Merge branch 'ggerganov:master' into master

.gitignore

@@ -1,6 +1,7 @@
 *.o
 *.a
 *.so
+*.bin
 .DS_Store
 .build/
 .cache/
@@ -39,6 +40,7 @@ models-mnt
 /perplexity
 /embedding
 /train-text-from-scratch
+/convert-llama2c-to-ggml
 /simple
 /benchmark-matmult
 /vdot
@@ -46,6 +48,7 @@ models-mnt
 /Pipfile
 /embd-input-test
 /libllama.so
+/llama-bench
 build-info.h
 arm_neon.h
 compile_commands.json
@@ -68,6 +71,7 @@ poetry.lock
 poetry.toml
 
 # Test binaries
+tests/test-grammar-parser
 tests/test-double-float
 tests/test-grad0
 tests/test-opt

CMakeLists.txt

@@ -69,7 +69,6 @@ option(LLAMA_BLAS                            "llama: use BLAS"
 set(LLAMA_BLAS_VENDOR "Generic" CACHE STRING "llama: BLAS library vendor")
 option(LLAMA_CUBLAS                          "llama: use CUDA"                                  OFF)
 #option(LLAMA_CUDA_CUBLAS                     "llama: use cuBLAS for prompt processing"          OFF)
-set(LLAMA_CUDA_MMQ_Y       "64" CACHE STRING "llama: y tile size for mmq CUDA kernels")
 option(LLAMA_CUDA_FORCE_DMMV                 "llama: use dmmv instead of mmvq CUDA kernels"     OFF)
 set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kernels")
 set(LLAMA_CUDA_MMV_Y        "1" CACHE STRING "llama: y block size for mmv CUDA kernels")
@@ -256,7 +255,6 @@ if (LLAMA_CUBLAS)
 #        if (LLAMA_CUDA_CUBLAS)
 #            add_compile_definitions(GGML_CUDA_CUBLAS)
 #        endif()
-        add_compile_definitions(GGML_CUDA_MMQ_Y=${LLAMA_CUDA_MMQ_Y})
         if (LLAMA_CUDA_FORCE_DMMV)
             add_compile_definitions(GGML_CUDA_FORCE_DMMV)
         endif()
@@ -298,7 +296,6 @@ if (LLAMA_METAL)
     find_library(FOUNDATION_LIBRARY         Foundation              REQUIRED)
     find_library(METAL_FRAMEWORK            Metal                   REQUIRED)
     find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)
-    find_library(METALPERFORMANCE_FRAMEWORK MetalPerformanceShaders REQUIRED)
 
     set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h)
 
@@ -315,7 +312,6 @@ if (LLAMA_METAL)
         ${FOUNDATION_LIBRARY}
         ${METAL_FRAMEWORK}
         ${METALKIT_FRAMEWORK}
-        ${METALPERFORMANCE_FRAMEWORK}
         )
 endif()
 
@@ -573,6 +569,16 @@ install(
         WORLD_READ
         WORLD_EXECUTE
     DESTINATION ${CMAKE_INSTALL_BINDIR})
+if (LLAMA_METAL)
+    install(
+        FILES ggml-metal.metal
+        PERMISSIONS
+            OWNER_READ
+            OWNER_WRITE
+            GROUP_READ
+            WORLD_READ
+        DESTINATION ${CMAKE_INSTALL_BINDIR})
+endif()
 
 #
 # programs, examples and tests

Makefile

@@ -1,8 +1,8 @@
 # Define the default target now so that it is always the first target
-BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch simple server embd-input-test
+BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple server embd-input-test llama-bench
 
 # Binaries only useful for tests
-TEST_TARGETS = tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0
+TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0
 
 default: $(BUILD_TARGETS)
 
@@ -283,7 +283,7 @@ endif # LLAMA_CLBLAST
 ifdef LLAMA_METAL
 	CFLAGS   += -DGGML_USE_METAL -DGGML_METAL_NDEBUG
 	CXXFLAGS += -DGGML_USE_METAL
-	LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders
+	LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit
 	OBJS     += ggml-metal.o
 endif # LLAMA_METAL
 
@@ -345,7 +345,7 @@ libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
 clean:
-	rm -vf *.o *.so *.dll main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state server simple vdot train-text-from-scratch embd-input-test build-info.h $(TEST_TARGETS)
+	rm -vf *.o *.so *.dll main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state server simple vdot train-text-from-scratch convert-llama2c-to-ggml embd-input-test llama-bench build-info.h $(TEST_TARGETS)
 
 #
 # Examples
@@ -388,6 +388,12 @@ embd-input-test: $(LIB_PRE)embdinput$(DSO_EXT) examples/embd-input/embd-input-te
 train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp    build-info.h ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
+convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp    build-info.h ggml.o llama.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+llama-bench: examples/llama-bench/llama-bench.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 build-info.h: $(wildcard .git/index) scripts/build-info.sh
 	@sh scripts/build-info.sh > $@.tmp
 	@if ! cmp -s $@.tmp $@; then \
@@ -409,6 +415,12 @@ benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o
 vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
+tests/test-llama-grammar: tests/test-llama-grammar.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+
+tests/test-grammar-parser: tests/test-grammar-parser.cpp examples/grammar-parser.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+
 tests/test-double-float: tests/test-double-float.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
 

README.md

@@ -9,13 +9,13 @@
 
 Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
-**Hot topics:**
+### 🚧 Incoming breaking change + refactoring:
 
-- Simple web chat example: https://github.com/ggerganov/llama.cpp/pull/1998
+See PR https://github.com/ggerganov/llama.cpp/pull/2398 for more info.
-- k-quants now support super-block size of 64: https://github.com/ggerganov/llama.cpp/pull/2001
+
-- New roadmap: https://github.com/users/ggerganov/projects/7
+To devs: avoid making big changes to `llama.h` / `llama.cpp` until merged
-- Azure CI brainstorming: https://github.com/ggerganov/llama.cpp/discussions/1985
+
-- p1 : LLM-based code completion engine at the edge : https://github.com/ggml-org/p1/discussions/1
+----
 
 <details>
   <summary>Table of Contents</summary>
@@ -96,8 +96,10 @@ as the main playground for developing new features for the [ggml](https://github
 - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
 - Node.js: [hlhr202/llama-node](https://github.com/hlhr202/llama-node)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
+- Rust: [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
 - C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp)
 - Scala 3: [donderom/llm4s](https://github.com/donderom/llm4s)
+- Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj)
 
 **UI:**
 
@@ -238,12 +240,17 @@ In order to build llama.cpp you have three different options.
     cmake --build . --config Release
     ```
 
-- Using `Zig`:
+- Using `Zig` (version 0.11 or later):
+
+    Building for optimization levels and CPU features can be accomplished using standard build arguments, for example AVX2, FMA, F16C,
+    it's also possible to cross compile for other operating systems and architectures:
 
     ```bash
-    zig build -Doptimize=ReleaseFast
+    zig build -Doptimize=ReleaseFast -Dtarget=x86_64-windows-gnu -Dcpu=x86_64+avx2+fma+f16c
     ```
 
+    The `zig targets` command will give you valid options to use.
+
 -   Using `gmake` (FreeBSD):
 
     1. Install and activate [DRM in FreeBSD](https://wiki.freebsd.org/Graphics)
@@ -408,7 +415,7 @@ Building the program with BLAS support may lead to some performance improvements
   |-------------------------|------------------------|---------|-------------|
   | LLAMA_CUDA_FORCE_DMMV   | Boolean                |   false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
   | LLAMA_CUDA_DMMV_X       | Positive integer >= 32 |      32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
-  | LLAMA_CUDA_MMV_Y        | Positive integer       |       1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. |
+  | LLAMA_CUDA_MMV_Y        | Positive integer       |       1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. |
   | LLAMA_CUDA_F16          | Boolean                |   false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. |
   | LLAMA_CUDA_KQUANTS_ITER | 1 or 2                 |       2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
 

build.zig

@@ -1,5 +1,6 @@
 // Compatible with Zig Version 0.11.0
 const std = @import("std");
+const ArrayList = std.ArrayList;
 const Compile = std.Build.Step.Compile;
 const ConfigHeader = std.Build.Step.ConfigHeader;
 const Mode = std.builtin.Mode;
@@ -10,11 +11,31 @@ const Maker = struct {
     target: CrossTarget,
     optimize: Mode,
     config_header: *ConfigHeader,
+    enable_lto: bool,
 
-    const cflags = .{"-std=c11"};
+    include_dirs: ArrayList([]const u8),
-    const cxxflags = .{"-std=c++11"};
+    cflags: ArrayList([]const u8),
+    cxxflags: ArrayList([]const u8),
+    objs: ArrayList(*Compile),
 
-    fn init(builder: *std.build.Builder) Maker {
+    fn addInclude(m: *Maker, dir: []const u8) !void {
+        try m.include_dirs.append(dir);
+    }
+    fn addProjectInclude(m: *Maker, path: []const []const u8) !void {
+        try m.addInclude(try m.builder.build_root.join(m.builder.allocator, path));
+    }
+    fn addCFlag(m: *Maker, flag: []const u8) !void {
+        try m.cflags.append(flag);
+    }
+    fn addCxxFlag(m: *Maker, flag: []const u8) !void {
+        try m.cxxflags.append(flag);
+    }
+    fn addFlag(m: *Maker, flag: []const u8) !void {
+        try m.addCFlag(flag);
+        try m.addCxxFlag(flag);
+    }
+
+    fn init(builder: *std.build.Builder) !Maker {
         const commit_hash = @embedFile(".git/refs/heads/master");
         const config_header = builder.addConfigHeader(
             .{ .style = .blank, .include_path = "build-info.h" },
@@ -23,58 +44,71 @@ const Maker = struct {
                 .BUILD_COMMIT = commit_hash[0 .. commit_hash.len - 1], // omit newline
             },
         );
-        return Maker{
+        var m = Maker{
             .builder = builder,
             .target = builder.standardTargetOptions(.{}),
             .optimize = builder.standardOptimizeOption(.{}),
             .config_header = config_header,
+            .enable_lto = false,
+            .include_dirs = ArrayList([]const u8).init(builder.allocator),
+            .cflags = ArrayList([]const u8).init(builder.allocator),
+            .cxxflags = ArrayList([]const u8).init(builder.allocator),
+            .objs = ArrayList(*Compile).init(builder.allocator),
         };
+        try m.addCFlag("-std=c11");
+        try m.addCxxFlag("-std=c++11");
+        try m.addProjectInclude(&.{});
+        try m.addProjectInclude(&.{"examples"});
+        return m;
     }
 
     fn obj(m: *const Maker, name: []const u8, src: []const u8) *Compile {
         const o = m.builder.addObject(.{ .name = name, .target = m.target, .optimize = m.optimize });
         if (std.mem.endsWith(u8, src, ".c")) {
-            o.addCSourceFiles(&.{src}, &cflags);
+            o.addCSourceFiles(&.{src}, m.cflags.items);
             o.linkLibC();
         } else {
-            o.addCSourceFiles(&.{src}, &cxxflags);
+            o.addCSourceFiles(&.{src}, m.cxxflags.items);
             o.linkLibCpp();
         }
-        o.addIncludePath(.{ .path = "." });
+        for (m.include_dirs.items) |i| o.addIncludePath(.{ .path = i });
-        o.addIncludePath(.{ .path = "./examples" });
+        o.want_lto = m.enable_lto;
         return o;
     }
 
     fn exe(m: *const Maker, name: []const u8, src: []const u8, deps: []const *Compile) *Compile {
         const e = m.builder.addExecutable(.{ .name = name, .target = m.target, .optimize = m.optimize });
-        e.addIncludePath(.{ .path = "." });
+        e.addCSourceFiles(&.{src}, m.cxxflags.items);
-        e.addIncludePath(.{ .path = "./examples" });
-        e.addCSourceFiles(&.{src}, &cxxflags);
         for (deps) |d| e.addObject(d);
+        for (m.objs.items) |o| e.addObject(o);
+        for (m.include_dirs.items) |i| e.addIncludePath(.{ .path = i });
         e.linkLibC();
         e.linkLibCpp();
         e.addConfigHeader(m.config_header);
         m.builder.installArtifact(e);
-
+        e.want_lto = m.enable_lto;
-        // Currently a bug is preventing correct linking for optimized builds for Windows:
-        // https://github.com/ziglang/zig/issues/15958
-        if (e.target.isWindows()) {
-            e.want_lto = false;
-        }
         return e;
     }
 };
 
-pub fn build(b: *std.build.Builder) void {
+pub fn build(b: *std.build.Builder) !void {
-    const make = Maker.init(b);
+    var make = try Maker.init(b);
+    make.enable_lto = b.option(bool, "lto", "Enable LTO optimization, (default: false)") orelse false;
+
+    if (b.option(bool, "k-quants", "Enable K-quants, (default: true)") orelse true) {
+        try make.addFlag("-DGGML_USE_K_QUANTS");
+        const k_quants = make.obj("k_quants", "k_quants.c");
+        try make.objs.append(k_quants);
+    }
 
     const ggml = make.obj("ggml", "ggml.c");
     const ggml_alloc = make.obj("ggml-alloc", "ggml-alloc.c");
     const llama = make.obj("llama", "llama.cpp");
     const common = make.obj("common", "examples/common.cpp");
+    const console = make.obj("common", "examples/console.cpp");
     const grammar_parser = make.obj("grammar-parser", "examples/grammar-parser.cpp");
 
-    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, llama, common, grammar_parser });
+    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, llama, common, console, grammar_parser });
     _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, llama });
     _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, llama, common });
     _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, llama, common });

examples/CMakeLists.txt

@@ -42,8 +42,10 @@ else()
     add_subdirectory(benchmark)
     add_subdirectory(baby-llama)
     add_subdirectory(train-text-from-scratch)
+    add_subdirectory(convert-llama2c-to-ggml)
     add_subdirectory(simple)
     add_subdirectory(embd-input)
+    add_subdirectory(llama-bench)
     if (LLAMA_METAL)
         add_subdirectory(metal)
     endif()

examples/common.cpp

@@ -283,6 +283,21 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.cfg_negative_prompt = argv[i];
+        } else if (arg == "--cfg-negative-prompt-file") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            std::ifstream file(argv[i]);
+            if (!file) {
+                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+                invalid_param = true;
+                break;
+            }
+            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.cfg_negative_prompt));
+            if (params.cfg_negative_prompt.back() == '\n') {
+                params.cfg_negative_prompt.pop_back();
+            }
         } else if (arg == "--cfg-scale") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -578,8 +593,10 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stdout, "                        or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'\n");
     fprintf(stdout, "  --grammar GRAMMAR     BNF-like grammar to constrain generations (see samples in grammars/ dir)\n");
     fprintf(stdout, "  --grammar-file FNAME  file to read grammar from\n");
-    fprintf(stdout, "  --cfg-negative-prompt PROMPT \n");
+    fprintf(stdout, "  --cfg-negative-prompt PROMPT\n");
     fprintf(stdout, "                        negative prompt to use for guidance. (default: empty)\n");
+    fprintf(stdout, "  --cfg-negative-prompt-file FNAME\n");
+    fprintf(stdout, "                        negative prompt file to use for guidance. (default: empty)\n");
     fprintf(stdout, "  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
     fprintf(stdout, "  --rope-scale N        RoPE context linear scaling factor, inverse of --rope-freq-scale (default: %g)\n", 1.0f/params.rope_freq_scale);
     fprintf(stdout, "  --rope-freq-base N    RoPE base frequency, used by NTK-aware scaling (default: %.1f)\n", params.rope_freq_base);

examples/convert-llama2c-to-ggml/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET convert-llama2c-to-ggml)
+add_executable(${TARGET} convert-llama2c-to-ggml.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/convert-llama2c-to-ggml/README.md

@@ -0,0 +1,26 @@
+## Convert llama2.c model to ggml
+
+This example reads weights from project [llama2.c](https://github.com/karpathy/llama2.c) and saves them in ggml compatible format. The vocab that is available in `models/ggml-vocab.bin` is used by default.
+
+To convert the model first download the models from the [llma2.c](https://github.com/karpathy/llama2.c) repository:
+
+`$ make -j`
+
+After successful compilation, following usage options are available:
+```
+usage: ./convert-llama2c-to-ggml [options]
+
+options:
+  -h, --help                       show this help message and exit
+  --copy-vocab-from-model FNAME    model path from which to copy vocab (default 'models/ggml-vocab.bin')
+  --llama2c-model FNAME            [REQUIRED] model path from which to load Karpathy's llama2.c model
+  --llama2c-output-model FNAME     model path to save the converted llama2.c model (default ak_llama_model.bin')
+```
+
+An example command is as follows:
+
+`$ ./convert-llama2c-to-ggml --copy-vocab-from-model <ggml-vocab.bin> --llama2c-model <llama2.c model path> --llama2c-output-model <ggml output model path>`
+
+Now you can use the model with command like:
+
+`$ ./main -m <ggml output model path> -p "One day, Lily met a Shoggoth" -n 500 -c 256 -eps 1e-5`

examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp

@@ -0,0 +1,825 @@
+#include "ggml.h"
+#include "llama.h"
+#include <unordered_map>
+#include <vector>
+#include <cassert>
+#include <climits>
+#include <cstring>
+#include <cstdarg>
+#include <ctime>
+#include <random>
+#include <stdexcept>
+#include <algorithm>
+#include <string>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+//////////////////////////////////////// llama2.c model structs and functions to load models, alloc memory etc.
+typedef struct {
+    int dim; // transformer dimension
+    int hidden_dim; // for ffn layers
+    int n_layers; // number of layers
+    int n_heads; // number of query heads
+    int n_kv_heads; // number of key/value heads (can be < query heads because of multiquery)
+    int vocab_size; // vocabulary size, usually 256 (byte-level)
+    int seq_len; // max sequence length
+} Config;
+
+typedef struct {
+    // token embedding table
+    float* token_embedding_table;    // (vocab_size, dim)
+    // weights for rmsnorms
+    float* rms_att_weight; // (layer, dim) rmsnorm weights
+    float* rms_ffn_weight; // (layer, dim)
+    // weights for matmuls
+    float* wq; // (layer, dim, dim)
+    float* wk; // (layer, dim, dim)
+    float* wv; // (layer, dim, dim)
+    float* wo; // (layer, dim, dim)
+    // weights for ffn
+    float* w1; // (layer, hidden_dim, dim)
+    float* w2; // (layer, dim, hidden_dim)
+    float* w3; // (layer, hidden_dim, dim)
+    // final rmsnorm
+    float* rms_final_weight; // (dim,)
+    // freq_cis for RoPE relatively positional embeddings
+    // float* freq_cis_real; // (seq_len, dim/2)
+    // float* freq_cis_imag; // (seq_len, dim/2)
+    // (optional) classifier weights for the logits, on the last layer
+    //float* wcls;
+} TransformerWeights;
+
+void malloc_weights(TransformerWeights* w, Config* p) {
+    // we calloc instead of malloc to keep valgrind happy
+    w->token_embedding_table = new float[p->vocab_size * p->dim]();
+    printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
+
+    w->rms_att_weight = new float[p->n_layers * p->dim]();
+    printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->rms_att_weight\n",__func__,p->n_layers, p->dim, p->n_layers * p->dim);
+
+    w->rms_ffn_weight = new float[p->n_layers * p->dim]();
+    printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->rms_ffn_weight\n",__func__,p->n_layers , p->dim, p->n_layers * p->dim);
+
+    w->wq = new float[p->n_layers * p->dim * p->dim]();
+    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wq\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+
+    w->wk = new float[p->n_layers * p->dim * p->dim]();
+    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wk\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+
+    w->wv = new float[p->n_layers * p->dim * p->dim]();
+    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wv\n",__func__, p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+
+    w->wo = new float[p->n_layers * p->dim * p->dim]();
+    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wo\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+
+    w->w1 = new float[p->n_layers * p->hidden_dim * p->dim]();
+    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->w1\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
+
+    w->w2 = new float[p->n_layers * p->hidden_dim * p->dim]();
+    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->w2\n",__func__,p->n_layers, p->dim, p->hidden_dim, p->n_layers * p->hidden_dim * p->dim);
+
+    w->w3 = new float[p->n_layers * p->hidden_dim * p->dim]();
+    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->w3\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
+
+    w->rms_final_weight = new float[p->dim]();
+    printf("[%s:AK] Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
+}
+
+int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f) {
+    if (fread(w->token_embedding_table, sizeof(float), p->vocab_size * p->dim, f) != static_cast<size_t>(p->vocab_size * p->dim)) return 1;
+    if (fread(w->rms_att_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim)) return 1;
+    if (fread(w->wq, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
+    if (fread(w->wk, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
+    if (fread(w->wv, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
+    if (fread(w->wo, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
+    if (fread(w->rms_ffn_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim)) return 1;
+    if (fread(w->w1, sizeof(float), p->n_layers * p->dim * p->hidden_dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->hidden_dim)) return 1;
+    if (fread(w->w2, sizeof(float), p->n_layers * p->hidden_dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->hidden_dim * p->dim)) return 1;
+    if (fread(w->w3, sizeof(float), p->n_layers * p->dim * p->hidden_dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->hidden_dim)) return 1;
+    if (fread(w->rms_final_weight, sizeof(float), p->dim, f) != static_cast<size_t>(p->dim)) return 1;
+    return 0;
+}
+
+void free_weights(TransformerWeights* w) {
+    delete w->token_embedding_table;
+    delete w->rms_att_weight;
+    delete w->rms_ffn_weight;
+    delete w->wq;
+    delete w->wk;
+    delete w->wv;
+    delete w->wo;
+    delete w->w1;
+    delete w->w2;
+    delete w->w3;
+    delete w->rms_final_weight;
+}
+
+void print_sample_weights(TransformerWeights *w){
+    printf("----- Quick print of first of the weight vales of all the variables\n");
+    printf("%f\n", w->token_embedding_table[0]);
+    printf("%f\n", w->rms_att_weight[0]);
+    printf("%f\n", w->rms_ffn_weight[0]);
+
+    printf("%f\n", w->wq[0]);
+    printf("%f\n", w->wk[0]);
+    printf("%f\n", w->wv[0]);
+    printf("%f\n", w->wo[0]);
+    printf("%f\n", w->w1[0]);
+    printf("%f\n", w->w2[0]);
+    printf("%f\n", w->w3[0]);
+    printf("%f\n", w->rms_att_weight[0]);
+}
+////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+//////////////////////////////////////// ggml structs and functions required to load models, configs and save the model.
+
+struct llama_vocab {
+    using id    = int32_t;
+    using token = std::string;
+
+    struct token_score {
+        token tok;
+        float score;
+    };
+
+    std::unordered_map<token, id> token_to_id;
+    std::vector<token_score> id_to_token;
+};
+
+struct my_llama_hparams {
+    uint32_t n_vocab = 32000;
+    uint32_t n_ctx   = 512;   // this is provided as user input?
+    uint32_t n_embd  = 4096;
+    uint32_t n_mult  = 4;
+    uint32_t n_head  = 32;
+    uint32_t n_layer = 32;
+    uint32_t n_rot   = 64;
+    bool operator!=(const my_llama_hparams& other) const {
+        return memcmp(this, &other, sizeof(my_llama_hparams));
+    }
+};
+
+struct my_llama_layer {
+    // normalization
+    struct ggml_tensor * attention_norm;
+
+    // attention
+    struct ggml_tensor * wq;
+    struct ggml_tensor * wk;
+    struct ggml_tensor * wv;
+    struct ggml_tensor * wo;
+
+    // normalization
+    struct ggml_tensor * ffn_norm;
+
+    // ff
+    struct ggml_tensor * w1;
+    struct ggml_tensor * w2;
+    struct ggml_tensor * w3;
+};
+
+struct my_llama_model {
+    struct ggml_context * ctx = NULL;
+
+    my_llama_hparams hparams;
+
+    struct ggml_tensor * tok_embeddings;
+
+    struct ggml_tensor * norm;
+    struct ggml_tensor * output;
+
+    std::vector<my_llama_layer> layers;
+
+    uint32_t train_its = 0;
+    uint32_t train_samples = 0;
+    uint32_t train_tokens = 0;
+};
+
+struct train_params {
+    const char * fn_vocab_model;
+    const char * fn_llama2c_model;
+    const char * fn_llama2c_output_model;
+    const char * fn_train_data;
+    const char * fn_checkpoint_in;
+    const char * fn_checkpoint_out;
+    const char * fn_model_out;
+
+    uint32_t seed;
+
+    int n_ctx;
+    int n_embd;
+    int n_mult;
+    int n_head;
+    int n_layer;
+    int n_rotmax;
+
+    int n_threads;
+    int n_batch;
+    int n_examples;
+    int n_predict;
+
+    int print_info_interval;
+    int print_details_interval;
+
+    bool samples_start_after_nl;
+    bool use_adam;
+    bool use_flash;
+    bool use_scratch;
+
+    // only adam
+    int   warmup;
+    int   cos_decay_steps;
+    float cos_decay_restart;
+    float cos_decay_alpha;
+
+    int   lbfgs_n_iter;
+    int   adam_n_iter;
+    float adam_alpha;
+    float adam_decay;
+
+    int mem_model_gb;
+    int mem_compute_gb;
+    int mem_compute0_gb;
+    int mem_compute1_gb;
+};
+
+uint32_t get_n_ff(const struct my_llama_hparams* hparams) {
+    const uint32_t n_ff = ((2*(4*hparams->n_embd)/3 + hparams->n_mult - 1)/hparams->n_mult)*hparams->n_mult;
+    return n_ff;
+}
+
+void print_params(struct my_llama_hparams * params) {
+    printf("%s: n_vocab: %d\n", __func__, params->n_vocab);
+    printf("%s: n_ctx:   %d\n", __func__, params->n_ctx);
+    printf("%s: n_embd:  %d\n", __func__, params->n_embd);
+    printf("%s: n_mult:  %d\n", __func__, params->n_mult);
+    printf("%s: n_head:  %d\n", __func__, params->n_head);
+    printf("%s: n_ff:    %d\n", __func__, get_n_ff(params));
+    printf("%s: n_layer: %d\n", __func__, params->n_layer);
+    printf("%s: n_rot:   %d\n", __func__, params->n_rot);
+}
+
+void init_model(struct my_llama_model * model) {
+    const auto & hparams = model->hparams;
+
+    const uint32_t n_embd  = hparams.n_embd;
+    const uint32_t n_layer = hparams.n_layer;
+    const uint32_t n_vocab = hparams.n_vocab;
+
+    const uint32_t n_ff = get_n_ff(&hparams);
+    struct ggml_context * ctx = model->ctx;
+
+    model->train_its = 0;
+    model->train_samples = 0;
+    model->train_tokens = 0;
+
+    model->tok_embeddings = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
+    printf("[%s:GG] Allocating [%d] x [%d] = [%d] float space for model->tok_embeddings\n",__func__,n_embd , n_vocab, n_embd * n_vocab);
+
+    model->norm           = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+    printf("[%s:GG] Allocating [%d] float space for model->norm\n",__func__,n_embd);
+
+    model->output         = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for model->output\n",__func__,n_embd, n_vocab, n_embd * n_vocab);
+
+    // printing the per-layer allocations here so we dont print in the for loop.
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wq for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wk for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wv for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wo for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+
+    printf("[%s:GG] Allocating [%d] float space for layer.ffn_norm for [%d] layers\n",__func__,n_embd, n_layer);
+
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.w1 for [%d] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.w2 for [%d] layers\n",__func__, n_embd, n_ff, n_ff * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.w3 for [%d] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
+
+    ggml_set_name(model->tok_embeddings, "tok_embeddings.weight");
+    ggml_set_name(model->norm,           "norm.weight");
+    ggml_set_name(model->output,         "output.weight");
+
+    model->layers.resize(n_layer);
+    for (uint32_t i = 0; i < n_layer; ++i) {
+        auto & layer = model->layers[i];
+
+        std::string layers_i = "layers." + std::to_string(i);
+
+        layer.attention_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        layer.wq = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
+        layer.wk = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
+        layer.wv = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
+        layer.wo = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
+
+        layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        layer.w1 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ff);
+        layer.w2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_ff, n_embd);
+        layer.w3 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ff);
+
+        ggml_set_name(layer.attention_norm, (layers_i + ".attention_norm.weight").c_str());
+
+        ggml_set_name(layer.wq, (layers_i + ".attention.wq.weight").c_str());
+        ggml_set_name(layer.wk, (layers_i + ".attention.wk.weight").c_str());
+        ggml_set_name(layer.wv, (layers_i + ".attention.wv.weight").c_str());
+        ggml_set_name(layer.wo, (layers_i + ".attention.wo.weight").c_str());
+
+        ggml_set_name(layer.ffn_norm, (layers_i + ".ffn_norm.weight").c_str());
+
+        ggml_format_name(layer.w1, "%s.feed_forward.w1.weight", layers_i.c_str());
+        ggml_format_name(layer.w2, "%s.feed_forward.w2.weight", layers_i.c_str());
+        ggml_format_name(layer.w3, "%s.feed_forward.w3.weight", layers_i.c_str());
+    }
+}
+
+float get_f32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
+    float * ptr = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
+    return *ptr;
+}
+
+int32_t get_i32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
+    int32_t * ptr = (int32_t *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
+    return *ptr;
+}
+
+void print_row(struct ggml_tensor * probs, int i) {
+    for (int k = 0; k < probs->ne[0]; ++k) {
+        float p = get_f32_2d(probs, k, i);
+        printf(" %f", p);
+    }
+    printf("\n");
+}
+
+void print_matrix(struct ggml_tensor * probs) {
+    assert(probs->n_dims == 2);
+    for (int i = 0; i < probs->ne[1]; ++i) {
+        for (int k = 0; k < probs->ne[0]; ++k) {
+            float p = get_f32_2d(probs, k, i);
+            printf(" %.2f", p);
+        }
+        printf("\n");
+    }
+}
+
+#ifdef __GNUC__
+#ifdef __MINGW32__
+__attribute__((format(gnu_printf, 1, 2)))
+#else
+__attribute__((format(printf, 1, 2)))
+#endif
+#endif
+static std::string format(const char * fmt, ...) {
+    va_list ap, ap2;
+    va_start(ap, fmt);
+    va_copy(ap2, ap);
+    int size = vsnprintf(NULL, 0, fmt, ap);
+    GGML_ASSERT(size >= 0 && size < INT_MAX);
+    std::vector<char> buf(size + 1);
+    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
+    GGML_ASSERT(size2 == size);
+    va_end(ap2);
+    va_end(ap);
+    return std::string(buf.data(), size);
+}
+
+struct llama_file {
+    // use FILE * so we don't have to re-open the file to mmap
+    FILE * fp;
+    size_t size;
+
+    llama_file(const char * fname, const char * mode) {
+        fp = std::fopen(fname, mode);
+        if (fp == NULL) {
+            size = 0;
+        } else {
+            seek(0, SEEK_END);
+            size = tell();
+            seek(0, SEEK_SET);
+        }
+    }
+
+    size_t tell() const {
+#ifdef _WIN32
+        __int64 ret = _ftelli64(fp);
+#else
+        long ret = std::ftell(fp);
+#endif
+        GGML_ASSERT(ret != -1); // this really shouldn't fail
+        return (size_t) ret;
+    }
+
+    void seek(size_t offset, int whence) {
+#ifdef _WIN32
+        int ret = _fseeki64(fp, (__int64) offset, whence);
+#else
+        int ret = std::fseek(fp, (long) offset, whence);
+#endif
+        GGML_ASSERT(ret == 0); // same
+    }
+
+    void read_raw(void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        std::size_t ret = std::fread(ptr, size, 1, fp);
+        if (ferror(fp)) {
+            throw std::runtime_error(format("read error: %s", strerror(errno)));
+        }
+        if (ret != 1) {
+            throw std::runtime_error(std::string("unexpectedly reached end of file"));
+        }
+    }
+
+    std::uint32_t read_u32() {
+        std::uint32_t ret;
+        read_raw(&ret, sizeof(ret));
+        return ret;
+    }
+    std::float_t read_f32() {
+        std::float_t ret;
+        read_raw(&ret, sizeof(ret));
+        return ret;
+    }
+
+    std::string read_string(std::uint32_t len) {
+        std::vector<char> chars(len);
+        read_raw(chars.data(), len);
+        return std::string(chars.data(), len);
+    }
+
+    void write_raw(const void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        size_t ret = std::fwrite(ptr, size, 1, fp);
+        if (ret != 1) {
+            throw std::runtime_error(format("write error: %s", strerror(errno)));
+        }
+    }
+
+    void write_u32(std::uint32_t val) {
+        write_raw(&val, sizeof(val));
+    }
+
+    ~llama_file() {
+        if (fp) {
+            std::fclose(fp);
+        }
+    }
+};
+
+void write_tensor(struct llama_file * file, struct ggml_tensor * tensor) {
+    if (tensor == NULL) {
+        file->write_u32(0);
+        file->write_u32(0);
+        file->write_u32(GGML_TYPE_F32);
+        file->seek((0-file->tell()) & 31, SEEK_CUR);
+        return;
+    }
+    const char * name = ggml_get_name(tensor);
+    uint32_t name_len = strlen(name);
+    uint32_t nd = tensor->n_dims;
+    uint32_t ne[4] = { (uint32_t)tensor->ne[0],
+                       (uint32_t)tensor->ne[1],
+                       (uint32_t)tensor->ne[2],
+                       (uint32_t)tensor->ne[3] };
+    file->write_u32(nd);
+    file->write_u32(name_len);
+    file->write_u32(tensor->type);
+    file->write_raw(ne, sizeof(ne[0]) * nd);
+    file->write_raw(name, name_len);
+    file->seek((0-file->tell()) & 31, SEEK_CUR);
+    file->write_raw(tensor->data, ggml_nbytes(tensor));
+}
+
+bool is_ggml_file(const char *filename) {
+    llama_file file(filename, "rb");
+    if (file.size < 4) {
+        return false;
+    }
+    uint32_t magic = file.read_u32();
+    return magic == LLAMA_FILE_MAGIC;
+}
+
+void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) {
+    // heuristic to infer whether vocab is from ggml or from llama2.c vocabulary
+    if (is_ggml_file(filename)) {
+
+        struct llama_context_params llama_params = llama_context_default_params();
+        llama_params.vocab_only = true;
+
+        struct llama_model * lmodel = llama_load_model_from_file(filename, llama_params);
+        struct llama_context * lctx = llama_new_context_with_model(lmodel, llama_params);
+
+        std::vector<const char *> strings;
+        std::vector<float> scores;
+        int n_vocab = llama_n_vocab(lctx);
+        strings.resize(n_vocab, NULL);
+        scores.resize(n_vocab, 0);
+        n_vocab = llama_get_vocab(lctx, strings.data(), scores.data(), n_vocab);
+        GGML_ASSERT(n_vocab == llama_n_vocab(lctx));
+        vocab->id_to_token.resize(n_vocab);
+        for (int i=0; i<n_vocab; ++i) {
+            std::string tok   = std::string(strings[i]);
+            float       score = scores[i];
+            vocab->id_to_token[i].tok   = tok;
+            vocab->id_to_token[i].score = score;
+            vocab->token_to_id.emplace(tok, i);
+        }
+        llama_free(lctx);
+        llama_free_model(lmodel);
+    } else { // assume llama2.c vocabulary
+        printf("Assuming llama2.c vocabulary since %s is not a ggml file\n", filename);
+        llama_file file(filename, "rb");
+        uint32_t n_vocab = config->vocab_size;
+        /* uint32_t max_token_length =  */ file.read_u32(); // unused
+        vocab->id_to_token.resize(n_vocab);
+        for (uint32_t i=0; i<n_vocab; ++i) {
+            float_t score = file.read_f32();
+            uint32_t len = file.read_u32();
+            std::string tok = file.read_string(len);
+            vocab->id_to_token[i].tok = tok;
+            vocab->id_to_token[i].score = score;
+            vocab->token_to_id.emplace(tok, i);
+        }
+    }
+}
+
+void stuff_karpathy_weights_into_gg(struct ggml_tensor * gg_weights, float * karpathy_weights){
+    int ct;
+    switch (gg_weights->n_dims){
+        case 1:
+            ct = 0;
+            for (int i0 = 0; i0 < gg_weights->ne[0]; i0++){
+                float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0]);
+                *ptr = karpathy_weights[ct];
+                ct++;
+            }
+            break;
+        case 2:
+            ct = 0;
+            for (int i1 = 0; i1 < gg_weights->ne[1]; i1++) {
+                for (int i0 = 0; i0 < gg_weights->ne[0]; i0++) {
+                    float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0] + i1*gg_weights->nb[1]);
+                    *ptr = karpathy_weights[ct];
+                    ct++;
+                }
+            }
+            break;
+        case 3:
+            ct = 0;
+            for (int i2 = 0; i2 < gg_weights->ne[2]; i2++) {
+                for (int i1 = 0; i1 < gg_weights->ne[1]; i1++) {
+                    for (int i0 = 0; i0 < gg_weights->ne[0]; i0++) {
+                        float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0] + i1*gg_weights->nb[1] + i2*gg_weights->nb[2]);
+                        *ptr = karpathy_weights[ct];
+                        ct++;
+                    }
+                }
+            }
+            break;
+    }
+}
+
+void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename) {
+    struct llama_file file(filename, "wb");
+    if (file.fp == NULL) {
+        return;
+    }
+    // write_magic
+    file.write_u32(LLAMA_FILE_MAGIC);   // magic
+    file.write_u32(LLAMA_FILE_VERSION); // version
+    // write_hparams
+    file.write_u32(model->hparams.n_vocab);
+    file.write_u32(model->hparams.n_embd);
+    file.write_u32(model->hparams.n_mult);
+    file.write_u32(model->hparams.n_head);
+    file.write_u32(model->hparams.n_layer);
+    file.write_u32(model->hparams.n_rot);
+    file.write_u32(LLAMA_FTYPE_ALL_F32);
+
+    // write_vocab - for now we are just writing the existing BPE voc. assuming karpathy's vocabulary is the same. idk.
+    uint32_t n_vocab = model->hparams.n_vocab;
+    for (uint32_t i = 0; i < n_vocab; i++) {
+        const auto & token_score = vocab->id_to_token.at(i);
+        file.write_u32((uint32_t) token_score.tok.size());
+        file.write_raw(token_score.tok.data(), token_score.tok.size());
+        file.write_raw(&token_score.score, sizeof(token_score.score));
+    }
+
+    // stuff AK weights into GG weights one by one.
+    // w->token_embedding_table -> model->tok_embeddings
+    // float*                   -> struct ggml_tensor
+    stuff_karpathy_weights_into_gg(model->tok_embeddings, w->token_embedding_table);
+    stuff_karpathy_weights_into_gg(model->output, w->token_embedding_table);
+
+    stuff_karpathy_weights_into_gg(model->norm, w->rms_final_weight);
+    //print_row(model->norm, 0);
+
+    // for rms-att-weight
+    int row_length = model->hparams.n_embd;
+    const auto & hparams = model->hparams;
+    //int n_ff = model->hparams.n_embd;
+    int n_ff = get_n_ff(&hparams);
+
+    for (uint32_t i = 0; i < model->hparams.n_layer; ++i){
+        auto & layer = model->layers[i];
+        // 1d
+        stuff_karpathy_weights_into_gg(layer.attention_norm, &w->rms_att_weight[i*row_length]);
+        stuff_karpathy_weights_into_gg(layer.ffn_norm      , &w->rms_ffn_weight[i*row_length]);
+
+        // from 3d matrix layer x dim x dim to 2d matrix dim x dim
+        stuff_karpathy_weights_into_gg(layer.wq            , &w->wq[i*row_length*row_length]);
+        stuff_karpathy_weights_into_gg(layer.wk            , &w->wk[i*row_length*row_length]);
+        stuff_karpathy_weights_into_gg(layer.wv            , &w->wv[i*row_length*row_length]);
+        stuff_karpathy_weights_into_gg(layer.wo            , &w->wo[i*row_length*row_length]);
+
+        stuff_karpathy_weights_into_gg(layer.w1            , &w->w1[i*row_length*n_ff]);
+        stuff_karpathy_weights_into_gg(layer.w2            , &w->w2[i*n_ff*row_length]);
+        stuff_karpathy_weights_into_gg(layer.w3            , &w->w3[i*row_length*n_ff]);
+    }
+    // write tensors
+    write_tensor(&file, model->tok_embeddings);
+    write_tensor(&file, model->norm);
+    write_tensor(&file, model->output); // ?
+    for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
+        auto & layer = model->layers[i];
+
+        write_tensor(&file, layer.attention_norm);
+        write_tensor(&file, layer.wq);
+        write_tensor(&file, layer.wk);
+        write_tensor(&file, layer.wv);
+        write_tensor(&file, layer.wo);
+        write_tensor(&file, layer.ffn_norm);
+        write_tensor(&file, layer.w1);
+        write_tensor(&file, layer.w2);
+        write_tensor(&file, layer.w3);
+    }
+}
+
+struct train_params get_default_train_params() {
+    struct train_params params;
+    params.fn_vocab_model    = "models/ggml-vocab.bin";
+    params.fn_llama2c_output_model = "ak_llama_model.bin";
+    params.fn_train_data     = "shakespeare.txt";
+    params.fn_checkpoint_in  = "checkpoint.bin";
+    params.fn_checkpoint_out = "checkpoint.bin";
+    params.fn_model_out      = "ggml-checkpoint-f32.bin";
+
+    params.seed       =   -1;
+
+    params.n_ctx      =  128;
+    params.n_embd     =  256;
+    params.n_mult     =  256;
+    params.n_head     =    8;
+    params.n_layer    =   16;
+    params.n_rotmax   =   64;
+
+    params.n_threads  =    6;
+    params.n_batch    =    8;
+    params.n_examples =    8;
+    params.n_predict  = 1024;
+
+    params.print_info_interval    = 1;
+    params.print_details_interval = 2;
+
+    params.samples_start_after_nl = false;
+    params.use_adam               = true;
+    params.use_flash              = true;
+    params.use_scratch            = true;
+
+    // only adam
+    params.warmup            =  100;
+    params.cos_decay_steps   = 1000;
+    params.cos_decay_restart = 1.1f;
+    params.cos_decay_alpha   = 0.0f;
+
+    params.lbfgs_n_iter      = 16;
+    params.adam_n_iter       = 16;
+    params.adam_alpha        = 1e-3f;
+    params.adam_decay        = 1e-3f;
+
+    params.mem_model_gb   = 2;
+    params.mem_compute_gb = 24;
+    params.mem_compute0_gb = 8;
+    params.mem_compute1_gb = 2;
+
+    return params;
+}
+
+void print_usage(int /*argc*/, char ** argv, const struct train_params * params) {
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help                       show this help message and exit\n");
+    fprintf(stderr, "  --copy-vocab-from-model FNAME    llama2.c vocabulary or ggml model path from which to copy vocab (default '%s')\n", params->fn_vocab_model);
+    fprintf(stderr, "  --llama2c-model FNAME            [REQUIRED] model path from which to load Karpathy's llama2.c model\n");
+    fprintf(stderr, "  --llama2c-output-model FNAME     model path to save the converted llama2.c model (default %s')\n", params->fn_llama2c_output_model);
+    fprintf(stderr, "\n");
+}
+
+bool params_parse(int argc, char ** argv, struct train_params * params) {
+    bool invalid_param = false;
+    bool reqd_param_found = false;
+    std::string arg;
+    struct train_params default_params = get_default_train_params();
+    const std::string arg_prefix = "--";
+
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+        if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
+            std::replace(arg.begin(), arg.end(), '_', '-');
+        }
+
+        if (arg == "--copy-vocab-from-model") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->fn_vocab_model = argv[i];
+        } else if (arg == "--llama2c-model") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            reqd_param_found = true;
+            params->fn_llama2c_model = argv[i];
+        } else if (arg == "--llama2c-output-model") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->fn_llama2c_output_model = argv[i];
+        } else if (arg == "-h" || arg == "--help") {
+            print_usage(argc, argv, &default_params);
+            exit(0);
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            print_usage(argc, argv, &default_params);
+            exit(1);
+        }
+    }
+    if (invalid_param) {
+        fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
+        print_usage(argc, argv, &default_params);
+        exit(1);
+    }
+    if (!reqd_param_found){
+        fprintf(stderr, "error: please specify a llama2.c .bin file to be converted with argument --llama2c-model\n");
+        print_usage(argc, argv, &default_params);
+        exit(1);
+    }
+
+    return true;
+}
+
+int main(int argc, char ** argv) {
+    struct train_params params = get_default_train_params();
+    if (!params_parse(argc, argv, &params)) {
+        return 1;
+    }
+    Config config;
+    TransformerWeights weights;
+    {
+        FILE *file = fopen(params.fn_llama2c_model, "rb");
+        if (!file) { printf("Unable to open the checkpoint file %s!\n", params.fn_llama2c_model); return 1; }
+        // read in the config header
+        if(fread(&config, sizeof(Config), 1, file) != 1) { return 1; }
+        // read in the Transformer weights
+        malloc_weights(&weights, &config);
+        if(checkpoint_init_weights(&weights, &config, file)) { return 1; }
+        fclose(file);
+    }
+
+    struct llama_vocab vocab;
+    load_vocab(params.fn_vocab_model, &config, &vocab);
+
+    struct my_llama_model model;
+    model.hparams.n_vocab = config.vocab_size; //llama_n_vocab(lctx);
+    model.hparams.n_ctx   = params.n_ctx;
+    model.hparams.n_embd  = config.dim; //params.n_embd;
+    model.hparams.n_mult  = 32;//params.n_mult;
+    model.hparams.n_head  = config.n_heads; //params.n_head;
+    model.hparams.n_layer = config.n_layers; //params.n_layer;
+    model.hparams.n_rot   = std::min((uint32_t)params.n_rotmax, model.hparams.n_embd / model.hparams.n_head);
+    print_params(&model.hparams);
+    struct ggml_init_params lcparams;
+    lcparams.mem_size   = 1024ll*1024ll*1024ll*((size_t) params.mem_model_gb);
+    lcparams.mem_buffer = NULL;
+    lcparams.no_alloc   = false;
+
+    model.ctx = ggml_init(lcparams);
+
+    init_model(&model);
+    save_as_llama_model(&vocab, &model, &weights, params.fn_llama2c_output_model);
+
+    printf("Saving llama.c model file %s in ggml format at %s\n", params.fn_llama2c_model, params.fn_llama2c_output_model);
+
+    ggml_free(model.ctx);
+    free_weights(&weights);
+    return 0;
+}

examples/llama-bench/CMakeLists.txt

@@ -0,0 +1,8 @@
+set(TARGET llama-bench)
+add_executable(${TARGET} llama-bench.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()

examples/llama-bench/llama-bench.cpp

@@ -0,0 +1,967 @@
+#include <algorithm>
+#include <array>
+#include <cassert>
+#include <chrono>
+#include <cinttypes>
+#include <cstring>
+#include <ctime>
+#include <iterator>
+#include <map>
+#include <numeric>
+#include <regex>
+#include <sstream>
+#include <stdio.h>
+#include <string>
+#include <vector>
+
+#include "ggml.h"
+#include "llama.h"
+#include "common.h"
+#include "build-info.h"
+#ifdef GGML_USE_CUBLAS
+#include "ggml-cuda.h"
+#endif
+
+// utils
+static uint64_t get_time_ns() {
+    using clock = std::chrono::high_resolution_clock;
+    return std::chrono::nanoseconds(clock::now().time_since_epoch()).count();
+}
+
+template<class T>
+static std::string join(const std::vector<T> & values, const std::string & delim) {
+    std::ostringstream str;
+    for (size_t i = 0; i < values.size(); i++) {
+        str << values[i];
+        if (i < values.size() - 1) {
+            str << delim;
+        }
+    }
+    return str.str();
+}
+
+template<class T>
+static std::vector<T> split(const std::string & str, char delim) {
+    std::vector<T> values;
+    std::istringstream str_stream(str);
+    std::string token;
+    while (std::getline(str_stream, token, delim)) {
+        T value;
+        std::istringstream token_stream(token);
+        token_stream >> value;
+        values.push_back(value);
+    }
+    return values;
+}
+
+template<typename T>
+static T avg(const std::vector<T> & v) {
+    if (v.empty()) {
+        return 0;
+    }
+    T sum = std::accumulate(v.begin(), v.end(), T(0));
+    return sum / (T)v.size();
+}
+
+template<typename T>
+static T stdev(const std::vector<T> & v) {
+    if (v.size() <= 1) {
+        return 0;
+    }
+    T mean = avg(v);
+    T sq_sum = std::inner_product(v.begin(), v.end(), v.begin(), T(0));
+    T stdev = std::sqrt(sq_sum / (T)(v.size() - 1) - mean * mean * (T)v.size() / (T)(v.size() - 1));
+    return stdev;
+}
+
+static bool ggml_cpu_has_metal() {
+#if defined(GGML_USE_METAL)
+    return true;
+#else
+    return false;
+#endif
+}
+
+static std::string get_cpu_info() {
+    std::string id;
+#ifdef __linux__
+    FILE * f = fopen("/proc/cpuinfo", "r");
+    if (f) {
+        char buf[1024];
+        while (fgets(buf, sizeof(buf), f)) {
+            if (strncmp(buf, "model name", 10) == 0) {
+                char * p = strchr(buf, ':');
+                if (p) {
+                    p++;
+                    while (std::isspace(*p)) {
+                        p++;
+                    }
+                    while (std::isspace(p[strlen(p) - 1])) {
+                        p[strlen(p) - 1] = '\0';
+                    }
+                    id = p;
+                    break;
+                }
+            }
+        }
+    }
+#endif
+    // TODO: other platforms
+    return id;
+}
+
+static std::string get_gpu_info() {
+    std::string id;
+#ifdef GGML_USE_CUBLAS
+    int count = ggml_cuda_get_device_count();
+    for (int i = 0; i < count; i++) {
+        char buf[128];
+        ggml_cuda_get_device_description(i, buf, sizeof(buf));
+        id += buf;
+        if (i < count - 1) {
+            id += "/";
+        }
+    }
+#endif
+    // TODO: other backends
+    return id;
+}
+
+// command line params
+enum output_formats {CSV, JSON, MARKDOWN, SQL};
+
+struct cmd_params {
+    std::vector<std::string> model;
+    std::vector<int> n_prompt;
+    std::vector<int> n_gen;
+    std::vector<int> n_batch;
+    std::vector<bool> f32_kv;
+    std::vector<int> n_threads;
+    std::vector<int> n_gpu_layers;
+    std::vector<int> main_gpu;
+    std::vector<bool> mul_mat_q;
+    std::vector<bool> low_vram;
+    std::vector<std::array<float, LLAMA_MAX_DEVICES>> tensor_split;
+    int reps;
+    bool verbose;
+    output_formats output_format;
+};
+
+static const cmd_params cmd_params_defaults = {
+    /* model         */ {"models/7B/ggml-model-q4_0.bin"},
+    /* n_prompt      */ {512},
+    /* n_gen         */ {128},
+    /* n_batch       */ {512},
+    /* f32_kv        */ {false},
+    /* n_threads     */ {get_num_physical_cores()},
+    /* n_gpu_layers  */ {99},
+    /* main_gpu      */ {0},
+    /* mul_mat_q     */ {true},
+    /* low_vram      */ {false},
+    /* tensor_split  */ {{}},
+    /* reps          */ 5,
+    /* verbose       */ false,
+    /* output_format */ MARKDOWN
+};
+
+static void print_usage(int /* argc */, char ** argv) {
+    fprintf(stdout, "usage: %s [options]\n", argv[0]);
+    fprintf(stdout, "\n");
+    fprintf(stdout, "options:\n");
+    fprintf(stdout, "  -h, --help\n");
+    fprintf(stdout, "  -m, --model <filename>            (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
+    fprintf(stdout, "  -p, --n-prompt <n>                (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
+    fprintf(stdout, "  -n, --n-gen <n>                   (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
+    fprintf(stdout, "  -b, --batch-size <n>              (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
+    fprintf(stdout, "  --memory-f32 <0|1>                (default: %s)\n", join(cmd_params_defaults.f32_kv, ",").c_str());
+    fprintf(stdout, "  -t, --threads <n>                 (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
+    fprintf(stdout, "  -ngl N, --n-gpu-layers <n>        (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
+    fprintf(stdout, "  -mg i, --main-gpu <n>             (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
+    fprintf(stdout, "  -lv, --low-vram <0|1>             (default: %s)\n", join(cmd_params_defaults.low_vram, ",").c_str());
+    fprintf(stdout, "  -mmq, --mul-mat-q <0|1>           (default: %s)\n", join(cmd_params_defaults.mul_mat_q, ",").c_str());
+    fprintf(stdout, "  -ts, --tensor_split <ts>                       \n");
+    fprintf(stdout, "  -r, --repetitions <n>             (default: %d)\n", cmd_params_defaults.reps);
+    fprintf(stdout, "  -o, --output <csv|json|md|sql>    (default: %s)\n", cmd_params_defaults.output_format == CSV ? "csv" : cmd_params_defaults.output_format == JSON ? "json" : "md");
+    fprintf(stdout, "  -v, --verbose                     (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
+    fprintf(stdout, "\n");
+    fprintf(stdout, "Multiple values can be given for each parameter by separating them with ',' or by repeating the parameter.\n");
+
+}
+
+static cmd_params parse_cmd_params(int argc, char ** argv) {
+    cmd_params params;
+    std::string arg;
+    bool invalid_param = false;
+    const std::string arg_prefix = "--";
+    const char split_delim = ',';
+
+    params.verbose = cmd_params_defaults.verbose;
+    params.output_format = cmd_params_defaults.output_format;
+    params.reps = cmd_params_defaults.reps;
+
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+        if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
+            std::replace(arg.begin(), arg.end(), '_', '-');
+        }
+
+        if (arg == "-h" || arg == "--help") {
+            print_usage(argc, argv);
+            exit(0);
+        } else if (arg == "-m" || arg == "--model") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<std::string>(argv[i], split_delim);
+            params.model.insert(params.model.end(), p.begin(), p.end());
+        } else if (arg == "-p" || arg == "--n-prompt") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<int>(argv[i], split_delim);
+            params.n_prompt.insert(params.n_prompt.end(), p.begin(), p.end());
+        } else if (arg == "-n" || arg == "--n-gen") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<int>(argv[i], split_delim);
+            params.n_gen.insert(params.n_gen.end(), p.begin(), p.end());
+        } else if (arg == "-b" || arg == "--batch-size") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<int>(argv[i], split_delim);
+            params.n_batch.insert(params.n_batch.end(), p.begin(), p.end());
+        } else if (arg == "--memory-f32") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<int>(argv[i], split_delim);
+            params.f32_kv.insert(params.f32_kv.end(), p.begin(), p.end());
+        } else if (arg == "-t" || arg == "--threads") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<int>(argv[i], split_delim);
+            params.n_threads.insert(params.n_threads.end(), p.begin(), p.end());
+        } else if (arg == "-ngl" || arg == "--n-gpu-layers") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<int>(argv[i], split_delim);
+            params.n_gpu_layers.insert(params.n_gpu_layers.end(), p.begin(), p.end());
+        } else if (arg == "-mg" || arg == "--main-gpu") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.main_gpu = split<int>(argv[i], split_delim);
+        } else if (arg == "-lv" || arg == "--low-vram") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<bool>(argv[i], split_delim);
+            params.low_vram.insert(params.low_vram.end(), p.begin(), p.end());
+        } else if (arg == "-mmq" || arg == "--mul-mat-q") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<bool>(argv[i], split_delim);
+            params.mul_mat_q.insert(params.mul_mat_q.end(), p.begin(), p.end());
+        } else if (arg == "-ts" || arg == "--tensor-split") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            for (auto ts : split<std::string>(argv[i], split_delim)) {
+                // split string by ; and /
+                const std::regex regex{R"([;/]+)"};
+                std::sregex_token_iterator it{ts.begin(), ts.end(), regex, -1};
+                std::vector<std::string> split_arg{it, {}};
+                GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES);
+
+                std::array<float, LLAMA_MAX_DEVICES> tensor_split;
+                for (size_t i = 0; i < LLAMA_MAX_DEVICES; ++i) {
+                    if (i < split_arg.size()) {
+                        tensor_split[i] = std::stof(split_arg[i]);
+                    } else {
+                        tensor_split[i] = 0.0f;
+                    }
+                }
+                params.tensor_split.push_back(tensor_split);
+            }
+        } else if (arg == "-r" || arg == "--repetitions") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.reps = std::stoi(argv[i]);
+        } else if (arg == "-o" || arg == "--output") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            if (argv[i] == std::string("csv")) {
+                params.output_format = CSV;
+            } else if (argv[i] == std::string("json")) {
+                params.output_format = JSON;
+            } else if (argv[i] == std::string("md")) {
+                params.output_format = MARKDOWN;
+            } else if (argv[i] == std::string("sql")) {
+                params.output_format = SQL;
+            } else {
+                invalid_param = true;
+                break;
+            }
+        } else if (arg == "-v" || arg == "--verbose") {
+            params.verbose = true;
+        } else {
+            invalid_param = true;
+            break;
+        }
+    }
+    if (invalid_param) {
+        fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
+        print_usage(argc, argv);
+        exit(1);
+    }
+
+    // set defaults
+    if (params.model.empty())        { params.model = cmd_params_defaults.model; }
+    if (params.n_prompt.empty())     { params.n_prompt = cmd_params_defaults.n_prompt; }
+    if (params.n_gen.empty())        { params.n_gen = cmd_params_defaults.n_gen; }
+    if (params.n_batch.empty())      { params.n_batch = cmd_params_defaults.n_batch; }
+    if (params.f32_kv.empty())       { params.f32_kv = cmd_params_defaults.f32_kv; }
+    if (params.n_gpu_layers.empty()) { params.n_gpu_layers = cmd_params_defaults.n_gpu_layers; }
+    if (params.main_gpu.empty())     { params.main_gpu = cmd_params_defaults.main_gpu; }
+    if (params.mul_mat_q.empty())    { params.mul_mat_q = cmd_params_defaults.mul_mat_q; }
+    if (params.low_vram.empty())     { params.low_vram = cmd_params_defaults.low_vram; }
+    if (params.tensor_split.empty()) { params.tensor_split = cmd_params_defaults.tensor_split; }
+    if (params.n_threads.empty())    { params.n_threads = cmd_params_defaults.n_threads; }
+
+    return params;
+}
+
+struct cmd_params_instance {
+    std::string model;
+    int n_prompt;
+    int n_gen;
+    int n_batch;
+    bool f32_kv;
+    int n_threads;
+    int n_gpu_layers;
+    int main_gpu;
+    bool mul_mat_q;
+    bool low_vram;
+    std::array<float, LLAMA_MAX_DEVICES> tensor_split;
+
+    llama_context_params to_llama_params() const {
+        llama_context_params lparams = llama_context_default_params();
+        lparams.n_ctx = n_prompt + n_gen;
+        lparams.n_batch = n_batch;
+        lparams.f16_kv = !f32_kv;
+        lparams.n_gpu_layers = n_gpu_layers;
+        lparams.main_gpu = main_gpu;
+        lparams.mul_mat_q = mul_mat_q;
+        lparams.low_vram = low_vram;
+        lparams.tensor_split = tensor_split.data();
+
+        return lparams;
+    }
+};
+
+static std::vector<cmd_params_instance> get_cmd_params_instances_int(const cmd_params & params, int n_gen, int n_prompt) {
+    std::vector<cmd_params_instance> instances;
+
+    for (const auto & m : params.model)
+    for (const auto & nb : params.n_batch)
+    for (const auto & fk : params.f32_kv)
+    for (const auto & nl : params.n_gpu_layers)
+    for (const auto & mg : params.main_gpu)
+    for (const auto & mmq : params.mul_mat_q)
+    for (const auto & lv : params.low_vram)
+    for (const auto & ts : params.tensor_split)
+    for (const auto & nt : params.n_threads) {
+        cmd_params_instance instance = {
+            /* .model        = */ m,
+            /* .n_prompt     = */ n_prompt,
+            /* .n_gen        = */ n_gen,
+            /* .n_batch      = */ nb,
+            /* .f32_kv       = */ fk,
+            /* .n_threads    = */ nt,
+            /* .n_gpu_layers = */ nl,
+            /* .main_gpu     = */ mg,
+            /* .mul_mat_q    = */ mmq,
+            /* .low_vram     = */ lv,
+            /* .tensor_split = */ ts,
+        };
+        instances.push_back(instance);
+    }
+    return instances;
+}
+
+static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_params & params) {
+    std::vector<cmd_params_instance> instances;
+
+    for (const auto & n_prompt : params.n_prompt) {
+        if (n_prompt == 0) {
+            continue;
+        }
+        auto instances_prompt = get_cmd_params_instances_int(params, 0, n_prompt);
+        instances.insert(instances.end(), instances_prompt.begin(), instances_prompt.end());
+    }
+
+    for (const auto & n_gen : params.n_gen) {
+        if (n_gen == 0) {
+            continue;
+        }
+        auto instances_gen = get_cmd_params_instances_int(params, n_gen, 0);
+        instances.insert(instances.end(), instances_gen.begin(), instances_gen.end());
+    }
+
+    return instances;
+}
+
+struct test {
+    static const std::string build_commit;
+    static const int build_number;
+    static const bool cuda;
+    static const bool opencl;
+    static const bool metal;
+    static const bool gpu_blas;
+    static const bool blas;
+    static const std::string cpu_info;
+    static const std::string gpu_info;
+    std::string model_filename;
+    std::string model_type;
+    int n_batch;
+    int n_threads;
+    bool f32_kv;
+    int n_gpu_layers;
+    int main_gpu;
+    bool mul_mat_q;
+    bool low_vram;
+    std::array<float, LLAMA_MAX_DEVICES> tensor_split;
+    int n_prompt;
+    int n_gen;
+    std::string test_time;
+    std::vector<uint64_t> samples_ns;
+
+    test(const cmd_params_instance & inst, const llama_model * lmodel, const llama_context * ctx) {
+        model_filename = inst.model;
+        char buf[128];
+        llama_model_type(lmodel, buf, sizeof(buf));
+        model_type = buf;
+        n_batch = inst.n_batch;
+        n_threads = inst.n_threads;
+        f32_kv = inst.f32_kv;
+        n_gpu_layers = inst.n_gpu_layers;
+        main_gpu = inst.main_gpu;
+        mul_mat_q = inst.mul_mat_q;
+        low_vram = inst.low_vram;
+        tensor_split = inst.tensor_split;
+        n_prompt = inst.n_prompt;
+        n_gen = inst.n_gen;
+        // RFC 3339 date-time format
+        time_t t = time(NULL);
+        std::strftime(buf, sizeof(buf), "%FT%TZ", gmtime(&t));
+        test_time = buf;
+
+        (void) ctx;
+    }
+
+    uint64_t avg_ns() const {
+        return ::avg(samples_ns);
+    }
+
+    uint64_t stdev_ns() const {
+        return ::stdev(samples_ns);
+    }
+
+    std::vector<double> get_ts() const {
+        int n_tokens = n_prompt + n_gen;
+        std::vector<double> ts;
+        std::transform(samples_ns.begin(), samples_ns.end(), std::back_inserter(ts), [n_tokens](uint64_t t) { return 1e9 * n_tokens / t; });
+        return ts;
+    }
+
+    double avg_ts() const {
+        return ::avg(get_ts());
+    }
+
+    double stdev_ts() const {
+        return ::stdev(get_ts());
+    }
+
+    static std::string get_backend() {
+        if (cuda) {
+            return "CUDA";
+        }
+        if (opencl) {
+            return "OpenCL";
+        }
+        if (metal) {
+            return "Metal";
+        }
+        if (gpu_blas) {
+            return "GPU BLAS";
+        }
+        if (blas) {
+            return "BLAS";
+        }
+        return "CPU";
+    }
+
+    static const std::vector<std::string> & get_fields() {
+        static const std::vector<std::string> fields = {
+            "build_commit", "build_number",
+            "cuda", "opencl", "metal", "gpu_blas", "blas",
+            "cpu_info", "gpu_info",
+            "model_filename", "model_type",
+            "n_batch", "n_threads", "f16_kv",
+            "n_gpu_layers", "main_gpu", "mul_mat_q", "low_vram", "tensor_split",
+            "n_prompt", "n_gen", "test_time",
+            "avg_ns", "stddev_ns",
+            "avg_ts", "stddev_ts"
+        };
+        return fields;
+    }
+
+    enum field_type {STRING, BOOL, INT, FLOAT};
+
+    static field_type get_field_type(const std::string & field) {
+        if (field == "build_number" || field == "n_batch" || field == "n_threads" ||
+            field == "n_gpu_layers" || field == "main_gpu" ||
+            field == "n_prompt" || field == "n_gen" ||
+            field == "avg_ns" || field == "stddev_ns") {
+            return INT;
+        }
+        if (field == "cuda" || field == "opencl" || field == "metal" || field == "gpu_blas" || field == "blas" ||
+            field == "f16_kv" || field == "mul_mat_q" || field == "low_vram") {
+            return BOOL;
+        }
+        if (field == "avg_ts" || field == "stddev_ts") {
+            return FLOAT;
+        }
+        return STRING;
+    }
+
+    std::vector<std::string> get_values() const {
+        std::string tensor_split_str;
+        int max_nonzero = 0;
+        for (int i = 0; i < LLAMA_MAX_DEVICES; i++) {
+            if (tensor_split[i] > 0) {
+                max_nonzero = i;
+            }
+        }
+        for (int i = 0; i <= max_nonzero; i++) {
+            char buf[32];
+            snprintf(buf, sizeof(buf), "%.2f", tensor_split[i]);
+            tensor_split_str += buf;
+            if (i < max_nonzero) {
+                tensor_split_str += "/";
+            }
+        }
+        std::vector<std::string> values = {
+            build_commit, std::to_string(build_number),
+            std::to_string(cuda), std::to_string(opencl), std::to_string(metal), std::to_string(gpu_blas), std::to_string(blas),
+            cpu_info, gpu_info,
+            model_filename, model_type,
+            std::to_string(n_batch), std::to_string(n_threads), std::to_string(!f32_kv),
+            std::to_string(n_gpu_layers), std::to_string(main_gpu), std::to_string(mul_mat_q), std::to_string(low_vram), tensor_split_str,
+            std::to_string(n_prompt), std::to_string(n_gen), test_time,
+            std::to_string(avg_ns()), std::to_string(stdev_ns()),
+            std::to_string(avg_ts()), std::to_string(stdev_ts())
+        };
+        return values;
+    }
+
+    std::map<std::string, std::string> get_map() const {
+        std::map<std::string, std::string> map;
+        auto fields = get_fields();
+        auto values = get_values();
+        std::transform(fields.begin(), fields.end(), values.begin(),
+                std::inserter(map, map.end()), std::make_pair<const std::string &, const std::string &>);
+        return map;
+    }
+};
+
+const std::string test::build_commit = BUILD_COMMIT;
+const int         test::build_number = BUILD_NUMBER;
+const bool        test::cuda         = !!ggml_cpu_has_cublas();
+const bool        test::opencl       = !!ggml_cpu_has_clblast();
+const bool        test::metal        = !!ggml_cpu_has_metal();
+const bool        test::gpu_blas     = !!ggml_cpu_has_gpublas();
+const bool        test::blas         = !!ggml_cpu_has_blas();
+const std::string test::cpu_info     = get_cpu_info();
+const std::string test::gpu_info     = get_gpu_info();
+
+struct printer {
+    FILE * fout;
+    virtual void print_header(const cmd_params & params) { (void) params; };
+    virtual void print_test(const test & t) = 0;
+    virtual void print_footer() { };
+};
+
+struct csv_printer : public printer {
+    static std::string escape_csv(const std::string & field) {
+        std::string escaped = "\"";
+        for (auto c : field) {
+            if (c == '"') {
+                escaped += "\"";
+            }
+            escaped += c;
+        }
+        escaped += "\"";
+        return escaped;
+    }
+
+    void print_header(const cmd_params & params) override  {
+        std::vector<std::string> fields = test::get_fields();
+        fprintf(fout, "%s\n", join(fields, ",").c_str());
+        (void) params;
+    }
+
+    void print_test(const test & t) override {
+        std::vector<std::string> values = t.get_values();
+        std::transform(values.begin(), values.end(), values.begin(), escape_csv);
+        fprintf(fout, "%s\n", join(values, ",").c_str());
+    }
+};
+
+struct json_printer : public printer {
+    bool first = true;
+
+    static std::string escape_json(const std::string & value) {
+        std::string escaped;
+        for (auto c : value) {
+            if (c == '"') {
+                escaped += "\\\"";
+            } else if (c == '\\') {
+                escaped += "\\\\";
+            } else  if (c <= 0x1f) {
+                char buf[8];
+                snprintf(buf, sizeof(buf), "\\u%04x", c);
+                escaped += buf;
+            } else {
+                escaped += c;
+            }
+        }
+        return escaped;
+    }
+
+    static std::string format_value(const std::string & field, const std::string & value) {
+        switch (test::get_field_type(field)) {
+            case test::STRING:
+                return "\"" + escape_json(value) + "\"";
+            case test::BOOL:
+                return value == "0" ? "false" : "true";
+            default:
+                return value;
+        }
+    }
+
+    void print_header(const cmd_params & params) override {
+        fprintf(fout, "[\n");
+        (void) params;
+    }
+
+    void print_fields(const std::vector<std::string> & fields, const std::vector<std::string> & values) {
+        assert(fields.size() == values.size());
+        for (size_t i = 0; i < fields.size(); i++) {
+            fprintf(fout, "    \"%s\": %s,\n", fields.at(i).c_str(), format_value(fields.at(i), values.at(i)).c_str());
+        }
+    }
+
+    void print_test(const test & t) override {
+        if (first) {
+            first = false;
+        } else {
+            fprintf(fout, ",\n");
+        }
+        fprintf(fout, "  {\n");
+        print_fields(test::get_fields(), t.get_values());
+        fprintf(fout, "    \"samples_ns\": [ %s ],\n", join(t.samples_ns, ", ").c_str());
+        fprintf(fout, "    \"samples_ts\": [ %s ]\n", join(t.get_ts(), ", ").c_str());
+        fprintf(fout, "  }");
+        fflush(fout);
+    }
+
+    void print_footer() override {
+        fprintf(fout, "\n]\n");
+    }
+};
+
+struct markdown_printer : public printer {
+    std::vector<std::string> fields;
+
+    static int get_field_width(const std::string & field) {
+        if (field == "model") {
+            return -30;
+        }
+        if (field == "t/s") {
+            return 15;
+        }
+        int width = std::max((int)field.length(), 10);
+
+        if (test::get_field_type(field) == test::STRING) {
+            return -width;
+        }
+        return width;
+    }
+
+    void print_header(const cmd_params & params) override {
+        // select fields to print
+        fields = { "model", "backend" };
+        bool is_cpu_backend = test::get_backend() == "CPU" || test::get_backend() == "BLAS";
+        if (!is_cpu_backend) {
+            fields.push_back("n_gpu_layers");
+        }
+        if (params.n_batch.size() > 1 || params.n_threads != cmd_params_defaults.n_threads || is_cpu_backend) {
+            fields.push_back("n_threads");
+        }
+        if (params.n_batch.size() > 1 || params.n_batch != cmd_params_defaults.n_batch) {
+            fields.push_back("n_batch");
+        }
+        if (params.f32_kv.size() > 1 || params.f32_kv != cmd_params_defaults.f32_kv) {
+            fields.push_back("f16_kv");
+        }
+        if (params.main_gpu.size() > 1 || params.main_gpu != cmd_params_defaults.main_gpu) {
+            fields.push_back("main_gpu");
+        }
+        if (params.mul_mat_q.size() > 1 || params.mul_mat_q != cmd_params_defaults.mul_mat_q) {
+            fields.push_back("mul_mat_q");
+        }
+        if (params.low_vram.size() > 1 || params.low_vram != cmd_params_defaults.low_vram) {
+            fields.push_back("low_vram");
+        }
+        if (params.tensor_split.size() > 1 || params.tensor_split != cmd_params_defaults.tensor_split) {
+            fields.push_back("tensor_split");
+        }
+        fields.push_back("test");
+        fields.push_back("t/s");
+
+        fprintf(fout, "|");
+        for (const auto & field : fields) {
+            fprintf(fout, " %*s |", get_field_width(field), field.c_str());
+        }
+        fprintf(fout, "\n");
+        fprintf(fout, "|");
+        for (const auto & field : fields) {
+            int width = get_field_width(field);
+            fprintf(fout, " %s%s |", std::string(std::abs(width) - 1, '-').c_str(), width > 0 ? ":" : "-");
+        }
+        fprintf(fout, "\n");
+    }
+
+    void print_test(const test & t) override {
+        std::map<std::string, std::string> vmap = t.get_map();
+
+        fprintf(fout, "|");
+        for (const auto & field : fields) {
+            std::string value;
+            if (field == "model") {
+                value = t.model_type;
+            } else if (field == "backend") {
+                value = test::get_backend();
+            } else if (field == "test") {
+                char buf[128];
+                if (t.n_prompt > 0 && t.n_gen == 0) {
+                    snprintf(buf, sizeof(buf), "pp %d", t.n_prompt);
+                } else if (t.n_gen > 0 && t.n_prompt == 0) {
+                    snprintf(buf, sizeof(buf), "tg %d", t.n_gen);
+                } else {
+                    assert(false);
+                    exit(1);
+                }
+                value = buf;
+            } else if (field == "t/s") {
+                char buf[128];
+                snprintf(buf, sizeof(buf), "%.2f ± %.2f", t.avg_ts(), t.stdev_ts());
+                value = buf;
+            } else if (vmap.find(field) != vmap.end()) {
+                value = vmap.at(field);
+            } else {
+                assert(false);
+                exit(1);
+            }
+
+            int width = get_field_width(field);
+            if (field == "t/s") {
+                // HACK: the utf-8 character is 2 bytes
+                width += 1;
+            }
+            fprintf(fout, " %*s |", width, value.c_str());
+        }
+        fprintf(fout, "\n");
+    }
+
+    void print_footer() override {
+        fprintf(fout, "\nbuild: %s (%d)\n", test::build_commit.c_str(), test::build_number);
+    }
+};
+
+struct sql_printer : public printer {
+    static std::string get_sql_field_type(const std::string & field) {
+        switch (test::get_field_type(field)) {
+            case test::STRING:
+                return "TEXT";
+            case test::BOOL:
+            case test::INT:
+                return "INTEGER";
+            case test::FLOAT:
+                return "REAL";
+            default:
+                assert(false);
+                exit(1);
+        }
+    }
+
+    void print_header(const cmd_params & params) override {
+        std::vector<std::string> fields = test::get_fields();
+        fprintf(fout, "CREATE TABLE IF NOT EXISTS test (\n");
+        for (size_t i = 0; i < fields.size(); i++) {
+            fprintf(fout, "  %s %s%s\n", fields.at(i).c_str(), get_sql_field_type(fields.at(i)).c_str(),  i < fields.size() - 1 ? "," : "");
+        }
+        fprintf(fout, ");\n");
+        fprintf(fout, "\n");
+        (void) params;
+    }
+
+    void print_test(const test & t) override {
+        fprintf(fout, "INSERT INTO test (%s) ", join(test::get_fields(), ", ").c_str());
+        fprintf(fout, "VALUES (");
+        std::vector<std::string> values = t.get_values();
+        for (size_t i = 0; i < values.size(); i++) {
+            fprintf(fout, "'%s'%s", values.at(i).c_str(), i < values.size() - 1 ? ", " : "");
+        }
+        fprintf(fout, ");\n");
+    }
+};
+
+static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_batch, int n_threads) {
+    std::vector<llama_token> tokens(n_batch, llama_token_bos());
+    int n_processed = 0;
+    while (n_processed < n_prompt) {
+        int n_tokens = std::min(n_prompt - n_processed, n_batch);
+        llama_eval(ctx, tokens.data(), n_tokens, n_past + n_processed, n_threads);
+        n_processed += n_tokens;
+    }
+}
+
+static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
+    llama_token token = llama_token_bos();
+    for (int i = 0; i < n_gen; i++) {
+        llama_eval(ctx, &token, 1, n_past + i, n_threads);
+    }
+}
+
+static void llama_null_log_callback(enum llama_log_level level, const char * text, void * user_data) {
+    (void) level;
+    (void) text;
+    (void) user_data;
+}
+
+int main(int argc, char ** argv) {
+#if !defined(NDEBUG)
+    fprintf(stderr, "warning: asserts enabled, performance may be affected\n");
+#endif
+
+#if (defined(_MSC_VER) && defined(_DEBUG)) || (!defined(_MSC_VER) && !defined(__OPTIMIZE__))
+    fprintf(stderr, "warning: debug build, performance may be affected\n");
+#endif
+
+#if defined(__SANITIZE_ADDRESS__) || defined(__SANITIZE_THREAD__)
+    fprintf(stderr, "warning: sanitizer enabled, performance may be affected\n");
+#endif
+
+    cmd_params params = parse_cmd_params(argc, argv);
+
+    // initialize llama.cpp
+    if (!params.verbose) {
+        llama_log_set(llama_null_log_callback, NULL);
+    }
+    bool numa = false;
+    llama_backend_init(numa);
+
+    // initialize printer
+    std::unique_ptr<printer> p;
+    switch (params.output_format) {
+        case CSV:
+            p.reset(new csv_printer());
+            break;
+        case JSON:
+            p.reset(new json_printer());
+            break;
+        case MARKDOWN:
+            p.reset(new markdown_printer());
+            break;
+        case SQL:
+            p.reset(new sql_printer());
+            break;
+        default:
+            assert(false);
+            exit(1);
+    }
+    p->fout = stdout;
+    p->print_header(params);
+
+    std::vector<cmd_params_instance> params_instances = get_cmd_params_instances(params);
+
+    for (const auto & inst : params_instances) {
+        // TODO: keep the model between tests when possible
+        llama_context_params lparams = inst.to_llama_params();
+
+        llama_model * lmodel  = llama_load_model_from_file(inst.model.c_str(), lparams);
+        if (lmodel == NULL) {
+            fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, inst.model.c_str());
+            return 1;
+        }
+
+        llama_context * ctx = llama_new_context_with_model(lmodel, lparams);
+        if (ctx == NULL) {
+            fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, inst.model.c_str());
+            llama_free_model(lmodel);
+            return 1;
+        }
+
+        test t(inst, lmodel, ctx);
+
+        // warmup run
+        test_gen(ctx, 1, 0, t.n_threads);
+
+        for (int i = 0; i < params.reps; i++) {
+            uint64_t t_start = get_time_ns();
+            if (t.n_prompt > 0) {
+                test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
+            }
+            if (t.n_gen > 0) {
+                test_gen(ctx, t.n_gen, t.n_prompt, t.n_threads);
+            }
+            uint64_t t_ns = get_time_ns() - t_start;
+            t.samples_ns.push_back(t_ns);
+        }
+
+        p->print_test(t);
+
+        llama_print_timings(ctx);
+
+        llama_free(ctx);
+        llama_free_model(lmodel);
+    }
+
+    p->print_footer();
+
+    llama_backend_free();
+
+    return 0;
+}

examples/perplexity/perplexity.cpp

@@ -88,7 +88,7 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
                 fprintf(stderr, "%d hours ", total_seconds / (60*60));
                 total_seconds = total_seconds % (60*60);
             }
-            fprintf(stderr, "%d minutes\n", total_seconds / 60);
+            fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
         }
 
         // We get the logits for all the tokens in the context window (params.n_ctx)

examples/server/README.md

@@ -17,6 +17,7 @@ Command line options:
 -   `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. Not recommended.
 -   `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped.
 -   `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed.
+-   `--numa`: Attempt optimizations that help on some NUMA systems.
 -   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 -   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
 -   `-to N`, `--timeout N`: Server read/write timeout in seconds. Default `600`.

examples/server/chat.mjs

@@ -1,5 +1,34 @@
 import * as readline from 'node:readline'
 import { stdin, stdout } from 'node:process'
+import { readFileSync } from 'node:fs'
+import { SchemaConverter }  from './public/json-schema-to-grammar.mjs'
+
+const args = process.argv.slice(2);
+const grammarJsonSchemaFile = args.find(
+    (_, index) => args[index - 1] === "--grammar-json-schema"
+);
+const grammarFile = args.find((_, index) => args[index - 1] === "--grammar");
+
+// Example usage: function,arguments
+const grammarJsonSchemaPropOrder = args.find(
+    (_, index) => args[index - 1] === "--grammar-json-schema-prop-order"
+);
+const propOrder = grammarJsonSchemaPropOrder
+    ? grammarJsonSchemaPropOrder
+          .split(",")
+          .reduce((acc, cur, index) => ({ ...acc, [cur]: index }), {})
+    : {};
+
+let grammar = null
+if (grammarJsonSchemaFile) {
+    const schema = JSON.parse(readFileSync(grammarJsonSchemaFile, 'utf-8'))
+    const converter = new SchemaConverter(propOrder)
+    converter.visit(schema, '')
+    grammar = converter.formatGrammar()
+}
+if (grammarFile) {
+    grammar = readFileSync(grammarFile, 'utf-8')
+}
 
 const API_URL = 'http://127.0.0.1:8080'
 
@@ -48,6 +77,7 @@ async function chat_completion(question) {
             n_keep: n_keep,
             n_predict: 256,
             stop: ["\n### Human:"], // stop completion after generating this
+            grammar,
             stream: true,
         })
     })

examples/server/json-schema-to-grammar.mjs.hpp

@@ -0,0 +1,311 @@
+unsigned char json_schema_to_grammar_mjs[] = {
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x53, 0x50, 0x41, 0x43, 0x45, 0x5f,
+  0x52, 0x55, 0x4c, 0x45, 0x20, 0x3d, 0x20, 0x27, 0x22, 0x20, 0x22, 0x3f,
+  0x27, 0x3b, 0x0a, 0x0a, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x50, 0x52,
+  0x49, 0x4d, 0x49, 0x54, 0x49, 0x56, 0x45, 0x5f, 0x52, 0x55, 0x4c, 0x45,
+  0x53, 0x20, 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x62, 0x6f, 0x6f, 0x6c,
+  0x65, 0x61, 0x6e, 0x3a, 0x20, 0x27, 0x28, 0x22, 0x74, 0x72, 0x75, 0x65,
+  0x22, 0x20, 0x7c, 0x20, 0x22, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x22, 0x29,
+  0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x6e,
+  0x75, 0x6d, 0x62, 0x65, 0x72, 0x3a, 0x20, 0x27, 0x28, 0x22, 0x2d, 0x22,
+  0x3f, 0x20, 0x28, 0x5b, 0x30, 0x2d, 0x39, 0x5d, 0x20, 0x7c, 0x20, 0x5b,
+  0x31, 0x2d, 0x39, 0x5d, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x5d, 0x2a, 0x29,
+  0x29, 0x20, 0x28, 0x22, 0x2e, 0x22, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x5d,
+  0x2b, 0x29, 0x3f, 0x20, 0x28, 0x5b, 0x65, 0x45, 0x5d, 0x20, 0x5b, 0x2d,
+  0x2b, 0x5d, 0x3f, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x5d, 0x2b, 0x29, 0x3f,
+  0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x69,
+  0x6e, 0x74, 0x65, 0x67, 0x65, 0x72, 0x3a, 0x20, 0x27, 0x28, 0x22, 0x2d,
+  0x22, 0x3f, 0x20, 0x28, 0x5b, 0x30, 0x2d, 0x39, 0x5d, 0x20, 0x7c, 0x20,
+  0x5b, 0x31, 0x2d, 0x39, 0x5d, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x5d, 0x2a,
+  0x29, 0x29, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x27, 0x2c, 0x0a, 0x20,
+  0x20, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x3a, 0x20, 0x60, 0x20, 0x22,
+  0x5c, 0x5c, 0x22, 0x22, 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x5b, 0x5e, 0x22, 0x5c, 0x5c, 0x5c, 0x5c, 0x5d, 0x20,
+  0x7c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x22, 0x5c,
+  0x5c, 0x5c, 0x5c, 0x22, 0x20, 0x28, 0x5b, 0x22, 0x5c, 0x5c, 0x5c, 0x5c,
+  0x2f, 0x62, 0x66, 0x6e, 0x72, 0x74, 0x5d, 0x20, 0x7c, 0x20, 0x22, 0x75,
+  0x22, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x61, 0x2d, 0x66, 0x41, 0x2d, 0x46,
+  0x5d, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x61, 0x2d, 0x66, 0x41, 0x2d, 0x46,
+  0x5d, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x61, 0x2d, 0x66, 0x41, 0x2d, 0x46,
+  0x5d, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x61, 0x2d, 0x66, 0x41, 0x2d, 0x46,
+  0x5d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x2a, 0x20,
+  0x22, 0x5c, 0x5c, 0x22, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x60,
+  0x2c, 0x0a, 0x20, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3a, 0x20, 0x27, 0x22,
+  0x6e, 0x75, 0x6c, 0x6c, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x27,
+  0x2c, 0x0a, 0x7d, 0x3b, 0x0a, 0x0a, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
+  0x49, 0x4e, 0x56, 0x41, 0x4c, 0x49, 0x44, 0x5f, 0x52, 0x55, 0x4c, 0x45,
+  0x5f, 0x43, 0x48, 0x41, 0x52, 0x53, 0x5f, 0x52, 0x45, 0x20, 0x3d, 0x20,
+  0x2f, 0x5b, 0x5e, 0x5c, 0x64, 0x41, 0x2d, 0x5a, 0x61, 0x2d, 0x7a, 0x2d,
+  0x5d, 0x2b, 0x2f, 0x67, 0x3b, 0x0a, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
+  0x47, 0x52, 0x41, 0x4d, 0x4d, 0x41, 0x52, 0x5f, 0x4c, 0x49, 0x54, 0x45,
+  0x52, 0x41, 0x4c, 0x5f, 0x45, 0x53, 0x43, 0x41, 0x50, 0x45, 0x5f, 0x52,
+  0x45, 0x20, 0x3d, 0x20, 0x2f, 0x5b, 0x5c, 0x6e, 0x5c, 0x72, 0x22, 0x5d,
+  0x2f, 0x67, 0x3b, 0x0a, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x47, 0x52,
+  0x41, 0x4d, 0x4d, 0x41, 0x52, 0x5f, 0x4c, 0x49, 0x54, 0x45, 0x52, 0x41,
+  0x4c, 0x5f, 0x45, 0x53, 0x43, 0x41, 0x50, 0x45, 0x53, 0x20, 0x3d, 0x20,
+  0x7b, 0x27, 0x5c, 0x72, 0x27, 0x3a, 0x20, 0x27, 0x5c, 0x5c, 0x72, 0x27,
+  0x2c, 0x20, 0x27, 0x5c, 0x6e, 0x27, 0x3a, 0x20, 0x27, 0x5c, 0x5c, 0x6e,
+  0x27, 0x2c, 0x20, 0x27, 0x22, 0x27, 0x3a, 0x20, 0x27, 0x5c, 0x5c, 0x22,
+  0x27, 0x7d, 0x3b, 0x0a, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20,
+  0x63, 0x6c, 0x61, 0x73, 0x73, 0x20, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61,
+  0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x72, 0x75, 0x63, 0x74, 0x6f,
+  0x72, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x29,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e,
+  0x5f, 0x70, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x20, 0x3d,
+  0x20, 0x70, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x20, 0x7c,
+  0x7c, 0x20, 0x7b, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68,
+  0x69, 0x73, 0x2e, 0x5f, 0x72, 0x75, 0x6c, 0x65, 0x73, 0x20, 0x3d, 0x20,
+  0x6e, 0x65, 0x77, 0x20, 0x4d, 0x61, 0x70, 0x28, 0x29, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x72, 0x75, 0x6c,
+  0x65, 0x73, 0x2e, 0x73, 0x65, 0x74, 0x28, 0x27, 0x73, 0x70, 0x61, 0x63,
+  0x65, 0x27, 0x2c, 0x20, 0x53, 0x50, 0x41, 0x43, 0x45, 0x5f, 0x52, 0x55,
+  0x4c, 0x45, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20,
+  0x5f, 0x66, 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x4c, 0x69, 0x74, 0x65, 0x72,
+  0x61, 0x6c, 0x28, 0x6c, 0x69, 0x74, 0x65, 0x72, 0x61, 0x6c, 0x29, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
+  0x65, 0x73, 0x63, 0x61, 0x70, 0x65, 0x64, 0x20, 0x3d, 0x20, 0x4a, 0x53,
+  0x4f, 0x4e, 0x2e, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x69, 0x66, 0x79,
+  0x28, 0x6c, 0x69, 0x74, 0x65, 0x72, 0x61, 0x6c, 0x29, 0x2e, 0x72, 0x65,
+  0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x47, 0x52, 0x41, 0x4d, 0x4d, 0x41, 0x52, 0x5f, 0x4c, 0x49, 0x54,
+  0x45, 0x52, 0x41, 0x4c, 0x5f, 0x45, 0x53, 0x43, 0x41, 0x50, 0x45, 0x5f,
+  0x52, 0x45, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x20,
+  0x3d, 0x3e, 0x20, 0x47, 0x52, 0x41, 0x4d, 0x4d, 0x41, 0x52, 0x5f, 0x4c,
+  0x49, 0x54, 0x45, 0x52, 0x41, 0x4c, 0x5f, 0x45, 0x53, 0x43, 0x41, 0x50,
+  0x45, 0x53, 0x5b, 0x6d, 0x5d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x29, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20,
+  0x60, 0x22, 0x24, 0x7b, 0x65, 0x73, 0x63, 0x61, 0x70, 0x65, 0x64, 0x7d,
+  0x22, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x5f,
+  0x61, 0x64, 0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x6e, 0x61, 0x6d, 0x65,
+  0x2c, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x65, 0x73, 0x63, 0x4e, 0x61, 0x6d,
+  0x65, 0x20, 0x3d, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x2e, 0x72, 0x65, 0x70,
+  0x6c, 0x61, 0x63, 0x65, 0x28, 0x49, 0x4e, 0x56, 0x41, 0x4c, 0x49, 0x44,
+  0x5f, 0x52, 0x55, 0x4c, 0x45, 0x5f, 0x43, 0x48, 0x41, 0x52, 0x53, 0x5f,
+  0x52, 0x45, 0x2c, 0x20, 0x27, 0x2d, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6b, 0x65, 0x79, 0x20, 0x3d, 0x20,
+  0x65, 0x73, 0x63, 0x4e, 0x61, 0x6d, 0x65, 0x3b, 0x0a, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f,
+  0x72, 0x75, 0x6c, 0x65, 0x73, 0x2e, 0x68, 0x61, 0x73, 0x28, 0x65, 0x73,
+  0x63, 0x4e, 0x61, 0x6d, 0x65, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, 0x68, 0x69, 0x73,
+  0x2e, 0x5f, 0x72, 0x75, 0x6c, 0x65, 0x73, 0x2e, 0x67, 0x65, 0x74, 0x28,
+  0x65, 0x73, 0x63, 0x4e, 0x61, 0x6d, 0x65, 0x29, 0x20, 0x3d, 0x3d, 0x3d,
+  0x20, 0x72, 0x75, 0x6c, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20,
+  0x6b, 0x65, 0x79, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20,
+  0x69, 0x20, 0x3d, 0x20, 0x30, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x77, 0x68, 0x69, 0x6c, 0x65, 0x20, 0x28, 0x74, 0x68, 0x69, 0x73,
+  0x2e, 0x5f, 0x72, 0x75, 0x6c, 0x65, 0x73, 0x2e, 0x68, 0x61, 0x73, 0x28,
+  0x60, 0x24, 0x7b, 0x65, 0x73, 0x63, 0x4e, 0x61, 0x6d, 0x65, 0x7d, 0x24,
+  0x7b, 0x69, 0x7d, 0x60, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x20, 0x2b, 0x3d, 0x20, 0x31, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x6b, 0x65, 0x79, 0x20, 0x3d, 0x20, 0x60, 0x24, 0x7b,
+  0x65, 0x73, 0x63, 0x4e, 0x61, 0x6d, 0x65, 0x7d, 0x24, 0x7b, 0x69, 0x7d,
+  0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x72, 0x75, 0x6c, 0x65,
+  0x73, 0x2e, 0x73, 0x65, 0x74, 0x28, 0x6b, 0x65, 0x79, 0x2c, 0x20, 0x72,
+  0x75, 0x6c, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65,
+  0x74, 0x75, 0x72, 0x6e, 0x20, 0x6b, 0x65, 0x79, 0x3b, 0x0a, 0x20, 0x20,
+  0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x76, 0x69, 0x73, 0x69, 0x74, 0x28, 0x73,
+  0x63, 0x68, 0x65, 0x6d, 0x61, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x29,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, 0x65, 0x20,
+  0x3d, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x74, 0x79, 0x70,
+  0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x72, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x20,
+  0x6e, 0x61, 0x6d, 0x65, 0x20, 0x7c, 0x7c, 0x20, 0x27, 0x72, 0x6f, 0x6f,
+  0x74, 0x27, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20,
+  0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x6f, 0x6e, 0x65, 0x4f,
+  0x66, 0x20, 0x7c, 0x7c, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e,
+  0x61, 0x6e, 0x79, 0x4f, 0x66, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x75, 0x6c,
+  0x65, 0x20, 0x3d, 0x20, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e,
+  0x6f, 0x6e, 0x65, 0x4f, 0x66, 0x20, 0x7c, 0x7c, 0x20, 0x73, 0x63, 0x68,
+  0x65, 0x6d, 0x61, 0x2e, 0x61, 0x6e, 0x79, 0x4f, 0x66, 0x29, 0x2e, 0x6d,
+  0x61, 0x70, 0x28, 0x28, 0x61, 0x6c, 0x74, 0x53, 0x63, 0x68, 0x65, 0x6d,
+  0x61, 0x2c, 0x20, 0x69, 0x29, 0x20, 0x3d, 0x3e, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x76, 0x69,
+  0x73, 0x69, 0x74, 0x28, 0x61, 0x6c, 0x74, 0x53, 0x63, 0x68, 0x65, 0x6d,
+  0x61, 0x2c, 0x20, 0x60, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x24,
+  0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3f, 0x20, 0x22, 0x2d, 0x22, 0x20,
+  0x3a, 0x20, 0x22, 0x22, 0x7d, 0x24, 0x7b, 0x69, 0x7d, 0x60, 0x29, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e,
+  0x28, 0x27, 0x20, 0x7c, 0x20, 0x27, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74,
+  0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x64, 0x64, 0x52, 0x75, 0x6c, 0x65,
+  0x28, 0x72, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x72,
+  0x75, 0x6c, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20,
+  0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, 0x66, 0x20, 0x28, 0x27, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x27, 0x20, 0x69, 0x6e, 0x20, 0x73, 0x63, 0x68, 0x65,
+  0x6d, 0x61, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e,
+  0x5f, 0x61, 0x64, 0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x72, 0x75, 0x6c,
+  0x65, 0x4e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e,
+  0x5f, 0x66, 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x4c, 0x69, 0x74, 0x65, 0x72,
+  0x61, 0x6c, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, 0x66, 0x20, 0x28, 0x27, 0x65,
+  0x6e, 0x75, 0x6d, 0x27, 0x20, 0x69, 0x6e, 0x20, 0x73, 0x63, 0x68, 0x65,
+  0x6d, 0x61, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x20, 0x3d,
+  0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x65, 0x6e, 0x75, 0x6d,
+  0x2e, 0x6d, 0x61, 0x70, 0x28, 0x76, 0x20, 0x3d, 0x3e, 0x20, 0x74, 0x68,
+  0x69, 0x73, 0x2e, 0x5f, 0x66, 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x4c, 0x69,
+  0x74, 0x65, 0x72, 0x61, 0x6c, 0x28, 0x76, 0x29, 0x29, 0x2e, 0x6a, 0x6f,
+  0x69, 0x6e, 0x28, 0x27, 0x20, 0x7c, 0x20, 0x27, 0x29, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20,
+  0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x64, 0x64, 0x52, 0x75, 0x6c,
+  0x65, 0x28, 0x72, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x2c, 0x20,
+  0x72, 0x75, 0x6c, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, 0x66, 0x20, 0x28, 0x73, 0x63,
+  0x68, 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d,
+  0x20, 0x27, 0x6f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x27, 0x20, 0x26, 0x26,
+  0x20, 0x27, 0x70, 0x72, 0x6f, 0x70, 0x65, 0x72, 0x74, 0x69, 0x65, 0x73,
+  0x27, 0x20, 0x69, 0x6e, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x29,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20,
+  0x54, 0x4f, 0x44, 0x4f, 0x3a, 0x20, 0x60, 0x72, 0x65, 0x71, 0x75, 0x69,
+  0x72, 0x65, 0x64, 0x60, 0x20, 0x6b, 0x65, 0x79, 0x77, 0x6f, 0x72, 0x64,
+  0x20, 0x28, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x70, 0x79, 0x74, 0x68, 0x6f,
+  0x6e, 0x20, 0x69, 0x6d, 0x70, 0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x61,
+  0x74, 0x69, 0x6f, 0x6e, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x4f, 0x72,
+  0x64, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f,
+  0x70, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70,
+  0x72, 0x6f, 0x70, 0x50, 0x61, 0x69, 0x72, 0x73, 0x20, 0x3d, 0x20, 0x4f,
+  0x62, 0x6a, 0x65, 0x63, 0x74, 0x2e, 0x65, 0x6e, 0x74, 0x72, 0x69, 0x65,
+  0x73, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x70, 0x72, 0x6f,
+  0x70, 0x65, 0x72, 0x74, 0x69, 0x65, 0x73, 0x29, 0x2e, 0x73, 0x6f, 0x72,
+  0x74, 0x28, 0x28, 0x61, 0x2c, 0x20, 0x62, 0x29, 0x20, 0x3d, 0x3e, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f,
+  0x20, 0x73, 0x6f, 0x72, 0x74, 0x20, 0x62, 0x79, 0x20, 0x70, 0x6f, 0x73,
+  0x69, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x69, 0x6e, 0x20, 0x70, 0x72, 0x6f,
+  0x70, 0x5f, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x20, 0x28, 0x69, 0x66, 0x20,
+  0x73, 0x70, 0x65, 0x63, 0x69, 0x66, 0x69, 0x65, 0x64, 0x29, 0x20, 0x74,
+  0x68, 0x65, 0x6e, 0x20, 0x62, 0x79, 0x20, 0x6b, 0x65, 0x79, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x41, 0x20, 0x3d, 0x20, 0x74, 0x79,
+  0x70, 0x65, 0x6f, 0x66, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64,
+  0x65, 0x72, 0x5b, 0x61, 0x5b, 0x30, 0x5d, 0x5d, 0x20, 0x3d, 0x3d, 0x3d,
+  0x20, 0x27, 0x6e, 0x75, 0x6d, 0x62, 0x65, 0x72, 0x27, 0x20, 0x3f, 0x20,
+  0x70, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x5b, 0x61, 0x5b,
+  0x30, 0x5d, 0x5d, 0x20, 0x3a, 0x20, 0x49, 0x6e, 0x66, 0x69, 0x6e, 0x69,
+  0x74, 0x79, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x42,
+  0x20, 0x3d, 0x20, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x70, 0x72,
+  0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x5b, 0x62, 0x5b, 0x30, 0x5d,
+  0x5d, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x6e, 0x75, 0x6d, 0x62, 0x65,
+  0x72, 0x27, 0x20, 0x3f, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64,
+  0x65, 0x72, 0x5b, 0x62, 0x5b, 0x30, 0x5d, 0x5d, 0x20, 0x3a, 0x20, 0x49,
+  0x6e, 0x66, 0x69, 0x6e, 0x69, 0x74, 0x79, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20,
+  0x6f, 0x72, 0x64, 0x65, 0x72, 0x41, 0x20, 0x2d, 0x20, 0x6f, 0x72, 0x64,
+  0x65, 0x72, 0x42, 0x20, 0x7c, 0x7c, 0x20, 0x61, 0x5b, 0x30, 0x5d, 0x2e,
+  0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x65, 0x43, 0x6f, 0x6d, 0x70, 0x61, 0x72,
+  0x65, 0x28, 0x62, 0x5b, 0x30, 0x5d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x20, 0x3d,
+  0x20, 0x27, 0x22, 0x7b, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x27,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x72, 0x6f, 0x70,
+  0x50, 0x61, 0x69, 0x72, 0x73, 0x2e, 0x66, 0x6f, 0x72, 0x45, 0x61, 0x63,
+  0x68, 0x28, 0x28, 0x5b, 0x70, 0x72, 0x6f, 0x70, 0x4e, 0x61, 0x6d, 0x65,
+  0x2c, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61,
+  0x5d, 0x2c, 0x20, 0x69, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x70, 0x72, 0x6f, 0x70, 0x52, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d,
+  0x65, 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x76, 0x69, 0x73,
+  0x69, 0x74, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x53, 0x63, 0x68, 0x65, 0x6d,
+  0x61, 0x2c, 0x20, 0x60, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x24,
+  0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3f, 0x20, 0x22, 0x2d, 0x22, 0x20,
+  0x3a, 0x20, 0x22, 0x22, 0x7d, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70, 0x4e,
+  0x61, 0x6d, 0x65, 0x7d, 0x60, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x69, 0x20, 0x3e, 0x20,
+  0x30, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x20, 0x2b, 0x3d, 0x20, 0x27,
+  0x20, 0x22, 0x2c, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x27, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x20,
+  0x2b, 0x3d, 0x20, 0x60, 0x20, 0x24, 0x7b, 0x74, 0x68, 0x69, 0x73, 0x2e,
+  0x5f, 0x66, 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x4c, 0x69, 0x74, 0x65, 0x72,
+  0x61, 0x6c, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x4e, 0x61, 0x6d, 0x65, 0x29,
+  0x7d, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x20, 0x22, 0x3a, 0x22, 0x20,
+  0x73, 0x70, 0x61, 0x63, 0x65, 0x20, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70,
+  0x52, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x7d, 0x60, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x20, 0x2b, 0x3d, 0x20,
+  0x27, 0x20, 0x22, 0x7d, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x27,
+  0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74,
+  0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x64,
+  0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x72, 0x75, 0x6c, 0x65, 0x4e, 0x61,
+  0x6d, 0x65, 0x2c, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x29, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, 0x66,
+  0x20, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, 0x65,
+  0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x61, 0x72, 0x72, 0x61, 0x79, 0x27,
+  0x20, 0x26, 0x26, 0x20, 0x27, 0x69, 0x74, 0x65, 0x6d, 0x73, 0x27, 0x20,
+  0x69, 0x6e, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x29, 0x20, 0x7b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x54, 0x4f,
+  0x44, 0x4f, 0x20, 0x60, 0x70, 0x72, 0x65, 0x66, 0x69, 0x78, 0x49, 0x74,
+  0x65, 0x6d, 0x73, 0x60, 0x20, 0x6b, 0x65, 0x79, 0x77, 0x6f, 0x72, 0x64,
+  0x20, 0x28, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x70, 0x79, 0x74, 0x68, 0x6f,
+  0x6e, 0x20, 0x69, 0x6d, 0x70, 0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x61,
+  0x74, 0x69, 0x6f, 0x6e, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x74, 0x65, 0x6d, 0x52, 0x75,
+  0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69,
+  0x73, 0x2e, 0x76, 0x69, 0x73, 0x69, 0x74, 0x28, 0x73, 0x63, 0x68, 0x65,
+  0x6d, 0x61, 0x2e, 0x69, 0x74, 0x65, 0x6d, 0x73, 0x2c, 0x20, 0x60, 0x24,
+  0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65,
+  0x20, 0x3f, 0x20, 0x22, 0x2d, 0x22, 0x20, 0x3a, 0x20, 0x22, 0x22, 0x7d,
+  0x69, 0x74, 0x65, 0x6d, 0x60, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x75, 0x6c, 0x65,
+  0x20, 0x3d, 0x20, 0x60, 0x22, 0x5b, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63,
+  0x65, 0x20, 0x28, 0x24, 0x7b, 0x69, 0x74, 0x65, 0x6d, 0x52, 0x75, 0x6c,
+  0x65, 0x4e, 0x61, 0x6d, 0x65, 0x7d, 0x20, 0x28, 0x22, 0x2c, 0x22, 0x20,
+  0x73, 0x70, 0x61, 0x63, 0x65, 0x20, 0x24, 0x7b, 0x69, 0x74, 0x65, 0x6d,
+  0x52, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x7d, 0x29, 0x2a, 0x29,
+  0x3f, 0x20, 0x22, 0x5d, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x60,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75,
+  0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x64, 0x64,
+  0x52, 0x75, 0x6c, 0x65, 0x28, 0x72, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d,
+  0x65, 0x2c, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x50, 0x52,
+  0x49, 0x4d, 0x49, 0x54, 0x49, 0x56, 0x45, 0x5f, 0x52, 0x55, 0x4c, 0x45,
+  0x53, 0x5b, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, 0x65,
+  0x5d, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x45,
+  0x72, 0x72, 0x6f, 0x72, 0x28, 0x60, 0x55, 0x6e, 0x72, 0x65, 0x63, 0x6f,
+  0x67, 0x6e, 0x69, 0x7a, 0x65, 0x64, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d,
+  0x61, 0x3a, 0x20, 0x24, 0x7b, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x73, 0x74,
+  0x72, 0x69, 0x6e, 0x67, 0x69, 0x66, 0x79, 0x28, 0x73, 0x63, 0x68, 0x65,
+  0x6d, 0x61, 0x29, 0x7d, 0x60, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65,
+  0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61,
+  0x64, 0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65,
+  0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x72, 0x6f, 0x6f, 0x74, 0x27, 0x20,
+  0x3f, 0x20, 0x27, 0x72, 0x6f, 0x6f, 0x74, 0x27, 0x20, 0x3a, 0x20, 0x73,
+  0x63, 0x68, 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, 0x65, 0x2c, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x50, 0x52, 0x49, 0x4d, 0x49,
+  0x54, 0x49, 0x56, 0x45, 0x5f, 0x52, 0x55, 0x4c, 0x45, 0x53, 0x5b, 0x73,
+  0x63, 0x68, 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, 0x65, 0x5d, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x66, 0x6f, 0x72,
+  0x6d, 0x61, 0x74, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x28, 0x29,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x67,
+  0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x20, 0x3d, 0x20, 0x27, 0x27, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x72,
+  0x75, 0x6c, 0x65, 0x73, 0x2e, 0x66, 0x6f, 0x72, 0x45, 0x61, 0x63, 0x68,
+  0x28, 0x28, 0x72, 0x75, 0x6c, 0x65, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65,
+  0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x20, 0x2b, 0x3d, 0x20,
+  0x60, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x20, 0x3a, 0x3a, 0x3d,
+  0x20, 0x24, 0x7b, 0x72, 0x75, 0x6c, 0x65, 0x7d, 0x5c, 0x6e, 0x60, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x67, 0x72, 0x61, 0x6d,
+  0x6d, 0x61, 0x72, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x7d, 0x0a
+};
+unsigned int json_schema_to_grammar_mjs_len = 3695;

examples/server/public/index.html

@@ -141,6 +141,7 @@
     } from '/index.js';
 
     import { llama } from '/completion.js';
+    import { SchemaConverter } from '/json-schema-to-grammar.mjs';
 
     const session = signal({
       prompt: "This is a conversation between user and llama, a friendly chatbot. respond in simple markdown.",
@@ -166,8 +167,139 @@
       mirostat: 0, // 0/1/2
       mirostat_tau: 5, // target entropy
       mirostat_eta: 0.1, // learning rate
+      grammar: '',
     })
 
+    /* START: Support for storing prompt templates and parameters in borwser LocalStorage */
+
+    const local_storage_storageKey = "llamacpp_server_local_storage";
+
+    function local_storage_setDataFromObject(tag, content) {
+      localStorage.setItem(local_storage_storageKey + '/' + tag, JSON.stringify(content));
+    }
+
+    function local_storage_setDataFromRawText(tag, content) {
+      localStorage.setItem(local_storage_storageKey + '/' + tag, content);
+    }
+
+    function local_storage_getDataAsObject(tag) {
+      const item = localStorage.getItem(local_storage_storageKey + '/' + tag);
+      if (!item) {
+        return null;
+      } else {
+        return JSON.parse(item);
+      }
+    }
+
+    function local_storage_getDataAsRawText(tag) {
+      const item = localStorage.getItem(local_storage_storageKey + '/' + tag);
+      if (!item) {
+        return null;
+      } else {
+        return item;
+      }
+    }
+
+    // create a container for user templates and settings
+
+    const savedUserTemplates = signal({})
+    const selectedUserTemplate = signal({ name: '', template: { session: {}, params: {} } })
+
+    // let's import locally saved templates and settings if there are any
+    // user templates and settings are stored in one object
+    // in form of { "templatename": "templatedata" } and { "settingstemplatename":"settingsdata" }
+
+    console.log('Importing saved templates')
+
+    let importedTemplates = local_storage_getDataAsObject('user_templates')
+
+    if (importedTemplates) {
+      // saved templates were successfuly imported.
+
+      console.log('Processing saved templates and updating default template')
+
+      //console.log(importedTemplates);
+      savedUserTemplates.value = importedTemplates;
+
+      //override default template
+      savedUserTemplates.value.default = { session: session.value, params: params.value }
+      local_storage_setDataFromObject('user_templates', savedUserTemplates.value)
+    } else {
+      // no saved templates detected.
+
+      console.log('Initializing LocalStorage and saving default template')
+
+      savedUserTemplates.value = { "default": { session: session.value, params: params.value } }
+      local_storage_setDataFromObject('user_templates', savedUserTemplates.value)
+    }
+
+    function userTemplateResetToDefault() {
+      console.log('Reseting themplate to default')
+      selectedUserTemplate.value.name = 'default';
+      selectedUserTemplate.value.data = savedUserTemplates.value['default'];
+    }
+
+    function userTemplateApply(t) {
+      session.value = t.data.session;
+      params.value = t.data.params;
+    }
+
+    function userTemplateResetToDefaultAndApply() {
+      userTemplateResetToDefault()
+      userTemplateApply(selectedUserTemplate.value)
+    }
+
+    function userTemplateLoadAndApplyAutosaved() {
+      // get autosaved last used template
+      let lastUsedTemplate = local_storage_getDataAsObject('user_templates_last')
+
+      if (lastUsedTemplate) {
+
+        console.log('Autosaved template found, restoring')
+
+        selectedUserTemplate.value = lastUsedTemplate
+      }
+      else {
+
+        console.log('No autosaved template found, using default template')
+        // no autosaved last used template was found, so load from default.
+
+        userTemplateResetToDefault()
+      }
+
+      console.log('Applying template')
+      // and update internal data from templates
+
+      userTemplateApply(selectedUserTemplate.value)
+    }
+
+    //console.log(savedUserTemplates.value)
+    //console.log(selectedUserTemplate.value)
+
+    function userTemplateAutosave() {
+      console.log('Template Autosave...')
+      if (selectedUserTemplate.value.name == 'default') {
+        // we don't want to save over default template, so let's create a new one
+        let newTemplateName = 'UserTemplate-' + Date.now().toString()
+        let newTemplate = { 'name': newTemplateName, 'data': { 'session': session.value, 'params': params.value } }
+
+        console.log('Saving as ' + newTemplateName)
+
+        // save in the autosave slot
+        local_storage_setDataFromObject('user_templates_last', newTemplate)
+
+        // and load it back and apply
+        userTemplateLoadAndApplyAutosaved()
+      } else {
+        local_storage_setDataFromObject('user_templates_last', { 'name': selectedUserTemplate.value.name, 'data': { 'session': session.value, 'params': params.value } })
+      }
+    }
+
+    console.log('Checking for autosaved last used template')
+    userTemplateLoadAndApplyAutosaved()
+
+    /* END: Support for storing prompt templates and parameters in browsers LocalStorage */
+
     const llamaStats = signal(null)
     const controller = signal(null)
 
@@ -304,6 +436,26 @@
       const updateParamsFloat = (el) => params.value = { ...params.value, [el.target.name]: parseFloat(el.target.value) }
       const updateParamsInt = (el) => params.value = { ...params.value, [el.target.name]: Math.floor(parseFloat(el.target.value)) }
 
+      const grammarJsonSchemaPropOrder = signal('')
+      const updateGrammarJsonSchemaPropOrder = (el) => grammarJsonSchemaPropOrder.value = el.target.value
+      const convertJSONSchemaGrammar = () => {
+        try {
+          const schema = JSON.parse(params.value.grammar)
+          const converter = new SchemaConverter(
+            grammarJsonSchemaPropOrder.value
+              .split(',')
+              .reduce((acc, cur, i) => ({...acc, [cur.trim()]: i}), {})
+          )
+          converter.visit(schema, '')
+          params.value = {
+            ...params.value,
+            grammar: converter.formatGrammar(),
+          }
+        } catch (e) {
+          alert(`Convert failed: ${e.message}`)
+        }
+      }
+
       const FloatField = ({label, max, min, name, step, value}) => {
         return html`
           <div>
@@ -324,8 +476,34 @@
         `
       };
 
+      const userTemplateReset = (e) => {
+        e.preventDefault();
+        userTemplateResetToDefaultAndApply()
+      }
+
+      const UserTemplateResetButton = () => {
+        if (selectedUserTemplate.value.name == 'default') {
+          return html`
+            <button disabled>Using default template</button>
+          `
+        }
+
+        return html`
+          <button onclick=${userTemplateReset}>Reset all to default</button>
+        `
+      };
+
+      useEffect(() => {
+        // autosave template on every change
+        userTemplateAutosave()
+      }, [session.value, params.value])
+
       return html`
         <form>
+          <fieldset>
+            <${UserTemplateResetButton}/>
+          </fieldset>
+
           <fieldset>
             <div>
               <label for="prompt">Prompt</label>
@@ -355,6 +533,13 @@
               <label for="template">Chat history template</label>
               <textarea id="template" name="historyTemplate" value="${session.value.historyTemplate}" rows=1 oninput=${updateSession}/>
             </div>
+
+            <div>
+              <label for="template">Grammar</label>
+              <textarea id="grammar" name="grammar" placeholder="Use gbnf or JSON Schema+convert" value="${params.value.grammar}" rows=4 oninput=${updateParams}/>
+              <input type="text" name="prop-order" placeholder="order: prop1,prop2,prop3" oninput=${updateGrammarJsonSchemaPropOrder} />
+              <button type="button" onclick=${convertJSONSchemaGrammar}>Convert JSON Schema</button>
+            </div>
           </fieldset>
 
           <fieldset class="two">

examples/server/public/json-schema-to-grammar.mjs

@@ -0,0 +1,112 @@
+const SPACE_RULE = '" "?';
+
+const PRIMITIVE_RULES = {
+  boolean: '("true" | "false") space',
+  number: '("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space',
+  integer: '("-"? ([0-9] | [1-9] [0-9]*)) space',
+  string: ` "\\"" (
+        [^"\\\\] |
+        "\\\\" (["\\\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+      )* "\\"" space`,
+  null: '"null" space',
+};
+
+const INVALID_RULE_CHARS_RE = /[^\dA-Za-z-]+/g;
+const GRAMMAR_LITERAL_ESCAPE_RE = /[\n\r"]/g;
+const GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"'};
+
+export class SchemaConverter {
+  constructor(propOrder) {
+    this._propOrder = propOrder || {};
+    this._rules = new Map();
+    this._rules.set('space', SPACE_RULE);
+  }
+
+  _formatLiteral(literal) {
+    const escaped = JSON.stringify(literal).replace(
+      GRAMMAR_LITERAL_ESCAPE_RE,
+      m => GRAMMAR_LITERAL_ESCAPES[m]
+    );
+    return `"${escaped}"`;
+  }
+
+  _addRule(name, rule) {
+    let escName = name.replace(INVALID_RULE_CHARS_RE, '-');
+    let key = escName;
+
+    if (this._rules.has(escName)) {
+      if (this._rules.get(escName) === rule) {
+        return key;
+      }
+
+      let i = 0;
+      while (this._rules.has(`${escName}${i}`)) {
+        i += 1;
+      }
+      key = `${escName}${i}`;
+    }
+
+    this._rules.set(key, rule);
+    return key;
+  }
+
+  visit(schema, name) {
+    const schemaType = schema.type;
+    const ruleName = name || 'root';
+
+    if (schema.oneOf || schema.anyOf) {
+      const rule = (schema.oneOf || schema.anyOf).map((altSchema, i) =>
+        this.visit(altSchema, `${name}${name ? "-" : ""}${i}`)
+      ).join(' | ');
+
+      return this._addRule(ruleName, rule);
+    } else if ('const' in schema) {
+      return this._addRule(ruleName, this._formatLiteral(schema.const));
+    } else if ('enum' in schema) {
+      const rule = schema.enum.map(v => this._formatLiteral(v)).join(' | ');
+      return this._addRule(ruleName, rule);
+    } else if (schemaType === 'object' && 'properties' in schema) {
+      // TODO: `required` keyword (from python implementation)
+      const propOrder = this._propOrder;
+      const propPairs = Object.entries(schema.properties).sort((a, b) => {
+        // sort by position in prop_order (if specified) then by key
+        const orderA = typeof propOrder[a[0]] === 'number' ? propOrder[a[0]] : Infinity;
+        const orderB = typeof propOrder[b[0]] === 'number' ? propOrder[b[0]] : Infinity;
+        return orderA - orderB || a[0].localeCompare(b[0]);
+      });
+
+      let rule = '"{" space';
+      propPairs.forEach(([propName, propSchema], i) => {
+        const propRuleName = this.visit(propSchema, `${name}${name ? "-" : ""}${propName}`);
+        if (i > 0) {
+          rule += ' "," space';
+        }
+        rule += ` ${this._formatLiteral(propName)} space ":" space ${propRuleName}`;
+      });
+      rule += ' "}" space';
+
+      return this._addRule(ruleName, rule);
+    } else if (schemaType === 'array' && 'items' in schema) {
+      // TODO `prefixItems` keyword (from python implementation)
+      const itemRuleName = this.visit(schema.items, `${name}${name ? "-" : ""}item`);
+      const rule = `"[" space (${itemRuleName} ("," space ${itemRuleName})*)? "]" space`;
+      return this._addRule(ruleName, rule);
+    } else {
+      if (!PRIMITIVE_RULES[schemaType]) {
+        throw new Error(`Unrecognized schema: ${JSON.stringify(schema)}`);
+      }
+      return this._addRule(
+        ruleName === 'root' ? 'root' : schemaType,
+        PRIMITIVE_RULES[schemaType]
+      );
+    }
+  }
+
+  formatGrammar() {
+    let grammar = '';
+    this._rules.forEach((rule, name) => {
+      grammar += `${name} ::= ${rule}\n`;
+    });
+    return grammar;
+  }
+}

examples/server/server.cpp

@@ -15,6 +15,7 @@
 #include "index.html.hpp"
 #include "index.js.hpp"
 #include "completion.js.hpp"
+#include "json-schema-to-grammar.mjs.hpp"
 
 #ifndef SERVER_VERBOSE
 #define SERVER_VERBOSE 1
@@ -668,6 +669,7 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     {
         fprintf(stdout, "  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
     }
+    fprintf(stdout, "  --numa                attempt optimizations that help on some NUMA systems\n");
 #ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
     fprintf(stdout, "  -ngl N, --n-gpu-layers N\n");
     fprintf(stdout, "                        number of layers to store in VRAM\n");
@@ -951,6 +953,10 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
         {
             params.use_mmap = false;
         }
+        else if (arg == "--numa")
+        {
+            params.numa = true;
+        }
         else if (arg == "--embedding")
         {
             params.embedding = true;
@@ -1019,7 +1025,7 @@ static json format_timings(llama_server_context &llama)
     assert(timings.n_eval == llama.num_tokens_predicted);
 
     return json{
-        {"prompt_n", timings.n_eval},
+        {"prompt_n", timings.n_p_eval},
         {"prompt_ms", timings.t_p_eval_ms},
         {"prompt_per_token_ms", timings.t_p_eval_ms / timings.n_p_eval},
         {"prompt_per_second", 1e3 / timings.t_p_eval_ms * timings.n_p_eval},
@@ -1048,7 +1054,6 @@ static json format_final_response(llama_server_context &llama, const std::string
         {"stopped_limit", llama.stopped_limit},
         {"stopping_word", llama.stopping_word},
         {"tokens_cached", llama.n_past},
-        {"tokens_predicted", llama.num_tokens_predicted},
         {"timings", format_timings(llama)},
     };
 
@@ -1226,6 +1231,12 @@ int main(int argc, char **argv)
         res.set_content(reinterpret_cast<const char*>(&completion_js), completion_js_len, "application/javascript");
         return false; });
 
+    // this is only called if no index.html is found in the public --path
+    svr.Get("/json-schema-to-grammar.mjs", [](const Request &, Response &res)
+            {
+        res.set_content(reinterpret_cast<const char*>(&json_schema_to_grammar_mjs), json_schema_to_grammar_mjs_len, "application/javascript");
+        return false; });
+
     svr.Post("/completion", [&llama](const Request &req, Response &res)
              {
         auto lock = llama.lock();

flake.nix

@@ -14,8 +14,6 @@
             with pkgs.darwin.apple_sdk_11_0.frameworks; [
               Accelerate
               MetalKit
-              MetalPerformanceShaders
-              MetalPerformanceShadersGraph
             ]
           else if isAarch32 && isDarwin then
             with pkgs.darwin.apple_sdk.frameworks; [

ggml-alloc.c

@@ -67,6 +67,8 @@ struct ggml_allocr {
     struct hash_node hash_table[GGML_GRAPH_HASHTABLE_SIZE];
     size_t max_size;
     bool measure;
+    int parse_seq[GGML_MAX_NODES];
+    bool has_parse_seq;
 
 #ifdef GGML_ALLOCATOR_DEBUG
     struct ggml_tensor * allocated_tensors[1024];
@@ -111,10 +113,10 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor)
 
     size_t max_avail = 0;
 
-    // find the best fitting free block
+    // find the best fitting free block besides the last block
     int best_fit_block = -1;
     size_t best_fit_size = SIZE_MAX;
-    for (int i = 0; i < alloc->n_free_blocks; i++) {
+    for (int i = 0; i < alloc->n_free_blocks - 1; i++) {
         struct free_block * block = &alloc->free_blocks[i];
         max_avail = MAX(max_avail, block->size);
         if (block->size >= size && block->size <= best_fit_size) {
@@ -126,11 +128,18 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor)
     AT_PRINTF("block %d\n", best_fit_block);
 
     if (best_fit_block == -1) {
+        // the last block is our last resort
+        struct free_block * block = &alloc->free_blocks[alloc->n_free_blocks - 1];
+        if (block->size >= size) {
+            best_fit_block = alloc->n_free_blocks - 1;
+            max_avail = MAX(max_avail, block->size);
+        } else {
             fprintf(stderr, "%s: not enough space in the buffer (needed %zu, largest block available %zu)\n",
                     __func__, size, max_avail);
             GGML_ASSERT(!"not enough space in the buffer");
         return;
         }
+    }
     struct free_block * block = &alloc->free_blocks[best_fit_block];
     void * addr = block->addr;
     block->addr = (char*)block->addr + size;
@@ -229,6 +238,17 @@ static void ggml_allocator_free_tensor(struct ggml_allocr * alloc, struct ggml_t
     alloc->n_free_blocks++;
 }
 
+void ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, int * list, int n) {
+    int pos = 0;
+    for (int i = 0; i < n; i++) {
+        if (list[i] != -1) {
+            alloc->parse_seq[pos] = list[i];
+            pos++;
+        }
+    }
+    alloc->has_parse_seq = true;
+}
+
 void ggml_allocr_reset(struct ggml_allocr * alloc) {
     alloc->n_free_blocks = 1;
     size_t align_offset = aligned_offset(alloc->data, 0, alloc->alignment);
@@ -248,6 +268,8 @@ struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment)
         /*.hash_table    = */ {{0}},
         /*.max_size      = */ 0,
         /*.measure       = */ false,
+        /*.parse_seq     = */ {0},
+        /*.has_parse_seq = */ false,
 #ifdef GGML_ALLOCATOR_DEBUG
         /*.allocated_tensors = */ = {0},
 #endif
@@ -275,6 +297,8 @@ struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
         /*.hash_table    = */ {{0}},
         /*.max_size      = */ 0,
         /*.measure       = */ true,
+        /*.parse_seq     = */ {0},
+        /*.has_parse_seq = */ false,
 #ifdef GGML_ALLOCATOR_DEBUG
         /*.allocated_tensors = */ = {0},
 #endif
@@ -473,7 +497,13 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
                 allocate_node(alloc, input);
             }
         }
-        for (int i = 0; i < gf->n_nodes; i++) {
+        for (int ind = 0; ind < gf->n_nodes; ind++) {
+            int i;
+            if (alloc->has_parse_seq) {
+                i = alloc->parse_seq[ind];
+            } else {
+                i = ind;
+            }
             struct ggml_tensor * node = gf->nodes[i];
 
             // allocate parents (leafs)

ggml-alloc.h

@@ -10,6 +10,10 @@ extern "C" {
 GGML_API struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment);
 GGML_API struct ggml_allocr * ggml_allocr_new_measure(size_t alignment);
 
+// tell the allocator to parse nodes following the order described in the list
+// you should call this if your graph are optimized to execute out-of-order
+GGML_API void   ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, int * list, int n);
+
 GGML_API void   ggml_allocr_free(struct ggml_allocr * alloc);
 GGML_API bool   ggml_allocr_is_measure(struct ggml_allocr * alloc);
 GGML_API void   ggml_allocr_reset(struct ggml_allocr * alloc);

ggml-cuda.h

@@ -8,29 +8,25 @@ extern "C" {
 
 #define GGML_CUDA_MAX_DEVICES       16
 
-void   ggml_init_cublas(void);
+GGML_API void   ggml_init_cublas(void);
-void   ggml_cuda_set_tensor_split(const float * tensor_split);
+GGML_API void * ggml_cuda_host_malloc(size_t size);
+GGML_API void   ggml_cuda_host_free(void * ptr);
 
-void   ggml_cuda_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+GGML_API bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+GGML_API void   ggml_cuda_set_tensor_split(const float * tensor_split);
-size_t ggml_cuda_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+GGML_API void   ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
-void   ggml_cuda_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
+GGML_API void   ggml_cuda_free_data(struct ggml_tensor * tensor);
+GGML_API void   ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
+GGML_API void   ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
+GGML_API void   ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
+GGML_API void   ggml_cuda_set_main_device(int main_device);
+GGML_API void   ggml_cuda_set_mul_mat_q(bool mul_mat_q);
+GGML_API void   ggml_cuda_set_scratch_size(size_t scratch_size);
+GGML_API void   ggml_cuda_free_scratch(void);
+GGML_API bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
 
-// TODO: export these with GGML_API
+GGML_API int    ggml_cuda_get_device_count(void);
-void * ggml_cuda_host_malloc(size_t size);
+GGML_API void   ggml_cuda_get_device_description(int device, char * description, size_t description_size);
-void   ggml_cuda_host_free(void * ptr);
-
-void   ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
-
-void   ggml_cuda_free_data(struct ggml_tensor * tensor);
-void   ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
-void   ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
-void   ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
-void   ggml_cuda_set_main_device(int main_device);
-void   ggml_cuda_set_mul_mat_q(bool mul_mat_q);
-void   ggml_cuda_set_scratch_size(size_t scratch_size);
-void   ggml_cuda_free_scratch(void);
-bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
 
 #ifdef  __cplusplus
 }

ggml-metal.h

@@ -63,10 +63,13 @@ void ggml_metal_get_tensor(struct ggml_metal_context * ctx, struct ggml_tensor *
 
 // try to find operations that can be run concurrently in the graph
 // you should run it again if the topology of your graph changes
-void ggml_metal_graph_find_concurrency(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
+void ggml_metal_graph_find_concurrency(struct ggml_metal_context * ctx, struct ggml_cgraph * gf, bool check_mem);
 
-// if the graph has been optimized for concurrently dispatch
+// if the graph has been optimized for concurrently dispatch, return length of the concur_list if optimized
-bool ggml_metal_if_optimized(struct ggml_metal_context * ctx);
+int ggml_metal_if_optimized(struct ggml_metal_context * ctx);
+
+// output the concur_list for ggml_alloc
+int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx);
 
 // same as ggml_graph_compute but uses Metal
 // creates gf->n_threads command buffers in parallel

ggml-metal.m

@@ -5,7 +5,6 @@
 #import <Foundation/Foundation.h>
 
 #import <Metal/Metal.h>
-#import <MetalPerformanceShaders/MetalPerformanceShaders.h>
 
 #undef MIN
 #undef MAX
@@ -79,6 +78,14 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(mul_mat_q4_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q5_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q6_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_f16_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_q4_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_q4_1_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_q2_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_q3_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_q4_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_q5_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_q6_K_f32);
     GGML_METAL_DECL_KERNEL(rope);
     GGML_METAL_DECL_KERNEL(alibi_f32);
     GGML_METAL_DECL_KERNEL(cpy_f32_f16);
@@ -110,13 +117,6 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
     ctx->n_buffers = 0;
     ctx->concur_list_len = 0;
 
-    // determine if we can use MPS
-    if (MPSSupportsMTLDevice(ctx->device)) {
-        fprintf(stderr, "%s: using MPS\n", __func__);
-    } else {
-        fprintf(stderr, "%s: not using MPS\n", __func__);
-        GGML_ASSERT(false && "MPS not supported");
-    }
 
 #if 0
     // compile from source string and show compile log
@@ -126,7 +126,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         ctx->library = [ctx->device newLibraryWithSource:msl_library_source options:nil error:&error];
         if (error) {
             fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
-            exit(1);
+            return NULL;
         }
     }
 #else
@@ -144,7 +144,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         NSString * src  = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:&error];
         if (error) {
             fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
-            exit(1);
+            return NULL;
         }
 
 #ifdef GGML_QKK_64
@@ -156,17 +156,22 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
 #endif
         if (error) {
             fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
-            exit(1);
+            return NULL;
         }
     }
 #endif
 
     // load kernels
     {
+        NSError * error = nil;
 #define GGML_METAL_ADD_KERNEL(name) \
         ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \
-        ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:nil]; \
+        ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:&error]; \
-        fprintf(stderr, "%s: loaded %-32s %16p\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name);
+        fprintf(stderr, "%s: loaded %-32s %16p\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name); \
+        if (error) { \
+            fprintf(stderr, "%s: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
+            return NULL; \
+        }
 
         GGML_METAL_ADD_KERNEL(add);
         GGML_METAL_ADD_KERNEL(add_row);
@@ -196,6 +201,14 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(mul_mat_q4_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q5_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q6_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mm_f16_f32);
+        GGML_METAL_ADD_KERNEL(mul_mm_q4_0_f32);
+        GGML_METAL_ADD_KERNEL(mul_mm_q4_1_f32);
+        GGML_METAL_ADD_KERNEL(mul_mm_q2_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mm_q3_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mm_q4_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mm_q5_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mm_q6_K_f32);
         GGML_METAL_ADD_KERNEL(rope);
         GGML_METAL_ADD_KERNEL(alibi_f32);
         GGML_METAL_ADD_KERNEL(cpy_f32_f16);
@@ -228,11 +241,12 @@ void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb) {
     ctx->n_cb = n_cb;
 }
 
-bool ggml_metal_if_optimized(struct ggml_metal_context * ctx) {
+int ggml_metal_if_optimized(struct ggml_metal_context * ctx) {
-    if (ctx->concur_list_len) {
+    return ctx->concur_list_len;
-        return true;
+}
-    }
+
-    return false;
+int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx) {
+    return ctx->concur_list;
 }
 
 // finds the Metal buffer that contains the tensor data on the GPU device
@@ -375,7 +389,7 @@ void ggml_metal_get_tensor(
 
 void ggml_metal_graph_find_concurrency(
         struct ggml_metal_context * ctx,
-        struct ggml_cgraph * gf) {
+        struct ggml_cgraph * gf, bool check_mem) {
     int search_depth = gf->n_nodes; //we only find concurrency in this range to avoid wasting too much time
     int nodes_unused[GGML_MAX_CONCUR];
 
@@ -422,7 +436,7 @@ void ggml_metal_graph_find_concurrency(
                         }
                     }
                 }
-                if (exe_flag) {
+                if (exe_flag && check_mem) {
                     // check if nodes[i]'s data will be overwritten by a node before nodes[i].
                     // if node[5] and node[3] write to the same memory region, then we can't issue node[5] before node[3]
                     int64_t data_start = (int64_t) gf->nodes[i]->data;
@@ -506,7 +520,7 @@ void ggml_metal_graph_compute(
 
             id<MTLCommandBuffer> command_buffer = command_buffers[cb_idx];
 
-            id<MTLComputeCommandEncoder> encoder = nil;
+            id<MTLComputeCommandEncoder> encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
 
             const int node_start =                                  (cb_idx + 0) * n_nodes_per_cb;
             const int node_end   = (cb_idx == n_cb - 1) ? n_nodes : (cb_idx + 1) * n_nodes_per_cb;
@@ -515,10 +529,6 @@ void ggml_metal_graph_compute(
                 const int i = has_concur ? ctx->concur_list[ind] : ind;
 
                 if (i == -1) {
-                    if (encoder == nil) {
-                        encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
-                        continue;
-                    }
                     [encoder memoryBarrierWithScope:MTLBarrierScopeBuffers];
                     continue;
                 }
@@ -592,10 +602,6 @@ void ggml_metal_graph_compute(
                         } break;
                     case GGML_OP_ADD:
                         {
-                            if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
-                            }
-
                             if (ggml_nelements(src1) == ne10) {
                                 // src1 is a row
                                 [encoder setComputePipelineState:ctx->pipeline_add_row];
@@ -613,10 +619,6 @@ void ggml_metal_graph_compute(
                         } break;
                     case GGML_OP_MUL:
                         {
-                            if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
-                            }
-
                             if (ggml_nelements(src1) == ne10) {
                                 // src1 is a row
                                 [encoder setComputePipelineState:ctx->pipeline_mul_row];
@@ -634,10 +636,6 @@ void ggml_metal_graph_compute(
                         } break;
                     case GGML_OP_SCALE:
                         {
-                            if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
-                            }
-
                             const float scale = *(const float *) src1->data;
 
                             [encoder setComputePipelineState:ctx->pipeline_scale];
@@ -653,10 +651,6 @@ void ggml_metal_graph_compute(
                         switch (ggml_get_unary_op(gf->nodes[i])) {
                             case GGML_UNARY_OP_SILU:
                                 {
-                                    if (encoder == nil) {
-                                        encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
-                                    }
-
                                     [encoder setComputePipelineState:ctx->pipeline_silu];
                                     [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                                     [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
@@ -667,10 +661,6 @@ void ggml_metal_graph_compute(
                                 } break;
                             case GGML_UNARY_OP_RELU:
                                 {
-                                    if (encoder == nil) {
-                                        encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
-                                    }
-
                                     [encoder setComputePipelineState:ctx->pipeline_relu];
                                     [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                                     [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
@@ -681,10 +671,6 @@ void ggml_metal_graph_compute(
                                 } break;
                             case GGML_UNARY_OP_GELU:
                                 {
-                                    if (encoder == nil) {
-                                        encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
-                                    }
-
                                     [encoder setComputePipelineState:ctx->pipeline_gelu];
                                     [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                                     [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
@@ -701,10 +687,6 @@ void ggml_metal_graph_compute(
                         } break;
                     case GGML_OP_SOFT_MAX:
                         {
-                            if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
-                            }
-
                             const int nth = 32;
 
                             [encoder setComputePipelineState:ctx->pipeline_soft_max];
@@ -719,10 +701,6 @@ void ggml_metal_graph_compute(
                         } break;
                     case GGML_OP_DIAG_MASK_INF:
                         {
-                            if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
-                            }
-
                             const int n_past = ((int32_t *)(dst->op_params))[0];
 
                             [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
@@ -740,53 +718,43 @@ void ggml_metal_graph_compute(
 
                             GGML_ASSERT(ne00 == ne10);
                             // GGML_ASSERT(ne02 == ne12); // Should be checked on individual data types until broadcast is implemented everywhere
+                            uint gqa = ne12/ne02;
                             GGML_ASSERT(ne03 == ne13);
 
+                            // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
+                            // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
                             if (ggml_is_contiguous(src0) &&
                                 ggml_is_contiguous(src1) &&
-                                (src0t == GGML_TYPE_F32 || src0t == GGML_TYPE_F16) && ne11 > 1) {
+                                src1t == GGML_TYPE_F32 &&
-
+                                [ctx->device supportsFamily:MTLGPUFamilyApple7] &&
-                                if (encoder != nil) {
+                                ne00%32 == 0 &&
-                                    [encoder endEncoding];
+                                ne11 > 1) {
-                                    encoder = nil;
+                                    switch (src0->type) {
+                                        case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32]; break;
+                                        case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break;
+                                        case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break;
+                                        case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q2_K_f32]; break;
+                                        case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q3_K_f32]; break;
+                                        case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_K_f32]; break;
+                                        case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_K_f32]; break;
+                                        case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q6_K_f32]; break;
+                                        default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
                                     }
-
+                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                                MPSDataType src0dt = src0t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
+                                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                                MPSDataType src1dt = src1t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
+                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-
+                                    [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
-                                // for F32 x F32 we use MPS
+                                    [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
-                                MPSMatrixDescriptor * desc0 = [MPSMatrixDescriptor
+                                    [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:5];
-                                    matrixDescriptorWithRows:ne01 columns:ne00 rowBytes:src0->nb[1] dataType:src0dt];
+                                    [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:6];
-
+                                    [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:7];
-                                MPSMatrixDescriptor * desc1 = [MPSMatrixDescriptor
+                                    [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:8];
-                                    matrixDescriptorWithRows:ne11 columns:ne10 rowBytes:src1->nb[1] dataType:src1dt];
+                                    [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:9];
-
+                                    [encoder setBytes:&gqa length:sizeof(gqa) atIndex:10];
-                                MPSMatrixDescriptor * desc  = [MPSMatrixDescriptor
+                                    [encoder setThreadgroupMemoryLength:8192 atIndex:0];
-                                    matrixDescriptorWithRows:ne1 columns:ne0 rowBytes:dst->nb[1] dataType:MPSDataTypeFloat32];
+                                    [encoder dispatchThreadgroups:MTLSizeMake( (ne11+31)/32, (ne01+63) / 64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
-
-                                MPSMatrixMultiplication * mul = [[MPSMatrixMultiplication alloc]
-                                    initWithDevice:ctx->device transposeLeft:false transposeRight:true
-                                        resultRows:ne11 resultColumns:ne01 interiorColumns:ne00 alpha:1.0 beta:0.0];
-
-                                // we need to do ne12 multiplications
-                                // TODO: is there a way to do this in parallel - currently very slow ..
-                                // TODO: might be possible to offload part of the computation to ANE using Accelerate's CBLAS
-                                for (int64_t i02 = 0; i02 < ne12; ++i02) {
-                                    size_t offs_src0_cur = offs_src0 + i02/(ne12/ne02)*nb02; // gqa not used for now
-                                    size_t offs_src1_cur = offs_src1 + i02*nb12;
-                                    size_t offs_dst_cur  = offs_dst  + i02*nb2;
-
-                                    MPSMatrix * mat_src0 = [[MPSMatrix alloc] initWithBuffer:id_src0 offset:offs_src0_cur descriptor:desc0];
-                                    MPSMatrix * mat_src1 = [[MPSMatrix alloc] initWithBuffer:id_src1 offset:offs_src1_cur descriptor:desc1];
-                                    MPSMatrix * mat_dst  = [[MPSMatrix alloc] initWithBuffer:id_dst  offset:offs_dst_cur  descriptor:desc ];
-
-                                    [mul encodeToCommandBuffer:command_buffer leftMatrix:mat_src1 rightMatrix:mat_src0 resultMatrix:mat_dst];
                                 }
-                            } else {
+                            else {
-                                if (encoder == nil) {
-                                    encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
-                                }
-
                                 int nth0 = 32;
                                 int nth1 = 1;
 
@@ -885,23 +853,24 @@ void ggml_metal_graph_compute(
                                 [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:14];
                                 [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:15];
                                 [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:16];
+                                [encoder setBytes:&gqa length:sizeof(gqa) atIndex:17];
 
                                 if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
                                     src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7) / 8, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7) / 8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
                                 else if (src0t == GGML_TYPE_Q3_K) {
 #ifdef GGML_QKK_64
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
 #else
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+3)/4, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
 #endif
                                 }
                                 else if (src0t == GGML_TYPE_Q5_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3) / 4, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3) / 4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
                                 else if (src0t == GGML_TYPE_Q6_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 } else {
                                     [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
                                     [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
@@ -910,10 +879,6 @@ void ggml_metal_graph_compute(
                         } break;
                     case GGML_OP_GET_ROWS:
                         {
-                            if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
-                            }
-
                             switch (src0->type) {
                                 case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_get_rows_f16]; break;
                                 case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
@@ -939,10 +904,6 @@ void ggml_metal_graph_compute(
                         } break;
                     case GGML_OP_RMS_NORM:
                         {
-                            if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
-                            }
-
                             float eps;
                             memcpy(&eps, dst->op_params, sizeof(float));
 
@@ -962,10 +923,6 @@ void ggml_metal_graph_compute(
                         } break;
                     case GGML_OP_NORM:
                         {
-                            if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
-                            }
-
                             const float eps = 1e-5f;
 
                             const int nth = 256;
@@ -984,10 +941,6 @@ void ggml_metal_graph_compute(
                         } break;
                     case GGML_OP_ALIBI:
                         {
-                            if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
-                            }
-
                             GGML_ASSERT((src0t == GGML_TYPE_F32));
 
                             const int n_past = ((int32_t *) dst->op_params)[0]; UNUSED(n_past);
@@ -1027,10 +980,6 @@ void ggml_metal_graph_compute(
                         } break;
                     case GGML_OP_ROPE:
                         {
-                            if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
-                            }
-
                             const int n_past = ((int32_t *) dst->op_params)[0];
                             const int n_dims = ((int32_t *) dst->op_params)[1];
                             const int mode   = ((int32_t *) dst->op_params)[2];
@@ -1071,10 +1020,6 @@ void ggml_metal_graph_compute(
                     case GGML_OP_CPY:
                     case GGML_OP_CONT:
                         {
-                            if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
-                            }
-
                             const int nth = 32;
 
                             switch (src0t) {

llama-util.h

@@ -271,20 +271,29 @@ struct llama_mmap {
             throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str()));
         }
 
-        #if _WIN32_WINNT >= _WIN32_WINNT_WIN8
         if (prefetch) {
-            // Advise the kernel to preload the mapped memory
+            // The PrefetchVirtualMemory API is only present on Windows 8 and above, so we
+            // will dynamically load it using GetProcAddress.
+            BOOL (WINAPI *pPrefetchVirtualMemory) (HANDLE, ULONG_PTR, PWIN32_MEMORY_RANGE_ENTRY, ULONG);
+            HMODULE hKernel32;
+
+            // This call is guaranteed to succeed.
+            hKernel32 = GetModuleHandleW(L"kernel32.dll");
+
+            // This call may fail if on a pre-Win8 system.
+            pPrefetchVirtualMemory = reinterpret_cast<decltype(pPrefetchVirtualMemory)> (GetProcAddress(hKernel32, "PrefetchVirtualMemory"));
+
+            if (pPrefetchVirtualMemory) {
+                // Advise the kernel to preload the mapped memory.
                 WIN32_MEMORY_RANGE_ENTRY range;
                 range.VirtualAddress = addr;
                 range.NumberOfBytes = (SIZE_T)size;
-            if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
+                if (!pPrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
                     fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
                             llama_format_win_err(GetLastError()).c_str());
                 }
             }
-        #else
+        }
-        #pragma message("warning: You are building for pre-Windows 8; prefetch not supported")
-        #endif // _WIN32_WINNT >= _WIN32_WINNT_WIN8
     }
 
     ~llama_mmap() {

llama.cpp

@@ -63,7 +63,7 @@ static void llama_log_callback_default(llama_log_level level, const char * text,
 #define LLAMA_LOG_ERROR(...) llama_log_internal(LLAMA_LOG_LEVEL_ERROR, __VA_ARGS__)
 
 
-#if !defined(GGML_USE_CUBLAS) && !defined(GGML_USE_METAL)
+#if !defined(GGML_USE_CUBLAS)
 #include "ggml-alloc.h"
 #define LLAMA_USE_ALLOCATOR
 #else
@@ -115,9 +115,9 @@ static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph *
 // memory sizes (calculated for n_batch == 512)
 //
 
-static const std::map<e_model, size_t> & MEM_REQ_SCRATCH0(int n_ctx)
+static std::map<e_model, size_t> MEM_REQ_SCRATCH0(int n_ctx)
 {
-    static std::map<e_model, size_t> k_sizes = {
+    std::map<e_model, size_t> k_sizes = {
         { MODEL_3B,   ((size_t) n_ctx / 16ull +  92ull) * MB },
         { MODEL_7B,   ((size_t) n_ctx / 16ull + 100ull) * MB },
         { MODEL_13B,  ((size_t) n_ctx / 12ull + 120ull) * MB },
@@ -984,7 +984,7 @@ int64_t llama_time_us() {
 // model loading
 //
 
-static const char *llama_file_version_name(llama_file_version version) {
+static const char * llama_file_version_name(llama_file_version version) {
     switch (version) {
         case LLAMA_FILE_VERSION_GGML: return "'ggml' (old version with low tokenizer quality and no mmap support)";
         case LLAMA_FILE_VERSION_GGMF_V1: return "ggmf v1 (old version with no mmap support)";
@@ -996,7 +996,7 @@ static const char *llama_file_version_name(llama_file_version version) {
     return "unknown";
 }
 
-static const char *llama_ftype_name(enum llama_ftype ftype) {
+const char * llama_ftype_name(enum llama_ftype ftype) {
     switch (ftype) {
         case LLAMA_FTYPE_ALL_F32:     return "all F32";
         case LLAMA_FTYPE_MOSTLY_F16:  return "mostly F16";
@@ -1021,7 +1021,7 @@ static const char *llama_ftype_name(enum llama_ftype ftype) {
     }
 }
 
-static const char *llama_model_type_name(e_model type) {
+static const char * llama_model_type_name(e_model type) {
     switch (type) {
         case MODEL_3B: return "3B";
         case MODEL_7B: return "7B";
@@ -1609,11 +1609,11 @@ static struct ggml_cgraph * llama_build_graph(
             ggml_set_name(Q, "Q");
 
             struct ggml_tensor * K =
-                ggml_permute(ctx0,
+                ggml_view_3d(ctx0, kv_self.k,
-                        ggml_reshape_3d(ctx0,
+                        n_embd_head, n_past + N, n_head_kv,
-                            ggml_view_1d(ctx0, kv_self.k, (n_past + N)*n_embd_gqa, il*n_ctx*ggml_element_size(kv_self.k)*n_embd_gqa),
+                        ggml_element_size(kv_self.k)*n_embd_gqa,
-                            n_embd_head, n_head_kv, n_past + N),
+                        ggml_element_size(kv_self.k)*n_embd_head,
-                        0, 2, 1, 3);
+                        ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il);
             offload_func_kq(K);
             ggml_set_name(K, "K");
 
@@ -1642,9 +1642,9 @@ static struct ggml_cgraph * llama_build_graph(
             struct ggml_tensor * V =
                 ggml_view_3d(ctx0, kv_self.v,
                         n_past + N, n_embd_head, n_head_kv,
-                        n_ctx*ggml_element_size(kv_self.v),
+                        ggml_element_size(kv_self.v)*n_ctx,
-                        n_ctx*ggml_element_size(kv_self.v)*n_embd_head,
+                        ggml_element_size(kv_self.v)*n_ctx*n_embd_head,
-                        n_ctx*ggml_element_size(kv_self.v)*n_embd_gqa*il);
+                        ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il);
             offload_func_v(V);
             ggml_set_name(V, "V");
 
@@ -1801,6 +1801,13 @@ static bool llama_eval_internal(
 
     LLAMA_ASSERT((!tokens && embd) || (tokens && !embd));
 
+    LLAMA_ASSERT(n_tokens > 0);
+    LLAMA_ASSERT(n_past >= 0);
+    LLAMA_ASSERT(n_threads > 0);
+    // TODO: keep the values of n_batch and n_ctx
+    // LLAMA_ASSERT(n_tokens <= n_batch);
+    // LLAMA_ASSERT(n_past + n_tokens <= n_ctx);
+
     const int64_t t_start_us = ggml_time_us();
 
 #ifdef GGML_USE_MPI
@@ -1848,11 +1855,7 @@ static bool llama_eval_internal(
 #endif
 
 #ifdef GGML_USE_METAL
-    if (lctx.ctx_metal && N == 1) {
+    if (lctx.ctx_metal) {
-        // TODO: disabled until #2413 is resolved
-        //if (!ggml_metal_if_optimized(lctx.ctx_metal)) {
-        //    ggml_metal_graph_find_concurrency(lctx.ctx_metal, gf);
-        //}
         ggml_metal_set_n_cb     (lctx.ctx_metal, n_threads);
         ggml_metal_graph_compute(lctx.ctx_metal, gf);
         ggml_metal_get_tensor   (lctx.ctx_metal, res);
@@ -1860,22 +1863,6 @@ static bool llama_eval_internal(
             ggml_metal_get_tensor(lctx.ctx_metal, embeddings);
         }
     } else {
-        // IMPORTANT:
-        // Since we don't have efficient Matrix x Matrix Metal multiplication yet, we fallback to vanilla
-        // ggml_graph_compute(). It uses Apple's Accelerate CBLAS API which takes advantage of the ANE or the AMX
-        // coprocessor.
-        //
-        // When we implement Matrix x Matrix Metal multiplication, we can avoid this branch.
-        // But for now, we have focused only on Matrix x Vector Metal multiplication.
-        //
-        // TODO: avoid these syncs via shared memory (ref #1696)
-        //
-        if (lctx.ctx_metal) {
-            // We need to sync the GPU KV cache with the CPU KV cache
-            ggml_metal_get_tensor(lctx.ctx_metal, kv_self.k);
-            ggml_metal_get_tensor(lctx.ctx_metal, kv_self.v);
-        }
-
         ggml_graph_compute_helper(lctx.work_buffer, gf, n_threads);
     }
 #else
@@ -2100,37 +2087,81 @@ static std::vector<llama_vocab::id> llama_tokenize(const llama_vocab & vocab, co
 // grammar - internal
 //
 
+struct llama_partial_utf8 {
+    uint32_t value;    // bit value so far (unshifted)
+    int      n_remain; // num bytes remaining; -1 indicates invalid sequence
+};
+
 struct llama_grammar {
     const std::vector<std::vector<llama_grammar_element>>   rules;
     std::vector<std::vector<const llama_grammar_element *>> stacks;
+
+    // buffer for partially generated UTF-8 sequence from accepted tokens
+    llama_partial_utf8                                      partial_utf8;
 };
 
 struct llama_grammar_candidate {
     size_t               index;
     const uint32_t     * code_points;
+    llama_partial_utf8   partial_utf8;
 };
 
-// NOTE: assumes valid utf8 (but checks for overrun)
+// Decodes a UTF-8 string which may end in an incomplete sequence. Adds a terminating 0 for use as
-// adds a terminating 0 for use as pointer
+// pointer. If an invalid sequence is encountered, returns `llama_partial_utf8.n_remain == -1`.
-std::vector<uint32_t> decode_utf8(const char * src) {
+std::pair<std::vector<uint32_t>, llama_partial_utf8> decode_utf8(
-    static const int      lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
+        const char         * src,
+        llama_partial_utf8   partial_start) {
+    static const int      lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 2, 2, 3, 4 };
     const char          * pos      = src;
     std::vector<uint32_t> code_points;
+    uint32_t              value    = partial_start.value;
+    int                   n_remain = partial_start.n_remain;
+
+    // continue previous decode, if applicable
+    while (*pos != 0 && n_remain > 0) {
+        uint8_t next_byte = static_cast<uint8_t>(*pos);
+        if ((next_byte >> 6) != 2) {
+            // invalid sequence, abort
+            code_points.push_back(0);
+            return std::make_pair(std::move(code_points), llama_partial_utf8{ 0, -1 });
+        }
+        value = (value << 6) + (next_byte & 0x3F);
+        ++pos;
+        --n_remain;
+    }
+
+    if (partial_start.n_remain > 0 && n_remain == 0) {
+        code_points.push_back(value);
+    }
+
+    // decode any subsequent utf-8 sequences, which may end in an incomplete one
     while (*pos != 0) {
         uint8_t  first_byte = static_cast<uint8_t>(*pos);
         uint8_t  highbits   = first_byte >> 4;
-        int      len        = lookup[highbits];
+                 n_remain   = lookup[highbits] - 1;
-        uint8_t  mask       = (1 << (8 - len)) - 1;
+
-        uint32_t value      = first_byte & mask;
+        if (n_remain < 0) {
-        const char * end    = pos + len; // may overrun!
+            // invalid sequence, abort
-        ++pos;
+            code_points.clear();
-        for ( ; pos < end && *pos != 0; ++pos) {
+            code_points.push_back(0);
-            value = (value << 6) + (static_cast<uint8_t>(*pos) & 0x3F);
+            return std::make_pair(std::move(code_points), llama_partial_utf8{ 0, n_remain });
         }
+
+        uint8_t  mask       = (1 << (7 - n_remain)) - 1;
+                 value      = first_byte & mask;
+        ++pos;
+        while (*pos != 0 && n_remain > 0) {
+            value = (value << 6) + (static_cast<uint8_t>(*pos) & 0x3F);
+            ++pos;
+            --n_remain;
+        }
+        if (n_remain == 0) {
             code_points.push_back(value);
         }
+    }
     code_points.push_back(0);
-    return code_points;
+
+    return std::make_pair(std::move(code_points), llama_partial_utf8{ value, n_remain });
 }
 
 // returns true iff pos points to the end of one of the definitions of a rule
@@ -2167,6 +2198,56 @@ static std::pair<bool, const llama_grammar_element *> llama_grammar_match_char(
     return std::make_pair(found == is_positive_char, pos);
 }
 
+// returns true iff some continuation of the given partial UTF-8 sequence could satisfy the char
+// range at pos (regular or inverse range)
+// asserts that pos is pointing to a char range element
+static bool llama_grammar_match_partial_char(
+        const llama_grammar_element * pos,
+        const llama_partial_utf8      partial_utf8) {
+
+    bool is_positive_char = pos->type == LLAMA_GRETYPE_CHAR;
+    LLAMA_ASSERT(is_positive_char || pos->type == LLAMA_GRETYPE_CHAR_NOT);
+
+    uint32_t partial_value = partial_utf8.value;
+    int      n_remain      = partial_utf8.n_remain;
+
+    // invalid sequence or 7-bit char split across 2 bytes (overlong)
+    if (n_remain < 0 || (n_remain == 1 && partial_value < 2)) {
+        return false;
+    }
+
+    // range of possible code points this partial UTF-8 sequence could complete to
+    uint32_t low  = partial_value << (n_remain * 6);
+    uint32_t high = low | ((1 << (n_remain * 6)) - 1);
+
+    if (low == 0) {
+        if (n_remain == 2) {
+            low = 1 << 11;
+        } else if (n_remain == 3) {
+            low = 1 << 16;
+        }
+    }
+
+    do {
+        if (pos[1].type == LLAMA_GRETYPE_CHAR_RNG_UPPER) {
+            // inclusive range, e.g. [a-z]
+            if (pos->value <= high && low <= pos[1].value) {
+                return is_positive_char;
+            }
+            pos += 2;
+        } else {
+            // exact char match, e.g. [a] or "a"
+            if (low <= pos->value && pos->value <= high) {
+                return is_positive_char;
+            }
+            pos += 1;
+        }
+    } while (pos->type == LLAMA_GRETYPE_CHAR_ALT);
+
+    return !is_positive_char;
+}
+
+
 // transforms a grammar pushdown stack into N possible stacks, all ending
 // at a character range (terminal element)
 static void llama_grammar_advance_stack(
@@ -2267,8 +2348,11 @@ static std::vector<llama_grammar_candidate> llama_grammar_reject_candidates_for_
     std::vector<llama_grammar_candidate> rejects;
 
     if (stack.empty()) {
-        // accept nothing; EOS is handled elsewhere
+        for (auto tok : candidates) {
-        rejects.insert(rejects.end(), candidates.begin(), candidates.end());
+            if (*tok.code_points != 0 || tok.partial_utf8.n_remain != 0) {
+                rejects.push_back(tok);
+            }
+        }
         return rejects;
     }
 
@@ -2276,10 +2360,15 @@ static std::vector<llama_grammar_candidate> llama_grammar_reject_candidates_for_
 
     std::vector<llama_grammar_candidate> next_candidates;
     for (auto tok : candidates) {
-        if (llama_grammar_match_char(stack_pos, tok.code_points[0]).first) {
+        if (*tok.code_points == 0) {
-            if (tok.code_points[1] != 0) {
+            // reached end of full codepoints in token, reject iff it ended in a partial sequence
-                next_candidates.push_back({ tok.index, tok.code_points + 1 });
+            // that cannot satisfy this position in grammar
+            if (tok.partial_utf8.n_remain != 0 &&
+                    !llama_grammar_match_partial_char(stack_pos, tok.partial_utf8)) {
+                rejects.push_back(tok);
             }
+        } else if (llama_grammar_match_char(stack_pos, *tok.code_points).first) {
+            next_candidates.push_back({ tok.index, tok.code_points + 1, tok.partial_utf8 });
         } else {
             rejects.push_back(tok);
         }
@@ -2297,7 +2386,7 @@ static std::vector<llama_grammar_candidate> llama_grammar_reject_candidates_for_
 
     auto next_rejects = llama_grammar_reject_candidates(rules, next_stacks, next_candidates);
     for (auto tok : next_rejects) {
-        rejects.push_back({ tok.index, tok.code_points - 1 });
+        rejects.push_back({ tok.index, tok.code_points - 1, tok.partial_utf8 });
     }
 
     return rejects;
@@ -2362,7 +2451,7 @@ struct llama_grammar * llama_grammar_init(
         }
     } while (true);
 
-    return new llama_grammar{ std::move(vec_rules), std::move(stacks) };
+    return new llama_grammar{ std::move(vec_rules), std::move(stacks), {} };
 }
 
 void llama_grammar_free(struct llama_grammar * grammar) {
@@ -2668,7 +2757,7 @@ void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * c
 
     const llama_token eos = llama_token_eos();
 
-    std::vector<std::vector<uint32_t>>   candidates_decoded;
+    std::vector<std::pair<std::vector<uint32_t>, llama_partial_utf8>> candidates_decoded;
     std::vector<llama_grammar_candidate>                              candidates_grammar;
 
     for (size_t i = 0; i < candidates->size; ++i) {
@@ -2681,8 +2770,10 @@ void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * c
         } else if (*str == 0) {
             candidates->data[i].logit = -INFINITY;
         } else {
-            candidates_decoded.push_back(decode_utf8(str));
+            candidates_decoded.push_back(decode_utf8(str, grammar->partial_utf8));
-            candidates_grammar.push_back({ i, candidates_decoded.back().data() });
+            candidates_grammar.push_back({
+                i, candidates_decoded.back().first.data(), candidates_decoded.back().second
+            });
         }
     }
 
@@ -2883,11 +2974,14 @@ void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar
     }
 
     const char * str = llama_token_to_str(ctx, token);
+
     // Note terminating 0 in decoded string
-    auto code_points = decode_utf8(str);
+    const auto   decoded     = decode_utf8(str, grammar->partial_utf8);
+    const auto & code_points = decoded.first;
     for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
         grammar->stacks = llama_grammar_accept(grammar->rules, grammar->stacks, *it);
     }
+    grammar->partial_utf8 = decoded.second;
     LLAMA_ASSERT(!grammar->stacks.empty());
 
     ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
@@ -3306,7 +3400,18 @@ struct llama_context * llama_new_context_with_model(
             int n_past = hparams.n_ctx - n_tokens;
             llama_token token = llama_token_bos(); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph
             ggml_cgraph * gf = llama_build_graph(*ctx, &token, NULL, n_tokens, n_past);
-
+#ifdef GGML_USE_METAL
+            if (params.n_gpu_layers > 0) {
+                ctx->ctx_metal = ggml_metal_init(1);
+                if (!ctx->ctx_metal) {
+                    LLAMA_LOG_ERROR("%s: ggml_metal_init() failed\n", __func__);
+                    llama_free(ctx);
+                    return NULL;
+                }
+                ggml_metal_graph_find_concurrency(ctx->ctx_metal, gf, false);
+                ggml_allocr_set_parse_seq(ctx->alloc, ggml_metal_get_concur_list(ctx->ctx_metal), ggml_metal_if_optimized(ctx->ctx_metal));
+            }
+#endif
             // measure memory requirements for the graph
             size_t alloc_size = ggml_allocr_alloc_graph(ctx->alloc, gf) + tensor_alignment;
 
@@ -3324,6 +3429,11 @@ struct llama_context * llama_new_context_with_model(
 
             ctx->buf_alloc.resize(alloc_size);
             ctx->alloc = ggml_allocr_new(ctx->buf_alloc.addr, ctx->buf_alloc.size, tensor_alignment);
+#ifdef GGML_USE_METAL
+            if (ctx->ctx_metal) {
+                ggml_allocr_set_parse_seq(ctx->alloc, ggml_metal_get_concur_list(ctx->ctx_metal), ggml_metal_if_optimized(ctx->ctx_metal));
+            }
+#endif
         }
 #else
         ctx->buf_compute.resize(MEM_REQ_EVAL().at(ctx->model.type) + ggml_graph_overhead());
@@ -3338,7 +3448,6 @@ struct llama_context * llama_new_context_with_model(
 #ifdef GGML_USE_METAL
     if (params.n_gpu_layers > 0) {
         // this allocates all Metal resources and memory buffers
-        ctx->ctx_metal = ggml_metal_init(1);
 
         void * data_ptr  = NULL;
         size_t data_size = 0;
@@ -3367,8 +3476,7 @@ struct llama_context * llama_new_context_with_model(
         LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "eval", ctx->buf_compute.addr, ctx->buf_compute.size, 0));
         LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "kv",   ctx->kv_self.buf.addr, ctx->kv_self.buf.size, 0));
 
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "scr0", ctx->buf_scratch[0].addr, ctx->buf_scratch[0].size, 0));
+        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "alloc", ctx->buf_alloc.addr, ctx->buf_alloc.size, 0));
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "scr1", ctx->buf_scratch[1].addr, ctx->buf_scratch[1].size, 0));
 #undef LLAMA_METAL_CHECK_BUF
     }
 #endif
@@ -4178,6 +4286,10 @@ int llama_n_embd(const struct llama_context * ctx) {
     return ctx->model.hparams.n_embd;
 }
 
+int llama_model_type(const struct llama_model * model, char * buf, size_t buf_size) {
+    return snprintf(buf, buf_size, "LLaMA %s %s", llama_model_type_name(model->type), llama_ftype_name(model->hparams.ftype));
+}
+
 int llama_get_vocab_from_model(
         const struct llama_model * model,
         const char * * strings,

llama.h

@@ -97,7 +97,7 @@ extern "C" {
     // If your logging mechanism cannot handle that, check if the last character is '\n' and strip it
     // if it exists.
     // It might not exist for progress report where '.' is output repeatedly.
-    typedef void (*llama_log_callback)(llama_log_level level, const char * text, void * user_data);
+    typedef void (*llama_log_callback)(enum llama_log_level level, const char * text, void * user_data);
 
     struct llama_context_params {
         uint32_t seed;         // RNG seed, -1 for random
@@ -353,6 +353,8 @@ extern "C" {
     LLAMA_API int llama_n_ctx_from_model  (const struct llama_model * model);
     LLAMA_API int llama_n_embd_from_model (const struct llama_model * model);
 
+    LLAMA_API int llama_model_type(const struct llama_model * model, char * buf, size_t buf_size);
+
     // Get the vocabulary as output parameters.
     // Returns number of results.
     LLAMA_API int llama_get_vocab(

scripts/get-wikitext-2.sh

@@ -0,0 +1,3 @@
+#!/bin/bash
+
+wget https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip

tests/CMakeLists.txt

@@ -11,5 +11,7 @@ llama_add_test(test-quantize-fns.cpp)
 llama_add_test(test-quantize-perf.cpp)
 llama_add_test(test-sampling.cpp)
 llama_add_test(test-tokenizer-0.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab.bin)
+llama_add_test(test-grammar-parser.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../examples/grammar-parser.cpp)
+llama_add_test(test-llama-grammar.cpp  ${CMAKE_CURRENT_SOURCE_DIR}/../examples/grammar-parser.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../examples/common.cpp)
 llama_add_test(test-grad0.cpp) # SLOW
 # llama_add_test(test-opt.cpp) # SLOW

tests/test-grammar-parser.cpp

@@ -0,0 +1,249 @@
+#ifdef NDEBUG
+#undef NDEBUG
+#endif
+
+#include "llama.h"
+#include "examples/grammar-parser.cpp"
+#include <cassert>
+
+int main()
+{
+    grammar_parser::parse_state parsed_grammar;
+
+    const char *grammar_bytes = R"""(root  ::= (expr "=" term "\n")+
+expr  ::= term ([-+*/] term)*
+term  ::= [0-9]+)""";
+
+    parsed_grammar = grammar_parser::parse(grammar_bytes);
+
+    std::vector<std::pair<std::string, uint32_t>> expected = {
+        {"expr", 2},
+        {"expr_5", 5},
+        {"expr_6", 6},
+        {"root", 0},
+        {"root_1", 1},
+        {"root_4", 4},
+        {"term", 3},
+        {"term_7", 7},
+    };
+
+    uint32_t index = 0;
+    for (auto it = parsed_grammar.symbol_ids.begin(); it != parsed_grammar.symbol_ids.end(); ++it)
+    {
+        std::string key = it->first;
+        uint32_t value = it->second;
+        std::pair<std::string, uint32_t> expected_pair = expected[index];
+
+        // pretty print error message before asserting
+        if (expected_pair.first != key || expected_pair.second != value)
+        {
+            fprintf(stderr, "expected_pair: %s, %d\n", expected_pair.first.c_str(), expected_pair.second);
+            fprintf(stderr, "actual_pair: %s, %d\n", key.c_str(), value);
+            fprintf(stderr, "expected_pair != actual_pair\n");
+        }
+
+        assert(expected_pair.first == key && expected_pair.second == value);
+
+        index++;
+    }
+    std::vector<llama_grammar_element> expected_rules = {
+        {LLAMA_GRETYPE_RULE_REF, 4},
+        {LLAMA_GRETYPE_END, 0},
+        {LLAMA_GRETYPE_RULE_REF, 2},
+        {LLAMA_GRETYPE_CHAR, 61},
+        {LLAMA_GRETYPE_RULE_REF, 3},
+        {LLAMA_GRETYPE_CHAR, 10},
+        {LLAMA_GRETYPE_END, 0},
+        {LLAMA_GRETYPE_RULE_REF, 3},
+        {LLAMA_GRETYPE_RULE_REF, 6},
+        {LLAMA_GRETYPE_END, 0},
+        {LLAMA_GRETYPE_RULE_REF, 7},
+        {LLAMA_GRETYPE_END, 0},
+        {LLAMA_GRETYPE_RULE_REF, 1},
+        {LLAMA_GRETYPE_RULE_REF, 4},
+        {LLAMA_GRETYPE_ALT, 0},
+        {LLAMA_GRETYPE_RULE_REF, 1},
+        {LLAMA_GRETYPE_END, 0},
+        {LLAMA_GRETYPE_CHAR, 45},
+        {LLAMA_GRETYPE_CHAR_ALT, 43},
+        {LLAMA_GRETYPE_CHAR_ALT, 42},
+        {LLAMA_GRETYPE_CHAR_ALT, 47},
+        {LLAMA_GRETYPE_RULE_REF, 3},
+        {LLAMA_GRETYPE_END, 0},
+        {LLAMA_GRETYPE_RULE_REF, 5},
+        {LLAMA_GRETYPE_RULE_REF, 6},
+        {LLAMA_GRETYPE_ALT, 0},
+        {LLAMA_GRETYPE_END, 0},
+        {LLAMA_GRETYPE_CHAR, 48},
+        {LLAMA_GRETYPE_CHAR_RNG_UPPER, 57},
+        {LLAMA_GRETYPE_RULE_REF, 7},
+        {LLAMA_GRETYPE_ALT, 0},
+        {LLAMA_GRETYPE_CHAR, 48},
+        {LLAMA_GRETYPE_CHAR_RNG_UPPER, 57},
+        {LLAMA_GRETYPE_END, 0},
+    };
+
+    index = 0;
+    for (auto rule : parsed_grammar.rules)
+    {
+        // compare rule to expected rule
+        for (uint32_t i = 0; i < rule.size(); i++)
+        {
+            llama_grammar_element element = rule[i];
+            llama_grammar_element expected_element = expected_rules[index];
+
+            // pretty print error message before asserting
+            if (expected_element.type != element.type || expected_element.value != element.value)
+            {
+                fprintf(stderr, "index: %d\n", index);
+                fprintf(stderr, "expected_element: %d, %d\n", expected_element.type, expected_element.value);
+                fprintf(stderr, "actual_element: %d, %d\n", element.type, element.value);
+                fprintf(stderr, "expected_element != actual_element\n");
+            }
+
+            assert(expected_element.type == element.type && expected_element.value == element.value);
+            index++;
+        }
+    }
+
+    const char *longer_grammar_bytes = R"""(
+    root  ::= (expr "=" ws term "\n")+
+    expr  ::= term ([-+*/] term)*
+    term  ::= ident | num | "(" ws expr ")" ws
+    ident ::= [a-z] [a-z0-9_]* ws
+    num   ::= [0-9]+ ws
+    ws    ::= [ \t\n]*
+    )""";
+
+    parsed_grammar = grammar_parser::parse(longer_grammar_bytes);
+
+    expected = {
+        {"expr", 2},
+        {"expr_6", 6},
+        {"expr_7", 7},
+        {"ident", 8},
+        {"ident_10", 10},
+        {"num", 9},
+        {"num_11", 11},
+        {"root", 0},
+        {"root_1", 1},
+        {"root_5", 5},
+        {"term", 4},
+        {"ws", 3},
+        {"ws_12", 12},
+    };
+
+    index = 0;
+    for (auto it = parsed_grammar.symbol_ids.begin(); it != parsed_grammar.symbol_ids.end(); ++it)
+    {
+        std::string key = it->first;
+        uint32_t value = it->second;
+        std::pair<std::string, uint32_t> expected_pair = expected[index];
+
+        // pretty print error message before asserting
+        if (expected_pair.first != key || expected_pair.second != value)
+        {
+            fprintf(stderr, "expected_pair: %s, %d\n", expected_pair.first.c_str(), expected_pair.second);
+            fprintf(stderr, "actual_pair: %s, %d\n", key.c_str(), value);
+            fprintf(stderr, "expected_pair != actual_pair\n");
+        }
+
+        assert(expected_pair.first == key && expected_pair.second == value);
+
+        index++;
+    }
+    expected_rules = {
+        {LLAMA_GRETYPE_RULE_REF, 5},
+        {LLAMA_GRETYPE_END, 0},
+        {LLAMA_GRETYPE_RULE_REF, 2},
+        {LLAMA_GRETYPE_CHAR, 61},
+        {LLAMA_GRETYPE_RULE_REF, 3},
+        {LLAMA_GRETYPE_RULE_REF, 4},
+        {LLAMA_GRETYPE_CHAR, 10},
+        {LLAMA_GRETYPE_END, 0},
+        {LLAMA_GRETYPE_RULE_REF, 4},
+        {LLAMA_GRETYPE_RULE_REF, 7},
+        {LLAMA_GRETYPE_END, 0},
+        {LLAMA_GRETYPE_RULE_REF, 12},
+        {LLAMA_GRETYPE_END, 0},
+        {LLAMA_GRETYPE_RULE_REF, 8},
+        {LLAMA_GRETYPE_ALT, 0},
+        {LLAMA_GRETYPE_RULE_REF, 9},
+        {LLAMA_GRETYPE_ALT, 0},
+        {LLAMA_GRETYPE_CHAR, 40},
+        {LLAMA_GRETYPE_RULE_REF, 3},
+        {LLAMA_GRETYPE_RULE_REF, 2},
+        {LLAMA_GRETYPE_CHAR, 41},
+        {LLAMA_GRETYPE_RULE_REF, 3},
+        {LLAMA_GRETYPE_END, 0},
+        {LLAMA_GRETYPE_RULE_REF, 1},
+        {LLAMA_GRETYPE_RULE_REF, 5},
+        {LLAMA_GRETYPE_ALT, 0},
+        {LLAMA_GRETYPE_RULE_REF, 1},
+        {LLAMA_GRETYPE_END, 0},
+        {LLAMA_GRETYPE_CHAR, 45},
+        {LLAMA_GRETYPE_CHAR_ALT, 43},
+        {LLAMA_GRETYPE_CHAR_ALT, 42},
+        {LLAMA_GRETYPE_CHAR_ALT, 47},
+        {LLAMA_GRETYPE_RULE_REF, 4},
+        {LLAMA_GRETYPE_END, 0},
+        {LLAMA_GRETYPE_RULE_REF, 6},
+        {LLAMA_GRETYPE_RULE_REF, 7},
+        {LLAMA_GRETYPE_ALT, 0},
+        {LLAMA_GRETYPE_END, 0},
+        {LLAMA_GRETYPE_CHAR, 97},
+        {LLAMA_GRETYPE_CHAR_RNG_UPPER, 122},
+        {LLAMA_GRETYPE_RULE_REF, 10},
+        {LLAMA_GRETYPE_RULE_REF, 3},
+        {LLAMA_GRETYPE_END, 0},
+        {LLAMA_GRETYPE_RULE_REF, 11},
+        {LLAMA_GRETYPE_RULE_REF, 3},
+        {LLAMA_GRETYPE_END, 0},
+        {LLAMA_GRETYPE_CHAR, 97},
+        {LLAMA_GRETYPE_CHAR_RNG_UPPER, 122},
+        {LLAMA_GRETYPE_CHAR_ALT, 48},
+        {LLAMA_GRETYPE_CHAR_RNG_UPPER, 57},
+        {LLAMA_GRETYPE_CHAR_ALT, 95},
+        {LLAMA_GRETYPE_RULE_REF, 10},
+        {LLAMA_GRETYPE_ALT, 0},
+        {LLAMA_GRETYPE_END, 0},
+        {LLAMA_GRETYPE_CHAR, 48},
+        {LLAMA_GRETYPE_CHAR_RNG_UPPER, 57},
+        {LLAMA_GRETYPE_RULE_REF, 11},
+        {LLAMA_GRETYPE_ALT, 0},
+        {LLAMA_GRETYPE_CHAR, 48},
+        {LLAMA_GRETYPE_CHAR_RNG_UPPER, 57},
+        {LLAMA_GRETYPE_END, 0},
+        {LLAMA_GRETYPE_CHAR, 32},
+        {LLAMA_GRETYPE_CHAR_ALT, 9},
+        {LLAMA_GRETYPE_CHAR_ALT, 10},
+        {LLAMA_GRETYPE_RULE_REF, 12},
+        {LLAMA_GRETYPE_ALT, 0},
+        {LLAMA_GRETYPE_END, 0},
+    };
+
+    index = 0;
+    for (auto rule : parsed_grammar.rules)
+    {
+        // compare rule to expected rule
+        for (uint32_t i = 0; i < rule.size(); i++)
+        {
+            llama_grammar_element element = rule[i];
+            llama_grammar_element expected_element = expected_rules[index];
+
+            // pretty print error message before asserting
+            if (expected_element.type != element.type || expected_element.value != element.value)
+            {
+                fprintf(stderr, "index: %d\n", index);
+                fprintf(stderr, "expected_element: %d, %d\n", expected_element.type, expected_element.value);
+                fprintf(stderr, "actual_element: %d, %d\n", element.type, element.value);
+                fprintf(stderr, "expected_element != actual_element\n");
+            }
+
+            assert(expected_element.type == element.type && expected_element.value == element.value);
+            index++;
+        }
+    }
+
+    return 0;
+}

tests/test-llama-grammar.cpp

@@ -0,0 +1,403 @@
+#ifdef NDEBUG
+#undef NDEBUG
+#endif
+
+#include "llama.cpp"
+#include "examples/common.cpp"
+#include "examples/grammar-parser.cpp"
+#include <cassert>
+
+int main()
+{
+    grammar_parser::parse_state parsed_grammar;
+
+    std::vector<std::pair<std::string, uint32_t>> expected = {
+        {"expr", 2},
+        {"expr_6", 6},
+        {"expr_7", 7},
+        {"ident", 8},
+        {"ident_10", 10},
+        {"num", 9},
+        {"num_11", 11},
+        {"root", 0},
+        {"root_1", 1},
+        {"root_5", 5},
+        {"term", 4},
+        {"ws", 3},
+        {"ws_12", 12},
+    };
+
+    std::vector<std::vector<llama_grammar_element>> expected_rules = {
+        {{LLAMA_GRETYPE_RULE_REF, 5}, {LLAMA_GRETYPE_END, 0}},
+        {
+            {LLAMA_GRETYPE_RULE_REF, 2},
+            {LLAMA_GRETYPE_CHAR, 61},
+            {LLAMA_GRETYPE_RULE_REF, 3},
+            {LLAMA_GRETYPE_RULE_REF, 4},
+            {LLAMA_GRETYPE_CHAR, 10},
+            {LLAMA_GRETYPE_END, 0},
+        },
+        {{LLAMA_GRETYPE_RULE_REF, 4}, {LLAMA_GRETYPE_RULE_REF, 7}, {LLAMA_GRETYPE_END, 0}},
+        {{LLAMA_GRETYPE_RULE_REF, 12}, {LLAMA_GRETYPE_END, 0}},
+        {
+            {LLAMA_GRETYPE_RULE_REF, 8},
+            {LLAMA_GRETYPE_ALT, 0},
+            {LLAMA_GRETYPE_RULE_REF, 9},
+            {LLAMA_GRETYPE_ALT, 0},
+            {LLAMA_GRETYPE_CHAR, 40},
+            {LLAMA_GRETYPE_RULE_REF, 3},
+            {LLAMA_GRETYPE_RULE_REF, 2},
+            {LLAMA_GRETYPE_CHAR, 41},
+            {LLAMA_GRETYPE_RULE_REF, 3},
+            {LLAMA_GRETYPE_END, 0},
+        },
+        {{LLAMA_GRETYPE_RULE_REF, 1}, {LLAMA_GRETYPE_RULE_REF, 5}, {LLAMA_GRETYPE_ALT, 0}, {LLAMA_GRETYPE_RULE_REF, 1}, {LLAMA_GRETYPE_END, 0}},
+        {
+            {LLAMA_GRETYPE_CHAR, 45},
+            {LLAMA_GRETYPE_CHAR_ALT, 43},
+            {LLAMA_GRETYPE_CHAR_ALT, 42},
+            {LLAMA_GRETYPE_CHAR_ALT, 47},
+            {LLAMA_GRETYPE_RULE_REF, 4},
+            {LLAMA_GRETYPE_END, 0},
+        },
+        {{LLAMA_GRETYPE_RULE_REF, 6}, {LLAMA_GRETYPE_RULE_REF, 7}, {LLAMA_GRETYPE_ALT, 0}, {LLAMA_GRETYPE_END, 0}},
+        {
+            {LLAMA_GRETYPE_CHAR, 97},
+            {LLAMA_GRETYPE_CHAR_RNG_UPPER, 122},
+            {LLAMA_GRETYPE_RULE_REF, 10},
+            {LLAMA_GRETYPE_RULE_REF, 3},
+            {LLAMA_GRETYPE_END, 0},
+        },
+        {{LLAMA_GRETYPE_RULE_REF, 11}, {LLAMA_GRETYPE_RULE_REF, 3}, {LLAMA_GRETYPE_END, 0}},
+        {
+            {LLAMA_GRETYPE_CHAR, 97},
+            {LLAMA_GRETYPE_CHAR_RNG_UPPER, 122},
+            {LLAMA_GRETYPE_CHAR_ALT, 48},
+            {LLAMA_GRETYPE_CHAR_RNG_UPPER, 57},
+            {LLAMA_GRETYPE_CHAR_ALT, 95},
+            {LLAMA_GRETYPE_RULE_REF, 10},
+            {LLAMA_GRETYPE_ALT, 0},
+            {LLAMA_GRETYPE_END, 0},
+        },
+        {
+            {LLAMA_GRETYPE_CHAR, 48},
+            {LLAMA_GRETYPE_CHAR_RNG_UPPER, 57},
+            {LLAMA_GRETYPE_RULE_REF, 11},
+            {LLAMA_GRETYPE_ALT, 0},
+            {LLAMA_GRETYPE_CHAR, 48},
+            {LLAMA_GRETYPE_CHAR_RNG_UPPER, 57},
+            {LLAMA_GRETYPE_END, 0},
+        },
+        {
+            {LLAMA_GRETYPE_CHAR, 32},
+            {LLAMA_GRETYPE_CHAR_ALT, 9},
+            {LLAMA_GRETYPE_CHAR_ALT, 10},
+            {LLAMA_GRETYPE_RULE_REF, 12},
+            {LLAMA_GRETYPE_ALT, 0},
+            {LLAMA_GRETYPE_END, 0},
+        },
+    };
+
+    for (auto pair : expected)
+    {
+        parsed_grammar.symbol_ids[pair.first] = pair.second;
+    }
+
+    for (auto rule : expected_rules)
+    {
+        parsed_grammar.rules.push_back({});
+        for (auto element : rule)
+        {
+            parsed_grammar.rules.back().push_back(element);
+        }
+    }
+
+    llama_grammar *grammar = NULL;
+    std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+    grammar = llama_grammar_init(
+        grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
+
+    std::vector<std::vector<llama_grammar_element>> expected_stacks = {
+        {
+            {LLAMA_GRETYPE_RULE_REF, 5},
+            {LLAMA_GRETYPE_CHAR, 61},
+            {LLAMA_GRETYPE_RULE_REF, 7},
+            {LLAMA_GRETYPE_CHAR, 97},
+        },
+        {
+            {LLAMA_GRETYPE_RULE_REF, 5},
+            {LLAMA_GRETYPE_CHAR, 61},
+            {LLAMA_GRETYPE_RULE_REF, 7},
+            {LLAMA_GRETYPE_RULE_REF, 3},
+            {LLAMA_GRETYPE_CHAR, 48},
+        },
+        {
+            {LLAMA_GRETYPE_RULE_REF, 5},
+            {LLAMA_GRETYPE_CHAR, 61},
+            {LLAMA_GRETYPE_RULE_REF, 7},
+            {LLAMA_GRETYPE_RULE_REF, 3},
+            {LLAMA_GRETYPE_CHAR, 48},
+        },
+        {
+            {LLAMA_GRETYPE_RULE_REF, 5},
+            {LLAMA_GRETYPE_CHAR, 61},
+            {LLAMA_GRETYPE_RULE_REF, 7},
+            {LLAMA_GRETYPE_CHAR, 40},
+        },
+        {
+            {LLAMA_GRETYPE_CHAR, 61},
+            {LLAMA_GRETYPE_RULE_REF, 7},
+            {LLAMA_GRETYPE_CHAR, 97},
+        },
+        {
+            {LLAMA_GRETYPE_CHAR, 61},
+            {LLAMA_GRETYPE_RULE_REF, 7},
+            {LLAMA_GRETYPE_RULE_REF, 3},
+            {LLAMA_GRETYPE_CHAR, 48},
+        },
+        {
+            {LLAMA_GRETYPE_CHAR, 61},
+            {LLAMA_GRETYPE_RULE_REF, 7},
+            {LLAMA_GRETYPE_RULE_REF, 3},
+            {LLAMA_GRETYPE_CHAR, 48},
+        },
+        {
+            {LLAMA_GRETYPE_CHAR, 61},
+            {LLAMA_GRETYPE_RULE_REF, 7},
+            {LLAMA_GRETYPE_CHAR, 40},
+        }};
+
+    auto index = 0;
+    for (auto stack : grammar->stacks)
+    {
+        // compare stack to expected_stack
+        for (uint32_t i = 0; i < stack.size(); i++)
+        {
+            auto element = stack[i];
+            auto expected_element = expected_stacks[index][i];
+
+            // pretty print error message before asserting
+            if (expected_element.type != element->type || expected_element.value != element->value)
+            {
+                fprintf(stderr, "index: %d\n", index);
+                fprintf(stderr, "expected_element: %d, %d\n", expected_element.type, expected_element.value);
+                fprintf(stderr, "actual_element: %d, %d\n", element->type, element->value);
+                fprintf(stderr, "expected_element != actual_element\n");
+            }
+
+            assert(expected_element.type == element->type && expected_element.value == element->value);
+        }
+        index++;
+    }
+
+    std::vector<std::vector<const llama_grammar_element *>> next_stacks;
+    std::vector<llama_grammar_candidate> next_candidates;
+    next_candidates.resize(24);
+
+    for (size_t i = 0; i < 24; ++i)
+    {
+        uint32_t *cp = new uint32_t[2]; // dynamically allocate memory for code_point
+        cp[0] = 37 + i;
+        cp[1] = 0;
+        next_candidates[i] = {i, cp, {}};
+    }
+
+    std::vector<std::vector<std::pair<uint32_t, uint16_t>>> expected_reject = {
+        {
+            {0, 37},
+            {1, 38},
+            {2, 39},
+            {3, 40},
+            {4, 41},
+            {5, 42},
+            {6, 43},
+            {7, 44},
+            {8, 45},
+            {9, 46},
+            {10, 47},
+            {11, 48},
+            {12, 49},
+            {13, 50},
+            {14, 51},
+            {15, 52},
+            {16, 53},
+            {17, 54},
+            {18, 55},
+            {19, 56},
+            {20, 57},
+            {21, 58},
+            {22, 59},
+            {23, 60},
+        },
+        {
+            {0, 37},
+            {1, 38},
+            {2, 39},
+            {3, 40},
+            {4, 41},
+            {5, 42},
+            {6, 43},
+            {7, 44},
+            {8, 45},
+            {9, 46},
+            {10, 47},
+            {21, 58},
+            {22, 59},
+            {23, 60},
+        },
+        {
+            {0, 37},
+            {1, 38},
+            {2, 39},
+            {3, 40},
+            {4, 41},
+            {5, 42},
+            {6, 43},
+            {7, 44},
+            {8, 45},
+            {9, 46},
+            {10, 47},
+            {21, 58},
+            {22, 59},
+            {23, 60},
+        },
+        {
+            {0, 37},
+            {1, 38},
+            {2, 39},
+            {4, 41},
+            {5, 42},
+            {6, 43},
+            {7, 44},
+            {8, 45},
+            {9, 46},
+            {10, 47},
+            {11, 48},
+            {12, 49},
+            {13, 50},
+            {14, 51},
+            {15, 52},
+            {16, 53},
+            {17, 54},
+            {18, 55},
+            {19, 56},
+            {20, 57},
+            {21, 58},
+            {22, 59},
+            {23, 60},
+        },
+        {
+            {0, 37},
+            {1, 38},
+            {2, 39},
+            {3, 40},
+            {4, 41},
+            {5, 42},
+            {6, 43},
+            {7, 44},
+            {8, 45},
+            {9, 46},
+            {10, 47},
+            {11, 48},
+            {12, 49},
+            {13, 50},
+            {14, 51},
+            {15, 52},
+            {16, 53},
+            {17, 54},
+            {18, 55},
+            {19, 56},
+            {20, 57},
+            {21, 58},
+            {22, 59},
+            {23, 60},
+        },
+        {
+            {0, 37},
+            {1, 38},
+            {2, 39},
+            {3, 40},
+            {4, 41},
+            {5, 42},
+            {6, 43},
+            {7, 44},
+            {8, 45},
+            {9, 46},
+            {10, 47},
+            {21, 58},
+            {22, 59},
+            {23, 60},
+        },
+        {
+            {0, 37},
+            {1, 38},
+            {2, 39},
+            {3, 40},
+            {4, 41},
+            {5, 42},
+            {6, 43},
+            {7, 44},
+            {8, 45},
+            {9, 46},
+            {10, 47},
+            {21, 58},
+            {22, 59},
+            {23, 60},
+        },
+        {
+            {0, 37},
+            {1, 38},
+            {2, 39},
+            {4, 41},
+            {5, 42},
+            {6, 43},
+            {7, 44},
+            {8, 45},
+            {9, 46},
+            {10, 47},
+            {11, 48},
+            {12, 49},
+            {13, 50},
+            {14, 51},
+            {15, 52},
+            {16, 53},
+            {17, 54},
+            {18, 55},
+            {19, 56},
+            {20, 57},
+            {21, 58},
+            {22, 59},
+            {23, 60},
+        },
+    };
+
+    std::vector<llama_grammar_candidate> rejects = llama_grammar_reject_candidates_for_stack(grammar->rules, grammar->stacks[0], next_candidates);
+
+    std::vector<std::vector<llama_grammar_candidate>> all_rejects;
+
+    for (std::size_t count = 0; count < grammar->stacks.size(); ++count)
+    {
+        rejects = llama_grammar_reject_candidates_for_stack(grammar->rules, grammar->stacks[count], next_candidates);
+        all_rejects.push_back(rejects);
+    }
+
+    index = 0;
+    for (auto rej : all_rejects)
+    {
+        for (uint32_t i = 0; i < rej.size(); i++)
+        {
+            auto element = rej[i];
+            auto expected_element = expected_reject[index][i];
+            assert(element.index == expected_element.first && *element.code_points == expected_element.second);
+        }
+        index++;
+    }
+
+    for (auto &candidate : next_candidates)
+    {
+        delete[] candidate.code_points;
+        candidate.code_points = nullptr;
+    }
+    delete grammar;
+    return 0;
+}