make : use -lfto=auto to avoid warnings and maintain perf

* Extend llama_kv_cache_seq_rm to allow matichng any sequence

* Replace llama_kv_cache_tokens_rm with llama_kv_cache_clear

Use llama_kv_cache_clear for cache clearing

Change calls to llama_kv_cache_tokens_rm that want to delete by position to use llama_kv_cache_seq_rm functionality

.github/ISSUE_TEMPLATE/bug.md

@@ -1,8 +1,7 @@
 ---
-name: Issue and enhancement template
-about: Used to report issues and request enhancements for llama.cpp
-title: "[User] Insert summary of your issue or enhancement.."
-labels: ''
+name: Bug template
+about: Used to report bugs in llama.cpp
+labels: ["bug-unconfirmed"]
 assignees: ''
 
 ---
@@ -46,7 +45,7 @@ $ g++ --version
 
 # Failure Information (for bugs)
 
-Please help provide information about the failure if this is a bug. If it is not a bug, please remove the rest of this template.
+Please help provide information about the failure / bug.
 
 # Steps to Reproduce
 

.github/ISSUE_TEMPLATE/enhancement.md

@@ -0,0 +1,28 @@
+---
+name: Enhancement template
+about: Used to request enhancements for llama.cpp
+labels: ["enhancement"]
+assignees: ''
+
+---
+
+# Prerequisites
+
+Please answer the following questions for yourself before submitting an issue.
+
+- [ ] I am running the latest code. Development is very rapid so there are no tagged versions as of now.
+- [ ] I carefully followed the [README.md](https://github.com/ggerganov/llama.cpp/blob/master/README.md).
+- [ ] I [searched using keywords relevant to my issue](https://docs.github.com/en/issues/tracking-your-work-with-issues/filtering-and-searching-issues-and-pull-requests) to make sure that I am creating a new issue that is not already open (or closed).
+- [ ] I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new bug or useful enhancement to share.
+
+# Feature Description
+
+Please provide a detailed written description of what you were trying to do, and what you expected `llama.cpp` to do as an enhancement.
+
+# Motivation
+
+Please provide a detailed written description of reasons why this feature is necessary and how it is useful to `llama.cpp` users.
+
+# Possible Implementation
+
+If you have an idea as to how it can be implemented, please write a detailed description. Feel free to give links to external sources or share visuals that might be helpful to understand the details better.

.github/workflows/build.yml

@@ -33,17 +33,17 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential gcc-8
+          sudo apt-get install build-essential gcc-8 g++-8
 
       - name: Build
         id: make_build
         run: |
-          CC=gcc-8 make -j $(nproc)
+          CC=gcc-8 CXX=g++-8 make -j $(nproc)
 
       - name: Test
         id: make_test
         run: |
-          CC=gcc-8 make tests -j $(nproc)
+          CC=gcc-8 CXX=g++-8 make tests -j $(nproc)
           make test -j $(nproc)
 
   ubuntu-latest-cmake:

.github/workflows/code-coverage.yml

@@ -15,13 +15,13 @@ jobs:
       - name: Dependencies
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential gcc-8 lcov
+          sudo apt-get install build-essential gcc-9 g++-9 lcov
 
       - name: Build
-        run: CC=gcc-8 make -j LLAMA_CODE_COVERAGE=1 tests
+        run: CC=gcc-9 CXX=g++-9 make -j LLAMA_CODE_COVERAGE=1 tests
 
       - name: Run tests
-        run: CC=gcc-8 make test
+        run: CC=gcc-9 CXX=g++-9 make test
 
       - name: Generate coverage report
         run: |

.gitignore

@@ -10,6 +10,7 @@
 *.gcno
 *.gcda
 *.dot
+*.bat
 *.metallib
 .DS_Store
 .build/

CMakeLists.txt

@@ -45,7 +45,7 @@ endif()
 # general
 option(LLAMA_STATIC                     "llama: static link libraries"                          OFF)
 option(LLAMA_NATIVE                     "llama: enable -march=native flag"                      ON)
-option(LLAMA_LTO                        "llama: enable link time optimization"                  OFF)
+option(LLAMA_LTO                        "llama: enable link time optimization"                  ON)
 
 # debug
 option(LLAMA_ALL_WARNINGS               "llama: enable all compiler warnings"                   ON)
@@ -82,6 +82,7 @@ set(LLAMA_BLAS_VENDOR "Generic" CACHE STRING "llama: BLAS library vendor")
 option(LLAMA_CUBLAS                          "llama: use CUDA"                                  OFF)
 #option(LLAMA_CUDA_CUBLAS                     "llama: use cuBLAS for prompt processing"          OFF)
 option(LLAMA_CUDA_FORCE_DMMV                 "llama: use dmmv instead of mmvq CUDA kernels"     OFF)
+option(LLAMA_CUDA_FORCE_MMQ                  "llama: use mmq kernels instead of cuBLAS"         OFF)
 set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kernels")
 set(LLAMA_CUDA_MMV_Y        "1" CACHE STRING "llama: y block size for mmv CUDA kernels")
 option(LLAMA_CUDA_F16                        "llama: use 16 bit floats for some calculations"   OFF)
@@ -93,7 +94,6 @@ option(LLAMA_CLBLAST                         "llama: use CLBlast"
 option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
 option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"                   OFF)
 option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
-option(LLAMA_K_QUANTS                        "llama: use k-quants"                              ON)
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
 
 option(LLAMA_BUILD_TESTS                "llama: build tests"    ${LLAMA_STANDALONE})
@@ -277,13 +277,8 @@ if (LLAMA_BLAS)
     endif()
 endif()
 
-if (LLAMA_K_QUANTS)
-    set(GGML_HEADERS_EXTRA k_quants.h)
-    set(GGML_SOURCES_EXTRA k_quants.c)
-    add_compile_definitions(GGML_USE_K_QUANTS)
-    if (LLAMA_QKK_64)
-        add_compile_definitions(GGML_QKK_64)
-    endif()
+if (LLAMA_QKK_64)
+    add_compile_definitions(GGML_QKK_64)
 endif()
 
 if (LLAMA_CUBLAS)
@@ -305,6 +300,9 @@ if (LLAMA_CUBLAS)
         if (LLAMA_CUDA_FORCE_DMMV)
             add_compile_definitions(GGML_CUDA_FORCE_DMMV)
         endif()
+        if (LLAMA_CUDA_FORCE_MMQ)
+            add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+        endif()
         add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
         add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
         if (DEFINED LLAMA_CUDA_DMMV_Y)
@@ -331,6 +329,7 @@ if (LLAMA_CUBLAS)
             set(CMAKE_CUDA_ARCHITECTURES "60;61;70") # needed for f16 CUDA intrinsics
         else()
             set(CMAKE_CUDA_ARCHITECTURES "52;61;70") # lowest CUDA 12 standard + lowest for integer intrinsics
+            #set(CMAKE_CUDA_ARCHITECTURES "") # use this to compile much faster, but only F16 models work
         endif()
     endif()
     message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
@@ -404,6 +403,9 @@ if (LLAMA_HIPBLAS)
         if (LLAMA_CUDA_FORCE_DMMV)
             target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_DMMV)
         endif()
+        if (LLAMA_CUDA_FORCE_MMQ)
+            target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_MMQ)
+        endif()
         target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
         target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
         target_compile_definitions(ggml-rocm PRIVATE K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
@@ -665,6 +667,8 @@ add_library(ggml OBJECT
             ggml-alloc.h
             ggml-backend.c
             ggml-backend.h
+            ggml-quants.c
+            ggml-quants.h
             ${GGML_SOURCES_CUDA} ${GGML_HEADERS_CUDA}
             ${GGML_SOURCES_OPENCL} ${GGML_HEADERS_OPENCL}
             ${GGML_SOURCES_METAL} ${GGML_HEADERS_METAL}

Makefile

@@ -120,11 +120,11 @@ MK_CXXFLAGS = -std=c++11 -fPIC
 
 # -Ofast tends to produce faster code, but may not be available for some compilers.
 ifdef LLAMA_FAST
-MK_CFLAGS        += -Ofast
+MK_CFLAGS        += -Ofast -flto=auto
 MK_HOST_CXXFLAGS += -Ofast
 MK_CUDA_CXXFLAGS += -O3
 else
-MK_CFLAGS        += -O3
+MK_CFLAGS        += -O3 -flto=auto
 MK_CXXFLAGS      += -O3
 endif
 
@@ -342,13 +342,9 @@ else
 	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
 endif
 
-ifndef LLAMA_NO_K_QUANTS
-	MK_CPPFLAGS += -DGGML_USE_K_QUANTS
-	OBJS     += k_quants.o
 ifdef LLAMA_QKK_64
 	MK_CPPFLAGS += -DGGML_QKK_64
 endif
-endif
 
 ifndef LLAMA_NO_ACCELERATE
 	# Mac OS - include Accelerate framework.
@@ -365,7 +361,7 @@ ifdef LLAMA_MPI
 	MK_CPPFLAGS += -DGGML_USE_MPI
 	MK_CFLAGS   += -Wno-cast-qual
 	MK_CXXFLAGS += -Wno-cast-qual
-	OBJS     += ggml-mpi.o
+	OBJS        += ggml-mpi.o
 endif # LLAMA_MPI
 
 ifdef LLAMA_OPENBLAS
@@ -382,7 +378,7 @@ endif # LLAMA_BLIS
 ifdef LLAMA_CUBLAS
 	MK_CPPFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
 	MK_LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
-	OBJS      += ggml-cuda.o
+	OBJS         += ggml-cuda.o
 	NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math
 ifdef LLAMA_CUDA_NVCC
 	NVCC = $(LLAMA_CUDA_NVCC)
@@ -397,6 +393,9 @@ endif # CUDA_DOCKER_ARCH
 ifdef LLAMA_CUDA_FORCE_DMMV
 	NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
 endif # LLAMA_CUDA_FORCE_DMMV
+ifdef LLAMA_CUDA_FORCE_MMQ
+	NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
+endif # LLAMA_CUDA_FORCE_MMQ
 ifdef LLAMA_CUDA_DMMV_X
 	NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
 else
@@ -494,11 +493,6 @@ ggml-mpi.o: ggml-mpi.c ggml-mpi.h
 	$(CC) $(CFLAGS) -c $< -o $@
 endif # LLAMA_MPI
 
-ifndef LLAMA_NO_K_QUANTS
-k_quants.o: k_quants.c k_quants.h
-	$(CC) $(CFLAGS) -c $< -o $@
-endif # LLAMA_NO_K_QUANTS
-
 # combine build flags with cmdline overrides
 override CFLAGS        := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(CFLAGS)
 override CXXFLAGS      := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS)
@@ -539,15 +533,18 @@ ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
 ggml-backend.o: ggml-backend.c ggml.h ggml-backend.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
-OBJS += ggml-alloc.o ggml-backend.o
+ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h
+	$(CC) $(CFLAGS)    -c $< -o $@
+
+OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o
 
 llama.o: llama.cpp ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-COMMON_H_DEPS = common/common.h common/sampling.h build-info.h common/log.h
-COMMON_DEPS   = $(COMMON_H_DEPS) common.o sampling.o grammar-parser.o
+COMMON_H_DEPS = common/common.h common/sampling.h common/log.h
+COMMON_DEPS   = common.o sampling.o grammar-parser.o
 
-common.o: common/common.cpp $(COMMON_H_DEPS)
+common.o: common/common.cpp build-info.h $(COMMON_H_DEPS)
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 sampling.o: common/sampling.cpp $(COMMON_H_DEPS)
@@ -605,8 +602,8 @@ embedding: examples/embedding/embedding.cpp                   build-info.h ggml.
 save-load-state: examples/save-load-state/save-load-state.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
-	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2)
+server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2) -Wno-cast-qual
 
 gguf: examples/gguf/gguf.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

Package.swift

@@ -42,13 +42,12 @@ let package = Package(
                 "llama.cpp",
                 "ggml-alloc.c",
                 "ggml-backend.c",
-                "k_quants.c",
+                "ggml-quants.c",
             ] + additionalSources,
             resources: resources,
             publicHeadersPath: "spm-headers",
             cSettings: [
                 .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
-                .define("GGML_USE_K_QUANTS"),
                 .define("GGML_USE_ACCELERATE")
                 // NOTE: NEW_LAPACK will required iOS version 16.4+
                 // We should consider add this in the future when we drop support for iOS 14

README.md

@@ -101,7 +101,7 @@ as the main playground for developing new features for the [ggml](https://github
 
 - Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
 - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
-- Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp), [hlhr202/llama-node](https://github.com/hlhr202/llama-node)
+- Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
 - Rust: [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
 - C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp)

build.zig

@@ -116,30 +116,26 @@ pub fn build(b: *std.build.Builder) !void {
     var make = try Maker.init(b);
     make.enable_lto = b.option(bool, "lto", "Enable LTO optimization, (default: false)") orelse false;
 
-    if (b.option(bool, "k-quants", "Enable K-quants, (default: true)") orelse true) {
-        try make.addFlag("-DGGML_USE_K_QUANTS");
-        const k_quants = make.obj("k_quants", "k_quants.c");
-        try make.objs.append(k_quants);
-    }
-
     const ggml = make.obj("ggml", "ggml.c");
     const ggml_alloc = make.obj("ggml-alloc", "ggml-alloc.c");
     const ggml_backend = make.obj("ggml-backend", "ggml-backend.c");
+    const ggml_quants = make.obj("ggml-quants", "ggml-quants.c");
     const llama = make.obj("llama", "llama.cpp");
     const common = make.obj("common", "common/common.cpp");
     const console = make.obj("console", "common/console.cpp");
     const sampling = make.obj("sampling", "common/sampling.cpp");
     const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.cpp");
     const train = make.obj("train", "common/train.cpp");
+    const clip = make.obj("clip", "examples/llava/clip.cpp");
 
-    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, sampling, console, grammar_parser });
-    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common });
-    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common });
-    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common });
-    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, train });
-    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, train });
+    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, sampling, console, grammar_parser });
+    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common });
+    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common });
+    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common });
+    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, train });
+    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, train });
 
-    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, sampling, grammar_parser });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, sampling, grammar_parser, clip });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }

common/common.cpp

@@ -224,6 +224,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             sparams.temp = std::stof(argv[i]);
+            sparams.temp = std::max(sparams.temp, 0.0f);
         } else if (arg == "--tfs") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -632,6 +633,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
         process_escapes(params.prompt);
         process_escapes(params.input_prefix);
         process_escapes(params.input_suffix);
+        process_escapes(sparams.cfg_negative_prompt);
         for (auto & antiprompt : params.antiprompt) {
             process_escapes(antiprompt);
         }
@@ -742,7 +744,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
 #endif // GGML_USE_CUBLAS
 #endif
     printf("  --verbose-prompt      print prompt before generation\n");
-    fprintf(stderr, "  --simple-io           use basic IO for better compatibility in subprocesses and limited consoles\n");
+    printf("  --simple-io           use basic IO for better compatibility in subprocesses and limited consoles\n");
     printf("  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
     printf("  --lora-scaled FNAME S apply LoRA adapter with user defined scaling S (implies --no-mmap)\n");
     printf("  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
@@ -879,15 +881,15 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
     }
 
     if (params.ignore_eos) {
-        params.sparams.logit_bias[llama_token_eos(lctx)] = -INFINITY;
+        params.sparams.logit_bias[llama_token_eos(model)] = -INFINITY;
     }
 
     {
         LOG("warming up the model with an empty run\n");
 
-        std::vector<llama_token> tmp = { llama_token_bos(lctx), llama_token_eos(lctx), };
+        std::vector<llama_token> tmp = { llama_token_bos(model), llama_token_eos(model), };
         llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
-        llama_kv_cache_tokens_rm(lctx, -1, -1);
+        llama_kv_cache_clear(lctx);
         llama_reset_timings(lctx);
     }
 
@@ -940,7 +942,7 @@ std::string llama_token_to_piece(const struct llama_context * ctx, llama_token t
 }
 
 std::string llama_detokenize_spm(llama_context * ctx, const std::vector<llama_token> & tokens) {
-    const llama_token bos_id = llama_token_bos(ctx);
+    const llama_token bos_id = llama_token_bos(llama_get_model(ctx));
 
     std::string piece;
     std::string result;
@@ -1185,7 +1187,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "hellaswag: %s # default: false\n", params.hellaswag ? "true" : "false");
     fprintf(stream, "hellaswag_tasks: %zu # default: 400\n", params.hellaswag_tasks);
 
-    const auto logit_bias_eos = sparams.logit_bias.find(llama_token_eos(lctx));
+    const auto logit_bias_eos = sparams.logit_bias.find(llama_token_eos(llama_get_model(lctx)));
     const bool ignore_eos = logit_bias_eos != sparams.logit_bias.end() && logit_bias_eos->second == -INFINITY;
     fprintf(stream, "ignore_eos: %s # default: false\n", ignore_eos ? "true" : "false");
 

common/log.h

@@ -97,22 +97,23 @@
     #define LOG_TEE_TARGET stderr
 #endif
 
+// NOTE: currently disabled as it produces too many log files
 // Utility to obtain "pid" like unique process id and use it when creating log files.
-inline std::string log_get_pid()
-{
-    static std::string pid;
-    if (pid.empty())
-    {
-        // std::this_thread::get_id() is the most portable way of obtaining a "process id"
-        //  it's not the same as "pid" but is unique enough to solve multiple instances
-        //  trying to write to the same log.
-        std::stringstream ss;
-        ss << std::this_thread::get_id();
-        pid = ss.str();
-    }
-
-    return pid;
-}
+//inline std::string log_get_pid()
+//{
+//    static std::string pid;
+//    if (pid.empty())
+//    {
+//        // std::this_thread::get_id() is the most portable way of obtaining a "process id"
+//        //  it's not the same as "pid" but is unique enough to solve multiple instances
+//        //  trying to write to the same log.
+//        std::stringstream ss;
+//        ss << std::this_thread::get_id();
+//        pid = ss.str();
+//    }
+//
+//    return pid;
+//}
 
 // Utility function for generating log file names with unique id based on thread id.
 //  invocation with log_filename_generator( "llama", "log" ) creates a string "llama.<number>.log"
@@ -126,8 +127,8 @@ inline std::string log_filename_generator_impl(const std::string & log_file_base
     std::stringstream buf;
 
     buf << log_file_basename;
-    buf << ".";
-    buf << log_get_pid();
+    //buf << ".";
+    //buf << log_get_pid();
     buf << ".";
     buf << log_file_extension;
 

common/sampling.cpp

@@ -147,7 +147,7 @@ llama_token llama_sampling_sample(
 
     // apply penalties
     if (!prev.empty()) {
-        const float nl_logit = logits[llama_token_nl(ctx_main)];
+        const float nl_logit = logits[llama_token_nl(llama_get_model(ctx_main))];
 
         llama_sample_repetition_penalties(ctx_main, &cur_p,
                 prev.data() + prev.size() - penalty_last_n,
@@ -155,7 +155,7 @@ llama_token llama_sampling_sample(
 
         if (!penalize_nl) {
             for (size_t idx = 0; idx < cur_p.size; idx++) {
-                if (cur_p.data[idx].id == llama_token_nl(ctx_main)) {
+                if (cur_p.data[idx].id == llama_token_nl(llama_get_model(ctx_main))) {
                     cur_p.data[idx].logit = nl_logit;
                     break;
                 }
@@ -167,8 +167,12 @@ llama_token llama_sampling_sample(
         llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
     }
 
-    if (temp <= 0) {
-        // greedy sampling
+    if (temp < 0.0) {
+        // greedy sampling, with probs
+        llama_sample_softmax(ctx_main, &cur_p);
+        id = cur_p.data[0].id;
+    } else if (temp == 0.0) {
+        // greedy sampling, no probs
         id = llama_sample_token_greedy(ctx_main, &cur_p);
     } else {
         if (mirostat == 1) {

common/train.cpp

@@ -236,8 +236,8 @@ int64_t get_example_targets_batch(
     int64_t used_samples = 0;
 
     ggml_set_f32(target_probs, 0.0f);
-    llama_token bos = llama_token_bos(lctx);
-    llama_token eos = llama_token_eos(lctx);
+    llama_token bos = llama_token_bos(llama_get_model(lctx));
+    llama_token eos = llama_token_eos(llama_get_model(lctx));
     // printf("%s: example_id=%d n_batch=%d n_train_samples=%zu\n", __func__, example_id, n_batch, n_train_samples);
     for (int k=0; k<n_batch; ++k) {
         // printf("%s: batch %d\n", __func__, k);
@@ -924,7 +924,7 @@ size_t tokenize_file(
         for (llama_token token=0; token < n_vocab; ++token) {
             max_token_text_size = std::max(
                 max_token_text_size,
-                strlen(llama_token_get_text(lctx, token)));
+                strlen(llama_token_get_text(llama_get_model(lctx), token)));
         }
 
         // upper bound of context byte length.

convert-baichuan-hf-to-gguf.py

@@ -110,7 +110,7 @@ print("gguf: loading model "+dir_model.name)
 with open(dir_model / "config.json", "r", encoding="utf-8") as f:
     hparams = json.load(f)
 print("hello print: ",hparams["architectures"][0])
-if hparams["architectures"][0] != "BaichuanForCausalLM":
+if hparams["architectures"][0] != "BaichuanForCausalLM" and hparams["architectures"][0] != "BaiChuanForCausalLM":
     print("Model architecture not supported: " + hparams["architectures"][0])
 
     sys.exit()
@@ -230,7 +230,7 @@ gguf_writer.add_token_list(tokens)
 gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
-special_vocab = gguf.SpecialVocab(dir_model)
+special_vocab = gguf.SpecialVocab(dir_model, n_vocab = len(tokens))
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert-bloom-hf-to-gguf.py

@@ -118,18 +118,27 @@ tokenizer = AutoTokenizer.from_pretrained(dir_model)
 vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
 assert max(tokenizer.vocab.values()) < vocab_size
 
+added_vocab = tokenizer.get_added_vocab()
 reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
 
 for i in range(vocab_size):
-    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
-    scores.append(0.0)  # dummy
-    toktypes.append(gguf.TokenType.NORMAL)
+    if i not in reverse_vocab:
+        tokens.append(f"[PAD{i}]")
+        toktypes.append(gguf.TokenType.USER_DEFINED)
+    elif reverse_vocab[i] in added_vocab:
+        tokens.append(reverse_vocab[i])
+        if tokenizer.added_tokens_decoder[i].special:
+            toktypes.append(gguf.TokenType.CONTROL)
+        else:
+            toktypes.append(gguf.TokenType.USER_DEFINED)
+    else:
+        tokens.append(reverse_vocab[i])
+        toktypes.append(gguf.TokenType.NORMAL)
 
 gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
-special_vocab = gguf.SpecialVocab(dir_model, load_merges=True)
+special_vocab = gguf.SpecialVocab(dir_model, load_merges=True, n_vocab = len(tokens))
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert-falcon-hf-to-gguf.py

@@ -152,7 +152,7 @@ gguf_writer.add_token_list(tokens)
 gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
-special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert-gptneox-hf-to-gguf.py

@@ -123,18 +123,27 @@ tokenizer = AutoTokenizer.from_pretrained(dir_model)
 vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
 assert max(tokenizer.vocab.values()) < vocab_size
 
+added_vocab = tokenizer.get_added_vocab()
 reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
 
 for i in range(vocab_size):
-    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
-    scores.append(0.0) # dummy
-    toktypes.append(gguf.TokenType.NORMAL)
+    if i not in reverse_vocab:
+        tokens.append(f"[PAD{i}]")
+        toktypes.append(gguf.TokenType.USER_DEFINED)
+    elif reverse_vocab[i] in added_vocab:
+        tokens.append(reverse_vocab[i])
+        if tokenizer.added_tokens_decoder[i].special:
+            toktypes.append(gguf.TokenType.CONTROL)
+        else:
+            toktypes.append(gguf.TokenType.USER_DEFINED)
+    else:
+        tokens.append(reverse_vocab[i])
+        toktypes.append(gguf.TokenType.NORMAL)
 
 gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
-special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert-llama-ggml-to-gguf.py

@@ -388,7 +388,9 @@ def handle_metadata(cfg, hp):
         cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir,
         cfg.vocabtype )
     # FIXME: Respect cfg.vocab_dir?
-    svocab = gguf.SpecialVocab(cfg.model_metadata_dir)
+    svocab = gguf.SpecialVocab(cfg.model_metadata_dir,
+        load_merges = cfg.vocabtype == 'bpe',
+        n_vocab = vocab.vocab_size)
     convert.check_vocab_size(params, vocab)
     return (params, vocab, svocab)
 

convert-mpt-hf-to-gguf.py

@@ -128,18 +128,27 @@ vocab_size = hparams["vocab_size"]
 # ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
 tokenizer = AutoTokenizer.from_pretrained(dir_model)
 
+added_vocab = tokenizer.get_added_vocab()
 reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
 
 for i in range(vocab_size):
-    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
-    scores.append(0.0) # dummy
-    toktypes.append(gguf.TokenType.NORMAL)
+    if i not in reverse_vocab:
+        tokens.append(f"[PAD{i}]")
+        toktypes.append(gguf.TokenType.USER_DEFINED)
+    elif reverse_vocab[i] in added_vocab:
+        tokens.append(reverse_vocab[i])
+        if tokenizer.added_tokens_decoder[i].special:
+            toktypes.append(gguf.TokenType.CONTROL)
+        else:
+            toktypes.append(gguf.TokenType.USER_DEFINED)
+    else:
+        tokens.append(reverse_vocab[i])
+        toktypes.append(gguf.TokenType.NORMAL)
 
 gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
-special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert-refact-hf-to-gguf.py

@@ -139,18 +139,27 @@ tokenizer = AutoTokenizer.from_pretrained(dir_model)
 vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
 assert max(tokenizer.vocab.values()) < vocab_size
 
+added_vocab = tokenizer.get_added_vocab()
 reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
 
 for i in range(vocab_size):
-    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
-    scores.append(0.0) # dummy
-    toktypes.append(gguf.TokenType.NORMAL)
+    if i not in reverse_vocab:
+        tokens.append(f"[PAD{i}]")
+        toktypes.append(gguf.TokenType.USER_DEFINED)
+    elif reverse_vocab[i] in added_vocab:
+        tokens.append(reverse_vocab[i])
+        if tokenizer.added_tokens_decoder[i].special:
+            toktypes.append(gguf.TokenType.CONTROL)
+        else:
+            toktypes.append(gguf.TokenType.USER_DEFINED)
+    else:
+        tokens.append(reverse_vocab[i])
+        toktypes.append(gguf.TokenType.NORMAL)
 
 gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
-special_vocab = gguf.SpecialVocab(dir_model, load_merges=True)
+special_vocab = gguf.SpecialVocab(dir_model, load_merges=True, n_vocab = len(tokens))
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert-starcoder-hf-to-gguf.py

@@ -111,18 +111,26 @@ tokenizer = AutoTokenizer.from_pretrained(dir_model)
 vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
 assert max(tokenizer.vocab.values()) < vocab_size
 
+added_vocab = tokenizer.get_added_vocab()
 reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
 
 for i in range(vocab_size):
-    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
-    scores.append(0.0) # dummy
-    toktypes.append(gguf.TokenType.NORMAL)
+    if i not in reverse_vocab:
+        tokens.append(f"[PAD{i}]")
+        toktypes.append(gguf.TokenType.USER_DEFINED)
+    elif reverse_vocab[i] in added_vocab:
+        tokens.append(reverse_vocab[i])
+        if tokenizer.added_tokens_decoder[i].special:
+            toktypes.append(gguf.TokenType.CONTROL)
+        else:
+            toktypes.append(gguf.TokenType.USER_DEFINED)
+    else:
+        tokens.append(reverse_vocab[i])
+        toktypes.append(gguf.TokenType.NORMAL)
 
 gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
-
-special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert.py

@@ -366,16 +366,19 @@ class SentencePieceVocab:
             added_tokens = {}
 
         vocab_size: int = self.sentencepiece_tokenizer.vocab_size()
-        expected_ids = list(range(vocab_size, vocab_size + len(added_tokens)))
-        actual_ids   = sorted(added_tokens.values())
-        if expected_ids != actual_ids:
-            raise Exception(f"Expected added token IDs to be sequential and start at {len(added_tokens)}; got {actual_ids}")
 
-        items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1])
-        self.added_tokens_list = [text for (text, idx) in items]
-        self.vocab_size_base: int = vocab_size
-        self.vocab_size: int = self.vocab_size_base + len(self.added_tokens_list)
-        self.fname_tokenizer = fname_tokenizer
+        new_tokens       = {id: piece for piece, id in added_tokens.items() if id >= vocab_size}
+        expected_new_ids = list(range(vocab_size, vocab_size + len(new_tokens)))
+        actual_new_ids   = sorted(new_tokens.keys())
+
+        if expected_new_ids != actual_new_ids:
+            raise ValueError(f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}")
+
+        # Token pieces that were added to the base vocabulary.
+        self.added_tokens_list  = [new_tokens[id] for id in actual_new_ids]
+        self.vocab_size_base    = vocab_size
+        self.vocab_size         = self.vocab_size_base + len(self.added_tokens_list)
+        self.fname_tokenizer    = fname_tokenizer
         self.fname_added_tokens = fname_added_tokens
 
     def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
@@ -1163,10 +1166,13 @@ def main(args_in: list[str] | None = None) -> None:
 
     vocab: Vocab
     if args.vocab_only:
-        assert args.outfile, "need --outfile if using --vocab-only"
+        if not args.outfile:
+            raise ValueError("need --outfile if using --vocab-only")
         # FIXME: Try to respect vocab_dir somehow?
         vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype)
-        special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, load_merges = args.vocabtype == 'bpe')
+        special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent,
+            load_merges = args.vocabtype == 'bpe',
+            n_vocab = vocab.vocab_size)
         outfile = args.outfile
         OutputFile.write_vocab_only(outfile, params, vocab, special_vocab)
         print(f"Wrote {outfile}")
@@ -1178,7 +1184,9 @@ def main(args_in: list[str] | None = None) -> None:
         vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent
         vocab = load_vocab(vocab_dir, args.vocabtype)
     # FIXME: Try to respect vocab_dir somehow?
-    special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, load_merges = args.vocabtype == 'bpe')
+    special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent,
+        load_merges = args.vocabtype == 'bpe',
+        n_vocab = vocab.vocab_size)
 
     model   = model_plus.model
     model   = convert_model_names(model, params)

examples/batched-bench/batched-bench.cpp

@@ -154,6 +154,10 @@ int main(int argc, char ** argv) {
         }
     }
 
+    LOG_TEE("\n");
+    LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, mmq = %d\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq);
+    LOG_TEE("\n");
+
     LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP",     "TG",     "B",    "N_KV",     "T_PP s",   "S_PP t/s", "T_TG s",   "S_TG t/s", "T s",      "S t/s");
     LOG_TEE("|%6s-|-%6s-|-%4s-|-%6s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|\n", "------", "------", "----", "------", "--------", "--------", "--------", "--------", "--------", "--------");
 
@@ -181,7 +185,7 @@ int main(int argc, char ** argv) {
 
                 const auto t_pp_start = ggml_time_us();
 
-                llama_kv_cache_tokens_rm(ctx, -1, -1);
+                llama_kv_cache_clear(ctx);
 
                 if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
                     LOG_TEE("%s: llama_decode() failed\n", __func__);

examples/batched/batched.cpp

@@ -11,7 +11,7 @@ int main(int argc, char ** argv) {
     gpt_params params;
 
     if (argc == 1 || argv[1][0] == '-') {
-        printf("usage: %s MODEL_PATH [PROMPT] [PARALLEL] [LEN]\n" , argv[0]);
+        printf("usage: %s MODEL_PATH [PROMPT] [PARALLEL] [LEN] [NGL]\n" , argv[0]);
         return 1 ;
     }
 
@@ -21,6 +21,9 @@ int main(int argc, char ** argv) {
     // total length of the sequences including the prompt
     int n_len = 32;
 
+    // number of layers to offload to the GPU
+    int n_gpu_layers = 0;
+
     if (argc >= 2) {
         params.model = argv[1];
     }
@@ -37,6 +40,10 @@ int main(int argc, char ** argv) {
         n_len = std::atoi(argv[4]);
     }
 
+    if (argc >= 6) {
+        n_gpu_layers = std::atoi(argv[5]);
+    }
+
     if (params.prompt.empty()) {
         params.prompt = "Hello my name is";
     }
@@ -49,7 +56,7 @@ int main(int argc, char ** argv) {
 
     llama_model_params model_params = llama_model_default_params();
 
-    // model_params.n_gpu_layers = 99; // offload all layers to the GPU
+    model_params.n_gpu_layers = n_gpu_layers;
 
     llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 
@@ -180,7 +187,7 @@ int main(int argc, char ** argv) {
             //const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
             // is it an end of stream? -> mark the stream as finished
-            if (new_token_id == llama_token_eos(ctx) || n_cur == n_len) {
+            if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
                 i_batch[i] = -1;
                 LOG_TEE("\n");
                 if (n_parallel > 1) {

examples/beam-search/beam-search.cpp

@@ -47,7 +47,7 @@ struct beam_search_callback_data {
 // In this case, end-of-beam (eob) is equivalent to end-of-sentence (eos) but this need not always be the same.
 // For example, eob can be flagged due to maximum token length, stop words, etc.
 static bool is_at_eob(const beam_search_callback_data & callback_data, const llama_token * tokens, size_t n_tokens) {
-    return n_tokens && tokens[n_tokens-1] == llama_token_eos(callback_data.ctx);
+    return n_tokens && tokens[n_tokens-1] == llama_token_eos(llama_get_model(callback_data.ctx));
 }
 
 // Function matching type llama_beam_search_callback_fn_t.

examples/infill/infill.cpp

@@ -246,14 +246,14 @@ int main(int argc, char ** argv) {
     if (suff_rm_leading_spc && inp_sfx[0] == space_token) {
         inp_sfx.erase(inp_sfx.begin());
     }
-    inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(ctx));
+    inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
     if (add_bos) {
-        inp_pfx.insert(inp_pfx.begin(), llama_token_bos(ctx));
+        inp_pfx.insert(inp_pfx.begin(), llama_token_bos(model));
     }
-    inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(ctx));
+    inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
     embd_inp = inp_pfx;
     embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
-    embd_inp.push_back(llama_token_middle(ctx));
+    embd_inp.push_back(llama_token_middle(model));
 
     LOG("prefix: \"%s\"\n", log_tostr(params.input_prefix));
     LOG("suffix: \"%s\"\n", log_tostr(params.input_suffix));
@@ -261,7 +261,7 @@ int main(int argc, char ** argv) {
 
     // Should not run without any tokens
     if (embd_inp.empty()) {
-        embd_inp.push_back(llama_token_bos(ctx));
+        embd_inp.push_back(llama_token_bos(model));
         LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
     }
 
@@ -577,10 +577,10 @@ int main(int argc, char ** argv) {
         if ((int) embd_inp.size() <= n_consumed) {
 
             // deal with eot token in infill mode
-            if ((llama_sampling_last(ctx_sampling) == llama_token_eot(ctx) || is_interacting) && params.interactive){
+            if ((llama_sampling_last(ctx_sampling) == llama_token_eot(model) || is_interacting) && params.interactive){
                 if(is_interacting && !params.interactive_first) {
                     // print an eot token
-                    printf("%s", llama_token_to_piece(ctx, llama_token_eot(ctx)).c_str());
+                    printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
                 }
                 fflush(stdout);
                 printf("\n");
@@ -627,14 +627,14 @@ int main(int argc, char ** argv) {
                 if (suff_rm_leading_spc && inp_sfx[0] == space_token) {
                     inp_sfx.erase(inp_sfx.begin());
                 }
-                inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(ctx));
+                inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
                 if (add_bos) {
-                    inp_pfx.insert(inp_pfx.begin(), llama_token_bos(ctx));
+                    inp_pfx.insert(inp_pfx.begin(), llama_token_bos(model));
                 }
-                inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(ctx));
+                inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
                 embd_inp = inp_pfx;
                 embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
-                embd_inp.push_back(llama_token_middle(ctx));
+                embd_inp.push_back(llama_token_middle(model));
                 embd.clear();
                 embd_guidance.clear();
                 n_remain = params.n_predict;
@@ -644,7 +644,7 @@ int main(int argc, char ** argv) {
                 is_interacting = false;
             }
             // deal with end of text token in interactive mode
-            else if (llama_sampling_last(ctx_sampling) == llama_token_eos(ctx)) {
+            else if (llama_sampling_last(ctx_sampling) == llama_token_eos(model)) {
                 LOG("found EOS token\n");
 
                 if (params.interactive) {
@@ -661,7 +661,7 @@ int main(int argc, char ** argv) {
 
                 if (params.input_prefix_bos) {
                     LOG("adding input prefix BOS token\n");
-                    embd_inp.push_back(llama_token_bos(ctx));
+                    embd_inp.push_back(llama_token_bos(model));
                 }
 
                 std::string buffer;
@@ -724,7 +724,7 @@ int main(int argc, char ** argv) {
         }
 
         // end of text token
-        if (!embd.empty() && embd.back() == llama_token_eos(ctx) && !params.interactive) {
+        if (!embd.empty() && embd.back() == llama_token_eos(model) && !params.interactive) {
             break;
         }
 
@@ -736,7 +736,7 @@ int main(int argc, char ** argv) {
         }
     }
     if (!params.interactive && n_remain <= 0) {
-        printf("%s", llama_token_to_piece(ctx, llama_token_eot(ctx)).c_str());
+        printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
         fflush(stdout);
     }
 

examples/llama-bench/llama-bench.cpp

@@ -933,7 +933,7 @@ struct sql_printer : public printer {
 };
 
 static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_batch, int n_threads) {
-    std::vector<llama_token> tokens(n_batch, llama_token_bos(ctx));
+    std::vector<llama_token> tokens(n_batch, llama_token_bos(llama_get_model(ctx)));
     int n_processed = 0;
 
     llama_set_n_threads(ctx, n_threads, n_threads);
@@ -946,7 +946,7 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_bat
 }
 
 static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
-    llama_token token = llama_token_bos(ctx);
+    llama_token token = llama_token_bos(llama_get_model(ctx));
 
     llama_set_n_threads(ctx, n_threads, n_threads);
 
@@ -1037,7 +1037,7 @@ int main(int argc, char ** argv) {
 
         test t(inst, lmodel, ctx);
 
-        llama_kv_cache_tokens_rm(ctx, -1, -1);
+        llama_kv_cache_clear(ctx);
 
         // warmup run
         if (t.n_prompt > 0) {
@@ -1048,7 +1048,7 @@ int main(int argc, char ** argv) {
         }
 
         for (int i = 0; i < params.reps; i++) {
-            llama_kv_cache_tokens_rm(ctx, -1, -1);
+            llama_kv_cache_clear(ctx);
 
             uint64_t t_start = get_time_ns();
             if (t.n_prompt > 0) {

examples/llava/CMakeLists.txt

@@ -1,7 +1,7 @@
 set(TARGET clip)
 add_library(${TARGET} clip.cpp clip.h)
 install(TARGETS ${TARGET} LIBRARY)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common ggml ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if (NOT MSVC)
     target_compile_options(${TARGET} PRIVATE -Wno-cast-qual) # stb_image.h

examples/llava/clip.cpp

@@ -610,8 +610,8 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         int idx_mean = get_key_idx(ctx, KEY_IMAGE_MEAN);
         int idx_std = get_key_idx(ctx, KEY_IMAGE_STD);
         for (int i = 0; i < 3; ++i) {
-            new_clip->image_mean[i] = *((float *)gguf_get_arr_data(ctx, idx_mean));
-            new_clip->image_std[i] = *((float *)gguf_get_arr_data(ctx, idx_std));
+            new_clip->image_mean[i] = *((const float *)gguf_get_arr_data(ctx, idx_mean));
+            new_clip->image_std[i] = *((const float *)gguf_get_arr_data(ctx, idx_std));
         }
 
         if (verbosity >= 2) {

examples/llava/llava-utils.h

@@ -137,7 +137,7 @@ inline llama_token sample_id(llama_context * ctx_llama, gpt_params & params) {
 inline const char * sample(struct llama_context * ctx_llama, gpt_params & params, int * n_past) {
     int id = sample_id(ctx_llama, params);
     static std::string ret;
-    if (id == llama_token_eos(ctx_llama)) {
+    if (id == llama_token_eos(llama_get_model(ctx_llama))) {
         ret = "</s>";
     } else {
         ret = llama_token_to_piece(ctx_llama, id);

examples/main-cmake-pkg/CMakeLists.txt

@@ -16,6 +16,8 @@ add_library(common OBJECT
     ${_common_path}/console.cpp
     ${_common_path}/grammar-parser.h
     ${_common_path}/grammar-parser.cpp
+    ${_common_path}/sampling.h
+    ${_common_path}/sampling.cpp
     )
 
 # WARNING: because build-info.h is auto-generated, it will only

examples/main/main.cpp

@@ -248,7 +248,7 @@ int main(int argc, char ** argv) {
 
     // Should not run without any tokens
     if (embd_inp.empty()) {
-        embd_inp.push_back(llama_token_bos(ctx));
+        embd_inp.push_back(llama_token_bos(model));
         LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
     }
 
@@ -298,7 +298,7 @@ int main(int argc, char ** argv) {
         }
 
         // remove any "future" tokens that we might have inherited from the previous session
-        llama_kv_cache_tokens_rm(ctx, n_matching_session_tokens, -1);
+        llama_kv_cache_seq_rm(ctx, -1, n_matching_session_tokens, -1);
     }
 
     LOGLN(
@@ -693,7 +693,7 @@ int main(int argc, char ** argv) {
             }
 
             // deal with end of text token in interactive mode
-            if (llama_sampling_last(ctx_sampling) == llama_token_eos(ctx)) {
+            if (llama_sampling_last(ctx_sampling) == llama_token_eos(model)) {
                 LOG("found EOS token\n");
 
                 if (params.interactive) {
@@ -720,7 +720,7 @@ int main(int argc, char ** argv) {
 
                 if (params.input_prefix_bos) {
                     LOG("adding input prefix BOS token\n");
-                    embd_inp.push_back(llama_token_bos(ctx));
+                    embd_inp.push_back(llama_token_bos(model));
                 }
 
                 std::string buffer;
@@ -761,6 +761,9 @@ int main(int argc, char ** argv) {
                         n_consumed = embd_inp.size();
                         embd_inp.insert(embd_inp.end(), inp_pfx.begin(), inp_pfx.end());
                     }
+                    if (params.escape) {
+                        process_escapes(buffer);
+                    }
 
                     const auto line_pfx = ::llama_tokenize(ctx, params.input_prefix, false, true);
                     const auto line_inp = ::llama_tokenize(ctx, buffer,              false, false);
@@ -801,7 +804,7 @@ int main(int argc, char ** argv) {
         }
 
         // end of text token
-        if (!embd.empty() && embd.back() == llama_token_eos(ctx) && !(params.instruct || params.interactive)) {
+        if (!embd.empty() && embd.back() == llama_token_eos(model) && !(params.instruct || params.interactive)) {
             LOG_TEE(" [end of text]\n");
             break;
         }

examples/parallel/parallel.cpp

@@ -347,7 +347,7 @@ int main(int argc, char ** argv) {
                 //        client.id, client.seq_id, id, client.n_decoded, client.i_batch, token_str.c_str());
 
                 if (client.n_decoded > 2 &&
-                        (id == llama_token_eos(ctx) ||
+                        (id == llama_token_eos(model) ||
                          (params.n_predict > 0 && client.n_decoded + client.n_prompt >= params.n_predict) ||
                          client.response.find("User:") != std::string::npos ||
                          client.response.find('\n') != std::string::npos)) {

examples/perplexity/perplexity.cpp

@@ -210,7 +210,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
         const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
-        llama_kv_cache_tokens_rm(ctx, -1, -1);
+        llama_kv_cache_clear(ctx);
 
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
@@ -227,7 +227,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
 
             // add BOS token for the first batch of each chunk
             if (add_bos && j == 0) {
-                tokens[batch_start] = llama_token_bos(ctx);
+                tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
             }
 
             const auto batch_logits = llama_get_logits(ctx);
@@ -339,7 +339,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
         const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
-        llama_kv_cache_tokens_rm(ctx, -1, -1);
+        llama_kv_cache_clear(ctx);
 
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
@@ -350,7 +350,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
 
             // add BOS token for the first batch of each chunk
             if (add_bos && j == 0) {
-                tokens[batch_start] = llama_token_bos(ctx);
+                tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
             }
 
             if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
@@ -573,7 +573,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         }
 
         // clear the KV cache
-        llama_kv_cache_tokens_rm(ctx, -1, -1);
+        llama_kv_cache_clear(ctx);
 
         auto logits = hellaswag_evaluate_tokens(ctx, query_embd, 0, params.n_batch, n_vocab);
         if (logits.empty()) {

examples/quantize/quantize.cpp

@@ -18,7 +18,6 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "Q4_1",   LLAMA_FTYPE_MOSTLY_Q4_1,   " 3.90G, +0.1585 ppl @ LLaMA-v1-7B", },
     { "Q5_0",   LLAMA_FTYPE_MOSTLY_Q5_0,   " 4.33G, +0.0683 ppl @ LLaMA-v1-7B", },
     { "Q5_1",   LLAMA_FTYPE_MOSTLY_Q5_1,   " 4.70G, +0.0349 ppl @ LLaMA-v1-7B", },
-#ifdef GGML_USE_K_QUANTS
     { "Q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K,   " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", },
     { "Q3_K",   LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" },
     { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", },
@@ -31,7 +30,6 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "Q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S, " 4.33G, +0.0400 ppl @ LLaMA-v1-7B", },
     { "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 4.45G, +0.0122 ppl @ LLaMA-v1-7B", },
     { "Q6_K",   LLAMA_FTYPE_MOSTLY_Q6_K,   " 5.15G, -0.0008 ppl @ LLaMA-v1-7B", },
-#endif
     { "Q8_0",   LLAMA_FTYPE_MOSTLY_Q8_0,   " 6.70G, +0.0004 ppl @ LLaMA-v1-7B", },
     { "F16",    LLAMA_FTYPE_MOSTLY_F16,    "13.00G              @ 7B", },
     { "F32",    LLAMA_FTYPE_ALL_F32,       "26.00G              @ 7B", },
@@ -70,13 +68,14 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
 }
 
 // usage:
-//  ./quantize [--allow-requantize] [--leave-output-tensor] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
+//  ./quantize [--allow-requantize] [--leave-output-tensor] [--pure] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
 //
 [[noreturn]]
 static void usage(const char * executable) {
-    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
+    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
     printf("  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
     printf("  --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
+    printf("  --pure: Disable k-quant mixtures and quantize all tensors to the same type\n");
     printf("\nAllowed quantization types:\n");
     for (auto & it : QUANT_OPTIONS) {
         if (it.name != "COPY") {
@@ -103,6 +102,8 @@ int main(int argc, char ** argv) {
             params.quantize_output_tensor = false;
         } else if (strcmp(argv[arg_idx], "--allow-requantize") == 0) {
             params.allow_requantize = true;
+        } else if (strcmp(argv[arg_idx], "--pure") == 0) {
+            params.pure = true;
         } else {
             usage(argv[0]);
         }

examples/server/CMakeLists.txt

@@ -6,7 +6,7 @@ install(TARGETS ${TARGET} RUNTIME)
 target_compile_definitions(${TARGET} PRIVATE
     SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
 )
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common llama clip ${CMAKE_THREAD_LIBS_INIT})
 if (WIN32)
     TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
 endif()

examples/server/README.md

@@ -24,6 +24,10 @@ Command line options:
 -   `--port`: Set the port to listen. Default: `8080`.
 -   `--path`: path from which to serve static files (default examples/server/public)
 -   `--embedding`: Enable embedding extraction, Default: disabled.
+-   `-np N`, `--parallel N`: Set the number of slots for process requests (default: 1)
+-   `-cb`, `--cont-batching`: enable continuous batching (a.k.a dynamic batching) (default: disabled)
+-   `-spf FNAME`, `--system-prompt-file FNAME` Set a file to load "a system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
+-   `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA.
 
 ## Build
 
@@ -158,6 +162,8 @@ node index.js
 
     `n_probs`: If greater than 0, the response also contains the probabilities of top N tokens for each generated token (default: 0)
 
+    `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `prompt`. You can determine the place of the image in the prompt as in the following: `USER:[img-12]Describe the image in detail.\nASSISTANT:` In this case, `[img-12]` will be replaced by the embeddings of the image id 12 in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 12}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
+
     *Result JSON:*
 
     Note: When using streaming mode (`stream`) only `content` and `stop` will be returned until end of completion.
@@ -188,6 +194,12 @@ node index.js
 
     `truncated`: Boolean indicating if the context size was exceeded during generation, i.e. the number of tokens provided in the prompt (`tokens_evaluated`) plus tokens generated (`tokens predicted`) exceeded the context size (`n_ctx`)
 
+    `slot_id`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot (default: -1)
+
+    `cache_prompt`: Save the prompt and generation for avoid reprocess entire prompt if a part of this isn't change (default: false)
+
+    `system_prompt`: Change the system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
+
 -   **POST** `/tokenize`: Tokenize a given text.
 
     *Options:*
@@ -218,8 +230,32 @@ node index.js
 
     It also accepts all the options of `/completion` except `stream` and `prompt`.
 
+-   **GET** `/props`: Return the required assistant name and anti-prompt to generate the prompt in case you have specified a system prompt for all slots.
+
 ## More examples
 
+### Change system prompt on runtime
+
+To use the server example to serve multiple chat-type clients while keeping the same system prompt, you can utilize the option `system_prompt` to achieve that. This only needs to be done once to establish it.
+
+`prompt`: Specify a context that you want all connecting clients to respect.
+
+`anti_prompt`: Specify the word you want to use to instruct the model to stop. This must be sent to each client through the `/props` endpoint.
+
+`assistant_name`: The bot's name is necessary for each customer to generate the prompt. This must be sent to each client through the `/props` endpoint.
+
+```json
+{
+    "system_prompt": {
+        "prompt": "Transcript of a never ending dialog, where the User interacts with an Assistant.\nThe Assistant is helpful, kind, honest, good at writing, and never fails to answer the User's requests immediately and with precision.\nUser: Recommend a nice restaurant in the area.\nAssistant: I recommend the restaurant \"The Golden Duck\". It is a 5 star restaurant with a great view of the city. The food is delicious and the service is excellent. The prices are reasonable and the portions are generous. The restaurant is located at 123 Main Street, New York, NY 10001. The phone number is (212) 555-1234. The hours are Monday through Friday from 11:00 am to 10:00 pm. The restaurant is closed on Saturdays and Sundays.\nUser: Who is Richard Feynman?\nAssistant: Richard Feynman was an American physicist who is best known for his work in quantum mechanics and particle physics. He was awarded the Nobel Prize in Physics in 1965 for his contributions to the development of quantum electrodynamics. He was a popular lecturer and author, and he wrote several books, including \"Surely You're Joking, Mr. Feynman!\" and \"What Do You Care What Other People Think?\".\nUser:",
+        "anti_prompt": "User:",
+        "assistant_name": "Assistant:"
+    }
+}
+```
+
+**NOTE**: You can do this automatically when starting the server by simply creating a .json file with these options and using the CLI option `-spf FNAME` or `--system-prompt-file FNAME`.
+
 ### Interactive mode
 
 Check the sample in [chat.mjs](chat.mjs).

examples/server/api_like_OAI.py

@@ -8,6 +8,7 @@ import json
 
 
 app = Flask(__name__)
+slot_id = -1
 
 parser = argparse.ArgumentParser(description="An example of using server.cpp with a similar API to OAI. It must be used together with server.cpp.")
 parser.add_argument("--chat-prompt", type=str, help="the top prompt in chat completions(default: 'A chat between a curious user and an artificial intelligence assistant. The assistant follows the given rules no matter what.\\n')", default='A chat between a curious user and an artificial intelligence assistant. The assistant follows the given rules no matter what.\\n')
@@ -77,7 +78,8 @@ def make_postData(body, chat=False, stream=False):
     if(is_present(body, "stop")): postData["stop"] += body["stop"]
     postData["n_keep"] = -1
     postData["stream"] = stream
-
+    postData["cache_prompt"] = True
+    postData["slot_id"] = slot_id
     return postData
 
 def make_resData(data, chat=False, promptToken=[]):
@@ -128,6 +130,7 @@ def make_resData_stream(data, chat=False, time_now = 0, start=False):
             }
         ]
     }
+    slot_id = data["slot_id"]
     if (chat):
         if (start):
             resData["choices"][0]["delta"] =  {

examples/server/chat.mjs

@@ -7,6 +7,11 @@ const args = process.argv.slice(2);
 const grammarJsonSchemaFile = args.find(
     (_, index) => args[index - 1] === "--grammar-json-schema"
 );
+
+const no_cached_prompt = args.find(
+    (_, index) => args[index - 1] === "--no-cache-prompt"
+) ?? "false";
+
 const grammarFile = args.find((_, index) => args[index - 1] === "--grammar");
 
 // Example usage: function,arguments
@@ -30,6 +35,9 @@ if (grammarFile) {
     grammar = readFileSync(grammarFile, 'utf-8')
 }
 
+// for cached prompt
+let slot_id = -1;
+
 const API_URL = 'http://127.0.0.1:8080'
 
 const chat = [
@@ -76,6 +84,8 @@ async function chat_completion(question) {
             top_p: 0.9,
             n_keep: n_keep,
             n_predict: 256,
+            cache_prompt: no_cached_prompt === "false",
+            slot_id: slot_id,
             stop: ["\n### Human:"], // stop completion after generating this
             grammar,
             stream: true,
@@ -92,6 +102,7 @@ async function chat_completion(question) {
         const t = Buffer.from(chunk).toString('utf8')
         if (t.startsWith('data: ')) {
             const message = JSON.parse(t.substring(6))
+            slot_id = message.slot_id
             answer += message.content
             process.stdout.write(message.content)
             if (message.stop) {

examples/server/public/index.html

@@ -125,6 +125,7 @@
       background-color: #222;
       color: #ddd;
     }
+
     code {
       font-family: monospace;
       padding: 0.1em 0.3em;
@@ -141,7 +142,8 @@
       display: inline;
     }
 
-    header, footer {
+    header,
+    footer {
       text-align: center;
     }
 
@@ -163,6 +165,7 @@
       0% {
         background-position: 0%;
       }
+
       100% {
         background-position: 100%;
       }
@@ -181,6 +184,7 @@
         --loading-color-1: #22222200;
         --loading-color-2: #222222ff;
       }
+
       .popover-content {
         background-color: black;
       }
@@ -194,6 +198,8 @@
 
     import { llama } from '/completion.js';
     import { SchemaConverter } from '/json-schema-to-grammar.mjs';
+    let selected_image = false;
+    var slot_id = -1;
 
     const session = signal({
       prompt: "This is a conversation between User and Llama, a friendly chatbot. Llama is helpful, kind, honest, good at writing, and never fails to answer any requests immediately and with precision.",
@@ -203,6 +209,7 @@
       type: "chat",  // "chat" | "completion"
       char: "Llama",
       user: "User",
+      image_selected: ''
     })
 
     const params = signal({
@@ -220,7 +227,9 @@
       mirostat_tau: 5, // target entropy
       mirostat_eta: 0.1, // learning rate
       grammar: '',
-      n_probs: 0, // no completion_probabilities
+      n_probs: 0, // no completion_probabilities,
+      image_data: [],
+      cache_prompt: true
     })
 
     /* START: Support for storing prompt templates and parameters in borwser LocalStorage */
@@ -270,6 +279,7 @@
       // saved templates were successfuly imported.
 
       console.log('Processing saved templates and updating default template')
+      params.value = { ...params.value, image_data: [] };
 
       //console.log(importedTemplates);
       savedUserTemplates.value = importedTemplates;
@@ -294,7 +304,9 @@
 
     function userTemplateApply(t) {
       session.value = t.data.session;
+      session.value = { ...session.value, image_selected: '' };
       params.value = t.data.params;
+      params.value = { ...params.value, image_data: [] };
     }
 
     function userTemplateResetToDefaultAndApply() {
@@ -385,20 +397,25 @@
         throw new Error("already running");
       }
       controller.value = new AbortController();
-      for await (const chunk of llama(prompt, llamaParams, {controller: controller.value})) {
+      for await (const chunk of llama(prompt, llamaParams, { controller: controller.value })) {
         const data = chunk.data;
 
         if (data.stop) {
           while (
             currentMessages.length > 0 &&
             currentMessages[currentMessages.length - 1].content.match(/\n$/) != null
-            ) {
+          ) {
             currentMessages.pop();
           }
           transcriptUpdate([...history, [char, currentMessages]])
           console.log("Completion finished: '", currentMessages.map(msg => msg.content).join(''), "', summary: ", data);
         } else {
           currentMessages.push(data);
+          slot_id = data.slot_id;
+          if (selected_image && !data.multimodal) {
+            alert("The server was not compiled for multimodal or the model projector can't be loaded.");
+            return;
+          }
           transcriptUpdate([...history, [char, currentMessages]])
         }
 
@@ -419,7 +436,7 @@
 
       transcriptUpdate([...session.value.transcript, ["{{user}}", msg]])
 
-      const prompt = template(session.value.template, {
+      let prompt = template(session.value.template, {
         message: msg,
         history: session.value.transcript.flatMap(
           ([name, data]) =>
@@ -434,9 +451,12 @@
             )
         ).join("\n"),
       });
-
+      if (selected_image) {
+        prompt = `A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:[img-10]${msg}\nASSISTANT:`;
+      }
       await runLlama(prompt, {
         ...params.value,
+        slot_id: slot_id,
         stop: ["</s>", template("{{char}}:"), template("{{user}}:")],
       }, "{{char}}");
     }
@@ -446,10 +466,11 @@
         console.log('already running...');
         return;
       }
-      const {prompt} = session.value;
+      const { prompt } = session.value;
       transcriptUpdate([...session.value.transcript, ["", prompt]]);
       await runLlama(prompt, {
         ...params.value,
+        slot_id: slot_id,
         stop: [],
       }, "");
     }
@@ -467,6 +488,27 @@
       transcriptUpdate([]);
     }
 
+    const uploadImage = (e) => {
+      e.preventDefault();
+      document.getElementById("fileInput").click();
+      document.getElementById("fileInput").addEventListener("change", function (event) {
+        const selectedFile = event.target.files[0];
+        if (selectedFile) {
+          const reader = new FileReader();
+          reader.onload = function () {
+            const image_data = reader.result;
+            session.value = { ...session.value, image_selected: image_data };
+            params.value = {
+              ...params.value, image_data: [
+                { data: image_data.replace(/data:image\/[^;]+;base64,/, ''), id: 10 }]
+            }
+          };
+          selected_image = true;
+          reader.readAsDataURL(selectedFile);
+        }
+      });
+    }
+
     function MessageInput() {
       const message = useSignal("")
 
@@ -497,6 +539,7 @@
           </div>
           <div class="right">
             <button type="submit" disabled=${generating.value}>Send</button>
+            <button onclick=${uploadImage}>Upload Image</button>
             <button onclick=${stop} disabled=${!generating.value}>Stop</button>
             <button onclick=${reset}>Reset</button>
           </div>
@@ -540,7 +583,7 @@
             data;
           message = html`<${Markdownish} text=${template(text)} />`
         }
-        if(user) {
+        if (user) {
           return html`<p key=${index}><strong>${template(user)}:</strong> ${message}</p>`
         } else {
           return html`<p key=${index}>${message}</p>`
@@ -549,6 +592,7 @@
 
       return html`
         <section id="chat" ref=${container}>
+          <img style="width: 60%;${!session.value.image_selected ? `display: none;` : ``}" src="${session.value.image_selected}"/>
           ${messages.flatMap(chatLine)}
         </section>`;
     };
@@ -567,7 +611,7 @@
           const converter = new SchemaConverter(
             grammarJsonSchemaPropOrder.value
               .split(',')
-              .reduce((acc, cur, i) => ({...acc, [cur.trim()]: i}), {})
+              .reduce((acc, cur, i) => ({ ...acc, [cur.trim()]: i }), {})
           )
           converter.visit(schema, '')
           params.value = {
@@ -579,7 +623,7 @@
         }
       }
 
-      const FloatField = ({label, max, min, name, step, value}) => {
+      const FloatField = ({ label, max, min, name, step, value }) => {
         return html`
           <div>
             <label for="${name}">${label}</label>
@@ -589,7 +633,7 @@
         `
       };
 
-      const IntField = ({label, max, min, name, value}) => {
+      const IntField = ({ label, max, min, name, value }) => {
         return html`
           <div>
             <label for="${name}">${label}</label>
@@ -672,7 +716,7 @@
             ${GrammarControl()}
           </fieldset>
       `
-    );
+      );
 
       const CompletionConfigForm = () => (
         html`
@@ -694,20 +738,20 @@
           ${session.value.type === 'chat' ? ChatConfigForm() : CompletionConfigForm()}
 
           <fieldset class="two">
-            ${IntField({label: "Predictions", max: 2048, min: -1, name: "n_predict", value: params.value.n_predict})}
-            ${FloatField({label: "Temperature", max: 1.5, min: 0.0, name: "temperature", step: 0.01, value: params.value.temperature})}
-            ${FloatField({label: "Penalize repeat sequence", max: 2.0, min: 0.0, name: "repeat_penalty", step: 0.01, value: params.value.repeat_penalty})}
-            ${IntField({label: "Consider N tokens for penalize", max: 2048, min: 0, name: "repeat_last_n", value: params.value.repeat_last_n})}
-            ${IntField({label: "Top-K sampling", max: 100, min: -1, name: "top_k", value: params.value.top_k})}
-            ${FloatField({label: "Top-P sampling", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p})}
+            ${IntField({ label: "Predictions", max: 2048, min: -1, name: "n_predict", value: params.value.n_predict })}
+            ${FloatField({ label: "Temperature", max: 1.5, min: 0.0, name: "temperature", step: 0.01, value: params.value.temperature })}
+            ${FloatField({ label: "Penalize repeat sequence", max: 2.0, min: 0.0, name: "repeat_penalty", step: 0.01, value: params.value.repeat_penalty })}
+            ${IntField({ label: "Consider N tokens for penalize", max: 2048, min: 0, name: "repeat_last_n", value: params.value.repeat_last_n })}
+            ${IntField({ label: "Top-K sampling", max: 100, min: -1, name: "top_k", value: params.value.top_k })}
+            ${FloatField({ label: "Top-P sampling", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p })}
           </fieldset>
           <details>
             <summary>More options</summary>
             <fieldset class="two">
-              ${FloatField({label: "TFS-Z", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z})}
-              ${FloatField({label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p})}
-              ${FloatField({label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty})}
-              ${FloatField({label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty})}
+              ${FloatField({ label: "TFS-Z", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })}
+              ${FloatField({ label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p })}
+              ${FloatField({ label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })}
+              ${FloatField({ label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })}
             </fieldset>
             <hr />
             <fieldset class="three">
@@ -716,11 +760,11 @@
                 <label><input type="radio" name="mirostat" value="1" checked=${params.value.mirostat == 1} oninput=${updateParamsInt} /> Mirostat v1</label>
                 <label><input type="radio" name="mirostat" value="2" checked=${params.value.mirostat == 2} oninput=${updateParamsInt} /> Mirostat v2</label>
               </div>
-              ${FloatField({label: "Mirostat tau", max: 10.0, min: 0.0, name: "mirostat_tau", step: 0.01, value: params.value.mirostat_tau})}
-              ${FloatField({label: "Mirostat eta", max: 1.0, min: 0.0, name: "mirostat_eta", step: 0.01, value: params.value.mirostat_eta})}
+              ${FloatField({ label: "Mirostat tau", max: 10.0, min: 0.0, name: "mirostat_tau", step: 0.01, value: params.value.mirostat_tau })}
+              ${FloatField({ label: "Mirostat eta", max: 1.0, min: 0.0, name: "mirostat_eta", step: 0.01, value: params.value.mirostat_eta })}
             </fieldset>
             <fieldset>
-              ${IntField({label: "Show Probabilities", max: 10, min: 0, name: "n_probs", value: params.value.n_probs})}
+              ${IntField({ label: "Show Probabilities", max: 10, min: 0, name: "n_probs", value: params.value.n_probs })}
             </fieldset>
           </details>
         </form>
@@ -759,20 +803,20 @@
         const popoverChildren = html`
           <div class="prob-set">
             ${probs.map((p, index) => {
-              return html`
+          return html`
                 <div
                   key=${index}
                   title=${`prob: ${p.prob}`}
                   style=${{
-                    padding: '0.3em',
-                    backgroundColor: p.tok_str === content ? probColor(p.prob) : 'transparent'
-                  }}
+              padding: '0.3em',
+              backgroundColor: p.tok_str === content ? probColor(p.prob) : 'transparent'
+            }}
                 >
                   <span>${p.tok_str}: </span>
                   <span>${Math.floor(p.prob * 100)}%</span>
                 </div>
               `
-            })}
+        })}
           </div>
         `
 
@@ -851,9 +895,9 @@
               ref=${popoverRef}
               class="popover-content"
               style=${{
-                top: position.value.top,
-                left: position.value.left,
-              }}
+            top: position.value.top,
+            left: position.value.left,
+          }}
             >
               ${props.popoverChildren}
             </div>
@@ -952,8 +996,11 @@
 </head>
 
 <body>
-  <div id="container"></div>
+  <div id="container">
+    <input type="file" id="fileInput" accept="image/*" style="display: none;">
+  </div>
   <div id="portal"></div>
 </body>
 
 </html>
+

examples/simple/simple.cpp

@@ -95,13 +95,8 @@ int main(int argc, char ** argv) {
     llama_batch batch = llama_batch_init(512, 0, 1);
 
     // evaluate the initial prompt
-    batch.n_tokens = tokens_list.size();
-
-    for (int32_t i = 0; i < batch.n_tokens; i++) {
-        batch.token[i]  = tokens_list[i];
-        batch.pos[i]    = i;
-        batch.seq_id[i] = 0;
-        batch.logits[i] = false;
+    for (size_t i = 0; i < tokens_list.size(); i++) {
+        llama_batch_add(batch, tokens_list[i], i, { 0 }, false);
     }
 
     // llama_decode will output logits only for the last token of the prompt
@@ -138,7 +133,7 @@ int main(int argc, char ** argv) {
             const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
             // is it an end of stream?
-            if (new_token_id == llama_token_eos(ctx) || n_cur == n_len) {
+            if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
                 LOG_TEE("\n");
 
                 break;
@@ -148,15 +143,10 @@ int main(int argc, char ** argv) {
             fflush(stdout);
 
             // prepare the next batch
-            batch.n_tokens = 0;
+            llama_batch_clear(batch);
 
             // push this new token for next evaluation
-            batch.token [batch.n_tokens] = new_token_id;
-            batch.pos   [batch.n_tokens] = n_cur;
-            batch.seq_id[batch.n_tokens] = 0;
-            batch.logits[batch.n_tokens] = true;
-
-            batch.n_tokens += 1;
+            llama_batch_add(batch, new_token_id, n_cur, { 0 }, true);
 
             n_decode += 1;
         }

examples/speculative/speculative.cpp

@@ -8,6 +8,9 @@
 #include <string>
 #include <vector>
 
+#define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  100
+#define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
+
 struct seq_draft {
     bool active   = false;
     bool drafting = false;
@@ -64,6 +67,33 @@ int main(int argc, char ** argv) {
     params.n_gpu_layers = params.n_gpu_layers_draft;
     std::tie(model_dft, ctx_dft) = llama_init_from_gpt_params(params);
 
+    {
+        const int n_vocab_tgt = llama_n_vocab(model_tgt);
+        const int n_vocab_dft = llama_n_vocab(model_dft);
+        const int vocab_diff  = n_vocab_tgt > n_vocab_dft
+            ? n_vocab_tgt - n_vocab_dft
+            : n_vocab_dft - n_vocab_tgt;
+
+        if (vocab_diff > SPEC_VOCAB_MAX_SIZE_DIFFERENCE) {
+            fprintf(stderr, "%s: error: draft model vocab must closely match target model to use speculation but ", __func__);
+            fprintf(stderr, "target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
+                    n_vocab_tgt, llama_n_vocab(model_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
+            return 1;
+        }
+
+        for (int i = SPEC_VOCAB_CHECK_START_TOKEN_ID; i < std::min(n_vocab_tgt, n_vocab_dft); ++i) {
+            const char * token_text_tgt = llama_token_get_text(model_tgt, i);
+            const char * token_text_dft = llama_token_get_text(model_dft, i);
+            if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
+                fprintf(stderr, "%s: error: draft model vocab must match target model to use speculation but ", __func__);
+                fprintf(stderr, "token %d content differs - target '%s', draft '%s'\n", i,
+                        llama_token_to_piece(ctx_tgt, i).c_str(),
+                        llama_token_to_piece(ctx_dft, i).c_str());
+                return 1;
+            }
+        }
+    }
+
     // tokenize the prompt
     std::vector<llama_token> inp;
     inp = ::llama_tokenize(ctx_tgt, params.prompt, true);
@@ -118,7 +148,7 @@ int main(int argc, char ** argv) {
     std::vector<seq_draft> drafts(n_seq_dft);
 
     params.sparams.grammar.clear(); // the draft samplers will copy the target sampler's grammar
-    params.sparams.temp = std::max(0.01f, params.sparams.temp);
+    params.sparams.temp = -1.0f;    // force greedy sampling with probs for the draft model
 
     for (int s = 0; s < n_seq_dft; ++s) {
         drafts[s].ctx_sampling = llama_sampling_init(params.sparams);
@@ -163,7 +193,7 @@ int main(int argc, char ** argv) {
             printf("%s", token_str.c_str());
             fflush(stdout);
 
-            if (id == llama_token_eos(ctx_tgt)) {
+            if (id == llama_token_eos(model_tgt)) {
                 has_eos = true;
             }
 
@@ -227,6 +257,7 @@ int main(int argc, char ** argv) {
             llama_batch_add  (batch_dft, id, n_past_dft, { 0 }, true);
 
             llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1);
+            // LOG("dft batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_dft, batch_dft).c_str());
             llama_decode         (ctx_dft, batch_dft);
 
             ++n_past_dft;
@@ -370,7 +401,7 @@ int main(int argc, char ** argv) {
                 llama_kv_cache_seq_cp(ctx_tgt, 0, s, -1, -1);
             }
 
-            //LOG("target batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_tgt, batch_tgt));
+            // LOG("target batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_tgt, batch_tgt).c_str());
             llama_decode(ctx_tgt, batch_tgt);
             ++n_past_tgt;
         }

flake.lock

@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1692913444,
-        "narHash": "sha256-1SvMQm2DwofNxXVtNWWtIcTh7GctEVrS/Xel/mdc6iY=",
+        "lastModified": 1698134075,
+        "narHash": "sha256-foCD+nuKzfh49bIoiCBur4+Fx1nozo+4C/6k8BYk4sg=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "18324978d632ffc55ef1d928e81630c620f4f447",
+        "rev": "8efd5d1e283604f75a808a20e6cde0ef313d07d4",
         "type": "github"
       },
       "original": {

flake.nix

@@ -51,6 +51,9 @@
         };
         llama-python =
           pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece ]);
+        # TODO(Green-Sky): find a better way to opt-into the heavy ml python runtime
+        llama-python-extra =
+          pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece torchWithoutCuda transformers ]);
         postPatch = ''
           substituteInPlace ./ggml-metal.m \
             --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
@@ -126,5 +129,9 @@
           buildInputs = [ llama-python ];
           packages = nativeBuildInputs ++ osSpecific;
         };
+        devShells.extra = pkgs.mkShell {
+          buildInputs = [ llama-python-extra ];
+          packages = nativeBuildInputs ++ osSpecific;
+        };
       });
 }

ggml-cuda.cu

@@ -29,6 +29,8 @@
 #define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
 #define cublasCreate hipblasCreate
 #define cublasGemmEx hipblasGemmEx
+#define cublasGemmBatchedEx hipblasGemmBatchedEx
+#define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx
 #define cublasHandle_t hipblasHandle_t
 #define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS
 #define cublasSetStream hipblasSetStream
@@ -85,6 +87,24 @@
 #define CC_OFFSET_AMD 1000000
 #define CC_RDNA2      (CC_OFFSET_AMD + 1030)
 
+// define this if you want to always fallback to MMQ kernels and not use cuBLAS for matrix multiplication
+// on modern hardware, using cuBLAS is recommended as it utilizes F16 tensor cores which are very performant
+// for large computational tasks. the drawback is that this requires some extra amount of VRAM:
+// -  7B quantum model: +100-200 MB
+// - 13B quantum model: +200-400 MB
+//
+//#define GGML_CUDA_FORCE_MMQ
+
+// TODO: improve this to be correct for more hardware
+//       for example, currently fails for GeForce GTX 1660 which is TURING arch (> VOLTA) but does not have tensor cores
+//       probably other such cases, and not sure what happens on AMD hardware
+#if !defined(GGML_CUDA_FORCE_MMQ)
+#define CUDA_USE_TENSOR_CORES
+#endif
+
+// max batch size to use MMQ kernels when tensor cores are available
+#define MMQ_MAX_BATCH_SIZE 32
+
 #if defined(GGML_USE_HIPBLAS)
 #define __CUDA_ARCH__ 1300
 
@@ -468,7 +488,6 @@ static int g_device_count = -1;
 static int g_main_device = 0;
 static int g_compute_capabilities[GGML_CUDA_MAX_DEVICES];
 static float g_tensor_split[GGML_CUDA_MAX_DEVICES] = {0};
-static bool g_mul_mat_q = true;
 
 static void * g_scratch_buffer = nullptr;
 static size_t g_scratch_size = 0; // disabled by default
@@ -3552,9 +3571,15 @@ static __device__ __forceinline__ void mul_mat_q(
 #define  MMQ_X_Q4_0_RDNA1  64
 #define  MMQ_Y_Q4_0_RDNA1  64
 #define NWARPS_Q4_0_RDNA1  8
+#if defined(CUDA_USE_TENSOR_CORES)
+#define  MMQ_X_Q4_0_AMPERE 4
+#define  MMQ_Y_Q4_0_AMPERE 32
+#define NWARPS_Q4_0_AMPERE 4
+#else
 #define  MMQ_X_Q4_0_AMPERE 64
 #define  MMQ_Y_Q4_0_AMPERE 128
 #define NWARPS_Q4_0_AMPERE 4
+#endif
 #define  MMQ_X_Q4_0_PASCAL 64
 #define  MMQ_Y_Q4_0_PASCAL 64
 #define NWARPS_Q4_0_PASCAL 8
@@ -3613,9 +3638,15 @@ template <bool need_check> static __global__ void
 #define  MMQ_X_Q4_1_RDNA1  64
 #define  MMQ_Y_Q4_1_RDNA1  64
 #define NWARPS_Q4_1_RDNA1  8
+#if defined(CUDA_USE_TENSOR_CORES)
+#define  MMQ_X_Q4_1_AMPERE 4
+#define  MMQ_Y_Q4_1_AMPERE 32
+#define NWARPS_Q4_1_AMPERE 4
+#else
 #define  MMQ_X_Q4_1_AMPERE 64
 #define  MMQ_Y_Q4_1_AMPERE 128
 #define NWARPS_Q4_1_AMPERE 4
+#endif
 #define  MMQ_X_Q4_1_PASCAL 64
 #define  MMQ_Y_Q4_1_PASCAL 64
 #define NWARPS_Q4_1_PASCAL 8
@@ -3676,9 +3707,15 @@ template <bool need_check> static __global__ void
 #define  MMQ_X_Q5_0_RDNA1  64
 #define  MMQ_Y_Q5_0_RDNA1  64
 #define NWARPS_Q5_0_RDNA1  8
+#if defined(CUDA_USE_TENSOR_CORES)
+#define  MMQ_X_Q5_0_AMPERE 4
+#define  MMQ_Y_Q5_0_AMPERE 32
+#define NWARPS_Q5_0_AMPERE 4
+#else
 #define  MMQ_X_Q5_0_AMPERE 128
 #define  MMQ_Y_Q5_0_AMPERE 64
 #define NWARPS_Q5_0_AMPERE 4
+#endif
 #define  MMQ_X_Q5_0_PASCAL 64
 #define  MMQ_Y_Q5_0_PASCAL 64
 #define NWARPS_Q5_0_PASCAL 8
@@ -3737,9 +3774,15 @@ template <bool need_check> static __global__ void
 #define  MMQ_X_Q5_1_RDNA1  64
 #define  MMQ_Y_Q5_1_RDNA1  64
 #define NWARPS_Q5_1_RDNA1  8
+#if defined(CUDA_USE_TENSOR_CORES)
+#define  MMQ_X_Q5_1_AMPERE 4
+#define  MMQ_Y_Q5_1_AMPERE 32
+#define NWARPS_Q5_1_AMPERE 4
+#else
 #define  MMQ_X_Q5_1_AMPERE 128
 #define  MMQ_Y_Q5_1_AMPERE 64
 #define NWARPS_Q5_1_AMPERE 4
+#endif
 #define  MMQ_X_Q5_1_PASCAL 64
 #define  MMQ_Y_Q5_1_PASCAL 64
 #define NWARPS_Q5_1_PASCAL 8
@@ -3798,9 +3841,15 @@ mul_mat_q5_1(
 #define  MMQ_X_Q8_0_RDNA1  64
 #define  MMQ_Y_Q8_0_RDNA1  64
 #define NWARPS_Q8_0_RDNA1  8
+#if defined(CUDA_USE_TENSOR_CORES)
+#define  MMQ_X_Q8_0_AMPERE 4
+#define  MMQ_Y_Q8_0_AMPERE 32
+#define NWARPS_Q8_0_AMPERE 4
+#else
 #define  MMQ_X_Q8_0_AMPERE 128
 #define  MMQ_Y_Q8_0_AMPERE 64
 #define NWARPS_Q8_0_AMPERE 4
+#endif
 #define  MMQ_X_Q8_0_PASCAL 64
 #define  MMQ_Y_Q8_0_PASCAL 64
 #define NWARPS_Q8_0_PASCAL 8
@@ -3859,9 +3908,15 @@ template <bool need_check> static __global__ void
 #define  MMQ_X_Q2_K_RDNA1  128
 #define  MMQ_Y_Q2_K_RDNA1  32
 #define NWARPS_Q2_K_RDNA1  8
+#if defined(CUDA_USE_TENSOR_CORES)
+#define  MMQ_X_Q2_K_AMPERE 4
+#define  MMQ_Y_Q2_K_AMPERE 32
+#define NWARPS_Q2_K_AMPERE 4
+#else
 #define  MMQ_X_Q2_K_AMPERE 64
 #define  MMQ_Y_Q2_K_AMPERE 128
 #define NWARPS_Q2_K_AMPERE 4
+#endif
 #define  MMQ_X_Q2_K_PASCAL 64
 #define  MMQ_Y_Q2_K_PASCAL 64
 #define NWARPS_Q2_K_PASCAL 8
@@ -3920,9 +3975,15 @@ mul_mat_q2_K(
 #define  MMQ_X_Q3_K_RDNA1  32
 #define  MMQ_Y_Q3_K_RDNA1  128
 #define NWARPS_Q3_K_RDNA1  8
+#if defined(CUDA_USE_TENSOR_CORES)
+#define  MMQ_X_Q3_K_AMPERE 4
+#define  MMQ_Y_Q3_K_AMPERE 32
+#define NWARPS_Q3_K_AMPERE 4
+#else
 #define  MMQ_X_Q3_K_AMPERE 128
 #define  MMQ_Y_Q3_K_AMPERE 128
 #define NWARPS_Q3_K_AMPERE 4
+#endif
 #define  MMQ_X_Q3_K_PASCAL 64
 #define  MMQ_Y_Q3_K_PASCAL 64
 #define NWARPS_Q3_K_PASCAL 8
@@ -3983,9 +4044,15 @@ template <bool need_check> static __global__ void
 #define  MMQ_X_Q4_K_RDNA1  32
 #define  MMQ_Y_Q4_K_RDNA1  64
 #define NWARPS_Q4_K_RDNA1  8
+#if defined(CUDA_USE_TENSOR_CORES)
+#define  MMQ_X_Q4_K_AMPERE 4
+#define  MMQ_Y_Q4_K_AMPERE 32
+#define NWARPS_Q4_K_AMPERE 4
+#else
 #define  MMQ_X_Q4_K_AMPERE 64
 #define  MMQ_Y_Q4_K_AMPERE 128
 #define NWARPS_Q4_K_AMPERE 4
+#endif
 #define  MMQ_X_Q4_K_PASCAL 64
 #define  MMQ_Y_Q4_K_PASCAL 64
 #define NWARPS_Q4_K_PASCAL 8
@@ -4046,9 +4113,15 @@ template <bool need_check> static __global__ void
 #define  MMQ_X_Q5_K_RDNA1  32
 #define  MMQ_Y_Q5_K_RDNA1  64
 #define NWARPS_Q5_K_RDNA1  8
+#if defined(CUDA_USE_TENSOR_CORES)
+#define  MMQ_X_Q5_K_AMPERE 4
+#define  MMQ_Y_Q5_K_AMPERE 32
+#define NWARPS_Q5_K_AMPERE 4
+#else
 #define  MMQ_X_Q5_K_AMPERE 64
 #define  MMQ_Y_Q5_K_AMPERE 128
 #define NWARPS_Q5_K_AMPERE 4
+#endif
 #define  MMQ_X_Q5_K_PASCAL 64
 #define  MMQ_Y_Q5_K_PASCAL 64
 #define NWARPS_Q5_K_PASCAL 8
@@ -4107,9 +4180,15 @@ mul_mat_q5_K(
 #define  MMQ_X_Q6_K_RDNA1  32
 #define  MMQ_Y_Q6_K_RDNA1  64
 #define NWARPS_Q6_K_RDNA1  8
+#if defined(CUDA_USE_TENSOR_CORES)
+#define  MMQ_X_Q6_K_AMPERE 4
+#define  MMQ_Y_Q6_K_AMPERE 32
+#define NWARPS_Q6_K_AMPERE 4
+#else
 #define  MMQ_X_Q6_K_AMPERE 64
 #define  MMQ_Y_Q6_K_AMPERE 64
 #define NWARPS_Q6_K_AMPERE 4
+#endif
 #define  MMQ_X_Q6_K_PASCAL 64
 #define  MMQ_Y_Q6_K_PASCAL 64
 #define NWARPS_Q6_K_PASCAL 8
@@ -4326,13 +4405,13 @@ static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
 
     const half * x = (const half *) vx;
 
-    const int row_x = blockDim.y*blockIdx.y + threadIdx.y;
-    const int channel = blockDim.z*blockIdx.z + threadIdx.z;
+    const int row_x     = blockDim.y*blockIdx.y + threadIdx.y;
+    const int channel   = blockDim.z*blockIdx.z + threadIdx.z;
     const int channel_x = channel / channel_x_divisor;
 
-    const int nrows_y = ncols_x;
+    const int nrows_y   = ncols_x;
     const int nrows_dst = nrows_x;
-    const int row_dst = row_x;
+    const int row_dst   = row_x;
 
     const int idst = channel*nrows_dst + row_dst;
 
@@ -4345,13 +4424,13 @@ static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
             break;
         }
 
-        const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x;
-        const float xi = __half2float(x[ix]);
-
         const int row_y = col_x;
 
+        const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x;
         const int iy = channel*nrows_y + row_y;
 
+        const float xi = __half2float(x[ix]);
+
         tmp += xi * y[iy];
     }
 
@@ -5661,11 +5740,21 @@ void ggml_init_cublas() {
         CUDA_CHECK(cudaGetDeviceCount(&g_device_count));
         GGML_ASSERT(g_device_count <= GGML_CUDA_MAX_DEVICES);
         int64_t total_vram = 0;
+#if defined(GGML_CUDA_FORCE_MMQ)
+        fprintf(stderr, "%s: GGML_CUDA_FORCE_MMQ:   yes\n", __func__);
+#else
+        fprintf(stderr, "%s: GGML_CUDA_FORCE_MMQ:   no\n", __func__);
+#endif
+#if defined(CUDA_USE_TENSOR_CORES)
+        fprintf(stderr, "%s: CUDA_USE_TENSOR_CORES: yes\n", __func__);
+#else
+        fprintf(stderr, "%s: CUDA_USE_TENSOR_CORES: no\n", __func__);
+#endif
         fprintf(stderr, "%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, g_device_count);
-        for (int64_t id = 0; id < g_device_count; ++id) {
+        for (int id = 0; id < g_device_count; ++id) {
             cudaDeviceProp prop;
             CUDA_CHECK(cudaGetDeviceProperties(&prop, id));
-            fprintf(stderr, "  Device %ld: %s, compute capability %d.%d\n", id, prop.name, prop.major, prop.minor);
+            fprintf(stderr, "  Device %d: %s, compute capability %d.%d\n", id, prop.name, prop.major, prop.minor);
 
             g_tensor_split[id] = total_vram;
             total_vram += prop.totalGlobalMem;
@@ -5675,15 +5764,15 @@ void ggml_init_cublas() {
             g_compute_capabilities[id] = 100*prop.major + 10*prop.minor;
 #endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
         }
-        for (int64_t id = 0; id < g_device_count; ++id) {
+        for (int id = 0; id < g_device_count; ++id) {
             g_tensor_split[id] /= total_vram;
         }
 
-        for (int64_t id = 0; id < g_device_count; ++id) {
+        for (int id = 0; id < g_device_count; ++id) {
             CUDA_CHECK(ggml_cuda_set_device(id));
 
             // create cuda streams
-            for (int64_t is = 0; is < MAX_STREAMS; ++is) {
+            for (int is = 0; is < MAX_STREAMS; ++is) {
                 CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams[id][is], cudaStreamNonBlocking));
             }
 
@@ -6252,16 +6341,15 @@ inline void ggml_cuda_op_mul_mat_cublas(
     const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
     const int64_t src1_padded_row_size, const cudaStream_t & stream) {
 
-    GGML_ASSERT(src0_dd_i != nullptr);
+    GGML_ASSERT(src0_dd_i  != nullptr);
     GGML_ASSERT(src1_ddf_i != nullptr);
-    GGML_ASSERT(dst_dd_i != nullptr);
-
+    GGML_ASSERT(dst_dd_i   != nullptr);
 
     const int64_t ne00 = src0->ne[0];
-
     const int64_t ne10 = src1->ne[0];
 
     const int64_t ne0 = dst->ne[0];
+
     const int64_t row_diff = row_high - row_low;
 
     int id;
@@ -6346,7 +6434,7 @@ inline void ggml_cuda_op_mul_mat_cublas(
             cublasSgemm(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
                     row_diff, src1_ncols, ne10,
                     &alpha, src0_ddf_i, ne00,
-                            src1_ddf_i,  ne10,
+                            src1_ddf_i, ne10,
                     &beta,  dst_dd_i,   ldc));
 
         if (src0_as != 0) {
@@ -7013,7 +7101,8 @@ static void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tens
 }
 
 static void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){
-    GGML_ASSERT(!ggml_is_contiguous(src0) && ggml_is_contiguous(src1));
+    GGML_ASSERT(!ggml_is_transposed(src0));
+    GGML_ASSERT(!ggml_is_transposed(src1));
     GGML_ASSERT(!ggml_is_permuted(src0));
     GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT);
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
@@ -7023,11 +7112,11 @@ static void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor
     const int64_t ne01 = src0->ne[1];
     const int64_t ne02 = src0->ne[2];
 
-    const int64_t ne12 = src1->ne[2];
-
     const int64_t nb01 = src0->nb[1];
     const int64_t nb02 = src0->nb[2];
 
+    const int64_t ne12 = src1->ne[2];
+
     CUDA_CHECK(ggml_cuda_set_device(g_main_device));
     cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
 
@@ -7046,27 +7135,200 @@ static void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor
     ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, main_stream);
 }
 
+static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(!ggml_is_transposed(src0));
+    GGML_ASSERT(!ggml_is_transposed(src1));
+
+    GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT);
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0]; GGML_UNUSED(ne00);
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[3];
+
+    const int64_t nb01 = src0->nb[1];
+    const int64_t nb02 = src0->nb[2]; GGML_UNUSED(nb02);
+    const int64_t nb03 = src0->nb[3]; GGML_UNUSED(nb03);
+
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+    const int64_t ne12 = src1->ne[2];
+    const int64_t ne13 = src1->ne[3];
+
+    const int64_t nb11 = src1->nb[1];
+    const int64_t nb12 = src1->nb[2]; GGML_UNUSED(nb12);
+    const int64_t nb13 = src1->nb[3]; GGML_UNUSED(nb13);
+
+    const int64_t ne1 = ggml_nelements(src1);
+    const int64_t ne  = ggml_nelements(dst);
+
+    CUDA_CHECK(ggml_cuda_set_device(g_main_device));
+    cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
+
+    int id;
+    CUDA_CHECK(cudaGetDevice(&id));
+    CUBLAS_CHECK(cublasSetStream(g_cublas_handles[id], main_stream));
+
+    ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
+    void * src0_ddq = src0_extra->data_device[g_main_device];
+    half * src0_as_f16 = (half *) src0_ddq;
+
+    ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
+    float * src1_ddf = (float *) src1_extra->data_device[g_main_device];
+
+    ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
+    float * dst_ddf = (float *) dst_extra->data_device[g_main_device];
+
+    // convert src1 to fp16
+    const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
+    GGML_ASSERT(to_fp16_cuda != nullptr);
+
+    size_t src1_as = 0;
+    half * src1_as_f16 = (half *) ggml_cuda_pool_malloc(ne1 * sizeof(half), &src1_as);
+    to_fp16_cuda(src1_ddf, src1_as_f16, ne1, main_stream);
+
+    size_t dst_as = 0;
+    half * dst_f16 = (half *) ggml_cuda_pool_malloc(ne * sizeof(half), &dst_as);
+
+    GGML_ASSERT(ne12 % ne02 == 0);
+    GGML_ASSERT(ne13 % ne03 == 0);
+
+    // broadcast factors
+    const int64_t r2 = ne12/ne02;
+    const int64_t r3 = ne13/ne03;
+
+    const half alpha_f16 = 1.0f;
+    const half beta_f16  = 0.0f;
+
+#if 0
+    // use cublasGemmEx
+    {
+        for (int i13 = 0; i13 < ne13; ++i13) {
+            for (int i12 = 0; i12 < ne12; ++i12) {
+                int i03 = i13 / r3;
+                int i02 = i12 / r2;
+
+                CUBLAS_CHECK(
+                        cublasGemmEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
+                            ne01, ne11, ne10,
+                            &alpha_f16, (const char *) src0_as_f16 + i02*src0->nb[2]   + i03*src0->nb[3]  , CUDA_R_16F, nb01/sizeof(half),
+                                        (const char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, CUDA_R_16F, nb11/sizeof(float),
+                            &beta_f16,  (      char *)     dst_f16 + i12* dst->nb[2]/2 + i13* dst->nb[3]/2, CUDA_R_16F, ne01,
+                            CUBLAS_COMPUTE_16F,
+                            CUBLAS_GEMM_DEFAULT_TENSOR_OP));
+            }
+        }
+    }
+#else
+    if (r2 == 1 && r3 == 1 && src0->nb[2]*src0->ne[2] == src0->nb[3] && src1->nb[2]*src1->ne[2] == src1->nb[3]) {
+        // there is no broadcast and src0, src1 are contiguous across dims 2, 3
+        // use cublasGemmStridedBatchedEx
+        CUBLAS_CHECK(
+        cublasGemmStridedBatchedEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
+                ne01, ne11, ne10,
+                &alpha_f16, (const char *) src0_as_f16, CUDA_R_16F, nb01/sizeof(half),  src0->nb[2]/sizeof(half),  // strideA
+                            (const char *) src1_as_f16, CUDA_R_16F, nb11/sizeof(float), src1->nb[2]/sizeof(float), // strideB
+                &beta_f16,  (      char *)     dst_f16, CUDA_R_16F, ne01,                dst->nb[2]/sizeof(float), // strideC
+                ne12*ne13,
+                CUBLAS_COMPUTE_16F,
+                CUBLAS_GEMM_DEFAULT_TENSOR_OP));
+    } else {
+        // use cublasGemmBatchedEx
+        // TODO: https://github.com/ggerganov/llama.cpp/pull/3749#discussion_r1369997000
+        const int ne23 = ne12*ne13;
+
+        // TODO: avoid this alloc
+        void ** ptrs = (void **) malloc(3*ne23*sizeof(void *));
+
+        for (int i13 = 0; i13 < ne13; ++i13) {
+            for (int i12 = 0; i12 < ne12; ++i12) {
+                int i03 = i13 / r3;
+                int i02 = i12 / r2;
+
+                ptrs[0*ne23 + i12 + i13*ne12] = (char *) src0_as_f16 + i02*src0->nb[2]   + i03*src0->nb[3];
+                ptrs[1*ne23 + i12 + i13*ne12] = (char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2;
+                ptrs[2*ne23 + i12 + i13*ne12] = (char *)     dst_f16 + i12* dst->nb[2]/2 + i13* dst->nb[3]/2;
+            }
+        }
+
+        // allocate device memory for pointers
+        void ** ptrs_as = nullptr;
+        CUDA_CHECK(cudaMalloc(&ptrs_as, 3*ne23*sizeof(void *)));
+
+        // TODO: this does not work for some reason -- not sure why?
+        //size_t ptrs_s = 0;
+        //ptrs_as = (void **) ggml_cuda_pool_malloc(3*ne23*sizeof(void *), &ptrs_s);
+
+        // copy pointers to device
+        CUDA_CHECK(cudaMemcpy(ptrs_as, ptrs, 3*ne23*sizeof(void *), cudaMemcpyHostToDevice));
+
+        free(ptrs);
+
+        CUBLAS_CHECK(
+        cublasGemmBatchedEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
+                ne01, ne11, ne10,
+                &alpha_f16, (const void **) (ptrs_as + 0*ne23), CUDA_R_16F, nb01/sizeof(half),
+                            (const void **) (ptrs_as + 1*ne23), CUDA_R_16F, nb11/sizeof(float),
+                &beta_f16,  (      void **) (ptrs_as + 2*ne23), CUDA_R_16F, ne01,
+                ne23,
+                CUBLAS_COMPUTE_16F,
+                CUBLAS_GEMM_DEFAULT_TENSOR_OP));
+
+        // free device memory for pointers
+        CUDA_CHECK(cudaFree(ptrs_as));
+        //ggml_cuda_pool_free(ptrs_as, ptrs_s);
+    }
+#endif
+
+    const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
+    to_fp32_cuda(dst_f16, dst_ddf, ne, main_stream);
+
+    ggml_cuda_pool_free(src1_as_f16, src1_as);
+    ggml_cuda_pool_free(dst_f16, dst_as);
+}
+
 static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    bool all_on_device = (src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT) &&
-        src1->backend == GGML_BACKEND_GPU && dst->backend == GGML_BACKEND_GPU;
+    const bool all_on_device =
+        (src0->backend == GGML_BACKEND_GPU) &&
+        (src1->backend == GGML_BACKEND_GPU) &&
+        ( dst->backend == GGML_BACKEND_GPU);
 
     int64_t min_compute_capability = INT_MAX;
     for (int64_t id = 0; id < g_device_count; ++id) {
-        if (min_compute_capability > g_compute_capabilities[id]
-                && g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) {
+        if (min_compute_capability > g_compute_capabilities[id] && g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) {
             min_compute_capability = g_compute_capabilities[id];
         }
     }
 
-    if (all_on_device && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
+#ifdef CUDA_USE_TENSOR_CORES
+    const bool use_tensor_cores = true;
+#else
+    const bool use_tensor_cores = false;
+#endif
+
+    // debug helpers
+    //printf("src0: %8d %8d %8d %8d\n", src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]);
+    //printf("      %8d %8d %8d %8d\n", src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]);
+    //printf("src1: %8d %8d %8d %8d\n", src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3]);
+    //printf("      %8d %8d %8d %8d\n", src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3]);
+    //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
+    //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
+
+    if (all_on_device && !use_tensor_cores && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
+        // KQ single-batch
         ggml_cuda_mul_mat_vec_p021(src0, src1, dst);
-    } else if (all_on_device && !ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && src1->ne[1] == 1) {
+    } else if (all_on_device && !use_tensor_cores && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
+        // KQV single-batch
         ggml_cuda_mul_mat_vec_nc(src0, src1, dst);
+    } else if (all_on_device && src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1)) {
+        // KQ + KQV multi-batch
+        ggml_cuda_mul_mat_mat_batched_cublas(src0, src1, dst);
     } else if (src0->type == GGML_TYPE_F32) {
         ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false);
     } else if (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16) {
         if (src1->ne[1] == 1 && src0->ne[0] % GGML_CUDA_DMMV_X == 0) {
-
 #ifdef GGML_CUDA_FORCE_DMMV
             const bool use_mul_mat_vec_q = false;
 #else
@@ -7079,7 +7341,15 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1
                 ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, false);
             }
         } else {
-            if (g_mul_mat_q && ggml_is_quantized(src0->type) && min_compute_capability >= MIN_CC_DP4A) {
+            bool use_mul_mat_q = min_compute_capability >= MIN_CC_DP4A && ggml_is_quantized(src0->type);
+
+            // when tensor cores are available, use them for large batch size
+            // ref: https://github.com/ggerganov/llama.cpp/pull/3776
+            if (use_tensor_cores && min_compute_capability >= CC_VOLTA && src1->ne[1] > MMQ_MAX_BATCH_SIZE) {
+                use_mul_mat_q = false;
+            }
+
+            if (use_mul_mat_q) {
                 ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_q, true);
             } else {
                 ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false);
@@ -7433,10 +7703,6 @@ void ggml_cuda_set_main_device(const int main_device) {
     }
 }
 
-void ggml_cuda_set_mul_mat_q(const bool mul_mat_q) {
-    g_mul_mat_q = mul_mat_q;
-}
-
 void ggml_cuda_set_scratch_size(const size_t scratch_size) {
     // this is a hack to not completely break llama.cpp when using multiple models or contexts simultaneously
     // it still won't always work as expected, but it's better than nothing

ggml-metal.m

@@ -62,6 +62,7 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(mul);
     GGML_METAL_DECL_KERNEL(mul_row); // TODO: avoid this extra kernel, instead extend the "mul" kernel to support broadcast
     GGML_METAL_DECL_KERNEL(scale);
+    GGML_METAL_DECL_KERNEL(scale_4);
     GGML_METAL_DECL_KERNEL(silu);
     GGML_METAL_DECL_KERNEL(relu);
     GGML_METAL_DECL_KERNEL(gelu);
@@ -209,6 +210,10 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
             GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
 
             NSString * sourcePath = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+            if (sourcePath == nil) {
+                GGML_METAL_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__);
+                sourcePath = @"ggml-metal.metal";
+            }
             GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [sourcePath UTF8String]);
             NSString * src = [NSString stringWithContentsOfFile:sourcePath encoding:NSUTF8StringEncoding error:&error];
             if (error) {
@@ -249,6 +254,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(mul);
         GGML_METAL_ADD_KERNEL(mul_row);
         GGML_METAL_ADD_KERNEL(scale);
+        GGML_METAL_ADD_KERNEL(scale_4);
         GGML_METAL_ADD_KERNEL(silu);
         GGML_METAL_ADD_KERNEL(relu);
         GGML_METAL_ADD_KERNEL(gelu);
@@ -347,6 +353,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(mul);
     GGML_METAL_DEL_KERNEL(mul_row);
     GGML_METAL_DEL_KERNEL(scale);
+    GGML_METAL_DEL_KERNEL(scale_4);
     GGML_METAL_DEL_KERNEL(silu);
     GGML_METAL_DEL_KERNEL(relu);
     GGML_METAL_DEL_KERNEL(gelu);
@@ -923,15 +930,20 @@ void ggml_metal_graph_compute(
 
                             const float scale = *(const float *) src1->data;
 
-                            [encoder setComputePipelineState:ctx->pipeline_scale];
+                            int64_t n = ggml_nelements(dst);
+
+                            if (n % 4 == 0) {
+                                n /= 4;
+                                [encoder setComputePipelineState:ctx->pipeline_scale_4];
+                            } else {
+                                [encoder setComputePipelineState:ctx->pipeline_scale];
+                            }
+
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
                             [encoder setBytes:&scale length:sizeof(scale) atIndex:2];
 
-                            const int64_t n = ggml_nelements(dst);
-                            GGML_ASSERT(n % 4 == 0);
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         } break;
                     case GGML_OP_UNARY:
                         switch (ggml_get_unary_op(gf->nodes[i])) {

ggml-metal.metal

@@ -125,9 +125,17 @@ kernel void kernel_mul_row(
 }
 
 kernel void kernel_scale(
+        device const float * src0,
+        device       float * dst,
+        constant     float & scale,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] * scale;
+}
+
+kernel void kernel_scale_4(
         device const float4 * src0,
         device       float4 * dst,
-        constant     float & scale,
+        constant     float  & scale,
         uint tpig[[thread_position_in_grid]]) {
     dst[tpig] = src0[tpig] * scale;
 }

ggml-quants.h

@@ -1,20 +1,14 @@
 #pragma once
 
+// This is a private API for quantization and dequantization
+// Should not be used directly, use ggml.h instead
+
 #include "ggml.h"
 
 #include <stdint.h>
 #include <assert.h>
 #include <stddef.h>
 
-// Super-block size
-#ifdef GGML_QKK_64
-#define QK_K 64
-#define K_SCALE_SIZE 4
-#else
-#define QK_K 256
-#define K_SCALE_SIZE 12
-#endif
-
 #ifndef static_assert
 #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
 #define static_assert(cond, msg) _Static_assert(cond, msg)
@@ -23,10 +17,66 @@
 #endif
 #endif
 
+#define QK4_0 32
+typedef struct {
+    ggml_fp16_t d;          // delta
+    uint8_t qs[QK4_0 / 2];  // nibbles / quants
+} block_q4_0;
+static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
+
+#define QK4_1 32
+typedef struct {
+    ggml_fp16_t d;          // delta
+    ggml_fp16_t m;          // min
+    uint8_t qs[QK4_1 / 2];  // nibbles / quants
+} block_q4_1;
+static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_fp16_t) + QK4_1 / 2, "wrong q4_1 block size/padding");
+
+#define QK5_0 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_0 / 2]; // nibbles / quants
+} block_q5_0;
+static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
+
+#define QK5_1 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    ggml_fp16_t m;         // min
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_1 / 2]; // nibbles / quants
+} block_q5_1;
+static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
+
+#define QK8_0 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    int8_t  qs[QK8_0];     // quants
+} block_q8_0;
+static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
+
+#define QK8_1 32
+typedef struct {
+    float d;               // delta
+    float s;               // d * sum(qs[i])
+    int8_t  qs[QK8_1];     // quants
+} block_q8_1;
+static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding");
+
 //
 // Super-block quantization structures
 //
 
+// Super-block size
+#ifdef GGML_QKK_64
+#define QK_K 64
+#define K_SCALE_SIZE 4
+#else
+#define QK_K 256
+#define K_SCALE_SIZE 12
+#endif
+
 // 2-bit quantization
 // weight is represented as x = a * q + b
 // 16 blocks of 16 elements each
@@ -127,6 +177,13 @@ static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_
 
 
 // Quantization
+void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k);
+void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k);
+void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k);
+void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k);
+void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k);
+void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int k);
+
 void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k);
 void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k);
 void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k);
@@ -134,6 +191,13 @@ void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict
 void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k);
 void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k);
 
+void quantize_row_q4_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q4_1(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_1(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_1(const float * restrict x, void * restrict y, int k);
+
 void quantize_row_q2_K(const float * restrict x, void * restrict y, int k);
 void quantize_row_q3_K(const float * restrict x, void * restrict y, int k);
 void quantize_row_q4_K(const float * restrict x, void * restrict y, int k);
@@ -142,6 +206,13 @@ void quantize_row_q6_K(const float * restrict x, void * restrict y, int k);
 void quantize_row_q8_K(const float * restrict x, void * restrict y, int k);
 
 // Dequantization
+void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k);
+void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int k);
+void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int k);
+void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int k);
+void dequantize_row_q8_0(const block_q8_0 * restrict x, float * restrict y, int k);
+//void dequantize_row_q8_1(const block_q8_1 * restrict x, float * restrict y, int k);
+
 void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k);
 void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k);
 void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k);
@@ -150,16 +221,14 @@ void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int
 void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k);
 
 // Dot product
+void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+
 void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-
-// Quantization with histogram collection
-size_t ggml_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist);
-size_t ggml_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist);
-size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist);
-size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist);
-size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist);
-

ggml.h

@@ -401,15 +401,16 @@ extern "C" {
         GGML_OP_ALIBI,
         GGML_OP_CLAMP,
         GGML_OP_CONV_1D,
-        GGML_OP_CONV_2D,
+        GGML_OP_CONV_1D_STAGE_0,  // internal
+        GGML_OP_CONV_1D_STAGE_1,  // internal
         GGML_OP_CONV_TRANSPOSE_1D,
+        GGML_OP_CONV_2D,
+        GGML_OP_CONV_2D_STAGE_0, // internal
+        GGML_OP_CONV_2D_STAGE_1, // internal
         GGML_OP_CONV_TRANSPOSE_2D,
         GGML_OP_POOL_1D,
         GGML_OP_POOL_2D,
 
-        GGML_OP_CONV_1D_STAGE_0,  // internal
-        GGML_OP_CONV_1D_STAGE_1,  // internal
-
         GGML_OP_UPSCALE, // nearest interpolate
 
         GGML_OP_FLASH_ATTN,
@@ -1020,9 +1021,9 @@ extern "C" {
             struct ggml_tensor  * b,
             float                 eps);
 
-    // A: n columns, m rows
-    // B: n columns, p rows  (i.e. we transpose it internally)
-    // result is m columns, p rows
+    // A: k columns, n rows => [ne03, ne02, n, k]
+    // B: k columns, m rows  (i.e. we transpose it internally) => [ne03 * x, ne02 * y, m, k]
+    // result is n columns, m rows => [ne03 * x, ne02 * y, m, n]
     GGML_API struct ggml_tensor * ggml_mul_mat(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1929,12 +1930,19 @@ extern "C" {
     // quantization
     //
 
+    // TODO: these would probably get removed in favor of the more general ggml_quantize_chunk
     GGML_API size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist);
     GGML_API size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist);
     GGML_API size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist);
     GGML_API size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * hist);
     GGML_API size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * hist);
 
+    GGML_API size_t ggml_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist);
+    GGML_API size_t ggml_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist);
+    GGML_API size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist);
+    GGML_API size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist);
+    GGML_API size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist);
+
     GGML_API size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist);
 
     //

gguf-py/gguf/gguf.py

@@ -987,12 +987,15 @@ class SpecialVocab:
     merges: list[str] = []
     special_token_types: tuple[str, ...] = ('bos', 'eos', 'unk', 'sep', 'pad')
     special_token_ids: dict[str, int] = {}
+    n_vocab: int | None = None
 
     def __init__(
         self, path: str | os.PathLike[str], load_merges: bool = False,
         special_token_types: tuple[str, ...] | None = None,
+        n_vocab: int | None = None,
     ):
         self.special_token_ids = {}
+        self.n_vocab = n_vocab
         self.load_merges = load_merges
         if special_token_types is not None:
             self.special_token_types = special_token_types
@@ -1002,6 +1005,16 @@ class SpecialVocab:
         if not self._try_load_from_tokenizer_json(path):
             self._try_load_from_config_json(path)
 
+    def _set_special_token(self, typ: str, tid: Any):
+        if not isinstance(tid, int) or tid < 0:
+            return
+        if self.n_vocab is None or tid < self.n_vocab:
+            self.special_token_ids[typ] = tid
+            return
+        print(f'gguf: WARNING: Special token type {typ}, id {tid} out of range, must be under {self.n_vocab} - skipping',
+            file = sys.stderr)
+
+
     def _try_load_from_tokenizer_json(self, path: Path) -> bool:
         tokenizer_file = path / 'tokenizer.json'
         if not tokenizer_file.is_file():
@@ -1029,10 +1042,11 @@ class SpecialVocab:
                 tc_content = entry_content
             else:
                 continue
-            for maybe_token_id in (atok.get('id') for atok in added_tokens if atok.get('content') == tc_content):
-                if isinstance(maybe_token_id, int) and maybe_token_id >= 0:
-                    self.special_token_ids[typ] = maybe_token_id
-                break
+            # We only need the first match here.
+            maybe_token_id = next((
+                atok.get('id') for atok in added_tokens
+                if atok.get('content') == tc_content), None)
+            self._set_special_token(typ, maybe_token_id)
         return True
 
     def _try_load_from_config_json(self, path: Path) -> bool:
@@ -1042,21 +1056,21 @@ class SpecialVocab:
         with open(config_file, encoding = 'utf-8') as f:
             config = json.load(f)
         for typ in self.special_token_types:
-            maybe_token_id = config.get(f'{typ}_token_id')
-            if isinstance(maybe_token_id, int) and maybe_token_id >= 0:
-                self.special_token_ids[typ] = maybe_token_id
+            self._set_special_token(typ, config.get(f'{typ}_token_id'))
         return True
 
-    def add_to_gguf(self, gw: GGUFWriter) -> None:
+    def add_to_gguf(self, gw: GGUFWriter, quiet: bool = False) -> None:
         if len(self.merges) > 0:
-            print(f'gguf: Adding {len(self.merges)} merge(s).')
+            if not quiet:
+                print(f'gguf: Adding {len(self.merges)} merge(s).')
             gw.add_token_merges(self.merges)
         for typ, tokid in self.special_token_ids.items():
             handler: Callable[[int], None] | None = getattr(gw, f'add_{typ}_token_id', None)
             if handler is None:
-                print(f'gguf: WARNING: No handler for special token type {typ} with id {tokid} - skipping')
+                print(f'gguf: WARNING: No handler for special token type {typ} with id {tokid} - skipping', file = sys.stderr)
                 continue
-            print(f'gguf: Setting special token type {typ} to {tokid}')
+            if not quiet:
+                print(f'gguf: Setting special token type {typ} to {tokid}')
             handler(tokid)
 
     def __repr__(self) -> str:

llama.cpp

@@ -19,13 +19,11 @@
 #ifdef GGML_USE_MPI
 #  include "ggml-mpi.h"
 #endif
-#ifdef GGML_USE_K_QUANTS
-#  ifndef QK_K
-#    ifdef GGML_QKK_64
-#      define QK_K 64
-#    else
-#      define QK_K 256
-#    endif
+#ifndef QK_K
+#  ifdef GGML_QKK_64
+#    define QK_K 64
+#  else
+#    define QK_K 256
 #  endif
 #endif
 
@@ -975,14 +973,15 @@ static void llama_nop(struct ggml_tensor * tensor) { // don't offload by default
     (void) tensor;
 }
 
-static std::string llama_token_to_str(const struct llama_context * ctx, llama_token token) {
+static std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) {
     std::vector<char> result(8, 0);
     const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
     if (n_tokens < 0) {
         result.resize(-n_tokens);
         int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
         GGML_ASSERT(check == -n_tokens);
-    } else {
+    }
+    else {
         result.resize(n_tokens);
     }
 
@@ -1202,10 +1201,10 @@ struct llama_vocab {
     id special_eot_id    = 32010;
 
     int find_bpe_rank(std::string token_left, std::string token_right) const {
-        replace_all(token_left,  " ",  "\u0120");
-        replace_all(token_left,  "\n", "\u010A");
-        replace_all(token_right, " ",  "\u0120");
-        replace_all(token_right, "\n", "\u010A");
+        GGML_ASSERT(token_left.find(" ") == std::string::npos);
+        GGML_ASSERT(token_left.find("\n") == std::string::npos);
+        GGML_ASSERT(token_right.find(" ") == std::string::npos);
+        GGML_ASSERT(token_right.find("\n") == std::string::npos);
 
         auto it = bpe_ranks.find(std::make_pair(token_left, token_right));
         if (it == bpe_ranks.end()) {
@@ -1467,17 +1466,12 @@ static int32_t llama_kv_cache_cell_max(const struct llama_kv_cache & cache) {
     return 0;
 }
 
-static void llama_kv_cache_tokens_rm(struct llama_kv_cache & cache, int32_t c0, int32_t c1) {
-    if (c0 < 0) c0 = 0;
-    if (c1 < 0) c1 = cache.size;
-
-    for (int32_t i = c0; i < c1; ++i) {
+static void llama_kv_cache_clear(struct llama_kv_cache & cache) {
+    for (int32_t i = 0; i < cache.size; ++i) {
         cache.cells[i].pos = -1;
         cache.cells[i].seq_id.clear();
     }
-
-    // Searching for a free slot can start here since we know it will be empty.
-    cache.head = uint32_t(c0);
+    cache.head = 0;
 }
 
 static void llama_kv_cache_seq_rm(
@@ -1491,8 +1485,14 @@ static void llama_kv_cache_seq_rm(
     if (p1 < 0) p1 = std::numeric_limits<llama_pos>::max();
 
     for (uint32_t i = 0; i < cache.size; ++i) {
-        if (cache.cells[i].has_seq_id(seq_id) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) {
-            cache.cells[i].seq_id.erase(seq_id);
+        if (cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) {
+            if (seq_id < 0) {
+                cache.cells[i].seq_id.clear();
+            } else if (cache.cells[i].has_seq_id(seq_id)) {
+                cache.cells[i].seq_id.erase(seq_id);
+            } else {
+                continue;
+            }
             if (cache.cells[i].seq_id.empty()) {
                 cache.cells[i].pos = -1;
                 if (new_head == cache.size) new_head = i;
@@ -1553,14 +1553,14 @@ static void llama_kv_cache_seq_shift(
 
     for (uint32_t i = 0; i < cache.size; ++i) {
         if (cache.cells[i].has_seq_id(seq_id) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) {
-            cache.cells[i].pos += delta;
+            cache.has_shift = true;
+            cache.cells[i].pos   += delta;
+            cache.cells[i].delta += delta;
+
             if (cache.cells[i].pos < 0) {
                 cache.cells[i].pos = -1;
                 cache.cells[i].seq_id.clear();
                 if (new_head == cache.size) new_head = i;
-            } else {
-                cache.has_shift = true;
-                cache.cells[i].delta = delta;
             }
         }
     }
@@ -1577,12 +1577,14 @@ static void llama_kv_cache_seq_shift(
 enum llama_fver {
     GGUF_FILE_VERSION_V1 = 1,
     GGUF_FILE_VERSION_V2 = 2,
+    GGUF_FILE_VERSION_V3 = 3,
 };
 
 static const char * llama_file_version_name(llama_fver version) {
     switch (version) {
         case GGUF_FILE_VERSION_V1: return "GGUF V1 (support until nov 2023)";
-        case GGUF_FILE_VERSION_V2: return "GGUF V2 (latest)";
+        case GGUF_FILE_VERSION_V2: return "GGUF V2";
+        case GGUF_FILE_VERSION_V3: return "GGUF V3 (latest)";
     }
 
     return "unknown";
@@ -2238,15 +2240,35 @@ static void llm_load_vocab(
     if (vocab.type == LLAMA_VOCAB_TYPE_SPM) {
         vocab.linefeed_id = llama_byte_to_token(vocab, '\n');
     } else {
-        vocab.linefeed_id = llama_tokenize_internal(vocab, "\u010A", false)[0];
+        const std::vector<int> ids = llama_tokenize_internal(vocab, "\u010A", false);
+        GGML_ASSERT(!ids.empty() && "model vocab missing newline token");
+        vocab.linefeed_id = ids[0];
     }
 
     // special tokens
-    GGUF_GET_KEY(ctx, vocab.special_bos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_BOS_ID));
-    GGUF_GET_KEY(ctx, vocab.special_eos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_EOS_ID));
-    GGUF_GET_KEY(ctx, vocab.special_unk_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_UNK_ID));
-    GGUF_GET_KEY(ctx, vocab.special_sep_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_SEP_ID));
-    GGUF_GET_KEY(ctx, vocab.special_pad_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_PAD_ID));
+    {
+        const std::vector<std::pair<enum llm_kv, int32_t &>> special_token_types = {
+            { LLM_KV_TOKENIZER_BOS_ID, vocab.special_bos_id },
+            { LLM_KV_TOKENIZER_EOS_ID, vocab.special_eos_id },
+            { LLM_KV_TOKENIZER_UNK_ID, vocab.special_unk_id },
+            { LLM_KV_TOKENIZER_SEP_ID, vocab.special_sep_id },
+            { LLM_KV_TOKENIZER_PAD_ID, vocab.special_pad_id },
+        };
+        for (const auto & it : special_token_types) {
+            const std::string & key = kv(std::get<0>(it));
+            int32_t & id = std::get<1>(it), old_id = id;
+
+            GGUF_GET_KEY(ctx, id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, key);
+            // Must be >= -1 and < vocab size. Since the key is unsigned, -1
+            // can only come from the default value, so there's no point in
+            // validating that.
+            if (size_t(id + 1) > vocab.id_to_token.size()) {
+                LLAMA_LOG_WARN("%s: bad special token: '%s' = %d, using default id %d\n",
+                    __func__, key.c_str(), id, old_id);
+                id = old_id;
+            }
+        }
+    }
 
     // build special tokens cache
     {
@@ -2672,8 +2694,8 @@ static void llm_load_tensors(
                 } break;
             case LLM_ARCH_STARCODER:
                 {
-                    model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU);
-                    model.pos_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, hparams.n_ctx_train}, GGML_BACKEND_CPU);
+                    model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab},             GGML_BACKEND_CPU);
+                    model.pos_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_POS_EMBD, "weight"),   {n_embd, hparams.n_ctx_train}, GGML_BACKEND_CPU);
 
                     // output
                     {
@@ -2724,19 +2746,19 @@ static void llm_load_tensors(
                         layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM,   "bias", i),   {n_embd}, backend);
 
                         layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split);
-                        layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa},         backend_split);
+                        layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa},         backend);
 
                         layer.wo   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd},   backend_split);
-                        layer.bo   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd},           backend_split);
+                        layer.bo   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd},           backend);
 
                         layer.ffn_norm   = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend);
                         layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, backend);
 
                         layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split);
-                        layer.b2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd},       backend_split);
+                        layer.b2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd},       backend);
 
                         layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, backend_split);
-                        layer.b3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff},           backend_split);
+                        layer.b3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff},           backend);
 
                         if (backend == GGML_BACKEND_GPU) {
                             vram_weights +=
@@ -4593,6 +4615,8 @@ static struct ggml_cgraph * llm_build_starcoder(
 
     const float norm_eps = hparams.f_norm_eps;
 
+    const int n_gpu_layers = model.n_gpu_layers;
+
     const int32_t n_tokens = batch.n_tokens;
     const int32_t n_kv     = ggml_allocr_is_measure(lctx.alloc) ? n_ctx            : kv_self.n;
     const int32_t kv_head  = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head;
@@ -4637,6 +4661,27 @@ static struct ggml_cgraph * llm_build_starcoder(
         }
     }
 
+    const int i_gpu_start = n_layer - n_gpu_layers;
+    (void) i_gpu_start;
+
+    // offload functions set the tensor output backend to GPU
+    // tensors are GPU-accelerated if any input or the output has been offloaded
+    offload_func_t offload_func_nr = llama_nop; // nr = non-repeating
+    offload_func_t offload_func_kq = llama_nop;
+    offload_func_t offload_func_v  = llama_nop;
+
+#ifdef GGML_USE_CUBLAS
+    if (n_gpu_layers > n_layer) {
+        offload_func_nr = ggml_cuda_assign_buffers_no_alloc;
+    }
+    if (n_gpu_layers > n_layer + 1) {
+        offload_func_v  = ggml_cuda_assign_buffers_no_alloc;
+    }
+    if (n_gpu_layers > n_layer + 2) {
+        offload_func_kq = ggml_cuda_assign_buffers_no_alloc;
+    }
+#endif // GGML_USE_CUBLAS
+
     {
         // Compute position embeddings.
         struct ggml_tensor * inp_positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
@@ -4662,6 +4707,7 @@ static struct ggml_cgraph * llm_build_starcoder(
     // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
     struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
     ggml_set_name(KQ_mask, "KQ_mask");
+    offload_func_kq(KQ_mask);
     ggml_allocr_alloc(lctx.alloc, KQ_mask);
     if (!ggml_allocr_is_measure(lctx.alloc)) {
         float * data = (float *) KQ_mask->data;
@@ -4685,44 +4731,67 @@ static struct ggml_cgraph * llm_build_starcoder(
     ggml_set_name(inpL, "inpL");
 
     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_no_alloc;
+        }
+#endif // GGML_USE_CUBLAS
+
         {
             // Norm
             cur = ggml_norm(ctx0, inpL, norm_eps);
+            offload_func(cur);
+
             cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].attn_norm), model.layers[il].attn_norm_b);
+            offload_func(cur);
         }
 
         {
             // Self Attention
-            cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wqkv, cur), model.layers[il].bqkv);
+            cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur);
+            offload_func_kq(cur);
 
-            struct ggml_tensor * tmpq = ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*n_embd);
-            struct ggml_tensor * tmpk = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], sizeof(float)*n_embd);
-            struct ggml_tensor * tmpv = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], sizeof(float)*(n_embd + n_embd_gqa));
+            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+            offload_func_kq(cur);
 
-            struct ggml_tensor * Qcur = tmpq;
+            struct ggml_tensor * tmpq = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+            struct ggml_tensor * tmpk = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+            struct ggml_tensor * tmpv = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+
+            ggml_set_name(tmpq, "tmpq");
+            ggml_set_name(tmpk, "tmpk");
+            ggml_set_name(tmpv, "tmpv");
+
+            offload_func_kq(tmpq);
+            offload_func_kq(tmpk);
+            offload_func_v (tmpv);
+
+            struct ggml_tensor * Qcur = ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, n_tokens);
             struct ggml_tensor * Kcur = tmpk;
 
             {
-                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_cont(ctx0, tmpv), n_embd_gqa, n_tokens));
+                struct ggml_tensor * Vcur = ggml_transpose(ctx0, tmpv);
+                offload_func_v(Vcur);
                 ggml_set_name(Vcur, "Vcur");
 
                 struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head));
+                offload_func_kq(k);
                 ggml_set_name(k, "k");
 
                 struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa,
                         (   n_ctx)*ggml_element_size(kv_self.v),
                         (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v));
+                offload_func_v(v);
+                ggml_set_name(v, "v");
 
                 ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
                 ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));
             }
 
-            struct ggml_tensor * Q =
-                ggml_permute(ctx0,
-                        ggml_cpy(ctx0,
-                            Qcur,
-                            ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd_head, n_head, n_tokens)),
-                        0, 2, 1, 3);
+            struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
+            offload_func_kq(Q);
             ggml_set_name(Q, "Q");
 
             struct ggml_tensor * K =
@@ -4731,23 +4800,28 @@ static struct ggml_cgraph * llm_build_starcoder(
                         ggml_element_size(kv_self.k)*n_embd_gqa,
                         ggml_element_size(kv_self.k)*n_embd_head,
                         ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il);
+            offload_func_kq(K);
             ggml_set_name(K, "K");
 
             // K * Q
             struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+            offload_func_kq(KQ);
             ggml_set_name(KQ, "KQ");
 
             // KQ_scaled = KQ / sqrt(n_embd_head)
             // KQ_scaled shape [n_past + n_tokens, n_tokens, n_head, 1]
             struct ggml_tensor * KQ_scaled = ggml_scale_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_add(ctx0, KQ_scaled, KQ_mask);
+            offload_func_kq(KQ_masked);
             ggml_set_name(KQ_masked, "KQ_masked");
 
             // KQ = soft_max(KQ_masked)
             struct ggml_tensor * KQ_soft_max = ggml_soft_max_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
@@ -4760,22 +4834,25 @@ static struct ggml_cgraph * llm_build_starcoder(
             ggml_set_name(V, "V");
 
             struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+            offload_func_v(KQV);
             ggml_set_name(KQV, "KQV");
 
-            // KQV_merged = KQV.permute(0, 2, 1, 3)
             struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+            offload_func_v(KQV_merged);
             ggml_set_name(KQV_merged, "KQV_merged");
 
-            // cur = KQV_merged.contiguous().view(n_embd, n_tokens)
             cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens);
+            offload_func_v(cur);
             ggml_set_name(cur, "KQV_merged_contiguous");
         }
 
         // Projection
         cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wo, cur), model.layers[il].bo);
+        offload_func(cur);
 
         // Add the input
         cur = ggml_add(ctx0, cur, inpL);
+        offload_func(cur);
 
         struct ggml_tensor * inpFF = cur;
 
@@ -4784,27 +4861,36 @@ static struct ggml_cgraph * llm_build_starcoder(
             // Norm
             {
                 cur = ggml_norm(ctx0, inpFF, norm_eps);
+                offload_func_nr(cur);
+
                 cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ffn_norm), model.layers[il].ffn_norm_b);
+                offload_func_nr(cur);
             }
 
             cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].w3, cur), model.layers[il].b3);
+            offload_func(cur);
 
             // GELU activation
             cur = ggml_gelu(ctx0, cur);
+            offload_func(cur);
 
             // Projection
             cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].w2, cur), model.layers[il].b2);
+            offload_func(cur);
         }
 
         inpL = ggml_add(ctx0, cur, inpFF);
+
     }
 
     // Output Norm
     {
         cur = ggml_norm(ctx0, inpL, norm_eps);
+        offload_func_nr(cur);
+
         cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.output_norm), model.output_norm_b);
+        ggml_set_name(cur, "result_norm");
     }
-    ggml_set_name(cur, "result_norm");
 
     cur = ggml_mul_mat(ctx0, model.output, cur);
     ggml_set_name(cur, "result_output");
@@ -5815,33 +5901,6 @@ static struct ggml_cgraph * llama_build_graph(
             GGML_ASSERT(false);
     }
 
-#if 1
-    for (int i = 0; i < result->n_nodes; ++i) {
-        struct ggml_tensor * node = result->nodes[i];
-        if (getenv("SKIP_KQ_ALL")) {
-            if (
-                    strcmp(node->name, "KQ")  == 0 ||
-                    strcmp(node->name, "KQ_scaled") == 0 ||
-                    strcmp(node->name, "KQ_masked") == 0 ||
-                    strcmp(node->name, "KQ_soft_max") == 0 ||
-                    strcmp(node->name, "KQV") == 0 ||
-                    false) {
-                //printf("skipping %s\n", dst->name);
-                node->op  = GGML_OP_NONE;
-            }
-        }
-        if (getenv("SKIP_KQ_KQV")) {
-            if (
-                    strcmp(node->name, "KQ")  == 0 ||
-                    strcmp(node->name, "KQV") == 0 ||
-                    false) {
-                //printf("skipping %s\n", dst->name);
-                node->op  = GGML_OP_NONE;
-            }
-        }
-    }
-#endif
-
     return result;
 }
 
@@ -5965,8 +6024,6 @@ static int llama_decode_internal(
         }
     }
 
-    ggml_cuda_set_mul_mat_q(cparams.mul_mat_q);
-
     // HACK: ggml-alloc may change the tensor backend when reusing a parent, so force output to be on the CPU here if needed
     if (!lctx.embedding.empty()) {
         embeddings->backend = GGML_BACKEND_CPU;
@@ -6017,11 +6074,20 @@ static int llama_decode_internal(
 #endif
 
     // update the kv ring buffer
-    lctx.kv_self.has_shift  = false;
-    lctx.kv_self.head      += n_tokens;
-    // Ensure kv cache head points to a valid index.
-    if (lctx.kv_self.head >= lctx.kv_self.size) {
-        lctx.kv_self.head = 0;
+    {
+        if (kv_self.has_shift) {
+            kv_self.has_shift = false;
+            for (uint32_t i = 0; i < kv_self.size; ++i) {
+                kv_self.cells[i].delta = 0;
+            }
+        }
+
+        kv_self.head += n_tokens;
+
+        // Ensure kv cache head points to a valid index.
+        if (kv_self.head >= kv_self.size) {
+            kv_self.head = 0;
+        }
     }
 
 #ifdef GGML_PERF
@@ -6130,11 +6196,10 @@ static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) {
 }
 
 static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch) {
+    static const char * hex = "0123456789ABCDEF";
     switch (llama_vocab_get_type(vocab)) {
     case LLAMA_VOCAB_TYPE_SPM: {
-        char buf[7];
-        int result = snprintf(buf, sizeof(buf), "<0x%02X>", ch);
-        GGML_ASSERT(0 <= result && result < 7);
+        const char buf[7] = { '<', '0', 'x', hex[ch >> 4], hex[ch & 15], '>', 0 };
         return vocab.token_to_id.at(buf);
     }
     case LLAMA_VOCAB_TYPE_BPE: {
@@ -7500,14 +7565,14 @@ void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * c
         }
     }
 
-    const llama_token eos = llama_token_eos(ctx);
+    const llama_token eos = llama_token_eos(&ctx->model);
 
     std::vector<std::pair<std::vector<uint32_t>, llama_partial_utf8>> candidates_decoded;
     std::vector<llama_grammar_candidate>                              candidates_grammar;
 
     for (size_t i = 0; i < candidates->size; ++i) {
         const llama_token id    = candidates->data[i].id;
-        const std::string piece = llama_token_to_str(ctx, id);
+        const std::string piece = llama_token_to_piece(ctx, id);
         if (id == eos) {
             if (!allow_eos) {
                 candidates->data[i].logit = -INFINITY;
@@ -7710,7 +7775,7 @@ llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_arra
 void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token) {
     const int64_t t_start_sample_us = ggml_time_us();
 
-    if (token == llama_token_eos(ctx)) {
+    if (token == llama_token_eos(&ctx->model)) {
         for (const auto & stack : grammar->stacks) {
             if (stack.empty()) {
                 return;
@@ -7719,7 +7784,7 @@ void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar
         GGML_ASSERT(false);
     }
 
-    const std::string piece = llama_token_to_str(ctx, token);
+    const std::string piece = llama_token_to_piece(ctx, token);
 
     // Note terminating 0 in decoded string
     const auto   decoded     = decode_utf8(piece.c_str(), grammar->partial_utf8);
@@ -7992,6 +8057,24 @@ struct no_init {
     no_init() { /* do nothing */ }
 };
 
+struct quantize_state_internal {
+    const llama_model                 & model;
+    const llama_model_quantize_params * params;
+
+    int n_attention_wv    = 0;
+    int n_feed_forward_w2 = 0;
+    int i_attention_wv    = 0;
+    int i_feed_forward_w2 = 0;
+
+    int n_k_quantized     = 0;
+    int n_fallback        = 0;
+
+    quantize_state_internal(const llama_model & model, const llama_model_quantize_params * params)
+        : model(model)
+        , params(params)
+        {}
+};
+
 static void llama_convert_tensor_internal(
     struct ggml_tensor * tensor, std::vector<no_init<float>> & output, std::vector<std::thread> & workers,
     const size_t nelements, const int nthread
@@ -8050,14 +8133,14 @@ static void llama_convert_tensor_internal(
     workers.clear();
 }
 
-#ifdef GGML_USE_K_QUANTS
 static ggml_type get_k_quant_type(
-    ggml_type new_type, const ggml_tensor * tensor, const llama_model & model, llama_ftype ftype, int * i_attention_wv,
-    int n_attention_wv, int * i_feed_forward_w2, int n_feed_forward_w2
+    quantize_state_internal & qs,
+    ggml_type new_type, const ggml_tensor * tensor, llama_ftype ftype
 ) {
     const std::string name = ggml_get_name(tensor);
     // TODO: avoid hardcoded tensor names - use the TN_* constants
-    const auto tn = LLM_TN(model.arch);
+    const llm_arch arch = qs.model.arch;
+    const auto       tn = LLM_TN(arch);
 
     auto use_more_bits = [](int i_layer, int num_layers) -> bool {
         return i_layer < num_layers/8 || i_layer >= 7*num_layers/8 || (i_layer - num_layers/8)%3 == 2;
@@ -8065,7 +8148,7 @@ static ggml_type get_k_quant_type(
 
     if (name == tn(LLM_TENSOR_OUTPUT, "weight")) {
         int nx = tensor->ne[0];
-        if (model.arch == LLM_ARCH_FALCON || nx % QK_K != 0) {
+        if (arch == LLM_ARCH_FALCON || nx % QK_K != 0) {
             new_type = GGML_TYPE_Q8_0;
         }
         else if (new_type != GGML_TYPE_Q8_0) {
@@ -8074,46 +8157,46 @@ static ggml_type get_k_quant_type(
     } else if (name.find("attn_v.weight") != std::string::npos) {
         if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
-            new_type = *i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
+            new_type = qs.i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
         }
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
         else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
-                use_more_bits(*i_attention_wv, n_attention_wv)) new_type = GGML_TYPE_Q6_K;
-        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && *i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
+                use_more_bits(qs.i_attention_wv, qs.n_attention_wv)) new_type = GGML_TYPE_Q6_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
         else if (QK_K == 64 && (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S) &&
-                (*i_attention_wv < n_attention_wv/8 || *i_attention_wv >= 7*n_attention_wv/8)) new_type = GGML_TYPE_Q6_K;
-        if (model.type == MODEL_70B) {
+                (qs.i_attention_wv < qs.n_attention_wv/8 || qs.i_attention_wv >= 7*qs.n_attention_wv/8)) new_type = GGML_TYPE_Q6_K;
+        if (qs.model.type == MODEL_70B) {
             // In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is
             // 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with
             // nearly negligible increase in model size by quantizing this tensor with more bits:
             if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K) new_type = GGML_TYPE_Q5_K;
         }
-        ++*i_attention_wv;
+        ++qs.i_attention_wv;
     } else if (name.find("ffn_down.weight") != std::string::npos) {
         if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
-            new_type = *i_feed_forward_w2 < 2 ? GGML_TYPE_Q5_K
-                     : model.arch != LLM_ARCH_FALCON || use_more_bits(*i_feed_forward_w2, n_feed_forward_w2) ? GGML_TYPE_Q4_K
+            new_type = qs.i_feed_forward_w2 < 2 ? GGML_TYPE_Q5_K
+                     : arch != LLM_ARCH_FALCON || use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2) ? GGML_TYPE_Q4_K
                      : GGML_TYPE_Q3_K;
         }
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) {
-            new_type = model.arch == LLM_ARCH_FALCON ? GGML_TYPE_Q4_K : GGML_TYPE_Q5_K;
+            new_type = arch == LLM_ARCH_FALCON ? GGML_TYPE_Q4_K : GGML_TYPE_Q5_K;
         }
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) {
-            if (model.arch == LLM_ARCH_FALCON) {
-                new_type = *i_feed_forward_w2 < 2 ? GGML_TYPE_Q6_K :
-                           use_more_bits(*i_feed_forward_w2, n_feed_forward_w2) ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
+            if (arch == LLM_ARCH_FALCON) {
+                new_type = qs.i_feed_forward_w2 < 2 ? GGML_TYPE_Q6_K :
+                           use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2) ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
             } else {
-                if (use_more_bits(*i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
+                if (use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
             }
         }
-        else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(*i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
-        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && model.arch != LLM_ARCH_FALCON && *i_feed_forward_w2 < 4) {
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && qs.i_feed_forward_w2 < 4) {
             new_type = GGML_TYPE_Q5_K;
         }
-        ++*i_feed_forward_w2;
+        ++qs.i_feed_forward_w2;
     } else if (name.find("attn_output.weight") != std::string::npos) {
-        if (model.arch != LLM_ARCH_FALCON) {
+        if (arch != LLM_ARCH_FALCON) {
             if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K  ) new_type = GGML_TYPE_Q3_K;
             else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) new_type = GGML_TYPE_Q4_K;
             else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
@@ -8140,25 +8223,27 @@ static ggml_type get_k_quant_type(
         int nx = tensor->ne[0];
         int ny = tensor->ne[1];
         if (nx % QK_K != 0) {
-            LLAMA_LOG_WARN("\n\n%s : tensor cols %d x %d are not divisible by %d, required for k-quants\n", __func__, nx, ny, QK_K);
+            LLAMA_LOG_WARN("\n\n%s : tensor cols %d x %d are not divisible by %d, required for %s", __func__, nx, ny, QK_K, ggml_type_name(new_type));
             convert_incompatible_tensor = true;
+        } else {
+            ++qs.n_k_quantized;
         }
     }
     if (convert_incompatible_tensor) {
-        if (name == tn(LLM_TENSOR_OUTPUT, "weight")) {
-            new_type = GGML_TYPE_F16; //fall back to F16 instead of just failing.
-            LLAMA_LOG_WARN("F16 will be used for this tensor instead.\n");
-        } else if (name == tn(LLM_TENSOR_TOKEN_EMBD, "weight")) {
-            new_type = GGML_TYPE_Q4_0; //fall back to Q4_0 instead of just failing.
-            LLAMA_LOG_WARN("Q4_0 will be used for this tensor instead.\n");
-        } else {
-            throw std::runtime_error("Unsupported tensor size encountered\n");
+        switch (new_type) {
+            case GGML_TYPE_Q2_K: new_type = GGML_TYPE_Q4_0; break;
+            case GGML_TYPE_Q3_K: new_type = GGML_TYPE_Q4_1; break;
+            case GGML_TYPE_Q4_K: new_type = GGML_TYPE_Q5_0; break;
+            case GGML_TYPE_Q5_K: new_type = GGML_TYPE_Q5_1; break;
+            case GGML_TYPE_Q6_K: new_type = GGML_TYPE_Q8_0; break;
+            default: throw std::runtime_error("\nUnsupported tensor size encountered\n");
         }
+        LLAMA_LOG_WARN(" - using fallback quantization %s\n", ggml_type_name(new_type));
+        ++qs.n_fallback;
     }
 
     return new_type;
 }
-#endif
 
 static void llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, const llama_model_quantize_params * params) {
     ggml_type quantized_type;
@@ -8173,7 +8258,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_F16:  quantized_type = GGML_TYPE_F16;  break;
         case LLAMA_FTYPE_ALL_F32:     quantized_type = GGML_TYPE_F32;  break;
 
-#ifdef GGML_USE_K_QUANTS
         // K-quants
         case LLAMA_FTYPE_MOSTLY_Q2_K:   quantized_type = GGML_TYPE_Q2_K; break;
         case LLAMA_FTYPE_MOSTLY_Q3_K_S:
@@ -8184,7 +8268,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_Q5_K_S:
         case LLAMA_FTYPE_MOSTLY_Q5_K_M: quantized_type = GGML_TYPE_Q5_K; break;
         case LLAMA_FTYPE_MOSTLY_Q6_K:   quantized_type = GGML_TYPE_Q6_K; break;
-#endif
+
         default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
     }
 
@@ -8211,6 +8295,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     llm_load_arch(ml, model);
     llm_load_hparams(ml, model);
 
+    struct quantize_state_internal qs(model, params);
+
     if (params->only_copy) {
         ftype = model.ftype;
     }
@@ -8223,10 +8309,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     gguf_set_val_u32(ctx_out, "general.quantization_version", GGML_QNT_VERSION);
     gguf_set_val_u32(ctx_out, "general.file_type", ftype);
 
-#ifdef GGML_USE_K_QUANTS
-    int n_attention_wv    = 0;
-    int n_feed_forward_w2 = 0;
-
     for (int i = 0; i < ml.n_tensors; ++i) {
         struct ggml_tensor * meta = ml.get_tensor_meta(i);
 
@@ -8234,21 +8316,17 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
 
         // TODO: avoid hardcoded tensor names - use the TN_* constants
         if (name.find("attn_v.weight") != std::string::npos || name.find("attn_qkv.weight") != std::string::npos) {
-            ++n_attention_wv;
+            ++qs.n_attention_wv;
         }
         else if (name.find("ffn_down.weight") != std::string::npos) {
-            ++n_feed_forward_w2;
+            ++qs.n_feed_forward_w2;
         }
     }
-    if (n_attention_wv != n_feed_forward_w2 || (uint32_t)n_attention_wv != model.hparams.n_layer) {
+    if (qs.n_attention_wv != qs.n_feed_forward_w2 || (uint32_t)qs.n_attention_wv != model.hparams.n_layer) {
         LLAMA_LOG_WARN("%s ============ Strange model: n_attention_wv = %d, n_feed_forward_w2 = %d, hparams.n_layer = %d\n",
-                __func__, n_attention_wv, n_feed_forward_w2, model.hparams.n_layer);
+                __func__, qs.n_attention_wv, qs.n_feed_forward_w2, model.hparams.n_layer);
     }
 
-    int i_attention_wv = 0;
-    int i_feed_forward_w2 = 0;
-#endif
-
     size_t total_size_org = 0;
     size_t total_size_new = 0;
     std::vector<int64_t> hist_all(1 << 4, 0);
@@ -8312,11 +8390,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
 
         if (quantize) {
             new_type = quantized_type;
-#ifdef GGML_USE_K_QUANTS
-            new_type = get_k_quant_type(
-                new_type, tensor, model, ftype, &i_attention_wv, n_attention_wv, &i_feed_forward_w2, n_feed_forward_w2
-            );
-#endif
+            if (!params->pure) {
+                new_type = get_k_quant_type(qs, new_type, tensor, ftype);
+            }
+
             // If we've decided to quantize to the same type the tensor is already
             // in then there's nothing to do.
             quantize = tensor->type != new_type;
@@ -8441,6 +8518,11 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
             LLAMA_LOG_INFO("\n");
         }
     }
+
+    if (qs.n_fallback > 0) {
+        LLAMA_LOG_WARN("%s: WARNING: %d of %d tensor(s) incompatible with k-quants and required fallback quantization\n",
+                __func__, qs.n_fallback, qs.n_k_quantized + qs.n_fallback);
+    }
 }
 
 static int llama_apply_lora_from_file_internal(
@@ -8765,6 +8847,7 @@ struct llama_model_quantize_params llama_model_quantize_default_params() {
         /*.allow_requantize            =*/ false,
         /*.quantize_output_tensor      =*/ true,
         /*.only_copy                   =*/ false,
+        /*.pure                        =*/ false,
     };
 
     return result;
@@ -8919,7 +9002,7 @@ struct llama_context * llama_new_context_with_model(
             // build worst-case graph
             int n_tokens = (int)std::min(cparams.n_ctx, cparams.n_batch);
             int n_past = cparams.n_ctx - n_tokens;
-            llama_token token = llama_token_bos(ctx); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph
+            llama_token token = llama_token_bos(&ctx->model); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph
             ggml_cgraph * gf = llama_build_graph(*ctx, llama_batch_get_one(&token, n_tokens, n_past, 0));
 
 #ifdef GGML_USE_METAL
@@ -9125,8 +9208,8 @@ int llama_get_kv_cache_token_count(const struct llama_context * ctx) {
     return ctx->kv_self.head;
 }
 
-void llama_kv_cache_tokens_rm(struct llama_context * ctx, int32_t c0, int32_t c1) {
-    llama_kv_cache_tokens_rm(ctx->kv_self, c0, c1);
+void llama_kv_cache_clear(struct llama_context * ctx) {
+    llama_kv_cache_clear(ctx->kv_self);
 }
 
 void llama_kv_cache_seq_rm(struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
@@ -9572,7 +9655,7 @@ int llama_eval(
                  llama_token * tokens,
                      int32_t   n_tokens,
                          int   n_past) {
-    llama_kv_cache_tokens_rm(ctx->kv_self, n_past, -1);
+    llama_kv_cache_seq_rm(ctx->kv_self, -1, n_past, -1);
 
     const int ret = llama_decode_internal(*ctx, llama_batch_get_one(tokens, n_tokens, n_past, 0));
     if (ret < 0) {
@@ -9587,7 +9670,7 @@ int llama_eval_embd(
                            float * embd,
                          int32_t   n_tokens,
                              int   n_past) {
-    llama_kv_cache_tokens_rm(ctx->kv_self, n_past, -1);
+    llama_kv_cache_seq_rm(ctx->kv_self, -1, n_past, -1);
 
     llama_batch batch = { n_tokens, nullptr, embd, nullptr, nullptr, nullptr, nullptr, n_past, 1, 0, };
 
@@ -9680,43 +9763,44 @@ float * llama_get_embeddings(struct llama_context * ctx) {
     return ctx->embedding.data();
 }
 
-const char * llama_token_get_text(const struct llama_context * ctx, llama_token token) {
-    return ctx->model.vocab.id_to_token[token].text.c_str();
+const char * llama_token_get_text(const struct llama_model * model, llama_token token) {
+    return model->vocab.id_to_token[token].text.c_str();
 }
 
-float llama_token_get_score(const struct llama_context * ctx, llama_token token) {
-    return ctx->model.vocab.id_to_token[token].score;
+float llama_token_get_score(const struct llama_model * model, llama_token token) {
+    return model->vocab.id_to_token[token].score;
 }
 
-llama_token_type llama_token_get_type(const struct llama_context * ctx, llama_token token) {
-    return ctx->model.vocab.id_to_token[token].type;
+llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token) {
+    return model->vocab.id_to_token[token].type;
 }
 
-llama_token llama_token_bos(const struct llama_context * ctx) {
-    return ctx->model.vocab.special_bos_id;
+llama_token llama_token_bos(const struct llama_model * model) {
+    return model->vocab.special_bos_id;
 }
 
-llama_token llama_token_eos(const struct llama_context * ctx) {
-    return ctx->model.vocab.special_eos_id;
+llama_token llama_token_eos(const struct llama_model * model) {
+    return model->vocab.special_eos_id;
 }
 
-llama_token llama_token_nl(const struct llama_context * ctx) {
-    return ctx->model.vocab.linefeed_id;
-}
-llama_token llama_token_prefix(const struct llama_context * ctx) {
-    return ctx->model.vocab.special_prefix_id;
+llama_token llama_token_nl(const struct llama_model * model) {
+    return model->vocab.linefeed_id;
 }
 
-llama_token llama_token_middle(const struct llama_context * ctx) {
-    return ctx->model.vocab.special_middle_id;
+llama_token llama_token_prefix(const struct llama_model * model) {
+    return model->vocab.special_prefix_id;
 }
 
-llama_token llama_token_suffix(const struct llama_context * ctx) {
-    return ctx->model.vocab.special_suffix_id;
+llama_token llama_token_middle(const struct llama_model * model) {
+    return model->vocab.special_middle_id;
 }
 
-llama_token llama_token_eot(const struct llama_context * ctx) {
-    return ctx->model.vocab.special_eot_id;
+llama_token llama_token_suffix(const struct llama_model * model) {
+    return model->vocab.special_suffix_id;
+}
+
+llama_token llama_token_eot(const struct llama_model * model) {
+    return model->vocab.special_eot_id;
 }
 
 int llama_tokenize(

llama.h

@@ -178,7 +178,7 @@ extern "C" {
         float rope_freq_scale; // RoPE frequency scaling factor, 0 = from model
 
         // Keep the booleans together to avoid misalignment during copy-by-value.
-        bool mul_mat_q;  // if true, use experimental mul_mat_q kernels
+        bool mul_mat_q;  // if true, use experimental mul_mat_q kernels (DEPRECATED - always true)
         bool f16_kv;     // use fp16 for KV cache, fp32 otherwise
         bool logits_all; // the llama_eval() call computes all logits, not just the last one
         bool embedding;  // embedding mode only
@@ -191,6 +191,7 @@ extern "C" {
         bool allow_requantize;       // allow quantizing non-f32/f16 tensors
         bool quantize_output_tensor; // quantize output.weight
         bool only_copy;              // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
+        bool pure;                   // disable k-quant mixtures and quantize all tensors to the same type
     } llama_model_quantize_params;
 
     // grammar types
@@ -333,17 +334,14 @@ extern "C" {
     LLAMA_API DEPRECATED(int llama_get_kv_cache_token_count(const struct llama_context * ctx),
             "avoid using this, it will be removed in the future, instead - count the tokens in user code");
 
-    // Remove all tokens data of cells in [c0, c1)
-    // c0 < 0 : [0,  c1]
-    // c1 < 0 : [c0, inf)
-    LLAMA_API void llama_kv_cache_tokens_rm(
-            struct llama_context * ctx,
-                         int32_t   c0,
-                         int32_t   c1);
+    // Clear the KV cache
+    LLAMA_API void llama_kv_cache_clear(
+            struct llama_context * ctx);
 
     // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
-    // p0 < 0 : [0,  p1]
-    // p1 < 0 : [p0, inf)
+    // seq_id < 0 : match any sequence
+    // p0 < 0     : [0,  p1]
+    // p1 < 0     : [p0, inf)
     LLAMA_API void llama_kv_cache_seq_rm(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
@@ -494,21 +492,22 @@ extern "C" {
     // Vocab
     //
 
-    LLAMA_API const char * llama_token_get_text(const struct llama_context * ctx, llama_token token);
+    LLAMA_API const char * llama_token_get_text(const struct llama_model * model, llama_token token);
 
-    LLAMA_API float llama_token_get_score(const struct llama_context * ctx, llama_token token);
+    LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token);
 
-    LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_context * ctx, llama_token token);
+    LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token);
 
     // Special tokens
-    LLAMA_API llama_token llama_token_bos(const struct llama_context * ctx);  // beginning-of-sentence
-    LLAMA_API llama_token llama_token_eos(const struct llama_context * ctx);  // end-of-sentence
-    LLAMA_API llama_token llama_token_nl (const struct llama_context * ctx);  // next-line
+    LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence
+    LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence
+    LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line
+
     // codellama infill tokens
-    LLAMA_API llama_token llama_token_prefix(const struct llama_context * ctx); // Beginning of infill prefix
-    LLAMA_API llama_token llama_token_middle(const struct llama_context * ctx); // Beginning of infill middle
-    LLAMA_API llama_token llama_token_suffix(const struct llama_context * ctx); // Beginning of infill suffix
-    LLAMA_API llama_token llama_token_eot   (const struct llama_context * ctx); // End of infill middle
+    LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix
+    LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle
+    LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix
+    LLAMA_API llama_token llama_token_eot   (const struct llama_model * model); // End of infill middle
 
     //
     // Tokenization
@@ -657,6 +656,7 @@ extern "C" {
                            float * mu);
 
     /// @details Selects the token with the highest probability.
+    ///          Does not compute the token probabilities. Use llama_sample_softmax() instead.
     LLAMA_API llama_token llama_sample_token_greedy(
             struct llama_context * ctx,
           llama_token_data_array * candidates);

tests/CMakeLists.txt

@@ -28,9 +28,14 @@ llama_build_executable(test-tokenizer-0-falcon.cpp)
 llama_test_executable (test-tokenizer-0-falcon test-tokenizer-0-falcon.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
 llama_build_executable(test-tokenizer-1-llama.cpp)
 llama_test_executable (test-tokenizer-1-llama test-tokenizer-1-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
+llama_test_executable(test-tokenizer-1-baichuan test-tokenizer-1-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-baichuan.gguf)
 llama_build_executable(test-tokenizer-1-bpe.cpp)
 llama_test_executable (test-tokenizer-1-falcon test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
 llama_test_executable(test-tokenizer-1-aquila test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf)
+llama_test_executable(test-tokenizer-1-mpt test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-mpt.gguf)
+llama_test_executable(test-tokenizer-1-gpt-neox test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-neox.gguf)
+llama_test_executable(test-tokenizer-1-refact test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf)
+llama_test_executable(test-tokenizer-1-starcoder test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf)
 llama_build_and_test_executable(test-grammar-parser.cpp)
 llama_build_and_test_executable(test-llama-grammar.cpp)
 llama_build_and_test_executable(test-grad0.cpp) # SLOW

tests/test-tokenizer-1-bpe.cpp

@@ -91,9 +91,19 @@ int main(int argc, char **argv) {
             }
         }
     }
-    // TODO: why doesn't this work for the full range of Unicodes?
+    // Restrict to assigned unicode planes
     // for (uint32_t cp = 0x10000; cp < 0x0010ffff; ++cp) {
-    for (uint32_t cp = 0x10000; cp < 0x00080000; ++cp) {
+    for (uint32_t cp = 0x10000; cp < 0x00040000; ++cp) {
+        std::string str = codepoint_to_utf8(cp);
+        std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
+        std::string check = llama_detokenize_bpe(ctx, tokens);
+        if (str != check) {
+            fprintf(stderr, "%s : error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
+                __func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
+            return 4;
+        }
+    }
+    for (uint32_t cp = 0x000e0000; cp < 0x0010ffff; ++cp) {
         std::string str = codepoint_to_utf8(cp);
         std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
         std::string check = llama_detokenize_bpe(ctx, tokens);
@@ -103,7 +113,6 @@ int main(int argc, char **argv) {
             return 4;
         }
     }
-
     llama_free_model(model);
     llama_free(ctx);