tests : adds simple llama grammar tests (#2618)

* adds simple llama grammar tests

* fix lint and add Makefile

* 0 terminate code_points

* avoid dangling pointers in candidate cleanup

* cleanup grammar at end of test

.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
@@ -68,6 +70,7 @@ 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
 
 # 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)
 
@@ -253,11 +253,6 @@ ifdef LLAMA_CUDA_KQUANTS_ITER
 else
 	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
 endif
-ifdef LLAMA_CUDA_MMQ_Y
-	NVCCFLAGS += -DGGML_CUDA_MMQ_Y=$(LLAMA_CUDA_MMQ_Y)
-else
-	NVCCFLAGS += -DGGML_CUDA_MMQ_Y=64
-endif # LLAMA_CUDA_MMQ_Y
 #ifdef LLAMA_CUDA_CUBLAS
 #	NVCCFLAGS += -DGGML_CUDA_CUBLAS
 #endif # LLAMA_CUDA_CUBLAS
@@ -288,7 +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
 
@@ -350,7 +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 build-info.h $(TEST_TARGETS)
 
 #
 # Examples
@@ -380,7 +375,7 @@ embedding: examples/embedding/embedding.cpp                   build-info.h ggml.
 save-load-state: examples/save-load-state/save-load-state.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp build-info.h ggml.o llama.o common.o $(OBJS)
+server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp build-info.h ggml.o llama.o common.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2)
 
 $(LIB_PRE)embdinput$(DSO_EXT): examples/embd-input/embd-input.h examples/embd-input/embd-input-lib.cpp build-info.h ggml.o llama.o common.o $(OBJS)
@@ -393,6 +388,9 @@ 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)
+
 build-info.h: $(wildcard .git/index) scripts/build-info.sh
 	@sh scripts/build-info.sh > $@.tmp
 	@if ! cmp -s $@.tmp $@; then \
@@ -414,6 +412,12 @@ 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

@@ -406,7 +406,6 @@ Building the program with BLAS support may lead to some performance improvements
 --->
   | Option                  | Legal values           | Default | Description |
   |-------------------------|------------------------|---------|-------------|
-  | LLAMA_CUDA_MMQ_Y        | Positive integer >= 32 |      64 | Tile size in y direction when using the custom CUDA kernels for prompt processing. Higher values can be faster depending on the amount of shared memory available. Power of 2 heavily recommended. |
   | 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. |

examples/CMakeLists.txt

@@ -42,6 +42,7 @@ 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)
     if (LLAMA_METAL)

examples/common.cpp

@@ -194,6 +194,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.rope_freq_scale = std::stof(argv[i]);
+        } else if (arg == "--rope-scale") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.rope_freq_scale = 1.0f/std::stof(argv[i]);
         } else if (arg == "--memory-f32") {
             params.memory_f16 = false;
         } else if (arg == "--top-p") {
@@ -537,7 +543,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stdout, "  --in-suffix STRING    string to suffix after user inputs with (default: empty)\n");
     fprintf(stdout, "  -f FNAME, --file FNAME\n");
     fprintf(stdout, "                        prompt file to start generation.\n");
-    fprintf(stdout, "  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity)\n", params.n_predict);
+    fprintf(stdout, "  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
     fprintf(stdout, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
     fprintf(stdout, "  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
     fprintf(stdout, "  -gqa N, --gqa N       grouped-query attention factor (TEMP!!! use 8 for LLaMAv2 70B) (default: %d)\n", params.n_gqa);
@@ -564,8 +570,9 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stdout, "  --cfg-negative-prompt PROMPT \n");
     fprintf(stdout, "                        negative prompt 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-freq-base N    RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
-    fprintf(stdout, "  --rope-freq-scale N   RoPE frequency scaling factor (default: %g)\n", params.rope_freq_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);
+    fprintf(stdout, "  --rope-freq-scale N   RoPE frequency linear scaling factor, inverse of --rope-scale (default: %g)\n", params.rope_freq_scale);
     fprintf(stdout, "  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
     fprintf(stdout, "  --no-penalize-nl      do not penalize newline token\n");
     fprintf(stdout, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");

examples/console.cpp

@@ -10,6 +10,9 @@
 #include <windows.h>
 #include <fcntl.h>
 #include <io.h>
+#ifndef ENABLE_VIRTUAL_TERMINAL_PROCESSING
+#define ENABLE_VIRTUAL_TERMINAL_PROCESSING 0x0004
+#endif
 #else
 #include <climits>
 #include <sys/ioctl.h>
@@ -68,9 +71,10 @@ namespace console {
             }
         }
         if (hConsole) {
-            // Enable ANSI colors on Windows 10+
-            if (advanced_display && !(dwMode & ENABLE_VIRTUAL_TERMINAL_PROCESSING)) {
-                SetConsoleMode(hConsole, dwMode | ENABLE_VIRTUAL_TERMINAL_PROCESSING);
+            // Check conditions combined to reduce nesting
+            if (advanced_display && !(dwMode & ENABLE_VIRTUAL_TERMINAL_PROCESSING) &&
+                !SetConsoleMode(hConsole, dwMode | ENABLE_VIRTUAL_TERMINAL_PROCESSING)) {
+                advanced_display = false;
             }
             // Set console output codepage to UTF8
             SetConsoleOutputCP(CP_UTF8);

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.vim

@@ -0,0 +1,132 @@
+" Requires an already running llama.cpp server
+" To install either copy or symlink to ~/.vim/autoload/llama.vim
+" Then start with either :call llama#doLlamaGen(),
+" or add a keybind to your vimrc such as
+" nnoremap Z :call llama#doLlamaGen()<CR>
+" Similarly, you could add an insert mode keybind with
+" inoremap <C-B> <Cmd>call llama#doLlamaGen()<CR>
+"
+" g:llama_api_url and g:llama_overrides can be configured in your .vimrc
+" let g:llama_api_url = "192.168.1.10:8080"
+" llama_overrides can also be set through buffer/window scopes. For instance
+" autocmd filetype python let b:llama_overrides = {"temp": 0.2}
+" Could be added to your .vimrc to automatically set a lower temperature when
+" editing a python script
+" Additionally, an override dict can be stored at the top of a file
+" !*{"stop": ["User:"]}
+" Could be added to the start of your chatlog.txt to set the stopping token
+" These parameter dicts are merged together from lowest to highest priority:
+" server default -> g:llama_overrides -> w:llama_overrides ->
+" b:llama_overrides -> in file (!*) overrides
+"
+" Sublists (like logit_bias and stop) are overridden, not merged
+" Example override:
+" !*{"logit_bias": [[13, -5], [2, false]], "temperature": 1, "top_k": 5, "top_p": 0.5, "n_predict": 256, "repeat_last_n": 256, "repeat_penalty": 1.17647}
+if !exists("g:llama_api_url")
+    let g:llama_api_url= "127.0.0.1:8080"
+endif
+if !exists("g:llama_overrides")
+   let g:llama_overrides = {}
+endif
+const s:querydata = {"n_predict": 256, "stop": [ "\n" ], "stream": v:true }
+const s:curlcommand = ['curl','--data-raw', "{\"prompt\":\"### System:\"}", '--silent', '--no-buffer', '--request', 'POST', '--url', g:llama_api_url .. '/completion', '--header', "Content-Type: application/json"]
+let s:linedict = {}
+
+func s:callbackHandler(bufn, channel, msg)
+   if len(a:msg) < 3
+      return
+   elseif a:msg[0] == "d"
+      let l:msg = a:msg[6:-1]
+   else
+      let l:msg = a:msg
+   endif
+   let l:decoded_msg = json_decode(l:msg)
+   let l:newtext = split(l:decoded_msg['content'], "\n", 1)
+   if len(l:newtext) > 0
+      call setbufline(a:bufn, s:linedict[a:bufn], getbufline(a:bufn, s:linedict[a:bufn])[0] .. newtext[0])
+   else
+      echo "nothing genned"
+   endif
+   if len(newtext) > 1
+      let l:failed = appendbufline(a:bufn, s:linedict[a:bufn], newtext[1:-1])
+      let s:linedict[a:bufn] = s:linedict[a:bufn] + len(newtext)-1
+   endif
+   if has_key(l:decoded_msg, "stop") && l:decoded_msg.stop
+       echo "Finished generation"
+   endif
+endfunction
+
+func llama#doLlamaGen()
+   if exists("b:job")
+      if job_status(b:job) == "run"
+         call job_stop(b:job)
+         return
+      endif
+   endif
+
+   let l:cbuffer = bufnr("%")
+   let s:linedict[l:cbuffer] = line('$')
+   let l:buflines = getbufline(l:cbuffer, 1, 1000)
+   let l:querydata = copy(s:querydata)
+   call extend(l:querydata, g:llama_overrides)
+   if exists("w:llama_overrides")
+      call extend(l:querydata, w:llama_overrides)
+   endif
+   if exists("b:llama_overrides")
+      call extend(l:querydata, b:llama_overrides)
+   endif
+   if l:buflines[0][0:1] == '!*'
+      let l:userdata = json_decode(l:buflines[0][2:-1])
+      call extend(l:querydata, l:userdata)
+      let l:buflines = l:buflines[1:-1]
+   endif
+   let l:querydata.prompt = join(l:buflines, "\n")
+   let l:curlcommand = copy(s:curlcommand)
+   let l:curlcommand[2] = json_encode(l:querydata)
+   let b:job = job_start(l:curlcommand, {"callback": function("s:callbackHandler", [l:cbuffer])})
+endfunction
+
+" Echos the tokkenization of the provided string , or cursor to end of word
+" Onus is placed on the user to include the preceding space
+func llama#tokenizeWord(...)
+    if (a:0 > 0)
+        let l:input = a:1
+    else
+        exe "normal \"*ye"
+        let l:input = @*
+    endif
+    let l:querydata = {"content": l:input}
+    let l:curlcommand = copy(s:curlcommand)
+    let l:curlcommand[2] = json_encode(l:querydata)
+    let l:curlcommand[8] = g:llama_api_url .. "/tokenize"
+   let s:token_job = job_start(l:curlcommand, {"callback": function("s:tokenizeWordCallback", [l:input])})
+endfunction
+
+func s:tokenizeWordCallback(plaintext, channel, msg)
+    echo '"' .. a:plaintext ..'" - ' .. string(json_decode(a:msg).tokens)
+endfunction
+
+
+" Echos the token count of the entire buffer (or provided string)
+" Example usage :echo llama#tokenCount()
+func llama#tokenCount(...)
+    if (a:0 > 0)
+        let l:buflines = a:1
+    else
+        let l:buflines = getline(1,1000)
+        if l:buflines[0][0:1] == '!*'
+            let l:buflines = l:buflines[1:-1]
+        endif
+        let l:buflines = join(l:buflines, "\n")
+    endif
+    let l:querydata = {"content": l:buflines}
+    let l:curlcommand = copy(s:curlcommand)
+    let l:curlcommand[2] = json_encode(l:querydata)
+    let l:curlcommand[8] = g:llama_api_url .. "/tokenize"
+   let s:token_job = job_start(l:curlcommand, {"callback": "s:tokenCountCallback"})
+endfunction
+
+func s:tokenCountCallback(channel, msg)
+    let resp = json_decode(a:msg)
+    echo len(resp.tokens)
+endfunction

examples/llm.vim

@@ -1,3 +1,5 @@
+" Basic plugin example
+
 function! Llm()
 
   let url = "http://127.0.0.1:8080/completion"
@@ -16,8 +18,10 @@ function! Llm()
   " Extract the content field from the response
   let content = json_decode(response).content
 
+  let split_newlines = split(content, '\n', 1)
+
   " Insert the content at the cursor position
-  call setline(line('.'), getline('.') . content)
+  call setline(line('.'), [ getline('.') . split_newlines[0] ] + split_newlines[1:])
 endfunction
 
 command! Llm call Llm()

examples/main/README.md

@@ -140,6 +140,12 @@ The `--ctx-size` option allows you to set the size of the prompt context used by
 
 -   `-c N, --ctx-size N`: Set the size of the prompt context (default: 512). The LLaMA models were built with a context of 2048, which will yield the best results on longer input/inference. However, increasing the context size beyond 2048 may lead to unpredictable results.
 
+### Extended Context Size
+
+Some fine-tuned models have extened the context length by scaling RoPE. For example, if the original pretrained model have a context length (max sequence length) of 4096 (4k) and the fine-tuned model have 32k. That is a scaling factor of 8, and should work by setting the above `--ctx-size` to 32768 (32k) and `--rope-scale` to 8.
+
+- `--rope-scale N`: Where N is the linear scaling factor used by the fine-tuned model.
+
 ### Keep Prompt
 
 The `--keep` option allows users to retain the original prompt when the model runs out of context, ensuring a connection to the initial instruction or conversation topic is maintained.
@@ -154,9 +160,13 @@ The following options allow you to control the text generation process and fine-
 
 ### Number of Tokens to Predict
 
--   `-n N, --n-predict N`: Set the number of tokens to predict when generating text (default: 128, -1 = infinity).
+-   `-n N, --n-predict N`: Set the number of tokens to predict when generating text (default: 128, -1 = infinity, -2 = until context filled)
 
-The `--n-predict` option controls the number of tokens the model generates in response to the input prompt. By adjusting this value, you can influence the length of the generated text. A higher value will result in longer text, while a lower value will produce shorter text. A value of -1 will cause text to be generated without limit.
+The `--n-predict` option controls the number of tokens the model generates in response to the input prompt. By adjusting this value, you can influence the length of the generated text. A higher value will result in longer text, while a lower value will produce shorter text.
+
+A value of -1 will enable infinite text generation, even though we have a finite context window. When the context window is full, some of the earlier tokens (half of the tokens after `--n-keep`) will be discarded. The context must then be re-evaluated before generation can resume. On large models and/or large context windows, this will result in significant pause in output.
+
+If the pause is undesirable, a value of -2 will stop generation immediately when the context is filled.
 
 It is important to note that the generated text may be shorter than the specified number of tokens if an End-of-Sequence (EOS) token or a reverse prompt is encountered. In interactive mode text generation will pause and control will be returned to the user. In non-interactive mode, the program will end. In both cases, the text generation may stop before reaching the specified `n-predict` value. If you want the model to keep going without ever producing End-of-Sequence on its own, you can use the `--ignore-eos` parameter.
 

examples/main/main.cpp

@@ -431,8 +431,12 @@ int main(int argc, char ** argv) {
             // - take the n_keep first tokens from the original prompt (via n_past)
             // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
             if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) > n_ctx) {
-                const int n_left = n_past - params.n_keep;
+                if (params.n_predict == -2) {
+                    fprintf(stderr, "\n\n%s: context full, stopping generation\n", __func__);
+                    break;
+                }
 
+                const int n_left = n_past - params.n_keep;
                 // always keep the first token - BOS
                 n_past = std::max(1, params.n_keep);
                 n_past_guidance = std::max(1, params.n_keep + guidance_offset);

examples/server/README.md

@@ -16,6 +16,7 @@ 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`.
@@ -151,6 +152,8 @@ node .
 
     `mirostat_eta`: Set the Mirostat learning rate, parameter eta (default: 0.1).
 
+    `grammar`: Set grammar for grammar-based sampling (default: no grammar)
+
     `seed`: Set the random number generator (RNG) seed (default: -1, -1 = random seed).
 
     `ignore_eos`: Ignore end of stream token and continue generating (default: false).

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.",
@@ -166,6 +167,7 @@
       mirostat: 0, // 0/1/2
       mirostat_tau: 5, // target entropy
       mirostat_eta: 0.1, // learning rate
+      grammar: '',
     })
 
     const llamaStats = signal(null)
@@ -304,6 +306,26 @@
       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: ${e.message}`)
+        }
+      }
+
       const FloatField = ({label, max, min, name, step, value}) => {
         return html`
           <div>
@@ -355,6 +377,13 @@
               <label for="template">Chat history template</label>
               <textarea id="template" name="historyTemplate" value="${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

@@ -1,6 +1,7 @@
 #include "common.h"
 #include "llama.h"
 #include "build-info.h"
+#include "grammar-parser.h"
 
 #ifndef NDEBUG
 // crash the server in debug mode, otherwise send an http 500 error
@@ -14,6 +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
@@ -195,6 +197,9 @@ struct llama_server_context
     llama_context *ctx = nullptr;
     gpt_params params;
 
+    grammar_parser::parse_state parsed_grammar;
+    llama_grammar *grammar = nullptr;
+
     bool truncated = false;
     bool stopped_eos = false;
     bool stopped_word = false;
@@ -226,6 +231,7 @@ struct llama_server_context
     void rewind()
     {
         params.antiprompt.clear();
+        params.grammar.clear();
         num_prompt_tokens = 0;
         num_tokens_predicted = 0;
         generated_text = "";
@@ -237,9 +243,13 @@ struct llama_server_context
         stopped_limit = false;
         stopping_word = "";
         multibyte_pending = 0;
-
         n_remain = 0;
         n_past = 0;
+
+        if (grammar != nullptr) {
+            llama_grammar_free(grammar);
+            grammar = nullptr;
+        }
     }
 
     bool loadModel(const gpt_params &params_)
@@ -257,6 +267,31 @@ struct llama_server_context
         return true;
     }
 
+    bool loadGrammar()
+    {
+        if (!params.grammar.empty()) {
+            parsed_grammar = grammar_parser::parse(params.grammar.c_str());
+            // will be empty (default) if there are parse errors
+            if (parsed_grammar.rules.empty()) {
+                LOG_ERROR("grammar parse error", {{"grammar", params.grammar}});
+                return false;
+            }
+            grammar_parser::print_grammar(stderr, parsed_grammar);
+
+            {
+                auto it = params.logit_bias.find(llama_token_eos());
+                if (it != params.logit_bias.end() && it->second == -INFINITY) {
+                    LOG_WARNING("EOS token is disabled, which will cause most grammars to fail", {});
+                }
+            }
+
+            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"));
+        }
+        return true;
+    }
+
     void loadPrompt()
     {
         params.prompt.insert(0, 1, ' '); // always add a first space
@@ -420,6 +455,10 @@ struct llama_server_context
                 logits[llama_token_nl()] = nl_logit;
             }
 
+            if (grammar != nullptr) {
+                llama_sample_grammar(ctx, &candidates_p, grammar);
+            }
+
             if (temp <= 0)
             {
                 // Greedy sampling
@@ -457,10 +496,15 @@ struct llama_server_context
                 }
             }
 
+            if (grammar != nullptr) {
+                llama_grammar_accept_token(ctx, grammar, result.tok);
+            }
+
             for (size_t i = 0; i < std::min(candidates_p.size, (size_t)n_probs); ++i)
             {
                 result.probs.push_back({candidates_p.data[i].id, candidates_p.data[i].p});
             }
+
             last_n_tokens.erase(last_n_tokens.begin());
             last_n_tokens.push_back(result.tok);
             num_tokens_predicted++;
@@ -623,6 +667,7 @@ 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");
@@ -897,6 +942,10 @@ 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;
@@ -947,6 +996,7 @@ static json format_generation_settings(llama_server_context &llama)
         {"stream", llama.stream},
         {"logit_bias", llama.params.logit_bias},
         {"n_probs", llama.params.n_probs},
+        {"grammar", llama.params.grammar},
     };
 }
 
@@ -964,7 +1014,7 @@ 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},
@@ -993,7 +1043,6 @@ 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)},
     };
 
@@ -1048,6 +1097,7 @@ static void parse_options_completion(const json &body, llama_server_context &lla
     llama.params.n_keep = body.value("n_keep", default_params.n_keep);
     llama.params.seed = body.value("seed", default_params.seed);
     llama.params.prompt = body.value("prompt", default_params.prompt);
+    llama.params.grammar = body.value("grammar", default_params.grammar);
     llama.params.n_probs = body.value("n_probs", default_params.n_probs);
 
     llama.params.logit_bias.clear();
@@ -1169,6 +1219,12 @@ 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();
@@ -1179,6 +1235,12 @@ int main(int argc, char **argv)
 
         parse_options_completion(json::parse(req.body), llama);
 
+        if (!llama.loadGrammar())
+        {
+            res.status = 400;
+            return;
+        }
+
         llama.loadPrompt();
         llama.beginCompletion();
 
@@ -1334,8 +1396,12 @@ int main(int argc, char **argv)
 
     svr.set_error_handler([](const Request &, Response &res)
                           {
-        res.set_content("File Not Found", "text/plain");
-        res.status = 404; });
+        if (res.status == 400) {
+            res.set_content("Invalid request", "text/plain");
+        } else {
+            res.set_content("File Not Found", "text/plain");
+            res.status = 404;
+        } });
 
     // set timeouts and change hostname and port
     svr.set_read_timeout(sparams.read_timeout);
@@ -1363,6 +1429,9 @@ int main(int argc, char **argv)
         return 1;
     }
 
+    if (llama.grammar != nullptr) {
+        llama_grammar_free(llama.grammar);
+    }
     llama_backend_free();
 
     return 0;

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) {
@@ -126,10 +128,17 @@ 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) {
-        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");
+        // 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;
@@ -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
@@ -394,6 +418,14 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
                     if (parent == NULL) {
                         break;
                     }
+
+                    // if the node's data is external, then we cannot re-use it
+                    if ((char *) parent->data < (char *) alloc->data ||
+                        (char *) parent->data >= ((char *) alloc->data + alloc->size)) {
+                        AT_PRINTF("not reusing parent %s for %s as %p is external\n", parent->name, node->name, parent->data);
+                        continue;
+                    }
+
                     struct hash_node * p_hn = hash_get(ht, parent);
                     if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && ggml_are_same_layout(node, parent)) {
                         if (ggml_is_view(parent)) {
@@ -465,7 +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-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
-bool ggml_metal_if_optimized(struct ggml_metal_context * ctx);
+// if the graph has been optimized for concurrently dispatch, return length of the concur_list if optimized
+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 +5,11 @@
 #import <Foundation/Foundation.h>
 
 #import <Metal/Metal.h>
-#import <MetalPerformanceShaders/MetalPerformanceShaders.h>
+
+#undef MIN
+#undef MAX
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
 
 #ifdef GGML_METAL_NDEBUG
 #define metal_printf(...)
@@ -15,6 +19,8 @@
 
 #define UNUSED(x) (void)(x)
 
+#define GGML_MAX_CONCUR (2*GGML_MAX_NODES)
+
 struct ggml_metal_buffer {
     const char * name;
 
@@ -36,7 +42,7 @@ struct ggml_metal_context {
     int n_buffers;
     struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
 
-    int concur_list[GGML_MAX_NODES];
+    int concur_list[GGML_MAX_CONCUR];
     int concur_list_len;
 
     // custom kernels
@@ -72,6 +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);
@@ -103,13 +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
@@ -119,7 +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
@@ -137,7 +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
@@ -149,17 +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]; \
-        fprintf(stderr, "%s: loaded %-32s %16p\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name);
+        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); \
+        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);
@@ -189,6 +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);
@@ -221,11 +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) {
-    if (ctx->concur_list_len) {
-        return true;
-    }
-    return false;
+int ggml_metal_if_optimized(struct ggml_metal_context * ctx) {
+    return ctx->concur_list_len;
+}
+
+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
@@ -368,17 +389,17 @@ 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_NODES];
+    int nodes_unused[GGML_MAX_CONCUR];
 
-    for (int i = 0; i < GGML_MAX_NODES; i++) {ctx->concur_list[i] = 0;}
-    for (int i = 0; i < gf->n_nodes; i++) {nodes_unused[i] = 1;}
+    for (int i = 0; i < GGML_MAX_CONCUR; i++) { ctx->concur_list[i] = 0; }
+    for (int i = 0; i < gf->n_nodes;     i++) { nodes_unused[i]     = 1; }
     ctx->concur_list_len = 0;
 
-    int n_left = gf->n_nodes;
-    int n_start = 0; // all nodes before n_start at nodes_unused array have been sorted and store back to ctx->concur_list
-    int level_pos = 0;  // at ctx->concur_list, the last layer (level) ends at level_pos
+    int n_left    = gf->n_nodes;
+    int n_start   = 0; // all nodes before n_start at nodes_unused array have been sorted and store back to ctx->concur_list
+    int level_pos = 0; // at ctx->concur_list, the last layer (level) ends at level_pos
 
     while (n_left > 0) {
         // number of nodes at a layer (that can be issued concurrently)
@@ -386,28 +407,40 @@ void ggml_metal_graph_find_concurrency(
         for (int i = n_start; i < ((n_start + search_depth > gf->n_nodes) ? gf->n_nodes : n_start + search_depth); i++) {
             if (nodes_unused[i]) {
                 // if the requirements for gf->nodes[i] are satisfied
-                int exe_flag=1;
+                int exe_flag = 1;
+
                 // scan all srcs
                 for (int src_ind = 0; src_ind < GGML_MAX_SRC; src_ind++) {
                     struct ggml_tensor * src_cur = gf->nodes[i]->src[src_ind];
                     if (src_cur) {
                         // if is leaf nodes it's satisfied.
-                        if (src_cur->op == GGML_OP_NONE && src_cur->grad == NULL) {continue;}
+                        // TODO: ggml_is_leaf()
+                        if (src_cur->op == GGML_OP_NONE && src_cur->grad == NULL) {
+                            continue;
+                        }
 
                         // otherwise this src should be the output from previous nodes.
                         int is_found = 0;
+
                         // scan 2*search_depth back because we inserted barrier.
-                        for (int j = ((level_pos - 2*search_depth) < 0 ? 0 : (level_pos - 2*search_depth)); j < level_pos; j++) {
-                            if (gf->nodes[ctx->concur_list[j]] == src_cur) {is_found = 1; break;}
+                        //for (int j = ((level_pos - 2*search_depth) < 0 ? 0 : (level_pos - 2*search_depth)); j < level_pos; j++) {
+                        for (int j = MAX(0, level_pos - 2*search_depth); j < level_pos; j++) {
+                            if (ctx->concur_list[j] >= 0 && gf->nodes[ctx->concur_list[j]] == src_cur) {
+                                is_found = 1;
+                                break;
+                            }
+                        }
+                        if (is_found == 0) {
+                            exe_flag = 0;
+                            break;
                         }
-                        if (is_found == 0) {exe_flag = 0; break;}
                     }
                 }
-                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;
-                    int64_t length = (int64_t) ggml_nbytes(gf->nodes[i]);
+                    int64_t length     = (int64_t) ggml_nbytes(gf->nodes[i]);
                     for (int j = n_start; j < i; j++) {
                         if (nodes_unused[j] && gf->nodes[j]->op != GGML_OP_RESHAPE \
                                             && gf->nodes[j]->op != GGML_OP_VIEW \
@@ -416,9 +449,9 @@ void ggml_metal_graph_find_concurrency(
                             if (((int64_t)gf->nodes[j]->data) >= data_start + length || \
                                 ((int64_t)gf->nodes[j]->data) + (int64_t) ggml_nbytes(gf->nodes[j]) <= data_start) {
                                 continue;
-                            } else {
-                                exe_flag = 0;
                             }
+
+                            exe_flag = 0;
                         }
                     }
                 }
@@ -435,11 +468,13 @@ void ggml_metal_graph_find_concurrency(
         ctx->concur_list[level_pos + concurrency] = -1;
         ctx->concur_list_len++;
         // jump all sorted nodes at nodes_bak
-        while (!nodes_unused[n_start]) {n_start++;}
+        while (!nodes_unused[n_start]) {
+            n_start++;
+        }
         level_pos += concurrency + 1;
     }
 
-    if (ctx->concur_list_len > GGML_MAX_NODES) {
+    if (ctx->concur_list_len > GGML_MAX_CONCUR) {
         fprintf(stderr, "%s: too many elements for metal ctx->concur_list!\n", __func__);
     }
 }
@@ -453,7 +488,7 @@ void ggml_metal_graph_compute(
     // else fallback to serial dispatch
     MTLComputePassDescriptor * edesc = MTLComputePassDescriptor.computePassDescriptor;
 
-    const bool has_concur = ctx->concur_list_len && ctx->concur_list_len <= GGML_MAX_NODES;
+    const bool has_concur = ctx->concur_list_len && ctx->concur_list_len <= GGML_MAX_CONCUR;
 
     const int n_nodes  = has_concur ? ctx->concur_list_len      : gf->n_nodes;
     edesc.dispatchType = has_concur ? MTLDispatchTypeConcurrent : MTLDispatchTypeSerial;
@@ -485,7 +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;
@@ -494,10 +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;
                 }
@@ -571,10 +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];
@@ -592,10 +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];
@@ -613,10 +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];
@@ -632,10 +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];
@@ -646,10 +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];
@@ -660,10 +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];
@@ -680,10 +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];
@@ -698,10 +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];
@@ -719,53 +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) {
-
-                                if (encoder != nil) {
-                                    [encoder endEncoding];
-                                    encoder = nil;
+                                src1t == GGML_TYPE_F32 &&
+                                [ctx->device supportsFamily:MTLGPUFamilyApple7] &&
+                                ne00%32 == 0 &&
+                                ne11 > 1) {
+                                    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];
+                                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                                    [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+                                    [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
+                                    [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:5];
+                                    [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:6];
+                                    [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:7];
+                                    [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:8];
+                                    [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:9];
+                                    [encoder setBytes:&gqa length:sizeof(gqa) atIndex:10];
+                                    [encoder setThreadgroupMemoryLength:8192 atIndex:0];
+                                    [encoder dispatchThreadgroups:MTLSizeMake( (ne11+31)/32, (ne01+63) / 64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
                                 }
-
-                                MPSDataType src0dt = src0t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
-                                MPSDataType src1dt = src1t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
-
-                                // for F32 x F32 we use MPS
-                                MPSMatrixDescriptor * desc0 = [MPSMatrixDescriptor
-                                    matrixDescriptorWithRows:ne01 columns:ne00 rowBytes:src0->nb[1] dataType:src0dt];
-
-                                MPSMatrixDescriptor * desc1 = [MPSMatrixDescriptor
-                                    matrixDescriptorWithRows:ne11 columns:ne10 rowBytes:src1->nb[1] dataType:src1dt];
-
-                                MPSMatrixDescriptor * desc  = [MPSMatrixDescriptor
-                                    matrixDescriptorWithRows:ne1 columns:ne0 rowBytes:dst->nb[1] dataType:MPSDataTypeFloat32];
-
-                                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 {
-                                if (encoder == nil) {
-                                    encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
-                                }
-
+                            else {
                                 int nth0 = 32;
                                 int nth1 = 1;
 
@@ -864,23 +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)];
@@ -889,10 +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;
@@ -918,10 +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));
 
@@ -941,10 +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;
@@ -963,10 +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);
@@ -1006,10 +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];
@@ -1050,10 +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) {

ggml.c

@@ -195,8 +195,8 @@ typedef void * thread_ret_t;
 #define GGML_ALIGNED_MALLOC(size)  _aligned_malloc(size, GGML_MEM_ALIGN)
 #define GGML_ALIGNED_FREE(ptr)     _aligned_free(ptr)
 #else
-inline static void* ggml_aligned_malloc(size_t size) {
-    void* aligned_memory = NULL;
+inline static void * ggml_aligned_malloc(size_t size) {
+    void * aligned_memory = NULL;
 #ifdef GGML_USE_METAL
     int result = posix_memalign(&aligned_memory, getpagesize(), size);
 #else
@@ -3811,7 +3811,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "CROSS_ENTROPY_LOSS_BACK",
 };
 
-static_assert(GGML_OP_COUNT == 59, "GGML_OP_COUNT != 59");
+static_assert(GGML_OP_COUNT == 62, "GGML_OP_COUNT != 62");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -3883,7 +3883,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "cross_entropy_loss_back(x,y)",
 };
 
-static_assert(GGML_OP_COUNT == 59, "GGML_OP_COUNT != 59");
+static_assert(GGML_OP_COUNT == 62, "GGML_OP_COUNT != 62");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
@@ -4110,7 +4110,7 @@ size_t ggml_nbytes(const struct ggml_tensor * tensor) {
     //
     // is enough, but just in case, adding the second part
 
-    return MAX(tensor->ne[3]*tensor->nb[3], (ggml_nelements(tensor)*GGML_TYPE_SIZE[tensor->type])/GGML_BLCK_SIZE[tensor->type]);
+    return GGML_PAD(MAX(tensor->ne[3]*tensor->nb[3], (ggml_nelements(tensor)*GGML_TYPE_SIZE[tensor->type])/GGML_BLCK_SIZE[tensor->type]), GGML_MEM_ALIGN);
 }
 
 size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split) {
@@ -4253,7 +4253,7 @@ static inline bool ggml_is_padded_1d(const struct ggml_tensor * tensor) {
         tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
 }
 
-static inline bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
+bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
     return
@@ -4602,7 +4602,7 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         /*.ne           =*/ { 1, 1, 1, 1 },
         /*.nb           =*/ { 0, 0, 0, 0 },
         /*.op           =*/ GGML_OP_NONE,
-        /*.op_params    =*/ {0},
+        /*.op_params    =*/ { 0 },
         /*.is_param     =*/ false,
         /*.grad         =*/ NULL,
         /*.src          =*/ { NULL },
@@ -4634,6 +4634,7 @@ static struct ggml_tensor * ggml_new_tensor_impl(
 }
 
 static void ggml_set_op_params(struct ggml_tensor * tensor, const void * params, size_t params_size) {
+    GGML_ASSERT(tensor != NULL); // silence -Warray-bounds warnings
     assert(params_size <= GGML_MAX_OP_PARAMS);
     memcpy(tensor->op_params, params, params_size);
 }
@@ -6439,7 +6440,7 @@ struct ggml_tensor * ggml_permute(
     result->src[0] = a;
 
     int32_t params[] = { axis0, axis1, axis2, axis3 };
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     return result;
 }
@@ -6565,7 +6566,7 @@ static struct ggml_tensor * ggml_diag_mask_inf_impl(
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     int32_t params[] = { n_past, inplace ? 1 : 0 };
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     result->op   = GGML_OP_DIAG_MASK_INF;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6605,7 +6606,7 @@ static struct ggml_tensor * ggml_diag_mask_zero_impl(
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     int32_t params[] = { n_past, inplace ? 1 : 0 };
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     result->op   = GGML_OP_DIAG_MASK_ZERO;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6721,9 +6722,9 @@ static struct ggml_tensor * ggml_rope_impl(
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     int32_t params[6] = { n_past, n_dims, mode, n_ctx };
-    memcpy(params + 4, &freq_base, sizeof(float));
+    memcpy(params + 4, &freq_base,  sizeof(float));
     memcpy(params + 5, &freq_scale, sizeof(float));
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     result->op   = GGML_OP_ROPE;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6797,7 +6798,7 @@ struct ggml_tensor * ggml_rope_back(
     struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
 
     int32_t params[] = { n_past, n_dims, mode, n_ctx };
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     result->op   = GGML_OP_ROPE_BACK;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6828,7 +6829,7 @@ struct ggml_tensor * ggml_alibi(
 
     int32_t op_params[3] = { n_past, n_head };
     memcpy(op_params + 2, &bias_max, sizeof(float));
-    ggml_set_op_params(result, &op_params, sizeof(op_params));
+    ggml_set_op_params(result, op_params, sizeof(op_params));
 
     result->op   = GGML_OP_ALIBI;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6855,7 +6856,7 @@ struct ggml_tensor * ggml_clamp(
     struct ggml_tensor * result = ggml_view_tensor(ctx, a);
 
     float params[] = { min, max };
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     result->op   = GGML_OP_CLAMP;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6890,10 +6891,10 @@ GGML_API struct ggml_tensor * ggml_conv_1d(
         ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0),
         a->ne[2], 1, 1,
     };
-    struct ggml_tensor* result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne);
 
     int32_t params[] = { s0, p0, d0 };
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     result->op = GGML_OP_CONV_1D;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6905,10 +6906,10 @@ GGML_API struct ggml_tensor * ggml_conv_1d(
 
 // ggml_conv_2d
 
-struct ggml_tensor* ggml_conv_2d(
-    struct ggml_context* ctx,
-    struct ggml_tensor * a,
-    struct ggml_tensor * b,
+struct ggml_tensor * ggml_conv_2d(
+    struct ggml_context * ctx,
+    struct ggml_tensor  * a,
+    struct ggml_tensor  * b,
     int                  s0,
     int                  s1,
     int                  p0,
@@ -6929,10 +6930,10 @@ struct ggml_tensor* ggml_conv_2d(
         ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1),
         a->ne[3], b->ne[3],
     };
-    struct ggml_tensor* result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
     int32_t params[] = { s0, s1, p0, p1, d0, d1 };
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     result->op = GGML_OP_CONV_2D;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6945,7 +6946,7 @@ struct ggml_tensor* ggml_conv_2d(
 
 // ggml_conv_1d_ph
 
-struct ggml_tensor* ggml_conv_1d_ph(
+struct ggml_tensor * ggml_conv_1d_ph(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         struct ggml_tensor  * b,
@@ -6963,7 +6964,7 @@ static int64_t ggml_calc_pool_output_size(int64_t ins, int ks, int s, int p) {
 
 // ggml_pool_1d
 
-struct ggml_tensor* ggml_pool_1d(
+struct ggml_tensor * ggml_pool_1d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         enum ggml_op_pool     op,
@@ -6982,10 +6983,10 @@ struct ggml_tensor* ggml_pool_1d(
         ggml_calc_pool_output_size(a->ne[0], k0, s0, p0),
         a->ne[1],
     };
-    struct ggml_tensor* result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne);
 
     int32_t params[] = { op, k0, s0, p0 };
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     result->op = GGML_OP_POOL_1D;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6996,7 +6997,7 @@ struct ggml_tensor* ggml_pool_1d(
 
 // ggml_pool_2d
 
-struct ggml_tensor* ggml_pool_2d(
+struct ggml_tensor * ggml_pool_2d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         enum ggml_op_pool     op,
@@ -7019,10 +7020,10 @@ struct ggml_tensor* ggml_pool_2d(
         ggml_calc_pool_output_size(a->ne[1], k1, s1, p1),
         a->ne[2],
     };
-    struct ggml_tensor* result = ggml_new_tensor(ctx, GGML_TYPE_F32, 3, ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 3, ne);
 
     int32_t params[] = { op, k0, k1, s0, s1, p0, p1 };
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     result->op = GGML_OP_POOL_2D;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -7190,7 +7191,7 @@ struct ggml_tensor * ggml_win_part(
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
     int32_t params[] = { npx, npy, w };
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     result->op   = GGML_OP_WIN_PART;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -7220,7 +7221,7 @@ struct ggml_tensor * ggml_win_unpart(
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 3, ne);
 
     int32_t params[] = { w };
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     result->op   = GGML_OP_WIN_UNPART;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -7349,7 +7350,7 @@ struct ggml_tensor * ggml_map_binary_inplace_f32(
     return ggml_map_binary_impl_f32(ctx, a, b, fun, true);
 }
 
-// ggml_map_custom1
+// ggml_map_custom1_f32
 
 static struct ggml_tensor * ggml_map_custom1_impl_f32(
         struct ggml_context          * ctx,
@@ -7366,7 +7367,7 @@ static struct ggml_tensor * ggml_map_custom1_impl_f32(
 
     ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
 
-    result->op = GGML_OP_MAP_CUSTOM1;
+    result->op = GGML_OP_MAP_CUSTOM1_F32;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
 
@@ -7387,7 +7388,7 @@ struct ggml_tensor * ggml_map_custom1_inplace_f32(
     return ggml_map_custom1_impl_f32(ctx, a, fun, true);
 }
 
-// ggml_map_custom2
+// ggml_map_custom2_f32
 
 static struct ggml_tensor * ggml_map_custom2_impl_f32(
         struct ggml_context          * ctx,
@@ -7405,7 +7406,7 @@ static struct ggml_tensor * ggml_map_custom2_impl_f32(
 
     ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
 
-    result->op = GGML_OP_MAP_CUSTOM2;
+    result->op = GGML_OP_MAP_CUSTOM2_F32;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
     result->src[1] = b;
@@ -7429,7 +7430,7 @@ struct ggml_tensor * ggml_map_custom2_inplace_f32(
     return ggml_map_custom2_impl_f32(ctx, a, b, fun, true);
 }
 
-// ggml_map_custom3
+// ggml_map_custom3_f32
 
 static struct ggml_tensor * ggml_map_custom3_impl_f32(
         struct ggml_context          * ctx,
@@ -7448,7 +7449,7 @@ static struct ggml_tensor * ggml_map_custom3_impl_f32(
 
     ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
 
-    result->op = GGML_OP_MAP_CUSTOM3;
+    result->op = GGML_OP_MAP_CUSTOM3_F32;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
     result->src[1] = b;
@@ -7475,6 +7476,190 @@ struct ggml_tensor * ggml_map_custom3_inplace_f32(
     return ggml_map_custom3_impl_f32(ctx, a, b, c, fun, true);
 }
 
+// ggml_map_custom1
+struct ggml_map_custom1_op_params {
+    ggml_custom1_op_t fun;
+    int n_tasks;
+    void * userdata;
+};
+
+static struct ggml_tensor * ggml_map_custom1_impl(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        const  ggml_custom1_op_t       fun,
+        int                            n_tasks,
+        void                         * userdata,
+        bool                           inplace) {
+    GGML_ASSERT(n_tasks == GGML_N_TASKS_MAX || n_tasks > 0);
+
+    bool is_node = false;
+
+    if (!inplace && a->grad) {
+        is_node = true;
+    }
+
+    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+
+    struct ggml_map_custom1_op_params params = {
+        /*.fun      =*/ fun,
+        /*.n_tasks  =*/ n_tasks,
+        /*.userdata =*/ userdata
+    };
+    ggml_set_op_params(result, (const void *) &params, sizeof(params));
+
+    result->op = GGML_OP_MAP_CUSTOM1;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = a;
+
+    return result;
+}
+
+struct ggml_tensor * ggml_map_custom1(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        const  ggml_custom1_op_t       fun,
+        int                            n_tasks,
+        void                         * userdata) {
+    return ggml_map_custom1_impl(ctx, a, fun, n_tasks, userdata, false);
+}
+
+struct ggml_tensor * ggml_map_custom1_inplace(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        const  ggml_custom1_op_t       fun,
+        int                            n_tasks,
+        void                         * userdata) {
+    return ggml_map_custom1_impl(ctx, a, fun, n_tasks, userdata, true);
+}
+
+// ggml_map_custom2
+
+struct ggml_map_custom2_op_params {
+    ggml_custom2_op_t fun;
+    int n_tasks;
+    void * userdata;
+};
+
+static struct ggml_tensor * ggml_map_custom2_impl(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        struct ggml_tensor           * b,
+        const  ggml_custom2_op_t       fun,
+        int                            n_tasks,
+        void                         * userdata,
+        bool                           inplace) {
+    GGML_ASSERT(n_tasks == GGML_N_TASKS_MAX || n_tasks > 0);
+
+    bool is_node = false;
+
+    if (!inplace && (a->grad || b->grad)) {
+        is_node = true;
+    }
+
+    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+
+    struct ggml_map_custom2_op_params params = {
+        /*.fun      =*/ fun,
+        /*.n_tasks  =*/ n_tasks,
+        /*.userdata =*/ userdata
+    };
+    ggml_set_op_params(result, (const void *) &params, sizeof(params));
+
+    result->op = GGML_OP_MAP_CUSTOM2;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = a;
+    result->src[1] = b;
+
+    return result;
+}
+
+struct ggml_tensor * ggml_map_custom2(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        struct ggml_tensor           * b,
+        const  ggml_custom2_op_t       fun,
+        int                            n_tasks,
+        void                         * userdata) {
+    return ggml_map_custom2_impl(ctx, a, b, fun, n_tasks, userdata, false);
+}
+
+struct ggml_tensor * ggml_map_custom2_inplace(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        struct ggml_tensor           * b,
+        const  ggml_custom2_op_t       fun,
+        int                            n_tasks,
+        void                         * userdata) {
+    return ggml_map_custom2_impl(ctx, a, b, fun, n_tasks, userdata, true);
+}
+
+// ggml_map_custom3
+
+struct ggml_map_custom3_op_params {
+    ggml_custom3_op_t fun;
+    int n_tasks;
+    void * userdata;
+};
+
+static struct ggml_tensor * ggml_map_custom3_impl(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        struct ggml_tensor           * b,
+        struct ggml_tensor           * c,
+        const  ggml_custom3_op_t       fun,
+        int                            n_tasks,
+        void                         * userdata,
+        bool                           inplace) {
+    GGML_ASSERT(n_tasks == GGML_N_TASKS_MAX || n_tasks > 0);
+
+    bool is_node = false;
+
+    if (!inplace && (a->grad || b->grad || c->grad)) {
+        is_node = true;
+    }
+
+    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+
+    struct ggml_map_custom3_op_params params = {
+        /*.fun      =*/ fun,
+        /*.n_tasks  =*/ n_tasks,
+        /*.userdata =*/ userdata
+    };
+    ggml_set_op_params(result, (const void *) &params, sizeof(params));
+
+    result->op = GGML_OP_MAP_CUSTOM3;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = a;
+    result->src[1] = b;
+    result->src[2] = c;
+
+    return result;
+}
+
+struct ggml_tensor * ggml_map_custom3(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        struct ggml_tensor           * b,
+        struct ggml_tensor           * c,
+        const  ggml_custom3_op_t       fun,
+        int                            n_tasks,
+        void                         * userdata) {
+    return ggml_map_custom3_impl(ctx, a, b, c, fun, n_tasks, userdata, false);
+}
+
+struct ggml_tensor * ggml_map_custom3_inplace(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        struct ggml_tensor           * b,
+        struct ggml_tensor           * c,
+        const  ggml_custom3_op_t       fun,
+        int                            n_tasks,
+        void                         * userdata) {
+    return ggml_map_custom3_impl(ctx, a, b, c, fun, n_tasks, userdata, true);
+}
+
+
+
 // ggml_cross_entropy_loss
 
 struct ggml_tensor * ggml_cross_entropy_loss(
@@ -9283,8 +9468,8 @@ static void ggml_compute_forward_sum_rows_f32(
     for (int64_t i3 = 0; i3 < ne03; i3++) {
         for (int64_t i2 = 0; i2 < ne02; i2++) {
             for (int64_t i1 = 0; i1 < ne01; i1++) {
-                float* src_row = (float *) ((char *) src0->data + i1*nb01 + i2*nb02 + i3*nb03);
-                float* dst_row = (float *) ((char *) dst->data  + i1*nb1  + i2*nb2  + i3*nb3);
+                float * src_row = (float *) ((char *) src0->data + i1*nb01 + i2*nb02 + i3*nb03);
+                float * dst_row = (float *) ((char *) dst->data  + i1*nb1  + i2*nb2  + i3*nb3);
                 float row_sum = 0;
                 ggml_vec_sum_f32(ne00, &row_sum, src_row);
                 dst_row[0] = row_sum;
@@ -10546,72 +10731,96 @@ static void ggml_compute_forward_mul_mat(
         return;
     }
 
-    // parallelize by src0 rows
-    const int64_t dr = (ne01 + nth - 1)/nth;
-
-    const int64_t ir10 = dr*ith;
-    const int64_t ir11 = MIN(ir10 + dr, ne01);
-
-    // src1 rows
-    const int64_t nr1 = ne11*ne12*ne13;
-
     const void * wdata    = (src1->type == vec_dot_type) ? src1->data : params->wdata;
     const size_t row_size = ne10*GGML_TYPE_SIZE[vec_dot_type]/GGML_BLCK_SIZE[vec_dot_type];
 
-    for (int64_t ir1 = 0; ir1 < nr1; ++ir1) {
-        const int64_t i13 = (ir1/(ne12*ne11));
-        const int64_t i12 = (ir1 - i13*ne12*ne11)/ne11;
-        const int64_t i11 = (ir1 - i13*ne12*ne11 - i12*ne11);
+    const int64_t nr0 = ne01;           // src0 rows
+    const int64_t nr1 = ne11*ne12*ne13; // src1 rows
 
-        const int64_t ir0 = (ir1/ne11)%(ne02*ne03);
-        const int64_t i03 = (ir0/(ne02));
-        // Hack for "Falcon multi-query-attention key stutter" / alternative to ggml_repeat2.
-        // See https://github.com/ggerganov/llama.cpp/issues/1602#issuecomment-1606087470:
-        // GG: this is likely the correct way to broadcast, though need some more thought
-        //     therefore leaving the comments to remind us for now
-        const int64_t i02 = (i12 / (ne12 / ne02));
-        // Original from PR/224 (and also essential/correct for non-broadcast matmuls in Falcon)
-        // const int64_t i02 = (ir0 - i03*ne02);
+    //printf("nr0 = %lld, nr1 = %lld\n", nr0, nr1);
 
-        const int64_t i1 = i11;
-        const int64_t i2 = i12;
-        const int64_t i3 = i13;
+    // distribute the thread work across the inner or outer loop based on which one is larger
 
-        const char * src0_row = (const char *) src0->data + (  0 + i02*nb02 + i03*nb03     );
+    const int64_t nth0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows
+    const int64_t nth1 = nr0 > nr1 ? 1 : nth; // parallelize by src1 rows
 
-        // desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides
-        //       if it is, then we have either copied the data to params->wdata and made it contiguous or we are using
-        //       the original src1 data pointer, so we should index using the indices directly
-        // TODO: this is a bit of a hack, we should probably have a better way to handle this
-        const char * src1_col = (const char *) wdata +
-            (src1_cont || src1->type != vec_dot_type
-             ? (i11      + i12*ne11 + i13*ne12*ne11)*row_size
-             : (i11*nb11 + i12*nb12 + i13*nb13));
+    const int64_t ith0 = ith % nth0;
+    const int64_t ith1 = ith / nth0;
 
-        float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3));
+    const int64_t dr0 = (nr0 + nth0 - 1)/nth0;
+    const int64_t dr1 = (nr1 + nth1 - 1)/nth1;
 
-        for (int64_t ir = ir10; ir < ir11; ++ir) {
-            vec_dot(ne00, &dst_col[ir], src0_row + ir*nb01, src1_col);
-        }
+    const int64_t ir010 = dr0*ith0;
+    const int64_t ir011 = MIN(ir010 + dr0, nr0);
+
+    const int64_t ir110 = dr1*ith1;
+    const int64_t ir111 = MIN(ir110 + dr1, nr1);
+
+    //printf("ir010 = %6lld, ir011 = %6lld, ir110 = %6lld, ir111 = %6lld\n", ir010, ir011, ir110, ir111);
+
+    // threads with no work simply yield (not sure if it helps)
+    if (ir010 >= ir011 || ir110 >= ir111) {
+        sched_yield();
+        return;
     }
 
-    //int64_t t1 = ggml_time_us();
-    //static int64_t acc = 0;
-    //acc += t1 - t0;
-    //if (t1 - t0 > 10) {
-    //    printf("\n");
-    //    printf("ne00 = %5d, ne01 = %5d, ne02 = %5d, ne03 = %5d\n", ne00, ne01, ne02, ne03);
-    //    printf("nb00 = %5d, nb01 = %5d, nb02 = %5d, nb03 = %5d\n", nb00, nb01, nb02, nb03);
-    //    printf("ne10 = %5d, ne11 = %5d, ne12 = %5d, ne13 = %5d\n", ne10, ne11, ne12, ne13);
+    assert(ne12 % ne02 == 0);
+    assert(ne13 % ne03 == 0);
 
-    //    printf("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX task %d/%d: %d us, acc = %d\n", ith, nth, (int) (t1 - t0), (int) acc);
-    //}
+    // broadcast factors
+    const int64_t r2 = ne12/ne02;
+    const int64_t r3 = ne13/ne03;
+
+    // block-tiling attempt
+    const int64_t blck_0 = 16;
+    const int64_t blck_1 = 16;
+
+    // attempt to reduce false-sharing (does not seem to make a difference)
+    float tmp[16];
+
+    for (int64_t iir1 = ir110; iir1 < ir111; iir1 += blck_1) {
+        for (int64_t iir0 = ir010; iir0 < ir011; iir0 += blck_0) {
+            for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir111; ++ir1) {
+                const int64_t i13 = (ir1/(ne12*ne11));
+                const int64_t i12 = (ir1 - i13*ne12*ne11)/ne11;
+                const int64_t i11 = (ir1 - i13*ne12*ne11 - i12*ne11);
+
+                // broadcast src0 into src1
+                const int64_t i03 = i13/r3;
+                const int64_t i02 = i12/r2;
+
+                const int64_t i1 = i11;
+                const int64_t i2 = i12;
+                const int64_t i3 = i13;
+
+                const char * src0_row = (const char *) src0->data + (0 + i02*nb02 + i03*nb03);
+
+                // desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides
+                //       if it is, then we have either copied the data to params->wdata and made it contiguous or we are using
+                //       the original src1 data pointer, so we should index using the indices directly
+                // TODO: this is a bit of a hack, we should probably have a better way to handle this
+                const char * src1_col = (const char *) wdata +
+                    (src1_cont || src1->type != vec_dot_type
+                     ? (i11      + i12*ne11 + i13*ne12*ne11)*row_size
+                     : (i11*nb11 + i12*nb12 + i13*nb13));
+
+                float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3));
+
+                //for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
+                //    vec_dot(ne00, &dst_col[ir0], src0_row + ir0*nb01, src1_col);
+                //}
+
+                for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
+                    vec_dot(ne00, &tmp[ir0 - iir0], src0_row + ir0*nb01, src1_col);
+                }
+                memcpy(&dst_col[iir0], tmp, (MIN(iir0 + blck_0, ir011) - iir0)*sizeof(float));
+            }
+        }
+    }
 }
 
-
 // ggml_compute_forward_out_prod
 
-
 static void ggml_compute_forward_out_prod_f32(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
@@ -12894,7 +13103,7 @@ static void ggml_compute_forward_pool_1d(
         const struct ggml_tensor * src0,
               struct ggml_tensor * dst) {
 
-    const int32_t* opts = (const int32_t*)dst->op_params;
+    const int32_t * opts = (const int32_t *)dst->op_params;
     enum ggml_op_pool op = opts[0];
     const int k0 = opts[1];
     const int s0 = opts[2];
@@ -14227,24 +14436,6 @@ static void ggml_compute_forward_map_custom1_f32(
     fun(dst, a);
 }
 
-
-static void ggml_compute_forward_map_custom1(
-        const struct ggml_compute_params * params,
-        const struct ggml_tensor * a,
-        struct ggml_tensor * dst,
-        const ggml_custom1_op_f32_t fun) {
-    switch (a->type) {
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_map_custom1_f32(params, a, dst, fun);
-            } break;
-        default:
-            {
-                GGML_ASSERT(false);
-            } break;
-    }
-}
-
 // ggml_compute_forward_map_custom2
 
 static void ggml_compute_forward_map_custom2_f32(
@@ -14263,24 +14454,6 @@ static void ggml_compute_forward_map_custom2_f32(
 }
 
 
-static void ggml_compute_forward_map_custom2(
-        const struct ggml_compute_params * params,
-        const struct ggml_tensor * a,
-        const struct ggml_tensor * b,
-        struct ggml_tensor * dst,
-        const ggml_custom2_op_f32_t fun) {
-    switch (a->type) {
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_map_custom2_f32(params, a, b, dst, fun);
-            } break;
-        default:
-            {
-                GGML_ASSERT(false);
-            } break;
-    }
-}
-
 // ggml_compute_forward_map_custom3
 
 static void ggml_compute_forward_map_custom3_f32(
@@ -14299,24 +14472,52 @@ static void ggml_compute_forward_map_custom3_f32(
     fun(dst, a, b, c);
 }
 
+// ggml_compute_forward_map_custom1
+
+static void ggml_compute_forward_map_custom1(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * a,
+              struct ggml_tensor * dst) {
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    struct ggml_map_custom1_op_params * p = (struct ggml_map_custom1_op_params *) dst->op_params;
+
+    p->fun(dst, a, params->ith, params->nth, p->userdata);
+}
+
+// ggml_compute_forward_map_custom2
+
+static void ggml_compute_forward_map_custom2(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * a,
+        const struct ggml_tensor * b,
+              struct ggml_tensor * dst) {
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    struct ggml_map_custom2_op_params * p = (struct ggml_map_custom2_op_params *) dst->op_params;
+
+    p->fun(dst, a, b, params->ith, params->nth, p->userdata);
+}
+
+// ggml_compute_forward_map_custom3
 
 static void ggml_compute_forward_map_custom3(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * a,
         const struct ggml_tensor * b,
         const struct ggml_tensor * c,
-        struct ggml_tensor * dst,
-        const ggml_custom3_op_f32_t fun) {
-    switch (a->type) {
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_map_custom3_f32(params, a, b, c, dst, fun);
-            } break;
-        default:
-            {
-                GGML_ASSERT(false);
-            } break;
+              struct ggml_tensor * dst) {
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+        return;
     }
+
+    struct ggml_map_custom3_op_params * p = (struct ggml_map_custom3_op_params *) dst->op_params;
+
+    p->fun(dst, a, b, c, params->ith, params->nth, p->userdata);
 }
 
 // ggml_compute_forward_cross_entropy_loss
@@ -14838,25 +15039,40 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
                 ggml_compute_forward_map_binary(params, tensor->src[0], tensor->src[1], tensor, fun);
             }
             break;
-        case GGML_OP_MAP_CUSTOM1:
+        case GGML_OP_MAP_CUSTOM1_F32:
             {
                 ggml_custom1_op_f32_t fun;
                 memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_custom1(params, tensor->src[0], tensor, fun);
+                ggml_compute_forward_map_custom1_f32(params, tensor->src[0], tensor, fun);
+            }
+            break;
+        case GGML_OP_MAP_CUSTOM2_F32:
+            {
+                ggml_custom2_op_f32_t fun;
+                memcpy(&fun, tensor->op_params, sizeof(fun));
+                ggml_compute_forward_map_custom2_f32(params, tensor->src[0], tensor->src[1], tensor, fun);
+            }
+            break;
+        case GGML_OP_MAP_CUSTOM3_F32:
+            {
+                ggml_custom3_op_f32_t fun;
+                memcpy(&fun, tensor->op_params, sizeof(fun));
+                ggml_compute_forward_map_custom3_f32(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor, fun);
+            }
+            break;
+        case GGML_OP_MAP_CUSTOM1:
+            {
+                ggml_compute_forward_map_custom1(params, tensor->src[0], tensor);
             }
             break;
         case GGML_OP_MAP_CUSTOM2:
             {
-                ggml_custom2_op_f32_t fun;
-                memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_custom2(params, tensor->src[0], tensor->src[1], tensor, fun);
+                ggml_compute_forward_map_custom2(params, tensor->src[0], tensor->src[1], tensor);
             }
             break;
         case GGML_OP_MAP_CUSTOM3:
             {
-                ggml_custom3_op_f32_t fun;
-                memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_custom3(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor, fun);
+                ggml_compute_forward_map_custom3(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor);
             }
             break;
         case GGML_OP_CROSS_ENTROPY_LOSS:
@@ -15664,6 +15880,9 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             } break;
         case GGML_OP_MAP_UNARY:
         case GGML_OP_MAP_BINARY:
+        case GGML_OP_MAP_CUSTOM1_F32:
+        case GGML_OP_MAP_CUSTOM2_F32:
+        case GGML_OP_MAP_CUSTOM3_F32:
         case GGML_OP_MAP_CUSTOM1:
         case GGML_OP_MAP_CUSTOM2:
         case GGML_OP_MAP_CUSTOM3:
@@ -16449,12 +16668,39 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
             case GGML_OP_WIN_UNPART:
             case GGML_OP_MAP_UNARY:
             case GGML_OP_MAP_BINARY:
-            case GGML_OP_MAP_CUSTOM1:
-            case GGML_OP_MAP_CUSTOM2:
-            case GGML_OP_MAP_CUSTOM3:
+            case GGML_OP_MAP_CUSTOM1_F32:
+            case GGML_OP_MAP_CUSTOM2_F32:
+            case GGML_OP_MAP_CUSTOM3_F32:
                 {
                     n_tasks = 1;
                 } break;
+            case GGML_OP_MAP_CUSTOM1:
+                {
+                    struct ggml_map_custom1_op_params * p = (struct ggml_map_custom1_op_params *) node->op_params;
+                    if (p->n_tasks == GGML_N_TASKS_MAX) {
+                        n_tasks = n_threads;
+                    } else {
+                        n_tasks = MIN(p->n_tasks, n_threads);
+                    }
+                } break;
+            case GGML_OP_MAP_CUSTOM2:
+                {
+                    struct ggml_map_custom2_op_params * p = (struct ggml_map_custom2_op_params *) node->op_params;
+                    if (p->n_tasks == GGML_N_TASKS_MAX) {
+                        n_tasks = n_threads;
+                    } else {
+                        n_tasks = MIN(p->n_tasks, n_threads);
+                    }
+                } break;
+            case GGML_OP_MAP_CUSTOM3:
+                {
+                    struct ggml_map_custom3_op_params * p = (struct ggml_map_custom3_op_params *) node->op_params;
+                    if (p->n_tasks == GGML_N_TASKS_MAX) {
+                        n_tasks = n_threads;
+                    } else {
+                        n_tasks = MIN(p->n_tasks, n_threads);
+                    }
+                } break;
             case GGML_OP_CROSS_ENTROPY_LOSS:
                 {
                     n_tasks = n_threads;

ggml.h

@@ -183,6 +183,15 @@
 #    define GGML_API
 #endif
 
+// TODO: support for clang
+#ifdef __GNUC__
+#    define GGML_DEPRECATED(func, hint) func __attribute__((deprecated(hint)))
+#elif defined(_MSC_VER)
+#    define GGML_DEPRECATED(func, hint) __declspec(deprecated(hint)) func
+#else
+#    define GGML_DEPRECATED(func, hint) func
+#endif
+
 #include <stdint.h>
 #include <stddef.h>
 #include <stdbool.h>
@@ -374,6 +383,10 @@ extern "C" {
         GGML_OP_MAP_UNARY,
         GGML_OP_MAP_BINARY,
 
+        GGML_OP_MAP_CUSTOM1_F32,
+        GGML_OP_MAP_CUSTOM2_F32,
+        GGML_OP_MAP_CUSTOM3_F32,
+
         GGML_OP_MAP_CUSTOM1,
         GGML_OP_MAP_CUSTOM2,
         GGML_OP_MAP_CUSTOM3,
@@ -570,6 +583,8 @@ extern "C" {
     GGML_API bool ggml_is_contiguous(const struct ggml_tensor * tensor);
     GGML_API bool ggml_is_permuted  (const struct ggml_tensor * tensor);
 
+    GGML_API bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
+
     // use this to compute the memory overhead of a tensor
     GGML_API size_t ggml_tensor_overhead(void);
 
@@ -1240,7 +1255,7 @@ extern "C" {
 
     // conv_1d with padding = half
     // alias for ggml_conv_1d(a, b, s, a->ne[0]/2, d)
-    GGML_API struct ggml_tensor* ggml_conv_1d_ph(
+    GGML_API struct ggml_tensor * ggml_conv_1d_ph(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
@@ -1253,7 +1268,7 @@ extern "C" {
         GGML_OP_POOL_COUNT,
     };
 
-    GGML_API struct ggml_tensor* ggml_pool_1d(
+    GGML_API struct ggml_tensor * ggml_pool_1d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             enum ggml_op_pool     op,
@@ -1261,7 +1276,7 @@ extern "C" {
             int                   s0, // stride
             int                   p0); // padding
 
-    GGML_API struct ggml_tensor* ggml_pool_2d(
+    GGML_API struct ggml_tensor * ggml_pool_2d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             enum ggml_op_pool     op,
@@ -1315,15 +1330,6 @@ extern "C" {
             int                   h0,
             int                   w);
 
-    // custom operators
-
-    typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);
-    typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
-
-    typedef void (*ggml_custom1_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *);
-    typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
-    typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
-
     GGML_API struct ggml_tensor * ggml_unary(
             struct ggml_context * ctx,
              struct ggml_tensor * a,
@@ -1334,63 +1340,138 @@ extern "C" {
         struct ggml_tensor  * a,
         enum ggml_unary_op op);
 
-    GGML_API struct ggml_tensor * ggml_map_unary_f32(
+    // custom operators
+
+    typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);
+    typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
+
+    typedef void (*ggml_custom1_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *);
+    typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
+    typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
+
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_f32(
             struct ggml_context        * ctx,
             struct ggml_tensor         * a,
-                   ggml_unary_op_f32_t   fun);
+                   ggml_unary_op_f32_t   fun),
+        "use ggml_map_custom1 instead");
 
-    GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32(
             struct ggml_context        * ctx,
             struct ggml_tensor         * a,
-                   ggml_unary_op_f32_t   fun);
+                   ggml_unary_op_f32_t   fun),
+        "use ggml_map_custom1_inplace instead");
 
-    GGML_API struct ggml_tensor * ggml_map_binary_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_f32(
             struct ggml_context         * ctx,
             struct ggml_tensor          * a,
             struct ggml_tensor          * b,
-                   ggml_binary_op_f32_t   fun);
+                   ggml_binary_op_f32_t   fun),
+        "use ggml_map_custom2 instead");
 
-    GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32(
             struct ggml_context         * ctx,
             struct ggml_tensor          * a,
             struct ggml_tensor          * b,
-                   ggml_binary_op_f32_t   fun);
+                   ggml_binary_op_f32_t   fun),
+        "use ggml_map_custom2_inplace instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom1_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
-                   ggml_custom1_op_f32_t   fun);
+                   ggml_custom1_op_f32_t   fun),
+        "use ggml_map_custom1 instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
-                   ggml_custom1_op_f32_t   fun);
+                   ggml_custom1_op_f32_t   fun),
+        "use ggml_map_custom1_inplace instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom2_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
             struct ggml_tensor           * b,
-                   ggml_custom2_op_f32_t   fun);
+                   ggml_custom2_op_f32_t   fun),
+        "use ggml_map_custom2 instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
             struct ggml_tensor           * b,
-                   ggml_custom2_op_f32_t   fun);
+                   ggml_custom2_op_f32_t   fun),
+        "use ggml_map_custom2_inplace instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom3_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
             struct ggml_tensor           * b,
             struct ggml_tensor           * c,
-                   ggml_custom3_op_f32_t   fun);
+                   ggml_custom3_op_f32_t   fun),
+        "use ggml_map_custom3 instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
             struct ggml_tensor           * b,
             struct ggml_tensor           * c,
-                   ggml_custom3_op_f32_t   fun);
+                   ggml_custom3_op_f32_t   fun),
+        "use ggml_map_custom3_inplace instead");
+
+    // custom operators v2
+
+    typedef void (*ggml_custom1_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, int ith, int nth, void * userdata);
+    typedef void (*ggml_custom2_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, int ith, int nth, void * userdata);
+    typedef void (*ggml_custom3_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, const struct ggml_tensor * c, int ith, int nth, void * userdata);
+
+    #define GGML_N_TASKS_MAX -1
+
+    GGML_API struct ggml_tensor * ggml_map_custom1(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            ggml_custom1_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom1_inplace(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            ggml_custom1_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom2(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            struct ggml_tensor    * b,
+            ggml_custom2_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom2_inplace(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            struct ggml_tensor    * b,
+            ggml_custom2_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom3(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            struct ggml_tensor    * b,
+            struct ggml_tensor    * c,
+            ggml_custom3_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom3_inplace(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            struct ggml_tensor    * b,
+            struct ggml_tensor    * c,
+            ggml_custom3_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
 
     // loss function
 

llama-util.h

@@ -219,7 +219,7 @@ struct llama_mmap {
         // prefetch/readahead impairs performance on NUMA systems
         if (numa) { prefetch = 0; }
 #ifdef __linux__
-        if (prefetch) { flags |= MAP_POPULATE; }
+        if (prefetch >= file->size) { flags |= MAP_POPULATE; }
 #endif
         addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
         if (addr == MAP_FAILED) {
@@ -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
-            WIN32_MEMORY_RANGE_ENTRY range;
-            range.VirtualAddress = addr;
-            range.NumberOfBytes = (SIZE_T)size;
-            if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
-                fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
-                        llama_format_win_err(GetLastError()).c_str());
+            // 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 (!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

@@ -56,7 +56,14 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-#if !defined(GGML_USE_CUBLAS) && !defined(GGML_USE_METAL)
+static void llama_log_internal(llama_log_level level, const char* format, ...);
+static void llama_log_callback_default(llama_log_level level, const char * text, void * user_data);
+#define LLAMA_LOG_INFO(...)  llama_log_internal(LLAMA_LOG_LEVEL_INFO , __VA_ARGS__)
+#define LLAMA_LOG_WARN(...)  llama_log_internal(LLAMA_LOG_LEVEL_WARN , __VA_ARGS__)
+#define LLAMA_LOG_ERROR(...) llama_log_internal(LLAMA_LOG_LEVEL_ERROR, __VA_ARGS__)
+
+
+#if !defined(GGML_USE_CUBLAS)
 #include "ggml-alloc.h"
 #define LLAMA_USE_ALLOCATOR
 #else
@@ -149,7 +156,7 @@ static const std::map<e_model, size_t> & MEM_REQ_EVAL()
 }
 
 // amount of VRAM needed per batch size to hold temporary results
-// the values for 3b and 65b are not derived from testing but instead chosen conservatively
+// the values for 3b are not derived from testing but instead chosen conservatively
 static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_BASE()
 {
     static std::map<e_model, size_t> k_sizes = {
@@ -157,14 +164,14 @@ static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_BASE()
         { MODEL_7B,   512ull * kB },
         { MODEL_13B,  640ull * kB },
         { MODEL_30B,  768ull * kB },
-        { MODEL_65B, 1536ull * kB },
-        { MODEL_70B, 1536ull * kB }, // TODO (likely can be reduced)
+        { MODEL_65B, 1280ull * kB },
+        { MODEL_70B, 1280ull * kB },
     };
     return k_sizes;
 }
 
 // amount of VRAM needed per batch size and context to hold temporary results
-// the values for 3b and 65b are not derived from testing but instead chosen conservatively
+// the values for 3b are not derived from testing but instead chosen conservatively
 static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_PER_CONTEXT()
 {
     static std::map<e_model, size_t> k_sizes = {
@@ -172,8 +179,8 @@ static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_PER_CONTEXT()
         { MODEL_7B,  128ull },
         { MODEL_13B, 160ull },
         { MODEL_30B, 208ull },
-        { MODEL_65B, 416ull },
-        { MODEL_70B, 416ull }, // TODO (likely can be reduced)
+        { MODEL_65B, 256ull },
+        { MODEL_70B, 256ull },
     };
     return k_sizes;
 }
@@ -438,6 +445,14 @@ struct llama_context {
     }
 };
 
+struct llama_state {
+    // We save the log callback globally
+    llama_log_callback log_callback = llama_log_callback_default;
+    void * log_callback_user_data = nullptr;
+};
+// global state
+static llama_state g_state;
+
 template <typename T>
 static T checked_mul(T a, T b) {
     T ret = a * b;
@@ -504,7 +519,7 @@ struct llama_file_loader {
 
     llama_file_loader(const char * fname, llama_load_tensors_map & tensors_map)
         : file(fname, "rb") {
-        fprintf(stderr, "llama.cpp: loading model from %s\n", fname);
+        LLAMA_LOG_INFO("llama.cpp: loading model from %s\n", fname);
         read_magic();
         read_hparams();
         read_vocab();
@@ -619,7 +634,7 @@ struct llama_file_saver {
     llama_file_loader * any_file_loader;
     llama_file_saver(const char * fname, llama_file_loader * any_file_loader, enum llama_ftype new_ftype)
         : file(fname, "wb"), any_file_loader(any_file_loader) {
-        fprintf(stderr, "llama.cpp: saving model to %s\n", fname);
+        LLAMA_LOG_INFO("llama.cpp: saving model to %s\n", fname);
         write_magic();
         write_hparams(new_ftype);
         write_vocab();
@@ -640,7 +655,7 @@ struct llama_file_saver {
     }
     void write_vocab() {
         if (any_file_loader->file_version == LLAMA_FILE_VERSION_GGML) {
-            fprintf(stderr, "llama.cpp: WARNING: input is an old file that doesn't have scores; will add dummy scores\n");
+            LLAMA_LOG_WARN("llama.cpp: WARNING: input is an old file that doesn't have scores; will add dummy scores\n");
         }
         uint32_t n_vocab = any_file_loader->hparams.n_vocab;
         for (uint32_t i = 0; i < n_vocab; i++) {
@@ -747,12 +762,12 @@ struct llama_model_loader {
 
     void load_all_data(llama_progress_callback progress_callback, void *  progress_callback_user_data, llama_mlock * lmlock) {
         size_t data_size = 0;
-        size_t prefetch_size = 0;
+        size_t prefetch_size = file_loader->file.size;
         size_t lock_size = 0;
         for (const llama_load_tensor & lt : tensors_map.tensors) {
             data_size += lt.size;
-            if (lt.ggml_tensor->backend == GGML_BACKEND_CPU) {
-                prefetch_size += lt.size;
+            if (lt.ggml_tensor->backend != GGML_BACKEND_CPU) {
+                prefetch_size -= lt.size;
             }
         }
 
@@ -831,7 +846,7 @@ struct llama_model_loader {
             uint8_t byte = lt.data[i];
             sum = byte + (sum << 6) + (sum << 16) - sum; // sdbm hash
         }
-        fprintf(stderr, "%s checksum: %#08x (%s, size %zu)\n", lt.name.c_str(), sum,
+        LLAMA_LOG_INFO("%s checksum: %#08x (%s, size %zu)\n", lt.name.c_str(), sum,
                 llama_format_tensor_shape(lt.ne).c_str(), lt.size);
     }
 
@@ -864,7 +879,7 @@ static bool kv_cache_init(
     cache.ctx = ggml_init(params);
 
     if (!cache.ctx) {
-        fprintf(stderr, "%s: failed to allocate memory for kv cache\n", __func__);
+        LLAMA_LOG_ERROR("%s: failed to allocate memory for kv cache\n", __func__);
         return false;
     }
 
@@ -1076,7 +1091,7 @@ static void llama_model_load_internal(
         LLAMA_ASSERT(hparams.n_head % n_gqa == 0);
         hparams.n_head_kv = hparams.n_head / n_gqa;
         if (model.type == e_model::MODEL_65B && n_gqa == 8) {
-            fprintf(stderr, "%s: warning: assuming 70B model based on GQA == %d\n", __func__, n_gqa);
+            LLAMA_LOG_WARN("%s: warning: assuming 70B model based on GQA == %d\n", __func__, n_gqa);
             model.type = e_model::MODEL_70B;
             hparams.f_ffn_mult = 1.3f; // from the params.json of the 70B model
         }
@@ -1092,22 +1107,22 @@ static void llama_model_load_internal(
     //const uint32_t n_ff = 28672;
 
     {
-        fprintf(stderr, "%s: format     = %s\n",   __func__, llama_file_version_name(file_version));
-        fprintf(stderr, "%s: n_vocab    = %u\n",   __func__, hparams.n_vocab);
-        fprintf(stderr, "%s: n_ctx      = %u\n",   __func__, hparams.n_ctx);
-        fprintf(stderr, "%s: n_embd     = %u\n",   __func__, hparams.n_embd);
-        fprintf(stderr, "%s: n_mult     = %u\n",   __func__, hparams.n_mult);
-        fprintf(stderr, "%s: n_head     = %u\n",   __func__, hparams.n_head);
-        fprintf(stderr, "%s: n_head_kv  = %u\n",   __func__, hparams.n_head_kv);
-        fprintf(stderr, "%s: n_layer    = %u\n",   __func__, hparams.n_layer);
-        fprintf(stderr, "%s: n_rot      = %u\n",   __func__, hparams.n_rot); // a.k.a. n_embd_head, n_head_dim
-        fprintf(stderr, "%s: n_gqa      = %u\n",   __func__, hparams.n_gqa());
-        fprintf(stderr, "%s: rnorm_eps  = %.1e\n", __func__, hparams.f_rms_norm_eps);
-        fprintf(stderr, "%s: n_ff       = %u\n",   __func__, n_ff);
-        fprintf(stderr, "%s: freq_base  = %.1f\n", __func__, hparams.rope_freq_base);
-        fprintf(stderr, "%s: freq_scale = %g\n",   __func__, hparams.rope_freq_scale);
-        fprintf(stderr, "%s: ftype      = %u (%s)\n", __func__, hparams.ftype, llama_ftype_name(hparams.ftype));
-        fprintf(stderr, "%s: model size = %s\n",   __func__, llama_model_type_name(model.type));
+        LLAMA_LOG_INFO("%s: format     = %s\n",   __func__, llama_file_version_name(file_version));
+        LLAMA_LOG_INFO("%s: n_vocab    = %u\n",   __func__, hparams.n_vocab);
+        LLAMA_LOG_INFO("%s: n_ctx      = %u\n",   __func__, hparams.n_ctx);
+        LLAMA_LOG_INFO("%s: n_embd     = %u\n",   __func__, hparams.n_embd);
+        LLAMA_LOG_INFO("%s: n_mult     = %u\n",   __func__, hparams.n_mult);
+        LLAMA_LOG_INFO("%s: n_head     = %u\n",   __func__, hparams.n_head);
+        LLAMA_LOG_INFO("%s: n_head_kv  = %u\n",   __func__, hparams.n_head_kv);
+        LLAMA_LOG_INFO("%s: n_layer    = %u\n",   __func__, hparams.n_layer);
+        LLAMA_LOG_INFO("%s: n_rot      = %u\n",   __func__, hparams.n_rot); // a.k.a. n_embd_head, n_head_dim
+        LLAMA_LOG_INFO("%s: n_gqa      = %u\n",   __func__, hparams.n_gqa());
+        LLAMA_LOG_INFO("%s: rnorm_eps  = %.1e\n", __func__, hparams.f_rms_norm_eps);
+        LLAMA_LOG_INFO("%s: n_ff       = %u\n",   __func__, n_ff);
+        LLAMA_LOG_INFO("%s: freq_base  = %.1f\n", __func__, hparams.rope_freq_base);
+        LLAMA_LOG_INFO("%s: freq_scale = %g\n",   __func__, hparams.rope_freq_scale);
+        LLAMA_LOG_INFO("%s: ftype      = %u (%s)\n", __func__, hparams.ftype, llama_ftype_name(hparams.ftype));
+        LLAMA_LOG_INFO("%s: model size = %s\n",   __func__, llama_model_type_name(model.type));
     }
 
     if (file_version < LLAMA_FILE_VERSION_GGJT_V2) {
@@ -1135,7 +1150,7 @@ static void llama_model_load_internal(
     size_t ctx_size;
     size_t mmapped_size;
     ml->calc_sizes(&ctx_size, &mmapped_size);
-    fprintf(stderr, "%s: ggml ctx size = %7.2f MB\n", __func__, ctx_size/1024.0/1024.0);
+    LLAMA_LOG_INFO("%s: ggml ctx size = %7.2f MB\n", __func__, ctx_size/1024.0/1024.0);
 
     // create the ggml context
     {
@@ -1160,13 +1175,13 @@ static void llama_model_load_internal(
     (void) main_gpu;
     (void) mul_mat_q;
 #if defined(GGML_USE_CUBLAS)
-    fprintf(stderr, "%s: using CUDA for GPU acceleration\n", __func__);
+    LLAMA_LOG_INFO("%s: using CUDA for GPU acceleration\n", __func__);
     ggml_cuda_set_main_device(main_gpu);
     ggml_cuda_set_mul_mat_q(mul_mat_q);
 #define LLAMA_BACKEND_OFFLOAD       GGML_BACKEND_GPU
 #define LLAMA_BACKEND_OFFLOAD_SPLIT GGML_BACKEND_GPU_SPLIT
 #elif defined(GGML_USE_CLBLAST)
-    fprintf(stderr, "%s: using OpenCL for GPU acceleration\n", __func__);
+    LLAMA_LOG_INFO("%s: using OpenCL for GPU acceleration\n", __func__);
 #define LLAMA_BACKEND_OFFLOAD       GGML_BACKEND_GPU
 #define LLAMA_BACKEND_OFFLOAD_SPLIT GGML_BACKEND_GPU
 #else
@@ -1271,14 +1286,14 @@ static void llama_model_load_internal(
         const size_t mem_required_state =
             scale*hparams.kv_size();
 
-        fprintf(stderr, "%s: mem required  = %7.2f MB (+ %7.2f MB per state)\n", __func__,
+        LLAMA_LOG_INFO("%s: mem required  = %7.2f MB (+ %7.2f MB per state)\n", __func__,
                 mem_required / 1024.0 / 1024.0, mem_required_state / 1024.0 / 1024.0);
 
         (void) vram_scratch;
         (void) n_batch;
 #ifdef GGML_USE_CUBLAS
         if (low_vram) {
-            fprintf(stderr, "%s: not allocating a VRAM scratch buffer due to low VRAM option\n", __func__);
+            LLAMA_LOG_INFO("%s: not allocating a VRAM scratch buffer due to low VRAM option\n", __func__);
             ggml_cuda_set_scratch_size(0); // disable scratch
         } else {
             const size_t vram_scratch_base = VRAM_REQ_SCRATCH_BASE().at(model.type);
@@ -1286,7 +1301,7 @@ static void llama_model_load_internal(
             vram_scratch = n_batch * (vram_scratch_base + n_ctx * vram_scratch_per_context);
             ggml_cuda_set_scratch_size(vram_scratch);
             if (n_gpu_layers > 0) {
-                fprintf(stderr, "%s: allocating batch_size x (%zd kB + n_ctx x %zd B) = %zd MB VRAM for the scratch buffer\n",
+                LLAMA_LOG_INFO("%s: allocating batch_size x (%zd kB + n_ctx x %zd B) = %zd MB VRAM for the scratch buffer\n",
                         __func__, vram_scratch_base / kB, vram_scratch_per_context,
                         (vram_scratch + MB - 1) / MB); // round up
             }
@@ -1296,9 +1311,9 @@ static void llama_model_load_internal(
 #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
         const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer));
 
-        fprintf(stderr, "%s: offloading %d repeating layers to GPU\n", __func__, n_gpu);
+        LLAMA_LOG_INFO("%s: offloading %d repeating layers to GPU\n", __func__, n_gpu);
         if (n_gpu_layers > (int) hparams.n_layer) {
-            fprintf(stderr, "%s: offloading non-repeating layers to GPU\n", __func__);
+            LLAMA_LOG_INFO("%s: offloading non-repeating layers to GPU\n", __func__);
         }
         size_t vram_kv_cache = 0;
 
@@ -1307,17 +1322,17 @@ static void llama_model_load_internal(
         const int max_offloadable_layers = low_vram ? hparams.n_layer + 1 : hparams.n_layer + 3;
         if (n_gpu_layers > (int) hparams.n_layer + 1) {
             if (low_vram) {
-                fprintf(stderr, "%s: cannot offload v cache to GPU due to low VRAM option\n", __func__);
+                LLAMA_LOG_INFO("%s: cannot offload v cache to GPU due to low VRAM option\n", __func__);
             } else {
-                fprintf(stderr, "%s: offloading v cache to GPU\n", __func__);
+                LLAMA_LOG_INFO("%s: offloading v cache to GPU\n", __func__);
                 vram_kv_cache += hparams.kv_size() / 2;
             }
         }
         if (n_gpu_layers > (int) hparams.n_layer + 2) {
             if (low_vram) {
-                fprintf(stderr, "%s: cannot offload k cache to GPU due to low VRAM option\n", __func__);
+                LLAMA_LOG_WARN("%s: cannot offload k cache to GPU due to low VRAM option\n", __func__);
             } else {
-                fprintf(stderr, "%s: offloading k cache to GPU\n", __func__);
+                LLAMA_LOG_INFO("%s: offloading k cache to GPU\n", __func__);
                 vram_kv_cache += hparams.kv_size() / 2;
             }
         }
@@ -1326,9 +1341,9 @@ static void llama_model_load_internal(
         const int max_offloadable_layers = hparams.n_layer + 1;
 #endif // GGML_USE_CUBLAS
 
-        fprintf(stderr, "%s: offloaded %d/%d layers to GPU\n",
+        LLAMA_LOG_INFO("%s: offloaded %d/%d layers to GPU\n",
                 __func__, std::min(n_gpu_layers, max_offloadable_layers), max_backend_supported_layers);
-        fprintf(stderr, "%s: total VRAM used: %zu MB\n",
+        LLAMA_LOG_INFO("%s: total VRAM used: %zu MB\n",
                 __func__, (vram_weights + vram_scratch + vram_kv_cache + MB - 1) / MB); // round up
 #else
         (void) n_gpu_layers;
@@ -1387,7 +1402,7 @@ static bool llama_model_load(
                                   use_mmap, use_mlock, vocab_only, progress_callback, progress_callback_user_data);
         return true;
     } catch (const std::exception & err) {
-        fprintf(stderr, "error loading model: %s\n", err.what());
+        LLAMA_LOG_ERROR("error loading model: %s\n", err.what());
         return false;
     }
 }
@@ -1751,7 +1766,7 @@ static struct ggml_cgraph * llama_build_graph(
     }
 
 #if 0
-    printf("\n%s: used_mem: eval ctx %.3f MB, scratch %.3f MB %.3f MB, work buf %.3f MB, n_past = %d, N = %d\n", __func__,
+    LLAMA_LOG_INFO("\n%s: used_mem: eval ctx %.3f MB, scratch %.3f MB %.3f MB, work buf %.3f MB, n_past = %d, N = %d\n", __func__,
             ggml_used_mem(ctx0)/1024.0/1024.0,
             lctx.get_buf_max_mem(0)/1024.0/1024.0,
             lctx.get_buf_max_mem(1)/1024.0/1024.0,
@@ -1812,7 +1827,7 @@ static bool llama_eval_internal(
     ggml_allocr_alloc_graph(lctx.alloc, gf);
 #endif
 
-    // fprintf(stderr, "graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);
+    // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);
 
     // for big prompts, if BLAS is enabled, it is better to use only one thread
     // otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance
@@ -1830,11 +1845,7 @@ static bool llama_eval_internal(
 #endif
 
 #ifdef GGML_USE_METAL
-    if (lctx.ctx_metal && N == 1) {
-        // TODO: disabled until #2413 is resolved
-        //if (!ggml_metal_if_optimized(lctx.ctx_metal)) {
-        //    ggml_metal_graph_find_concurrency(lctx.ctx_metal, gf);
-        //}
+    if (lctx.ctx_metal) {
         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);
@@ -1842,22 +1853,6 @@ 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
@@ -1999,7 +1994,7 @@ struct llama_tokenizer {
             left_sym.n += right_sym.n;
             right_sym.n = 0;
 
-            //printf("left = '%*s' size = %zu\n", (int) left_sym.n, left_sym.text, bigram.size);
+            //LLAMA_LOG_INFO("left = '%*s' size = %zu\n", (int) left_sym.n, left_sym.text, bigram.size);
 
             // remove the right sym from the chain
             left_sym.next = right_sym.next;
@@ -3007,7 +3002,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         tensor.data = read_data.addr;
         model_loader->load_data_for(tensor);
 
-        printf("[%4zu/%4zu] %36s - %16s, type = %6s, ",
+        LLAMA_LOG_INFO("[%4zu/%4zu] %36s - %16s, type = %6s, ",
                ++idx, model_loader->tensors_map.tensors.size(),
                tensor.name.c_str(), llama_format_tensor_shape(tensor.ne).c_str(),
                ggml_type_name(tensor.type));
@@ -3029,7 +3024,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
             new_type = tensor.type;
             new_data = tensor.data;
             new_size = tensor.size;
-            printf("size = %8.3f MB\n", tensor.size/1024.0/1024.0);
+            LLAMA_LOG_INFO("size = %8.3f MB\n", tensor.size/1024.0/1024.0);
         } else {
             new_type = quantized_type;
 #ifdef GGML_USE_K_QUANTS
@@ -3064,17 +3059,17 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                 int nx = tensor.ne.at(0);
                 int ny = tensor.ne.at(1);
                 if (nx % QK_K != 0 || ny % QK_K != 0) {
-                    fprintf(stderr, "\n\nTensor sizes %d x %d are not divisible by %d, required for k-quants.\n",nx,ny,QK_K);
+                    LLAMA_LOG_INFO("\n\nTensor sizes %d x %d are not divisible by %d, required for k-quants.\n",nx,ny,QK_K);
                     convert_incompatible_tensor = true;
                 }
             }
             if (convert_incompatible_tensor) {
                 if (tensor.name == "output.weight") {
                     new_type = GGML_TYPE_F16; //fall back to F16 instead of just failing.
-                    fprintf(stderr, "F16 will be used for this tensor instead.\n");
+                    LLAMA_LOG_WARN("F16 will be used for this tensor instead.\n");
                 } else if (tensor.name == "tok_embeddings.weight") {
                     new_type = GGML_TYPE_Q4_0; //fall back to Q4_0 instead of just failing.
-                    fprintf(stderr, "Q4_0 will be used for this tensor instead.\n");
+                    LLAMA_LOG_WARN("Q4_0 will be used for this tensor instead.\n");
                 } else {
                     throw std::runtime_error("Unsupported tensor size encountered\n");
                 }
@@ -3094,7 +3089,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                 f32_data = (float *) f32_conv_buf.addr;
             }
 
-            printf("quantizing to %s .. ", ggml_type_name(new_type));
+            LLAMA_LOG_INFO("quantizing to %s .. ", ggml_type_name(new_type));
             fflush(stdout);
 
             work.resize(nelements * 4); // upper bound on size
@@ -3144,7 +3139,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                 }
             }
 
-            printf("size = %8.2f MB -> %8.2f MB | hist: ", tensor.size/1024.0/1024.0, new_size/1024.0/1024.0);
+            LLAMA_LOG_INFO("size = %8.2f MB -> %8.2f MB | hist: ", tensor.size/1024.0/1024.0, new_size/1024.0/1024.0);
             int64_t tot_count = 0;
             for (size_t i = 0; i < hist_cur.size(); i++) {
                 hist_all[i] += hist_cur[i];
@@ -3153,18 +3148,18 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
 
             if (tot_count > 0) {
                 for (size_t i = 0; i < hist_cur.size(); i++) {
-                    printf("%5.3f ", hist_cur[i] / float(nelements));
+                    LLAMA_LOG_INFO("%5.3f ", hist_cur[i] / float(nelements));
                 }
             }
-            printf("\n");
+            LLAMA_LOG_INFO("\n");
         }
         total_size_org += tensor.size;
         total_size_new += new_size;
         file_saver.write_tensor(tensor, new_type, new_data, new_size);
     }
 
-    printf("%s: model size  = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
-    printf("%s: quant size  = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0);
+    LLAMA_LOG_INFO("%s: model size  = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
+    LLAMA_LOG_INFO("%s: quant size  = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0);
 
     {
         int64_t sum_all = 0;
@@ -3173,11 +3168,11 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         }
 
         if (sum_all > 0) {
-            printf("%s: hist: ", __func__);
+            LLAMA_LOG_INFO("%s: hist: ", __func__);
             for (size_t i = 0; i < hist_all.size(); i++) {
-                printf("%5.3f ", hist_all[i] / float(sum_all));
+                LLAMA_LOG_INFO("%5.3f ", hist_all[i] / float(sum_all));
             }
-            printf("\n");
+            LLAMA_LOG_INFO("\n");
         }
     }
 }
@@ -3201,8 +3196,8 @@ struct llama_model * llama_load_model_from_file(
                 params.main_gpu, params.tensor_split, params.mul_mat_q, params.rope_freq_base, params.rope_freq_scale,params.low_vram,
                 memory_type, params.use_mmap, params.use_mlock, params.vocab_only, params.progress_callback,
                 params.progress_callback_user_data)) {
+        LLAMA_LOG_ERROR("%s: failed to load model\n", __func__);
         delete model;
-        fprintf(stderr, "%s: failed to load model\n", __func__);
         return nullptr;
     }
 
@@ -3235,10 +3230,9 @@ struct llama_context * llama_new_context_with_model(
             unsigned percentage = (unsigned) (100 * progress);
             while (percentage > *cur_percentage_p) {
                 *cur_percentage_p = percentage;
-                fprintf(stderr, ".");
-                fflush(stderr);
+                LLAMA_LOG_INFO(".");
                 if (percentage >= 100) {
-                    fprintf(stderr, "\n");
+                    LLAMA_LOG_INFO("\n");
                 }
             }
         };
@@ -3252,14 +3246,14 @@ struct llama_context * llama_new_context_with_model(
     // reserve memory for context buffers
     if (!params.vocab_only) {
         if (!kv_cache_init(ctx->model.hparams, ctx->kv_self, memory_type, ctx->model.hparams.n_ctx, params.n_gpu_layers)) {
-            fprintf(stderr, "%s: kv_cache_init() failed for self-attention cache\n", __func__);
+            LLAMA_LOG_ERROR("%s: kv_cache_init() failed for self-attention cache\n", __func__);
             llama_free(ctx);
             return nullptr;
         }
 
         {
             const size_t memory_size = ggml_nbytes(ctx->kv_self.k) + ggml_nbytes(ctx->kv_self.v);
-            fprintf(stderr, "%s: kv self size  = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0);
+            LLAMA_LOG_INFO("%s: kv self size  = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0);
         }
 
         const auto & hparams = ctx->model.hparams;
@@ -3289,24 +3283,40 @@ 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;
 
-            fprintf(stderr, "%s: compute buffer total size = %7.2f MB\n", __func__, (ctx->buf_compute.size + alloc_size) / 1024.0 / 1024.0);
+            LLAMA_LOG_INFO("%s: compute buffer total size = %7.2f MB\n", __func__, (ctx->buf_compute.size + alloc_size) / 1024.0 / 1024.0);
 
             // debug - for comparison with scratch buffer
             //size_t prev_req =
             //    MEM_REQ_SCRATCH0(hparams.n_ctx).at(ctx->model.type) +
             //    MEM_REQ_SCRATCH1().at(ctx->model.type) +
             //    MEM_REQ_EVAL().at(ctx->model.type);
-            //fprintf(stderr, "%s: (debug) equivalent with scratch buffer = %7.2f MB\n", __func__, prev_req / 1024.0 / 1024.0);
+            //LLAMA_LOG_INFO("%s: (debug) equivalent with scratch buffer = %7.2f MB\n", __func__, prev_req / 1024.0 / 1024.0);
 
             // recreate allocator with exact memory requirements
             ggml_allocr_free(ctx->alloc);
 
             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());
@@ -3321,7 +3331,6 @@ 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;
@@ -3336,13 +3345,13 @@ struct llama_context * llama_new_context_with_model(
 
         const size_t max_size = ggml_get_max_tensor_size(ctx->model.ctx);
 
-        fprintf(stderr, "%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0);
+        LLAMA_LOG_INFO("%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0);
 
-#define LLAMA_METAL_CHECK_BUF(result)                                          \
-    if (!(result)) {                                                           \
-        fprintf(stderr, "%s: failed to add buffer\n", __func__);               \
-        llama_free(ctx);                                                       \
-        return NULL;                                                           \
+#define LLAMA_METAL_CHECK_BUF(result)                            \
+    if (!(result)) {                                             \
+        LLAMA_LOG_ERROR("%s: failed to add buffer\n", __func__); \
+        llama_free(ctx);                                         \
+        return NULL;                                             \
     }
 
         LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "data", data_ptr, data_size, max_size));
@@ -3350,8 +3359,7 @@ 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, "scr1", ctx->buf_scratch[1].addr, ctx->buf_scratch[1].size, 0));
+        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "alloc", ctx->buf_alloc.addr, ctx->buf_alloc.size, 0));
 #undef LLAMA_METAL_CHECK_BUF
     }
 #endif
@@ -3396,19 +3404,19 @@ int llama_model_quantize(
         llama_model_quantize_internal(fname_inp, fname_out, params);
         return 0;
     } catch (const std::exception & err) {
-        fprintf(stderr, "%s: failed to quantize: %s\n", __func__, err.what());
+        LLAMA_LOG_ERROR("%s: failed to quantize: %s\n", __func__, err.what());
         return 1;
     }
 }
 
 int llama_apply_lora_from_file_internal(const struct llama_model & model, const char * path_lora, const char * path_base_model, int n_threads) {
-    fprintf(stderr, "%s: applying lora adapter from '%s' - please wait ...\n", __func__, path_lora);
+    LLAMA_LOG_INFO("%s: applying lora adapter from '%s' - please wait ...\n", __func__, path_lora);
 
     const int64_t t_start_lora_us = ggml_time_us();
 
     auto fin = std::ifstream(path_lora, std::ios::binary);
     if (!fin) {
-        fprintf(stderr, "%s: failed to open '%s'\n", __func__, path_lora);
+        LLAMA_LOG_ERROR("%s: failed to open '%s'\n", __func__, path_lora);
         return 1;
     }
 
@@ -3417,14 +3425,14 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
         uint32_t magic;
         fin.read((char *) &magic, sizeof(magic));
         if (magic != LLAMA_FILE_MAGIC_GGLA) {
-            fprintf(stderr, "%s: bad file magic\n", __func__);
+            LLAMA_LOG_ERROR("%s: bad file magic\n", __func__);
             return 1;
         }
         uint32_t format_version;
         fin.read((char *) &format_version, sizeof(format_version));
 
         if (format_version != 1) {
-            fprintf(stderr, "%s: unsupported file version\n", __func__ );
+            LLAMA_LOG_ERROR("%s: unsupported file version\n", __func__ );
             return 1;
         }
     }
@@ -3435,7 +3443,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
     fin.read((char *) &lora_alpha, sizeof(lora_alpha));
     float scaling = (float)lora_alpha / (float)lora_r;
 
-    fprintf(stderr, "%s: r = %d, alpha = %d, scaling = %.2f\n", __func__, lora_r, lora_alpha, scaling);
+    LLAMA_LOG_INFO("%s: r = %d, alpha = %d, scaling = %.2f\n", __func__, lora_r, lora_alpha, scaling);
 
 
     // create a temporary ggml context to store the lora tensors
@@ -3461,7 +3469,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
     ggml_context * base_ctx = NULL;
     llama_buffer base_buf;
     if (path_base_model) {
-        fprintf(stderr, "%s: loading base model from '%s'\n", __func__, path_base_model);
+        LLAMA_LOG_INFO("%s: loading base model from '%s'\n", __func__, path_base_model);
         model_loader.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true));
 
         size_t ctx_size;
@@ -3518,17 +3526,17 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
         const std::string lora_suffix = ".lora";
         size_t pos = name.rfind(lora_suffix);
         if (pos == std::string::npos) {
-            fprintf(stderr, "%s: error: '%s' is not a lora tensor\n", __func__, name.c_str());
+            LLAMA_LOG_ERROR("%s: error: '%s' is not a lora tensor\n", __func__, name.c_str());
             return 1;
         }
 
         std::string lora_type = name.substr(pos + lora_suffix.length());
         std::string base_name = name;
         base_name.erase(pos);
-        // fprintf(stderr, "%s: %s => %s (lora type %s) ", __func__, name.c_str(),base_name.c_str(), lora_type.c_str());
+        // LLAMA_LOG_INFO("%s: %s => %s (lora type %s) \n", __func__, name.c_str(),base_name.c_str(), lora_type.c_str());
 
         if (model_tensors.find(base_name) == model_tensors.end()) {
-            fprintf(stderr, "%s: unknown tensor '%s' in lora adapter\n", __func__, name.data());
+            LLAMA_LOG_ERROR("%s: unknown tensor '%s' in lora adapter\n", __func__, name.data());
             return 1;
         }
 
@@ -3539,7 +3547,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
             case 1: wtype = GGML_TYPE_F16;  break;
             default:
                     {
-                        fprintf(stderr, "%s: invalid tensor data type '%d'\n",
+                        LLAMA_LOG_ERROR("%s: invalid tensor data type '%d'\n",
                                 __func__, ftype);
                         return false;
                     }
@@ -3549,7 +3557,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
             lora_tensor = ggml_new_tensor_2d(lora_ctx, wtype, ne[0], ne[1]);
         }
         else {
-            fprintf(stderr, "%s: unsupported tensor dimension %d\n", __func__, n_dims);
+            LLAMA_LOG_ERROR("%s: unsupported tensor dimension %d\n", __func__, n_dims);
             return 1;
         }
         ggml_set_name(lora_tensor, "lora_tensor");
@@ -3587,7 +3595,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
             if (model_loader) {
                 // load from base model
                 if (model_loader->tensors_map.name_to_idx.find(base_name) == model_loader->tensors_map.name_to_idx.end()) {
-                    fprintf(stderr, "%s: error: tensor '%s' not found in base model\n", __func__, base_name.c_str());
+                    LLAMA_LOG_ERROR("%s: error: tensor '%s' not found in base model\n", __func__, base_name.c_str());
                     return 1;
                 }
                 size_t idx = model_loader->tensors_map.name_to_idx[base_name];
@@ -3603,8 +3611,8 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
 
             if (ggml_is_quantized(base_t->type)) {
                 if (!warned) {
-                    fprintf(stderr, "%s: warning: using a lora adapter with a quantized model may result in poor quality, "
-                                    "use a f16 or f32 base model with --lora-base\n", __func__);
+                    LLAMA_LOG_WARN("%s: warning: using a lora adapter with a quantized model may result in poor quality, "
+                                   "use a f16 or f32 base model with --lora-base\n", __func__);
                     warned = true;
                 }
             }
@@ -3618,8 +3626,8 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
             ggml_set_name(loraB, "loraB");
 
             if (base_t->ne[0] != loraA->ne[1] || base_t->ne[1] != loraB->ne[1]) {
-                fprintf(stderr, "%s: incompatible tensor dimensions (%" PRId64 " and %" PRId64 ");"
-                               " are you sure that this adapter is for this model?\n", __func__, base_t->ne[0], loraA->ne[1]);
+                LLAMA_LOG_ERROR("%s: incompatible tensor dimensions (%" PRId64 " and %" PRId64 ");"
+                                " are you sure that this adapter is for this model?\n", __func__, base_t->ne[0], loraA->ne[1]);
                 return 1;
             }
 
@@ -3664,7 +3672,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
 
             n_tensors++;
             if (n_tensors % 4 == 0) {
-                fprintf(stderr, ".");
+                LLAMA_LOG_INFO(".");
             }
         }
     }
@@ -3676,7 +3684,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
     }
 
     const int64_t t_lora_us = ggml_time_us() - t_start_lora_us;
-    fprintf(stderr, " done (%.2f ms)\n", t_lora_us / 1000.0);
+    LLAMA_LOG_INFO(" done (%.2f ms)\n", t_lora_us / 1000.0);
 
     return 0;
 }
@@ -3685,7 +3693,7 @@ int llama_apply_lora_from_file(struct llama_context * ctx, const char * path_lor
     try {
         return llama_apply_lora_from_file_internal(ctx->model, path_lora, path_base_model, n_threads);
     } catch (const std::exception & err) {
-        fprintf(stderr, "%s: failed to apply lora adapter: %s\n", __func__, err.what());
+        LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what());
         return 1;
     }
 }
@@ -3694,7 +3702,7 @@ int llama_model_apply_lora_from_file(const struct llama_model * model, const cha
     try {
         return llama_apply_lora_from_file_internal(*model, path_lora, path_base_model, n_threads);
     } catch (const std::exception & err) {
-        fprintf(stderr, "%s: failed to apply lora adapter: %s\n", __func__, err.what());
+        LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what());
         return 1;
     }
 }
@@ -3976,7 +3984,7 @@ static bool llama_load_session_file_internal(struct llama_context * ctx, const c
         const uint32_t version = file.read_u32();
 
         if (magic != LLAMA_SESSION_MAGIC || version != LLAMA_SESSION_VERSION) {
-            fprintf(stderr, "%s : unknown (magic, version) for session file: %08x, %08x\n", __func__, magic, version);
+            LLAMA_LOG_ERROR("%s : unknown (magic, version) for session file: %08x, %08x\n", __func__, magic, version);
             return false;
         }
 
@@ -3984,7 +3992,7 @@ static bool llama_load_session_file_internal(struct llama_context * ctx, const c
         file.read_raw(&session_hparams, sizeof(llama_hparams));
 
         if (session_hparams != ctx->model.hparams) {
-            fprintf(stderr, "%s : model hparams didn't match from session file!\n", __func__);
+            LLAMA_LOG_INFO("%s : model hparams didn't match from session file!\n", __func__);
             return false;
         }
     }
@@ -3994,7 +4002,7 @@ static bool llama_load_session_file_internal(struct llama_context * ctx, const c
         const uint32_t n_token_count = file.read_u32();
 
         if (n_token_count > n_token_capacity) {
-            fprintf(stderr, "%s : token count in session file exceeded capacity! %u > %zu\n", __func__, n_token_count, n_token_capacity);
+            LLAMA_LOG_ERROR("%s : token count in session file exceeded capacity! %u > %zu\n", __func__, n_token_count, n_token_capacity);
             return false;
         }
 
@@ -4008,7 +4016,7 @@ static bool llama_load_session_file_internal(struct llama_context * ctx, const c
         const size_t n_state_size_max = llama_get_state_size(ctx);
 
         if (n_state_size_cur > n_state_size_max) {
-            fprintf(stderr, "%s : the state size in session file is too big! max %zu, got %zu\n", __func__, n_state_size_max, n_state_size_cur);
+            LLAMA_LOG_ERROR("%s : the state size in session file is too big! max %zu, got %zu\n", __func__, n_state_size_max, n_state_size_cur);
             return false;
         }
 
@@ -4025,7 +4033,7 @@ bool llama_load_session_file(struct llama_context * ctx, const char * path_sessi
     try {
         return llama_load_session_file_internal(ctx, path_session, tokens_out, n_token_capacity, n_token_count_out);
     } catch (const std::exception & err) {
-        fprintf(stderr, "error loading session file: %s\n", err.what());
+        LLAMA_LOG_ERROR("error loading session file: %s\n", err.what());
         return false;
     }
 }
@@ -4056,7 +4064,7 @@ int llama_eval(
                          int   n_past,
                          int   n_threads) {
     if (!llama_eval_internal(*ctx, tokens, nullptr, n_tokens, n_past, n_threads, nullptr)) {
-        fprintf(stderr, "%s: failed to eval\n", __func__);
+        LLAMA_LOG_ERROR("%s: failed to eval\n", __func__);
         return 1;
     }
 
@@ -4078,7 +4086,7 @@ int llama_eval_embd(
                              int   n_past,
                              int   n_threads) {
     if (!llama_eval_internal(*ctx, nullptr, embd, n_tokens, n_past, n_threads, nullptr)) {
-        fprintf(stderr, "%s: failed to eval\n", __func__);
+        LLAMA_LOG_ERROR("%s: failed to eval\n", __func__);
         return 1;
     }
 
@@ -4099,7 +4107,7 @@ int llama_eval_export(struct llama_context * ctx, const char * fname) {
     const std::vector<llama_token> tmp(n_batch, llama_token_bos());
 
     if (!llama_eval_internal(*ctx, tmp.data(), nullptr, tmp.size(), n_ctx, 1, fname)) {
-        fprintf(stderr, "%s: failed to eval\n", __func__);
+        LLAMA_LOG_ERROR("%s: failed to eval\n", __func__);
         return 1;
     }
 
@@ -4115,7 +4123,7 @@ int llama_tokenize_with_model(
     auto res = llama_tokenize(model->vocab, text, add_bos);
 
     if (n_max_tokens < (int) res.size()) {
-        fprintf(stderr, "%s: too many tokens\n", __func__);
+        LLAMA_LOG_ERROR("%s: too many tokens\n", __func__);
         return -((int) res.size());
     }
 
@@ -4232,15 +4240,15 @@ struct llama_timings llama_get_timings(struct llama_context * ctx) {
 void llama_print_timings(struct llama_context * ctx) {
     const llama_timings timings = llama_get_timings(ctx);
 
-    fprintf(stderr, "\n");
-    fprintf(stderr, "%s:        load time = %8.2f ms\n", __func__, timings.t_load_ms);
-    fprintf(stderr, "%s:      sample time = %8.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
+    LLAMA_LOG_INFO("\n");
+    LLAMA_LOG_INFO("%s:        load time = %8.2f ms\n", __func__, timings.t_load_ms);
+    LLAMA_LOG_INFO("%s:      sample time = %8.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
             __func__, timings.t_sample_ms, timings.n_sample, timings.t_sample_ms / timings.n_sample, 1e3 / timings.t_sample_ms * timings.n_sample);
-    fprintf(stderr, "%s: prompt eval time = %8.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
+    LLAMA_LOG_INFO("%s: prompt eval time = %8.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
             __func__, timings.t_p_eval_ms, timings.n_p_eval, timings.t_p_eval_ms / timings.n_p_eval, 1e3 / timings.t_p_eval_ms * timings.n_p_eval);
-    fprintf(stderr, "%s:        eval time = %8.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
+    LLAMA_LOG_INFO("%s:        eval time = %8.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
             __func__, timings.t_eval_ms, timings.n_eval, timings.t_eval_ms / timings.n_eval, 1e3 / timings.t_eval_ms * timings.n_eval);
-    fprintf(stderr, "%s:       total time = %8.2f ms\n", __func__, (timings.t_end_ms - timings.t_start_ms));
+    LLAMA_LOG_INFO("%s:       total time = %8.2f ms\n", __func__, (timings.t_end_ms - timings.t_start_ms));
 }
 
 void llama_reset_timings(struct llama_context * ctx) {
@@ -4276,3 +4284,44 @@ const char * llama_print_system_info(void) {
 const std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx) {
     return ctx->model.tensors_by_name;
 }
+
+
+void llama_log_set(llama_log_callback log_callback, void * user_data) {
+    g_state.log_callback = log_callback ? log_callback : llama_log_callback_default;
+    g_state.log_callback_user_data = user_data;
+}
+
+#if defined(_MSC_VER) && !defined(vsnprintf)
+#define vsnprintf _vsnprintf
+#endif
+
+static void llama_log_internal_v(llama_log_level level, const char * format, va_list args) {
+    va_list args_copy;
+    va_copy(args_copy, args);
+    char buffer[128];
+    int len = vsnprintf(buffer, 128, format, args);
+    if (len < 128) {
+        g_state.log_callback(level, buffer, g_state.log_callback_user_data);
+    } else {
+        char* buffer2 = new char[len+1];
+        vsnprintf(buffer2, len+1, format, args_copy);
+        buffer2[len] = 0;
+        g_state.log_callback(level, buffer2, g_state.log_callback_user_data);
+        delete[] buffer2;
+    }
+    va_end(args_copy);
+}
+
+static void llama_log_internal(llama_log_level level, const char * format, ...) {
+    va_list args;
+    va_start(args, format);
+    llama_log_internal_v(level, format, args);
+    va_end(args);
+}
+
+static void llama_log_callback_default(llama_log_level level, const char * text, void * user_data) {
+    (void) level;
+    (void) user_data;
+    fputs(text, stderr);
+    fflush(stderr);
+}

llama.h

@@ -86,7 +86,20 @@ extern "C" {
 
     typedef void (*llama_progress_callback)(float progress, void *ctx);
 
-   struct llama_context_params {
+    enum llama_log_level {
+        LLAMA_LOG_LEVEL_ERROR = 2,
+        LLAMA_LOG_LEVEL_WARN  = 3,
+        LLAMA_LOG_LEVEL_INFO  = 4
+    };
+
+    // Signature for logging events
+    // Note that text includes the new line character at the end for most events.
+    // 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)(enum llama_log_level level, const char * text, void * user_data);
+
+    struct llama_context_params {
         uint32_t seed;         // RNG seed, -1 for random
         int32_t  n_ctx;        // text context
         int32_t  n_batch;      // prompt processing batch size
@@ -195,6 +208,10 @@ extern "C" {
         int32_t n_eval;
     };
 
+    // Set callback for all future logging events.
+    // If this is not called, or NULL is supplied, everything is output on stderr.
+    LLAMA_API void llama_log_set(llama_log_callback log_callback, void * user_data);
+
     LLAMA_API int llama_max_devices();
 
     LLAMA_API struct llama_context_params llama_context_default_params();

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;
+}