store(lctx, ctx0, gf, Kcur, Vcur, n_tokens, kv_head, cb, il); - struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, tmpv, n_embd_gqa, n_tokens)); - offload_func_v(Vcur); - ggml_set_name(Vcur, "Vcur"); - - struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head)); - offload_func_kq(k); - ggml_set_name(k, "k"); - - struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa, - ( n_ctx)*ggml_element_size(kv_self.v), - (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v)); - offload_func_v(v); - ggml_set_name(v, "v"); - - // important: storing RoPE-ed version of K in the KV cache! - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); - } - - struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); - offload_func_kq(Q); - ggml_set_name(Q, "Q"); - - struct ggml_tensor * K = - ggml_view_3d(ctx0, kv_self.k, - n_embd_head, n_kv, n_head_kv, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il); - offload_func_kq(K); - ggml_set_name(K, "K"); - - // K * Q - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - offload_func_kq(KQ); - ggml_set_name(KQ, "KQ"); - - // KQ_scaled = KQ / sqrt(n_embd_head) - // KQ_scaled shape [n_kv, n_tokens, n_head, 1] - struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale); - offload_func_kq(KQ_scaled); - ggml_set_name(KQ_scaled, "KQ_scaled"); - - // KQ_masked = mask_past(KQ_scaled) - struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled, KQ_mask); - offload_func_kq(KQ_masked); - ggml_set_name(KQ_masked, "KQ_masked"); - - // KQ = soft_max(KQ_masked) - struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked); - offload_func_v(KQ_soft_max); - ggml_set_name(KQ_soft_max, "KQ_soft_max"); - - // split cached V into n_head heads - struct ggml_tensor * V = - ggml_view_3d(ctx0, kv_self.v, - n_kv, n_embd_head, n_head_kv, - ggml_element_size(kv_self.v)*n_ctx, - ggml_element_size(kv_self.v)*n_ctx*n_embd_head, - ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il); - offload_func_v(V); - ggml_set_name(V, "V"); - -#if 1 - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); - offload_func_v(KQV); - ggml_set_name(KQV, "KQV"); -#else - // make V contiguous in memory to speed up the matmul, however we waste time on the copy - // on M1 this is faster for the perplexity computation, but ~5% slower for the single-token generation - // is there a better way? - struct ggml_tensor * V_cont = ggml_cpy(ctx0, V, ggml_new_tensor_3d(ctx0, kv_self.v->type, n_ctx, n_embd_head, n_head)); - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_cont, KQ_soft_max); -#endif - - // KQV_merged = KQV.permute(0, 2, 1, 3) - struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - offload_func_v(KQV_merged); - ggml_set_name(KQV_merged, "KQV_merged"); - - // cur = KQV_merged.contiguous().view(n_embd, n_tokens) - cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens); - offload_func_v(cur); - ggml_set_name(cur, "KQV_merged_contiguous"); - - // projection (no bias) - cur = ggml_mul_mat(ctx0, - model.layers[il].wo, - cur); - offload_func(cur); - ggml_set_name(cur, "result_wo"); + cur = llm_build_kqv(lctx, ctx0, cur, + model.layers[il].wo, NULL, + Qcur, KQ_scale, KQ_mask, n_tokens, n_kv, -1.0f, cb, il); + cb(cur, "kqv_out", il); } - struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA); - offload_func(inpFF); - ggml_set_name(inpFF, "inpFF"); + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); // feed-forward network { - // norm - { - cur = ggml_rms_norm(ctx0, inpFF, norm_rms_eps); - offload_func(cur); - ggml_set_name(cur, "rms_norm_1"); + cur = llm_build_norm(ctx0, ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, norm_rms_eps, cb, il); + cb(cur, "ffn_norm", il); - // cur = cur*ffn_norm(broadcasted) - cur = ggml_mul(ctx0, cur, model.layers[il].ffn_norm); - offload_func(cur); - ggml_set_name(cur, "ffn_norm"); - } - - struct ggml_tensor * tmp = ggml_mul_mat(ctx0, - model.layers[il].w3, - cur); - offload_func(tmp); - ggml_set_name(tmp, "result_w3"); - - cur = ggml_mul_mat(ctx0, - model.layers[il].w1, - cur); - offload_func(cur); - ggml_set_name(cur, "result_w1"); - - // SILU activation - cur = ggml_silu(ctx0, cur); - offload_func(cur); - ggml_set_name(cur, "silu"); - - cur = ggml_mul(ctx0, cur, tmp); - offload_func(cur); - ggml_set_name(cur, "silu_x_result_w3"); - - cur = ggml_mul_mat(ctx0, - model.layers[il].w2, - cur); - offload_func(cur); - ggml_set_name(cur, "result_w2"); + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, + model.layers[il].ffn_gate, NULL, + model.layers[il].ffn_down, NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); } - cur = ggml_add(ctx0, cur, inpFF); - offload_func(cur); - ggml_set_name(cur, "inpFF_+_result_w2"); + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); // input for next layer inpL = cur; @@ -3432,21 +3569,14 @@ static struct ggml_cgraph * llm_build_llama( cur = inpL; - // norm - { - cur = ggml_rms_norm(ctx0, cur, norm_rms_eps); - offload_func_nr(cur); - ggml_set_name(cur, "rms_norm_2"); - - // cur = cur*norm(broadcasted) - cur = ggml_mul(ctx0, cur, model.output_norm); - // offload_func_nr(cur); // TODO CPU + GPU mirrored backend - ggml_set_name(cur, "result_norm"); - } + cur = llm_build_norm(ctx0, cur, + model.output_norm, NULL, + LLM_NORM_RMS, norm_rms_eps, cb, -1); + cb(cur, "result_norm", -1); // lm_head cur = ggml_mul_mat(ctx0, model.output, cur); - ggml_set_name(cur, "result_output"); + cb(cur, "result_output", -1); ggml_build_forward_expand(gf, cur); @@ -3457,7 +3587,9 @@ static struct ggml_cgraph * llm_build_llama( static struct ggml_cgraph * llm_build_baichaun( llama_context & lctx, - const llama_batch & batch) { + const llama_batch & batch, + const llm_build_cb & cb, + bool worst_case) { const auto & model = lctx.model; const auto & hparams = model.hparams; const auto & cparams = lctx.cparams; @@ -3472,7 +3604,6 @@ static struct ggml_cgraph * llm_build_baichaun( const int64_t n_head = hparams.n_head; const int64_t n_head_kv = hparams.n_head_kv; const int64_t n_embd_head = hparams.n_embd_head(); - const int64_t n_embd_gqa = hparams.n_embd_gqa(); GGML_ASSERT(n_embd_head == hparams.n_rot); @@ -3480,13 +3611,11 @@ static struct ggml_cgraph * llm_build_baichaun( const float freq_scale = cparams.rope_freq_scale; const float norm_rms_eps = hparams.f_norm_rms_eps; - const int n_gpu_layers = model.n_gpu_layers; - const int32_t n_tokens = batch.n_tokens; - const int32_t n_kv = ggml_allocr_is_measure(lctx.alloc) ? n_ctx : kv_self.n; - const int32_t kv_head = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head; + const int32_t n_kv = worst_case ? n_ctx : kv_self.n; + const int32_t kv_head = worst_case ? n_ctx - n_tokens : kv_self.head; - const bool do_rope_shift = ggml_allocr_is_measure(lctx.alloc) || kv_self.has_shift; + const bool do_rope_shift = worst_case || kv_self.has_shift; auto & buf_compute = lctx.buf_compute; @@ -3503,331 +3632,91 @@ static struct ggml_cgraph * llm_build_baichaun( struct ggml_tensor * cur; struct ggml_tensor * inpL; - if (batch.token) { - struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + inpL = llm_build_inp_embd(ctx0, batch, model.tok_embd, n_embd, n_tokens, cb); + cb(inpL, "inp_embd", -1); - ggml_allocr_alloc(lctx.alloc, inp_tokens); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens)); - } - ggml_set_name(inp_tokens, "inp_tokens"); - - inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens); - } else { -#ifdef GGML_USE_MPI - GGML_ASSERT(false && "not implemented"); -#endif - - inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens); - - ggml_allocr_alloc(lctx.alloc, inpL); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inpL->data, batch.embd, n_tokens * n_embd * ggml_element_size(inpL)); - } - } - - const int i_gpu_start = n_layer - n_gpu_layers; - (void) i_gpu_start; - - // offload functions set the tensor output backend to GPU - // tensors are GPU-accelerated if any input or the output has been offloaded - offload_func_t offload_func_nr = llama_nop; // nr = non-repeating - offload_func_t offload_func_kq = llama_nop; - offload_func_t offload_func_v = llama_nop; - -#ifdef GGML_USE_CUBLAS - if (n_gpu_layers > n_layer) { - offload_func_nr = ggml_cuda_assign_buffers_no_alloc; - } - if (n_gpu_layers > n_layer + 1) { - offload_func_v = ggml_cuda_assign_buffers_no_alloc; - } - if (n_gpu_layers > n_layer + 2) { - offload_func_kq = ggml_cuda_assign_buffers_no_alloc; - } -#endif // GGML_USE_CUBLAS + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + cb(inp_pos, "inp_pos", -1); // KQ_scale struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)"); - ggml_allocr_alloc(lctx.alloc, KQ_scale); - if (!ggml_allocr_is_measure(lctx.alloc)) { - ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head)); - } + cb(KQ_scale, "KQ_scale", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); - offload_func_kq(KQ_mask); - ggml_set_name(KQ_mask, "KQ_mask"); - ggml_allocr_alloc(lctx.alloc, KQ_mask); - if (!ggml_allocr_is_measure(lctx.alloc)) { - float * data = (float *) KQ_mask->data; - memset(data, 0, ggml_nbytes(KQ_mask)); - - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - const llama_pos pos = batch.pos[j]; - const llama_seq_id seq_id = batch.seq_id[j][0]; - - for (int i = 0; i < n_kv; ++i) { - if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) { - data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY; - } - } - } - } - } - - // KQ_pos - contains the positions - struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); - offload_func_kq(KQ_pos); - ggml_set_name(KQ_pos, "KQ_pos"); - ggml_allocr_alloc(lctx.alloc, KQ_pos); - if (!ggml_allocr_is_measure(lctx.alloc)) { - int * data = (int *) KQ_pos->data; - for (int i = 0; i < n_tokens; ++i) { - data[i] = batch.pos[i]; - } - } + cb(KQ_mask, "KQ_mask", -1); // shift the entire K-cache if needed if (do_rope_shift) { - struct ggml_tensor * K_shift = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_ctx); - offload_func_kq(K_shift); - ggml_set_name(K_shift, "K_shift"); - ggml_allocr_alloc(lctx.alloc, K_shift); - if (!ggml_allocr_is_measure(lctx.alloc)) { - int * data = (int *) K_shift->data; - for (int i = 0; i < n_ctx; ++i) { - data[i] = kv_self.cells[i].delta; - } - } - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * tmp = - ggml_rope_custom_inplace(ctx0, - ggml_view_3d(ctx0, kv_self.k, - n_embd_head, n_head_kv, n_ctx, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il), - K_shift, n_embd_head, 0, 0, freq_base, freq_scale); - offload_func_kq(tmp); - ggml_build_forward_expand(gf, tmp); - } + llm_build_k_shift(lctx, ctx0, gf, n_embd_head, LLM_ROPE, cb); } for (int il = 0; il < n_layer; ++il) { - ggml_format_name(inpL, "layer_inp_%d", il); - - offload_func_t offload_func = llama_nop; - -#ifdef GGML_USE_CUBLAS - if (il >= i_gpu_start) { - offload_func = ggml_cuda_assign_buffers_no_alloc; - } -#endif // GGML_USE_CUBLAS - struct ggml_tensor * inpSA = inpL; - // norm - { - cur = ggml_rms_norm(ctx0, inpL, norm_rms_eps); - offload_func(cur); - ggml_set_name(cur, "rms_norm_0"); - - // cur = cur*attn_norm(broadcasted) - cur = ggml_mul(ctx0, cur, model.layers[il].attn_norm); - offload_func(cur); - ggml_set_name(cur, "attention_norm_0"); - } + cur = llm_build_norm(ctx0, inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, norm_rms_eps, cb, il); + cb(cur, "attn_norm", il); // self-attention { - // compute Q and K and RoPE them - struct ggml_tensor * tmpk = ggml_mul_mat(ctx0, model.layers[il].wk, cur); - offload_func_kq(tmpk); - ggml_set_name(tmpk, "tmpk"); + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); - struct ggml_tensor * tmpq = ggml_mul_mat(ctx0, model.layers[il].wq, cur); - offload_func_kq(tmpq); - ggml_set_name(tmpq, "tmpq"); + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); - struct ggml_tensor * Kcur; - struct ggml_tensor * Qcur; switch (model.type) { case MODEL_7B: - Kcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens), KQ_pos, n_embd_head, 0, 0, freq_base, freq_scale); - Qcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, n_tokens), KQ_pos, n_embd_head, 0, 0, freq_base, freq_scale); + Qcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, n_embd_head, 0, 0, freq_base, freq_scale); + Kcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, n_embd_head, 0, 0, freq_base, freq_scale); break; case MODEL_13B: - Kcur = ggml_reshape_3d(ctx0, tmpk, n_embd/n_head, n_head, n_tokens); - Qcur = ggml_reshape_3d(ctx0, tmpq, n_embd/n_head, n_head, n_tokens); + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd/n_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd/n_head, n_head, n_tokens); break; default: GGML_ASSERT(false); } + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); - offload_func_kq(Kcur); - ggml_set_name(Kcur, "Kcur"); + llm_build_kv_store(lctx, ctx0, gf, Kcur, Vcur, n_tokens, kv_head, cb, il); - offload_func_kq(Qcur); - ggml_set_name(Qcur, "Qcur"); + // apply ALiBi for 13B model + const float alibi_bias_max = model.type == MODEL_13B ? 8.0f : -1.0f; - // store key and value to memory - { - // compute the transposed [n_tokens, n_embd] V matrix - - struct ggml_tensor * tmpv = ggml_mul_mat(ctx0, model.layers[il].wv, cur); - offload_func_v(tmpv); - ggml_set_name(tmpv, "tmpv"); - - struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, tmpv, n_embd_gqa, n_tokens)); - offload_func_v(Vcur); - ggml_set_name(Vcur, "Vcur"); - - struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head)); - offload_func_kq(k); - ggml_set_name(k, "k"); - - struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa, - ( n_ctx)*ggml_element_size(kv_self.v), - (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v)); - offload_func_v(v); - ggml_set_name(v, "v"); - - // important: storing RoPE-ed version of K in the KV cache! - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); - } - - struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); - offload_func_kq(Q); - ggml_set_name(Q, "Q"); - - struct ggml_tensor * K = - ggml_view_3d(ctx0, kv_self.k, - n_embd_head, n_kv, n_head_kv, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il); - offload_func_kq(K); - ggml_set_name(K, "K"); - - // K * Q - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - offload_func_kq(KQ); - ggml_set_name(KQ, "KQ"); - - // KQ_scaled = KQ / sqrt(n_embd_head) - // KQ_scaled shape [n_past + n_tokens, n_tokens, n_head, 1] - struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale); - offload_func_kq(KQ_scaled); - ggml_set_name(KQ_scaled, "KQ_scaled"); - - struct ggml_tensor * KQ_masked; - struct ggml_tensor * KQ_scaled_alibi; - - switch (model.type) { - case MODEL_7B: - KQ_masked = ggml_add(ctx0, KQ_scaled, KQ_mask); - break; - case MODEL_13B: - // TODO: replace with ggml_add() - KQ_scaled_alibi = ggml_alibi(ctx0, KQ_scaled, /*n_past*/ 0, n_head, 8); - ggml_set_name(KQ_scaled_alibi, "KQ_scaled_alibi"); - KQ_masked = ggml_add(ctx0, KQ_scaled_alibi, KQ_mask); - break; - default: - GGML_ASSERT(false); - } - - // KQ = soft_max(KQ_masked) - struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked); - offload_func_v(KQ_soft_max); - ggml_set_name(KQ_soft_max, "KQ_soft_max"); - - // split cached V into n_head heads - struct ggml_tensor * V = - ggml_view_3d(ctx0, kv_self.v, - n_kv, n_embd_head, n_head_kv, - ggml_element_size(kv_self.v)*n_ctx, - ggml_element_size(kv_self.v)*n_ctx*n_embd_head, - ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il); - offload_func_v(V); - ggml_set_name(V, "V"); - - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); - offload_func_v(KQV); - ggml_set_name(KQV, "KQV"); - - // KQV_merged = KQV.permute(0, 2, 1, 3) - struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - offload_func_v(KQV_merged); - ggml_set_name(KQV_merged, "KQV_merged"); - - // cur = KQV_merged.contiguous().view(n_embd, n_tokens) - cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens); - offload_func_v(cur); - ggml_set_name(cur, "KQV_merged_contiguous"); - - // projection (no bias) - cur = ggml_mul_mat(ctx0, - model.layers[il].wo, - cur); - offload_func(cur); - ggml_set_name(cur, "result_wo"); + cur = llm_build_kqv(lctx, ctx0, cur, + model.layers[il].wo, NULL, + Qcur, KQ_scale, KQ_mask, n_tokens, n_kv, alibi_bias_max, cb, il); + cb(cur, "kqv_out", il); } - struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA); - offload_func(inpFF); - ggml_set_name(inpFF, "inpFF"); + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); // feed-forward network { - // norm - { - cur = ggml_rms_norm(ctx0, inpFF, norm_rms_eps); - offload_func(cur); - ggml_set_name(cur, "rms_norm_1"); + cur = llm_build_norm(ctx0, ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, norm_rms_eps, cb, il); + cb(cur, "ffn_norm", il); - // cur = cur*ffn_norm(broadcasted) - cur = ggml_mul(ctx0, cur, model.layers[il].ffn_norm); - offload_func(cur); - ggml_set_name(cur, "ffn_norm"); - } - - struct ggml_tensor * tmp = ggml_mul_mat(ctx0, - model.layers[il].w3, - cur); - offload_func(tmp); - ggml_set_name(tmp, "result_w3"); - - cur = ggml_mul_mat(ctx0, - model.layers[il].w1, - cur); - offload_func(cur); - ggml_set_name(cur, "result_w1"); - - // SILU activation - cur = ggml_silu(ctx0, cur); - offload_func(cur); - ggml_set_name(cur, "silu"); - - cur = ggml_mul(ctx0, cur, tmp); - offload_func(cur); - ggml_set_name(cur, "silu_x_result_w3"); - - cur = ggml_mul_mat(ctx0, - model.layers[il].w2, - cur); - offload_func(cur); - ggml_set_name(cur, "result_w2"); + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, + model.layers[il].ffn_gate, NULL, + model.layers[il].ffn_down, NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); } - cur = ggml_add(ctx0, cur, inpFF); - offload_func(cur); - ggml_set_name(cur, "inpFF_+_result_w2"); + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); // input for next layer inpL = cur; @@ -3835,366 +3724,14 @@ static struct ggml_cgraph * llm_build_baichaun( cur = inpL; - // norm - { - cur = ggml_rms_norm(ctx0, cur, norm_rms_eps); - offload_func_nr(cur); - ggml_set_name(cur, "rms_norm_2"); - - // cur = cur*norm(broadcasted) - cur = ggml_mul(ctx0, cur, model.output_norm); - // offload_func_nr(cur); // TODO CPU + GPU mirrored backend - ggml_set_name(cur, "result_norm"); - } + cur = llm_build_norm(ctx0, cur, + model.output_norm, NULL, + LLM_NORM_RMS, norm_rms_eps, cb, -1); + cb(cur, "result_norm", -1); // lm_head cur = ggml_mul_mat(ctx0, model.output, cur); - ggml_set_name(cur, "result_output"); - - ggml_build_forward_expand(gf, cur); - - ggml_free(ctx0); - - return gf; -} - -static struct ggml_cgraph * llm_build_refact( - llama_context & lctx, - const llama_batch & batch) { - const auto & model = lctx.model; - const auto & hparams = model.hparams; - const auto & cparams = lctx.cparams; - - const auto & kv_self = lctx.kv_self; - - GGML_ASSERT(!!kv_self.ctx); - - const int64_t n_embd = hparams.n_embd; - const int64_t n_layer = hparams.n_layer; - const int64_t n_ctx = cparams.n_ctx; - const int64_t n_head = hparams.n_head; - const int64_t n_head_kv = hparams.n_head_kv; - const int64_t n_embd_head = hparams.n_embd_head(); - const int64_t n_embd_gqa = hparams.n_embd_gqa(); - - const float norm_rms_eps = hparams.f_norm_rms_eps; - - const int n_gpu_layers = model.n_gpu_layers; - - const int32_t n_tokens = batch.n_tokens; - const int32_t n_kv = ggml_allocr_is_measure(lctx.alloc) ? n_ctx : kv_self.n; - const int32_t kv_head = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head; - - // printf("n_kv = %d\n", n_kv); - - auto & buf_compute = lctx.buf_compute; - - struct ggml_init_params params = { - /*.mem_size =*/ buf_compute.size, - /*.mem_buffer =*/ buf_compute.data, - /*.no_alloc =*/ true, - }; - - struct ggml_context * ctx0 = ggml_init(params); - - ggml_cgraph * gf = ggml_new_graph(ctx0); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - if (batch.token) { - struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); - - ggml_allocr_alloc(lctx.alloc, inp_tokens); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens)); - } - ggml_set_name(inp_tokens, "inp_tokens"); - - inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens); - } else { -#ifdef GGML_USE_MPI - GGML_ASSERT(false && "not implemented"); -#endif - - inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens); - - ggml_allocr_alloc(lctx.alloc, inpL); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inpL->data, batch.embd, n_tokens * n_embd * ggml_element_size(inpL)); - } - } - - const int i_gpu_start = n_layer - n_gpu_layers; - (void) i_gpu_start; - - // offload functions set the tensor output backend to GPU - // tensors are GPU-accelerated if any input or the output has been offloaded - offload_func_t offload_func_nr = llama_nop; // nr = non-repeating - offload_func_t offload_func_kq = llama_nop; - offload_func_t offload_func_v = llama_nop; - -#ifdef GGML_USE_CUBLAS - if (n_gpu_layers > n_layer) { - offload_func_nr = ggml_cuda_assign_buffers_no_alloc; - } - if (n_gpu_layers > n_layer + 1) { - offload_func_v = ggml_cuda_assign_buffers_no_alloc; - } - if (n_gpu_layers > n_layer + 2) { - offload_func_kq = ggml_cuda_assign_buffers_no_alloc; - } -#endif // GGML_USE_CUBLAS - - // KQ_scale - struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)"); - ggml_allocr_alloc(lctx.alloc, KQ_scale); - if (!ggml_allocr_is_measure(lctx.alloc)) { - ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd_head))); - } - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); - offload_func_kq(KQ_mask); - ggml_set_name(KQ_mask, "KQ_mask"); - ggml_allocr_alloc(lctx.alloc, KQ_mask); - if (!ggml_allocr_is_measure(lctx.alloc)) { - float * data = (float *) KQ_mask->data; - memset(data, 0, ggml_nbytes(KQ_mask)); - - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - const llama_pos pos = batch.pos[j]; - const llama_seq_id seq_id = batch.seq_id[j][0]; - - for (int i = 0; i < n_kv; ++i) { - if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) { - data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY; - } - } - } - } - } - - for (int il = 0; il < n_layer; ++il) { - ggml_format_name(inpL, "layer_inp_%d", il); - - offload_func_t offload_func = llama_nop; - -#ifdef GGML_USE_CUBLAS - if (il >= i_gpu_start) { - offload_func = ggml_cuda_assign_buffers_no_alloc; - } -#endif // GGML_USE_CUBLAS - - struct ggml_tensor * inpSA = inpL; - - // norm - { - cur = ggml_rms_norm(ctx0, inpL, norm_rms_eps); - offload_func(cur); - ggml_set_name(cur, "rms_norm_0"); - - // cur = cur*attn_norm(broadcasted) - cur = ggml_mul(ctx0, cur, model.layers[il].attn_norm); - offload_func(cur); - ggml_set_name(cur, "attention_norm_0"); - } - - // self-attention - { - // compute Q and K - struct ggml_tensor * tmpk = ggml_mul_mat(ctx0, model.layers[il].wk, cur); - offload_func_kq(tmpk); - ggml_set_name(tmpk, "tmpk"); - - struct ggml_tensor * tmpq = ggml_mul_mat(ctx0, model.layers[il].wq, cur); - offload_func_kq(tmpq); - ggml_set_name(tmpq, "tmpq"); - - struct ggml_tensor * Kcur = ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens); - offload_func_kq(Kcur); - ggml_set_name(Kcur, "Kcur"); - - struct ggml_tensor * Qcur = ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, n_tokens); - offload_func_kq(Qcur); - ggml_set_name(Qcur, "Qcur"); - - // store key and value to memory - { - // compute the transposed [n_tokens, n_embd] V matrix - - struct ggml_tensor * tmpv = ggml_mul_mat(ctx0, model.layers[il].wv, cur); - offload_func_v(tmpv); - ggml_set_name(tmpv, "tmpv"); - - struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, tmpv, n_embd_gqa, n_tokens)); - offload_func_v(Vcur); - ggml_set_name(Vcur, "Vcur"); - - struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head)); - offload_func_kq(k); - ggml_set_name(k, "k"); - - struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa, - ( n_ctx)*ggml_element_size(kv_self.v), - (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v)); - offload_func_v(v); - ggml_set_name(v, "v"); - - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); - } - - struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); - offload_func_kq(Q); - ggml_set_name(Q, "Q"); - - struct ggml_tensor * K = - ggml_view_3d(ctx0, kv_self.k, - n_embd_head, n_kv, n_head_kv, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il); - offload_func_kq(K); - ggml_set_name(K, "K"); - - // K * Q - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - offload_func_kq(KQ); - ggml_set_name(KQ, "KQ"); - - // KQ_scaled = KQ / sqrt(n_embd_head) - // KQ_scaled shape [n_kv, n_tokens, n_head, 1] - struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale); - offload_func_kq(KQ_scaled); - ggml_set_name(KQ_scaled, "KQ_scaled"); - - // KQ_masked = mask_past(KQ_scaled) - struct ggml_tensor * KQ_scaled_alibi = ggml_alibi(ctx0, KQ_scaled, /*n_past*/ 0, n_head, 8); - ggml_set_name(KQ_scaled_alibi, "KQ_scaled_alibi"); - - struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled_alibi, KQ_mask); - offload_func_kq(KQ_masked); - ggml_set_name(KQ_masked, "KQ_masked"); - - // KQ = soft_max(KQ_masked) - struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked); - offload_func_v(KQ_soft_max); - ggml_set_name(KQ_soft_max, "KQ_soft_max"); - - // split cached V into n_head heads - struct ggml_tensor * V = - ggml_view_3d(ctx0, kv_self.v, - n_kv, n_embd_head, n_head_kv, - ggml_element_size(kv_self.v)*n_ctx, - ggml_element_size(kv_self.v)*n_ctx*n_embd_head, - ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il); - offload_func_v(V); - ggml_set_name(V, "V"); - -#if 1 - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); - offload_func_v(KQV); - ggml_set_name(KQV, "KQV"); -#else - // make V contiguous in memory to speed up the matmul, however we waste time on the copy - // on M1 this is faster for the perplexity computation, but ~5% slower for the single-token generation - // is there a better way? - struct ggml_tensor * V_cont = ggml_cpy(ctx0, V, ggml_new_tensor_3d(ctx0, kv_self.v->type, n_ctx, n_embd_head, n_head)); - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_cont, KQ_soft_max); -#endif - - // KQV_merged = KQV.permute(0, 2, 1, 3) - struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - offload_func_v(KQV_merged); - ggml_set_name(KQV_merged, "KQV_merged"); - - // cur = KQV_merged.contiguous().view(n_embd, n_tokens) - cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens); - offload_func_v(cur); - ggml_set_name(cur, "KQV_merged_contiguous"); - - // projection (no bias) - cur = ggml_mul_mat(ctx0, - model.layers[il].wo, - cur); - offload_func(cur); - ggml_set_name(cur, "result_wo"); - } - - struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA); - offload_func(inpFF); - ggml_set_name(inpFF, "inpFF"); - - // feed-forward network - { - // norm - { - cur = ggml_rms_norm(ctx0, inpFF, norm_rms_eps); - offload_func(cur); - ggml_set_name(cur, "rms_norm_1"); - - // cur = cur*ffn_norm(broadcasted) - cur = ggml_mul(ctx0, cur, model.layers[il].ffn_norm); - offload_func(cur); - ggml_set_name(cur, "ffn_norm"); - } - - struct ggml_tensor * tmp = ggml_mul_mat(ctx0, - model.layers[il].w3, - cur); - offload_func(tmp); - ggml_set_name(tmp, "result_w3"); - - cur = ggml_mul_mat(ctx0, - model.layers[il].w1, - cur); - offload_func(cur); - ggml_set_name(cur, "result_w1"); - - // SILU activation - cur = ggml_silu(ctx0, cur); - offload_func(cur); - ggml_set_name(cur, "silu"); - - cur = ggml_mul(ctx0, cur, tmp); - offload_func(cur); - ggml_set_name(cur, "silu_x_result_w3"); - - cur = ggml_mul_mat(ctx0, - model.layers[il].w2, - cur); - offload_func(cur); - ggml_set_name(cur, "result_w2"); - } - - cur = ggml_add(ctx0, cur, inpFF); - offload_func(cur); - ggml_set_name(cur, "inpFF_+_result_w2"); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - // norm - { - cur = ggml_rms_norm(ctx0, cur, norm_rms_eps); - offload_func_nr(cur); - ggml_set_name(cur, "rms_norm_2"); - - // cur = cur*norm(broadcasted) - cur = ggml_mul(ctx0, cur, model.output_norm); - // offload_func_nr(cur); // TODO CPU + GPU mirrored backend - ggml_set_name(cur, "result_norm"); - } - - // lm_head - cur = ggml_mul_mat(ctx0, model.output, cur); - ggml_set_name(cur, "result_output"); + cb(cur, "result_output", -1); ggml_build_forward_expand(gf, cur); @@ -4205,7 +3742,9 @@ static struct ggml_cgraph * llm_build_refact( static struct ggml_cgraph * llm_build_falcon( llama_context & lctx, - const llama_batch & batch) { + const llama_batch & batch, + const llm_build_cb & cb, + bool worst_case) { const auto & model = lctx.model; const auto & hparams = model.hparams; const auto & cparams = lctx.cparams; @@ -4228,13 +3767,11 @@ static struct ggml_cgraph * llm_build_falcon( const float freq_scale = cparams.rope_freq_scale; const float norm_eps = hparams.f_norm_eps; - const int n_gpu_layers = model.n_gpu_layers; - const int32_t n_tokens = batch.n_tokens; - const int32_t n_kv = ggml_allocr_is_measure(lctx.alloc) ? n_ctx : kv_self.n; - const int32_t kv_head = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head; + const int32_t n_kv = worst_case ? n_ctx : kv_self.n; + const int32_t kv_head = worst_case ? n_ctx - n_tokens : kv_self.head; - const bool do_rope_shift = ggml_allocr_is_measure(lctx.alloc) || kv_self.has_shift; + const bool do_rope_shift = worst_case || kv_self.has_shift; //printf("kv_head = %d, n_kv = %d, n_tokens = %d, n_ctx = %d, is_measure = %d, has_shift = %d\n", // kv_head, n_kv, n_tokens, n_ctx, ggml_allocr_is_measure(lctx.alloc), kv_self.has_shift); @@ -4254,294 +3791,94 @@ static struct ggml_cgraph * llm_build_falcon( struct ggml_tensor * cur; struct ggml_tensor * inpL; - if (batch.token) { - struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + inpL = llm_build_inp_embd(ctx0, batch, model.tok_embd, n_embd, n_tokens, cb); + cb(inpL, "inp_embd", -1); - ggml_allocr_alloc(lctx.alloc, inp_tokens); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens)); - } - ggml_set_name(inp_tokens, "inp_tokens"); - - inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens); - } else { -#ifdef GGML_USE_MPI - GGML_ASSERT(false && "not implemented"); -#endif - - inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens); - - ggml_allocr_alloc(lctx.alloc, inpL); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inpL->data, batch.embd, n_tokens * n_embd * ggml_element_size(inpL)); - } - } - - const int i_gpu_start = n_layer - n_gpu_layers; - (void) i_gpu_start; - - // offload functions set the tensor output backend to GPU - // tensors are GPU-accelerated if any input or the output has been offloaded - offload_func_t offload_func_nr = llama_nop; // nr = non-repeating - offload_func_t offload_func_kq = llama_nop; - offload_func_t offload_func_v = llama_nop; - -#ifdef GGML_USE_CUBLAS - if (n_gpu_layers > n_layer) { - offload_func_nr = ggml_cuda_assign_buffers_no_alloc; - } - if (n_gpu_layers > n_layer + 1) { - offload_func_v = ggml_cuda_assign_buffers_no_alloc; - } - if (n_gpu_layers > n_layer + 2) { - offload_func_kq = ggml_cuda_assign_buffers_no_alloc; - } -#endif // GGML_USE_CUBLAS + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + cb(inp_pos, "inp_pos", -1); // KQ_scale struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)"); - ggml_allocr_alloc(lctx.alloc, KQ_scale); - if (!ggml_allocr_is_measure(lctx.alloc)) { - ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head)); - } + cb(KQ_scale, "KQ_scale", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); - offload_func_kq(KQ_mask); - ggml_set_name(KQ_mask, "KQ_mask"); - ggml_allocr_alloc(lctx.alloc, KQ_mask); - if (!ggml_allocr_is_measure(lctx.alloc)) { - float * data = (float *) KQ_mask->data; - memset(data, 0, ggml_nbytes(KQ_mask)); - - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - const llama_pos pos = batch.pos[j]; - const llama_seq_id seq_id = batch.seq_id[j][0]; - - for (int i = 0; i < n_kv; ++i) { - if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) { - data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY; - } - } - } - } - } - - // KQ_pos - contains the positions - struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); - offload_func_kq(KQ_pos); - ggml_set_name(KQ_pos, "KQ_pos"); - ggml_allocr_alloc(lctx.alloc, KQ_pos); - if (!ggml_allocr_is_measure(lctx.alloc)) { - int * data = (int *) KQ_pos->data; - for (int i = 0; i < n_tokens; ++i) { - data[i] = batch.pos[i]; - } - } + cb(KQ_mask, "KQ_mask", -1); // shift the entire K-cache if needed if (do_rope_shift) { - struct ggml_tensor * K_shift = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_ctx); - offload_func_kq(K_shift); - ggml_set_name(K_shift, "K_shift"); - ggml_allocr_alloc(lctx.alloc, K_shift); - if (!ggml_allocr_is_measure(lctx.alloc)) { - int * data = (int *) K_shift->data; - for (int i = 0; i < n_ctx; ++i) { - data[i] = kv_self.cells[i].delta; - } - } - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * tmp = - ggml_rope_custom_inplace(ctx0, - ggml_view_3d(ctx0, kv_self.k, - n_embd_head, n_head_kv, n_ctx, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il), - K_shift, n_embd_head, 2, 0, freq_base, freq_scale); - offload_func_kq(tmp); - ggml_build_forward_expand(gf, tmp); - } + llm_build_k_shift(lctx, ctx0, gf, n_embd_head, LLM_ROPE_NEOX, cb); } for (int il = 0; il < n_layer; ++il) { struct ggml_tensor * attn_norm; - offload_func_t offload_func = llama_nop; - -#ifdef GGML_USE_CUBLAS - if (il >= i_gpu_start) { - offload_func = ggml_cuda_assign_buffers_no_alloc; - } -#endif // GGML_USE_CUBLAS + attn_norm = llm_build_norm(ctx0, inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, norm_eps, cb, il); + cb(attn_norm, "attn_norm", il); // self-attention - // TODO: refactor into common function (shared with LLaMA) { - attn_norm = ggml_norm(ctx0, inpL, norm_eps); - offload_func(attn_norm); - - attn_norm = ggml_add(ctx0, - ggml_mul(ctx0, attn_norm, model.layers[il].attn_norm), - model.layers[il].attn_norm_b); - offload_func(attn_norm->src[0]); - offload_func(attn_norm); - - if (model.layers[il].attn_norm_2) { // Falcon-40B - cur = ggml_norm(ctx0, inpL, norm_eps); - offload_func(cur); - - cur = ggml_add(ctx0, - ggml_mul(ctx0, cur, model.layers[il].attn_norm_2), - model.layers[il].attn_norm_2_b); - offload_func(cur->src[0]); - offload_func(cur); - } else { // Falcon 7B + if (model.layers[il].attn_norm_2) { + // Falcon-40B + cur = llm_build_norm(ctx0, attn_norm, + model.layers[il].attn_norm_2, + model.layers[il].attn_norm_2_b, + LLM_NORM, norm_eps, cb, il); + cb(cur, "attn_norm_2", il); + } else { cur = attn_norm; } - // compute QKV - cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); - offload_func_kq(cur); + cb(cur, "wqkv", il); - // Note that the strides for Kcur, Vcur are set up so that the - // resulting views are misaligned with the tensor's storage - // (by applying the K/V offset we shift the tensor's original - // view to stick out behind the viewed QKV tensor's allocated - // memory, so to say). This is ok because no actual accesses - // happen to that out-of-range memory, but it can require some - // trickery when trying to accurately dump these views for - // debugging. + struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - const size_t wsize = ggml_type_size(cur->type); + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); - // TODO: these 2 ggml_conts are technically not needed, but we add them until CUDA support for - // non-contiguous views is added for the rope operator - struct ggml_tensor * tmpq = ggml_cont(ctx0, ggml_view_3d( - ctx0, cur, n_embd_head, n_head, n_tokens, - wsize * n_embd_head, - wsize * n_embd_head * (n_head + 2 * n_head_kv), - 0)); - offload_func_kq(tmpq); - - struct ggml_tensor * tmpk = ggml_cont(ctx0, ggml_view_3d( - ctx0, cur, n_embd_head, n_head_kv, n_tokens, - wsize * n_embd_head, - wsize * n_embd_head * (n_head + 2 * n_head_kv), - wsize * n_embd_head * n_head)); - offload_func_kq(tmpk); - - struct ggml_tensor * tmpv = ggml_view_3d( - ctx0, cur, n_embd_head, n_head_kv, n_tokens, - wsize * n_embd_head, - wsize * n_embd_head * (n_head + 2 * n_head_kv), - wsize * n_embd_head * (n_head + n_head_kv)); - offload_func_v(tmpv); + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); // using mode = 2 for neox mode - struct ggml_tensor * Qcur = ggml_rope_custom(ctx0, tmpq, KQ_pos, n_embd_head, 2, 0, freq_base, freq_scale); - offload_func_kq(Qcur); - struct ggml_tensor * Kcur = ggml_rope_custom(ctx0, tmpk, KQ_pos, n_embd_head, 2, 0, freq_base, freq_scale); - offload_func_kq(Kcur); + Qcur = ggml_rope_custom(ctx0, Qcur, inp_pos, n_embd_head, 2, 0, freq_base, freq_scale); + cb(Qcur, "Qcur", il); - { - struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_cont(ctx0, tmpv), n_embd_gqa, n_tokens)); - offload_func_v(Vcur); - offload_func_v(Vcur->src[0]->src[0]); - ggml_set_name(Vcur, "Vcur"); + Kcur = ggml_rope_custom(ctx0, Kcur, inp_pos, n_embd_head, 2, 0, freq_base, freq_scale); + cb(Kcur, "Kcur", il); - struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head)); - offload_func_kq(k); - ggml_set_name(k, "k"); + llm_build_kv_store(lctx, ctx0, gf, Kcur, Vcur, n_tokens, kv_head, cb, il); - struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa, - ( n_ctx)*ggml_element_size(kv_self.v), - (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v)); - offload_func_v(v); - - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); - } - - struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); - offload_func_kq(Q); - ggml_set_name(Q, "Q"); - - struct ggml_tensor * K = - ggml_view_3d(ctx0, kv_self.k, - n_embd_head, n_kv, n_head_kv, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il); - offload_func_kq(K); - ggml_set_name(K, "K"); - - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - offload_func_kq(KQ); - ggml_set_name(KQ, "KQ"); - - struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale); - offload_func_kq(KQ_scaled); - ggml_set_name(KQ_scaled, "KQ_scaled"); - - struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled, KQ_mask); - offload_func_kq(KQ_masked); - ggml_set_name(KQ_masked, "KQ_masked"); - - struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked); - offload_func_v(KQ_soft_max); - ggml_set_name(KQ_soft_max, "KQ_soft_max"); - - struct ggml_tensor * V = - ggml_view_3d(ctx0, kv_self.v, - n_kv, n_embd_head, n_head_kv, - ggml_element_size(kv_self.v)*n_ctx, - ggml_element_size(kv_self.v)*n_ctx*n_embd_head, - ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il); - offload_func_v(V); - ggml_set_name(V, "V"); - - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); - offload_func_v(KQV); - ggml_set_name(KQV, "KQV"); - - struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - offload_func_v(KQV_merged); - ggml_set_name(KQV_merged, "KQV_merged"); - - cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens); - offload_func_v(cur); - ggml_set_name(cur, "KQV_merged_contiguous"); - - cur = ggml_mul_mat(ctx0, model.layers[il].wo, cur); - offload_func(cur); - ggml_set_name(cur, "result_wo"); + cur = llm_build_kqv(lctx, ctx0, attn_norm, + model.layers[il].wo, NULL, + Qcur, KQ_scale, KQ_mask, n_tokens, n_kv, -1.0f, cb, il); + cb(cur, "kqv_out", il); } - struct ggml_tensor * attn_out = cur; + struct ggml_tensor * ffn_inp = cur; // feed forward { - struct ggml_tensor * inpFF = attn_norm; - - cur = ggml_mul_mat(ctx0, model.layers[il].w3, inpFF); - offload_func(cur); - - cur = ggml_gelu(ctx0, cur); - offload_func(cur); - cur = ggml_mul_mat(ctx0, model.layers[il].w2, cur); - offload_func(cur); + cur = llm_build_ffn(ctx0, attn_norm, // !! use the attn norm, not the result + model.layers[il].ffn_up, NULL, + NULL, NULL, + model.layers[il].ffn_down, NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); + cb(cur, "ffn_out", il); } - cur = ggml_add(ctx0, cur, attn_out); - offload_func(cur); + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); + cur = ggml_add(ctx0, cur, inpL); - offload_func(cur); + cb(cur, "l_out", il); // input for next layer inpL = cur; @@ -4550,18 +3887,14 @@ static struct ggml_cgraph * llm_build_falcon( cur = inpL; // norm - { - cur = ggml_norm(ctx0, cur, norm_eps); - offload_func_nr(cur); - - cur = ggml_add(ctx0, - ggml_mul(ctx0, cur, model.output_norm), - model.output_norm_b); - ggml_set_name(cur, "result_norm"); - } + cur = llm_build_norm(ctx0, cur, + model.output_norm, + model.output_norm_b, + LLM_NORM, norm_eps, cb, -1); + cb(cur, "result_norm", -1); cur = ggml_mul_mat(ctx0, model.output, cur); - ggml_set_name(cur, "result_output"); + cb(cur, "result_output", -1); ggml_build_forward_expand(gf, cur); @@ -4572,7 +3905,9 @@ static struct ggml_cgraph * llm_build_falcon( static struct ggml_cgraph * llm_build_starcoder( llama_context & lctx, - const llama_batch & batch) { + const llama_batch & batch, + const llm_build_cb & cb, + bool worst_case) { const auto & model = lctx.model; const auto & hparams = model.hparams; const auto & cparams = lctx.cparams; @@ -4585,7 +3920,6 @@ static struct ggml_cgraph * llm_build_starcoder( const int64_t n_layer = hparams.n_layer; const int64_t n_ctx = cparams.n_ctx; const int64_t n_head = hparams.n_head; - const int64_t n_head_kv = hparams.n_head_kv; const int64_t n_embd_head = hparams.n_embd_head(); const int64_t n_embd_gqa = hparams.n_embd_gqa(); @@ -4594,8 +3928,8 @@ static struct ggml_cgraph * llm_build_starcoder( const float norm_eps = hparams.f_norm_eps; const int32_t n_tokens = batch.n_tokens; - const int32_t n_kv = ggml_allocr_is_measure(lctx.alloc) ? n_ctx : kv_self.n; - const int32_t kv_head = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head; + const int32_t n_kv = worst_case ? n_ctx : kv_self.n; + const int32_t kv_head = worst_case ? n_ctx - n_tokens : kv_self.head; auto & buf_compute = lctx.buf_compute; @@ -4610,204 +3944,95 @@ static struct ggml_cgraph * llm_build_starcoder( ggml_cgraph * gf = ggml_new_graph(ctx0); struct ggml_tensor * cur; - struct ggml_tensor * token; - struct ggml_tensor * position; + struct ggml_tensor * pos; struct ggml_tensor * inpL; - if (batch.token) { - struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + inpL = llm_build_inp_embd(ctx0, batch, model.tok_embd, n_embd, n_tokens, cb); + cb(inpL, "inp_embd", -1); - ggml_allocr_alloc(lctx.alloc, inp_tokens); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens)); - } - ggml_set_name(inp_tokens, "inp_tokens"); - - token = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens); - } else { -#ifdef GGML_USE_MPI - GGML_ASSERT(false && "not implemented"); -#endif - - token = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens); - - ggml_allocr_alloc(lctx.alloc, token); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(token->data, batch.embd, n_tokens * n_embd * ggml_element_size(token)); - } - } - - { - // Compute position embeddings. - struct ggml_tensor * inp_positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); - ggml_allocr_alloc(lctx.alloc, inp_positions); - if (!ggml_allocr_is_measure(lctx.alloc)) { - for (int i = 0; i < n_tokens; ++i) { - ((int32_t *) inp_positions->data)[i] = batch.pos[i]; - } - } - ggml_set_name(inp_positions, "inp_positions"); - - position = ggml_get_rows(ctx0, model.pos_embeddings, inp_positions); - } + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + cb(inp_pos, "inp_pos", -1); // KQ_scale struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)"); - ggml_allocr_alloc(lctx.alloc, KQ_scale); - if (!ggml_allocr_is_measure(lctx.alloc)) { - ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head)); - } + cb(KQ_scale, "KQ_scale", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); - ggml_set_name(KQ_mask, "KQ_mask"); - ggml_allocr_alloc(lctx.alloc, KQ_mask); - if (!ggml_allocr_is_measure(lctx.alloc)) { - float * data = (float *) KQ_mask->data; - memset(data, 0, ggml_nbytes(KQ_mask)); + cb(KQ_mask, "KQ_mask", -1); - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - const llama_pos pos = batch.pos[j]; - const llama_seq_id seq_id = batch.seq_id[j][0]; + pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); + cb(pos, "pos_embd", -1); - for (int i = 0; i < n_kv; ++i) { - if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) { - data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY; - } - } - } - } - } - - inpL = ggml_add(ctx0, token, position); - ggml_set_name(inpL, "inpL"); + inpL = ggml_add(ctx0, inpL, pos); + cb(inpL, "inpL", -1); for (int il = 0; il < n_layer; ++il) { + cur = llm_build_norm(ctx0, inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, norm_eps, cb, il); + cb(cur, "attn_norm", il); + + // self-attention { - // Norm - cur = ggml_norm(ctx0, inpL, norm_eps); - cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].attn_norm), model.layers[il].attn_norm_b); + cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + + llm_build_kv_store(lctx, ctx0, gf, Kcur, Vcur, n_tokens, kv_head, cb, il); + + cur = llm_build_kqv(lctx, ctx0, cur, + model.layers[il].wo, model.layers[il].bo, + Qcur, KQ_scale, KQ_mask, n_tokens, n_kv, -1.0f, cb, il); + cb(cur, "kqv_out", il); } - { - // Self Attention - cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wqkv, cur), model.layers[il].bqkv); - - struct ggml_tensor * tmpq = ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*n_embd); - struct ggml_tensor * tmpk = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], sizeof(float)*n_embd); - struct ggml_tensor * tmpv = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], sizeof(float)*(n_embd + n_embd_gqa)); - - struct ggml_tensor * Qcur = tmpq; - struct ggml_tensor * Kcur = tmpk; - - { - struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_cont(ctx0, tmpv), n_embd_gqa, n_tokens)); - ggml_set_name(Vcur, "Vcur"); - - struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head)); - ggml_set_name(k, "k"); - - struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa, - ( n_ctx)*ggml_element_size(kv_self.v), - (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v)); - - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); - } - - struct ggml_tensor * Q = - ggml_permute(ctx0, - ggml_cpy(ctx0, - Qcur, - ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd_head, n_head, n_tokens)), - 0, 2, 1, 3); - ggml_set_name(Q, "Q"); - - struct ggml_tensor * K = - ggml_view_3d(ctx0, kv_self.k, - n_embd_head, n_kv, n_head_kv, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il); - ggml_set_name(K, "K"); - - // K * Q - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - ggml_set_name(KQ, "KQ"); - - // KQ_scaled = KQ / sqrt(n_embd_head) - // KQ_scaled shape [n_past + n_tokens, n_tokens, n_head, 1] - struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, KQ_scale); - ggml_set_name(KQ_scaled, "KQ_scaled"); - - // KQ_masked = mask_past(KQ_scaled) - struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled, KQ_mask); - ggml_set_name(KQ_masked, "KQ_masked"); - - // KQ = soft_max(KQ_masked) - struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked); - ggml_set_name(KQ_soft_max, "KQ_soft_max"); - - // split cached V into n_head heads - struct ggml_tensor * V = - ggml_view_3d(ctx0, kv_self.v, - n_kv, n_embd_head, n_head_kv, - ggml_element_size(kv_self.v)*n_ctx, - ggml_element_size(kv_self.v)*n_ctx*n_embd_head, - ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il); - ggml_set_name(V, "V"); - - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); - ggml_set_name(KQV, "KQV"); - - // KQV_merged = KQV.permute(0, 2, 1, 3) - struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - ggml_set_name(KQV_merged, "KQV_merged"); - - // cur = KQV_merged.contiguous().view(n_embd, n_tokens) - cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens); - ggml_set_name(cur, "KQV_merged_contiguous"); - } - - // Projection - cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wo, cur), model.layers[il].bo); - - // Add the input - cur = ggml_add(ctx0, cur, inpL); - - struct ggml_tensor * inpFF = cur; + // add the input + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); // FF { - // Norm - { - cur = ggml_norm(ctx0, inpFF, norm_eps); - cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ffn_norm), model.layers[il].ffn_norm_b); - } + cur = llm_build_norm(ctx0, ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, norm_eps, cb, il); + cb(cur, "ffn_norm", il); - cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].w3, cur), model.layers[il].b3); - - // GELU activation - cur = ggml_gelu(ctx0, cur); - - // Projection - cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].w2, cur), model.layers[il].b2); + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, + NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, + LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); + cb(cur, "ffn_out", il); } - inpL = ggml_add(ctx0, cur, inpFF); + inpL = ggml_add(ctx0, cur, ffn_inp); + cb(inpL, "l_out", il); } - // Output Norm - { - cur = ggml_norm(ctx0, inpL, norm_eps); - cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.output_norm), model.output_norm_b); - } - ggml_set_name(cur, "result_norm"); + cur = llm_build_norm(ctx0, inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, norm_eps, cb, -1); + cb(cur, "result_norm", -1); cur = ggml_mul_mat(ctx0, model.output, cur); - ggml_set_name(cur, "result_output"); + cb(cur, "result_output", -1); ggml_build_forward_expand(gf, cur); ggml_free(ctx0); @@ -4817,7 +4042,9 @@ static struct ggml_cgraph * llm_build_starcoder( static struct ggml_cgraph * llm_build_persimmon( llama_context & lctx, - const llama_batch & batch) { + const llama_batch & batch, + const llm_build_cb & cb, + bool worst_case) { const auto & model = lctx.model; const auto & hparams = model.hparams; @@ -4826,29 +4053,27 @@ static struct ggml_cgraph * llm_build_persimmon( GGML_ASSERT(!!kv_self.ctx); const auto & cparams = lctx.cparams; + const int64_t n_embd = hparams.n_embd; const int64_t n_layer = hparams.n_layer; const int64_t n_ctx = cparams.n_ctx; const int64_t n_head_kv = hparams.n_head_kv; const int64_t n_head = hparams.n_head; const int64_t n_embd_head = hparams.n_embd_head(); - const int64_t n_embd_gqa = hparams.n_embd_gqa(); - const size_t n_rot = n_embd_head / 2; + const int64_t n_rot = n_embd_head / 2; const float freq_base = cparams.rope_freq_base; const float freq_scale = cparams.rope_freq_scale; - const float norm_eps = hparams.f_norm_eps; - - const int n_gpu_layers = model.n_gpu_layers; - + const float norm_eps = hparams.f_norm_eps; const int32_t n_tokens = batch.n_tokens; - const int32_t n_kv = ggml_allocr_is_measure(lctx.alloc) ? n_ctx : kv_self.n; - const int32_t kv_head = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head; + const int32_t n_kv = worst_case ? n_ctx : kv_self.n; + const int32_t kv_head = worst_case ? n_ctx - n_tokens : kv_self.head; - const bool do_rope_shift = ggml_allocr_is_measure(lctx.alloc) || kv_self.has_shift; + const bool do_rope_shift = worst_case || kv_self.has_shift; auto & buf_compute = lctx.buf_compute; + struct ggml_init_params params = { /*.mem_size =*/ buf_compute.size, /*.mem_buffer =*/ buf_compute.data, @@ -4862,146 +4087,77 @@ static struct ggml_cgraph * llm_build_persimmon( struct ggml_tensor * cur; struct ggml_tensor * inpL; - if (batch.token) { - struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + inpL = llm_build_inp_embd(ctx0, batch, model.tok_embd, n_embd, n_tokens, cb); + cb(inpL, "imp_embd", -1); + + struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + cb(inp_pos, "inp_pos", -1); - ggml_allocr_alloc(lctx.alloc, inp_tokens); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens)); - } - ggml_set_name(inp_tokens, "inp_tokens"); - inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens); - } else { - inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens); - ggml_allocr_alloc(lctx.alloc, inpL); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inpL->data, batch.embd, n_tokens * n_embd * ggml_element_size(inpL)); - } - } - const int i_gpu_start = n_layer - n_gpu_layers; - (void) i_gpu_start; - offload_func_t offload_func_nr = llama_nop; // nr = non-repeating - offload_func_t offload_func_kq = llama_nop; - offload_func_t offload_func_v = llama_nop; // KQ_scale struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - ggml_allocr_alloc(lctx.alloc, KQ_scale); - if (!ggml_allocr_is_measure(lctx.alloc)) { - ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd_head))); - } - ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)"); + cb(KQ_scale, "KQ_scale", -1); + struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); - offload_func_kq(KQ_mask); - ggml_set_name(KQ_mask, "KQ_mask"); - ggml_allocr_alloc(lctx.alloc, KQ_mask); + cb(KQ_mask, "KQ_mask", -1); - if (!ggml_allocr_is_measure(lctx.alloc)) { - float * data = (float *) KQ_mask->data; - memset(data, 0, ggml_nbytes(KQ_mask)); - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - const llama_pos pos = batch.pos[j]; - const llama_seq_id seq_id = batch.seq_id[j][0]; - for (int i = 0; i < n_kv; ++i) { - if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) { - data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY; - } - } - } - } - } - - struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); - offload_func_kq(KQ_pos); - ggml_set_name(KQ_pos, "KQ_pos"); - ggml_allocr_alloc(lctx.alloc, KQ_pos); - if (!ggml_allocr_is_measure(lctx.alloc)) { - int * data = (int *) KQ_pos->data; - for (int i = 0; i < n_tokens; ++i) { - data[i] = batch.pos[i]; - } - } if (do_rope_shift) { - struct ggml_tensor * K_shift = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_ctx); - offload_func_kq(K_shift); - ggml_set_name(K_shift, "K_shift"); - ggml_allocr_alloc(lctx.alloc, K_shift); - if (!ggml_allocr_is_measure(lctx.alloc)) { - int * data = (int *) K_shift->data; - for (int i = 0; i < n_ctx; ++i) { - data[i] = kv_self.cells[i].delta; - } - } - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * tmp = - // we rotate only the first n_rot dimensions. - ggml_rope_custom_inplace(ctx0, - ggml_view_3d(ctx0, kv_self.k, - n_rot, n_head, n_ctx, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*(n_embd_head*n_ctx*il) - ), - K_shift, n_rot, 2, 0, freq_base, freq_scale); - offload_func_kq(tmp); - ggml_build_forward_expand(gf, tmp); - } + llm_build_k_shift(lctx, ctx0, gf, n_rot, LLM_ROPE_NEOX, cb); } - for (int il=0; il < n_layer; ++il) { + + for (int il = 0; il < n_layer; ++il) { struct ggml_tensor * residual = inpL; - offload_func_t offload_func = llama_nop; - { - cur = ggml_norm(ctx0, inpL, norm_eps); - offload_func(cur); - cur = ggml_mul(ctx0, cur, model.layers[il].attn_norm); - offload_func(cur); - cur = ggml_add(ctx0, cur, model.layers[il].attn_norm_b); - offload_func(cur); - ggml_format_name(cur, "input_layernorm_%d", il); - } + + cur = llm_build_norm(ctx0, inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, norm_eps, cb, il); + cb(cur, "attn_norm", il); + // self attention { cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); - offload_func_kq(cur); + cb(cur, "wqkv", il); + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - offload_func_kq(cur); + cb(cur, "bqkv", il); // split qkv GGML_ASSERT(n_head_kv == n_head); - ggml_set_name(cur, format("qkv_%d", il).c_str()); + struct ggml_tensor * tmpqkv = ggml_reshape_4d(ctx0, cur, n_embd_head, 3, n_head, n_tokens); - offload_func_kq(tmpqkv); + cb(tmpqkv, "tmpqkv", il); + struct ggml_tensor * tmpqkv_perm = ggml_cont(ctx0, ggml_permute(ctx0, tmpqkv, 0, 3, 1, 2)); - offload_func_kq(tmpqkv_perm); - ggml_format_name(tmpqkv_perm, "tmpqkv_perm_%d", il); + cb(tmpqkv_perm, "tmpqkv", il); + struct ggml_tensor * tmpq = ggml_view_3d( ctx0, tmpqkv_perm, n_embd_head, n_head, n_tokens, ggml_element_size(tmpqkv_perm) * n_embd_head, ggml_element_size(tmpqkv_perm) * n_embd_head * n_head, 0 ); - offload_func_kq(tmpq); + cb(tmpq, "tmpq", il); + struct ggml_tensor * tmpk = ggml_view_3d( ctx0, tmpqkv_perm, n_embd_head, n_head, n_tokens, ggml_element_size(tmpqkv_perm) * n_embd_head, ggml_element_size(tmpqkv_perm) * n_embd_head * n_head, ggml_element_size(tmpqkv_perm) * n_embd_head * n_head * n_tokens ); - offload_func_kq(tmpk); - // Q/K Layernorm - tmpq = ggml_norm(ctx0, tmpq, norm_eps); - offload_func_kq(tmpq); - tmpq = ggml_mul(ctx0, tmpq, model.layers[il].attn_q_norm); - offload_func_kq(tmpq); - tmpq = ggml_add(ctx0, tmpq, model.layers[il].attn_q_norm_b); - offload_func_kq(tmpq); + cb(tmpk, "tmpk", il); - tmpk = ggml_norm(ctx0, tmpk, norm_eps); - offload_func_v(tmpk); - tmpk = ggml_mul(ctx0, tmpk, model.layers[il].attn_k_norm); - offload_func_v(tmpk); - tmpk = ggml_add(ctx0, tmpk, model.layers[il].attn_k_norm_b); - offload_func_v(tmpk); + // Q/K Layernorm + tmpq = llm_build_norm(ctx0, tmpq, + model.layers[il].attn_q_norm, + model.layers[il].attn_q_norm_b, + LLM_NORM, norm_eps, cb, il); + cb(tmpq, "tmpq", il); + + tmpk = llm_build_norm(ctx0, tmpk, + model.layers[il].attn_k_norm, + model.layers[il].attn_k_norm_b, + LLM_NORM, norm_eps, cb, il); + cb(tmpk, "tmpk", il); // RoPE the first n_rot of q/k, pass the other half, and concat. struct ggml_tensor * qrot = ggml_view_3d( @@ -5010,16 +4166,15 @@ static struct ggml_cgraph * llm_build_persimmon( ggml_element_size(tmpq) * n_embd_head * n_head, 0 ); - offload_func_kq(qrot); - ggml_format_name(qrot, "qrot_%d", il); + cb(qrot, "qrot", il); + struct ggml_tensor * krot = ggml_view_3d( ctx0, tmpk, n_rot, n_head, n_tokens, ggml_element_size(tmpk) * n_embd_head, ggml_element_size(tmpk) * n_embd_head * n_head, 0 ); - offload_func_kq(krot); - ggml_format_name(krot, "krot_%d", il); + cb(krot, "krot", il); // get the second half of tmpq, e.g tmpq[n_rot:, :, :] struct ggml_tensor * qpass = ggml_view_3d( @@ -5028,193 +4183,117 @@ static struct ggml_cgraph * llm_build_persimmon( ggml_element_size(tmpq) * n_embd_head * n_head, ggml_element_size(tmpq) * n_rot ); - offload_func_kq(qpass); - ggml_format_name(qpass, "qpass_%d", il); + cb(qpass, "qpass", il); + struct ggml_tensor * kpass = ggml_view_3d( ctx0, tmpk, n_rot, n_head, n_tokens, ggml_element_size(tmpk) * n_embd_head, ggml_element_size(tmpk) * n_embd_head * n_head, ggml_element_size(tmpk) * n_rot ); - offload_func_kq(kpass); - ggml_format_name(kpass, "kpass_%d", il); + cb(kpass, "kpass", il); - struct ggml_tensor * qrotated = ggml_rope_custom( - ctx0, qrot, KQ_pos, n_rot, 2, 0, freq_base, freq_scale + struct ggml_tensor * qrotated = ggml_rope_custom( + ctx0, qrot, inp_pos, n_rot, 2, 0, freq_base, freq_scale ); - offload_func_kq(qrotated); + cb(qrotated, "qrotated", il); + struct ggml_tensor * krotated = ggml_rope_custom( - ctx0, krot, KQ_pos, n_rot, 2, 0, freq_base, freq_scale + ctx0, krot, inp_pos, n_rot, 2, 0, freq_base, freq_scale ); - offload_func_kq(krotated); + cb(krotated, "krotated", il); + // ggml currently only supports concatenation on dim=2 // so we need to permute qrot, qpass, concat, then permute back. qrotated = ggml_cont(ctx0, ggml_permute(ctx0, qrotated, 2, 1, 0, 3)); - offload_func_kq(qrotated); + cb(qrotated, "qrotated", il); + krotated = ggml_cont(ctx0, ggml_permute(ctx0, krotated, 2, 1, 0, 3)); - offload_func_kq(krotated); + cb(krotated, "krotated", il); qpass = ggml_cont(ctx0, ggml_permute(ctx0, qpass, 2, 1, 0, 3)); - offload_func_kq(qpass); + cb(qpass, "qpass", il); + kpass = ggml_cont(ctx0, ggml_permute(ctx0, kpass, 2, 1, 0, 3)); - offload_func_kq(kpass); + cb(kpass, "kpass", il); struct ggml_tensor * Qcur = ggml_concat(ctx0, qrotated, qpass); - offload_func_kq(Qcur); + cb(Qcur, "Qcur", il); + struct ggml_tensor * Kcur = ggml_concat(ctx0, krotated, kpass); - offload_func_kq(Kcur); + cb(Kcur, "Kcur", il); struct ggml_tensor * Q = ggml_cont(ctx0, ggml_permute(ctx0, Qcur, 1, 2, 0, 3)); - offload_func_kq(Q); + cb(Q, "Q", il); Kcur = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 2, 1, 0, 3)); - offload_func_kq(Kcur); - { - struct ggml_tensor * tmpv = ggml_view_3d( - ctx0, tmpqkv_perm, n_embd_head, n_head, n_tokens, - ggml_element_size(tmpqkv_perm) * n_embd_head, - ggml_element_size(tmpqkv_perm) * n_embd_head * n_head, - ggml_element_size(tmpqkv_perm) * n_embd_head * n_head * n_tokens * 2 + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = ggml_view_3d( + ctx0, tmpqkv_perm, n_embd_head, n_head, n_tokens, + ggml_element_size(tmpqkv_perm) * n_embd_head, + ggml_element_size(tmpqkv_perm) * n_embd_head * n_head, + ggml_element_size(tmpqkv_perm) * n_embd_head * n_head * n_tokens * 2 ); - offload_func_v(tmpv); - // store K, V in cache - struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, tmpv, n_embd_gqa, n_tokens)); - offload_func_v(Vcur); - ggml_set_name(Vcur, "Vcur"); + cb(Vcur, "Vcur", il); - struct ggml_tensor * k = ggml_view_1d( - ctx0, kv_self.k, n_tokens*n_embd_gqa, - (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head) - ); - offload_func_kq(k); - ggml_set_name(k, "k"); + llm_build_kv_store(lctx, ctx0, gf, Kcur, Vcur, n_tokens, kv_head, cb, il); - struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa, - ( n_ctx)*ggml_element_size(kv_self.v), - (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v)); - offload_func_v(v); - ggml_set_name(v, "v"); - - // important: storing RoPE-ed version of K in the KV cache! - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); - } - struct ggml_tensor * K = ggml_view_3d(ctx0, kv_self.k, - n_embd_head, n_kv, n_head_kv, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il); - - offload_func_kq(K); - ggml_format_name(K, "K_%d", il); - - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - offload_func_kq(KQ); - ggml_set_name(KQ, "KQ"); - - struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale); - offload_func_kq(KQ_scaled); - ggml_set_name(KQ_scaled, "KQ_scaled"); - - struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled, KQ_mask); - offload_func_kq(KQ_masked); - ggml_set_name(KQ_masked, "KQ_masked"); - - struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked); - offload_func_kq(KQ_soft_max); - ggml_set_name(KQ_soft_max, "KQ_soft_max"); - - struct ggml_tensor * V = - ggml_view_3d(ctx0, kv_self.v, - n_kv, n_embd_head, n_head_kv, - ggml_element_size(kv_self.v)*n_ctx, - ggml_element_size(kv_self.v)*n_ctx*n_embd_head, - ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il); - offload_func_v(V); - ggml_set_name(V, "V"); - - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); - offload_func_v(KQV); - ggml_set_name(KQV, "KQV"); - - struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - offload_func_v(KQV_merged); - ggml_set_name(KQV_merged, "KQV_merged"); - - cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens); - offload_func_v(cur); - ggml_set_name(cur, "KQV_merged_contiguous"); - - cur = ggml_mul_mat(ctx0, model.layers[il].wo, cur); - offload_func(cur); - cur = ggml_add(ctx0, cur, model.layers[il].bo); - offload_func(cur); - ggml_set_name(cur, "result_wo"); + // TODO: not tested, could be broken + cur = llm_build_kqv(lctx, ctx0, Q, + model.layers[il].wo, model.layers[il].bo, + Q, KQ_scale, KQ_mask, n_tokens, n_kv, -1.0f, cb, il); + cb(cur, "kqv_out", il); } - struct ggml_tensor * inpFF = ggml_add(ctx0, residual, cur); - offload_func(inpFF); - ggml_set_name(inpFF, "inpFF"); + struct ggml_tensor * ffn_inp = ggml_add(ctx0, residual, cur); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network { - // MLP - { - // Norm - cur = ggml_norm(ctx0, inpFF, norm_eps); - offload_func(cur); - cur = ggml_add(ctx0, - ggml_mul(ctx0, cur, model.layers[il].ffn_norm), - model.layers[il].ffn_norm_b - ); - ggml_set_name(cur, "ffn_norm"); - offload_func(cur); - } - cur = ggml_mul_mat(ctx0, model.layers[il].w3, cur); - offload_func(cur); + cur = llm_build_norm(ctx0, ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, norm_eps, cb, il); + cb(cur, "ffn_norm", il); - cur = ggml_add(ctx0, cur, model.layers[il].b3); - offload_func(cur); - ggml_set_name(cur, "result_ffn_up"); - - cur = ggml_sqr(ctx0, ggml_relu(ctx0, cur)); - ggml_set_name(cur, "result_ffn_act"); - offload_func(cur); - offload_func(cur->src[0]); - - cur = ggml_mul_mat(ctx0, model.layers[il].w2, cur); - offload_func(cur); - cur = ggml_add(ctx0, - cur, - model.layers[il].b2); - offload_func(cur); - ggml_set_name(cur, "outFF"); + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, + NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, + LLM_FFN_RELU_SQR, LLM_FFN_SEQ, cb, il); + cb(cur, "ffn_out", il); } - cur = ggml_add(ctx0, cur, inpFF); - offload_func(cur); - ggml_set_name(cur, "inpFF_+_outFF"); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); + inpL = cur; } + cur = inpL; - { - cur = ggml_norm(ctx0, cur, norm_eps); - offload_func_nr(cur); - cur = ggml_mul(ctx0, cur, model.output_norm); - offload_func_nr(cur); - cur = ggml_add(ctx0, cur, model.output_norm_b); - // offload_func_nr(cur); + cur = llm_build_norm(ctx0, cur, + model.output_norm, + model.output_norm_b, + LLM_NORM, norm_eps, cb, -1); + cb(cur, "result_norm", -1); - ggml_set_name(cur, "result_norm"); - } cur = ggml_mul_mat(ctx0, model.output, cur); - ggml_set_name(cur, "result_output"); + cb(cur, "result_output", -1); + ggml_build_forward_expand(gf, cur); + ggml_free(ctx0); + return gf; } -static struct ggml_cgraph * llm_build_bloom( +static struct ggml_cgraph * llm_build_refact( llama_context & lctx, - const llama_batch & batch) { + const llama_batch & batch, + const llm_build_cb & cb, + bool worst_case) { const auto & model = lctx.model; const auto & hparams = model.hparams; const auto & cparams = lctx.cparams; @@ -5229,6 +4308,133 @@ static struct ggml_cgraph * llm_build_bloom( const int64_t n_head = hparams.n_head; const int64_t n_head_kv = hparams.n_head_kv; const int64_t n_embd_head = hparams.n_embd_head(); + + const float norm_rms_eps = hparams.f_norm_rms_eps; + + const int32_t n_tokens = batch.n_tokens; + const int32_t n_kv = worst_case ? n_ctx : kv_self.n; + const int32_t kv_head = worst_case ? n_ctx - n_tokens : kv_self.head; + + auto & buf_compute = lctx.buf_compute; + + struct ggml_init_params params = { + /*.mem_size =*/ buf_compute.size, + /*.mem_buffer =*/ buf_compute.data, + /*.no_alloc =*/ true, + }; + + struct ggml_context * ctx0 = ggml_init(params); + + ggml_cgraph * gf = ggml_new_graph(ctx0); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, batch, model.tok_embd, n_embd, n_tokens, cb); + cb(inpL, "inp_embd", -1); + + // KQ_scale + struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); + cb(KQ_scale, "KQ_scale", -1); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + cb(KQ_mask, "KQ_mask", -1); + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + cur = llm_build_norm(ctx0, inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, norm_rms_eps, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + cb(Kcur, "Kcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + cb(Qcur, "Qcur", il); + + llm_build_kv_store(lctx, ctx0, gf, Kcur, Vcur, n_tokens, kv_head, cb, il); + + cur = llm_build_kqv(lctx, ctx0, Qcur, + model.layers[il].wo, NULL, + Qcur, KQ_scale, KQ_mask, n_tokens, n_kv, 8.0f, cb, il); + cb(cur, "kqv_out", il); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = llm_build_norm(ctx0, ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, norm_rms_eps, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, + model.layers[il].ffn_gate, NULL, + model.layers[il].ffn_down, NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = llm_build_norm(ctx0, cur, + model.output_norm, NULL, + LLM_NORM_RMS, norm_rms_eps, cb, -1); + cb(cur, "result_norm", -1); + + // lm_head + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + ggml_free(ctx0); + + return gf; +} + +static struct ggml_cgraph * llm_build_bloom( + llama_context & lctx, + const llama_batch & batch, + const llm_build_cb & cb, + bool worst_case) { + const auto & model = lctx.model; + const auto & hparams = model.hparams; + const auto & cparams = lctx.cparams; + + const auto & kv_self = lctx.kv_self; + + GGML_ASSERT(!!kv_self.ctx); + + const int64_t n_embd = hparams.n_embd; + const int64_t n_layer = hparams.n_layer; + const int64_t n_ctx = cparams.n_ctx; + const int64_t n_head = hparams.n_head; + const int64_t n_embd_head = hparams.n_embd_head(); const int64_t n_embd_gqa = hparams.n_embd_gqa(); GGML_ASSERT(n_embd_head == hparams.n_rot); @@ -5236,8 +4442,8 @@ static struct ggml_cgraph * llm_build_bloom( const float norm_eps = hparams.f_norm_eps; const int32_t n_tokens = batch.n_tokens; - const int32_t n_kv = ggml_allocr_is_measure(lctx.alloc) ? n_ctx : kv_self.n; - const int32_t kv_head = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head; + const int32_t n_kv = worst_case ? n_ctx : kv_self.n; + const int32_t kv_head = worst_case ? n_ctx - n_tokens : kv_self.head; auto & buf_compute = lctx.buf_compute; @@ -5254,198 +4460,90 @@ static struct ggml_cgraph * llm_build_bloom( ggml_cgraph * gf = ggml_new_graph(ctx0); struct ggml_tensor * cur; - struct ggml_tensor * token; struct ggml_tensor * inpL; - if (batch.token) { - struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); - - ggml_allocr_alloc(lctx.alloc, inp_tokens); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens)); - } - ggml_set_name(inp_tokens, "inp_tokens"); - - token = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens); - } else { -#ifdef GGML_USE_MPI - GGML_ASSERT(false && "not implemented"); -#endif - - token = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens); - - ggml_allocr_alloc(lctx.alloc, token); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(token->data, batch.embd, n_tokens * n_embd * ggml_element_size(token)); - } - } + inpL = llm_build_inp_embd(ctx0, batch, model.tok_embd, n_embd, n_tokens, cb); + cb(inpL, "inp_embd", -1); // KQ_scale struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)"); - ggml_allocr_alloc(lctx.alloc, KQ_scale); - if (!ggml_allocr_is_measure(lctx.alloc)) { - ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head)); - } + cb(KQ_scale, "KQ_scale", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); - ggml_set_name(KQ_mask, "KQ_mask"); - ggml_allocr_alloc(lctx.alloc, KQ_mask); - if (!ggml_allocr_is_measure(lctx.alloc)) { - float * data = (float *) KQ_mask->data; - memset(data, 0, ggml_nbytes(KQ_mask)); + cb(KQ_mask, "KQ_mask", -1); - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - const llama_pos pos = batch.pos[j]; - const llama_seq_id seq_id = batch.seq_id[j][0]; - - for (int i = 0; i < n_kv; ++i) { - if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) { - data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY; - } - } - } - } - } - - // norm - { - inpL = ggml_norm(ctx0, token, norm_eps); - inpL = ggml_add(ctx0, ggml_mul(ctx0, inpL, model.tok_norm), model.tok_norm_b); - } - - ggml_set_name(inpL, "inpL"); + inpL = llm_build_norm(ctx0, inpL, + model.tok_norm, + model.tok_norm_b, + LLM_NORM, norm_eps, cb, -1); + cb(inpL, "inp_norm", -1); for (int il = 0; il < n_layer; ++il) { + cur = llm_build_norm(ctx0, inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, norm_eps, cb, il); + cb(cur, "attn_norm", il); + + // self-attention { - // Norm - cur = ggml_norm(ctx0, inpL, norm_eps); - cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].attn_norm), model.layers[il].attn_norm_b); + cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + + llm_build_kv_store(lctx, ctx0, gf, Kcur, Vcur, n_tokens, kv_head, cb, il); + + cur = llm_build_kqv(lctx, ctx0, Qcur, + model.layers[il].wo, model.layers[il].bo, + Qcur, KQ_scale, KQ_mask, n_tokens, n_kv, 8.0f, cb, il); + cb(cur, "kqv_out", il); } - { - // Self Attention - cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wqkv, cur), model.layers[il].bqkv); - - struct ggml_tensor * tmpq = ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*n_embd); - struct ggml_tensor * tmpk = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], sizeof(float)*n_embd); - struct ggml_tensor * tmpv = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], sizeof(float)*(n_embd + n_embd_gqa)); - - struct ggml_tensor * Qcur = tmpq; - struct ggml_tensor * Kcur = tmpk; - - // store key and value to memory - { - struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_cont(ctx0, tmpv), n_embd_gqa, n_tokens)); - ggml_set_name(Vcur, "Vcur"); - - struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head)); - ggml_set_name(k, "k"); - - struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa, - ( n_ctx)*ggml_element_size(kv_self.v), - (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v)); - - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); - } - - struct ggml_tensor * Q = - ggml_permute(ctx0, - ggml_cpy(ctx0, - Qcur, - ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd_head, n_head, n_tokens)), - 0, 2, 1, 3); - ggml_set_name(Q, "Q"); - - struct ggml_tensor * K = - ggml_view_3d(ctx0, kv_self.k, - n_embd_head, n_kv, n_head_kv, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il); - ggml_set_name(K, "K"); - - // K * Q - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - ggml_set_name(KQ, "KQ"); - - // KQ_scaled = KQ / sqrt(n_embd_head) - // KQ_scaled shape [n_past + n_tokens, n_tokens, n_head, 1] - struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, KQ_scale); - ggml_set_name(KQ_scaled, "KQ_scaled"); - - struct ggml_tensor * KQ_scaled_alibi = ggml_alibi(ctx0, KQ_scaled, /*n_past*/ kv_head, n_head, 8); - ggml_set_name(KQ_scaled_alibi, "KQ_scaled_alibi"); - - // KQ_masked = mask_past(KQ_scaled) - struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled_alibi, KQ_mask); - ggml_set_name(KQ_masked, "KQ_masked"); - - // KQ = soft_max(KQ_masked) - struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked); - ggml_set_name(KQ_soft_max, "KQ_soft_max"); - - // split cached V into n_head heads - struct ggml_tensor * V = - ggml_view_3d(ctx0, kv_self.v, - n_kv, n_embd_head, n_head_kv, - ggml_element_size(kv_self.v)*n_ctx, - ggml_element_size(kv_self.v)*n_ctx*n_embd_head, - ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il); - ggml_set_name(V, "V"); - - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); - ggml_set_name(KQV, "KQV"); - - // KQV_merged = KQV.permute(0, 2, 1, 3) - struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - ggml_set_name(KQV_merged, "KQV_merged"); - - // cur = KQV_merged.contiguous().view(n_embd, n_tokens) - cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens); - ggml_set_name(cur, "KQV_merged_contiguous"); - } - - // Projection - cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wo, cur), model.layers[il].bo); - // Add the input - cur = ggml_add(ctx0, cur, inpL); - - struct ggml_tensor * inpFF = cur; + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); // FF { - // Norm - { - cur = ggml_norm(ctx0, inpFF, norm_eps); - cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ffn_norm), model.layers[il].ffn_norm_b); - } + cur = llm_build_norm(ctx0, ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, norm_eps, cb, il); + cb(cur, "ffn_norm", il); - cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].w3, cur), model.layers[il].b3); - - // GELU activation - cur = ggml_gelu(ctx0, cur); - - // Projection - cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].w2, cur), model.layers[il].b2); + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, + NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, + LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); + cb(cur, "ffn_out", il); } - inpL = ggml_add(ctx0, cur, inpFF); + inpL = ggml_add(ctx0, cur, ffn_inp); + cb(inpL, "l_out", il); } - // Output Norm - { - cur = ggml_norm(ctx0, inpL, norm_eps); - cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.output_norm), model.output_norm_b); - } - ggml_set_name(cur, "result_norm"); + cur = llm_build_norm(ctx0, inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, norm_eps, cb, -1); + cb(cur, "result_norm", -1); cur = ggml_mul_mat(ctx0, model.output, cur); - ggml_set_name(cur, "result_output"); + cb(cur, "result_output", -1); ggml_build_forward_expand(gf, cur); @@ -5456,7 +4554,9 @@ static struct ggml_cgraph * llm_build_bloom( static struct ggml_cgraph * llm_build_mpt( llama_context & lctx, - const llama_batch & batch) { + const llama_batch & batch, + const llm_build_cb & cb, + bool worst_case) { const auto & model = lctx.model; const auto & hparams = model.hparams; const auto & cparams = lctx.cparams; @@ -5469,7 +4569,6 @@ static struct ggml_cgraph * llm_build_mpt( const int64_t n_layer = hparams.n_layer; const int64_t n_ctx = cparams.n_ctx; const int64_t n_head = hparams.n_head; - const int64_t n_head_kv = hparams.n_head_kv; const int64_t n_embd_head = hparams.n_embd_head(); const int64_t n_embd_gqa = hparams.n_embd_gqa(); @@ -5477,11 +4576,9 @@ static struct ggml_cgraph * llm_build_mpt( const float clamp_kqv = hparams.f_clamp_kqv; const float max_alibi_bias = hparams.f_max_alibi_bias; - const int n_gpu_layers = model.n_gpu_layers; - const int32_t n_tokens = batch.n_tokens; - const int32_t n_kv = ggml_allocr_is_measure(lctx.alloc) ? n_ctx : kv_self.n; - const int32_t kv_head = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head; + const int32_t n_kv = worst_case ? n_ctx : kv_self.n; + const int32_t kv_head = worst_case ? n_ctx - n_tokens : kv_self.head; auto & buf_compute = lctx.buf_compute; @@ -5500,269 +4597,93 @@ static struct ggml_cgraph * llm_build_mpt( struct ggml_tensor * cur; struct ggml_tensor * inpL; - //int warmup = 0; - if (batch.token) { - struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); - - ggml_allocr_alloc(lctx.alloc, inp_tokens); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens)); - //warmup = ((uint32_t*) inp_tokens->data)[0] == 0; - } - - ggml_set_name(inp_tokens, "inp_tokens"); - - inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens); - } else { -#ifdef GGML_USE_MPI - GGML_ASSERT(false && "not implemented"); -#endif - - inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens); - - ggml_allocr_alloc(lctx.alloc, inpL); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inpL->data, batch.embd, n_tokens * n_embd * ggml_element_size(inpL)); - } - } - - const int i_gpu_start = n_layer - n_gpu_layers; - (void) i_gpu_start; - - // offload functions set the tensor output backend to GPU - // tensors are GPU-accelerated if any input or the output has been offloaded - offload_func_t offload_func_nr = llama_nop; // nr = non-repeating - offload_func_t offload_func_kq = llama_nop; - offload_func_t offload_func_v = llama_nop; - -#ifdef GGML_USE_CUBLAS - if (n_gpu_layers > n_layer) { - offload_func_nr = ggml_cuda_assign_buffers_no_alloc; - } - if (n_gpu_layers > n_layer + 1) { - offload_func_v = ggml_cuda_assign_buffers_no_alloc; - } - if (n_gpu_layers > n_layer + 2) { - offload_func_kq = ggml_cuda_assign_buffers_no_alloc; - } -#endif // GGML_USE_CUBLAS + inpL = llm_build_inp_embd(ctx0, batch, model.tok_embd, n_embd, n_tokens, cb); + cb(inpL, "inp_embd", -1); // KQ_scale struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)"); - ggml_allocr_alloc(lctx.alloc, KQ_scale); - if (!ggml_allocr_is_measure(lctx.alloc)) { - ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head)); - } + cb(KQ_scale, "KQ_scale", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); - offload_func_kq(KQ_mask); - ggml_set_name(KQ_mask, "KQ_mask"); - ggml_allocr_alloc(lctx.alloc, KQ_mask); - if (!ggml_allocr_is_measure(lctx.alloc)) { - float * data = (float *) KQ_mask->data; - memset(data, 0, ggml_nbytes(KQ_mask)); - - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - const llama_pos pos = batch.pos[j]; - const llama_seq_id seq_id = batch.seq_id[j][0]; - - for (int i = 0; i < n_kv; ++i) { - if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) { - data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY; - } - } - } - } - } + cb(KQ_mask, "KQ_mask", -1); for (int il = 0; il < n_layer; ++il) { struct ggml_tensor * attn_norm; - offload_func_t offload_func = llama_nop; - -#ifdef GGML_USE_CUBLAS - if (il >= i_gpu_start) { - offload_func = ggml_cuda_assign_buffers_no_alloc; - } -#endif // GGML_USE_CUBLAS + attn_norm = llm_build_norm(ctx0, inpL, + model.layers[il].attn_norm, + NULL, + LLM_NORM, norm_eps, cb, il); + cb(attn_norm, "attn_norm", il); // self-attention - // TODO: refactor into common function (shared with LLaMA) { - attn_norm = ggml_norm(ctx0, inpL, norm_eps); - offload_func(attn_norm); - - attn_norm = ggml_mul(ctx0, attn_norm, model.layers[il].attn_norm); - offload_func(attn_norm); - - if (1) { - cur = attn_norm; - } - - // compute QKV + cur = attn_norm; cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); - offload_func_kq(cur); + cb(cur, "wqkv", il); if (clamp_kqv > 0.0f) { cur = ggml_clamp(ctx0, cur, -clamp_kqv, clamp_kqv); - offload_func_kq(cur); + cb(cur, "wqkv_clamped", il); } - const size_t wsize = ggml_type_size(cur->type); + struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - struct ggml_tensor * Qcur = ggml_view_3d( - ctx0, cur, n_embd_head, n_head, n_tokens, - wsize * n_embd_head, - wsize * n_embd_head * (n_head + 2 * n_head_kv), - 0); - offload_func_kq(Qcur); + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); - struct ggml_tensor * Kcur = ggml_view_3d( - ctx0, cur, n_embd_head, n_head_kv, n_tokens, - wsize * n_embd_head, - wsize * n_embd_head * (n_head + 2 * n_head_kv), - wsize * n_embd_head * n_head); - offload_func_kq(Kcur); + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - struct ggml_tensor * tmpv = ggml_view_3d( - ctx0, cur, n_embd_head, n_head_kv, n_tokens, - wsize * n_embd_head, - wsize * n_embd_head * (n_head + 2 * n_head_kv), - wsize * n_embd_head * (n_head + n_head_kv)); - offload_func_kq(Kcur); + llm_build_kv_store(lctx, ctx0, gf, Kcur, Vcur, n_tokens, kv_head, cb, il); - ggml_set_name(Qcur, "Qcur"); - ggml_set_name(Kcur, "Kcur"); - - { - struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_cont(ctx0, tmpv), n_embd_gqa, n_tokens)); - offload_func_v(Vcur); - offload_func_v(Vcur->src[0]->src[0]); - ggml_set_name(Vcur, "Vcur"); - - struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head)); - offload_func_kq(k); - ggml_set_name(k, "k"); - - struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa, - ( n_ctx)*ggml_element_size(kv_self.v), - (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v)); - offload_func_v(v); - - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); - } - - struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); - offload_func_kq(Q); - ggml_set_name(Q, "Q"); - - struct ggml_tensor * K = - ggml_view_3d(ctx0, kv_self.k, - n_embd_head, n_kv, n_head_kv, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il); - offload_func_kq(K); - ggml_set_name(K, "K"); - - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - offload_func_kq(KQ); - ggml_set_name(KQ, "KQ"); - - struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale); - offload_func_kq(KQ_scaled); - ggml_set_name(KQ_scaled, "KQ_scaled"); - - // TODO: replace with ggml_add() - struct ggml_tensor * KQ_scaled_alibi = - ggml_alibi(ctx0, KQ_scaled, 0, n_head, max_alibi_bias); - offload_func_kq(KQ_scaled_alibi); - ggml_set_name(KQ_scaled_alibi, "KQ_scaled_alibi"); - - struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled_alibi, KQ_mask); - offload_func_kq(KQ_masked); - ggml_set_name(KQ_masked, "KQ_masked"); - - struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked); - offload_func_v(KQ_soft_max); - ggml_set_name(KQ_soft_max, "KQ_soft_max"); - - struct ggml_tensor * V = - ggml_view_3d(ctx0, kv_self.v, - n_kv, n_embd_head, n_head_kv, - ggml_element_size(kv_self.v)*n_ctx, - ggml_element_size(kv_self.v)*n_ctx*n_embd_head, - ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il); - offload_func_v(V); - ggml_set_name(V, "V"); - - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); - offload_func_v(KQV); - ggml_set_name(KQV, "KQV"); - - struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - offload_func_v(KQV_merged); - ggml_set_name(KQV_merged, "KQV_merged"); - - cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens); - offload_func_v(cur); - ggml_set_name(cur, "KQV_merged_contiguous"); - - cur = ggml_mul_mat(ctx0, model.layers[il].wo, cur); - offload_func(cur); - ggml_set_name(cur, "result_wo"); + cur = llm_build_kqv(lctx, ctx0, Qcur, + model.layers[il].wo, NULL, + Qcur, KQ_scale, KQ_mask, n_tokens, n_kv, max_alibi_bias, cb, il); + cb(cur, "kqv_out", il); } // Add the input - cur = ggml_add(ctx0, cur, inpL); - offload_func(cur); - - struct ggml_tensor * attn_out = cur; + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); // feed forward { - // Norm - { - cur = ggml_norm(ctx0, attn_out, norm_eps); - offload_func(cur); + cur = llm_build_norm(ctx0, ffn_inp, + model.layers[il].ffn_norm, + NULL, + LLM_NORM, norm_eps, cb, il); + cb(cur, "ffn_norm", il); - cur = ggml_mul(ctx0, cur, model.layers[il].ffn_norm); - offload_func(cur); - } - - cur = ggml_mul_mat(ctx0, model.layers[il].w3, cur); - offload_func(cur); - - cur = ggml_gelu(ctx0, cur); - offload_func(cur); - cur = ggml_mul_mat(ctx0, model.layers[il].w2, cur); - offload_func(cur); + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, + NULL, NULL, + model.layers[il].ffn_down, NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); + cb(cur, "ffn_out", il); } - cur = ggml_add(ctx0, cur, attn_out); - offload_func(cur); + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); + // input for next layer inpL = cur; } cur = inpL; - // norm - { - cur = ggml_norm(ctx0, cur, norm_eps); - offload_func_nr(cur); - - cur = ggml_mul(ctx0, cur, model.output_norm); - ggml_set_name(cur, "result_norm"); - } + cur = llm_build_norm(ctx0, cur, + model.output_norm, + NULL, + LLM_NORM, norm_eps, cb, -1); + cb(cur, "result_norm", -1); cur = ggml_mul_mat(ctx0, model.output, cur); - ggml_set_name(cur, "result_output"); + cb(cur, "result_output", -1); ggml_build_forward_expand(gf, cur); @@ -5771,50 +4692,494 @@ static struct ggml_cgraph * llm_build_mpt( return gf; } +// +// tensor offloading helpers +// +// TODO: will be removed with backend v2 + +enum llm_offload_func_e { + OFFLOAD_FUNC_NOP, + OFFLOAD_FUNC, + OFFLOAD_FUNC_KQ, + OFFLOAD_FUNC_V, + OFFLOAD_FUNC_NR, + OFFLOAD_FUNC_EMB, + OFFLOAD_FUNC_OUT, +}; + +// TODO: will be removed with backend v2 +struct llm_offload_trie { + struct node { + ~node() { + for (int i = 0; i < 256; ++i) { + if (children[i]) { + delete children[i]; + } + } + } + + node * children[256] = { nullptr }; + llm_offload_func_e func = OFFLOAD_FUNC_NOP; + }; + + llm_offload_trie() { + root = new node; + } + + llm_offload_trie(const std::unordered_map & map) { + root = new node; + + for (const auto & kv : map) { + add(kv.first, kv.second); + } + } + + ~llm_offload_trie() { + delete root; + } + + void add(const char * name, llm_offload_func_e func) { + node * cur = root; + + for (int i = 0; ; ++i) { + const uint8_t c = name[i]; + + if (!c) { + break; + } + + if (!cur->children[c]) { + cur->children[c] = new node; + } + + cur = cur->children[c]; + } + + cur->func = func; + } + + llm_offload_func_e find(const char * name) const { + const node * cur = root; + + for (int i = 0; ; ++i) { + const uint8_t c = name[i]; + + if (!c) { + break; + } + + if (!cur->children[c]) { + return OFFLOAD_FUNC_NOP; + } + + cur = cur->children[c]; + } + + return cur->func; + } + + node * root = nullptr; +}; + +// TODO: will be removed with backend v2 +static const std::unordered_map k_offload_map = { + //{ "inp_tokens", OFFLOAD_FUNC_NR }, // TODO: missing K-quants get_rows kernel + //{ "inp_embd", OFFLOAD_FUNC_NR }, // TODO: missing K-quants get_rows kernel + { "pos_embd", OFFLOAD_FUNC_NR }, + + { "inp_pos", OFFLOAD_FUNC_KQ }, // this is often used for KQ ops (e.g. rope) + { "KQ_scale", OFFLOAD_FUNC_KQ }, + { "KQ_mask", OFFLOAD_FUNC_KQ }, + { "K_shift", OFFLOAD_FUNC_KQ }, + { "K_shifted", OFFLOAD_FUNC_KQ }, + + { "inp_norm", OFFLOAD_FUNC_NR }, + { "inp_norm_w", OFFLOAD_FUNC_NR }, + { "inp_norm_wb", OFFLOAD_FUNC_NR }, + + { "norm", OFFLOAD_FUNC }, + { "norm_w", OFFLOAD_FUNC }, + { "norm_wb", OFFLOAD_FUNC }, + + { "attn_norm", OFFLOAD_FUNC }, + { "attn_norm_2", OFFLOAD_FUNC }, + + { "wqkv", OFFLOAD_FUNC_KQ }, + { "bqkv", OFFLOAD_FUNC_KQ }, + { "wqkv_clamped", OFFLOAD_FUNC_KQ }, + + { "tmpk", OFFLOAD_FUNC_KQ }, + { "tmpq", OFFLOAD_FUNC_KQ }, + { "tmpv", OFFLOAD_FUNC_V }, + { "Kcur", OFFLOAD_FUNC_KQ }, + { "Qcur", OFFLOAD_FUNC_KQ }, + { "Vcur", OFFLOAD_FUNC_V }, + + { "krot", OFFLOAD_FUNC_KQ }, + { "qrot", OFFLOAD_FUNC_KQ }, + { "kpass", OFFLOAD_FUNC_KQ }, + { "qpass", OFFLOAD_FUNC_KQ }, + { "krotated", OFFLOAD_FUNC_KQ }, + { "qrotated", OFFLOAD_FUNC_KQ }, + + { "q", OFFLOAD_FUNC_KQ }, + { "k", OFFLOAD_FUNC_KQ }, + { "kq", OFFLOAD_FUNC_KQ }, + { "kq_scaled", OFFLOAD_FUNC_KQ }, + { "kq_scaled_alibi", OFFLOAD_FUNC_KQ }, + { "kq_masked", OFFLOAD_FUNC_KQ }, + { "kq_soft_max", OFFLOAD_FUNC_V }, + { "v", OFFLOAD_FUNC_V }, + { "kqv", OFFLOAD_FUNC_V }, + { "kqv_merged", OFFLOAD_FUNC_V }, + { "kqv_merged_cont", OFFLOAD_FUNC_V }, + { "kqv_wo", OFFLOAD_FUNC_V }, + { "kqv_out", OFFLOAD_FUNC_V }, + + { "ffn_inp", OFFLOAD_FUNC }, + { "ffn_norm", OFFLOAD_FUNC }, + + { "ffn_up", OFFLOAD_FUNC }, + { "ffn_up_b", OFFLOAD_FUNC }, + { "ffn_gate", OFFLOAD_FUNC }, + { "ffn_gate_b", OFFLOAD_FUNC }, + { "ffn_gate_par", OFFLOAD_FUNC }, + { "ffn_down", OFFLOAD_FUNC }, + { "ffn_down_b", OFFLOAD_FUNC }, + { "ffn_out", OFFLOAD_FUNC }, + + { "ffn_silu", OFFLOAD_FUNC }, + { "ffn_gelu", OFFLOAD_FUNC }, + { "ffn_relu", OFFLOAD_FUNC }, + { "ffn_sqr(relu)", OFFLOAD_FUNC }, + + { "l_out", OFFLOAD_FUNC }, + + { "result_norm", OFFLOAD_FUNC_EMB }, + { "result_output", OFFLOAD_FUNC_OUT }, +}; + +static llm_offload_trie k_offload_func_trie(k_offload_map); + static struct ggml_cgraph * llama_build_graph( llama_context & lctx, const llama_batch & batch) { const auto & model = lctx.model; + // check if we should build the worst-case graph (for memory measurement) + const bool worst_case = ggml_allocr_is_measure(lctx.alloc); + + // keep track of the input that has already been allocated + bool alloc_inp_tokens = false; + bool alloc_inp_embd = false; + bool alloc_inp_pos = false; + bool alloc_inp_KQ_scale = false; + bool alloc_inp_KQ_mask = false; + bool alloc_inp_K_shift = false; + +#ifdef GGML_USE_CUBLAS + const bool do_offload = true; +#else + const bool do_offload = true; // TODO: set to false after finishing refactoring +#endif + + int n_non_view = 0; // number of non-view tensors that have been processed by the callback + + // this callback allows us to apply custom logic to each tensor (e.g. ggml-alloc, offloading, etc.) + // TODO: will be removed with backend v2 + llm_build_cb cb = [&](struct ggml_tensor * cur, const char * name, int il) { + if (il >= 0) { + ggml_format_name(cur, "%s-%d", name, il); + } else { + ggml_set_name(cur, name); + } + + // + // allocate input tensors and set input data + // + // TODO: will be removed with backend v2 + + if (!alloc_inp_tokens && strcmp(name, "inp_tokens") == 0) { + ggml_allocr_alloc(lctx.alloc, cur); + + if (!ggml_allocr_is_measure(lctx.alloc) && batch.token) { + const int64_t n_tokens = cur->ne[0]; + + memcpy(cur->data, batch.token, n_tokens*ggml_element_size(cur)); + } + + alloc_inp_tokens = true; + } + + if (!alloc_inp_embd && strcmp(name, "inp_embd") == 0) { + ggml_allocr_alloc(lctx.alloc, cur); + + if (!ggml_allocr_is_measure(lctx.alloc) && batch.embd) { + const int64_t n_embd = cur->ne[0]; + const int64_t n_tokens = cur->ne[1]; + + memcpy(cur->data, batch.embd, n_tokens*n_embd*ggml_element_size(cur)); + } + + alloc_inp_embd = true; + } + + if (!alloc_inp_pos && strcmp(name, "inp_pos") == 0) { + ggml_allocr_alloc(lctx.alloc, cur); + + if (!ggml_allocr_is_measure(lctx.alloc) && batch.pos) { + const int64_t n_tokens = cur->ne[0]; + + int32_t * data = (int32_t *) cur->data; + + for (int i = 0; i < n_tokens; ++i) { + data[i] = batch.pos[i]; + } + } + + alloc_inp_pos = true; + } + + if (!alloc_inp_KQ_scale && strcmp(name, "KQ_scale") == 0) { + ggml_allocr_alloc(lctx.alloc, cur); + + if (!ggml_allocr_is_measure(lctx.alloc)) { + const int64_t n_embd_head = model.hparams.n_embd_head(); + ggml_set_f32(cur, 1.0f/sqrtf(float(n_embd_head))); + } + + alloc_inp_KQ_scale = true; + } + + if (!alloc_inp_KQ_mask && strcmp(name, "KQ_mask") == 0) { + ggml_allocr_alloc(lctx.alloc, cur); + + if (!ggml_allocr_is_measure(lctx.alloc)) { + const int64_t n_kv = cur->ne[0]; + const int64_t n_tokens = cur->ne[1]; + + float * data = (float *) cur->data; + memset(data, 0, ggml_nbytes(cur)); + + for (int h = 0; h < 1; ++h) { + for (int j = 0; j < n_tokens; ++j) { + const llama_pos pos = batch.pos[j]; + const llama_seq_id seq_id = batch.seq_id[j][0]; + + for (int i = 0; i < n_kv; ++i) { + if (!lctx.kv_self.cells[i].has_seq_id(seq_id) || lctx.kv_self.cells[i].pos > pos) { + data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY; + } + } + } + } + } + + alloc_inp_KQ_mask = true; + } + + if (!alloc_inp_K_shift && strcmp(name, "K_shift") == 0) { + ggml_allocr_alloc(lctx.alloc, cur); + + if (!ggml_allocr_is_measure(lctx.alloc)) { + const int64_t n_ctx = cur->ne[0]; + + int32_t * data = (int32_t *) cur->data; + + for (int i = 0; i < n_ctx; ++i) { + data[i] = lctx.kv_self.cells[i].delta; + } + } + + alloc_inp_K_shift = true; + } + + // view tensors are not processed further + if (cur->view_src != nullptr) { + return; + } + + if (cur->op != GGML_OP_NONE) { + n_non_view++; + } + + // + // offload layers + // + // TODO: will be removed with backend v2 + +//#define LLAMA_OFFLOAD_DEBUG + + if (!do_offload) { + return; + } + + const int n_layer = model.hparams.n_layer; + + const int n_gpu_layers = model.n_gpu_layers; + const int i_gpu_start = n_layer - n_gpu_layers; + + // should we offload the final norm? yes if we are not computing embeddings + const bool offload_emb = lctx.embedding.empty(); + + static const std::unordered_map> k_offload_func_name = { + { OFFLOAD_FUNC_NOP, "CPU" }, + { OFFLOAD_FUNC_OUT, "CPU" }, +#ifdef GGML_USE_CUBLAS + { OFFLOAD_FUNC, "GPU (CUDA)" }, + { OFFLOAD_FUNC_KQ, "GPU (CUDA) KQ" }, + { OFFLOAD_FUNC_V, "GPU (CUDA) V" }, + { OFFLOAD_FUNC_NR, "GPU (CUDA) NR" }, + { OFFLOAD_FUNC_EMB, "GPU (CUDA) EMB" }, +#else + { OFFLOAD_FUNC, "CPU" }, + { OFFLOAD_FUNC_KQ, "CPU" }, + { OFFLOAD_FUNC_V, "CPU" }, + { OFFLOAD_FUNC_NR, "CPU" }, + { OFFLOAD_FUNC_EMB, "CPU" }, +#endif // GGML_USE_CUBLAS + }; + + // check the global map for what offload function to use for this tensor + llm_offload_func_e func_e = k_offload_func_trie.find(name); + + if (func_e == OFFLOAD_FUNC_NOP) { +#ifdef LLAMA_OFFLOAD_DEBUG + // if a tensor hasn't been offloaded, we warn the user + if (worst_case) { + LLAMA_LOG_WARN("%s: %32s: not offloaded (ref: %s)\n", __func__, + cur->name, "https://github.com/ggerganov/llama.cpp/pull/3837"); + } +#endif + + return; + } + + // count the number of layers and respect the provided n_gpu_layers + switch (func_e) { + case OFFLOAD_FUNC_NOP: + case OFFLOAD_FUNC_OUT: + break; + case OFFLOAD_FUNC: + if (n_gpu_layers < n_layer) { + if (il < i_gpu_start) { + func_e = OFFLOAD_FUNC_NOP; + } + } + break; + case OFFLOAD_FUNC_NR: + if (n_gpu_layers <= n_layer + 0) { + func_e = OFFLOAD_FUNC_NOP; + } + break; + case OFFLOAD_FUNC_V: + if (n_gpu_layers <= n_layer + 1) { + func_e = OFFLOAD_FUNC_NOP; + } + break; + case OFFLOAD_FUNC_KQ: + if (n_gpu_layers <= n_layer + 2) { + func_e = OFFLOAD_FUNC_NOP; + } + break; + case OFFLOAD_FUNC_EMB: + if (!offload_emb || n_gpu_layers < n_layer) { + func_e = OFFLOAD_FUNC_NOP; + } + break; + default: GGML_ASSERT(false); + } + + offload_func_t func = ggml_offload_nop; + + // this is needed for compatibility with Metal for example +#ifdef GGML_USE_CUBLAS + static offload_func_t ggml_offload_gpu = ggml_cuda_assign_buffers_no_alloc; +#else + static offload_func_t ggml_offload_gpu = ggml_offload_nop; +#endif + + switch (func_e) { + case OFFLOAD_FUNC_NOP: + case OFFLOAD_FUNC_OUT: func = ggml_offload_nop; break; + case OFFLOAD_FUNC: + case OFFLOAD_FUNC_KQ: + case OFFLOAD_FUNC_V: + case OFFLOAD_FUNC_NR: + case OFFLOAD_FUNC_EMB: func = ggml_offload_gpu; break; + default: GGML_ASSERT(false); + } + + // apply offload function to the tensor + func(cur); + +#ifdef LLAMA_OFFLOAD_DEBUG + if (worst_case) { + LLAMA_LOG_INFO("%s: %32s: %s\n", __func__, cur->name, k_offload_func_name.at(func_e).c_str()); + } +#endif + }; + struct ggml_cgraph * result = NULL; switch (model.arch) { case LLM_ARCH_LLAMA: { - result = llm_build_llama(lctx, batch); + result = llm_build_llama(lctx, batch, cb, worst_case); } break; case LLM_ARCH_BAICHUAN: { - result = llm_build_baichaun(lctx, batch); + result = llm_build_baichaun(lctx, batch, cb, worst_case); } break; case LLM_ARCH_FALCON: { - result = llm_build_falcon(lctx, batch); + result = llm_build_falcon(lctx, batch, cb, worst_case); } break; case LLM_ARCH_STARCODER: { - result = llm_build_starcoder(lctx, batch); + result = llm_build_starcoder(lctx, batch, cb, worst_case); } break; case LLM_ARCH_PERSIMMON: { - result = llm_build_persimmon(lctx, batch); + result = llm_build_persimmon(lctx, batch, cb, worst_case); } break; case LLM_ARCH_REFACT: { - result = llm_build_refact(lctx, batch); + result = llm_build_refact(lctx, batch, cb, worst_case); } break; case LLM_ARCH_BLOOM: { - result = llm_build_bloom(lctx, batch); + result = llm_build_bloom(lctx, batch, cb, worst_case); } break; case LLM_ARCH_MPT: { - result = llm_build_mpt(lctx, batch); + result = llm_build_mpt(lctx, batch, cb, worst_case); } break; default: GGML_ASSERT(false); } + if (worst_case) { + int n_non_view_total = 0; + + for (int i = 0; i < result->n_nodes; ++i) { + if (result->nodes[i]->view_src == nullptr) { + n_non_view_total++; + } + } + + LLAMA_LOG_INFO("%s: non-view tensors processed: %d/%d\n", __func__, n_non_view, n_non_view_total); + + if (n_non_view != n_non_view_total) { + LLAMA_LOG_WARN("%s: ****************************************************************\n", __func__); + LLAMA_LOG_WARN("%s: not all non-view tensors have been processed with a callback\n", __func__); + LLAMA_LOG_WARN("%s: this can indicate an inefficiency in the graph implementation\n", __func__); + LLAMA_LOG_WARN("%s: build with LLAMA_OFFLOAD_DEBUG for more info\n", __func__); + LLAMA_LOG_WARN("%s: ref: https://github.com/ggerganov/llama.cpp/pull/3837\n", __func__); + LLAMA_LOG_WARN("%s: ****************************************************************\n", __func__); + } + } + return result; } @@ -5938,8 +5303,6 @@ static int llama_decode_internal( } } - ggml_cuda_set_mul_mat_q(cparams.mul_mat_q); - // HACK: ggml-alloc may change the tensor backend when reusing a parent, so force output to be on the CPU here if needed if (!lctx.embedding.empty()) { embeddings->backend = GGML_BACKEND_CPU; @@ -5959,11 +5322,13 @@ static int llama_decode_internal( } // If all tensors can be run on the GPU then using more than 1 thread is detrimental. - const bool full_offload_supported = model.arch == LLM_ARCH_LLAMA || + const bool full_offload_supported = + model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_BAICHUAN || - model.arch == LLM_ARCH_FALCON || - model.arch == LLM_ARCH_REFACT || + model.arch == LLM_ARCH_FALCON || + model.arch == LLM_ARCH_REFACT || model.arch == LLM_ARCH_MPT; + const bool fully_offloaded = model.n_gpu_layers >= (int) hparams.n_layer + 3; if (ggml_cpu_has_cublas() && full_offload_supported && fully_offloaded) { n_threads = 1; @@ -5990,11 +5355,20 @@ static int llama_decode_internal( #endif // update the kv ring buffer - lctx.kv_self.has_shift = false; - lctx.kv_self.head += n_tokens; - // Ensure kv cache head points to a valid index. - if (lctx.kv_self.head >= lctx.kv_self.size) { - lctx.kv_self.head = 0; + { + if (kv_self.has_shift) { + kv_self.has_shift = false; + for (uint32_t i = 0; i < kv_self.size; ++i) { + kv_self.cells[i].delta = 0; + } + } + + kv_self.head += n_tokens; + + // Ensure kv cache head points to a valid index. + if (kv_self.head >= kv_self.size) { + kv_self.head = 0; + } } #ifdef GGML_PERF @@ -6009,6 +5383,8 @@ static int llama_decode_internal( //} // extract logits + // TODO: do not compute and extract logits if only embeddings are needed + // need to update the graphs to skip "result_output" { auto & logits_out = lctx.logits; @@ -6103,11 +5479,10 @@ static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) { } static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch) { + static const char * hex = "0123456789ABCDEF"; switch (llama_vocab_get_type(vocab)) { case LLAMA_VOCAB_TYPE_SPM: { - char buf[7]; - int result = snprintf(buf, sizeof(buf), "<0x%02X>", ch); - GGML_ASSERT(0 <= result && result < 7); + const char buf[7] = { '<', '0', 'x', hex[ch >> 4], hex[ch & 15], '>', 0 }; return vocab.token_to_id.at(buf); } case LLAMA_VOCAB_TYPE_BPE: { @@ -7276,6 +6651,32 @@ void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * can } } +void llama_sample_min_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep) { + if (p <= 0.0f || !candidates->size) { + return; + } + + llama_sample_softmax(ctx, candidates); + + const int64_t t_start_sample_us = ggml_time_us(); + + float scale = candidates->data[0].p; // scale by max prob + size_t i = 1; // first token always matches + + for (; i < candidates->size; ++i) { + if (candidates->data[i].p < p * scale && i >= min_keep) { + break; // prob too small + } + } + + // Resize the output vector to keep only the matching tokens + candidates->size = i; + + if (ctx) { + ctx->t_sample_us += ggml_time_us() - t_start_sample_us; + } +} + void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep) { if (z >= 1.0f || candidates->size <= 2) { return; @@ -7473,14 +6874,14 @@ void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * c } } - const llama_token eos = llama_token_eos(ctx); + const llama_token eos = llama_token_eos(&ctx->model); std::vector, llama_partial_utf8>> candidates_decoded; std::vector candidates_grammar; for (size_t i = 0; i < candidates->size; ++i) { const llama_token id = candidates->data[i].id; - const std::string piece = llama_token_to_str(ctx, id); + const std::string piece = llama_token_to_piece(ctx, id); if (id == eos) { if (!allow_eos) { candidates->data[i].logit = -INFINITY; @@ -7683,7 +7084,7 @@ llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_arra void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token) { const int64_t t_start_sample_us = ggml_time_us(); - if (token == llama_token_eos(ctx)) { + if (token == llama_token_eos(&ctx->model)) { for (const auto & stack : grammar->stacks) { if (stack.empty()) { return; @@ -7692,7 +7093,7 @@ void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar GGML_ASSERT(false); } - const std::string piece = llama_token_to_str(ctx, token); + const std::string piece = llama_token_to_piece(ctx, token); // Note terminating 0 in decoded string const auto decoded = decode_utf8(piece.c_str(), grammar->partial_utf8); @@ -7965,6 +7366,24 @@ struct no_init { no_init() { /* do nothing */ } }; +struct quantize_state_internal { + const llama_model & model; + const llama_model_quantize_params * params; + + int n_attention_wv = 0; + int n_feed_forward_w2 = 0; + int i_attention_wv = 0; + int i_feed_forward_w2 = 0; + + int n_k_quantized = 0; + int n_fallback = 0; + + quantize_state_internal(const llama_model & model, const llama_model_quantize_params * params) + : model(model) + , params(params) + {} +}; + static void llama_convert_tensor_internal( struct ggml_tensor * tensor, std::vector> & output, std::vector & workers, const size_t nelements, const int nthread @@ -8023,14 +7442,14 @@ static void llama_convert_tensor_internal( workers.clear(); } -#ifdef GGML_USE_K_QUANTS static ggml_type get_k_quant_type( - ggml_type new_type, const ggml_tensor * tensor, const llama_model & model, llama_ftype ftype, int * i_attention_wv, - int n_attention_wv, int * i_feed_forward_w2, int n_feed_forward_w2 + quantize_state_internal & qs, + ggml_type new_type, const ggml_tensor * tensor, llama_ftype ftype ) { const std::string name = ggml_get_name(tensor); // TODO: avoid hardcoded tensor names - use the TN_* constants - const auto tn = LLM_TN(model.arch); + const llm_arch arch = qs.model.arch; + const auto tn = LLM_TN(arch); auto use_more_bits = [](int i_layer, int num_layers) -> bool { return i_layer < num_layers/8 || i_layer >= 7*num_layers/8 || (i_layer - num_layers/8)%3 == 2; @@ -8038,7 +7457,7 @@ static ggml_type get_k_quant_type( if (name == tn(LLM_TENSOR_OUTPUT, "weight")) { int nx = tensor->ne[0]; - if (model.arch == LLM_ARCH_FALCON || nx % QK_K != 0) { + if (arch == LLM_ARCH_FALCON || nx % QK_K != 0) { new_type = GGML_TYPE_Q8_0; } else if (new_type != GGML_TYPE_Q8_0) { @@ -8047,46 +7466,46 @@ static ggml_type get_k_quant_type( } else if (name.find("attn_v.weight") != std::string::npos) { if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) { - new_type = *i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K; + new_type = qs.i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K; } else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K; else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) && - use_more_bits(*i_attention_wv, n_attention_wv)) new_type = GGML_TYPE_Q6_K; - else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && *i_attention_wv < 4) new_type = GGML_TYPE_Q5_K; + use_more_bits(qs.i_attention_wv, qs.n_attention_wv)) new_type = GGML_TYPE_Q6_K; + else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K; else if (QK_K == 64 && (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S) && - (*i_attention_wv < n_attention_wv/8 || *i_attention_wv >= 7*n_attention_wv/8)) new_type = GGML_TYPE_Q6_K; - if (model.type == MODEL_70B) { + (qs.i_attention_wv < qs.n_attention_wv/8 || qs.i_attention_wv >= 7*qs.n_attention_wv/8)) new_type = GGML_TYPE_Q6_K; + if (qs.model.type == MODEL_70B) { // In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is // 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with // nearly negligible increase in model size by quantizing this tensor with more bits: if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K) new_type = GGML_TYPE_Q5_K; } - ++*i_attention_wv; + ++qs.i_attention_wv; } else if (name.find("ffn_down.weight") != std::string::npos) { if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) { - new_type = *i_feed_forward_w2 < 2 ? GGML_TYPE_Q5_K - : model.arch != LLM_ARCH_FALCON || use_more_bits(*i_feed_forward_w2, n_feed_forward_w2) ? GGML_TYPE_Q4_K + new_type = qs.i_feed_forward_w2 < 2 ? GGML_TYPE_Q5_K + : arch != LLM_ARCH_FALCON || use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2) ? GGML_TYPE_Q4_K : GGML_TYPE_Q3_K; } else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) { - new_type = model.arch == LLM_ARCH_FALCON ? GGML_TYPE_Q4_K : GGML_TYPE_Q5_K; + new_type = arch == LLM_ARCH_FALCON ? GGML_TYPE_Q4_K : GGML_TYPE_Q5_K; } else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) { - if (model.arch == LLM_ARCH_FALCON) { - new_type = *i_feed_forward_w2 < 2 ? GGML_TYPE_Q6_K : - use_more_bits(*i_feed_forward_w2, n_feed_forward_w2) ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K; + if (arch == LLM_ARCH_FALCON) { + new_type = qs.i_feed_forward_w2 < 2 ? GGML_TYPE_Q6_K : + use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2) ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K; } else { - if (use_more_bits(*i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K; + if (use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K; } } - else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(*i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K; - else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && model.arch != LLM_ARCH_FALCON && *i_feed_forward_w2 < 4) { + else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K; + else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && qs.i_feed_forward_w2 < 4) { new_type = GGML_TYPE_Q5_K; } - ++*i_feed_forward_w2; + ++qs.i_feed_forward_w2; } else if (name.find("attn_output.weight") != std::string::npos) { - if (model.arch != LLM_ARCH_FALCON) { + if (arch != LLM_ARCH_FALCON) { if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K ) new_type = GGML_TYPE_Q3_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) new_type = GGML_TYPE_Q4_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K; @@ -8113,25 +7532,27 @@ static ggml_type get_k_quant_type( int nx = tensor->ne[0]; int ny = tensor->ne[1]; if (nx % QK_K != 0) { - LLAMA_LOG_WARN("\n\n%s : tensor cols %d x %d are not divisible by %d, required for k-quants\n", __func__, nx, ny, QK_K); + LLAMA_LOG_WARN("\n\n%s : tensor cols %d x %d are not divisible by %d, required for %s", __func__, nx, ny, QK_K, ggml_type_name(new_type)); convert_incompatible_tensor = true; + } else { + ++qs.n_k_quantized; } } if (convert_incompatible_tensor) { - if (name == tn(LLM_TENSOR_OUTPUT, "weight")) { - new_type = GGML_TYPE_F16; //fall back to F16 instead of just failing. - LLAMA_LOG_WARN("F16 will be used for this tensor instead.\n"); - } else if (name == tn(LLM_TENSOR_TOKEN_EMBD, "weight")) { - new_type = GGML_TYPE_Q4_0; //fall back to Q4_0 instead of just failing. - LLAMA_LOG_WARN("Q4_0 will be used for this tensor instead.\n"); - } else { - throw std::runtime_error("Unsupported tensor size encountered\n"); + switch (new_type) { + case GGML_TYPE_Q2_K: new_type = GGML_TYPE_Q4_0; break; + case GGML_TYPE_Q3_K: new_type = GGML_TYPE_Q4_1; break; + case GGML_TYPE_Q4_K: new_type = GGML_TYPE_Q5_0; break; + case GGML_TYPE_Q5_K: new_type = GGML_TYPE_Q5_1; break; + case GGML_TYPE_Q6_K: new_type = GGML_TYPE_Q8_0; break; + default: throw std::runtime_error("\nUnsupported tensor size encountered\n"); } + LLAMA_LOG_WARN(" - using fallback quantization %s\n", ggml_type_name(new_type)); + ++qs.n_fallback; } return new_type; } -#endif static void llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, const llama_model_quantize_params * params) { ggml_type quantized_type; @@ -8146,7 +7567,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_MOSTLY_F16: quantized_type = GGML_TYPE_F16; break; case LLAMA_FTYPE_ALL_F32: quantized_type = GGML_TYPE_F32; break; -#ifdef GGML_USE_K_QUANTS // K-quants case LLAMA_FTYPE_MOSTLY_Q2_K: quantized_type = GGML_TYPE_Q2_K; break; case LLAMA_FTYPE_MOSTLY_Q3_K_S: @@ -8157,7 +7577,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_MOSTLY_Q5_K_S: case LLAMA_FTYPE_MOSTLY_Q5_K_M: quantized_type = GGML_TYPE_Q5_K; break; case LLAMA_FTYPE_MOSTLY_Q6_K: quantized_type = GGML_TYPE_Q6_K; break; -#endif + default: throw std::runtime_error(format("invalid output file type %d\n", ftype)); } @@ -8184,6 +7604,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s llm_load_arch(ml, model); llm_load_hparams(ml, model); + struct quantize_state_internal qs(model, params); + if (params->only_copy) { ftype = model.ftype; } @@ -8196,10 +7618,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s gguf_set_val_u32(ctx_out, "general.quantization_version", GGML_QNT_VERSION); gguf_set_val_u32(ctx_out, "general.file_type", ftype); -#ifdef GGML_USE_K_QUANTS - int n_attention_wv = 0; - int n_feed_forward_w2 = 0; - for (int i = 0; i < ml.n_tensors; ++i) { struct ggml_tensor * meta = ml.get_tensor_meta(i); @@ -8207,21 +7625,17 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s // TODO: avoid hardcoded tensor names - use the TN_* constants if (name.find("attn_v.weight") != std::string::npos || name.find("attn_qkv.weight") != std::string::npos) { - ++n_attention_wv; + ++qs.n_attention_wv; } else if (name.find("ffn_down.weight") != std::string::npos) { - ++n_feed_forward_w2; + ++qs.n_feed_forward_w2; } } - if (n_attention_wv != n_feed_forward_w2 || (uint32_t)n_attention_wv != model.hparams.n_layer) { + if (qs.n_attention_wv != qs.n_feed_forward_w2 || (uint32_t)qs.n_attention_wv != model.hparams.n_layer) { LLAMA_LOG_WARN("%s ============ Strange model: n_attention_wv = %d, n_feed_forward_w2 = %d, hparams.n_layer = %d\n", - __func__, n_attention_wv, n_feed_forward_w2, model.hparams.n_layer); + __func__, qs.n_attention_wv, qs.n_feed_forward_w2, model.hparams.n_layer); } - int i_attention_wv = 0; - int i_feed_forward_w2 = 0; -#endif - size_t total_size_org = 0; size_t total_size_new = 0; std::vector hist_all(1 << 4, 0); @@ -8285,11 +7699,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s if (quantize) { new_type = quantized_type; -#ifdef GGML_USE_K_QUANTS - new_type = get_k_quant_type( - new_type, tensor, model, ftype, &i_attention_wv, n_attention_wv, &i_feed_forward_w2, n_feed_forward_w2 - ); -#endif + if (!params->pure) { + new_type = get_k_quant_type(qs, new_type, tensor, ftype); + } + // If we've decided to quantize to the same type the tensor is already // in then there's nothing to do. quantize = tensor->type != new_type; @@ -8414,6 +7827,11 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s LLAMA_LOG_INFO("\n"); } } + + if (qs.n_fallback > 0) { + LLAMA_LOG_WARN("%s: WARNING: %d of %d tensor(s) incompatible with k-quants and required fallback quantization\n", + __func__, qs.n_fallback, qs.n_k_quantized + qs.n_fallback); + } } static int llama_apply_lora_from_file_internal( @@ -8578,14 +7996,14 @@ static int llama_apply_lora_from_file_internal( ggml_tensor * dest_t = model_tensors[base_name]; - offload_func_t offload_func = llama_nop; - offload_func_t offload_func_force_inplace = llama_nop; + offload_func_t offload_func = ggml_offload_nop; + offload_func_t offload_func_force_inplace = ggml_offload_nop; #ifdef GGML_USE_CUBLAS if (dest_t->backend == GGML_BACKEND_GPU || dest_t->backend == GGML_BACKEND_GPU_SPLIT) { if (dest_t->type != GGML_TYPE_F16) { throw std::runtime_error(format( - "%s: error: the simultaneous use of LoRAs and GPU acceleration is only supported for f16 models", __func__)); + "%s: error: the simultaneous use of LoRAs and GPU acceleration is only supported for f16 models. dest_t->type: %d", __func__, dest_t->type)); } offload_func = ggml_cuda_assign_buffers; offload_func_force_inplace = ggml_cuda_assign_buffers_force_inplace; @@ -8738,6 +8156,7 @@ struct llama_model_quantize_params llama_model_quantize_default_params() { /*.allow_requantize =*/ false, /*.quantize_output_tensor =*/ true, /*.only_copy =*/ false, + /*.pure =*/ false, }; return result; @@ -8892,7 +8311,7 @@ struct llama_context * llama_new_context_with_model( // build worst-case graph int n_tokens = (int)std::min(cparams.n_ctx, cparams.n_batch); int n_past = cparams.n_ctx - n_tokens; - llama_token token = llama_token_bos(ctx); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph + llama_token token = llama_token_bos(&ctx->model); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph ggml_cgraph * gf = llama_build_graph(*ctx, llama_batch_get_one(&token, n_tokens, n_past, 0)); #ifdef GGML_USE_METAL @@ -9098,8 +8517,8 @@ int llama_get_kv_cache_token_count(const struct llama_context * ctx) { return ctx->kv_self.head; } -void llama_kv_cache_tokens_rm(struct llama_context * ctx, int32_t c0, int32_t c1) { - llama_kv_cache_tokens_rm(ctx->kv_self, c0, c1); +void llama_kv_cache_clear(struct llama_context * ctx) { + llama_kv_cache_clear(ctx->kv_self); } void llama_kv_cache_seq_rm(struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1) { @@ -9545,7 +8964,7 @@ int llama_eval( llama_token * tokens, int32_t n_tokens, int n_past) { - llama_kv_cache_tokens_rm(ctx->kv_self, n_past, -1); + llama_kv_cache_seq_rm(ctx->kv_self, -1, n_past, -1); const int ret = llama_decode_internal(*ctx, llama_batch_get_one(tokens, n_tokens, n_past, 0)); if (ret < 0) { @@ -9560,7 +8979,7 @@ int llama_eval_embd( float * embd, int32_t n_tokens, int n_past) { - llama_kv_cache_tokens_rm(ctx->kv_self, n_past, -1); + llama_kv_cache_seq_rm(ctx->kv_self, -1, n_past, -1); llama_batch batch = { n_tokens, nullptr, embd, nullptr, nullptr, nullptr, nullptr, n_past, 1, 0, }; @@ -9653,43 +9072,44 @@ float * llama_get_embeddings(struct llama_context * ctx) { return ctx->embedding.data(); } -const char * llama_token_get_text(const struct llama_context * ctx, llama_token token) { - return ctx->model.vocab.id_to_token[token].text.c_str(); +const char * llama_token_get_text(const struct llama_model * model, llama_token token) { + return model->vocab.id_to_token[token].text.c_str(); } -float llama_token_get_score(const struct llama_context * ctx, llama_token token) { - return ctx->model.vocab.id_to_token[token].score; +float llama_token_get_score(const struct llama_model * model, llama_token token) { + return model->vocab.id_to_token[token].score; } -llama_token_type llama_token_get_type(const struct llama_context * ctx, llama_token token) { - return ctx->model.vocab.id_to_token[token].type; +llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token) { + return model->vocab.id_to_token[token].type; } -llama_token llama_token_bos(const struct llama_context * ctx) { - return ctx->model.vocab.special_bos_id; +llama_token llama_token_bos(const struct llama_model * model) { + return model->vocab.special_bos_id; } -llama_token llama_token_eos(const struct llama_context * ctx) { - return ctx->model.vocab.special_eos_id; +llama_token llama_token_eos(const struct llama_model * model) { + return model->vocab.special_eos_id; } -llama_token llama_token_nl(const struct llama_context * ctx) { - return ctx->model.vocab.linefeed_id; -} -llama_token llama_token_prefix(const struct llama_context * ctx) { - return ctx->model.vocab.special_prefix_id; +llama_token llama_token_nl(const struct llama_model * model) { + return model->vocab.linefeed_id; } -llama_token llama_token_middle(const struct llama_context * ctx) { - return ctx->model.vocab.special_middle_id; +llama_token llama_token_prefix(const struct llama_model * model) { + return model->vocab.special_prefix_id; } -llama_token llama_token_suffix(const struct llama_context * ctx) { - return ctx->model.vocab.special_suffix_id; +llama_token llama_token_middle(const struct llama_model * model) { + return model->vocab.special_middle_id; } -llama_token llama_token_eot(const struct llama_context * ctx) { - return ctx->model.vocab.special_eot_id; +llama_token llama_token_suffix(const struct llama_model * model) { + return model->vocab.special_suffix_id; +} + +llama_token llama_token_eot(const struct llama_model * model) { + return model->vocab.special_eot_id; } int llama_tokenize( diff --git a/llama.h b/llama.h index 306f5b383..75fe391ef 100644 --- a/llama.h +++ b/llama.h @@ -178,7 +178,7 @@ extern "C" { float rope_freq_scale; // RoPE frequency scaling factor, 0 = from model // Keep the booleans together to avoid misalignment during copy-by-value. - bool mul_mat_q; // if true, use experimental mul_mat_q kernels + bool mul_mat_q; // if true, use experimental mul_mat_q kernels (DEPRECATED - always true) bool f16_kv; // use fp16 for KV cache, fp32 otherwise bool logits_all; // the llama_eval() call computes all logits, not just the last one bool embedding; // embedding mode only @@ -191,6 +191,7 @@ extern "C" { bool allow_requantize; // allow quantizing non-f32/f16 tensors bool quantize_output_tensor; // quantize output.weight bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored + bool pure; // disable k-quant mixtures and quantize all tensors to the same type } llama_model_quantize_params; // grammar types @@ -333,17 +334,14 @@ extern "C" { LLAMA_API DEPRECATED(int llama_get_kv_cache_token_count(const struct llama_context * ctx), "avoid using this, it will be removed in the future, instead - count the tokens in user code"); - // Remove all tokens data of cells in [c0, c1) - // c0 < 0 : [0, c1] - // c1 < 0 : [c0, inf) - LLAMA_API void llama_kv_cache_tokens_rm( - struct llama_context * ctx, - int32_t c0, - int32_t c1); + // Clear the KV cache + LLAMA_API void llama_kv_cache_clear( + struct llama_context * ctx); // Removes all tokens that belong to the specified sequence and have positions in [p0, p1) - // p0 < 0 : [0, p1] - // p1 < 0 : [p0, inf) + // seq_id < 0 : match any sequence + // p0 < 0 : [0, p1] + // p1 < 0 : [p0, inf) LLAMA_API void llama_kv_cache_seq_rm( struct llama_context * ctx, llama_seq_id seq_id, @@ -494,21 +492,22 @@ extern "C" { // Vocab // - LLAMA_API const char * llama_token_get_text(const struct llama_context * ctx, llama_token token); + LLAMA_API const char * llama_token_get_text(const struct llama_model * model, llama_token token); - LLAMA_API float llama_token_get_score(const struct llama_context * ctx, llama_token token); + LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token); - LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_context * ctx, llama_token token); + LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token); // Special tokens - LLAMA_API llama_token llama_token_bos(const struct llama_context * ctx); // beginning-of-sentence - LLAMA_API llama_token llama_token_eos(const struct llama_context * ctx); // end-of-sentence - LLAMA_API llama_token llama_token_nl (const struct llama_context * ctx); // next-line + LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence + LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence + LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line + // codellama infill tokens - LLAMA_API llama_token llama_token_prefix(const struct llama_context * ctx); // Beginning of infill prefix - LLAMA_API llama_token llama_token_middle(const struct llama_context * ctx); // Beginning of infill middle - LLAMA_API llama_token llama_token_suffix(const struct llama_context * ctx); // Beginning of infill suffix - LLAMA_API llama_token llama_token_eot (const struct llama_context * ctx); // End of infill middle + LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix + LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle + LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix + LLAMA_API llama_token llama_token_eot (const struct llama_model * model); // End of infill middle // // Tokenization @@ -599,6 +598,13 @@ extern "C" { float p, size_t min_keep); + /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841 + LLAMA_API void llama_sample_min_p( + struct llama_context * ctx, + llama_token_data_array * candidates, + float p, + size_t min_keep); + /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/. LLAMA_API void llama_sample_tail_free( struct llama_context * ctx, @@ -657,6 +663,7 @@ extern "C" { float * mu); /// @details Selects the token with the highest probability. + /// Does not compute the token probabilities. Use llama_sample_softmax() instead. LLAMA_API llama_token llama_sample_token_greedy( struct llama_context * ctx, llama_token_data_array * candidates); diff --git a/models/ggml-vocab-baichuan.gguf b/models/ggml-vocab-baichuan.gguf new file mode 100644 index 000000000..7caaf8239 Binary files /dev/null and b/models/ggml-vocab-baichuan.gguf differ diff --git a/models/ggml-vocab-gpt-neox.gguf b/models/ggml-vocab-gpt-neox.gguf new file mode 100644 index 000000000..b9af16845 Binary files /dev/null and b/models/ggml-vocab-gpt-neox.gguf differ diff --git a/models/ggml-vocab-mpt.gguf b/models/ggml-vocab-mpt.gguf new file mode 100644 index 000000000..6affa34bd Binary files /dev/null and b/models/ggml-vocab-mpt.gguf differ diff --git a/models/ggml-vocab-refact.gguf b/models/ggml-vocab-refact.gguf new file mode 100644 index 000000000..8f26cfb76 Binary files /dev/null and b/models/ggml-vocab-refact.gguf differ diff --git a/models/ggml-vocab-starcoder.gguf b/models/ggml-vocab-starcoder.gguf new file mode 100644 index 000000000..a52983fdb Binary files /dev/null and b/models/ggml-vocab-starcoder.gguf differ diff --git a/scripts/server-llm.sh b/scripts/server-llm.sh new file mode 100644 index 000000000..7bf0929bb --- /dev/null +++ b/scripts/server-llm.sh @@ -0,0 +1,391 @@ +#!/bin/bash +# +# Helper script for deploying llama.cpp server with a single Bash command +# +# - Works on Linux and macOS +# - Supports: CPU, CUDA, Metal, OpenCL +# - Can run all GGUF models from HuggingFace +# - Can serve requests in parallel +# - Always builds latest llama.cpp from GitHub +# +# Limitations +# +# - Chat templates are poorly supported (base models recommended) +# - Might be unstable! +# +# Usage: +# ./server-llm.sh [--port] [--repo] [--wtype] [--backend] [--gpu-id] [--n-parallel] [--n-kv] [--verbose] +# +# --port: port number, default is 8888 +# --repo: path to a repo containing GGUF model files +# --wtype: weights type (f16, q8_0, q4_0, q4_1), default is user-input +# --backend: cpu, cuda, metal, opencl, depends on the OS +# --gpu-id: gpu id, default is 0 +# --n-parallel: number of parallel requests, default is 8 +# --n-kv: KV cache size, default is 4096 +# --verbose: verbose output +# +# Example: +# +# bash -c "$(curl -s https://ggml.ai/server-llm.sh)" +# + +set -e + +# required utils: curl, git, make +if ! command -v curl &> /dev/null; then + printf "[-] curl not found\n" + exit 1 +fi +if ! command -v git &> /dev/null; then + printf "[-] git not found\n" + exit 1 +fi +if ! command -v make &> /dev/null; then + printf "[-] make not found\n" + exit 1 +fi + +# parse arguments +port=8888 +repo="" +wtype="" +backend="cpu" + +# if macOS, use metal backend by default +if [[ "$OSTYPE" == "darwin"* ]]; then + backend="metal" +elif command -v nvcc &> /dev/null; then + backend="cuda" +fi + +gpu_id=0 +n_parallel=8 +n_kv=4096 +verbose=0 + +function print_usage { + printf "Usage:\n" + printf " ./server-llm.sh [--port] [--repo] [--wtype] [--backend] [--gpu-id] [--n-parallel] [--n-kv] [--verbose]\n\n" + printf " --port: port number, default is 8888\n" + printf " --repo: path to a repo containing GGUF model files\n" + printf " --wtype: weights type (f16, q8_0, q4_0, q4_1), default is user-input\n" + printf " --backend: cpu, cuda, metal, opencl, depends on the OS\n" + printf " --gpu-id: gpu id, default is 0\n" + printf " --n-parallel: number of parallel requests, default is 8\n" + printf " --n-kv: KV cache size, default is 4096\n" + printf " --verbose: verbose output\n\n" + printf "Example:\n\n" + printf ' bash -c "$(curl -s https://ggml.ai/server-llm.sh)"\n\n' +} + +while [[ $# -gt 0 ]]; do + key="$1" + case $key in + --port) + port="$2" + shift + shift + ;; + --repo) + repo="$2" + shift + shift + ;; + --wtype) + wtype="$2" + shift + shift + ;; + --backend) + backend="$2" + shift + shift + ;; + --gpu-id) + gpu_id="$2" + shift + shift + ;; + --n-parallel) + n_parallel="$2" + shift + shift + ;; + --n-kv) + n_kv="$2" + shift + shift + ;; + --verbose) + verbose=1 + shift + ;; + --help) + print_usage + exit 0 + ;; + *) + echo "Unknown argument: $key" + print_usage + exit 1 + ;; + esac +done + +# available weights types +wtypes=("F16" "Q8_0" "Q4_0" "Q4_1" "Q5_0" "Q5_1" "Q6_K" "Q5_K_M" "Q5_K_S" "Q4_K_M" "Q4_K_S" "Q3_K_L" "Q3_K_M" "Q3_K_S" "Q2_K") + +wfiles=() +for wt in "${wtypes[@]}"; do + wfiles+=("") +done + +# sample repos +repos=( + "https://huggingface.co/TheBloke/Llama-2-7B-GGUF" + "https://huggingface.co/TheBloke/Llama-2-13B-GGUF" + "https://huggingface.co/TheBloke/Llama-2-70B-GGUF" + "https://huggingface.co/TheBloke/CodeLlama-7B-GGUF" + "https://huggingface.co/TheBloke/CodeLlama-13B-GGUF" + "https://huggingface.co/TheBloke/CodeLlama-34B-GGUF" + "https://huggingface.co/TheBloke/Mistral-7B-v0.1-GGUF" + "https://huggingface.co/TheBloke/zephyr-7B-beta-GGUF" + "https://huggingface.co/TheBloke/OpenHermes-2-Mistral-7B-GGUF" + "https://huggingface.co/TheBloke/CausalLM-7B-GGUF" +) + +printf "\n" +printf "[I] This is a helper script for deploying llama.cpp's server on this machine.\n\n" +printf " Based on the options that follow, the script might download a model file\n" +printf " from the internet, which can be a few GBs in size. The script will also\n" +printf " build the latest llama.cpp source code from GitHub, which can be unstable.\n" +printf "\n" +printf " Upon success, an HTTP server will be started and it will serve the selected\n" +printf " model using llama.cpp for demonstration purposes.\n" +printf "\n" +printf " Please note:\n" +printf "\n" +printf " - All new data will be stored in the current folder\n" +printf " - The server will be listening on all network interfaces\n" +printf " - The server will run with default settings which are not always optimal\n" +printf " - Do not judge the quality of a model based on the results from this script\n" +printf " - Do not use this script to benchmark llama.cpp\n" +printf " - Do not use this script in production\n" +printf " - This script is only for demonstration purposes\n" +printf "\n" +printf " If you don't know what you are doing, please press Ctrl-C to abort now\n" +printf "\n" +printf " Press Enter to continue ...\n\n" + +read + +if [[ -z "$repo" ]]; then + printf "[+] No repo provided from the command line\n" + printf " Please select a number from the list below or enter an URL:\n\n" + + is=0 + for r in "${repos[@]}"; do + printf " %2d) %s\n" $is "$r" + is=$((is+1)) + done + + # ask for repo until index of sample repo is provided or an URL + while [[ -z "$repo" ]]; do + printf "\n Or choose one from: https://huggingface.co/models?sort=trending&search=gguf\n\n" + read -p "[+] Select repo: " repo + + # check if the input is a number + if [[ "$repo" =~ ^[0-9]+$ ]]; then + if [[ "$repo" -ge 0 && "$repo" -lt ${#repos[@]} ]]; then + repo="${repos[$repo]}" + else + printf "[-] Invalid repo index: %s\n" "$repo" + repo="" + fi + elif [[ "$repo" =~ ^https?:// ]]; then + repo="$repo" + else + printf "[-] Invalid repo URL: %s\n" "$repo" + repo="" + fi + done +fi + +# remove suffix +repo=$(echo "$repo" | sed -E 's/\/tree\/main$//g') + +printf "[+] Checking for GGUF model files in %s\n" "$repo" + +# find GGUF files in the source +# TODO: better logic +model_tree="${repo%/}/tree/main" +model_files=$(curl -s "$model_tree" | grep -i "\\.gguf" | sed -E 's/.*(.*)<\/span><\/a>/\1/g') + +# list all files in the provided git repo +printf "[+] Model files:\n\n" +for file in $model_files; do + # determine iw by grepping the filename with wtypes + iw=-1 + is=0 + for wt in "${wtypes[@]}"; do + # uppercase + ufile=$(echo "$file" | tr '[:lower:]' '[:upper:]') + if [[ "$ufile" =~ "$wt" ]]; then + iw=$is + break + fi + is=$((is+1)) + done + + if [[ $iw -eq -1 ]]; then + continue + fi + + wfiles[$iw]="$file" + + have=" " + if [[ -f "$file" ]]; then + have="*" + fi + + printf " %2d) %s %s\n" $iw "$have" "$file" +done + +# ask for weights type until provided and available +while [[ -z "$wtype" ]]; do + printf "\n" + read -p "[+] Select weight type: " wtype + wfile="${wfiles[$wtype]}" + + if [[ -z "$wfile" ]]; then + printf "[-] Invalid weight type: %s\n" "$wtype" + wtype="" + fi +done + +printf "[+] Selected weight type: %s (%s)\n" "$wtype" "$wfile" + +url="${repo%/}/resolve/main/$wfile" + +# check file if the model has been downloaded before +chk="$wfile.chk" + +# check if we should download the file +# - if $wfile does not exist +# - if $wfile exists but $chk does not exist +# - if $wfile exists and $chk exists but $wfile is newer than $chk +# TODO: better logic using git lfs info + +do_download=0 + +if [[ ! -f "$wfile" ]]; then + do_download=1 +elif [[ ! -f "$chk" ]]; then + do_download=1 +elif [[ "$wfile" -nt "$chk" ]]; then + do_download=1 +fi + +if [[ $do_download -eq 1 ]]; then + printf "[+] Downloading weights from %s\n" "$url" + + # download the weights file + curl -o "$wfile" -# -L "$url" + + # create a check file if successful + if [[ $? -eq 0 ]]; then + printf "[+] Creating check file %s\n" "$chk" + touch "$chk" + fi +else + printf "[+] Using cached weights %s\n" "$wfile" +fi + +# get latest llama.cpp and build + +printf "[+] Downloading latest llama.cpp\n" + +llama_cpp_dir="__llama_cpp_port_${port}__" + +if [[ -d "$llama_cpp_dir" && ! -f "$llama_cpp_dir/__ggml_script__" ]]; then + # if the dir exists and there isn't a file "__ggml_script__" in it, abort + printf "[-] Directory %s already exists\n" "$llama_cpp_dir" + printf "[-] Please remove it and try again\n" + exit 1 +elif [[ -d "$llama_cpp_dir" ]]; then + printf "[+] Directory %s already exists\n" "$llama_cpp_dir" + printf "[+] Using cached llama.cpp\n" + + cd "$llama_cpp_dir" + git reset --hard + git fetch + git checkout origin/master + + cd .. +else + printf "[+] Cloning llama.cpp\n" + + git clone https://github.com/ggerganov/llama.cpp "$llama_cpp_dir" +fi + +# mark that that the directory is made by this script +touch "$llama_cpp_dir/__ggml_script__" + +if [[ $verbose -eq 1 ]]; then + set -x +fi + +# build +cd "$llama_cpp_dir" + +make clean + +log="--silent" +if [[ $verbose -eq 1 ]]; then + log="" +fi + +if [[ "$backend" == "cuda" ]]; then + printf "[+] Building with CUDA backend\n" + LLAMA_CUBLAS=1 make -j server $log +elif [[ "$backend" == "cpu" ]]; then + printf "[+] Building with CPU backend\n" + make -j server $log +elif [[ "$backend" == "metal" ]]; then + printf "[+] Building with Metal backend\n" + make -j server $log +elif [[ "$backend" == "opencl" ]]; then + printf "[+] Building with OpenCL backend\n" + LLAMA_CLBLAST=1 make -j server $log +else + printf "[-] Unknown backend: %s\n" "$backend" + exit 1 +fi + +# run the server + +printf "[+] Running server\n" + +args="" +if [[ "$backend" == "cuda" ]]; then + export CUDA_VISIBLE_DEVICES=$gpu_id + args="-ngl 999" +elif [[ "$backend" == "cpu" ]]; then + args="-ngl 0" +elif [[ "$backend" == "metal" ]]; then + args="-ngl 999" +elif [[ "$backend" == "opencl" ]]; then + args="-ngl 999" +else + printf "[-] Unknown backend: %s\n" "$backend" + exit 1 +fi + +if [[ $verbose -eq 1 ]]; then + args="$args --verbose" +fi + +./server -m "../$wfile" --host 0.0.0.0 --port "$port" -c $n_kv -np "$n_parallel" $args + +exit 0 diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt index 61407e573..6757ad1cc 100644 --- a/tests/CMakeLists.txt +++ b/tests/CMakeLists.txt @@ -28,9 +28,14 @@ llama_build_executable(test-tokenizer-0-falcon.cpp) llama_test_executable (test-tokenizer-0-falcon test-tokenizer-0-falcon.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf) llama_build_executable(test-tokenizer-1-llama.cpp) llama_test_executable (test-tokenizer-1-llama test-tokenizer-1-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf) +llama_test_executable(test-tokenizer-1-baichuan test-tokenizer-1-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-baichuan.gguf) llama_build_executable(test-tokenizer-1-bpe.cpp) llama_test_executable (test-tokenizer-1-falcon test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf) llama_test_executable(test-tokenizer-1-aquila test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf) +llama_test_executable(test-tokenizer-1-mpt test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-mpt.gguf) +llama_test_executable(test-tokenizer-1-gpt-neox test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-neox.gguf) +llama_test_executable(test-tokenizer-1-refact test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf) +llama_test_executable(test-tokenizer-1-starcoder test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf) llama_build_and_test_executable(test-grammar-parser.cpp) llama_build_and_test_executable(test-llama-grammar.cpp) llama_build_and_test_executable(test-grad0.cpp) # SLOW diff --git a/tests/test-double-float.cpp b/tests/test-double-float.cpp index afd7bf77f..753dae911 100644 --- a/tests/test-double-float.cpp +++ b/tests/test-double-float.cpp @@ -4,7 +4,7 @@ #undef NDEBUG #include -#if !defined(__riscv) && !defined(__s390__) +#if !defined(__riscv) && !defined(__s390__) && !defined(__ARM_NEON) #include #endif #include diff --git a/tests/test-quantize-fns.cpp b/tests/test-quantize-fns.cpp index 884af4054..a2459a286 100644 --- a/tests/test-quantize-fns.cpp +++ b/tests/test-quantize-fns.cpp @@ -129,6 +129,13 @@ int main(int argc, char * argv[]) { ggml_type type = (ggml_type) i; ggml_type_traits_t qfns = ggml_internal_get_type_traits(type); + // deprecated - skip + if (qfns.blck_size == 0) { + continue; + } + + printf("Testing %s\n", ggml_type_name((ggml_type) i)); + if (qfns.from_float && qfns.to_float) { const float total_error = total_quantization_error(qfns, test_size, test_data.data()); const float max_quantization_error = diff --git a/tests/test-tokenizer-1-bpe.cpp b/tests/test-tokenizer-1-bpe.cpp index 85a59a14d..386530f23 100644 --- a/tests/test-tokenizer-1-bpe.cpp +++ b/tests/test-tokenizer-1-bpe.cpp @@ -91,9 +91,19 @@ int main(int argc, char **argv) { } } } - // TODO: why doesn't this work for the full range of Unicodes? + // Restrict to assigned unicode planes // for (uint32_t cp = 0x10000; cp < 0x0010ffff; ++cp) { - for (uint32_t cp = 0x10000; cp < 0x00080000; ++cp) { + for (uint32_t cp = 0x10000; cp < 0x00040000; ++cp) { + std::string str = codepoint_to_utf8(cp); + std::vector tokens = llama_tokenize(ctx, str, false); + std::string check = llama_detokenize_bpe(ctx, tokens); + if (str != check) { + fprintf(stderr, "%s : error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n", + __func__, cp, check.c_str(), check.length(), str.c_str(), str.length()); + return 4; + } + } + for (uint32_t cp = 0x000e0000; cp < 0x0010ffff; ++cp) { std::string str = codepoint_to_utf8(cp); std::vector tokens = llama_tokenize(ctx, str, false); std::string check = llama_detokenize_bpe(ctx, tokens); @@ -103,7 +113,6 @@ int main(int argc, char **argv) { return 4; } } - llama_free_model(model); llama_free(ctx);