imatrix: guard Q4_0/Q5_0 against ffn_down craziness

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>

.github/workflows/nix-ci-aarch64.yml

@@ -0,0 +1,55 @@
+name: Nix aarch64 builds
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix']
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix']
+
+jobs:
+  nix-build-aarch64:
+    if: ${{ vars.CACHIX_NAME != '' }}
+    runs-on: ubuntu-latest
+    steps:
+    - name: Checkout repository
+      uses: actions/checkout@v4
+    - name: Install QEMU
+      # Copy-paste from https://github.com/orgs/community/discussions/8305#discussioncomment-5888654
+      run: |
+        sudo apt-get update
+        sudo apt-get install -y qemu-user-static qemu-system-aarch64
+        sudo usermod -a -G kvm $USER
+    - name: Install Nix
+      uses: DeterminateSystems/nix-installer-action@v9
+      with:
+        github-token: ${{ secrets.GITHUB_TOKEN }}
+        extra-conf: |
+          extra-platforms = aarch64-linux
+          extra-system-features = nixos-test kvm
+          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+    - uses: DeterminateSystems/magic-nix-cache-action@v2
+      with:
+        upstream-cache: https://${{ matrix.cachixName }}.cachix.org
+    - name: Set-up cachix to push the results to
+      uses: cachix/cachix-action@v13
+      with:
+        authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
+        name: ${{ vars.CACHIX_NAME }}
+    - name: Show all output paths
+      run: >
+          nix run github:nix-community/nix-eval-jobs
+          -- --gc-roots-dir gcroot
+          --flake
+          ".#packages.aarch64-linux"
+    - name: Build
+      run: >
+          nix run github:Mic92/nix-fast-build
+          -- --skip-cached --no-nom
+          --systems aarch64-linux
+          --flake
+          ".#checks.aarch64-linux"

.github/workflows/nix-ci.yml

@@ -69,44 +69,3 @@ jobs:
           -- --skip-cached --no-nom
           --flake
           ".#checks.$(nix eval --raw --impure --expr builtins.currentSystem)"
-  nix-build-aarch64:
-    if: ${{ vars.CACHIX_NAME != '' }}
-    runs-on: ubuntu-latest
-    steps:
-    - name: Checkout repository
-      uses: actions/checkout@v4
-    - name: Install QEMU
-      # Copy-paste from https://github.com/orgs/community/discussions/8305#discussioncomment-5888654
-      run: |
-        sudo apt-get install -y qemu-user-static qemu-system-aarch64
-        sudo usermod -a -G kvm $USER
-    - name: Install Nix
-      uses: DeterminateSystems/nix-installer-action@v9
-      with:
-        github-token: ${{ secrets.GITHUB_TOKEN }}
-        extra-conf: |
-          extra-platforms = aarch64-linux
-          extra-system-features = nixos-test kvm
-          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
-    - uses: DeterminateSystems/magic-nix-cache-action@v2
-      with:
-        upstream-cache: https://${{ matrix.cachixName }}.cachix.org
-    - name: Set-up cachix to push the results to
-      uses: cachix/cachix-action@v13
-      with:
-        authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
-        name: ${{ vars.CACHIX_NAME }}
-    - name: Show all output paths
-      run: >
-          nix run github:nix-community/nix-eval-jobs
-          -- --gc-roots-dir gcroot
-          --flake
-          ".#packages.aarch64-linux"
-    - name: Build
-      run: >
-          nix run github:Mic92/nix-fast-build
-          -- --skip-cached --no-nom
-          --systems aarch64-linux
-          --flake
-          ".#checks.aarch64-linux"

.github/workflows/nix-flake-update.yml

@@ -19,4 +19,4 @@ jobs:
           pr-labels: |
             nix
           pr-reviewers: philiptaron,SomeoneSerge
-          token: ${{ secrets.GITHUB_TOKEN }}
+          token: ${{ secrets.FLAKE_TOKEN }}

.gitignore

@@ -43,6 +43,7 @@ models-mnt
 /embedding
 /gguf
 /gguf-llama-simple
+/imatrix
 /infill
 /libllama.so
 /llama-bench
@@ -51,6 +52,7 @@ models-mnt
 /lookup
 /main
 /metal
+/passkey
 /perplexity
 /q8dot
 /quantize

CMakeLists.txt

@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.13)  # for add_link_options
+cmake_minimum_required(VERSION 3.14)  # for add_link_options and implicit target directories.
 project("llama.cpp" C CXX)
 
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
@@ -76,6 +76,10 @@ if (NOT MSVC)
     option(LLAMA_F16C                        "llama: enable F16C"                               ${INS_ENB})
 endif()
 
+if (WIN32)
+    option(LLAMA_WIN_VER                     "llama: Windows Version"                           0x602)
+endif()
+
 # 3rd party libs
 option(LLAMA_ACCELERATE                      "llama: enable Accelerate framework"               ON)
 option(LLAMA_BLAS                            "llama: use BLAS"                                  OFF)
@@ -590,6 +594,13 @@ if (NOT MSVC)
     endif()
 endif()
 
+function(add_compile_option_cpp ARG)
+    # Adds a compile option to C/C++ only, but not for Cuda.
+    # Use, e.g., for CPU-architecture flags.
+    add_compile_options($<$<COMPILE_LANGUAGE:CXX>:${ARG}>)
+    add_compile_options($<$<COMPILE_LANGUAGE:C>:${ARG}>)
+endfunction()
+
 if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64") OR ("${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "arm64"))
     message(STATUS "ARM detected")
     if (MSVC)
@@ -624,8 +635,7 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GE
             include(cmake/FindSIMD.cmake)
         endif ()
         if (LLAMA_AVX512)
-            add_compile_options($<$<COMPILE_LANGUAGE:C>:/arch:AVX512>)
-            add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX512>)
+            add_compile_option_cpp(/arch:AVX512)
             # MSVC has no compile-time flags enabling specific
             # AVX512 extensions, neither it defines the
             # macros corresponding to the extensions.
@@ -639,37 +649,35 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GE
                 add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
             endif()
         elseif (LLAMA_AVX2)
-            add_compile_options($<$<COMPILE_LANGUAGE:C>:/arch:AVX2>)
-            add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX2>)
+            add_compile_option_cpp(/arch:AVX2)
         elseif (LLAMA_AVX)
-            add_compile_options($<$<COMPILE_LANGUAGE:C>:/arch:AVX>)
-            add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX>)
+            add_compile_option_cpp(/arch:AVX)
         endif()
     else()
         if (LLAMA_NATIVE)
-            add_compile_options(-march=native)
+            add_compile_option_cpp(-march=native)
         endif()
         if (LLAMA_F16C)
-            add_compile_options(-mf16c)
+            add_compile_option_cpp(-mf16c)
         endif()
         if (LLAMA_FMA)
-            add_compile_options(-mfma)
+            add_compile_option_cpp(-mfma)
         endif()
         if (LLAMA_AVX)
-            add_compile_options(-mavx)
+            add_compile_option_cpp(-mavx)
         endif()
         if (LLAMA_AVX2)
-            add_compile_options(-mavx2)
+            add_compile_option_cpp(-mavx2)
         endif()
         if (LLAMA_AVX512)
-            add_compile_options(-mavx512f)
-            add_compile_options(-mavx512bw)
+            add_compile_option_cpp(-mavx512f)
+            add_compile_option_cpp(-mavx512bw)
         endif()
         if (LLAMA_AVX512_VBMI)
-            add_compile_options(-mavx512vbmi)
+            add_compile_option_cpp(-mavx512vbmi)
         endif()
         if (LLAMA_AVX512_VNNI)
-            add_compile_options(-mavx512vnni)
+            add_compile_option_cpp(-mavx512vnni)
         endif()
     endif()
 elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
@@ -686,7 +694,7 @@ endif()
 
 if (MINGW)
     # Target Windows 8 for PrefetchVirtualMemory
-    add_compile_definitions(_WIN32_WINNT=0x602)
+    add_compile_definitions(_WIN32_WINNT=${LLAMA_WIN_VER})
 endif()
 
 #

Makefile

@@ -1,8 +1,8 @@
 # Define the default target now so that it is always the first target
 BUILD_TARGETS = \
-	main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
+	main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
 	simple batched batched-bench save-load-state server gguf llama-bench libllava.a llava-cli baby-llama beam-search  \
-	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup tests/test-c.o
+	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = \
@@ -43,10 +43,6 @@ ifeq ($(UNAME_S),Darwin)
 	endif
 endif
 
-ifneq '' '$(or $(filter clean,$(MAKECMDGOALS)),$(LLAMA_METAL))'
-BUILD_TARGETS += metal
-endif
-
 default: $(BUILD_TARGETS)
 
 test: $(TEST_TARGETS)
@@ -614,6 +610,9 @@ quantize-stats: examples/quantize-stats/quantize-stats.cpp    build-info.o ggml.
 perplexity: examples/perplexity/perplexity.cpp                ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
+imatrix: examples/imatrix/imatrix.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 embedding: examples/embedding/embedding.cpp                   ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
@@ -665,10 +664,8 @@ lookahead: examples/lookahead/lookahead.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS
 lookup: examples/lookup/lookup.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-ifdef LLAMA_METAL
-metal: examples/metal/metal.cpp ggml.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
-endif
+passkey: examples/passkey/passkey.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 ifeq ($(UNAME_S),Darwin)
 swift: examples/batched.swift

Package.swift

@@ -14,14 +14,14 @@ let package = Package(
         .library(name: "llama", targets: ["llama"]),
     ],
     dependencies: [
-        .package(url: "https://github.com/ggerganov/ggml.git", .branch("master"))
+        .package(url: "https://github.com/ggerganov/ggml.git", .branch("release"))
     ],
     targets: [
         .target(
             name: "llama",
             dependencies: ["ggml"],
             path: ".",
-            exclude: [],
+            exclude: ["ggml-metal.metal"],
             sources: [
                 "llama.cpp",
             ],

README.md

@@ -10,6 +10,7 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 ### Hot topics
 
+- New SOTA quantized models, including pure 2-bits: https://huggingface.co/ikawrakow
 - Collecting Apple Silicon performance stats:
   - M-series: https://github.com/ggerganov/llama.cpp/discussions/4167
   - A-series: https://github.com/ggerganov/llama.cpp/discussions/4508
@@ -118,6 +119,7 @@ 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)
+- JS/TS (llama.cpp server client): [lgrammel/modelfusion](https://modelfusion.dev/integration/model-provider/llamacpp)
 - 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)
@@ -135,6 +137,7 @@ as the main playground for developing new features for the [ggml](https://github
 - [semperai/amica](https://github.com/semperai/amica)
 - [psugihara/FreeChat](https://github.com/psugihara/FreeChat)
 - [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal)
+- [iohub/collama](https://github.com/iohub/coLLaMA)
 
 ---
 

awq-py/README.md

@@ -43,7 +43,7 @@ Example for llama model
 # For llama7b and llama2 models
 python convert.py models/llama-7b/ --awq-path awq_cache/llama-7b-w4-g128.pt --outfile models/llama_7b_fp16.gguf
 # For mistral and mpt models
-python convert-hf-to-gguf.py models/mpt-7b/ --awq-path awq_cache/llama-7b-w4-g128.pt --outfile models/mpt_7b_fp16.gguf
+python convert-hf-to-gguf.py models/mpt-7b/ --awq-path awq_cache/mpt-7b-w4-g128.pt --outfile models/mpt_7b_fp16.gguf
 ```
 
 ## Quantize

common/common.cpp

@@ -220,6 +220,20 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.n_ctx = std::stoi(argv[i]);
+        } else if (arg == "--grp-attn-n" || arg == "-gan") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+
+            params.grp_attn_n = std::stoi(argv[i]);
+        } else if (arg == "--grp-attn-w" || arg == "-gaw") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+
+            params.grp_attn_w = std::stoi(argv[i]);
         } else if (arg == "--rope-freq-base") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -529,9 +543,8 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
             params.n_gpu_layers = std::stoi(argv[i]);
-#else
+#ifndef LLAMA_SUPPORTS_GPU_OFFLOAD
             fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n");
             fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
 #endif
@@ -540,9 +553,8 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
             params.n_gpu_layers_draft = std::stoi(argv[i]);
-#else
+#ifndef LLAMA_SUPPORTS_GPU_OFFLOAD
             fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers-draft option will be ignored\n");
             fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
 #endif
@@ -551,25 +563,44 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-#ifdef GGML_USE_CUBLAS
             params.main_gpu = std::stoi(argv[i]);
-#else
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.\n");
-#endif
+#ifndef GGML_USE_CUBLAS
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting the main GPU has no effect.\n");
+#endif // GGML_USE_CUBLAS
+        } else if (arg == "--split-mode" || arg == "-sm") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            std::string arg_next = argv[i];
+            if (arg_next == "none") {
+                params.split_mode = LLAMA_SPLIT_NONE;
+            } else if (arg_next == "layer") {
+                params.split_mode = LLAMA_SPLIT_LAYER;
+            } else if (arg_next == "row") {
+                params.split_mode = LLAMA_SPLIT_ROW;
+            } else {
+                invalid_param = true;
+                break;
+            }
+#ifndef GGML_USE_CUBLAS
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting the split mode has no effect.\n");
+#endif // GGML_USE_CUBLAS
         } else if (arg == "--tensor-split" || arg == "-ts") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-#ifdef GGML_USE_CUBLAS
             std::string arg_next = argv[i];
 
             // split string by , and /
             const std::regex regex{R"([,/]+)"};
             std::sregex_token_iterator it{arg_next.begin(), arg_next.end(), regex, -1};
             std::vector<std::string> split_arg{it, {}};
-            GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES);
-
+            if (split_arg.size() >= LLAMA_MAX_DEVICES) {
+                invalid_param = true;
+                break;
+            }
             for (size_t i = 0; i < LLAMA_MAX_DEVICES; ++i) {
                 if (i < split_arg.size()) {
                     params.tensor_split[i] = std::stof(split_arg[i]);
@@ -577,14 +608,8 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                     params.tensor_split[i] = 0.0f;
                 }
             }
-#else
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
-#endif // GGML_USE_CUBLAS
-        } else if (arg == "--no-mul-mat-q" || arg == "-nommq") {
-#ifdef GGML_USE_CUBLAS
-            params.mul_mat_q = false;
-#else
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Disabling mul_mat_q kernels has no effect.\n");
+#ifndef GGML_USE_CUBLAS
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting a tensor split has no effect.\n");
 #endif // GGML_USE_CUBLAS
         } else if (arg == "--no-mmap") {
             params.use_mmap = false;
@@ -592,6 +617,8 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
             params.numa = true;
         } else if (arg == "--verbose-prompt") {
             params.verbose_prompt = true;
+        } else if (arg == "--no-display-prompt") {
+            params.display_prompt = false;
         } else if (arg == "-r" || arg == "--reverse-prompt") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -616,6 +643,12 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.ppl_stride = std::stoi(argv[i]);
+        } else if (arg == "-ptc" || arg == "--print-token-count") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_print = std::stoi(argv[i]);
         } else if (arg == "--ppl-output-type") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -798,7 +831,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("\n");
     printf("options:\n");
     printf("  -h, --help            show this help message and exit\n");
-    printf("      --version         show version and build info\n");
+    printf("  --version             show version and build info\n");
     printf("  -i, --interactive     run in interactive mode\n");
     printf("  --interactive-first   run in interactive mode and wait for input right away\n");
     printf("  -ins, --instruct      run in instruction mode (use with Alpaca models)\n");
@@ -895,16 +928,22 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("                        number of layers to store in VRAM\n");
     printf("  -ngld N, --n-gpu-layers-draft N\n");
     printf("                        number of layers to store in VRAM for the draft model\n");
-    printf("  -ts SPLIT --tensor-split SPLIT\n");
-    printf("                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    printf("  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
-#ifdef GGML_USE_CUBLAS
-    printf("  -nommq, --no-mul-mat-q\n");
-    printf("                        use " GGML_CUBLAS_NAME " instead of custom mul_mat_q " GGML_CUDA_NAME " kernels.\n");
-    printf("                        Not recommended since this is both slower and uses more VRAM.\n");
-#endif // GGML_USE_CUBLAS
+    printf("  -sm SPLIT_MODE, --split-mode SPLIT_MODE\n");
+    printf("                        how to split the model across multiple GPUs, one of:\n");
+    printf("                          - none: use one GPU only\n");
+    printf("                          - layer (default): split layers and KV across GPUs\n");
+    printf("                          - row: split rows across GPUs\n");
+    printf("  -ts SPLIT, --tensor-split SPLIT\n");
+    printf("                        fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1\n");
+    printf("  -mg i, --main-gpu i   the GPU to use for the model (with split-mode = none),\n");
+    printf("                        or for intermediate results and KV (with split-mode = row) (default: %d)\n", params.main_gpu);
 #endif
-    printf("  --verbose-prompt      print prompt before generation\n");
+    printf("  --verbose-prompt      print a verbose prompt before generation (default: %s)\n", params.verbose_prompt ? "true" : "false");
+    printf("  --no-display-prompt   don't print prompt at generation (default: %s)\n", !params.display_prompt ? "true" : "false");
+    printf("  -gan N, --grp-attn-n N\n");
+    printf("                        group-attention factor (default: %d)\n", params.grp_attn_n);
+    printf("  -gaw N, --grp-attn-w N\n");
+    printf("                        group-attention width (default: %.1f)\n", (double)params.grp_attn_w);
     printf("  -dkvc, --dump-kv-cache\n");
     printf("                        verbose print of the KV cache\n");
     printf("  -nkvo, --no-kv-offload\n");
@@ -926,6 +965,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  --override-kv KEY=TYPE:VALUE\n");
     printf("                        advanced option to override model metadata by key. may be specified multiple times.\n");
     printf("                        types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n");
+    printf("  -ptc N, --print-token-count N\n");
+    printf("                        print token count every N tokens (default: %d)\n", params.n_print);
     printf("\n");
 #ifndef LOG_DISABLE_LOGS
     log_print_usage();
@@ -1015,6 +1056,7 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
         mparams.n_gpu_layers = params.n_gpu_layers;
     }
     mparams.main_gpu        = params.main_gpu;
+    mparams.split_mode      = params.split_mode;
     mparams.tensor_split    = params.tensor_split;
     mparams.use_mmap        = params.use_mmap;
     mparams.use_mlock       = params.use_mlock;
@@ -1029,6 +1071,9 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
 }
 
 static ggml_type kv_cache_type_from_str(const std::string & s) {
+    if (s == "f32") {
+        return GGML_TYPE_F32;
+    }
     if (s == "f16") {
         return GGML_TYPE_F16;
     }
@@ -1540,6 +1585,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p);
     fprintf(stream, "typical_p: %f # default: 1.0\n", sparams.typical_p);
     fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false");
+    fprintf(stream, "display_prompt: %s # default: true\n", params.display_prompt ? "true" : "false");
 }
 
 //

common/common.h

@@ -59,9 +59,13 @@ struct gpt_params {
     float   p_split                         = 0.1f;  // speculative decoding split probability
     int32_t n_gpu_layers                    = -1;    // number of layers to store in VRAM (-1 - use default)
     int32_t n_gpu_layers_draft              = -1;    // number of layers to store in VRAM for the draft model (-1 - use default)
+    llama_split_mode split_mode             = LLAMA_SPLIT_LAYER; // how to split the model across GPUs
     int32_t main_gpu                        = 0;     // the GPU that is used for scratch and small tensors
     float   tensor_split[LLAMA_MAX_DEVICES] = {0};   // how split tensors should be distributed across GPUs
     int32_t n_beams                         = 0;     // if non-zero then use beam search of given width.
+    int32_t grp_attn_n                      = 1;     // group-attention factor
+    int32_t grp_attn_w                      = 512;   // group-attention width
+    int32_t n_print                         = -1;    // print token count every n tokens (-1 = disabled)
     float   rope_freq_base                  = 0.0f;  // RoPE base frequency
     float   rope_freq_scale                 = 0.0f;  // RoPE frequency scaling factor
     float   yarn_ext_factor                 = -1.0f; // YaRN extrapolation mix factor
@@ -122,6 +126,7 @@ struct gpt_params {
     bool use_mlock         = false; // use mlock to keep model in memory
     bool numa              = false; // attempt optimizations that help on some NUMA systems
     bool verbose_prompt    = false; // print prompt tokens before generation
+    bool display_prompt    = true;  // print prompt before generation
     bool infill            = false; // use infill mode
     bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
     bool no_kv_offload     = false; // disable KV offloading
@@ -240,4 +245,3 @@ void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size = 80);
 
 // Dump the KV cache view showing individual sequences in each cell (long output).
 void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
-

common/sampling.cpp

@@ -190,6 +190,11 @@ static llama_token llama_sampling_sample_impl(
         logits[it->first] += it->second;
     }
 
+    if (ctx_cfg) {
+        float * logits_guidance = llama_get_logits_ith(ctx_cfg, idx);
+        llama_sample_apply_guidance(ctx_main, logits, logits_guidance, params.cfg_scale);
+    }
+
     cur.clear();
 
     for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
@@ -198,10 +203,6 @@ static llama_token llama_sampling_sample_impl(
 
     llama_token_data_array cur_p = { cur.data(), cur.size(), false };
 
-    if (ctx_cfg) {
-        llama_sample_classifier_free_guidance(ctx_main, &cur_p, ctx_cfg, params.cfg_scale);
-    }
-
     // apply penalties
     const auto& penalty_tokens = params.use_penalty_prompt_tokens ? params.penalty_prompt_tokens : prev;
     const int penalty_tokens_used_size = std::min((int)penalty_tokens.size(), penalty_last_n);

convert-hf-to-gguf.py

@@ -23,6 +23,15 @@ if 'NO_LOCAL_GGUF' not in os.environ:
 import gguf
 
 
+# check for any of the given keys in the dictionary and return the value of the first key found
+def get_key_opts(d, keys):
+    for k in keys:
+        if k in d:
+            return d[k]
+    print(f"Could not find any of {keys}")
+    sys.exit()
+
+
 ###### MODEL DEFINITIONS ######
 
 class SentencePieceTokenTypes(IntEnum):
@@ -257,10 +266,11 @@ class Model:
                 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)
+                if hasattr(tokenizer, "added_tokens_decoder"):
+                    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)
@@ -817,10 +827,17 @@ class PersimmonModel(Model):
         hidden_size = self.hparams["hidden_size"]
 
         self.gguf_writer.add_name('persimmon-8b-chat')
+        self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
         self.gguf_writer.add_embedding_length(hidden_size)
         self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
-        self.gguf_writer.add_rope_dimension_count(hidden_size // head_count)
+
+        # NOTE: not sure about this change - why does the model not have a rope dimension count when it is smaller
+        #       than the head size?
+        #       ref: https://github.com/ggerganov/llama.cpp/pull/4889
+        # self.gguf_writer.add_rope_dimension_count(hidden_size // head_count)
+        self.gguf_writer.add_rope_dimension_count(hidden_size // head_count // 2)
+
         self.gguf_writer.add_head_count(head_count)
         self.gguf_writer.add_head_count_kv(head_count_kv)
         self.gguf_writer.add_rope_freq_base(self.hparams["rope_theta"])
@@ -1061,17 +1078,22 @@ class GPT2Model(Model):
 
 class Phi2Model(Model):
     def set_gguf_parameters(self):
-        block_count = self.hparams["n_layer"]
+        block_count = get_key_opts(self.hparams, ["num_hidden_layers", "n_layer"])
+
+        rot_pct = get_key_opts(self.hparams, ["partial_rotary_factor"])
+        n_embd = get_key_opts(self.hparams, ["hidden_size", "n_embd"])
+        n_head = get_key_opts(self.hparams, ["num_attention_heads", "n_head"])
 
         self.gguf_writer.add_name("Phi2")
-        self.gguf_writer.add_context_length(self.hparams["n_positions"])
-        self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
-        self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
+        self.gguf_writer.add_context_length(get_key_opts(self.hparams, ["n_positions", "max_position_embeddings"]))
+
+        self.gguf_writer.add_embedding_length(n_embd)
+        self.gguf_writer.add_feed_forward_length(4 * n_embd)
         self.gguf_writer.add_block_count(block_count)
-        self.gguf_writer.add_head_count(self.hparams["n_head"])
-        self.gguf_writer.add_head_count_kv(self.hparams["n_head"])
-        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
-        self.gguf_writer.add_rope_dimension_count(self.hparams["rotary_dim"])
+        self.gguf_writer.add_head_count(n_head)
+        self.gguf_writer.add_head_count_kv(n_head)
+        self.gguf_writer.add_layer_norm_eps(get_key_opts(self.hparams, ["layer_norm_epsilon", "layer_norm_eps"]))
+        self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
         self.gguf_writer.add_file_type(self.ftype)
         self.gguf_writer.add_add_bos_token(False)
 

examples/CMakeLists.txt

@@ -31,13 +31,12 @@ else()
     add_subdirectory(quantize-stats)
     add_subdirectory(save-load-state)
     add_subdirectory(simple)
+    add_subdirectory(passkey)
     add_subdirectory(speculative)
     add_subdirectory(lookahead)
     add_subdirectory(lookup)
     add_subdirectory(train-text-from-scratch)
-    if (LLAMA_METAL)
-        add_subdirectory(metal)
-    endif()
+    add_subdirectory(imatrix)
     if (LLAMA_BUILD_SERVER)
         add_subdirectory(server)
     endif()

examples/batched-bench/batched-bench.cpp

@@ -88,7 +88,10 @@ int main(int argc, char ** argv) {
 
     llama_model_params model_params = llama_model_default_params();
 
+    const std::vector<float> t_split (LLAMA_MAX_DEVICES, 0.0f);
+
     model_params.n_gpu_layers = n_gpu_layers;
+    model_params.tensor_split = t_split.data();
 
     llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 

examples/batched/batched.cpp

@@ -69,6 +69,7 @@ int main(int argc, char ** argv) {
 
     std::vector<llama_token> tokens_list;
     tokens_list = ::llama_tokenize(model, params.prompt, true);
+
     const int n_kv_req = tokens_list.size() + (n_len - tokens_list.size())*n_parallel;
 
     // initialize the context

examples/benchmark/benchmark-matmult.cpp

@@ -194,7 +194,7 @@ int main(int argc, char ** argv)  {
     // Set up a the benchmark matrices
     // printf("Creating new tensor q11 & Running quantize\n");
     struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
-    ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements, hist_cur.data());
+    ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements/m11->ne[0], m11->ne[0], hist_cur.data(), nullptr);
 
     // Set up a the compute graph
     // printf("Creating new tensor q31\n");
@@ -207,7 +207,7 @@ int main(int argc, char ** argv)  {
     // Set up a second graph computation to make sure we override the CPU cache lines
     // printf("Creating new tensor q12 & Running quantize\n");
     struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
-    ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements, hist_cur.data());
+    ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements/m12->ne[0], m12->ne[0], hist_cur.data(), nullptr);
 
     // printf("Creating new tensor q32\n");
     struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2);

examples/export-lora/export-lora.cpp

@@ -245,9 +245,8 @@ static struct lora_data * load_lora(struct lora_info * info) {
     params_ggml.no_alloc   = true;
     result->ctx = ggml_init(params_ggml);
 
-    uint32_t LLAMA_FILE_MAGIC_LORA = 0x67676C61; // 'ggla'
     uint32_t magic   = file.read_u32();
-    if (magic != LLAMA_FILE_MAGIC_LORA) {
+    if (magic != LLAMA_FILE_MAGIC_GGLA) {
         die_fmt("unexpected lora header file magic in '%s'", info->filename.c_str());
     }
     uint32_t version = file.read_u32();

examples/imatrix/CMakeLists.txt

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

examples/imatrix/imatrix.cpp

@@ -0,0 +1,380 @@
+#include "common.h"
+#include "llama.h"
+
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <sstream>
+#include <thread>
+#include <mutex>
+#include <vector>
+#include <fstream>
+#include <unordered_map>
+#include <algorithm>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+struct Stats {
+    std::vector<float> values;
+    int ncall = 0;
+};
+
+struct StatParams {
+    std::string ofile = "imatrix.dat";
+    int         n_output_frequency = 10;
+    int         verbosity = 1;
+    bool        collect_output_weight = false;
+};
+
+class IMatrixCollector {
+public:
+    IMatrixCollector() = default;
+    void set_parameters(StatParams&& params) { m_params = std::move(params); }
+    void collect_imatrix(const struct ggml_tensor * src0, const struct ggml_tensor * src1);
+    void save_imatrix() const;
+private:
+    std::unordered_map<std::string, Stats> m_stats;
+    StatParams                             m_params;
+    std::mutex                             m_mutex;
+    int                                    m_last_call = 0;
+};
+
+void IMatrixCollector::collect_imatrix(const struct ggml_tensor * src0, const struct ggml_tensor * src1) {
+    if (src1->ne[1] < 16 || src1->type != GGML_TYPE_F32) return;
+    if (!(strncmp(src0->name, "blk.", 4) == 0 || (m_params.collect_output_weight && strcmp(src0->name, "output.weight") == 0))) return;
+    std::lock_guard<std::mutex> lock(m_mutex);
+    auto& e = m_stats[src0->name];
+    if (e.values.empty()) {
+        e.values.resize(src1->ne[0], 0);
+    }
+    else if (e.values.size() != (size_t)src1->ne[0]) {
+        fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", src0->name, (int)e.values.size(), (int)src1->ne[0]);
+        exit(1); //GGML_ASSERT(false);
+    }
+    ++e.ncall;
+    if (m_params.verbosity > 1) {
+        printf("%s[%d]: %s, %d x %d, %d\n",__func__,m_last_call,src0->name,(int)src1->ne[0],(int)src1->ne[1],(int)src1->type);
+    }
+    for (int row = 0; row < (int)src1->ne[1]; ++row) {
+        const float * x = (const float *)src1->data + row * src1->ne[0];
+        for (int j = 0; j < (int)src1->ne[0]; ++j) {
+            e.values[j] += x[j]*x[j];
+        }
+    }
+    if (e.ncall > m_last_call) {
+        m_last_call = e.ncall;
+        if (m_last_call % m_params.n_output_frequency == 0) {
+            save_imatrix();
+        }
+    }
+}
+
+void IMatrixCollector::save_imatrix() const {
+    const char * fname = m_params.ofile.empty() ? "imatrix.dat" : m_params.ofile.c_str();
+    std::ofstream out(fname, std::ios::binary);
+    int n_entries = m_stats.size();
+    out.write((const char*)&n_entries, sizeof(n_entries));
+    for (auto& p : m_stats) {
+        int len = p.first.size();
+        out.write((const char*)&len, sizeof(len));
+        out.write(p.first.c_str(), len);
+        out.write((const char*)&p.second.ncall, sizeof(p.second.ncall));
+        int nval = p.second.values.size();
+        out.write((const char*)&nval, sizeof(nval));
+        if (nval > 0) out.write((const char*)p.second.values.data(), nval*sizeof(float));
+    }
+    if (m_params.verbosity > 0) {
+        fprintf(stderr, "\n%s: stored collected data after %d chunks in %s\n",__func__,m_last_call,fname);
+    }
+}
+
+static IMatrixCollector g_collector;
+
+static void ik_collect_imatrix(const struct ggml_tensor * src0, const struct ggml_tensor * src1) {
+    g_collector.collect_imatrix(src0, src1);
+}
+
+
+struct results_log_softmax {
+    double log_softmax;
+    float  logit;
+    float  prob;
+};
+
+static std::vector<float> softmax(const std::vector<float>& logits) {
+    std::vector<float> probs(logits.size());
+    float max_logit = logits[0];
+    for (float v : logits) {
+        max_logit = std::max(max_logit, v);
+    }
+    double sum_exp = 0.0;
+    for (size_t i = 0; i < logits.size(); i++) {
+        // Subtract the maximum logit value from the current logit value for numerical stability
+        const float logit = logits[i] - max_logit;
+        const float exp_logit = expf(logit);
+        sum_exp += exp_logit;
+        probs[i] = exp_logit;
+    }
+    for (size_t i = 0; i < probs.size(); i++) {
+        probs[i] /= sum_exp;
+    }
+    return probs;
+}
+
+static results_log_softmax log_softmax(int n_vocab, const float * logits, int tok) {
+    float max_logit = logits[0];
+    for (int i = 1; i < n_vocab; ++i) {
+        max_logit = std::max(max_logit, logits[i]);
+    }
+    double sum_exp = 0.0;
+    for (int i = 0; i < n_vocab; ++i) {
+        sum_exp += expf(logits[i] - max_logit);
+    }
+    return {logits[tok] - max_logit - log(sum_exp), logits[tok], expf(logits[tok] - max_logit) / (float) sum_exp};
+}
+
+static void process_logits(
+    int n_vocab, const float * logits, const int * tokens, int n_token, std::vector<std::thread> & workers,
+    double & nll, double & nll2, float * logit_history, float * prob_history
+) {
+    std::mutex mutex;
+    int counter = 0;
+    auto compute = [&mutex, &counter, &nll, &nll2, logit_history, prob_history, n_vocab, logits, tokens, n_token] () {
+        double local_nll  = 0;
+        double local_nll2 = 0;
+        while (true) {
+            std::unique_lock<std::mutex> lock(mutex);
+            int i = counter++;
+            if (i >= n_token) {
+                nll += local_nll; nll2 += local_nll2;
+                break;
+            }
+            lock.unlock();
+            const results_log_softmax results = log_softmax(n_vocab, logits + i*n_vocab, tokens[i+1]);
+            const double v = -results.log_softmax;
+            local_nll += v;
+            local_nll2 += v*v;
+
+            logit_history[i] = results.logit;
+            prob_history[i]  = results.prob;
+        }
+    };
+    for (auto & w : workers) {
+        w = std::thread(compute);
+    }
+    compute();
+    for (auto & w : workers) {
+        w.join();
+    }
+}
+
+static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
+
+    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+    const int n_ctx = llama_n_ctx(ctx);
+
+    auto tim1 = std::chrono::high_resolution_clock::now();
+    fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
+
+    std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
+
+    auto tim2 = std::chrono::high_resolution_clock::now();
+    fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
+
+    if (int(tokens.size()) < 2*n_ctx) {
+        fprintf(stderr, "%s: you need at least %d tokens for a context of %d tokens\n",__func__,2*n_ctx,
+                n_ctx);
+        fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
+        return false;
+    }
+
+    std::vector<float> logit_history;
+    logit_history.resize(tokens.size());
+
+    std::vector<float> prob_history;
+    prob_history.resize(tokens.size());
+
+    const int n_chunk_max = tokens.size() / n_ctx;
+
+    const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+    const int n_batch = params.n_batch;
+
+    int count = 0;
+    double nll = 0.0;
+    double nll2 = 0.0;
+
+    fprintf(stderr, "%s: computing over %d chunks with batch_size %d\n", __func__, n_chunk, n_batch);
+
+    std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
+
+    for (int i = 0; i < n_chunk; ++i) {
+        const int start =     i * n_ctx;
+        const int end   = start + n_ctx;
+
+        const int num_batches = (n_ctx + n_batch - 1) / n_batch;
+
+        std::vector<float> logits;
+
+        const auto t_start = std::chrono::high_resolution_clock::now();
+
+        // clear the KV cache
+        llama_kv_cache_clear(ctx);
+
+        for (int j = 0; j < num_batches; ++j) {
+            const int batch_start = start + j * n_batch;
+            const int batch_size  = std::min(end - batch_start, n_batch);
+
+            // save original token and restore it after eval
+            const auto token_org = tokens[batch_start];
+
+            // add BOS token for the first batch of each chunk
+            if (add_bos && j == 0) {
+                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))) {
+                fprintf(stderr, "%s : failed to eval\n", __func__);
+                return false;
+            }
+
+            // restore the original token in case it was set to BOS
+            tokens[batch_start] = token_org;
+
+            const auto * batch_logits = llama_get_logits(ctx);
+            logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
+        }
+
+        const auto t_end = std::chrono::high_resolution_clock::now();
+
+        if (i == 0) {
+            const float t_total = std::chrono::duration<float>(t_end - t_start).count();
+            fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
+            int total_seconds = (int)(t_total * n_chunk);
+            if (total_seconds >= 60*60) {
+                fprintf(stderr, "%d hours ", total_seconds / (60*60));
+                total_seconds = total_seconds % (60*60);
+            }
+            fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
+        }
+
+        const int first = n_ctx/2;
+        process_logits(n_vocab, logits.data() + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
+                       workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first);
+        count += n_ctx - first - 1;
+
+        printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
+        fflush(stdout);
+    }
+    printf("\n");
+
+    nll2 /= count;
+    nll /= count;
+    const double ppl = exp(nll);
+    nll2 -= nll * nll;
+    if (nll2 > 0) {
+        nll2 = sqrt(nll2/(count-1));
+        printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl);
+    } else {
+        printf("Unexpected negative standard deviation of log(prob)\n");
+    }
+
+    return true;
+}
+
+int main(int argc, char ** argv) {
+
+    StatParams sparams;
+    std::vector<char*> args;
+    args.push_back(argv[0]);
+    int iarg = 1;
+    for (; iarg < argc-1; ++iarg) {
+        std::string arg{argv[iarg]};
+        if (arg == "-o" || arg == "--output-file") {
+            sparams.ofile = argv[++iarg];
+        }
+        else if (arg == "-ofreq" || arg == "--output-frequency") {
+            sparams.n_output_frequency = std::stoi(argv[++iarg]);
+        }
+        else if (arg == "-ow" || arg == "--output-weight") {
+            sparams.collect_output_weight = std::stoi(argv[++iarg]);
+        }
+        else if (arg == "--verbosity") {
+            sparams.verbosity = std::stoi(argv[++iarg]);
+        } else {
+            args.push_back(argv[iarg]);
+        }
+    }
+    if (iarg < argc) {
+        args.push_back(argv[iarg]);
+    }
+
+    gpt_params params;
+    params.n_batch = 512;
+    if (!gpt_params_parse(args.size(), args.data(), params)) {
+        return 1;
+    }
+
+    g_collector.set_parameters(std::move(sparams));
+
+    ggml_set_imatrix_collection(ik_collect_imatrix);
+
+    params.logits_all = true;
+    params.n_batch = std::min(params.n_batch, params.n_ctx);
+
+    print_build_info();
+
+    if (params.seed == LLAMA_DEFAULT_SEED) {
+        params.seed = time(NULL);
+    }
+
+    fprintf(stderr, "%s: seed  = %u\n", __func__, params.seed);
+
+    std::mt19937 rng(params.seed);
+    if (params.random_prompt) {
+        params.prompt = gpt_random_prompt(rng);
+    }
+
+    llama_backend_init(params.numa);
+
+    llama_model * model;
+    llama_context * ctx;
+
+    // load the model and apply lora adapter, if any
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    if (model == NULL) {
+        fprintf(stderr, "%s: error: unable to load model\n", __func__);
+        return 1;
+    }
+
+    const int n_ctx_train = llama_n_ctx_train(model);
+    if (params.n_ctx > n_ctx_train) {
+        fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
+                __func__, n_ctx_train, params.n_ctx);
+    }
+
+    // print system information
+    {
+        fprintf(stderr, "\n");
+        fprintf(stderr, "%s\n", get_system_info(params).c_str());
+    }
+
+    bool OK = compute_imatrix(ctx, params);
+    if (!OK) {
+        return 1;
+    }
+
+    g_collector.save_imatrix();
+
+    llama_print_timings(ctx);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    return 0;
+}

examples/llama-bench/llama-bench.cpp

@@ -128,6 +128,25 @@ static std::string get_gpu_info() {
 // command line params
 enum output_formats {CSV, JSON, MARKDOWN, SQL};
 
+static const char * output_format_str(output_formats format) {
+    switch (format) {
+        case CSV:      return "csv";
+        case JSON:     return "json";
+        case MARKDOWN: return "md";
+        case SQL:      return "sql";
+        default: GGML_ASSERT(!"invalid output format");
+    }
+}
+
+static const char * split_mode_str(llama_split_mode mode) {
+    switch (mode) {
+        case LLAMA_SPLIT_NONE:  return "none";
+        case LLAMA_SPLIT_LAYER: return "layer";
+        case LLAMA_SPLIT_ROW:   return "row";
+        default: GGML_ASSERT(!"invalid split mode");
+    }
+}
+
 struct cmd_params {
     std::vector<std::string> model;
     std::vector<int> n_prompt;
@@ -137,7 +156,9 @@ struct cmd_params {
     std::vector<ggml_type> type_v;
     std::vector<int> n_threads;
     std::vector<int> n_gpu_layers;
+    std::vector<llama_split_mode> split_mode;
     std::vector<int> main_gpu;
+    std::vector<bool> no_kv_offload;
     std::vector<bool> mul_mat_q;
     std::vector<std::array<float, LLAMA_MAX_DEVICES>> tensor_split;
     int reps;
@@ -154,7 +175,9 @@ static const cmd_params cmd_params_defaults = {
     /* type_v        */ {GGML_TYPE_F16},
     /* n_threads     */ {get_num_physical_cores()},
     /* n_gpu_layers  */ {99},
+    /* split_mode    */ {LLAMA_SPLIT_LAYER},
     /* main_gpu      */ {0},
+    /* no_kv_offload */ {false},
     /* mul_mat_q     */ {true},
     /* tensor_split  */ {{}},
     /* reps          */ 5,
@@ -167,20 +190,22 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("\n");
     printf("options:\n");
     printf("  -h, --help\n");
-    printf("  -m, --model <filename>            (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
-    printf("  -p, --n-prompt <n>                (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
-    printf("  -n, --n-gen <n>                   (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
-    printf("  -b, --batch-size <n>              (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
-    printf("  -ctk <t>, --cache-type-k <t>      (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
-    printf("  -ctv <t>, --cache-type-v <t>      (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
-    printf("  -t, --threads <n>                 (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
-    printf("  -ngl, --n-gpu-layers <n>          (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
-    printf("  -mg, --main-gpu <i>               (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
-    printf("  -mmq, --mul-mat-q <0|1>           (default: %s)\n", join(cmd_params_defaults.mul_mat_q, ",").c_str());
-    printf("  -ts, --tensor_split <ts0/ts1/..>               \n");
-    printf("  -r, --repetitions <n>             (default: %d)\n", cmd_params_defaults.reps);
-    printf("  -o, --output <csv|json|md|sql>    (default: %s)\n", cmd_params_defaults.output_format == CSV ? "csv" : cmd_params_defaults.output_format == JSON ? "json" : cmd_params_defaults.output_format == MARKDOWN ? "md" : "sql");
-    printf("  -v, --verbose                     (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
+    printf("  -m, --model <filename>              (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
+    printf("  -p, --n-prompt <n>                  (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
+    printf("  -n, --n-gen <n>                     (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
+    printf("  -b, --batch-size <n>                (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
+    printf("  -ctk <t>, --cache-type-k <t>        (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
+    printf("  -ctv <t>, --cache-type-v <t>        (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
+    printf("  -t, --threads <n>                   (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
+    printf("  -ngl, --n-gpu-layers <n>            (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
+    printf("  -sm, --split-mode <none|layer|row>  (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str());
+    printf("  -mg, --main-gpu <i>                 (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
+    printf("  -nkvo, --no-kv-offload <0|1>        (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
+    printf("  -mmq, --mul-mat-q <0|1>             (default: %s)\n", join(cmd_params_defaults.mul_mat_q, ",").c_str());
+    printf("  -ts, --tensor_split <ts0/ts1/..>    (default: 0)\n");
+    printf("  -r, --repetitions <n>               (default: %d)\n", cmd_params_defaults.reps);
+    printf("  -o, --output <csv|json|md|sql>      (default: %s)\n", output_format_str(cmd_params_defaults.output_format));
+    printf("  -v, --verbose                       (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
     printf("\n");
     printf("Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
 }
@@ -303,12 +328,41 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = split<int>(argv[i], split_delim);
             params.n_gpu_layers.insert(params.n_gpu_layers.end(), p.begin(), p.end());
+        } else if (arg == "-sm" || arg == "--split-mode") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<std::string>(argv[i], split_delim);
+            std::vector<llama_split_mode> modes;
+            for (const auto & m : p) {
+                llama_split_mode mode;
+                if (m == "none") {
+                    mode = LLAMA_SPLIT_NONE;
+                } else if (m == "layer") {
+                    mode = LLAMA_SPLIT_LAYER;
+                } else if (m == "row") {
+                    mode = LLAMA_SPLIT_ROW;
+                } else {
+                    invalid_param = true;
+                    break;
+                }
+                modes.push_back(mode);
+            }
+            params.split_mode.insert(params.split_mode.end(), modes.begin(), modes.end());
         } else if (arg == "-mg" || arg == "--main-gpu") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
             params.main_gpu = split<int>(argv[i], split_delim);
+        } else if (arg == "-nkvo" || arg == "--no-kv-offload") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<bool>(argv[i], split_delim);
+            params.no_kv_offload.insert(params.no_kv_offload.end(), p.begin(), p.end());
         } else if (arg == "-mmq" || arg == "--mul-mat-q") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -382,7 +436,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.type_k.empty())       { params.type_k = cmd_params_defaults.type_k; }
     if (params.type_v.empty())       { params.type_v = cmd_params_defaults.type_v; }
     if (params.n_gpu_layers.empty()) { params.n_gpu_layers = cmd_params_defaults.n_gpu_layers; }
+    if (params.split_mode.empty())   { params.split_mode = cmd_params_defaults.split_mode; }
     if (params.main_gpu.empty())     { params.main_gpu = cmd_params_defaults.main_gpu; }
+    if (params.no_kv_offload.empty()){ params.no_kv_offload = cmd_params_defaults.no_kv_offload; }
     if (params.mul_mat_q.empty())    { params.mul_mat_q = cmd_params_defaults.mul_mat_q; }
     if (params.tensor_split.empty()) { params.tensor_split = cmd_params_defaults.tensor_split; }
     if (params.n_threads.empty())    { params.n_threads = cmd_params_defaults.n_threads; }
@@ -399,7 +455,9 @@ struct cmd_params_instance {
     ggml_type type_v;
     int n_threads;
     int n_gpu_layers;
+    llama_split_mode split_mode;
     int main_gpu;
+    bool no_kv_offload;
     bool mul_mat_q;
     std::array<float, LLAMA_MAX_DEVICES> tensor_split;
 
@@ -407,6 +465,7 @@ struct cmd_params_instance {
         llama_model_params mparams = llama_model_default_params();
 
         mparams.n_gpu_layers = n_gpu_layers;
+        mparams.split_mode = split_mode;
         mparams.main_gpu = main_gpu;
         mparams.tensor_split = tensor_split.data();
 
@@ -416,6 +475,7 @@ struct cmd_params_instance {
     bool equal_mparams(const cmd_params_instance & other) const {
         return model == other.model &&
                n_gpu_layers == other.n_gpu_layers &&
+               split_mode == other.split_mode &&
                main_gpu == other.main_gpu &&
                tensor_split == other.tensor_split;
     }
@@ -428,54 +488,26 @@ struct cmd_params_instance {
         cparams.type_k = type_k;
         cparams.type_v = type_v;
         cparams.mul_mat_q = mul_mat_q;
+        cparams.offload_kqv = !no_kv_offload;
 
         return cparams;
     }
 };
 
-static std::vector<cmd_params_instance> get_cmd_params_instances_int(const cmd_params & params, int n_gen, int n_prompt) {
-    std::vector<cmd_params_instance> instances;
-
-    for (const auto & m : params.model)
-    for (const auto & nl : params.n_gpu_layers)
-    for (const auto & mg : params.main_gpu)
-    for (const auto & ts : params.tensor_split)
-    for (const auto & nb : params.n_batch)
-    for (const auto & tk : params.type_k)
-    for (const auto & tv : params.type_v)
-    for (const auto & mmq : params.mul_mat_q)
-    for (const auto & nt : params.n_threads) {
-        cmd_params_instance instance = {
-            /* .model        = */ m,
-            /* .n_prompt     = */ n_prompt,
-            /* .n_gen        = */ n_gen,
-            /* .n_batch      = */ nb,
-            /* .type_k       = */ tk,
-            /* .type_v       = */ tv,
-            /* .n_threads    = */ nt,
-            /* .n_gpu_layers = */ nl,
-            /* .main_gpu     = */ mg,
-            /* .mul_mat_q    = */ mmq,
-            /* .tensor_split = */ ts,
-        };
-        instances.push_back(instance);
-    }
-    return instances;
-}
-
 static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_params & params) {
     std::vector<cmd_params_instance> instances;
 
-#if 1
     // this ordering minimizes the number of times that each model needs to be reloaded
     for (const auto & m : params.model)
     for (const auto & nl : params.n_gpu_layers)
+    for (const auto & sm : params.split_mode)
     for (const auto & mg : params.main_gpu)
     for (const auto & ts : params.tensor_split)
     for (const auto & nb : params.n_batch)
     for (const auto & tk : params.type_k)
     for (const auto & tv : params.type_v)
     for (const auto & mmq : params.mul_mat_q)
+    for (const auto & nkvo : params.no_kv_offload)
     for (const auto & nt : params.n_threads) {
         for (const auto & n_prompt : params.n_prompt) {
             if (n_prompt == 0) {
@@ -490,7 +522,9 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
                 /* .n_gpu_layers = */ nl,
+                /* .split_mode   = */ sm,
                 /* .main_gpu     = */ mg,
+                /* .no_kv_offload= */ nkvo,
                 /* .mul_mat_q    = */ mmq,
                 /* .tensor_split = */ ts,
             };
@@ -510,31 +544,15 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
                 /* .n_gpu_layers = */ nl,
+                /* .split_mode   = */ sm,
                 /* .main_gpu     = */ mg,
+                /* .no_kv_offload= */ nkvo,
                 /* .mul_mat_q    = */ mmq,
                 /* .tensor_split = */ ts,
             };
             instances.push_back(instance);
         }
     }
-#else
-    // this ordering separates the prompt and generation tests
-    for (const auto & n_prompt : params.n_prompt) {
-        if (n_prompt == 0) {
-            continue;
-        }
-        auto instances_prompt = get_cmd_params_instances_int(params, 0, n_prompt);
-        instances.insert(instances.end(), instances_prompt.begin(), instances_prompt.end());
-    }
-
-    for (const auto & n_gen : params.n_gen) {
-        if (n_gen == 0) {
-            continue;
-        }
-        auto instances_gen = get_cmd_params_instances_int(params, n_gen, 0);
-        instances.insert(instances.end(), instances_gen.begin(), instances_gen.end());
-    }
-#endif
 
     return instances;
 }
@@ -558,7 +576,9 @@ struct test {
     ggml_type type_k;
     ggml_type type_v;
     int n_gpu_layers;
+    llama_split_mode split_mode;
     int main_gpu;
+    bool no_kv_offload;
     bool mul_mat_q;
     std::array<float, LLAMA_MAX_DEVICES> tensor_split;
     int n_prompt;
@@ -578,7 +598,9 @@ struct test {
         type_k = inst.type_k;
         type_v = inst.type_v;
         n_gpu_layers = inst.n_gpu_layers;
+        split_mode = inst.split_mode;
         main_gpu = inst.main_gpu;
+        no_kv_offload = inst.no_kv_offload;
         mul_mat_q = inst.mul_mat_q;
         tensor_split = inst.tensor_split;
         n_prompt = inst.n_prompt;
@@ -640,7 +662,9 @@ struct test {
             "cpu_info", "gpu_info",
             "model_filename", "model_type", "model_size", "model_n_params",
             "n_batch", "n_threads", "type_k", "type_v",
-            "n_gpu_layers", "main_gpu", "mul_mat_q", "tensor_split",
+            "n_gpu_layers", "split_mode",
+            "main_gpu", "no_kv_offload",
+            "mul_mat_q", "tensor_split",
             "n_prompt", "n_gen", "test_time",
             "avg_ns", "stddev_ns",
             "avg_ts", "stddev_ts"
@@ -659,7 +683,7 @@ struct test {
             return INT;
         }
         if (field == "cuda" || field == "opencl" || field == "metal" || field == "gpu_blas" || field == "blas" ||
-            field == "f16_kv" || field == "mul_mat_q") {
+            field == "f16_kv" || field == "no_kv_offload" || field == "mul_mat_q") {
             return BOOL;
         }
         if (field == "avg_ts" || field == "stddev_ts") {
@@ -690,7 +714,9 @@ struct test {
             cpu_info, gpu_info,
             model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
             std::to_string(n_batch), std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
-            std::to_string(n_gpu_layers), std::to_string(main_gpu), std::to_string(mul_mat_q), tensor_split_str,
+            std::to_string(n_gpu_layers), split_mode_str(split_mode),
+            std::to_string(main_gpu), std::to_string(no_kv_offload),
+            std::to_string(mul_mat_q), tensor_split_str,
             std::to_string(n_prompt), std::to_string(n_gen), test_time,
             std::to_string(avg_ns()), std::to_string(stdev_ns()),
             std::to_string(avg_ts()), std::to_string(stdev_ts())
@@ -845,12 +871,18 @@ struct markdown_printer : public printer {
         if (field == "n_gpu_layers") {
             return "ngl";
         }
+        if (field == "split_mode") {
+            return "sm";
+        }
         if (field == "n_threads") {
             return "threads";
         }
         if (field == "mul_mat_q") {
             return "mmq";
         }
+        if (field == "no_kv_offload") {
+            return "nkvo";
+        }
         if (field == "tensor_split") {
             return "ts";
         }
@@ -882,9 +914,15 @@ struct markdown_printer : public printer {
         if (params.main_gpu.size() > 1 || params.main_gpu != cmd_params_defaults.main_gpu) {
             fields.push_back("main_gpu");
         }
+        if (params.split_mode.size() > 1 || params.split_mode != cmd_params_defaults.split_mode) {
+            fields.push_back("split_mode");
+        }
         if (params.mul_mat_q.size() > 1 || params.mul_mat_q != cmd_params_defaults.mul_mat_q) {
             fields.push_back("mul_mat_q");
         }
+        if (params.no_kv_offload.size() > 1 || params.no_kv_offload != cmd_params_defaults.no_kv_offload) {
+            fields.push_back("no_kv_offload");
+        }
         if (params.tensor_split.size() > 1 || params.tensor_split != cmd_params_defaults.tensor_split) {
             fields.push_back("tensor_split");
         }

examples/llama.swiftui/README.md

@@ -1,7 +1,12 @@
-# llama.swiftui
+# llama.cpp/examples/llama.swiftui
 
-Local inference of llama.cpp on an iPhone.
-So far I only tested with starcoder 1B model, but it can most likely handle 7B models as well.
+Local inference of llama.cpp on an iPhone. This is a sample app that can be used as a starting
+point for more advanced projects.
+
+For usage instructions and performance stats, check the following discussion: https://github.com/ggerganov/llama.cpp/discussions/4508
+
+![image](https://github.com/ggerganov/llama.cpp/assets/1991296/2b40284f-8421-47a2-b634-74eece09a299)
+
+Video demonstration:
 
 https://github.com/bachittle/llama.cpp/assets/39804642/e290827a-4edb-4093-9642-2a5e399ec545
-

examples/llama.swiftui/llama.cpp.swift/LibLlama.swift

@@ -1,8 +1,5 @@
 import Foundation
-
-// To use this in your own project, add llama.cpp as a swift package dependency
-// and uncomment this import line.
-// import llama
+import llama
 
 enum LlamaError: Error {
     case couldNotInitializeContext

examples/llama.swiftui/llama.cpp.swift/bridging-header.h

@@ -1,5 +0,0 @@
-//
-//  Use this file to import your target's public headers that you would like to expose to Swift.
-//
-
-#import "llama.h"

examples/llama.swiftui/llama.swiftui.xcodeproj/project.pbxproj

@@ -7,52 +7,33 @@
 	objects = {
 
 /* Begin PBXBuildFile section */
-		542376082B0D9BFB008E6A1C /* ggml-quants.c in Sources */ = {isa = PBXBuildFile; fileRef = 542376072B0D9BFB008E6A1C /* ggml-quants.c */; settings = {COMPILER_FLAGS = "-O3"; }; };
-		5423760B2B0D9C4B008E6A1C /* ggml-backend.c in Sources */ = {isa = PBXBuildFile; fileRef = 5423760A2B0D9C4B008E6A1C /* ggml-backend.c */; settings = {COMPILER_FLAGS = "-O3"; }; };
-		542EA09D2AC8723900A8AEE9 /* ggml.c in Sources */ = {isa = PBXBuildFile; fileRef = 542EA09B2AC8723900A8AEE9 /* ggml.c */; settings = {COMPILER_FLAGS = "-DGGML_USE_ACCELERATE -DGGML_USE_METAL -DGGML_USE_K_QUANTS -O3"; }; };
-		542EA0A02AC8725700A8AEE9 /* ggml-alloc.c in Sources */ = {isa = PBXBuildFile; fileRef = 542EA09F2AC8725700A8AEE9 /* ggml-alloc.c */; settings = {COMPILER_FLAGS = "-O3"; }; };
-		542EA0A32AC8729100A8AEE9 /* llama.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 542EA0A12AC8729100A8AEE9 /* llama.cpp */; settings = {COMPILER_FLAGS = "-DGGML_USE_K_QUANTS -DGGML_USE_METAL -O3"; }; };
 		549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 549479CA2AC9E16000E0F78B /* Metal.framework */; };
-		549479CD2AC9E42A00E0F78B /* ggml-metal.m in Sources */ = {isa = PBXBuildFile; fileRef = 549479C52AC9E0F200E0F78B /* ggml-metal.m */; settings = {COMPILER_FLAGS = "-fno-objc-arc -DGGML_SWIFT -DGGML_USE_METAL -O3"; }; };
+		79E1D9CD2B4CD16E005F8E46 /* InputButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = 79E1D9CC2B4CD16E005F8E46 /* InputButton.swift */; };
 		7FA3D2B32B2EA2F600543F92 /* DownloadButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */; };
 		8A1C83772AC328BD0096AF73 /* llama_swiftuiApp.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */; };
 		8A1C83792AC328BD0096AF73 /* ContentView.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A1C83782AC328BD0096AF73 /* ContentView.swift */; };
 		8A1C837B2AC328BE0096AF73 /* Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 8A1C837A2AC328BE0096AF73 /* Assets.xcassets */; };
-		8A1C837E2AC328BE0096AF73 /* Preview Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 8A1C837D2AC328BE0096AF73 /* Preview Assets.xcassets */; };
 		8A39BE0A2AC7601100BFEB40 /* Accelerate.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 8A39BE092AC7601000BFEB40 /* Accelerate.framework */; };
 		8A3F84242AC4C891005E2EE8 /* models in Resources */ = {isa = PBXBuildFile; fileRef = 8A3F84232AC4C891005E2EE8 /* models */; };
 		8A907F332AC7138A006146EA /* LibLlama.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A907F322AC7134E006146EA /* LibLlama.swift */; };
 		8A9F7C4D2AC332EE008AE1EA /* LlamaState.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */; };
-		F1FE20DC2B465C4500B45541 /* ggml-metal.metal in Resources */ = {isa = PBXBuildFile; fileRef = 549479C82AC9E10B00E0F78B /* ggml-metal.metal */; };
+		DF810E132B4A5BA200301144 /* llama in Frameworks */ = {isa = PBXBuildFile; productRef = DF810E122B4A5BA200301144 /* llama */; };
 		F1FE20E22B465ECA00B45541 /* LoadCustomButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = F1FE20E12B465EC900B45541 /* LoadCustomButton.swift */; };
 /* End PBXBuildFile section */
 
 /* Begin PBXFileReference section */
-		542376062B0D9BEA008E6A1C /* ggml-quants.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-quants.h"; path = "../../ggml-quants.h"; sourceTree = "<group>"; };
-		542376072B0D9BFB008E6A1C /* ggml-quants.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-quants.c"; path = "../../ggml-quants.c"; sourceTree = "<group>"; };
-		542376092B0D9C40008E6A1C /* ggml-backend.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-backend.h"; path = "../../ggml-backend.h"; sourceTree = "<group>"; };
-		5423760A2B0D9C4B008E6A1C /* ggml-backend.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-backend.c"; path = "../../ggml-backend.c"; sourceTree = "<group>"; };
-		542EA09B2AC8723900A8AEE9 /* ggml.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = ggml.c; path = ../../ggml.c; sourceTree = "<group>"; };
-		542EA09C2AC8723900A8AEE9 /* ggml.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = ggml.h; path = ../../ggml.h; sourceTree = "<group>"; };
-		542EA09E2AC8725700A8AEE9 /* ggml-alloc.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-alloc.h"; path = "../../ggml-alloc.h"; sourceTree = "<group>"; };
-		542EA09F2AC8725700A8AEE9 /* ggml-alloc.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-alloc.c"; path = "../../ggml-alloc.c"; sourceTree = "<group>"; };
-		542EA0A12AC8729100A8AEE9 /* llama.cpp */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.cpp; name = llama.cpp; path = ../../llama.cpp; sourceTree = "<group>"; };
-		542EA0A22AC8729100A8AEE9 /* llama.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = llama.h; path = ../../llama.h; sourceTree = "<group>"; };
-		549479C52AC9E0F200E0F78B /* ggml-metal.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; name = "ggml-metal.m"; path = "../../ggml-metal.m"; sourceTree = "<group>"; };
-		549479C62AC9E0F200E0F78B /* ggml-metal.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-metal.h"; path = "../../ggml-metal.h"; sourceTree = "<group>"; };
-		549479C82AC9E10B00E0F78B /* ggml-metal.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; name = "ggml-metal.metal"; path = "../../ggml-metal.metal"; sourceTree = "<group>"; };
 		549479CA2AC9E16000E0F78B /* Metal.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = Metal.framework; path = System/Library/Frameworks/Metal.framework; sourceTree = SDKROOT; };
+		79E1D9CC2B4CD16E005F8E46 /* InputButton.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = InputButton.swift; sourceTree = "<group>"; };
 		7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = DownloadButton.swift; sourceTree = "<group>"; };
-		8A08D20A2AC73B1500FE6CD4 /* bridging-header.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = "bridging-header.h"; sourceTree = "<group>"; };
 		8A1C83732AC328BD0096AF73 /* llama.swiftui.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = llama.swiftui.app; sourceTree = BUILT_PRODUCTS_DIR; };
 		8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = llama_swiftuiApp.swift; sourceTree = "<group>"; };
 		8A1C83782AC328BD0096AF73 /* ContentView.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ContentView.swift; sourceTree = "<group>"; };
 		8A1C837A2AC328BE0096AF73 /* Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = Assets.xcassets; sourceTree = "<group>"; };
-		8A1C837D2AC328BE0096AF73 /* Preview Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = "Preview Assets.xcassets"; sourceTree = "<group>"; };
 		8A39BE092AC7601000BFEB40 /* Accelerate.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = Accelerate.framework; path = System/Library/Frameworks/Accelerate.framework; sourceTree = SDKROOT; };
 		8A3F84232AC4C891005E2EE8 /* models */ = {isa = PBXFileReference; lastKnownFileType = folder; name = models; path = llama.swiftui/Resources/models; sourceTree = "<group>"; };
 		8A907F322AC7134E006146EA /* LibLlama.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LibLlama.swift; sourceTree = "<group>"; };
 		8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LlamaState.swift; sourceTree = "<group>"; };
+		DF2D2FE72B4A59BE00FCB72D /* llama.cpp */ = {isa = PBXFileReference; lastKnownFileType = wrapper; name = llama.cpp; path = ../..; sourceTree = "<group>"; };
 		F1FE20E12B465EC900B45541 /* LoadCustomButton.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LoadCustomButton.swift; sourceTree = "<group>"; };
 /* End PBXFileReference section */
 
@@ -61,6 +42,7 @@
 			isa = PBXFrameworksBuildPhase;
 			buildActionMask = 2147483647;
 			files = (
+				DF810E132B4A5BA200301144 /* llama in Frameworks */,
 				549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */,
 				8A39BE0A2AC7601100BFEB40 /* Accelerate.framework in Frameworks */,
 			);
@@ -69,30 +51,10 @@
 /* End PBXFrameworksBuildPhase section */
 
 /* Begin PBXGroup section */
-		8A08D1F62AC7383900FE6CD4 /* llama.cpp */ = {
-			isa = PBXGroup;
-			children = (
-				5423760A2B0D9C4B008E6A1C /* ggml-backend.c */,
-				542376092B0D9C40008E6A1C /* ggml-backend.h */,
-				542376062B0D9BEA008E6A1C /* ggml-quants.h */,
-				542376072B0D9BFB008E6A1C /* ggml-quants.c */,
-				549479C82AC9E10B00E0F78B /* ggml-metal.metal */,
-				549479C62AC9E0F200E0F78B /* ggml-metal.h */,
-				549479C52AC9E0F200E0F78B /* ggml-metal.m */,
-				542EA09B2AC8723900A8AEE9 /* ggml.c */,
-				542EA09C2AC8723900A8AEE9 /* ggml.h */,
-				542EA09F2AC8725700A8AEE9 /* ggml-alloc.c */,
-				542EA09E2AC8725700A8AEE9 /* ggml-alloc.h */,
-				542EA0A12AC8729100A8AEE9 /* llama.cpp */,
-				542EA0A22AC8729100A8AEE9 /* llama.h */,
-			);
-			name = llama.cpp;
-			sourceTree = "<group>";
-		};
 		8A1C836A2AC328BD0096AF73 = {
 			isa = PBXGroup;
 			children = (
-				8A08D1F62AC7383900FE6CD4 /* llama.cpp */,
+				DF2D2FE72B4A59BE00FCB72D /* llama.cpp */,
 				8A907F312AC7134E006146EA /* llama.cpp.swift */,
 				8A3F84232AC4C891005E2EE8 /* models */,
 				8A1C83752AC328BD0096AF73 /* llama.swiftui */,
@@ -117,19 +79,10 @@
 				8A9F7C4A2AC332BF008AE1EA /* UI */,
 				8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */,
 				8A1C837A2AC328BE0096AF73 /* Assets.xcassets */,
-				8A1C837C2AC328BE0096AF73 /* Preview Content */,
 			);
 			path = llama.swiftui;
 			sourceTree = "<group>";
 		};
-		8A1C837C2AC328BE0096AF73 /* Preview Content */ = {
-			isa = PBXGroup;
-			children = (
-				8A1C837D2AC328BE0096AF73 /* Preview Assets.xcassets */,
-			);
-			path = "Preview Content";
-			sourceTree = "<group>";
-		};
 		8A39BE082AC7601000BFEB40 /* Frameworks */ = {
 			isa = PBXGroup;
 			children = (
@@ -157,7 +110,6 @@
 		8A907F312AC7134E006146EA /* llama.cpp.swift */ = {
 			isa = PBXGroup;
 			children = (
-				8A08D20A2AC73B1500FE6CD4 /* bridging-header.h */,
 				8A907F322AC7134E006146EA /* LibLlama.swift */,
 			);
 			path = llama.cpp.swift;
@@ -169,6 +121,7 @@
 				7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */,
 				8A1C83782AC328BD0096AF73 /* ContentView.swift */,
 				F1FE20E12B465EC900B45541 /* LoadCustomButton.swift */,
+				79E1D9CC2B4CD16E005F8E46 /* InputButton.swift */,
 			);
 			path = UI;
 			sourceTree = "<group>";
@@ -198,6 +151,7 @@
 			);
 			name = llama.swiftui;
 			packageProductDependencies = (
+				DF810E122B4A5BA200301144 /* llama */,
 			);
 			productName = llama.swiftui;
 			productReference = 8A1C83732AC328BD0096AF73 /* llama.swiftui.app */;
@@ -244,9 +198,7 @@
 			isa = PBXResourcesBuildPhase;
 			buildActionMask = 2147483647;
 			files = (
-				F1FE20DC2B465C4500B45541 /* ggml-metal.metal in Resources */,
 				8A3F84242AC4C891005E2EE8 /* models in Resources */,
-				8A1C837E2AC328BE0096AF73 /* Preview Assets.xcassets in Resources */,
 				8A1C837B2AC328BE0096AF73 /* Assets.xcassets in Resources */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
@@ -258,18 +210,13 @@
 			isa = PBXSourcesBuildPhase;
 			buildActionMask = 2147483647;
 			files = (
-				542376082B0D9BFB008E6A1C /* ggml-quants.c in Sources */,
-				549479CD2AC9E42A00E0F78B /* ggml-metal.m in Sources */,
 				F1FE20E22B465ECA00B45541 /* LoadCustomButton.swift in Sources */,
-				542EA09D2AC8723900A8AEE9 /* ggml.c in Sources */,
 				8A907F332AC7138A006146EA /* LibLlama.swift in Sources */,
-				542EA0A32AC8729100A8AEE9 /* llama.cpp in Sources */,
 				8A9F7C4D2AC332EE008AE1EA /* LlamaState.swift in Sources */,
 				8A1C83792AC328BD0096AF73 /* ContentView.swift in Sources */,
 				8A1C83772AC328BD0096AF73 /* llama_swiftuiApp.swift in Sources */,
 				7FA3D2B32B2EA2F600543F92 /* DownloadButton.swift in Sources */,
-				542EA0A02AC8725700A8AEE9 /* ggml-alloc.c in Sources */,
-				5423760B2B0D9C4B008E6A1C /* ggml-backend.c in Sources */,
+				79E1D9CD2B4CD16E005F8E46 /* InputButton.swift in Sources */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
 		};
@@ -399,12 +346,10 @@
 			isa = XCBuildConfiguration;
 			buildSettings = {
 				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
-				ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
 				CLANG_ENABLE_MODULES = YES;
 				CODE_SIGN_STYLE = Automatic;
 				CURRENT_PROJECT_VERSION = 1;
-				DEVELOPMENT_ASSET_PATHS = "\"llama.swiftui/Preview Content\"";
-				DEVELOPMENT_TEAM = STLSG3FG8Q;
+				DEVELOPMENT_TEAM = K5UQJPP73A;
 				ENABLE_PREVIEWS = YES;
 				GENERATE_INFOPLIST_FILE = YES;
 				INFOPLIST_KEY_UIApplicationSceneManifest_Generation = YES;
@@ -420,11 +365,12 @@
 				MARKETING_VERSION = 1.0;
 				PRODUCT_BUNDLE_IDENTIFIER = "com.bachittle.llama-swift";
 				PRODUCT_NAME = "$(TARGET_NAME)";
+				SUPPORTED_PLATFORMS = "iphoneos iphonesimulator xros xrsimulator";
+				SUPPORTS_XR_DESIGNED_FOR_IPHONE_IPAD = NO;
 				SWIFT_EMIT_LOC_STRINGS = YES;
-				SWIFT_OBJC_BRIDGING_HEADER = "llama.cpp.swift/bridging-header.h";
 				SWIFT_OPTIMIZATION_LEVEL = "-Onone";
 				SWIFT_VERSION = 5.0;
-				TARGETED_DEVICE_FAMILY = "1,2";
+				TARGETED_DEVICE_FAMILY = "1,2,7";
 			};
 			name = Debug;
 		};
@@ -432,12 +378,10 @@
 			isa = XCBuildConfiguration;
 			buildSettings = {
 				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
-				ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
 				CLANG_ENABLE_MODULES = YES;
 				CODE_SIGN_STYLE = Automatic;
 				CURRENT_PROJECT_VERSION = 1;
-				DEVELOPMENT_ASSET_PATHS = "\"llama.swiftui/Preview Content\"";
-				DEVELOPMENT_TEAM = STLSG3FG8Q;
+				DEVELOPMENT_TEAM = K5UQJPP73A;
 				ENABLE_PREVIEWS = YES;
 				GENERATE_INFOPLIST_FILE = YES;
 				INFOPLIST_KEY_UIApplicationSceneManifest_Generation = YES;
@@ -453,10 +397,11 @@
 				MARKETING_VERSION = 1.0;
 				PRODUCT_BUNDLE_IDENTIFIER = "com.bachittle.llama-swift";
 				PRODUCT_NAME = "$(TARGET_NAME)";
+				SUPPORTED_PLATFORMS = "iphoneos iphonesimulator xros xrsimulator";
+				SUPPORTS_XR_DESIGNED_FOR_IPHONE_IPAD = NO;
 				SWIFT_EMIT_LOC_STRINGS = YES;
-				SWIFT_OBJC_BRIDGING_HEADER = "llama.cpp.swift/bridging-header.h";
 				SWIFT_VERSION = 5.0;
-				TARGETED_DEVICE_FAMILY = "1,2";
+				TARGETED_DEVICE_FAMILY = "1,2,7";
 			};
 			name = Release;
 		};
@@ -482,6 +427,13 @@
 			defaultConfigurationName = Release;
 		};
 /* End XCConfigurationList section */
+
+/* Begin XCSwiftPackageProductDependency section */
+		DF810E122B4A5BA200301144 /* llama */ = {
+			isa = XCSwiftPackageProductDependency;
+			productName = llama;
+		};
+/* End XCSwiftPackageProductDependency section */
 	};
 	rootObject = 8A1C836B2AC328BD0096AF73 /* Project object */;
 }

examples/llama.swiftui/llama.swiftui/Assets.xcassets/AccentColor.colorset/Contents.json

@@ -1,11 +0,0 @@
-{
-  "colors" : [
-    {
-      "idiom" : "universal"
-    }
-  ],
-  "info" : {
-    "author" : "xcode",
-    "version" : 1
-  }
-}

examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift

@@ -1,9 +1,19 @@
 import Foundation
 
+struct Model: Identifiable {
+    var id = UUID()
+    var name: String
+    var url: String
+    var filename: String
+    var status: String?
+}
+
 @MainActor
 class LlamaState: ObservableObject {
     @Published var messageLog = ""
     @Published var cacheCleared = false
+    @Published var downloadedModels: [Model] = []
+    @Published var undownloadedModels: [Model] = []
     let NS_PER_S = 1_000_000_000.0
 
     private var llamaContext: LlamaContext?
@@ -13,23 +23,102 @@ class LlamaState: ObservableObject {
     }
 
     init() {
+        loadModelsFromDisk()
+        loadDefaultModels()
+    }
+
+    private func loadModelsFromDisk() {
+        do {
+            let documentsURL = getDocumentsDirectory()
+            let modelURLs = try FileManager.default.contentsOfDirectory(at: documentsURL, includingPropertiesForKeys: nil, options: [.skipsHiddenFiles, .skipsSubdirectoryDescendants])
+            for modelURL in modelURLs {
+                let modelName = modelURL.deletingPathExtension().lastPathComponent
+                downloadedModels.append(Model(name: modelName, url: "", filename: modelURL.lastPathComponent, status: "downloaded"))
+            }
+        } catch {
+            print("Error loading models from disk: \(error)")
+        }
+    }
+
+    private func loadDefaultModels() {
         do {
             try loadModel(modelUrl: defaultModelUrl)
         } catch {
             messageLog += "Error!\n"
         }
+
+        for model in defaultModels {
+            let fileURL = getDocumentsDirectory().appendingPathComponent(model.filename)
+            if FileManager.default.fileExists(atPath: fileURL.path) {
+
+            } else {
+                var undownloadedModel = model
+                undownloadedModel.status = "download"
+                undownloadedModels.append(undownloadedModel)
+            }
+        }
     }
 
+    func getDocumentsDirectory() -> URL {
+        let paths = FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)
+        return paths[0]
+    }
+    private let defaultModels: [Model] = [
+        Model(name: "TinyLlama-1.1B (Q4_0, 0.6 GiB)",url: "https://huggingface.co/TheBloke/TinyLlama-1.1B-1T-OpenOrca-GGUF/resolve/main/tinyllama-1.1b-1t-openorca.Q4_0.gguf?download=true",filename: "tinyllama-1.1b-1t-openorca.Q4_0.gguf", status: "download"),
+        Model(
+            name: "TinyLlama-1.1B Chat (Q8_0, 1.1 GiB)",
+            url: "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF/resolve/main/tinyllama-1.1b-chat-v1.0.Q8_0.gguf?download=true",
+            filename: "tinyllama-1.1b-chat-v1.0.Q8_0.gguf", status: "download"
+        ),
+
+        Model(
+            name: "TinyLlama-1.1B (F16, 2.2 GiB)",
+            url: "https://huggingface.co/ggml-org/models/resolve/main/tinyllama-1.1b/ggml-model-f16.gguf?download=true",
+            filename: "tinyllama-1.1b-f16.gguf", status: "download"
+        ),
+
+        Model(
+            name: "Phi-2.7B (Q4_0, 1.6 GiB)",
+            url: "https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q4_0.gguf?download=true",
+            filename: "phi-2-q4_0.gguf", status: "download"
+        ),
+
+        Model(
+            name: "Phi-2.7B (Q8_0, 2.8 GiB)",
+            url: "https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q8_0.gguf?download=true",
+            filename: "phi-2-q8_0.gguf", status: "download"
+        ),
+
+        Model(
+            name: "Mistral-7B-v0.1 (Q4_0, 3.8 GiB)",
+            url: "https://huggingface.co/TheBloke/Mistral-7B-v0.1-GGUF/resolve/main/mistral-7b-v0.1.Q4_0.gguf?download=true",
+            filename: "mistral-7b-v0.1.Q4_0.gguf", status: "download"
+        ),
+        Model(
+            name: "OpenHermes-2.5-Mistral-7B (Q3_K_M, 3.52 GiB)",
+            url: "https://huggingface.co/TheBloke/OpenHermes-2.5-Mistral-7B-GGUF/resolve/main/openhermes-2.5-mistral-7b.Q3_K_M.gguf?download=true",
+            filename: "openhermes-2.5-mistral-7b.Q3_K_M.gguf", status: "download"
+        )
+    ]
     func loadModel(modelUrl: URL?) throws {
         if let modelUrl {
             messageLog += "Loading model...\n"
             llamaContext = try LlamaContext.create_context(path: modelUrl.path())
             messageLog += "Loaded model \(modelUrl.lastPathComponent)\n"
+
+            // Assuming that the model is successfully loaded, update the downloaded models
+            updateDownloadedModels(modelName: modelUrl.lastPathComponent, status: "downloaded")
         } else {
             messageLog += "Load a model from the list below\n"
         }
     }
 
+
+    private func updateDownloadedModels(modelName: String, status: String) {
+        undownloadedModels.removeAll { $0.name == modelName }
+    }
+
+
     func complete(text: String) async {
         guard let llamaContext else {
             return

examples/llama.swiftui/llama.swiftui/Preview Content/Preview Assets.xcassets/Contents.json

@@ -1,6 +0,0 @@
-{
-  "info" : {
-    "author" : "xcode",
-    "version" : 1
-  }
-}

examples/llama.swiftui/llama.swiftui/UI/ContentView.swift

@@ -2,115 +2,57 @@ import SwiftUI
 
 struct ContentView: View {
     @StateObject var llamaState = LlamaState()
-
     @State private var multiLineText = ""
-
-    private static func cleanupModelCaches() {
-        // Delete all models (*.gguf)
-        let fileManager = FileManager.default
-        let documentsUrl =  FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)[0]
-        do {
-            let fileURLs = try fileManager.contentsOfDirectory(at: documentsUrl, includingPropertiesForKeys: nil)
-            for fileURL in fileURLs {
-                if fileURL.pathExtension == "gguf" {
-                    try fileManager.removeItem(at: fileURL)
-                }
-            }
-        } catch {
-            print("Error while enumerating files \(documentsUrl.path): \(error.localizedDescription)")
-        }
-    }
+    @State private var showingHelp = false    // To track if Help Sheet should be shown
 
     var body: some View {
-        VStack {
-            ScrollView(.vertical, showsIndicators: true) {
-                Text(llamaState.messageLog)
-                .font(.system(size: 12))
-                .frame(maxWidth: .infinity, alignment: .leading)
+        NavigationView {
+            VStack {
+                ScrollView(.vertical, showsIndicators: true) {
+                    Text(llamaState.messageLog)
+                        .font(.system(size: 12))
+                        .frame(maxWidth: .infinity, alignment: .leading)
+                        .padding()
+                        .onTapGesture {
+                            UIApplication.shared.sendAction(#selector(UIResponder.resignFirstResponder), to: nil, from: nil, for: nil)
+                        }
+                }
+
+                TextEditor(text: $multiLineText)
+                    .frame(height: 80)
+                    .padding()
+                    .border(Color.gray, width: 0.5)
+
+                HStack {
+                    Button("Send") {
+                        sendText()
+                    }
+
+                    Button("Bench") {
+                        bench()
+                    }
+
+                    Button("Clear") {
+                        clear()
+                    }
+
+                    Button("Copy") {
+                        UIPasteboard.general.string = llamaState.messageLog
+                    }
+                }
+                .buttonStyle(.bordered)
                 .padding()
-                .onTapGesture {
-                    UIApplication.shared.sendAction(#selector(UIResponder.resignFirstResponder), to: nil, from: nil, for: nil)
-                }
-            }
 
-            TextEditor(text: $multiLineText)
-                .frame(height: 80)
+                NavigationLink(destination: DrawerView(llamaState: llamaState)) {
+                    Text("View Models")
+                }
                 .padding()
-                .border(Color.gray, width: 0.5)
 
-            HStack {
-                Button("Send") {
-                    sendText()
-                }
-
-                Button("Bench") {
-                    bench()
-                }
-
-                Button("Clear") {
-                    clear()
-                }
-
-                Button("Copy") {
-                    UIPasteboard.general.string = llamaState.messageLog
-                }
-            }.buttonStyle(.bordered)
-
-            VStack(alignment: .leading) {
-                DownloadButton(
-                    llamaState: llamaState,
-                    modelName: "TinyLlama-1.1B (Q4_0, 0.6 GiB)",
-                    modelUrl: "https://huggingface.co/TheBloke/TinyLlama-1.1B-1T-OpenOrca-GGUF/resolve/main/tinyllama-1.1b-1t-openorca.Q4_0.gguf?download=true",
-                    filename: "tinyllama-1.1b-1t-openorca.Q4_0.gguf"
-                )
-
-                DownloadButton(
-                    llamaState: llamaState,
-                    modelName: "TinyLlama-1.1B (Q8_0, 1.1 GiB)",
-                    modelUrl: "https://huggingface.co/TheBloke/TinyLlama-1.1B-1T-OpenOrca-GGUF/resolve/main/tinyllama-1.1b-1t-openorca.Q8_0.gguf?download=true",
-                    filename: "tinyllama-1.1b-1t-openorca.Q8_0.gguf"
-                )
-
-                DownloadButton(
-                    llamaState: llamaState,
-                    modelName: "TinyLlama-1.1B (F16, 2.2 GiB)",
-                    modelUrl: "https://huggingface.co/ggml-org/models/resolve/main/tinyllama-1.1b/ggml-model-f16.gguf?download=true",
-                    filename: "tinyllama-1.1b-f16.gguf"
-                )
-
-                DownloadButton(
-                    llamaState: llamaState,
-                    modelName: "Phi-2.7B (Q4_0, 1.6 GiB)",
-                    modelUrl: "https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q4_0.gguf?download=true",
-                    filename: "phi-2-q4_0.gguf"
-                )
-
-                DownloadButton(
-                    llamaState: llamaState,
-                    modelName: "Phi-2.7B (Q8_0, 2.8 GiB)",
-                    modelUrl: "https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q8_0.gguf?download=true",
-                    filename: "phi-2-q8_0.gguf"
-                )
-
-                DownloadButton(
-                    llamaState: llamaState,
-                    modelName: "Mistral-7B-v0.1 (Q4_0, 3.8 GiB)",
-                    modelUrl: "https://huggingface.co/TheBloke/Mistral-7B-v0.1-GGUF/resolve/main/mistral-7b-v0.1.Q4_0.gguf?download=true",
-                    filename: "mistral-7b-v0.1.Q4_0.gguf"
-                )
-
-                Button("Clear downloaded models") {
-                    ContentView.cleanupModelCaches()
-                    llamaState.cacheCleared = true
-                }
-
-                LoadCustomButton(llamaState: llamaState)
             }
-            .padding(.top, 4)
-            .font(.system(size: 12))
-            .frame(maxWidth: .infinity, alignment: .leading)
+            .padding()
+            .navigationBarTitle("Model Settings", displayMode: .inline)
+
         }
-        .padding()
     }
 
     func sendText() {
@@ -131,8 +73,73 @@ struct ContentView: View {
             await llamaState.clear()
         }
     }
+    struct DrawerView: View {
+
+        @ObservedObject var llamaState: LlamaState
+        @State private var showingHelp = false
+        func delete(at offsets: IndexSet) {
+            offsets.forEach { offset in
+                let model = llamaState.downloadedModels[offset]
+                let fileURL = getDocumentsDirectory().appendingPathComponent(model.filename)
+                do {
+                    try FileManager.default.removeItem(at: fileURL)
+                } catch {
+                    print("Error deleting file: \(error)")
+                }
+            }
+
+            // Remove models from downloadedModels array
+            llamaState.downloadedModels.remove(atOffsets: offsets)
+        }
+
+        func getDocumentsDirectory() -> URL {
+            let paths = FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)
+            return paths[0]
+        }
+        var body: some View {
+            List {
+                Section(header: Text("Download Models From Hugging Face")) {
+                    HStack {
+                        InputButton(llamaState: llamaState)
+                    }
+                }
+                Section(header: Text("Downloaded Models")) {
+                    ForEach(llamaState.downloadedModels) { model in
+                        DownloadButton(llamaState: llamaState, modelName: model.name, modelUrl: model.url, filename: model.filename)
+                    }
+                    .onDelete(perform: delete)
+                }
+                Section(header: Text("Default Models")) {
+                    ForEach(llamaState.undownloadedModels) { model in
+                        DownloadButton(llamaState: llamaState, modelName: model.name, modelUrl: model.url, filename: model.filename)
+                    }
+                }
+
+            }
+            .listStyle(GroupedListStyle())
+            .navigationBarTitle("Model Settings", displayMode: .inline).toolbar {
+                ToolbarItem(placement: .navigationBarTrailing) {
+                    Button("Help") {
+                        showingHelp = true
+                    }
+                }
+            }.sheet(isPresented: $showingHelp) {    // Sheet for help modal
+                VStack(alignment: .leading) {
+                    VStack(alignment: .leading) {
+                        Text("1. Make sure the model is in GGUF Format")
+                               .padding()
+                        Text("2. Copy the download link of the quantized model")
+                               .padding()
+                    }
+                    Spacer()
+                   }
+            }
+        }
+    }
 }
 
-//#Preview {
-//    ContentView()
-//}
+struct ContentView_Previews: PreviewProvider {
+    static var previews: some View {
+        ContentView()
+    }
+}

examples/llama.swiftui/llama.swiftui/UI/DownloadButton.swift

@@ -53,6 +53,8 @@ struct DownloadButton: View {
 
                     llamaState.cacheCleared = false
 
+                    let model = Model(name: modelName, url: modelUrl, filename: filename, status: "downloaded")
+                    llamaState.downloadedModels.append(model)
                     status = "downloaded"
                 }
             } catch let err {

examples/llama.swiftui/llama.swiftui/UI/InputButton.swift

@@ -0,0 +1,131 @@
+import SwiftUI
+
+struct InputButton: View {
+    @ObservedObject var llamaState: LlamaState
+    @State private var inputLink: String = ""
+    @State private var status: String = "download"
+    @State private var filename: String = ""
+
+    @State private var downloadTask: URLSessionDownloadTask?
+    @State private var progress = 0.0
+    @State private var observation: NSKeyValueObservation?
+
+    private static func extractModelInfo(from link: String) -> (modelName: String, filename: String)? {
+        guard let url = URL(string: link),
+              let lastPathComponent = url.lastPathComponent.components(separatedBy: ".").first,
+              let modelName = lastPathComponent.components(separatedBy: "-").dropLast().joined(separator: "-").removingPercentEncoding,
+              let filename = lastPathComponent.removingPercentEncoding else {
+            return nil
+        }
+
+        return (modelName, filename)
+    }
+
+    private static func getFileURL(filename: String) -> URL {
+        FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)[0].appendingPathComponent(filename)
+    }
+
+    private func download() {
+        guard let extractedInfo = InputButton.extractModelInfo(from: inputLink) else {
+            // Handle invalid link or extraction failure
+            return
+        }
+
+        let (modelName, filename) = extractedInfo
+        self.filename = filename  // Set the state variable
+
+        status = "downloading"
+        print("Downloading model \(modelName) from \(inputLink)")
+        guard let url = URL(string: inputLink) else { return }
+        let fileURL = InputButton.getFileURL(filename: filename)
+
+        downloadTask = URLSession.shared.downloadTask(with: url) { temporaryURL, response, error in
+            if let error = error {
+                print("Error: \(error.localizedDescription)")
+                return
+            }
+
+            guard let response = response as? HTTPURLResponse, (200...299).contains(response.statusCode) else {
+                print("Server error!")
+                return
+            }
+
+            do {
+                if let temporaryURL = temporaryURL {
+                    try FileManager.default.copyItem(at: temporaryURL, to: fileURL)
+                    print("Writing to \(filename) completed")
+
+                    llamaState.cacheCleared = false
+
+                    let model = Model(name: modelName, url: self.inputLink, filename: filename, status: "downloaded")
+                    llamaState.downloadedModels.append(model)
+                    status = "downloaded"
+                }
+            } catch let err {
+                print("Error: \(err.localizedDescription)")
+            }
+        }
+
+        observation = downloadTask?.progress.observe(\.fractionCompleted) { progress, _ in
+            self.progress = progress.fractionCompleted
+        }
+
+        downloadTask?.resume()
+    }
+
+    var body: some View {
+        VStack {
+            HStack {
+                TextField("Paste Quantized Download Link", text: $inputLink)
+                    .textFieldStyle(RoundedBorderTextFieldStyle())
+
+                Button(action: {
+                    downloadTask?.cancel()
+                    status = "download"
+                }) {
+                    Text("Cancel")
+                }
+            }
+
+            if status == "download" {
+                Button(action: download) {
+                    Text("Download Custom Model")
+                }
+            } else if status == "downloading" {
+                Button(action: {
+                    downloadTask?.cancel()
+                    status = "download"
+                }) {
+                    Text("Downloading \(Int(progress * 100))%")
+                }
+            } else if status == "downloaded" {
+                Button(action: {
+                    let fileURL = InputButton.getFileURL(filename: self.filename)
+                    if !FileManager.default.fileExists(atPath: fileURL.path) {
+                        download()
+                        return
+                    }
+                    do {
+                        try llamaState.loadModel(modelUrl: fileURL)
+                    } catch let err {
+                        print("Error: \(err.localizedDescription)")
+                    }
+                }) {
+                    Text("Load Custom Model")
+                }
+            } else {
+                Text("Unknown status")
+            }
+        }
+        .onDisappear() {
+            downloadTask?.cancel()
+        }
+        .onChange(of: llamaState.cacheCleared) { newValue in
+            if newValue {
+                downloadTask?.cancel()
+                let fileURL = InputButton.getFileURL(filename: self.filename)
+                status = FileManager.default.fileExists(atPath: fileURL.path) ? "downloaded" : "download"
+            }
+        }
+    }
+}

examples/llava/clip.cpp

@@ -126,24 +126,7 @@ static struct ggml_tensor * get_tensor(struct ggml_context * ctx, const std::str
 }
 
 static std::string get_ftype(int ftype) {
-    switch (ftype) {
-    case 0:
-        return "f32";
-    case 1:
-        return "f16";
-    case 2:
-        return "q4_0";
-    case 3:
-        return "q4_1";
-    case 6:
-        return "q5_0";
-    case 7:
-        return "q5_1";
-    case 8:
-        return "q8_0";
-    default:
-        throw std::runtime_error(format("%s: Unrecognized file type: %d\n", __func__, ftype));
-    }
+    return ggml_type_name(static_cast<ggml_type>(ftype));
 }
 
 //
@@ -533,6 +516,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
     buffer_size += n_tensors * 128 /* CLIP PADDING */;
 
     clip_ctx * new_clip = new clip_ctx;
+
 #ifdef GGML_USE_CUBLAS
     new_clip->backend = ggml_backend_cuda_init(0);
     printf("%s: CLIP using CUDA backend\n", __func__);
@@ -543,6 +527,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
     printf("%s: CLIP using Metal backend\n", __func__);
 #endif
 
+
     if (!new_clip->backend) {
         new_clip->backend = ggml_backend_cpu_init();
         printf("%s: CLIP using CPU backend\n", __func__);
@@ -931,26 +916,8 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
 
     ggml_type type = GGML_TYPE_Q4_1;
 
-    switch (itype) {
-        case 2:
-            type = GGML_TYPE_Q4_0;
-            break;
-        case 3:
-            type = GGML_TYPE_Q4_1;
-            break;
-        case 6:
-            type = GGML_TYPE_Q5_0;
-            break;
-        case 7:
-            type = GGML_TYPE_Q5_1;
-            break;
-        case 8:
-            type = GGML_TYPE_Q8_0;
-            break;
-        default:
-            fprintf(stderr, "%s: invalid quantization type %d\n", __func__, itype);
-            return false;
-    };
+    assert(itype < GGML_TYPE_COUNT);
+    type = static_cast<ggml_type>(itype);
 
     auto * ctx_clip = clip_model_load(fname_inp, 2);
 
@@ -1010,6 +977,10 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
 
         if (quantize) {
             new_type = type;
+            if (new_type >= GGML_TYPE_Q2_K && name.find("embd") != std::string::npos) {
+                new_type = GGML_TYPE_Q8_0; // ggml_get_rows needs non K type
+                // fprintf(stderr, "%s: quantizing %s to %s\n", __func__, name.c_str(), ggml_type_name(new_type));
+            }
             const size_t n_elms = ggml_nelements(cur);
             float * f32_data;
 
@@ -1054,6 +1025,21 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
                 case GGML_TYPE_Q8_0: {
                     new_size = ggml_quantize_q8_0(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
                 } break;
+                case GGML_TYPE_Q2_K: {
+                    new_size = ggml_quantize_q2_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q3_K: {
+                    new_size = ggml_quantize_q3_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q4_K: {
+                    new_size = ggml_quantize_q4_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q5_K: {
+                    new_size = ggml_quantize_q5_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q6_K: {
+                    new_size = ggml_quantize_q6_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
                 default: {
                     fprintf(stderr, "%s: unsupported quantization type %d\n", __func__, new_type);
                     return false;

examples/llava/llava-cli.cpp

@@ -243,6 +243,9 @@ int main(int argc, char ** argv) {
     }
 
     auto image_embed = load_image(ctx_llava, &params);
+    if (!image_embed) {
+        return 1;
+    }
 
     // process the prompt
     process_prompt(ctx_llava, image_embed, &params, params.prompt);

examples/main/main.cpp

@@ -439,6 +439,21 @@ int main(int argc, char ** argv) {
     LOG_TEE("sampling: \n%s\n", llama_sampling_print(sparams).c_str());
     LOG_TEE("sampling order: \n%s\n", llama_sampling_order_print(sparams).c_str());
     LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
+
+    // group-attention state
+    // number of grouped KV tokens so far (used only if params.grp_attn_n > 1)
+    int ga_i = 0;
+
+    const int ga_n = params.grp_attn_n;
+    const int ga_w = params.grp_attn_w;
+
+    if (ga_n != 1) {
+        GGML_ASSERT(ga_n > 0                    && "grp_attn_n must be positive");                     // NOLINT
+        GGML_ASSERT(ga_w % ga_n == 0            && "grp_attn_w must be a multiple of grp_attn_n");     // NOLINT
+      //GGML_ASSERT(n_ctx_train % ga_w == 0     && "n_ctx_train must be a multiple of grp_attn_w");    // NOLINT
+      //GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * grp_attn_n"); // NOLINT
+        LOG_TEE("self-extend: n_ctx_train = %d, grp_attn_n = %d, grp_attn_w = %d\n", n_ctx_train, ga_n, ga_w);
+    }
     LOG_TEE("\n\n");
 
     if (params.interactive) {
@@ -462,6 +477,7 @@ int main(int argc, char ** argv) {
 
     bool is_antiprompt        = false;
     bool input_echo           = true;
+    bool display              = true;
     bool need_to_save_session = !path_session.empty() && n_matching_session_tokens < embd_inp.size();
 
     int n_past             = 0;
@@ -476,6 +492,7 @@ int main(int argc, char ** argv) {
 
     // the first thing we will do is to output the prompt, so set color accordingly
     console::set_display(console::prompt);
+    display = params.display_prompt;
 
     std::vector<llama_token> embd;
     std::vector<llama_token> embd_guidance;
@@ -485,7 +502,7 @@ int main(int argc, char ** argv) {
     while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
         // predict
         if (!embd.empty()) {
-            // Note: n_ctx - 4 here is to match the logic for commandline prompt handling via
+            // Note: (n_ctx - 4) here is to match the logic for commandline prompt handling via
             // --prompt or --file which uses the same value.
             int max_embd_size = n_ctx - 4;
 
@@ -500,37 +517,61 @@ int main(int argc, char ** argv) {
                 fflush(stdout);
             }
 
-            // infinite text generation via context swapping
-            // if we run out of context:
-            // - take the n_keep first tokens from the original prompt (via n_past)
-            // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
-            if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) > n_ctx) {
-                if (params.n_predict == -2) {
-                    LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
-                    break;
+            if (ga_n == 1) {
+                // infinite text generation via context shifting
+                // if we run out of context:
+                // - take the n_keep first tokens from the original prompt (via n_past)
+                // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
+                if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) > n_ctx) {
+                    if (params.n_predict == -2) {
+                        LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
+                        break;
+                    }
+
+                    const int n_left    = n_past - params.n_keep - 1;
+                    const int n_discard = n_left/2;
+
+                    LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
+                            n_past, n_left, n_ctx, params.n_keep, n_discard);
+
+                    llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                    llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+
+                    n_past -= n_discard;
+
+                    if (ctx_guidance) {
+                        n_past_guidance -= n_discard;
+                    }
+
+                    LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
+
+                    LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
+
+                    LOG("clear session path\n");
+                    path_session.clear();
                 }
+            } else {
+                // context extension via Self-Extend
+                while (n_past >= ga_i + ga_w) {
+                    const int ib = (ga_n*ga_i)/ga_w;
+                    const int bd = (ga_w/ga_n)*(ga_n - 1);
+                    const int dd = (ga_w/ga_n) - ib*bd - ga_w;
 
-                const int n_left    = n_past - params.n_keep - 1;
-                const int n_discard = n_left/2;
+                    LOG("\n");
+                    LOG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i, n_past, ib*bd, ga_i + ib*bd, n_past + ib*bd);
+                    LOG("div:   [%6d, %6d] / %6d -> [%6d, %6d]\n", ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n, (ga_i + ib*bd)/ga_n, (ga_i + ib*bd + ga_w)/ga_n);
+                    LOG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i + ib*bd + ga_w, n_past + ib*bd, dd, ga_i + ib*bd + ga_w + dd, n_past + ib*bd + dd);
 
-                LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
-                    n_past, n_left, n_ctx, params.n_keep, n_discard);
+                    llama_kv_cache_seq_shift(ctx, 0, ga_i,                n_past,              ib*bd);
+                    llama_kv_cache_seq_div  (ctx, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
+                    llama_kv_cache_seq_shift(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
 
-                llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
-                llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+                    n_past -= bd;
 
-                n_past -= n_discard;
+                    ga_i += ga_w/ga_n;
 
-                if (ctx_guidance) {
-                    n_past_guidance -= n_discard;
+                    LOG("\nn_past_old = %d, n_past = %d, ga_i = %d\n\n", n_past + bd, n_past, ga_i);
                 }
-
-                LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
-
-                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
-
-                LOG("clear session path\n");
-                path_session.clear();
             }
 
             // try to reuse a matching prefix from the loaded session instead of re-eval (via n_past)
@@ -611,6 +652,10 @@ int main(int argc, char ** argv) {
                 n_past += n_eval;
 
                 LOG("n_past = %d\n", n_past);
+                // Display total tokens alongside total time
+                if (params.n_print > 0 && n_past % params.n_print == 0) {
+                    LOG_TEE("\n\033[31mTokens consumed so far = %d / %d \033[0m\n", n_past, n_ctx);
+                }
             }
 
             if (!embd.empty() && !path_session.empty()) {
@@ -664,7 +709,7 @@ int main(int argc, char ** argv) {
         }
 
         // display text
-        if (input_echo) {
+        if (input_echo && display) {
             for (auto id : embd) {
                 const std::string token_str = llama_token_to_piece(ctx, id);
                 printf("%s", token_str.c_str());
@@ -681,6 +726,7 @@ int main(int argc, char ** argv) {
         // reset color to default if there is no pending user input
         if (input_echo && (int) embd_inp.size() == n_consumed) {
             console::set_display(console::reset);
+            display = true;
         }
 
         // if not currently processing queued inputs;
@@ -753,6 +799,7 @@ int main(int argc, char ** argv) {
 
                 // color user input only
                 console::set_display(console::user_input);
+                display = params.display_prompt;
 
                 std::string line;
                 bool another_line = true;
@@ -763,6 +810,7 @@ int main(int argc, char ** argv) {
 
                 // done taking input, reset color
                 console::set_display(console::reset);
+                display = true;
 
                 // Add tokens to embd only if the input buffer is non-empty
                 // Entering a empty line lets the user pass control back

examples/metal/CMakeLists.txt

@@ -1,4 +0,0 @@
-set(TEST_TARGET metal)
-add_executable(${TEST_TARGET} metal.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TEST_TARGET} PRIVATE ggml)

examples/metal/metal.cpp

@@ -1,103 +0,0 @@
-// Evaluate a statically exported ggml computation graph with Metal
-//
-// - First, export a LLaMA graph:
-//
-//  $ ./bin/main -m ../models/7B/ggml-model-q4_0.gguf --export
-//
-// - Run this tool to evaluate the exported graph:
-//
-//  $ ./bin/metal llama.ggml
-//
-// The purpose of this tool is mostly for debugging and demonstration purposes.
-// The main limitation of exporting computation graphs is that their sizes are static which often
-// can be a problem for real-world applications.
-//
-
-#include "ggml.h"
-#include "ggml-metal.h"
-
-#include <cstdio>
-#include <cstring>
-#include <cstdlib>
-
-int main(int argc, char ** argv) {
-    ggml_time_init();
-
-    if (argc != 2) {
-        fprintf(stderr, "Usage: %s llama.ggml\n", argv[0]);
-        return -1;
-    }
-
-    const char * fname_cgraph = argv[1];
-
-    // load the compute graph
-    struct ggml_context * ctx_data = NULL;
-    struct ggml_context * ctx_eval = NULL;
-
-    struct ggml_cgraph * gf = ggml_graph_import(fname_cgraph, &ctx_data, &ctx_eval);
-
-    // this allocates all Metal resources and memory buffers
-    auto * ctx_metal = ggml_metal_init(1);
-
-    const size_t max_size_data = ggml_get_max_tensor_size(ctx_data);
-    const size_t max_size_eval = ggml_get_max_tensor_size(ctx_eval);
-    ggml_metal_add_buffer(ctx_metal, "data", ggml_get_mem_buffer(ctx_data), ggml_get_mem_size(ctx_data), max_size_data);
-    ggml_metal_add_buffer(ctx_metal, "eval", ggml_get_mem_buffer(ctx_eval), ggml_get_mem_size(ctx_eval), max_size_eval);
-
-    // main
-    {
-        struct ggml_tensor * input = ggml_graph_get_tensor(gf, "embd");
-        *(int32_t *) input->data = 1; // BOS
-
-        ggml_metal_set_tensor(ctx_metal, input);
-
-        // warmup
-        ggml_metal_graph_compute(ctx_metal, gf);
-
-        const int n_iter = 16;
-
-        const int64_t t0 = ggml_time_us();
-
-        // the actual inference happens here
-        for (int i = 0; i < n_iter; ++i) {
-            ggml_metal_graph_compute(ctx_metal, gf);
-        }
-
-        const int64_t t1 = ggml_time_us();
-
-        printf("time: %.2f ms, %.2f ms/tok\n", (t1 - t0) / 1000.0, (t1 - t0) / 1000.0 / n_iter);
-    }
-
-    // debug output
-    {
-        struct ggml_tensor * logits = gf->nodes[gf->n_nodes - 1];
-        ggml_metal_get_tensor(ctx_metal, logits);
-
-        float * ptr = (float *) ggml_get_data(logits);
-
-        printf("logits: ");
-        for (int i = 0; i < 10; i++) {
-            printf("%8.4f ", ptr[i]);
-        }
-        printf("\n");
-        int imax = 0;
-        double sum = 0.0;
-        double vmax = -1e9;
-        for (int i = 0; i < 32000; i++) {
-            sum += (double) ptr[i];
-            if (ptr[i] > vmax) {
-                vmax = ptr[i];
-                imax = i;
-            }
-        }
-        printf("sum: %f, imax = %d, vmax = %f\n", sum, imax, vmax);
-    }
-
-    ggml_metal_free(ctx_metal);
-
-    ggml_free(ctx_data);
-    ggml_free(ctx_eval);
-
-    return 0;
-}
-

examples/passkey/CMakeLists.txt

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

examples/passkey/README.md

@@ -0,0 +1,12 @@
+# llama.cpp/example/passkey
+
+See the following PRs for more info:
+
+- https://github.com/ggerganov/llama.cpp/pull/3856
+- https://github.com/ggerganov/llama.cpp/pull/4810
+
+### Usage
+
+```bash
+make -j && ./passkey ./models/llama-7b-v2/ggml-model-f16.gguf 250
+```

examples/passkey/passkey.cpp

@@ -0,0 +1,296 @@
+#include "common.h"
+#include "llama.h"
+
+#include <cmath>
+#include <cstdio>
+#include <string>
+#include <vector>
+
+int main(int argc, char ** argv) {
+    gpt_params params;
+
+    if (argc == 1 || argv[1][0] == '-') {
+        printf("usage: %s MODEL_PATH N_JUNK N_GRP I_POS SEED\n" , argv[0]);
+        return 1 ;
+    }
+
+    int seed = -1;
+
+    int n_junk = 250; // number of times to repeat the junk text
+    int n_keep = 32;  // number of tokens in the prompt prefix
+    int n_grp  = 1;   // if more than 1 - perform LongLM SelfExtend
+    int i_pos  = -1;  // position of the passkey in the junk text
+
+    if (argc >= 2) {
+        params.model = argv[1];
+    }
+
+    if (argc >= 3) {
+        n_junk = std::stoi(argv[2]);
+    }
+
+    if (argc >= 4) {
+        n_grp = std::stoi(argv[3]);
+    }
+
+    if (argc >= 5) {
+        i_pos = std::stoi(argv[4]);
+    }
+
+    if (argc >= 6) {
+        seed = std::stoi(argv[5]);
+    }
+
+    if (seed == -1) {
+        seed = time(NULL);
+    }
+
+    srand(seed);
+
+    if (i_pos == -1) {
+        i_pos = rand() % n_junk;
+    }
+
+    const std::string prompt_prefix = "There is an important info hidden inside a lot of irrelevant text. Find it and memorize them. I will quiz you about the important information there.";
+    const std::string prompt_suffix = " What is the pass key? The pass key is";
+
+    // generate junk text
+    params.prompt = prompt_prefix;
+
+    const int passkey = rand() % 50000 + 1;
+
+    for (int i = 0; i < n_junk; i++) {
+        if (i % n_junk == i_pos) {
+            params.prompt += " The pass key is " + std::to_string(passkey) + ". Remember it. " + std::to_string(passkey) + " is the pass key.";
+        }
+
+        params.prompt += " The grass is green. The sky is blue. The sun is yellow. Here we go. There and back again.";
+    }
+
+    params.prompt += prompt_suffix;
+
+    // init LLM
+
+    llama_backend_init(params.numa);
+
+    // initialize the model
+
+    llama_model_params model_params = llama_model_default_params();
+
+    model_params.n_gpu_layers = 99; // offload all layers to the GPU
+
+    llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
+
+    if (model == NULL) {
+        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
+        return 1;
+    }
+
+    // initialize the context
+
+    llama_context_params ctx_params = llama_context_default_params();
+
+    ctx_params.seed    = seed;
+    ctx_params.n_ctx   = llama_n_ctx_train(model)*n_grp + n_keep;
+    ctx_params.n_batch = 512;
+    ctx_params.n_threads       = params.n_threads;
+    ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+
+    GGML_ASSERT(ctx_params.n_batch % n_grp == 0 && "n_batch must be divisible by n_grp");
+
+    llama_context * ctx = llama_new_context_with_model(model, ctx_params);
+
+    if (ctx == NULL) {
+        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+        return 1;
+    }
+
+    // tokenize the prompt
+    std::vector<llama_token> tokens_list;
+    tokens_list = ::llama_tokenize(ctx, params.prompt, true);
+
+    // tokenize the prefix and use it as a sink
+    const int n_tokens_prefix = ::llama_tokenize(ctx, prompt_prefix, true).size();
+
+    const int n_tokens_all = tokens_list.size();
+
+    // we leave a margin of 16 tokens for the generated text - it should contain just the passkey
+    const int n_predict = 16;
+
+    // total length of the sequences including the prompt
+    const int n_len = n_tokens_all + n_predict;
+
+    const int n_ctx       = llama_n_ctx(ctx) - n_keep;
+    const int n_kv_req    = llama_n_ctx(ctx);
+    const int n_batch     = ctx_params.n_batch;
+    const int n_batch_grp = ctx_params.n_batch/n_grp;
+
+    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d, n_grp = %d, n_batch = %d\n", __func__, n_len, n_ctx, n_kv_req, n_grp, n_batch);
+
+    // print the prompt token-by-token
+
+    LOG_TEE("\n");
+    LOG_TEE("prefix tokens: %d\n", n_tokens_prefix);
+    LOG_TEE("prompt tokens: %d\n", n_tokens_all);
+    //LOG_TEE("prompt: %s\n", params.prompt.c_str());
+
+    llama_batch batch = llama_batch_init(512, 0, 1);
+
+    int n_past = 0;
+
+    // fill the KV cache
+    for (int i = 0; i < n_ctx; i += n_batch) {
+        if (i > 0 && n_grp > 1) {
+            // if SelfExtend is enabled, we compress the position from the last batch by a factor of n_grp
+            const int ib = i/n_batch - 1;
+            const int bd = n_batch_grp*(n_grp - 1);
+
+            llama_kv_cache_seq_shift(ctx, 0, n_past - n_batch,         n_past,         ib*bd);
+            llama_kv_cache_seq_div  (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
+
+            n_past -= bd;
+        }
+
+        llama_batch_clear(batch);
+
+        for (int j = 0; j < n_batch && i + j < n_tokens_all; j++) {
+            llama_batch_add(batch, tokens_list[i + j], n_past++, { 0 }, false);
+        }
+
+        if (i + n_batch >= n_tokens_all) {
+            batch.logits[batch.n_tokens - 1] = true;
+        }
+
+        if (llama_decode(ctx, batch) != 0) {
+            LOG_TEE("%s: llama_decode() failed\n", __func__);
+            return 1;
+        }
+
+        LOG_TEE("%s: processed: [%6d, %6d)\n", __func__, i, std::min(i + n_batch, n_tokens_all));
+
+        if (i + n_batch >= n_tokens_all) {
+            break;
+        }
+    }
+
+    for (int i = n_ctx; i < n_tokens_all; i += n_batch) {
+        const int n_discard = n_batch;
+
+        LOG_TEE("%s: shifting KV cache with %d\n", __func__, n_discard);
+
+        llama_kv_cache_seq_rm   (ctx, 0, n_keep            , n_keep + n_discard);
+        llama_kv_cache_seq_shift(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+
+        n_past -= n_discard;
+
+        llama_batch_clear(batch);
+
+        for (int j = 0; j < n_batch && i + j < n_tokens_all; j++) {
+            llama_batch_add(batch, tokens_list[i + j], n_past++, { 0 }, false);
+        }
+
+        if (i + n_batch >= n_tokens_all) {
+            batch.logits[batch.n_tokens - 1] = true;
+        }
+
+        if (llama_decode(ctx, batch) != 0) {
+            LOG_TEE("%s: llama_decode() failed\n", __func__);
+            return 1;
+        }
+
+        LOG_TEE("%s: processed: [%6d, %6d)\n", __func__, i, std::min(i + n_batch, n_tokens_all));
+    }
+
+    {
+        const int n_discard = n_past - n_ctx + n_predict;
+
+        if (n_discard > 0) {
+            LOG_TEE("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard);
+
+            llama_kv_cache_seq_rm   (ctx, 0, n_keep            , n_keep + n_discard);
+            llama_kv_cache_seq_shift(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+
+            n_past -= n_discard;
+        }
+    }
+
+    LOG_TEE("\n");
+    LOG_TEE("%s: passkey = %d, inserted at position %d / %d (token pos: ~%d)\n", __func__, passkey, i_pos, n_junk, (i_pos * n_tokens_all) / n_junk);
+    LOG_TEE("\n");
+
+    // main loop
+
+    int n_cur    = n_tokens_all;
+    int n_decode = 0;
+
+    LOG_TEE("%s", prompt_suffix.c_str());
+    fflush(stdout);
+
+    const auto t_main_start = ggml_time_us();
+
+    while (n_cur <= n_len) {
+        // sample the next token
+        {
+            auto   n_vocab = llama_n_vocab(model);
+            auto * logits  = llama_get_logits_ith(ctx, batch.n_tokens - 1);
+
+            std::vector<llama_token_data> candidates;
+            candidates.reserve(n_vocab);
+
+            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+                candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
+            }
+
+            llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+
+            // sample the most likely token
+            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(model) || n_cur == n_len) {
+                LOG_TEE("\n");
+
+                break;
+            }
+
+            LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
+            fflush(stdout);
+
+            n_decode += 1;
+
+            // prepare the next batch
+            llama_batch_clear(batch);
+
+            // push this new token for next evaluation
+            llama_batch_add(batch, new_token_id, n_past++, { 0 }, true);
+        }
+
+        n_cur += 1;
+
+        // evaluate the current batch with the transformer model
+        if (llama_decode(ctx, batch)) {
+            fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
+            return 1;
+        }
+    }
+
+    LOG_TEE("\n");
+
+    const auto t_main_end = ggml_time_us();
+
+    LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
+            __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
+
+    llama_print_timings(ctx);
+
+    fprintf(stderr, "\n");
+
+    llama_batch_free(batch);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    return 0;
+}

examples/pydantic-models-to-grammar-examples.py

@@ -0,0 +1,136 @@
+# Function calling example using pydantic models.
+
+import json
+from enum import Enum
+from typing import Union, Optional
+
+import requests
+from pydantic import BaseModel, Field
+
+import importlib
+from pydantic_models_to_grammar import generate_gbnf_grammar_and_documentation
+
+# Function to get completion on the llama.cpp server with grammar.
+def create_completion(prompt, grammar):
+    headers = {"Content-Type": "application/json"}
+    data = {"prompt": prompt, "grammar": grammar}
+
+    response = requests.post("http://127.0.0.1:8080/completion", headers=headers, json=data)
+    data = response.json()
+
+    print(data["content"])
+    return data["content"]
+
+
+# A function for the agent to send a message to the user.
+class SendMessageToUser(BaseModel):
+    """
+    Send a message to the User.
+    """
+    chain_of_thought: str = Field(..., description="Your chain of thought while sending the message.")
+    message: str = Field(..., description="Message you want to send to the user.")
+
+    def run(self):
+        print(self.message)
+
+
+# Enum for the calculator function.
+class MathOperation(Enum):
+    ADD = "add"
+    SUBTRACT = "subtract"
+    MULTIPLY = "multiply"
+    DIVIDE = "divide"
+
+
+# Very simple calculator tool for the agent.
+class Calculator(BaseModel):
+    """
+    Perform a math operation on two numbers.
+    """
+    number_one: Union[int, float] = Field(..., description="First number.")
+    operation: MathOperation = Field(..., description="Math operation to perform.")
+    number_two: Union[int, float] = Field(..., description="Second number.")
+
+    def run(self):
+        if self.operation == MathOperation.ADD:
+            return self.number_one + self.number_two
+        elif self.operation == MathOperation.SUBTRACT:
+            return self.number_one - self.number_two
+        elif self.operation == MathOperation.MULTIPLY:
+            return self.number_one * self.number_two
+        elif self.operation == MathOperation.DIVIDE:
+            return self.number_one / self.number_two
+        else:
+            raise ValueError("Unknown operation.")
+
+
+# Here the grammar gets generated by passing the available function models to generate_gbnf_grammar_and_documentation function. This also generates a documentation usable by the LLM.
+# pydantic_model_list is the list of pydanitc models
+# outer_object_name is an optional name for an outer object around the actual model object. Like a "function" object with "function_parameters" which contains the actual model object. If None, no outer object will be generated
+# outer_object_content is the name of outer object content.
+# model_prefix is the optional prefix for models in the documentation. (Default="Output Model")
+# fields_prefix is the prefix for the model fields in the documentation. (Default="Output Fields")
+gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation(
+    pydantic_model_list=[SendMessageToUser, Calculator], outer_object_name="function",
+    outer_object_content="function_parameters", model_prefix="Function", fields_prefix="Parameters")
+
+print(gbnf_grammar)
+print(documentation)
+
+system_message = "You are an advanced AI, tasked to assist the user by calling functions in JSON format. The following are the available functions and their parameters and types:\n\n" + documentation
+
+user_message = "What is 42 * 42?"
+prompt = f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{user_message}<|im_end|>\n<|im_start|>assistant"
+
+text = create_completion(prompt=prompt, grammar=gbnf_grammar)
+# This should output something like this:
+# {
+#     "function": "calculator",
+#     "function_parameters": {
+#         "number_one": 42,
+#         "operation": "multiply",
+#         "number_two": 42
+#     }
+# }
+function_dictionary = json.loads(text)
+if function_dictionary["function"] == "calculator":
+    function_parameters = {**function_dictionary["function_parameters"]}
+
+    print(Calculator(**function_parameters).run())
+    # This should output: 1764
+
+
+# A example structured output based on pydantic models. The LLM will create an entry for a Book database out of an unstructured text.
+class Category(Enum):
+    """
+    The category of the book.
+    """
+    Fiction = "Fiction"
+    NonFiction = "Non-Fiction"
+
+
+class Book(BaseModel):
+    """
+    Represents an entry about a book.
+    """
+    title: str = Field(..., description="Title of the book.")
+    author: str = Field(..., description="Author of the book.")
+    published_year: Optional[int] = Field(..., description="Publishing year of the book.")
+    keywords: list[str] = Field(..., description="A list of keywords.")
+    category: Category = Field(..., description="Category of the book.")
+    summary: str = Field(..., description="Summary of the book.")
+
+
+# We need no additional parameters other than our list of pydantic models.
+gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation([Book])
+
+system_message = "You are an advanced AI, tasked to create a dataset entry in JSON for a Book. The following is the expected output model:\n\n" + documentation
+
+text = """The Feynman Lectures on Physics is a physics textbook based on some lectures by Richard Feynman, a Nobel laureate who has sometimes been called "The Great Explainer". The lectures were presented before undergraduate students at the California Institute of Technology (Caltech), during 1961–1963. The book's co-authors are Feynman, Robert B. Leighton, and Matthew Sands."""
+prompt = f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{text}<|im_end|>\n<|im_start|>assistant"
+
+text = create_completion(prompt=prompt, grammar=gbnf_grammar)
+
+json_data = json.loads(text)
+
+print(Book(**json_data))

examples/quantize/quantize.cpp

@@ -5,6 +5,10 @@
 #include <cstring>
 #include <vector>
 #include <string>
+#include <unordered_map>
+#include <fstream>
+#include <cmath>
+#include <algorithm>
 
 struct quant_option {
     std::string name;
@@ -17,7 +21,10 @@ 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", },
+    { "IQ2_XXS",LLAMA_FTYPE_MOSTLY_IQ2_XXS," 2.06 bpw quantization",            },
+    { "IQ2_XS", LLAMA_FTYPE_MOSTLY_IQ2_XS, " 2.31 bpw quantization",            },
     { "Q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K,   " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", },
+    { "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.16G, +9.0634 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", },
     { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.07G, +0.2496 ppl @ LLaMA-v1-7B", },
@@ -71,10 +78,14 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
 //
 [[noreturn]]
 static void usage(const char * executable) {
-    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
+    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] 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("  --imatrix file_name: use data in file_name as importance matrix for quant optimizations\n");
+    printf("  --include-weights tensor_name: use importance matrix for this/these tensor(s)\n");
+    printf("  --exclude-weights tensor_name: use importance matrix for this/these tensor(s)\n");
+    printf("Note: --include-weights and --exclude-weights cannot be used together\n");
     printf("\nAllowed quantization types:\n");
     for (auto & it : QUANT_OPTIONS) {
         if (it.name != "COPY") {
@@ -82,11 +93,93 @@ static void usage(const char * executable) {
         } else {
             printf("          ");
         }
-        printf("%-6s : %s\n", it.name.c_str(), it.desc.c_str());
+        printf("%-7s : %s\n", it.name.c_str(), it.desc.c_str());
     }
     exit(1);
 }
 
+static void load_imatrix(const std::string& imatrix_file, std::unordered_map<std::string, std::vector<float>>& imatrix_data) {
+    std::ifstream in(imatrix_file.c_str(), std::ios::binary);
+    if (!in) {
+        printf("%s: failed to open %s\n",__func__,imatrix_file.c_str());
+        return;
+    }
+    int n_entries;
+    in.read((char*)&n_entries, sizeof(n_entries));
+    if (in.fail() || n_entries < 1) {
+        printf("%s: no data in file %s\n", __func__, imatrix_file.c_str());
+        return;
+    }
+    for (int i = 0; i < n_entries; ++i) {
+        int len; in.read((char *)&len, sizeof(len));
+        std::vector<char> name_as_vec(len+1);
+        in.read((char *)name_as_vec.data(), len);
+        if (in.fail()) {
+            printf("%s: failed reading name for entry %d from %s\n",__func__,i+1,imatrix_file.c_str());
+            return;
+        }
+        name_as_vec[len] = 0;
+        std::string name{name_as_vec.data()};
+        auto& e = imatrix_data[std::move(name)];
+        int ncall;
+        in.read((char*)&ncall, sizeof(ncall));
+        int nval;
+        in.read((char *)&nval, sizeof(nval));
+        if (in.fail() || nval < 1) {
+            printf("%s: failed reading number of values for entry %d\n",__func__,i);
+            imatrix_data = {};
+            return;
+        }
+        e.resize(nval);
+        in.read((char*)e.data(), nval*sizeof(float));
+        if (in.fail()) {
+            printf("%s: failed reading data for entry %d\n",__func__,i);
+            imatrix_data = {};
+            return;
+        }
+        if (ncall > 0) {
+            for (auto& v : e) v /= ncall;
+        }
+    }
+    printf("%s: loaded %d importance matrix entries from %s\n",__func__,int(imatrix_data.size()),imatrix_file.c_str());
+}
+
+static void prepare_imatrix(const std::string& imatrix_file,
+        const std::vector<std::string>& included_weights,
+        const std::vector<std::string>& excluded_weights,
+        std::unordered_map<std::string, std::vector<float>>& imatrix_data) {
+    if (!imatrix_file.empty()) {
+        load_imatrix(imatrix_file, imatrix_data);
+    }
+    if (imatrix_data.empty()) {
+        return;
+    }
+    if (!excluded_weights.empty()) {
+        for (auto& name : excluded_weights) {
+            for (auto it = imatrix_data.begin(); it != imatrix_data.end(); ) {
+                auto pos = it->first.find(name);
+                if (pos != std::string::npos) it = imatrix_data.erase(it);
+                else ++it;
+            }
+        }
+    }
+    if (!included_weights.empty()) {
+        std::unordered_map<std::string, std::vector<float>> tmp;
+        for (auto& name : included_weights) {
+            for (auto& e : imatrix_data) {
+                auto pos = e.first.find(name);
+                if (pos != std::string::npos) {
+                    tmp.emplace(std::move(e));
+                }
+            }
+        }
+        imatrix_data = std::move(tmp);
+    }
+    if (!imatrix_data.empty()) {
+        printf("%s: have %d importance matrix entries\n", __func__, int(imatrix_data.size()));
+    }
+}
+
 int main(int argc, char ** argv) {
     if (argc < 3) {
         usage(argv[0]);
@@ -95,6 +188,8 @@ int main(int argc, char ** argv) {
     llama_model_quantize_params params = llama_model_quantize_default_params();
 
     int arg_idx = 1;
+    std::string imatrix_file;
+    std::vector<std::string> included_weights, excluded_weights;
 
     for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
         if (strcmp(argv[arg_idx], "--leave-output-tensor") == 0) {
@@ -103,14 +198,42 @@ int main(int argc, char ** argv) {
             params.allow_requantize = true;
         } else if (strcmp(argv[arg_idx], "--pure") == 0) {
             params.pure = true;
+        } else if (strcmp(argv[arg_idx], "--imatrix") == 0) {
+            if (arg_idx < argc-1) {
+                imatrix_file = argv[++arg_idx];
+            } else {
+                usage(argv[0]);
+            }
+        } else if (strcmp(argv[arg_idx], "--include-weights") == 0) {
+            if (arg_idx < argc-1) {
+                included_weights.push_back(argv[++arg_idx]);
+            } else {
+                usage(argv[0]);
+            }
+        } else if (strcmp(argv[arg_idx], "--exclude-weights") == 0) {
+            if (arg_idx < argc-1) {
+                excluded_weights.push_back(argv[++arg_idx]);
+            } else {
+                usage(argv[0]);
+            }
         } else {
             usage(argv[0]);
         }
     }
 
     if (argc - arg_idx < 2) {
+        printf("%s: bad arguments\n", argv[0]);
         usage(argv[0]);
     }
+    if (!included_weights.empty() && !excluded_weights.empty()) {
+        usage(argv[0]);
+    }
+
+    std::unordered_map<std::string, std::vector<float>> imatrix_data;
+    prepare_imatrix(imatrix_file, included_weights, excluded_weights, imatrix_data);
+    if (!imatrix_data.empty()) {
+        params.imatrix = &imatrix_data;
+    }
 
     llama_backend_init(false);
 
@@ -162,6 +285,13 @@ int main(int argc, char ** argv) {
         }
     }
 
+    if ((params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) && imatrix_data.empty()) {
+        fprintf(stderr, "\n===============================================================================================\n");
+        fprintf(stderr, "Please do not use IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");
+        fprintf(stderr, "===============================================================================================\n\n\n");
+        return 1;
+    }
+
     print_build_info();
 
     fprintf(stderr, "%s: quantizing '%s' to '%s' as %s", __func__, fname_inp.c_str(), fname_out.c_str(), ftype_str.c_str());

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

@@ -45,13 +45,13 @@ int main(int argc, char ** argv) {
     // save state (rng, logits, embedding and kv_cache) to file
     {
         std::vector<uint8_t> state_mem(llama_get_state_size(ctx));
+        const size_t written = llama_copy_state_data(ctx, state_mem.data());
 
-        {
-            FILE *fp_write = fopen("dump_state.bin", "wb");
-            llama_copy_state_data(ctx, state_mem.data()); // could also copy directly to memory mapped file
-            fwrite(state_mem.data(), 1, state_mem.size(), fp_write);
-            fclose(fp_write);
-        }
+        FILE *fp_write = fopen("dump_state.bin", "wb");
+        fwrite(state_mem.data(), 1, written, fp_write);
+        fclose(fp_write);
+
+        fprintf(stderr, "%s : serialized state into %zd out of a maximum of %zd bytes\n", __func__, written, state_mem.size());
     }
 
     // save state (last tokens)
@@ -100,18 +100,17 @@ int main(int argc, char ** argv) {
         std::vector<uint8_t> state_mem(llama_get_state_size(ctx2));
 
         FILE * fp_read = fopen("dump_state.bin", "rb");
+        const size_t read = fread(state_mem.data(), 1, state_mem.size(), fp_read);
+        fclose(fp_read);
 
-        const size_t ret = fread(state_mem.data(), 1, state_mem.size(), fp_read);
-        if (ret != state_mem.size()) {
+        if (read != llama_set_state_data(ctx2, state_mem.data())) {
             fprintf(stderr, "\n%s : failed to read state\n", __func__);
             llama_free(ctx2);
             llama_free_model(model);
             return 1;
         }
 
-        llama_set_state_data(ctx2, state_mem.data());
-
-        fclose(fp_read);
+        fprintf(stderr, "%s : deserialized state from %zd out of a maximum of %zd bytes\n", __func__, read, state_mem.size());
     }
 
     // restore state (last tokens)

examples/server/README.md

@@ -23,6 +23,8 @@ Command line options:
 -   `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`.
 -   `--port`: Set the port to listen. Default: `8080`.
 -   `--path`: path from which to serve static files (default examples/server/public)
+-   `--api-key`: Set an api key for request authorization. By default the server responds to every request. With an api key set, the requests must have the Authorization header set with the api key as Bearer token. May be used multiple times to enable multiple valid keys.
+-   `--api-key-file`: path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access. May be used in conjunction with `--api-key`'s.
 -   `--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)
@@ -109,6 +111,10 @@ node index.js
 ```
 
 ## API Endpoints
+- **GET** `/health`: Returns the current state of the server:
+    - `{"status": "loading model"}` if the model is still being loaded.
+    - `{"status": "error"}` if the model failed to load.
+    - `{"status": "ok"}` if the model is successfully loaded and the server is ready for further requests mentioned below.
 
 -   **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
 
@@ -174,35 +180,44 @@ node index.js
 
     `system_prompt`: Change the system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
 
-    *Result JSON:*
+### Result JSON:
 
-    Note: When using streaming mode (`stream`) only `content` and `stop` will be returned until end of completion.
+* Note: When using streaming mode (`stream`) only `content` and `stop` will be returned until end of completion.
 
-    `content`: Completion result as a string (excluding `stopping_word` if any). In case of streaming mode, will contain the next token as a string.
 
-    `stop`: Boolean for use with `stream` to check whether the generation has stopped (Note: This is not related to stopping words array `stop` from input options)
+- `completion_probabilities`: An array of token probabilities for each completion. The array's length is `n_predict`. Each item in the array has the following structure:
 
-    `generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model`
+```
+{
+  "content": "<the token selected by the model>",
+  "probs": [
+    {
+      "prob": float,
+      "tok_str": "<most likely token>"
+    },
+    {
+      "prob": float,
+      "tok_str": "<second most likely tonen>"
+    },
+    ...
+  ]
+},
+```
+Notice that each `probs` is an array of length `n_probs`.
 
-    `model`: The path to the model loaded with `-m`
-
-    `prompt`: The provided `prompt`
-
-    `stopped_eos`: Indicating whether the completion has stopped because it encountered the EOS token
-
-    `stopped_limit`: Indicating whether the completion stopped because `n_predict` tokens were generated before stop words or EOS was encountered
-
-    `stopped_word`: Indicating whether the completion stopped due to encountering a stopping word from `stop` JSON array provided
-
-    `stopping_word`: The stopping word encountered which stopped the generation (or "" if not stopped due to a stopping word)
-
-    `timings`: Hash of timing information about the completion such as the number of tokens `predicted_per_second`
-
-    `tokens_cached`: Number of tokens from the prompt which could be re-used from previous completion (`n_past`)
-
-    `tokens_evaluated`: Number of tokens evaluated in total from the prompt
-
-    `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`)
+- `content`: Completion result as a string (excluding `stopping_word` if any). In case of streaming mode, will contain the next token as a string.
+- `stop`: Boolean for use with `stream` to check whether the generation has stopped (Note: This is not related to stopping words array `stop` from input options)
+- `generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model`
+- `model`: The path to the model loaded with `-m`
+- `prompt`: The provided `prompt`
+- `stopped_eos`: Indicating whether the completion has stopped because it encountered the EOS token
+- `stopped_limit`: Indicating whether the completion stopped because `n_predict` tokens were generated before stop words or EOS was encountered
+- `stopped_word`: Indicating whether the completion stopped due to encountering a stopping word from `stop` JSON array provided
+- `stopping_word`: The stopping word encountered which stopped the generation (or "" if not stopped due to a stopping word)
+- `timings`: Hash of timing information about the completion such as the number of tokens `predicted_per_second`
+- `tokens_cached`: Number of tokens from the prompt which could be re-used from previous completion (`n_past`)
+- `tokens_evaluated`: Number of tokens evaluated in total from the prompt
+- `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`)
 
 -   **POST** `/tokenize`: Tokenize a given text.
 

examples/server/server.cpp

@@ -26,6 +26,7 @@
 #include <mutex>
 #include <chrono>
 #include <condition_variable>
+#include <atomic>
 
 #ifndef SERVER_VERBOSE
 #define SERVER_VERBOSE 1
@@ -38,7 +39,7 @@ using json = nlohmann::json;
 struct server_params
 {
     std::string hostname = "127.0.0.1";
-    std::string api_key;
+    std::vector<std::string> api_keys;
     std::string public_path = "examples/server/public";
     int32_t port = 8080;
     int32_t read_timeout = 600;
@@ -146,9 +147,15 @@ static std::vector<uint8_t> base64_decode(const std::string & encoded_string)
 // parallel
 //
 
+enum server_state {
+    SERVER_STATE_LOADING_MODEL,  // Server is starting up, model not fully loaded yet
+    SERVER_STATE_READY,          // Server is ready and model is loaded
+    SERVER_STATE_ERROR           // An error occurred, load_model failed
+};
+
 enum task_type {
-    COMPLETION_TASK,
-    CANCEL_TASK
+    TASK_TYPE_COMPLETION,
+    TASK_TYPE_CANCEL,
 };
 
 struct task_server {
@@ -447,8 +454,14 @@ struct llama_client_slot
     }
 
     bool has_budget(gpt_params &global_params) {
+        if (params.n_predict == -1 && global_params.n_predict == -1)
+        {
+            return true; // limitless
+        }
+
         n_remaining = -1;
-        if(params.n_predict != -1)
+
+        if (params.n_predict != -1)
         {
             n_remaining = params.n_predict - n_decoded;
         }
@@ -456,7 +469,8 @@ struct llama_client_slot
         {
             n_remaining = global_params.n_predict - n_decoded;
         }
-        return n_remaining > 0 || n_remaining == -1; // no budget || limitless
+
+        return n_remaining > 0; // no budget
     }
 
     bool available() const {
@@ -1102,7 +1116,7 @@ struct llama_server_context
         }
 
         // check the limits
-        if (slot.n_decoded > 2 && slot.has_next_token && !slot.has_budget(params))
+        if (slot.n_decoded > 0 && slot.has_next_token && !slot.has_budget(params))
         {
             slot.stopped_limit = true;
             slot.has_next_token = false;
@@ -1166,8 +1180,9 @@ struct llama_server_context
         return slot.images.size() > 0;
     }
 
-    void send_error(task_server& task, std::string error)
+    void send_error(task_server& task, const std::string &error)
     {
+        LOG_TEE("task %i - error: %s\n", task.id, error.c_str());
         std::unique_lock<std::mutex> lock(mutex_results);
         task_result res;
         res.id = task.id;
@@ -1336,14 +1351,17 @@ struct llama_server_context
             res.result_json["model"] = slot.oaicompat_model;
         }
 
+        queue_results.push_back(res);
+        condition_results.notify_all();
+
+        // done with results, unlock
+        lock.unlock();
+
         // parent multitask, if any, needs to be updated
         if (slot.multitask_id != -1)
         {
             update_multi_task(slot.multitask_id, slot.task_id, res);
         }
-
-        queue_results.push_back(res);
-        condition_results.notify_all();
     }
 
     void send_embedding(llama_client_slot &slot)
@@ -1388,11 +1406,11 @@ struct llama_server_context
         task.data = std::move(data);
         task.infill_mode = infill;
         task.embedding_mode = embedding;
-        task.type = COMPLETION_TASK;
+        task.type = TASK_TYPE_COMPLETION;
         task.multitask_id = multitask_id;
 
         // when a completion task's prompt array is not a singleton, we split it into multiple requests
-        if (task.data.at("prompt").size() > 1)
+        if (task.data.count("prompt") && task.data.at("prompt").size() > 1)
         {
             lock.unlock(); // entering new func scope
             return split_multiprompt_task(task);
@@ -1510,7 +1528,7 @@ struct llama_server_context
         std::unique_lock<std::mutex> lock(mutex_tasks);
         task_server task;
         task.id = id_gen++;
-        task.type = CANCEL_TASK;
+        task.type = TASK_TYPE_CANCEL;
         task.target_id = task_id;
         queue_tasks.push_back(task);
         condition_tasks.notify_one();
@@ -1546,26 +1564,36 @@ struct llama_server_context
             queue_tasks.erase(queue_tasks.begin());
             switch (task.type)
             {
-                case COMPLETION_TASK: {
+                case TASK_TYPE_COMPLETION: {
                     llama_client_slot *slot = get_slot(json_value(task.data, "slot_id", -1));
                     if (slot == nullptr)
                     {
                         LOG_TEE("slot unavailable\n");
                         // send error result
                         send_error(task, "slot unavailable");
-                        return;
+                        break;
                     }
 
                     if (task.data.contains("system_prompt"))
                     {
+                        if (!all_slots_are_idle) {
+                            send_error(task, "system prompt can only be updated when all slots are idle");
+                            break;
+                        }
                         process_system_prompt_data(task.data["system_prompt"]);
+
+                        // reset cache_tokens for all slots
+                        for (llama_client_slot &slot : slots)
+                        {
+                            slot.cache_tokens.clear();
+                        }
                     }
 
                     slot->reset();
 
-                    slot->infill = task.infill_mode;
-                    slot->embedding = task.embedding_mode;
-                    slot->task_id = task.id;
+                    slot->infill       = task.infill_mode;
+                    slot->embedding    = task.embedding_mode;
+                    slot->task_id      = task.id;
                     slot->multitask_id = task.multitask_id;
 
                     if (!launch_slot_with_data(slot, task.data))
@@ -1575,7 +1603,7 @@ struct llama_server_context
                         break;
                     }
                 } break;
-                case CANCEL_TASK: { // release slot linked with the task id
+                case TASK_TYPE_CANCEL: { // release slot linked with the task id
                     for (auto & slot : slots)
                     {
                         if (slot.task_id == task.target_id)
@@ -1589,6 +1617,7 @@ struct llama_server_context
         }
 
         // remove finished multitasks from the queue of multitasks, and add the corresponding result to the result queue
+        std::vector<task_result> agg_results;
         auto queue_iterator = queue_multitasks.begin();
         while (queue_iterator != queue_multitasks.end())
         {
@@ -1609,8 +1638,9 @@ struct llama_server_context
                 }
                 aggregate_result.result_json = json{ "results", result_jsons };
 
-                std::lock_guard<std::mutex> lock(mutex_results);
-                queue_results.push_back(aggregate_result);
+
+                agg_results.push_back(aggregate_result);
+
                 condition_results.notify_all();
 
                 queue_iterator = queue_multitasks.erase(queue_iterator);
@@ -1620,14 +1650,20 @@ struct llama_server_context
                 ++queue_iterator;
             }
         }
+
+        // done with tasks, unlock
+        lock.unlock();
+
+        // copy aggregate results of complete multi-tasks to the results queue
+        std::lock_guard<std::mutex> lock_results(mutex_results);
+        queue_results.insert(queue_results.end(), agg_results.begin(), agg_results.end());
     }
 
     bool update_slots() {
         // attend tasks
         process_tasks();
 
-        // update the system prompt wait until all slots are idle state
-        if (system_need_update && all_slots_are_idle)
+        if (system_need_update)
         {
             LOG_TEE("updating system prompt\n");
             update_system_prompt();
@@ -1703,7 +1739,6 @@ struct llama_server_context
 
             llama_batch_add(batch, slot.sampled, system_tokens.size() + slot.n_past, { slot.id }, true);
 
-            slot.n_decoded += 1;
             slot.n_past += 1;
         }
 
@@ -1718,7 +1753,8 @@ struct llama_server_context
                 const bool has_prompt = slot.prompt.is_array() || (slot.prompt.is_string() && !slot.prompt.get<std::string>().empty()) || !slot.images.empty();
 
                 // empty prompt passed -> release the slot and send empty response
-                if (slot.state == IDLE && slot.command == LOAD_PROMPT && !has_prompt)
+                // note: infill mode allows empty prompt
+                if (slot.state == IDLE && slot.command == LOAD_PROMPT && !has_prompt && !slot.infill)
                 {
                     slot.release();
                     slot.print_timings();
@@ -1821,7 +1857,7 @@ struct llama_server_context
 
                     slot.cache_tokens = prompt_tokens;
 
-                    if (slot.n_past == slot.num_prompt_tokens)
+                    if (slot.n_past == slot.num_prompt_tokens && slot.n_past > 0)
                     {
                         // we have to evaluate at least 1 token to generate logits.
                         LOG_TEE("slot %d : we have to evaluate at least 1 token to generate logits\n", slot.id);
@@ -1921,6 +1957,7 @@ struct llama_server_context
 
                 llama_sampling_accept(slot.ctx_sampling, ctx, id, true);
 
+                slot.n_decoded += 1;
                 if (slot.n_decoded == 1)
                 {
                     slot.t_start_genereration = ggml_time_us();
@@ -1990,12 +2027,15 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
 #ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
     printf("  -ngl N, --n-gpu-layers N\n");
     printf("                        number of layers to store in VRAM\n");
+    printf("  -sm SPLIT_MODE, --split-mode SPLIT_MODE\n");
+    printf("                        how to split the model across multiple GPUs, one of:\n");
+    printf("                          - none: use one GPU only\n");
+    printf("                          - layer (default): split layers and KV across GPUs\n");
+    printf("                          - row: split rows across GPUs\n");
     printf("  -ts SPLIT --tensor-split SPLIT\n");
-    printf("                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    printf("  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
-    printf("  -nommq, --no-mul-mat-q\n");
-    printf("                        use cuBLAS instead of custom mul_mat_q CUDA kernels.\n");
-    printf("                        Not recommended since this is both slower and uses more VRAM.\n");
+    printf("                        fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1\n");
+    printf("  -mg i, --main-gpu i   the GPU to use for the model (with split-mode = none),\n");
+    printf("                        or for intermediate results and KV (with split-mode = row)\n");
 #endif
     printf("  -m FNAME, --model FNAME\n");
     printf("                        model path (default: %s)\n", params.model.c_str());
@@ -2007,6 +2047,7 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("  --port PORT           port to listen (default  (default: %d)\n", sparams.port);
     printf("  --path PUBLIC_PATH    path from which to serve static files (default %s)\n", sparams.public_path.c_str());
     printf("  --api-key API_KEY     optional api key to enhance server security. If set, requests must include this key for access.\n");
+    printf("  --api-key-file FNAME  path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access.\n");
     printf("  -to N, --timeout N    server read/write timeout in seconds (default: %d)\n", sparams.read_timeout);
     printf("  --embedding           enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled");
     printf("  -np N, --parallel N   number of slots for process requests (default: %d)\n", params.n_parallel);
@@ -2067,7 +2108,28 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
                 invalid_param = true;
                 break;
             }
-            sparams.api_key = argv[i];
+            sparams.api_keys.push_back(argv[i]);
+        }
+        else if (arg == "--api-key-file")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            std::ifstream key_file(argv[i]);
+            if (!key_file) {
+                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+                invalid_param = true;
+                break;
+            }
+            std::string key;
+            while (std::getline(key_file, key)) {
+               if (key.size() > 0) {
+                   sparams.api_keys.push_back(key);
+               }
+            }
+            key_file.close();
         }
         else if (arg == "--timeout" || arg == "-to")
         {
@@ -2216,6 +2278,33 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
                         "See main README.md for information on enabling GPU BLAS support",
                         {{"n_gpu_layers", params.n_gpu_layers}});
 #endif
+        }
+        else if (arg == "--split-mode" || arg == "-sm")
+        {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            std::string arg_next = argv[i];
+            if (arg_next == "none")
+            {
+                params.split_mode = LLAMA_SPLIT_NONE;
+            }
+            else if (arg_next == "layer")
+            {
+                params.split_mode = LLAMA_SPLIT_LAYER;
+            }
+            else if (arg_next == "row")
+            {
+                params.split_mode = LLAMA_SPLIT_ROW;
+            }
+            else {
+                invalid_param = true;
+                break;
+            }
+#ifndef GGML_USE_CUBLAS
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting the split mode has no effect.\n");
+#endif // GGML_USE_CUBLAS
         }
         else if (arg == "--tensor-split" || arg == "-ts")
         {
@@ -2446,7 +2535,6 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
     }
 }
 
-
 static std::string random_string()
 {
     static const std::string str("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz");
@@ -2502,7 +2590,7 @@ json oaicompat_completion_params_parse(
     //
     // https://platform.openai.com/docs/api-reference/chat/create
     llama_sampling_params default_sparams;
-    llama_params["model"]             = json_value(body, "model", std::string("uknown"));
+    llama_params["model"]             = json_value(body, "model", std::string("unknown"));
     llama_params["prompt"]            = format_chatml(body["messages"]); // OpenAI 'messages' to llama.cpp 'prompt'
     llama_params["cache_prompt"]      = json_value(body, "cache_prompt", false);
     llama_params["temperature"]       = json_value(body, "temperature", 0.0);
@@ -2574,8 +2662,8 @@ static json format_final_response_oaicompat(const json &request, const task_resu
             {"object", streaming ? "chat.completion.chunk" : "chat.completion"},
             {"usage",
                 json{{"completion_tokens", num_tokens_predicted},
-                    {"prompt_tokens", num_prompt_tokens},
-                    {"total_tokens", num_tokens_predicted + num_prompt_tokens}}},
+                     {"prompt_tokens",     num_prompt_tokens},
+                     {"total_tokens",      num_tokens_predicted + num_prompt_tokens}}},
             {"id", gen_chatcmplid()}};
 
     if (server_verbose) {
@@ -2783,20 +2871,131 @@ int main(int argc, char **argv)
                                 {"system_info", llama_print_system_info()},
                             });
 
-    // load the model
-    if (!llama.load_model(params))
+    httplib::Server svr;
+
+    std::atomic<server_state> state{SERVER_STATE_LOADING_MODEL};
+
+    svr.set_default_headers({{"Server", "llama.cpp"}});
+
+    // CORS preflight
+    svr.Options(R"(.*)", [](const httplib::Request &req, httplib::Response &res) {
+        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
+        res.set_header("Access-Control-Allow-Credentials", "true");
+        res.set_header("Access-Control-Allow-Methods", "POST");
+        res.set_header("Access-Control-Allow-Headers", "*");
+    });
+
+    svr.Get("/health", [&](const httplib::Request&, httplib::Response& res) {
+        server_state current_state = state.load();
+        switch(current_state) {
+            case SERVER_STATE_READY:
+                res.set_content(R"({"status": "ok"})", "application/json");
+                res.status = 200; // HTTP OK
+                break;
+            case SERVER_STATE_LOADING_MODEL:
+                res.set_content(R"({"status": "loading model"})", "application/json");
+                res.status = 503; // HTTP Service Unavailable
+                break;
+            case SERVER_STATE_ERROR:
+                res.set_content(R"({"status": "error", "error": "Model failed to load"})", "application/json");
+                res.status = 500; // HTTP Internal Server Error
+                break;
+        }
+    });
+
+    svr.set_logger(log_server_request);
+
+    svr.set_exception_handler([](const httplib::Request &, httplib::Response &res, std::exception_ptr ep)
+            {
+                const char fmt[] = "500 Internal Server Error\n%s";
+                char buf[BUFSIZ];
+                try
+                {
+                    std::rethrow_exception(std::move(ep));
+                }
+                catch (std::exception &e)
+                {
+                    snprintf(buf, sizeof(buf), fmt, e.what());
+                }
+                catch (...)
+                {
+                    snprintf(buf, sizeof(buf), fmt, "Unknown Exception");
+                }
+                res.set_content(buf, "text/plain; charset=utf-8");
+                res.status = 500;
+            });
+
+    svr.set_error_handler([](const httplib::Request &, httplib::Response &res)
+            {
+                if (res.status == 401)
+                {
+                    res.set_content("Unauthorized", "text/plain; charset=utf-8");
+                }
+                if (res.status == 400)
+                {
+                    res.set_content("Invalid request", "text/plain; charset=utf-8");
+                }
+                else if (res.status == 404)
+                {
+                    res.set_content("File Not Found", "text/plain; charset=utf-8");
+                    res.status = 404;
+                }
+            });
+
+    // set timeouts and change hostname and port
+    svr.set_read_timeout (sparams.read_timeout);
+    svr.set_write_timeout(sparams.write_timeout);
+
+    if (!svr.bind_to_port(sparams.hostname, sparams.port))
     {
+        fprintf(stderr, "\ncouldn't bind to server socket: hostname=%s port=%d\n\n", sparams.hostname.c_str(), sparams.port);
         return 1;
     }
 
-    llama.initialize();
+    // Set the base directory for serving static files
+    svr.set_base_dir(sparams.public_path);
 
-    httplib::Server svr;
+    // to make it ctrl+clickable:
+    LOG_TEE("\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);
+
+    std::unordered_map<std::string, std::string> log_data;
+    log_data["hostname"] = sparams.hostname;
+    log_data["port"] = std::to_string(sparams.port);
+
+    if (sparams.api_keys.size() == 1) {
+        log_data["api_key"] = "api_key: ****" + sparams.api_keys[0].substr(sparams.api_keys[0].length() - 4);
+    } else if (sparams.api_keys.size() > 1) {
+        log_data["api_key"] = "api_key: " + std::to_string(sparams.api_keys.size()) + " keys loaded";
+    }
+
+    LOG_INFO("HTTP server listening", log_data);
+    // run the HTTP server in a thread - see comment below
+    std::thread t([&]()
+            {
+                if (!svr.listen_after_bind())
+                {
+                    state.store(SERVER_STATE_ERROR);
+                    return 1;
+                }
+
+                return 0;
+            });
+
+    // load the model
+    if (!llama.load_model(params))
+    {
+        state.store(SERVER_STATE_ERROR);
+        return 1;
+    } else {
+        llama.initialize();
+        state.store(SERVER_STATE_READY);
+        LOG_INFO("model loaded", {});
+    }
 
     // Middleware for API key validation
     auto validate_api_key = [&sparams](const httplib::Request &req, httplib::Response &res) -> bool {
         // If API key is not set, skip validation
-        if (sparams.api_key.empty()) {
+        if (sparams.api_keys.empty()) {
             return true;
         }
 
@@ -2805,7 +3004,7 @@ int main(int argc, char **argv)
         std::string prefix = "Bearer ";
         if (auth_header.substr(0, prefix.size()) == prefix) {
             std::string received_api_key = auth_header.substr(prefix.size());
-            if (received_api_key == sparams.api_key) {
+            if (std::find(sparams.api_keys.begin(), sparams.api_keys.end(), received_api_key) != sparams.api_keys.end()) {
                 return true; // API key is valid
             }
         }
@@ -2819,10 +3018,6 @@ int main(int argc, char **argv)
         return false;
     };
 
-    svr.set_default_headers({{"Server", "llama.cpp"},
-                             {"Access-Control-Allow-Origin", "*"},
-                             {"Access-Control-Allow-Headers", "content-type"}});
-
     // this is only called if no index.html is found in the public --path
     svr.Get("/", [](const httplib::Request &, httplib::Response &res)
             {
@@ -2851,9 +3046,9 @@ int main(int argc, char **argv)
                 return false;
             });
 
-    svr.Get("/props", [&llama](const httplib::Request & /*req*/, httplib::Response &res)
+    svr.Get("/props", [&llama](const httplib::Request & req, httplib::Response &res)
             {
-                res.set_header("Access-Control-Allow-Origin", "*");
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 json data = {
                     { "user_name",      llama.name_user.c_str() },
                     { "assistant_name", llama.name_assistant.c_str() }
@@ -2863,6 +3058,7 @@ int main(int argc, char **argv)
 
     svr.Post("/completion", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res)
             {
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 if (!validate_api_key(req, res)) {
                     return;
                 }
@@ -2930,10 +3126,9 @@ int main(int argc, char **argv)
                 }
             });
 
-
-
-    svr.Get("/v1/models", [&params](const httplib::Request&, httplib::Response& res)
+    svr.Get("/v1/models", [&params](const httplib::Request& req, httplib::Response& res)
             {
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 std::time_t t = std::time(0);
 
                 json models = {
@@ -2951,9 +3146,11 @@ int main(int argc, char **argv)
                 res.set_content(models.dump(), "application/json; charset=utf-8");
             });
 
+
     // TODO: add mount point without "/v1" prefix -- how?
     svr.Post("/v1/chat/completions", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res)
             {
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 if (!validate_api_key(req, res)) {
                     return;
                 }
@@ -3027,6 +3224,7 @@ int main(int argc, char **argv)
 
     svr.Post("/infill", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res)
             {
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 if (!validate_api_key(req, res)) {
                     return;
                 }
@@ -3099,6 +3297,7 @@ int main(int argc, char **argv)
 
     svr.Post("/tokenize", [&llama](const httplib::Request &req, httplib::Response &res)
             {
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 const json body = json::parse(req.body);
                 std::vector<llama_token> tokens;
                 if (body.count("content") != 0)
@@ -3111,6 +3310,7 @@ int main(int argc, char **argv)
 
     svr.Post("/detokenize", [&llama](const httplib::Request &req, httplib::Response &res)
             {
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 const json body = json::parse(req.body);
                 std::string content;
                 if (body.count("tokens") != 0)
@@ -3125,6 +3325,7 @@ int main(int argc, char **argv)
 
     svr.Post("/embedding", [&llama](const httplib::Request &req, httplib::Response &res)
             {
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 const json body = json::parse(req.body);
                 json prompt;
                 if (body.count("content") != 0)
@@ -3150,81 +3351,6 @@ int main(int argc, char **argv)
                 return res.set_content(result.result_json.dump(), "application/json; charset=utf-8");
             });
 
-    svr.set_logger(log_server_request);
-
-    svr.set_exception_handler([](const httplib::Request &, httplib::Response &res, std::exception_ptr ep)
-            {
-                const char fmt[] = "500 Internal Server Error\n%s";
-                char buf[BUFSIZ];
-                try
-                {
-                    std::rethrow_exception(std::move(ep));
-                }
-                catch (std::exception &e)
-                {
-                    snprintf(buf, sizeof(buf), fmt, e.what());
-                }
-                catch (...)
-                {
-                    snprintf(buf, sizeof(buf), fmt, "Unknown Exception");
-                }
-                res.set_content(buf, "text/plain; charset=utf-8");
-                res.status = 500;
-            });
-
-    svr.set_error_handler([](const httplib::Request &, httplib::Response &res)
-            {
-                if (res.status == 401)
-                {
-                    res.set_content("Unauthorized", "text/plain; charset=utf-8");
-                }
-                if (res.status == 400)
-                {
-                    res.set_content("Invalid request", "text/plain; charset=utf-8");
-                }
-                else if (res.status == 404)
-                {
-                    res.set_content("File Not Found", "text/plain; charset=utf-8");
-                    res.status = 404;
-                }
-            });
-
-    // set timeouts and change hostname and port
-    svr.set_read_timeout (sparams.read_timeout);
-    svr.set_write_timeout(sparams.write_timeout);
-
-    if (!svr.bind_to_port(sparams.hostname, sparams.port))
-    {
-        fprintf(stderr, "\ncouldn't bind to server socket: hostname=%s port=%d\n\n", sparams.hostname.c_str(), sparams.port);
-        return 1;
-    }
-
-    // Set the base directory for serving static files
-    svr.set_base_dir(sparams.public_path);
-
-    // to make it ctrl+clickable:
-    LOG_TEE("\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);
-
-    std::unordered_map<std::string, std::string> log_data;
-    log_data["hostname"] = sparams.hostname;
-    log_data["port"] = std::to_string(sparams.port);
-
-    if (!sparams.api_key.empty()) {
-        log_data["api_key"] = "api_key: ****" + sparams.api_key.substr(sparams.api_key.length() - 4);
-    }
-
-    LOG_INFO("HTTP server listening", log_data);
-    // run the HTTP server in a thread - see comment below
-    std::thread t([&]()
-            {
-                if (!svr.listen_after_bind())
-                {
-                    return 1;
-                }
-
-                return 0;
-            });
-
     // GG: if I put the main loop inside a thread, it crashes on the first request when build in Debug!?
     //     "Bus error: 10" - this is on macOS, it does not crash on Linux
     //std::thread t2([&]()

ggml-alloc.c

@@ -102,8 +102,6 @@ void ggml_tallocr_alloc(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
         }
     }
 
-    AT_PRINTF("block %d\n", best_fit_block);
-
     if (best_fit_block == -1) {
         // the last block is our last resort
         struct free_block * block = &alloc->free_blocks[alloc->n_free_blocks - 1];
@@ -117,6 +115,7 @@ void ggml_tallocr_alloc(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
             return;
         }
     }
+
     struct free_block * block = &alloc->free_blocks[best_fit_block];
     void * addr = block->addr;
     block->addr = (char*)block->addr + size;
@@ -129,6 +128,8 @@ void ggml_tallocr_alloc(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
         }
     }
 
+    AT_PRINTF("block %d, addr %p\n", best_fit_block, addr);
+
     tensor->data = addr;
     tensor->buffer = alloc->buffer;
     if (!alloc->measure) {
@@ -229,6 +230,7 @@ void ggml_tallocr_reset(ggml_tallocr_t alloc) {
         alloc->free_blocks[0].size = SIZE_MAX/2; // restrict maximum size of a measure allocator to half size_t max to avoid overflows
     } else {
         alloc->free_blocks[0].size = ggml_backend_buffer_get_size(alloc->buffer) - align_offset;
+        ggml_backend_buffer_reset(alloc->buffer);
     }
 }
 
@@ -263,9 +265,9 @@ ggml_tallocr_t ggml_tallocr_new_measure(size_t alignment) {
     return alloc;
 }
 
-ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backend) {
+ggml_tallocr_t ggml_tallocr_new_measure_from_buft(struct ggml_backend_buffer_type * buft) {
     // create a backend buffer to get the correct tensor allocation sizes
-    ggml_backend_buffer_t buffer = ggml_backend_alloc_buffer(backend, 1);
+    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, 1);
 
     // TODO: move alloc initialization to a common ggml_tallocr_new_impl function
     ggml_tallocr_t alloc = ggml_tallocr_new_from_buffer(buffer);
@@ -275,13 +277,22 @@ ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backe
     return alloc;
 }
 
-ggml_tallocr_t ggml_tallocr_new_from_backend(struct ggml_backend * backend, size_t size) {
-    ggml_backend_buffer_t buffer = ggml_backend_alloc_buffer(backend, size);
+ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backend) {
+    return ggml_tallocr_new_measure_from_buft(ggml_backend_get_default_buffer_type(backend));
+}
+
+ggml_tallocr_t ggml_tallocr_new_from_buft(struct ggml_backend_buffer_type * buft, size_t size) {
+    // create a backend buffer to get the correct tensor allocation sizes
+    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, size);
     ggml_tallocr_t alloc = ggml_tallocr_new_from_buffer(buffer);
     alloc->buffer_owned = true;
     return alloc;
 }
 
+ggml_tallocr_t ggml_tallocr_new_from_backend(struct ggml_backend * backend, size_t size) {
+    return ggml_tallocr_new_from_buft(ggml_backend_get_default_buffer_type(backend), size);
+}
+
 ggml_tallocr_t ggml_tallocr_new_from_buffer(struct ggml_backend_buffer * buffer) {
     ggml_tallocr_t alloc = (ggml_tallocr_t)malloc(sizeof(struct ggml_tallocr));
 
@@ -779,10 +790,21 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
 
     if (nbytes == 0) {
         // all the tensors in the context are already allocated
+#ifndef NDEBUG
+        fprintf(stderr, "%s: all tensors in the context are already allocated\n", __func__);
+#endif
         return NULL;
     }
 
     ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, nbytes);
+    if (buffer == NULL) {
+        // failed to allocate buffer
+#ifndef NDEBUG
+        fprintf(stderr, "%s: failed to allocate buffer\n", __func__);
+#endif
+        return NULL;
+    }
+
     ggml_tallocr_t tallocr = ggml_tallocr_new_from_buffer(buffer);
 
     for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {

ggml-alloc.h

@@ -52,8 +52,10 @@ typedef struct ggml_tallocr * ggml_tallocr_t;
 
 GGML_API ggml_tallocr_t ggml_tallocr_new(void * data, size_t size, size_t alignment);
 GGML_API ggml_tallocr_t ggml_tallocr_new_measure(size_t alignment);
-GGML_API ggml_tallocr_t ggml_tallocr_new_from_buffer(struct ggml_backend_buffer * buffer);
+GGML_API ggml_tallocr_t ggml_tallocr_new_from_buft(struct ggml_backend_buffer_type * buft, size_t size);
 GGML_API ggml_tallocr_t ggml_tallocr_new_from_backend(struct ggml_backend * backend, size_t size); // allocates an owned buffer
+GGML_API ggml_tallocr_t ggml_tallocr_new_from_buffer(struct ggml_backend_buffer * buffer);
+GGML_API ggml_tallocr_t ggml_tallocr_new_measure_from_buft(struct ggml_backend_buffer_type * buft);
 GGML_API ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backend);
 
 GGML_API struct ggml_backend_buffer * ggml_tallocr_get_buffer(ggml_tallocr_t talloc);

ggml-backend-impl.h

@@ -16,9 +16,10 @@ extern "C" {
     typedef void * ggml_backend_buffer_type_context_t;
 
     struct ggml_backend_buffer_type_i {
+        const char *          (*get_name)        (ggml_backend_buffer_type_t buft);
         ggml_backend_buffer_t (*alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
         size_t                (*get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
-        size_t                (*get_alloc_size)  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
+        size_t                (*get_alloc_size)  (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
         bool                  (*supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
         // check if tensor data is in host memory
         // should be equivalent to supports_backend(buft, ggml_backend_cpu_init())
@@ -34,16 +35,15 @@ extern "C" {
     typedef void * ggml_backend_buffer_context_t;
 
     struct ggml_backend_buffer_i {
-        void   (*free_buffer)    (ggml_backend_buffer_t buffer);
-        //void     (*reset)      (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
-        void * (*get_base)       (ggml_backend_buffer_t buffer);
-        void   (*init_tensor)    (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-        void   (*set_tensor)     (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void   (*get_tensor)     (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        // (optional) copy tensor between different buffer-type, allow for single-copy tranfers
-        void   (*cpy_tensor_from)(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void   (*cpy_tensor_to)  (ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void   (*clear)          (ggml_backend_buffer_t buffer, uint8_t value);
+        const char * (*get_name)   (ggml_backend_buffer_t buffer);
+        void         (*free_buffer)(ggml_backend_buffer_t buffer);
+        void *       (*get_base)   (ggml_backend_buffer_t buffer);
+        void         (*init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+        void         (*set_tensor) (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void         (*get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        bool         (*cpy_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); // dst is in the buffer, src may be in any buffer
+        void         (*clear)      (ggml_backend_buffer_t buffer, uint8_t value);
+        void         (*reset)      (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
     };
 
     struct ggml_backend_buffer {
@@ -51,6 +51,7 @@ extern "C" {
         ggml_backend_buffer_type_t    buft;
         ggml_backend_buffer_context_t context;
         size_t size;
+        enum ggml_backend_buffer_usage usage;
     };
 
     ggml_backend_buffer_t ggml_backend_buffer_init(
@@ -59,6 +60,8 @@ extern "C" {
                    ggml_backend_buffer_context_t   context,
                    size_t                          size);
 
+    // do not use directly, use ggml_backend_tensor_copy instead
+    bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
 
     //
     // Backend
@@ -74,22 +77,20 @@ extern "C" {
         // buffer allocation
         ggml_backend_buffer_type_t (*get_default_buffer_type)(ggml_backend_t backend);
 
-        // (optional) asynchroneous tensor data access
+        // (optional) asynchronous tensor data access
         void (*set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
         void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        bool (*cpy_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * src, struct ggml_tensor * dst);
 
-        // (optional) asynchroneous tensor copy
-        void (*cpy_tensor_from_async)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void (*cpy_tensor_to_async)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
-
+        // (optional) complete all pending operations
         void (*synchronize)(ggml_backend_t backend);
 
         // compute graph with a plan
-        ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+        ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
         void                      (*graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
         void                      (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
 
-        // compute graph without a plan
+        // compute graph without a plan (async)
         bool (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
 
         // check if the backend supports an operation
@@ -102,7 +103,6 @@ extern "C" {
         ggml_backend_context_t context;
     };
 
-
     //
     // Backend registry
     //

ggml-backend.h

@@ -17,22 +17,31 @@ extern "C" {
     //
 
     // buffer type
-    GGML_API ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size);
-    GGML_API size_t ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft);
-    GGML_API size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
-    GGML_API bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
-    GGML_API bool ggml_backend_buft_is_host         (ggml_backend_buffer_type_t buft);
+    GGML_API const char *          ggml_backend_buft_name            (ggml_backend_buffer_type_t buft);
+    GGML_API ggml_backend_buffer_t ggml_backend_buft_alloc_buffer    (ggml_backend_buffer_type_t buft, size_t size);
+    GGML_API size_t                ggml_backend_buft_get_alignment   (ggml_backend_buffer_type_t buft);
+    GGML_API size_t                ggml_backend_buft_get_alloc_size  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
+    GGML_API bool                  ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
+    GGML_API bool                  ggml_backend_buft_is_host         (ggml_backend_buffer_type_t buft);
 
     // buffer
-    GGML_API void   ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
-    GGML_API void * ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
-    GGML_API void   ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API void   ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
-    GGML_API bool   ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
-    GGML_API ggml_backend_buffer_type_t ggml_backend_buffer_type(ggml_backend_buffer_t buffer);
+    enum ggml_backend_buffer_usage {
+        GGML_BACKEND_BUFFER_USAGE_ANY = 0,
+        GGML_BACKEND_BUFFER_USAGE_WEIGHTS = 1,
+    };
+
+    GGML_API const char *               ggml_backend_buffer_name          (ggml_backend_buffer_t buffer);
+    GGML_API void                       ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
+    GGML_API void *                     ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
+    GGML_API size_t                     ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
+    GGML_API void                       ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API size_t                     ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
+    GGML_API size_t                     ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API void                       ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
+    GGML_API bool                       ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
+    GGML_API void                       ggml_backend_buffer_set_usage     (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
+    GGML_API ggml_backend_buffer_type_t ggml_backend_buffer_get_type      (ggml_backend_buffer_t buffer);
+    GGML_API void                       ggml_backend_buffer_reset         (ggml_backend_buffer_t buffer);
 
     //
     // Backend
@@ -140,23 +149,24 @@ extern "C" {
     typedef struct ggml_backend_sched * ggml_backend_sched_t;
 
     // Initialize a backend scheduler
-    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends);
-
-    GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched);
-
+    GGML_API ggml_backend_sched_t  ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size);
+    GGML_API void                  ggml_backend_sched_free(ggml_backend_sched_t sched);
     // Initialize backend buffers from a measure graph
-    GGML_API void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
+    GGML_API void                  ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
+    // Get the number of splits of the last graph
+    GGML_API int                   ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
 
     GGML_API ggml_tallocr_t        ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend);
     GGML_API ggml_backend_buffer_t ggml_backend_sched_get_buffer (ggml_backend_sched_t sched, ggml_backend_t backend);
 
-    GGML_API void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
+    GGML_API void                  ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
+    GGML_API ggml_backend_t        ggml_backend_sched_get_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
 
-    // Allocate a graph on the backend scheduler
-    GGML_API void ggml_backend_sched_graph_compute(
-            ggml_backend_sched_t sched,
-            struct ggml_cgraph * graph);
+    // Allocate and compute graph on the backend scheduler
+    GGML_API void                  ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
 
+    // Reset all assignments and allocators - must be called before using the sched allocators to allocate inputs
+    GGML_API void                  ggml_backend_sched_reset(ggml_backend_sched_t sched);
 
     //
     // Utils
@@ -176,7 +186,7 @@ extern "C" {
     typedef bool (*ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
 
     // Compare the output of two backends
-    GGML_API void ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
+    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
 
     // Tensor initialization
     GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);

ggml-cuda.h

@@ -27,22 +27,6 @@ GGML_API void * ggml_cuda_host_malloc(size_t size);
 GGML_API void   ggml_cuda_host_free(void * ptr);
 
 GGML_API bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API void   ggml_cuda_set_tensor_split(const float * tensor_split);
-GGML_API void   ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
-GGML_API void   ggml_cuda_free_data(struct ggml_tensor * tensor);
-
-GGML_API void   ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
-GGML_API void   ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
-GGML_API void   ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
-
-GGML_API void   ggml_cuda_assign_buffers_no_alloc(struct ggml_tensor * tensor);
-GGML_API void   ggml_cuda_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset);
-GGML_API void   ggml_cuda_copy_to_device(struct ggml_tensor * tensor);
-
-GGML_API void   ggml_cuda_set_main_device(int main_device);
-GGML_API void   ggml_cuda_set_mul_mat_q(bool mul_mat_q);
-GGML_API void   ggml_cuda_set_scratch_size(size_t scratch_size);
-GGML_API void   ggml_cuda_free_scratch(void);
 GGML_API bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
 
 GGML_API int    ggml_cuda_get_device_count(void);
@@ -52,13 +36,17 @@ GGML_API void   ggml_cuda_get_device_description(int device, char * description,
 GGML_API ggml_backend_t ggml_backend_cuda_init(int device);
 
 GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
-GGML_API int  ggml_backend_cuda_get_device(ggml_backend_t backend);
 
 GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
-
-// pinned host buffer for use with CPU backend for faster copies between CPU and GPU
+// split tensor buffer that splits matrices by rows across multiple devices
+GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
+// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
 GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
 
+GGML_API int  ggml_backend_cuda_get_device_count(void);
+GGML_API void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
+GGML_API void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
+
 #ifdef  __cplusplus
 }
 #endif

ggml-impl.h

@@ -228,6 +228,8 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 #define GGML_HASHTABLE_FULL ((size_t)-1)
 #define GGML_HASHTABLE_ALREADY_EXISTS ((size_t)-2)
 
+struct ggml_hash_set ggml_hash_set_new(size_t size);
+
 bool   ggml_hash_contains      (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
 
 // returns GGML_HASHTABLE_FULL if table is full, otherwise the current index of the key or where it should be inserted

ggml-metal.h

@@ -36,64 +36,13 @@ struct ggml_cgraph;
 extern "C" {
 #endif
 
-//
-// internal API
-// temporary exposed to user-code
-//
-
-struct ggml_metal_context;
-
-void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
-
-// number of command buffers to use
-struct ggml_metal_context * ggml_metal_init(int n_cb);
-void ggml_metal_free(struct ggml_metal_context * ctx);
-
-void * ggml_metal_host_malloc(size_t n);
-void   ggml_metal_host_free  (void * data);
-
-// set the number of command buffers to use
-void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb);
-
-// creates a mapping between a host memory buffer and a device memory buffer
-// - make sure to map all buffers used in the graph before calling ggml_metal_graph_compute
-// - the mapping is used during computation to determine the arguments of the compute kernels
-// - you don't need to keep the host memory buffer allocated as it is never accessed by Metal
-// - max_size specifies the maximum size of a tensor and is used to create shared views such
-//   that it is guaranteed that the tensor will fit in at least one of the views
-//
-bool ggml_metal_add_buffer(
-        struct ggml_metal_context * ctx,
-                       const char * name,
-                             void * data,
-                           size_t   size,
-                           size_t   max_size);
-
-// set data from host memory into the device
-void ggml_metal_set_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
-
-// get data from the device into host memory
-void ggml_metal_get_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
-
-// try to find operations that can be run concurrently in the graph
-// you should run it again if the topology of your graph changes
-void ggml_metal_graph_find_concurrency(struct ggml_metal_context * ctx, struct ggml_cgraph * gf, bool check_mem);
-
-// if the graph has been optimized for concurrently dispatch, return length of the concur_list if optimized
-int ggml_metal_if_optimized(struct ggml_metal_context * ctx);
-
-// output the concur_list for ggml_alloc
-int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx);
-
-// same as ggml_graph_compute but uses Metal
-// creates gf->n_threads command buffers in parallel
-bool ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
-
 //
 // backend API
 // user-code should use only these functions
 //
 
+GGML_API void ggml_backend_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
+
 GGML_API ggml_backend_t ggml_backend_metal_init(void);
 
 GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);

ggml-metal.metal

@@ -2446,6 +2446,19 @@ typedef struct {
 } block_q6_K;
 // 210 bytes / block
 
+typedef struct {
+    half d;
+    uint16_t qs[QK_K/8];
+} block_iq2_xxs;
+// 66 bytes / block for QK_K = 256, so 2.0625 bpw
+
+typedef struct {
+    half d;
+    uint16_t qs[QK_K/8];
+    uint8_t  scales[QK_K/32];
+} block_iq2_xs;
+// 74 bytes / block for QK_K = 256, so 2.3125 bpw
+
 //====================================== dot products =========================
 
 void kernel_mul_mv_q2_K_f32_impl(
@@ -3468,6 +3481,495 @@ kernel void kernel_mul_mv_q6_K_f32(
     kernel_mul_mv_q6_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, tgpig, tiisg, sgitg);
 }
 
+// ======================= "True" 2-bit
+
+constexpr constant static uint64_t iq2xxs_grid[256] = {
+    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
+    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x08080808082b0808,
+    0x08080808082b082b, 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819,
+    0x0808080819081908, 0x0808080819190808, 0x0808080819192b08, 0x08080808192b0819,
+    0x08080808192b1908, 0x080808082b080808, 0x080808082b08082b, 0x080808082b082b2b,
+    0x080808082b2b082b, 0x0808081908080819, 0x0808081908081908, 0x0808081908190808,
+    0x0808081908191919, 0x0808081919080808, 0x080808192b081908, 0x080808192b192b08,
+    0x0808082b08080808, 0x0808082b0808082b, 0x0808082b082b082b, 0x0808082b2b08082b,
+    0x0808190808080819, 0x0808190808081908, 0x0808190808190808, 0x08081908082b0819,
+    0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, 0x0808190819082b08,
+    0x08081908192b0808, 0x080819082b080819, 0x080819082b081908, 0x080819082b190808,
+    0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, 0x0808191908082b08,
+    0x08081919082b0808, 0x080819191908192b, 0x08081919192b2b19, 0x080819192b080808,
+    0x080819192b190819, 0x0808192b08082b19, 0x0808192b08190808, 0x0808192b19080808,
+    0x0808192b2b081908, 0x0808192b2b2b1908, 0x08082b0808080808, 0x08082b0808081919,
+    0x08082b0808082b08, 0x08082b0808191908, 0x08082b08082b2b08, 0x08082b0819080819,
+    0x08082b0819081908, 0x08082b0819190808, 0x08082b081919082b, 0x08082b082b082b08,
+    0x08082b1908081908, 0x08082b1919080808, 0x08082b2b0808082b, 0x08082b2b08191908,
+    0x0819080808080819, 0x0819080808081908, 0x0819080808190808, 0x08190808082b0819,
+    0x0819080819080808, 0x08190808192b0808, 0x081908082b081908, 0x081908082b190808,
+    0x081908082b191919, 0x0819081908080808, 0x0819081908082b08, 0x08190819082b0808,
+    0x0819081919190808, 0x0819081919192b2b, 0x081908192b080808, 0x0819082b082b1908,
+    0x0819082b19081919, 0x0819190808080808, 0x0819190808082b08, 0x08191908082b0808,
+    0x08191908082b1919, 0x0819190819082b19, 0x081919082b080808, 0x0819191908192b08,
+    0x08191919192b082b, 0x0819192b08080808, 0x0819192b0819192b, 0x08192b0808080819,
+    0x08192b0808081908, 0x08192b0808190808, 0x08192b0819080808, 0x08192b082b080819,
+    0x08192b1908080808, 0x08192b1908081919, 0x08192b192b2b0808, 0x08192b2b19190819,
+    0x082b080808080808, 0x082b08080808082b, 0x082b080808082b2b, 0x082b080819081908,
+    0x082b0808192b0819, 0x082b08082b080808, 0x082b08082b08082b, 0x082b0819082b2b19,
+    0x082b081919082b08, 0x082b082b08080808, 0x082b082b0808082b, 0x082b190808080819,
+    0x082b190808081908, 0x082b190808190808, 0x082b190819080808, 0x082b19081919192b,
+    0x082b191908080808, 0x082b191919080819, 0x082b1919192b1908, 0x082b192b2b190808,
+    0x082b2b0808082b08, 0x082b2b08082b0808, 0x082b2b082b191908, 0x082b2b2b19081908,
+    0x1908080808080819, 0x1908080808081908, 0x1908080808190808, 0x1908080808192b08,
+    0x19080808082b0819, 0x19080808082b1908, 0x1908080819080808, 0x1908080819082b08,
+    0x190808081919192b, 0x19080808192b0808, 0x190808082b080819, 0x190808082b081908,
+    0x190808082b190808, 0x1908081908080808, 0x19080819082b0808, 0x19080819192b0819,
+    0x190808192b080808, 0x190808192b081919, 0x1908082b08080819, 0x1908082b08190808,
+    0x1908082b19082b08, 0x1908082b1919192b, 0x1908082b192b2b08, 0x1908190808080808,
+    0x1908190808082b08, 0x19081908082b0808, 0x190819082b080808, 0x190819082b192b19,
+    0x190819190819082b, 0x19081919082b1908, 0x1908192b08080808, 0x19082b0808080819,
+    0x19082b0808081908, 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919,
+    0x19082b1908080808, 0x19082b1919192b08, 0x19082b19192b0819, 0x19082b192b08082b,
+    0x19082b2b19081919, 0x19082b2b2b190808, 0x1919080808080808, 0x1919080808082b08,
+    0x1919080808190819, 0x1919080808192b19, 0x19190808082b0808, 0x191908082b080808,
+    0x191908082b082b08, 0x1919081908081908, 0x191908191908082b, 0x191908192b2b1908,
+    0x1919082b2b190819, 0x191919082b190808, 0x191919082b19082b, 0x1919191908082b2b,
+    0x1919192b08080819, 0x1919192b19191908, 0x19192b0808080808, 0x19192b0808190819,
+    0x19192b0808192b19, 0x19192b08192b1908, 0x19192b1919080808, 0x19192b2b08082b08,
+    0x192b080808081908, 0x192b080808190808, 0x192b080819080808, 0x192b0808192b2b08,
+    0x192b081908080808, 0x192b081919191919, 0x192b082b08192b08, 0x192b082b192b0808,
+    0x192b190808080808, 0x192b190808081919, 0x192b191908190808, 0x192b19190819082b,
+    0x192b19192b081908, 0x192b2b081908082b, 0x2b08080808080808, 0x2b0808080808082b,
+    0x2b08080808082b2b, 0x2b08080819080819, 0x2b0808082b08082b, 0x2b08081908081908,
+    0x2b08081908192b08, 0x2b08081919080808, 0x2b08082b08190819, 0x2b08190808080819,
+    0x2b08190808081908, 0x2b08190808190808, 0x2b08190808191919, 0x2b08190819080808,
+    0x2b081908192b0808, 0x2b08191908080808, 0x2b0819191908192b, 0x2b0819192b191908,
+    0x2b08192b08082b19, 0x2b08192b19080808, 0x2b08192b192b0808, 0x2b082b080808082b,
+    0x2b082b1908081908, 0x2b082b2b08190819, 0x2b19080808081908, 0x2b19080808190808,
+    0x2b190808082b1908, 0x2b19080819080808, 0x2b1908082b2b0819, 0x2b1908190819192b,
+    0x2b1908192b080808, 0x2b19082b19081919, 0x2b19190808080808, 0x2b191908082b082b,
+    0x2b19190819081908, 0x2b19191919190819, 0x2b192b082b080819, 0x2b192b19082b0808,
+    0x2b2b08080808082b, 0x2b2b080819190808, 0x2b2b08082b081919, 0x2b2b081908082b19,
+    0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, 0x2b2b2b1908081908,
+};
+
+constexpr constant static uint64_t iq2xs_grid[512] = {
+    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
+    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b,
+    0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919,
+    0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b,
+    0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919,
+    0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808,
+    0x080808082b08082b, 0x080808082b081919, 0x080808082b082b08, 0x080808082b190819,
+    0x080808082b191908, 0x080808082b192b19, 0x080808082b2b0808, 0x0808081908080819,
+    0x0808081908081908, 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808,
+    0x080808190819082b, 0x0808081908191919, 0x0808081908192b08, 0x0808081908192b2b,
+    0x08080819082b0819, 0x08080819082b1908, 0x0808081919080808, 0x080808191908082b,
+    0x0808081919081919, 0x0808081919082b08, 0x0808081919190819, 0x0808081919191908,
+    0x08080819192b0808, 0x08080819192b2b08, 0x080808192b080819, 0x080808192b081908,
+    0x080808192b190808, 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b08081919,
+    0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, 0x0808082b082b0808,
+    0x0808082b19080819, 0x0808082b19081908, 0x0808082b19190808, 0x0808082b19191919,
+    0x0808082b2b080808, 0x0808082b2b082b2b, 0x0808190808080819, 0x0808190808081908,
+    0x080819080808192b, 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b,
+    0x0808190808191919, 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908,
+    0x0808190819080808, 0x080819081908082b, 0x0808190819081919, 0x0808190819082b08,
+    0x0808190819190819, 0x0808190819191908, 0x080819081919192b, 0x08081908192b0808,
+    0x080819082b080819, 0x080819082b081908, 0x080819082b190808, 0x0808191908080808,
+    0x080819190808082b, 0x0808191908081919, 0x0808191908082b08, 0x0808191908190819,
+    0x0808191908191908, 0x08081919082b0808, 0x0808191919080819, 0x0808191919081908,
+    0x0808191919190808, 0x08081919192b0819, 0x080819192b080808, 0x0808192b08080819,
+    0x0808192b08081908, 0x0808192b08190808, 0x0808192b082b192b, 0x0808192b19080808,
+    0x0808192b1908082b, 0x0808192b2b081908, 0x08082b0808080808, 0x08082b080808082b,
+    0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808082b2b, 0x08082b0808190819,
+    0x08082b0808191908, 0x08082b08082b0808, 0x08082b08082b1919, 0x08082b0819080819,
+    0x08082b0819081908, 0x08082b0819190808, 0x08082b0819192b08, 0x08082b082b080808,
+    0x08082b082b2b0808, 0x08082b082b2b2b2b, 0x08082b1908080819, 0x08082b1908081908,
+    0x08082b1908190808, 0x08082b1919080808, 0x08082b192b080819, 0x08082b192b082b19,
+    0x08082b2b08080808, 0x08082b2b082b0808, 0x08082b2b082b2b08, 0x08082b2b2b19192b,
+    0x08082b2b2b2b0808, 0x0819080808080819, 0x0819080808081908, 0x081908080808192b,
+    0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, 0x0819080808191919,
+    0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, 0x0819080819080808,
+    0x081908081908082b, 0x0819080819081919, 0x0819080819082b08, 0x0819080819190819,
+    0x0819080819191908, 0x08190808192b0808, 0x08190808192b2b2b, 0x081908082b080819,
+    0x081908082b081908, 0x081908082b190808, 0x0819081908080808, 0x081908190808082b,
+    0x0819081908081919, 0x0819081908082b08, 0x0819081908190819, 0x0819081908191908,
+    0x08190819082b0808, 0x0819081919080819, 0x0819081919081908, 0x0819081919190808,
+    0x081908192b080808, 0x081908192b191908, 0x081908192b19192b, 0x0819082b08080819,
+    0x0819082b08081908, 0x0819082b0808192b, 0x0819082b08190808, 0x0819082b19080808,
+    0x0819082b192b0808, 0x0819190808080808, 0x081919080808082b, 0x0819190808081919,
+    0x0819190808082b08, 0x0819190808190819, 0x0819190808191908, 0x08191908082b0808,
+    0x0819190819080819, 0x0819190819081908, 0x0819190819082b19, 0x0819190819190808,
+    0x08191908192b1908, 0x081919082b080808, 0x0819191908080819, 0x0819191908081908,
+    0x0819191908190808, 0x0819191919080808, 0x0819192b08080808, 0x0819192b08191908,
+    0x0819192b19082b19, 0x08192b0808080819, 0x08192b0808081908, 0x08192b0808190808,
+    0x08192b080819082b, 0x08192b0819080808, 0x08192b0819191908, 0x08192b082b08192b,
+    0x08192b1908080808, 0x08192b1908081919, 0x08192b19192b192b, 0x08192b2b19190819,
+    0x08192b2b2b2b2b19, 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919,
+    0x082b080808082b08, 0x082b080808082b2b, 0x082b080808190819, 0x082b080808191908,
+    0x082b0808082b0808, 0x082b080819080819, 0x082b080819081908, 0x082b080819190808,
+    0x082b08082b080808, 0x082b08082b2b0808, 0x082b081908080819, 0x082b081908081908,
+    0x082b081908190808, 0x082b081919080808, 0x082b081919082b08, 0x082b0819192b1919,
+    0x082b082b08080808, 0x082b082b082b082b, 0x082b082b2b080808, 0x082b082b2b2b2b08,
+    0x082b190808080819, 0x082b190808081908, 0x082b190808190808, 0x082b1908082b2b19,
+    0x082b190819080808, 0x082b191908080808, 0x082b191919080819, 0x082b19191919082b,
+    0x082b19192b192b19, 0x082b192b08080819, 0x082b192b08192b2b, 0x082b192b2b2b192b,
+    0x082b2b0808080808, 0x082b2b0808082b08, 0x082b2b0808082b2b, 0x082b2b08082b0808,
+    0x082b2b0819191919, 0x082b2b082b082b08, 0x082b2b082b2b082b, 0x082b2b19192b2b08,
+    0x082b2b192b190808, 0x082b2b2b08082b08, 0x082b2b2b082b0808, 0x082b2b2b2b08082b,
+    0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, 0x1908080808081908,
+    0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, 0x190808080819082b,
+    0x1908080808191919, 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908,
+    0x1908080819080808, 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08,
+    0x1908080819082b2b, 0x1908080819190819, 0x1908080819191908, 0x19080808192b0808,
+    0x19080808192b1919, 0x190808082b080819, 0x190808082b081908, 0x190808082b190808,
+    0x1908081908080808, 0x190808190808082b, 0x1908081908081919, 0x1908081908082b08,
+    0x1908081908190819, 0x1908081908191908, 0x19080819082b0808, 0x1908081919080819,
+    0x1908081919081908, 0x1908081919190808, 0x190808192b080808, 0x190808192b081919,
+    0x190808192b2b082b, 0x1908082b08080819, 0x1908082b08081908, 0x1908082b08190808,
+    0x1908082b0819082b, 0x1908082b082b2b19, 0x1908082b19080808, 0x1908190808080808,
+    0x190819080808082b, 0x1908190808081919, 0x1908190808082b08, 0x1908190808190819,
+    0x1908190808191908, 0x1908190808192b19, 0x19081908082b0808, 0x1908190819080819,
+    0x1908190819081908, 0x1908190819190808, 0x190819082b080808, 0x190819082b191908,
+    0x1908191908080819, 0x1908191908081908, 0x1908191908190808, 0x19081919082b1908,
+    0x1908191919080808, 0x190819192b192b2b, 0x1908192b08080808, 0x1908192b08082b2b,
+    0x1908192b19081908, 0x1908192b19190808, 0x19082b0808080819, 0x19082b0808081908,
+    0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, 0x19082b0819191908,
+    0x19082b08192b082b, 0x19082b1908080808, 0x19082b1908190819, 0x19082b1919081908,
+    0x19082b1919190808, 0x19082b19192b2b19, 0x19082b2b08081908, 0x1919080808080808,
+    0x191908080808082b, 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819,
+    0x1919080808191908, 0x19190808082b0808, 0x19190808082b2b08, 0x1919080819080819,
+    0x1919080819081908, 0x1919080819190808, 0x191908082b080808, 0x1919081908080819,
+    0x1919081908081908, 0x1919081908190808, 0x1919081908191919, 0x1919081919080808,
+    0x191908191908082b, 0x1919082b08080808, 0x1919082b19081908, 0x1919082b2b2b2b2b,
+    0x1919190808080819, 0x1919190808081908, 0x1919190808190808, 0x19191908082b0819,
+    0x1919190819080808, 0x19191908192b0808, 0x191919082b080819, 0x191919082b2b0819,
+    0x1919191908080808, 0x1919191908082b08, 0x191919192b080808, 0x191919192b082b08,
+    0x1919192b082b0819, 0x1919192b192b2b08, 0x1919192b2b2b0819, 0x19192b0808080808,
+    0x19192b0808191908, 0x19192b0819080819, 0x19192b0819190808, 0x19192b082b192b19,
+    0x19192b1908192b2b, 0x19192b1919080808, 0x19192b191908082b, 0x19192b2b2b081919,
+    0x192b080808080819, 0x192b080808081908, 0x192b080808190808, 0x192b080819080808,
+    0x192b080819191908, 0x192b0808192b082b, 0x192b08082b08192b, 0x192b08082b2b2b19,
+    0x192b081908080808, 0x192b082b082b1908, 0x192b082b19082b2b, 0x192b082b2b19082b,
+    0x192b190808080808, 0x192b19080819192b, 0x192b191908190808, 0x192b191919080808,
+    0x192b191919081919, 0x192b19192b2b1908, 0x192b2b0808080819, 0x192b2b08192b2b2b,
+    0x192b2b19082b1919, 0x192b2b2b0808192b, 0x192b2b2b19191908, 0x192b2b2b192b082b,
+    0x2b08080808080808, 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08,
+    0x2b08080808190819, 0x2b08080808191908, 0x2b080808082b0808, 0x2b080808082b2b2b,
+    0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808082b080808,
+    0x2b0808082b08082b, 0x2b0808082b2b2b08, 0x2b0808082b2b2b2b, 0x2b08081908080819,
+    0x2b08081908081908, 0x2b0808190808192b, 0x2b08081908190808, 0x2b08081919080808,
+    0x2b08081919190819, 0x2b08081919192b19, 0x2b08082b08080808, 0x2b08082b082b0808,
+    0x2b08082b2b080808, 0x2b08082b2b08082b, 0x2b08082b2b2b0808, 0x2b08082b2b2b2b08,
+    0x2b08190808080819, 0x2b08190808081908, 0x2b08190808190808, 0x2b0819080819082b,
+    0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, 0x2b0819082b082b19,
+    0x2b08191908080808, 0x2b08191919081908, 0x2b0819192b2b1919, 0x2b08192b08192b08,
+    0x2b08192b192b2b2b, 0x2b082b0808080808, 0x2b082b0808082b08, 0x2b082b08082b1919,
+    0x2b082b0819192b2b, 0x2b082b082b080808, 0x2b082b082b08082b, 0x2b082b082b2b2b08,
+    0x2b082b190808192b, 0x2b082b2b082b082b, 0x2b082b2b2b080808, 0x2b082b2b2b082b08,
+    0x2b082b2b2b19192b, 0x2b082b2b2b2b2b08, 0x2b19080808080819, 0x2b19080808081908,
+    0x2b19080808190808, 0x2b19080819080808, 0x2b1908081919192b, 0x2b1908082b081908,
+    0x2b19081908080808, 0x2b190819082b082b, 0x2b190819192b1908, 0x2b19082b1919192b,
+    0x2b19082b2b082b19, 0x2b19190808080808, 0x2b19190808081919, 0x2b19190819081908,
+    0x2b19190819190808, 0x2b19190819192b08, 0x2b191919082b2b19, 0x2b1919192b190808,
+    0x2b1919192b19082b, 0x2b19192b19080819, 0x2b192b0819190819, 0x2b192b082b2b192b,
+    0x2b192b1919082b19, 0x2b192b2b08191919, 0x2b192b2b192b0808, 0x2b2b080808080808,
+    0x2b2b08080808082b, 0x2b2b080808082b08, 0x2b2b080808082b2b, 0x2b2b0808082b0808,
+    0x2b2b0808082b2b2b, 0x2b2b08082b2b0808, 0x2b2b081919190819, 0x2b2b081919192b19,
+    0x2b2b08192b2b192b, 0x2b2b082b08080808, 0x2b2b082b0808082b, 0x2b2b082b08082b08,
+    0x2b2b082b082b2b2b, 0x2b2b082b2b080808, 0x2b2b082b2b2b0808, 0x2b2b190819080808,
+    0x2b2b19082b191919, 0x2b2b192b192b1919, 0x2b2b192b2b192b08, 0x2b2b2b0808082b2b,
+    0x2b2b2b08082b0808, 0x2b2b2b08082b082b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b0808,
+    0x2b2b2b082b2b2b08, 0x2b2b2b1908081908, 0x2b2b2b192b081908, 0x2b2b2b192b08192b,
+    0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b,
+};
+
+constexpr constant static uint8_t ksigns_iq2xs[128] = {
+      0, 129, 130,   3, 132,   5,   6, 135, 136,   9,  10, 139,  12, 141, 142,  15,
+    144,  17,  18, 147,  20, 149, 150,  23,  24, 153, 154,  27, 156,  29,  30, 159,
+    160,  33,  34, 163,  36, 165, 166,  39,  40, 169, 170,  43, 172,  45,  46, 175,
+     48, 177, 178,  51, 180,  53,  54, 183, 184,  57,  58, 187,  60, 189, 190,  63,
+    192,  65,  66, 195,  68, 197, 198,  71,  72, 201, 202,  75, 204,  77,  78, 207,
+     80, 209, 210,  83, 212,  85,  86, 215, 216,  89,  90, 219,  92, 221, 222,  95,
+     96, 225, 226,  99, 228, 101, 102, 231, 232, 105, 106, 235, 108, 237, 238, 111,
+    240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255,
+};
+
+constexpr constant static uint8_t kmask_iq2xs[8] = {1, 2, 4, 8, 16, 32, 64, 128};
+
+void kernel_mul_mv_iq2_xxs_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int ib_row = first_row * nb;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    device const block_iq2_xxs * x = (device const block_iq2_xxs *) src0 + ib_row + offset0;
+    device const float         * y = (device const float         *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[32];
+    float sumf[N_DST]={0.f}, all_sum;
+
+    const int nb32 = nb * (QK_K / 32);
+
+    threadgroup uint64_t * values = (threadgroup uint64_t *)shared_values;
+    threadgroup uint8_t  * shared_signs = (threadgroup uint8_t *)(values + 256);
+    {
+        int nval = 4;
+        int pos  = (32*sgitg + tiisg)*nval;
+        for (int i = 0; i < nval; ++i) values[pos + i] = iq2xxs_grid[pos + i];
+        nval = 2;
+        pos  = (32*sgitg + tiisg)*nval;
+        for (int i = 0; i < nval; ++i) shared_signs[pos+i] = ksigns_iq2xs[pos+i];
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+#if QK_K == 256
+    const int ix = tiisg;
+
+    device const float * y4 = y + 32 * ix;
+
+    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
+
+        for (int i = 0; i < 32; ++i) {
+            yl[i] = y4[i];
+        }
+
+        const int ibl = ib32 / (QK_K / 32);
+        const int ib  = ib32 % (QK_K / 32);
+
+        device const block_iq2_xxs * xr = x + ibl;
+        device const uint16_t * q2 = xr->qs + 4 * ib;
+        device const half * dh = &xr->d;
+
+        for (int row = 0; row < N_DST; row++) {
+
+            const float db = dh[0];
+            device const uint8_t * aux8 = (device const uint8_t *)q2;
+            const uint32_t aux32 = q2[2] | (q2[3] << 16);
+            const float d = db * (0.5f + (aux32 >> 28));
+
+            float sum = 0;
+            for (int l = 0; l < 4; ++l) {
+                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + aux8[l]);
+                const uint8_t signs = shared_signs[(aux32 >> 7*l) & 127];
+                for (int j = 0; j < 8; ++j) {
+                    sum += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+                }
+            }
+            sumf[row] += d * sum;
+
+            dh += nb*sizeof(block_iq2_xxs)/2;
+            q2 += nb*sizeof(block_iq2_xxs)/2;
+        }
+
+        y4 += 32 * 32;
+    }
+#else
+    // TODO
+#endif
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.25f;
+        }
+    }
+}
+
+[[host_name("kernel_mul_mv_iq2_xxs_f32")]]
+kernel void kernel_mul_mv_iq2_xxs_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_iq2_xxs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+}
+
+void kernel_mul_mv_iq2_xs_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int ib_row = first_row * nb;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    device const block_iq2_xs * x = (device const block_iq2_xs *) src0 + ib_row + offset0;
+    device const float        * y = (device const float        *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[32];
+    float sumf[N_DST]={0.f}, all_sum;
+
+    const int nb32 = nb * (QK_K / 32);
+
+    threadgroup uint64_t * values = (threadgroup uint64_t *)shared_values;
+    threadgroup uint8_t  * shared_signs = (threadgroup uint8_t *)(values + 512);
+    {
+        int nval = 8;
+        int pos  = (32*sgitg + tiisg)*nval;
+        for (int i = 0; i < nval; ++i) values[pos + i] = iq2xs_grid[pos + i];
+        nval = 2;
+        pos  = (32*sgitg + tiisg)*nval;
+        for (int i = 0; i < nval; ++i) shared_signs[pos+i] = ksigns_iq2xs[pos+i];
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+#if QK_K == 256
+    const int ix = tiisg;
+
+    device const float * y4 = y + 32 * ix;
+
+    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
+
+        for (int i = 0; i < 32; ++i) {
+            yl[i] = y4[i];
+        }
+
+        const int ibl = ib32 / (QK_K / 32);
+        const int ib  = ib32 % (QK_K / 32);
+
+        device const block_iq2_xs * xr = x + ibl;
+        device const uint16_t * q2 = xr->qs + 4 * ib;
+        device const uint8_t  * sc = xr->scales + ib;
+        device const half * dh = &xr->d;
+
+        for (int row = 0; row < N_DST; row++) {
+
+            const float db = dh[0];
+            const uint8_t ls1 = sc[0] & 0xf;
+            const uint8_t ls2 = sc[0] >>  4;
+            const float d1 = db * (0.5f + ls1);
+            const float d2 = db * (0.5f + ls2);
+
+            float sum1 = 0, sum2 = 0;
+            for (int l = 0; l < 2; ++l) {
+                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + (q2[l] & 511));
+                const uint8_t signs = shared_signs[(q2[l] >> 9)];
+                for (int j = 0; j < 8; ++j) {
+                    sum1 += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+                }
+            }
+            for (int l = 2; l < 4; ++l) {
+                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + (q2[l] & 511));
+                const uint8_t signs = shared_signs[(q2[l] >> 9)];
+                for (int j = 0; j < 8; ++j) {
+                    sum2 += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+                }
+            }
+            sumf[row] += d1 * sum1 + d2 * sum2;
+
+            dh += nb*sizeof(block_iq2_xs)/2;
+            q2 += nb*sizeof(block_iq2_xs)/2;
+            sc += nb*sizeof(block_iq2_xs);
+        }
+
+        y4 += 32 * 32;
+    }
+#else
+    // TODO
+#endif
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.25f;
+        }
+    }
+}
+
+[[host_name("kernel_mul_mv_iq2_xs_f32")]]
+kernel void kernel_mul_mv_iq2_xs_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_iq2_xs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+}
+
 //============================= templates and their specializations =============================
 
 // NOTE: this is not dequantizing - we are simply fitting the template
@@ -3620,8 +4122,8 @@ void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg
     uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4];
     int16_t  dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2)
                                : (scale_2&kmask2) | ((scale_1&kmask1) << 4);
-    half dl = il<8 ? d_all * (dl_int - 32.h) : d_all * (dl_int / 16.h - 32.h);
-    const half ml = 4.h * dl;
+    float dl = il<8 ? d_all * (dl_int - 32.f) : d_all * (dl_int / 16.f - 32.f);
+    const float ml = 4.f * dl;
 
     il = (il/2) & 3;
     const half    coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
@@ -3688,7 +4190,7 @@ void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg
     uint8_t ul = 1 << (il/2);
     il = il & 3;
     const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
-    const float d = il < 2 ? xb->d : xb->d / 16.h;
+    const float d = il < 2 ? xb->d : xb->d / 16.f;
     const float min = xb->dmin;
     const float dl = d * sc[0];
     const float ml = min * sc[1];
@@ -3721,17 +4223,17 @@ void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg
 #if QK_K == 256
     ql = ql + 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
     qh = qh + 32*(il/8) + 16*(il&1);
-    half sc = scales[(il%2) + 2 * ((il/2))];
+    float sc = scales[(il%2) + 2 * ((il/2))];
     il = (il/2) & 3;
 #else
     ql = ql + 16 * (il&1);
-    half sc = scales[il];
+    float sc = scales[il];
 #endif
     const uint16_t  kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
     const uint16_t  kmask2 = il>1 ? 0xF0              : 0x0F;
-    const half        coef = il>1 ? 1.f/16.h          : 1.h;
-    const half ml = d_all * sc * 32.h;
-    const half dl = d_all * sc * coef;
+    const float       coef = il>1 ? 1.f/16.f          : 1.f;
+    const float ml = d_all * sc * 32.f;
+    const float dl = d_all * sc * coef;
     for (int i = 0; i < 16; ++i) {
         const half q = il&1 ? ((ql[i] & kmask2) | ((qh[i] & kmask1) << 2))
                             : ((ql[i] & kmask2) | ((qh[i] & kmask1) << 4));
@@ -3739,6 +4241,52 @@ void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg
     }
 }
 
+template <typename type4x4>
+void dequantize_iq2_xxs(device const block_iq2_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    // each block of 32 needs 2 uint32_t's for the quants & scale, so 4 uint16_t's.
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const uint32_t aux32_g = q2[0] | (q2[1] << 16);
+    const uint32_t aux32_s = q2[2] | (q2[3] << 16);
+    thread const uint8_t * aux8 = (thread const uint8_t *)&aux32_g;
+    const float dl = d * (0.5f + (aux32_s >> 28)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
+    uint8_t signs = ksigns_iq2xs[(aux32_s >> 14*il) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
+    signs = ksigns_iq2xs[(aux32_s >> (14*il+7)) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_xs(device const block_iq2_xs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+0] & 511));
+    uint8_t signs = ksigns_iq2xs[q2[2*il+0] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+1] & 511));
+    signs = ksigns_iq2xs[q2[2*il+1] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
 kernel void kernel_get_rows(
         device const  void * src0,
@@ -4278,6 +4826,8 @@ template [[host_name("kernel_get_rows_q3_K")]] kernel get_rows_t kernel_get_rows
 template [[host_name("kernel_get_rows_q4_K")]] kernel get_rows_t kernel_get_rows<block_q4_K, QK_NL, dequantize_q4_K>;
 template [[host_name("kernel_get_rows_q5_K")]] kernel get_rows_t kernel_get_rows<block_q5_K, QK_NL, dequantize_q5_K>;
 template [[host_name("kernel_get_rows_q6_K")]] kernel get_rows_t kernel_get_rows<block_q6_K, QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_get_rows_iq2_xxs")]] kernel get_rows_t kernel_get_rows<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
+template [[host_name("kernel_get_rows_iq2_xs")]]  kernel get_rows_t kernel_get_rows<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
 
 //
 // matrix-matrix multiplication
@@ -4314,6 +4864,8 @@ template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_mm_t kernel_mul_mm<b
 template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_K, QK_NL, dequantize_q4_K>;
 template [[host_name("kernel_mul_mm_q5_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_K, QK_NL, dequantize_q5_K>;
 template [[host_name("kernel_mul_mm_q6_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q6_K, QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
+template [[host_name("kernel_mul_mm_iq2_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
 
 //
 // indirect matrix-matrix multiplication
@@ -4362,6 +4914,8 @@ template [[host_name("kernel_mul_mm_id_q3_K_f32")]] kernel mat_mm_id_t kernel_mu
 template [[host_name("kernel_mul_mm_id_q4_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q4_K, QK_NL, dequantize_q4_K>;
 template [[host_name("kernel_mul_mm_id_q5_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q5_K, QK_NL, dequantize_q5_K>;
 template [[host_name("kernel_mul_mm_id_q6_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q6_K, QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
+template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
 
 //
 // matrix-vector multiplication
@@ -5134,3 +5688,133 @@ kernel void kernel_mul_mv_id_q6_K_f32(
         tiisg,
         sgitg);
 }
+
+[[host_name("kernel_mul_mv_id_iq2_xxs_f32")]]
+kernel void kernel_mul_mv_id_iq2_xxs_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         float * dst,
+        constant    uint64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant    uint64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        threadgroup int8_t   * shared_values [[threadgroup(0)]],
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_iq2_xxs_f32_impl(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        dst + bid*ne0,
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        shared_values,
+        tgpig,
+        tiisg,
+        sgitg);
+}
+
+[[host_name("kernel_mul_mv_id_iq2_xs_f32")]]
+kernel void kernel_mul_mv_id_iq2_xs_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         float * dst,
+        constant    uint64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant    uint64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        threadgroup int8_t   * shared_values [[threadgroup(0)]],
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_iq2_xs_f32_impl(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        dst + bid*ne0,
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        shared_values,
+        tgpig,
+        tiisg,
+        sgitg);
+}

ggml-opencl.cpp

@@ -1,5 +1,6 @@
 #include "ggml.h"
 #include "ggml-opencl.h"
+#include "ggml-backend-impl.h"
 
 #include <array>
 #include <atomic>
@@ -10,7 +11,7 @@
 #include <sstream>
 #include <vector>
 
-#define CL_TARGET_OPENCL_VERSION 110
+#define CL_TARGET_OPENCL_VERSION 120
 #include <clblast.h>
 
 #if defined(_MSC_VER)
@@ -929,6 +930,12 @@ static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, co
 }
 
 void ggml_cl_init(void) {
+    static bool initialized = false;
+    if (initialized) {
+        return;
+    }
+    initialized = true;
+
     cl_int err;
 
     struct cl_device;
@@ -1483,8 +1490,8 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
     } else {
         d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size);
     }
-    cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
-    cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
+    cl_mem d_Y = src1->backend == GGML_BACKEND_GPU ? (cl_mem) src1->extra : ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
+    cl_mem d_D =  dst->backend == GGML_BACKEND_GPU ? (cl_mem)  dst->extra : ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
 
     size_t x_offset = 0;
 
@@ -1501,7 +1508,9 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
 
                 for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
                     // copy src1 to device
-                    CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
+                    if (src1->backend == GGML_BACKEND_CPU) {
+                        CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
+                    }
 
                     CL_CHECK(clFinish(queue));
 
@@ -1522,8 +1531,10 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
                     }
 
                     // copy dst to host
-                    float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
-                    CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
+                    if (dst->backend == GGML_BACKEND_CPU) {
+                        float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
+                        CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
+                    }
                 }
             }
         }
@@ -1532,8 +1543,12 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
     if (src0->backend != GGML_BACKEND_GPU) {
         ggml_cl_pool_free(d_X, x_size);
     }
-    ggml_cl_pool_free(d_Y, y_size);
-    ggml_cl_pool_free(d_D, d_size);
+    if (src1->backend != GGML_BACKEND_GPU) {
+        ggml_cl_pool_free(d_Y, y_size);
+    }
+    if (dst->backend != GGML_BACKEND_GPU) {
+        ggml_cl_pool_free(d_D, d_size);
+    }
 }
 
 static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t wsize) {
@@ -1598,6 +1613,8 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
                     CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
                 }
 
+                // FIXME: convert on device
+
                 for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
                     // convert src1 to fp16
                     // TODO: use multiple threads
@@ -1643,11 +1660,13 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
                     }
 
                     // copy dst to host, then convert to float
-                    CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
-
-                    float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
-
-                    ggml_fp16_to_fp32_row(tmp, d, d_ne);
+                    if (dst->backend == GGML_BACKEND_CPU) {
+                        CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
+                        float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
+                        ggml_fp16_to_fp32_row(tmp, d, d_ne);
+                    } else {
+                        // FIXME: convert dst to fp32 on device
+                    }
                 }
             }
         }
@@ -1801,7 +1820,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
 }
 
 
-bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
+bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, const struct ggml_tensor * dst) {
     const int64_t ne10 = src1->ne[0];
 
     const int64_t ne0 = dst->ne[0];
@@ -1895,3 +1914,291 @@ void ggml_cl_transform_tensor(void * data, ggml_tensor * tensor) {
     tensor->extra = dst;
     GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 }
+
+// ggml-backend
+
+// buffer
+
+struct ggml_backend_opencl_buffer_context {
+    ~ggml_backend_opencl_buffer_context() {
+        if (buffer) {
+            clReleaseMemObject(buffer);
+        }
+        for (auto * sub_buffer : sub_buffers) {
+            clReleaseMemObject(sub_buffer);
+        }
+    }
+
+    cl_mem buffer;
+    std::vector<cl_mem> sub_buffers;
+};
+
+static void * const cl_ptr_base = (void *)(uintptr_t) 0x1000;
+
+static const char * ggml_backend_opencl_buffer_get_name(ggml_backend_buffer_t buffer) {
+    return "OpenCL";
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_opencl_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+    delete ctx;
+}
+
+static void * ggml_backend_opencl_buffer_get_base(ggml_backend_buffer_t buffer) {
+    return cl_ptr_base;
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_opencl_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
+    if (tensor->view_src != NULL && tensor->view_offs == 0) {
+        tensor->extra = tensor->view_src->extra;
+    } else {
+        ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+        cl_buffer_region region = {(size_t)((char *)tensor->data - (char *)cl_ptr_base), ggml_nbytes(tensor)};
+        cl_int err;
+        cl_mem sub_buffer = clCreateSubBuffer(ctx->buffer, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        ctx->sub_buffers.push_back(sub_buffer);
+        tensor->extra = sub_buffer;
+    }
+    tensor->backend = GGML_BACKEND_GPU;
+}
+
+static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    cl_mem tensor_buffer = (cl_mem) tensor->extra;
+    CL_CHECK(clEnqueueWriteBuffer(queue, tensor_buffer, true, offset, size, data, 0, NULL, NULL));
+    CL_CHECK(clFinish(queue));
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    cl_mem tensor_buffer = (cl_mem) tensor->extra;
+    CL_CHECK(clEnqueueReadBuffer(queue, tensor_buffer, true, offset, size, data, 0, NULL, NULL));
+    CL_CHECK(clFinish(queue));
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_opencl_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+    CL_CHECK(clEnqueueFillBuffer(queue, ctx->buffer, &value, sizeof(value), 0, buffer->size, 0, NULL, NULL));
+    CL_CHECK(clFinish(queue));
+}
+
+static void ggml_backend_opencl_buffer_reset(ggml_backend_buffer_t buffer) {
+    ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+    for (auto * sub_buffer : ctx->sub_buffers) {
+        clReleaseMemObject(sub_buffer);
+    }
+    ctx->sub_buffers.clear();
+}
+
+static ggml_backend_buffer_i ggml_backend_opencl_buffer_interface = {
+    /* .get_name        = */ ggml_backend_opencl_buffer_get_name,
+    /* .free_buffer     = */ ggml_backend_opencl_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_opencl_buffer_get_base,
+    /* .init_tensor     = */ ggml_backend_opencl_buffer_init_tensor,
+    /* .set_tensor      = */ ggml_backend_opencl_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_opencl_buffer_get_tensor,
+    /* .cpy_tensor      = */ NULL,
+    /* .clear           = */ ggml_backend_opencl_buffer_clear,
+    /* .reset           = */ ggml_backend_opencl_buffer_reset,
+};
+
+// buffer type
+
+static const char * ggml_backend_opencl_buffer_type_name(ggml_backend_buffer_type_t buffer_type) {
+    return "OpenCL";
+
+    GGML_UNUSED(buffer_type);
+}
+
+static ggml_backend_buffer_t ggml_backend_opencl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buffer_type, size_t size) {
+    ggml_cl_init();
+
+    cl_int err;
+    cl_mem mem = clCreateBuffer(context, CL_MEM_READ_WRITE, size, NULL, &err);
+    if (err != CL_SUCCESS) {
+        fprintf(stderr, "%s: failed to allocate %.2f MiB\n", __func__, size / 1024.0 / 1024.0);
+        return nullptr;
+    }
+
+    ggml_backend_opencl_buffer_context * ctx = new ggml_backend_opencl_buffer_context{mem, {}};
+
+    return ggml_backend_buffer_init(buffer_type, ggml_backend_opencl_buffer_interface, ctx, size);
+}
+
+static size_t ggml_backend_opencl_buffer_type_get_alignment(ggml_backend_buffer_type_t buffer_type) {
+    // FIXME: not thread safe, device may not be initialized yet
+    static cl_uint alignment = -1;
+    if (alignment == (cl_uint)-1) {
+        ggml_cl_init();
+        clGetDeviceInfo(device, CL_DEVICE_MEM_BASE_ADDR_ALIGN, sizeof(cl_uint), &alignment, NULL);
+    }
+    return alignment;
+
+    GGML_UNUSED(buffer_type);
+}
+
+static bool ggml_backend_opencl_buffer_type_supports_backend(ggml_backend_buffer_type_t buffer_type, ggml_backend_t backend) {
+    //return ggml_backend_is_opencl(backend); // opencl must be used through the cpu backend
+    return ggml_backend_is_cpu(backend);
+
+    GGML_UNUSED(buffer_type);
+}
+
+static ggml_backend_buffer_type_i ggml_backend_opencl_buffer_type_interface = {
+    /* .get_name         = */ ggml_backend_opencl_buffer_type_name,
+    /* .alloc_buffer     = */ ggml_backend_opencl_buffer_type_alloc_buffer,
+    /* .get_alignment    = */ ggml_backend_opencl_buffer_type_get_alignment,
+    /* .get_alloc_size   = */ NULL,
+    /* .supports_backend = */ ggml_backend_opencl_buffer_type_supports_backend,
+    /* .is_host          = */ NULL,
+};
+
+
+ggml_backend_buffer_type_t ggml_backend_opencl_buffer_type() {
+    static ggml_backend_buffer_type buffer_type = {
+        /* .iface   = */ ggml_backend_opencl_buffer_type_interface,
+        /* .context = */ nullptr,
+    };
+
+    return &buffer_type;
+}
+
+#if 0
+// host buffer type
+
+static const char * ggml_backend_opencl_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
+    return "CL_Host";
+
+    GGML_UNUSED(buft);
+}
+
+static const char * ggml_backend_opencl_host_buffer_name(ggml_backend_buffer_t buffer) {
+    return "CL_Host";
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_opencl_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_cl_host_free(buffer->context);
+}
+
+static ggml_backend_buffer_t ggml_backend_opencl_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    void * ptr = ggml_cl_host_malloc(size);
+
+    if (ptr == nullptr) {
+        // fallback to cpu buffer
+        return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
+    buffer->buft = buft;
+    buffer->iface.get_name = ggml_backend_opencl_host_buffer_name;
+    buffer->iface.free_buffer = ggml_backend_opencl_host_buffer_free_buffer;
+
+    return buffer;
+}
+
+ggml_backend_buffer_type_t ggml_backend_opencl_host_buffer_type() {
+    static struct ggml_backend_buffer_type ggml_backend_opencl_buffer_type_host = {
+        /* .iface    = */ {
+            /* .get_name         = */ ggml_backend_opencl_host_buffer_type_name,
+            /* .alloc_buffer     = */ ggml_backend_opencl_host_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
+            /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
+            /* .supports_backend = */ ggml_backend_cpu_buffer_type()->iface.supports_backend,
+            /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
+        },
+        /* .context  = */ nullptr,
+    };
+
+    return &ggml_backend_opencl_buffer_type_host;
+}
+
+// backend
+
+static const char * ggml_backend_opencl_name(ggml_backend_t backend) {
+    return "OpenCL";
+
+    GGML_UNUSED(backend);
+}
+
+static void ggml_backend_opencl_free(ggml_backend_t backend) {
+    GGML_UNUSED(backend);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_opencl_get_default_buffer_type(ggml_backend_t backend) {
+    return ggml_backend_opencl_buffer_type();
+
+    GGML_UNUSED(backend);
+}
+
+static bool ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggml_cgraph * graph) {
+    for (int i = 0; i < graph->n_nodes; ++i) {
+        ggml_tensor * node = graph->nodes[i];
+        switch (node->op) {
+            case GGML_OP_MUL_MAT:
+                ggml_cl_mul_mat(node->src[0], node->src[1], node, nullptr, 0);
+                break;
+            case GGML_OP_MUL:
+                ggml_cl_mul(node->src[0], node->src[1], node);
+                break;
+            default:
+                GGML_ASSERT(false);
+        }
+    }
+
+    return true;
+
+    GGML_UNUSED(backend);
+}
+
+static bool ggml_backend_opencl_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
+    switch (op->op) {
+        case GGML_OP_MUL_MAT:
+            return ggml_cl_can_mul_mat(op->src[0], op->src[1], op);
+        case GGML_OP_MUL:
+            // return ggml_can_repeat_rows(op->src[1], op->src[0]);
+            return true;
+        default:
+            return false;
+    }
+
+    GGML_UNUSED(backend);
+}
+
+static ggml_backend_i opencl_backend_i = {
+    /* .get_name                = */ ggml_backend_opencl_name,
+    /* .free                    = */ ggml_backend_opencl_free,
+    /* .get_default_buffer_type = */ ggml_backend_opencl_get_default_buffer_type,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_from_async   = */ NULL,
+    /* .cpy_tensor_to_async     = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_opencl_graph_compute,
+    /* .supports_op             = */ ggml_backend_opencl_supports_op,
+};
+
+ggml_backend_t ggml_backend_opencl_init() {
+    ggml_backend_t backend = new ggml_backend {
+        /* .interface = */ opencl_backend_i,
+        /* .context   = */ nullptr
+    };
+
+    return backend;
+}
+
+bool ggml_backend_is_opencl(ggml_backend_t backend) {
+    return backend && backend->iface.get_name == ggml_backend_opencl_name;
+}
+#endif

ggml-opencl.h

@@ -1,6 +1,7 @@
 #pragma once
 
 #include "ggml.h"
+#include "ggml-backend.h"
 
 #ifdef  __cplusplus
 extern "C" {
@@ -9,17 +10,26 @@ extern "C" {
 GGML_API void ggml_cl_init(void);
 
 GGML_API void   ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API bool   ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+GGML_API bool   ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, const struct ggml_tensor * dst);
 GGML_API size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
 GGML_API void   ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
 
-GGML_API void * ggml_cl_host_malloc(size_t size);
-GGML_API void   ggml_cl_host_free(void * ptr);
+// GGML_API void * ggml_cl_host_malloc(size_t size);
+// GGML_API void   ggml_cl_host_free(void * ptr);
 
 GGML_API void ggml_cl_free_data(const struct ggml_tensor* tensor);
 
 GGML_API void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor);
 
+// backend API
+
+// GGML_API ggml_backend_t ggml_backend_opencl_init(void);
+
+// GGML_API bool ggml_backend_is_opencl(ggml_backend_t backend);
+
+GGML_API ggml_backend_buffer_type_t ggml_backend_opencl_buffer_type(void);
+// GGML_API ggml_backend_buffer_type_t ggml_backend_opencl_host_buffer_type(void);
+
 #ifdef  __cplusplus
 }
 #endif

ggml-quants.h

@@ -165,6 +165,22 @@ typedef struct {
 } block_q8_K;
 static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
 
+// (Almost) "true" 2-bit quantization.
+// Due to the need to use blocks as per ggml dsign, it ends up using
+// 2.0625 bpw because of the 16-bit scale for each block of 256.
+typedef struct {
+    ggml_fp16_t d;
+    uint16_t qs[QK_K/8];
+} block_iq2_xxs;
+static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding");
+
+// 2.3125 bpw quants
+typedef struct {
+    ggml_fp16_t d;
+    uint16_t qs[QK_K/8];
+    uint8_t  scales[QK_K/32];
+} block_iq2_xs;
+static_assert(sizeof(block_iq2_xs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
 
 // Quantization
 void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k);
@@ -209,6 +225,8 @@ void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int
 void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int k);
 void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k);
 void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k);
+void dequantize_row_iq2_xxs(const block_iq2_xxs * restrict x, float * restrict y, int k);
+void dequantize_row_iq2_xs (const block_iq2_xs  * 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);
@@ -222,3 +240,20 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx,
 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);
+void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_iq2_xs_q8_K (int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+
+//
+// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
+//
+size_t quantize_iq2_xxs(const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_iq2_xs (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_q2_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_q3_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_q4_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_q5_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_q6_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_q4_0   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_q4_1   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_q5_0   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_q5_1   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);

ggml.c

@@ -132,7 +132,7 @@ void ggml_print_backtrace(void) {
             "-ex", "bt -frame-info source-and-location",
             "-ex", "detach",
             "-ex", "quit",
-            NULL);
+            (char *) NULL);
     } else {
         waitpid(pid, NULL, 0);
     }
@@ -394,6 +394,12 @@ static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
 static void ggml_vec_dot_f32(const int n, float * restrict s, const float * restrict x, const float * restrict y);
 static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t * restrict x, ggml_fp16_t * restrict y);
 
+ggml_collect_imatrix_t g_imatrix_collect = NULL;
+
+void ggml_set_imatrix_collection(ggml_collect_imatrix_t imatrix_collect) {
+    g_imatrix_collect = imatrix_collect;
+}
+
 static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
     [GGML_TYPE_I8] = {
         .type_name                = "i8",
@@ -573,6 +579,28 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .vec_dot                  = ggml_vec_dot_q6_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
     },
+    [GGML_TYPE_IQ2_XXS] = {
+        .type_name                = "iq2_xxs",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_iq2_xxs),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq2_xxs,
+        .from_float               = NULL,
+        .from_float_reference     = NULL,
+        .vec_dot                  = ggml_vec_dot_iq2_xxs_q8_K,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+    },
+    [GGML_TYPE_IQ2_XS] = {
+        .type_name                = "iq2_xs",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_iq2_xs),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq2_xs,
+        .from_float               = NULL,
+        .from_float_reference     = NULL,
+        .vec_dot                  = ggml_vec_dot_iq2_xs_q8_K,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+    },
     [GGML_TYPE_Q8_K] = {
         .type_name                = "q8_K",
         .blck_size                = QK_K,
@@ -2111,6 +2139,8 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
         case GGML_FTYPE_MOSTLY_Q4_K:          wtype = GGML_TYPE_Q4_K;  break;
         case GGML_FTYPE_MOSTLY_Q5_K:          wtype = GGML_TYPE_Q5_K;  break;
         case GGML_FTYPE_MOSTLY_Q6_K:          wtype = GGML_TYPE_Q6_K;  break;
+        case GGML_FTYPE_MOSTLY_IQ2_XXS:       wtype = GGML_TYPE_IQ2_XXS;  break;
+        case GGML_FTYPE_MOSTLY_IQ2_XS:        wtype = GGML_TYPE_IQ2_XS;   break;
         case GGML_FTYPE_UNKNOWN:              wtype = GGML_TYPE_COUNT; break;
         case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break;
     }
@@ -2324,6 +2354,10 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
 }
 
 void ggml_free(struct ggml_context * ctx) {
+    if (ctx == NULL) {
+        return;
+    }
+
     // make this function thread safe
     ggml_critical_section_start();
 
@@ -4299,13 +4333,13 @@ struct ggml_tensor * ggml_set_2d_inplace(
 static struct ggml_tensor * ggml_cpy_impl(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        struct ggml_tensor  * b,
-        bool inplace) {
+        struct ggml_tensor  * b) {
     GGML_ASSERT(ggml_nelements(a) == ggml_nelements(b));
 
     bool is_node = false;
 
-    if (!inplace && (a->grad || b->grad)) {
+    if (a->grad || b->grad) {
+        // inplace is false and either one have a grad
         is_node = true;
     }
 
@@ -4329,29 +4363,38 @@ struct ggml_tensor * ggml_cpy(
         struct ggml_context * ctx,
         struct ggml_tensor * a,
         struct ggml_tensor * b) {
-    return ggml_cpy_impl(ctx, a, b, false);
+    return ggml_cpy_impl(ctx, a, b);
 }
 
-struct ggml_tensor * ggml_cpy_inplace(
+struct ggml_tensor * ggml_cast(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        struct ggml_tensor * b) {
-    return ggml_cpy_impl(ctx, a, b, true);
+        struct ggml_tensor  * a,
+        enum   ggml_type      type) {
+    bool is_node = false;
+
+    struct ggml_tensor * result = ggml_new_tensor(ctx, type, GGML_MAX_DIMS, a->ne);
+    ggml_format_name(result, "%s (copy)", a->name);
+
+    result->op   = GGML_OP_CPY;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = a;
+    result->src[1] = result;
+
+    return result;
 }
 
 // ggml_cont
 
 static struct ggml_tensor * ggml_cont_impl(
         struct ggml_context * ctx,
-        struct ggml_tensor  * a,
-        bool inplace) {
+        struct ggml_tensor  * a) {
     bool is_node = false;
 
-    if (!inplace && a->grad) {
+    if (a->grad) {
         is_node = true;
     }
 
-    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+    struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
     ggml_format_name(result, "%s (cont)", a->name);
 
     result->op   = GGML_OP_CONT;
@@ -4364,13 +4407,7 @@ static struct ggml_tensor * ggml_cont_impl(
 struct ggml_tensor * ggml_cont(
         struct ggml_context * ctx,
         struct ggml_tensor * a) {
-    return ggml_cont_impl(ctx, a, false);
-}
-
-struct ggml_tensor * ggml_cont_inplace(
-        struct ggml_context * ctx,
-        struct ggml_tensor * a) {
-    return ggml_cont_impl(ctx, a, true);
+    return ggml_cont_impl(ctx, a);
 }
 
 // make contiguous, with new shape
@@ -7436,6 +7473,8 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
             {
                 ggml_compute_forward_add_q_f32(params, src0, src1, dst);
             } break;
@@ -7700,6 +7739,8 @@ static void ggml_compute_forward_add1(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
             {
                 ggml_compute_forward_add1_q_f32(params, src0, src1, dst);
             } break;
@@ -7814,6 +7855,8 @@ static void ggml_compute_forward_acc(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
         default:
             {
                 GGML_ASSERT(false);
@@ -9747,6 +9790,10 @@ static void ggml_compute_forward_mul_mat(
     const int ith = params->ith;
     const int nth = params->nth;
 
+    if (ith == 1 && g_imatrix_collect) {
+        g_imatrix_collect(src0, src1);
+    }
+
     const enum ggml_type type = src0->type;
 
     const bool src1_cont = ggml_is_contiguous(src1);
@@ -10050,6 +10097,10 @@ static void ggml_compute_forward_mul_mat_id(
 
         const struct ggml_tensor * src0_cur = dst->src[cur_a + 2];
 
+        if (ith == 1 && g_imatrix_collect) {
+            g_imatrix_collect(src0_cur, src1);
+        }
+
         const void * wdata    = (src1->type == vec_dot_type) ? src1->data : params->wdata;
         const size_t row_size = ggml_row_size(vec_dot_type, ne10);
 
@@ -10455,6 +10506,8 @@ static void ggml_compute_forward_out_prod(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
             {
                 ggml_compute_forward_out_prod_q_f32(params, src0, src1, dst);
             } break;
@@ -10629,6 +10682,8 @@ static void ggml_compute_forward_set(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
         default:
             {
                 GGML_ASSERT(false);
@@ -10823,6 +10878,8 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
             {
                 ggml_compute_forward_get_rows_q(params, src0, src1, dst);
             } break;
@@ -11459,6 +11516,8 @@ static void ggml_compute_forward_alibi(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_Q8_K:
         case GGML_TYPE_I8:
         case GGML_TYPE_I16:
@@ -11533,6 +11592,8 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_Q8_K:
         case GGML_TYPE_I8:
         case GGML_TYPE_I16:
@@ -11577,6 +11638,21 @@ static float ggml_rope_yarn_corr_dim(int n_dims, int n_orig_ctx, float n_rot, fl
     return n_dims * logf(n_orig_ctx / (n_rot * 2 * (float)M_PI)) / (2 * logf(base));
 }
 
+static void ggml_rope_cache_init(
+     float theta_base, float freq_scale, float corr_dims[2], int64_t ne0, float ext_factor, float mscale,
+     float * cache, float sin_sign, float theta_scale
+) {
+    float theta = theta_base;
+    for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
+        rope_yarn(
+            theta, freq_scale, corr_dims, i0, ext_factor, mscale, &cache[i0 + 0], &cache[i0 + 1]
+        );
+        cache[i0 + 1] *= sin_sign;
+
+        theta *= theta_scale;
+    }
+}
+
 void ggml_rope_yarn_corr_dims(
     int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]
 ) {
@@ -11659,6 +11735,12 @@ static void ggml_compute_forward_rope_f32(
     for (int64_t i3 = 0; i3 < ne3; i3++) {
         for (int64_t i2 = 0; i2 < ne2; i2++) {
             const int64_t p = pos[i2];
+
+            float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith;
+            if (!is_glm && !is_neox) { // TODO: cache sin/cos for glm, neox
+                ggml_rope_cache_init(p, freq_scale, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+            }
+
             for (int64_t i1 = 0; i1 < ne1; i1++) {
                 if (ir++ < ir0) continue;
                 if (ir   > ir1) break;
@@ -11692,18 +11774,13 @@ static void ggml_compute_forward_rope_f32(
                     }
                 } else if (!is_neox) {
                     for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
-                        float cos_theta, sin_theta;
-                        rope_yarn(
-                            theta_base, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta
-                        );
-                        sin_theta *= sin_sign;
+                        const float cos_theta = cache[i0 + 0];
+                        const float sin_theta = cache[i0 + 1];
 
                         // zeta scaling for xPos only:
                         float zeta = xpos_base != 0.0f ? powf((i0 + 0.4f * ne0) / (1.4f * ne0), p / xpos_base) : 1.0f;
                         if (xpos_down) zeta = 1.0f / zeta;
 
-                        theta_base *= theta_scale;
-
                         const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
                               float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
 
@@ -11827,6 +11904,12 @@ static void ggml_compute_forward_rope_f16(
     for (int64_t i3 = 0; i3 < ne3; i3++) {
         for (int64_t i2 = 0; i2 < ne2; i2++) {
             const int64_t p = pos[i2];
+
+            float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith;
+            if (!is_glm && !is_neox) { // TODO: cache sin/cos for glm, neox
+                ggml_rope_cache_init(p, freq_scale, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+            }
+
             for (int64_t i1 = 0; i1 < ne1; i1++) {
                 if (ir++ < ir0) continue;
                 if (ir   > ir1) break;
@@ -11860,13 +11943,8 @@ static void ggml_compute_forward_rope_f16(
                     }
                 } else if (!is_neox) {
                     for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
-                        float cos_theta, sin_theta;
-                        rope_yarn(
-                            theta_base, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta
-                        );
-                        sin_theta *= sin_sign;
-
-                        theta_base *= theta_scale;
+                        const float cos_theta = cache[i0 + 0];
+                        const float sin_theta = cache[i0 + 1];
 
                         const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
                               ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
@@ -14831,7 +14909,7 @@ size_t ggml_hash_find_or_insert(struct ggml_hash_set hash_set, struct ggml_tenso
     return i;
 }
 
-static struct ggml_hash_set ggml_hash_set_new(size_t size) {
+struct ggml_hash_set ggml_hash_set_new(size_t size) {
     size = ggml_hash_size(size);
     struct ggml_hash_set result;
     result.size = size;
@@ -16580,7 +16658,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
     return GGML_EXIT_SUCCESS;
 }
 
-struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
+struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threads) {
     if (n_threads <= 0) {
         n_threads = GGML_DEFAULT_N_THREADS;
     }
@@ -16642,14 +16720,15 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                 } break;
             case GGML_OP_MUL_MAT_ID:
                 {
+                    cur = 0;
                     const struct ggml_tensor * src0 = node->src[2];
                     const struct ggml_tensor * src1 = node->src[1];
                     const enum ggml_type vec_dot_type = type_traits[src0->type].vec_dot_type;
                     if (src1->type != vec_dot_type) {
-                        cur = ggml_row_size(vec_dot_type, ggml_nelements(src1));
+                        cur += ggml_row_size(vec_dot_type, ggml_nelements(src1));
                     }
                     const int n_as = ggml_get_op_params_i32(node, 1);
-                    cur = GGML_PAD(cur, sizeof(int64_t));        // align
+                    cur += GGML_PAD(cur, sizeof(int64_t));       // align
                     cur += n_as * sizeof(int64_t);               // matrix_row_counts
                     cur += n_as * src1->ne[1] * sizeof(int64_t); // matrix_rows
                 } break;
@@ -16660,6 +16739,7 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                     }
                 } break;
             case GGML_OP_SOFT_MAX:
+            case GGML_OP_ROPE:
                 {
                     cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks;
                 } break;
@@ -18585,32 +18665,47 @@ size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t *
     return (n/QK8_0*sizeof(block_q8_0));
 }
 
-size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist) {
+size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start,
+        int nrows, int n_per_row, int64_t * hist, const float * imatrix) {
+    (void)imatrix;
     size_t result = 0;
+    int n = nrows * n_per_row;
     switch (type) {
         case GGML_TYPE_Q4_0:
             {
                 GGML_ASSERT(start % QK4_0 == 0);
-                block_q4_0 * block = (block_q4_0*)dst + start / QK4_0;
-                result = ggml_quantize_q4_0(src + start, block, n, n, hist);
+                GGML_ASSERT(start % n_per_row == 0);
+                size_t start_row = start / n_per_row;
+                size_t row_size = ggml_row_size(type, n_per_row);
+                result = quantize_q4_0(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
+                GGML_ASSERT(result == row_size * nrows);
             } break;
         case GGML_TYPE_Q4_1:
             {
                 GGML_ASSERT(start % QK4_1 == 0);
-                block_q4_1 * block = (block_q4_1*)dst + start / QK4_1;
-                result = ggml_quantize_q4_1(src + start, block, n, n, hist);
+                GGML_ASSERT(start % n_per_row == 0);
+                size_t start_row = start / n_per_row;
+                size_t row_size = ggml_row_size(type, n_per_row);
+                result = quantize_q4_1(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
+                GGML_ASSERT(result == row_size * nrows);
             } break;
         case GGML_TYPE_Q5_0:
             {
                 GGML_ASSERT(start % QK5_0 == 0);
-                block_q5_0 * block = (block_q5_0*)dst + start / QK5_0;
-                result = ggml_quantize_q5_0(src + start, block, n, n, hist);
+                GGML_ASSERT(start % n_per_row == 0);
+                size_t start_row = start / n_per_row;
+                size_t row_size = ggml_row_size(type, n_per_row);
+                result = quantize_q5_0(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
+                GGML_ASSERT(result == row_size * nrows);
             } break;
         case GGML_TYPE_Q5_1:
             {
                 GGML_ASSERT(start % QK5_1 == 0);
-                block_q5_1 * block = (block_q5_1*)dst + start / QK5_1;
-                result = ggml_quantize_q5_1(src + start, block, n, n, hist);
+                GGML_ASSERT(start % n_per_row == 0);
+                size_t start_row = start / n_per_row;
+                size_t row_size = ggml_row_size(type, n_per_row);
+                result = quantize_q5_1(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
+                GGML_ASSERT(result == row_size * nrows);
             } break;
         case GGML_TYPE_Q8_0:
             {
@@ -18621,32 +18716,67 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
         case GGML_TYPE_Q2_K:
             {
                 GGML_ASSERT(start % QK_K == 0);
-                block_q2_K * block = (block_q2_K*)dst + start / QK_K;
-                result = ggml_quantize_q2_K(src + start, block, n, n, hist);
+                GGML_ASSERT(start % n_per_row == 0);
+                size_t start_row = start / n_per_row;
+                size_t row_size = ggml_row_size(type, n_per_row);
+                result = quantize_q2_K(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
+                GGML_ASSERT(result == row_size * nrows);
             } break;
         case GGML_TYPE_Q3_K:
             {
                 GGML_ASSERT(start % QK_K == 0);
-                block_q3_K * block = (block_q3_K*)dst + start / QK_K;
-                result = ggml_quantize_q3_K(src + start, block, n, n, hist);
+                GGML_ASSERT(start % n_per_row == 0);
+                size_t start_row = start / n_per_row;
+                size_t row_size = ggml_row_size(type, n_per_row);
+                result = quantize_q3_K(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
+                GGML_ASSERT(result == row_size * nrows);
             } break;
         case GGML_TYPE_Q4_K:
             {
                 GGML_ASSERT(start % QK_K == 0);
-                block_q4_K * block = (block_q4_K*)dst + start / QK_K;
-                result = ggml_quantize_q4_K(src + start, block, n, n, hist);
+                GGML_ASSERT(start % n_per_row == 0);
+                size_t start_row = start / n_per_row;
+                size_t row_size = ggml_row_size(type, n_per_row);
+                result = quantize_q4_K(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
+                GGML_ASSERT(result == row_size * nrows);
             } break;
         case GGML_TYPE_Q5_K:
             {
                 GGML_ASSERT(start % QK_K == 0);
-                block_q5_K * block = (block_q5_K*)dst + start / QK_K;
-                result = ggml_quantize_q5_K(src + start, block, n, n, hist);
+                GGML_ASSERT(start % n_per_row == 0);
+                size_t start_row = start / n_per_row;
+                size_t row_size = ggml_row_size(type, n_per_row);
+                result = quantize_q5_K(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
+                GGML_ASSERT(result == row_size * nrows);
             } break;
         case GGML_TYPE_Q6_K:
             {
                 GGML_ASSERT(start % QK_K == 0);
-                block_q6_K * block = (block_q6_K*)dst + start / QK_K;
-                result = ggml_quantize_q6_K(src + start, block, n, n, hist);
+                GGML_ASSERT(start % n_per_row == 0);
+                size_t start_row = start / n_per_row;
+                size_t row_size = ggml_row_size(type, n_per_row);
+                result = quantize_q6_K(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
+                GGML_ASSERT(result == row_size * nrows);
+            } break;
+        case GGML_TYPE_IQ2_XXS:
+            {
+                GGML_ASSERT(start % QK_K == 0);
+                GGML_ASSERT(start % n_per_row == 0);
+                GGML_ASSERT(imatrix);
+                size_t start_row = start / n_per_row;
+                size_t row_size = ggml_row_size(type, n_per_row);
+                result = quantize_iq2_xxs(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
+                GGML_ASSERT(result == row_size * nrows);
+            } break;
+        case GGML_TYPE_IQ2_XS:
+            {
+                GGML_ASSERT(start % QK_K == 0);
+                GGML_ASSERT(start % n_per_row == 0);
+                GGML_ASSERT(imatrix);
+                size_t start_row = start / n_per_row;
+                size_t row_size = ggml_row_size(type, n_per_row);
+                result = quantize_iq2_xs(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
+                GGML_ASSERT(result == row_size * nrows);
             } break;
         case GGML_TYPE_F16:
             {
@@ -19003,8 +19133,8 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
                 (int64_t) info->ne[3];
 
             if (ne % ggml_blck_size(info->type) != 0) {
-                fprintf(stderr, "%s: tensor '%s' number of elements (%" PRId64 ") is not a multiple of block size (%d)\n",
-                        __func__, info->name.data, ne, ggml_blck_size(info->type));
+                fprintf(stderr, "%s: tensor '%s' of type %d (%s) number of elements (%" PRId64 ") is not a multiple of block size (%d)\n",
+                        __func__, info->name.data, (int)info->type, ggml_type_name(info->type), ne, ggml_blck_size(info->type));
                 fclose(file);
                 gguf_free(ctx);
                 return NULL;
@@ -19110,7 +19240,7 @@ void gguf_free(struct gguf_context * ctx) {
 
     if (ctx->kv) {
         // free string memory - not great..
-        for (uint32_t i = 0; i < ctx->header.n_kv; ++i) {
+        for (uint64_t i = 0; i < ctx->header.n_kv; ++i) {
             struct gguf_kv * kv = &ctx->kv[i];
 
             if (kv->key.data) {
@@ -19126,7 +19256,7 @@ void gguf_free(struct gguf_context * ctx) {
             if (kv->type == GGUF_TYPE_ARRAY) {
                 if (kv->value.arr.data) {
                     if (kv->value.arr.type == GGUF_TYPE_STRING) {
-                        for (uint32_t j = 0; j < kv->value.arr.n; ++j) {
+                        for (uint64_t j = 0; j < kv->value.arr.n; ++j) {
                             struct gguf_str * str = &((struct gguf_str *) kv->value.arr.data)[j];
                             if (str->data) {
                                 free(str->data);
@@ -19142,7 +19272,7 @@ void gguf_free(struct gguf_context * ctx) {
     }
 
     if (ctx->infos) {
-        for (uint32_t i = 0; i < ctx->header.n_tensors; ++i) {
+        for (uint64_t i = 0; i < ctx->header.n_tensors; ++i) {
             struct gguf_tensor_info * info = &ctx->infos[i];
 
             if (info->name.data) {

ggml.h

@@ -218,7 +218,9 @@
 #define GGML_MAX_PARAMS         2048
 #define GGML_MAX_CONTEXTS       64
 #define GGML_MAX_SRC            10
+#ifndef GGML_MAX_NAME
 #define GGML_MAX_NAME           64
+#endif
 #define GGML_MAX_OP_PARAMS      64
 #define GGML_DEFAULT_N_THREADS  4
 #define GGML_DEFAULT_GRAPH_SIZE 2048
@@ -339,6 +341,8 @@ extern "C" {
         GGML_TYPE_Q5_K = 13,
         GGML_TYPE_Q6_K = 14,
         GGML_TYPE_Q8_K = 15,
+        GGML_TYPE_IQ2_XXS = 16,
+        GGML_TYPE_IQ2_XS  = 17,
         GGML_TYPE_I8,
         GGML_TYPE_I16,
         GGML_TYPE_I32,
@@ -373,6 +377,8 @@ extern "C" {
         GGML_FTYPE_MOSTLY_Q4_K = 12, // except 1d tensors
         GGML_FTYPE_MOSTLY_Q5_K = 13, // except 1d tensors
         GGML_FTYPE_MOSTLY_Q6_K = 14, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ2_XXS = 15, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ2_XS  = 16, // except 1d tensors
     };
 
     // available tensor operations:
@@ -1159,22 +1165,16 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
-    // a -> b, in-place, return view(b)
-    GGML_API struct ggml_tensor * ggml_cpy_inplace(
+    GGML_API struct ggml_tensor * ggml_cast(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
+            enum   ggml_type      type);
 
     // make contiguous
     GGML_API struct ggml_tensor * ggml_cont(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
-    // make contiguous, in-place
-    GGML_API struct ggml_tensor * ggml_cont_inplace(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a);
-
     // make contiguous, with new shape
     GGML_API struct ggml_tensor * ggml_cont_1d(
             struct ggml_context * ctx,
@@ -1847,8 +1847,8 @@ extern "C" {
 
     // ggml_graph_plan() has to be called before ggml_graph_compute()
     // when plan.work_size > 0, caller must allocate memory for plan.work_data
-    GGML_API struct ggml_cplan ggml_graph_plan   (struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/);
-    GGML_API int               ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
+    GGML_API struct ggml_cplan ggml_graph_plan   (const struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/);
+    GGML_API int               ggml_graph_compute(      struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
 
     // same as ggml_graph_compute() but the work data is allocated as a part of the context
     // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
@@ -2068,7 +2068,18 @@ extern "C" {
     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);
+    GGML_API size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst,
+            int start, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+
+    // These are needed for IQ2_XS and IQ2_XXS quantizations
+    GGML_API void ggml_init_iq2_quantization(enum ggml_type type);
+    GGML_API void ggml_deinit_iq2_quantization(enum ggml_type type);
+
+    //
+    // Importance matrix
+    //
+    typedef void(*ggml_collect_imatrix_t)(const struct ggml_tensor * src0, const struct ggml_tensor * src1);
+    GGML_API void ggml_set_imatrix_collection(ggml_collect_imatrix_t imatrix_collect);
 
     //
     // gguf

gguf-py/gguf/constants.py

@@ -389,6 +389,9 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.OUTPUT,
         MODEL_TENSOR.ATTN_NORM,
         MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
         MODEL_TENSOR.ATTN_OUT,
         MODEL_TENSOR.FFN_NORM,
         MODEL_TENSOR.FFN_DOWN,

gguf-py/gguf/tensor_mapping.py

@@ -57,6 +57,7 @@ class TensorNameMap:
             "transformer.norm_f",                      # mpt
             "ln_f",                                    # refact bloom qwen gpt2
             "language_model.encoder.final_layernorm",  # persimmon
+            "model.final_layernorm",                   # persimmon
             "lm_head.ln",                              # phi2
         ),
 
@@ -98,6 +99,7 @@ class TensorNameMap:
             "transformer.h.{bid}.self_attention.query_key_value",                  # falcon
             "h.{bid}.self_attention.query_key_value",                              # bloom
             "language_model.encoder.layers.{bid}.self_attention.query_key_value",  # persimmon
+            "model.layers.{bid}.self_attn.query_key_value",                        # persimmon
             "h.{bid}.attn.c_attn",                                                 # gpt2
             "transformer.h.{bid}.mixer.Wqkv",                                      # phi2
         ),
@@ -141,6 +143,7 @@ class TensorNameMap:
             "encoder.layer.{bid}.attention.output.dense",                # bert
             "transformer.h.{bid}.attn.out_proj",                         # gpt-j
             "language_model.encoder.layers.{bid}.self_attention.dense",  # persimmon
+            "model.layers.{bid}.self_attn.dense",                        # persimmon
             "h.{bid}.attn.c_proj",                                       # gpt2
             "transformer.h.{bid}.mixer.out_proj",                        # phi2
             "model.layers.layers.{bid}.self_attn.o_proj",                # plamo
@@ -184,9 +187,11 @@ class TensorNameMap:
             "encoder.layer.{bid}.intermediate.dense",                 # bert
             "transformer.h.{bid}.mlp.fc_in",                          # gpt-j
             "language_model.encoder.layers.{bid}.mlp.dense_h_to_4h",  # persimmon
+            "model.layers.{bid}.mlp.dense_h_to_4h",                   # persimmon
             "transformer.h.{bid}.mlp.w1",                             # qwen
             "h.{bid}.mlp.c_fc",                                       # gpt2
             "transformer.h.{bid}.mlp.fc1",                            # phi2
+            "model.layers.{bid}.mlp.fc1",                             # phi2
             "model.layers.layers.{bid}.mlp.up_proj",                  # plamo
         ),
 
@@ -225,8 +230,10 @@ class TensorNameMap:
             "encoder.layer.{bid}.output.dense",                       # bert
             "transformer.h.{bid}.mlp.fc_out",                         # gpt-j
             "language_model.encoder.layers.{bid}.mlp.dense_4h_to_h",  # persimmon
+            "model.layers.{bid}.mlp.dense_4h_to_h",                   # persimmon
             "h.{bid}.mlp.c_proj",                                     # gpt2
             "transformer.h.{bid}.mlp.fc2",                            # phi2
+            "model.layers.{bid}.mlp.fc2",                             # phi2
             "model.layers.layers.{bid}.mlp.down_proj",                # plamo
         ),
 
@@ -237,10 +244,12 @@ class TensorNameMap:
 
         MODEL_TENSOR.ATTN_Q_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.q_layernorm",
+            "model.layers.{bid}.self_attn.q_layernorm",                       # persimmon
         ),
 
         MODEL_TENSOR.ATTN_K_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.k_layernorm",
+            "model.layers.{bid}.self_attn.k_layernorm",                       # persimmon
         ),
 
         MODEL_TENSOR.ROPE_FREQS: (

llama.h

@@ -43,7 +43,7 @@
 #define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
 
 #define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
-#define LLAMA_SESSION_VERSION 3
+#define LLAMA_SESSION_VERSION 4
 
 #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
 // Defined when llama.cpp is compiled with support for offloading model layers to GPU.
@@ -103,6 +103,9 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_Q5_K_S        = 16, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q5_K_M        = 17, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q6_K          = 18, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ2_XXS       = 19, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ2_XS        = 20, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q2_K_S        = 21, // except 1d tensors
 
         LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };
@@ -115,6 +118,12 @@ extern "C" {
         LLAMA_ROPE_SCALING_MAX_VALUE   = LLAMA_ROPE_SCALING_YARN,
     };
 
+    enum llama_split_mode {
+        LLAMA_SPLIT_NONE    = 0, // single GPU
+        LLAMA_SPLIT_LAYER   = 1, // split layers and KV across GPUs
+        LLAMA_SPLIT_ROW     = 2, // split rows across GPUs
+    };
+
     typedef struct llama_token_data {
         llama_token id; // token id
         float logit;    // log-odds of the token
@@ -177,8 +186,16 @@ extern "C" {
 
     struct llama_model_params {
         int32_t n_gpu_layers; // number of layers to store in VRAM
-        int32_t main_gpu;     // the GPU that is used for scratch and small tensors
-        const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES)
+        enum llama_split_mode split_mode; // how to split the model across multiple GPUs
+
+        // main_gpu interpretation depends on split_mode:
+        // LLAMA_SPLIT_NONE: the GPU that is used for the entire model
+        // LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results
+        // LLAMA_SPLIT_LAYER: ignored
+        int32_t main_gpu;
+
+        // proportion of the model (layers or rows) to offload to each GPU, size: LLAMA_MAX_DEVICES
+        const float * tensor_split;
 
         // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
         // If the provided progress_callback returns true, model loading continues.
@@ -232,6 +249,7 @@ extern "C" {
         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
+        void * imatrix;              // pointer to importance matrix data
     } llama_model_quantize_params;
 
     // grammar types
@@ -484,6 +502,17 @@ extern "C" {
                        llama_pos   p1,
                        llama_pos   delta);
 
+    // Integer division of the positions by factor of `d > 1`
+    // If the KV cache is RoPEd, the KV data is updated accordingly
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_kv_cache_seq_div(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1,
+                             int   d);
+
     //
     // State / sessions
     //
@@ -685,14 +714,21 @@ extern "C" {
                            float   penalty_present);
 
     /// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806
-    /// @param candidates A vector of `llama_token_data` containing the candidate tokens, the logits must be directly extracted from the original generation context without being sorted.
-    /// @params guidance_ctx A separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context.
-    /// @params scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance.
-    LLAMA_API void llama_sample_classifier_free_guidance(
+    /// @param logits Logits extracted from the original generation context.
+    /// @param logits_guidance Logits extracted from a separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context.
+    /// @param scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance.
+    LLAMA_API void llama_sample_apply_guidance(
+              struct llama_context * ctx,
+                             float * logits,
+                             float * logits_guidance,
+                             float   scale);
+
+    LLAMA_API DEPRECATED(void llama_sample_classifier_free_guidance(
               struct llama_context * ctx,
             llama_token_data_array * candidates,
               struct llama_context * guidance_ctx,
-                             float   scale);
+                             float   scale),
+              "use llama_sample_apply_guidance() instead");
 
     /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
     LLAMA_API void llama_sample_softmax(

scripts/compare-llama-bench.py

@@ -0,0 +1,374 @@
+#!/usr/bin/env python3
+
+import argparse
+import heapq
+import sys
+import os
+from glob import glob
+import sqlite3
+
+try:
+    import git
+    from tabulate import tabulate
+except ImportError as e:
+    print("ERROR: the following Python libraries are required: GitPython, tabulate.")
+    raise e
+
+# Properties by which to differentiate results per commit:
+KEY_PROPERTIES = [
+    "cpu_info", "gpu_info", "n_gpu_layers", "main_gpu", "cuda", "opencl", "metal", "gpu_blas",
+    "blas", "model_filename", "model_type", "model_size", "model_n_params", "n_batch", "n_threads",
+    "type_k", "type_v", "no_kv_offload", "mul_mat_q", "tensor_split", "n_prompt", "n_gen"
+]
+
+# Properties that are boolean and are converted to Yes/No for the table:
+BOOL_PROPERTIES = ["cuda", "opencl", "metal", "gpu_blas", "blas"]
+
+# Header names for the table:
+PRETTY_NAMES = {
+    "cuda": "CUDA", "opencl": "OpenCL", "metal": "Metal", "gpu_blas": "GPU BLAS", "blas": "BLAS",
+    "cpu_info": "CPU", "gpu_info": "GPU", "model_filename": "File", "model_type": "Model",
+    "model_size": "Model Size [GiB]", "model_n_params": "Num. of Parameters",
+    "n_batch": "Batch size", "n_threads": "Threads", "type_k": "K type", "type_v": "V type",
+    "n_gpu_layers": "GPU layers", "main_gpu": "Main GPU", "no_kv_offload": "NKVO",
+    "mul_mat_q": "MMQ", "tensor_split": "Tensor split"
+}
+
+DEFAULT_SHOW = ["model_type"]  # Always show these properties by default.
+DEFAULT_HIDE = ["model_filename"]  # Always hide these properties by default.
+GPU_NAME_STRIP = ["NVIDIA GeForce ", "Tesla ", "AMD Radeon "]  # Strip prefixes for smaller tables.
+MODEL_SUFFIX_REPLACE = {" - Small": "_S", " - Medium": "_M", " - Large": "_L"}
+
+DESCRIPTION = """Creates tables from llama-bench data written to an SQLite database. Example usage (Linux):
+
+$ git checkout master
+$ make clean && make llama-bench
+$ ./llama-bench -o sql | sqlite3 llama-bench.sqlite
+$ git checkout some_branch
+$ make clean && make llama-bench
+$ ./llama-bench -o sql | sqlite3 llama-bench.sqlite
+$ ./scripts/compare-llama-bench.py
+
+Performance numbers from multiple runs per commit are averaged WITHOUT being weighted by the --repetitions parameter of llama-bench.
+"""
+
+parser = argparse.ArgumentParser(
+    description=DESCRIPTION, formatter_class=argparse.RawDescriptionHelpFormatter)
+help_b = (
+    "The baseline commit to compare performance to. "
+    "Accepts either a branch name, tag name, or commit hash. "
+    "Defaults to latest master commit with data."
+)
+parser.add_argument("-b", "--baseline", help=help_b)
+help_c = (
+    "The commit whose performance is to be compared to the baseline. "
+    "Accepts either a branch name, tag name, or commit hash. "
+    "Defaults to the non-master commit for which llama-bench was run most recently."
+)
+parser.add_argument("-c", "--compare", help=help_c)
+help_i = (
+    "Input SQLite file for comparing commits. "
+    "Defaults to 'llama-bench.sqlite' in the current working directory. "
+    "If no such file is found and there is exactly one .sqlite file in the current directory, "
+    "that file is instead used as input."
+)
+parser.add_argument("-i", "--input", help=help_i)
+help_o = (
+    "Output format for the table. "
+    "Defaults to 'pipe' (GitHub compatible). "
+    "Also supports e.g. 'latex' or 'mediawiki'. "
+    "See tabulate documentation for full list."
+)
+parser.add_argument("-o", "--output", help=help_o, default="pipe")
+help_s = (
+    "Columns to add to the table. "
+    "Accepts a comma-separated list of values. "
+    f"Legal values: {', '.join(KEY_PROPERTIES[:-2])}. "
+    "Defaults to model name (model_type) and CPU and/or GPU name (cpu_info, gpu_info) "
+    "plus any column where not all data points are the same. "
+    "If the columns are manually specified, then the results for each unique combination of the "
+    "specified values are averaged WITHOUT weighing by the --repetitions parameter of llama-bench."
+)
+parser.add_argument("-s", "--show", help=help_s)
+
+known_args, unknown_args = parser.parse_known_args()
+
+if unknown_args:
+    print(f"ERROR: Received unknown args: {unknown_args}.")
+    print()
+    parser.print_help()
+    sys.exit(1)
+
+input_file = known_args.input
+if input_file is None and os.path.exists("./llama-bench.sqlite"):
+    input_file = "llama-bench.sqlite"
+if input_file is None:
+    sqlite_files = glob("*.sqlite")
+    if len(sqlite_files) == 1:
+        input_file = sqlite_files[0]
+
+if input_file is None:
+    print("ERROR: Cannot find a suitable input file, please provide one.")
+    print()
+    parser.print_help()
+    sys.exit(1)
+
+connection = sqlite3.connect(input_file)
+cursor = connection.cursor()
+builds = cursor.execute("SELECT DISTINCT build_commit FROM test;").fetchall()
+
+try:
+    repo = git.Repo(".", search_parent_directories=True)
+except git.exc.InvalidGitRepositoryError:
+    repo = None
+
+
+def find_parent_in_data(commit):
+    """Helper function to find the most recent parent measured in number of commits for which there is data."""
+    heap = [(0, commit)]
+    seen_hexsha8 = set()
+    while heap:
+        depth, current_commit = heapq.heappop(heap)
+        current_hexsha8 = commit.hexsha[:8]
+        if (current_hexsha8,) in builds:
+            return current_hexsha8
+        for parent in commit.parents:
+            parent_hexsha8 = parent.hexsha[:8]
+            if parent_hexsha8 not in seen_hexsha8:
+                seen_hexsha8.add(parent_hexsha8)
+                heapq.heappush(heap, (depth + 1, parent))
+    return None
+
+
+def get_all_parent_hexsha8s(commit):
+    """Helper function to recursively get hexsha8 values for all parents of a commit."""
+    unvisited = [commit]
+    visited   = []
+
+    while unvisited:
+        current_commit = unvisited.pop(0)
+        visited.append(current_commit.hexsha[:8])
+        for parent in current_commit.parents:
+            if parent.hexsha[:8] not in visited:
+                unvisited.append(parent)
+
+    return visited
+
+
+def get_commit_name(hexsha8):
+    """Helper function to find a human-readable name for a commit if possible."""
+    if repo is None:
+        return hexsha8
+    for h in repo.heads:
+        if h.commit.hexsha[:8] == hexsha8:
+            return h.name
+    for t in repo.tags:
+        if t.commit.hexsha[:8] == hexsha8:
+            return t.name
+    return hexsha8
+
+
+def get_commit_hexsha8(name):
+    """Helper function to search for a commit given a human-readable name."""
+    if repo is None:
+        return None
+    for h in repo.heads:
+        if h.name == name:
+            return h.commit.hexsha[:8]
+    for t in repo.tags:
+        if t.name == name:
+            return t.commit.hexsha[:8]
+    return None
+
+
+hexsha8_baseline = name_baseline = None
+
+# If the user specified a baseline, try to find a commit for it:
+if known_args.baseline is not None:
+    if (known_args.baseline,) in builds:
+        hexsha8_baseline = known_args.baseline
+    if hexsha8_baseline is None:
+        hexsha8_baseline = get_commit_hexsha8(known_args.baseline)
+        name_baseline = known_args.baseline
+    if hexsha8_baseline is None:
+        print(f"ERROR: cannot find data for baseline={known_args.baseline}.")
+        sys.exit(1)
+# Otherwise, search for the most recent parent of master for which there is data:
+elif repo is not None:
+    hexsha8_baseline = find_parent_in_data(repo.heads.master.commit)
+
+    if hexsha8_baseline is None:
+        print("ERROR: No baseline was provided and did not find data for any master branch commits.")
+        print()
+        parser.print_help()
+        sys.exit(1)
+else:
+    print(
+        "ERROR: No baseline was provided and the current working directory "
+        "is not part of a git repository from which a baseline could be inferred."
+    )
+    print()
+    parser.print_help()
+    sys.exit(1)
+
+
+name_baseline = get_commit_name(hexsha8_baseline)
+
+hexsha8_compare = name_compare = None
+
+# If the user has specified a compare value, try to find a corresponding commit:
+if known_args.compare is not None:
+    if (known_args.compare,) in builds:
+        hexsha8_compare = known_args.compare
+    if hexsha8_compare is None:
+        hexsha8_compare = get_commit_hexsha8(known_args.compare)
+        name_compare = known_args.compare
+    if hexsha8_compare is None:
+        print(f"ERROR: cannot find data for baseline={known_args.compare}.")
+        sys.exit(1)
+# Otherwise, search for the commit for llama-bench was most recently run
+# and that is not a parent of master:
+elif repo is not None:
+    hexsha8s_master = get_all_parent_hexsha8s(repo.heads.master.commit)
+    builds_timestamp = cursor.execute(
+        "SELECT build_commit, test_time FROM test ORDER BY test_time;").fetchall()
+    for (hexsha8, _) in reversed(builds_timestamp):
+        if hexsha8 not in hexsha8s_master:
+            hexsha8_compare = hexsha8
+            break
+
+    if hexsha8_compare is None:
+        print("ERROR: No compare target was provided and did not find data for any non-master commits.")
+        print()
+        parser.print_help()
+        sys.exit(1)
+else:
+    print(
+        "ERROR: No compare target was provided and the current working directory "
+        "is not part of a git repository from which a compare target could be inferred."
+    )
+    print()
+    parser.print_help()
+    sys.exit(1)
+
+name_compare = get_commit_name(hexsha8_compare)
+
+
+def get_rows(properties):
+    """
+    Helper function that gets table rows for some list of properties.
+    Rows are created by combining those where all provided properties are equal.
+    The resulting rows are then grouped by the provided properties and the t/s values are averaged.
+    The returned rows are unique in terms of property combinations.
+    """
+    select_string = ", ".join(
+        [f"tb.{p}" for p in properties] + ["tb.n_prompt", "tb.n_gen", "AVG(tb.avg_ts)", "AVG(tc.avg_ts)"])
+    equal_string = " AND ".join(
+        [f"tb.{p} = tc.{p}" for p in KEY_PROPERTIES] + [
+            f"tb.build_commit = '{hexsha8_baseline}'", f"tc.build_commit = '{hexsha8_compare}'"]
+    )
+    group_order_string = ", ".join([f"tb.{p}" for p in properties] + ["tb.n_gen", "tb.n_prompt"])
+    query = (f"SELECT {select_string} FROM test tb JOIN test tc ON {equal_string} "
+             f"GROUP BY {group_order_string} ORDER BY {group_order_string};")
+    return cursor.execute(query).fetchall()
+
+
+# If the user provided columns to group the results by, use them:
+if known_args.show is not None:
+    show = known_args.show.split(",")
+    unknown_cols = []
+    for prop in show:
+        if prop not in KEY_PROPERTIES[:-2]:  # Last two values are n_prompt, n_gen.
+            unknown_cols.append(prop)
+    if unknown_cols:
+        print(f"ERROR: Unknown values for --show: {', '.join(unknown_cols)}")
+        print()
+        parser.print_usage()
+        sys.exit(1)
+    rows_show = get_rows(show)
+# Otherwise, select those columns where the values are not all the same:
+else:
+    rows_full = get_rows(KEY_PROPERTIES)
+    properties_different = []
+    for i, kp_i in enumerate(KEY_PROPERTIES):
+        if kp_i in DEFAULT_SHOW or kp_i == "n_prompt" or kp_i == "n_gen":
+            continue
+        for row_full in rows_full:
+            if row_full[i] != rows_full[0][i]:
+                properties_different.append(kp_i)
+                break
+
+    show = []
+    # Show CPU and/or GPU by default even if the hardware for all results is the same:
+    if "gpu_blas" not in properties_different and "n_gpu_layers" not in properties_different:
+        gpu_blas = bool(rows_full[0][KEY_PROPERTIES.index("gpu_blas")])
+        ngl = int(rows_full[0][KEY_PROPERTIES.index("n_gpu_layers")])
+
+        if not gpu_blas or ngl != 99 and "cpu_info" not in properties_different:
+            show.append("cpu_info")
+        if gpu_blas and "gpu_info" not in properties_different:
+            show.append("gpu_info")
+
+    show += properties_different
+
+    index_default = 0
+    for prop in ["cpu_info", "gpu_info", "n_gpu_layers", "main_gpu"]:
+        if prop in show:
+            index_default += 1
+    show = show[:index_default] + DEFAULT_SHOW + show[index_default:]
+    for prop in DEFAULT_HIDE:
+        try:
+            show.remove(prop)
+        except ValueError:
+            pass
+    rows_show = get_rows(show)
+
+table = []
+for row in rows_show:
+    n_prompt = int(row[-4])
+    n_gen    = int(row[-3])
+    assert n_prompt == 0 or n_gen == 0
+    test_name = f"tg{n_gen}" if n_prompt == 0 else f"pp{n_prompt}"
+    #           Regular columns    test name    avg t/s values              Speedup
+    #            VVVVVVVVVVVVV     VVVVVVVVV    VVVVVVVVVVVVVV              VVVVVVV
+    table.append(list(row[:-4]) + [test_name] + list(row[-2:]) + [float(row[-1]) / float(row[-2])])
+
+# Some a-posteriori fixes to make the table contents prettier:
+for bool_property in BOOL_PROPERTIES:
+    if bool_property in show:
+        ip = show.index(bool_property)
+        for row_table in table:
+            row_table[ip] = "Yes" if int(row_table[ip]) == 1 else "No"
+
+if "model_type" in show:
+    ip = show.index("model_type")
+    for (old, new) in MODEL_SUFFIX_REPLACE.items():
+        for row_table in table:
+            row_table[ip] = row_table[ip].replace(old, new)
+
+if "model_size" in show:
+    ip = show.index("model_size")
+    for row_table in table:
+        row_table[ip] = float(row_table[ip]) / 1024 ** 3
+
+if "gpu_info" in show:
+    ip = show.index("gpu_info")
+    for row_table in table:
+        for gns in GPU_NAME_STRIP:
+            row_table[ip] = row_table[ip].replace(gns, "")
+
+        gpu_names = row_table[ip].split("/")
+        num_gpus = len(gpu_names)
+        all_names_the_same = len(set(gpu_names)) == 1
+        if len(gpu_names) >= 2 and all_names_the_same:
+            row_table[ip] = f"{num_gpus}x {gpu_names[0]}"
+
+headers  = [PRETTY_NAMES[p] for p in show]
+headers += ["Test", f"t/s {name_baseline}", f"t/s {name_compare}", "Speedup"]
+
+print(tabulate(
+    table,
+    headers=headers,
+    floatfmt=".2f",
+    tablefmt=known_args.output
+))

scripts/get-pg.sh

@@ -0,0 +1,70 @@
+#!/bin/bash
+
+function usage {
+    echo "usage: <n>$0"
+    echo "note: n is the number of essays to download"
+    echo "for specific n, the resulting pg.txt file will have the following number of tokens:"
+    echo "n   | tokens"
+    echo "--- | ---"
+    echo "1   | 6230"
+    echo "2   | 23619"
+    echo "5   | 25859"
+    echo "10  | 36888"
+    echo "15  | 50188"
+    echo "20  | 59094"
+    echo "25  | 88764"
+    echo "30  | 103121"
+    echo "32  | 108338"
+    echo "35  | 113403"
+    echo "40  | 127699"
+    echo "45  | 135896"
+    exit 1
+}
+
+function has_cmd {
+    if ! [ -x "$(command -v $1)" ]; then
+        echo "error: $1 is not available" >&2
+        exit 1
+    fi
+}
+
+# check for: curl, html2text, tail, sed, fmt
+has_cmd curl
+has_cmd html2text
+has_cmd tail
+has_cmd sed
+
+if [ $# -ne 1 ]; then
+    usage
+fi
+
+n=$1
+
+# get urls
+urls="$(curl http://www.aaronsw.com/2002/feeds/pgessays.rss | grep html | sed -e "s/.*http/http/" | sed -e "s/html.*/html/" | head -n $n)"
+
+printf "urls:\n%s\n" "$urls"
+
+if [ -f pg.txt ]; then
+    rm pg.txt
+fi
+
+c=1
+for url in $urls; do
+    echo "processing $url"
+
+    cc=$(printf "%03d" $c)
+
+    curl -L $url | html2text | tail -n +4 | sed -E "s/^[[:space:]]+//g" | fmt -w 80 >> pg-$cc-one.txt
+    cat pg-$cc-one.txt >> pg.txt
+
+    cp -v pg.txt pg-$cc-all.txt
+    c=$((c+1))
+
+    # don't flood the server
+    sleep 1
+done
+
+echo "done. data in pg.txt"
+
+exit 0

scripts/sync-ggml-am.sh

@@ -5,7 +5,7 @@
 # Usage:
 #
 #   $ cd /path/to/llama.cpp
-#   $ ./scripts/sync-ggml-am.sh
+#   $ ./scripts/sync-ggml-am.sh -skip hash0,hash1,hash2...
 #
 
 set -e
@@ -24,6 +24,11 @@ fi
 lc=$(cat $SRC_LLAMA/scripts/sync-ggml.last)
 echo "Syncing ggml changes since commit $lc"
 
+to_skip=""
+if [ "$1" == "-skip" ]; then
+    to_skip=$2
+fi
+
 cd $SRC_GGML
 
 git log --oneline $lc..HEAD
@@ -40,6 +45,13 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
 fi
 
 while read c; do
+    if [ -n "$to_skip" ]; then
+        if [[ $to_skip == *"$c"* ]]; then
+            echo "Skipping $c"
+            continue
+        fi
+    fi
+
     git format-patch -k $c~1..$c --stdout -- \
         include/ggml/ggml*.h \
         src/ggml*.h \

scripts/sync-ggml.last

@@ -1 +1 @@
-f96711108d55bdbbd277e6be07204dce6a94fb93
+b306d6e996ec0ace77118fa5098822cdc7f9c88f

spm-headers/ggml.h

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

tests/test-backend-ops.cpp

@@ -56,7 +56,7 @@ static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float m
         GGML_ASSERT(size % ggml_blck_size(tensor->type) == 0);
         std::vector<uint8_t> dataq(ggml_row_size(tensor->type, size));
         int64_t hist[16];
-        ggml_quantize_chunk(tensor->type, data.data(), dataq.data(), 0, size, hist);
+        ggml_quantize_chunk(tensor->type, data.data(), dataq.data(), 0, size/tensor->ne[0], tensor->ne[0], hist, nullptr);
         ggml_backend_tensor_set(tensor, dataq.data(), 0, dataq.size());
     } else if (tensor->type == GGML_TYPE_I8 || tensor->type == GGML_TYPE_I16 || tensor->type == GGML_TYPE_I32) {
         // This is going to create some weird integers though.
@@ -376,6 +376,11 @@ struct test_case {
 
         // allocate
         ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(ctx, backend1);
+        if (buf == NULL) {
+            printf("failed to allocate tensors [%s] ", ggml_backend_name(backend1));
+            ggml_free(ctx);
+            return false;
+        }
 
         // build graph
         ggml_build_forward_expand(gf, out);
@@ -450,7 +455,7 @@ struct test_case {
 
             double err = nmse(f1.data(), f2.data(), f1.size());
             if (err > ud->max_err) {
-                printf("[%s] NMSE = %f ", ggml_op_desc(t1), err);
+                printf("[%s] NMSE = %.9f > %.9f ", ggml_op_desc(t1), err, ud->max_err);
                 //for (int i = 0; i < (int) f1.size(); i++) {
                 //    printf("%5d %9.6f %9.6f, diff = %9.6f\n", i, f1[i], f2[i], f1[i] - f2[i]);
                 //}
@@ -463,19 +468,23 @@ struct test_case {
             GGML_UNUSED(index);
         };
 
-        ggml_backend_compare_graph_backend(backend1, backend2, gf, callback, &ud);
+        const bool cmp_ok = ggml_backend_compare_graph_backend(backend1, backend2, gf, callback, &ud);
 
-        if (ud.ok) {
-            printf("\033[1;32mOK\033[0m\n");
-        } else {
-            printf("\033[1;31mFAIL\033[0m\n");
+        if (!cmp_ok) {
+            printf("compare failed ");
         }
 
         ggml_backend_buffer_free(buf);
 
         ggml_free(ctx);
 
-        return ud.ok;
+        if (ud.ok && cmp_ok) {
+            printf("\033[1;32mOK\033[0m\n");
+            return true;
+        }
+
+        printf("\033[1;31mFAIL\033[0m\n");
+        return false;
     }
 
     bool eval_perf(ggml_backend_t backend, const char * op_name) {
@@ -519,6 +528,11 @@ struct test_case {
 
         // allocate
         ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(ctx, backend);
+        if (buf == NULL) {
+            printf("failed to allocate tensors\n");
+            ggml_free(ctx);
+            return false;
+        }
 
         // randomize tensors
         initialize_tensors(ctx);
@@ -1449,6 +1463,7 @@ struct test_moe : public test_case {
 
 static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op_name) {
     std::vector<std::unique_ptr<test_case>> test_cases;
+    std::default_random_engine rng(0);
 
     const ggml_type all_types[] = {
         GGML_TYPE_F32, GGML_TYPE_F16,
@@ -1583,7 +1598,19 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10, 10,  1}, 5));
     test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10, 10, 10}, 5));
 
-    test_cases.emplace_back(new test_soft_max());
+    std::uniform_int_distribution<> dist_ne1(1, 50);
+    int exponent = 1;
+    while (exponent < (1 << 17)) {
+        std::uniform_int_distribution<> dist_ne0(exponent, 2*exponent);
+
+        for (int n = 0; n < 10; ++n) {
+            int64_t ne0 = dist_ne0(rng);
+            int64_t ne1 = dist_ne1(rng);
+            test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0, ne1, 1, 1}));
+        }
+
+        exponent <<= 1;
+    }
 
     for (ggml_type type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
         test_cases.emplace_back(new test_rope(type, {128,  32, 10, 1}, 128, 0, 512)); // llama 7B

tests/test-quantize-fns.cpp

@@ -134,6 +134,12 @@ int main(int argc, char * argv[]) {
             continue;
         }
 
+        const ggml_type ei = (ggml_type)i;
+        if (ei == GGML_TYPE_IQ2_XXS || ei == GGML_TYPE_IQ2_XS) {
+            printf("Skip %s due to missing quantization functionality\n", ggml_type_name(ei));
+            continue;
+        }
+
         printf("Testing %s\n", ggml_type_name((ggml_type) i));
 
         if (qfns.from_float && qfns.to_float) {