diff --git a/.github/ISSUE_TEMPLATE/bug.md b/.github/ISSUE_TEMPLATE/bug.md index c003fe7c1..ce69e6395 100644 --- a/.github/ISSUE_TEMPLATE/bug.md +++ b/.github/ISSUE_TEMPLATE/bug.md @@ -6,179 +6,4 @@ 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. - -# Expected Behavior - -Please provide a detailed written description of what you were trying to do, and what you expected `llama.cpp` to do. - -# Current Behavior - -Please provide a detailed written description of what `llama.cpp` did, instead. - -# Environment and Context - -Please provide detailed information about your computer setup. This is important in case the issue is not reproducible except for under certain specific conditions. - -* Physical (or virtual) hardware you are using, e.g. for Linux: - -`$ lscpu` - -* Operating System, e.g. for Linux: - -`$ uname -a` - -* SDK version, e.g. for Linux: - -``` -$ python3 --version -$ make --version -$ g++ --version -``` - -# Failure Information (for bugs) - -Please help provide information about the failure / bug. - -# Steps to Reproduce - -Please provide detailed steps for reproducing the issue. We are not sitting in front of your screen, so the more detail the better. - -1. step 1 -2. step 2 -3. step 3 -4. etc. - -# Failure Logs - -Please include any relevant log snippets or files. If it works under one configuration but not under another, please provide logs for both configurations and their corresponding outputs so it is easy to see where behavior changes. - -Also, please try to **avoid using screenshots** if at all possible. Instead, copy/paste the console output and use [Github's markdown](https://docs.github.com/en/get-started/writing-on-github/getting-started-with-writing-and-formatting-on-github/basic-writing-and-formatting-syntax) to cleanly format your logs for easy readability. - -Example environment info: -``` -llama.cpp$ git log | head -1 -commit 2af23d30434a677c6416812eea52ccc0af65119c - -llama.cpp$ lscpu | egrep "AMD|Flags" -Vendor ID: AuthenticAMD -Model name: AMD Ryzen Threadripper 1950X 16-Core Processor -Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid amd_dcm aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb hw_pstate ssbd ibpb vmmcall fsgsbase bmi1 avx2 smep bmi2 rdseed adx smap clflushopt sha_ni xsaveopt xsavec xgetbv1 xsaves clzero irperf xsaveerptr arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif overflow_recov succor smca sme sev -Virtualization: AMD-V - -llama.cpp$ python3 --version -Python 3.10.9 - -llama.cpp$ pip list | egrep "torch|numpy|sentencepiece" -numpy 1.24.2 -numpydoc 1.5.0 -sentencepiece 0.1.97 -torch 1.13.1 -torchvision 0.14.1 - -llama.cpp$ make --version | head -1 -GNU Make 4.3 - -$ md5sum ./models/65B/ggml-model-q4_0.bin -dbdd682cce80e2d6e93cefc7449df487 ./models/65B/ggml-model-q4_0.bin -``` - -Example run with the Linux command [perf](https://www.brendangregg.com/perf.html) -``` -llama.cpp$ perf stat ./main -m ./models/65B/ggml-model-q4_0.bin -t 16 -n 1024 -p "Please close your issue when it has been answered." -main: seed = 1679149377 -llama_model_load: loading model from './models/65B/ggml-model-q4_0.bin' - please wait ... -llama_model_load: n_vocab = 32000 -llama_model_load: n_ctx = 512 -llama_model_load: n_embd = 8192 -llama_model_load: n_mult = 256 -llama_model_load: n_head = 64 -llama_model_load: n_layer = 80 -llama_model_load: n_rot = 128 -llama_model_load: f16 = 2 -llama_model_load: n_ff = 22016 -llama_model_load: n_parts = 8 -llama_model_load: ggml ctx size = 41477.73 MB -llama_model_load: memory_size = 2560.00 MB, n_mem = 40960 -llama_model_load: loading model part 1/8 from './models/65B/ggml-model-q4_0.bin' -llama_model_load: .......................................................................................... done -llama_model_load: model size = 4869.09 MB / num tensors = 723 -llama_model_load: loading model part 2/8 from './models/65B/ggml-model-q4_0.bin.1' -llama_model_load: .......................................................................................... done -llama_model_load: model size = 4869.09 MB / num tensors = 723 -llama_model_load: loading model part 3/8 from './models/65B/ggml-model-q4_0.bin.2' -llama_model_load: .......................................................................................... done -llama_model_load: model size = 4869.09 MB / num tensors = 723 -llama_model_load: loading model part 4/8 from './models/65B/ggml-model-q4_0.bin.3' -llama_model_load: .......................................................................................... done -llama_model_load: model size = 4869.09 MB / num tensors = 723 -llama_model_load: loading model part 5/8 from './models/65B/ggml-model-q4_0.bin.4' -llama_model_load: .......................................................................................... done -llama_model_load: model size = 4869.09 MB / num tensors = 723 -llama_model_load: loading model part 6/8 from './models/65B/ggml-model-q4_0.bin.5' -llama_model_load: .......................................................................................... done -llama_model_load: model size = 4869.09 MB / num tensors = 723 -llama_model_load: loading model part 7/8 from './models/65B/ggml-model-q4_0.bin.6' -llama_model_load: .......................................................................................... done -llama_model_load: model size = 4869.09 MB / num tensors = 723 -llama_model_load: loading model part 8/8 from './models/65B/ggml-model-q4_0.bin.7' -llama_model_load: .......................................................................................... done -llama_model_load: model size = 4869.09 MB / num tensors = 723 - -system_info: n_threads = 16 / 32 | AVX = 1 | AVX2 = 1 | AVX512 = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | VSX = 0 | - -main: prompt: 'Please close your issue when it has been answered.' -main: number of tokens in prompt = 11 - 1 -> '' - 12148 -> 'Please' - 3802 -> ' close' - 596 -> ' your' - 2228 -> ' issue' - 746 -> ' when' - 372 -> ' it' - 756 -> ' has' - 1063 -> ' been' - 7699 -> ' answered' - 29889 -> '.' - -sampling parameters: temp = 0.800000, top_k = 40, top_p = 0.950000, repeat_last_n = 64, repeat_penalty = 1.300000 - - -Please close your issue when it has been answered. -@duncan-donut: I'm trying to figure out what kind of "support" you need for this script and why, exactly? Is there a question about how the code works that hasn't already been addressed in one or more comments below this ticket, or are we talking something else entirely like some sorta bugfixing job because your server setup is different from mine?? -I can understand if your site needs to be running smoothly and you need help with a fix of sorts but there should really be nothing wrong here that the code itself could not handle. And given that I'm getting reports about how it works perfectly well on some other servers, what exactly are we talking? A detailed report will do wonders in helping us get this resolved for ya quickly so please take your time and describe the issue(s) you see as clearly & concisely as possible!! -@duncan-donut: I'm not sure if you have access to cPanel but you could try these instructions. It is worth a shot! Let me know how it goes (or what error message, exactly!) when/if ya give that code a go? [end of text] - - -main: mem per token = 71159620 bytes -main: load time = 19309.95 ms -main: sample time = 168.62 ms -main: predict time = 223895.61 ms / 888.47 ms per token -main: total time = 246406.42 ms - - Performance counter stats for './main -m ./models/65B/ggml-model-q4_0.bin -t 16 -n 1024 -p Please close your issue when it has been answered.': - - 3636882.89 msec task-clock # 14.677 CPUs utilized - 13509 context-switches # 3.714 /sec - 2436 cpu-migrations # 0.670 /sec - 10476679 page-faults # 2.881 K/sec - 13133115082869 cycles # 3.611 GHz (16.77%) - 29314462753 stalled-cycles-frontend # 0.22% frontend cycles idle (16.76%) - 10294402631459 stalled-cycles-backend # 78.39% backend cycles idle (16.74%) - 23479217109614 instructions # 1.79 insn per cycle - # 0.44 stalled cycles per insn (16.76%) - 2353072268027 branches # 647.002 M/sec (16.77%) - 1998682780 branch-misses # 0.08% of all branches (16.76%) - - 247.802177522 seconds time elapsed - - 3618.573072000 seconds user - 18.491698000 seconds sys -``` +Please include information about your system, the steps to reproduce the bug, and the version of llama.cpp that you are using. If possible, please provide a minimal code example that reproduces the bug. diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml index a7165a38f..87904b75e 100644 --- a/.github/workflows/docker.yml +++ b/.github/workflows/docker.yml @@ -69,6 +69,19 @@ jobs: docker-images: true swap-storage: true + - name: Determine tag name + id: tag + shell: bash + run: | + BUILD_NUMBER="$(git rev-list --count HEAD)" + SHORT_HASH="$(git rev-parse --short=7 HEAD)" + if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then + echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT + else + SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-') + echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT + fi + - name: Build and push Docker image (versioned) if: github.event_name == 'push' uses: docker/build-push-action@v4 @@ -85,5 +98,5 @@ jobs: context: . push: ${{ github.event_name == 'push' }} platforms: ${{ matrix.config.platforms }} - tags: "ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }}" + tags: "ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }},ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}" file: ${{ matrix.config.dockerfile }} diff --git a/.gitignore b/.gitignore index 76b3d2861..def74a1e9 100644 --- a/.gitignore +++ b/.gitignore @@ -48,6 +48,7 @@ models-mnt /llama-bench /llava-cli /lookahead +/lookup /main /metal /perplexity diff --git a/CMakeLists.txt b/CMakeLists.txt index 6fc6508c5..545aab267 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -302,6 +302,8 @@ if (LLAMA_CUBLAS) set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt) endif() + set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cuda_driver) + if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES) # 52 == lowest CUDA 12 standard # 60 == f16 CUDA intrinsics diff --git a/Makefile b/Makefile index 68df7702a..28c6d79bc 100644 --- a/Makefile +++ b/Makefile @@ -2,7 +2,7 @@ BUILD_TARGETS = \ main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \ simple batched batched-bench save-load-state server gguf llama-bench libllava.a llava-cli baby-llama beam-search \ - speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead tests/test-c.o + speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup tests/test-c.o # Binaries only useful for tests TEST_TARGETS = \ @@ -282,8 +282,17 @@ endif ifneq ($(filter aarch64%,$(UNAME_M)),) # Apple M1, M2, etc. # Raspberry Pi 3, 4, Zero 2 (64-bit) + # Nvidia Jetson MK_CFLAGS += -mcpu=native MK_CXXFLAGS += -mcpu=native + JETSON_RELEASE_INFO = $(shell jetson_release) + ifdef JETSON_RELEASE_INFO + ifneq ($(filter TX2%,$(JETSON_RELEASE_INFO)),) + JETSON_EOL_MODULE_DETECT = 1 + CC = aarch64-unknown-linux-gnu-gcc + cxx = aarch64-unknown-linux-gnu-g++ + endif + endif endif ifneq ($(filter armv6%,$(UNAME_M)),) @@ -357,15 +366,16 @@ ifdef LLAMA_BLIS endif # LLAMA_BLIS ifdef LLAMA_CUBLAS - MK_CPPFLAGS += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(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$(CUDA_PATH)/targets/x86_64-linux/lib + MK_CPPFLAGS += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include -I/usr/local/cuda/targets/aarch64-linux/include + MK_LDFLAGS += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib -L/usr/local/cuda/targets/aarch64-linux/lib -L/usr/lib/wsl/lib OBJS += ggml-cuda.o - MK_NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math - + MK_NVCCFLAGS = -use_fast_math +ifndef JETSON_EOL_MODULE_DETECT + MK_NVCCFLAGS += --forward-unknown-to-host-compiler +endif # JETSON_EOL_MODULE_DETECT ifdef LLAMA_DEBUG MK_NVCCFLAGS += -lineinfo -endif - +endif # LLAMA_DEBUG ifdef LLAMA_CUDA_NVCC NVCC = $(LLAMA_CUDA_NVCC) else @@ -417,7 +427,11 @@ ifdef LLAMA_CUDA_CCBIN MK_NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN) endif ggml-cuda.o: ggml-cuda.cu ggml-cuda.h +ifdef JETSON_EOL_MODULE_DETECT + $(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@ +else $(NVCC) $(BASE_CXXFLAGS) $(NVCCFLAGS) -Wno-pedantic -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@ +endif # JETSON_EOL_MODULE_DETECT endif # LLAMA_CUBLAS ifdef LLAMA_CLBLAST @@ -452,6 +466,9 @@ ifdef LLAMA_HIPBLAS LLAMA_CUDA_MMV_Y ?= 1 LLAMA_CUDA_KQUANTS_ITER ?= 2 MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS +ifdef LLAMA_HIP_UMA + MK_CPPFLAGS += -DGGML_HIP_UMA +endif # LLAMA_HIP_UMA MK_LDFLAGS += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib MK_LDFLAGS += -lhipblas -lamdhip64 -lrocblas HIPFLAGS += $(addprefix --offload-arch=,$(GPU_TARGETS)) @@ -645,6 +662,9 @@ parallel: examples/parallel/parallel.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) lookahead: examples/lookahead/lookahead.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) +lookup: examples/lookup/lookup.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) + ifdef LLAMA_METAL metal: examples/metal/metal.cpp ggml.o $(OBJS) $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) diff --git a/README.md b/README.md index 73fe59bb4..3b202a336 100644 --- a/README.md +++ b/README.md @@ -102,6 +102,7 @@ as the main playground for developing new features for the [ggml](https://github - [x] [Deepseek models](https://huggingface.co/models?search=deepseek-ai/deepseek) - [x] [Qwen models](https://huggingface.co/models?search=Qwen/Qwen) - [x] [Mixtral MoE](https://huggingface.co/models?search=mistral-ai/Mixtral) +- [x] [PLaMo-13B](https://github.com/ggerganov/llama.cpp/pull/3557) **Multimodal models:** @@ -123,6 +124,7 @@ as the main playground for developing new features for the [ggml](https://github - Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj) - React Native: [mybigday/llama.rn](https://github.com/mybigday/llama.rn) - Java: [kherud/java-llama.cpp](https://github.com/kherud/java-llama.cpp) +- Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig) **UI:** @@ -131,6 +133,7 @@ as the main playground for developing new features for the [ggml](https://github - [withcatai/catai](https://github.com/withcatai/catai) - [semperai/amica](https://github.com/semperai/amica) - [psugihara/FreeChat](https://github.com/psugihara/FreeChat) +- [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal) --- @@ -395,6 +398,9 @@ Building the program with BLAS support may lead to some performance improvements - #### cuBLAS This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads). + + For Jetson user, if you have Jetson Orin, you can try this: [Offical Support](https://www.jetson-ai-lab.com/tutorial_text-generation.html). If you are using an old model(nano/TX2), need some additional operations before compiling. + - Using `make`: ```bash make LLAMA_CUBLAS=1 @@ -439,7 +445,13 @@ Building the program with BLAS support may lead to some performance improvements && cmake --build build -- -j 16 ``` On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DLLAMA_HIP_UMA=ON"`. - However, this hurts performance for non-integrated GPUs. + However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs). + + - Using `make` (example for target gfx1030, build with 16 CPU threads): + ```bash + make -j16 LLAMA_HIPBLAS=1 LLAMA_HIP_UMA=1 AMDGPU_TARGETS=gxf1030 + ``` + - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU): ```bash set PATH=%HIP_PATH%\bin;%PATH% diff --git a/common/common.h b/common/common.h index e87ce1133..9659aa045 100644 --- a/common/common.h +++ b/common/common.h @@ -51,7 +51,7 @@ struct gpt_params { int32_t n_ctx = 512; // context size int32_t n_batch = 512; // batch size for prompt processing (must be >=32 to use BLAS) int32_t n_keep = 0; // number of tokens to keep from initial prompt - int32_t n_draft = 16; // number of tokens to draft during speculative decoding + int32_t n_draft = 8; // number of tokens to draft during speculative decoding int32_t n_chunks = -1; // max number of chunks to process (-1 = unlimited) int32_t n_parallel = 1; // number of parallel sequences to decode int32_t n_sequences = 1; // number of sequences to decode @@ -240,3 +240,4 @@ void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size = 80); // Dump the KV cache view showing individual sequences in each cell (long output). void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40); + diff --git a/common/sampling.cpp b/common/sampling.cpp index f4e76df31..8e45909f1 100644 --- a/common/sampling.cpp +++ b/common/sampling.cpp @@ -149,11 +149,12 @@ static void sampler_queue( } } -llama_token llama_sampling_sample( +static llama_token llama_sampling_sample_impl( struct llama_sampling_context * ctx_sampling, struct llama_context * ctx_main, struct llama_context * ctx_cfg, - const int idx) { + const int idx, + bool is_resampling) { // Add a parameter to indicate if we are resampling const llama_sampling_params & params = ctx_sampling->params; const int n_vocab = llama_n_vocab(llama_get_model(ctx_main)); @@ -173,8 +174,17 @@ llama_token llama_sampling_sample( llama_token id = 0; + // Get a pointer to the logits float * logits = llama_get_logits_ith(ctx_main, idx); + // Declare original_logits at the beginning of the function scope + std::vector original_logits; + + if (!is_resampling) { + // Only make a copy of the original logits if we are not in the resampling phase, not sure if I actually have to do this. + original_logits = std::vector(logits, logits + llama_n_vocab(llama_get_model(ctx_main))); + } + // apply params.logit_bias map for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) { logits[it->first] += it->second; @@ -193,12 +203,14 @@ llama_token llama_sampling_sample( } // apply penalties - if (!prev.empty()) { + const auto& penalty_tokens = params.use_penalty_prompt_tokens ? params.penalty_prompt_tokens : prev; + const int penalty_tokens_used_size = std::min((int)penalty_tokens.size(), penalty_last_n); + if (penalty_tokens_used_size) { 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, - penalty_last_n, penalty_repeat, penalty_freq, penalty_present); + penalty_tokens.data() + penalty_tokens.size() - penalty_tokens_used_size, + penalty_tokens_used_size, penalty_repeat, penalty_freq, penalty_present); if (!penalize_nl) { for (size_t idx = 0; idx < cur_p.size; idx++) { @@ -210,7 +222,8 @@ llama_token llama_sampling_sample( } } - if (ctx_sampling->grammar != NULL) { + // If we are in the resampling phase, apply grammar checks before sampling logic + if (is_resampling && ctx_sampling->grammar != NULL) { llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar); } @@ -252,9 +265,40 @@ llama_token llama_sampling_sample( } } + if (ctx_sampling->grammar != NULL && !is_resampling) { + // Create an array with a single token data element for the sampled id + llama_token_data single_token_data = {id, logits[id], 0.0f}; + llama_token_data_array single_token_data_array = { &single_token_data, 1, false }; + + // Apply grammar constraints to the single token + llama_sample_grammar(ctx_main, &single_token_data_array, ctx_sampling->grammar); + + // Check if the token is valid according to the grammar by seeing if its logit has been set to -INFINITY + bool is_valid = single_token_data_array.data[0].logit != -INFINITY; + + // If the token is not valid according to the grammar, perform resampling + if (!is_valid) { + LOG("Resampling because token %d: '%s' does not meet grammar rules\n", id, llama_token_to_piece(ctx_main, id).c_str()); + + // Restore logits from the copy + std::copy(original_logits.begin(), original_logits.end(), logits); + + return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, true); // Pass true for is_resampling + } + } + return id; } +llama_token llama_sampling_sample( + struct llama_sampling_context * ctx_sampling, + struct llama_context * ctx_main, + struct llama_context * ctx_cfg, + const int idx) { + // Call the implementation function with is_resampling set to false by default + return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, false); +} + void llama_sampling_accept( struct llama_sampling_context * ctx_sampling, struct llama_context * ctx_main, diff --git a/common/sampling.h b/common/sampling.h index fdfa9eed1..f16ef97e3 100644 --- a/common/sampling.h +++ b/common/sampling.h @@ -36,6 +36,9 @@ typedef struct llama_sampling_params { float cfg_scale = 1.f; // how strong is guidance std::unordered_map logit_bias; // logit bias for specific tokens + + std::vector penalty_prompt_tokens; + bool use_penalty_prompt_tokens = false; } llama_sampling_params; // general sampler context diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index e71a96c48..303d08170 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -184,6 +184,8 @@ class Model: return MixtralModel if model_architecture == "PhiForCausalLM": return Phi2Model + if model_architecture == "PlamoForCausalLM": + return PlamoModel return Model def _is_model_safetensors(self) -> bool: @@ -225,6 +227,8 @@ class Model: return gguf.MODEL_ARCH.LLAMA if arch == "PhiForCausalLM": return gguf.MODEL_ARCH.PHI2 + if arch == "PlamoForCausalLM": + return gguf.MODEL_ARCH.PLAMO raise NotImplementedError(f'Architecture "{arch}" not supported!') @@ -1002,11 +1006,91 @@ class Phi2Model(Model): self.gguf_writer.add_add_bos_token(False) +class PlamoModel(Model): + def set_vocab(self): + self._set_vocab_sentencepiece() + + def set_gguf_parameters(self): + hparams = self.hparams + block_count = hparams["num_hidden_layers"] + + self.gguf_writer.add_name("PLaMo") + self.gguf_writer.add_context_length(4096) # not in config.json + self.gguf_writer.add_embedding_length(hparams["hidden_size"]) + self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"]) + self.gguf_writer.add_block_count(block_count) + self.gguf_writer.add_head_count(hparams["num_attention_heads"]) + self.gguf_writer.add_head_count_kv(5) # hparams["num_key_value_heads"]) is wrong + self.gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"]) + + def shuffle_attn_q_weight(self, data_torch): + assert data_torch.size() == (5120, 5120) + data_torch = data_torch.reshape(8, 5, 128, 5120) + data_torch = torch.permute(data_torch, (1, 0, 2, 3)) + data_torch = torch.reshape(data_torch, (5120, 5120)) + return data_torch + + def shuffle_attn_output_weight(self, data_torch): + assert data_torch.size() == (5120, 5120) + data_torch = data_torch.reshape(5120, 8, 5, 128) + data_torch = torch.permute(data_torch, (0, 2, 1, 3)) + data_torch = torch.reshape(data_torch, (5120, 5120)) + return data_torch + + def write_tensors(self): + block_count = self.hparams.get("num_layers", self.hparams.get("num_hidden_layers")) + tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count) + + for name, data_torch in self.get_tensors(): + if "self_attn.rotary_emb.inv_freq" in name: + continue + + # map tensor names + new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) + if new_name is None: + print(f"Can not map tensor {name!r}") + sys.exit() + + # shuffle for broadcasting of gqa in ggml_mul_mat + if new_name.endswith("attn_q.weight"): + data_torch = self.shuffle_attn_q_weight(data_torch) + elif new_name.endswith("attn_output.weight"): + data_torch = self.shuffle_attn_output_weight(data_torch) + + old_dtype = data_torch.dtype + + # convert any unsupported data types to float32 + if data_torch.dtype not in (torch.float16, torch.float32): + data_torch = data_torch.to(torch.float32) + + data = data_torch.squeeze().numpy() + + n_dims = len(data.shape) + data_dtype = data.dtype + + # if f32 desired, convert any float16 to float32 + if self.ftype == 0 and data_dtype == np.float16: + data = data.astype(np.float32) + + # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 + if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1: + data = data.astype(np.float32) + + # if f16 desired, convert any float32 2-dim weight tensors to float16 + if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: + data = data.astype(np.float16) + + print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + + self.gguf_writer.add_tensor(new_name, data) + + ###### CONVERSION LOGIC ###### def parse_args() -> argparse.Namespace: - parser = argparse.ArgumentParser(description="Convert a huggingface model to a GGML compatible file") + parser = argparse.ArgumentParser( + description="Convert a huggingface model to a GGML compatible file") parser.add_argument( "--vocab-only", action="store_true", help="extract only the vocab", diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt index 6744944fd..4cc13d6e9 100644 --- a/examples/CMakeLists.txt +++ b/examples/CMakeLists.txt @@ -33,6 +33,7 @@ else() add_subdirectory(simple) add_subdirectory(speculative) add_subdirectory(lookahead) + add_subdirectory(lookup) add_subdirectory(train-text-from-scratch) if (LLAMA_METAL) add_subdirectory(metal) diff --git a/examples/lookup/CMakeLists.txt b/examples/lookup/CMakeLists.txt new file mode 100644 index 000000000..c060b8f56 --- /dev/null +++ b/examples/lookup/CMakeLists.txt @@ -0,0 +1,5 @@ +set(TARGET lookup) +add_executable(${TARGET} lookup.cpp) +install(TARGETS ${TARGET} RUNTIME) +target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_11) diff --git a/examples/lookup/README.md b/examples/lookup/README.md new file mode 100644 index 000000000..5bfb0de93 --- /dev/null +++ b/examples/lookup/README.md @@ -0,0 +1,13 @@ +# llama.cpp/examples/lookup + +Demonstration of Prompt Lookup Decoding + +https://github.com/apoorvumang/prompt-lookup-decoding + +The key parameters for lookup decoding are `ngram_min`, `ngram_max` and `n_draft`. The first two determine the size of the ngrams to search for in the prompt for a match. The latter specifies how many subsequent tokens to draft if a match is found. + +More info: + +https://github.com/ggerganov/llama.cpp/pull/4484 +https://github.com/ggerganov/llama.cpp/issues/4226 + diff --git a/examples/lookup/lookup.cpp b/examples/lookup/lookup.cpp new file mode 100644 index 000000000..d8de7dd38 --- /dev/null +++ b/examples/lookup/lookup.cpp @@ -0,0 +1,230 @@ +#include "common.h" +#include "llama.h" + +#include +#include +#include +#include + +int main(int argc, char ** argv){ + gpt_params params; + + if (!gpt_params_parse(argc, argv, params)) { + return 1; + } + + // max/min n-grams size to search for in prompt + const int ngram_max = 4; + const int ngram_min = 1; + + // length of the candidate / draft sequence, if match is found + const int n_draft = params.n_draft; + + const bool dump_kv_cache = params.dump_kv_cache; + +#ifndef LOG_DISABLE_LOGS + log_set_target(log_filename_generator("lookup", "log")); + LOG_TEE("Log start\n"); + log_dump_cmdline(argc, argv); +#endif // LOG_DISABLE_LOGS + + // init llama.cpp + llama_backend_init(params.numa); + + llama_model * model = NULL; + llama_context * ctx = NULL; + + // load the model + std::tie(model, ctx) = llama_init_from_gpt_params(params); + + // tokenize the prompt + const bool add_bos = llama_should_add_bos_token(model); + LOG("add_bos tgt: %d\n", add_bos); + + std::vector inp; + inp = ::llama_tokenize(ctx, params.prompt, add_bos, true); + + const int max_context_size = llama_n_ctx(ctx); + const int max_tokens_list_size = max_context_size - 4; + + if ((int) inp.size() > max_tokens_list_size) { + fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size); + return 1; + } + + fprintf(stderr, "\n\n"); + + for (auto id : inp) { + fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str()); + } + + fflush(stderr); + + const int n_input = inp.size(); + + const auto t_enc_start = ggml_time_us(); + + llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1, 0, 0)); + llama_decode(ctx, llama_batch_get_one(&inp.back(), 1, n_input - 1, 0)); + + const auto t_enc_end = ggml_time_us(); + + int n_predict = 0; + int n_drafted = 0; + int n_accept = 0; + + int n_past = inp.size(); + + bool has_eos = false; + + struct llama_sampling_context * ctx_sampling = llama_sampling_init(params.sparams); + + std::vector draft; + + llama_batch batch_tgt = llama_batch_init(params.n_ctx, 0, 1); + + // debug + struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, 1); + + const auto t_dec_start = ggml_time_us(); + + while (true) { + // debug + if (dump_kv_cache) { + llama_kv_cache_view_update(ctx, &kvc_view); + dump_kv_cache_view_seqs(kvc_view, 40); + } + + // print current draft sequence + LOG("drafted %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, draft).c_str()); + + int i_dft = 0; + while (true) { + // sample from the target model + llama_token id = llama_sampling_sample(ctx_sampling, ctx, NULL, i_dft); + + llama_sampling_accept(ctx_sampling, ctx, id, true); + + const std::string token_str = llama_token_to_piece(ctx, id); + + if (!params.use_color) { + printf("%s", token_str.c_str()); + } + + if (id == llama_token_eos(model)) { + has_eos = true; + } + + ++n_predict; + + // check if the target token matches the draft + if (i_dft < (int) draft.size() && id == draft[i_dft]) { + LOG("the sampled target token matches the %dth drafted token (%d, '%s') - accepted\n", i_dft, id, token_str.c_str()); + ++n_accept; + ++n_past; + ++i_dft; + inp.push_back(id); + + if (params.use_color) { + // color accepted draft token + printf("\033[34m%s\033[0m", token_str.c_str()); + fflush(stdout); + } + continue; + } + + if (params.use_color) { + printf("%s", token_str.c_str()); + } + fflush(stdout); + + + LOG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", id, token_str.c_str()); + + draft.clear(); + draft.push_back(id); + inp.push_back(id); + break; + } + + if ((params.n_predict > 0 && n_predict > params.n_predict) || has_eos) { + break; + } + + // KV cache management + // clean the cache of draft tokens that weren't accepted + llama_kv_cache_seq_rm(ctx, 0, n_past, -1); + + llama_batch_clear(batch_tgt); + llama_batch_add(batch_tgt, draft[0], n_past, { 0 }, true); + + // generate n_pred tokens through prompt lookup + auto prompt_lookup = [&]() -> void { + int inp_size = inp.size(); + for (int ngram_size = ngram_max ; ngram_size > ngram_min; --ngram_size){ + const llama_token * ngram = &inp[inp_size - ngram_size]; + + for (int i = 0; i <= (int) inp_size - (ngram_size * 2); ++i) { + bool match = true; + for (int j = 0; j < ngram_size; ++j) { + if (inp[i + j] != ngram[j]) { + match = false; + break; + } + } + + if (match) { + const int startIdx = i + ngram_size; + const int endIdx = startIdx + n_draft; + if (endIdx < inp_size) { + for (int j = startIdx; j < endIdx; ++j) { + LOG(" - draft candidate %d: %d\n", j, inp[j]); + draft.push_back(inp[j]); + llama_batch_add(batch_tgt, inp[j], n_past + (j - startIdx) + 1, { 0 }, true); + ++n_drafted; + } + return; + } + } + } + } + return; + }; + + prompt_lookup(); + + llama_decode(ctx, batch_tgt); + ++n_past; + + draft.erase(draft.begin()); + } + + auto t_dec_end = ggml_time_us(); + + LOG_TEE("\n\n"); + + LOG_TEE("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input, (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f)); + LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f)); + + LOG_TEE("\n"); + LOG_TEE("n_draft = %d\n", n_draft); + LOG_TEE("n_predict = %d\n", n_predict); + LOG_TEE("n_drafted = %d\n", n_drafted); + LOG_TEE("n_accept = %d\n", n_accept); + LOG_TEE("accept = %.3f%%\n", 100.0f * n_accept / n_drafted); + + LOG_TEE("\ntarget:\n"); + llama_print_timings(ctx); + + llama_sampling_free(ctx_sampling); + llama_batch_free(batch_tgt); + + llama_free(ctx); + llama_free_model(model); + + llama_backend_free(); + + fprintf(stderr, "\n\n"); + + return 0; +} diff --git a/examples/server/README.md b/examples/server/README.md index 0751b9612..f1e586a1c 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -148,6 +148,8 @@ node index.js `frequency_penalty`: Repeat alpha frequency penalty (default: 0.0, 0.0 = disabled); + `penalty_prompt`: This will replace the `prompt` for the purpose of the penalty evaluation. Can be either `null`, a string or an array of numbers representing tokens (default: `null` = use the original `prompt`). + `mirostat`: Enable Mirostat sampling, controlling perplexity during text generation (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0). `mirostat_tau`: Set the Mirostat target entropy, parameter tau (default: 5.0). diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 04038530f..72dfe452c 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -761,6 +761,42 @@ struct llama_server_context slot->prompt = ""; } + slot->sparams.penalty_prompt_tokens.clear(); + slot->sparams.use_penalty_prompt_tokens = false; + const auto &penalty_prompt = data.find("penalty_prompt"); + if (penalty_prompt != data.end()) + { + if (penalty_prompt->is_string()) + { + const auto penalty_prompt_string = penalty_prompt->get(); + auto penalty_tokens = llama_tokenize(model, penalty_prompt_string, false); + slot->sparams.penalty_prompt_tokens.swap(penalty_tokens); + if (slot->params.n_predict > 0) + { + slot->sparams.penalty_prompt_tokens.reserve(slot->sparams.penalty_prompt_tokens.size() + slot->params.n_predict); + } + slot->sparams.use_penalty_prompt_tokens = true; + } + else if (penalty_prompt->is_array()) + { + const auto n_tokens = penalty_prompt->size(); + slot->sparams.penalty_prompt_tokens.reserve(n_tokens + std::max(0, slot->params.n_predict)); + const int n_vocab = llama_n_vocab(model); + for (const auto &penalty_token : *penalty_prompt) + { + if (penalty_token.is_number_integer()) + { + const auto tok = penalty_token.get(); + if (tok >= 0 && tok < n_vocab) + { + slot->sparams.penalty_prompt_tokens.push_back(tok); + } + } + } + slot->sparams.use_penalty_prompt_tokens = true; + } + } + slot->sparams.logit_bias.clear(); if (json_value(data, "ignore_eos", false)) @@ -992,6 +1028,12 @@ struct llama_server_context slot.generated_text += token_str; slot.has_next_token = true; + if (slot.ctx_sampling->params.use_penalty_prompt_tokens && result.tok != -1) + { + // we can change penalty_prompt_tokens because it is always created from scratch each request + slot.ctx_sampling->params.penalty_prompt_tokens.push_back(result.tok); + } + // check if there is incomplete UTF-8 character at the end bool incomplete = false; for (unsigned i = 1; i < 5 && i <= slot.generated_text.size(); ++i) @@ -1183,6 +1225,8 @@ struct llama_server_context {"repeat_penalty", slot.sparams.penalty_repeat}, {"presence_penalty", slot.sparams.penalty_present}, {"frequency_penalty", slot.sparams.penalty_freq}, + {"penalty_prompt_tokens", slot.sparams.penalty_prompt_tokens}, + {"use_penalty_prompt_tokens", slot.sparams.use_penalty_prompt_tokens}, {"mirostat", slot.sparams.mirostat}, {"mirostat_tau", slot.sparams.mirostat_tau}, {"mirostat_eta", slot.sparams.mirostat_eta}, diff --git a/ggml-alloc.c b/ggml-alloc.c index a97436b17..a27dd54b0 100644 --- a/ggml-alloc.c +++ b/ggml-alloc.c @@ -72,7 +72,7 @@ static void remove_allocated_tensor(ggml_tallocr_t alloc, struct ggml_tensor * t // check if a tensor is allocated by this buffer static bool ggml_tallocr_is_own(ggml_tallocr_t alloc, const struct ggml_tensor * tensor) { - return tensor->buffer == alloc->buffer; + return tensor->buffer == alloc->buffer && (!tensor->view_src || tensor->view_src->buffer == alloc->buffer); } static bool ggml_is_view(struct ggml_tensor * t) { diff --git a/ggml-backend.c b/ggml-backend.c index 0c8c9ec43..526ce732b 100644 --- a/ggml-backend.c +++ b/ggml-backend.c @@ -297,7 +297,7 @@ static void ggml_backend_registry_init(void) { void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data) { GGML_ASSERT(ggml_backend_registry_count < GGML_MAX_BACKENDS_REG); - int id = ggml_backend_registry_count; + size_t id = ggml_backend_registry_count; ggml_backend_registry[id] = (struct ggml_backend_reg) { /* .name = */ {0}, @@ -330,6 +330,8 @@ size_t ggml_backend_reg_find_by_name(const char * name) { return i; } } + + // not found return SIZE_MAX; } @@ -340,15 +342,15 @@ ggml_backend_t ggml_backend_reg_init_backend_from_str(const char * backend_str) const char * params = strchr(backend_str, ':'); char backend_name[128]; if (params == NULL) { - strcpy(backend_name, backend_str); + snprintf(backend_name, sizeof(backend_name), "%s", backend_str); params = ""; } else { - strncpy(backend_name, backend_str, params - backend_str); - backend_name[params - backend_str] = '\0'; + snprintf(backend_name, sizeof(backend_name), "%.*s", (int)(params - backend_str), backend_str); params++; } size_t backend_i = ggml_backend_reg_find_by_name(backend_name); + if (backend_i == SIZE_MAX) { fprintf(stderr, "%s: backend %s not found\n", __func__, backend_name); return NULL; @@ -396,18 +398,12 @@ static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) { } static void ggml_backend_cpu_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { - GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds"); - GGML_ASSERT(tensor->data != NULL && "tensor not allocated"); - memcpy((char *)tensor->data + offset, data, size); GGML_UNUSED(buffer); } static void ggml_backend_cpu_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { - GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds"); - GGML_ASSERT(tensor->data != NULL && "tensor not allocated"); - memcpy(data, (const char *)tensor->data + offset, size); GGML_UNUSED(buffer); diff --git a/ggml-cuda.cu b/ggml-cuda.cu index ac91ee12e..abad9cc39 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -68,8 +68,9 @@ #define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault) #endif #define cudaMemcpy hipMemcpy -#define cudaMemcpy2DAsync hipMemcpy2DAsync #define cudaMemcpyAsync hipMemcpyAsync +#define cudaMemcpyPeerAsync hipMemcpyPeerAsync +#define cudaMemcpy2DAsync hipMemcpy2DAsync #define cudaMemcpyDeviceToDevice hipMemcpyDeviceToDevice #define cudaMemcpyDeviceToHost hipMemcpyDeviceToHost #define cudaMemcpyHostToDevice hipMemcpyHostToDevice @@ -86,10 +87,29 @@ #define cudaStream_t hipStream_t #define cudaSuccess hipSuccess #define __trap abort +#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS +#define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED +#define CUBLAS_STATUS_ALLOC_FAILED HIPBLAS_STATUS_ALLOC_FAILED +#define CUBLAS_STATUS_INVALID_VALUE HIPBLAS_STATUS_INVALID_VALUE +#define CUBLAS_STATUS_ARCH_MISMATCH HIPBLAS_STATUS_ARCH_MISMATCH +#define CUBLAS_STATUS_MAPPING_ERROR HIPBLAS_STATUS_MAPPING_ERROR +#define CUBLAS_STATUS_EXECUTION_FAILED HIPBLAS_STATUS_EXECUTION_FAILED +#define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR +#define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED #else #include +#include #include #include + +#if CUDART_VERSION < 11020 +#define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED +#define CUBLAS_TF32_TENSOR_OP_MATH CUBLAS_TENSOR_OP_MATH +#define CUBLAS_COMPUTE_16F CUDA_R_16F +#define CUBLAS_COMPUTE_32F CUDA_R_32F +#define cublasComputeType_t cudaDataType_t +#endif // CUDART_VERSION < 11020 + #endif // defined(GGML_USE_HIPBLAS) #include "ggml-cuda.h" @@ -144,7 +164,7 @@ static __device__ __forceinline__ int __vsubss4(const int a, const int b) { const int8x4_t vb = reinterpret_cast(b); #if __has_builtin(__builtin_elementwise_sub_sat) const int8x4_t c = __builtin_elementwise_sub_sat(va, vb); - return reinterpret_cast(c); + return reinterpret_cast(c); #else int8x4_t c; int16_t tmp; @@ -155,7 +175,7 @@ static __device__ __forceinline__ int __vsubss4(const int a, const int b) { if(tmp < std::numeric_limits::min()) tmp = std::numeric_limits::min(); c[i] = tmp; } - return reinterpret_cast(c); + return reinterpret_cast(c); #endif // __has_builtin(__builtin_elementwise_sub_sat) } @@ -193,45 +213,59 @@ static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) { static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size"); -#define CUDA_CHECK(err) \ - do { \ - cudaError_t err_ = (err); \ - if (err_ != cudaSuccess) { \ - int id; \ - cudaGetDevice(&id); \ - fprintf(stderr, "\nCUDA error %d at %s:%d: %s\n", err_, __FILE__, __LINE__, \ - cudaGetErrorString(err_)); \ - fprintf(stderr, "current device: %d\n", id); \ - GGML_ASSERT(!"CUDA error"); \ - } \ +[[noreturn]] +static void ggml_cuda_error(const char * stmt, const char * func, const char * file, const int line, const char * msg) { + int id = -1; // in case cudaGetDevice fails + cudaGetDevice(&id); + + fprintf(stderr, "CUDA error: %s\n", msg); + fprintf(stderr, " current device: %d, in function %s at %s:%d\n", id, func, file, line); + fprintf(stderr, " %s\n", stmt); + // abort with GGML_ASSERT to get a stack trace + GGML_ASSERT(!"CUDA error"); +} + +#define CUDA_CHECK_GEN(err, success, error_fn) \ + do { \ + auto err_ = (err); \ + if (err_ != (success)) { \ + ggml_cuda_error(#err, __func__, __FILE__, __LINE__, error_fn(err_)); \ + } \ } while (0) +#define CUDA_CHECK(err) CUDA_CHECK_GEN(err, cudaSuccess, cudaGetErrorString) + #if CUDART_VERSION >= 12000 -#define CUBLAS_CHECK(err) \ - do { \ - cublasStatus_t err_ = (err); \ - if (err_ != CUBLAS_STATUS_SUCCESS) { \ - int id; \ - cudaGetDevice(&id); \ - fprintf(stderr, "\ncuBLAS error %d at %s:%d: %s\n", \ - err_, __FILE__, __LINE__, cublasGetStatusString(err_)); \ - fprintf(stderr, "current device: %d\n", id); \ - GGML_ASSERT(!"cuBLAS error"); \ - } \ - } while (0) + static const char * cublas_get_error_str(const cublasStatus_t err) { + return cublasGetStatusString(err); + } #else -#define CUBLAS_CHECK(err) \ - do { \ - cublasStatus_t err_ = (err); \ - if (err_ != CUBLAS_STATUS_SUCCESS) { \ - int id; \ - cudaGetDevice(&id); \ - fprintf(stderr, "\ncuBLAS error %d at %s:%d\n", err_, __FILE__, __LINE__); \ - fprintf(stderr, "current device: %d\n", id); \ - GGML_ASSERT(!"cuBLAS error"); \ - } \ - } while (0) -#endif // CUDART_VERSION >= 11 + static const char * cublas_get_error_str(const cublasStatus_t err) { + switch (err) { + case CUBLAS_STATUS_SUCCESS: return "CUBLAS_STATUS_SUCCESS"; + case CUBLAS_STATUS_NOT_INITIALIZED: return "CUBLAS_STATUS_NOT_INITIALIZED"; + case CUBLAS_STATUS_ALLOC_FAILED: return "CUBLAS_STATUS_ALLOC_FAILED"; + case CUBLAS_STATUS_INVALID_VALUE: return "CUBLAS_STATUS_INVALID_VALUE"; + case CUBLAS_STATUS_ARCH_MISMATCH: return "CUBLAS_STATUS_ARCH_MISMATCH"; + case CUBLAS_STATUS_MAPPING_ERROR: return "CUBLAS_STATUS_MAPPING_ERROR"; + case CUBLAS_STATUS_EXECUTION_FAILED: return "CUBLAS_STATUS_EXECUTION_FAILED"; + case CUBLAS_STATUS_INTERNAL_ERROR: return "CUBLAS_STATUS_INTERNAL_ERROR"; + case CUBLAS_STATUS_NOT_SUPPORTED: return "CUBLAS_STATUS_NOT_SUPPORTED"; + default: return "unknown error"; + } + } +#endif // CUDART_VERSION >= 12000 + +#define CUBLAS_CHECK(err) CUDA_CHECK_GEN(err, CUBLAS_STATUS_SUCCESS, cublas_get_error_str) + +#if !defined(GGML_USE_HIPBLAS) +static const char * cu_get_error_str(CUresult err) { + const char * err_str; + cuGetErrorString(err, &err_str); + return err_str; +} +#define CU_CHECK(err) CUDA_CHECK_GEN(err, CUDA_SUCCESS, cu_get_error_str) +#endif #if CUDART_VERSION >= 11100 #define GGML_CUDA_ASSUME(x) __builtin_assume(x) @@ -287,10 +321,10 @@ typedef void (*ggml_cuda_func_t)(const ggml_tensor * src0, const ggml_tensor * s typedef void (*ggml_cuda_op_mul_mat_t)( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, 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); + const int64_t src1_padded_row_size, cudaStream_t stream); typedef void (*ggml_cuda_op_flatten_t)( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream); + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream); // QK = number of values after dequantization // QR = QK / number of values before dequantization @@ -496,22 +530,29 @@ struct ggml_tensor_extra_gpu { // this is faster on Windows // probably because the Windows CUDA libraries forget to make this check before invoking the drivers -inline cudaError_t ggml_cuda_set_device(const int device) { +static void ggml_cuda_set_device(const int device) { int current_device; CUDA_CHECK(cudaGetDevice(¤t_device)); if (device == current_device) { - return cudaSuccess; + return; } - return cudaSetDevice(device); + CUDA_CHECK(cudaSetDevice(device)); } 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}; +struct cuda_device_capabilities { + int cc; // compute capability + bool vmm; // virtual memory support + size_t vmm_granularity; // granularity of virtual memory +}; + +static cuda_device_capabilities g_device_caps[GGML_CUDA_MAX_DEVICES] = { {0, false, 0} }; + static void * g_scratch_buffer = nullptr; static size_t g_scratch_size = 0; // disabled by default static size_t g_scratch_offset = 0; @@ -553,6 +594,7 @@ static __device__ __forceinline__ float warp_reduce_max(float x) { static __device__ __forceinline__ float op_repeat(const float a, const float b) { return b; + GGML_UNUSED(a); } static __device__ __forceinline__ float op_add(const float a, const float b) { @@ -674,7 +716,7 @@ static __global__ void silu_f32(const float * x, float * dst, const int k) { dst[i] = x[i] / (1.0f + expf(-x[i])); } -static __global__ void gelu_quick_f32(const float *x, float *dst, int k) { +static __global__ void gelu_quick_f32(const float * x, float * dst, int k) { const float GELU_QUICK_COEF = -1.702f; const int i = blockDim.x*blockIdx.x + threadIdx.x; if (i >= k) { @@ -683,7 +725,7 @@ static __global__ void gelu_quick_f32(const float *x, float *dst, int k) { dst[i] = x[i] * (1.0f / (1.0f + expf(GELU_QUICK_COEF * x[i]))); } -static __global__ void tanh_f32(const float *x, float *dst, int k) { +static __global__ void tanh_f32(const float * x, float * dst, int k) { const int i = blockDim.x*blockIdx.x + threadIdx.x; if (i >= k) { return; @@ -700,7 +742,7 @@ static __global__ void relu_f32(const float * x, float * dst, const int k) { dst[i] = fmaxf(x[i], 0); } -static __global__ void leaky_relu_f32(const float *x, float *dst, const int k, const float negative_slope) { +static __global__ void leaky_relu_f32(const float * x, float * dst, const int k, const float negative_slope) { const int i = blockDim.x*blockIdx.x + threadIdx.x; if (i >= k) { return; @@ -753,7 +795,7 @@ static __global__ void norm_f32(const float * x, float * dst, const int ncols, c } } -static __global__ void concat_f32(const float *x,const float *y, float *dst, const int ne0, const int ne02) { +static __global__ void concat_f32(const float * x,const float * y, float * dst, const int ne0, const int ne02) { int nidx = threadIdx.x + blockIdx.x * blockDim.x; if (nidx >= ne0) { return; @@ -778,7 +820,7 @@ static __global__ void concat_f32(const float *x,const float *y, float *dst, c } } -static __global__ void upscale_f32(const float *x, float *dst, const int ne00, const int nb02, const int scale_factor) { +static __global__ void upscale_f32(const float * x, float * dst, const int ne00, const int nb02, const int scale_factor) { int ne0 = ne00 * scale_factor; int nidx = threadIdx.x + blockIdx.x * blockDim.x; if (nidx >= ne0) { @@ -798,7 +840,7 @@ static __global__ void upscale_f32(const float *x, float *dst, const int ne00, dst[offset_dst] = x[offset_src]; } -static __global__ void pad_f32(const float *x, float *dst, const int ne0, const int ne00, const int ne01, const int ne02) { +static __global__ void pad_f32(const float * x, float * dst, const int ne0, const int ne00, const int ne01, const int ne02) { int nidx = threadIdx.x + blockIdx.x * blockDim.x; if (nidx >= ne0) { return; @@ -4700,7 +4742,6 @@ static __global__ void mul_mat_p021_f16_f32( const int row_y = col_x; - // y is not transposed but permuted const int iy = channel*nrows_y + row_y; @@ -5266,17 +5307,17 @@ static __global__ void im2col_f32_f16( const int ky = (i - kd) / OW; const int ix = i % OW; - const int iiw = ix * s0 + kx * d0 - p0; - const int iih = blockIdx.y * s1 + ky * d1 - p1; + const int64_t iiw = ix * s0 + kx * d0 - p0; + const int64_t iih = blockIdx.y * s1 + ky * d1 - p1; - const int offset_dst = + const int64_t offset_dst = (blockIdx.y * OW + ix) * CHW + (blockIdx.z * (KW * KH) + ky * KW + kx); if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { dst[offset_dst] = __float2half(0.0f); } else { - const int offset_src = blockIdx.z * offset_delta; + const int64_t offset_src = blockIdx.z * offset_delta; dst[offset_dst] = __float2half(x[offset_src + iih * IW + iiw]); } } @@ -5375,7 +5416,7 @@ struct bin_bcast_cuda { cne[3] = 1; }; - auto collapse_nb = [](size_t cnb[], int64_t cne[]) { + auto collapse_nb = [](size_t cnb[], const int64_t cne[]) { cnb[1] *= cne[1]; cnb[2] *= cne[2]; cnb[3] *= cne[3]; @@ -5868,7 +5909,7 @@ static void ggml_mul_mat_q4_0_q8_1_cuda( int id; CUDA_CHECK(cudaGetDevice(&id)); - const int compute_capability = g_compute_capabilities[id]; + const int compute_capability = g_device_caps[id].cc; int mmq_x, mmq_y, nwarps; if (compute_capability >= CC_RDNA2) { @@ -5913,7 +5954,7 @@ static void ggml_mul_mat_q4_1_q8_1_cuda( int id; CUDA_CHECK(cudaGetDevice(&id)); - const int compute_capability = g_compute_capabilities[id]; + const int compute_capability = g_device_caps[id].cc; int mmq_x, mmq_y, nwarps; if (compute_capability >= CC_RDNA2) { @@ -5958,7 +5999,7 @@ static void ggml_mul_mat_q5_0_q8_1_cuda( int id; CUDA_CHECK(cudaGetDevice(&id)); - const int compute_capability = g_compute_capabilities[id]; + const int compute_capability = g_device_caps[id].cc; int mmq_x, mmq_y, nwarps; if (compute_capability >= CC_RDNA2) { @@ -6003,7 +6044,7 @@ static void ggml_mul_mat_q5_1_q8_1_cuda( int id; CUDA_CHECK(cudaGetDevice(&id)); - const int compute_capability = g_compute_capabilities[id]; + const int compute_capability = g_device_caps[id].cc; int mmq_x, mmq_y, nwarps; if (compute_capability >= CC_RDNA2) { @@ -6048,7 +6089,7 @@ static void ggml_mul_mat_q8_0_q8_1_cuda( int id; CUDA_CHECK(cudaGetDevice(&id)); - const int compute_capability = g_compute_capabilities[id]; + const int compute_capability = g_device_caps[id].cc; int mmq_x, mmq_y, nwarps; if (compute_capability >= CC_RDNA2) { @@ -6093,7 +6134,7 @@ static void ggml_mul_mat_q2_K_q8_1_cuda( int id; CUDA_CHECK(cudaGetDevice(&id)); - const int compute_capability = g_compute_capabilities[id]; + const int compute_capability = g_device_caps[id].cc; int mmq_x, mmq_y, nwarps; if (compute_capability >= CC_RDNA2) { @@ -6140,7 +6181,7 @@ static void ggml_mul_mat_q3_K_q8_1_cuda( int id; CUDA_CHECK(cudaGetDevice(&id)); - const int compute_capability = g_compute_capabilities[id]; + const int compute_capability = g_device_caps[id].cc; int mmq_x, mmq_y, nwarps; if (compute_capability >= CC_RDNA2) { @@ -6186,7 +6227,7 @@ static void ggml_mul_mat_q4_K_q8_1_cuda( int id; CUDA_CHECK(cudaGetDevice(&id)); - const int compute_capability = g_compute_capabilities[id]; + const int compute_capability = g_device_caps[id].cc; int mmq_x, mmq_y, nwarps; if (compute_capability >= CC_RDNA2) { @@ -6231,7 +6272,7 @@ static void ggml_mul_mat_q5_K_q8_1_cuda( int id; CUDA_CHECK(cudaGetDevice(&id)); - const int compute_capability = g_compute_capabilities[id]; + const int compute_capability = g_device_caps[id].cc; int mmq_x, mmq_y, nwarps; if (compute_capability >= CC_RDNA2) { @@ -6276,7 +6317,7 @@ static void ggml_mul_mat_q6_K_q8_1_cuda( int id; CUDA_CHECK(cudaGetDevice(&id)); - const int compute_capability = g_compute_capabilities[id]; + const int compute_capability = g_device_caps[id].cc; int mmq_x, mmq_y, nwarps; if (compute_capability >= CC_RDNA2) { @@ -6536,30 +6577,30 @@ struct scoped_spin_lock { scoped_spin_lock& operator=(const scoped_spin_lock&) = delete; }; -struct cuda_buffer { +static std::atomic_flag g_cuda_pool_lock = ATOMIC_FLAG_INIT; + +// #define DEBUG_CUDA_MALLOC +struct ggml_cuda_buffer { void * ptr = nullptr; size_t size = 0; }; -static cuda_buffer g_cuda_buffer_pool[GGML_CUDA_MAX_DEVICES][MAX_CUDA_BUFFERS]; -static std::atomic_flag g_cuda_pool_lock = ATOMIC_FLAG_INIT; +static ggml_cuda_buffer g_cuda_buffer_pool[GGML_CUDA_MAX_DEVICES][MAX_CUDA_BUFFERS]; +static size_t g_cuda_pool_size[GGML_CUDA_MAX_DEVICES] = {0}; -static void * ggml_cuda_pool_malloc(size_t size, size_t * actual_size) { +static void * ggml_cuda_pool_malloc_leg(int device, size_t size, size_t * actual_size) { scoped_spin_lock lock(g_cuda_pool_lock); - int id; - CUDA_CHECK(cudaGetDevice(&id)); #ifdef DEBUG_CUDA_MALLOC int nnz = 0; - size_t max_size = 0, tot_size = 0; + size_t max_size = 0; #endif size_t best_diff = 1ull << 36; int ibest = -1; for (int i = 0; i < MAX_CUDA_BUFFERS; ++i) { - cuda_buffer& b = g_cuda_buffer_pool[id][i]; + ggml_cuda_buffer& b = g_cuda_buffer_pool[device][i]; if (b.ptr != nullptr) { #ifdef DEBUG_CUDA_MALLOC ++nnz; - tot_size += b.size; if (b.size > max_size) max_size = b.size; #endif if (b.size >= size) { @@ -6579,32 +6620,32 @@ static void * ggml_cuda_pool_malloc(size_t size, size_t * actual_size) { } } if (ibest >= 0) { - cuda_buffer& b = g_cuda_buffer_pool[id][ibest]; + ggml_cuda_buffer& b = g_cuda_buffer_pool[device][ibest]; void * ptr = b.ptr; *actual_size = b.size; b.ptr = nullptr; b.size = 0; return ptr; } -#ifdef DEBUG_CUDA_MALLOC - fprintf(stderr, "%s: %d buffers, max_size = %u MB, tot_size = %u MB, requested %u MB\n", __func__, nnz, - (uint32_t)(max_size/1024/1024), (uint32_t)(tot_size/1024/1024), (uint32_t)(size/1024/1024)); -#endif void * ptr; size_t look_ahead_size = (size_t) (1.05 * size); look_ahead_size = 256 * ((look_ahead_size + 255)/256); + ggml_cuda_set_device(device); CUDA_CHECK(cudaMalloc((void **) &ptr, look_ahead_size)); *actual_size = look_ahead_size; + g_cuda_pool_size[device] += look_ahead_size; +#ifdef DEBUG_CUDA_MALLOC + fprintf(stderr, "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, id, nnz, + (uint32_t)(max_size/1024/1024), (uint32_t)(g_cuda_pool_size[id]/1024/1024), (uint32_t)(size/1024/1024)); +#endif return ptr; } -static void ggml_cuda_pool_free(void * ptr, size_t size) { +static void ggml_cuda_pool_free_leg(int device, void * ptr, size_t size) { scoped_spin_lock lock(g_cuda_pool_lock); - int id; - CUDA_CHECK(cudaGetDevice(&id)); for (int i = 0; i < MAX_CUDA_BUFFERS; ++i) { - cuda_buffer& b = g_cuda_buffer_pool[id][i]; + ggml_cuda_buffer& b = g_cuda_buffer_pool[device][i]; if (b.ptr == nullptr) { b.ptr = ptr; b.size = size; @@ -6612,9 +6653,149 @@ static void ggml_cuda_pool_free(void * ptr, size_t size) { } } fprintf(stderr, "WARNING: cuda buffer pool full, increase MAX_CUDA_BUFFERS\n"); + ggml_cuda_set_device(device); CUDA_CHECK(cudaFree(ptr)); + g_cuda_pool_size[device] -= size; } +#if !defined(GGML_USE_HIPBLAS) +// pool with virtual memory +static CUdeviceptr g_cuda_pool_addr[GGML_CUDA_MAX_DEVICES] = {0}; +static size_t g_cuda_pool_used[GGML_CUDA_MAX_DEVICES] = {0}; +static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 36; // 64 GB + +static void * ggml_cuda_pool_malloc_vmm(int device, size_t size, size_t * actual_size) { + scoped_spin_lock lock(g_cuda_pool_lock); + + // round up the allocation size to the alignment to ensure that all allocations are aligned for all data types + const size_t alignment = 128; + size = alignment * ((size + alignment - 1) / alignment); + + size_t avail = g_cuda_pool_size[device] - g_cuda_pool_used[device]; + + if (size > avail) { + // round up to the next multiple of the granularity + size_t reserve_size = size - avail; + const size_t granularity = g_device_caps[device].vmm_granularity; + reserve_size = granularity * ((reserve_size + granularity - 1) / granularity); + + GGML_ASSERT(g_cuda_pool_size[device] + reserve_size <= CUDA_POOL_VMM_MAX_SIZE); + + // allocate more physical memory + CUmemAllocationProp prop = {}; + prop.type = CU_MEM_ALLOCATION_TYPE_PINNED; + prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE; + prop.location.id = device; + CUmemGenericAllocationHandle handle; + CU_CHECK(cuMemCreate(&handle, reserve_size, &prop, 0)); + + // reserve virtual address space (if not already reserved) + if (g_cuda_pool_addr[device] == 0) { + CU_CHECK(cuMemAddressReserve(&g_cuda_pool_addr[device], CUDA_POOL_VMM_MAX_SIZE, 0, 0, 0)); + } + + // map at the end of the pool + CU_CHECK(cuMemMap(g_cuda_pool_addr[device] + g_cuda_pool_size[device], reserve_size, 0, handle, 0)); + + // the memory allocation handle is no longer needed after mapping + CU_CHECK(cuMemRelease(handle)); + + // set access + CUmemAccessDesc access = {}; + access.location.type = CU_MEM_LOCATION_TYPE_DEVICE; + access.location.id = device; + access.flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE; + CU_CHECK(cuMemSetAccess(g_cuda_pool_addr[device] + g_cuda_pool_size[device], reserve_size, &access, 1)); + + // add to the pool + g_cuda_pool_size[device] += reserve_size; + + //printf("cuda pool[%d]: size increased to %llu MB (reserved %llu MB)\n", + // id, (unsigned long long) (g_cuda_pool_size[id]/1024/1024), + // (unsigned long long) (reserve_size/1024/1024)); + } + + GGML_ASSERT(g_cuda_pool_addr[device] != 0); + + void * ptr = (void *) (g_cuda_pool_addr[device] + g_cuda_pool_used[device]); + *actual_size = size; + g_cuda_pool_used[device] += size; + +#ifdef DEBUG_CUDA_MALLOC + printf("cuda pool[%d]: allocated %llu bytes at %llx [%s]\n", id, (unsigned long long) size, ptr); +#endif + + return ptr; +} + +static void ggml_cuda_pool_free_vmm(int device, void * ptr, size_t size) { + scoped_spin_lock lock(g_cuda_pool_lock); + +#ifdef DEBUG_CUDA_MALLOC + printf("cuda pool[%d]: freed %llu bytes at %llx\n", id, (unsigned long long) size, ptr); +#endif + + g_cuda_pool_used[device] -= size; + + // all deallocations must be in reverse order of the allocations + GGML_ASSERT(ptr == (void *) (g_cuda_pool_addr[device] + g_cuda_pool_used[device])); +} + +static void * ggml_cuda_pool_malloc(int device, size_t size, size_t * actual_size) { + if (g_device_caps[device].vmm) { + return ggml_cuda_pool_malloc_vmm(device, size, actual_size); + } else { + return ggml_cuda_pool_malloc_leg(device, size, actual_size); + } +} + +static void ggml_cuda_pool_free(int device, void * ptr, size_t size) { + if (g_device_caps[device].vmm) { + ggml_cuda_pool_free_vmm(device, ptr, size); + } else { + ggml_cuda_pool_free_leg(device, ptr, size); + } +} +#else +#define ggml_cuda_pool_malloc ggml_cuda_pool_malloc_leg +#define ggml_cuda_pool_free ggml_cuda_pool_free_leg +#endif // !defined(GGML_USE_HIPBLAS) + +template +struct cuda_pool_alloc { + int device = -1; + T * ptr = nullptr; + size_t actual_size = 0; + + // size is in number of elements + T * alloc(size_t size) { + GGML_ASSERT(ptr == nullptr); + CUDA_CHECK(cudaGetDevice(&device)); + ptr = (T *) ggml_cuda_pool_malloc(device, size * sizeof(T), &this->actual_size); + return ptr; + } + + cuda_pool_alloc(size_t size) { + alloc(size); + } + + ~cuda_pool_alloc() { + if (ptr != nullptr) { + ggml_cuda_pool_free(device, ptr, actual_size); + } + } + + T * get() { + return ptr; + } + + cuda_pool_alloc() = default; + cuda_pool_alloc(const cuda_pool_alloc &) = delete; + cuda_pool_alloc(cuda_pool_alloc &&) = delete; + cuda_pool_alloc& operator=(const cuda_pool_alloc &) = delete; + cuda_pool_alloc& operator=(cuda_pool_alloc &&) = delete; +}; + static bool g_cublas_loaded = false; bool ggml_cublas_loaded(void) { @@ -6653,16 +6834,33 @@ void ggml_init_cublas() { #endif fprintf(stderr, "%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, g_device_count); for (int id = 0; id < g_device_count; ++id) { + int device_vmm = 0; + +#if !defined(GGML_USE_HIPBLAS) + CUdevice device; + CU_CHECK(cuDeviceGet(&device, id)); + CU_CHECK(cuDeviceGetAttribute(&device_vmm, CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED, device)); + + if (device_vmm) { + CUmemAllocationProp alloc_prop = {}; + alloc_prop.type = CU_MEM_ALLOCATION_TYPE_PINNED; + alloc_prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE; + alloc_prop.location.id = id; + CU_CHECK(cuMemGetAllocationGranularity(&g_device_caps[id].vmm_granularity, &alloc_prop, CU_MEM_ALLOC_GRANULARITY_RECOMMENDED)); + } +#endif // !defined(GGML_USE_HIPBLAS) + g_device_caps[id].vmm = !!device_vmm; + cudaDeviceProp prop; CUDA_CHECK(cudaGetDeviceProperties(&prop, id)); - fprintf(stderr, " Device %d: %s, compute capability %d.%d\n", id, prop.name, prop.major, prop.minor); + fprintf(stderr, " Device %d: %s, compute capability %d.%d, VMM: %s\n", id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no"); g_tensor_split[id] = total_vram; total_vram += prop.totalGlobalMem; #if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) - g_compute_capabilities[id] = 100*prop.major + 10*prop.minor + CC_OFFSET_AMD; + g_device_caps[id].cc = 100*prop.major + 10*prop.minor + CC_OFFSET_AMD; #else - g_compute_capabilities[id] = 100*prop.major + 10*prop.minor; + g_device_caps[id].cc = 100*prop.major + 10*prop.minor; #endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) } for (int id = 0; id < g_device_count; ++id) { @@ -6670,7 +6868,7 @@ void ggml_init_cublas() { } for (int id = 0; id < g_device_count; ++id) { - CUDA_CHECK(ggml_cuda_set_device(id)); + ggml_cuda_set_device(id); // create cuda streams for (int is = 0; is < MAX_STREAMS; ++is) { @@ -6722,8 +6920,7 @@ void * ggml_cuda_host_malloc(size_t size) { void * ptr = nullptr; cudaError_t err = cudaMallocHost((void **) &ptr, size); if (err != cudaSuccess) { - // The allocation error can be bypassed. A null ptr will assigned out of this function. - // This can fixed the OOM error in WSL. + // clear the error cudaGetLastError(); fprintf(stderr, "WARNING: failed to allocate %.2f MB of pinned memory: %s\n", size/1024.0/1024.0, cudaGetErrorString(err)); @@ -6786,7 +6983,7 @@ static cudaError_t ggml_cuda_cpy_tensor_2d( static void ggml_cuda_op_get_rows( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_d, const float * src1_d, float * dst_d, const cudaStream_t & stream) { + const float * src0_d, const float * src1_d, float * dst_d, cudaStream_t stream) { GGML_ASSERT(src1->type == GGML_TYPE_I32); GGML_ASSERT(dst->type == GGML_TYPE_F32); @@ -6828,9 +7025,9 @@ static void ggml_cuda_op_get_rows( } template -inline void ggml_cuda_op_bin_bcast( +static void ggml_cuda_op_bin_bcast( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src1->type == GGML_TYPE_F32); @@ -6849,7 +7046,7 @@ inline void ggml_cuda_op_bin_bcast( static void ggml_cuda_op_repeat( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_d, const float * src1_d, float * dst_d, const cudaStream_t & main_stream) { + const float * src0_d, const float * src1_d, float * dst_d, cudaStream_t main_stream) { ggml_cuda_op_bin_bcast>(dst, src0, dst, nullptr, src0_d, dst_d, main_stream); @@ -6857,16 +7054,16 @@ static void ggml_cuda_op_repeat( (void) src1_d; } -inline void ggml_cuda_op_add( +static void ggml_cuda_op_add( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { ggml_cuda_op_bin_bcast>(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream); } -inline void ggml_cuda_op_acc( +static void ggml_cuda_op_acc( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT(src1->type == GGML_TYPE_F32); @@ -6883,23 +7080,23 @@ inline void ggml_cuda_op_acc( (void) dst; } -inline void ggml_cuda_op_mul( +static void ggml_cuda_op_mul( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { ggml_cuda_op_bin_bcast>(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream); } -inline void ggml_cuda_op_div( +static void ggml_cuda_op_div( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { ggml_cuda_op_bin_bcast>(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream); } -inline void ggml_cuda_op_gelu( +static void ggml_cuda_op_gelu( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); @@ -6911,9 +7108,9 @@ inline void ggml_cuda_op_gelu( (void) src1_dd; } -inline void ggml_cuda_op_silu( +static void ggml_cuda_op_silu( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); @@ -6925,9 +7122,9 @@ inline void ggml_cuda_op_silu( (void) src1_dd; } -inline void ggml_cuda_op_gelu_quick( +static void ggml_cuda_op_gelu_quick( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); @@ -6939,9 +7136,9 @@ inline void ggml_cuda_op_gelu_quick( (void) src1_dd; } -inline void ggml_cuda_op_tanh( +static void ggml_cuda_op_tanh( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); @@ -6953,9 +7150,9 @@ inline void ggml_cuda_op_tanh( (void) src1_dd; } -inline void ggml_cuda_op_relu( +static void ggml_cuda_op_relu( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); @@ -6967,9 +7164,9 @@ inline void ggml_cuda_op_relu( (void) src1_dd; } -inline void ggml_cuda_op_leaky_relu( +static void ggml_cuda_op_leaky_relu( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); @@ -6984,9 +7181,9 @@ inline void ggml_cuda_op_leaky_relu( (void) src1_dd; } -inline void ggml_cuda_op_sqr( +static void ggml_cuda_op_sqr( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); @@ -6998,9 +7195,9 @@ inline void ggml_cuda_op_sqr( (void) src1_dd; } -inline void ggml_cuda_op_norm( +static void ggml_cuda_op_norm( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); @@ -7018,10 +7215,9 @@ inline void ggml_cuda_op_norm( (void) src1_dd; } - -inline void ggml_cuda_op_group_norm( +static void ggml_cuda_op_group_norm( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); @@ -7035,9 +7231,9 @@ inline void ggml_cuda_op_group_norm( (void) src1_dd; } -inline void ggml_cuda_op_concat( +static void ggml_cuda_op_concat( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT(src1->type == GGML_TYPE_F32); @@ -7051,9 +7247,9 @@ inline void ggml_cuda_op_concat( (void) dst; } -inline void ggml_cuda_op_upscale( +static void ggml_cuda_op_upscale( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT(dst->type == GGML_TYPE_F32); @@ -7068,9 +7264,9 @@ inline void ggml_cuda_op_upscale( (void) src1_dd; } -inline void ggml_cuda_op_pad( +static void ggml_cuda_op_pad( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT(dst->type == GGML_TYPE_F32); @@ -7085,9 +7281,9 @@ inline void ggml_cuda_op_pad( (void) src1_dd; } -inline void ggml_cuda_op_rms_norm( +static void ggml_cuda_op_rms_norm( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); @@ -7105,10 +7301,10 @@ inline void ggml_cuda_op_rms_norm( (void) src1_dd; } -inline void ggml_cuda_op_mul_mat_q( +static void ggml_cuda_op_mul_mat_q( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, 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) { + const int64_t src1_padded_row_size, cudaStream_t stream) { const int64_t ne00 = src0->ne[0]; @@ -7170,13 +7366,13 @@ inline void ggml_cuda_op_mul_mat_q( static int64_t get_row_rounding(ggml_type type) { int64_t min_compute_capability = INT_MAX; int64_t max_compute_capability = INT_MIN; - for (int64_t id = 0; id < g_device_count; ++id) { + for (int id = 0; id < g_device_count; ++id) { if (g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) { - if (min_compute_capability > g_compute_capabilities[id]) { - min_compute_capability = g_compute_capabilities[id]; + if (min_compute_capability > g_device_caps[id].cc) { + min_compute_capability = g_device_caps[id].cc; } - if (max_compute_capability < g_compute_capabilities[id]) { - max_compute_capability = g_compute_capabilities[id]; + if (max_compute_capability < g_device_caps[id].cc) { + max_compute_capability = g_device_caps[id].cc; } } } @@ -7228,10 +7424,10 @@ static int64_t get_row_rounding(ggml_type type) { #endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) } -inline void ggml_cuda_op_mul_mat_vec_q( +static void ggml_cuda_op_mul_mat_vec_q( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, 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) { + const int64_t src1_padded_row_size, cudaStream_t stream) { GGML_ASSERT(ggml_nrows(src1) == 1); @@ -7281,18 +7477,18 @@ inline void ggml_cuda_op_mul_mat_vec_q( (void) src1_padded_row_size; } -inline void ggml_cuda_op_dequantize_mul_mat_vec( +static void ggml_cuda_op_dequantize_mul_mat_vec( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, 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) { + const int64_t src1_padded_row_size, cudaStream_t stream) { const int64_t ne00 = src0->ne[0]; const int64_t row_diff = row_high - row_low; // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics #ifdef GGML_CUDA_F16 - size_t ash; - dfloat * src1_dfloat = nullptr; // dfloat == half + cuda_pool_alloc src1_dfloat_a; + half * src1_dfloat = nullptr; // dfloat == half bool src1_convert_f16 = src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 || @@ -7300,7 +7496,7 @@ inline void ggml_cuda_op_dequantize_mul_mat_vec( src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16; if (src1_convert_f16) { - src1_dfloat = (half *) ggml_cuda_pool_malloc(ne00*sizeof(half), &ash); + src1_dfloat = src1_dfloat_a.alloc(ne00); ggml_cpy_f32_f16_cuda((const char *) src1_ddf_i, (char *) src1_dfloat, ne00, ne00, 1, sizeof(float), 0, 0, ne00, 1, sizeof(half), 0, 0, stream); @@ -7348,12 +7544,6 @@ inline void ggml_cuda_op_dequantize_mul_mat_vec( break; } -#ifdef GGML_CUDA_F16 - if (src1_convert_f16) { - ggml_cuda_pool_free(src1_dfloat, ash); - } -#endif // GGML_CUDA_F16 - (void) src1; (void) dst; (void) src1_ddq_i; @@ -7361,10 +7551,10 @@ inline void ggml_cuda_op_dequantize_mul_mat_vec( (void) src1_padded_row_size; } -inline void ggml_cuda_op_mul_mat_cublas( +static void ggml_cuda_op_mul_mat_cublas( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, 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) { + const int64_t src1_padded_row_size, cudaStream_t stream) { GGML_ASSERT(src0_dd_i != nullptr); GGML_ASSERT(src1_ddf_i != nullptr); @@ -7384,33 +7574,30 @@ inline void ggml_cuda_op_mul_mat_cublas( // ldc == nrows of the matrix that cuBLAS writes into int ldc = dst->backend == GGML_BACKEND_GPU && id == g_main_device ? ne0 : row_diff; - const int compute_capability = g_compute_capabilities[id]; + const int compute_capability = g_device_caps[id].cc; if (compute_capability >= CC_VOLTA && (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT) { // convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32 - half * src0_as_f16 = nullptr; - size_t src0_as = 0; + cuda_pool_alloc src0_as_f16; if (src0->type != GGML_TYPE_F16) { const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src0->type); GGML_ASSERT(to_fp16_cuda != nullptr); size_t ne = row_diff*ne00; - src0_as_f16 = (half *) ggml_cuda_pool_malloc(ne * sizeof(half), &src0_as); - to_fp16_cuda(src0_dd_i, src0_as_f16, ne, stream); + src0_as_f16.alloc(ne); + to_fp16_cuda(src0_dd_i, src0_as_f16.get(), ne, stream); } - const half * src0_ptr = src0->type == GGML_TYPE_F16 ? (const half *) src0_dd_i : src0_as_f16; + const half * src0_ptr = src0->type == GGML_TYPE_F16 ? (const half *) src0_dd_i : src0_as_f16.get(); - half * src1_as_f16 = nullptr; - size_t src1_as = 0; + cuda_pool_alloc src1_as_f16; if (src1->type != GGML_TYPE_F16) { const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type); GGML_ASSERT(to_fp16_cuda != nullptr); size_t ne = src1_ncols*ne10; - src1_as_f16 = (half *) ggml_cuda_pool_malloc(ne * sizeof(half), &src1_as); - to_fp16_cuda(src1_ddf_i, src1_as_f16, ne, stream); + src1_as_f16.alloc(ne); + to_fp16_cuda(src1_ddf_i, src1_as_f16.get(), ne, stream); } - const half * src1_ptr = src1->type == GGML_TYPE_F16 ? (const half *) src1_ddf_i : src1_as_f16; - size_t dst_as = 0; - half * dst_f16 = (half *) ggml_cuda_pool_malloc(row_diff*src1_ncols * sizeof(half), &dst_as); + const half * src1_ptr = src1->type == GGML_TYPE_F16 ? (const half *) src1_ddf_i : src1_as_f16.get(); + cuda_pool_alloc dst_f16(row_diff*src1_ncols); const half alpha_f16 = 1.0f; const half beta_f16 = 0.0f; @@ -7419,36 +7606,25 @@ inline void ggml_cuda_op_mul_mat_cublas( CUBLAS_CHECK( cublasGemmEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N, row_diff, src1_ncols, ne10, - &alpha_f16, src0_ptr, CUDA_R_16F, ne00, - src1_ptr, CUDA_R_16F, ne10, - &beta_f16, dst_f16, CUDA_R_16F, ldc, + &alpha_f16, src0_ptr, CUDA_R_16F, ne00, + src1_ptr, CUDA_R_16F, ne10, + &beta_f16, dst_f16.get(), CUDA_R_16F, ldc, CUBLAS_COMPUTE_16F, CUBLAS_GEMM_DEFAULT_TENSOR_OP)); const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16); - to_fp32_cuda(dst_f16, dst_dd_i, row_diff*src1_ncols, stream); - - ggml_cuda_pool_free(dst_f16, dst_as); - - if (src0_as != 0) { - ggml_cuda_pool_free(src0_as_f16, src0_as); - } - - if (src1_as != 0) { - ggml_cuda_pool_free(src1_as_f16, src1_as); - } + to_fp32_cuda(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream); } else { - float * src0_ddq_as_f32 = nullptr; - size_t src0_as = 0; + cuda_pool_alloc src0_ddq_as_f32; if (src0->type != GGML_TYPE_F32) { const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(src0->type); GGML_ASSERT(to_fp32_cuda != nullptr); - src0_ddq_as_f32 = (float *) ggml_cuda_pool_malloc(row_diff*ne00 * sizeof(float), &src0_as); // NOLINT - to_fp32_cuda(src0_dd_i, src0_ddq_as_f32, row_diff*ne00, stream); + src0_ddq_as_f32.alloc(row_diff*ne00); + to_fp32_cuda(src0_dd_i, src0_ddq_as_f32.get(), row_diff*ne00, stream); } - const float * src0_ddf_i = src0->type == GGML_TYPE_F32 ? (const float *) src0_dd_i : src0_ddq_as_f32; + const float * src0_ddf_i = src0->type == GGML_TYPE_F32 ? (const float *) src0_dd_i : src0_ddq_as_f32.get(); const float alpha = 1.0f; const float beta = 0.0f; @@ -7460,10 +7636,6 @@ inline void ggml_cuda_op_mul_mat_cublas( &alpha, src0_ddf_i, ne00, src1_ddf_i, ne10, &beta, dst_dd_i, ldc)); - - if (src0_as != 0) { - ggml_cuda_pool_free(src0_ddq_as_f32, src0_as); - } } (void) dst; @@ -7471,9 +7643,9 @@ inline void ggml_cuda_op_mul_mat_cublas( (void) src1_padded_row_size; } -inline void ggml_cuda_op_rope( +static void ggml_cuda_op_rope( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); @@ -7551,9 +7723,9 @@ inline void ggml_cuda_op_rope( (void) src1_dd; } -inline void ggml_cuda_op_alibi( +static void ggml_cuda_op_alibi( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); @@ -7582,9 +7754,9 @@ inline void ggml_cuda_op_alibi( (void) src1_dd; } -inline void ggml_cuda_op_im2col( +static void ggml_cuda_op_im2col( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F16); GGML_ASSERT(src1->type == GGML_TYPE_F32); @@ -7617,10 +7789,9 @@ inline void ggml_cuda_op_im2col( (void) src0_dd; } - -inline void ggml_cuda_op_sum_rows( +static void ggml_cuda_op_sum_rows( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); @@ -7635,9 +7806,9 @@ inline void ggml_cuda_op_sum_rows( (void) src1_dd; } -inline void ggml_cuda_op_argsort( +static void ggml_cuda_op_argsort( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_I32); @@ -7654,9 +7825,9 @@ inline void ggml_cuda_op_argsort( (void) src1_dd; } -inline void ggml_cuda_op_diag_mask_inf( +static void ggml_cuda_op_diag_mask_inf( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); @@ -7674,9 +7845,9 @@ inline void ggml_cuda_op_diag_mask_inf( (void) src1_dd; } -inline void ggml_cuda_op_soft_max( +static void ggml_cuda_op_soft_max( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); @@ -7695,14 +7866,15 @@ inline void ggml_cuda_op_soft_max( (void) dst; } -inline void ggml_cuda_op_scale( +static void ggml_cuda_op_scale( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); - const float scale = ((float *) dst->op_params)[0]; + float scale; + memcpy(&scale, dst->op_params, sizeof(float)); scale_f32_cuda(src0_dd, dst_dd, scale, ggml_nelements(src0), main_stream); CUDA_CHECK(cudaGetLastError()); @@ -7712,9 +7884,9 @@ inline void ggml_cuda_op_scale( (void) src1_dd; } -inline void ggml_cuda_op_clamp( +static void ggml_cuda_op_clamp( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); @@ -7754,18 +7926,17 @@ static void ggml_cuda_op_flatten(const ggml_tensor * src0, const ggml_tensor * s float * src1_ddf = nullptr; float * dst_ddf = nullptr; - // as = actual size - size_t src0_asf = 0; - size_t src1_asf = 0; - size_t dst_asf = 0; + cuda_pool_alloc src0_f; + cuda_pool_alloc src1_f; + cuda_pool_alloc dst_f; ggml_cuda_set_device(g_main_device); - const cudaStream_t main_stream = g_cudaStreams[g_main_device][0]; + cudaStream_t main_stream = g_cudaStreams[g_main_device][0]; if (src0_on_device) { src0_ddf = (float *) src0_extra->data_device[g_main_device]; } else { - src0_ddf = (float *) ggml_cuda_pool_malloc(ggml_nbytes(src0), &src0_asf); + src0_ddf = src0_f.alloc(ggml_nelements(src0)); CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_ddf, src0, 0, 0, 0, nrows0, main_stream)); } @@ -7773,14 +7944,14 @@ static void ggml_cuda_op_flatten(const ggml_tensor * src0, const ggml_tensor * s if (src1_on_device) { src1_ddf = (float *) src1_extra->data_device[g_main_device]; } else { - src1_ddf = (float *) ggml_cuda_pool_malloc(ggml_nbytes(src1), &src1_asf); + src1_ddf = src1_f.alloc(ggml_nelements(src1)); CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src1_ddf, src1, 0, 0, 0, nrows1, main_stream)); } } if (dst_on_device) { dst_ddf = (float *) dst_extra->data_device[g_main_device]; } else { - dst_ddf = (float *) ggml_cuda_pool_malloc(ggml_nbytes(dst), &dst_asf); + dst_ddf = dst_f.alloc(ggml_nelements(dst)); } // do the computation @@ -7792,16 +7963,6 @@ static void ggml_cuda_op_flatten(const ggml_tensor * src0, const ggml_tensor * s CUDA_CHECK(cudaMemcpyAsync(dst->data, dst_ddf, ggml_nbytes(dst), cudaMemcpyDeviceToHost, main_stream)); } - if (src0_asf > 0) { - ggml_cuda_pool_free(src0_ddf, src0_asf); - } - if (src1_asf > 0) { - ggml_cuda_pool_free(src1_ddf, src1_asf); - } - if (dst_asf > 0) { - ggml_cuda_pool_free(dst_ddf, dst_asf); - } - if (dst->backend == GGML_BACKEND_CPU) { CUDA_CHECK(cudaDeviceSynchronize()); } @@ -7818,12 +7979,12 @@ static void ggml_cuda_set_peer_access(const int n_tokens) { #ifdef NDEBUG for (int id = 0; id < g_device_count; ++id) { - CUDA_CHECK(ggml_cuda_set_device(id)); + ggml_cuda_set_device(id); CUDA_CHECK(cudaDeviceSynchronize()); } for (int id = 0; id < g_device_count; ++id) { - CUDA_CHECK(ggml_cuda_set_device(id)); + ggml_cuda_set_device(id); for (int id_other = 0; id_other < g_device_count; ++id_other) { if (id == id_other) { @@ -7857,7 +8018,6 @@ static void ggml_cuda_op_mul_mat( const int64_t ne01 = src0->ne[1]; const int64_t ne02 = src0->ne[2]; const int64_t ne03 = src0->ne[3]; - const int64_t nrows0 = ggml_nrows(src0); const int64_t ne10 = src1->ne[0]; const int64_t ne11 = src1->ne[1]; @@ -7900,27 +8060,29 @@ static void ggml_cuda_op_mul_mat( GGML_ASSERT(!(split && ne03 > 1)); GGML_ASSERT(!(split && ne02 < ne12)); - // dd = data device - char * src0_dd[GGML_CUDA_MAX_DEVICES] = {nullptr}; - float * src1_ddf[GGML_CUDA_MAX_DEVICES] = {nullptr}; // float - char * src1_ddq[GGML_CUDA_MAX_DEVICES] = {nullptr}; // q8_1 - float * dst_dd[GGML_CUDA_MAX_DEVICES] = {nullptr}; + struct dev_data { + cuda_pool_alloc src0_dd_alloc; + cuda_pool_alloc src1_ddf_alloc; + cuda_pool_alloc src1_ddq_alloc; + cuda_pool_alloc dst_dd_alloc; - // as = actual size - size_t src0_as[GGML_CUDA_MAX_DEVICES] = {0}; - size_t src1_asf[GGML_CUDA_MAX_DEVICES] = {0}; - size_t src1_asq[GGML_CUDA_MAX_DEVICES] = {0}; - size_t dst_as[GGML_CUDA_MAX_DEVICES] = {0}; + char * src0_dd = nullptr; + float * src1_ddf = nullptr; // float + char * src1_ddq = nullptr; // q8_1 + float * dst_dd = nullptr; - int64_t row_low[GGML_CUDA_MAX_DEVICES]; - int64_t row_high[GGML_CUDA_MAX_DEVICES]; + int64_t row_low; + int64_t row_high; + }; + + dev_data dev[GGML_CUDA_MAX_DEVICES]; int used_devices = 0; - for (int64_t id = 0; id < g_device_count; ++id) { + for (int id = 0; id < g_device_count; ++id) { // by default, use all rows - row_low[id] = 0; - row_high[id] = ne01; + dev[id].row_low = 0; + dev[id].row_high = ne01; // for multi GPU, get the row boundaries from tensor split // and round to mul_mat_q tile sizes @@ -7928,19 +8090,23 @@ static void ggml_cuda_op_mul_mat( const int64_t rounding = get_row_rounding(src0->type); if (id != 0) { - row_low[id] = ne01*g_tensor_split[id]; - row_low[id] -= row_low[id] % rounding; + dev[id].row_low = ne01*g_tensor_split[id]; + if (dev[id].row_low < ne01) { + dev[id].row_low -= dev[id].row_low % rounding; + } } if (id != g_device_count - 1) { - row_high[id] = ne01*g_tensor_split[id + 1]; - row_high[id] -= row_high[id] % rounding; + dev[id].row_high = ne01*g_tensor_split[id + 1]; + if (dev[id].row_high < ne01) { + dev[id].row_high -= dev[id].row_high % rounding; + } } } } - for (int64_t id = 0; id < g_device_count; ++id) { - if ((!split && id != g_main_device) || row_low[id] == row_high[id]) { + for (int id = 0; id < g_device_count; ++id) { + if ((!split && id != g_main_device) || dev[id].row_low == dev[id].row_high) { continue; } @@ -7950,42 +8116,41 @@ static void ggml_cuda_op_mul_mat( const bool dst_on_device = dst->backend == GGML_BACKEND_GPU && id == g_main_device; ggml_cuda_set_device(id); - const cudaStream_t stream = g_cudaStreams[id][0]; + cudaStream_t stream = g_cudaStreams[id][0]; if (src0_on_device && src0_is_contiguous) { - src0_dd[id] = (char *) src0_extra->data_device[id]; + dev[id].src0_dd = (char *) src0_extra->data_device[id]; } else { - // const size_t size_src0_ddq = split ? (row_high[id]-row_low[id])*ne00 * src0_ts/src0_bs : ggml_nbytes(src0); - src0_dd[id] = (char *) ggml_cuda_pool_malloc(ggml_nbytes(src0), &src0_as[id]); + dev[id].src0_dd = dev[id].src0_dd_alloc.alloc(ggml_nbytes(src0)); } if (src1_on_device && src1_is_contiguous) { - src1_ddf[id] = (float *) src1_extra->data_device[id]; + dev[id].src1_ddf = (float *) src1_extra->data_device[id]; } else { - src1_ddf[id] = (float *) ggml_cuda_pool_malloc(ggml_nbytes(src1), &src1_asf[id]); + dev[id].src1_ddf = dev[id].src1_ddf_alloc.alloc(ggml_nelements(src1)); } if (convert_src1_to_q8_1) { - src1_ddq[id] = (char *) ggml_cuda_pool_malloc(nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs, &src1_asq[id]); + dev[id].src1_ddq = dev[id].src1_ddq_alloc.alloc(nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs); if (src1_on_device && src1_is_contiguous) { - quantize_row_q8_1_cuda(src1_ddf[id], src1_ddq[id], ne10, nrows1, src1_padded_col_size, stream); + quantize_row_q8_1_cuda(dev[id].src1_ddf, dev[id].src1_ddq, ne10, nrows1, src1_padded_col_size, stream); CUDA_CHECK(cudaGetLastError()); } } if (dst_on_device) { - dst_dd[id] = (float *) dst_extra->data_device[id]; + dev[id].dst_dd = (float *) dst_extra->data_device[id]; } else { - const size_t size_dst_ddf = split ? (row_high[id]-row_low[id])*ne1*sizeof(float) : ggml_nbytes(dst); - dst_dd[id] = (float *) ggml_cuda_pool_malloc(size_dst_ddf, &dst_as[id]); + const size_t size_dst_ddf = split ? (dev[id].row_high - dev[id].row_low)*ne1 : ggml_nelements(dst); + dev[id].dst_dd = dev[id].dst_dd_alloc.alloc(size_dst_ddf); } } // if multiple devices are used they need to wait for the main device // here an event is recorded that signals that the main device has finished calculating the input data if (split && used_devices > 1) { - CUDA_CHECK(ggml_cuda_set_device(g_main_device)); + ggml_cuda_set_device(g_main_device); CUDA_CHECK(cudaEventRecord(src0_extra->events[g_main_device][0], g_cudaStreams[g_main_device][0])); } @@ -7994,17 +8159,17 @@ static void ggml_cuda_op_mul_mat( const int64_t is = split ? (src1_col_0/src1_col_stride) % MAX_STREAMS : 0; const int64_t src1_ncols = src1_col_0 + src1_col_stride > ne11 ? ne11 - src1_col_0 : src1_col_stride; - for (int64_t id = 0; id < g_device_count; ++id) { - if ((!split && id != g_main_device) || row_low[id] == row_high[id]) { + for (int id = 0; id < g_device_count; ++id) { + if ((!split && id != g_main_device) || dev[id].row_low == dev[id].row_high) { continue; } const bool src1_on_device = src1->backend == GGML_BACKEND_GPU && id == g_main_device; const bool dst_on_device = dst->backend == GGML_BACKEND_GPU && id == g_main_device; - const int64_t row_diff = row_high[id] - row_low[id]; + const int64_t row_diff = dev[id].row_high - dev[id].row_low; ggml_cuda_set_device(id); - const cudaStream_t stream = g_cudaStreams[id][is]; + cudaStream_t stream = g_cudaStreams[id][is]; // wait for main GPU data if necessary if (split && (id != g_main_device || is != 0)) { @@ -8018,34 +8183,34 @@ static void ggml_cuda_op_mul_mat( const size_t src1_ddq_i_offset = (i0*ne11 + src1_col_0) * src1_padded_col_size*q8_1_ts/q8_1_bs; // for split tensors the data begins at i0 == i0_offset_low - char * src0_dd_i = src0_dd[id] + (i0/i02_divisor) * (ne01*ne00*src0_ts)/src0_bs; - float * src1_ddf_i = src1_ddf[id] + (i0*ne11 + src1_col_0) * ne10; - char * src1_ddq_i = src1_ddq[id] + src1_ddq_i_offset; - float * dst_dd_i = dst_dd[id] + (i0*ne1 + src1_col_0) * (dst_on_device ? ne0 : row_diff); + char * src0_dd_i = dev[id].src0_dd + (i0/i02_divisor) * (ne01*ne00*src0_ts)/src0_bs; + float * src1_ddf_i = dev[id].src1_ddf + (i0*ne11 + src1_col_0) * ne10; + char * src1_ddq_i = dev[id].src1_ddq + src1_ddq_i_offset; + float * dst_dd_i = dev[id].dst_dd + (i0*ne1 + src1_col_0) * (dst_on_device ? ne0 : row_diff); // the main device memory buffer can be on VRAM scratch, with space for all partial results // in that case an offset on dst_ddf_i is needed if (dst->backend == GGML_BACKEND_GPU && id == g_main_device) { - dst_dd_i += row_low[id]; // offset is 0 if no tensor split + dst_dd_i += dev[id].row_low; // offset is 0 if no tensor split } // copy src0, src1 to device if necessary if (src1->backend == GGML_BACKEND_GPU && src1_is_contiguous) { if (id != g_main_device) { if (convert_src1_to_q8_1) { - char * src1_ddq_i_source = src1_ddq[g_main_device] + src1_ddq_i_offset; - CUDA_CHECK(cudaMemcpyAsync(src1_ddq_i, src1_ddq_i_source, src1_ncols*src1_padded_col_size*q8_1_ts/q8_1_bs, - cudaMemcpyDeviceToDevice, stream)); + char * src1_ddq_i_source = dev[g_main_device].src1_ddq + src1_ddq_i_offset; + CUDA_CHECK(cudaMemcpyPeerAsync(src1_ddq_i, id, src1_ddq_i_source, g_main_device, + src1_ncols*src1_padded_col_size*q8_1_ts/q8_1_bs, stream)); } else { float * src1_ddf_i_source = (float *) src1_extra->data_device[g_main_device]; src1_ddf_i_source += (i0*ne11 + src1_col_0) * ne10; - CUDA_CHECK(cudaMemcpyAsync(src1_ddf_i, src1_ddf_i_source, src1_ncols*ne10*sizeof(float), - cudaMemcpyDeviceToDevice, stream)); + CUDA_CHECK(cudaMemcpyPeerAsync(src1_ddf_i, id, src1_ddf_i_source, g_main_device, + src1_ncols*ne10*sizeof(float), stream)); } } } else if (src1->backend == GGML_BACKEND_CPU || (src1_on_device && !src1_is_contiguous)) { CUDA_CHECK(ggml_cuda_cpy_tensor_2d( - src1_ddf_i, src1, i03, i02, src1_col_0, src1_col_0+src1_ncols, stream)); + src1_ddf_i, src1, i03, i02, src1_col_0, src1_col_0+src1_ncols, stream)); } else { GGML_ASSERT(false); } @@ -8056,12 +8221,12 @@ static void ggml_cuda_op_mul_mat( } if (src1_col_0 == 0 && (!src0_on_device || !src0_is_contiguous) && i02 % i02_divisor == 0) { - CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_dd_i, src0, i03, i02/i02_divisor, row_low[id], row_high[id], stream)); + CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_dd_i, src0, i03, i02/i02_divisor, dev[id].row_low, dev[id].row_high, stream)); } // do the computation op(src0, src1, dst, src0_dd_i, src1_ddf_i, src1_ddq_i, dst_dd_i, - row_low[id], row_high[id], src1_ncols, src1_padded_col_size, stream); + dev[id].row_low, dev[id].row_high, src1_ncols, src1_padded_col_size, stream); CUDA_CHECK(cudaGetLastError()); // copy dst to host or other device if necessary @@ -8085,9 +8250,25 @@ static void ggml_cuda_op_mul_mat( // If dst is a vector with ne0 == 1 then you don't have to do this but it still produces correct results. float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3); GGML_ASSERT(dst->nb[1] == ne0*sizeof(float)); - dhf_dst_i += src1_col_0*ne0 + row_low[id]; - CUDA_CHECK(cudaMemcpy2DAsync(dhf_dst_i, ne0*sizeof(float), dst_dd_i, row_diff*sizeof(float), - row_diff*sizeof(float), src1_ncols, kind, stream)); + dhf_dst_i += src1_col_0*ne0 + dev[id].row_low; +#if !defined(GGML_USE_HIPBLAS) + if (kind == cudaMemcpyDeviceToDevice) { + // cudaMemcpy2DAsync may fail with copies between vmm pools of different devices + cudaMemcpy3DPeerParms p = {}; + p.dstDevice = g_main_device; + p.dstPtr = make_cudaPitchedPtr(dhf_dst_i, ne0*sizeof(float), row_diff, src1_ncols); + p.srcDevice = id; + p.srcPtr = make_cudaPitchedPtr(dst_dd_i, row_diff*sizeof(float), row_diff, src1_ncols); + p.extent = make_cudaExtent(row_diff*sizeof(float), src1_ncols, 1); + CUDA_CHECK(cudaMemcpy3DPeerAsync(&p, stream)); + } else +#endif + { + CUDA_CHECK(cudaMemcpy2DAsync(dhf_dst_i, ne0*sizeof(float), + dst_dd_i, row_diff*sizeof(float), + row_diff*sizeof(float), src1_ncols, + kind, stream)); + } } else { float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3); GGML_ASSERT(dst->nb[1] == ne0*sizeof(float)); @@ -8104,35 +8285,14 @@ static void ggml_cuda_op_mul_mat( } } - for (int64_t id = 0; id < g_device_count; ++id) { - if ((!split && id != g_main_device) || row_low[id] == row_high[id]) { - continue; - } - CUDA_CHECK(ggml_cuda_set_device(id)); - - // free buffers again when done - if (src0_as[id] > 0) { - ggml_cuda_pool_free(src0_dd[id], src0_as[id]); - } - if (src1_asf[id] > 0) { - ggml_cuda_pool_free(src1_ddf[id], src1_asf[id]); - } - if (src1_asq[id] > 0) { - ggml_cuda_pool_free(src1_ddq[id], src1_asq[id]); - } - if (dst_as[id] > 0) { - ggml_cuda_pool_free(dst_dd[id], dst_as[id]); - } - } - // main device waits for all other devices to be finished if (split && g_device_count > 1) { int64_t is_max = (ne11 + MUL_MAT_SRC1_COL_STRIDE - 1) / MUL_MAT_SRC1_COL_STRIDE; is_max = is_max <= MAX_STREAMS ? is_max : MAX_STREAMS; - CUDA_CHECK(ggml_cuda_set_device(g_main_device)); - for (int64_t id = 0; id < g_device_count; ++id) { - if (row_low[id] == row_high[id]) { + ggml_cuda_set_device(g_main_device); + for (int id = 0; id < g_device_count; ++id) { + if (dev[id].row_low == dev[id].row_high) { continue; } for (int64_t is = 0; is < is_max; ++is) { @@ -8142,7 +8302,7 @@ static void ggml_cuda_op_mul_mat( } if (dst->backend == GGML_BACKEND_CPU) { - CUDA_CHECK(ggml_cuda_set_device(g_main_device)); + ggml_cuda_set_device(g_main_device); CUDA_CHECK(cudaDeviceSynchronize()); } } @@ -8252,7 +8412,7 @@ static void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tens const int64_t ne12 = src1->ne[2]; - CUDA_CHECK(ggml_cuda_set_device(g_main_device)); + ggml_cuda_set_device(g_main_device); cudaStream_t main_stream = g_cudaStreams[g_main_device][0]; ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; @@ -8284,7 +8444,7 @@ static void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor const int64_t ne12 = src1->ne[2]; - CUDA_CHECK(ggml_cuda_set_device(g_main_device)); + ggml_cuda_set_device(g_main_device); cudaStream_t main_stream = g_cudaStreams[g_main_device][0]; ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; @@ -8355,7 +8515,7 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const const int64_t ne1 = ggml_nelements(src1); const int64_t ne = ggml_nelements(dst); - CUDA_CHECK(ggml_cuda_set_device(g_main_device)); + ggml_cuda_set_device(g_main_device); cudaStream_t main_stream = g_cudaStreams[g_main_device][0]; CUBLAS_CHECK(cublasSetStream(g_cublas_handles[g_main_device], main_stream)); @@ -8374,14 +8534,11 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const 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); + cuda_pool_alloc src1_as_f16(ne1); + to_fp16_cuda(src1_ddf, src1_as_f16.get(), ne1, main_stream); - size_t dst_as = 0; - - half * dst_f16 = nullptr; - char * dst_t = nullptr; + cuda_pool_alloc dst_f16; + char * dst_t; cublasComputeType_t cu_compute_type = CUBLAS_COMPUTE_16F; cudaDataType_t cu_data_type = CUDA_R_16F; @@ -8400,8 +8557,7 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const const void * beta = &beta_f16; if (dst->op_params[0] == GGML_PREC_DEFAULT) { - dst_f16 = (half *) ggml_cuda_pool_malloc(ne * sizeof(half), &dst_as); - dst_t = (char *) dst_f16; + dst_t = (char *) dst_f16.alloc(ne); nbd2 /= sizeof(float) / sizeof(half); nbd3 /= sizeof(float) / sizeof(half); @@ -8448,9 +8604,9 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const CUBLAS_CHECK( cublasGemmStridedBatchedEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N, ne01, ne11, ne10, - alpha, (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, ( char *) dst_t, cu_data_type, ne01, dst->nb[2]/sizeof(float), // strideC + alpha, (const char *) src0_as_f16, CUDA_R_16F, nb01/sizeof(half), src0->nb[2]/sizeof(half), // strideA + (const char *) src1_as_f16.get(), CUDA_R_16F, nb11/sizeof(float), src1->nb[2]/sizeof(float), // strideB + beta, ( char *) dst_t, cu_data_type, ne01, dst->nb[2]/sizeof(float), // strideC ne12*ne13, cu_compute_type, CUBLAS_GEMM_DEFAULT_TENSOR_OP)); @@ -8458,19 +8614,13 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const // use cublasGemmBatchedEx const int ne23 = ne12*ne13; - const void ** ptrs_src = nullptr; - void ** ptrs_dst = nullptr; - - size_t ptrs_src_s = 0; - size_t ptrs_dst_s = 0; - - ptrs_src = (const void **) ggml_cuda_pool_malloc(2*ne23*sizeof(void *), &ptrs_src_s); - ptrs_dst = ( void **) ggml_cuda_pool_malloc(1*ne23*sizeof(void *), &ptrs_dst_s); + cuda_pool_alloc ptrs_src(2*ne23); + cuda_pool_alloc< void *> ptrs_dst(1*ne23); dim3 block_dims(ne13, ne12); k_compute_batched_ptrs<<<1, block_dims, 0, main_stream>>>( - src0_as_f16, src1_as_f16, dst_t, - ptrs_src, ptrs_dst, + src0_as_f16, src1_as_f16.get(), dst_t, + ptrs_src.get(), ptrs_dst.get(), ne12, ne13, ne23, nb02, nb03, @@ -8482,30 +8632,19 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const CUBLAS_CHECK( cublasGemmBatchedEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N, ne01, ne11, ne10, - alpha, (const void **) (ptrs_src + 0*ne23), CUDA_R_16F, nb01/sizeof(half), - (const void **) (ptrs_src + 1*ne23), CUDA_R_16F, nb11/sizeof(float), - beta, ( void **) (ptrs_dst + 0*ne23), cu_data_type, ne01, + alpha, (const void **) (ptrs_src.get() + 0*ne23), CUDA_R_16F, nb01/sizeof(half), + (const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F, nb11/sizeof(float), + beta, ( void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne01, ne23, cu_compute_type, CUBLAS_GEMM_DEFAULT_TENSOR_OP)); - - if (ptrs_src_s != 0) { - ggml_cuda_pool_free(ptrs_src, ptrs_src_s); - } - if (ptrs_dst_s != 0) { - ggml_cuda_pool_free(ptrs_dst, ptrs_dst_s); - } } #endif if (dst->op_params[0] == GGML_PREC_DEFAULT) { 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(dst_f16, dst_as); + to_fp32_cuda(dst_f16.get(), dst_ddf, ne, main_stream); } - - ggml_cuda_pool_free(src1_as_f16, src1_as); } static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -8517,9 +8656,9 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1 const bool split = src0->backend == GGML_BACKEND_GPU_SPLIT; 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)) { - min_compute_capability = g_compute_capabilities[id]; + for (int id = 0; id < g_device_count; ++id) { + if (min_compute_capability > g_device_caps[id].cc && g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) { + min_compute_capability = g_device_caps[id].cc; } } @@ -8660,7 +8799,7 @@ static void ggml_cuda_mul_mat_id_cublas(ggml_tensor * dst) { const int64_t ne1 = ggml_nelements(src1); const int64_t ne = ggml_nelements(dst); - CUDA_CHECK(ggml_cuda_set_device(g_main_device)); + ggml_cuda_set_device(g_main_device); cudaStream_t main_stream = g_cudaStreams[g_main_device][0]; CUBLAS_CHECK(cublasSetStream(g_cublas_handles[g_main_device], main_stream)); @@ -8778,7 +8917,7 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s std::vector ids_host(ggml_nbytes(ids)); - const cudaStream_t stream = g_cudaStreams[g_main_device][0]; + cudaStream_t stream = g_cudaStreams[g_main_device][0]; if (ids->backend == GGML_BACKEND_GPU) { const char * ids_dev = (const char *)((const ggml_tensor_extra_gpu *)ids->extra)->data_device[g_main_device]; @@ -8832,17 +8971,16 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s ggml_cuda_mul_mat(src0_row, &src1_row, &dst_row); } } else { - size_t as_src1, as_dst; - char * src1_contiguous = (char *) ggml_cuda_pool_malloc(sizeof(float)*ggml_nelements(src1), &as_src1); - char * dst_contiguous = (char *) ggml_cuda_pool_malloc(sizeof(float)*ggml_nelements(dst), &as_dst); + cuda_pool_alloc src1_contiguous(sizeof(float)*ggml_nelements(src1)); + cuda_pool_alloc dst_contiguous(sizeof(float)*ggml_nelements(dst)); - src1_row_extra.data_device[g_main_device] = src1_contiguous; - dst_row_extra.data_device[g_main_device] = dst_contiguous; + src1_row_extra.data_device[g_main_device] = src1_contiguous.get(); + dst_row_extra.data_device[g_main_device] = dst_contiguous.get(); const cudaMemcpyKind src1_kind = src1->backend == GGML_BACKEND_CPU ? cudaMemcpyHostToDevice : cudaMemcpyDeviceToDevice; const cudaMemcpyKind dst_kind = dst->backend == GGML_BACKEND_CPU ? - cudaMemcpyHostToDevice : cudaMemcpyDeviceToDevice; + cudaMemcpyDeviceToHost : cudaMemcpyDeviceToDevice; for (int32_t row_id = 0; row_id < n_as; ++row_id) { const struct ggml_tensor * src0_row = dst->src[row_id + 2]; @@ -8857,7 +8995,7 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s GGML_ASSERT(row_id >= 0 && row_id < n_as); - CUDA_CHECK(cudaMemcpyAsync(src1_contiguous + num_src1_rows*nb11, src1_original + i01*nb11, + CUDA_CHECK(cudaMemcpyAsync(src1_contiguous.get() + num_src1_rows*nb11, src1_original + i01*nb11, nb11, src1_kind, stream)); num_src1_rows++; } @@ -8889,14 +9027,11 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s GGML_ASSERT(row_id >= 0 && row_id < n_as); - CUDA_CHECK(cudaMemcpyAsync(dst_original + i01*nb1, dst_contiguous + num_src1_rows*nb1, + CUDA_CHECK(cudaMemcpyAsync(dst_original + i01*nb1, dst_contiguous.get() + num_src1_rows*nb1, nb1, dst_kind, stream)); num_src1_rows++; } } - - ggml_cuda_pool_free(src1_contiguous, as_src1); - ggml_cuda_pool_free(dst_contiguous, as_dst); } if (dst->backend == GGML_BACKEND_CPU) { @@ -8938,7 +9073,7 @@ static void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, gg const int64_t nb11 = src1->nb[1]; const int64_t nb12 = src1->nb[2]; - CUDA_CHECK(ggml_cuda_set_device(g_main_device)); + ggml_cuda_set_device(g_main_device); cudaStream_t main_stream = g_cudaStreams[g_main_device][0]; const ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; @@ -9028,7 +9163,7 @@ void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) { ggml_tensor_extra_gpu * extra = new struct ggml_tensor_extra_gpu; memset(extra, 0, sizeof(*extra)); - for (int64_t id = 0; id < g_device_count; ++id) { + for (int id = 0; id < g_device_count; ++id) { if (backend == GGML_BACKEND_GPU && id != g_main_device) { continue; } @@ -9099,15 +9234,14 @@ void ggml_cuda_free_data(struct ggml_tensor * tensor) { ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; - for (int64_t id = 0; id < g_device_count; ++id) { + for (int id = 0; id < g_device_count; ++id) { + ggml_cuda_set_device(id); if (extra->data_device[id] != nullptr) { - CUDA_CHECK(ggml_cuda_set_device(id)); CUDA_CHECK(cudaFree(extra->data_device[id])); } for (int64_t is = 0; is < MAX_STREAMS; ++is) { if (extra->events[id][is] != nullptr) { - CUDA_CHECK(ggml_cuda_set_device(id)); CUDA_CHECK(cudaEventDestroy(extra->events[id][is])); } } @@ -9161,7 +9295,7 @@ static void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scra force_inplace; const size_t size = ggml_nbytes(tensor); - CUDA_CHECK(ggml_cuda_set_device(g_main_device)); + ggml_cuda_set_device(g_main_device); if (inplace && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT)) { ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->src[0]->extra; char * src0_ddc = (char *) src0_extra->data_device[g_main_device]; @@ -9238,7 +9372,7 @@ void ggml_cuda_copy_to_device(struct ggml_tensor * tensor) { GGML_ASSERT(ggml_is_contiguous(tensor)); ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; - CUDA_CHECK(ggml_cuda_set_device(g_main_device)); + ggml_cuda_set_device(g_main_device); CUDA_CHECK(cudaMemcpy(extra->data_device[g_main_device], tensor->data, ggml_nbytes(tensor), cudaMemcpyHostToDevice)); } @@ -9662,12 +9796,16 @@ ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) { // host buffer type static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) { - CUDA_CHECK(cudaFreeHost(buffer->context)); + ggml_cuda_host_free(buffer->context); } static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { - void * ptr; - CUDA_CHECK(cudaMallocHost(&ptr, size)); + void * ptr = ggml_cuda_host_malloc(size); + + if (ptr == nullptr) { + // fallback to cpu buffer + return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size); + } // FIXME: this is a hack to avoid having to implement a new buffer type ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size); diff --git a/ggml-quants.c b/ggml-quants.c index 0e8163a16..a15a24048 100644 --- a/ggml-quants.c +++ b/ggml-quants.c @@ -3677,7 +3677,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri const uint8x16_t mins = vshrq_n_u8(mins_and_scales, 4); const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums); - const ggml_int16x8x2_t mins16 = {vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mins))), vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mins)))}; + const ggml_int16x8x2_t mins16 = {{vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mins))), vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mins)))}}; const int32x4_t s0 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[0]), vget_low_s16 (q8sums.val[0])), vmull_s16(vget_high_s16(mins16.val[0]), vget_high_s16(q8sums.val[0]))); const int32x4_t s1 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[1]), vget_low_s16 (q8sums.val[1])), @@ -6626,7 +6626,7 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums); const int8x16_t scales = vld1q_s8(scale); - const ggml_int16x8x2_t q6scales = {vmovl_s8(vget_low_s8(scales)), vmovl_s8(vget_high_s8(scales))}; + const ggml_int16x8x2_t q6scales = {{vmovl_s8(vget_low_s8(scales)), vmovl_s8(vget_high_s8(scales))}}; const int32x4_t prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[0]), vget_low_s16 (q6scales.val[0])), vmull_s16(vget_high_s16(q8sums.val[0]), vget_high_s16(q6scales.val[0]))), diff --git a/ggml.c b/ggml.c index f27920a2d..ed56e60a8 100644 --- a/ggml.c +++ b/ggml.c @@ -4041,7 +4041,6 @@ static struct ggml_tensor * ggml_group_norm_impl( result->op = GGML_OP_GROUP_NORM; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; // TODO: maybe store epsilon here? return result; } @@ -5541,7 +5540,6 @@ static struct ggml_tensor * ggml_upscale_impl( result->op_params[0] = scale_factor; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; return result; } @@ -5846,7 +5844,6 @@ struct ggml_tensor * ggml_get_rel_pos( result->op = GGML_OP_GET_REL_POS; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; return result; } @@ -10335,7 +10332,8 @@ static void ggml_compute_forward_scale_f32( } // scale factor - const float v = *(float *) dst->op_params; + float v; + memcpy(&v, dst->op_params, sizeof(float)); const int ith = params->ith; const int nth = params->nth; @@ -15152,7 +15150,8 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor { // necessary for llama if (src0->grad) { - const float s = ((float *) tensor->op_params)[0]; + float s; + memcpy(&s, tensor->op_params, sizeof(float)); src0->grad = ggml_add_or_set(ctx, @@ -15335,6 +15334,8 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor const int n_past = ((int32_t *) tensor->op_params)[0]; src0->grad = ggml_add_or_set(ctx, src0->grad, + /* ggml_diag_mask_inf_impl() shouldn't be here */ + /* ref: https://github.com/ggerganov/llama.cpp/pull/4203#discussion_r1412377992 */ ggml_diag_mask_zero_impl(ctx, tensor->grad, n_past, false), zero_table); } @@ -17452,9 +17453,9 @@ static void ggml_opt_acc_grad(int np, struct ggml_tensor * const ps[], float * g } // -// ADAM +// Using AdamW - ref: https://arxiv.org/pdf/1711.05101v3.pdf // -// ref: https://arxiv.org/pdf/1412.6980.pdf +// (Original Adam - ref: https://arxiv.org/pdf/1412.6980.pdf) // static enum ggml_opt_result ggml_opt_adam( @@ -19347,7 +19348,7 @@ void gguf_set_kv(struct gguf_context * ctx, struct gguf_context * src) { data[j] = ((struct gguf_str *)src->kv[i].value.arr.data)[j].data; } gguf_set_arr_str(ctx, src->kv[i].key.data, data, src->kv[i].value.arr.n); - free(data); + free((void *)data); } else if (src->kv[i].value.arr.type == GGUF_TYPE_ARRAY) { GGML_ASSERT(false && "nested arrays not supported"); } else { diff --git a/ggml.h b/ggml.h index 75918502b..67d6bc4f1 100644 --- a/ggml.h +++ b/ggml.h @@ -255,6 +255,8 @@ #define GGML_UNREACHABLE() GGML_ASSERT(!"statement should not be reached") #elif defined(__GNUC__) #define GGML_UNREACHABLE() __builtin_unreachable() +#elif defined(_MSC_VER) +#define GGML_UNREACHABLE() __assume(0) #else #define GGML_UNREACHABLE() ((void) 0) #endif @@ -484,7 +486,8 @@ extern "C" { enum ggml_log_level { GGML_LOG_LEVEL_ERROR = 2, GGML_LOG_LEVEL_WARN = 3, - GGML_LOG_LEVEL_INFO = 4 + GGML_LOG_LEVEL_INFO = 4, + GGML_LOG_LEVEL_DEBUG = 5 }; // ggml object diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 390dca049..4cd87cdda 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -96,6 +96,7 @@ class MODEL_ARCH(IntEnum): STABLELM = auto() QWEN = auto() PHI2 = auto() + PLAMO = auto() class MODEL_TENSOR(IntEnum): @@ -142,6 +143,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.STABLELM: "stablelm", MODEL_ARCH.QWEN: "qwen", MODEL_ARCH.PHI2: "phi2", + MODEL_ARCH.PLAMO: "plamo", } TENSOR_NAMES: dict[MODEL_TENSOR, str] = { @@ -349,6 +351,21 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN, MODEL_TENSOR.FFN_UP, ], + MODEL_ARCH.PLAMO: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], MODEL_ARCH.GPT2: [ # TODO ], diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 6fcbdbc1c..446c6b688 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -79,6 +79,7 @@ class TensorNameMap: "language_model.encoder.layers.{bid}.input_layernorm", # persimmon "model.layers.{bid}.ln1", # yi "transformer.h.{bid}.ln", # phi2 + "model.layers.layers.{bid}.norm", # plamo ), # Attention norm 2 @@ -99,26 +100,29 @@ class TensorNameMap: # Attention query MODEL_TENSOR.ATTN_Q: ( - "model.layers.{bid}.self_attn.q_proj", # llama-hf - "layers.{bid}.attention.wq", # llama-pth - "encoder.layer.{bid}.attention.self.query", # bert - "transformer.h.{bid}.attn.q_proj", # gpt-j + "model.layers.{bid}.self_attn.q_proj", # llama-hf + "layers.{bid}.attention.wq", # llama-pth + "encoder.layer.{bid}.attention.self.query", # bert + "transformer.h.{bid}.attn.q_proj", # gpt-j + "model.layers.layers.{bid}.self_attn.q_proj", # plamo ), # Attention key MODEL_TENSOR.ATTN_K: ( - "model.layers.{bid}.self_attn.k_proj", # llama-hf - "layers.{bid}.attention.wk", # llama-pth - "encoder.layer.{bid}.attention.self.key", # bert - "transformer.h.{bid}.attn.k_proj", # gpt-j + "model.layers.{bid}.self_attn.k_proj", # llama-hf + "layers.{bid}.attention.wk", # llama-pth + "encoder.layer.{bid}.attention.self.key", # bert + "transformer.h.{bid}.attn.k_proj", # gpt-j + "model.layers.layers.{bid}.self_attn.k_proj", # plamo ), # Attention value MODEL_TENSOR.ATTN_V: ( - "model.layers.{bid}.self_attn.v_proj", # llama-hf - "layers.{bid}.attention.wv", # llama-pth - "encoder.layer.{bid}.attention.self.value", # bert - "transformer.h.{bid}.attn.v_proj", # gpt-j + "model.layers.{bid}.self_attn.v_proj", # llama-hf + "layers.{bid}.attention.wv", # llama-pth + "encoder.layer.{bid}.attention.self.value", # bert + "transformer.h.{bid}.attn.v_proj", # gpt-j + "model.layers.layers.{bid}.self_attn.v_proj", # plamo ), # Attention output @@ -134,12 +138,14 @@ class TensorNameMap: "transformer.h.{bid}.attn.out_proj", # gpt-j "language_model.encoder.layers.{bid}.self_attention.dense", # persimmon "transformer.h.{bid}.mixer.out_proj", # phi2 + "model.layers.layers.{bid}.self_attn.o_proj", # plamo ), # Rotary embeddings MODEL_TENSOR.ATTN_ROT_EMBD: ( - "model.layers.{bid}.self_attn.rotary_emb.inv_freq", # llama-hf - "layers.{bid}.attention.inner_attention.rope.freqs", # llama-pth + "model.layers.{bid}.self_attn.rotary_emb.inv_freq", # llama-hf + "layers.{bid}.attention.inner_attention.rope.freqs", # llama-pth + "model.layers.layers.{bid}.self_attn.rotary_emb.inv_freq", # plamo ), # Feed-forward norm @@ -174,6 +180,7 @@ class TensorNameMap: "language_model.encoder.layers.{bid}.mlp.dense_h_to_4h", # persimmon "transformer.h.{bid}.mlp.w1", # qwen "transformer.h.{bid}.mlp.fc1", # phi2 + "model.layers.layers.{bid}.mlp.up_proj", # plamo ), MODEL_TENSOR.FFN_UP_EXP: ( @@ -186,6 +193,7 @@ class TensorNameMap: "model.layers.{bid}.mlp.gate_proj", # llama-hf refact "layers.{bid}.feed_forward.w1", # llama-pth "transformer.h.{bid}.mlp.w2", # qwen + "model.layers.layers.{bid}.mlp.gate_proj", # plamo ), MODEL_TENSOR.FFN_GATE_EXP: ( @@ -206,6 +214,7 @@ class TensorNameMap: "transformer.h.{bid}.mlp.fc_out", # gpt-j "language_model.encoder.layers.{bid}.mlp.dense_4h_to_h", # persimmon "transformer.h.{bid}.mlp.fc2", # phi2 + "model.layers.layers.{bid}.mlp.down_proj", # plamo ), MODEL_TENSOR.FFN_DOWN_EXP: ( diff --git a/llama.cpp b/llama.cpp index d6c192441..4aa59c4c0 100644 --- a/llama.cpp +++ b/llama.cpp @@ -198,6 +198,7 @@ enum llm_arch { LLM_ARCH_STABLELM, LLM_ARCH_QWEN, LLM_ARCH_PHI2, + LLM_ARCH_PLAMO, LLM_ARCH_UNKNOWN, }; @@ -216,6 +217,7 @@ static std::map LLM_ARCH_NAMES = { { LLM_ARCH_STABLELM, "stablelm" }, { LLM_ARCH_QWEN, "qwen" }, { LLM_ARCH_PHI2, "phi2" }, + { LLM_ARCH_PLAMO, "plamo" }, }; enum llm_kv { @@ -567,6 +569,24 @@ static std::map> LLM_TENSOR_NAMES = { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_PLAMO, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, { LLM_ARCH_UNKNOWN, @@ -778,7 +798,7 @@ struct llama_file { throw std::runtime_error(format("read error: %s", strerror(errno))); } if (ret != 1) { - throw std::runtime_error(std::string("unexpectedly reached end of file")); + throw std::runtime_error("unexpectedly reached end of file"); } } @@ -931,29 +951,29 @@ struct llama_mmap { #elif defined(_WIN32) static constexpr bool SUPPORTED = true; - llama_mmap(struct llama_file * file, bool prefetch = true, bool numa = false) { - (void) numa; + llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1, bool numa = false) { + GGML_UNUSED(numa); size = file->size; HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp)); HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL); - DWORD error = GetLastError(); if (hMapping == NULL) { + DWORD error = GetLastError(); throw std::runtime_error(format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str())); } addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0); - error = GetLastError(); + DWORD error = GetLastError(); CloseHandle(hMapping); if (addr == NULL) { throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str())); } - if (prefetch) { + if (prefetch > 0) { // PrefetchVirtualMemory is only present on Windows 8 and above, so we dynamically load it BOOL (WINAPI *pPrefetchVirtualMemory) (HANDLE, ULONG_PTR, PWIN32_MEMORY_RANGE_ENTRY, ULONG); HMODULE hKernel32 = GetModuleHandleW(L"kernel32.dll"); @@ -965,9 +985,9 @@ struct llama_mmap { // advise the kernel to preload the mapped memory WIN32_MEMORY_RANGE_ENTRY range; range.VirtualAddress = addr; - range.NumberOfBytes = (SIZE_T)size; + range.NumberOfBytes = (SIZE_T) std::min(size, prefetch); if (!pPrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) { - fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n", + LLAMA_LOG_WARN("warning: PrefetchVirtualMemory failed: %s\n", llama_format_win_err(GetLastError()).c_str()); } } @@ -982,26 +1002,26 @@ struct llama_mmap { ~llama_mmap() { if (!UnmapViewOfFile(addr)) { - fprintf(stderr, "warning: UnmapViewOfFile failed: %s\n", + LLAMA_LOG_WARN("warning: UnmapViewOfFile failed: %s\n", llama_format_win_err(GetLastError()).c_str()); } } #else static constexpr bool SUPPORTED = false; - llama_mmap(struct llama_file * file, bool prefetch = true, bool numa = false) { - (void) file; - (void) prefetch; - (void) numa; + llama_mmap(struct llama_file * file, size_t prefetch = -1, bool numa = false) { + GGML_UNUSED(file); + GGML_UNUSED(prefetch); + GGML_UNUSED(numa); - throw std::runtime_error(std::string("mmap not supported")); + throw std::runtime_error("mmap not supported"); } - void unmap(size_t offset, size_t len) { - (void) offset; - (void) len; + void unmap_fragment(size_t first, size_t last) { + GGML_UNUSED(first); + GGML_UNUSED(last); - throw std::runtime_error(std::string("mmap not supported")); + throw std::runtime_error("mmap not supported"); } #endif }; @@ -1177,21 +1197,27 @@ static std::string llama_token_to_piece(const struct llama_context * ctx, llama_ } static ggml_backend_buffer_type_t llama_default_buffer_type(int n_gpu_layers) { + ggml_backend_buffer_type_t buft = nullptr; + #ifdef GGML_USE_METAL if (n_gpu_layers > 0) { - return ggml_backend_metal_buffer_type(); + buft = ggml_backend_metal_buffer_type(); } #elif defined(GGML_USE_CUBLAS) && defined(LLAMA_GGML_BACKEND_CUDA_TEST) if (n_gpu_layers > 0) { - return ggml_backend_cuda_buffer_type(0); + buft = ggml_backend_cuda_buffer_type(0); } #elif defined(GGML_USE_CUBLAS) - return ggml_backend_cuda_host_buffer_type(); + buft = ggml_backend_cuda_host_buffer_type(); #elif defined(GGML_USE_CPU_HBM) - return ggml_backend_cpu_hbm_buffer_type(); + buft = ggml_backend_cpu_hbm_buffer_type(); #endif - return ggml_backend_cpu_buffer_type(); + if (buft == nullptr) { + buft = ggml_backend_cpu_buffer_type(); + } + + return buft; GGML_UNUSED(n_gpu_layers); } @@ -1275,7 +1301,7 @@ struct llama_hparams { if (this->rope_finetuned != other.rope_finetuned) return true; if (this->n_yarn_orig_ctx != other.n_yarn_orig_ctx) return true; - const float EPSILON = 1e-9; + const float EPSILON = 1e-9f; if (!is_float_close(this->f_norm_eps, other.f_norm_eps, EPSILON)) return true; if (!is_float_close(this->f_norm_rms_eps, other.f_norm_rms_eps, EPSILON)) return true; @@ -2372,7 +2398,8 @@ struct llama_model_loader { } } - void load_all_data(struct ggml_context * ctx, llama_progress_callback progress_callback, void * progress_callback_user_data, ggml_backend_buffer_t buf_mmap, llama_mlock * lmlock) const { + // Returns false if cancelled by progress_callback + bool load_all_data(struct ggml_context * ctx, llama_progress_callback progress_callback, void * progress_callback_user_data, ggml_backend_buffer_t buf_mmap, llama_mlock * lmlock) const { size_t size_data = 0; for (int i = 0; i < gguf_get_n_tensors(ctx_gguf); i++) { @@ -2404,7 +2431,9 @@ struct llama_model_loader { GGML_ASSERT(cur); // unused tensors should have been caught by load_data already if (progress_callback) { - progress_callback((float) size_done / size_data, progress_callback_user_data); + if (!progress_callback((float) size_done / size_data, progress_callback_user_data)) { + return false; + } } const size_t offs = file_offset(ggml_get_name(cur)); @@ -2466,8 +2495,11 @@ struct llama_model_loader { } if (progress_callback) { - progress_callback(1.0f, progress_callback_user_data); + // Even though the model is done loading, we still honor + // cancellation since we need to free allocations. + return progress_callback(1.0f, progress_callback_user_data); } + return true; } }; @@ -2737,6 +2769,15 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_PLAMO: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + case 40: model.type = e_model::MODEL_13B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; default: (void)0; } @@ -3044,7 +3085,8 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) { if (vocab.linefeed_id != -1) { LLAMA_LOG_INFO( "%s: LF token = %d '%s'\n", __func__, vocab.linefeed_id, vocab.id_to_token[vocab.linefeed_id].text.c_str() ); } } -static void llm_load_tensors( +// Returns false if cancelled by progress_callback +static bool llm_load_tensors( llama_model_loader & ml, llama_model & model, int n_gpu_layers, @@ -3617,6 +3659,51 @@ static void llm_load_tensors( layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend); } } break; + case LLM_ARCH_PLAMO: + { + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + + // output + { + ggml_backend_type backend_norm; + ggml_backend_type backend_output; + + if (n_gpu_layers > int(n_layer)) { + backend_norm = llama_backend_offload; + backend_output = llama_backend_offload_split; + } else { + backend_norm = GGML_BACKEND_CPU; + backend_output = GGML_BACKEND_CPU; + } + + model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); + model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); + } + + const uint32_t n_ff = hparams.n_ff; + + const int i_gpu_start = n_layer - n_gpu_layers; + + model.layers.resize(n_layer); + + for (uint32_t i = 0; i < n_layer; ++i) { + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); + + layer.wq = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, backend_split); + layer.wk = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, backend_split); + layer.wv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, 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.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); + } + } break; default: throw std::runtime_error("unknown architecture"); } @@ -3722,16 +3809,20 @@ static void llm_load_tensors( model.tensors_by_name.emplace_back(ggml_get_name(cur), cur); } - ml.load_all_data(ctx, progress_callback, progress_callback_user_data, buf_mmap, use_mlock ? &model.mlock_mmap : NULL); + if (!ml.load_all_data(ctx, progress_callback, progress_callback_user_data, buf_mmap, use_mlock ? &model.mlock_mmap : NULL)) { + return false; + } model.mapping = std::move(ml.mapping); // loading time will be recalculate after the first eval, so // we take page faults deferred by mmap() into consideration model.t_load_us = ggml_time_us() - model.t_start_us; + return true; } -static bool llama_model_load(const std::string & fname, llama_model & model, const llama_model_params & params) { +// Returns 0 on success, -1 on error, and -2 on cancellation via llama_progress_callback +static int llama_model_load(const std::string & fname, llama_model & model, const llama_model_params & params) { try { llama_model_loader ml(fname, params.use_mmap, params.kv_overrides); @@ -3749,19 +3840,21 @@ static bool llama_model_load(const std::string & fname, llama_model & model, con if (params.vocab_only) { LLAMA_LOG_INFO("%s: vocab only - skipping tensors\n", __func__); - return true; + return 0; } - llm_load_tensors( + if (!llm_load_tensors( ml, model, params.n_gpu_layers, params.main_gpu, params.tensor_split, params.use_mlock, params.progress_callback, params.progress_callback_user_data - ); + )) { + return -2; + } } catch (const std::exception & err) { LLAMA_LOG_ERROR("error loading model: %s\n", err.what()); - return false; + return -1; } - return true; + return 0; } // @@ -5536,6 +5629,109 @@ struct llm_build_context { return gf; } + + struct ggml_cgraph * build_plamo() { + struct ggml_cgraph * gf = ggml_new_graph(ctx0); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + cb(inpL, "inp_embd", -1); + + // 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_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); + + // shift the entire K-cache if needed + if (do_rope_shift) { + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, n_embd_head, freq_base, freq_scale, cb); + } + + for (int il = 0; il < n_layer; ++il) { + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); + + struct ggml_tensor * attention_norm = cur; + + // self-attention + { + // compute Q and K and RoPE them + 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); + + Qcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + n_embd_head, 2, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + n_embd_head, 2, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + cb(Kcur, "Kcur", il); + + llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); + + cur = llm_build_kqv(ctx0, model, hparams, kv_self, + model.layers[il].wo, NULL, + Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + cb(cur, "kqv_out", il); + } + struct ggml_tensor * sa_out = cur; + + cur = attention_norm; + + // feed-forward network + { + 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, sa_out); + cb(cur, "l_out", il); + + cur = ggml_add(ctx0, cur, inpL); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, NULL, + LLM_NORM_RMS, 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); + + return gf; + } }; // @@ -6046,6 +6242,10 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_phi2(); } break; + case LLM_ARCH_PLAMO: + { + result = llm.build_plamo(); + } break; default: GGML_ASSERT(false); } @@ -9141,11 +9341,18 @@ struct llama_model * llama_load_model_from_file( LLAMA_LOG_INFO("\n"); } } + return true; }; } - if (!llama_model_load(path_model, *model, params)) { - LLAMA_LOG_ERROR("%s: failed to load model\n", __func__); + int status = llama_model_load(path_model, *model, params); + GGML_ASSERT(status <= 0); + if (status < 0) { + if (status == -1) { + LLAMA_LOG_ERROR("%s: failed to load model\n", __func__); + } else if (status == -2) { + LLAMA_LOG_INFO("%s: cancelled model load\n", __func__); + } delete model; return nullptr; } @@ -9312,7 +9519,8 @@ struct llama_context * llama_new_context_with_model( ctx->alloc = ggml_allocr_new_from_buffer(ctx->buf_alloc); #if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST) if (model->n_gpu_layers > 0) { - ggml_cuda_set_scratch_size(alloc_size); + // the CPU buffer adds this padding in case the malloc buffer is not aligned, so we need to do the same for the GPU buffer, since we use the same offsets + ggml_cuda_set_scratch_size(alloc_size + 64); LLAMA_LOG_INFO("%s: VRAM scratch buffer: %.2f MiB\n", __func__, alloc_size / 1024.0 / 1024.0); // calculate total VRAM usage @@ -10274,7 +10482,7 @@ int llama_token_to_piece(const struct llama_model * model, llama_token token, ch std::string result = model->vocab.id_to_token[token].text; llama_unescape_whitespace(result); if (length < (int) result.length()) { - return -result.length(); + return -(int) result.length(); } memcpy(buf, result.c_str(), result.length()); return result.length(); @@ -10304,7 +10512,7 @@ int llama_token_to_piece(const struct llama_model * model, llama_token token, ch std::string result = model->vocab.id_to_token[token].text; result = llama_decode_text(result); if (length < (int) result.length()) { - return -result.length(); + return -(int) result.length(); } memcpy(buf, result.c_str(), result.length()); return result.length(); diff --git a/llama.h b/llama.h index 0be4b1337..af76bae2d 100644 --- a/llama.h +++ b/llama.h @@ -127,7 +127,7 @@ extern "C" { bool sorted; } llama_token_data_array; - typedef void (*llama_progress_callback)(float progress, void *ctx); + typedef bool (*llama_progress_callback)(float progress, void *ctx); // Input data for llama_decode // A llama_batch object can contain input about one or many sequences @@ -180,7 +180,9 @@ extern "C" { int32_t main_gpu; // the GPU that is used for scratch and small tensors const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES) - // called with a progress value between 0 and 1, pass NULL to disable + // Called with a progress value between 0.0 and 1.0. Pass NULL to disable. + // If the provided progress_callback returns true, model loading continues. + // If it returns false, model loading is immediately aborted. llama_progress_callback progress_callback; // context pointer passed to the progress callback diff --git a/tests/test-grad0.cpp b/tests/test-grad0.cpp index 14914def5..8ff76c891 100644 --- a/tests/test-grad0.cpp +++ b/tests/test-grad0.cpp @@ -883,9 +883,6 @@ int main(int argc, const char ** argv) { srand(seed); const int nargs = 1; - int64_t ne2[4]; - ne2[0] = 1; - for (int ndims = 1; ndims <= 2; ++ndims) { x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);