Allow cmake to build ggml as a library (#1896)

* Allow cmake to build ggml as a library

* A ggml_static library will be created

* When BUILD_SHARED_LIBS is enabled, ggml_shared will also be built

.gitignore

@@ -22,6 +22,7 @@ build-metal/
 build-no-accel/
 build-sanitize-addr/
 build-sanitize-thread/
+out/
 
 models/*
 *.bin
@@ -32,6 +33,7 @@ models/*
 /result
 /perplexity
 /embedding
+/train-text-from-scratch
 /benchmark-matmult
 /vdot
 /Pipfile
@@ -40,6 +42,7 @@ models/*
 build-info.h
 arm_neon.h
 compile_commands.json
+CMakeSettings.json
 
 __pycache__
 

CMakeLists.txt

@@ -70,6 +70,7 @@ set(LLAMA_BLAS_VENDOR "Generic" CACHE STRING "llama: BLAS library vendor")
 option(LLAMA_CUBLAS                          "llama: use cuBLAS"                                OFF)
 set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kernels")
 set(LLAMA_CUDA_DMMV_Y       "1" CACHE STRING "llama: y block size for dmmv CUDA kernels")
+set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K")
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
 option(LLAMA_METAL                           "llama: use Metal"                                 OFF)
 option(LLAMA_K_QUANTS                        "llama: use k-quants"                              ON)
@@ -158,17 +159,64 @@ if (LLAMA_BLAS)
     if ($(CMAKE_VERSION) VERSION_GREATER_EQUAL 3.22)
         set(BLA_SIZEOF_INTEGER 8)
     endif()
+
     set(BLA_VENDOR ${LLAMA_BLAS_VENDOR})
     find_package(BLAS)
+
     if (BLAS_FOUND)
         message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}")
 
+        if ("${BLAS_INCLUDE_DIRS}" STREQUAL "")
+            # BLAS_INCLUDE_DIRS is missing in FindBLAS.cmake.
+            # see https://gitlab.kitware.com/cmake/cmake/-/issues/20268
+            find_package(PkgConfig REQUIRED)
+            if (${LLAMA_BLAS_VENDOR} MATCHES "Generic")
+                pkg_check_modules(DepBLAS REQUIRED blas)
+            elseif (${LLAMA_BLAS_VENDOR} MATCHES "OpenBLAS")
+                pkg_check_modules(DepBLAS REQUIRED openblas)
+            elseif (${LLAMA_BLAS_VENDOR} MATCHES "FLAME")
+                pkg_check_modules(DepBLAS REQUIRED blis)
+            elseif (${LLAMA_BLAS_VENDOR} MATCHES "ATLAS")
+                pkg_check_modules(DepBLAS REQUIRED blas-atlas)
+            elseif (${LLAMA_BLAS_VENDOR} MATCHES "FlexiBLAS")
+                pkg_check_modules(DepBLAS REQUIRED flexiblas_api)
+            elseif (${LLAMA_BLAS_VENDOR} MATCHES "Intel")
+                # all Intel* libraries share the same include path
+                pkg_check_modules(DepBLAS REQUIRED mkl-sdl)
+            elseif (${LLAMA_BLAS_VENDOR} MATCHES "NVHPC")
+                # this doesn't provide pkg-config
+                # suggest to assign BLAS_INCLUDE_DIRS on your own
+                if ("${NVHPC_VERSION}" STREQUAL "")
+                    message(WARNING "Better to set NVHPC_VERSION")
+                else()
+                    set(DepBLAS_FOUND ON)
+                    set(DepBLAS_INCLUDE_DIRS "/opt/nvidia/hpc_sdk/${CMAKE_SYSTEM_NAME}_${CMAKE_SYSTEM_PROCESSOR}/${NVHPC_VERSION}/math_libs/include")
+                endif()
+            endif()
+            if (DepBLAS_FOUND)
+                set(BLAS_INCLUDE_DIRS ${DepBLAS_INCLUDE_DIRS})
+            else()
+                message(WARNING "BLAS_INCLUDE_DIRS neither been provided nor been automatically"
+                " detected by pkgconfig, trying to find cblas.h from possible paths...")
+                find_path(BLAS_INCLUDE_DIRS
+                    NAMES cblas.h
+                    HINTS
+                        /usr/include
+                        /usr/local/include
+                        /usr/include/openblas
+                        /opt/homebrew/opt/openblas/include
+                        /usr/local/opt/openblas/include
+                        /usr/include/x86_64-linux-gnu/openblas/include
+                )
+            endif()
+        endif()
+
+        message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
         add_compile_options(${BLAS_LINKER_FLAGS})
         add_compile_definitions(GGML_USE_OPENBLAS)
         set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ${BLAS_LIBRARIES})
+        set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${BLAS_INCLUDE_DIRS})
 
-        message("${BLAS_LIBRARIES} ${BLAS_INCLUDE_DIRS}")
-        include_directories(${BLAS_INCLUDE_DIRS})
     else()
         message(WARNING "BLAS not found, please refer to "
         "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
@@ -190,6 +238,7 @@ if (LLAMA_CUBLAS)
         add_compile_definitions(GGML_USE_CUBLAS)
         add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
         add_compile_definitions(GGML_CUDA_DMMV_Y=${LLAMA_CUDA_DMMV_Y})
+        add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
 
         if (LLAMA_STATIC)
             set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
@@ -408,12 +457,14 @@ add_library(ggml OBJECT
             ${GGML_SOURCES_EXTRA}
             )
 
-target_include_directories(ggml PUBLIC .)
+target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})
 target_compile_features(ggml PUBLIC c_std_11) # don't bump
 target_link_libraries(ggml PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
 
+add_library(ggml_static STATIC $<TARGET_OBJECTS:ggml>)
 if (BUILD_SHARED_LIBS)
     set_target_properties(ggml PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    add_library(ggml_shared SHARED $<TARGET_OBJECTS:ggml>)
 endif()
 
 add_library(llama

Makefile

@@ -1,5 +1,5 @@
 # Define the default target now so that it is always the first target
-BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot
+BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch simple
 
 ifdef LLAMA_BUILD_SERVER
 	BUILD_TARGETS += server
@@ -171,6 +171,11 @@ ifdef LLAMA_CUDA_DMMV_Y
 else
 	NVCCFLAGS += -DGGML_CUDA_DMMV_Y=1
 endif # LLAMA_CUDA_DMMV_Y
+ifdef LLAMA_CUDA_KQUANTS_ITER
+	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
+else
+	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
+endif
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@
 endif # LLAMA_CUBLAS
@@ -259,7 +264,7 @@ libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
 clean:
-	rm -vf *.o main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state server vdot build-info.h
+	rm -vf *.o *.so main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state server vdot train-text-from-scratch build-info.h
 
 #
 # Examples
@@ -271,6 +276,12 @@ main: examples/main/main.cpp                                  build-info.h ggml.
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
+simple: examples/simple/simple.cpp                            build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	@echo
+	@echo '====  Run ./simple -h for help.  ===='
+	@echo
+
 quantize: examples/quantize/quantize.cpp                      build-info.h ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
@@ -289,6 +300,9 @@ save-load-state: examples/save-load-state/save-load-state.cpp build-info.h ggml.
 server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS)
 
+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)
+
 build-info.h: $(wildcard .git/index) scripts/build-info.sh
 	@sh scripts/build-info.sh > $@.tmp
 	@if ! cmp -s $@.tmp $@; then \

Package.swift

@@ -11,6 +11,7 @@ let package = Package(
         .target(
             name: "llama",
             path: ".",
+            exclude: ["ggml-metal.metal"],
             sources: ["ggml.c", "llama.cpp"],
             publicHeadersPath: "spm-headers",
             cSettings: [.unsafeFlags(["-Wno-shorten-64-to-32"]), .define("GGML_USE_ACCELERATE")],

examples/CMakeLists.txt

@@ -37,6 +37,7 @@ else()
     add_subdirectory(save-load-state)
     add_subdirectory(benchmark)
     add_subdirectory(baby-llama)
+    add_subdirectory(train-text-from-scratch)
     if (LLAMA_METAL)
         add_subdirectory(metal)
     endif()

examples/baby-llama/baby-llama.cpp

@@ -4,6 +4,10 @@
 #include <random>
 #include <cstring>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 float frand() {
     return (float)rand()/(float)RAND_MAX;
 }
@@ -79,34 +83,39 @@ struct ggml_tensor * randomize_tensor_normal(
         int ndims,
         const int64_t ne[],
         struct random_normal_distribution * rnd) {
+    float scale = 1.0; // xavier
     switch (ndims) {
         case 1:
+            scale /= sqrtf(ne[0]);
             for (int i0 = 0; i0 < ne[0]; i0++) {
-                ((float *)tensor->data)[i0] = frand_normal(rnd);
+                ((float *)tensor->data)[i0] = scale * frand_normal(rnd);
             }
             break;
         case 2:
+            scale /= sqrtf(ne[0]+ne[1]);
             for (int i1 = 0; i1 < ne[1]; i1++) {
                 for (int i0 = 0; i0 < ne[0]; i0++) {
-                    ((float *)tensor->data)[i1*ne[0] + i0] = frand_normal(rnd);
+                    ((float *)tensor->data)[i1*ne[0] + i0] = scale * frand_normal(rnd);
                 }
             }
             break;
         case 3:
+            scale /= sqrtf(ne[0]+ne[1]);
             for (int i2 = 0; i2 < ne[2]; i2++) {
                 for (int i1 = 0; i1 < ne[1]; i1++) {
                     for (int i0 = 0; i0 < ne[0]; i0++) {
-                        ((float *)tensor->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand_normal(rnd);
+                        ((float *)tensor->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = scale * frand_normal(rnd);
                     }
                 }
             }
             break;
         case 4:
+            scale /= sqrtf(ne[0]+ne[1]);
             for (int i3 = 0; i3 < ne[3]; i3++) {
                 for (int i2 = 0; i2 < ne[2]; i2++) {
                     for (int i1 = 0; i1 < ne[1]; i1++) {
                         for (int i0 = 0; i0 < ne[0]; i0++) {
-                            ((float *)tensor->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand_normal(rnd);
+                            ((float *)tensor->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = scale * frand_normal(rnd);
                         }
                     }
                 }
@@ -148,8 +157,8 @@ struct llama_hparams_lora {
     uint32_t n_rot   = 64;
     uint32_t n_lora  = 64;
 
-    bool operator!=(const llama_hparams & other) const {
-        return memcmp(this, &other, sizeof(llama_hparams));
+    bool operator!=(const llama_hparams_lora & other) const {
+        return memcmp(this, &other, sizeof(llama_hparams_lora)) != 0;
     }
 };
 
@@ -1465,7 +1474,7 @@ struct ggml_tensor * square_error_loss(struct ggml_context * ctx, struct ggml_te
 }
 
 struct ggml_tensor * cross_entropy_loss(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) {
-    const float eps = 1e-3;
+    const float eps = 1e-3f;
     return
         ggml_sum(ctx,
             ggml_neg(ctx,

examples/benchmark/benchmark-matmult.cpp

@@ -16,6 +16,10 @@
 #include <iterator>
 #include <algorithm>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 float tensor_sum_elements(const ggml_tensor * tensor) {
     float sum = 0;
     if (tensor->type==GGML_TYPE_F32) {
@@ -29,9 +33,9 @@ float tensor_sum_elements(const ggml_tensor * tensor) {
 }
 
 void tensor_dump(const ggml_tensor * tensor, const char * name) {
-    printf("%15s: type = %i (%5s) ne = %5d x %5d x %5d, nb = (%5li, %5li, %5li) - ", name,
+    printf("%15s: type = %i (%5s) ne = %5" PRIi64 " x %5" PRIi64 " x %5" PRIi64 ", nb = (%5zi, %5zi, %5zi) - ", name,
         tensor->type, ggml_type_name(tensor->type),
-        (int) tensor->ne[0], (int) tensor->ne[1], (int) tensor->ne[2], tensor->nb[0], tensor->nb[1], tensor->nb[2]);
+        tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->nb[0], tensor->nb[1], tensor->nb[2]);
     float sum = tensor_sum_elements(tensor);
     printf("Sum of tensor %s is %6.2f\n", name, sum);
 }
@@ -120,7 +124,7 @@ int main(int argc, char ** argv)  {
     ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
     ctx_size += 1024*1024*16;
 
-    printf("Allocating Memory of size %li bytes, %li MB\n",ctx_size, (ctx_size/1024/1024));
+    printf("Allocating Memory of size %zi bytes, %zi MB\n",ctx_size, (ctx_size/1024/1024));
 
     struct ggml_init_params params = {
         /*.mem_size   =*/ ctx_size,

examples/chat-vicuna.sh

@@ -0,0 +1,41 @@
+#!/bin/bash
+
+set -e
+
+cd "$(dirname "$0")/.." || exit
+
+MODEL="${MODEL:-./models/ggml-vic13b-uncensored-q5_0.bin}"
+PROMPT_TEMPLATE=${PROMPT_TEMPLATE:-./prompts/chat.txt}
+USER_NAME="### Human"
+AI_NAME="### Assistant"
+
+# Adjust to the number of CPU cores you want to use.
+N_THREAD="${N_THREAD:-8}"
+# Number of tokens to predict (made it larger than default because we want a long interaction)
+N_PREDICTS="${N_PREDICTS:-2048}"
+
+# Note: you can also override the generation options by specifying them on the command line:
+# For example, override the context size by doing: ./chatLLaMa --ctx_size 1024
+GEN_OPTIONS="${GEN_OPTIONS:---ctx_size 2048 --temp 0.7 --top_k 40 --top_p 0.5 --repeat_last_n 256 --batch_size 1024 --repeat_penalty 1.17647}"
+
+DATE_TIME=$(date +%H:%M)
+DATE_YEAR=$(date +%Y)
+
+PROMPT_FILE=$(mktemp -t llamacpp_prompt.XXXXXXX.txt)
+
+sed -e "s/\[\[USER_NAME\]\]/$USER_NAME/g" \
+    -e "s/\[\[AI_NAME\]\]/$AI_NAME/g" \
+    -e "s/\[\[DATE_TIME\]\]/$DATE_TIME/g" \
+    -e "s/\[\[DATE_YEAR\]\]/$DATE_YEAR/g" \
+     $PROMPT_TEMPLATE > $PROMPT_FILE
+
+# shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS
+./bin/main $GEN_OPTIONS \
+  --model "$MODEL" \
+  --threads "$N_THREAD" \
+  --n_predict "$N_PREDICTS" \
+  --color --interactive \
+  --file ${PROMPT_FILE} \
+  --reverse-prompt "### Human:" \
+  --in-prefix ' ' \
+  "$@"

examples/common.cpp

@@ -28,6 +28,10 @@
 #include <wchar.h>
 #endif
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 int32_t get_num_physical_cores() {
 #ifdef __linux__
     // enumerate the set of thread siblings, num entries is num cores
@@ -331,6 +335,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             }
 #else
       fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
+#endif // GGML_USE_CUBLAS
+        } else if (arg == "--low-vram" || arg == "-lv") {
+#ifdef GGML_USE_CUBLAS
+            params.low_vram = true;
+#else
+      fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
 #endif // GGML_USE_CUBLAS
         } else if (arg == "--no-mmap") {
             params.use_mmap = false;
@@ -367,7 +377,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 } else {
                     throw std::exception();
                 }
-            } catch (const std::exception &e) {
+            } catch (const std::exception&) {
                 invalid_param = true;
                 break;
             }
@@ -406,6 +416,14 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
         gpt_print_usage(argc, argv, default_params);
         exit(1);
     }
+
+#ifdef GGML_USE_CUBLAS
+    if (!params.lora_adapter.empty() && params.n_gpu_layers > 0) {
+        fprintf(stderr, "%s: error: the simultaneous use of LoRAs and GPU acceleration is not supported", __func__);
+        exit(1);
+    }
+#endif // GGML_USE_CUBLAS
+
     if (escape_prompt) {
         process_escapes(params.prompt);
     }
@@ -479,6 +497,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  -ts SPLIT --tensor-split SPLIT\n");
     fprintf(stderr, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
     fprintf(stderr, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n" );
+    fprintf(stderr, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n" );
 #endif
     fprintf(stderr, "  --mtest               compute maximum memory usage\n");
     fprintf(stderr, "  --export              export the computation graph to 'llama.ggml'\n");
@@ -528,6 +547,7 @@ struct llama_context * llama_init_from_gpt_params(const gpt_params & params) {
     lparams.n_gpu_layers = params.n_gpu_layers;
     lparams.main_gpu     = params.main_gpu;
     memcpy(lparams.tensor_split, params.tensor_split, LLAMA_MAX_DEVICES*sizeof(float));
+    lparams.low_vram     = params.low_vram;
     lparams.seed         = params.seed;
     lparams.f16_kv       = params.memory_f16;
     lparams.use_mmap     = params.use_mmap;

examples/common.h

@@ -21,15 +21,16 @@
 int32_t get_num_physical_cores();
 
 struct gpt_params {
-    int32_t seed                           = -1;   // RNG seed
-    int32_t n_threads                      = get_num_physical_cores();
-    int32_t n_predict                      = -1;   // new tokens to predict
-    int32_t n_ctx                          = 512;  // context size
-    int32_t n_batch                        = 512;  // batch size for prompt processing (must be >=32 to use BLAS)
-    int32_t n_keep                         = 0;    // number of tokens to keep from initial prompt
-    int32_t n_gpu_layers                   = 0;    // number of layers to store in VRAM
-    int32_t main_gpu                       = 0;    // the GPU that is used for scratch and small tensors
+    int32_t seed                            = -1;  // RNG seed
+    int32_t n_threads                       = get_num_physical_cores();
+    int32_t n_predict                       = -1;  // new tokens to predict
+    int32_t n_ctx                           = 512; // context size
+    int32_t n_batch                         = 512; // batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_keep                          = 0;   // number of tokens to keep from initial prompt
+    int32_t n_gpu_layers                    = 0;   // number of layers to store in VRAM
+    int32_t main_gpu                        = 0;   // the GPU that is used for scratch and small tensors
     float   tensor_split[LLAMA_MAX_DEVICES] = {0}; // how split tensors should be distributed across GPUs
+    bool    low_vram                        = 0;   // if true, reduce VRAM usage at the cost of performance
 
     // sampling parameters
     std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens

examples/embedding/embedding.cpp

@@ -4,6 +4,10 @@
 
 #include <ctime>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 int main(int argc, char ** argv) {
     gpt_params params;
 

examples/main/README.md

@@ -288,5 +288,6 @@ These options provide extra functionality and customization when running the LLa
 -   `-ngl N, --n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
 -   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
 -   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
+-   `-lv, --low-vram`: Do not allocate a VRAM scratch buffer for holding temporary results. Reduces VRAM usage at the cost of performance, particularly prompt processing speed. Requires cuBLAS.
 -   `--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.

examples/main/main.cpp

@@ -23,11 +23,17 @@
 #include <unistd.h>
 #elif defined (_WIN32)
 #define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
 #define NOMINMAX
+#endif
 #include <windows.h>
 #include <signal.h>
 #endif
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 static console_state con_st;
 static llama_context ** g_ctx;
 
@@ -348,7 +354,7 @@ int main(int argc, char ** argv) {
             if ((int)embd.size() > max_embd_size) {
                 auto skipped_tokens = embd.size() - max_embd_size;
                 console_set_color(con_st, CONSOLE_COLOR_ERROR);
-                printf("<<input too long: skipped %ld token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
+                printf("<<input too long: skipped %" PRIu64 "  token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
                 console_set_color(con_st, CONSOLE_COLOR_DEFAULT);
                 fflush(stdout);
                 embd.resize(max_embd_size);

examples/perplexity/perplexity.cpp

@@ -5,6 +5,10 @@
 #include <cmath>
 #include <ctime>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 std::vector<float> softmax(const std::vector<float>& logits) {
     std::vector<float> probs(logits.size());
     float max_logit = logits[0];

examples/quantize-stats/quantize-stats.cpp

@@ -19,6 +19,10 @@
 #include <thread>
 #include <mutex>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 struct quantize_stats_params {
     std::string model = "models/7B/ggml-model-f16.bin";
     bool verbose = false;

examples/save-load-state/save-load-state.cpp

@@ -37,7 +37,7 @@ int main(int argc, char ** argv) {
     // init
     auto ctx = llama_init_from_file(params.model.c_str(), lparams);
     auto tokens = std::vector<llama_token>(params.n_ctx);
-    auto n_prompt_tokens = llama_tokenize(ctx, params.prompt.c_str(), tokens.data(), tokens.size(), true);
+    auto n_prompt_tokens = llama_tokenize(ctx, params.prompt.c_str(), tokens.data(), int(tokens.size()), true);
 
     if (n_prompt_tokens < 1) {
         fprintf(stderr, "%s : failed to tokenize prompt\n", __func__);

examples/server/README.md

@@ -16,6 +16,10 @@ This example allow you to have a llama.cpp http server to interact from a web pa
 To get started right away, run the following command, making sure to use the correct path for the model you have:
 
 #### Unix-based systems (Linux, macOS, etc.):
+Make sure to build with the server option on
+```bash
+LLAMA_BUILD_SERVER=1 make
+```
 
 ```bash
 ./server -m models/7B/ggml-model.bin --ctx_size 2048
@@ -289,6 +293,7 @@ Test();
 -   `-ngl N, --n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
 -   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
 -   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
+-   `-lv, --low-vram`: Do not allocate a VRAM scratch buffer for holding temporary results. Reduces VRAM usage at the cost of performance, particularly prompt processing speed. Requires cuBLAS.
 -   `--embedding`: Enable the embedding mode. **Completion function doesn't work in this mode**.
 -   `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`;
 -   `--port`: Set the port to listen. Default: `8080`.

examples/server/server.cpp

@@ -405,6 +405,7 @@ void server_print_usage(int /*argc*/, char **argv, const gpt_params &params)
   fprintf(stderr, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
   fprintf(stderr, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
   fprintf(stderr, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n" );
+  fprintf(stderr, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n" );
 #endif
   fprintf(stderr, "  -m FNAME, --model FNAME\n");
   fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
@@ -537,6 +538,14 @@ bool server_params_parse(int argc, char **argv, server_params &sparams, gpt_para
       }
 #else
       fprintf(stderr, "WARNING: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
+#endif // GGML_USE_CUBLAS
+    }
+    else if (arg == "--low-vram" || arg == "-lv")
+    {
+#ifdef GGML_USE_CUBLAS
+      params.low_vram = true;
+#else
+      fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
 #endif // GGML_USE_CUBLAS
     }
     else if (arg == "--main-gpu" || arg == "-mg")

examples/simple/CMakeLists.txt

@@ -0,0 +1,7 @@
+set(TARGET simple)
+add_executable(${TARGET} simple.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()

examples/simple/simple.cpp

@@ -0,0 +1,177 @@
+#ifndef _GNU_SOURCE
+#define _GNU_SOURCE
+#endif
+
+#include "common.h"
+#include "llama.h"
+#include "build-info.h"
+
+#include <cassert>
+#include <cinttypes>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <vector>
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+#include <signal.h>
+#include <unistd.h>
+#elif defined (_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#define NOMINMAX
+#include <windows.h>
+#include <signal.h>
+#endif
+
+
+
+int main(int argc, char ** argv)
+{
+    gpt_params params;
+
+    //---------------------------------
+    // Print help :
+    //---------------------------------
+
+    if ( argc == 1 || argv[1][0] == '-' )
+    {
+        printf( "usage: %s MODEL_PATH [PROMPT]\n" , argv[0] );
+        return 1 ;
+    }
+
+    //---------------------------------
+    // Load parameters :
+    //---------------------------------
+
+    if ( argc >= 2 )
+    {
+        params.model = argv[1];
+    }
+
+    if ( argc >= 3 )
+    {
+        params.prompt = argv[2];
+    }
+
+    if ( params.prompt.empty() )
+    {
+        params.prompt = "Hello my name is";
+    }
+
+    //---------------------------------
+    // Init LLM :
+    //---------------------------------
+
+    llama_init_backend();
+
+    llama_context * ctx ;
+
+    ctx = llama_init_from_gpt_params( params );
+
+    if ( ctx == NULL )
+    {
+        fprintf( stderr , "%s: error: unable to load model\n" , __func__ );
+        return 1;
+    }
+
+    //---------------------------------
+    // Tokenize the prompt :
+    //---------------------------------
+
+    std::vector<llama_token> tokens_list;
+    tokens_list = ::llama_tokenize( ctx , params.prompt , true );
+
+    const int max_context_size     = llama_n_ctx( ctx );
+    const int max_tokens_list_size = max_context_size - 4 ;
+
+    if ( (int)tokens_list.size() > max_tokens_list_size )
+    {
+        fprintf( stderr , "%s: error: prompt too long (%d tokens, max %d)\n" ,
+             __func__ , (int)tokens_list.size() , max_tokens_list_size );
+        return 1;
+    }
+
+    fprintf( stderr, "\n\n" );
+
+    // Print the tokens from the prompt :
+
+    for( auto id : tokens_list )
+    {
+        printf( "%s" , llama_token_to_str( ctx , id ) );
+    }
+
+    fflush(stdout);
+
+
+    //---------------------------------
+    // Main prediction loop :
+    //---------------------------------
+
+    // The LLM keeps a contextual cache memory of previous token evaluation.
+    // Usually, once this cache is full, it is required to recompute a compressed context based on previous
+    // tokens (see "infinite text generation via context swapping" in the main example), but in this minimalist
+    // example, we will just stop the loop once this cache is full or once an end of stream is detected.
+
+    while ( llama_get_kv_cache_token_count( ctx ) < max_context_size )
+    {
+        //---------------------------------
+        // Evaluate the tokens :
+        //---------------------------------
+
+        if ( llama_eval( ctx , tokens_list.data() , tokens_list.size() , llama_get_kv_cache_token_count( ctx ) , params.n_threads ) )
+        {
+            fprintf( stderr,  "%s : failed to eval\n" , __func__ );
+            return 1;
+        }
+
+        tokens_list.clear();
+
+        //---------------------------------
+        // Select the best prediction :
+        //---------------------------------
+
+        llama_token new_token_id = 0;
+
+        auto logits  = llama_get_logits( ctx );
+        auto n_vocab = llama_n_vocab( ctx ); // the size of the LLM vocabulary (in tokens)
+
+        std::vector<llama_token_data> candidates;
+        candidates.reserve( n_vocab );
+
+        for( llama_token token_id = 0 ; token_id < n_vocab ; token_id++ )
+        {
+            candidates.emplace_back( llama_token_data{ token_id , logits[ token_id ] , 0.0f } );
+        }
+
+        llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+
+        // Select it using the "Greedy sampling" method :
+        new_token_id = llama_sample_token_greedy( ctx , &candidates_p );
+
+
+        // is it an end of stream ?
+        if ( new_token_id == llama_token_eos() )
+        {
+            fprintf(stderr, " [end of text]\n");
+            break;
+        }
+
+        // Print the new token :
+        printf( "%s" , llama_token_to_str( ctx , new_token_id ) );
+        fflush( stdout );
+
+        // Push this new token for next evaluation :
+        tokens_list.push_back( new_token_id );
+
+    } // wend of main loop
+
+    llama_free( ctx );
+
+    return 0;
+}
+
+// EOF

examples/train-text-from-scratch/CMakeLists.txt

@@ -0,0 +1,4 @@
+set(TARGET train-text-from-scratch)
+add_executable(${TARGET} train-text-from-scratch.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/train-text-from-scratch/README.md

@@ -0,0 +1,22 @@
+# train-text-from-scratch
+
+Basic usage instructions:
+
+```bash
+# get training data
+wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/shakespeare.txt
+
+# train
+./bin/train-text-from-scratch \
+        --vocab-model ../models/ggml-vocab.bin \
+        --ctx 64 --embd 256 --head 8 --layer 16 \
+        --checkpoint-in  chk-shakespeare-256x16.bin \
+        --checkpoint-out chk-shakespeare-256x16.bin \
+        --model-out ggml-shakespeare-256x16-f32.bin \
+        --train-data "shakespeare.txt" \
+        -t 6 -b 16 -n 32 --seed 1 --adam-iter 16 \
+        --print-details-interval 0 --predict 16 --use-flash
+
+# predict
+./bin/main -m ggml-shakespeare-256x16-f32.bin
+```

ggml-cuda.h

@@ -28,8 +28,10 @@ void   ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
 
 void   ggml_cuda_free_data(struct ggml_tensor * tensor);
 void   ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
+void   ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
 void   ggml_cuda_set_main_device(int main_device);
 void   ggml_cuda_set_scratch_size(size_t scratch_size);
+void   ggml_cuda_free_scratch(void);
 bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
 
 #ifdef  __cplusplus

ggml-metal.h

@@ -55,6 +55,7 @@ void ggml_metal_set_tensor(struct ggml_metal_context * ctx, struct ggml_tensor *
 void ggml_metal_get_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
 
 // same as ggml_graph_compute but uses Metal
+// creates gf->n_threads command buffers in parallel
 void ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
 
 #ifdef __cplusplus

ggml-metal.m

@@ -284,528 +284,551 @@ void ggml_metal_get_tensor(
 
 void ggml_metal_graph_compute(
         struct ggml_metal_context * ctx,
-             struct ggml_cgraph * gf) {
+               struct ggml_cgraph * gf) {
     metal_printf("%s: evaluating graph\n", __func__);
 
-    size_t offs_src0 = 0;
-    size_t offs_src1 = 0;
-    size_t offs_dst  = 0;
+    // create multiple command buffers and enqueue them
+    // then, we encode the graph into the command buffers in parallel
 
-    id<MTLCommandBuffer> command_buffer  = [ctx->queue commandBuffer];
-    id<MTLComputeCommandEncoder> encoder = nil;
+    const int n_cb = gf->n_threads;
 
-    for (int i = 0; i < gf->n_nodes; ++i) {
-        //metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
+    NSMutableArray * command_buffers = [NSMutableArray arrayWithCapacity:n_cb];
 
-        struct ggml_tensor * src0 = gf->nodes[i]->src0;
-        struct ggml_tensor * src1 = gf->nodes[i]->src1;
-        struct ggml_tensor * dst  = gf->nodes[i];
+    for (int i = 0; i < n_cb; ++i) {
+        command_buffers[i] = [ctx->queue commandBuffer];
 
-        const int64_t  ne00 = src0 ? src0->ne[0] : 0;
-        const int64_t  ne01 = src0 ? src0->ne[1] : 0;
-        const int64_t  ne02 = src0 ? src0->ne[2] : 0;
-        const int64_t  ne03 = src0 ? src0->ne[3] : 0;
+        // enqueue the command buffers in order to specify their execution order
+        [command_buffers[i] enqueue];
+    }
 
-        const uint64_t nb00 = src0 ? src0->nb[0] : 0;
-        const uint64_t nb01 = src0 ? src0->nb[1] : 0;
-        const uint64_t nb02 = src0 ? src0->nb[2] : 0;
-        const uint64_t nb03 = src0 ? src0->nb[3] : 0;
+    // TODO: is this the best way to start threads?
+    dispatch_queue_t queue = dispatch_queue_create("llama.cpp", DISPATCH_QUEUE_CONCURRENT);
 
-        const int64_t  ne10 = src1 ? src1->ne[0] : 0;
-        const int64_t  ne11 = src1 ? src1->ne[1] : 0;
-        const int64_t  ne12 = src1 ? src1->ne[2] : 0;
-        const int64_t  ne13 = src1 ? src1->ne[3] : 0; UNUSED(ne13);
+    for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) {
+        const int n_nodes_per_cb = (gf->n_nodes + n_cb - 1) / n_cb;
 
-        const uint64_t nb10 = src1 ? src1->nb[0] : 0;
-        const uint64_t nb11 = src1 ? src1->nb[1] : 0;
-        const uint64_t nb12 = src1 ? src1->nb[2] : 0;
-        const uint64_t nb13 = src1 ? src1->nb[3] : 0; UNUSED(nb13);
+        dispatch_async(queue, ^{
+            size_t offs_src0 = 0;
+            size_t offs_src1 = 0;
+            size_t offs_dst  = 0;
 
-        const int64_t  ne0  = dst ? dst->ne[0] : 0;
-        const int64_t  ne1  = dst ? dst->ne[1] : 0;
-        const int64_t  ne2  = dst ? dst->ne[2] : 0;
-        const int64_t  ne3  = dst ? dst->ne[3] : 0;
+            id<MTLCommandBuffer> command_buffer = command_buffers[cb_idx];
 
-        const uint64_t nb0  = dst ? dst->nb[0] : 0;
-        const uint64_t nb1  = dst ? dst->nb[1] : 0;
-        const uint64_t nb2  = dst ? dst->nb[2] : 0;
-        const uint64_t nb3  = dst ? dst->nb[3] : 0;
+            id<MTLComputeCommandEncoder> encoder = nil;
 
-        const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
-        const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
-        const enum ggml_type dstt  = dst  ? dst->type  : GGML_TYPE_COUNT;
+            const int node_start =                                      (cb_idx + 0) * n_nodes_per_cb;
+            const int node_end   = (cb_idx == n_cb - 1) ? gf->n_nodes : (cb_idx + 1) * n_nodes_per_cb;
 
-        id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(ctx, src0, &offs_src0) : nil;
-        id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(ctx, src1, &offs_src1) : nil;
-        id<MTLBuffer> id_dst  = dst  ? ggml_metal_get_buffer(ctx, dst,  &offs_dst)  : nil;
+            for (int i = node_start; i < node_end; ++i) {
+                metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
 
-        //metal_printf("%s: op - %s\n", __func__, ggml_op_name(dst->op));
-        //if (src0) {
-        //    metal_printf("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
-        //            ggml_is_contiguous(src0), src0->name);
-        //}
-        //if (src1) {
-        //    metal_printf("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
-        //            ggml_is_contiguous(src1), src1->name);
-        //}
-        //if (dst) {
-        //    metal_printf("%s: dst  - %4s [%5lld, %5lld, %5lld], 1, %s\n",  __func__, ggml_type_name(dstt),  ne0,  ne1,  ne2,
-        //            dst->name);
-        //}
+                struct ggml_tensor * src0 = gf->nodes[i]->src0;
+                struct ggml_tensor * src1 = gf->nodes[i]->src1;
+                struct ggml_tensor * dst  = gf->nodes[i];
 
-        switch (dst->op) {
-            case GGML_OP_RESHAPE:
-            case GGML_OP_VIEW:
-            case GGML_OP_TRANSPOSE:
-            case GGML_OP_PERMUTE:
-                {
-                    // noop
-                } break;
-            case GGML_OP_ADD:
-                {
-                    if (encoder == nil) {
-                        encoder = [command_buffer computeCommandEncoder];
-                    }
+                const int64_t  ne00 = src0 ? src0->ne[0] : 0;
+                const int64_t  ne01 = src0 ? src0->ne[1] : 0;
+                const int64_t  ne02 = src0 ? src0->ne[2] : 0;
+                const int64_t  ne03 = src0 ? src0->ne[3] : 0;
 
-                    [encoder setComputePipelineState:ctx->pipeline_add];
-                    [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];
+                const uint64_t nb00 = src0 ? src0->nb[0] : 0;
+                const uint64_t nb01 = src0 ? src0->nb[1] : 0;
+                const uint64_t nb02 = src0 ? src0->nb[2] : 0;
+                const uint64_t nb03 = src0 ? src0->nb[3] : 0;
 
-                    const int64_t n = ggml_nelements(dst);
+                const int64_t  ne10 = src1 ? src1->ne[0] : 0;
+                const int64_t  ne11 = src1 ? src1->ne[1] : 0;
+                const int64_t  ne12 = src1 ? src1->ne[2] : 0;
+                const int64_t  ne13 = src1 ? src1->ne[3] : 0; UNUSED(ne13);
 
-                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                } break;
-            case GGML_OP_MUL:
-                {
-                    if (encoder == nil) {
-                        encoder = [command_buffer computeCommandEncoder];
-                    }
+                const uint64_t nb10 = src1 ? src1->nb[0] : 0;
+                const uint64_t nb11 = src1 ? src1->nb[1] : 0;
+                const uint64_t nb12 = src1 ? src1->nb[2] : 0;
+                const uint64_t nb13 = src1 ? src1->nb[3] : 0; UNUSED(nb13);
 
-                    if (ggml_nelements(src1) == ne10) {
-                        // src1 is a row
-                        [encoder setComputePipelineState:ctx->pipeline_mul_row];
-                    } else {
-                        [encoder setComputePipelineState:ctx->pipeline_mul];
-                    }
-                    [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];
+                const int64_t  ne0  = dst ? dst->ne[0] : 0;
+                const int64_t  ne1  = dst ? dst->ne[1] : 0;
+                const int64_t  ne2  = dst ? dst->ne[2] : 0;
+                const int64_t  ne3  = dst ? dst->ne[3] : 0;
 
-                    const int64_t n = ggml_nelements(dst);
+                const uint64_t nb0  = dst ? dst->nb[0] : 0;
+                const uint64_t nb1  = dst ? dst->nb[1] : 0;
+                const uint64_t nb2  = dst ? dst->nb[2] : 0;
+                const uint64_t nb3  = dst ? dst->nb[3] : 0;
 
-                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                } break;
-            case GGML_OP_SCALE:
-                {
-                    if (encoder == nil) {
-                        encoder = [command_buffer computeCommandEncoder];
-                    }
+                const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
+                const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
+                const enum ggml_type dstt  = dst  ? dst->type  : GGML_TYPE_COUNT;
 
-                    const float scale = *(const float *) src1->data;
+                id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(ctx, src0, &offs_src0) : nil;
+                id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(ctx, src1, &offs_src1) : nil;
+                id<MTLBuffer> id_dst  = dst  ? ggml_metal_get_buffer(ctx, dst,  &offs_dst)  : nil;
 
-                    [encoder setComputePipelineState:ctx->pipeline_scale];
-                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                    [encoder setBytes:&scale length:sizeof(scale) atIndex:2];
+                //metal_printf("%s: op - %s\n", __func__, ggml_op_name(dst->op));
+                //if (src0) {
+                //    metal_printf("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
+                //            ggml_is_contiguous(src0), src0->name);
+                //}
+                //if (src1) {
+                //    metal_printf("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
+                //            ggml_is_contiguous(src1), src1->name);
+                //}
+                //if (dst) {
+                //    metal_printf("%s: dst  - %4s [%5lld, %5lld, %5lld], 1, %s\n",  __func__, ggml_type_name(dstt),  ne0,  ne1,  ne2,
+                //            dst->name);
+                //}
 
-                    const int64_t n = ggml_nelements(dst);
+                switch (dst->op) {
+                    case GGML_OP_RESHAPE:
+                    case GGML_OP_VIEW:
+                    case GGML_OP_TRANSPOSE:
+                    case GGML_OP_PERMUTE:
+                        {
+                            // noop
+                        } break;
+                    case GGML_OP_ADD:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
 
-                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                } break;
-            case GGML_OP_SILU:
-                {
-                    if (encoder == nil) {
-                        encoder = [command_buffer computeCommandEncoder];
-                    }
+                            [encoder setComputePipelineState:ctx->pipeline_add];
+                            [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 setComputePipelineState:ctx->pipeline_silu];
-                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            const int64_t n = ggml_nelements(dst);
 
-                    const int64_t n = ggml_nelements(dst);
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_MUL:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
 
-                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                } break;
-            case GGML_OP_RELU:
-                {
-                    if (encoder == nil) {
-                        encoder = [command_buffer computeCommandEncoder];
-                    }
+                            if (ggml_nelements(src1) == ne10) {
+                                // src1 is a row
+                                [encoder setComputePipelineState:ctx->pipeline_mul_row];
+                            } else {
+                                [encoder setComputePipelineState:ctx->pipeline_mul];
+                            }
+                            [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 setComputePipelineState:ctx->pipeline_relu];
-                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            const int64_t n = ggml_nelements(dst);
 
-                    const int64_t n = ggml_nelements(dst);
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_SCALE:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
 
-                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                } break;
-            case GGML_OP_GELU:
-            {
-                    if (encoder == nil) {
-                        encoder = [command_buffer computeCommandEncoder];
-                    }
+                            const float scale = *(const float *) src1->data;
 
-                    [encoder setComputePipelineState:ctx->pipeline_gelu];
-                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setComputePipelineState:ctx->pipeline_scale];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBytes:&scale length:sizeof(scale) atIndex:2];
 
-                    const int64_t n = ggml_nelements(dst);
+                            const int64_t n = ggml_nelements(dst);
 
-                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-            } break;
-            case GGML_OP_SOFT_MAX:
-                {
-                    if (encoder == nil) {
-                        encoder = [command_buffer computeCommandEncoder];
-                    }
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_SILU:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
 
-                    const int nth = 32;
+                            [encoder setComputePipelineState:ctx->pipeline_silu];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
 
-                    [encoder setComputePipelineState:ctx->pipeline_soft_max];
-                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                    [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
-                    [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
-                    [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
-                    [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
+                            const int64_t n = ggml_nelements(dst);
 
-                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                } break;
-            case GGML_OP_DIAG_MASK_INF:
-                {
-                    if (encoder == nil) {
-                        encoder = [command_buffer computeCommandEncoder];
-                    }
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_RELU:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
 
-                    const int n_past = ((int32_t *)(src1->data))[0];
+                            [encoder setComputePipelineState:ctx->pipeline_relu];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
 
-                    [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
-                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                    [encoder setBytes:&ne00   length:sizeof(ne00) atIndex:2];
-                    [encoder setBytes:&ne01   length:sizeof(ne01) atIndex:3];
-                    [encoder setBytes:&n_past length:sizeof(int)  atIndex:4];
+                            const int64_t n = ggml_nelements(dst);
 
-                    [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                } break;
-            case GGML_OP_MUL_MAT:
-                {
-                    // TODO: needs to be updated after PR: https://github.com/ggerganov/ggml/pull/224
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_GELU:
+                    {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
 
-                    GGML_ASSERT(ne00 == ne10);
-                    GGML_ASSERT(ne02 == ne12);
+                            [encoder setComputePipelineState:ctx->pipeline_gelu];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
 
-                    if (ggml_is_contiguous(src0) &&
-                        ggml_is_contiguous(src1) &&
-                        (src0t == GGML_TYPE_F32 || src0t == GGML_TYPE_F16) && ne11 > 1) {
+                            const int64_t n = ggml_nelements(dst);
 
-                        if (encoder != nil) {
-                            [encoder endEncoding];
-                            encoder = nil;
-                        }
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                    } break;
+                    case GGML_OP_SOFT_MAX:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
 
-                        MPSDataType src0dt = src0t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
-                        MPSDataType src1dt = src1t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
+                            const int nth = 32;
 
-                        // for F32 x F32 we use MPS
-                        MPSMatrixDescriptor * desc0 = [MPSMatrixDescriptor
-                            matrixDescriptorWithRows:ne01 columns:ne00 rowBytes:src0->nb[1] dataType:src0dt];
+                            [encoder setComputePipelineState:ctx->pipeline_soft_max];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
+                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
+                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
+                            [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
 
-                        MPSMatrixDescriptor * desc1 = [MPSMatrixDescriptor
-                            matrixDescriptorWithRows:ne11 columns:ne10 rowBytes:src1->nb[1] dataType:src1dt];
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                        } break;
+                    case GGML_OP_DIAG_MASK_INF:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
 
-                        MPSMatrixDescriptor * desc  = [MPSMatrixDescriptor
-                            matrixDescriptorWithRows:ne1 columns:ne0 rowBytes:dst->nb[1] dataType:MPSDataTypeFloat32];
+                            const int n_past = ((int32_t *)(src1->data))[0];
 
-                        MPSMatrixMultiplication * mul = [[MPSMatrixMultiplication alloc]
-                            initWithDevice:ctx->device transposeLeft:false transposeRight:true
-                                resultRows:ne11 resultColumns:ne01 interiorColumns:ne00 alpha:1.0 beta:0.0];
+                            [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBytes:&ne00   length:sizeof(ne00) atIndex:2];
+                            [encoder setBytes:&ne01   length:sizeof(ne01) atIndex:3];
+                            [encoder setBytes:&n_past length:sizeof(int)  atIndex:4];
 
-                        // we need to do ne02 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 < ne02; ++i02) {
-                            size_t offs_src0_cur = offs_src0 + i02*nb02;
-                            size_t offs_src1_cur = offs_src1 + i02*nb12;
-                            size_t offs_dst_cur  = offs_dst  + i02*nb2;
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_MUL_MAT:
+                        {
+                            // TODO: needs to be updated after PR: https://github.com/ggerganov/ggml/pull/224
 
-                            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 ];
+                            GGML_ASSERT(ne00 == ne10);
+                            GGML_ASSERT(ne02 == ne12);
 
-                            [mul encodeToCommandBuffer:command_buffer leftMatrix:mat_src1 rightMatrix:mat_src0 resultMatrix:mat_dst];
-                        }
-                    } else {
-                        if (encoder == nil) {
-                            encoder = [command_buffer computeCommandEncoder];
-                        }
+                            if (ggml_is_contiguous(src0) &&
+                                ggml_is_contiguous(src1) &&
+                                (src0t == GGML_TYPE_F32 || src0t == GGML_TYPE_F16) && ne11 > 1) {
 
-                        int nth0 = 32;
-                        int nth1 = 1;
-
-                        // use custom matrix x vector kernel
-                        switch (src0t) {
-                            case GGML_TYPE_F16:
-                                {
-                                    GGML_ASSERT(ne02 == ne12);
-
-                                    nth0 = 64;
-                                    nth1 = 1;
-                                    [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
-                                } break;
-                            case GGML_TYPE_Q4_0:
-                                {
-                                    GGML_ASSERT(ne02 == 1);
-                                    GGML_ASSERT(ne12 == 1);
-
-                                    nth0 = 8;
-                                    nth1 = 8;
-                                    [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_0_f32];
-                                } break;
-                            case GGML_TYPE_Q4_1:
-                                {
-                                    GGML_ASSERT(ne02 == 1);
-                                    GGML_ASSERT(ne12 == 1);
-
-                                    nth0 = 8;
-                                    nth1 = 8;
-                                    [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_1_f32];
-                                } break;
-                            case GGML_TYPE_Q2_K:
-                                {
-                                    GGML_ASSERT(ne02 == 1);
-                                    GGML_ASSERT(ne12 == 1);
-
-                                    nth0 = 4;
-                                    nth1 = 16;
-                                    [encoder setComputePipelineState:ctx->pipeline_mul_mat_q2_k_f32];
-                                } break;
-                            case GGML_TYPE_Q3_K:
-                                {
-                                    GGML_ASSERT(ne02 == 1);
-                                    GGML_ASSERT(ne12 == 1);
-
-                                    nth0 = 4;
-                                    nth1 = 16;
-                                    [encoder setComputePipelineState:ctx->pipeline_mul_mat_q3_k_f32];
-                                } break;
-                            case GGML_TYPE_Q4_K:
-                                {
-                                    GGML_ASSERT(ne02 == 1);
-                                    GGML_ASSERT(ne12 == 1);
-
-                                    nth0 = 4;
-                                    nth1 = 16;
-                                    [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_k_f32];
-                                } break;
-                            case GGML_TYPE_Q5_K:
-                                {
-                                    GGML_ASSERT(ne02 == 1);
-                                    GGML_ASSERT(ne12 == 1);
-
-                                    nth0 = 4;
-                                    nth1 = 16;
-                                    [encoder setComputePipelineState:ctx->pipeline_mul_mat_q5_k_f32];
-                                } break;
-                            case GGML_TYPE_Q6_K:
-                                {
-                                    GGML_ASSERT(ne02 == 1);
-                                    GGML_ASSERT(ne12 == 1);
-
-                                    nth0 = 4;
-                                    nth1 = 16;
-                                    [encoder setComputePipelineState:ctx->pipeline_mul_mat_q6_k_f32];
-                                } break;
-                            default:
-                                {
-                                    fprintf(stderr, "Asserting on type %d\n",(int)src0t);
-                                    GGML_ASSERT(false && "not implemented");
+                                if (encoder != nil) {
+                                    [encoder endEncoding];
+                                    encoder = nil;
                                 }
-                        };
 
+                                MPSDataType src0dt = src0t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
+                                MPSDataType src1dt = src1t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
 
-                        [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:&ne01 length:sizeof(ne01) atIndex:4];
-                        [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:5];
-                        [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:6];
-                        [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:7];
-                        [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:8];
-                        [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:9];
-                        [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:10];
-                        [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:11];
-                        [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:12];
-                        [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:13];
-                        [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:14];
+                                // for F32 x F32 we use MPS
+                                MPSMatrixDescriptor * desc0 = [MPSMatrixDescriptor
+                                    matrixDescriptorWithRows:ne01 columns:ne00 rowBytes:src0->nb[1] dataType:src0dt];
 
-                        if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1) {
-                            [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                        }
-                        else if (src0t == GGML_TYPE_Q2_K ||
-                                 src0t == GGML_TYPE_Q3_K ||
-                                 src0t == GGML_TYPE_Q4_K ||
-                                 src0t == GGML_TYPE_Q5_K ||
-                                 src0t == GGML_TYPE_Q6_K) {
-                            [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                        } else {
-                            [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                        }
-                    }
-                } break;
-            case GGML_OP_GET_ROWS:
-                {
-                    if (encoder == nil) {
-                        encoder = [command_buffer computeCommandEncoder];
-                    }
+                                MPSMatrixDescriptor * desc1 = [MPSMatrixDescriptor
+                                    matrixDescriptorWithRows:ne11 columns:ne10 rowBytes:src1->nb[1] dataType:src1dt];
 
-                    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;
-                        case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_1]; break;
-                        case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q2_k]; break;
-                        case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q3_k]; break;
-                        case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_k]; break;
-                        case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q5_k]; break;
-                        case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q6_k]; break;
-                        default: GGML_ASSERT(false && "not implemented");
-                    }
+                                MPSMatrixDescriptor * desc  = [MPSMatrixDescriptor
+                                    matrixDescriptorWithRows:ne1 columns:ne0 rowBytes:dst->nb[1] dataType:MPSDataTypeFloat32];
 
-                    [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:&(src0->ne[0]) length:sizeof( int64_t) atIndex:3];
-                    [encoder setBytes:&(src0->nb[1]) length:sizeof(uint64_t) atIndex:4];
-                    [encoder setBytes:&(dst->nb[1])  length:sizeof(uint64_t) atIndex:5];
+                                MPSMatrixMultiplication * mul = [[MPSMatrixMultiplication alloc]
+                                    initWithDevice:ctx->device transposeLeft:false transposeRight:true
+                                        resultRows:ne11 resultColumns:ne01 interiorColumns:ne00 alpha:1.0 beta:0.0];
 
-                    const int64_t n = ggml_nelements(src1);
+                                // we need to do ne02 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 < ne02; ++i02) {
+                                    size_t offs_src0_cur = offs_src0 + i02*nb02;
+                                    size_t offs_src1_cur = offs_src1 + i02*nb12;
+                                    size_t offs_dst_cur  = offs_dst  + i02*nb2;
 
-                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                } break;
-            case GGML_OP_RMS_NORM:
-                {
-                    if (encoder == nil) {
-                        encoder = [command_buffer computeCommandEncoder];
-                    }
+                                    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 ];
 
-                    const float eps = 1e-6f;
+                                    [mul encodeToCommandBuffer:command_buffer leftMatrix:mat_src1 rightMatrix:mat_src0 resultMatrix:mat_dst];
+                                }
+                            } else {
+                                if (encoder == nil) {
+                                    encoder = [command_buffer computeCommandEncoder];
+                                }
 
-                    const int nth = 256;
+                                int nth0 = 32;
+                                int nth1 = 1;
 
-                    [encoder setComputePipelineState:ctx->pipeline_rms_norm];
-                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                    [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
-                    [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3];
-                    [encoder setBytes:&eps  length:sizeof(   float) atIndex:4];
-                    [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
+                                // use custom matrix x vector kernel
+                                switch (src0t) {
+                                    case GGML_TYPE_F16:
+                                        {
+                                            GGML_ASSERT(ne02 == ne12);
 
-                    const int64_t nrows = ggml_nrows(src0);
+                                            nth0 = 64;
+                                            nth1 = 1;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
+                                        } break;
+                                    case GGML_TYPE_Q4_0:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
 
-                    [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                } break;
-            case GGML_OP_ROPE:
-                {
-                    if (encoder == nil) {
-                        encoder = [command_buffer computeCommandEncoder];
-                    }
+                                            nth0 = 8;
+                                            nth1 = 8;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_0_f32];
+                                        } break;
+                                    case GGML_TYPE_Q4_1:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
 
-                    const int n_dims = ((int32_t *) src1->data)[1];
-                    const int mode   = ((int32_t *) src1->data)[2];
+                                            nth0 = 8;
+                                            nth1 = 8;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_1_f32];
+                                        } break;
+                                    case GGML_TYPE_Q2_K:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
 
-                    const int n_past = ((int32_t *)(src1->data))[0];
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q2_k_f32];
+                                        } break;
+                                    case GGML_TYPE_Q3_K:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
 
-                    [encoder setComputePipelineState:ctx->pipeline_rope];
-                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                    [encoder setBytes:&ne00   length:sizeof( int64_t) atIndex:2];
-                    [encoder setBytes:&ne01   length:sizeof( int64_t) atIndex:3];
-                    [encoder setBytes:&ne02   length:sizeof( int64_t) atIndex:4];
-                    [encoder setBytes:&ne03   length:sizeof( int64_t) atIndex:5];
-                    [encoder setBytes:&nb00   length:sizeof(uint64_t) atIndex:6];
-                    [encoder setBytes:&nb01   length:sizeof(uint64_t) atIndex:7];
-                    [encoder setBytes:&nb02   length:sizeof(uint64_t) atIndex:8];
-                    [encoder setBytes:&nb03   length:sizeof(uint64_t) atIndex:9];
-                    [encoder setBytes:&ne0    length:sizeof( int64_t) atIndex:10];
-                    [encoder setBytes:&ne1    length:sizeof( int64_t) atIndex:11];
-                    [encoder setBytes:&ne2    length:sizeof( int64_t) atIndex:12];
-                    [encoder setBytes:&ne3    length:sizeof( int64_t) atIndex:13];
-                    [encoder setBytes:&nb0    length:sizeof(uint64_t) atIndex:14];
-                    [encoder setBytes:&nb1    length:sizeof(uint64_t) atIndex:15];
-                    [encoder setBytes:&nb2    length:sizeof(uint64_t) atIndex:16];
-                    [encoder setBytes:&nb3    length:sizeof(uint64_t) atIndex:17];
-                    [encoder setBytes:&n_past length:sizeof(     int) atIndex:18];
-                    [encoder setBytes:&n_dims length:sizeof(     int) atIndex:19];
-                    [encoder setBytes:&mode   length:sizeof(     int) atIndex:20];
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q3_k_f32];
+                                        } break;
+                                    case GGML_TYPE_Q4_K:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
 
-                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                } break;
-            case GGML_OP_CPY:
-                {
-                    if (encoder == nil) {
-                        encoder = [command_buffer computeCommandEncoder];
-                    }
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_k_f32];
+                                        } break;
+                                    case GGML_TYPE_Q5_K:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
 
-                    const int nth = 32;
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q5_k_f32];
+                                        } break;
+                                    case GGML_TYPE_Q6_K:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
 
-                    switch (src0t) {
-                        case GGML_TYPE_F32:
-                            {
-                                switch (dstt) {
-                                    case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f16]; break;
-                                    case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f32]; break;
-                                    default: GGML_ASSERT(false && "not implemented");
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q6_k_f32];
+                                        } break;
+                                    default:
+                                        {
+                                            fprintf(stderr, "Asserting on type %d\n",(int)src0t);
+                                            GGML_ASSERT(false && "not implemented");
+                                        }
                                 };
-                            } break;
-                        default: GGML_ASSERT(false && "not implemented");
-                    }
 
-                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                    [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
-                    [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
-                    [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
-                    [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
-                    [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
-                    [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
-                    [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
-                    [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
-                    [encoder setBytes:&ne0  length:sizeof( int64_t) atIndex:10];
-                    [encoder setBytes:&ne1  length:sizeof( int64_t) atIndex:11];
-                    [encoder setBytes:&ne2  length:sizeof( int64_t) atIndex:12];
-                    [encoder setBytes:&ne3  length:sizeof( int64_t) atIndex:13];
-                    [encoder setBytes:&nb0  length:sizeof(uint64_t) atIndex:14];
-                    [encoder setBytes:&nb1  length:sizeof(uint64_t) atIndex:15];
-                    [encoder setBytes:&nb2  length:sizeof(uint64_t) atIndex:16];
-                    [encoder setBytes:&nb3  length:sizeof(uint64_t) atIndex:17];
+                                [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:&ne01 length:sizeof(ne01) atIndex:4];
+                                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:5];
+                                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:6];
+                                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:7];
+                                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:8];
+                                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:9];
+                                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:10];
+                                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:11];
+                                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:12];
+                                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:13];
+                                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:14];
 
-                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                } break;
-            default:
-                fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
-                GGML_ASSERT(false);
-        }
+                                if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1) {
+                                    [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
+                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
+                                else if (src0t == GGML_TYPE_Q2_K ||
+                                         src0t == GGML_TYPE_Q3_K ||
+                                         src0t == GGML_TYPE_Q4_K ||
+                                         src0t == GGML_TYPE_Q5_K ||
+                                         src0t == GGML_TYPE_Q6_K) {
+                                    [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
+                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                } else {
+                                    [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
+                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
+                            }
+                        } break;
+                    case GGML_OP_GET_ROWS:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            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;
+                                case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_1]; break;
+                                case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q2_k]; break;
+                                case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q3_k]; break;
+                                case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_k]; break;
+                                case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q5_k]; break;
+                                case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q6_k]; break;
+                                default: GGML_ASSERT(false && "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:&(src0->ne[0]) length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&(src0->nb[1]) length:sizeof(uint64_t) atIndex:4];
+                            [encoder setBytes:&(dst->nb[1])  length:sizeof(uint64_t) atIndex:5];
+
+                            const int64_t n = ggml_nelements(src1);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_RMS_NORM:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            const float eps = 1e-6f;
+
+                            const int nth = 256;
+
+                            [encoder setComputePipelineState:ctx->pipeline_rms_norm];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
+                            [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3];
+                            [encoder setBytes:&eps  length:sizeof(   float) atIndex:4];
+                            [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
+
+                            const int64_t nrows = ggml_nrows(src0);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                        } break;
+                    case GGML_OP_ROPE:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            const int n_dims = ((int32_t *) src1->data)[1];
+                            const int mode   = ((int32_t *) src1->data)[2];
+
+                            const int n_past = ((int32_t *)(src1->data))[0];
+
+                            [encoder setComputePipelineState:ctx->pipeline_rope];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBytes:&ne00   length:sizeof( int64_t) atIndex:2];
+                            [encoder setBytes:&ne01   length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&ne02   length:sizeof( int64_t) atIndex:4];
+                            [encoder setBytes:&ne03   length:sizeof( int64_t) atIndex:5];
+                            [encoder setBytes:&nb00   length:sizeof(uint64_t) atIndex:6];
+                            [encoder setBytes:&nb01   length:sizeof(uint64_t) atIndex:7];
+                            [encoder setBytes:&nb02   length:sizeof(uint64_t) atIndex:8];
+                            [encoder setBytes:&nb03   length:sizeof(uint64_t) atIndex:9];
+                            [encoder setBytes:&ne0    length:sizeof( int64_t) atIndex:10];
+                            [encoder setBytes:&ne1    length:sizeof( int64_t) atIndex:11];
+                            [encoder setBytes:&ne2    length:sizeof( int64_t) atIndex:12];
+                            [encoder setBytes:&ne3    length:sizeof( int64_t) atIndex:13];
+                            [encoder setBytes:&nb0    length:sizeof(uint64_t) atIndex:14];
+                            [encoder setBytes:&nb1    length:sizeof(uint64_t) atIndex:15];
+                            [encoder setBytes:&nb2    length:sizeof(uint64_t) atIndex:16];
+                            [encoder setBytes:&nb3    length:sizeof(uint64_t) atIndex:17];
+                            [encoder setBytes:&n_past length:sizeof(     int) atIndex:18];
+                            [encoder setBytes:&n_dims length:sizeof(     int) atIndex:19];
+                            [encoder setBytes:&mode   length:sizeof(     int) atIndex:20];
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_CPY:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            const int nth = 32;
+
+                            switch (src0t) {
+                                case GGML_TYPE_F32:
+                                    {
+                                        switch (dstt) {
+                                            case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f16]; break;
+                                            case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f32]; break;
+                                            default: GGML_ASSERT(false && "not implemented");
+                                        };
+                                    } break;
+                                default: GGML_ASSERT(false && "not implemented");
+                            }
+
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
+                            [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
+                            [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
+                            [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
+                            [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
+                            [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
+                            [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
+                            [encoder setBytes:&ne0  length:sizeof( int64_t) atIndex:10];
+                            [encoder setBytes:&ne1  length:sizeof( int64_t) atIndex:11];
+                            [encoder setBytes:&ne2  length:sizeof( int64_t) atIndex:12];
+                            [encoder setBytes:&ne3  length:sizeof( int64_t) atIndex:13];
+                            [encoder setBytes:&nb0  length:sizeof(uint64_t) atIndex:14];
+                            [encoder setBytes:&nb1  length:sizeof(uint64_t) atIndex:15];
+                            [encoder setBytes:&nb2  length:sizeof(uint64_t) atIndex:16];
+                            [encoder setBytes:&nb3  length:sizeof(uint64_t) atIndex:17];
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                        } break;
+                    default:
+                        fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+                        GGML_ASSERT(false);
+                }
+            }
+
+            if (encoder != nil) {
+                [encoder endEncoding];
+                encoder = nil;
+            }
+
+            [command_buffer commit];
+        });
     }
 
-    if (encoder != nil) {
-        [encoder endEncoding];
-        encoder = nil;
-    }
+    // wait for all threads to finish
+    dispatch_barrier_sync(queue, ^{});
 
-    [command_buffer commit];
-    [command_buffer waitUntilCompleted];
-
-    {
-        const double time_elapsed = [command_buffer GPUEndTime] - [command_buffer GPUStartTime];
-        UNUSED(time_elapsed);
-
-        metal_printf("%s: time elapsed = %f ms\n", __func__, time_elapsed * 1000.0);
-    }
+    [command_buffers[n_cb - 1] waitUntilCompleted];
 }

ggml-opencl.cpp

@@ -15,7 +15,7 @@
 
 #include "ggml.h"
 
-#define CL_DMMV_BLOCK_SIZE 32;
+#define CL_DMMV_BLOCK_SIZE 32
 
 #define MULTILINE_QUOTE(...) #__VA_ARGS__
 static std::string program_source = MULTILINE_QUOTE(
@@ -59,6 +59,46 @@ struct __attribute__ ((packed)) block_q8_0
     int8_t qs[QK8_0];
 };
 
+struct __attribute__((packed)) block_q2_K
+{
+    uint8_t scales[16];
+    uint8_t qs[64];
+    half d;
+    half dmin;
+};
+
+struct __attribute__((packed)) block_q3_K
+{
+    uint8_t hmask[32];
+    uint8_t qs[64];
+    uint8_t scales[12];
+    half d;
+};
+
+struct __attribute__((packed)) block_q4_K
+{
+    half d;
+    half dmin;
+    uint8_t scales[12];
+    uint8_t qs[128];
+};
+
+struct __attribute__((packed)) block_q5_K
+{
+    half d;
+    half dmin;
+    uint8_t scales[12];
+    uint8_t qh[32];
+    uint8_t qs[128];
+};
+
+struct __attribute__((packed)) block_q6_K
+{
+    uint8_t ql[128];
+    uint8_t qh[64];
+    int8_t scales[16];
+    half d;
+};
 
 __kernel void convert_fp16_to_fp32(__global half* x, __global float* y) {
     const uint i = get_global_id(0);
@@ -131,8 +171,314 @@ void convert_f16(__global half* x, const int ib, const int iqs, float* v0, float
     *v0 = vload_half(0, &x[ib + 0]);
     *v1 = vload_half(0, &x[ib + 1]);
 }
+
+inline void get_scale_min_k4(int j, const __global uint8_t *q, uint8_t *d, uint8_t *m)
+{
+    if (j < 4)
+    {
+        *d = q[j] & 63;
+        *m = q[j + 4] & 63;
+    }
+    else
+    {
+        *d = (q[j + 4] & 0xF) | ((q[j - 4] >> 6) << 4);
+        *m = (q[j + 4] >> 4) | ((q[j - 0] >> 6) << 4);
+    }
+}
+
+__kernel void dequantize_block_q2_K(__global const struct block_q2_K *x, __global float *yy)
+{
+    const int i = get_group_id(0);
+    const int tid = get_local_id(0);
+    const int n = tid / 32;
+    const int l = tid - 32 * n;
+    const int is = 8 * n + l / 16;
+
+    const uint8_t q = x[i].qs[32 * n + l];
+    __global float *y = yy + i * 256 + 128 * n;
+
+    const float dall = vload_half(0, &x[i].d);
+    const float dmin = vload_half(0, &x[i].dmin);
+
+    y[l + 0] = dall * (x[i].scales[is + 0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is + 0] >> 4);
+    y[l + 32] = dall * (x[i].scales[is + 2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is + 2] >> 4);
+    y[l + 64] = dall * (x[i].scales[is + 4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is + 4] >> 4);
+    y[l + 96] = dall * (x[i].scales[is + 6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is + 6] >> 4);
+}
+
+__kernel void dequantize_block_q3_K(__global const struct block_q3_K *x, __global float *yy)
+{
+    int r = get_local_id(0) / 4;
+    int i = get_group_id(0);
+    int tid = r / 2;
+    int is0 = r % 2;
+    int l0 = 16 * is0 + 4 * (get_local_id(0) % 4);
+    int n = tid / 4;
+    int j = tid - 4 * n;
+
+    uint8_t m = 1 << (4 * n + j);
+    int is = 8 * n + 2 * j + is0;
+    int shift = 2 * j;
+
+    int8_t us = is < 4 ? (x[i].scales[is - 0] & 0xF) | (((x[i].scales[is + 8] >> 0) & 3) << 4)
+              : is < 8 ? (x[i].scales[is - 0] & 0xF) | (((x[i].scales[is + 4] >> 2) & 3) << 4)
+              : is < 12  ? (x[i].scales[is - 8] >> 4) | (((x[i].scales[is + 0] >> 4) & 3) << 4)
+              : (x[i].scales[is - 8] >> 4) | (((x[i].scales[is - 4] >> 6) & 3) << 4);
+    float d_all = vload_half(0, &x[i].d);
+    float dl = d_all * (us - 32);
+
+    __global float *y = yy + i * 256 + 128 * n + 32 * j;
+    const __global uint8_t *q = x[i].qs + 32 * n;
+    const __global uint8_t *hm = x[i].hmask;
+
+    for (int l = l0; l < l0 + 4; ++l)
+        y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4));
+}
+
+__kernel void dequantize_block_q4_K(__global const struct block_q4_K *x, __global float *yy)
+{
+    const int i = get_group_id(0);
+    const int tid = get_local_id(0);
+    const int il = tid / 8;
+    const int ir = tid % 8;
+    const int is = 2 * il;
+    const int n = 4;
+
+    __global float *y = yy + i * 256 + 64 * il + n * ir;
+
+    const float dall = vload_half(0, &x[i].d);
+    const float dmin = vload_half(0, &x[i].dmin);
+
+    __global const uint8_t *q = x[i].qs + 32 * il + n * ir;
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[i].scales, &sc, &m);
+    float d1 = dall * sc;
+    float m1 = dmin * m;
+    get_scale_min_k4(is + 1, x[i].scales, &sc, &m);
+    float d2 = dall * sc;
+    float m2 = dmin * m;
+    for (int l = 0; l < n; ++l)
+    {
+        y[l + 0] = d1 * (q[l] & 0xF) - m1;
+        y[l + 32] = d2 * (q[l] >> 4) - m2;
+    }
+}
+
+__kernel void dequantize_block_q5_K(__global const struct block_q5_K *x, __global float *yy)
+{
+    const int i = get_group_id(0);
+    const int tid = get_local_id(0);
+    const int il = tid / 16;
+    const int ir = tid % 16;
+    const int is = 2 * il;
+
+    __global float *y = yy + i * 256 + 64 * il + 2 * ir;
+
+    const float dall = vload_half(0, &x[i].d);
+    const float dmin = vload_half(0, &x[i].dmin);
+
+    __global const uint8_t *ql = x[i].qs + 32 * il + 2 * ir;
+    __global const uint8_t *qh = x[i].qh + 2 * ir;
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[i].scales, &sc, &m);
+    const float d1 = dall * sc;
+    const float m1 = dmin * m;
+    get_scale_min_k4(is + 1, x[i].scales, &sc, &m);
+    const float d2 = dall * sc;
+    const float m2 = dmin * m;
+
+    uint8_t hm = 1 << (2 * il);
+    y[0] = d1 * ((ql[0] & 0xF) + (qh[0] & hm ? 16 : 0)) - m1;
+    y[1] = d1 * ((ql[1] & 0xF) + (qh[1] & hm ? 16 : 0)) - m1;
+    hm <<= 1;
+    y[32] = d2 * ((ql[0] >> 4) + (qh[0] & hm ? 16 : 0)) - m2;
+    y[33] = d2 * ((ql[1] >> 4) + (qh[1] & hm ? 16 : 0)) - m2;
+}
+
+__kernel void dequantize_block_q6_K(__global const struct block_q6_K *x, __global float *yy)
+{
+    const int i = get_group_id(0);
+    const int tid = get_local_id(0);
+    const int ip = tid / 32;
+    const int il = tid - 32 * ip;
+    const int is = 8 * ip + il / 16;
+
+    __global float *y = yy + i * 256 + 128 * ip + il;
+
+    const float d = vload_half(0, &x[i].d);
+
+    __global const uint8_t *ql = x[i].ql + 64 * ip + il;
+    const uint8_t qh = x[i].qh[32 * ip + il];
+    __global const int8_t *sc = x[i].scales + is;
+
+    y[0] = d * sc[0] * ((int8_t)((ql[0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
+    y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
+    y[64] = d * sc[4] * ((int8_t)((ql[0] >> 4) | (((qh >> 4) & 3) << 4)) - 32);
+    y[96] = d * sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32);
+}
+
+
+void vec_dot_q2_K(__global const struct block_q2_K* x, const int ib, const int iqs, const __global float *yy, float *result) {
+
+    int n = iqs / 128;
+    int r = iqs - 128 * n;
+    int l = r / 8;
+
+    __global const float *y = yy + 128 * n + l;
+    __global const uint8_t *q = x[ib].qs + 32 * n + l;
+    __global const uint8_t *s = x[ib].scales + 8 * n;
+
+    const float dall = vload_half(0, &x[ib].d);
+    const float dmin = vload_half(0, &x[ib].dmin);
+
+    float sum = y[  0] * (dall * ((s[0] & 0xF) * ((q[ 0] >> 0) & 3)) - dmin * (s[0] >> 4))
+              + y[ 32] * (dall * ((s[2] & 0xF) * ((q[ 0] >> 2) & 3)) - dmin * (s[2] >> 4))
+              + y[ 64] * (dall * ((s[4] & 0xF) * ((q[ 0] >> 4) & 3)) - dmin * (s[4] >> 4))
+              + y[ 96] * (dall * ((s[6] & 0xF) * ((q[ 0] >> 6) & 3)) - dmin * (s[6] >> 4))
+              + y[ 16] * (dall * ((s[1] & 0xF) * ((q[16] >> 0) & 3)) - dmin * (s[1] >> 4))
+              + y[ 48] * (dall * ((s[3] & 0xF) * ((q[16] >> 2) & 3)) - dmin * (s[3] >> 4))
+              + y[ 80] * (dall * ((s[5] & 0xF) * ((q[16] >> 4) & 3)) - dmin * (s[5] >> 4))
+              + y[112] * (dall * ((s[7] & 0xF) * ((q[16] >> 6) & 3)) - dmin * (s[7] >> 4));
+
+    *result = sum;
+}
+
+void vec_dot_q3_K(__global const struct block_q3_K* x, const int ib, const int iqs, const __global float *yy, float *result) {
+
+    const uint32_t kmask1 = 0x03030303;
+    const uint32_t kmask2 = 0x0f0f0f0f;
+
+    uint32_t aux[3];
+    uint32_t utmp[4];
+
+    int n = iqs/128;
+    int r = iqs - 128*n;
+    int l = r/8;
+
+    __global const float   * y = yy + 128*n + l;
+    __global const uint8_t * q = x[ib].qs + 32*n + l;
+    __global const uint8_t * hm = x[ib].hmask + l;
+    const int8_t * s = (const int8_t *)utmp + 8*n;
+
+    aux[0] = x[ib].scales[0] | x[ib].scales[1] << 8 | x[ib].scales[2] << 16 | x[ib].scales[3] << 24;
+    aux[1] = x[ib].scales[4] | x[ib].scales[5] << 8 | x[ib].scales[6] << 16 | x[ib].scales[7] << 24;
+    aux[2] = x[ib].scales[8] | x[ib].scales[9] << 8 | x[ib].scales[10] << 16 | x[ib].scales[11] << 24;
+
+    utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
+    utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
+    utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
+    utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
+
+    const float dall = vload_half(0, &x[ib].d);
+    const uint8_t m = 1 << (4*n);
+
+    float sum = y[  0] * (s[0] - 32) * (((q[ 0] >> 0) & 3) - (hm[ 0] & (m << 0) ? 0 : 4))
+              + y[ 32] * (s[2] - 32) * (((q[ 0] >> 2) & 3) - (hm[ 0] & (m << 1) ? 0 : 4))
+              + y[ 64] * (s[4] - 32) * (((q[ 0] >> 4) & 3) - (hm[ 0] & (m << 2) ? 0 : 4))
+              + y[ 96] * (s[6] - 32) * (((q[ 0] >> 6) & 3) - (hm[ 0] & (m << 3) ? 0 : 4))
+              + y[ 16] * (s[1] - 32) * (((q[16] >> 0) & 3) - (hm[16] & (m << 0) ? 0 : 4))
+              + y[ 48] * (s[3] - 32) * (((q[16] >> 2) & 3) - (hm[16] & (m << 1) ? 0 : 4))
+              + y[ 80] * (s[5] - 32) * (((q[16] >> 4) & 3) - (hm[16] & (m << 2) ? 0 : 4))
+              + y[112] * (s[7] - 32) * (((q[16] >> 6) & 3) - (hm[16] & (m << 3) ? 0 : 4));
+
+    *result = sum * dall;
+
+}
+
+void vec_dot_q4_K(__global const struct block_q4_K* x, const int ib, const int iqs, const __global float *yy, float *result) {
+
+    const int j  = iqs / 64;        // j  is in 0...3
+    const int ir = (iqs - 64*j)/2;  // ir is in 0...28 in steps of 4
+    const int is = 2*j;             // is is in 0...6 in steps of 2
+
+    __global const float   * y = yy + 64*j + ir;
+    __global const uint8_t * q = x[ib].qs + 32*j + ir;
+
+    const float dall = vload_half(0, &x[ib].d);
+    const float dmin = vload_half(0, &x[ib].dmin);
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[ib].scales, &sc, &m);
+    const float d1 = dall * sc;
+    const float m1 = dmin * m;
+    get_scale_min_k4(is + 1, x[ib].scales, &sc, &m);
+    const float d2 = dall * sc;
+    const float m2 = dmin * m;
+
+    float sum = 0;
+    for (int k = 0; k < 4; ++k) {
+        sum += y[k +  0] * (d1 * (q[k] & 0xF) - m1);
+        sum += y[k + 32] * (d2 * (q[k] >>  4) - m2);
+    }
+
+    *result = sum;
+}
+
+void vec_dot_q5_K(__global const struct block_q5_K* x, const int ib, const int iqs, const __global float *yy, float *result) {
+
+    const int j  = iqs / 64;
+    const int ir = (iqs - 64*j)/2;
+    const int is = 2*j;
+
+    __global const float   * y  = yy + 64*j + ir;
+    __global const uint8_t * ql = x[ib].qs + 32*j + ir;
+    __global const uint8_t * qh = x[ib].qh + ir;
+
+    const float dall = vload_half(0, &x[ib].d);
+    const float dmin = vload_half(0, &x[ib].dmin);
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[ib].scales, &sc, &m);
+    const float d1 = dall * sc;
+    const float m1 = dmin * m;
+    get_scale_min_k4(is + 1, x[ib].scales, &sc, &m);
+    const float d2 = dall * sc;
+    const float m2 = dmin * m;
+
+    uint8_t hm  = 1 << is;
+    float sum = 0;
+    for (int k = 0; k < 4; ++k) {
+        sum += y[k +  0] * (d1 * ((ql[k] & 0xF) + (qh[k] & hm ? 16 : 0)) - m1);
+    }
+    hm <<= 1;
+    for (int k = 0; k < 4; ++k) {
+        sum += y[k + 32] * (d2 * ((ql[k] >>  4) + (qh[k] & hm ? 16 : 0)) - m2);
+    }
+    *result = sum;
+
+}
+
+void vec_dot_q6_K(__global const struct block_q6_K* x, const int ib, const int iqs, const __global float *yy, float *result) {
+
+
+    const int ip = iqs / 128;        // 0 or 1
+    const int il = (iqs - 128*ip)/8; // 0...15
+    const int is = 8*ip;
+
+    __global const float * y = yy + 128*ip + il;
+
+    const float d = vload_half(0, &x[ib].d);
+
+    __global const uint8_t * ql = x[ib].ql + 64*ip + il;
+    __global const uint8_t * qh = x[ib].qh + 32*ip + il;
+    __global const int8_t  * sc = x[ib].scales + is;
+
+    *result = y[  0] * d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh[ 0] >> 0) & 3) << 4)) - 32)
+           + y[ 32] * d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh[ 0] >> 2) & 3) << 4)) - 32)
+           + y[ 64] * d * sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh[ 0] >> 4) & 3) << 4)) - 32)
+           + y[ 96] * d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh[ 0] >> 6) & 3) << 4)) - 32)
+           + y[ 16] * d * sc[1] * ((int8_t)((ql[16] & 0xF) | (((qh[16] >> 0) & 3) << 4)) - 32)
+           + y[ 48] * d * sc[3] * ((int8_t)((ql[48] & 0xF) | (((qh[16] >> 2) & 3) << 4)) - 32)
+           + y[ 80] * d * sc[5] * ((int8_t)((ql[16]  >> 4) | (((qh[16] >> 4) & 3) << 4)) - 32)
+           + y[112] * d * sc[7] * ((int8_t)((ql[48]  >> 4) | (((qh[16] >> 6) & 3) << 4)) - 32);
+
+}
+
 );
 
+
 std::string dequant_template = MULTILINE_QUOTE(
 __kernel void KERNEL_NAME(__global X_TYPE* x, __global float* y) {
     const int i = get_group_id(0)*get_local_size(0) + get_local_id(0)*2;
@@ -160,7 +506,7 @@ __kernel void KERNEL_NAME(__global X_TYPE* x, __global float* y) {
 std::string dequant_mul_mat_vec_template = MULTILINE_QUOTE(
 __kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float* y, __global float* dst, const int ncols) {
     const int block_size = get_local_size(0);
-    const int row = get_global_id(0) / block_size;
+    const int row = get_group_id(0);
     const int tid = get_local_id(0);
 
     const uint qk = QUANT_K;
@@ -199,6 +545,45 @@ __kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float
 }
 );
 
+std::string dequant_mul_mat_vec_k_template = MULTILINE_QUOTE(
+__kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float* y, __global float* dst, const int ncols) {
+    const int block_size = get_local_size(0);
+    const int row = get_group_id(0);
+    const int tid = get_local_id(0);
+
+    const int iter_stride = 256;
+    const int vals_per_iter = iter_stride / block_size;
+    const int num_blocks_per_row = ncols / 256;
+    const int ib0 = row*num_blocks_per_row;
+
+    tmp[tid] = 0;
+
+    for (int i = 0; i < ncols; i += iter_stride) {
+        const int col = i + vals_per_iter*tid;
+        const int ib = ib0 + col/256; // x block index
+        const int iqs = col%256; // x quant index
+        const int iybs = col - col%256; // y block start index
+
+        // dequantize
+        float v;
+        DOT_KERNEL(x, ib, iqs, y + iybs, &v);
+        tmp[tid] += v;
+    }
+
+    // sum up partial sums and write back result
+    barrier(CLK_LOCAL_MEM_FENCE);
+    for (int s=block_size/2; s>0; s>>=1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    if (tid == 0) {
+        dst[row] = tmp[0];
+    }
+}
+);
+
 std::string mul_template = MULTILINE_QUOTE(
 __kernel void KERNEL_NAME(__global TYPE* x, const int x_offset, __global TYPE* y, const int y_offset, __global TYPE* dst, const int dst_offset, const int ky) {
     const int i = get_group_id(0)*get_local_size(0) + get_local_id(0);
@@ -260,6 +645,18 @@ std::array<std::string, 2> mul_str_values = {
     "mul_f32", "float"
 };
 
+std::array<std::string, 3> dmmv_k_str_keys = {
+    "KERNEL_NAME", "X_TYPE", "DOT_KERNEL"
+};
+
+std::array<std::string, 15> dmmv_k_str_values = {
+    "dequantize_mul_mat_vec_q2_K", "struct block_q2_K", "vec_dot_q2_K",
+    "dequantize_mul_mat_vec_q3_K", "struct block_q3_K", "vec_dot_q3_K",
+    "dequantize_mul_mat_vec_q4_K", "struct block_q4_K", "vec_dot_q4_K",
+    "dequantize_mul_mat_vec_q5_K", "struct block_q5_K", "vec_dot_q5_K",
+    "dequantize_mul_mat_vec_q6_K", "struct block_q6_K", "vec_dot_q6_K",
+};
+
 std::string& replace(std::string& s, const std::string& from, const std::string& to) {
     size_t pos = 0;
     while ((pos = s.find(from, pos)) != std::string::npos) {
@@ -289,6 +686,14 @@ std::string generate_kernels() {
         }
         src << mul_kernel << '\n';
     }
+    for (size_t i = 0; i < dmmv_k_str_values.size(); i += dmmv_k_str_keys.size()) {
+        std::string dmmv_k_kernel = dequant_mul_mat_vec_k_template;
+        for (size_t j = 0; j < dmmv_k_str_keys.size(); j++) {
+            replace(dmmv_k_kernel, dmmv_k_str_keys[j], dmmv_k_str_values[i + j]);
+        }
+        src << dmmv_k_kernel << '\n';
+    }
+
     return src.str();
 }
 
@@ -300,6 +705,8 @@ static cl_program program;
 static cl_kernel convert_row_f16_cl;
 static cl_kernel dequantize_row_q4_0_cl, dequantize_row_q4_1_cl, dequantize_row_q5_0_cl, dequantize_row_q5_1_cl, dequantize_row_q8_0_cl;
 static cl_kernel dequantize_mul_mat_vec_q4_0_cl, dequantize_mul_mat_vec_q4_1_cl, dequantize_mul_mat_vec_q5_0_cl, dequantize_mul_mat_vec_q5_1_cl, dequantize_mul_mat_vec_q8_0_cl, convert_mul_mat_vec_f16_cl;
+static cl_kernel dequantize_block_q2_k_cl, dequantize_block_q3_k_cl, dequantize_block_q4_k_cl, dequantize_block_q5_k_cl, dequantize_block_q6_k_cl;
+static cl_kernel dequantize_mul_mat_vec_q2_K_cl, dequantize_mul_mat_vec_q3_K_cl, dequantize_mul_mat_vec_q4_K_cl, dequantize_mul_mat_vec_q5_K_cl, dequantize_mul_mat_vec_q6_K_cl;
 static cl_kernel mul_f32_cl;
 static bool fp16_support;
 
@@ -529,6 +936,12 @@ void ggml_cl_init(void) {
     CL_CHECK((dequantize_row_q5_0_cl = clCreateKernel(program, "dequantize_row_q5_0", &err), err));
     CL_CHECK((dequantize_row_q5_1_cl = clCreateKernel(program, "dequantize_row_q5_1", &err), err));
     CL_CHECK((dequantize_row_q8_0_cl = clCreateKernel(program, "dequantize_row_q8_0", &err), err));
+    CL_CHECK((dequantize_row_q8_0_cl = clCreateKernel(program, "dequantize_row_q8_0", &err), err));
+    CL_CHECK((dequantize_block_q2_k_cl = clCreateKernel(program, "dequantize_block_q2_K", &err), err));
+    CL_CHECK((dequantize_block_q3_k_cl = clCreateKernel(program, "dequantize_block_q3_K", &err), err));
+    CL_CHECK((dequantize_block_q4_k_cl = clCreateKernel(program, "dequantize_block_q4_K", &err), err));
+    CL_CHECK((dequantize_block_q5_k_cl = clCreateKernel(program, "dequantize_block_q5_K", &err), err));
+    CL_CHECK((dequantize_block_q6_k_cl = clCreateKernel(program, "dequantize_block_q6_K", &err), err));
 
     // dequant mul mat kernel
     CL_CHECK((dequantize_mul_mat_vec_q4_0_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q4_0", &err), err));
@@ -537,6 +950,11 @@ void ggml_cl_init(void) {
     CL_CHECK((dequantize_mul_mat_vec_q5_1_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q5_1", &err), err));
     CL_CHECK((dequantize_mul_mat_vec_q8_0_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q8_0", &err), err));
     CL_CHECK((convert_mul_mat_vec_f16_cl = clCreateKernel(program, "convert_mul_mat_vec_f16", &err), err));
+    CL_CHECK((dequantize_mul_mat_vec_q2_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q2_K", &err), err));
+    CL_CHECK((dequantize_mul_mat_vec_q3_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q3_K", &err), err));
+    CL_CHECK((dequantize_mul_mat_vec_q4_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q4_K", &err), err));
+    CL_CHECK((dequantize_mul_mat_vec_q5_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q5_K", &err), err));
+    CL_CHECK((dequantize_mul_mat_vec_q6_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q6_K", &err), err));
 
     // mul kernel
     CL_CHECK((mul_f32_cl = clCreateKernel(program, "mul_f32", &err), err));
@@ -554,6 +972,16 @@ static cl_kernel* ggml_get_to_fp32_cl(ggml_type type) {
             return &dequantize_row_q5_1_cl;
         case GGML_TYPE_Q8_0:
             return &dequantize_row_q8_0_cl;
+        case GGML_TYPE_Q2_K:
+            return &dequantize_block_q2_k_cl;
+        case GGML_TYPE_Q3_K:
+            return &dequantize_block_q3_k_cl;
+        case GGML_TYPE_Q4_K:
+            return &dequantize_block_q4_k_cl;
+        case GGML_TYPE_Q5_K:
+            return &dequantize_block_q5_k_cl;
+        case GGML_TYPE_Q6_K:
+            return &dequantize_block_q6_k_cl;
         case GGML_TYPE_F16:
             return &convert_row_f16_cl;
         default:
@@ -561,6 +989,50 @@ static cl_kernel* ggml_get_to_fp32_cl(ggml_type type) {
     }
 }
 
+static size_t ggml_cl_global_denom(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+            return 1;
+        case GGML_TYPE_Q2_K:
+        case GGML_TYPE_Q3_K:
+            return 4;
+        case GGML_TYPE_Q4_K:
+            return 8;
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
+            return 4;
+        case GGML_TYPE_F16:
+        default:
+            return 1;
+    }
+}
+
+static size_t ggml_cl_local_size(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+            return 0;
+        case GGML_TYPE_Q2_K:
+        case GGML_TYPE_Q3_K:
+            return 64;
+        case GGML_TYPE_Q4_K:
+            return 32;
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
+            return 64;
+        case GGML_TYPE_F16:
+        default:
+            return 0;
+    }
+}
+
 static cl_kernel* ggml_get_dequantize_mul_mat_vec_cl(ggml_type type) {
     switch (type) {
         case GGML_TYPE_Q4_0:
@@ -575,6 +1047,16 @@ static cl_kernel* ggml_get_dequantize_mul_mat_vec_cl(ggml_type type) {
             return &dequantize_mul_mat_vec_q8_0_cl;
         case GGML_TYPE_F16:
             return &convert_mul_mat_vec_f16_cl;
+        case GGML_TYPE_Q2_K:
+            return &dequantize_mul_mat_vec_q2_K_cl;
+        case GGML_TYPE_Q3_K:
+            return &dequantize_mul_mat_vec_q3_K_cl;
+        case GGML_TYPE_Q4_K:
+            return &dequantize_mul_mat_vec_q4_K_cl;
+        case GGML_TYPE_Q5_K:
+            return &dequantize_mul_mat_vec_q5_K_cl;
+        case GGML_TYPE_Q6_K:
+            return &dequantize_mul_mat_vec_q6_K_cl;
         default:
             return nullptr;
     }
@@ -1017,6 +1499,9 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
     cl_kernel* dmmv = ggml_get_dequantize_mul_mat_vec_cl(type);
     GGML_ASSERT(to_fp32_cl != nullptr);
 
+    const size_t global_denom = ggml_cl_global_denom(type);
+    const size_t local = ggml_cl_local_size(type);
+
     size_t ev_idx = 0;
     std::vector<cl_event> events;
 
@@ -1049,10 +1534,10 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
                 CL_CHECK(clEnqueueNDRangeKernel(queue, *dmmv, 1, NULL, &global, &local, events.size() - 1, events.data(), events.data() + ev_idx++));
             } else { // general dequantization kernel + CLBlast matrix matrix multiplication
                 // convert src0 to fp32 on device
-                const size_t global = x_ne;
+                const size_t global = x_ne / global_denom;
                 CL_CHECK(clSetKernelArg(*to_fp32_cl, 0, sizeof(cl_mem), &d_Q));
                 CL_CHECK(clSetKernelArg(*to_fp32_cl, 1, sizeof(cl_mem), &d_X));
-                CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, NULL, &global, NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
+                CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, NULL, &global, local > 0 ? &local : NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
 
                 // copy src1 to device
                 CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, NULL));

ggml.h

@@ -296,6 +296,7 @@ extern "C" {
         GGML_OP_SUM_ROWS,
         GGML_OP_MEAN,
         GGML_OP_REPEAT,
+        GGML_OP_REPEAT_BACK,
         GGML_OP_ABS,
         GGML_OP_SGN,
         GGML_OP_NEG,
@@ -309,6 +310,7 @@ extern "C" {
         GGML_OP_RMS_NORM_BACK,
 
         GGML_OP_MUL_MAT,
+        GGML_OP_OUT_PROD,
 
         GGML_OP_SCALE,
         GGML_OP_SET,
@@ -324,6 +326,7 @@ extern "C" {
         GGML_OP_DIAG_MASK_INF,
         GGML_OP_DIAG_MASK_ZERO,
         GGML_OP_SOFT_MAX,
+        GGML_OP_SOFT_MAX_BACK,
         GGML_OP_ROPE,
         GGML_OP_ROPE_BACK,
         GGML_OP_ALIBI,
@@ -333,10 +336,14 @@ extern "C" {
 
         GGML_OP_FLASH_ATTN,
         GGML_OP_FLASH_FF,
+        GGML_OP_FLASH_ATTN_BACK,
 
         GGML_OP_MAP_UNARY,
         GGML_OP_MAP_BINARY,
 
+        GGML_OP_CROSS_ENTROPY_LOSS,
+        GGML_OP_CROSS_ENTROPY_LOSS_BACK,
+
         GGML_OP_COUNT,
     };
 
@@ -478,6 +485,7 @@ extern "C" {
 
     GGML_API bool ggml_is_transposed(const struct ggml_tensor * tensor);
     GGML_API bool ggml_is_contiguous(const struct ggml_tensor * tensor);
+    GGML_API bool ggml_is_permuted  (const struct ggml_tensor * tensor);
 
     // use this to compute the memory overhead of a tensor
     GGML_API size_t ggml_tensor_overhead(void);
@@ -574,6 +582,11 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_add1_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     GGML_API struct ggml_tensor * ggml_acc(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -645,6 +658,11 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_repeat_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     GGML_API struct ggml_tensor * ggml_abs(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -698,14 +716,22 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
-    // A: m rows, n columns
-    // B: p rows, n columns (i.e. we transpose it internally)
+    // A: n columns, m rows
+    // B: n columns, p rows  (i.e. we transpose it internally)
     // result is m columns, p rows
     GGML_API struct ggml_tensor * ggml_mul_mat(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // A: m columns, n rows,
+    // B: p columns, n rows,
+    // result is m columns, p rows
+    GGML_API struct ggml_tensor * ggml_out_prod(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     //
     // operations on tensors without backpropagation
     //
@@ -916,6 +942,17 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_soft_max_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_soft_max_back_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     // rotary position embedding
     // if mode & 1 == 1, skip n_past elements
     // if mode & 2 == 1, GPT-NeoX style
@@ -982,6 +1019,14 @@ extern "C" {
             struct ggml_tensor  * v,
             bool                  masked);
 
+    GGML_API struct ggml_tensor * ggml_flash_attn_back(
+           struct ggml_context * ctx,
+           struct ggml_tensor  * q,
+           struct ggml_tensor  * k,
+           struct ggml_tensor  * v,
+           struct ggml_tensor  * d,
+           bool                  masked);
+
     GGML_API struct ggml_tensor * ggml_flash_ff(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1005,6 +1050,19 @@ extern "C" {
             struct ggml_tensor          * b,
                    ggml_binary_op_f32_t   fun);
 
+    // loss function
+
+    GGML_API struct ggml_tensor * ggml_cross_entropy_loss(
+            struct ggml_context         * ctx,
+            struct ggml_tensor          * a,
+            struct ggml_tensor          * b);
+
+    GGML_API struct ggml_tensor * ggml_cross_entropy_loss_back(
+            struct ggml_context         * ctx,
+            struct ggml_tensor          * a,
+            struct ggml_tensor          * b,
+            struct ggml_tensor          * c);
+
     //
     // automatic differentiation
     //
@@ -1099,6 +1157,8 @@ extern "C" {
         struct {
             int n_iter;
 
+            float sched; // schedule multiplier (fixed, decay or warmup)
+            float decay; // weight decay for AdamW, use 0.0f to disable
             float alpha; // learning rate
             float beta1;
             float beta2;
@@ -1123,6 +1183,49 @@ extern "C" {
         } lbfgs;
     };
 
+    struct ggml_opt_context {
+        struct ggml_context * ctx;
+        struct ggml_opt_params params;
+
+        int iter;
+        int64_t nx; // number of parameter elements
+
+        bool just_initialized;
+
+        struct {
+            struct ggml_tensor * x;  // view of the parameters
+            struct ggml_tensor * g1; // gradient
+            struct ggml_tensor * g2; // gradient squared
+            struct ggml_tensor * m;  // first moment
+            struct ggml_tensor * v;  // second moment
+            struct ggml_tensor * mh; // first moment hat
+            struct ggml_tensor * vh; // second moment hat
+            struct ggml_tensor * pf; // past function values
+            float fx_best;
+            float fx_prev;
+            int n_no_improvement;
+        } adam;
+
+        struct {
+            struct ggml_tensor * x;    // current parameters
+            struct ggml_tensor * xp;   // previous parameters
+            struct ggml_tensor * g;    // current gradient
+            struct ggml_tensor * gp;   // previous gradient
+            struct ggml_tensor * d;    // search direction
+            struct ggml_tensor * pf;   // past function values
+            struct ggml_tensor * lmal; // the L-BFGS memory alpha
+            struct ggml_tensor * lmys; // the L-BFGS memory ys
+            struct ggml_tensor * lms;  // the L-BFGS memory s
+            struct ggml_tensor * lmy;  // the L-BFGS memory y
+            float fx_best;
+            float step;
+            int j;
+            int k;
+            int end;
+            int n_no_improvement;
+        } lbfgs;
+    };
+
     GGML_API struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type);
 
     // optimize the function defined by the tensor f
@@ -1131,6 +1234,27 @@ extern "C" {
             struct ggml_opt_params params,
             struct ggml_tensor * f);
 
+    // initialize optimizer context
+    GGML_API void ggml_opt_init(
+            struct ggml_context * ctx,
+            struct ggml_opt_context * opt,
+            struct ggml_opt_params params,
+            int64_t nx);
+
+    // continue optimizing the function defined by the tensor f
+    GGML_API enum ggml_opt_result ggml_opt_resume(
+            struct ggml_context * ctx,
+            struct ggml_opt_context * opt,
+            struct ggml_tensor * f);
+
+    // continue optimizing the function defined by the tensor f
+    GGML_API enum ggml_opt_result ggml_opt_resume_g(
+            struct ggml_context * ctx,
+            struct ggml_opt_context * opt,
+            struct ggml_tensor * f,
+            struct ggml_cgraph * gf,
+            struct ggml_cgraph * gb);
+
     //
     // quantization
     //

llama.cpp

@@ -40,6 +40,10 @@
 #include <sstream>
 #include <numeric>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 #define LLAMA_USE_SCRATCH
 #define LLAMA_MAX_SCRATCH_BUFFERS 16
 
@@ -165,6 +169,11 @@ struct llama_kv_cache {
         if (ctx) {
             ggml_free(ctx);
         }
+
+#ifdef GGML_USE_CUBLAS
+        ggml_cuda_free_data(k);
+        ggml_cuda_free_data(v);
+#endif // GGML_USE_CUBLAS
     }
 };
 
@@ -210,6 +219,7 @@ struct llama_model {
         for (size_t i = 0; i < tensors_by_name.size(); ++i) {
             ggml_cuda_free_data(tensors_by_name[i].second);
         }
+        ggml_cuda_free_scratch();
 #elif defined(GGML_USE_CLBLAST)
         for (size_t i = 0; i < tensors_by_name.size(); ++i) {
             ggml_cl_free_data(tensors_by_name[i].second);
@@ -867,7 +877,8 @@ static bool kv_cache_init(
         const struct llama_hparams & hparams,
              struct llama_kv_cache & cache,
                          ggml_type   wtype,
-                               int   n_ctx) {
+                               int   n_ctx,
+                               int   n_gpu_layers) {
     const int n_embd  = hparams.n_embd;
     const int n_layer = hparams.n_layer;
 
@@ -893,6 +904,15 @@ static bool kv_cache_init(
     ggml_set_name(cache.k, "cache_k");
     ggml_set_name(cache.v, "cache_v");
 
+#ifdef GGML_USE_CUBLAS
+    if (n_gpu_layers > n_layer + 1) {
+        ggml_cuda_assign_buffers_no_scratch(cache.v);
+    }
+    if (n_gpu_layers > n_layer + 2) {
+        ggml_cuda_assign_buffers_no_scratch(cache.k);
+    }
+#endif // GGML_USE_CUBLAS
+
     return true;
 }
 
@@ -903,6 +923,7 @@ struct llama_context_params llama_context_default_params() {
         /*.gpu_layers                  =*/ 0,
         /*.main_gpu                    =*/ 0,
         /*.tensor_split                =*/ {0},
+        /*.low_vram                    =*/ false,
         /*.seed                        =*/ -1,
         /*.f16_kv                      =*/ true,
         /*.logits_all                  =*/ false,
@@ -1011,6 +1032,7 @@ static void llama_model_load_internal(
         int n_gpu_layers,
         int main_gpu,
         const float * tensor_split,
+        bool low_vram,
         ggml_type memory_type,
         bool use_mmap,
         bool use_mlock,
@@ -1036,6 +1058,12 @@ static void llama_model_load_internal(
             case 40: model.type = e_model::MODEL_13B; break;
             case 60: model.type = e_model::MODEL_30B; break;
             case 80: model.type = e_model::MODEL_65B; break;
+            default:
+                {
+                    if (hparams.n_layer < 32) {
+                        model.type = e_model::MODEL_7B;
+                    }
+                } break;
         }
 
         hparams.n_ctx = n_ctx;
@@ -1131,18 +1159,34 @@ static void llama_model_load_internal(
         ml->ggml_ctx = ctx;
 
         model.tok_embeddings = ml->get_tensor("tok_embeddings.weight", {n_embd, n_vocab}, GGML_BACKEND_CPU);
-        model.norm           = ml->get_tensor("norm.weight",           {n_embd},          GGML_BACKEND_CPU);
 
         // "output" tensor
         {
+            ggml_backend backend_norm;
             ggml_backend backend_output;
             if (n_gpu_layers > int(n_layer)) { // NOLINT
+                // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
+                // on Windows however this is detrimental unless everything is on the GPU
+#ifndef _WIN32
+                backend_norm = low_vram ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
+#else
+                backend_norm = low_vram || n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
+#endif // _WIN32
+
                 backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT;
             } else {
+                backend_norm = GGML_BACKEND_CPU;
                 backend_output = GGML_BACKEND_CPU;
             }
 
+            model.norm   = ml->get_tensor("norm.weight",   {n_embd},          backend_norm);
             model.output = ml->get_tensor("output.weight", {n_embd, n_vocab}, backend_output);
+            if (backend_norm == GGML_BACKEND_GPU) {
+                vram_weights += ggml_nbytes(model.norm);
+            }
+            if (backend_output == GGML_BACKEND_GPU_SPLIT) {
+                vram_weights += ggml_nbytes(model.output);
+            }
         }
 
         const int i_gpu_start = n_layer - n_gpu_layers;
@@ -1200,23 +1244,49 @@ static void llama_model_load_internal(
                 mem_required / 1024.0 / 1024.0, mem_required_state / 1024.0 / 1024.0);
 
         (void) vram_scratch;
+        (void) n_batch;
 #ifdef GGML_USE_CUBLAS
-        vram_scratch = n_batch * MB;
-        ggml_cuda_set_scratch_size(vram_scratch);
-        if (n_gpu_layers > 0) {
-            fprintf(stderr, "%s: allocating batch_size x 1 MB = %ld MB VRAM for the scratch buffer\n",
-                    __func__, vram_scratch / MB);
+        if (low_vram) {
+            fprintf(stderr, "%s: not allocating a VRAM scratch buffer due to low VRAM option\n", __func__);
+            ggml_cuda_set_scratch_size(0); // disable scratch
+        } else {
+            vram_scratch = n_batch * MB;
+            ggml_cuda_set_scratch_size(vram_scratch);
+            if (n_gpu_layers > 0) {
+                fprintf(stderr, "%s: allocating batch_size x 1 MB = %ld MB VRAM for the scratch buffer\n",
+                        __func__, vram_scratch / MB);
+            }
         }
 #endif // GGML_USE_CUBLAS
 #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 layers to GPU\n", __func__, n_gpu);
+        fprintf(stderr, "%s: offloading %d repeating layers to GPU\n", __func__, n_gpu);
         if (n_gpu_layers > (int) hparams.n_layer) {
-            fprintf(stderr, "%s: offloading output layer to GPU\n", __func__);
+            fprintf(stderr, "%s: offloading non-repeating layers to GPU\n", __func__);
         }
+        size_t vram_kv_cache = 0;
+        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__);
+            } else {
+                fprintf(stderr, "%s: offloading v cache to GPU\n", __func__);
+                vram_kv_cache += MEM_REQ_KV_SELF().at(model.type) / 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__);
+            } else {
+                fprintf(stderr, "%s: offloading k cache to GPU\n", __func__);
+                vram_kv_cache += MEM_REQ_KV_SELF().at(model.type) / 2;
+            }
+        }
+        const int max_offloadable_layers = low_vram ? hparams.n_layer + 1 : hparams.n_layer + 3;
+        fprintf(stderr, "%s: offloaded %d/%d layers to GPU\n",
+                __func__, std::min(n_gpu_layers, max_offloadable_layers), hparams.n_layer + 3);
         fprintf(stderr, "%s: total VRAM used: %zu MB\n",
-                __func__, (vram_weights + vram_scratch + MB - 1) / MB); // round up
+                __func__, (vram_weights + vram_scratch + vram_kv_cache + MB - 1) / MB); // round up
 #else
         (void) n_gpu_layers;
 #endif
@@ -1227,6 +1297,7 @@ static void llama_model_load_internal(
         model.tensors_by_name.emplace_back(lt.name, lt.ggml_tensor);
     }
 
+    (void) tensor_split;
 #if defined(GGML_USE_CUBLAS)
     {
         ggml_cuda_set_tensor_split(tensor_split);
@@ -1254,6 +1325,7 @@ static bool llama_model_load(
         int n_gpu_layers,
         int main_gpu,
         float * tensor_split,
+        bool low_vram,
         ggml_type memory_type,
         bool use_mmap,
         bool use_mlock,
@@ -1261,7 +1333,7 @@ static bool llama_model_load(
         llama_progress_callback progress_callback,
         void *progress_callback_user_data) {
     try {
-        llama_model_load_internal(fname, lctx, n_ctx, n_batch, n_gpu_layers, main_gpu, tensor_split, memory_type,
+        llama_model_load_internal(fname, lctx, n_ctx, n_batch, n_gpu_layers, main_gpu, tensor_split, low_vram, memory_type,
                                   use_mmap, use_mlock, vocab_only, progress_callback, progress_callback_user_data);
         return true;
     } catch (const std::exception & err) {
@@ -1337,12 +1409,33 @@ static bool llama_eval_internal(
     const int i_gpu_start = n_layer - n_gpu_layers;
     (void) i_gpu_start;
 
+    // offload functions set the tensor output backend to GPU
+    // tensors are GPU-accelerated if any input or the output has been offloaded
+    //
+    // with the low VRAM option VRAM scratch is disabled in llama_load_model_internal
+    // in that case ggml_cuda_assign_buffers has no effect
+    offload_func_t offload_func_nr = llama_nop; // nr = non-repeating
+    offload_func_t offload_func_kq = llama_nop;
+    offload_func_t offload_func_v  = llama_nop;
+
+#ifdef GGML_USE_CUBLAS
+        if (n_gpu_layers > n_layer) {
+            offload_func_nr = ggml_cuda_assign_buffers;
+        }
+        if (n_gpu_layers > n_layer + 1) {
+            offload_func_v  = ggml_cuda_assign_buffers;
+        }
+        if (n_gpu_layers > n_layer + 2) {
+            offload_func_kq = ggml_cuda_assign_buffers;
+        }
+#endif // GGML_USE_CUBLAS
+
     for (int il = 0; il < n_layer; ++il) {
         offload_func_t offload_func = llama_nop;
 
 #ifdef GGML_USE_CUBLAS
         if (il >= i_gpu_start) {
-            offload_func = ggml_cuda_assign_buffers; // sets the output backend to GPU
+            offload_func = ggml_cuda_assign_buffers;
         }
 #endif // GGML_USE_CUBLAS
 
@@ -1365,31 +1458,42 @@ static bool llama_eval_internal(
         // self-attention
         {
             // compute Q and K and RoPE them
-            struct ggml_tensor * tmpq = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
-            // offload_func(tmpq);
-            ggml_set_name(tmpq, "tmpq");
-
             struct ggml_tensor * tmpk = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
-            // offload_func(tmpk);
+            offload_func_kq(tmpk);
             ggml_set_name(tmpk, "tmpk");
 
+            struct ggml_tensor * tmpq = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+            offload_func_kq(tmpq);
+            ggml_set_name(tmpq, "tmpq");
+
             struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd/n_head, n_head, N), n_past, n_rot, 0);
+            offload_func_kq(Kcur);
             ggml_set_name(Kcur, "Kcur");
 
             struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd/n_head, n_head, N), n_past, n_rot, 0);
+            offload_func_kq(Qcur);
             ggml_set_name(Qcur, "Qcur");
 
             // store key and value to memory
             {
                 // compute the transposed [N, n_embd] V matrix
-                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wv, cur), n_embd, N));
+
+                struct ggml_tensor * tmpv = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+                offload_func_v(tmpv);
+                ggml_set_name(tmpv, "tmpv");
+
+                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, tmpv, n_embd, N));
+                offload_func_v(Vcur);
                 ggml_set_name(Vcur, "Vcur");
 
                 struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past));
+                offload_func_kq(k);
                 ggml_set_name(k, "k");
+
                 struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd,
                         (   n_ctx)*ggml_element_size(kv_self.v),
                         (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v));
+                offload_func_v(v);
                 ggml_set_name(v, "v");
 
                 // important: storing RoPE-ed version of K in the KV cache!
@@ -1401,6 +1505,7 @@ static bool llama_eval_internal(
                 ggml_permute(ctx0,
                         Qcur,
                         0, 2, 1, 3);
+            offload_func_kq(Q);
             ggml_set_name(Q, "Q");
 
             struct ggml_tensor * K =
@@ -1409,10 +1514,12 @@ static bool llama_eval_internal(
                             ggml_view_1d(ctx0, kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kv_self.k)*n_embd),
                             n_embd/n_head, n_head, n_past + N),
                         0, 2, 1, 3);
+            offload_func_kq(K);
             ggml_set_name(K, "K");
 
             // K * Q
             struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+            offload_func_kq(KQ);
             ggml_set_name(KQ, "KQ");
 
             // KQ_scaled = KQ / sqrt(n_embd/n_head)
@@ -1421,14 +1528,17 @@ static bool llama_eval_internal(
 
             // KQ_scaled shape [n_past + N, N, n_head, 1]
             struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, KQ_scale);
+            offload_func_kq(KQ_scaled);
             ggml_set_name(KQ_scaled, "KQ_scaled");
 
             // KQ_masked = mask_past(KQ_scaled)
             struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
+            offload_func_kq(KQ_masked);
             ggml_set_name(KQ_masked, "KQ_masked");
 
             // KQ = soft_max(KQ_masked)
             struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
+            offload_func_v(KQ_soft_max);
             ggml_set_name(KQ_soft_max, "KQ_soft_max");
 
             // split cached V into n_head heads
@@ -1438,10 +1548,12 @@ static bool llama_eval_internal(
                         n_ctx*ggml_element_size(kv_self.v),
                         n_ctx*ggml_element_size(kv_self.v)*n_embd/n_head,
                         il*n_ctx*ggml_element_size(kv_self.v)*n_embd);
+            offload_func_v(V);
             ggml_set_name(V, "V");
 
 #if 1
             struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+            offload_func_v(KQV);
             ggml_set_name(KQV, "KQV");
 #else
             // make V contiguous in memory to speed up the matmul, however we waste time on the copy
@@ -1453,12 +1565,14 @@ static bool llama_eval_internal(
 
             // KQV_merged = KQV.permute(0, 2, 1, 3)
             struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+            offload_func_v(KQV_merged);
             ggml_set_name(KQV_merged, "KQV_merged");
 
             // cur = KQV_merged.contiguous().view(n_embd, N)
             cur = ggml_cpy(ctx0,
                     KQV_merged,
                     ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
+            offload_func_v(cur);
             ggml_set_name(cur, "KQV_merged_contiguous");
 
             // projection (no bias)
@@ -1470,7 +1584,6 @@ static bool llama_eval_internal(
         }
 
         lctx.use_buf(ctx0, 1);
-        //ggml_cuda_set_scratch(1);
 
         struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA);
         offload_func(inpFF);
@@ -1528,32 +1641,24 @@ static bool llama_eval_internal(
     }
 
     lctx.use_buf(ctx0, 0);
-    //ggml_cuda_set_scratch(0);
 
     // used at the end to optionally extract the embeddings
     struct ggml_tensor * embeddings = NULL;
 
-    offload_func_t offload_func = llama_nop;
-
-#ifdef GGML_USE_CUBLAS
-        if (n_gpu_layers > n_layer) {
-            offload_func = ggml_cuda_assign_buffers; // sets the output backend to GPU
-        }
-#endif // GGML_USE_CUBLAS
 
     // norm
     {
         cur = ggml_rms_norm(ctx0, inpL);
-        offload_func(cur);
+        offload_func_nr(cur);
         ggml_set_name(cur, "rms_norm_inpL");
 
         cur = ggml_rms_norm(ctx0, cur);
-        offload_func(cur);
+        offload_func_nr(cur);
         ggml_set_name(cur, "rms_norm_after");
 
         // cur = cur*norm(broadcasted)
         cur = ggml_mul(ctx0, cur, model.norm);
-        offload_func(cur);
+        // offload_func_nr(cur); // TODO CPU + GPU mirrored backend
         ggml_set_name(cur, "result_norm");
 
         embeddings = cur;
@@ -2161,6 +2266,10 @@ llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_tok
         return -log2f(candidate.p) > *mu;
     }));
 
+    if (candidates->size == 0) {
+        candidates->size = 1;
+    }
+
     // Normalize the probabilities of the remaining words
     llama_sample_softmax(ctx, candidates);
 
@@ -2540,8 +2649,8 @@ struct llama_context * llama_init_from_file(
 
     ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32;
 
-    if (!llama_model_load(path_model, *ctx, params.n_ctx, params.n_batch, params.n_gpu_layers,
-                params.main_gpu, params.tensor_split, memory_type, params.use_mmap, params.use_mlock,
+    if (!llama_model_load(path_model, *ctx, params.n_ctx, params.n_batch, params.n_gpu_layers, params.main_gpu,
+                params.tensor_split, params.low_vram, memory_type, params.use_mmap, params.use_mlock,
                 params.vocab_only, params.progress_callback, params.progress_callback_user_data)) {
         fprintf(stderr, "%s: failed to load model\n", __func__);
         llama_free(ctx);
@@ -2550,7 +2659,7 @@ struct llama_context * llama_init_from_file(
 
     // reserve memory for context buffers
     if (!params.vocab_only) {
-        if (!kv_cache_init(ctx->model.hparams, ctx->model.kv_self, memory_type, ctx->model.hparams.n_ctx)) {
+        if (!kv_cache_init(ctx->model.hparams, ctx->model.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_free(ctx);
             return nullptr;
@@ -3287,6 +3396,19 @@ int llama_n_embd(const struct llama_context * ctx) {
     return ctx->model.hparams.n_embd;
 }
 
+int llama_get_vocab(
+        const struct llama_context * ctx,
+        const char * * strings,
+        float  * scores,
+        int capacity) {
+    int n = std::min(capacity, (int) ctx->vocab.id_to_token.size());
+    for (int i = 0; i<n; ++i) {
+        strings[i] = ctx->vocab.id_to_token[i].tok.c_str();
+        scores[i]  = ctx->vocab.id_to_token[i].score;
+    }
+    return n;
+}
+
 float * llama_get_logits(struct llama_context * ctx) {
     return ctx->logits.data();
 }

llama.h

@@ -77,6 +77,7 @@ extern "C" {
         int n_gpu_layers;                      // number of layers to store in VRAM
         int main_gpu;                          // the GPU that is used for scratch and small tensors
         float tensor_split[LLAMA_MAX_DEVICES]; // how to split layers across multiple GPUs
+        bool low_vram;                         // if true, reduce VRAM usage at the cost of performance
         int seed;                              // RNG seed, -1 for random
 
         bool f16_kv;     // use fp16 for KV cache
@@ -220,6 +221,14 @@ extern "C" {
     LLAMA_API int llama_n_ctx  (const struct llama_context * ctx);
     LLAMA_API int llama_n_embd (const struct llama_context * ctx);
 
+    // Get the vocabulary as output parameters.
+    // Returns number of results.
+    LLAMA_API int llama_get_vocab(
+            const struct llama_context * ctx,
+                          const char * * strings,
+                                 float * scores,
+                                   int   capacity);
+
     // Token logits obtained from the last call to llama_eval()
     // The logits for the last token are stored in the last row
     // Can be mutated in order to change the probabilities of the next token
@@ -235,9 +244,9 @@ extern "C" {
     LLAMA_API const char * llama_token_to_str(const struct llama_context * ctx, llama_token token);
 
     // Special tokens
-    LLAMA_API llama_token llama_token_bos();
-    LLAMA_API llama_token llama_token_eos();
-    LLAMA_API llama_token llama_token_nl();
+    LLAMA_API llama_token llama_token_bos();  // beginning-of-sentence
+    LLAMA_API llama_token llama_token_eos();  // end-of-sentence
+    LLAMA_API llama_token llama_token_nl();   // next-line
 
     // Sampling functions
 

pocs/vdot/vdot.cpp

@@ -10,6 +10,10 @@
 
 #include <ggml.h>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 constexpr int kVecSize = 1 << 18;
 
 float drawFromGaussianPdf(std::mt19937& rndm) {

spm-headers/ggml.h

@@ -0,0 +1 @@
+../ggml.h

tests/test-grad0.c

@@ -5,7 +5,7 @@
 #include <stdlib.h>
 #include <assert.h>
 
-#define MAX_NARGS 2
+#define MAX_NARGS 3
 
 #undef MIN
 #undef MAX
@@ -1090,6 +1090,25 @@ int main(int argc, const char ** argv) {
             }
         }
 
+        // cross_entropy_loss
+        {
+            const int nargs = 1;
+
+            int64_t ne2[4];
+            get_random_dims(ne2, 4);
+
+            for (int ndims = 1; ndims <= 3; ++ndims) {
+                x[0] = get_random_tensor(ctx0, ndims, ne2, -1.0f, 1.0f);
+                x[1] = get_random_tensor(ctx0, ndims, ne2, 0.0f, 1.0f);
+                ggml_set_param(ctx0, x[0]);
+
+                struct ggml_tensor * f = ggml_sum(ctx0, ggml_cross_entropy_loss(ctx0, x[0], x[1]));
+
+                check_gradient("cross_entropy_loss", ctx0, x, f, ndims, nargs, 1e-1f, 1e-2f, INFINITY);
+                // finite differences regularly fails!
+            }
+        }
+
         // rope
         {
             const int nargs = 1;
@@ -1124,6 +1143,45 @@ int main(int argc, const char ** argv) {
             }
         }
 
+        // flash_attn
+        {
+            const int nargs = 3;
+
+            int64_t ne2[4];
+
+            get_random_dims(ne2, 4);
+            int64_t D = ne2[0];
+            int64_t N = ne2[1];
+            int64_t M = ne2[2] + N;
+            int64_t B = ne2[3];
+
+            for (int masked = 0; masked <= 1; ++masked) {
+                for (int ndims = 2; ndims <= 4; ++ndims) {
+                    int64_t neq[4] = { D, N, B, ne[3] };
+                    int64_t nek[4] = { D, M, B, ne[3] };
+                    int64_t nev[4] = { M, D, B, ne[3] };
+                    if (ndims == 2) {
+                        neq[2] = 1; neq[3] = 1;
+                        nek[2] = 1; nek[3] = 1;
+                        nev[2] = 1; nev[3] = 1;
+                    } else if (ndims == 3) {
+                        neq[3] = 1;
+                        nek[3] = 1;
+                        nev[3] = 1;
+                    }
+                    x[0] = get_random_tensor(ctx0, ndims, neq, -0.1250f, 0.1250f);
+                    x[1] = get_random_tensor(ctx0, ndims, nek, -0.1250f, 0.1250f);
+                    x[2] = get_random_tensor(ctx0, ndims, nev, -0.1250f, 0.1250f);
+                    ggml_set_param(ctx0, x[0]);
+                    ggml_set_param(ctx0, x[1]);
+                    ggml_set_param(ctx0, x[2]);
+
+                    struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
+
+                    check_gradient("flash_attn", ctx0, x, f, ndims, nargs, 1.5e-4f, INFINITY, 3.5f);
+                }
+            }
+        }
         ggml_free(ctx0);
     }
 

tests/test-quantize-fns.cpp

@@ -9,12 +9,15 @@
 #include <string>
 #include <vector>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
 
-const float MAX_QUANTIZATION_REFERENCE_ERROR = 0.0001;
-const float MAX_QUANTIZATION_TOTAL_ERROR = 0.002;
-const float MAX_QUANTIZATION_TOTAL_ERROR_2BITS = 0.0075;
-const float MAX_QUANTIZATION_TOTAL_ERROR_3BITS = 0.0040;
-const float MAX_DOT_PRODUCT_ERROR = 0.02;
+const float MAX_QUANTIZATION_REFERENCE_ERROR = 0.0001f;
+const float MAX_QUANTIZATION_TOTAL_ERROR = 0.002f;
+const float MAX_QUANTIZATION_TOTAL_ERROR_2BITS = 0.0075f;
+const float MAX_QUANTIZATION_TOTAL_ERROR_3BITS = 0.0040f;
+const float MAX_DOT_PRODUCT_ERROR = 0.02f;
 
 const char* RESULT_STR[] = {"ok", "FAILED"};
 

tests/test-quantize-perf.cpp

@@ -13,6 +13,10 @@
 #include <string>
 #include <vector>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 #define MAX_ALIGNMENT 64
 #define QK 32
 #define WARMUP 5

tests/test-sampling.cpp

@@ -176,27 +176,27 @@ void test_frequency_presence_penalty(
 int main(void) {
     ggml_time_init();
 
-    test_top_k({0.1, 0.2, 0.3, 0.4}, {0.4}, 1);
-    test_top_k({0.1, 0.2, 0.3, 0.4}, {0.4, 0.3, 0.2}, 3);
+    test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f}, 1);
+    test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f}, 3);
 
-    test_top_p({0.1, 0.2, 0.3, 0.4}, {0.4}, 0);
-    test_top_p({0.1, 0.2, 0.3, 0.4}, {0.4, 0.3}, 0.7);
-    test_top_p({0.1, 0.2, 0.3, 0.4}, {0.4, 0.3, 0.2, 0.1}, 1);
+    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f}, 0);
+    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f}, 0.7f);
+    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1);
 
-    test_tfs({0.1, 0.15, 0.2, 0.25, 0.3}, {0.3}, 0.25);
-    test_tfs({0.1, 0.15, 0.2, 0.25, 0.3}, {0.3, 0.25}, 0.75);
-    test_tfs({0.1, 0.15, 0.2, 0.25, 0.3}, {0.3, 0.25}, 0.99);
+    test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f}, 0.25f);
+    test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.75f);
+    test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.99f);
 
-    test_typical({0.97, 0.01, 0.01, 0.01}, {0.97}, 0.5);
-    test_typical({0.4, 0.2, 0.2, 0.2}, {0.2, 0.2, 0.2}, 0.5);
+    test_typical({0.97f, 0.01f, 0.01f, 0.01f}, {0.97f}, 0.5f);
+    test_typical({0.4f, 0.2f, 0.2f, 0.2f}, {0.2f, 0.2f, 0.2f}, 0.5f);
 
-    test_repetition_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0}, {0.25, 0.25, 0.25, 0.25, 0}, 50.0);
-    test_repetition_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0, 1, 2}, {0.5, 0.5, 0, 0, 0}, 50.0);
-    test_repetition_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0, 1, 2, 0, 0}, {0.5, 0.5, 0, 0, 0}, 50.0);
+    test_repetition_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0}, {0.25f, 0.25f, 0.25f, 0.25f, 0}, 50.0f);
+    test_repetition_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2}, {0.5f, 0.5f, 0, 0, 0}, 50.0f);
+    test_repetition_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.5f, 0.5f, 0, 0, 0}, 50.0f);
 
-    test_frequency_presence_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0},             {0.249997, 0.249997, 0.249997, 0.249997, 0.000011}, 5.0, 5.0);
-    test_frequency_presence_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0, 1, 2},       {0.499966, 0.499966, 0.000023, 0.000023, 0.000023}, 5.0, 5.0);
-    test_frequency_presence_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0, 1, 2, 0, 0}, {0.499977, 0.499977, 0.000023, 0.000023, 0.000000}, 5.0, 5.0);
+    test_frequency_presence_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0},             {0.249997f, 0.249997f, 0.249997f, 0.249997f, 0.000011f}, 5.0f, 5.0f);
+    test_frequency_presence_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2},       {0.499966f, 0.499966f, 0.000023f, 0.000023f, 0.000023f}, 5.0f, 5.0f);
+    test_frequency_presence_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.499977f, 0.499977f, 0.000023f, 0.000023f, 0.000000f}, 5.0f, 5.0f);
 
     printf("OK\n");
 }

tests/test-tokenizer-0.cpp

@@ -53,7 +53,7 @@ int main(int argc, char **argv) {
 
     for (const auto & test_kv : k_tests()) {
         std::vector<llama_token> res(test_kv.first.size());
-        const int n = llama_tokenize(ctx, test_kv.first.c_str(), res.data(), res.size(), true);
+        const int n = llama_tokenize(ctx, test_kv.first.c_str(), res.data(), int(res.size()), true);
         res.resize(n);
 
         bool correct = res.size() == test_kv.second.size();