server : chunked prefill support

ggml-ci

.github/workflows/build.yml

@@ -552,44 +552,35 @@ jobs:
             -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
 
-  macOS-latest-swift:
-    runs-on: macos-latest
-
-    strategy:
-      matrix:
-        destination: ['generic/platform=macOS', 'generic/platform=iOS', 'generic/platform=tvOS']
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-
-      - name: Dependencies
-        id: depends
-        continue-on-error: true
-        run: |
-          brew update
-
-      - name: Build llama.cpp with CMake
-        id: cmake_build
-        run: |
-          sysctl -a
-          mkdir build
-          cd build
-          cmake -G Xcode .. \
-            -DGGML_METAL_USE_BF16=ON \
-            -DGGML_METAL_EMBED_LIBRARY=ON \
-            -DLLAMA_BUILD_EXAMPLES=OFF \
-            -DLLAMA_BUILD_TESTS=OFF \
-            -DLLAMA_BUILD_SERVER=OFF \
-            -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64"
-          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
-          sudo cmake --install . --config Release
-
-      - name: xcodebuild for swift package
-        id: xcodebuild
-        run: |
-          xcodebuild -scheme llama-Package -destination "${{ matrix.destination }}"
+# TODO: tmp disabled. see for possible re-enable:
+#       https://github.com/ggerganov/llama.cpp/pull/10525
+#  macOS-latest-swift:
+#    runs-on: macos-latest
+#
+#    strategy:
+#      matrix:
+#        destination: ['generic/platform=macOS', 'generic/platform=iOS', 'generic/platform=tvOS']
+#
+#    steps:
+#      - name: Clone
+#        id: checkout
+#        uses: actions/checkout@v4
+#
+#      - name: Dependencies
+#        id: depends
+#        continue-on-error: true
+#        run: |
+#          brew update
+#
+#      - name: xcodebuild for swift package
+#        id: xcodebuild
+#        run: |
+#          xcodebuild -scheme llama -destination "${{ matrix.destination }}"
+#
+#      - name: Build Swift Example
+#        id: make_build_swift_example
+#        run: |
+#            make swift
 
   windows-msys2:
     runs-on: windows-latest
@@ -1113,29 +1104,6 @@ jobs:
       - name: Checkout code
         uses: actions/checkout@v4
 
-      - name: Build
-        id: cmake_build
-        run: |
-          sysctl -a
-          mkdir build
-          cd build
-          cmake -G Xcode .. \
-            -DGGML_METAL_USE_BF16=ON \
-            -DGGML_METAL_EMBED_LIBRARY=ON \
-            -DLLAMA_BUILD_EXAMPLES=OFF \
-            -DLLAMA_BUILD_TESTS=OFF \
-            -DLLAMA_BUILD_SERVER=OFF \
-            -DCMAKE_SYSTEM_NAME=iOS \
-            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
-            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
-          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
-          sudo cmake --install . --config Release
-
-      - name: xcodebuild for swift package
-        id: xcodebuild
-        run: |
-          xcodebuild -scheme llama-Package -destination 'generic/platform=iOS'
-
       - name: Build Xcode project
         run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' build
 
@@ -1163,6 +1131,23 @@ jobs:
 
           ./gradlew build --no-daemon
 
+#  freeBSD-latest:
+#    runs-on: macos-12
+#    steps:
+#    - name: Clone
+#      uses: actions/checkout@v4
+#
+#    - name: Build
+#      uses: cross-platform-actions/action@v0.19.0
+#      with:
+#        operating_system: freebsd
+#        version: '13.2'
+#        hypervisor: 'qemu'
+#        run: |
+#            sudo pkg update
+#            sudo pkg install -y gmake automake autoconf pkgconf llvm15 openblas
+#            gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j `sysctl -n hw.ncpu`
+
   release:
     if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
 

.github/workflows/server.yml

@@ -79,7 +79,7 @@ jobs:
       # Setup nodejs (to be used for verifying bundled index.html)
       - uses: actions/setup-node@v4
         with:
-          node-version: '22.11.0'
+          node-version: 22
 
       - name: Verify bundled index.html
         id: verify_server_index_html

CMakeLists.txt

@@ -46,9 +46,11 @@ if (WIN32)
     add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
 endif()
 
-if (MSVC)
-    add_compile_options("$<$<COMPILE_LANGUAGE:C>:/utf-8>")
-    add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/utf-8>")
+if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC")
+    add_compile_options("$<$<COMPILE_LANGUAGE:C>:/source-charset:utf-8>")
+    add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/source-charset:utf-8>")
+    add_compile_options("$<$<COMPILE_LANGUAGE:C>:/execution-charset:utf-8>")
+    add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/execution-charset:utf-8>")
 endif()
 
 #

CMakePresets.json

@@ -31,13 +31,6 @@
     { "name": "sycl_f16", "hidden": true, "cacheVariables": { "GGML_SYCL_F16":    "ON" } },
     { "name": "vulkan",   "hidden": true, "cacheVariables": { "GGML_VULKAN":      "ON" } },
 
-    {
-        "name": "x64-windows-llvm", "hidden": true,
-        "cacheVariables": {
-            "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/x64-windows-llvm.cmake"
-        }
-    },
-
     {
         "name": "arm64-windows-msvc", "hidden": true,
         "architecture": { "value": "arm64",    "strategy": "external" },
@@ -77,11 +70,6 @@
     { "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc",  "reldbg" ] },
     { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc",  "reldbg", "static" ] },
 
-    { "name": "x64-windows-llvm-debug", "inherits": [ "base", "x64-windows-llvm", "debug" ] },
-    { "name": "x64-windows-llvm-release", "inherits": [ "base", "x64-windows-llvm", "release" ] },
-    { "name": "x64-windows-llvm-reldbg", "inherits": [ "base", "x64-windows-llvm", "reldbg" ] },
-    { "name": "x64-windows-llvm+static-release", "inherits": [ "base", "x64-windows-llvm", "reldbg", "static" ] },
-
     { "name": "x64-windows-msvc-debug", "inherits": [ "base", "debug" ] },
     { "name": "x64-windows-msvc-release", "inherits": [ "base", "reldbg" ] },
     { "name": "x64-windows-msvc+static-release", "inherits": [ "base", "reldbg", "static" ] },

Package.swift

@@ -2,6 +2,60 @@
 
 import PackageDescription
 
+var sources = [
+    "src/llama.cpp",
+    "src/llama-vocab.cpp",
+    "src/llama-grammar.cpp",
+    "src/llama-sampling.cpp",
+    "src/unicode.cpp",
+    "src/unicode-data.cpp",
+    "ggml/src/ggml.c",
+    "ggml/src/ggml-alloc.c",
+    "ggml/src/ggml-backend.cpp",
+    "ggml/src/ggml-backend-reg.cpp",
+    "ggml/src/ggml-cpu/ggml-cpu.c",
+    "ggml/src/ggml-cpu/ggml-cpu.cpp",
+    "ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp",
+    "ggml/src/ggml-cpu/ggml-cpu-hbm.cpp",
+    "ggml/src/ggml-cpu/ggml-cpu-quants.c",
+    "ggml/src/ggml-cpu/ggml-cpu-traits.cpp",
+    "ggml/src/ggml-threading.cpp",
+    "ggml/src/ggml-quants.c",
+]
+
+var resources: [Resource] = []
+var linkerSettings: [LinkerSetting] = []
+var cSettings: [CSetting] =  [
+    .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
+    .unsafeFlags(["-fno-objc-arc"]),
+    .headerSearchPath("ggml/src"),
+    .headerSearchPath("ggml/src/ggml-cpu"),
+    // NOTE: NEW_LAPACK will required iOS version 16.4+
+    // We should consider add this in the future when we drop support for iOS 14
+    // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
+    // .define("ACCELERATE_NEW_LAPACK"),
+    // .define("ACCELERATE_LAPACK_ILP64")
+    .define("GGML_USE_CPU"),
+]
+
+
+#if canImport(Darwin)
+sources.append("ggml/src/ggml-common.h")
+sources.append("ggml/src/ggml-metal/ggml-metal.m")
+resources.append(.process("ggml/src/ggml-metal/ggml-metal.metal"))
+linkerSettings.append(.linkedFramework("Accelerate"))
+cSettings.append(
+    contentsOf: [
+        .define("GGML_USE_ACCELERATE"),
+        .define("GGML_USE_METAL"),
+    ]
+)
+#endif
+
+#if os(Linux)
+    cSettings.append(.define("_GNU_SOURCE"))
+#endif
+
 let package = Package(
     name: "llama",
     platforms: [
@@ -14,6 +68,26 @@ let package = Package(
         .library(name: "llama", targets: ["llama"]),
     ],
     targets: [
-        .systemLibrary(name: "llama", pkgConfig: "llama"),
-    ]
+        .target(
+            name: "llama",
+            path: ".",
+            exclude: [
+               "build",
+               "cmake",
+               "examples",
+               "scripts",
+               "models",
+               "tests",
+               "CMakeLists.txt",
+               "Makefile",
+               "ggml/src/ggml-metal-embed.metal"
+            ],
+            sources: sources,
+            resources: resources,
+            publicHeadersPath: "spm-headers",
+            cSettings: cSettings,
+            linkerSettings: linkerSettings
+        )
+    ],
+    cxxLanguageStandard: .cxx17
 )

Sources/llama/llama.h

@@ -1,4 +0,0 @@
-#pragma once
-
-#include <llama.h>
-

Sources/llama/module.modulemap

@@ -1,5 +0,0 @@
-module llama [system] {
-    header "llama.h"
-    link "llama"
-    export *
-}

cmake/llama.pc.in

@@ -6,5 +6,5 @@ includedir=${prefix}/include
 Name: llama
 Description: Port of Facebook's LLaMA model in C/C++
 Version: @PROJECT_VERSION@
-Libs: -L${libdir} -lggml  -lggml-base -lllama
+Libs: -L${libdir} -lllama
 Cflags: -I${includedir}

cmake/x64-windows-llvm.cmake

@@ -1,11 +0,0 @@
-set( CMAKE_SYSTEM_NAME Windows )
-set( CMAKE_SYSTEM_PROCESSOR x86_64 )
-
-set( CMAKE_C_COMPILER    clang )
-set( CMAKE_CXX_COMPILER  clang++ )
-
-set( arch_c_flags "-march=native" )
-
-set( CMAKE_C_FLAGS_INIT   "${arch_c_flags}" )
-set( CMAKE_CXX_FLAGS_INIT "${arch_c_flags}" )
-

common/arg.cpp

@@ -591,7 +591,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.ctx_shift = false;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_IMATRIX}).set_env("LLAMA_ARG_NO_CONTEXT_SHIFT"));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_CONTEXT_SHIFT"));
     add_opt(common_arg(
         {"--chunks"}, "N",
         string_format("max number of chunks to process (default: %d, -1 = all)", params.n_chunks),
@@ -1711,13 +1711,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.public_path = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_STATIC_PATH"));
-    add_opt(common_arg(
-        {"--no-webui"},
-        string_format("Disable the Web UI (default: %s)", params.webui ? "enabled" : "disabled"),
-        [](common_params & params) {
-            params.webui = false;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_WEBUI"));
     add_opt(common_arg(
         {"--embedding", "--embeddings"},
         string_format("restrict to only support embedding use case; use only with dedicated embedding models (default: %s)", params.embedding ? "enabled" : "disabled"),
@@ -2083,35 +2076,35 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, int value) {
             params.speculative.n_max = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_DRAFT_MAX"));
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"--draft-min", "--draft-n-min"}, "N",
         string_format("minimum number of draft tokens to use for speculative decoding (default: %d)", params.speculative.n_min),
         [](common_params & params, int value) {
             params.speculative.n_min = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_DRAFT_MIN"));
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"--draft-p-split"}, "P",
         string_format("speculative decoding split probability (default: %.1f)", (double)params.speculative.p_split),
         [](common_params & params, const std::string & value) {
             params.speculative.p_split = std::stof(value);
         }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}).set_env("LLAMA_ARG_DRAFT_P_SPLIT"));
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
     add_opt(common_arg(
         {"--draft-p-min"}, "P",
         string_format("minimum speculative decoding probability (greedy) (default: %.1f)", (double)params.speculative.p_min),
         [](common_params & params, const std::string & value) {
             params.speculative.p_min = std::stof(value);
         }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_DRAFT_P_MIN"));
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"-cd", "--ctx-size-draft"}, "N",
         string_format("size of the prompt context for the draft model (default: %d, 0 = loaded from model)", params.speculative.n_ctx),
         [](common_params & params, int value) {
             params.speculative.n_ctx = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CTX_SIZE_DRAFT"));
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"-devd", "--device-draft"}, "<dev1,dev2,..>",
         "comma-separated list of devices to use for offloading the draft model (none = don't offload)\n"
@@ -2131,14 +2124,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                 fprintf(stderr, "warning: consult docs/build.md for compilation instructions\n");
             }
         }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_N_GPU_LAYERS_DRAFT"));
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"-md", "--model-draft"}, "FNAME",
         "draft model for speculative decoding (default: unused)",
         [](common_params & params, const std::string & value) {
             params.speculative.model = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODEL_DRAFT"));
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
 
     return ctx_arg;
 }

docs/build.md

@@ -57,13 +57,6 @@ cmake --build build --config Release
     ```
     Building for arm64 can also be done with the MSVC compiler with the build-arm64-windows-MSVC preset, or the standard CMake build instructions. However, note that the MSVC compiler does not support inline ARM assembly code, used e.g. for the accelerated Q4_0_N_M CPU kernels.
 
-    For building with ninja generator and clang compiler as default:
-      -set path:set LIB=C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22621.0\um\x64;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.41.34120\lib\x64\uwp;C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22621.0\ucrt\x64
-      ```bash
-      cmake --preset x64-windows-llvm-release
-      cmake --build build-x64-windows-llvm-release
-      ```
-
 ## BLAS Build
 
 Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Using BLAS doesn't affect the generation performance. There are currently several different BLAS implementations available for build and use:

examples/CMakeLists.txt

@@ -20,12 +20,7 @@ else()
     add_subdirectory(batched)
     add_subdirectory(embedding)
     add_subdirectory(eval-callback)
-
-    if (NOT WIN32)
-        # disabled on Windows because it uses internal functions not exported with LLAMA_API
-        add_subdirectory(gbnf-validator)
-    endif()
-
+    add_subdirectory(gbnf-validator)
     add_subdirectory(gguf-hash)
     add_subdirectory(gguf-split)
     add_subdirectory(gguf)
@@ -56,10 +51,7 @@ else()
         add_subdirectory(convert-llama2c-to-ggml)
         add_subdirectory(cvector-generator)
         add_subdirectory(export-lora)
-        if (NOT WIN32)
-            # disabled on Windows because it uses internal functions not exported with LLAMA_API
-            add_subdirectory(quantize-stats)
-        endif()
+        add_subdirectory(quantize-stats)
         add_subdirectory(llava)
         if (GGML_RPC)
             add_subdirectory(rpc)

examples/gguf-split/gguf-split.cpp

@@ -287,7 +287,7 @@ struct split_strategy {
     }
 
     void print_info() {
-        printf("n_split: %zu\n", ctx_outs.size());
+        printf("n_split: %ld\n", ctx_outs.size());
         int i_split = 0;
         for (auto & ctx_out : ctx_outs) {
             // re-calculate the real gguf size for each split (= metadata size + total size of all tensors)
@@ -297,7 +297,7 @@ struct split_strategy {
                 total_size += ggml_nbytes(t);
             }
             total_size = total_size / 1000 / 1000; // convert to megabytes
-            printf("split %05d: n_tensors = %d, total_size = %zuM\n", i_split + 1, gguf_get_n_tensors(ctx_out), total_size);
+            printf("split %05d: n_tensors = %d, total_size = %ldM\n", i_split + 1, gguf_get_n_tensors(ctx_out), total_size);
             i_split++;
         }
     }

examples/llama-bench/llama-bench.cpp

@@ -1521,7 +1521,7 @@ int main(int argc, char ** argv) {
     for (const auto & inst : params_instances) {
         params_idx++;
         if (params.progress) {
-            fprintf(stderr, "llama-bench: benchmark %d/%zu: starting\n", params_idx, params_count);
+            fprintf(stderr, "llama-bench: benchmark %d/%ld: starting\n", params_idx, params_count);
         }
         // keep the same model between tests when possible
         if (!lmodel || !prev_inst || !inst.equal_mparams(*prev_inst)) {
@@ -1573,14 +1573,14 @@ int main(int argc, char ** argv) {
         // warmup run
         if (t.n_prompt > 0) {
             if (params.progress) {
-                fprintf(stderr, "llama-bench: benchmark %d/%zu: warmup prompt run\n", params_idx, params_count);
+                fprintf(stderr, "llama-bench: benchmark %d/%ld: warmup prompt run\n", params_idx, params_count);
             }
             //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads);
             test_prompt(ctx, t.n_prompt, t.n_batch, t.n_threads);
         }
         if (t.n_gen > 0) {
             if (params.progress) {
-                fprintf(stderr, "llama-bench: benchmark %d/%zu: warmup generation run\n", params_idx, params_count);
+                fprintf(stderr, "llama-bench: benchmark %d/%ld: warmup generation run\n", params_idx, params_count);
             }
             test_gen(ctx, 1, t.n_threads);
         }
@@ -1592,14 +1592,14 @@ int main(int argc, char ** argv) {
 
             if (t.n_prompt > 0) {
                 if (params.progress) {
-                    fprintf(stderr, "llama-bench: benchmark %d/%zu: prompt run %d/%d\n", params_idx, params_count,
+                    fprintf(stderr, "llama-bench: benchmark %d/%ld: prompt run %d/%d\n", params_idx, params_count,
                             i + 1, params.reps);
                 }
                 test_prompt(ctx, t.n_prompt, t.n_batch, t.n_threads);
             }
             if (t.n_gen > 0) {
                 if (params.progress) {
-                    fprintf(stderr, "llama-bench: benchmark %d/%zu: generation run %d/%d\n", params_idx, params_count,
+                    fprintf(stderr, "llama-bench: benchmark %d/%ld: generation run %d/%d\n", params_idx, params_count,
                             i + 1, params.reps);
                 }
                 test_gen(ctx, t.n_gen, t.n_threads);

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

@@ -210,20 +210,20 @@ actor LlamaContext {
 
             llama_kv_cache_clear(context)
 
-            let t_pp_start = DispatchTime.now().uptimeNanoseconds / 1000;
+            let t_pp_start = ggml_time_us()
 
             if llama_decode(context, batch) != 0 {
                 print("llama_decode() failed during prompt")
             }
             llama_synchronize(context)
 
-            let t_pp_end = DispatchTime.now().uptimeNanoseconds / 1000;
+            let t_pp_end = ggml_time_us()
 
             // bench text generation
 
             llama_kv_cache_clear(context)
 
-            let t_tg_start = DispatchTime.now().uptimeNanoseconds / 1000;
+            let t_tg_start = ggml_time_us()
 
             for i in 0..<tg {
                 llama_batch_clear(&batch)
@@ -238,7 +238,7 @@ actor LlamaContext {
                 llama_synchronize(context)
             }
 
-            let t_tg_end = DispatchTime.now().uptimeNanoseconds / 1000;
+            let t_tg_end = ggml_time_us()
 
             llama_kv_cache_clear(context)
 

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

@@ -7,7 +7,6 @@
 	objects = {
 
 /* Begin PBXBuildFile section */
-		1809696D2D05A39F00400EE8 /* llama in Frameworks */ = {isa = PBXBuildFile; productRef = 1809696C2D05A39F00400EE8 /* llama */; };
 		549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 549479CA2AC9E16000E0F78B /* Metal.framework */; };
 		79E1D9CD2B4CD16E005F8E46 /* InputButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = 79E1D9CC2B4CD16E005F8E46 /* InputButton.swift */; };
 		7FA3D2B32B2EA2F600543F92 /* DownloadButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */; };
@@ -18,6 +17,7 @@
 		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 */; };
+		DF810E132B4A5BA200301144 /* llama in Frameworks */ = {isa = PBXBuildFile; productRef = DF810E122B4A5BA200301144 /* llama */; };
 		F1FE20E22B465ECA00B45541 /* LoadCustomButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = F1FE20E12B465EC900B45541 /* LoadCustomButton.swift */; };
 /* End PBXBuildFile section */
 
@@ -42,7 +42,7 @@
 			isa = PBXFrameworksBuildPhase;
 			buildActionMask = 2147483647;
 			files = (
-				1809696D2D05A39F00400EE8 /* llama in Frameworks */,
+				DF810E132B4A5BA200301144 /* llama in Frameworks */,
 				549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */,
 				8A39BE0A2AC7601100BFEB40 /* Accelerate.framework in Frameworks */,
 			);
@@ -151,7 +151,7 @@
 			);
 			name = llama.swiftui;
 			packageProductDependencies = (
-				1809696C2D05A39F00400EE8 /* llama */,
+				DF810E122B4A5BA200301144 /* llama */,
 			);
 			productName = llama.swiftui;
 			productReference = 8A1C83732AC328BD0096AF73 /* llama.swiftui.app */;
@@ -429,7 +429,7 @@
 /* End XCConfigurationList section */
 
 /* Begin XCSwiftPackageProductDependency section */
-		1809696C2D05A39F00400EE8 /* llama */ = {
+		DF810E122B4A5BA200301144 /* llama */ = {
 			isa = XCSwiftPackageProductDependency;
 			productName = llama;
 		};

examples/quantize/README.md

@@ -81,7 +81,7 @@ Several quantization methods are supported. They differ in the resulting model d
   - [#4930 - imatrix for all k-quants](https://github.com/ggerganov/llama.cpp/pull/4930)
   - [#4951 - imatrix on the GPU](https://github.com/ggerganov/llama.cpp/pull/4957)
   - [#4969 - imatrix for legacy quants](https://github.com/ggerganov/llama.cpp/pull/4969)
-  - [#4996 - k-quants tuning](https://github.com/ggerganov/llama.cpp/pull/4996)
+  - [#4996 - k-qunats tuning](https://github.com/ggerganov/llama.cpp/pull/4996)
   - [#5060 - Q3_K_XS](https://github.com/ggerganov/llama.cpp/pull/5060)
   - [#5196 - 3-bit i-quants](https://github.com/ggerganov/llama.cpp/pull/5196)
   - [quantization tuning](https://github.com/ggerganov/llama.cpp/pull/5320), [another one](https://github.com/ggerganov/llama.cpp/pull/5334), and [another one](https://github.com/ggerganov/llama.cpp/pull/5361)

examples/retrieval/retrieval.cpp

@@ -143,7 +143,7 @@ int main(int argc, char ** argv) {
         std::vector<chunk> file_chunk = chunk_file(context_file, params.chunk_size, params.chunk_separator);
         chunks.insert(chunks.end(), file_chunk.begin(), file_chunk.end());
     }
-    LOG_INF("Number of chunks: %zu\n", chunks.size());
+    LOG_INF("Number of chunks: %ld\n", chunks.size());
 
     llama_backend_init();
     llama_numa_init(params.numa);

examples/server/README.md

@@ -146,7 +146,6 @@ The project is under active development, and we are [looking for feedback and co
 | `--host HOST` | ip address to listen (default: 127.0.0.1)<br/>(env: LLAMA_ARG_HOST) |
 | `--port PORT` | port to listen (default: 8080)<br/>(env: LLAMA_ARG_PORT) |
 | `--path PATH` | path to serve static files from (default: )<br/>(env: LLAMA_ARG_STATIC_PATH) |
-| `--no-webui` | disable the Web UI<br/>(env: LLAMA_ARG_NO_WEBUI) |
 | `--embedding, --embeddings` | restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)<br/>(env: LLAMA_ARG_EMBEDDINGS) |
 | `--reranking, --rerank` | enable reranking endpoint on server (default: disabled)<br/>(env: LLAMA_ARG_RERANKING) |
 | `--api-key KEY` | API key to use for authentication (default: none)<br/>(env: LLAMA_API_KEY) |
@@ -303,23 +302,23 @@ mkdir llama-client
 cd llama-client
 ```
 
-Create an index.js file and put this inside:
+Create a index.js file and put this inside:
 
 ```javascript
-const prompt = "Building a website can be done in 10 simple steps:"
+const prompt = `Building a website can be done in 10 simple steps:`;
 
-async function test() {
+async function Test() {
     let response = await fetch("http://127.0.0.1:8080/completion", {
-        method: "POST",
+        method: 'POST',
         body: JSON.stringify({
             prompt,
-            n_predict: 64,
+            n_predict: 512,
         })
     })
     console.log((await response.json()).content)
 }
 
-test()
+Test()
 ```
 
 And run it:
@@ -381,7 +380,7 @@ Multiple prompts are also supported. In this case, the completion result will be
 `n_keep`: Specify the number of tokens from the prompt to retain when the context size is exceeded and tokens need to be discarded. The number excludes the BOS token.
 By default, this value is set to `0`, meaning no tokens are kept. Use `-1` to retain all tokens from the prompt.
 
-`stream`: Allows receiving each predicted token in real-time instead of waiting for the completion to finish (uses a different response format). To enable this, set to `true`.
+`stream`: It allows receiving each predicted token in real-time instead of waiting for the completion to finish. To enable this, set to `true`.
 
 `stop`: Specify a JSON array of stopping strings.
 These words will not be included in the completion, so make sure to add them to the prompt for the next iteration. Default: `[]`
@@ -442,11 +441,11 @@ These words will not be included in the completion, so make sure to add them to
 
 `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["dry", "top_k", "typ_p", "top_p", "min_p", "xtc", "temperature"]` - these are all the available values.
 
-`timings_per_token`: Include prompt processing and text generation speed information in each response.  Default: `false`
+    `timings_per_token`: Include prompt processing and text generation speed information in each response.  Default: `false`
 
 **Response format**
 
-- Note: In streaming mode (`stream`), only `content` and `stop` will be returned until end of completion. Responses are sent using the [Server-sent events](https://html.spec.whatwg.org/multipage/server-sent-events.html) standard. Note: the browser's `EventSource` interface cannot be used due to its lack of `POST` request support.
+- Note: When using streaming mode (`stream`), only `content` and `stop` will be returned until end of completion.
 
 - `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:
 

examples/server/server.cpp

@@ -392,7 +392,7 @@ struct server_task_result {
         return false;
     }
     virtual bool is_stop() {
-        // only used by server_task_result_cmpl_*
+        // only used by server_task_result_cmpl_partial
         return false;
     }
     virtual int get_index() {
@@ -478,20 +478,14 @@ struct server_task_result_cmpl_final : server_task_result {
         return index;
     }
 
-    virtual bool is_stop() override {
-        return true; // in stream mode, final responses are considered stop
-    }
-
     virtual json to_json() override {
-        return oaicompat
-            ? (stream ? to_json_oaicompat_chat_stream() : to_json_oaicompat_chat())
-            : to_json_non_oaicompat();
+        return oaicompat ? to_json_oaicompat_chat() : to_json_non_oaicompat();
     }
 
     json to_json_non_oaicompat() {
         json res = json {
             {"index",               index},
-            {"content",             stream ? "" : content}, // in stream mode, content is already in last partial chunk
+            {"content",             content},
             {"id_slot",             id_slot},
             {"stop",                true},
             {"model",               oaicompat_model},
@@ -552,46 +546,18 @@ struct server_task_result_cmpl_final : server_task_result {
 
         return res;
     }
-
-    json to_json_oaicompat_chat_stream() {
-        std::time_t t = std::time(0);
-        std::string finish_reason = "length";
-        if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) {
-            finish_reason = "stop";
-        }
-
-        json choices = json::array({json{{"finish_reason", finish_reason},
-                                        {"index", 0},
-                                        {"delta", json::object()}}});
-
-        json ret = json {
-            {"choices", choices},
-            {"created", t},
-            {"id",      oaicompat_cmpl_id},
-            {"model",   oaicompat_model},
-            {"object",  "chat.completion.chunk"},
-            {"usage", json {
-                {"completion_tokens", n_decoded},
-                {"prompt_tokens",     n_prompt_tokens},
-                {"total_tokens",      n_decoded + n_prompt_tokens},
-            }},
-        };
-
-        if (timings.prompt_n >= 0) {
-            ret.push_back({"timings", timings.to_json()});
-        }
-
-        return ret;
-    }
 };
 
 struct server_task_result_cmpl_partial : server_task_result {
     int index = 0;
     std::string content;
 
+    bool truncated;
     int32_t n_decoded;
     int32_t n_prompt_tokens;
 
+    stop_type stop = STOP_TYPE_NONE;
+
     std::vector<completion_token_output> probs_output;
     result_timings timings;
 
@@ -607,19 +573,20 @@ struct server_task_result_cmpl_partial : server_task_result {
     }
 
     virtual bool is_stop() override {
-        return false; // in stream mode, partial responses are not considered stop
+        return stop != STOP_TYPE_NONE;
     }
 
     virtual json to_json() override {
-        return oaicompat ? to_json_oaicompat() : to_json_non_oaicompat();
-    }
-
-    json to_json_non_oaicompat() {
+        if (oaicompat) {
+            return to_json_oaicompat();
+        }
+        bool is_stop = stop != STOP_TYPE_NONE;
         // non-OAI-compat JSON
         json res = json {
             {"index",            index},
             {"content",          content},
-            {"stop",             false},
+            {"stop_type",        stop_type_to_str(stop)},
+            {"stop",             is_stop},
             {"id_slot",          id_slot},
             {"tokens_predicted", n_decoded},
             {"tokens_evaluated", n_prompt_tokens},
@@ -631,54 +598,72 @@ struct server_task_result_cmpl_partial : server_task_result {
         if (!probs_output.empty()) {
             res["completion_probabilities"] = completion_token_output::probs_vector_to_json(probs_output);
         }
+        if (is_stop) {
+            res.push_back({"truncated", truncated});
+        }
         return res;
     }
 
     json to_json_oaicompat() {
         bool first = n_decoded == 0;
+
+        std::string finish_reason;
+        if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) {
+            finish_reason = "stop";
+        } else if (stop == STOP_TYPE_LIMIT) {
+            finish_reason = "length";
+        }
+
         std::time_t t = std::time(0);
+
         json choices;
 
-        if (first) {
-            if (content.empty()) {
-                choices = json::array({json{{"finish_reason", nullptr},
-                                            {"index", 0},
-                                            {"delta", json{{"role", "assistant"}}}}});
-            } else {
-                // We have to send this as two updates to conform to openai behavior
-                json initial_ret = json{{"choices", json::array({json{
-                                        {"finish_reason", nullptr},
+        if (!finish_reason.empty()) {
+            choices = json::array({json{{"finish_reason", finish_reason},
                                         {"index", 0},
-                                        {"delta", json{
-                                            {"role", "assistant"}
-                                        }}}})},
-                            {"created", t},
-                            {"id", oaicompat_cmpl_id},
-                            {"model", oaicompat_model},
-                            {"object", "chat.completion.chunk"}};
-
-                json second_ret = json{
-                            {"choices", json::array({json{{"finish_reason", nullptr},
-                                                            {"index", 0},
-                                                            {"delta", json{
-                                                            {"content", content}}}
-                                                            }})},
-                            {"created", t},
-                            {"id", oaicompat_cmpl_id},
-                            {"model", oaicompat_model},
-                            {"object", "chat.completion.chunk"}};
-
-                return std::vector<json>({initial_ret, second_ret});
-            }
+                                        {"delta", json::object()}}});
         } else {
-            choices = json::array({json{
-                {"finish_reason", nullptr},
-                {"index", 0},
-                {"delta",
-                json{
-                    {"content", content},
-                }},
-            }});
+            if (first) {
+                if (content.empty()) {
+                    choices = json::array({json{{"finish_reason", nullptr},
+                                                {"index", 0},
+                                                {"delta", json{{"role", "assistant"}}}}});
+                } else {
+                    // We have to send this as two updates to conform to openai behavior
+                    json initial_ret = json{{"choices", json::array({json{
+                                            {"finish_reason", nullptr},
+                                            {"index", 0},
+                                            {"delta", json{
+                                                {"role", "assistant"}
+                                            }}}})},
+                                {"created", t},
+                                {"id", oaicompat_cmpl_id},
+                                {"model", oaicompat_model},
+                                {"object", "chat.completion.chunk"}};
+
+                    json second_ret = json{
+                                {"choices", json::array({json{{"finish_reason", nullptr},
+                                                                {"index", 0},
+                                                                {"delta", json{
+                                                                {"content", content}}}
+                                                                }})},
+                                {"created", t},
+                                {"id", oaicompat_cmpl_id},
+                                {"model", oaicompat_model},
+                                {"object", "chat.completion.chunk"}};
+
+                    return std::vector<json>({initial_ret, second_ret});
+                }
+            } else {
+                choices = json::array({json{
+                    {"finish_reason", nullptr},
+                    {"index", 0},
+                    {"delta",
+                    json{
+                        {"content", content},
+                    }},
+                }});
+            }
         }
 
         json ret = json {
@@ -693,6 +678,14 @@ struct server_task_result_cmpl_partial : server_task_result {
             ret.push_back({"timings", timings.to_json()});
         }
 
+        if (!finish_reason.empty()) {
+            ret.push_back({"usage", json {
+                {"completion_tokens", n_decoded},
+                {"prompt_tokens",     n_prompt_tokens},
+                {"total_tokens",      n_decoded + n_prompt_tokens},
+            }});
+        }
+
         return std::vector<json>({ret});
     }
 };
@@ -1895,9 +1888,12 @@ struct server_context {
         res->index   = slot.index;
         res->content = tkn.text_to_send;
 
+        res->truncated       = slot.truncated;
         res->n_decoded       = slot.n_decoded;
         res->n_prompt_tokens = slot.n_prompt_tokens;
 
+        res->stop = slot.stop;
+
         res->verbose           = slot.params.verbose;
         res->oaicompat         = slot.params.oaicompat;
         res->oaicompat_chat    = slot.params.oaicompat_chat;
@@ -1928,6 +1924,12 @@ struct server_context {
     }
 
     void send_final_response(server_slot & slot) {
+        if (slot.params.stream) {
+            // if in stream mode, send the last partial response
+            send_partial_response(slot, {0, "", {}});
+            return;
+        }
+
         auto res = std::make_unique<server_task_result_cmpl_final>();
         res->id              = slot.id_task;
         res->id_slot         = slot.id;
@@ -1946,7 +1948,6 @@ struct server_context {
         res->stop            = slot.stop;
 
         res->verbose           = slot.params.verbose;
-        res->stream            = slot.params.stream;
         res->oaicompat         = slot.params.oaicompat;
         res->oaicompat_chat    = slot.params.oaicompat_chat;
         res->oaicompat_model   = slot.params.oaicompat_model;
@@ -2099,10 +2100,7 @@ struct server_context {
                 return;
             }
 
-            GGML_ASSERT(
-                dynamic_cast<server_task_result_cmpl_partial*>(result.get()) != nullptr
-                || dynamic_cast<server_task_result_cmpl_final*>(result.get()) != nullptr
-            );
+            GGML_ASSERT(dynamic_cast<server_task_result_cmpl_partial*>(result.get()) != nullptr);
             if (!result_handler(result)) {
                 cancel_tasks(id_tasks);
                 break;
@@ -2420,6 +2418,14 @@ struct server_context {
         int32_t n_batch  = llama_n_batch(ctx);
         int32_t n_ubatch = llama_n_ubatch(ctx);
 
+        // there are currently slots with ongoing text generation
+        const bool is_tg = batch.n_tokens > 0;
+
+        // limit the batch to avoid blocking the processing
+        if (is_tg) {
+            n_batch = 32; // TODO: configurable
+        }
+
         // track if this is an embedding or non-embedding batch
         // if we've added sampled tokens above, we are in non-embedding mode
         // -1: none, 0: non-embedding, 1: embedding
@@ -2428,6 +2434,18 @@ struct server_context {
 
         // next, batch any pending prompts without exceeding n_batch
         if (params_base.cont_batching || batch.n_tokens == 0) {
+            // count how many slots are currently processing prompt
+            int n_slots_pp = 0;
+            for (auto & slot : slots) {
+                if (slot.state == SLOT_STATE_PROCESSING_PROMPT || slot.state == SLOT_STATE_STARTED) {
+                    n_slots_pp++;
+                }
+            }
+
+            // determine the chunk size of the chunk prefill
+            // a slot cannot submit more than this number of tokens in a single batch if other slots are processing
+            const int32_t n_chunk_pp = std::max(n_slots_pp > 0 ? (n_batch / n_slots_pp) : n_batch, 8);
+
             for (auto & slot : slots) {
                 // this slot still has a prompt to be processed
                 if (slot.state == SLOT_STATE_PROCESSING_PROMPT || slot.state == SLOT_STATE_STARTED) {
@@ -2611,8 +2629,10 @@ struct server_context {
                     // remove the non-common part from the cache
                     slot.cache_tokens.resize(slot.n_past);
 
+                    int n_cur = 0;
+
                     // add prompt tokens for processing in the current batch
-                    while (slot.n_past < slot.n_prompt_tokens && batch.n_tokens < n_batch) {
+                    while (slot.n_past < slot.n_prompt_tokens && batch.n_tokens < n_batch && n_cur < n_chunk_pp) {
                         common_batch_add(batch, prompt_tokens[slot.n_past], slot.n_past, { slot.id }, false);
 
                         if (slot.params.cache_prompt) {
@@ -2621,6 +2641,8 @@ struct server_context {
 
                         slot.n_prompt_tokens_processed++;
                         slot.n_past++;
+
+                        n_cur++;
                     }
 
                     SLT_INF(slot, "prompt processing progress, n_past = %d, n_tokens = %d, progress = %f\n", slot.n_past, batch.n_tokens, (float) slot.n_prompt_tokens_processed / slot.n_prompt_tokens);
@@ -3484,11 +3506,6 @@ int main(int argc, char ** argv) {
         json data = json::parse(req.body);
 
         // validate input
-        if (data.contains("prompt") && !data.at("prompt").is_string()) {
-            // prompt is optional
-            res_error(res, format_error_response("\"prompt\" must be a string", ERROR_TYPE_INVALID_REQUEST));
-        }
-
         if (!data.contains("input_prefix")) {
             res_error(res, format_error_response("\"input_prefix\" is required", ERROR_TYPE_INVALID_REQUEST));
         }
@@ -3498,11 +3515,9 @@ int main(int argc, char ** argv) {
         }
 
         if (data.contains("input_extra") && !data.at("input_extra").is_array()) {
-            // input_extra is optional
             res_error(res, format_error_response("\"input_extra\" must be an array of {\"filename\": string, \"text\": string}", ERROR_TYPE_INVALID_REQUEST));
             return;
         }
-
         json input_extra = json_value(data, "input_extra", json::array());
         for (const auto & chunk : input_extra) {
             // { "text": string, "filename": string }
@@ -3518,21 +3533,6 @@ int main(int argc, char ** argv) {
         }
         data["input_extra"] = input_extra; // default to empty array if it's not exist
 
-        std::string prompt = json_value(data, "prompt", std::string());
-        std::vector<llama_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.ctx, prompt, true, true);
-        SRV_DBG("creating infill tasks, n_prompts = %d\n", (int) tokenized_prompts.size());
-        data["prompt"] = format_infill(
-            ctx_server.ctx,
-            data.at("input_prefix"),
-            data.at("input_suffix"),
-            data.at("input_extra"),
-            ctx_server.params_base.n_batch,
-            ctx_server.params_base.n_predict,
-            ctx_server.slots[0].n_ctx, // TODO: there should be a better way
-            ctx_server.params_base.spm_infill,
-            tokenized_prompts[0]
-        );
-
         return handle_completions_generic(SERVER_TASK_TYPE_INFILL, data, res);
     };
 
@@ -3815,24 +3815,20 @@ int main(int argc, char ** argv) {
     // Router
     //
 
-    if (!params.webui) {
-        LOG_INF("Web UI is disabled\n");
-    } else {
-        // register static assets routes
-        if (!params.public_path.empty()) {
-            // Set the base directory for serving static files
-            bool is_found = svr->set_mount_point("/", params.public_path);
-            if (!is_found) {
-                LOG_ERR("%s: static assets path not found: %s\n", __func__, params.public_path.c_str());
-                return 1;
-            }
-        } else {
-            // using embedded static index.html
-            svr->Get("/", [](const httplib::Request &, httplib::Response & res) {
-                res.set_content(reinterpret_cast<const char*>(index_html), index_html_len, "text/html; charset=utf-8");
-                return false;
-            });
+    // register static assets routes
+    if (!params.public_path.empty()) {
+        // Set the base directory for serving static files
+        bool is_found = svr->set_mount_point("/", params.public_path);
+        if (!is_found) {
+            LOG_ERR("%s: static assets path not found: %s\n", __func__, params.public_path.c_str());
+            return 1;
         }
+    } else {
+        // using embedded static index.html
+        svr->Get("/", [](const httplib::Request &, httplib::Response & res) {
+            res.set_content(reinterpret_cast<const char*>(index_html), index_html_len, "text/html; charset=utf-8");
+            return false;
+        });
     }
 
     // register API routes

examples/server/tests/unit/test_basic.py

@@ -1,5 +1,4 @@
 import pytest
-import requests
 from utils import *
 
 server = ServerPreset.tinyllama2()
@@ -77,20 +76,3 @@ def test_load_split_model():
     })
     assert res.status_code == 200
     assert match_regex("(little|girl)+", res.body["content"])
-
-
-def test_no_webui():
-    global server
-    # default: webui enabled
-    server.start()
-    url = f"http://{server.server_host}:{server.server_port}"
-    res = requests.get(url)
-    assert res.status_code == 200
-    assert "<html>" in res.text
-    server.stop()
-
-    # with --no-webui
-    server.no_webui = True
-    server.start()
-    res = requests.get(url)
-    assert res.status_code == 404

examples/server/tests/unit/test_completion.py

@@ -42,16 +42,10 @@ def test_completion_stream(prompt: str, n_predict: int, re_content: str, n_promp
     })
     content = ""
     for data in res:
-        assert "stop" in data and type(data["stop"]) == bool
         if data["stop"]:
             assert data["timings"]["prompt_n"] == n_prompt
             assert data["timings"]["predicted_n"] == n_predicted
             assert data["truncated"] == truncated
-            assert data["stop_type"] == "limit"
-            assert "generation_settings" in data
-            assert server.n_predict is not None
-            assert data["generation_settings"]["n_predict"] == min(n_predict, server.n_predict)
-            assert data["generation_settings"]["seed"] == server.seed
             assert match_regex(re_content, content)
         else:
             content += data["content"]

examples/server/tests/unit/test_infill.py

@@ -13,28 +13,28 @@ def test_infill_without_input_extra():
     global server
     server.start()
     res = server.make_request("POST", "/infill", data={
-        "input_prefix": "#include <cstdio>\n#include \"llama.h\"\n\nint main() {\n",
-        "prompt": "    int n_threads = llama_",
+        "prompt": "Complete this",
+        "input_prefix": "#include <cstdio>\n#include \"llama.h\"\n\nint main() {\n    int n_threads = llama_",
         "input_suffix": "}\n",
     })
     assert res.status_code == 200
-    assert match_regex("(Ann|small|shiny)+", res.body["content"])
+    assert match_regex("(One|day|she|saw|big|scary|bird)+", res.body["content"])
 
 
 def test_infill_with_input_extra():
     global server
     server.start()
     res = server.make_request("POST", "/infill", data={
+        "prompt": "Complete this",
         "input_extra": [{
             "filename": "llama.h",
             "text": "LLAMA_API int32_t llama_n_threads();\n"
         }],
-        "input_prefix": "#include <cstdio>\n#include \"llama.h\"\n\nint main() {\n",
-        "prompt": "    int n_threads = llama_",
+        "input_prefix": "#include <cstdio>\n#include \"llama.h\"\n\nint main() {\n    int n_threads = llama_",
         "input_suffix": "}\n",
     })
     assert res.status_code == 200
-    assert match_regex("(Dad|excited|park)+", res.body["content"])
+    assert match_regex("(cuts|Jimmy|mom|came|into|the|room)+", res.body["content"])
 
 
 @pytest.mark.parametrize("input_extra", [
@@ -48,30 +48,10 @@ def test_invalid_input_extra_req(input_extra):
     global server
     server.start()
     res = server.make_request("POST", "/infill", data={
+        "prompt": "Complete this",
         "input_extra": [input_extra],
-        "input_prefix": "#include <cstdio>\n#include \"llama.h\"\n\nint main() {\n",
-        "prompt": "    int n_threads = llama_",
+        "input_prefix": "#include <cstdio>\n#include \"llama.h\"\n\nint main() {\n    int n_threads = llama_",
         "input_suffix": "}\n",
     })
     assert res.status_code == 400
     assert "error" in res.body
-
-
-@pytest.mark.skipif(not is_slow_test_allowed(), reason="skipping slow test")
-def test_with_qwen_model():
-    global server
-    server.model_file = None
-    server.model_hf_repo = "ggml-org/Qwen2.5-Coder-1.5B-IQ3_XXS-GGUF"
-    server.model_hf_file = "qwen2.5-coder-1.5b-iq3_xxs-imat.gguf"
-    server.start(timeout_seconds=600)
-    res = server.make_request("POST", "/infill", data={
-        "input_extra": [{
-            "filename": "llama.h",
-            "text": "LLAMA_API int32_t llama_n_threads();\n"
-        }],
-        "input_prefix": "#include <cstdio>\n#include \"llama.h\"\n\nint main() {\n",
-        "prompt": "    int n_threads = llama_",
-        "input_suffix": "}\n",
-    })
-    assert res.status_code == 200
-    assert res.body["content"] == "n_threads();\n    printf(\"Number of threads: %d\\n\", n_threads);\n    return 0;\n"

examples/server/tests/utils.py

@@ -72,7 +72,6 @@ class ServerProcess:
     disable_ctx_shift: int | None = False
     draft_min: int | None = None
     draft_max: int | None = None
-    no_webui: bool | None = None
 
     # session variables
     process: subprocess.Popen | None = None
@@ -159,8 +158,6 @@ class ServerProcess:
             server_args.extend(["--draft-max", self.draft_max])
         if self.draft_min:
             server_args.extend(["--draft-min", self.draft_min])
-        if self.no_webui:
-            server_args.append("--no-webui")
 
         args = [str(arg) for arg in [server_path, *server_args]]
         print(f"bench: starting server with: {' '.join(args)}")
@@ -374,6 +371,3 @@ def match_regex(regex: str, text: str) -> bool:
         ).search(text)
         is not None
     )
-
-def is_slow_test_allowed():
-    return os.environ.get("SLOW_TESTS") == "1" or os.environ.get("SLOW_TESTS") == "ON"

examples/server/utils.hpp

@@ -333,7 +333,7 @@ static std::string llama_get_chat_template(const struct llama_model * model) {
     if (res < 2) {
         return "";
     } else {
-        std::vector<char> model_template(res + 1, 0);
+        std::vector<char> model_template(res, 0);
         llama_model_meta_val_str(model, template_key.c_str(), model_template.data(), model_template.size());
         return std::string(model_template.data(), model_template.size() - 1);
     }

examples/server/webui/index.html

@@ -15,7 +15,7 @@
       <!-- sidebar -->
       <div class="drawer-side h-screen lg:h-screen z-50 lg:max-w-64">
         <label for="toggle-drawer" aria-label="close sidebar" class="drawer-overlay"></label>
-        <div class="flex flex-col bg-base-200 min-h-full max-w-64 py-4 px-4">
+        <div class="flex flex-col bg-base-200 min-h-full max-w-[calc(100vw-2em)] py-4 px-4">
           <div class="flex flex-row items-center justify-between mb-4 mt-4">
             <h2 class="font-bold ml-4">Conversations</h2>
 
@@ -120,25 +120,51 @@
             {{ messages.length === 0 ? 'Send a message to start' : '' }}
           </div>
           <div v-for="msg in messages" class="group">
-            <message-bubble
-              :config="config"
-              :msg="msg"
-              :key="msg.id"
-              :is-generating="isGenerating"
-              :edit-user-msg-and-regenerate="editUserMsgAndRegenerate"
-              :regenerate-msg="regenerateMsg"></message-bubble>
+            <div :class="{
+              'chat': true,
+              'chat-start': msg.role !== 'user',
+              'chat-end': msg.role === 'user',
+            }">
+              <div :class="{
+                'chat-bubble markdown': true,
+                'chat-bubble-base-300': msg.role !== 'user',
+              }">
+                <!-- textarea for editing message -->
+                <template v-if="editingMsg && editingMsg.id === msg.id">
+                  <textarea
+                    class="textarea textarea-bordered bg-base-100 text-base-content w-[calc(90vw-8em)] lg:w-96"
+                    v-model="msg.content"></textarea>
+                  <br/>
+                  <button class="btn btn-ghost mt-2 mr-2" @click="editingMsg = null">Cancel</button>
+                  <button class="btn mt-2" @click="editUserMsgAndRegenerate(msg)">Submit</button>
+                </template>
+                <!-- render message as markdown -->
+                <vue-markdown v-else :source="msg.content" />
+              </div>
+            </div>
+
+            <!-- actions for each message -->
+            <div :class="{'text-right': msg.role === 'user'}" class="mx-4 mt-2 mb-2">
+              <!-- user message -->
+              <button v-if="msg.role === 'user'" class="badge btn-mini show-on-hover" @click="editingMsg = msg" :disabled="isGenerating">
+                ✍️ Edit
+              </button>
+              <!-- assistant message -->
+              <button v-if="msg.role === 'assistant'" class="badge btn-mini show-on-hover mr-2" @click="regenerateMsg(msg)" :disabled="isGenerating">
+                🔄 Regenerate
+              </button>
+              <button v-if="msg.role === 'assistant'" class="badge btn-mini show-on-hover mr-2" @click="copyMsg(msg)" :disabled="isGenerating">
+                📋 Copy
+              </button>
+            </div>
           </div>
 
           <!-- pending (ongoing) assistant message -->
-          <div id="pending-msg" class="group">
-            <message-bubble
-              v-if="pendingMsg"
-              :config="config"
-              :msg="pendingMsg"
-              :key="pendingMsg.id"
-              :is-generating="isGenerating"
-              :edit-user-msg-and-regenerate="() => {}"
-              :regenerate-msg="() => {}"></message-bubble>
+          <div id="pending-msg" class="chat chat-start">
+            <div v-if="pendingMsg" class="chat-bubble markdown chat-bubble-base-300">
+              <span v-if="!pendingMsg.content" class="loading loading-dots loading-md"></span>
+              <vue-markdown v-else :source="pendingMsg.content" />
+            </div>
           </div>
         </div>
 
@@ -201,10 +227,6 @@
           <details class="collapse collapse-arrow bg-base-200 mb-2 overflow-visible">
             <summary class="collapse-title font-bold">Advanced config</summary>
             <div class="collapse-content">
-              <div class="flex flex-row items-center mb-2">
-                <input type="checkbox" class="checkbox" v-model="config.showTokensPerSecond" />
-                <span class="ml-4">Show tokens per second</span>
-              </div>
               <label class="form-control mb-2">
                 <!-- Custom parameters input -->
                 <div class="label inline">Custom JSON config (For more info, refer to <a class="underline" href="https://github.com/ggerganov/llama.cpp/blob/master/examples/server/README.md" target="_blank" rel="noopener noreferrer">server documentation</a>)</div>
@@ -225,66 +247,6 @@
 
   </div>
 
-
-  <!-- Template to be used as message bubble -->
-  <template id="message-bubble">
-    <div :class="{
-      'chat': true,
-      'chat-start': msg.role !== 'user',
-      'chat-end': msg.role === 'user',
-    }">
-      <div :class="{
-        'chat-bubble markdown': true,
-        'chat-bubble-base-300': msg.role !== 'user',
-      }">
-        <!-- textarea for editing message -->
-        <template v-if="editingContent !== null">
-          <textarea
-            class="textarea textarea-bordered bg-base-100 text-base-content w-[calc(90vw-8em)] lg:w-96"
-            v-model="editingContent"></textarea>
-          <br/>
-          <button class="btn btn-ghost mt-2 mr-2" @click="editingContent = null">Cancel</button>
-          <button class="btn mt-2" @click="editMsg()">Submit</button>
-        </template>
-        <template v-else>
-          <!-- show loading dots for pending message -->
-          <span v-if="msg.content === null" class="loading loading-dots loading-md"></span>
-          <!-- render message as markdown -->
-          <vue-markdown v-else :source="msg.content"></vue-markdown>
-          <!-- render timings if enabled -->
-          <div class="dropdown dropdown-hover dropdown-top mt-2" v-if="timings && config.showTokensPerSecond">
-            <div tabindex="0" role="button" class="cursor-pointer font-semibold text-sm opacity-60">Speed: {{ timings.predicted_per_second.toFixed(1) }} t/s</div>
-            <div class="dropdown-content bg-base-100 z-10 w-64 p-2 shadow mt-4">
-              <b>Prompt</b><br/>
-              - Tokens: {{ timings.prompt_n }}<br/>
-              - Time: {{ timings.prompt_ms }} ms<br/>
-              - Speed: {{ timings.prompt_per_second.toFixed(1) }} t/s<br/>
-              <b>Generation</b><br/>
-              - Tokens: {{ timings.predicted_n }}<br/>
-              - Time: {{ timings.predicted_ms }} ms<br/>
-              - Speed: {{ timings.predicted_per_second.toFixed(1) }} t/s<br/>
-            </div>
-          </div>
-        </template>
-      </div>
-    </div>
-    <!-- actions for each message -->
-    <div :class="{'text-right': msg.role === 'user', 'opacity-0': isGenerating}" class="mx-4 mt-2 mb-2">
-      <!-- user message -->
-      <button v-if="msg.role === 'user'" class="badge btn-mini show-on-hover" @click="editingContent = msg.content" :disabled="isGenerating">
-        ✍️ Edit
-      </button>
-      <!-- assistant message -->
-      <button v-if="msg.role === 'assistant'" class="badge btn-mini show-on-hover mr-2" @click="regenerateMsg(msg)" :disabled="isGenerating">
-        🔄 Regenerate
-      </button>
-      <button v-if="msg.role === 'assistant'" class="badge btn-mini show-on-hover mr-2" @click="copyMsg()" :disabled="isGenerating">
-        📋 Copy
-      </button>
-    </div>
-  </template>
-
-
   <!-- Template to be used by settings modal -->
   <template id="settings-modal-short-input">
     <label class="input input-bordered join-item grow flex items-center gap-2 mb-2">

examples/server/webui/package-lock.json

@@ -13,7 +13,6 @@
         "markdown-it": "^14.1.0",
         "postcss": "^8.4.49",
         "tailwindcss": "^3.4.15",
-        "textlinestream": "^1.1.1",
         "vite-plugin-singlefile": "^2.0.3",
         "vue": "^3.5.13"
       },
@@ -2678,12 +2677,6 @@
         "node": ">=14.0.0"
       }
     },
-    "node_modules/textlinestream": {
-      "version": "1.1.1",
-      "resolved": "https://registry.npmjs.org/textlinestream/-/textlinestream-1.1.1.tgz",
-      "integrity": "sha512-iBHbi7BQxrFmwZUQJsT0SjNzlLLsXhvW/kg7EyOMVMBIrlnj/qYofwo1LVLZi+3GbUEo96Iu2eqToI2+lZoAEQ==",
-      "license": "MIT"
-    },
     "node_modules/uc.micro": {
       "version": "2.1.0",
       "resolved": "https://registry.npmjs.org/uc.micro/-/uc.micro-2.1.0.tgz",

examples/server/webui/package.json

@@ -17,7 +17,6 @@
     "markdown-it": "^14.1.0",
     "postcss": "^8.4.49",
     "tailwindcss": "^3.4.15",
-    "textlinestream": "^1.1.1",
     "vite-plugin-singlefile": "^2.0.3",
     "vue": "^3.5.13"
   }

examples/server/webui/src/completion.js

@@ -0,0 +1,225 @@
+const paramDefaults = {
+  stream: true,
+  temperature: 0.2,
+};
+
+let generation_settings = null;
+
+export class CompletionError extends Error {
+  constructor(message, name, data) {
+    super(message);
+    this.name = name;
+  }
+};
+
+// Completes the prompt as a generator. Recommended for most use cases.
+//
+// Example:
+//
+//    import { llama } from '/completion.js'
+//
+//    const request = llama("Tell me a joke", {n_predict: 800})
+//    for await (const chunk of request) {
+//      document.write(chunk.data.content)
+//    }
+//
+export async function* llama(prompt, params = {}, config = {}) {
+  let controller = config.controller;
+  const api_url = config.api_url?.replace(/\/+$/, '') || "";
+
+  if (!controller) {
+    controller = new AbortController();
+  }
+
+  const completionParams = { ...paramDefaults, ...params, prompt };
+
+  const response = await fetch(`${api_url}${config.endpoint || '/completion'}`, {
+    method: 'POST',
+    body: JSON.stringify(completionParams),
+    headers: {
+      'Connection': 'keep-alive',
+      'Content-Type': 'application/json',
+      'Accept': 'text/event-stream',
+      ...(params.api_key ? {'Authorization': `Bearer ${params.api_key}`} : {})
+    },
+    signal: controller.signal,
+  });
+
+  const status = response.status;
+  if (status !== 200) {
+    try {
+      const body = await response.json();
+      if (body && body.error && body.error.message) {
+        throw new CompletionError(body.error.message, 'ServerError');
+      }
+    } catch (err) {
+      throw new CompletionError(err.message, 'ServerError');
+    }
+  }
+
+  const reader = response.body.getReader();
+  const decoder = new TextDecoder();
+
+  let content = "";
+  let leftover = ""; // Buffer for partially read lines
+
+  try {
+    let cont = true;
+
+    while (cont) {
+      const result = await reader.read();
+      if (result.done) {
+        break;
+      }
+
+      // Add any leftover data to the current chunk of data
+      const text = leftover + decoder.decode(result.value);
+
+      // Check if the last character is a line break
+      const endsWithLineBreak = text.endsWith('\n');
+
+      // Split the text into lines
+      let lines = text.split('\n');
+
+      // If the text doesn't end with a line break, then the last line is incomplete
+      // Store it in leftover to be added to the next chunk of data
+      if (!endsWithLineBreak) {
+        leftover = lines.pop();
+      } else {
+        leftover = ""; // Reset leftover if we have a line break at the end
+      }
+
+      // Parse all sse events and add them to result
+      const regex = /^(\S+):\s(.*)$/gm;
+      for (const line of lines) {
+        const match = regex.exec(line);
+        if (match) {
+          result[match[1]] = match[2];
+          if (result.data === '[DONE]') {
+            cont = false;
+            break;
+          }
+
+          // since we know this is llama.cpp, let's just decode the json in data
+          if (result.data) {
+            result.data = JSON.parse(result.data);
+            content += result.data.content;
+
+            // yield
+            yield result;
+
+            // if we got a stop token from server, we will break here
+            if (result.data.stop) {
+              if (result.data.generation_settings) {
+                generation_settings = result.data.generation_settings;
+              }
+              cont = false;
+              break;
+            }
+          }
+          if (result.error) {
+            try {
+              result.error = JSON.parse(result.error);
+              if (result.error.message.includes('slot unavailable')) {
+                // Throw an error to be caught by upstream callers
+                throw new Error('slot unavailable');
+              } else {
+                console.error(`llama.cpp error [${result.error.code} - ${result.error.type}]: ${result.error.message}`);
+              }
+            } catch(e) {
+              console.error(`llama.cpp error ${result.error}`)
+            }
+          }
+        }
+      }
+    }
+  } catch (e) {
+    if (e.name !== 'AbortError') {
+      console.error("llama error: ", e);
+    }
+    throw e;
+  }
+  finally {
+    controller.abort();
+  }
+
+  return content;
+}
+
+// Call llama, return an event target that you can subscribe to
+//
+// Example:
+//
+//    import { llamaEventTarget } from '/completion.js'
+//
+//    const conn = llamaEventTarget(prompt)
+//    conn.addEventListener("message", (chunk) => {
+//      document.write(chunk.detail.content)
+//    })
+//
+export const llamaEventTarget = (prompt, params = {}, config = {}) => {
+  const eventTarget = new EventTarget();
+  (async () => {
+    let content = "";
+    for await (const chunk of llama(prompt, params, config)) {
+      if (chunk.data) {
+        content += chunk.data.content;
+        eventTarget.dispatchEvent(new CustomEvent("message", { detail: chunk.data }));
+      }
+      if (chunk.data.generation_settings) {
+        eventTarget.dispatchEvent(new CustomEvent("generation_settings", { detail: chunk.data.generation_settings }));
+      }
+      if (chunk.data.timings) {
+        eventTarget.dispatchEvent(new CustomEvent("timings", { detail: chunk.data.timings }));
+      }
+    }
+    eventTarget.dispatchEvent(new CustomEvent("done", { detail: { content } }));
+  })();
+  return eventTarget;
+}
+
+// Call llama, return a promise that resolves to the completed text. This does not support streaming
+//
+// Example:
+//
+//     llamaPromise(prompt).then((content) => {
+//       document.write(content)
+//     })
+//
+//     or
+//
+//     const content = await llamaPromise(prompt)
+//     document.write(content)
+//
+export const llamaPromise = (prompt, params = {}, config = {}) => {
+  return new Promise(async (resolve, reject) => {
+    let content = "";
+    try {
+      for await (const chunk of llama(prompt, params, config)) {
+        content += chunk.data.content;
+      }
+      resolve(content);
+    } catch (error) {
+      reject(error);
+    }
+  });
+};
+
+/**
+ * (deprecated)
+ */
+export const llamaComplete = async (params, controller, callback) => {
+  for await (const chunk of llama(params.prompt, params, { controller })) {
+    callback(chunk);
+  }
+}
+
+// Get the model info from the server. This is useful for getting the context window and so on.
+export const llamaModelInfo = async (config = {}) => {
+  if (!generation_settings) {
+    const api_url = config.api_url?.replace(/\/+$/, '') || "";
+    const props = await fetch(`${api_url}/props`).then(r => r.json());
+    generation_settings = props.default_generation_settings;
+  }
+  return generation_settings;
+}

examples/server/webui/src/main.js

@@ -1,25 +1,21 @@
 import './styles.css';
 import { createApp, defineComponent, shallowRef, computed, h } from 'vue/dist/vue.esm-bundler.js';
+import { llama } from './completion.js';
 import MarkdownIt from 'markdown-it';
-import TextLineStream from 'textlinestream';
-
-const isDev = import.meta.env.MODE === 'development';
 
 // utility functions
 const isString = (x) => !!x.toLowerCase;
-const isBoolean = (x) => x === true || x === false;
-const isNumeric = (n) => !isString(n) && !isNaN(n) && !isBoolean(n);
+const isNumeric = (n) => !isString(n) && !isNaN(n);
 const escapeAttr = (str) => str.replace(/>/g, '>').replace(/"/g, '"');
 const copyStr = (str) => navigator.clipboard.writeText(str);
 
 // constants
 const BASE_URL = localStorage.getItem('base') // for debugging
-  || (new URL('.', document.baseURI).href).toString().replace(/\/$/, ''); // for production
+  || (new URL('.', document.baseURI).href).toString(); // for production
 const CONFIG_DEFAULT = {
   // Note: in order not to introduce breaking changes, please keep the same data type (number, string, etc) if you want to change the default value. Do not use null or undefined for default value.
   apiKey: '',
   systemMessage: 'You are a helpful assistant.',
-  showTokensPerSecond: false,
   // make sure these default values are in sync with `common.h`
   samplers: 'dkypmxt',
   temperature: 0.8,
@@ -105,48 +101,6 @@ const SettingsModalShortInput = defineComponent({
   },
 });
 
-// message bubble component
-const MessageBubble = defineComponent({
-  components: {
-    VueMarkdown
-  },
-  template: document.getElementById('message-bubble').innerHTML,
-  props: {
-    config: Object,
-    msg: Object,
-    isGenerating: Boolean,
-    editUserMsgAndRegenerate: Function,
-    regenerateMsg: Function,
-  },
-  data() {
-    return {
-      editingContent: null,
-    };
-  },
-  computed: {
-    timings() {
-      if (!this.msg.timings) return null;
-      return {
-        ...this.msg.timings,
-        prompt_per_second: this.msg.timings.prompt_n / (this.msg.timings.prompt_ms / 1000),
-        predicted_per_second: this.msg.timings.predicted_n / (this.msg.timings.predicted_ms / 1000),
-      };
-    }
-  },
-  methods: {
-    copyMsg() {
-      copyStr(this.msg.content);
-    },
-    editMsg() {
-      this.editUserMsgAndRegenerate({
-        ...this.msg,
-        content: this.editingContent,
-      });
-      this.editingContent = null;
-    },
-  },
-});
-
 // coversations is stored in localStorage
 // format: { [convId]: { id: string, lastModified: number, messages: [...] } }
 // convId is a string prefixed with 'conv-'
@@ -238,29 +192,10 @@ const chatScrollToBottom = (requiresNearBottom) => {
   }
 };
 
-// wrapper for SSE
-async function* sendSSEPostRequest(url, fetchOptions) {
-  const res = await fetch(url, fetchOptions);
-  const lines = res.body
-    .pipeThrough(new TextDecoderStream())
-    .pipeThrough(new TextLineStream());
-  for await (const line of lines) {
-    if (isDev) console.log({line});
-    if (line.startsWith('data:') && !line.endsWith('[DONE]')) {
-      const data = JSON.parse(line.slice(5));
-      yield data;
-    } else if (line.startsWith('error:')) {
-      const data = JSON.parse(line.slice(6));
-      throw new Error(data.message || 'Unknown error');
-    }
-  }
-};
-
 const mainApp = createApp({
   components: {
     VueMarkdown,
     SettingsModalShortInput,
-    MessageBubble,
   },
   data() {
     return {
@@ -274,6 +209,7 @@ const mainApp = createApp({
       selectedTheme: StorageUtils.getTheme(),
       config: StorageUtils.getConfig(),
       showConfigDialog: false,
+      editingMsg: null,
       // const
       themes: THEMES,
       configDefault: {...CONFIG_DEFAULT},
@@ -290,15 +226,6 @@ const mainApp = createApp({
     });
     resizeObserver.observe(pendingMsgElem);
   },
-  watch: {
-    viewingConvId: function(val, oldVal) {
-      if (val != oldVal) {
-        this.fetchMessages();
-        chatScrollToBottom();
-        this.hideSidebar();
-      }
-    }
-  },
   methods: {
     hideSidebar() {
       document.getElementById('toggle-drawer').checked = false;
@@ -310,10 +237,18 @@ const mainApp = createApp({
     newConversation() {
       if (this.isGenerating) return;
       this.viewingConvId = StorageUtils.getNewConvId();
+      this.editingMsg = null;
+      this.fetchMessages();
+      chatScrollToBottom();
+      this.hideSidebar();
     },
     setViewingConv(convId) {
       if (this.isGenerating) return;
       this.viewingConvId = convId;
+      this.editingMsg = null;
+      this.fetchMessages();
+      chatScrollToBottom();
+      this.hideSidebar();
     },
     deleteConv(convId) {
       if (this.isGenerating) return;
@@ -321,6 +256,7 @@ const mainApp = createApp({
         StorageUtils.remove(convId);
         if (this.viewingConvId === convId) {
           this.viewingConvId = StorageUtils.getNewConvId();
+          this.editingMsg = null;
         }
         this.fetchConversation();
         this.fetchMessages();
@@ -355,6 +291,7 @@ const mainApp = createApp({
       this.fetchConversation();
       this.fetchMessages();
       this.inputMsg = '';
+      this.editingMsg = null;
       this.generateMessage(currConvId);
       chatScrollToBottom();
     },
@@ -362,6 +299,7 @@ const mainApp = createApp({
       if (this.isGenerating) return;
       this.pendingMsg = { id: Date.now()+1, role: 'assistant', content: null };
       this.isGenerating = true;
+      this.editingMsg = null;
 
       try {
         const abortController = new AbortController();
@@ -392,21 +330,17 @@ const mainApp = createApp({
           dry_allowed_length: this.config.dry_allowed_length,
           dry_penalty_last_n: this.config.dry_penalty_last_n,
           max_tokens: this.config.max_tokens,
-          timings_per_token: !!this.config.showTokensPerSecond,
           ...(this.config.custom.length ? JSON.parse(this.config.custom) : {}),
+          ...(this.config.apiKey ? { api_key: this.config.apiKey } : {}),
         };
-        const chunks = sendSSEPostRequest(`${BASE_URL}/v1/chat/completions`, {
-          method: 'POST',
-          headers: {
-            'Content-Type': 'application/json',
-            'Authorization': this.config.apiKey ? `Bearer ${this.config.apiKey}` : undefined,
-          },
-          body: JSON.stringify(params),
-          signal: abortController.signal,
-        });
-        for await (const chunk of chunks) {
-          const stop = chunk.stop;
-          const addedContent = chunk.choices[0].delta.content;
+        const config = {
+          controller: abortController,
+          api_url: BASE_URL,
+          endpoint: '/chat/completions',
+        };
+        for await (const chunk of llama(prompt, params, config)) {
+          const stop = chunk.data.stop;
+          const addedContent = chunk.data.choices[0].delta.content;
           const lastContent = this.pendingMsg.content || '';
           if (addedContent) {
             this.pendingMsg = {
@@ -415,16 +349,6 @@ const mainApp = createApp({
               content: lastContent + addedContent,
             };
           }
-          const timings = chunk.timings;
-          if (timings && this.config.showTokensPerSecond) {
-            // only extract what's really needed, to save some space
-            this.pendingMsg.timings = {
-              prompt_n: timings.prompt_n,
-              prompt_ms: timings.prompt_ms,
-              predicted_n: timings.predicted_n,
-              predicted_ms: timings.predicted_ms,
-            };
-          }
         }
 
         StorageUtils.appendMsg(currConvId, this.pendingMsg);
@@ -463,10 +387,14 @@ const mainApp = createApp({
       this.fetchMessages();
       this.generateMessage(currConvId);
     },
+    copyMsg(msg) {
+      copyStr(msg.content);
+    },
     editUserMsgAndRegenerate(msg) {
       if (this.isGenerating) return;
       const currConvId = this.viewingConvId;
       const newContent = msg.content;
+      this.editingMsg = null;
       StorageUtils.filterAndKeepMsgs(currConvId, (m) => m.id < msg.id);
       StorageUtils.appendMsg(currConvId, {
         id: Date.now(),

examples/tokenize/tokenize.cpp

@@ -394,7 +394,7 @@ int main(int raw_argc, char ** raw_argv) {
     }
 
     if (show_token_count) {
-        printf("Total number of tokens: %zu\n", tokens.size());
+        printf("Total number of tokens: %ld\n", tokens.size());
     }
     // silence valgrind
     llama_free(ctx);

ggml/CMakeLists.txt

@@ -32,13 +32,6 @@ else()
     endif()
 endif()
 
-# remove the lib prefix on win32 mingw
-if (WIN32)
-    set(CMAKE_STATIC_LIBRARY_PREFIX "")
-    set(CMAKE_SHARED_LIBRARY_PREFIX "")
-    set(CMAKE_SHARED_MODULE_PREFIX  "")
-endif()
-
 option(BUILD_SHARED_LIBS "ggml: build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
 option(GGML_BACKEND_DL   "ggml: build backends as dynamic libraries (requires BUILD_SHARED_LIBS)" OFF)
 

ggml/include/ggml-backend.h

@@ -228,7 +228,6 @@ extern "C" {
     GGML_API void               ggml_backend_unload(ggml_backend_reg_t reg);
     // Load all known backends from dynamic libraries
     GGML_API void               ggml_backend_load_all(void);
-    GGML_API void               ggml_backend_load_all_from_path(const char * dir_path);
 
     //
     // Backend scheduler

ggml/src/CMakeLists.txt

@@ -194,6 +194,11 @@ endif()
 
 if (WIN32)
     add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
+
+    if (BUILD_SHARED_LIBS)
+        # TODO: should not use this
+        set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
+    endif()
 endif()
 
 # ggml

ggml/src/ggml-backend-reg.cpp

@@ -449,21 +449,11 @@ static std::string backend_filename_suffix() {
 #endif
 }
 
-static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent, const char * user_search_path) {
+static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent) {
     // enumerate all the files that match [lib]ggml-name-*.[so|dll] in the search paths
      // TODO: search system paths
+    std::vector<std::string> search_paths = { "./", get_executable_path() };
     std::string file_prefix = backend_filename_prefix() + name + "-";
-    std::vector<std::string> search_paths;
-    if (user_search_path == nullptr) {
-        search_paths.push_back("./");
-        search_paths.push_back(get_executable_path());
-    } else {
-#if defined(_WIN32)
-        search_paths.push_back(std::string(user_search_path) + "\\");
-#else
-        search_paths.push_back(std::string(user_search_path) + "/");
-#endif
-    }
 
     int best_score = 0;
     std::string best_path;
@@ -519,25 +509,21 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
 }
 
 void ggml_backend_load_all() {
-    ggml_backend_load_all_from_path(nullptr);
-}
-
-void ggml_backend_load_all_from_path(const char * dir_path) {
 #ifdef NDEBUG
     bool silent = true;
 #else
     bool silent = false;
 #endif
 
-    ggml_backend_load_best("blas", silent, dir_path);
-    ggml_backend_load_best("cann", silent, dir_path);
-    ggml_backend_load_best("cuda", silent, dir_path);
-    ggml_backend_load_best("hip", silent, dir_path);
-    ggml_backend_load_best("kompute", silent, dir_path);
-    ggml_backend_load_best("metal", silent, dir_path);
-    ggml_backend_load_best("rpc", silent, dir_path);
-    ggml_backend_load_best("sycl", silent, dir_path);
-    ggml_backend_load_best("vulkan", silent, dir_path);
-    ggml_backend_load_best("musa", silent, dir_path);
-    ggml_backend_load_best("cpu", silent, dir_path);
+    ggml_backend_load_best("blas", silent);
+    ggml_backend_load_best("cann", silent);
+    ggml_backend_load_best("cuda", silent);
+    ggml_backend_load_best("hip", silent);
+    ggml_backend_load_best("kompute", silent);
+    ggml_backend_load_best("metal", silent);
+    ggml_backend_load_best("rpc", silent);
+    ggml_backend_load_best("sycl", silent);
+    ggml_backend_load_best("vulkan", silent);
+    ggml_backend_load_best("musa", silent);
+    ggml_backend_load_best("cpu", silent);
 }

ggml/src/ggml-common.h

@@ -473,7 +473,7 @@ GGML_TABLE_BEGIN(uint8_t, ksigns_iq2xs, 128)
     240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255,
 GGML_TABLE_END()
 
-//#if __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A // lowest compute capability for integer intrinsics
+//#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
 GGML_TABLE_BEGIN(uint64_t, ksigns64, 128)
     0x0000000000000000, 0xff000000000000ff, 0xff0000000000ff00, 0x000000000000ffff,
     0xff00000000ff0000, 0x0000000000ff00ff, 0x0000000000ffff00, 0xff00000000ffffff,

ggml/src/ggml-cpu/amx/amx.cpp

@@ -122,7 +122,7 @@ static const char * ggml_backend_amx_buffer_type_get_name(ggml_backend_buffer_ty
 }
 
 static ggml_backend_buffer_t ggml_backend_amx_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    void * data = ggml_aligned_malloc(size);
+    void * data = aligned_alloc(TENSOR_ALIGNMENT, size);
     if (data == NULL) {
         fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size);
         return NULL;

ggml/src/ggml-cpu/ggml-cpu.c

@@ -126,7 +126,8 @@ struct ggml_arm_arch_features_type {
 #endif
 #include <windows.h>
 
-#if defined(_MSC_VER) && !defined(__clang__)
+
+#if !defined(__clang__)
 #define GGML_CACHE_ALIGN __declspec(align(GGML_CACHE_LINE))
 
 typedef volatile LONG atomic_int;
@@ -12944,7 +12945,7 @@ static thread_ret_t ggml_graph_compute_secondary_thread(void* data);
 #include "windows.h"
 
 // TODO: support > 64 CPUs
-static bool ggml_thread_apply_affinity(bool * mask) {
+bool ggml_thread_apply_affinity(bool * mask) {
     HANDLE    h = GetCurrentThread();
     uint64_t  bitmask = 0ULL;
 

ggml/src/ggml-cuda/common.cuh

@@ -41,28 +41,28 @@
 #define CUDART_HMAX   11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed)
 #define CUDART_HMASK  12000 // CUDA 12.0, min. ver. for half2 -> uint mask comparisons
 
-#define GGML_CUDA_CC_PASCAL     600
-#define GGML_CUDA_CC_DP4A       610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
-#define GGML_CUDA_CC_VOLTA      700
-#define GGML_CUDA_CC_TURING     750
-#define GGML_CUDA_CC_AMPERE     800
-#define GGML_CUDA_CC_OFFSET_AMD 1000000
+#define CC_PASCAL     600
+#define MIN_CC_DP4A   610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
+#define CC_VOLTA      700
+#define CC_TURING     750
+#define CC_AMPERE     800
+#define CC_OFFSET_AMD 1000000
 
 // GCN/CNDA, wave size is 64
-#define GGML_CUDA_CC_GCN4       (GGML_CUDA_CC_OFFSET_AMD + 803)  // Tonga, Fiji, Polaris, minimum for fast fp16
-#define GGML_CUDA_CC_VEGA       (GGML_CUDA_CC_OFFSET_AMD + 900)  // Vega56/64, minimum for fp16 dual issue
-#define GGML_CUDA_CC_VEGA20     (GGML_CUDA_CC_OFFSET_AMD + 906)  // MI50/Radeon VII, minimum for dp4a
-#define GGML_CUDA_CC_CDNA       (GGML_CUDA_CC_OFFSET_AMD + 908)  // MI100, minimum for MFMA, acc registers
-#define GGML_CUDA_CC_CDNA2      (GGML_CUDA_CC_OFFSET_AMD + 910)  // MI210, minimum acc register renameing
-#define GGML_CUDA_CC_CDNA3      (GGML_CUDA_CC_OFFSET_AMD + 942)  // MI300
+#define CC_GCN4       (CC_OFFSET_AMD + 803)  // Tonga, Fiji, Polaris, minimum for fast fp16
+#define CC_VEGA       (CC_OFFSET_AMD + 900)  // Vega56/64, minimum for fp16 dual issue
+#define CC_VEGA20     (CC_OFFSET_AMD + 906)  // MI50/Radeon VII, minimum for dp4a
+#define CC_CDNA       (CC_OFFSET_AMD + 908)  // MI100, minimum for MFMA, acc registers
+#define CC_CDNA2      (CC_OFFSET_AMD + 910)  // MI210, minimum acc register renameing
+#define CC_CDNA3      (CC_OFFSET_AMD + 942)  // MI300
 
 // RNDA removes MFMA, dp4a, xnack, acc registers, wave size is 32
-#define GGML_CUDA_CC_RDNA1      (GGML_CUDA_CC_OFFSET_AMD + 1010) // RX 5000
-#define GGML_CUDA_CC_RDNA2      (GGML_CUDA_CC_OFFSET_AMD + 1030) // RX 6000, minimum for dp4a
-#define GGML_CUDA_CC_RDNA3      (GGML_CUDA_CC_OFFSET_AMD + 1100) // RX 7000, minimum for WMMA
+#define CC_RDNA1      (CC_OFFSET_AMD + 1010) // RX 5000
+#define CC_RDNA2      (CC_OFFSET_AMD + 1030) // RX 6000, minimum for dp4a
+#define CC_RDNA3      (CC_OFFSET_AMD + 1100) // RX 7000, minimum for WMMA
 
-#define GGML_CUDA_CC_QY1        210
-#define GGML_CUDA_CC_QY2        220
+#define CC_QY1        210
+#define CC_QY2        220
 
 #define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
 
@@ -131,36 +131,36 @@ typedef float dfloat; // dequantize float
 typedef float2 dfloat2;
 #endif // GGML_CUDA_F16
 
-#if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
+#if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
 #define FP16_AVAILABLE
-#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
+#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
 
 #if defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
 #define FAST_FP16_AVAILABLE
 #endif // defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
 #define FP16_MMA_AVAILABLE
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
 #define INT8_MMA_AVAILABLE
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
 
-#if !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1)
+#if !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= CC_QY1)
 #define FLASH_ATTN_AVAILABLE
-#endif // !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1)
+#endif // !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= CC_QY1)
 
 static constexpr bool fast_fp16_available(const int cc) {
-    return cc >= GGML_CUDA_CC_PASCAL && cc != 610;
+    return cc >= CC_PASCAL && cc != 610;
 }
 
 static constexpr bool fp16_mma_available(const int cc) {
-    return cc < GGML_CUDA_CC_OFFSET_AMD && cc >= GGML_CUDA_CC_VOLTA;
+    return cc < CC_OFFSET_AMD && cc >= CC_VOLTA;
 }
 
 static constexpr bool int8_mma_available(const int cc) {
-    return cc < GGML_CUDA_CC_OFFSET_AMD && cc >= GGML_CUDA_CC_TURING;
+    return cc < CC_OFFSET_AMD && cc >= CC_TURING;
 }
 
 [[noreturn]]
@@ -187,7 +187,7 @@ static __device__ void no_device_code(
 #endif // __CUDA_ARCH__
 
 static __device__ __forceinline__ int warp_reduce_sum(int x) {
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
     return __reduce_add_sync(0xffffffff, x);
 #else
 #pragma unroll
@@ -195,7 +195,7 @@ static __device__ __forceinline__ int warp_reduce_sum(int x) {
         x += __shfl_xor_sync(0xffffffff, x, offset, 32);
     }
     return x;
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
 }
 
 static __device__ __forceinline__ float warp_reduce_sum(float x) {
@@ -284,7 +284,7 @@ static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const hal
 }
 
 static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
 #pragma unroll
    for (int offset = 16; offset > 0; offset >>= 1) {
        x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, offset, 32));
@@ -293,7 +293,7 @@ static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
 #else
    GGML_UNUSED(x);
    NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
 }
 
 #if CUDART_VERSION < CUDART_HMASK
@@ -333,13 +333,13 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
 
 #else // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 
-#if __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A
+#if __CUDA_ARCH__ >= MIN_CC_DP4A
     return __dp4a(a, b, c);
-#else // __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A
+#else // __CUDA_ARCH__ >= MIN_CC_DP4A
     const int8_t * a8 = (const int8_t *) &a;
     const int8_t * b8 = (const int8_t *) &b;
     return c + a8[0]*b8[0] + a8[1]*b8[1] + a8[2]*b8[2] + a8[3]*b8[3];
-#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A
+#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 
 #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 }

ggml/src/ggml-cuda/concat.cu

@@ -94,9 +94,7 @@ static void concat_f32_cuda(const float * x, const float * y, float * dst, int n
 }
 
 // non-contiguous kernel (slow)
-template <int dim>
-static __global__ void __launch_bounds__(CUDA_CONCAT_BLOCK_SIZE)
-    concat_f32_non_cont(
+static __global__ void concat_f32_non_cont(
         const char * src0,
         const char * src1,
               char * dst,
@@ -123,28 +121,22 @@ static __global__ void __launch_bounds__(CUDA_CONCAT_BLOCK_SIZE)
           uint64_t   nb0,
           uint64_t   nb1,
           uint64_t   nb2,
-          uint64_t   nb3){
-    static_assert(dim >= 0 && dim <= 3);
-
+          uint64_t   nb3,
+          int32_t   dim) {
     const int64_t i3 = blockIdx.z;
     const int64_t i2 = blockIdx.y;
     const int64_t i1 = blockIdx.x;
 
+    int64_t o[4] = {0, 0, 0, 0};
+    o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
+
     const float * x;
 
-    for (int64_t i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
+    for (int i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
         if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
             x = (const float *)(src0 + (i3       )*nb03 + (i2       )*nb02 + (i1       )*nb01 + (i0       )*nb00);
         } else {
-            if constexpr (dim == 0) {
-                x = (const float *) (src1 + i3 * nb13 + i2 * nb12 + i1 * nb11 + (i0 - ne00) * nb10);
-            } else if constexpr (dim == 1) {
-                x = (const float *) (src1 + i3 * nb13 + i2 * nb12 + (i1 - ne01) * nb11 + i0 * nb10);
-            } else if constexpr (dim == 2) {
-                x = (const float *) (src1 + i3 * nb13 + (i2 - ne02) * nb12 + i1 * nb11 + i0 * nb10);
-            } else if constexpr (dim == 3) {
-                x = (const float *) (src1 + (i3 - ne03) * nb13 + i2 * nb12 + i1 * nb11 + i0 * nb10);
-            }
+            x = (const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10);
         }
 
         float * y = (float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
@@ -190,32 +182,15 @@ void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
         }
     } else {
         dim3 grid_dim(dst->ne[1], dst->ne[2], dst->ne[3]);
-        auto launch_kernel = [&](auto dim) {
-            concat_f32_non_cont<dim><<<grid_dim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(
-                (const char *) src0->data, (const char *) src1->data, (char *) dst->data,
+        concat_f32_non_cont<<<grid_dim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(
+                (const char *)src0->data,
+                (const char *)src1->data,
+                (      char *)dst->data,
                 src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
                 src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
                 src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
                 src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3],
-                dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
-                dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3]);
-        };
-        switch (dim) {
-            case 0:
-                launch_kernel(std::integral_constant<int, 0>{});
-                break;
-            case 1:
-                launch_kernel(std::integral_constant<int, 1>{});
-                break;
-            case 2:
-                launch_kernel(std::integral_constant<int, 2>{});
-                break;
-            case 3:
-                launch_kernel(std::integral_constant<int, 3>{});
-                break;
-            default:
-                GGML_ABORT("Invalid dim: %d", dim);
-                break;
-        }
+                dst->ne[0],  dst->ne[1],  dst->ne[2],  dst->ne[3],
+                dst->nb[0],  dst->nb[1],  dst->nb[2],  dst->nb[3], dim);
     }
 }

ggml/src/ggml-cuda/convert.cu

@@ -26,7 +26,7 @@ static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __
 
 template <bool need_check>
 static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, half * __restrict__ y, const int64_t k) {
-#if __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
+#if __CUDA_ARCH__ >= CC_PASCAL
     constexpr int nint = CUDA_Q8_0_NE_ALIGN/sizeof(int) + WARP_SIZE;
 
     const int64_t   i0 = CUDA_Q8_0_NE_ALIGN*blockIdx.x;
@@ -64,7 +64,7 @@ static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, h
     GGML_UNUSED(y);
     GGML_UNUSED(k);
     NO_DEVICE_CODE;
-#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
+#endif // __CUDA_ARCH__ >= CC_PASCAL
 }
 
 template<typename dst_t>
@@ -599,7 +599,7 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
         case GGML_TYPE_Q5_1:
             return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
         case GGML_TYPE_Q8_0:
-            if (ggml_cuda_info().devices[ggml_cuda_get_device()].cc >= GGML_CUDA_CC_PASCAL) {
+            if (ggml_cuda_info().devices[ggml_cuda_get_device()].cc >= CC_PASCAL) {
                 return dequantize_block_q8_0_f16_cuda;
             }
             return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;

ggml/src/ggml-cuda/fattn.cu

@@ -304,7 +304,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
     const enum ggml_prec prec = ggml_flash_attn_ext_get_prec(KQV);
 
     // On AMD the tile kernels perform poorly, use the vec kernel instead:
-    if (cc >= GGML_CUDA_CC_OFFSET_AMD) {
+    if (cc >= CC_OFFSET_AMD) {
         if (prec == GGML_PREC_DEFAULT && fast_fp16_available(cc)) {
             ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
         } else {

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -177,7 +177,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
         info.devices[id].smpb  = prop.sharedMemPerBlock;
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
         info.devices[id].smpbo = prop.sharedMemPerBlock;
-        info.devices[id].cc = 100*prop.major + 10*prop.minor + GGML_CUDA_CC_OFFSET_AMD;
+        info.devices[id].cc = 100*prop.major + 10*prop.minor + CC_OFFSET_AMD;
 #else
         info.devices[id].smpbo = prop.sharedMemPerBlockOptin;
         info.devices[id].cc = 100*prop.major + 10*prop.minor;
@@ -1081,7 +1081,7 @@ static void ggml_cuda_op_mul_mat_cublas(
 
     const int compute_capability = ggml_cuda_info().devices[id].cc;
 
-    if (compute_capability >= GGML_CUDA_CC_VOLTA && (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT) {
+    if (compute_capability >= CC_VOLTA && (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT) {
         // convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32
         ggml_cuda_pool_alloc<half> src0_as_f16(ctx.pool(id));
         if (src0->type != GGML_TYPE_F16) {
@@ -1108,7 +1108,7 @@ static void ggml_cuda_op_mul_mat_cublas(
         const half beta_f16 = 0.0f;
 
         cublasComputeType_t cu_compute_type = CUBLAS_COMPUTE_16F;
-        if (ggml_cuda_info().devices[ctx.device].cc == GGML_CUDA_CC_CDNA) {
+        if (ggml_cuda_info().devices[ctx.device].cc == CC_CDNA) {
             cu_compute_type = CUBLAS_COMPUTE_32F;
         }
 
@@ -1612,7 +1612,7 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
     cublasComputeType_t cu_compute_type = CUBLAS_COMPUTE_16F;
     cudaDataType_t      cu_data_type    = CUDA_R_16F;
 
-    if (ggml_cuda_info().devices[ctx.device].cc == GGML_CUDA_CC_CDNA) {
+    if (ggml_cuda_info().devices[ctx.device].cc == CC_CDNA) {
         cu_compute_type = CUBLAS_COMPUTE_32F;
     }
 
@@ -2357,7 +2357,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
     std::vector<void *> ggml_cuda_cpy_fn_ptrs;
 
     if (cuda_ctx->cuda_graph->graph == nullptr) {
-        if (ggml_cuda_info().devices[cuda_ctx->device].cc < GGML_CUDA_CC_AMPERE) {
+        if (ggml_cuda_info().devices[cuda_ctx->device].cc < CC_AMPERE) {
             cuda_ctx->cuda_graph->disable_due_to_gpu_arch = true;
 #ifndef NDEBUG
             GGML_LOG_DEBUG("%s: disabling CUDA graphs due to GPU architecture\n", __func__);
@@ -3028,7 +3028,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 return true;
             }
             const int cc = ggml_cuda_info().devices[dev_ctx->device].cc;
-            return cc >= GGML_CUDA_CC_VOLTA && cc < GGML_CUDA_CC_OFFSET_AMD && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16;
+            return cc >= CC_VOLTA && cc < CC_OFFSET_AMD && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16;
         }
         case GGML_OP_CROSS_ENTROPY_LOSS:
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:

ggml/src/ggml-cuda/mma.cuh

@@ -171,7 +171,7 @@ struct mma_int_C_I16J8 {
 
     __device__ __forceinline__ void mma_K4(const mma_int_A_I16K4 & mma_A, const mma_int_B_J8K4 & mma_B) {
 #ifdef INT8_MMA_AVAILABLE
-#if __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#if __CUDA_ARCH__ >= CC_AMPERE
         asm("mma.sync.aligned.m16n8k16.row.col.s32.s8.s8.s32 {%0, %1, %2, %3}, {%4, %5}, {%6}, {%0, %1, %2, %3};"
             : "+r"(x[0]), "+r"(x[1]), "+r"(x[2]), "+r"(x[3])
             : "r"(mma_A.x[0]), "r"(mma_A.x[1]), "r"(mma_B.x[0]));
@@ -183,7 +183,7 @@ struct mma_int_C_I16J8 {
         asm("mma.sync.aligned.m8n8k16.row.col.s32.s8.s8.s32 {%0, %1}, {%2}, {%3}, {%0, %1};"
             : "+r"(x[2]), "+r"(x[3])
             : "r"(mma_A.x[1]), "r"(mma_B.x[0]));
-#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#endif // __CUDA_ARCH__ >= CC_AMPERE
 #else
         GGML_UNUSED(mma_A);
         GGML_UNUSED(mma_B);
@@ -193,7 +193,7 @@ struct mma_int_C_I16J8 {
 
     __device__ __forceinline__ void mma_K8(const mma_int_A_I16K8 & mma_A, const mma_int_B_J8K8 & mma_B) {
 #ifdef INT8_MMA_AVAILABLE
-#if __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#if __CUDA_ARCH__ >= CC_AMPERE
         asm("mma.sync.aligned.m16n8k32.row.col.s32.s8.s8.s32 {%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%0, %1, %2, %3};"
             : "+r"(x[0]), "+r"(x[1]), "+r"(x[2]), "+r"(x[3])
             : "r"(mma_A.x[0]), "r"(mma_A.x[1]), "r"(mma_A.x[2]), "r"(mma_A.x[3]), "r"(mma_B.x[0]), "r"(mma_B.x[1]));
@@ -211,7 +211,7 @@ struct mma_int_C_I16J8 {
         asm("mma.sync.aligned.m8n8k16.row.col.s32.s8.s8.s32 {%0, %1}, {%2}, {%3}, {%0, %1};"
             : "+r"(x[2]), "+r"(x[3])
             : "r"(mma_A.x[3]), "r"(mma_B.x[1]));
-#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#endif // __CUDA_ARCH__ >= CC_AMPERE
 #else
         GGML_UNUSED(mma_A);
         GGML_UNUSED(mma_B);

ggml/src/ggml-cuda/mmq.cu

@@ -27,7 +27,7 @@ void ggml_cuda_op_mul_mat_q(
     // The stream-k decomposition is only faster for recent NVIDIA GPUs.
     // Also its fixup needs to allocate a temporary buffer in the memory pool.
     // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
-    const bool use_stream_k = compute_capability >= GGML_CUDA_CC_VOLTA && compute_capability < GGML_CUDA_CC_OFFSET_AMD && src1_ncols == ne11;
+    const bool use_stream_k = compute_capability >= CC_VOLTA && compute_capability < CC_OFFSET_AMD && src1_ncols == ne11;
     const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst, use_stream_k};
 
     switch (src0->type) {
@@ -136,7 +136,7 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
         return true;
     }
 
-    if (cc < GGML_CUDA_CC_DP4A) {
+    if (cc < MIN_CC_DP4A) {
         return false;
     }
 
@@ -144,9 +144,9 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
     return true;
 #endif //GGML_CUDA_FORCE_MMQ
 
-    if (cc < GGML_CUDA_CC_OFFSET_AMD) {
-        return cc < GGML_CUDA_CC_VOLTA || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+    if (cc < CC_OFFSET_AMD) {
+        return cc < CC_VOLTA || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
     }
 
-    return (cc < GGML_CUDA_CC_RDNA3 && cc != GGML_CUDA_CC_CDNA && cc != GGML_CUDA_CC_VEGA20) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+    return (cc < CC_RDNA3 && cc != CC_CDNA && cc != CC_VEGA20) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
 }

ggml/src/ggml-cuda/mmq.cuh

@@ -89,9 +89,9 @@ struct tile_x_sizes {
 static constexpr int get_mmq_x_max_host(const int cc) {
     return int8_mma_available(cc) ? 128 :
 #ifdef GGML_CUDA_FORCE_MMQ
-        cc >= GGML_CUDA_CC_VOLTA && cc < GGML_CUDA_CC_OFFSET_AMD ? 128                     : 64;
+        cc >= CC_VOLTA && cc < CC_OFFSET_AMD ? 128                     : 64;
 #else
-        cc >= GGML_CUDA_CC_VOLTA && cc < GGML_CUDA_CC_OFFSET_AMD ? MMQ_DP4A_MAX_BATCH_SIZE : 64;
+        cc >= CC_VOLTA && cc < CC_OFFSET_AMD ? MMQ_DP4A_MAX_BATCH_SIZE : 64;
 #endif // GGML_CUDA_FORCE_MMQ
 }
 
@@ -104,23 +104,23 @@ static constexpr __device__ int get_mmq_x_max_device() {
     return 128;
 #else // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 
-#if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#if __CUDA_ARCH__ >= CC_VOLTA
 #ifdef GGML_CUDA_FORCE_MMQ
     return MMQ_DP4A_MAX_BATCH_SIZE;
 #else // GGML_CUDA_FORCE_MMQ
     return 128;
 #endif // GGML_CUDA_FORCE_MMQ
-#else // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#else // __CUDA_ARCH__ >= CC_VOLTA
 
     return 64;
-#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#endif // __CUDA_ARCH__ >= CC_VOLTA
 
 #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 #endif // INT8_MMA_AVAILABLE
 }
 
 static constexpr int get_mmq_y_host(const int cc) {
-    return cc >= GGML_CUDA_CC_OFFSET_AMD ? (cc == GGML_CUDA_CC_RDNA1 ? 64 : 128) : (cc >= GGML_CUDA_CC_VOLTA ? 128 : 64);
+    return cc >= CC_OFFSET_AMD ? (cc == CC_RDNA1 ? 64 : 128) : (cc >= CC_VOLTA ? 128 : 64);
 }
 
 static constexpr __device__ int get_mmq_y_device() {
@@ -131,11 +131,11 @@ static constexpr __device__ int get_mmq_y_device() {
     return 128;
 #endif // defined RDNA1
 #else
-#if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#if __CUDA_ARCH__ >= CC_VOLTA
     return 128;
 #else
     return 64;
-#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#endif // __CUDA_ARCH__ >= CC_VOLTA
 #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 }
 
@@ -2574,11 +2574,11 @@ template <ggml_type type, int mmq_x, int nwarps, bool need_check>
     __launch_bounds__(WARP_SIZE*nwarps, 2)
 #endif // defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
 #else
-#if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#if __CUDA_ARCH__ >= CC_VOLTA
     __launch_bounds__(WARP_SIZE*nwarps, 1)
 #else
     __launch_bounds__(WARP_SIZE*nwarps, 2)
-#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#endif // __CUDA_ARCH__ >= CC_VOLTA
 #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 static __global__ void mul_mat_q(
     const char * __restrict__ x, const char * __restrict__ yc, float * __restrict__ dst, float * __restrict__ tmp_fixup,
@@ -2594,7 +2594,7 @@ static __global__ void mul_mat_q(
     constexpr int mmq_y = get_mmq_y_device();
 
     // On AMD or old CUDA the performance with stream-k was worse, use conventional tiling instead:
-#if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA
+#if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
     {
         constexpr bool fixup = false;
         mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
@@ -2602,7 +2602,7 @@ static __global__ void mul_mat_q(
                 blockIdx.x, blockIdx.y, 0, ne00/qk);
         return;
     }
-#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA
+#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
 
     const     int64_t blocks_per_ne00 = ne00 / qk;
     constexpr int     blocks_per_iter = MMQ_ITER_K / qk;
@@ -2825,7 +2825,7 @@ void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cuda
     const int mmq_x_max = get_mmq_x_max_host(cc);
     const int mmq_y = get_mmq_y_host(cc);
     const int block_num_y = (args.ne01 + mmq_y - 1) / mmq_y;
-    const bool use_stream_k = cc >= GGML_CUDA_CC_VOLTA && cc < GGML_CUDA_CC_OFFSET_AMD;
+    const bool use_stream_k = cc >= CC_VOLTA && cc < CC_OFFSET_AMD;
 
     int mmq_x_best  = 0;
     int nparts_best = INT_MAX;

ggml/src/ggml-cuda/mmv.cu

@@ -57,7 +57,7 @@ static __global__ void mul_mat_vec(
     if (block_size > WARP_SIZE) {
         buf_iw[tid/WARP_SIZE] = sumf;
         __syncthreads();
-        if (tid >= WARP_SIZE) {
+        if (tid > WARP_SIZE) {
             return;
         }
         sumf = buf_iw[tid];

ggml/src/ggml-cuda/mmvq.cu

@@ -142,7 +142,7 @@ static void mul_mat_vec_q_cuda(
     int64_t nwarps = 1;
     int64_t rows_per_cuda_block = 1;
 
-    if (ggml_cuda_info().devices[id].cc < GGML_CUDA_CC_CDNA || ggml_cuda_info().devices[id].cc == GGML_CUDA_CC_RDNA1) { // NVIDIA and AMD older than RDNA2 but not CDNA
+    if (ggml_cuda_info().devices[id].cc < CC_CDNA || ggml_cuda_info().devices[id].cc == CC_RDNA1) { // NVIDIA and AMD older than RDNA2 but not CDNA
         switch(ncols_y) {
             case 1:
                 nwarps = 4;

ggml/src/ggml-cuda/sum.cu

@@ -3,6 +3,8 @@
 #endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700
 
 #ifdef USE_CUB
+// On Windows CUB uses libraries with variables called CC_PASCAL which conflict with the define in common.cuh.
+// For this reason CUB must be included BEFORE anything else.
 #include <cub/cub.cuh>
 using namespace cub;
 #endif // USE_CUB

ggml/src/ggml-impl.h

@@ -74,8 +74,8 @@ static inline int ggml_up(int n, int m) {
 //
 
 GGML_ATTRIBUTE_FORMAT(2, 3)
-GGML_API void ggml_log_internal        (enum ggml_log_level level, const char * format, ...);
-GGML_API void ggml_log_callback_default(enum ggml_log_level level, const char * text, void * user_data);
+void ggml_log_internal        (enum ggml_log_level level, const char * format, ...);
+void ggml_log_callback_default(enum ggml_log_level level, const char * text, void * user_data);
 
 #define GGML_LOG(...)       ggml_log_internal(GGML_LOG_LEVEL_NONE , __VA_ARGS__)
 #define GGML_LOG_INFO(...)  ggml_log_internal(GGML_LOG_LEVEL_INFO , __VA_ARGS__)
@@ -304,8 +304,8 @@ struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph, int i0, int i1);
 
 // Memory allocation
 
-GGML_API void * ggml_aligned_malloc(size_t size);
-GGML_API void ggml_aligned_free(void * ptr, size_t size);
+void * ggml_aligned_malloc(size_t size);
+void ggml_aligned_free(void * ptr, size_t size);
 
 // FP16 to FP32 conversion
 

ggml/src/ggml-threading.h

@@ -1,13 +1,11 @@
 #pragma once
 
-#include "ggml.h"
-
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-GGML_API void ggml_critical_section_start(void);
-GGML_API void ggml_critical_section_end(void);
+void ggml_critical_section_start(void);
+void ggml_critical_section_end(void);
 
 #ifdef __cplusplus
 }

ggml/src/ggml-vulkan/CMakeLists.txt

@@ -8,20 +8,6 @@ if (Vulkan_FOUND)
                              ../../include/ggml-vulkan.h
                             )
 
-    # Compile a test shader to determine whether GL_NV_cooperative_matrix2 is supported.
-    # If it's not, there will be an error to stderr.
-    # If it's supported, set a define to indicate that we should compile those shaders
-    execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_coopmat2_support.comp"
-                    OUTPUT_VARIABLE glslc_output
-                    ERROR_VARIABLE glslc_error)
-
-    if (${glslc_error} MATCHES ".*extension not supported: GL_NV_cooperative_matrix2.*")
-        message(STATUS "GL_NV_cooperative_matrix2 not supported by glslc")
-    else()
-        message(STATUS "GL_NV_cooperative_matrix2 supported by glslc")
-        add_compile_definitions(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
-    endif()
-
     target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan)
     target_include_directories(ggml-vulkan PRIVATE ${CMAKE_CURRENT_BINARY_DIR})
 
@@ -81,7 +67,7 @@ if (Vulkan_FOUND)
             --target-cpp ${_ggml_vk_source}
             --no-clean
 
-        DEPENDS ${_ggml_vk_shader_deps} ${_ggml_vk_genshaders_cmd}
+        DEPENDS ${_ggml_vk_shader_deps}
         COMMENT "Generate vulkan shaders"
     )
 

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -44,6 +44,12 @@
 
 #define MAX_VK_BUFFERS 256
 
+#ifndef K_QUANTS_PER_ITERATION
+#define K_QUANTS_PER_ITERATION 1
+#else
+static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
+#endif
+
 #define VK_CHECK(err, msg)                                          \
     do {                                                            \
         vk::Result err_ = (err);                                    \
@@ -162,12 +168,7 @@ struct vk_device_struct {
     uint32_t subgroup_size;
     uint32_t shader_core_count;
     bool uma;
-    bool float_controls_rte_fp16;
-
-    bool subgroup_size_control;
-    uint32_t subgroup_min_size;
-    uint32_t subgroup_max_size;
-    bool subgroup_require_full_support;
+    bool coopmat2;
 
     bool coopmat_support;
     bool coopmat_acc_f32_support;
@@ -175,7 +176,6 @@ struct vk_device_struct {
     uint32_t coopmat_m;
     uint32_t coopmat_n;
     uint32_t coopmat_k;
-    bool coopmat2;
 
     size_t idx;
 
@@ -427,7 +427,7 @@ static_assert(sizeof(vk_op_unary_push_constants) <= 128, "sizeof(vk_op_unary_pus
 // and a shift:
 //
 // n/d = (mulhi(n, mp) + n) >> L;
-static void init_fastdiv_values(uint32_t d, uint32_t &mp, uint32_t &L)
+void init_fastdiv_values(uint32_t d, uint32_t &mp, uint32_t &L)
 {
     // compute L = ceil(log2(d));
     L = 0;
@@ -439,7 +439,6 @@ static void init_fastdiv_values(uint32_t d, uint32_t &mp, uint32_t &L)
 }
 
 template <typename T> void init_pushconst_fastdiv(T &p) {
-    GGML_UNUSED(p);
     static_assert(!std::is_const<T>::value, "unexpected type");
 }
 
@@ -754,12 +753,8 @@ static uint32_t compile_count = 0;
 static std::mutex compile_count_mutex;
 static std::condition_variable compile_count_cond;
 
-static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipeline, const std::string name, size_t spv_size, const void* spv_data, const std::string entrypoint,
-                                         uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t> specialization_constants,
-                                         uint32_t align, bool disable_robustness, bool require_full_subgroups, uint32_t required_subgroup_size) {
-    VK_LOG_DEBUG("ggml_vk_create_pipeline(" << device->name << ", " << name << ", " << entrypoint << ", " << parameter_count << ", " << push_constant_size <<
-                 ", (" << wg_denoms[0] << "," << wg_denoms[1] << "," << wg_denoms[2] << "), specialization_constants, " << align <<
-                 ", " << disable_robustness << ", " << require_full_subgroups << ", " << required_subgroup_size << ")");
+static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipeline, const std::string name, size_t spv_size, const void* spv_data, const std::string entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t> specialization_constants, uint32_t align, bool disable_robustness) {
+    VK_LOG_DEBUG("ggml_vk_create_pipeline(" << device->name << ", " << name << ", " << entrypoint << ", " << parameter_count << ", " << push_constant_size << ", (" << wg_denoms[0] << "," << wg_denoms[1] << "," << wg_denoms[2] << "), specialization_constants, " << align << ")");
     GGML_ASSERT(parameter_count > 0);
     GGML_ASSERT(wg_denoms[0] > 0 && wg_denoms[1] > 0 && wg_denoms[2] > 0); // NOLINT
 
@@ -818,28 +813,14 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
         specialization_constants.data()
     );
 
-    vk::PipelineShaderStageCreateFlags pipeline_shader_stage_create_flags{};
-
-    if (device->subgroup_require_full_support && require_full_subgroups) {
-        pipeline_shader_stage_create_flags |= vk::PipelineShaderStageCreateFlagBits::eRequireFullSubgroupsEXT;
-    }
-
     vk::PipelineShaderStageCreateInfo pipeline_shader_create_info(
-            pipeline_shader_stage_create_flags,
+            vk::PipelineShaderStageCreateFlags(),
             vk::ShaderStageFlagBits::eCompute,
             pipeline->shader_module,
             entrypoint.c_str(),
             &specialization_info);
-
-    vk::PipelineShaderStageRequiredSubgroupSizeCreateInfoEXT pipeline_shader_stage_required_subgroup_size_create_info;
-    pipeline_shader_stage_required_subgroup_size_create_info.requiredSubgroupSize = required_subgroup_size;
-    if (device->subgroup_size_control && required_subgroup_size > 0) {
-        GGML_ASSERT(device->subgroup_min_size <= required_subgroup_size && required_subgroup_size <= device->subgroup_max_size);
-        pipeline_shader_create_info.setPNext(&pipeline_shader_stage_required_subgroup_size_create_info);
-    }
-
     vk::ComputePipelineCreateInfo compute_pipeline_create_info(
-        vk::PipelineCreateFlags{},
+        vk::PipelineCreateFlags(),
         pipeline_shader_create_info,
         pipeline->layout);
 
@@ -1519,9 +1500,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     device->pipeline_matmul_id_f32 = std::make_shared<vk_matmul_pipeline_struct>();
 
     std::vector<std::future<void>> compiles;
-    auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint,
-                                              uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, const std::vector<uint32_t>& specialization_constants,
-                                              uint32_t align, bool disable_robustness = false, bool require_full_subgroups = false, uint32_t required_subgroup_size = 0) {
+    auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, const std::vector<uint32_t>& specialization_constants, uint32_t align, bool disable_robustness = false) {
         {
             // wait until fewer than N compiles are in progress
             uint32_t N = std::max(1u, std::thread::hardware_concurrency());
@@ -1531,11 +1510,10 @@ static void ggml_vk_load_shaders(vk_device& device) {
             }
             compile_count++;
         }
-        compiles.push_back(std::async(ggml_vk_create_pipeline_func, std::ref(device), std::ref(pipeline), name, spv_size, spv_data, entrypoint,
-                                      parameter_count, push_constant_size, wg_denoms, specialization_constants, align, disable_robustness, require_full_subgroups, required_subgroup_size));
+        compiles.push_back(std::async(ggml_vk_create_pipeline_func, std::ref(device), std::ref(pipeline), name, spv_size, spv_data, entrypoint, parameter_count, push_constant_size, wg_denoms, specialization_constants, align, disable_robustness));
     };
 
-#if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+#if defined(VK_NV_cooperative_matrix2)
     if (device->coopmat2) {
 
         auto const &fa_wg_denoms = [&](uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::array<uint32_t, 3> {
@@ -1633,64 +1611,45 @@ static void ggml_vk_load_shaders(vk_device& device) {
 #undef CREATE_MM
 #undef CREATE_MM2
     } else
-#endif  // defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+#endif  // defined(VK_NV_cooperative_matrix2)
     if (device->coopmat_support) {
         // Create 6 variants, {s,m,l}x{unaligned,aligned}
 #define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
         if (device->mul_mat ## ID ## _l) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _coopmat_len, NAMELC ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _coopmat_len, NAMELC ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1);   \
         if (device->mul_mat ## ID ## _m) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _coopmat_len, NAMELC ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _coopmat_len, NAMELC ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1);   \
         if (device->mul_mat ## ID ## _s) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _coopmat_len, NAMELC ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _coopmat_len, NAMELC ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1);   \
         if (device->mul_mat ## ID ## _l) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _coopmat_len, NAMELC ## _aligned ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _coopmat_len, NAMELC ## _aligned ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align);   \
         if (device->mul_mat ## ID ## _m) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _coopmat_len, NAMELC ## _aligned ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _coopmat_len, NAMELC ## _aligned ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align);   \
         if (device->mul_mat ## ID ## _s) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _coopmat_len, NAMELC ## _aligned ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _coopmat_len, NAMELC ## _aligned ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align);   \
 
         // Create 2 variants, {f16,f32} accumulator
 #define CREATE_MM2(PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
-        if (device->coopmat_acc_f16_support) { \
-            CREATE_MM(PIPELINE_NAME . f16acc, NAMELC, _f16acc, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
-        } \
-        if (device->coopmat_acc_f32_support) { \
-            CREATE_MM(PIPELINE_NAME . f32acc, NAMELC, , WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
-        } \
+        CREATE_MM(PIPELINE_NAME . f16acc, NAMELC, _f16acc, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
+        CREATE_MM(PIPELINE_NAME . f32acc, NAMELC, , WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
 
         CREATE_MM(pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
         CREATE_MM(pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
         CREATE_MM2(pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
         CREATE_MM2(pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
 
-        if (device->coopmat_acc_f16_support) {
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
 
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-        } else {
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-        }
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
 
         // If there's not enough shared memory for row_ids and the result tile, don't create these pipelines.
         if (device->mul_mat_id_s || device->mul_mat_id_m || device->mul_mat_id_l) {
@@ -1698,35 +1657,19 @@ static void ggml_vk_load_shaders(vk_device& device) {
             CREATE_MM2(pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
             CREATE_MM2(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
 
-            if (device->coopmat_acc_f16_support) {
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
 
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            } else {
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            }
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         }
-#undef CREATE_MM2
 #undef CREATE_MM
     } else if (device->fp16) {
         // Create 6 variants, {s,m,l}x{unaligned,aligned}
@@ -1744,11 +1687,6 @@ static void ggml_vk_load_shaders(vk_device& device) {
         if (device->mul_mat ## ID ## _s) \
             ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align);   \
 
-        // Create 2 variants, {f16,f32} accumulator
-#define CREATE_MM2(PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
-        CREATE_MM(PIPELINE_NAME . f16acc, NAMELC, _f16acc, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
-        CREATE_MM(PIPELINE_NAME . f32acc, NAMELC, , WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
-
         CREATE_MM(pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
         CREATE_MM(pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
         CREATE_MM2(pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
@@ -1786,7 +1724,6 @@ static void ggml_vk_load_shaders(vk_device& device) {
             CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
             CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         }
-#undef CREATE_MM2
 #undef CREATE_MM
     } else {
         // Create 6 variants, {s,m,l}x{unaligned,aligned}
@@ -1841,6 +1778,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
             CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f32acc, matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
             CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc, matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         }
+#undef CREATE_MM2
 #undef CREATE_MM
     }
 
@@ -1853,10 +1791,10 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32_f32", mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32_f32", mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32_f32", mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f32_f32", mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f32_f32", mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f32_f32", mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f32_f32", mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f32_f32", mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
 
@@ -1867,10 +1805,10 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f16_f32", mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f16_f32", mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f16_f32", mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f16_f32", mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f16_f32", mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f16_f32", mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f16_f32", mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f16_f32", mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size}, 1, true);
 
@@ -1881,10 +1819,10 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
 
@@ -1983,26 +1921,17 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16_wg512, "soft_max_f32_f16_wg512", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f32, "rope_norm_f32", rope_norm_f32_len, rope_norm_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f32, "rope_neox_f32", rope_neox_f32_len, rope_neox_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f16, "rope_norm_f16", rope_norm_f16_len, rope_norm_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
 
-    if (device->float_controls_rte_fp16) {
-        ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f16, "rope_norm_f16", rope_norm_f16_rte_len, rope_norm_f16_rte_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_rte_len, rope_neox_f16_rte_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-    } else {
-        ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f16, "rope_norm_f16", rope_norm_f16_len, rope_norm_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_len, rope_neox_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-    }
+    ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f32, "rope_neox_f32", rope_neox_f32_len, rope_neox_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_len, rope_neox_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_argsort_f32, "argsort_f32", argsort_f32_len, argsort_f32_data, "main", 2, sizeof(vk_op_argsort_push_constants), {1024, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_im2col_f32, "im2col_f32", im2col_f32_len, im2col_f32_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1);
-    if (device->float_controls_rte_fp16) {
-        ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_rte_len, im2col_f32_f16_rte_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1);
-    } else {
-        ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_len, im2col_f32_f16_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1);
-    }
+    ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_len, im2col_f32_f16_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_timestep_embedding_f32, "timestep_embedding_f32", timestep_embedding_f32_len, timestep_embedding_f32_data, "main", 2, sizeof(vk_op_timestep_embedding_push_constants), {256, 1, 1}, {}, 1);
 
@@ -2064,8 +1993,6 @@ static vk_device ggml_vk_get_device(size_t idx) {
                 amd_shader_core_properties2 = true;
             } else if (strcmp("VK_EXT_pipeline_robustness", properties.extensionName) == 0) {
                 pipeline_robustness = true;
-            } else if (strcmp("VK_EXT_subgroup_size_control", properties.extensionName) == 0) {
-                device->subgroup_size_control = true;
             } else if (strcmp("VK_KHR_cooperative_matrix", properties.extensionName) == 0 &&
                        !getenv("GGML_VK_DISABLE_COOPMAT")) {
                 device->coopmat_support = true;
@@ -2085,15 +2012,11 @@ static vk_device ggml_vk_get_device(size_t idx) {
         vk::PhysicalDeviceDriverProperties driver_props;
         vk::PhysicalDeviceShaderSMBuiltinsPropertiesNV sm_props;
         vk::PhysicalDeviceShaderCoreProperties2AMD amd_shader_core_properties2_props;
-        vk::PhysicalDeviceVulkan12Properties vk12_props;
-        vk::PhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_size_control_props;
-
         props2.pNext = &props3;
         props3.pNext = &subgroup_props;
         subgroup_props.pNext = &driver_props;
-        driver_props.pNext = &vk12_props;
 
-        VkBaseOutStructure * last_struct = (VkBaseOutStructure *)&vk12_props;
+        VkBaseOutStructure * last_struct = (VkBaseOutStructure *)&driver_props;
 
         if (maintenance4_support) {
             last_struct->pNext = (VkBaseOutStructure *)&props4;
@@ -2107,10 +2030,6 @@ static vk_device ggml_vk_get_device(size_t idx) {
             last_struct->pNext = (VkBaseOutStructure *)&amd_shader_core_properties2_props;
             last_struct = (VkBaseOutStructure *)&amd_shader_core_properties2_props;
         }
-        if (device->subgroup_size_control) {
-            last_struct->pNext = (VkBaseOutStructure *)&subgroup_size_control_props;
-            last_struct = (VkBaseOutStructure *)&subgroup_size_control_props;
-        }
 
 #if defined(VK_NV_cooperative_matrix2)
         vk::PhysicalDeviceCooperativeMatrix2PropertiesNV coopmat2_props;
@@ -2143,13 +2062,12 @@ static vk_device ggml_vk_get_device(size_t idx) {
         } else {
             device->shader_core_count = 0;
         }
-        device->float_controls_rte_fp16 = vk12_props.shaderRoundingModeRTEFloat16;
 
         const bool force_disable_f16 = getenv("GGML_VK_DISABLE_F16") != nullptr;
 
         device->fp16 = !force_disable_f16 && fp16_storage && fp16_compute;
 
-        if (device->vendor_id == VK_VENDOR_ID_INTEL || (device->vendor_id == VK_VENDOR_ID_AMD && (driver_props.driverID == vk::DriverId::eAmdProprietary || driver_props.driverID == vk::DriverId::eAmdOpenSource))) {
+        if (device->vendor_id == VK_VENDOR_ID_INTEL || (props2.properties.vendorID == VK_VENDOR_ID_AMD && driver_props.driverID == vk::DriverId::eAmdProprietary)) {
             // Intel drivers don't support coopmat properly yet
             // Only RADV supports coopmat properly on AMD
             device->coopmat_support = false;
@@ -2205,17 +2123,6 @@ static vk_device ggml_vk_get_device(size_t idx) {
             device_extensions.push_back("VK_EXT_pipeline_robustness");
         }
 
-        VkPhysicalDeviceSubgroupSizeControlFeaturesEXT subgroup_size_control_features;
-        subgroup_size_control_features.pNext = nullptr;
-        subgroup_size_control_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT;
-        subgroup_size_control_features.computeFullSubgroups = false;
-        subgroup_size_control_features.subgroupSizeControl = false;
-
-        if (device->subgroup_size_control) {
-            last_struct->pNext = (VkBaseOutStructure *)&subgroup_size_control_features;
-            last_struct = (VkBaseOutStructure *)&subgroup_size_control_features;
-        }
-
         VkPhysicalDeviceCooperativeMatrixFeaturesKHR coopmat_features;
         coopmat_features.pNext = nullptr;
         coopmat_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_FEATURES_KHR;
@@ -2243,21 +2150,10 @@ static vk_device ggml_vk_get_device(size_t idx) {
 
         device->pipeline_robustness = pl_robustness_features.pipelineRobustness;
 
-        device->subgroup_size_control = device->subgroup_size_control &&
-                (subgroup_size_control_props.requiredSubgroupSizeStages & vk::ShaderStageFlagBits::eCompute) &&
-                subgroup_size_control_features.subgroupSizeControl;
-
-        if (device->subgroup_size_control) {
-            device->subgroup_min_size = subgroup_size_control_props.minSubgroupSize;
-            device->subgroup_max_size = subgroup_size_control_props.maxSubgroupSize;
-            device->subgroup_require_full_support = subgroup_size_control_features.computeFullSubgroups;
-            device_extensions.push_back("VK_EXT_subgroup_size_control");
-        }
-
         device->coopmat_support = device->coopmat_support && coopmat_features.cooperativeMatrix;
 
         if (coopmat2_support) {
-#if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+#if defined(VK_NV_cooperative_matrix2)
             if (coopmat2_features.cooperativeMatrixWorkgroupScope &&
                 coopmat2_features.cooperativeMatrixFlexibleDimensions &&
                 coopmat2_features.cooperativeMatrixReductions &&
@@ -2403,7 +2299,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
                 }
             }
 
-            if (device->coopmat_m == 0 || !device->coopmat_acc_f32_support) {
+            if (device->coopmat_m == 0) {
                 // No suitable matmul mode found
                 GGML_LOG_DEBUG("ggml_vulkan: WARNING: No suitable matrix core mode found. Disabling matrix cores.\n");
                 device->coopmat_support = false;
@@ -2518,25 +2414,18 @@ static void ggml_vk_print_gpu_info(size_t idx) {
     bool fp16_storage = false;
     bool fp16_compute = false;
     bool coopmat_support = false;
-    bool coopmat2_support = false;
 
     for (auto properties : ext_props) {
         if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) {
             fp16_storage = true;
         } else if (strcmp("VK_KHR_shader_float16_int8", properties.extensionName) == 0) {
             fp16_compute = true;
-        } else if (strcmp("VK_KHR_cooperative_matrix", properties.extensionName) == 0 &&
-                   !getenv("GGML_VK_DISABLE_COOPMAT")) {
+        } else if (strcmp("VK_KHR_cooperative_matrix", properties.extensionName) == 0) {
             coopmat_support = true;
-#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
-        } else if (strcmp("VK_NV_cooperative_matrix2", properties.extensionName) == 0 &&
-                   !getenv("GGML_VK_DISABLE_COOPMAT2")) {
-            coopmat2_support = true;
-#endif
         }
     }
 
-    if (props2.properties.vendorID == VK_VENDOR_ID_INTEL || (props2.properties.vendorID == VK_VENDOR_ID_AMD && (driver_props.driverID == vk::DriverId::eAmdProprietary || driver_props.driverID == vk::DriverId::eAmdOpenSource))) {
+    if (props2.properties.vendorID == VK_VENDOR_ID_INTEL || (props2.properties.vendorID == VK_VENDOR_ID_AMD && driver_props.driverID == vk::DriverId::eAmdProprietary)) {
         // Intel drivers don't support coopmat properly yet
         // Only RADV supports coopmat properly on AMD
         coopmat_support = false;
@@ -2583,11 +2472,9 @@ static void ggml_vk_print_gpu_info(size_t idx) {
 
     coopmat_support = coopmat_support && coopmat_features.cooperativeMatrix;
 
-    std::string matrix_cores = coopmat2_support ? "NV_coopmat2" : coopmat_support ? "KHR_coopmat" : "none";
-
     std::string device_name = props2.properties.deviceName.data();
-    GGML_LOG_DEBUG("ggml_vulkan: %zu = %s (%s) | uma: %d | fp16: %d | warp size: %zu | matrix cores: %s\n",
-              idx, device_name.c_str(), driver_props.driverName.data(), uma, fp16, subgroup_size, matrix_cores.c_str());
+    GGML_LOG_DEBUG("ggml_vulkan: %zu = %s (%s) | uma: %d | fp16: %d | warp size: %zu | matrix cores: %d\n",
+              idx, device_name.c_str(), driver_props.driverName.data(), uma, fp16, subgroup_size, coopmat_support);
 
     if (props2.properties.deviceType == vk::PhysicalDeviceType::eCpu) {
         GGML_LOG_DEBUG("ggml_vulkan: Warning: Device type is CPU. This is probably not the device you want.\n");
@@ -2823,7 +2710,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
     if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
         return ctx->device->pipeline_matmul_f32_f16;
     }
-    if (prec == GGML_PREC_DEFAULT && ctx->device->fp16 && !(ctx->device->coopmat_support && !ctx->device->coopmat_acc_f16_support)) {
+    if (prec == GGML_PREC_DEFAULT && ctx->device->fp16) {
         if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
             return ctx->device->pipeline_matmul_f16_f32.f16acc;
         }
@@ -2898,7 +2785,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
     if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
         return ctx->device->pipeline_matmul_id_f32;
     }
-    if (prec == GGML_PREC_DEFAULT && ctx->device->fp16 && !(ctx->device->coopmat_support && !ctx->device->coopmat_acc_f16_support)) {
+    if (prec == GGML_PREC_DEFAULT && ctx->device->fp16) {
         if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
             return ctx->device->pipeline_matmul_id_f16_f32.f16acc;
         }
@@ -3523,7 +3410,7 @@ static uint32_t ggml_vk_guess_split_k(ggml_backend_vk_context * ctx, int m, int
     return split_k;
 }
 
-static vk_pipeline ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned) {
+static vk_pipeline ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned, ggml_type type_a) {
     VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline(" << m << ", " << n << ", " << aligned << ")");
 
     if (ctx->device->coopmat2) {
@@ -3545,9 +3432,9 @@ static vk_pipeline ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx,
     return aligned ? mmp->a_l : mmp->l;
 }
 
-static uint32_t ggml_vk_guess_matmul_pipeline_align(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n) {
+static uint32_t ggml_vk_guess_matmul_pipeline_align(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, ggml_type type_a) {
     VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline_align(" << m << ", " << n << ")");
-    return ggml_vk_guess_matmul_pipeline(ctx, mmp, m, n, true)->align;
+    return ggml_vk_guess_matmul_pipeline(ctx, mmp, m, n, true, type_a)->align;
 }
 
 static void ggml_vk_matmul(
@@ -3677,7 +3564,6 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context&
         (uint32_t)tensor->ne[0], (uint32_t)tensor->ne[1], (uint32_t)tensor->ne[2], (uint32_t)tensor->ne[3],                       1                   , (uint32_t)tensor->ne[0]                   , (uint32_t)(tensor->ne[0] * tensor->ne[1]) , (uint32_t)(tensor->ne[0] * tensor->ne[1] * tensor->ne[2]),
         0,
         0.0f, 0.0f,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     };
     init_pushconst_fastdiv(pc);
     ggml_vk_sync_buffers(subctx);
@@ -3751,10 +3637,10 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     const int y_ne = ne11 * ne10;
     const int d_ne = ne11 * ne01;
 
-    const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11));
+    const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11, src0->type));
     const bool aligned = ne10 == kpad && ne01 > 8 && ne11 > 8;
 
-    vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned);
+    vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned, src0->type);
 
     const uint32_t split_k = ggml_vk_guess_split_k(ctx, ne01, ne11, ne10, pipeline);
 
@@ -5458,8 +5344,7 @@ static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context& subctx, con
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
-        op_params[0], 0.0f,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+        op_params[0], 0.0f
     }, dryrun);
 }
 
@@ -5473,7 +5358,6 @@ static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context& subctx, const
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         0.0f, 0.0f,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     }, dryrun);
 }
 
@@ -5487,7 +5371,6 @@ static void ggml_vk_sin(ggml_backend_vk_context * ctx, vk_context& subctx, const
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         0.0f, 0.0f,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     }, dryrun);
 }
 
@@ -5501,7 +5384,6 @@ static void ggml_vk_cos(ggml_backend_vk_context * ctx, vk_context& subctx, const
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         0.0f, 0.0f,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     }, dryrun);
 }
 
@@ -5516,7 +5398,6 @@ static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context& subctx, con
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         op_params[0], op_params[1],
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     }, dryrun);
 }
 
@@ -5530,7 +5411,6 @@ static void ggml_vk_pad(ggml_backend_vk_context * ctx, vk_context& subctx, const
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         0.0f, 0.0f,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     }, dryrun);
 }
 
@@ -5544,7 +5424,6 @@ static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context& subctx, co
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         0.0f, 0.0f,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     }, dryrun);
 }
 
@@ -5559,7 +5438,6 @@ static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         d_offset,
         0.0f, 0.0f,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     }, dryrun);
 }
 

ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp

@@ -9,8 +9,8 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
-        const uint ib = gl_WorkGroupID.x * 256 + wgy;
-        if (ib >= p.M * p.K / QUANT_K) {
+        const uint i = gl_WorkGroupID.x * 256 + wgy;
+        if (i >= p.M * p.K / QUANT_K) {
             return;
         }
 
@@ -20,49 +20,37 @@ void main() {
         const uint is = 2 * il;
         const uint n = 4;
 
-        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].d.y);
+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
 
-        const uint y_idx = ib * QUANT_K + 64 * il + n * ir;
+        const uint y_idx = i * QUANT_K + 64 * il + n * ir;
         const uint qs_idx = 32*il + n * ir;
 
-        uint scidx0 = (is < 4) ? is : (is + 4);
-        uint scidx1 = (is < 4) ? is : (is - 4);
-        uint scidxmask1 = (is < 4) ? 0x30 : 0xC0;
-        uint scidxshift1 = (is < 4) ? 0 : 2;
-        uint mbidx0 = is + 4;
-        uint mbidx1 = (is < 4) ? is + 4 : is;
-        uint mbidxmask0 = (is < 4) ? 0xF : 0xF0;
-        uint mbidxshift0 = (is < 4) ? 0 : 4;
-        uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
-        uint mbidxshift1 = (is < 4) ? 0 : 2;
-
-        uint8_t sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1));
-        uint8_t mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
-
+        uint8_t sc;
+        uint8_t m;
+        if (is < 4) {
+            sc = uint8_t(data_a[i].scales[is] & 63);
+            m  = uint8_t(data_a[i].scales[is + 4] & 63);
+        } else {
+            sc = uint8_t((data_a[i].scales[is + 4] & 0xF) | ((data_a[i].scales[is - 4] >> 6) << 4));
+            m  = uint8_t((data_a[i].scales[is + 4] >>  4) | ((data_a[i].scales[is    ] >> 6) << 4));
+        }
         const FLOAT_TYPE d1 = dall * sc;
-        const FLOAT_TYPE m1 = dmin * mbyte;
-
-        scidx0 = (is < 4) ? is + 1 : (is + 5);
-        scidx1 = (is < 4) ? is + 1 : (is - 3);
-        scidxmask1 = (is < 4) ? 0x30 : 0xC0;
-        scidxshift1 = (is < 4) ? 0 : 2;
-        mbidx0 = is + 5;
-        mbidx1 = (is < 4) ? is + 5 : is + 1;
-        mbidxmask0 = (is < 4) ? 0xF : 0xF0;
-        mbidxshift0 = (is < 4) ? 0 : 4;
-        mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
-        mbidxshift1 = (is < 4) ? 0 : 2;
-
-        sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1));
-        mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
+        const FLOAT_TYPE m1 = dmin * m;
 
+        if (is < 4) {
+            sc = uint8_t(data_a[i].scales[is + 1] & 63);
+            m  = uint8_t(data_a[i].scales[is + 5] & 63);
+        } else {
+            sc = uint8_t((data_a[i].scales[is + 5] & 0xF) | ((data_a[i].scales[is - 3] >> 6) << 4));
+            m  = uint8_t((data_a[i].scales[is + 5] >>  4) | ((data_a[i].scales[is + 1] >> 6) << 4));
+        }
         const FLOAT_TYPE d2 = dall * sc;
-        const FLOAT_TYPE m2 = dmin * mbyte;
+        const FLOAT_TYPE m2 = dmin * m;
 
         [[unroll]] for (uint l = 0; l < n; ++l) {
-            data_b[y_idx + l     ] = D_TYPE(d1 * FLOAT_TYPE(data_a[ib].qs[qs_idx + l] & 0xF) - m1);
-            data_b[y_idx + l + 32] = D_TYPE(d2 * FLOAT_TYPE(data_a[ib].qs[qs_idx + l] >>  4) - m2);
+            data_b[y_idx + l     ] = D_TYPE(d1 * FLOAT_TYPE(data_a[i].qs[qs_idx + l] & 0xF) - m1);
+            data_b[y_idx + l + 32] = D_TYPE(d2 * FLOAT_TYPE(data_a[i].qs[qs_idx + l] >>  4) - m2);
         }
     }
 }

ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp

@@ -9,8 +9,8 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
-        const uint ib = gl_WorkGroupID.x * 256 + wgy;
-        if (ib >= p.M * p.K / QUANT_K) {
+        const uint i = gl_WorkGroupID.x * 256 + wgy;
+        if (i >= p.M * p.K / QUANT_K) {
             return;
         }
 
@@ -19,52 +19,40 @@ void main() {
         const uint ir = tid % 16;
         const uint is = 2 * il;
 
-        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].d.y);
+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
 
-        const uint y_idx = ib * QUANT_K + 64 * il + 2 * ir;
+        const uint y_idx = i * QUANT_K + 64 * il + 2 * ir;
         const uint qs_idx = 32*il + 2 * ir;
         const uint qh_idx = 2 * ir;
 
-        uint scidx0 = (is < 4) ? is : (is + 4);
-        uint scidx1 = (is < 4) ? is : (is - 4);
-        uint scidxmask1 = (is < 4) ? 0x30 : 0xC0;
-        uint scidxshift1 = (is < 4) ? 0 : 2;
-        uint mbidx0 = is + 4;
-        uint mbidx1 = (is < 4) ? is + 4 : is;
-        uint mbidxmask0 = (is < 4) ? 0xF : 0xF0;
-        uint mbidxshift0 = (is < 4) ? 0 : 4;
-        uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
-        uint mbidxshift1 = (is < 4) ? 0 : 2;
-
-        uint8_t sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1));
-        uint8_t mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
-
+        uint8_t sc;
+        uint8_t m;
+        if (is < 4) {
+            sc = uint8_t(data_a[i].scales[is] & 63);
+            m  = uint8_t(data_a[i].scales[is + 4] & 63);
+        } else {
+            sc = uint8_t((data_a[i].scales[is + 4] & 0xF) | ((data_a[i].scales[is - 4] >> 6) << 4));
+            m  = uint8_t((data_a[i].scales[is + 4] >>  4) | ((data_a[i].scales[is    ] >> 6) << 4));
+        }
         const FLOAT_TYPE d1 = dall * sc;
-        const FLOAT_TYPE m1 = dmin * mbyte;
-
-        scidx0 = (is < 4) ? is + 1 : (is + 5);
-        scidx1 = (is < 4) ? is + 1 : (is - 3);
-        scidxmask1 = (is < 4) ? 0x30 : 0xC0;
-        scidxshift1 = (is < 4) ? 0 : 2;
-        mbidx0 = is + 5;
-        mbidx1 = (is < 4) ? is + 5 : is + 1;
-        mbidxmask0 = (is < 4) ? 0xF : 0xF0;
-        mbidxshift0 = (is < 4) ? 0 : 4;
-        mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
-        mbidxshift1 = (is < 4) ? 0 : 2;
-
-        sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1));
-        mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
+        const FLOAT_TYPE m1 = dmin * m;
 
+        if (is < 4) {
+            sc = uint8_t(data_a[i].scales[is + 1] & 63);
+            m  = uint8_t(data_a[i].scales[is + 5] & 63);
+        } else {
+            sc = uint8_t((data_a[i].scales[is + 5] & 0xF) | ((data_a[i].scales[is - 3] >> 6) << 4));
+            m  = uint8_t((data_a[i].scales[is + 5] >>  4) | ((data_a[i].scales[is + 1] >> 6) << 4));
+        }
         const FLOAT_TYPE d2 = dall * sc;
-        const FLOAT_TYPE m2 = dmin * mbyte;
+        const FLOAT_TYPE m2 = dmin * m;
 
         const uint8_t hm1 = uint8_t(1 << (2 * il    ));
         const uint8_t hm2 = uint8_t(1 << (2 * il + 1));
-        data_b[y_idx     ] = D_TYPE(d1 * FLOAT_TYPE((data_a[ib].qs[qs_idx    ] & 0xF) + (((data_a[ib].qh[qh_idx    ] & hm1) != 0) ? 16 : 0)) - m1);
-        data_b[y_idx +  1] = D_TYPE(d1 * FLOAT_TYPE((data_a[ib].qs[qs_idx + 1] & 0xF) + (((data_a[ib].qh[qh_idx + 1] & hm1) != 0) ? 16 : 0)) - m1);
-        data_b[y_idx + 32] = D_TYPE(d2 * FLOAT_TYPE((data_a[ib].qs[qs_idx    ]  >> 4) + (((data_a[ib].qh[qh_idx    ] & hm2) != 0) ? 16 : 0)) - m2);
-        data_b[y_idx + 33] = D_TYPE(d2 * FLOAT_TYPE((data_a[ib].qs[qs_idx + 1]  >> 4) + (((data_a[ib].qh[qh_idx + 1] & hm2) != 0) ? 16 : 0)) - m2);
+        data_b[y_idx     ] = D_TYPE(d1 * FLOAT_TYPE((data_a[i].qs[qs_idx    ] & 0xF) + (((data_a[i].qh[qh_idx    ] & hm1) != 0) ? 16 : 0)) - m1);
+        data_b[y_idx +  1] = D_TYPE(d1 * FLOAT_TYPE((data_a[i].qs[qs_idx + 1] & 0xF) + (((data_a[i].qh[qh_idx + 1] & hm1) != 0) ? 16 : 0)) - m1);
+        data_b[y_idx + 32] = D_TYPE(d2 * FLOAT_TYPE((data_a[i].qs[qs_idx    ]  >> 4) + (((data_a[i].qh[qh_idx    ] & hm2) != 0) ? 16 : 0)) - m2);
+        data_b[y_idx + 33] = D_TYPE(d2 * FLOAT_TYPE((data_a[i].qs[qs_idx + 1]  >> 4) + (((data_a[i].qh[qh_idx + 1] & hm2) != 0) ? 16 : 0)) - m2);
     }
 }

ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp

@@ -1,11 +1,6 @@
 #version 450
 
 #extension GL_EXT_shader_16bit_storage : require
-#extension GL_EXT_spirv_intrinsics: enable
-
-#if RTE16
-spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits
-#endif
 
 layout (push_constant) uniform parameter
 {

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp

@@ -2,6 +2,8 @@
 #extension GL_EXT_shader_16bit_storage : require
 #extension GL_EXT_shader_8bit_storage : require
 
+#define K_QUANTS_PER_ITERATION 2
+
 #ifdef MUL_MAT_ID
 #define EXPERT_COUNT 8
 #endif

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp

@@ -3,11 +3,9 @@
 
 #include "mul_mat_vec_base.comp"
 
-layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
 
-layout (constant_id = 0) const uint BLOCK_SIZE = 32;
-
-shared FLOAT_TYPE tmp[BLOCK_SIZE];
+shared FLOAT_TYPE tmp[32];
 
 void main() {
     const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
@@ -22,25 +20,22 @@ void main() {
     const uint num_blocks_per_row = p.ncols / QUANT_K;
     const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
 
-    // 16 threads are used to process each block
-    const uint it_size = gl_WorkGroupSize.x/16;
-    const uint tid = gl_LocalInvocationID.x;
-    const uint itid = tid%16;  // 0...16
-    const uint ix  = tid/16;
+    const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
+    const uint ix  = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
 
-    const uint step = 8;
+    const uint step = 16/K_QUANTS_PER_ITERATION;            // 16 or 8
 
-    const uint v_im = itid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
-    const uint v_in = itid - step*v_im;                      // 0...15 or 0...7
+    const uint v_im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const uint v_in = tid - step*v_im;                      // 0...15 or 0...7
 
-    const uint l0 = 2*v_in;                                  // 0...15
+    const uint l0 = K_QUANTS_PER_ITERATION*v_in;            // 0...15
     const uint q_offset = 32*v_im + l0;
     const uint s_offset = 8*v_im;
     const uint y_offset = 128*v_im + l0;
 
     FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
 
-    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
         const uint y_idx = i * QUANT_K + y_offset;
 
         f16vec2 d = data_a[ib0 + i].d;
@@ -76,7 +71,7 @@ void main() {
 
         FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
         FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
-        [[unroll]] for (int l = 0; l < 2; ++l) {
+        [[unroll]] for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
             sum1 = fma(FLOAT_TYPE(b0[l]),   FLOAT_TYPE(s0_lo4[0]) * FLOAT_TYPE((qs0[l]  >> 0) & 3),
                    fma(FLOAT_TYPE(b16[l]),  FLOAT_TYPE(s0_lo4[1]) * FLOAT_TYPE((qs16[l] >> 0) & 3),
                    fma(FLOAT_TYPE(b32[l]),  FLOAT_TYPE(s0_lo4[2]) * FLOAT_TYPE((qs0[l]  >> 2) & 3),
@@ -101,7 +96,7 @@ void main() {
 
     // sum up partial sums and write back result
     barrier();
-    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
+    [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
         if (tid < s) {
             tmp[tid] += tmp[tid + s];
         }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp

@@ -3,11 +3,9 @@
 
 #include "mul_mat_vec_base.comp"
 
-layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
 
-layout (constant_id = 0) const uint BLOCK_SIZE = 32;
-
-shared FLOAT_TYPE tmp[BLOCK_SIZE];
+shared FLOAT_TYPE tmp[32];
 
 void main() {
     const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
@@ -22,20 +20,17 @@ void main() {
     const uint num_blocks_per_row = p.ncols / QUANT_K;
     const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
 
-    // 16 threads are used to process each block
-    const uint it_size = gl_WorkGroupSize.x/16;
-    const uint tid = gl_LocalInvocationID.x;
-    const uint itid = tid%16;  // 0...16
-    const uint ix  = tid/16;
+    const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
+    const uint ix  = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
 
-    const uint step = 8;
+    const uint step = 16/K_QUANTS_PER_ITERATION;            // 16 or 8
 
-    const uint v_im = itid/step;                            // 0 or 1. 0 computes 0..., 1 computes 128...
-    const uint v_in = itid - step*v_im;                     // 0...15 or 0...7
+    const uint v_im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const uint v_in = tid - step*v_im;                      // 0...15 or 0...7
 
     const uint8_t m = uint8_t(1 << (4 * v_im));
 
-    const uint l0 = 2*v_in;                                // 0...15
+    const uint l0 = K_QUANTS_PER_ITERATION*v_in;            // 0...15
     const uint q_offset = 32*v_im + l0;
     const uint y_offset = 128*v_im + l0;
 
@@ -43,7 +38,7 @@ void main() {
 
     const uint s_shift = 4 * v_im;
 
-    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
         const uint y_idx = i * QUANT_K + y_offset;
 
         const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
@@ -71,7 +66,7 @@ void main() {
         u8vec2 s10 = unpack8(s10_16);
 
         FLOAT_TYPE sum = FLOAT_TYPE(0.0);
-        [[unroll]] for (int l = 0; l < 2; ++l) {
+        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
             sum = fma(FLOAT_TYPE(b0[l])   * FLOAT_TYPE(int8_t(((s0[0] >> s_shift) & 0xF) | ((s8[0]  >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 0)) != 0) ? 0 : 4)),
                   fma(FLOAT_TYPE(b32[l])  * FLOAT_TYPE(int8_t(((s2[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 1)) != 0) ? 0 : 4)),
                   fma(FLOAT_TYPE(b64[l])  * FLOAT_TYPE(int8_t(((s4[0] >> s_shift) & 0xF) | ((s8[0]  >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 2)) != 0) ? 0 : 4)),
@@ -88,7 +83,7 @@ void main() {
 
     // sum up partial sums and write back result
     barrier();
-    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
+    [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
         if (tid < s) {
             tmp[tid] += tmp[tid + s];
         }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp

@@ -4,12 +4,11 @@
 
 #include "mul_mat_vec_base.comp"
 
-layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
 
-layout (constant_id = 0) const uint BLOCK_SIZE = 32;
-
-shared FLOAT_TYPE tmp[BLOCK_SIZE];
+shared FLOAT_TYPE tmp[32];
 
+// This shader assumes K_QUANTS_PER_ITERATION == 2 for alignment of loads
 void main() {
     const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
 
@@ -23,17 +22,14 @@ void main() {
     const uint num_blocks_per_row = p.ncols / QUANT_K;
     const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
 
-    // 16 threads are used to process each block
-    const uint it_size = gl_WorkGroupSize.x/16;
-    const uint tid = gl_LocalInvocationID.x;
-    const uint itid = tid%16;  // 0...16
-    const uint ix  = tid/16;
+    const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
+    const uint ix  = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
 
-    const uint step = 4;
+    const uint step = 8/K_QUANTS_PER_ITERATION;             // 8 or 4
 
-    const uint il = itid/step;                               // 0...3
-    const uint ir = itid - step*il;                          // 0...7 or 0...3
-    const uint n =  4;
+    const uint il = tid/step;                               // 0...3
+    const uint ir = tid - step*il;                          // 0...7 or 0...3
+    const uint n =  2 * K_QUANTS_PER_ITERATION;             // 2 or 4
 
     const uint v_im = il / 2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
     const uint v_in = il % 2;
@@ -44,7 +40,7 @@ void main() {
 
     FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
 
-    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
         const uint y1_idx = i * QUANT_K + y_offset;
         const uint y2_idx = y1_idx + 128;
 
@@ -119,7 +115,7 @@ void main() {
 
     // sum up partial sums and write back result
     barrier();
-    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
+    [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
         if (tid < s) {
             tmp[tid] += tmp[tid + s];
         }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp

@@ -4,11 +4,9 @@
 
 #include "mul_mat_vec_base.comp"
 
-layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
 
-layout (constant_id = 0) const uint BLOCK_SIZE = 32;
-
-shared FLOAT_TYPE tmp[BLOCK_SIZE];
+shared FLOAT_TYPE tmp[32];
 
 void main() {
     const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
@@ -23,14 +21,11 @@ void main() {
     const uint num_blocks_per_row = p.ncols / QUANT_K;
     const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
 
-    // 16 threads are used to process each block
-    const uint it_size = gl_WorkGroupSize.x/16;
-    const uint tid = gl_LocalInvocationID.x;
-    const uint itid = tid%16;  // 0...16
-    const uint ix  = tid/16;
+    const uint tid = gl_LocalInvocationID.x/2;  // 0...31 or 0...16
+    const uint ix  = gl_LocalInvocationID.x%2;  // 0 or 0, 1
 
-    const uint il = itid/4;                          // 0...3
-    const uint ir = itid - 4*il;                     // 0...7 or 0...3
+    const uint il = tid/4;                           // 0...3
+    const uint ir = tid - 4*il;                      // 0...7 or 0...3
 
     const uint v_im = il / 2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
     const uint v_in = il % 2;
@@ -41,7 +36,7 @@ void main() {
 
     FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
 
-    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += 2) {
         const uint y1_idx = i * QUANT_K + y_offset;
         const uint y2_idx = y1_idx + 128;
 
@@ -148,7 +143,7 @@ void main() {
 
     // sum up partial sums and write back result
     barrier();
-    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
+    [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
         if (tid < s) {
             tmp[tid] += tmp[tid + s];
         }

ggml/src/ggml-vulkan/vulkan-shaders/rope_head.comp

@@ -1,11 +1,6 @@
 #include "types.comp"
 
 #extension GL_EXT_shader_16bit_storage : require
-#extension GL_EXT_spirv_intrinsics: enable
-
-#if RTE16
-spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits
-#endif
 
 layout(local_size_x = 1, local_size_y = 256, local_size_z = 1) in;
 

ggml/src/ggml-vulkan/vulkan-shaders/tanh.comp

@@ -16,5 +16,6 @@ void main() {
     if (i >= p.KX) {
         return;
     }
-    data_d[i] = D_TYPE(1. - 2. / (exp(2.*data_a[i]) + 1.));
+
+    data_d[i] = D_TYPE(tanh(data_a[i]));
 }

ggml/src/ggml-vulkan/vulkan-shaders/test_coopmat2_support.comp

@@ -1,7 +0,0 @@
-#version 460
-
-#extension GL_NV_cooperative_matrix2 : require
-
-void main()
-{
-}

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -206,13 +206,10 @@ void string_to_spv_func(const std::string& _name, const std::string& in_fname, c
 
     std::string target_env = (name.find("_cm2") != std::string::npos) ? "--target-env=vulkan1.3" : "--target-env=vulkan1.2";
 
-    // disable spirv-opt for coopmat shaders for https://github.com/ggerganov/llama.cpp/issues/10734
-    std::string opt_level = coopmat ? "" : "-O";
-
     #ifdef _WIN32
-        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, opt_level, "\"" + in_path + "\"", "-o", "\"" + out_fname + "\""};
+        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, "-O", "\"" + in_path + "\"", "-o", "\"" + out_fname + "\""};
     #else
-        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, opt_level, in_path, "-o",  out_fname};
+        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, "-O", in_path, "-o",  out_fname};
     #endif
 
     #ifdef GGML_VULKAN_SHADER_DEBUG_INFO
@@ -345,14 +342,14 @@ void process_shaders() {
         matmul_shaders(true, matmul_id, true, false, false);
         matmul_shaders(true, matmul_id, true, false, true);
 
-#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+#if defined(VK_NV_cooperative_matrix2)
         // Coopmat2, fp32acc and fp16acc
         matmul_shaders(true, matmul_id, false, true, false);
         matmul_shaders(true, matmul_id, false, true, true);
 #endif
     }
 
-#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+#if defined(VK_NV_cooperative_matrix2)
     // flash attention
     for (const auto& f16acc : {false, true}) {
         std::string acctype = f16acc ? "float16_t" : "float";
@@ -461,11 +458,9 @@ void process_shaders() {
 
     string_to_spv("rope_norm_f32", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("rope_norm_f16", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
-    string_to_spv("rope_norm_f16_rte", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
 
     string_to_spv("rope_neox_f32", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("rope_neox_f16", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
-    string_to_spv("rope_neox_f16_rte", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
 
     string_to_spv("argsort_f32", "argsort.comp", {{"A_TYPE", "float"}});
 
@@ -473,7 +468,6 @@ void process_shaders() {
 
     string_to_spv("im2col_f32", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("im2col_f32_f16", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}}));
-    string_to_spv("im2col_f32_f16_rte", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}}));
 
     string_to_spv("timestep_embedding_f32", "timestep_embedding.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 

gguf-py/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "gguf"
-version = "0.11.0"
+version = "0.10.0"
 description = "Read and write ML models in GGUF for GGML"
 authors = ["GGML <ggml@ggml.ai>"]
 packages = [

include/llama.h

@@ -456,7 +456,6 @@ extern "C" {
     // Functions to access the model's GGUF metadata scalar values
     // - The functions return the length of the string on success, or -1 on failure
     // - The output string is always null-terminated and cleared on failure
-    // - When retrieving a string, an extra byte must be allocated to account for the null terminator
     // - GGUF array values are not supported by these functions
 
     // Get metadata value as a string by key name

src/CMakeLists.txt

@@ -1,3 +1,10 @@
+# TODO: should not use this
+if (WIN32)
+    if (BUILD_SHARED_LIBS)
+        set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
+    endif()
+endif()
+
 llama_add_compile_flags()
 
 #

src/llama.cpp

@@ -1794,7 +1794,7 @@ private:
         DWORD bufLen = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
                                     NULL, error_code, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&lpMsgBuf, 0, NULL);
         if (!bufLen) {
-            ret = format("Win32 error code: %lx", error_code);
+            ret = format("Win32 error code: %s", error_code);
         } else {
             ret = lpMsgBuf;
             LocalFree(lpMsgBuf);
@@ -2132,7 +2132,7 @@ struct llama_mmap {
             HMODULE hKernel32 = GetModuleHandleW(L"kernel32.dll");
 
             // may fail on pre-Windows 8 systems
-            pPrefetchVirtualMemory = (decltype(pPrefetchVirtualMemory))(void *) GetProcAddress(hKernel32, "PrefetchVirtualMemory");
+            pPrefetchVirtualMemory = reinterpret_cast<decltype(pPrefetchVirtualMemory)> (GetProcAddress(hKernel32, "PrefetchVirtualMemory"));
 
             if (pPrefetchVirtualMemory) {
                 // advise the kernel to preload the mapped memory
@@ -21577,7 +21577,7 @@ float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i) {
                 throw std::runtime_error(format("negative index out of range [0, %d)", ctx->n_outputs));
             }
         } else if ((size_t) i >= ctx->output_ids.size()) {
-            throw std::runtime_error(format("out of range [0, %zu)", ctx->output_ids.size()));
+            throw std::runtime_error(format("out of range [0, %lu)", ctx->output_ids.size()));
         } else {
             j = ctx->output_ids[i];
         }

tests/CMakeLists.txt

@@ -84,50 +84,38 @@ llama_test(test-tokenizer-0 NAME test-tokenizer-0-qwen2             ARGS ${CMAKE
 llama_test(test-tokenizer-0 NAME test-tokenizer-0-refact            ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf)
 llama_test(test-tokenizer-0 NAME test-tokenizer-0-starcoder         ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf)
 
+# build test-tokenizer-1-bpe target once and add many tests
+add_executable(test-tokenizer-1-bpe test-tokenizer-1-bpe.cpp)
+target_link_libraries(test-tokenizer-1-bpe PRIVATE common)
+install(TARGETS test-tokenizer-1-bpe RUNTIME)
 
-if (NOT WIN32)
-    # these tests are disabled on Windows because they use internal functions not exported with LLAMA_API
-    llama_target_and_test(test-sampling.cpp)
-    llama_target_and_test(test-grammar-parser.cpp)
-    llama_target_and_test(test-grammar-integration.cpp)
-    llama_target_and_test(test-llama-grammar.cpp)
-    # TODO: disabled on loongarch64 because the ggml-ci node lacks Python 3.8
-    if (NOT ${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
-        llama_target_and_test(test-json-schema-to-grammar.cpp   WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/..)
-        target_include_directories(test-json-schema-to-grammar PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/../examples/server)
-    endif()
+# TODO: disabled due to slowness
+#llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-aquila    ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf)
+#llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-falcon    ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
+#llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-gpt-2     ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-2.gguf)
+#llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-gpt-neox  ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-neox.gguf)
+#llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-llama-bpe ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-bpe.gguf --ignore-merges)
+#llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-mpt       ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-mpt.gguf)
+#llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-refact    ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf)
+#llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-starcoder ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf)
 
+# build test-tokenizer-1-spm target once and add many tests
+add_executable(test-tokenizer-1-spm test-tokenizer-1-spm.cpp)
+target_link_libraries(test-tokenizer-1-spm PRIVATE common)
+install(TARGETS test-tokenizer-1-spm RUNTIME)
 
-    # build test-tokenizer-1-bpe target once and add many tests
-    add_executable(test-tokenizer-1-bpe test-tokenizer-1-bpe.cpp)
-    target_link_libraries(test-tokenizer-1-bpe PRIVATE common)
-    install(TARGETS test-tokenizer-1-bpe RUNTIME)
-
-    # TODO: disabled due to slowness
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-aquila    ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf)
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-falcon    ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-gpt-2     ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-2.gguf)
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-gpt-neox  ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-neox.gguf)
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-llama-bpe ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-bpe.gguf --ignore-merges)
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-mpt       ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-mpt.gguf)
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-refact    ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf)
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-starcoder ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf)
-
-    # build test-tokenizer-1-spm target once and add many tests
-    add_executable(test-tokenizer-1-spm test-tokenizer-1-spm.cpp)
-    target_link_libraries(test-tokenizer-1-spm PRIVATE common)
-    install(TARGETS test-tokenizer-1-spm RUNTIME)
-
-    llama_test(test-tokenizer-1-spm  NAME test-tokenizer-1-llama-spm ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-spm.gguf)
-    #llama_test(test-tokenizer-1-spm  NAME test-tokenizer-1-baichuan  ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-baichuan.gguf)
-
-    # llama_target_and_test(test-double-float.cpp) # SLOW
-endif()
+llama_test(test-tokenizer-1-spm  NAME test-tokenizer-1-llama-spm ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-spm.gguf)
+#llama_test(test-tokenizer-1-spm  NAME test-tokenizer-1-baichuan  ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-baichuan.gguf)
 
+# llama_target_and_test(test-double-float.cpp) # SLOW
 llama_target_and_test(test-log.cpp)
 llama_target_and_test(test-arg-parser.cpp)
+llama_target_and_test(test-sampling.cpp)
 llama_target_and_test(test-chat-template.cpp)
 
+llama_target_and_test(test-grammar-parser.cpp)
+llama_target_and_test(test-grammar-integration.cpp)
+llama_target_and_test(test-llama-grammar.cpp)
 # llama_target_and_test(test-opt.cpp) # SLOW
 llama_target_and_test(test-backend-ops.cpp)
 
@@ -142,6 +130,11 @@ if (NOT GGML_BACKEND_DL)
     llama_target_and_test(test-rope.cpp)
 endif()
 
+# TODO: disabled on loongarch64 because the ggml-ci node lacks Python 3.8
+if (NOT ${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
+    llama_target_and_test(test-json-schema-to-grammar.cpp   WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/..)
+    target_include_directories(test-json-schema-to-grammar PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/../examples/server)
+endif()
 
 # dummy executable - not installed
 get_filename_component(TEST_TARGET test-c.c NAME_WE)