Merge remote-tracking branch 'origin/master' into tool-call

.devops/tools.sh

@@ -8,11 +8,11 @@ arg1="$1"
 shift
 
 if [[ "$arg1" == '--convert' || "$arg1" == '-c' ]]; then
-    python3 ./convert_hf_to_gguf.py "$@"
+    exec python3 ./convert_hf_to_gguf.py "$@"
 elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then
-    ./llama-quantize "$@"
+    exec ./llama-quantize "$@"
 elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then
-    ./llama-cli "$@"
+    exec ./llama-cli "$@"
 elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
     echo "Converting PTH to GGML..."
     for i in `ls $1/$2/ggml-model-f16.bin*`; do
@@ -20,11 +20,11 @@ elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
             echo "Skip model quantization, it already exists: ${i/f16/q4_0}"
         else
             echo "Converting PTH to GGML: $i into ${i/f16/q4_0}..."
-            ./llama-quantize "$i" "${i/f16/q4_0}" q4_0
+            exec ./llama-quantize "$i" "${i/f16/q4_0}" q4_0
         fi
     done
 elif [[ "$arg1" == '--server' || "$arg1" == '-s' ]]; then
-    ./llama-server "$@"
+    exec ./llama-server "$@"
 else
     echo "Unknown command: $arg1"
     echo "Available commands: "

.github/workflows/build.yml

@@ -662,6 +662,8 @@ jobs:
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'msvc-arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
+          - build: 'llvm-arm64-opencl-adreno'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON'
 
     steps:
       - name: Clone
@@ -703,6 +705,28 @@ jobs:
         run: |
           choco install ninja
 
+      - name: Install OpenCL Headers and Libs
+        id: install_opencl
+        if: ${{ matrix.build == 'llvm-arm64-opencl-adreno' }}
+        run: |
+          git clone https://github.com/KhronosGroup/OpenCL-Headers
+          cd OpenCL-Headers
+          mkdir build && cd build
+          cmake .. `
+            -DBUILD_TESTING=OFF `
+            -DOPENCL_HEADERS_BUILD_TESTING=OFF `
+            -DOPENCL_HEADERS_BUILD_CXX_TESTS=OFF `
+            -DCMAKE_INSTALL_PREFIX="$env:RUNNER_TEMP/opencl-arm64-release"
+          cmake --build . --target install
+          git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader
+          cd OpenCL-ICD-Loader
+          mkdir build-arm64-release && cd build-arm64-release
+          cmake .. `
+            -A arm64 `
+            -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" `
+            -DCMAKE_INSTALL_PREFIX="$env:RUNNER_TEMP/opencl-arm64-release"
+          cmake --build . --target install --config release
+
       - name: Build
         id: cmake_build
         run: |
@@ -732,7 +756,7 @@ jobs:
       - name: Test
         id: cmake_test
         # not all machines have native AVX-512
-        if: ${{ matrix.build != 'msvc-arm64' && matrix.build != 'llvm-arm64' && matrix.build != 'kompute-x64' && matrix.build != 'vulkan-x64' && (matrix.build != 'avx512-x64' || env.HAS_AVX512F == '1') }}
+        if: ${{ matrix.build != 'msvc-arm64' && matrix.build != 'llvm-arm64' && matrix.build != 'llvm-arm64-opencl-adreno' && matrix.build != 'kompute-x64' && matrix.build != 'vulkan-x64' && (matrix.build != 'avx512-x64' || env.HAS_AVX512F == '1') }}
         run: |
           cd build
           ctest -L main -C Release --verbose --timeout 900

.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
+          node-version: '22.11.0'
 
       - name: Verify bundled index.html
         id: verify_server_index_html

CODEOWNERS

@@ -1,3 +1,5 @@
 # collaborators can optionally add themselves here to indicate their availability for reviewing related PRs
 
-ci/ @ggerganov
+/ci/ @ggerganov
+/.devops/ @ngxson
+/examples/server/ @ngxson

Makefile

@@ -22,6 +22,7 @@ BUILD_TARGETS = \
 	llama-infill \
 	llama-llava-cli \
 	llama-minicpmv-cli\
+	llama-qwen2vl-cli\
 	llama-lookahead \
 	llama-lookup \
 	llama-lookup-create \
@@ -1411,6 +1412,14 @@ llama-minicpmv-cli: examples/llava/minicpmv-cli.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
 
+llama-qwen2vl-cli: examples/llava/qwen2vl-cli.cpp \
+	examples/llava/llava.cpp \
+	examples/llava/llava.h \
+	examples/llava/clip.cpp \
+	examples/llava/clip.h \
+	$(OBJ_ALL)
+	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
+
 ifeq ($(UNAME_S),Darwin)
 swift: examples/batched.swift
 	(cd examples/batched.swift; make build)

README.md

@@ -110,6 +110,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [Mini CPM](https://huggingface.co/models?search=MiniCPM)
 - [x] [Moondream](https://huggingface.co/vikhyatk/moondream2)
 - [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
+- [x] [Qwen2-VL](https://huggingface.co/collections/Qwen/qwen2-vl-66cee7455501d7126940800d)
 
 </details>
 
@@ -433,6 +434,20 @@ To learn more about model quantization, [read this documentation](examples/quant
 
     </details>
 
+## [`llama-run`](examples/run)
+
+#### A comprehensive example for running `llama.cpp` models. Useful for inferencing. Used with RamaLama [^3].
+
+- <details>
+    <summary>Run a model with a specific prompt (by default it's pulled from Ollama registry)</summary>
+
+    ```bash
+    llama-run granite-code
+    ```
+
+    </details>
+
+[^3]: [https://github.com/containers/ramalama](RamaLama)
 
 ## [`llama-simple`](examples/simple)
 

common/CMakeLists.txt

@@ -84,7 +84,7 @@ set(LLAMA_COMMON_EXTRA_LIBS build_info)
 # Use curl to download model url
 if (LLAMA_CURL)
     find_package(CURL REQUIRED)
-    add_definitions(-DLLAMA_USE_CURL)
+    target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_CURL)
     include_directories(${CURL_INCLUDE_DIRS})
     find_library(CURL_LIBRARY curl REQUIRED)
     set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARY})

common/arg.cpp

@@ -145,6 +145,35 @@ static void common_params_handle_model_default(common_params & params) {
     }
 }
 
+const std::vector<ggml_type> kv_cache_types = {
+    GGML_TYPE_F32,
+    GGML_TYPE_F16,
+    GGML_TYPE_BF16,
+    GGML_TYPE_Q8_0,
+    GGML_TYPE_Q4_0,
+    GGML_TYPE_Q4_1,
+    GGML_TYPE_IQ4_NL,
+    GGML_TYPE_Q5_0,
+    GGML_TYPE_Q5_1,
+};
+
+static ggml_type kv_cache_type_from_str(const std::string & s) {
+    for (const auto & type : kv_cache_types) {
+        if (ggml_type_name(type) == s) {
+            return type;
+        }
+    }
+    throw std::runtime_error("Unsupported cache type: " + s);
+}
+
+static std::string get_all_kv_cache_types() {
+    std::ostringstream msg;
+    for (const auto & type : kv_cache_types) {
+        msg << ggml_type_name(type) << (&type == &kv_cache_types.back() ? "" : ", ");
+    }
+    return msg.str();
+}
+
 //
 // CLI argument parsing functions
 //
@@ -591,7 +620,7 @@ 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}).set_env("LLAMA_ARG_NO_CONTEXT_SHIFT"));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_IMATRIX}).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),
@@ -1174,18 +1203,28 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_env("LLAMA_ARG_NO_KV_OFFLOAD"));
     add_opt(common_arg(
         {"-ctk", "--cache-type-k"}, "TYPE",
-        string_format("KV cache data type for K (default: %s)", params.cache_type_k.c_str()),
+        string_format(
+            "KV cache data type for K\n"
+            "allowed values: %s\n"
+            "(default: %s)",
+            get_all_kv_cache_types().c_str(),
+            ggml_type_name(params.cache_type_k)
+        ),
         [](common_params & params, const std::string & value) {
-            // TODO: get the type right here
+            params.cache_type_k = kv_cache_type_from_str(value);
-            params.cache_type_k = value;
         }
     ).set_env("LLAMA_ARG_CACHE_TYPE_K"));
     add_opt(common_arg(
         {"-ctv", "--cache-type-v"}, "TYPE",
-        string_format("KV cache data type for V (default: %s)", params.cache_type_v.c_str()),
+        string_format(
+            "KV cache data type for V\n"
+            "allowed values: %s\n"
+            "(default: %s)",
+            get_all_kv_cache_types().c_str(),
+            ggml_type_name(params.cache_type_v)
+        ),
         [](common_params & params, const std::string & value) {
-            // TODO: get the type right here
+            params.cache_type_v = kv_cache_type_from_str(value);
-            params.cache_type_v = value;
         }
     ).set_env("LLAMA_ARG_CACHE_TYPE_V"));
     add_opt(common_arg(
@@ -1711,6 +1750,13 @@ 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"),
@@ -2111,35 +2157,35 @@ 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_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_DRAFT_MAX"));
     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_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_DRAFT_MIN"));
     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_examples({LLAMA_EXAMPLE_SPECULATIVE}).set_env("LLAMA_ARG_DRAFT_P_SPLIT"));
     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_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_DRAFT_P_MIN"));
     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_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CTX_SIZE_DRAFT"));
     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"
@@ -2159,14 +2205,14 @@ 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_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_N_GPU_LAYERS_DRAFT"));
     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_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODEL_DRAFT"));
 
     return ctx_arg;
 }

common/common.cpp

@@ -1054,38 +1054,6 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     return mparams;
 }
 
-static ggml_type kv_cache_type_from_str(const std::string & s) {
-    if (s == "f32") {
-        return GGML_TYPE_F32;
-    }
-    if (s == "f16") {
-        return GGML_TYPE_F16;
-    }
-    if (s == "bf16") {
-        return GGML_TYPE_BF16;
-    }
-    if (s == "q8_0") {
-        return GGML_TYPE_Q8_0;
-    }
-    if (s == "q4_0") {
-        return GGML_TYPE_Q4_0;
-    }
-    if (s == "q4_1") {
-        return GGML_TYPE_Q4_1;
-    }
-    if (s == "iq4_nl") {
-        return GGML_TYPE_IQ4_NL;
-    }
-    if (s == "q5_0") {
-        return GGML_TYPE_Q5_0;
-    }
-    if (s == "q5_1") {
-        return GGML_TYPE_Q5_1;
-    }
-
-    throw std::runtime_error("Unsupported cache type: " + s);
-}
-
 struct llama_context_params common_context_params_to_llama(const common_params & params) {
     auto cparams = llama_context_default_params();
 
@@ -1120,8 +1088,8 @@ struct llama_context_params common_context_params_to_llama(const common_params &
         cparams.pooling_type  = LLAMA_POOLING_TYPE_RANK;
     }
 
-    cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
+    cparams.type_k = params.cache_type_k;
-    cparams.type_v = kv_cache_type_from_str(params.cache_type_v);
+    cparams.type_v = params.cache_type_v;
 
     return cparams;
 }
@@ -1147,12 +1115,6 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_p
 #define CURL_MAX_RETRY 3
 #define CURL_RETRY_DELAY_SECONDS 2
 
-
-static bool starts_with(const std::string & str, const std::string & prefix) {
-    // While we wait for C++20's std::string::starts_with...
-    return str.rfind(prefix, 0) == 0;
-}
-
 static bool curl_perform_with_retry(const std::string& url, CURL* curl, int max_attempts, int retry_delay_seconds) {
     int remaining_attempts = max_attempts;
 

common/common.h

@@ -43,9 +43,9 @@ using llama_tokens = std::vector<llama_token>;
 
 // build info
 extern int LLAMA_BUILD_NUMBER;
-extern char const * LLAMA_COMMIT;
+extern const char * LLAMA_COMMIT;
-extern char const * LLAMA_COMPILER;
+extern const char * LLAMA_COMPILER;
-extern char const * LLAMA_BUILD_TARGET;
+extern const char * LLAMA_BUILD_TARGET;
 
 struct common_control_vector_load_info;
 
@@ -293,8 +293,8 @@ struct common_params {
     bool warmup            = true;  // warmup run
     bool check_tensors     = false; // validate tensor data
 
-    std::string cache_type_k = "f16"; // KV cache data type for the K
+    ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
-    std::string cache_type_v = "f16"; // KV cache data type for the V
+    ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V
 
     // multimodal models (see examples/llava)
     std::string mmproj = "";        // path to multimodal projector                                         // NOLINT
@@ -446,6 +446,11 @@ std::vector<std::string> string_split<std::string>(const std::string & input, ch
     return parts;
 }
 
+static bool string_starts_with(const std::string & str,
+                               const std::string & prefix) {  // While we wait for C++20's std::string::starts_with...
+    return str.rfind(prefix, 0) == 0;
+}
+
 bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
 void string_process_escapes(std::string & input);
 

convert_hf_to_gguf.py

@@ -2001,6 +2001,29 @@ class Qwen2Model(Model):
                 self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
 
 
+@Model.register("Qwen2VLForConditionalGeneration")
+class Qwen2VLModel(Model):
+    model_arch = gguf.MODEL_ARCH.QWEN2VL
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        mrope_section = self.hparams["rope_scaling"]["mrope_section"]
+        mrope_section += [0] * max(0, 4 - len(mrope_section))
+        self.gguf_writer.add_rope_dimension_sections(mrope_section)
+
+    def set_vocab(self):
+        try:
+            self._set_vocab_sentencepiece()
+        except FileNotFoundError:
+            self._set_vocab_gpt2()
+
+    def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
+        for name, data in super().get_tensors():
+            if name.startswith("visual."):
+                continue
+            yield name, data
+
+
 @Model.register("Qwen2MoeForCausalLM")
 class Qwen2MoeModel(Model):
     model_arch = gguf.MODEL_ARCH.QWEN2MOE

examples/CMakeLists.txt

@@ -20,7 +20,12 @@ 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(gguf-hash)
     add_subdirectory(gguf-split)
     add_subdirectory(gguf)
@@ -46,12 +51,16 @@ else()
     add_subdirectory(speculative)
     add_subdirectory(speculative-simple)
     add_subdirectory(tokenize)
+    add_subdirectory(gen-docs)
     if (NOT GGML_BACKEND_DL)
         # these examples use the backends directly and cannot be built with dynamic loading
         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(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: %ld\n", ctx_outs.size());
+        printf("n_split: %zu\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 = %ldM\n", i_split + 1, gguf_get_n_tensors(ctx_out), total_size);
+            printf("split %05d: n_tensors = %d, total_size = %zuM\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/%ld: starting\n", params_idx, params_count);
+            fprintf(stderr, "llama-bench: benchmark %d/%zu: 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/%ld: warmup prompt run\n", params_idx, params_count);
+                fprintf(stderr, "llama-bench: benchmark %d/%zu: 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/%ld: warmup generation run\n", params_idx, params_count);
+                fprintf(stderr, "llama-bench: benchmark %d/%zu: 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/%ld: prompt run %d/%d\n", params_idx, params_count,
+                    fprintf(stderr, "llama-bench: benchmark %d/%zu: 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/%ld: generation run %d/%d\n", params_idx, params_count,
+                    fprintf(stderr, "llama-bench: benchmark %d/%zu: generation run %d/%d\n", params_idx, params_count,
                             i + 1, params.reps);
                 }
                 test_gen(ctx, t.n_gen, t.n_threads);

examples/llava/CMakeLists.txt

@@ -43,3 +43,10 @@ set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-minicpmv-cli)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+set(TARGET llama-qwen2vl-cli)
+add_executable(${TARGET} qwen2vl-cli.cpp)
+set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-qwen2vl-cli)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/llava/clip.cpp

@@ -102,7 +102,9 @@ static std::string format(const char * fmt, ...) {
 #define KEY_HAS_LLAVA_PROJ      "clip.has_llava_projector"
 #define KEY_HAS_MINICPMV_PROJ   "clip.has_minicpmv_projector"
 #define KEY_MINICPMV_VERSION    "clip.minicpmv_version"
+#define KEY_HAS_QWEN2VL_MERGER  "clip.has_qwen2vl_merger"
 #define KEY_USE_GELU            "clip.use_gelu"
+#define KEY_USE_SILU            "clip.use_silu"
 #define KEY_N_EMBD              "clip.%s.embedding_length"
 #define KEY_N_FF                "clip.%s.feed_forward_length"
 #define KEY_N_BLOCK             "clip.%s.block_count"
@@ -129,7 +131,8 @@ static std::string format(const char * fmt, ...) {
 #define TN_TOKEN_EMBD      "%s.token_embd.weight"
 #define TN_POS_EMBD        "%s.position_embd.weight"
 #define TN_CLASS_EMBD      "v.class_embd"
-#define TN_PATCH_EMBD      "v.patch_embd.weight"
+#define TN_PATCH_EMBD      "v.patch_embd.weight"  // not rename tensor with ".0" postfix for backwrad compat
+#define TN_PATCH_EMBD_1    "v.patch_embd.weight.1"
 #define TN_PATCH_BIAS      "v.patch_embd.bias"
 #define TN_ATTN_K          "%s.blk.%d.attn_k.%s"
 #define TN_ATTN_Q          "%s.blk.%d.attn_q.%s"
@@ -163,6 +166,7 @@ enum projector_type {
     PROJECTOR_TYPE_LDP,
     PROJECTOR_TYPE_LDPV2,
     PROJECTOR_TYPE_RESAMPLER,
+    PROJECTOR_TYPE_MERGER,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -171,6 +175,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_LDP, "ldp" },
     { PROJECTOR_TYPE_LDPV2, "ldpv2"},
     { PROJECTOR_TYPE_RESAMPLER, "resampler"},
+    { PROJECTOR_TYPE_MERGER, "qwen2vl_merger"},
 };
 
 
@@ -463,7 +468,8 @@ struct clip_vision_model {
 
     // embeddings
     struct ggml_tensor * class_embedding;
-    struct ggml_tensor * patch_embeddings;
+    struct ggml_tensor * patch_embeddings_0;
+    struct ggml_tensor * patch_embeddings_1;  // second Conv2D kernel when we decouple Conv3D along temproal dimension (Qwen2VL)
     struct ggml_tensor * patch_bias;
     struct ggml_tensor * position_embeddings;
 
@@ -553,6 +559,7 @@ struct clip_ctx {
     bool has_vision_encoder  = false;
     bool has_llava_projector = false;
     bool has_minicpmv_projector = false;
+    bool has_qwen2vl_merger = false;
     int minicpmv_version = 2;
 
     struct clip_vision_model vision_model;
@@ -561,6 +568,7 @@ struct clip_ctx {
     float image_mean[3];
     float image_std[3];
     bool use_gelu = false;
+    bool use_silu = false;
     int32_t ftype = 1;
 
     bool has_class_embedding = true;
@@ -606,14 +614,26 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             image_size_height = imgs->data->ny;
         }
     }
+    else if (ctx->has_qwen2vl_merger) {
+        // use the image's native resolution when image is avaible
+        if (is_inf) {
+        // if (imgs->data->nx && imgs->data->ny) {
+            image_size_width  = imgs->data->nx;
+            image_size_height = imgs->data->ny;
+        }
+    }
     const int patch_size           = hparams.patch_size;
     const int num_patches          = ((image_size_width / patch_size) * (image_size_height / patch_size));
+    const int patches_w            = image_size_width / patch_size;
+    const int patches_h            = image_size_height / patch_size;
     const int num_positions        = num_patches + (ctx->has_class_embedding ? 1 : 0);
+    const int num_position_ids     = ctx->has_qwen2vl_merger ? num_positions * 4 : num_positions;
     const int hidden_size          = hparams.hidden_size;
     const int n_head               = hparams.n_head;
     const int d_head               = hidden_size / n_head;
     int n_layer                    = hparams.n_layer;
     const float eps                = hparams.eps;
+    int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4};
 
     const int batch_size = imgs->size;
 
@@ -634,10 +654,30 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     ggml_set_name(inp_raw, "inp_raw");
     ggml_set_input(inp_raw);
 
-    struct ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+    struct ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
 
+    if (ctx->has_qwen2vl_merger) {
+        GGML_ASSERT(image_size_width % (patch_size * 2) == 0);
+        GGML_ASSERT(image_size_height % (patch_size * 2) == 0);
+
+        auto inp_1 = ggml_conv_2d(ctx0, model.patch_embeddings_1, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+        inp = ggml_add(ctx0, inp, inp_1);
+        inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 2, 0, 3));  // [w, h, c, b] -> [c, w, h, b]
+        inp = ggml_reshape_4d(
+            ctx0, inp,
+            hidden_size * 2, patches_w / 2, patches_h, batch_size);
+        inp = ggml_reshape_4d(
+            ctx0, inp,
+            hidden_size * 2, patches_w / 2, 2, batch_size * (patches_h / 2));
+        inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 0, 2, 1, 3));
+        inp = ggml_reshape_3d(
+            ctx0, inp,
+            hidden_size, patches_w * patches_h, batch_size);
+    }
+    else {
         inp = ggml_reshape_3d(ctx0, inp, num_patches, hidden_size, batch_size);
         inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 0, 2, 3));
+    }
 
     if (ctx->has_patch_bias) {
         // inp = ggml_add(ctx0, inp, ggml_repeat(ctx0, model.patch_bias, inp));
@@ -659,12 +699,14 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         }
     }
 
-    struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions);
+    struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_position_ids);
     ggml_set_name(positions, "positions");
     ggml_set_input(positions);
 
+    if (!ctx->has_qwen2vl_merger) { // qwen2vl use rope position embedding
         embeddings =
             ggml_add(ctx0, embeddings, ggml_get_rows(ctx0, model.position_embeddings, positions));
+    }
 
     if (ctx->has_minicpmv_projector) {
         int pos_w = image_size_width/patch_size;
@@ -688,7 +730,8 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     }
 
     // loop over layers
-    if (ctx->has_minicpmv_projector) {
+    if (ctx->has_minicpmv_projector || ctx->has_qwen2vl_merger) {
+        // TODO: figure out why we doing thing in this way ???
         n_layer += 1;
     }
     for (int il = 0; il < n_layer - 1; il++) {
@@ -710,8 +753,13 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             struct ggml_tensor * Q =
                 ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].q_w, cur), model.layers[il].q_b);
 
-            Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head));
             Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, num_positions, batch_size);
+            if (ctx->has_qwen2vl_merger) {
+                Q = ggml_rope_multi(
+                    ctx0, Q, positions, nullptr,
+                    d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1);
+            }
+            Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head));
             Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
             Q = ggml_reshape_3d(ctx0, Q, d_head, num_positions, n_head * batch_size);
 
@@ -719,6 +767,11 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
                 ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].k_w, cur), model.layers[il].k_b);
 
             K = ggml_reshape_4d(ctx0, K, d_head, n_head, num_positions, batch_size);
+            if (ctx->has_qwen2vl_merger) {
+                K = ggml_rope_multi(
+                    ctx0, K, positions, nullptr,
+                    d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1);
+            }
             K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3));
             K = ggml_reshape_3d(ctx0, K, d_head, num_positions, n_head * batch_size);
 
@@ -758,6 +811,8 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
 
         if (ctx->use_gelu) {
             cur = ggml_gelu_inplace(ctx0, cur);
+        } else if (ctx->use_silu) {
+            cur = ggml_silu_inplace(ctx0, cur);
         } else {
             cur = ggml_gelu_quick_inplace(ctx0, cur);
         }
@@ -769,6 +824,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         cur = ggml_add(ctx0, embeddings, cur);
 
         embeddings = cur;
+
     }
 
     // post-layernorm
@@ -1030,6 +1086,19 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             GGML_ASSERT(false);
         }
     }
+    else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
+        embeddings = ggml_reshape_3d(ctx0, embeddings, hidden_size * 4, num_positions / 4, batch_size);
+
+        embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
+        embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
+
+        // GELU activation
+        embeddings = ggml_gelu(ctx0, embeddings);
+
+        // Second linear layer
+        embeddings = ggml_mul_mat(ctx0, model.mm_1_w, embeddings);
+        embeddings = ggml_add(ctx0, embeddings, model.mm_1_b);
+    }
 
     // build the graph
     ggml_build_forward_expand(gf, embeddings);
@@ -1206,6 +1275,10 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             new_clip->minicpmv_version = gguf_get_val_i32(ctx, idx);
         }
 
+        idx = gguf_find_key(ctx, KEY_HAS_QWEN2VL_MERGER);
+        if (idx != -1) {
+            new_clip->has_qwen2vl_merger = gguf_get_val_bool(ctx, idx);
+        }
         // GGML_ASSERT(new_clip->has_llava_projector); // see monatis/clip.cpp for image and/or text encoding for semantic search
 
         GGML_ASSERT(new_clip->has_vision_encoder);
@@ -1214,6 +1287,13 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         idx = get_key_idx(ctx, KEY_USE_GELU);
         new_clip->use_gelu = gguf_get_val_bool(ctx, idx);
 
+        try {
+            idx = get_key_idx(ctx, KEY_USE_SILU);
+            new_clip->use_silu = gguf_get_val_bool(ctx, idx);
+        } catch (std::runtime_error & /*e*/) {
+            new_clip->use_silu = false;
+        }
+
         if (verbosity >= 1) {
             LOG_INF("%s: text_encoder:   %d\n", __func__, new_clip->has_text_encoder);
             LOG_INF("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
@@ -1389,11 +1469,16 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
 
         try {
-            vision_model.patch_embeddings    = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD);
+            vision_model.patch_embeddings_0    = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD);
             vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v"));
         } catch(const std::exception& /*e*/) {
             LOG_ERR("%s: failed to load vision model tensors\n", __func__);
         }
+        try {
+            vision_model.patch_embeddings_1    = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD_1);
+        } catch(const std::exception& /*e*/) {
+            new_clip->has_qwen2vl_merger = false;
+        }
 
         // LLaVA projection
         if (new_clip->proj_type == PROJECTOR_TYPE_MLP || new_clip->proj_type == PROJECTOR_TYPE_MLP_NORM) {
@@ -1481,6 +1566,12 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             vision_model.mm_model_ln_post_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "weight"));
             vision_model.mm_model_ln_post_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "bias"));
         }
+        else if (new_clip->proj_type == PROJECTOR_TYPE_MERGER) {
+            vision_model.mm_0_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "weight"));
+            vision_model.mm_0_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "bias"));
+            vision_model.mm_1_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight"));
+            vision_model.mm_1_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias"));
+        }
         else {
             std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type];
             throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
@@ -1519,6 +1610,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         new_clip->compute_alloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(new_clip->backend));
         clip_image_f32_batch batch;
         batch.size = 1;
+        batch.data = nullptr;
         ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch, nullptr, false);
         ggml_gallocr_reserve(new_clip->compute_alloc, gf);
         size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_clip->compute_alloc, 0);
@@ -1532,6 +1624,10 @@ void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size
     ctx_clip->load_image_size = load_image_size;
 }
 
+struct clip_image_size * clip_get_load_image_size(struct clip_ctx * ctx_clip) {
+    return ctx_clip->load_image_size;
+}
+
 struct clip_image_size * clip_image_size_init() {
     struct clip_image_size * load_image_size = new struct clip_image_size();
     load_image_size->width = 448;
@@ -1984,6 +2080,23 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
         }
         return true;
     }
+    else if (ctx->has_qwen2vl_merger) {
+        clip_image_u8 * resized = clip_image_u8_init();
+        auto patch_size = clip_patch_size(ctx) * 2;
+        int nx = ceil((float)img->nx / patch_size) * patch_size;
+        int ny = ceil((float)img->ny / patch_size) * patch_size;
+        bicubic_resize(*img, *resized, nx, ny);
+
+        res_imgs->data = new clip_image_f32[1];
+        // clip_image_f32 * res = clip_image_f32_init();
+        normalize_image_u8_to_f32(resized, res_imgs->data, ctx->image_mean, ctx->image_std);
+        // res_imgs->data[0] = *res;
+        res_imgs->size = 1;
+
+        // clip_image_f32_free(res);
+        clip_image_u8_free(resized);
+        return true;
+    }
 
     bool pad_to_square = true;
     if (!ctx->has_vision_encoder) {
@@ -2173,6 +2286,13 @@ size_t clip_embd_nbytes(const struct clip_ctx * ctx) {
     return clip_n_patches(ctx) * clip_n_mmproj_embd(ctx) * sizeof(float);
 }
 
+size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w) {
+    clip_image_f32 img;
+    img.nx = img_w;
+    img.ny = img_h;
+    return clip_n_patches_by_img(ctx, &img) * clip_n_mmproj_embd(ctx) * sizeof(float);
+}
+
 int32_t clip_image_size(const struct clip_ctx * ctx) {
     return ctx->vision_model.hparams.image_size;
 }
@@ -2194,6 +2314,13 @@ const int32_t * clip_image_grid(const struct clip_ctx * ctx) {
 }
 
 int clip_n_patches(const struct clip_ctx * ctx) {
+    clip_image_f32 img;
+    img.nx = ctx->vision_model.hparams.image_size;
+    img.ny = ctx->vision_model.hparams.image_size;
+    return clip_n_patches_by_img(ctx, &img);
+}
+
+int clip_n_patches_by_img(const struct clip_ctx * ctx, struct clip_image_f32 * img) {
     const auto & params = ctx->vision_model.hparams;
 
     int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size);
@@ -2207,6 +2334,11 @@ int clip_n_patches(const struct clip_ctx * ctx) {
         else if (ctx->minicpmv_version == 3) {
             n_patches = 64;
         }
+    } else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
+        int patch_size = params.patch_size * 2;
+        int x_patch = img->nx / patch_size + (int)(img->nx % patch_size > 0);
+        int y_patch = img->ny / patch_size + (int)(img->ny % patch_size > 0);
+        n_patches = x_patch * y_patch;
     }
 
     return n_patches;
@@ -2335,7 +2467,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
     const int image_size = hparams.image_size;
     int image_size_width  = image_size;
     int image_size_height = image_size;
-    if (ctx->has_minicpmv_projector) {
+    if (ctx->has_minicpmv_projector | ctx->has_qwen2vl_merger) {
         image_size_width  = imgs->data[0].nx;
         image_size_height = imgs->data[0].ny;
     }
@@ -2355,7 +2487,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         for (size_t i = 0; i < imgs->size; i++) {
             const int nx = imgs->data[i].nx;
             const int ny = imgs->data[i].ny;
-            if (!ctx->has_minicpmv_projector) {
+            if (!(ctx->has_minicpmv_projector | ctx->has_qwen2vl_merger)) {
                 GGML_ASSERT(nx == image_size && ny == image_size);
             }
 
@@ -2413,9 +2545,9 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             auto pos_embed_t = get_2d_sincos_pos_embed(embed_dim, std::make_pair(pos_w, pos_h));
 
             float * pos_embed_data = (float *)malloc(ggml_nbytes(pos_embed));
-            for(int i=0;i<pos_w * pos_h;++i){
+            for(int i=0;i < pos_w * pos_h; ++i){
-                for(int j=0;j<embed_dim;++j){
+                for(int j=0; j < embed_dim; ++j){
-                    pos_embed_data[i*embed_dim+j]=pos_embed_t[i][j];
+                    pos_embed_data[i * embed_dim + j] = pos_embed_t[i][j];
                 }
             }
 
@@ -2435,7 +2567,34 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             }
         }
 
+        if (ctx->has_qwen2vl_merger) {
+            struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
+
+            const int pw = image_size_width / patch_size;
+            const int ph = image_size_height / patch_size;
+            int* positions_data = (int*)malloc(ggml_nbytes(positions));
+
+            int ptr = 0;
+            for (int y = 0; y < ph; y+=2)
             {
+                for (int x = 0; x < pw; x+=2)
+                {
+                    for (int dy = 0; dy < 2; dy++) {
+                        for (int dx = 0; dx < 2; dx++) {
+                            positions_data[ptr]                 = y + dy;
+                            positions_data[num_patches + ptr]     = x + dx;
+                            positions_data[num_patches * 2 + ptr] = y + dy;
+                            positions_data[num_patches * 3 + ptr] = x + dx;
+                            ptr++;
+                        }
+                    }
+                }
+            }
+
+            ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
+            free(positions_data);
+        }
+        else {
             struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
 
             int* positions_data = (int*)malloc(ggml_nbytes(positions));
@@ -2444,7 +2603,6 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             }
             ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
             free(positions_data);
-        }
 
             {
                 struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
@@ -2456,6 +2614,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
                 free(patches_data);
             }
         }
+    }
 
     if (ggml_backend_is_cpu(ctx->backend)) {
         ggml_backend_cpu_set_n_threads(ctx->backend, n_threads);
@@ -2626,6 +2785,9 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
             return 3584;
         }
     }
+    if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
+        return ctx->vision_model.mm_1_b->ne[0];
+    }
 
     std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type];
     throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
@@ -2637,3 +2799,21 @@ int clip_is_minicpmv(const struct clip_ctx * ctx) {
     }
     return 0;
 }
+
+bool clip_is_qwen2vl(const struct clip_ctx * ctx) {
+    return ctx->has_qwen2vl_merger;
+}
+
+
+bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec) {
+    clip_image_f32 clip_img;
+    clip_img.buf.resize(h * w * 3);
+    for (int i = 0; i < h*w*3; i++)
+    {
+        clip_img.buf[i] = img[i];
+    }
+    clip_img.nx = w;
+    clip_img.ny = h;
+    clip_image_encode(ctx, n_threads, &clip_img, vec);
+    return true;
+}

examples/llava/clip.h

@@ -45,6 +45,7 @@ CLIP_API struct clip_ctx * clip_model_load_cpu(const char * fname, int verbosity
 CLIP_API void clip_free(struct clip_ctx * ctx);
 
 CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
+CLIP_API size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w);
 
 CLIP_API int32_t clip_image_size (const struct clip_ctx * ctx);
 CLIP_API int32_t clip_patch_size (const struct clip_ctx * ctx);
@@ -56,10 +57,12 @@ CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);
 CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
 
 CLIP_API int clip_n_patches        (const struct clip_ctx * ctx);
-CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
+CLIP_API int clip_n_patches_by_img (const struct clip_ctx * ctx, struct clip_image_f32 * img);
+CLIP_API int clip_n_mmproj_embd    (const struct clip_ctx * ctx);
 
 CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip);
 CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size);
+CLIP_API struct clip_image_size * clip_get_load_image_size(struct clip_ctx * ctx_clip);
 
 CLIP_API struct clip_image_size * clip_image_size_init();
 CLIP_API struct clip_image_u8  * clip_image_u8_init ();
@@ -86,6 +89,9 @@ CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, cons
 CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype);
 
 CLIP_API int clip_is_minicpmv(const struct clip_ctx * ctx);
+CLIP_API bool clip_is_qwen2vl(const struct clip_ctx * ctx);
+
+CLIP_API bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec);
 
 #ifdef __cplusplus
 }

examples/llava/llava.cpp

@@ -259,18 +259,24 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
 
     const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip);
 
-    if (clip_is_minicpmv(ctx_clip)) {
+    if (clip_is_minicpmv(ctx_clip) || clip_is_qwen2vl(ctx_clip)) {
         std::vector<float *> image_embd_v;
         image_embd_v.resize(img_res_v.size);
         struct clip_image_size * load_image_size = clip_image_size_init();
+
         for (size_t i = 0; i < img_res_v.size; i++) {
             const int64_t t_img_enc_step_start_us = ggml_time_us();
-            image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip));
+            image_embd_v[i] = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, img_res_v.data[i].nx, img_res_v.data[i].ny));
             int patch_size=14;
             load_image_size->width = img_res_v.data[i].nx;
             load_image_size->height = img_res_v.data[i].ny;
             clip_add_load_image_size(ctx_clip, load_image_size);
+
             bool encoded = false;
+            if (clip_is_qwen2vl(ctx_clip)) {
+                encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]);
+            }
+            else {
                 int has_minicpmv_projector = clip_is_minicpmv(ctx_clip);
                 if (has_minicpmv_projector == 2) {
                     encoded = clip_image_encode(ctx_clip, n_threads, only_v2_5_reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]);
@@ -278,6 +284,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
                 else if (has_minicpmv_projector == 3) {
                     encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]);
                 }
+            }
+
             if (!encoded) {
                 LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
                 return false;
@@ -290,8 +298,11 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
 
         int n_img_pos_out = 0;
         for (size_t i = 0; i < image_embd_v.size(); i++) {
-            std::memcpy(image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip), image_embd_v[i], clip_embd_nbytes(ctx_clip));
+            std::memcpy(
-            n_img_pos_out += clip_n_patches(ctx_clip);
+                image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip),
+                image_embd_v[i],
+                clip_embd_nbytes_by_img(ctx_clip, img_res_v.data[i].nx, img_res_v.data[i].ny));
+            n_img_pos_out += clip_n_patches_by_img(ctx_clip, &img_res_v.data[i]);
         }
         *n_img_pos = n_img_pos_out;
         for (size_t i = 0; i < image_embd_v.size(); i++) {
@@ -387,7 +398,13 @@ bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, co
     if (clip_is_minicpmv(ctx_clip)) {
         num_max_patches = 10;
     }
-    float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model
+    float * image_embd;
+    if (clip_is_qwen2vl(ctx_clip)) {
+        // qwen2vl don't split image into chunks, so `num_max_patches` is not needed.
+        image_embd = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, img->nx, img->ny));
+    } else {
+        image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model
+    }
     if (!image_embd) {
         LOG_ERR("Unable to allocate memory for image embeddings\n");
         return false;

examples/llava/qwen2_vl_surgery.py

@@ -0,0 +1,158 @@
+import argparse
+from typing import Dict
+
+import torch
+import numpy as np
+from gguf import *
+from transformers import (
+    Qwen2VLForConditionalGeneration,
+    Qwen2VLProcessor,
+    AutoProcessor,
+    Qwen2VLConfig
+)
+
+
+VISION = "clip.vision"
+
+
+def k(raw_key: str, arch: str) -> str:
+    return raw_key.format(arch=arch)
+
+
+def to_gguf_name(name: str) -> str:
+    og = name
+    name = name.replace("text_model", "t").replace("vision_model", "v")
+    name = name.replace("blocks", "blk").replace("embeddings.", "")
+    name = name.replace("attn.", "attn_")
+    name = name.replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("proj.", "out.")
+    # name = name.replace("layrnorm", "ln").replace("layer_norm", "ln").replace("layernorm", "ln")
+    name = name.replace("norm1", "ln1").replace("norm2", "ln2")
+    name = name.replace("merger.mlp", 'mm')
+    print(f"[to_gguf_name] {og} --> {name}")
+    return name
+
+
+def find_vision_tensors(qwen2vl, dtype) -> Dict[str, np.ndarray]:
+    vision_model = qwen2vl.visual
+    tensor_map = {}
+    for name, ten in vision_model.state_dict().items():
+        ten = ten.numpy()
+        if 'qkv' in name:
+            if ten.ndim == 2: # weight
+                c3, _ = ten.shape
+            else:             # bias
+                c3 = ten.shape[0]
+            assert c3 % 3 == 0
+            c = c3 // 3
+            wq = ten[:c]
+            wk = ten[c: c * 2]
+            wv = ten[c * 2:]
+            tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "q")] = wq
+            tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "k")] = wk
+            tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "v")] = wv
+        elif 'merger' in name:
+            if name.endswith("ln_q.weight"):
+                tensor_map['v.post_ln.weight'] = ten
+            elif name.endswith("ln_q.bias"):
+                tensor_map['v.post_ln.bias'] = ten
+            else:
+                # "merger.mlp.%d.weight/bias" --> "mm.%d.weight/bias"
+                tensor_map[to_gguf_name(name)] = ten
+        elif 'patch_embed.proj.weight' in name:
+            # NOTE: split Conv3D into Conv2Ds
+            c1, c2, kt, kh, kw = ten.shape
+            assert kt == 2, "Current implmentation only support temporal_patch_size of 2"
+            tensor_map["v.patch_embd.weight"] = ten[:, :, 0, ...]
+            tensor_map["v.patch_embd.weight.1"] = ten[:, :, 1, ...]
+        else:
+            tensor_map[to_gguf_name(f"vision_model.{name}")] = ten
+
+    for new_name, ten in tensor_map.items():
+        if ten.ndim <= 1 or new_name.endswith("_norm.weight"):
+            tensor_map[new_name] = ten.astype(np.float32)
+        else:
+            tensor_map[new_name] = ten.astype(dtype)
+    tensor_map["v.position_embd.weight"] = np.zeros([10, 10], dtype=np.float32)  # dummy tensor, just here as a placeholder
+    return tensor_map
+
+
+def main(args):
+    if args.data_type == 'fp32':
+        dtype = torch.float32
+        np_dtype = np.float32
+        ftype = 0
+    elif args.data_type == 'fp16':
+        dtype = torch.float32
+        np_dtype = np.float16
+        ftype = 1
+    else:
+        raise ValueError()
+
+    model_name = args.model_name
+    print("model_name: ", model_name)
+    qwen2vl = Qwen2VLForConditionalGeneration.from_pretrained(
+        model_name, torch_dtype=dtype, device_map="cpu"
+    )
+    cfg: Qwen2VLConfig = qwen2vl.config  # type: ignore[reportAssignmentType]
+    vcfg = cfg.vision_config
+
+    if os.path.isdir(model_name):
+        if model_name.endswith(os.sep):
+            model_name = model_name[:-1]
+        model_name = os.path.basename(model_name)
+    fname_out = f"{model_name.replace('/', '-').lower()}-vision.gguf"
+
+    fout = GGUFWriter(path=fname_out, arch="clip")
+    fout.add_description("image encoder for Qwen2VL")
+
+    fout.add_file_type(ftype)
+    fout.add_bool("clip.has_text_encoder", False)
+    fout.add_bool("clip.has_vision_encoder", True)
+    fout.add_bool("clip.has_qwen2vl_merger", True)
+    fout.add_string("clip.projector_type", "qwen2vl_merger")
+
+    print(cfg.vision_config)
+    if 'silu' in cfg.vision_config.hidden_act.lower():
+        fout.add_bool("clip.use_silu", True)
+        fout.add_bool("clip.use_gelu", False)
+    elif 'gelu' in cfg.vision_config.hidden_act.lower():
+        fout.add_bool("clip.use_silu", False)
+        fout.add_bool("clip.use_gelu", 'quick' not in cfg.vision_config.hidden_act.lower())
+    else:
+        raise ValueError()
+
+    tensor_map = find_vision_tensors(qwen2vl, np_dtype)
+    for name, data in tensor_map.items():
+        fout.add_tensor(name, data)
+
+    fout.add_uint32("clip.vision.patch_size", vcfg.patch_size)
+    fout.add_uint32("clip.vision.image_size", 14 * 40)  # some reasonable size that is divable by (14*2)
+    fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), vcfg.embed_dim)
+    fout.add_uint32("clip.vision.projection_dim", vcfg.hidden_size)
+    fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), vcfg.num_heads)
+    fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
+    fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), vcfg.depth)
+    fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, VISION), 0)  # not sure what this does, put 0 here as a placeholder
+    fout.add_name(model_name)
+    """
+    HACK: Since vision rope related parameter aren't stored in the `Qwen2VLConfig,
+            it will be hardcoded in the `clip_image_build_graph` from `clip.cpp`.
+    """
+
+    processor: Qwen2VLProcessor = AutoProcessor.from_pretrained(model_name)
+    fout.add_array("clip.vision.image_mean", processor.image_processor.image_mean) # type: ignore[reportAttributeAccessIssue]
+    fout.add_array("clip.vision.image_std", processor.image_processor.image_std) # type: ignore[reportAttributeAccessIssue]
+
+    fout.write_header_to_file()
+    fout.write_kv_data_to_file()
+    fout.write_tensors_to_file()
+    fout.close()
+    print("save model as: ", fname_out)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("model_name", nargs='?', default="Qwen/Qwen2-VL-2B-Instruct")
+    parser.add_argument("--data_type", nargs='?', choices=['fp32', 'fp16'], default="fp32")
+    args = parser.parse_args()
+    main(args)

examples/llava/qwen2vl-cli.cpp

@@ -0,0 +1,581 @@
+#include "arg.h"
+#include "base64.hpp"
+#include "log.h"
+#include "common.h"
+#include "sampling.h"
+#include "clip.h"
+#include "llava.h"
+#include "llama.h"
+#include "ggml.h"
+
+#ifdef GGML_USE_CUDA
+#include "ggml-cuda.h"
+#endif
+#ifdef NDEBUG
+#include "ggml-alloc.h"
+#include "ggml-backend.h"
+#endif
+
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <vector>
+#include <algorithm>
+#include <iostream>
+#include <fstream>
+
+
+static bool qwen2vl_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed,
+                                     int n_batch, int * n_past, int * st_pos_id, struct clip_image_size * image_size) {
+    int n_embd  = llama_n_embd(llama_get_model(ctx_llama));
+    const int patch_size = 14 * 2;
+    const int ph = image_size->height / patch_size + (image_size->height % patch_size > 0);
+    const int pw = image_size->width / patch_size + (image_size->width % patch_size > 0);
+    auto img_tokens = image_embed->n_image_pos;
+    // llama_pos mrope_pos[img_tokens * 4];
+    std::vector<llama_pos> mrope_pos;
+    mrope_pos.resize(img_tokens * 4);
+
+    for (int y = 0; y < ph; y++)
+    {
+        for (int x = 0; x < pw; x++)
+        {
+            int i = y * pw + x;
+            mrope_pos[i] = *st_pos_id;
+            mrope_pos[i + img_tokens] = *st_pos_id + y;
+            mrope_pos[i + img_tokens * 2] = *st_pos_id + x;
+            mrope_pos[i + img_tokens * 3] = 0;
+        }
+    }
+    *st_pos_id += std::max(pw, ph);
+
+    int processed = 0;
+    std::vector<llama_pos> batch_mrope_pos;
+    batch_mrope_pos.resize(img_tokens * 4);
+
+    for (int i = 0; i < img_tokens; i += n_batch) {
+        int n_eval = img_tokens - i;
+        if (n_eval > n_batch) {
+            n_eval = n_batch;
+        }
+
+        // llama_pos batch_mrope_pos[n_eval * 4];
+        std::fill(batch_mrope_pos.begin(), batch_mrope_pos.end(), 0);
+        memcpy(batch_mrope_pos.data(), &mrope_pos[processed], n_eval * sizeof(llama_pos));
+        memcpy(&batch_mrope_pos[n_eval * 1], &mrope_pos[img_tokens * 1 + processed], n_eval * sizeof(llama_pos));
+        memcpy(&batch_mrope_pos[n_eval * 2], &mrope_pos[img_tokens * 2 + processed], n_eval * sizeof(llama_pos));
+        memcpy(&batch_mrope_pos[n_eval * 3], &mrope_pos[img_tokens * 3 + processed], n_eval * sizeof(llama_pos));
+
+        llama_batch batch = {
+            int32_t(n_eval),                // n_tokens
+            nullptr,                        // token
+            (image_embed->embed+i*n_embd),  // embed
+            batch_mrope_pos.data(),         // pos
+            nullptr,  // n_seq_id
+            nullptr,  // seq_id
+            nullptr,  // logits
+        };
+
+        if (llama_decode(ctx_llama, batch)) {
+            LOG_ERR("%s : failed to eval\n", __func__);
+            return false;
+        }
+        *n_past += n_eval;
+        processed += n_eval;
+    }
+    return true;
+}
+
+
+static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past, int * st_pos_id) {
+    int N = (int) tokens.size();
+    std::vector<llama_pos> pos;
+    for (int i = 0; i < N; i += n_batch) {
+        int n_eval = (int) tokens.size() - i;
+        if (n_eval > n_batch) {
+            n_eval = n_batch;
+        }
+        auto batch = llama_batch_get_one(&tokens[i], n_eval);
+        // TODO: add mrope pos ids somewhere else
+        pos.resize(batch.n_tokens * 4);
+        std::fill(pos.begin(), pos.end(), 0);
+        for (int j = 0; j < batch.n_tokens * 3; j ++) {
+            pos[j] = *st_pos_id + (j % batch.n_tokens);
+        }
+        batch.pos = pos.data();
+
+        if (llama_decode(ctx_llama, batch)) {
+            LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
+            return false;
+        }
+        *n_past += n_eval;
+        *st_pos_id += n_eval;
+    }
+    return true;
+}
+
+static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past, int * st_pos_id) {
+    std::vector<llama_token> tokens;
+    tokens.push_back(id);
+    return eval_tokens(ctx_llama, tokens, 1, n_past, st_pos_id);
+}
+
+static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, int * st_pos_id, bool add_bos){
+    std::string              str2     = str;
+    std::vector<llama_token> embd_inp = common_tokenize(ctx_llama, str2, add_bos, true);
+    eval_tokens(ctx_llama, embd_inp, n_batch, n_past, st_pos_id);
+    return true;
+}
+
+static const char * sample(struct common_sampler * smpl,
+                           struct llama_context * ctx_llama,
+                           int * n_past, int * st_pos_id) {
+    const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
+    common_sampler_accept(smpl, id, true);
+    static std::string ret;
+    if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
+        ret = "</s>";
+    } else {
+        ret = common_token_to_piece(ctx_llama, id);
+    }
+    eval_id(ctx_llama, id, n_past, st_pos_id);
+    return ret.c_str();
+}
+
+static const char* IMG_BASE64_TAG_BEGIN = "<img src=\"data:image/jpeg;base64,";
+static const char* IMG_BASE64_TAG_END = "\">";
+
+static void find_image_tag_in_prompt(const std::string& prompt, size_t& begin_out, size_t& end_out) {
+    begin_out = prompt.find(IMG_BASE64_TAG_BEGIN);
+    end_out = prompt.find(IMG_BASE64_TAG_END, (begin_out == std::string::npos) ? 0UL : begin_out);
+}
+
+static bool prompt_contains_image(const std::string& prompt) {
+    size_t begin, end;
+    find_image_tag_in_prompt(prompt, begin, end);
+    return (begin != std::string::npos);
+}
+
+// replaces the base64 image tag in the prompt with `replacement`
+static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip_ctx * ctx_clip, int n_threads, const std::string& prompt) {
+    size_t img_base64_str_start, img_base64_str_end;
+    find_image_tag_in_prompt(prompt, img_base64_str_start, img_base64_str_end);
+    if (img_base64_str_start == std::string::npos || img_base64_str_end == std::string::npos) {
+        LOG_ERR("%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
+        return NULL;
+    }
+
+    auto base64_bytes_start = img_base64_str_start + strlen(IMG_BASE64_TAG_BEGIN);
+    auto base64_bytes_count = img_base64_str_end - base64_bytes_start;
+    auto base64_str = prompt.substr(base64_bytes_start, base64_bytes_count );
+
+    auto required_bytes = base64::required_encode_size(base64_str.size());
+    auto img_bytes = std::vector<unsigned char>(required_bytes);
+    base64::decode(base64_str.begin(), base64_str.end(), img_bytes.begin());
+
+    auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, img_bytes.data(), img_bytes.size());
+    if (!embed) {
+        LOG_ERR("%s: could not load image from base64 string.\n", __func__);
+        return NULL;
+    }
+
+    return embed;
+}
+
+static std::string remove_image_from_prompt(const std::string& prompt, const char * replacement = "") {
+    size_t begin, end;
+    find_image_tag_in_prompt(prompt, begin, end);
+    if (begin == std::string::npos || end == std::string::npos) {
+        return prompt;
+    }
+    auto pre = prompt.substr(0, begin);
+    auto post = prompt.substr(end + strlen(IMG_BASE64_TAG_END));
+    return pre + replacement + post;
+}
+
+struct llava_context {
+    struct clip_ctx * ctx_clip = NULL;
+    struct llama_context * ctx_llama = NULL;
+    struct llama_model * model = NULL;
+};
+
+static void print_usage(int, char ** argv) {
+    LOG("\n example usage:\n");
+    LOG("\n     %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
+    LOG("\n note: a lower temperature value like 0.1 is recommended for better quality.\n");
+}
+
+static struct llava_image_embed * load_image(llava_context * ctx_llava, common_params * params, const std::string & fname) {
+
+    // load and preprocess the image
+    llava_image_embed * embed = NULL;
+    auto prompt = params->prompt;
+    if (prompt_contains_image(prompt)) {
+        if (!params->image.empty()) {
+            LOG_INF("using base64 encoded image instead of command line image path\n");
+        }
+        embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->cpuparams.n_threads, prompt);
+        if (!embed) {
+            LOG_ERR("%s: can't load image from prompt\n", __func__);
+            return NULL;
+        }
+        params->prompt = remove_image_from_prompt(prompt);
+    } else {
+        embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->cpuparams.n_threads, fname.c_str());
+        if (!embed) {
+            fprintf(stderr, "%s: is %s really an image file?\n", __func__, fname.c_str());
+            return NULL;
+        }
+    }
+
+    return embed;
+}
+
+static void process_prompt(struct llava_context * ctx_llava, struct llava_image_embed * image_embed, common_params * params, const std::string & prompt) {
+    int n_past = 0;
+    int cur_pos_id = 0;
+
+    const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict;
+
+    std::string system_prompt, user_prompt;
+    size_t image_pos = prompt.find("<|vision_start|>");
+    if (image_pos != std::string::npos) {
+        // new templating mode: Provide the full prompt including system message and use <image> as a placeholder for the image
+        system_prompt = prompt.substr(0, image_pos);
+        user_prompt = prompt.substr(image_pos + std::string("<|vision_pad|>").length());
+        LOG_INF("system_prompt: %s\n", system_prompt.c_str());
+        if (params->verbose_prompt) {
+            auto tmp = common_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
+            for (int i = 0; i < (int) tmp.size(); i++) {
+                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+            }
+        }
+        LOG_INF("user_prompt: %s\n", user_prompt.c_str());
+        if (params->verbose_prompt) {
+            auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
+            for (int i = 0; i < (int) tmp.size(); i++) {
+                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+            }
+        }
+    } else {
+        // llava-1.5 native mode
+        system_prompt = "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n<|vision_start|>";
+        user_prompt = "<|vision_end|>" + prompt + "<|im_end|>\n<|im_start|>assistant\n";
+        if (params->verbose_prompt) {
+            auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
+            for (int i = 0; i < (int) tmp.size(); i++) {
+                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+            }
+        }
+    }
+
+    eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, &cur_pos_id, true);
+    if (image_embed != nullptr) {
+        auto image_size = clip_get_load_image_size(ctx_llava->ctx_clip);
+        qwen2vl_eval_image_embed(ctx_llava->ctx_llama, image_embed, params->n_batch, &n_past, &cur_pos_id, image_size);
+    }
+    eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, &cur_pos_id, false);
+
+    // generate the response
+
+    LOG("\n");
+
+    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sampling);
+    if (!smpl) {
+        LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__);
+        exit(1);
+    }
+
+    std::string response = "";
+    for (int i = 0; i < max_tgt_len; i++) {
+        const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past, &cur_pos_id);
+        response += tmp;
+        if (strcmp(tmp, "</s>") == 0) break;
+        if (strstr(tmp, "###")) break; // Yi-VL behavior
+        LOG("%s", tmp);
+        if (strstr(response.c_str(), "<|im_end|>")) break; // Yi-34B llava-1.6 - for some reason those decode not as the correct token (tokenizer works)
+        if (strstr(response.c_str(), "<|im_start|>")) break; // Yi-34B llava-1.6
+        if (strstr(response.c_str(), "USER:")) break; // mistral llava-1.6
+
+        fflush(stdout);
+    }
+
+    common_sampler_free(smpl);
+    LOG("\n");
+}
+
+static struct llama_model * llava_init(common_params * params) {
+    llama_backend_init();
+    llama_numa_init(params->numa);
+
+    llama_model_params model_params = common_model_params_to_llama(*params);
+
+    llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
+    if (model == NULL) {
+        LOG_ERR("%s: unable to load model\n" , __func__);
+        return NULL;
+    }
+    return model;
+}
+
+static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
+    const char * clip_path = params->mmproj.c_str();
+
+    auto prompt = params->prompt;
+    if (prompt.empty()) {
+        prompt = "describe the image in detail.";
+    }
+
+    auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
+
+
+    llama_context_params ctx_params = common_context_params_to_llama(*params);
+    ctx_params.n_ctx           = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings
+
+    llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
+
+    if (ctx_llama == NULL) {
+        LOG_ERR("%s: failed to create the llama_context\n" , __func__);
+        return NULL;
+    }
+
+    auto * ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
+
+    ctx_llava->ctx_llama = ctx_llama;
+    ctx_llava->ctx_clip = ctx_clip;
+    ctx_llava->model = model;
+    return ctx_llava;
+}
+
+static void llava_free(struct llava_context * ctx_llava) {
+    if (ctx_llava->ctx_clip) {
+        clip_free(ctx_llava->ctx_clip);
+        ctx_llava->ctx_clip = NULL;
+    }
+
+    llama_free(ctx_llava->ctx_llama);
+    llama_free_model(ctx_llava->model);
+    llama_backend_free();
+}
+
+#ifndef NDEBUG
+
+static void debug_test_mrope_2d() {
+    // 1. Initialize backend
+    ggml_backend_t backend = NULL;
+    std::string backend_name = "";
+#ifdef GGML_USE_CUDA
+    fprintf(stderr, "%s: using CUDA backend\n", __func__);
+    backend = ggml_backend_cuda_init(0); // init device 0
+    backend_name = "cuda";
+    if (!backend) {
+        fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
+    }
+#endif
+    // if there aren't GPU Backends fallback to CPU backend
+    if (!backend) {
+        backend = ggml_backend_cpu_init();
+        backend_name = "cpu";
+    }
+
+    // Calculate the size needed to allocate
+    size_t ctx_size = 0;
+    ctx_size += 2 * ggml_tensor_overhead(); // tensors
+    // no need to allocate anything else!
+
+    // 2. Allocate `ggml_context` to store tensor data
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ ctx_size,
+        /*.mem_buffer =*/ NULL,
+        /*.no_alloc   =*/ true, // the tensors will be allocated later by ggml_backend_alloc_ctx_tensors()
+    };
+    struct ggml_context * ctx = ggml_init(params);
+
+    struct ggml_tensor * inp_raw = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, 128, 12, 30);
+    ggml_set_name(inp_raw, "inp_raw");
+    ggml_set_input(inp_raw);
+
+    struct ggml_tensor * pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 30 * 4);
+    ggml_set_name(pos, "pos");
+    ggml_set_input(pos);
+
+    std::vector<float> dummy_q;
+    dummy_q.resize(128 * 12 * 30);
+    std::fill(dummy_q.begin(), dummy_q.end(), 0.1);
+    // memcpy(inp_raw->data, dummy_q.data(), 128 * 12 * 30 * ggml_element_size(inp_raw));
+
+    std::vector<int> pos_id;
+    pos_id.resize(30 * 4);
+    for (int i = 0; i < 30; i ++) {
+        pos_id[i] = i;
+        pos_id[i + 30] = i + 10;
+        pos_id[i + 60] = i + 20;
+        pos_id[i + 90] = i + 30;
+    }
+    int sections[4] = {32, 32, 0, 0};
+
+    // 4. Allocate a `ggml_backend_buffer` to store all tensors
+    ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx, backend);
+
+    // 5. Copy tensor data from main memory (RAM) to backend buffer
+    ggml_backend_tensor_set(inp_raw, dummy_q.data(), 0, ggml_nbytes(inp_raw));
+    ggml_backend_tensor_set(pos, pos_id.data(), 0, ggml_nbytes(pos));
+
+    // 6. Create a `ggml_cgraph` for mul_mat operation
+    struct ggml_cgraph * gf = NULL;
+    struct ggml_context * ctx_cgraph = NULL;
+
+    // create a temporally context to build the graph
+    struct ggml_init_params params0 = {
+        /*.mem_size   =*/ ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead(),
+        /*.mem_buffer =*/ NULL,
+        /*.no_alloc   =*/ true, // the tensors will be allocated later by ggml_gallocr_alloc_graph()
+    };
+    ctx_cgraph = ggml_init(params0);
+    gf = ggml_new_graph(ctx_cgraph);
+
+    struct ggml_tensor * result0 = ggml_rope_multi(
+        ctx_cgraph, inp_raw, pos, nullptr,
+        128/2, sections, LLAMA_ROPE_TYPE_VISION, 32768, 1000000, 1,
+        0, 1, 32, 1);
+
+    // Add "result" tensor and all of its dependencies to the cgraph
+    ggml_build_forward_expand(gf, result0);
+
+    // 7. Create a `ggml_gallocr` for cgraph computation
+    ggml_gallocr_t allocr = ggml_gallocr_new(ggml_backend_get_default_buffer_type(backend));
+    ggml_gallocr_alloc_graph(allocr, gf);
+
+    // 9. Run the computation
+    int n_threads = 1; // Optional: number of threads to perform some operations with multi-threading
+    if (ggml_backend_is_cpu(backend)) {
+        ggml_backend_cpu_set_n_threads(backend, n_threads);
+    }
+    ggml_backend_graph_compute(backend, gf);
+
+    // 10. Retrieve results (output tensors)
+    // in this example, output tensor is always the last tensor in the graph
+    struct ggml_tensor * result = result0;
+    // struct ggml_tensor * result = gf->nodes[gf->n_nodes - 1];
+    float * result_data = (float *)malloc(ggml_nbytes(result));
+    // because the tensor data is stored in device buffer, we need to copy it back to RAM
+    ggml_backend_tensor_get(result, result_data, 0, ggml_nbytes(result));
+    const std::string bin_file = "mrope_2d_" + backend_name +".bin";
+    std::ofstream outFile(bin_file, std::ios::binary);
+
+    if (outFile.is_open()) {
+        outFile.write(reinterpret_cast<const char*>(result_data), ggml_nbytes(result));
+        outFile.close();
+        std::cout << "Data successfully written to " + bin_file << std::endl;
+    } else {
+        std::cerr << "Error opening file!" << std::endl;
+    }
+
+    free(result_data);
+    // 11. Free memory and exit
+    ggml_free(ctx_cgraph);
+    ggml_gallocr_free(allocr);
+    ggml_free(ctx);
+    ggml_backend_buffer_free(buffer);
+    ggml_backend_free(backend);
+}
+
+static void debug_dump_img_embed(struct llava_context * ctx_llava) {
+    int n_embd  = llama_n_embd(llama_get_model(ctx_llava->ctx_llama));
+    int ne = n_embd * 4;
+    float vals[56 * 56 * 3];
+    // float embd[ne];
+    std::vector<float> embd;
+    embd.resize(ne);
+
+    for (int i = 0; i < 56*56; i++)
+    {
+        for (int c = 0; c < 3; c++)
+            vals[i * 3 + c] = (float)(i % (56 * 56)) / (56*56);
+    }
+
+    clip_encode_float_image(ctx_llava->ctx_clip, 16, vals, 56, 56, embd.data());
+
+    std::ofstream outFile("img_embed.bin", std::ios::binary);
+    if (outFile.is_open()) {
+        outFile.write(reinterpret_cast<const char*>(embd.data()), ne * sizeof(float));
+
+        outFile.close();
+        std::cout << "Data successfully written to mrope.bin" << std::endl;
+    } else {
+        std::cerr << "Error opening file!" << std::endl;
+    }
+}
+
+#endif
+
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    common_params params;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, print_usage)) {
+        return 1;
+    }
+
+    common_init();
+
+    if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
+        print_usage(argc, argv);
+        return 1;
+    }
+
+    auto * model = llava_init(&params);
+    if (model == NULL) {
+        fprintf(stderr, "%s: error: failed to init llava model\n", __func__);
+        return 1;
+    }
+
+    if (prompt_contains_image(params.prompt)) {
+        auto * ctx_llava = llava_init_context(&params, model);
+
+        auto * image_embed = load_image(ctx_llava, &params, "");
+
+        // process the prompt
+        process_prompt(ctx_llava, image_embed, &params, params.prompt);
+
+        llama_perf_context_print(ctx_llava->ctx_llama);
+        llava_image_embed_free(image_embed);
+        ctx_llava->model = NULL;
+        llava_free(ctx_llava);
+#ifndef NDEBUG
+    } else if (params.image[0].empty()) {
+        auto ctx_llava = llava_init_context(&params, model);
+
+        debug_test_mrope_2d();
+        debug_dump_img_embed(ctx_llava);
+
+        llama_perf_context_print(ctx_llava->ctx_llama);
+        ctx_llava->model = NULL;
+        llava_free(ctx_llava);
+#endif
+    } else {
+        for (auto & image : params.image) {
+            auto * ctx_llava = llava_init_context(&params, model);
+
+            auto * image_embed = load_image(ctx_llava, &params, image);
+            if (!image_embed) {
+                LOG_ERR("%s: failed to load image %s. Terminating\n\n", __func__, image.c_str());
+                return 1;
+            }
+
+            // process the prompt
+            process_prompt(ctx_llava, image_embed, &params, params.prompt);
+
+            llama_perf_context_print(ctx_llava->ctx_llama);
+            llava_image_embed_free(image_embed);
+            ctx_llava->model = NULL;
+            llava_free(ctx_llava);
+        }
+    }
+
+    llama_free_model(model);
+
+    return 0;
+}

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-qunats tuning](https://github.com/ggerganov/llama.cpp/pull/4996)
+  - [#4996 - k-quants 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: %ld\n", chunks.size());
+    LOG_INF("Number of chunks: %zu\n", chunks.size());
 
     llama_backend_init();
     llama_numa_init(params.numa);

examples/run/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-run)
 add_executable(${TARGET} run.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/run/README.md

@@ -3,5 +3,45 @@
 The purpose of this example is to demonstrate a minimal usage of llama.cpp for running models.
 
 ```bash
-./llama-run Meta-Llama-3.1-8B-Instruct.gguf
+llama-run granite-code
+...
+
+```bash
+llama-run -h
+Description:
+  Runs a llm
+
+Usage:
+  llama-run [options] model [prompt]
+
+Options:
+  -c, --context-size <value>
+      Context size (default: 2048)
+  -n, --ngl <value>
+      Number of GPU layers (default: 0)
+  -h, --help
+      Show help message
+
+Commands:
+  model
+      Model is a string with an optional prefix of
+      huggingface:// (hf://), ollama://, https:// or file://.
+      If no protocol is specified and a file exists in the specified
+      path, file:// is assumed, otherwise if a file does not exist in
+      the specified path, ollama:// is assumed. Models that are being
+      pulled are downloaded with .partial extension while being
+      downloaded and then renamed as the file without the .partial
+      extension when complete.
+
+Examples:
+  llama-run llama3
+  llama-run ollama://granite-code
+  llama-run ollama://smollm:135m
+  llama-run hf://QuantFactory/SmolLM-135M-GGUF/SmolLM-135M.Q2_K.gguf
+  llama-run huggingface://bartowski/SmolLM-1.7B-Instruct-v0.2-GGUF/SmolLM-1.7B-Instruct-v0.2-IQ3_M.gguf
+  llama-run https://example.com/some-file1.gguf
+  llama-run some-file2.gguf
+  llama-run file://some-file3.gguf
+  llama-run --ngl 99 some-file4.gguf
+  llama-run --ngl 99 some-file5.gguf Hello World
 ...

examples/run/run.cpp

@@ -1,113 +1,333 @@
 #if defined(_WIN32)
-#include <windows.h>
+#    include <windows.h>
 #else
-#include <unistd.h>
+#    include <unistd.h>
 #endif
 
-#include <climits>
+#if defined(LLAMA_USE_CURL)
+#    include <curl/curl.h>
+#endif
+
+#include <cstdarg>
 #include <cstdio>
 #include <cstring>
+#include <filesystem>
 #include <iostream>
 #include <sstream>
 #include <string>
-#include <unordered_map>
 #include <vector>
 
+#include "common.h"
+#include "json.hpp"
 #include "llama-cpp.h"
 
-typedef std::unique_ptr<char[]> char_array_ptr;
+#define printe(...)                   \
+    do {                              \
+        fprintf(stderr, __VA_ARGS__); \
+    } while (0)
 
-struct Argument {
+class Opt {
-    std::string flag;
-    std::string help_text;
-};
-
-struct Options {
-    std::string model_path, prompt_non_interactive;
-    int ngl = 99;
-    int n_ctx = 2048;
-};
-
-class ArgumentParser {
   public:
-    ArgumentParser(const char * program_name) : program_name(program_name) {}
+    int init(int argc, const char ** argv) {
-
+        construct_help_str_();
-    void add_argument(const std::string & flag, std::string & var, const std::string & help_text = "") {
+        // Parse arguments
-        string_args[flag] = &var;
+        if (parse(argc, argv)) {
-        arguments.push_back({flag, help_text});
+            printe("Error: Failed to parse arguments.\n");
+            help();
+            return 1;
         }
 
-    void add_argument(const std::string & flag, int & var, const std::string & help_text = "") {
+        // If help is requested, show help and exit
-        int_args[flag] = &var;
+        if (help_) {
-        arguments.push_back({flag, help_text});
+            help();
+            return 2;
+        }
+
+        return 0;  // Success
+    }
+
+    std::string model_;
+    std::string user_;
+    int         context_size_ = 2048, ngl_ = -1;
+
+  private:
+    std::string help_str_;
+    bool        help_ = false;
+
+    void construct_help_str_() {
+        help_str_ =
+            "Description:\n"
+            "  Runs a llm\n"
+            "\n"
+            "Usage:\n"
+            "  llama-run [options] model [prompt]\n"
+            "\n"
+            "Options:\n"
+            "  -c, --context-size <value>\n"
+            "      Context size (default: " +
+            std::to_string(context_size_);
+        help_str_ +=
+            ")\n"
+            "  -n, --ngl <value>\n"
+            "      Number of GPU layers (default: " +
+            std::to_string(ngl_);
+        help_str_ +=
+            ")\n"
+            "  -h, --help\n"
+            "      Show help message\n"
+            "\n"
+            "Commands:\n"
+            "  model\n"
+            "      Model is a string with an optional prefix of \n"
+            "      huggingface:// (hf://), ollama://, https:// or file://.\n"
+            "      If no protocol is specified and a file exists in the specified\n"
+            "      path, file:// is assumed, otherwise if a file does not exist in\n"
+            "      the specified path, ollama:// is assumed. Models that are being\n"
+            "      pulled are downloaded with .partial extension while being\n"
+            "      downloaded and then renamed as the file without the .partial\n"
+            "      extension when complete.\n"
+            "\n"
+            "Examples:\n"
+            "  llama-run llama3\n"
+            "  llama-run ollama://granite-code\n"
+            "  llama-run ollama://smollm:135m\n"
+            "  llama-run hf://QuantFactory/SmolLM-135M-GGUF/SmolLM-135M.Q2_K.gguf\n"
+            "  llama-run huggingface://bartowski/SmolLM-1.7B-Instruct-v0.2-GGUF/SmolLM-1.7B-Instruct-v0.2-IQ3_M.gguf\n"
+            "  llama-run https://example.com/some-file1.gguf\n"
+            "  llama-run some-file2.gguf\n"
+            "  llama-run file://some-file3.gguf\n"
+            "  llama-run --ngl 99 some-file4.gguf\n"
+            "  llama-run --ngl 99 some-file5.gguf Hello World\n";
     }
 
     int parse(int argc, const char ** argv) {
+        int positional_args_i = 0;
         for (int i = 1; i < argc; ++i) {
-            std::string arg = argv[i];
+            if (strcmp(argv[i], "-c") == 0 || strcmp(argv[i], "--context-size") == 0) {
-            if (string_args.count(arg)) {
+                if (i + 1 >= argc) {
-                if (i + 1 < argc) {
-                    *string_args[arg] = argv[++i];
-                } else {
-                    fprintf(stderr, "error: missing value for %s\n", arg.c_str());
-                    print_usage();
                     return 1;
                 }
-            } else if (int_args.count(arg)) {
+
-                if (i + 1 < argc) {
+                context_size_ = std::atoi(argv[++i]);
-                    if (parse_int_arg(argv[++i], *int_args[arg]) != 0) {
+            } else if (strcmp(argv[i], "-n") == 0 || strcmp(argv[i], "--ngl") == 0) {
-                        fprintf(stderr, "error: invalid value for %s: %s\n", arg.c_str(), argv[i]);
+                if (i + 1 >= argc) {
-                        print_usage();
                     return 1;
                 }
+
+                ngl_ = std::atoi(argv[++i]);
+            } else if (strcmp(argv[i], "-h") == 0 || strcmp(argv[i], "--help") == 0) {
+                help_ = true;
+                return 0;
+            } else if (!positional_args_i) {
+                ++positional_args_i;
+                model_ = argv[i];
+            } else if (positional_args_i == 1) {
+                ++positional_args_i;
+                user_ = argv[i];
             } else {
-                    fprintf(stderr, "error: missing value for %s\n", arg.c_str());
+                user_ += " " + std::string(argv[i]);
-                    print_usage();
-                    return 1;
-                }
-            } else {
-                fprintf(stderr, "error: unrecognized argument %s\n", arg.c_str());
-                print_usage();
-                return 1;
             }
         }
 
-        if (string_args["-m"]->empty()) {
+        return model_.empty();  // model_ is the only required value
-            fprintf(stderr, "error: -m is required\n");
+    }
-            print_usage();
+
+    void help() const { printf("%s", help_str_.c_str()); }
+};
+
+struct progress_data {
+    size_t file_size = 0;
+    std::chrono::steady_clock::time_point start_time = std::chrono::steady_clock::now();
+    bool   printed   = false;
+};
+
+struct FileDeleter {
+    void operator()(FILE * file) const {
+        if (file) {
+            fclose(file);
+        }
+    }
+};
+
+typedef std::unique_ptr<FILE, FileDeleter> FILE_ptr;
+
+#ifdef LLAMA_USE_CURL
+class CurlWrapper {
+  public:
+    int init(const std::string & url, const std::vector<std::string> & headers, const std::string & output_file,
+             const bool progress, std::string * response_str = nullptr) {
+        std::string output_file_partial;
+        curl = curl_easy_init();
+        if (!curl) {
             return 1;
         }
 
+        progress_data data;
+        FILE_ptr      out;
+        if (!output_file.empty()) {
+            output_file_partial = output_file + ".partial";
+            out.reset(fopen(output_file_partial.c_str(), "ab"));
+        }
+
+        set_write_options(response_str, out);
+        data.file_size = set_resume_point(output_file_partial);
+        set_progress_options(progress, data);
+        set_headers(headers);
+        perform(url);
+        if (!output_file.empty()) {
+            std::filesystem::rename(output_file_partial, output_file);
+        }
+
         return 0;
     }
 
+    ~CurlWrapper() {
+        if (chunk) {
+            curl_slist_free_all(chunk);
+        }
+
+        if (curl) {
+            curl_easy_cleanup(curl);
+        }
+    }
+
   private:
-    const char * program_name;
+    CURL *              curl  = nullptr;
-    std::unordered_map<std::string, std::string *> string_args;
+    struct curl_slist * chunk = nullptr;
-    std::unordered_map<std::string, int *> int_args;
-    std::vector<Argument> arguments;
 
-    int parse_int_arg(const char * arg, int & value) {
+    void set_write_options(std::string * response_str, const FILE_ptr & out) {
-        char * end;
+        if (response_str) {
-        const long val = std::strtol(arg, &end, 10);
+            curl_easy_setopt(curl, CURLOPT_WRITEFUNCTION, capture_data);
-        if (*end == '\0' && val >= INT_MIN && val <= INT_MAX) {
+            curl_easy_setopt(curl, CURLOPT_WRITEDATA, response_str);
-            value = static_cast<int>(val);
+        } else {
+            curl_easy_setopt(curl, CURLOPT_WRITEFUNCTION, write_data);
+            curl_easy_setopt(curl, CURLOPT_WRITEDATA, out.get());
+        }
+    }
+
+    size_t set_resume_point(const std::string & output_file) {
+        size_t file_size = 0;
+        if (std::filesystem::exists(output_file)) {
+            file_size = std::filesystem::file_size(output_file);
+            curl_easy_setopt(curl, CURLOPT_RESUME_FROM_LARGE, static_cast<curl_off_t>(file_size));
+        }
+
+        return file_size;
+    }
+
+    void set_progress_options(bool progress, progress_data & data) {
+        if (progress) {
+            curl_easy_setopt(curl, CURLOPT_NOPROGRESS, 0L);
+            curl_easy_setopt(curl, CURLOPT_XFERINFODATA, &data);
+            curl_easy_setopt(curl, CURLOPT_XFERINFOFUNCTION, progress_callback);
+        }
+    }
+
+    void set_headers(const std::vector<std::string> & headers) {
+        if (!headers.empty()) {
+            if (chunk) {
+                curl_slist_free_all(chunk);
+                chunk = 0;
+            }
+
+            for (const auto & header : headers) {
+                chunk = curl_slist_append(chunk, header.c_str());
+            }
+
+            curl_easy_setopt(curl, CURLOPT_HTTPHEADER, chunk);
+        }
+    }
+
+    void perform(const std::string & url) {
+        CURLcode res;
+        curl_easy_setopt(curl, CURLOPT_URL, url.c_str());
+        curl_easy_setopt(curl, CURLOPT_FOLLOWLOCATION, 1L);
+        curl_easy_setopt(curl, CURLOPT_DEFAULT_PROTOCOL, "https");
+        curl_easy_setopt(curl, CURLOPT_FAILONERROR, 1L);
+        res = curl_easy_perform(curl);
+        if (res != CURLE_OK) {
+            printe("curl_easy_perform() failed: %s\n", curl_easy_strerror(res));
+        }
+    }
+
+    static std::string human_readable_time(double seconds) {
+        int hrs  = static_cast<int>(seconds) / 3600;
+        int mins = (static_cast<int>(seconds) % 3600) / 60;
+        int secs = static_cast<int>(seconds) % 60;
+
+        std::ostringstream out;
+        if (hrs > 0) {
+            out << hrs << "h " << std::setw(2) << std::setfill('0') << mins << "m " << std::setw(2) << std::setfill('0')
+                << secs << "s";
+        } else if (mins > 0) {
+            out << mins << "m " << std::setw(2) << std::setfill('0') << secs << "s";
+        } else {
+            out << secs << "s";
+        }
+
+        return out.str();
+    }
+
+    static std::string human_readable_size(curl_off_t size) {
+        static const char * suffix[] = { "B", "KB", "MB", "GB", "TB" };
+        char         length   = sizeof(suffix) / sizeof(suffix[0]);
+        int          i        = 0;
+        double       dbl_size = size;
+        if (size > 1024) {
+            for (i = 0; (size / 1024) > 0 && i < length - 1; i++, size /= 1024) {
+                dbl_size = size / 1024.0;
+            }
+        }
+
+        std::ostringstream out;
+        out << std::fixed << std::setprecision(2) << dbl_size << " " << suffix[i];
+        return out.str();
+    }
+
+    static int progress_callback(void * ptr, curl_off_t total_to_download, curl_off_t now_downloaded, curl_off_t,
+                                 curl_off_t) {
+        progress_data * data = static_cast<progress_data *>(ptr);
+        if (total_to_download <= 0) {
             return 0;
         }
-        return 1;
+
+        total_to_download += data->file_size;
+        const curl_off_t now_downloaded_plus_file_size = now_downloaded + data->file_size;
+        const curl_off_t percentage                    = (now_downloaded_plus_file_size * 100) / total_to_download;
+        const curl_off_t pos                           = (percentage / 5);
+        std::string progress_bar;
+        for (int i = 0; i < 20; ++i) {
+            progress_bar.append((i < pos) ? "█" : " ");
         }
 
-    void print_usage() const {
+        // Calculate download speed and estimated time to completion
-        printf("\nUsage:\n");
+        const auto                          now             = std::chrono::steady_clock::now();
-        printf("  %s [OPTIONS]\n\n", program_name);
+        const std::chrono::duration<double> elapsed_seconds = now - data->start_time;
-        printf("Options:\n");
+        const double                        speed           = now_downloaded / elapsed_seconds.count();
-        for (const auto & arg : arguments) {
+        const double                        estimated_time  = (total_to_download - now_downloaded) / speed;
-            printf("  %-10s %s\n", arg.flag.c_str(), arg.help_text.c_str());
+        printe("\r%ld%% |%s| %s/%s  %.2f MB/s  %s      ", percentage, progress_bar.c_str(),
+               human_readable_size(now_downloaded).c_str(), human_readable_size(total_to_download).c_str(),
+               speed / (1024 * 1024), human_readable_time(estimated_time).c_str());
+        fflush(stderr);
+        data->printed = true;
+
+        return 0;
     }
 
-        printf("\n");
+    // Function to write data to a file
+    static size_t write_data(void * ptr, size_t size, size_t nmemb, void * stream) {
+        FILE * out = static_cast<FILE *>(stream);
+        return fwrite(ptr, size, nmemb, out);
+    }
+
+    // Function to capture data into a string
+    static size_t capture_data(void * ptr, size_t size, size_t nmemb, void * stream) {
+        std::string * str = static_cast<std::string *>(stream);
+        str->append(static_cast<char *>(ptr), size * nmemb);
+        return size * nmemb;
     }
 };
+#endif
 
 class LlamaData {
   public:
@@ -115,14 +335,16 @@ class LlamaData {
     llama_sampler_ptr               sampler;
     llama_context_ptr               context;
     std::vector<llama_chat_message> messages;
+    std::vector<std::string>        msg_strs;
+    std::vector<char>               fmtted;
 
-    int init(const Options & opt) {
+    int init(Opt & opt) {
-        model = initialize_model(opt.model_path, opt.ngl);
+        model = initialize_model(opt);
         if (!model) {
             return 1;
         }
 
-        context = initialize_context(model, opt.n_ctx);
+        context = initialize_context(model, opt.context_size_);
         if (!context) {
             return 1;
         }
@@ -132,14 +354,122 @@ class LlamaData {
     }
 
   private:
-    // Initializes the model and returns a unique pointer to it
+#ifdef LLAMA_USE_CURL
-    llama_model_ptr initialize_model(const std::string & model_path, const int ngl) {
+    int download(const std::string & url, const std::vector<std::string> & headers, const std::string & output_file,
-        llama_model_params model_params = llama_model_default_params();
+                 const bool progress, std::string * response_str = nullptr) {
-        model_params.n_gpu_layers = ngl;
+        CurlWrapper curl;
+        if (curl.init(url, headers, output_file, progress, response_str)) {
+            return 1;
+        }
 
-        llama_model_ptr model(llama_load_model_from_file(model_path.c_str(), model_params));
+        return 0;
+    }
+#else
+    int download(const std::string &, const std::vector<std::string> &, const std::string &, const bool,
+                 std::string * = nullptr) {
+        printe("%s: llama.cpp built without libcurl, downloading from an url not supported.\n", __func__);
+        return 1;
+    }
+#endif
+
+    int huggingface_dl(const std::string & model, const std::vector<std::string> headers, const std::string & bn) {
+        // Find the second occurrence of '/' after protocol string
+        size_t pos = model.find('/');
+        pos        = model.find('/', pos + 1);
+        if (pos == std::string::npos) {
+            return 1;
+        }
+
+        const std::string hfr = model.substr(0, pos);
+        const std::string hff = model.substr(pos + 1);
+        const std::string url = "https://huggingface.co/" + hfr + "/resolve/main/" + hff;
+        return download(url, headers, bn, true);
+    }
+
+    int ollama_dl(std::string & model, const std::vector<std::string> headers, const std::string & bn) {
+        if (model.find('/') == std::string::npos) {
+            model = "library/" + model;
+        }
+
+        std::string model_tag = "latest";
+        size_t      colon_pos = model.find(':');
+        if (colon_pos != std::string::npos) {
+            model_tag = model.substr(colon_pos + 1);
+            model     = model.substr(0, colon_pos);
+        }
+
+        std::string manifest_url = "https://registry.ollama.ai/v2/" + model + "/manifests/" + model_tag;
+        std::string manifest_str;
+        const int   ret = download(manifest_url, headers, "", false, &manifest_str);
+        if (ret) {
+            return ret;
+        }
+
+        nlohmann::json manifest = nlohmann::json::parse(manifest_str);
+        std::string    layer;
+        for (const auto & l : manifest["layers"]) {
+            if (l["mediaType"] == "application/vnd.ollama.image.model") {
+                layer = l["digest"];
+                break;
+            }
+        }
+
+        std::string blob_url = "https://registry.ollama.ai/v2/" + model + "/blobs/" + layer;
+        return download(blob_url, headers, bn, true);
+    }
+
+    std::string basename(const std::string & path) {
+        const size_t pos = path.find_last_of("/\\");
+        if (pos == std::string::npos) {
+            return path;
+        }
+
+        return path.substr(pos + 1);
+    }
+
+    int remove_proto(std::string & model_) {
+        const std::string::size_type pos = model_.find("://");
+        if (pos == std::string::npos) {
+            return 1;
+        }
+
+        model_ = model_.substr(pos + 3);  // Skip past "://"
+        return 0;
+    }
+
+    int resolve_model(std::string & model_) {
+        const std::string              bn      = basename(model_);
+        const std::vector<std::string> headers = { "--header",
+                                                   "Accept: application/vnd.docker.distribution.manifest.v2+json" };
+        int                            ret     = 0;
+        if (string_starts_with(model_, "file://") || std::filesystem::exists(bn)) {
+            remove_proto(model_);
+        } else if (string_starts_with(model_, "hf://") || string_starts_with(model_, "huggingface://")) {
+            remove_proto(model_);
+            ret = huggingface_dl(model_, headers, bn);
+        } else if (string_starts_with(model_, "ollama://")) {
+            remove_proto(model_);
+            ret = ollama_dl(model_, headers, bn);
+        } else if (string_starts_with(model_, "https://")) {
+            download(model_, headers, bn, true);
+        } else {
+            ret = ollama_dl(model_, headers, bn);
+        }
+
+        model_ = bn;
+
+        return ret;
+    }
+
+    // Initializes the model and returns a unique pointer to it
+    llama_model_ptr initialize_model(Opt & opt) {
+        ggml_backend_load_all();
+        llama_model_params model_params = llama_model_default_params();
+        model_params.n_gpu_layers       = opt.ngl_ >= 0 ? opt.ngl_ : model_params.n_gpu_layers;
+        resolve_model(opt.model_);
+        llama_model_ptr model(llama_load_model_from_file(opt.model_.c_str(), model_params));
         if (!model) {
-            fprintf(stderr, "%s: error: unable to load model\n", __func__);
+            printe("%s: error: unable to load model from file: %s\n", __func__, opt.model_.c_str());
         }
 
         return model;
@@ -150,10 +480,9 @@ class LlamaData {
         llama_context_params ctx_params = llama_context_default_params();
         ctx_params.n_ctx                = n_ctx;
         ctx_params.n_batch              = n_ctx;
-
         llama_context_ptr context(llama_new_context_with_model(model.get(), ctx_params));
         if (!context) {
-            fprintf(stderr, "%s: error: failed to create the llama_context\n", __func__);
+            printe("%s: error: failed to create the llama_context\n", __func__);
         }
 
         return context;
@@ -170,23 +499,22 @@ class LlamaData {
     }
 };
 
-// Add a message to `messages` and store its content in `owned_content`
+// Add a message to `messages` and store its content in `msg_strs`
-static void add_message(const char * role, const std::string & text, LlamaData & llama_data,
+static void add_message(const char * role, const std::string & text, LlamaData & llama_data) {
-                        std::vector<char_array_ptr> & owned_content) {
+    llama_data.msg_strs.push_back(std::move(text));
-    char_array_ptr content(new char[text.size() + 1]);
+    llama_data.messages.push_back({ role, llama_data.msg_strs.back().c_str() });
-    std::strcpy(content.get(), text.c_str());
-    llama_data.messages.push_back({role, content.get()});
-    owned_content.push_back(std::move(content));
 }
 
 // Function to apply the chat template and resize `formatted` if needed
-static int apply_chat_template(const LlamaData & llama_data, std::vector<char> & formatted, const bool append) {
+static int apply_chat_template(LlamaData & llama_data, const bool append) {
-    int result = llama_chat_apply_template(llama_data.model.get(), nullptr, llama_data.messages.data(),
+    int result = llama_chat_apply_template(
-                                           llama_data.messages.size(), append, formatted.data(), formatted.size());
+        llama_data.model.get(), nullptr, llama_data.messages.data(), llama_data.messages.size(), append,
-    if (result > static_cast<int>(formatted.size())) {
+        append ? llama_data.fmtted.data() : nullptr, append ? llama_data.fmtted.size() : 0);
-        formatted.resize(result);
+    if (append && result > static_cast<int>(llama_data.fmtted.size())) {
+        llama_data.fmtted.resize(result);
         result = llama_chat_apply_template(llama_data.model.get(), nullptr, llama_data.messages.data(),
-                                           llama_data.messages.size(), append, formatted.data(), formatted.size());
+                                           llama_data.messages.size(), append, llama_data.fmtted.data(),
+                                           llama_data.fmtted.size());
     }
 
     return result;
@@ -199,7 +527,8 @@ static int tokenize_prompt(const llama_model_ptr & model, const std::string & pr
     prompt_tokens.resize(n_prompt_tokens);
     if (llama_tokenize(model.get(), prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), true,
                        true) < 0) {
-        GGML_ABORT("failed to tokenize the prompt\n");
+        printe("failed to tokenize the prompt\n");
+        return -1;
     }
 
     return n_prompt_tokens;
@@ -211,7 +540,7 @@ static int check_context_size(const llama_context_ptr & ctx, const llama_batch &
     const int n_ctx_used = llama_get_kv_cache_used_cells(ctx.get());
     if (n_ctx_used + batch.n_tokens > n_ctx) {
         printf("\033[0m\n");
-        fprintf(stderr, "context size exceeded\n");
+        printe("context size exceeded\n");
         return 1;
     }
 
@@ -223,7 +552,8 @@ static int convert_token_to_string(const llama_model_ptr & model, const llama_to
     char buf[256];
     int  n = llama_token_to_piece(model.get(), token_id, buf, sizeof(buf), 0, true);
     if (n < 0) {
-        GGML_ABORT("failed to convert token to piece\n");
+        printe("failed to convert token to piece\n");
+        return 1;
     }
 
     piece = std::string(buf, n);
@@ -238,19 +568,19 @@ static void print_word_and_concatenate_to_response(const std::string & piece, st
 
 // helper function to evaluate a prompt and generate a response
 static int generate(LlamaData & llama_data, const std::string & prompt, std::string & response) {
-    std::vector<llama_token> prompt_tokens;
+    std::vector<llama_token> tokens;
-    const int n_prompt_tokens = tokenize_prompt(llama_data.model, prompt, prompt_tokens);
+    if (tokenize_prompt(llama_data.model, prompt, tokens) < 0) {
-    if (n_prompt_tokens < 0) {
         return 1;
     }
 
     // prepare a batch for the prompt
-    llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size());
+    llama_batch batch = llama_batch_get_one(tokens.data(), tokens.size());
     llama_token new_token_id;
     while (true) {
         check_context_size(llama_data.context, batch);
         if (llama_decode(llama_data.context.get(), batch)) {
-            GGML_ABORT("failed to decode\n");
+            printe("failed to decode\n");
+            return 1;
         }
 
         // sample the next token, check is it an end of generation?
@@ -273,22 +603,9 @@ static int generate(LlamaData & llama_data, const std::string & prompt, std::str
     return 0;
 }
 
-static int parse_arguments(const int argc, const char ** argv, Options & opt) {
-    ArgumentParser parser(argv[0]);
-    parser.add_argument("-m", opt.model_path, "model");
-    parser.add_argument("-p", opt.prompt_non_interactive, "prompt");
-    parser.add_argument("-c", opt.n_ctx, "context_size");
-    parser.add_argument("-ngl", opt.ngl, "n_gpu_layers");
-    if (parser.parse(argc, argv)) {
-        return 1;
-    }
-
-    return 0;
-}
-
 static int read_user_input(std::string & user) {
     std::getline(std::cin, user);
-    return user.empty();  // Indicate an error or empty input
+    return user.empty();  // Should have data in happy path
 }
 
 // Function to generate a response based on the prompt
@@ -296,7 +613,7 @@ static int generate_response(LlamaData & llama_data, const std::string & prompt,
     // Set response color
     printf("\033[33m");
     if (generate(llama_data, prompt, response)) {
-        fprintf(stderr, "failed to generate response\n");
+        printe("failed to generate response\n");
         return 1;
     }
 
@@ -306,11 +623,10 @@ static int generate_response(LlamaData & llama_data, const std::string & prompt,
 }
 
 // Helper function to apply the chat template and handle errors
-static int apply_chat_template_with_error_handling(const LlamaData & llama_data, std::vector<char> & formatted,
+static int apply_chat_template_with_error_handling(LlamaData & llama_data, const bool append, int & output_length) {
-                                                   const bool is_user_input, int & output_length) {
+    const int new_len = apply_chat_template(llama_data, append);
-    const int new_len = apply_chat_template(llama_data, formatted, is_user_input);
     if (new_len < 0) {
-        fprintf(stderr, "failed to apply the chat template\n");
+        printe("failed to apply the chat template\n");
         return -1;
     }
 
@@ -319,49 +635,56 @@ static int apply_chat_template_with_error_handling(const LlamaData & llama_data,
 }
 
 // Helper function to handle user input
-static bool handle_user_input(std::string & user_input, const std::string & prompt_non_interactive) {
+static int handle_user_input(std::string & user_input, const std::string & user_) {
-    if (!prompt_non_interactive.empty()) {
+    if (!user_.empty()) {
-        user_input = prompt_non_interactive;
+        user_input = user_;
-        return true;  // No need for interactive input
+        return 0;  // No need for interactive input
     }
 
-    printf("\033[32m> \033[0m");
+    printf(
-    return !read_user_input(user_input);  // Returns false if input ends the loop
+        "\r                                                                       "
+        "\r\033[32m> \033[0m");
+    return read_user_input(user_input);  // Returns true if input ends the loop
 }
 
 // Function to tokenize the prompt
-static int chat_loop(LlamaData & llama_data, std::string & prompt_non_interactive) {
+static int chat_loop(LlamaData & llama_data, const std::string & user_) {
-    std::vector<char_array_ptr> owned_content;
-    std::vector<char> fmtted(llama_n_ctx(llama_data.context.get()));
     int prev_len = 0;
-
+    llama_data.fmtted.resize(llama_n_ctx(llama_data.context.get()));
     while (true) {
         // Get user input
         std::string user_input;
-        if (!handle_user_input(user_input, prompt_non_interactive)) {
+        while (handle_user_input(user_input, user_)) {
-            break;
         }
 
-        add_message("user", prompt_non_interactive.empty() ? user_input : prompt_non_interactive, llama_data,
+        add_message("user", user_.empty() ? user_input : user_, llama_data);
-                    owned_content);
-
         int new_len;
-        if (apply_chat_template_with_error_handling(llama_data, fmtted, true, new_len) < 0) {
+        if (apply_chat_template_with_error_handling(llama_data, true, new_len) < 0) {
             return 1;
         }
 
-        std::string prompt(fmtted.begin() + prev_len, fmtted.begin() + new_len);
+        std::string prompt(llama_data.fmtted.begin() + prev_len, llama_data.fmtted.begin() + new_len);
         std::string response;
         if (generate_response(llama_data, prompt, response)) {
             return 1;
         }
+
+        if (!user_.empty()) {
+            break;
         }
+
+        add_message("assistant", response, llama_data);
+        if (apply_chat_template_with_error_handling(llama_data, false, prev_len) < 0) {
+            return 1;
+        }
+    }
+
     return 0;
 }
 
 static void log_callback(const enum ggml_log_level level, const char * text, void *) {
     if (level == GGML_LOG_LEVEL_ERROR) {
-        fprintf(stderr, "%s", text);
+        printe("%s", text);
     }
 }
 
@@ -382,17 +705,20 @@ static std::string read_pipe_data() {
 }
 
 int main(int argc, const char ** argv) {
-    Options opt;
+    Opt       opt;
-    if (parse_arguments(argc, argv, opt)) {
+    const int ret = opt.init(argc, argv);
+    if (ret == 2) {
+        return 0;
+    } else if (ret) {
         return 1;
     }
 
     if (!is_stdin_a_terminal()) {
-        if (!opt.prompt_non_interactive.empty()) {
+        if (!opt.user_.empty()) {
-            opt.prompt_non_interactive += "\n\n";
+            opt.user_ += "\n\n";
         }
 
-        opt.prompt_non_interactive += read_pipe_data();
+        opt.user_ += read_pipe_data();
     }
 
     llama_log_set(log_callback, nullptr);
@@ -401,7 +727,7 @@ int main(int argc, const char ** argv) {
         return 1;
     }
 
-    if (chat_loop(llama_data, opt.prompt_non_interactive)) {
+    if (chat_loop(llama_data, opt.user_)) {
         return 1;
     }
 

examples/server/README.md

@@ -62,8 +62,8 @@ The project is under active development, and we are [looking for feedback and co
 | `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: 32.0)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
 | `-dkvc, --dump-kv-cache` | verbose print of the KV cache |
 | `-nkvo, --no-kv-offload` | disable KV offload<br/>(env: LLAMA_ARG_NO_KV_OFFLOAD) |
-| `-ctk, --cache-type-k TYPE` | KV cache data type for K (default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_K) |
+| `-ctk, --cache-type-k TYPE` | KV cache data type for K<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_K) |
-| `-ctv, --cache-type-v TYPE` | KV cache data type for V (default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_V) |
+| `-ctv, --cache-type-v TYPE` | KV cache data type for V<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_V) |
 | `-dt, --defrag-thold N` | KV cache defragmentation threshold (default: 0.1, < 0 - disabled)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
 | `-np, --parallel N` | number of parallel sequences to decode (default: 1)<br/>(env: LLAMA_ARG_N_PARALLEL) |
 | `--mlock` | force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
@@ -138,6 +138,7 @@ The project is under active development, and we are [looking for feedback and co
 | -------- | ----------- |
 | `--no-context-shift` | disables context shift on inifinite text generation (default: disabled)<br/>(env: LLAMA_ARG_NO_CONTEXT_SHIFT) |
 | `-sp, --special` | special tokens output enabled (default: false) |
+| `--no-warmup` | skip warming up the model with an empty run |
 | `--spm-infill` | use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled) |
 | `--pooling {none,mean,cls,last,rank}` | pooling type for embeddings, use model default if unspecified<br/>(env: LLAMA_ARG_POOLING) |
 | `-cb, --cont-batching` | enable continuous batching (a.k.a dynamic batching) (default: enabled)<br/>(env: LLAMA_ARG_CONT_BATCHING) |
@@ -146,6 +147,7 @@ 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 (default: enabled)<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) |
@@ -163,13 +165,13 @@ The project is under active development, and we are [looking for feedback and co
 | `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>list of built-in templates:<br/>chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, exaone3, gemma, granite, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, monarch, openchat, orion, phi3, rwkv-world, vicuna, vicuna-orca, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
 | `-sps, --slot-prompt-similarity SIMILARITY` | how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled)<br/> |
 | `--lora-init-without-apply` | load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled) |
-| `--draft-max, --draft, --draft-n N` | number of tokens to draft for speculative decoding (default: 16) |
+| `--draft-max, --draft, --draft-n N` | number of tokens to draft for speculative decoding (default: 16)<br/>(env: LLAMA_ARG_DRAFT_MAX) |
-| `--draft-min, --draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 5) |
+| `--draft-min, --draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 5)<br/>(env: LLAMA_ARG_DRAFT_MIN) |
-| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.9) |
+| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.9)<br/>(env: LLAMA_ARG_DRAFT_P_MIN) |
-| `-cd, --ctx-size-draft N` | size of the prompt context for the draft model (default: 0, 0 = loaded from model) |
+| `-cd, --ctx-size-draft N` | size of the prompt context for the draft model (default: 0, 0 = loaded from model)<br/>(env: LLAMA_ARG_CTX_SIZE_DRAFT) |
 | `-devd, --device-draft <dev1,dev2,..>` | comma-separated list of devices to use for offloading the draft model (none = don't offload)<br/>use --list-devices to see a list of available devices |
-| `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | number of layers to store in VRAM for the draft model |
+| `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | number of layers to store in VRAM for the draft model<br/>(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) |
-| `-md, --model-draft FNAME` | draft model for speculative decoding (default: unused) |
+| `-md, --model-draft FNAME` | draft model for speculative decoding (default: unused)<br/>(env: LLAMA_ARG_MODEL_DRAFT) |
 
 
 Note: If both command line argument and environment variable are both set for the same param, the argument will take precedence over env var.
@@ -302,23 +304,23 @@ mkdir llama-client
 cd llama-client
 ```
 
-Create a index.js file and put this inside:
+Create an 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: 512,
+            n_predict: 64,
         })
     })
     console.log((await response.json()).content)
 }
 
-Test()
+test()
 ```
 
 And run it:
@@ -380,7 +382,7 @@ 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`: It allows receiving each predicted token in real-time instead of waiting for the completion to finish. To enable this, set to `true`.
+`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`.
 
 `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: `[]`
@@ -441,11 +443,11 @@ 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: When using streaming mode (`stream`), only `content` and `stop` will be returned until end of completion.
+- 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.
 
 - `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

@@ -1145,9 +1145,9 @@ struct server_slot {
 
         SLT_INF(*this,
                 "\n"
-                "\rprompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n"
+                "prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n"
-                "\r       eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n"
+                "       eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n"
-                "\r      total time = %10.2f ms / %5d tokens\n",
+                "      total time = %10.2f ms / %5d tokens\n",
                 t_prompt_processing, n_prompt_tokens_processed, t_prompt, n_prompt_second,
                 t_token_generation, n_decoded, t_gen, n_gen_second,
                 t_prompt_processing + t_token_generation, n_prompt_tokens_processed + n_decoded);
@@ -4000,6 +4000,9 @@ 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
@@ -4015,6 +4018,7 @@ int main(int argc, char ** argv) {
                 return false;
             });
         }
+    }
 
     // register API routes
     svr->Get ("/health",              handle_health); // public endpoint (no API key check)

examples/server/tests/unit/test_basic.py

@@ -1,4 +1,5 @@
 import pytest
+import requests
 from utils import *
 
 server = ServerPreset.tinyllama2()
@@ -76,3 +77,20 @@ 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/utils.py

@@ -74,6 +74,7 @@ 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
@@ -164,6 +165,8 @@ 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)}")

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-[calc(100vw-2em)] py-4 px-4">
+        <div class="flex flex-col bg-base-200 min-h-full max-w-64 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>
 
@@ -120,51 +120,25 @@
             {{ messages.length === 0 ? 'Send a message to start' : '' }}
           </div>
           <div v-for="msg in messages" class="group">
-            <div :class="{
+            <message-bubble
-              'chat': true,
+              :config="config"
-              'chat-start': msg.role !== 'user',
+              :msg="msg"
-              'chat-end': msg.role === 'user',
+              :key="msg.id"
-            }">
+              :is-generating="isGenerating"
-              <div :class="{
+              :edit-user-msg-and-regenerate="editUserMsgAndRegenerate"
-                'chat-bubble markdown': true,
+              :regenerate-msg="regenerateMsg"></message-bubble>
-                '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="chat chat-start">
+          <div id="pending-msg" class="group">
-            <div v-if="pendingMsg" class="chat-bubble markdown chat-bubble-base-300">
+            <message-bubble
-              <span v-if="!pendingMsg.content" class="loading loading-dots loading-md"></span>
+              v-if="pendingMsg"
-              <vue-markdown v-else :source="pendingMsg.content" />
+              :config="config"
-            </div>
+              :msg="pendingMsg"
+              :key="pendingMsg.id"
+              :is-generating="isGenerating"
+              :edit-user-msg-and-regenerate="() => {}"
+              :regenerate-msg="() => {}"></message-bubble>
           </div>
         </div>
 
@@ -227,6 +201,10 @@
           <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>
@@ -247,6 +225,66 @@
 
   </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

@@ -13,6 +13,7 @@
         "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"
       },
@@ -2677,6 +2678,12 @@
         "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

@@ -17,6 +17,7 @@
     "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

@@ -1,225 +0,0 @@
-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

@@ -1,21 +1,25 @@
 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 isNumeric = (n) => !isString(n) && !isNaN(n);
+const isBoolean = (x) => x === true || x === false;
+const isNumeric = (n) => !isString(n) && !isNaN(n) && !isBoolean(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(); // for production
+  || (new URL('.', document.baseURI).href).toString().replace(/\/$/, ''); // 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,
@@ -101,6 +105,48 @@ 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-'
@@ -192,10 +238,29 @@ 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 {
@@ -209,7 +274,6 @@ const mainApp = createApp({
       selectedTheme: StorageUtils.getTheme(),
       config: StorageUtils.getConfig(),
       showConfigDialog: false,
-      editingMsg: null,
       // const
       themes: THEMES,
       configDefault: {...CONFIG_DEFAULT},
@@ -226,6 +290,15 @@ 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;
@@ -237,18 +310,10 @@ 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;
@@ -256,7 +321,6 @@ const mainApp = createApp({
         StorageUtils.remove(convId);
         if (this.viewingConvId === convId) {
           this.viewingConvId = StorageUtils.getNewConvId();
-          this.editingMsg = null;
         }
         this.fetchConversation();
         this.fetchMessages();
@@ -291,7 +355,6 @@ const mainApp = createApp({
       this.fetchConversation();
       this.fetchMessages();
       this.inputMsg = '';
-      this.editingMsg = null;
       this.generateMessage(currConvId);
       chatScrollToBottom();
     },
@@ -299,7 +362,6 @@ 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();
@@ -330,17 +392,21 @@ 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 config = {
+        const chunks = sendSSEPostRequest(`${BASE_URL}/v1/chat/completions`, {
-          controller: abortController,
+          method: 'POST',
-          api_url: BASE_URL,
+          headers: {
-          endpoint: '/chat/completions',
+            'Content-Type': 'application/json',
-        };
+            'Authorization': this.config.apiKey ? `Bearer ${this.config.apiKey}` : undefined,
-        for await (const chunk of llama(prompt, params, config)) {
+          },
-          const stop = chunk.data.stop;
+          body: JSON.stringify(params),
-          const addedContent = chunk.data.choices[0].delta.content;
+          signal: abortController.signal,
+        });
+        for await (const chunk of chunks) {
+          const stop = chunk.stop;
+          const addedContent = chunk.choices[0].delta.content;
           const lastContent = this.pendingMsg.content || '';
           if (addedContent) {
             this.pendingMsg = {
@@ -349,6 +415,16 @@ 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);
@@ -387,14 +463,10 @@ 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: %ld\n", tokens.size());
+        printf("Total number of tokens: %zu\n", tokens.size());
     }
     // silence valgrind
     llama_free(ctx);

ggml/CMakeLists.txt

@@ -32,6 +32,13 @@ 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)
 
@@ -172,6 +179,11 @@ set   (GGML_SYCL_TARGET "INTEL" CACHE STRING
 set   (GGML_SYCL_DEVICE_ARCH "" CACHE STRING
                                             "ggml: sycl device architecture")
 
+option(GGML_OPENCL                          "ggml: use OpenCL"                                OFF)
+option(GGML_OPENCL_PROFILING                "ggml: use OpenCL profiling (increases overhead)" OFF)
+option(GGML_OPENCL_EMBED_KERNELS            "ggml: embed kernels"                             ON)
+option(GGML_OPENCL_USE_ADRENO_KERNELS       "ggml: use optimized kernels for Adreno"          ON)
+
 # extra artifacts
 option(GGML_BUILD_TESTS    "ggml: build tests"    ${GGML_STANDALONE})
 option(GGML_BUILD_EXAMPLES "ggml: build examples" ${GGML_STANDALONE})

ggml/include/ggml-backend.h

@@ -228,6 +228,7 @@ 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/include/ggml-opencl.h

@@ -0,0 +1,26 @@
+#ifndef GGML_OPENCL_H
+#define GGML_OPENCL_H
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+//
+// backend API
+//
+GGML_BACKEND_API ggml_backend_t ggml_backend_opencl_init(void);
+GGML_BACKEND_API bool ggml_backend_is_opencl(ggml_backend_t backend);
+
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_opencl_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_opencl_host_buffer_type(void);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_opencl_reg(void);
+
+#ifdef  __cplusplus
+}
+#endif
+
+#endif // GGML_OPENCL_H

ggml/include/ggml.h

@@ -238,6 +238,8 @@
 #define GGML_EXIT_ABORTED 1
 
 #define GGML_ROPE_TYPE_NEOX   2
+#define GGML_ROPE_TYPE_MROPE  8
+#define GGML_ROPE_TYPE_VISION 24
 
 #define GGUF_MAGIC "GGUF"
 
@@ -1443,6 +1445,22 @@ extern "C" {
             float                 beta_fast,
             float                 beta_slow);
 
+    GGML_API struct ggml_tensor * ggml_rope_multi(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            struct ggml_tensor  * c,
+            int                   n_dims,
+            int                   sections[4],
+            int                   mode,
+            int                   n_ctx_orig,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow);
+
     // in-place, returns view(a)
     GGML_API struct ggml_tensor * ggml_rope_ext_inplace(
             struct ggml_context * ctx,

ggml/src/CMakeLists.txt

@@ -194,11 +194,6 @@ 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
@@ -313,6 +308,7 @@ ggml_add_backend(MUSA)
 ggml_add_backend(RPC)
 ggml_add_backend(SYCL)
 ggml_add_backend(Vulkan)
+ggml_add_backend(OpenCL)
 
 foreach (target ggml-base ggml)
     target_include_directories(${target} PUBLIC    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../include> $<INSTALL_INTERFACE:include>)

ggml/src/ggml-backend-reg.cpp

@@ -46,6 +46,10 @@
 #include "ggml-vulkan.h"
 #endif
 
+#ifdef GGML_USE_OPENCL
+#include "ggml-opencl.h"
+#endif
+
 #ifdef GGML_USE_BLAS
 #include "ggml-blas.h"
 #endif
@@ -146,6 +150,9 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_VULKAN
         register_backend(ggml_backend_vk_reg());
 #endif
+#ifdef GGML_USE_OPENCL
+        register_backend(ggml_backend_opencl_reg());
+#endif
 #ifdef GGML_USE_CANN
         register_backend(ggml_backend_cann_reg());
 #endif
@@ -449,11 +456,21 @@ static std::string backend_filename_suffix() {
 #endif
 }
 
-static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent) {
+static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent, const char * user_search_path) {
     // 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;
@@ -463,7 +480,8 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent)
         if (!fs::exists(search_path)) {
             continue;
         }
-        for (const auto & entry : fs::directory_iterator(search_path)) {
+        fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied);
+        for (const auto & entry : dir_it) {
             if (entry.is_regular_file()) {
                 std::string filename = entry.path().filename().string();
                 std::string ext = entry.path().extension().string();
@@ -509,21 +527,26 @@ 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);
+    ggml_backend_load_best("blas", silent, dir_path);
-    ggml_backend_load_best("cann", silent);
+    ggml_backend_load_best("cann", silent, dir_path);
-    ggml_backend_load_best("cuda", silent);
+    ggml_backend_load_best("cuda", silent, dir_path);
-    ggml_backend_load_best("hip", silent);
+    ggml_backend_load_best("hip", silent, dir_path);
-    ggml_backend_load_best("kompute", silent);
+    ggml_backend_load_best("kompute", silent, dir_path);
-    ggml_backend_load_best("metal", silent);
+    ggml_backend_load_best("metal", silent, dir_path);
-    ggml_backend_load_best("rpc", silent);
+    ggml_backend_load_best("rpc", silent, dir_path);
-    ggml_backend_load_best("sycl", silent);
+    ggml_backend_load_best("sycl", silent, dir_path);
-    ggml_backend_load_best("vulkan", silent);
+    ggml_backend_load_best("vulkan", silent, dir_path);
-    ggml_backend_load_best("musa", silent);
+    ggml_backend_load_best("opencl", silent, dir_path);
-    ggml_backend_load_best("cpu", silent);
+    ggml_backend_load_best("musa", silent, dir_path);
+    ggml_backend_load_best("cpu", silent, dir_path);
 }

ggml/src/ggml-cann/ggml-cann.cpp

@@ -1747,6 +1747,15 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             if (*ext_factor != 0) {
                 return false;
             }
+
+            const int mode = ((const int32_t *) op->op_params)[2];
+            if (mode & GGML_ROPE_TYPE_MROPE) {
+                return false;
+            }
+            if (mode & GGML_ROPE_TYPE_VISION) {
+                return false;
+            }
+
             return true;
         }
         case GGML_OP_UPSCALE: {

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__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+//#if __CUDA_ARCH__ >= GGML_CUDA_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 = aligned_alloc(TENSOR_ALIGNMENT, size);
+    void * data = ggml_aligned_malloc(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

@@ -126,8 +126,7 @@ 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;
@@ -469,7 +468,7 @@ const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type
     for (int i = 0; i < offset; ++i) {                  \
         (x)[i] = vaddq_f32((x)[i], (x)[offset+i]);      \
     }                                                   \
-    (res) = GGML_F32x4_REDUCE_ONE((x)[0]);         \
+    (res) = (ggml_float) GGML_F32x4_REDUCE_ONE((x)[0]); \
 }
 
 #define GGML_F32_VEC        GGML_F32x4
@@ -2396,7 +2395,7 @@ static void ggml_init_arm_arch_features(void) {
     uint32_t hwcap2 = getauxval(AT_HWCAP2);
 
     ggml_arm_arch_features.has_neon = !!(hwcap & HWCAP_ASIMD);
-    ggml_arm_arch_features.has_dotprod = !!(hwcap && HWCAP_ASIMDDP);
+    ggml_arm_arch_features.has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
     ggml_arm_arch_features.has_i8mm = !!(hwcap2 & HWCAP2_I8MM);
     ggml_arm_arch_features.has_sve  = !!(hwcap & HWCAP_SVE);
 
@@ -9134,6 +9133,64 @@ static void ggml_rope_cache_init(
     }
 }
 
+static void ggml_mrope_cache_init(
+     float theta_base_t, float theta_base_h, float theta_base_w, float theta_base_e, int sections[4], bool indep_sects,
+     float freq_scale, const float * freq_factors, float corr_dims[2], int64_t ne0, float ext_factor, float mscale,
+     float * cache, float sin_sign, float theta_scale) {
+    // ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py
+    float theta_t = theta_base_t;
+    float theta_h = theta_base_h;
+    float theta_w = theta_base_w;
+    float theta_e = theta_base_e;  // extra position id for vision encoder
+    int sect_dims = sections[0] + sections[1] + sections[2] + sections[3];
+    int sec_w = sections[1] + sections[0];
+    int sec_e = sections[2] + sec_w;
+    GGML_ASSERT(sect_dims <= ne0);
+
+    for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
+        const float ff = freq_factors ? freq_factors[i0/2] : 1.0f;
+
+        int sector = (i0 / 2) % sect_dims;
+        if (indep_sects) {
+            // compute theta independently for each dim sections
+            // (i.e. reset corresponding theta when `i0` go from one section to another)
+            if (sector == 0) {
+                theta_t = theta_base_t;
+            }
+            else if (sector == sections[0]) {
+                theta_h = theta_base_h;;
+            }
+            else if (sector == sec_w) {
+                theta_w = theta_base_w;
+            }
+            else if (sector == sec_e) {
+                theta_e = theta_base_e;
+            }
+        }
+
+        float theta = theta_t;
+        if (sector >= sections[0] && sector < sec_w) {
+            theta = theta_h;
+        }
+        else if (sector >= sec_w && sector < sec_w + sections[2]) {
+            theta = theta_w;
+        }
+        else if (sector >= sec_w + sections[2]) {
+            theta = theta_e;
+        }
+
+        rope_yarn(
+            theta/ff, freq_scale, corr_dims, i0, ext_factor, mscale, &cache[i0 + 0], &cache[i0 + 1]
+        );
+        cache[i0 + 1] *= sin_sign;
+
+        theta_t *= theta_scale;
+        theta_w *= theta_scale;
+        theta_h *= theta_scale;
+        theta_e *= theta_scale;
+    }
+}
+
 static void ggml_compute_forward_rope_f32(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst,
@@ -9144,6 +9201,7 @@ static void ggml_compute_forward_rope_f32(
     const struct ggml_tensor * src2 = dst->src[2];
 
     float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
+    int sections[4];
 
     //const int n_past     = ((int32_t *) dst->op_params)[0];
     const int n_dims     = ((int32_t *) dst->op_params)[1];
@@ -9157,6 +9215,7 @@ static void ggml_compute_forward_rope_f32(
     memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
     memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
     memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
+    memcpy(&sections,    (int32_t *) dst->op_params + 11, sizeof(int)*4);
 
     GGML_TENSOR_UNARY_OP_LOCALS
 
@@ -9189,6 +9248,16 @@ static void ggml_compute_forward_rope_f32(
     ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+    const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;  // ggml_rope_multi, multimodal rotary position embedding
+    const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
+
+    if (is_mrope) {
+        GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0);
+    }
+
+    if (is_vision) {
+        GGML_ASSERT(n_dims == ne0/2);
+    }
 
     const float * freq_factors = NULL;
     if (src2 != NULL) {
@@ -9204,30 +9273,44 @@ static void ggml_compute_forward_rope_f32(
 
     const int32_t * pos = (const int32_t *) src1->data;
 
-    for (int64_t i3 = 0; i3 < ne3; i3++) {
+    for (int64_t i3 = 0; i3 < ne3; i3++) { // batch
-        for (int64_t i2 = 0; i2 < ne2; i2++) {
+        for (int64_t i2 = 0; i2 < ne2; i2++) { // seq-len
-            const int64_t p = pos[i2];
 
             float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith;
+            if (!is_mrope) {
+                const int64_t p = pos[i2];
                 ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+            }
+            else {
+                const int64_t p_t = pos[i2];
+                const int64_t p_h = pos[i2 + ne2];
+                const int64_t p_w = pos[i2 + ne2 * 2];
+                const int64_t p_e = pos[i2 + ne2 * 3];
+                ggml_mrope_cache_init(
+                    p_t, p_h, p_w, p_e, sections, is_vision,
+                    freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+            }
 
-            for (int64_t i1 = 0; i1 < ne1; i1++) {
+            for (int64_t i1 = 0; i1 < ne1; i1++) { // attn-heads
                 if (ir++ < ir0) continue;
                 if (ir   > ir1) break;
 
-                if (!is_neox) {
+                if (is_neox || is_mrope) {
+                    if (is_vision){
                         for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
+                            const int64_t ic = i0/2;
+
                             const float cos_theta = cache[i0 + 0];
                             const float sin_theta = cache[i0 + 1];
 
-                        const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+                            const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-                              float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+                            float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
 
                             const float x0 = src[0];
-                        const float x1 = src[1];
+                            const float x1 = src[n_dims];
 
                             dst_data[0]      = x0*cos_theta - x1*sin_theta;
-                        dst_data[1] = x0*sin_theta + x1*cos_theta;
+                            dst_data[n_dims] = x0*sin_theta + x1*cos_theta;
                         }
                     } else {
                         for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
@@ -9246,7 +9329,40 @@ static void ggml_compute_forward_rope_f32(
                             dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta;
                         }
                     }
+                } else {
+                    for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
+                        const float cos_theta = cache[i0 + 0];
+                        const float sin_theta = cache[i0 + 1];
 
+                        const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+                              float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+                        const float x0 = src[0];
+                        const float x1 = src[1];
+
+                        dst_data[0] = x0*cos_theta - x1*sin_theta;
+                        dst_data[1] = x0*sin_theta + x1*cos_theta;
+                    }
+                }
+
+                if (is_vision) {
+                    for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
+                        const int64_t ic = i0/2;
+
+                        const float cos_theta = cache[i0 + 0];
+                        const float sin_theta = cache[i0 + 1];
+
+                        const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+                        float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+                        const float x0 = src[0];
+                        const float x1 = src[n_dims];
+
+                        dst_data[0]      = x0*cos_theta - x1*sin_theta;
+                        dst_data[n_dims] = x0*sin_theta + x1*cos_theta;
+                    }
+                } else {
+                    // fill the remain channels with data from src tensor
                     for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
                         const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
                         float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
@@ -9257,6 +9373,7 @@ static void ggml_compute_forward_rope_f32(
                 }
             }
         }
+    }
 }
 
 // TODO: deduplicate f16/f32 code
@@ -9270,6 +9387,7 @@ static void ggml_compute_forward_rope_f16(
     const struct ggml_tensor * src2 = dst->src[2];
 
     float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
+    int sections[4];
 
     //const int n_past     = ((int32_t *) dst->op_params)[0];
     const int n_dims     = ((int32_t *) dst->op_params)[1];
@@ -9282,6 +9400,8 @@ static void ggml_compute_forward_rope_f16(
     memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
     memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
     memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
+    memcpy(&sections,    (int32_t *) dst->op_params + 11, sizeof(int)*4);
+
 
     GGML_TENSOR_UNARY_OP_LOCALS
 
@@ -9314,6 +9434,16 @@ static void ggml_compute_forward_rope_f16(
     ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+    const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
+    const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
+
+    if (is_mrope) {
+        GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0);
+    }
+
+    if (is_vision) {
+        GGML_ASSERT(n_dims == ne0/2);
+    }
 
     const float * freq_factors = NULL;
     if (src2 != NULL) {
@@ -9331,28 +9461,42 @@ static void ggml_compute_forward_rope_f16(
 
     for (int64_t i3 = 0; i3 < ne3; i3++) {
         for (int64_t i2 = 0; i2 < ne2; i2++) {
-            const int64_t p = pos[i2];
 
             float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith;
+            if (!is_mrope) {
+                const int64_t p = pos[i2];
                 ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+            }
+            else {
+                const int64_t p_t = pos[i2];
+                const int64_t p_h = pos[i2 + ne2];
+                const int64_t p_w = pos[i2 + ne2 * 2];
+                const int64_t p_e = pos[i2 + ne2 * 3];
+                ggml_mrope_cache_init(
+                    p_t, p_h, p_w, p_e, sections, is_vision,
+                    freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+            }
 
             for (int64_t i1 = 0; i1 < ne1; i1++) {
                 if (ir++ < ir0) continue;
                 if (ir   > ir1) break;
 
-                if (!is_neox) {
+                if (is_neox || is_mrope) {
+                    if (is_vision) {
                         for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
+                            const int64_t ic = i0/2;
+
                             const float cos_theta = cache[i0 + 0];
                             const float sin_theta = cache[i0 + 1];
 
-                        const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+                            const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-                              ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+                            ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
 
                             const float x0 = GGML_FP16_TO_FP32(src[0]);
-                        const float x1 = GGML_FP16_TO_FP32(src[1]);
+                            const float x1 = GGML_FP16_TO_FP32(src[n_dims]);
 
                             dst_data[0]      = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
-                        dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
+                            dst_data[n_dims] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
                         }
                     } else {
                         for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
@@ -9371,7 +9515,39 @@ static void ggml_compute_forward_rope_f16(
                             dst_data[n_dims/2] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
                         }
                     }
+                } else {
+                    for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
+                        const float cos_theta = cache[i0 + 0];
+                        const float sin_theta = cache[i0 + 1];
 
+                        const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+                              ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+                        const float x0 = GGML_FP16_TO_FP32(src[0]);
+                        const float x1 = GGML_FP16_TO_FP32(src[1]);
+
+                        dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
+                        dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
+                    }
+                }
+
+                if (is_vision) {
+                    for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
+                        const int64_t ic = i0/2;
+
+                        const float cos_theta = cache[i0 + 0];
+                        const float sin_theta = cache[i0 + 1];
+
+                        const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+                        ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+                        const float x0 = GGML_FP16_TO_FP32(src[0]);
+                        const float x1 = GGML_FP16_TO_FP32(src[n_dims]);
+
+                        dst_data[0]      = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
+                        dst_data[n_dims] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
+                    }
+                } else {
                     for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
                         const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
                         ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
@@ -9382,6 +9558,7 @@ static void ggml_compute_forward_rope_f16(
                 }
             }
         }
+    }
 }
 
 static void ggml_compute_forward_rope(
@@ -12945,7 +13122,7 @@ static thread_ret_t ggml_graph_compute_secondary_thread(void* data);
 #include "windows.h"
 
 // TODO: support > 64 CPUs
-bool ggml_thread_apply_affinity(bool * mask) {
+static 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 CC_PASCAL     600
+#define GGML_CUDA_CC_PASCAL     600
-#define MIN_CC_DP4A   610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
+#define GGML_CUDA_CC_DP4A       610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
-#define CC_VOLTA      700
+#define GGML_CUDA_CC_VOLTA      700
-#define CC_TURING     750
+#define GGML_CUDA_CC_TURING     750
-#define CC_AMPERE     800
+#define GGML_CUDA_CC_AMPERE     800
-#define CC_OFFSET_AMD 1000000
+#define GGML_CUDA_CC_OFFSET_AMD 1000000
 
 // GCN/CNDA, wave size is 64
-#define CC_GCN4       (CC_OFFSET_AMD + 803)  // Tonga, Fiji, Polaris, minimum for fast fp16
+#define GGML_CUDA_CC_GCN4       (GGML_CUDA_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 GGML_CUDA_CC_VEGA       (GGML_CUDA_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 GGML_CUDA_CC_VEGA20     (GGML_CUDA_CC_OFFSET_AMD + 906)  // MI50/Radeon VII, minimum for dp4a
-#define CC_CDNA       (CC_OFFSET_AMD + 908)  // MI100, minimum for MFMA, acc registers
+#define GGML_CUDA_CC_CDNA       (GGML_CUDA_CC_OFFSET_AMD + 908)  // MI100, minimum for MFMA, acc registers
-#define CC_CDNA2      (CC_OFFSET_AMD + 910)  // MI210, minimum acc register renameing
+#define GGML_CUDA_CC_CDNA2      (GGML_CUDA_CC_OFFSET_AMD + 910)  // MI210, minimum acc register renameing
-#define CC_CDNA3      (CC_OFFSET_AMD + 942)  // MI300
+#define GGML_CUDA_CC_CDNA3      (GGML_CUDA_CC_OFFSET_AMD + 942)  // MI300
 
 // RNDA removes MFMA, dp4a, xnack, acc registers, wave size is 32
-#define CC_RDNA1      (CC_OFFSET_AMD + 1010) // RX 5000
+#define GGML_CUDA_CC_RDNA1      (GGML_CUDA_CC_OFFSET_AMD + 1010) // RX 5000
-#define CC_RDNA2      (CC_OFFSET_AMD + 1030) // RX 6000, minimum for dp4a
+#define GGML_CUDA_CC_RDNA2      (GGML_CUDA_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_RDNA3      (GGML_CUDA_CC_OFFSET_AMD + 1100) // RX 7000, minimum for WMMA
 
-#define CC_QY1        210
+#define GGML_CUDA_CC_QY1        210
-#define CC_QY2        220
+#define GGML_CUDA_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__ >= CC_PASCAL
+#if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 #define FP16_AVAILABLE
-#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
+#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= GGML_CUDA_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__ >= CC_VOLTA
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
 #define FP16_MMA_AVAILABLE
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
+#endif // !(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_TURING
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
 #define INT8_MMA_AVAILABLE
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
 
-#if !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= CC_QY1)
+#if !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1)
 #define FLASH_ATTN_AVAILABLE
-#endif // !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= CC_QY1)
+#endif // !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1)
 
 static constexpr bool fast_fp16_available(const int cc) {
-    return cc >= CC_PASCAL && cc != 610;
+    return cc >= GGML_CUDA_CC_PASCAL && cc != 610;
 }
 
 static constexpr bool fp16_mma_available(const int cc) {
-    return cc < CC_OFFSET_AMD && cc >= CC_VOLTA;
+    return cc < GGML_CUDA_CC_OFFSET_AMD && cc >= GGML_CUDA_CC_VOLTA;
 }
 
 static constexpr bool int8_mma_available(const int cc) {
-    return cc < CC_OFFSET_AMD && cc >= CC_TURING;
+    return cc < GGML_CUDA_CC_OFFSET_AMD && cc >= GGML_CUDA_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__ >= CC_AMPERE
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_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__ >= CC_AMPERE
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_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__ >= CC_PASCAL
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_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__ >= CC_PASCAL
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_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__ >= MIN_CC_DP4A
+#if __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A
     return __dp4a(a, b, c);
-#else // __CUDA_ARCH__ >= MIN_CC_DP4A
+#else // __CUDA_ARCH__ >= GGML_CUDA_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__ >= MIN_CC_DP4A
+#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A
 
 #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 }

ggml/src/ggml-cuda/concat.cu

@@ -94,7 +94,9 @@ static void concat_f32_cuda(const float * x, const float * y, float * dst, int n
 }
 
 // non-contiguous kernel (slow)
-static __global__ void concat_f32_non_cont(
+template <int dim>
+static __global__ void __launch_bounds__(CUDA_CONCAT_BLOCK_SIZE)
+    concat_f32_non_cont(
         const char * src0,
         const char * src1,
               char * dst,
@@ -121,22 +123,28 @@ static __global__ void concat_f32_non_cont(
           uint64_t   nb0,
           uint64_t   nb1,
           uint64_t   nb2,
-          uint64_t   nb3,
+          uint64_t   nb3){
-          int32_t   dim) {
+    static_assert(dim >= 0 && dim <= 3);
+
     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 (int i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
+    for (int64_t 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 {
-            x = (const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10);
+            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);
+            }
         }
 
         float * y = (float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
@@ -182,15 +190,32 @@ 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]);
-        concat_f32_non_cont<<<grid_dim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(
+        auto launch_kernel = [&](auto dim) {
-                (const char *)src0->data,
+            concat_f32_non_cont<dim><<<grid_dim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(
-                (const char *)src1->data,
+                (const char *) src0->data, (const char *) src1->data, (char *) dst->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], dim);
+                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;
+        }
     }
 }

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__ >= CC_PASCAL
+#if __CUDA_ARCH__ >= GGML_CUDA_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__ >= CC_PASCAL
+#endif // __CUDA_ARCH__ >= GGML_CUDA_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 >= CC_PASCAL) {
+            if (ggml_cuda_info().devices[ggml_cuda_get_device()].cc >= GGML_CUDA_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 >= CC_OFFSET_AMD) {
+    if (cc >= GGML_CUDA_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 + CC_OFFSET_AMD;
+        info.devices[id].cc = 100*prop.major + 10*prop.minor + GGML_CUDA_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 >= 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 >= 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) {
         // 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 == CC_CDNA) {
+        if (ggml_cuda_info().devices[ctx.device].cc == GGML_CUDA_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 == CC_CDNA) {
+    if (ggml_cuda_info().devices[ctx.device].cc == GGML_CUDA_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 < CC_AMPERE) {
+        if (ggml_cuda_info().devices[cuda_ctx->device].cc < GGML_CUDA_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 >= CC_VOLTA && cc < CC_OFFSET_AMD && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16;
+            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;
         }
         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__ >= CC_AMPERE
+#if __CUDA_ARCH__ >= GGML_CUDA_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__ >= CC_AMPERE
+#endif // __CUDA_ARCH__ >= GGML_CUDA_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__ >= CC_AMPERE
+#if __CUDA_ARCH__ >= GGML_CUDA_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__ >= CC_AMPERE
+#endif // __CUDA_ARCH__ >= GGML_CUDA_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 >= CC_VOLTA && compute_capability < CC_OFFSET_AMD && src1_ncols == ne11;
+    const bool use_stream_k = compute_capability >= GGML_CUDA_CC_VOLTA && compute_capability < GGML_CUDA_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 < MIN_CC_DP4A) {
+    if (cc < GGML_CUDA_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 < CC_OFFSET_AMD) {
+    if (cc < GGML_CUDA_CC_OFFSET_AMD) {
-        return cc < CC_VOLTA || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+        return cc < GGML_CUDA_CC_VOLTA || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
     }
 
-    return (cc < CC_RDNA3 && cc != CC_CDNA && cc != CC_VEGA20) || 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;
 }

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 >= CC_VOLTA && cc < CC_OFFSET_AMD ? 128                     : 64;
+        cc >= GGML_CUDA_CC_VOLTA && cc < GGML_CUDA_CC_OFFSET_AMD ? 128                     : 64;
 #else
-        cc >= CC_VOLTA && cc < CC_OFFSET_AMD ? MMQ_DP4A_MAX_BATCH_SIZE : 64;
+        cc >= GGML_CUDA_CC_VOLTA && cc < GGML_CUDA_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__ >= CC_VOLTA
+#if __CUDA_ARCH__ >= GGML_CUDA_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__ >= CC_VOLTA
+#else // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
 
     return 64;
-#endif // __CUDA_ARCH__ >= CC_VOLTA
+#endif // __CUDA_ARCH__ >= GGML_CUDA_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 >= CC_OFFSET_AMD ? (cc == CC_RDNA1 ? 64 : 128) : (cc >= CC_VOLTA ? 128 : 64);
+    return cc >= GGML_CUDA_CC_OFFSET_AMD ? (cc == GGML_CUDA_CC_RDNA1 ? 64 : 128) : (cc >= GGML_CUDA_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__ >= CC_VOLTA
+#if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
     return 128;
 #else
     return 64;
-#endif // __CUDA_ARCH__ >= CC_VOLTA
+#endif // __CUDA_ARCH__ >= GGML_CUDA_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__ >= CC_VOLTA
+#if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
     __launch_bounds__(WARP_SIZE*nwarps, 1)
 #else
     __launch_bounds__(WARP_SIZE*nwarps, 2)
-#endif // __CUDA_ARCH__ >= CC_VOLTA
+#endif // __CUDA_ARCH__ >= GGML_CUDA_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__ < CC_VOLTA
+#if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < GGML_CUDA_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__ < CC_VOLTA
+#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < GGML_CUDA_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 >= CC_VOLTA && cc < CC_OFFSET_AMD;
+    const bool use_stream_k = cc >= GGML_CUDA_CC_VOLTA && cc < GGML_CUDA_CC_OFFSET_AMD;
 
     int mmq_x_best  = 0;
     int nparts_best = INT_MAX;

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 < CC_CDNA || ggml_cuda_info().devices[id].cc == CC_RDNA1) { // NVIDIA and AMD older than RDNA2 but not CDNA
+    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
         switch(ncols_y) {
             case 1:
                 nwarps = 4;

ggml/src/ggml-cuda/rope.cu

@@ -4,6 +4,11 @@ struct rope_corr_dims {
     float v[2];
 };
 
+
+struct mrope_sections {
+    int v[4];
+};
+
 static __device__ float rope_yarn_ramp(const float low, const float high, const int i0) {
     const float y = (i0 / 2 - low) / max(0.001f, high - low);
     return 1.0f - min(1.0f, max(0.0f, y));
@@ -108,6 +113,105 @@ static __global__ void rope_neox(
     dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta;
 }
 
+template<typename T, bool has_ff>
+static __global__ void rope_multi(
+    const T * x, T * dst, int ne0, int ne2, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, const float * freq_factors, mrope_sections sections) {
+    const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
+
+    if (i0 >= ne0) {
+        return;
+    }
+
+    const int row = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i0 >= n_dims) {
+        const int i = row*ne0 + i0;
+
+        dst[i + 0] = x[i + 0];
+        dst[i + 1] = x[i + 1];
+
+        return;
+    }
+
+    const int i  = row*ne0 + i0/2;
+    const int i2 = row/p_delta_rows;
+
+    int sect_dims = sections.v[0] + sections.v[1] + sections.v[2] + sections.v[3];
+    int sec_w = sections.v[1] + sections.v[0];
+    int sector = (i0 / 2) % sect_dims;
+
+    float theta_base = 0.0;
+    if (sector < sections.v[0]) {
+        theta_base = pos[i2]*powf(theta_scale, i0/2.0f);
+    }
+    else if (sector >= sections.v[0] && sector < sec_w) {
+        theta_base = pos[i2 + ne2 * 1]*powf(theta_scale, i0/2.0f);
+    }
+    else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
+        theta_base = pos[i2 + ne2 * 2]*powf(theta_scale, i0/2.0f);
+    }
+    else if (sector >= sec_w + sections.v[2]) {
+        theta_base = pos[i2 + ne2 * 3]*powf(theta_scale, i0/2.0f);
+    }
+
+    const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
+
+    float cos_theta;
+    float sin_theta;
+
+    rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
+
+    const float x0 = x[i + 0];
+    const float x1 = x[i + n_dims/2];
+
+    dst[i + 0]        = x0*cos_theta - x1*sin_theta;
+    dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta;
+}
+
+template<typename T, bool has_ff>
+static __global__ void rope_vision(
+    const T * x, T * dst, int ne0, int ne2, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, const float * freq_factors, mrope_sections sections) {
+    const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
+
+    if (i0 >= ne0) {
+        return;
+    }
+
+    const int row = blockDim.x*blockIdx.x + threadIdx.x;
+
+    const int i  = row*ne0 + i0/2;
+    const int i2 = row/p_delta_rows; // i2-th tokens
+
+    int sect_dims = sections.v[0] + sections.v[1];
+    int sec_w = sections.v[1] + sections.v[0];
+    int sector = (i0 / 2) % sect_dims;
+
+    float theta_base = 0.0;
+    if (sector < sections.v[0]) {
+        const int p = sector;
+        theta_base = pos[i2]*powf(theta_scale, p);
+    }
+    else if (sector >= sections.v[0] && sector < sec_w) {
+        const int p = sector - sections.v[0];
+        theta_base = pos[i2 + ne2]*powf(theta_scale, p);
+    }
+
+    const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
+
+    float cos_theta;
+    float sin_theta;
+
+    rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
+
+    const float x0 = x[i + 0];
+    const float x1 = x[i + n_dims];
+
+    dst[i + 0]      = x0*cos_theta - x1*sin_theta;
+    dst[i + n_dims] = x0*sin_theta + x1*cos_theta;
+}
+
 template<typename T>
 static void rope_norm_cuda(
     const T * x, T * dst, int ne0, int n_dims, int nr, const int32_t * pos, float freq_scale, int p_delta_rows,
@@ -156,6 +260,56 @@ static void rope_neox_cuda(
     }
 }
 
+template<typename T>
+static void rope_multi_cuda(
+    const T * x, T * dst, int ne0, int ne2, int n_dims, int nr, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, mrope_sections sections, cudaStream_t stream) {
+    GGML_ASSERT(ne0 % 2 == 0);
+    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
+    const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
+    const dim3 block_nums(nr, n_blocks_x, 1);
+
+    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+
+    if (freq_factors == nullptr) {
+        rope_multi<T, false><<<block_nums, block_dims, 0, stream>>>(
+                x, dst, ne0, ne2, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
+                theta_scale, freq_factors, sections
+                );
+    } else {
+        rope_multi<T, true><<<block_nums, block_dims, 0, stream>>>(
+                x, dst, ne0, ne2, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
+                theta_scale, freq_factors, sections
+                );
+    }
+}
+
+template<typename T>
+static void rope_vision_cuda(
+    const T * x, T * dst, int ne0, int ne2, int n_dims, int nr, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, mrope_sections sections, cudaStream_t stream) {
+    GGML_ASSERT(ne0 % 2 == 0);
+    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
+    const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
+    const dim3 block_nums(nr, n_blocks_x, 1);
+    // break down (head_dim, heads, seq) into (CUDA_ROPE_BLOCK_SIZE, x, heads * seq)
+    // where x ~= ceil(head_dim / CUDA_ROPE_BLOCK_SIZE);
+
+    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+
+    if (freq_factors == nullptr) {
+        rope_vision<T, false><<<block_nums, block_dims, 0, stream>>>(
+                x, dst, ne0, ne2, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
+                theta_scale, freq_factors, sections
+                );
+    } else {
+        rope_vision<T, true><<<block_nums, block_dims, 0, stream>>>(
+                x, dst, ne0, ne2, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
+                theta_scale, freq_factors, sections
+                );
+    }
+}
+
 static void rope_norm_cuda_f16(
     const half * x, half * dst, int ne0, int n_dims, int nr, const int32_t * pos, float freq_scale, int p_delta_rows,
     float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream) {
@@ -185,6 +339,38 @@ static void rope_neox_cuda_f32(
     rope_neox_cuda<float>(x, dst, ne0, n_dims, nr, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, stream);
 }
 
+static void rope_multi_cuda_f16(
+    const half * x, half * dst, int ne0, int ne2, int n_dims, int nr, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, mrope_sections sections, cudaStream_t stream
+) {
+
+    rope_multi_cuda<half>(x, dst, ne0, ne2, n_dims, nr, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, stream);
+}
+
+static void rope_multi_cuda_f32(
+    const float * x, float * dst, int ne0, int ne2, int n_dims, int nr, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, mrope_sections sections, cudaStream_t stream
+) {
+
+    rope_multi_cuda<float>(x, dst, ne0, ne2, n_dims, nr, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, stream);
+}
+
+static void rope_vision_cuda_f16(
+    const half * x, half * dst, int ne0, int ne2, int n_dims, int nr, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, mrope_sections sections, cudaStream_t stream
+) {
+
+    rope_vision_cuda<half>(x, dst, ne0, ne2, n_dims, nr, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, stream);
+}
+
+static void rope_vision_cuda_f32(
+    const float * x, float * dst, int ne0, int ne2, int n_dims, int nr, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, mrope_sections sections, cudaStream_t stream
+) {
+
+    rope_vision_cuda<float>(x, dst, ne0, ne2, n_dims, nr, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, stream);
+}
+
 void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
@@ -201,8 +387,9 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     GGML_ASSERT( dst->type == GGML_TYPE_F32 ||  dst->type == GGML_TYPE_F16);
     GGML_ASSERT(src0->type == dst->type);
 
-    const int64_t ne00 = src0->ne[0];
+    const int64_t ne00 = src0->ne[0]; // head dims
-    const int64_t ne01 = src0->ne[1];
+    const int64_t ne01 = src0->ne[1]; // num heads
+    const int64_t ne02 = src0->ne[2]; // num heads
     const int64_t nr = ggml_nrows(src0);
 
     //const int n_past     = ((int32_t *) dst->op_params)[0];
@@ -210,6 +397,7 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const int mode       = ((int32_t *) dst->op_params)[2];
     //const int n_ctx      = ((int32_t *) dst->op_params)[3];
     const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
+    mrope_sections sections;
 
     // RoPE alteration for extended context
     float freq_base;
@@ -225,8 +413,19 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
     memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
     memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
+    memcpy(&sections.v,  (int32_t *) dst->op_params + 11, sizeof(int)*4);
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+    const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
+    const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
+
+    if (is_mrope) {
+        GGML_ASSERT(sections.v[0] > 0 || sections.v[1] > 0 || sections.v[2] > 0);
+    }
+
+    if (is_vision) {
+        GGML_ASSERT(n_dims == ne00/2);
+    }
 
     const int32_t * pos = (const int32_t *) src1_d;
 
@@ -253,6 +452,34 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
         } else {
             GGML_ABORT("fatal error");
         }
+    } else if (is_mrope && !is_vision) {
+        if (src0->type == GGML_TYPE_F32) {
+            rope_multi_cuda_f32(
+                (const float *)src0_d, (float *)dst_d, ne00, ne02, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, freq_factors, sections, stream
+            );
+        } else if (src0->type == GGML_TYPE_F16) {
+            rope_multi_cuda_f16(
+                (const half *)src0_d, (half *)dst_d, ne00, ne02, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, freq_factors, sections, stream
+            );
+        } else {
+            GGML_ABORT("fatal error");
+        }
+    } else if (is_vision) {
+        if (src0->type == GGML_TYPE_F32) {
+            rope_vision_cuda_f32(
+                (const float *)src0_d, (float *)dst_d, ne00, ne02, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, freq_factors, sections, stream
+            );
+        } else if (src0->type == GGML_TYPE_F16) {
+            rope_vision_cuda_f16(
+                (const half *)src0_d, (half *)dst_d, ne00, ne02, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, freq_factors, sections, stream
+            );
+        } else {
+            GGML_ABORT("fatal error");
+        }
     } else {
         if (src0->type == GGML_TYPE_F32) {
             rope_norm_cuda_f32(

ggml/src/ggml-cuda/sum.cu

@@ -3,8 +3,6 @@
 #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)
-void ggml_log_internal        (enum ggml_log_level level, const char * format, ...);
+GGML_API 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);
+GGML_API 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
 
-void * ggml_aligned_malloc(size_t size);
+GGML_API void * ggml_aligned_malloc(size_t size);
-void ggml_aligned_free(void * ptr, size_t size);
+GGML_API void ggml_aligned_free(void * ptr, size_t size);
 
 // FP16 to FP32 conversion
 

ggml/src/ggml-kompute/ggml-kompute.cpp

@@ -1419,8 +1419,18 @@ static bool ggml_backend_kompute_device_supports_op(ggml_backend_dev_t dev, cons
         case GGML_OP_SOFT_MAX:
         case GGML_OP_RMS_NORM:
         case GGML_OP_NORM:
-        case GGML_OP_ROPE:
             return true;
+        case GGML_OP_ROPE:
+            {
+                const int mode = ((const int32_t *) op->op_params)[2];
+                if (mode & GGML_ROPE_TYPE_MROPE) {
+                    return false;
+                }
+                if (mode & GGML_ROPE_TYPE_VISION) {
+                    return false;
+                }
+                return true;
+            }
         case GGML_OP_DUP:
         case GGML_OP_CPY:
         case GGML_OP_CONT:

ggml/src/ggml-metal/ggml-metal.m

@@ -1125,8 +1125,18 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
             return has_simdgroup_reduction && (op->ne[0] % 4 == 0);
         case GGML_OP_ARGMAX:
         case GGML_OP_NORM:
-        case GGML_OP_ROPE:
             return true;
+        case GGML_OP_ROPE:
+            {
+                const int mode = ((const int32_t *) op->op_params)[2];
+                if (mode & GGML_ROPE_TYPE_MROPE) {
+                    return false;
+                }
+                if (mode & GGML_ROPE_TYPE_VISION) {
+                    return false;
+                }
+                return true;
+            }
         case GGML_OP_IM2COL:
             return op->src[0]->type == GGML_TYPE_F16;
         case GGML_OP_POOL_1D:
@@ -3026,7 +3036,9 @@ static void ggml_metal_encode_node(
             } break;
         case GGML_OP_ROPE:
             {
-                GGML_ASSERT(ne10 == ne02);
+                // make sure we have one or more position id(ne10) per token(ne02)
+                GGML_ASSERT(ne10 % ne02 == 0);
+                GGML_ASSERT(ne10 >= ne02);
 
                 const int nth = MIN(1024, ne00);
 

ggml/src/ggml-opencl/CMakeLists.txt

@@ -0,0 +1,147 @@
+find_package(OpenCL REQUIRED)
+find_package(Python3 REQUIRED)
+
+set(TARGET_NAME ggml-opencl)
+
+ggml_add_backend_library(${TARGET_NAME}
+                         ggml-opencl.cpp
+                         ../../include/ggml-opencl.h)
+target_link_libraries(${TARGET_NAME} PRIVATE ${OpenCL_LIBRARIES})
+target_include_directories(${TARGET_NAME} PRIVATE ${OpenCL_INCLUDE_DIRS})
+
+if (GGML_OPENCL_PROFILING)
+    message(STATUS "OpenCL profiling enabled (increases CPU overhead)")
+    add_compile_definitions(GGML_OPENCL_PROFILING)
+endif ()
+
+add_compile_definitions(GGML_OPENCL_SOA_Q)
+
+if (GGML_OPENCL_USE_ADRENO_KERNELS)
+    message(STATUS "OpenCL will use matmul kernels optimized for Adreno")
+    add_compile_definitions(GGML_OPENCL_USE_ADRENO_KERNELS)
+endif ()
+
+if (GGML_OPENCL_EMBED_KERNELS)
+    add_compile_definitions(GGML_OPENCL_EMBED_KERNELS)
+
+    set(OPENCL_CL_SOURCE_EMBED         "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl.cl.h")
+    set(OPENCL_MM_CL_SOURCE_EMBED      "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_mm.cl.h")
+    set(OPENCL_CVT_CL_SOURCE_EMBED     "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_cvt.cl.h")
+
+    set(OPENCL_GEMV_NOSHUFFLE_SOURCE_EMBED             "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_gemv_noshuffle.cl.h")
+    set(OPENCL_GEMV_NOSHUFFLE_GENERAL_SOURCE_EMBED     "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_gemv_noshuffle_general.cl.h")
+    set(OPENCL_MUL_MAT_Ab_Bi_8x4_SOURCE_EMBED          "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_mul_mat_Ab_Bi_8x4.cl.h")
+    set(OPENCL_TRANSPOSE_16_SOURCE_EMBED               "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_transpose_16.cl.h")
+    set(OPENCL_TRANSPOSE_32_SOURCE_EMBED               "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_transpose_32.cl.h")
+    set(OPENCL_TRANSPOSE_32_16_SOURCE_EMBED            "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_transpose_32_16.cl.h")
+
+    set(EMBED_KERNEL_SCRIPT             "${CMAKE_CURRENT_SOURCE_DIR}/kernels/embed_kernel.py")
+    file(MAKE_DIRECTORY                 "${CMAKE_BINARY_DIR}/autogenerated")
+
+    include_directories("${CMAKE_BINARY_DIR}/autogenerated")
+
+    # Python must be accessible from command line
+    add_custom_command(
+        OUTPUT ${OPENCL_CL_SOURCE_EMBED}
+        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
+            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl.cl
+            ${OPENCL_CL_SOURCE_EMBED}
+        DEPENDS kernels/ggml-opencl.cl ${EMBED_KERNEL_SCRIPT}
+        COMMENT "Generate ggml-opencl.cl.h"
+    )
+
+    add_custom_command(
+        OUTPUT ${OPENCL_MM_CL_SOURCE_EMBED}
+        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
+            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_mm.cl
+            ${OPENCL_MM_CL_SOURCE_EMBED}
+        DEPENDS kernels/ggml-opencl_mm.cl ${EMBED_KERNEL_SCRIPT}
+        COMMENT "Generate ggml-opencl_mm.cl.h"
+    )
+
+    add_custom_command(
+        OUTPUT ${OPENCL_CVT_CL_SOURCE_EMBED}
+        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
+            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_cvt.cl
+            ${OPENCL_CVT_CL_SOURCE_EMBED}
+        DEPENDS kernels/ggml-opencl_cvt.cl ${EMBED_KERNEL_SCRIPT}
+        COMMENT "Generate ggml-opencl_cvt.cl.h"
+    )
+
+    add_custom_command(
+        OUTPUT ${OPENCL_GEMV_NOSHUFFLE_SOURCE_EMBED}
+        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
+            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_gemv_noshuffle.cl
+            ${OPENCL_GEMV_NOSHUFFLE_SOURCE_EMBED}
+        DEPENDS kernels/ggml-opencl_gemv_noshuffle.cl ${EMBED_KERNEL_SCRIPT}
+        COMMENT "Generate ggml-opencl_gemv_noshuffle.cl.h"
+    )
+
+    add_custom_command(
+        OUTPUT ${OPENCL_GEMV_NOSHUFFLE_GENERAL_SOURCE_EMBED}
+        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
+            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_gemv_noshuffle_general.cl
+            ${OPENCL_GEMV_NOSHUFFLE_GENERAL_SOURCE_EMBED}
+        DEPENDS kernels/ggml-opencl_gemv_noshuffle_general.cl ${EMBED_KERNEL_SCRIPT}
+        COMMENT "Generate ggml-opencl_gemv_noshuffle_general.cl.h"
+    )
+
+    add_custom_command(
+        OUTPUT ${OPENCL_MUL_MAT_Ab_Bi_8x4_SOURCE_EMBED}
+        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
+            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_mul_mat_Ab_Bi_8x4.cl
+            ${OPENCL_MUL_MAT_Ab_Bi_8x4_SOURCE_EMBED}
+        DEPENDS kernels/ggml-opencl_mul_mat_Ab_Bi_8x4.cl ${EMBED_KERNEL_SCRIPT}
+        COMMENT "Generate ggml-opencl_mul_mat_Ab_Bi_8x4.cl.cl.h"
+    )
+
+    add_custom_command(
+        OUTPUT ${OPENCL_TRANSPOSE_16_SOURCE_EMBED}
+        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
+            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_transpose_16.cl
+            ${OPENCL_TRANSPOSE_16_SOURCE_EMBED}
+        DEPENDS kernels/ggml-opencl_transpose_16.cl ${EMBED_KERNEL_SCRIPT}
+        COMMENT "Generate ggml-opencl_transpose_16.cl.h"
+    )
+
+    add_custom_command(
+        OUTPUT ${OPENCL_TRANSPOSE_32_SOURCE_EMBED}
+        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
+            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_transpose_32.cl
+            ${OPENCL_TRANSPOSE_32_SOURCE_EMBED}
+        DEPENDS kernels/ggml-opencl_transpose_32.cl ${EMBED_KERNEL_SCRIPT}
+        COMMENT "Generate ggml-opencl_transpose_32.cl.h"
+    )
+
+    add_custom_command(
+        OUTPUT ${OPENCL_TRANSPOSE_32_16_SOURCE_EMBED}
+        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
+            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_transpose_32_16.cl
+            ${OPENCL_TRANSPOSE_32_16_SOURCE_EMBED}
+        DEPENDS kernels/ggml-opencl_transpose_32_16.cl ${EMBED_KERNEL_SCRIPT}
+        COMMENT "Generate ggml-opencl_transpose_32_16.cl.h"
+    )
+
+    target_sources(${TARGET_NAME} PRIVATE
+                   ${OPENCL_CL_SOURCE_EMBED}
+                   ${OPENCL_MM_CL_SOURCE_EMBED}
+                   ${OPENCL_CVT_CL_SOURCE_EMBED}
+                   ${OPENCL_GEMV_NOSHUFFLE_SOURCE_EMBED}
+                   ${OPENCL_GEMV_NOSHUFFLE_GENERAL_SOURCE_EMBED}
+                   ${OPENCL_MUL_MAT_Ab_Bi_8x4_SOURCE_EMBED}
+                   ${OPENCL_TRANSPOSE_16_SOURCE_EMBED}
+                   ${OPENCL_TRANSPOSE_32_SOURCE_EMBED}
+                   ${OPENCL_TRANSPOSE_32_16_SOURCE_EMBED})
+else ()
+    # copy ggml-opencl.cl to bin directory
+    configure_file(kernels/ggml-opencl.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl.cl COPYONLY)
+    configure_file(kernels/ggml-opencl_mm.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_mm.cl COPYONLY)
+    configure_file(kernels/ggml-opencl_cvt.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_cvt.cl COPYONLY)
+
+    configure_file(kernels/ggml-opencl_gemv_noshuffle.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_gemv_noshuffle.cl COPYONLY)
+    configure_file(kernels/ggml-opencl_gemv_noshuffle_general.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_gemv_noshuffle_general.cl COPYONLY)
+    configure_file(kernels/ggml-opencl_mul_mat_Ab_Bi_8x4.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_mul_mat_Ab_Bi_8x4.cl COPYONLY)
+    configure_file(kernels/ggml-opencl_transpose_16.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_transpose_16.cl COPYONLY)
+    configure_file(kernels/ggml-opencl_transpose_32.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_transpose_32.cl COPYONLY)
+    configure_file(kernels/ggml-opencl_transpose_32_16.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_transpose_32_16.cl COPYONLY)
+endif ()

ggml/src/ggml-opencl/kernels/embed_kernel.py

@@ -0,0 +1,26 @@
+#
+
+import sys
+import logging
+logger = logging.getLogger("opencl-embed-kernel")
+
+
+def main():
+    logging.basicConfig(level=logging.INFO)
+
+    if len(sys.argv) != 3:
+        logger.info("Usage: python embed_kernel.py <input_file> <output_file>")
+        sys.exit(1)
+
+    ifile = open(sys.argv[1], "r")
+    ofile = open(sys.argv[2], "w")
+
+    for i in ifile:
+        ofile.write('R"({})"\n'.format(i))
+
+    ifile.close()
+    ofile.close()
+
+
+if __name__ == "__main__":
+    main()

ggml/src/ggml-opencl/kernels/ggml-opencl_cvt.cl

@@ -0,0 +1,106 @@
+//------------------------------------------------------------------------------
+// This file is contains additional kernels for data conversion.
+// These kernels are used when loading the model, so its performance is less
+// important.
+//------------------------------------------------------------------------------
+#ifdef cl_khr_fp16
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#elif defined(cl_amd_fp16)
+#pragma OPENCL EXTENSION cl_amd_fp16 : enable
+#else
+#error "Half precision floating point not supportedby OpenCL implementation on your device."
+#endif
+
+#ifdef cl_khr_subgroups
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#elif defined(cl_intel_subgroups)
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#error "Subgroup not supported on your device."
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+// Always use subgroup size of 32 on Intel.
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+// Always use subgroups size of 64 on Adreno.
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#else
+// TODO: do not know how to choose subgroup size on other GPUs.
+#error "Selecting subgroup size is not supported on your device."
+#endif
+
+#define QK4_0                   32
+#define QR4_0                   2
+#define QK4_1                   32
+#define QR4_1                   2
+#define QK5_0                   32
+#define QR5_0                   2
+#define QK5_1                   32
+#define QR5_1                   2
+#define QK8_0                   32
+#define QR8_0                   1
+#define QK_K                    256
+#define K_QUANTS_PER_ITERATION  2
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+//------------------------------------------------------------------------------
+// mul_vec_q_n_f32_flat_noshuffle
+//
+// This variation uses flat arrays (struct of arrays, SOA) representation for
+// quant tensors. It also uses non shuffled bit order for weights.
+//
+// The shuffled version is kept in the original file because moving it here
+// seems to result in worse performance for adreno.
+//------------------------------------------------------------------------------
+
+kernel void kernel_convert_block_q4_0_noshuffle(
+    global struct block_q4_0 * src0,
+    global uchar * dst_q,
+    global half  * dst_d
+) {
+    global struct block_q4_0 * b = (global struct block_q4_0 *) src0 + get_global_id(0);
+    global uchar * q = (global uchar *) dst_q + QK4_0/2*get_global_id(0);
+    global half  * d = (global half *) dst_d + get_global_id(0);
+
+    *d = b->d;
+    for (int i = 0; i < QK4_0/4; ++i) {
+        uchar x0 = b->qs[2*i + 0];
+        uchar x1 = b->qs[2*i + 1];
+
+        q[i + 0      ] = convert_uchar(x0 & 0x0F) | convert_uchar((x1 & 0x0F) << 4);
+        q[i + QK4_0/4] = convert_uchar((x0 & 0xF0) >> 4) | convert_uchar(x1 & 0xF0);
+
+#ifdef ADRENO_GPU
+        // Workaround for adreno - must have the following printf statement for
+        // the kernel to work properly. Otherwise it produces incorrect result.
+        // convert_uchar above also seems necessary.
+        // Compare against a large number so that it does not print anything.
+        // get_sub_group_local_id() also works.
+        if (get_global_id(0) == 65536*4096) {
+            printf("%04x - %02x\n", *(global ushort*)d, ((x0 & 0xF0) >> 4) | (x1 & 0xF0));
+        }
+#endif
+    }
+}

ggml/src/ggml-opencl/kernels/ggml-opencl_gemv_noshuffle.cl

@@ -0,0 +1,265 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#pragma OPENCL EXTENSION cl_qcom_subgroup_uniform_load: enable
+#pragma OPENCL EXTENSION cl_qcom_subgroup_constant_load: enable
+#pragma OPENCL EXTENSION cl_qcom_extra_vector_types : enable
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+
+// assume
+#define QK4_0 32
+#define N_SIMDGROUP 4
+
+#define dequantizeBlockAccum_ns_sgbroadcast_1_hi(total_sums, bits4, scale, y) \
+    float shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 0); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 0); \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 0); \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 0); \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 0); \
+    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 0); \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 0); \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 0); \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 1); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 1); \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 1); \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 1); \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 1); \
+    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 1); \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 1); \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 1); \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_1_lo(total_sums, bits4, scale, y) \
+    shared_y = sub_group_broadcast(y.s0, 2); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 2); \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 2); \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 2); \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 2); \
+    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 2); \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 2); \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 2); \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 3); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 3); \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 3); \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 3); \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 3); \
+    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 3); \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 3); \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 3); \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_8_hi(total_sums, bits4, scale, y) \
+    float8 shared_y; \
+    shared_y = sub_group_broadcast(y, 0); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 1); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_8_lo(total_sums, bits4, scale, y) \
+    shared_y = sub_group_broadcast(y, 2); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 3); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \
+
+
+__attribute__((qcom_reqd_sub_group_size("full")))
+__kernel void kernel_gemv_noshuffle(
+        __read_only  image1d_buffer_t src0_q,  // quantized A
+        global half2  * src0_d,  // A scales
+        __read_only  image1d_buffer_t src1,    // B
+        ulong offset1,            // offset to B (0)
+        global float * dst,     // C
+        ulong offsetd,            // offset to C (0)
+        uint K,               // K
+        int ne01,               // M
+        int ne02,               // 1
+        int ne10,               // K
+        int ne12,               // 1
+        int ne0,                // M
+        int ne1,                // N
+        int r2,                 // 1
+        int r3)
+{
+    uint groupId = get_local_id(1);
+    uint gid     = get_global_id(0);
+    ushort slid    = get_sub_group_local_id();
+
+    __private uint4     regA;
+    __private half2     regS;
+    __private float8    regB;
+
+    __private float2 totalSum = (float2)(0.0f);
+
+    // loop along K in block granularity, skip 4 blocks every iter
+    for (uint k = groupId; k < (K / QK4_0); k += N_SIMDGROUP) {
+        regS = src0_d[gid + k * LINE_STRIDE_A]; // each fiber loads scale of two rows
+        // first 4 fibers in each wave load 8 B values to its private scope
+        if (slid < 4) {
+            regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));
+            regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));
+        }
+
+        // load half weights for two blocks in consecutive rows
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAT
+        dequantizeBlockAccum_ns_sgbroadcast_8_hi(totalSum, as_ushort8(regA), regS, regB);
+#else
+        dequantizeBlockAccum_ns_sgbroadcast_1_hi(totalSum, as_ushort8(regA), regS, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAT
+
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAT
+        dequantizeBlockAccum_ns_sgbroadcast_8_lo(totalSum, as_ushort8(regA), regS, regB);
+#else
+        dequantizeBlockAccum_ns_sgbroadcast_1_lo(totalSum, as_ushort8(regA), regS, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAT
+    }
+
+    // reduction in local memory, assumes #wave=4
+    __local float2 reduceLM[SIMDGROUP_WIDTH * 3];
+    if (groupId == 1) reduceLM[SIMDGROUP_WIDTH * 0 + slid] = totalSum;
+    if (groupId == 2) reduceLM[SIMDGROUP_WIDTH * 1 + slid] = totalSum;
+    if (groupId == 3) reduceLM[SIMDGROUP_WIDTH * 2 + slid] = totalSum;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 0 + slid];
+    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 1 + slid];
+    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 2 + slid];
+
+    // 2 outputs per fiber in wave 0
+    if (groupId == 0) {
+        dst = (global float*)((global char*)dst + offsetd);
+        vstore2(totalSum, 0, &(dst[gid * 2]));
+    }
+
+}

ggml/src/ggml-opencl/kernels/ggml-opencl_gemv_noshuffle_general.cl

@@ -0,0 +1,271 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#pragma OPENCL EXTENSION cl_qcom_subgroup_uniform_load: enable
+#pragma OPENCL EXTENSION cl_qcom_subgroup_constant_load: enable
+#pragma OPENCL EXTENSION cl_qcom_extra_vector_types : enable
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+
+// assume
+#define QK4_0 32
+#define N_SIMDGROUP 4
+
+#define dequantizeBlockAccum_ns_sgbroadcast_1_hi(total_sums, bits4, scale, y) \
+    float shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 0); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 0); \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 0); \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 0); \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 0); \
+    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 0); \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 0); \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 0); \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 1); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 1); \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 1); \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 1); \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 1); \
+    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 1); \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 1); \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 1); \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_1_lo(total_sums, bits4, scale, y) \
+    shared_y = sub_group_broadcast(y.s0, 2); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 2); \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 2); \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 2); \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 2); \
+    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 2); \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 2); \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 2); \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 3); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 3); \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 3); \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 3); \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 3); \
+    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 3); \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 3); \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 3); \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_8_hi(total_sums, bits4, scale, y) \
+    float8 shared_y; \
+    shared_y = sub_group_broadcast(y, 0); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 1); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_8_lo(total_sums, bits4, scale, y) \
+    shared_y = sub_group_broadcast(y, 2); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 3); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \
+
+
+__attribute__((qcom_reqd_sub_group_size("full")))
+__kernel void kernel_gemv_noshuffle(
+        __read_only  image1d_buffer_t src0_q,  // quantized A
+        global half2  * src0_d,  // A scales
+        __read_only  image1d_buffer_t src1,    // B
+        ulong offset1,            // offset to B (0)
+        global float * dst,     // C
+        ulong offsetd,            // offset to C (0)
+        int ne00,               // K
+        int ne01,               // M
+        int ne02,               // 1
+        int ne10,               // K
+        int ne12,               // 1
+        int ne0,                // M
+        int ne1,                // N
+        int r2,                 // 1
+        int r3)
+{
+    uint groupId = get_local_id(1);
+    uint gid     = get_global_id(0);
+    ushort slid    = get_sub_group_local_id();
+
+    uint K = ne00;
+    uint M = ne01;
+
+    uint LINE_STRIDE_A = M / 2;
+    uint BLOCK_STRIDE_A = N_SIMDGROUP * M;
+
+    __private uint4     regA;
+    __private half2     regS;
+    __private float8    regB;
+
+    __private float2 totalSum = (float2)(0.0f);
+
+    // loop along K in block granularity, skip 4 blocks every iter
+    for (uint k = groupId; k < (K / QK4_0); k += N_SIMDGROUP) {
+        regS = src0_d[gid + k * LINE_STRIDE_A]; // each fiber loads scale of two rows
+        // first 4 fibers in each wave load 8 B values to its private scope
+        if (slid < 4) {
+            regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));
+            regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));
+        }
+
+        // load half weights for two blocks in consecutive rows
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAT
+        dequantizeBlockAccum_ns_sgbroadcast_8_hi(totalSum, as_ushort8(regA), regS, regB);
+#else
+        dequantizeBlockAccum_ns_sgbroadcast_1_hi(totalSum, as_ushort8(regA), regS, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAT
+
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAT
+        dequantizeBlockAccum_ns_sgbroadcast_8_lo(totalSum, as_ushort8(regA), regS, regB);
+#else
+        dequantizeBlockAccum_ns_sgbroadcast_1_lo(totalSum, as_ushort8(regA), regS, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAT
+    }
+
+    // reduction in local memory, assumes #wave=4
+    __local float2 reduceLM[SIMDGROUP_WIDTH * 3];
+    if (groupId == 1) reduceLM[SIMDGROUP_WIDTH * 0 + slid] = totalSum;
+    if (groupId == 2) reduceLM[SIMDGROUP_WIDTH * 1 + slid] = totalSum;
+    if (groupId == 3) reduceLM[SIMDGROUP_WIDTH * 2 + slid] = totalSum;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 0 + slid];
+    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 1 + slid];
+    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 2 + slid];
+
+    // 2 outputs per fiber in wave 0
+    if (groupId == 0) {
+        dst = (global float*)((global char*)dst + offsetd);
+        vstore2(totalSum, 0, &(dst[gid * 2]));
+    }
+
+}

ggml/src/ggml-opencl/kernels/ggml-opencl_mul_mat_Ab_Bi_8x4.cl

@@ -0,0 +1,130 @@
+// src0_q, src0_d, src1 are transposed as a preprocessing step
+// 4-bit weights are transposed in groups of 4 (unsigned short int)
+// consider weights originally "next to each other", now "on top of each other"
+// each fiber computes a 8x4 tile of output elements
+// using unshuffled weights
+
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+
+__attribute__((qcom_reqd_sub_group_size("full")))
+kernel void kernel_mul_mat_Ab_Bi_8x4(
+        global const ushort * src0_q,       // quantized A
+        global const half  * src0_d,        // A scales
+        __read_only image1d_buffer_t src1,  // B (1d image)
+        global float * dst,                 // C
+        int m,                              // M
+        int n,                              // N with padding
+        int k,                              // K
+        int n_no_padding                    // N without padding
+) {
+
+    int m_4 = m >> 2;
+    int n_4 = n >> 2;
+
+    int gy = get_global_id(0);
+    int gx = get_global_id(1);
+    int gx_2 = gx << 2;
+
+    half8 c0 = 0, c1 = 0, c2 = 0, c3 = 0; // 8x4 output elements
+    half8 B; // registers for activations
+    half4 dequantized_weights; // registers for dequantized weights
+    __global const ushort* weight_ptr = src0_q + gx_2; // pointer for weights
+    __global const half* scale_ptr = src0_d + gx_2; // pointer for scales
+
+    for(int i=0; i<k; i+=4){ //loop through K dimension
+
+        B.s0123 = read_imageh(src1, gy*2 + (i)*(n_4));
+        B.s4567 = read_imageh(src1, gy*2 + (i)*(n_4)+1);
+
+        // keep (i/4) and (i/32) in parenthesis, rounds down
+        // load 4 consecutive groups of 4 weights
+        ushort4 bits4 = vload4(0, weight_ptr + (i/4)*(m)); // (i/4) because weights grouped in 4s
+
+        // load 4 consecutive scales
+        half4 scale = vload4(0, scale_ptr + (i/32)*(m));// (i/32) because 1 scale per 32 elements
+
+        // j=0
+        dequantized_weights.s0 = ((bits4.s0 & (0x000F)) - 8) * scale.s0; // dequantize a row of the 16 weights
+        dequantized_weights.s1 = ((bits4.s1 & (0x000F)) - 8) * scale.s1;
+        dequantized_weights.s2 = ((bits4.s2 & (0x000F)) - 8) * scale.s2;
+        dequantized_weights.s3 = ((bits4.s3 & (0x000F)) - 8) * scale.s3;
+        c0 += B * dequantized_weights.s0; // vector-scalar multiplication to accumulate
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=1
+        B.s0123 = read_imageh(src1, gy*2 + (i+1)*(n_4));
+        B.s4567 = read_imageh(src1, gy*2 + (i+1)*(n_4)+1);
+        dequantized_weights.s0 = (((bits4.s0 & (0x00F0)) >> 4) - 8) * scale.s0; // dequantize a row of the 16 weights
+        dequantized_weights.s1 = (((bits4.s1 & (0x00F0)) >> 4) - 8) * scale.s1;
+        dequantized_weights.s2 = (((bits4.s2 & (0x00F0)) >> 4) - 8) * scale.s2;
+        dequantized_weights.s3 = (((bits4.s3 & (0x00F0)) >> 4) - 8) * scale.s3;
+        c0 += B * dequantized_weights.s0; //vector-scalar multiplication to accumulate
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=2
+        B.s0123 = read_imageh(src1, gy*2 + (i+2)*(n_4));
+        B.s4567 = read_imageh(src1, gy*2 + (i+2)*(n_4)+1);
+        dequantized_weights.s0 = (((bits4.s0 & (0x0F00)) >> 8) - 8) * scale.s0; // dequantize a row of the 16 weights
+        dequantized_weights.s1 = (((bits4.s1 & (0x0F00)) >> 8) - 8) * scale.s1;
+        dequantized_weights.s2 = (((bits4.s2 & (0x0F00)) >> 8) - 8) * scale.s2;
+        dequantized_weights.s3 = (((bits4.s3 & (0x0F00)) >> 8) - 8) * scale.s3;
+        c0 += B * dequantized_weights.s0; // vector-scalar multiplication to accumulate
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=3
+        B.s0123 = read_imageh(src1, gy*2 + (i+3)*(n_4));
+        B.s4567 = read_imageh(src1, gy*2 + (i+3)*(n_4)+1);
+        dequantized_weights.s0 = (((bits4.s0 & (0xF000)) >> 12) - 8) * scale.s0; // dequantize a row of the 16 weights
+        dequantized_weights.s1 = (((bits4.s1 & (0xF000)) >> 12) - 8) * scale.s1;
+        dequantized_weights.s2 = (((bits4.s2 & (0xF000)) >> 12) - 8) * scale.s2;
+        dequantized_weights.s3 = (((bits4.s3 & (0xF000)) >> 12) - 8) * scale.s3;
+        c0 += B * dequantized_weights.s0; // vector-scalar multiplication to accumulate
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+    }
+
+    int idx = (gy<<3)*m + (gx<<2); // vectorized store 16 elements
+
+    // conditional check if store is to a valid location. Required when N is not a multiple of 8
+    // if statements allow registers to be reused for each store
+    // provides a performance boost due to reduced register footprint, which increases number of concurrent waves
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s0, c1.s0, c2.s0, c3.s0), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s1, c1.s1, c2.s1, c3.s1), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s2, c1.s2, c2.s2, c3.s2), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s3, c1.s3, c2.s3, c3.s3), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s4, c1.s4, c2.s4, c3.s4), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s5, c1.s5, c2.s5, c3.s5), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s6, c1.s6, c2.s6, c3.s6), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s7, c1.s7, c2.s7, c3.s7), 0, dst + idx);
+    }
+}

ggml/src/ggml-opencl/kernels/ggml-opencl_transpose_16.cl

@@ -0,0 +1,32 @@
+// 16-bit transpose, loading/storing an 8x8 tile of elements
+
+kernel void kernel_transpose_16(
+    __read_only image1d_buffer_t input,
+    __write_only image1d_buffer_t output,
+    const uint rows,
+    const uint cols
+) {
+
+    const int i = get_global_id(0);
+    const int j = get_global_id(1);
+    const int i_3 = i<<3;
+    const int j_3 = j<<3;
+
+    ushort8 temp0 = as_ushort8(read_imagef(input, (j_3+0)*cols+i));
+    ushort8 temp1 = as_ushort8(read_imagef(input, (j_3+1)*cols+i));
+    ushort8 temp2 = as_ushort8(read_imagef(input, (j_3+2)*cols+i));
+    ushort8 temp3 = as_ushort8(read_imagef(input, (j_3+3)*cols+i));
+    ushort8 temp4 = as_ushort8(read_imagef(input, (j_3+4)*cols+i));
+    ushort8 temp5 = as_ushort8(read_imagef(input, (j_3+5)*cols+i));
+    ushort8 temp6 = as_ushort8(read_imagef(input, (j_3+6)*cols+i));
+    ushort8 temp7 = as_ushort8(read_imagef(input, (j_3+7)*cols+i));
+
+    write_imagef(output, (i_3+0)*rows+j, as_float4((ushort8)(temp0.s0, temp1.s0, temp2.s0, temp3.s0, temp4.s0, temp5.s0, temp6.s0, temp7.s0)));
+    write_imagef(output, (i_3+1)*rows+j, as_float4((ushort8)(temp0.s1, temp1.s1, temp2.s1, temp3.s1, temp4.s1, temp5.s1, temp6.s1, temp7.s1)));
+    write_imagef(output, (i_3+2)*rows+j, as_float4((ushort8)(temp0.s2, temp1.s2, temp2.s2, temp3.s2, temp4.s2, temp5.s2, temp6.s2, temp7.s2)));
+    write_imagef(output, (i_3+3)*rows+j, as_float4((ushort8)(temp0.s3, temp1.s3, temp2.s3, temp3.s3, temp4.s3, temp5.s3, temp6.s3, temp7.s3)));
+    write_imagef(output, (i_3+4)*rows+j, as_float4((ushort8)(temp0.s4, temp1.s4, temp2.s4, temp3.s4, temp4.s4, temp5.s4, temp6.s4, temp7.s4)));
+    write_imagef(output, (i_3+5)*rows+j, as_float4((ushort8)(temp0.s5, temp1.s5, temp2.s5, temp3.s5, temp4.s5, temp5.s5, temp6.s5, temp7.s5)));
+    write_imagef(output, (i_3+6)*rows+j, as_float4((ushort8)(temp0.s6, temp1.s6, temp2.s6, temp3.s6, temp4.s6, temp5.s6, temp6.s6, temp7.s6)));
+    write_imagef(output, (i_3+7)*rows+j, as_float4((ushort8)(temp0.s7, temp1.s7, temp2.s7, temp3.s7, temp4.s7, temp5.s7, temp6.s7, temp7.s7)));
+}

ggml/src/ggml-opencl/kernels/ggml-opencl_transpose_32.cl

@@ -0,0 +1,25 @@
+// 32-bit transpose, loading/storing a 4x4 tile of elements
+
+kernel void kernel_transpose_32(
+    __read_only image1d_buffer_t input,
+    __write_only image1d_buffer_t output,
+    const uint rows,
+    const uint cols
+) {
+
+    const int i = get_global_id(0);
+    const int j = get_global_id(1);
+    const int i_2 = i<<2;
+    const int j_2 = j<<2;
+
+    float4 temp0 = read_imagef(input, (j_2+0)*cols+i);
+    float4 temp1 = read_imagef(input, (j_2+1)*cols+i);
+    float4 temp2 = read_imagef(input, (j_2+2)*cols+i);
+    float4 temp3 = read_imagef(input, (j_2+3)*cols+i);
+
+    write_imagef(output, (i_2+0)*rows+j, (float4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0));
+    write_imagef(output, (i_2+1)*rows+j, (float4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1));
+    write_imagef(output, (i_2+2)*rows+j, (float4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2));
+    write_imagef(output, (i_2+3)*rows+j, (float4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
+
+}

ggml/src/ggml-opencl/kernels/ggml-opencl_transpose_32_16.cl

@@ -0,0 +1,35 @@
+// 32-bit transpose, loading/storing a 4x4 tile of elements
+// Only used for activations
+// converts to FP16
+// also adds zero padding for non multiple of 8 prompt lengths
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+kernel void kernel_transpose_32_16(__read_only image1d_buffer_t input, __write_only image1d_buffer_t output, const uint rows, const uint cols, const uint padded_rows) {
+
+    const int i = get_global_id(0);
+    const int j = get_global_id(1);
+    const int i_2 = i<<2;
+    const int j_2 = j<<2;
+    half4 temp0 = {0,0,0,0}; // initialize outputs to 0
+    half4 temp1 = {0,0,0,0};
+    half4 temp2 = {0,0,0,0};
+    half4 temp3 = {0,0,0,0};
+
+    if((j_2+0)*cols+i*4+3 < rows*cols*16){ // only load from a valid location. Otherwise keep register data as 0
+        temp0 = read_imageh(input, (j_2+0)*cols+i);
+    }
+    if((j_2+1)*cols+i*4+3 < rows*cols*16){
+        temp1 = read_imageh(input, (j_2+1)*cols+i);
+    }
+    if((j_2+2)*cols+i*4+3 < rows*cols*16){
+        temp2 = read_imageh(input, (j_2+2)*cols+i);
+    }
+    if((j_2+3)*cols+i*4+3 < rows*cols*16){
+        temp3 = read_imageh(input, (j_2+3)*cols+i);
+    }
+
+    write_imageh(output, (i_2+0)*padded_rows+j, (half4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0)); // no conditionals for output, includes zero padding
+    write_imageh(output, (i_2+1)*padded_rows+j, (half4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1));
+    write_imageh(output, (i_2+2)*padded_rows+j, (half4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2));
+    write_imageh(output, (i_2+3)*padded_rows+j, (half4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
+}

ggml/src/ggml-sycl/common.cpp

@@ -11,6 +11,7 @@
 //
 
 #include "common.hpp"
+#include "ggml-impl.h"
 
 int get_current_device_id() {
   return dpct::dev_mgr::instance().current_device_id();
@@ -28,11 +29,7 @@ void* ggml_sycl_host_malloc(size_t size) try {
 
   if (err != 0) {
     // clear the error
-    fprintf(
+    GGML_LOG_ERROR("WARNING: failed to allocate %.2f MB of pinned memory: %s\n", size / 1024.0 / 1024.0,    "syclGetErrorString is not supported");
-        stderr,
-        "WARNING: failed to allocate %.2f MB of pinned memory: %s\n",
-        size / 1024.0 / 1024.0,
-        "syclGetErrorString is not supported");
     return nullptr;
   }
 
@@ -66,18 +63,12 @@ int64_t downsample_sycl_global_range(int64_t accumulate_block_num, int64_t block
 void ggml_sycl_op_flatten(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
                                  const ggml_tensor *src1, ggml_tensor *dst,
                                  const ggml_sycl_op_flatten_t op) try {
-    const int64_t nrows0 = ggml_nrows(src0);
 
     const bool use_src1 = src1 != nullptr;
-    const int64_t nrows1 = use_src1 ? ggml_nrows(src1) : 1;
 
     GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
     GGML_ASSERT(              dst->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
 
-    ggml_tensor_extra_gpu * src0_extra =            (ggml_tensor_extra_gpu *) src0->extra;
-    ggml_tensor_extra_gpu * src1_extra = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr;
-    ggml_tensor_extra_gpu * dst_extra  =            (ggml_tensor_extra_gpu *)  dst->extra;
-
     // dd = data device
     float * src0_ddf = (float *) src0->data;
     float * src1_ddf = use_src1 ? (float *) src1->data : nullptr;

ggml/src/ggml-sycl/common.hpp

@@ -626,6 +626,7 @@ struct bin_bcast_sycl {
                     });
             }
         }
+        GGML_UNUSED(ctx);
     }
 };
 

ggml/src/ggml-sycl/concat.cpp

@@ -47,7 +47,7 @@ static void concat_f32_dim1(const float *x, const float *y, float *dst,
   // operation
   int offset_dst = nidx + item_ct1.get_group(1) * ne0 +
                    item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1);
-  if (item_ct1.get_group(1) < ne01) { // src0
+  if (item_ct1.get_group(1) < (size_t) ne01) { // src0
     int offset_src =
         nidx + item_ct1.get_group(1) * ne0 + item_ct1.get_group(0) * ne0 * ne01;
     dst[offset_dst] = x[offset_src];
@@ -70,7 +70,7 @@ static void concat_f32_dim2(const float *x, const float *y, float *dst,
   // operation
   int offset_dst = nidx + item_ct1.get_group(1) * ne0 +
                    item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1);
-  if (item_ct1.get_group(0) < ne02) { // src0
+  if (item_ct1.get_group(0) < (size_t) ne02) { // src0
     int offset_src = nidx + item_ct1.get_group(1) * ne0 +
                      item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1);
     dst[offset_dst] = x[offset_src];

ggml/src/ggml-sycl/convert.cpp

@@ -424,7 +424,7 @@ static void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y,
     const int64_t global_id = item_ct1.get_local_id(2) + work_group_size * item_ct1.get_group(2);
 
     // make each work-item deal with more elements since sycl global range can not exceed max int
-    const src_t * x = (src_t *) vx;
+    const src_t * x = (const src_t *) vx;
     for (int64_t i = global_id; i < k; i += work_group_size * item_ct1.get_group_range(2)) {
         y[i] = x[i];
     }

ggml/src/ggml-sycl/dmmv.cpp

@@ -1015,9 +1015,9 @@ void ggml_sycl_op_dequantize_mul_mat_vec(
             break;
     }
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_ddq_i;
+    GGML_UNUSED(src1_ddq_i);
-    (void) src1_ncols;
+    GGML_UNUSED(src1_ncols);
-    (void) src1_padded_row_size;
+    GGML_UNUSED(src1_padded_row_size);
 }

ggml/src/ggml-sycl/dpct/helper.hpp

@@ -1237,7 +1237,7 @@ namespace dpct
 
             std::map<byte_t *, allocation>::iterator get_map_iterator(const void *ptr)
             {
-                auto it = m_map.upper_bound((byte_t *)ptr);
+                auto it = m_map.upper_bound(const_cast<byte_t *>(reinterpret_cast<const byte_t *>(ptr)));
                 if (it == m_map.end())
                 {
                     // Not a virtual pointer.

ggml/src/ggml-sycl/element_wise.cpp

@@ -237,7 +237,7 @@ void upscale_f32(const float  *x, float *dst, const int nb00, const int nb01,
     int i02 = i12 / sf2;
     int i03 = i13 / sf3;
 
-    dst[index] = *(float *)((char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00);
+    dst[index] = *(const float *)((const char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00);
 }
 
 void pad_f32(const float  *x, float *dst, const int ne0, const int ne00, const int ne01, const int ne02,
@@ -251,8 +251,7 @@ void pad_f32(const float  *x, float *dst, const int ne0, const int ne00, const i
     // operation
     int offset_dst = nidx + item_ct1.get_group(1) * ne0 +
                      item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1);
-    if (nidx < ne00 && item_ct1.get_group(1) < ne01 &&
+    if (nidx < ne00 && item_ct1.get_group(1) < (size_t) ne01 && item_ct1.get_group(0) < (size_t) ne02) {
-        item_ct1.get_group(0) < ne02) {
         int offset_src = nidx + item_ct1.get_group(1) * ne00 +
                          item_ct1.get_group(0) * ne00 * ne01;
             dst[offset_dst] = x[offset_src];
@@ -520,9 +519,10 @@ inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     silu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_gelu(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
@@ -535,9 +535,10 @@ inline void ggml_sycl_op_gelu(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     gelu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 inline void ggml_sycl_op_gelu_quick(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
                                     const ggml_tensor *src1, ggml_tensor *dst,
@@ -550,9 +551,10 @@ inline void ggml_sycl_op_gelu_quick(ggml_backend_sycl_context & ctx, const ggml_
 
     gelu_quick_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_tanh(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
@@ -564,9 +566,10 @@ inline void ggml_sycl_op_tanh(ggml_backend_sycl_context & ctx, const ggml_tensor
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
     tanh_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_relu(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
@@ -579,9 +582,10 @@ inline void ggml_sycl_op_relu(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     relu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_hardsigmoid(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -595,9 +599,10 @@ inline void ggml_sycl_op_hardsigmoid(ggml_backend_sycl_context & ctx, const ggml
 
     hardsigmoid_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_hardswish(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -610,9 +615,10 @@ inline void ggml_sycl_op_hardswish(ggml_backend_sycl_context & ctx, const ggml_t
 
     hardswish_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_exp(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -625,9 +631,10 @@ inline void ggml_sycl_op_exp(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     exp_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_log(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -640,9 +647,10 @@ inline void ggml_sycl_op_log(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     log_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_sigmoid(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -655,9 +663,10 @@ inline void ggml_sycl_op_sigmoid(ggml_backend_sycl_context & ctx, const ggml_ten
 
     sigmoid_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_sqrt(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -670,9 +679,10 @@ inline void ggml_sycl_op_sqrt(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     sqrt_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_sin(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -685,9 +695,10 @@ inline void ggml_sycl_op_sin(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     sin_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_cos(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -700,9 +711,10 @@ inline void ggml_sycl_op_cos(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     cos_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_step(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -715,9 +727,10 @@ inline void ggml_sycl_op_step(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     step_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_neg(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -730,9 +743,10 @@ inline void ggml_sycl_op_neg(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     neg_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_leaky_relu(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -749,9 +763,10 @@ inline void ggml_sycl_op_leaky_relu(ggml_backend_sycl_context & ctx, const ggml_
 
     leaky_relu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), negative_slope, main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
@@ -764,9 +779,10 @@ inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     sqr_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -787,9 +803,10 @@ inline void ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, const ggml_ten
                      dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], sf0, sf1, sf2, sf3,
                      main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
@@ -805,9 +822,10 @@ inline void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, const ggml_tensor
         src0->ne[0], src0->ne[1], src0->ne[2],
         dst->ne[0], dst->ne[1], dst->ne[2], main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
@@ -827,7 +845,8 @@ inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     acc_f32_sycl(src0_dd, src1_dd, dst_dd, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, main_stream);
 
-    (void) dst;
+    GGML_UNUSED(dst);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_add(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,

ggml/src/ggml-sycl/gemm.hpp

@@ -51,8 +51,8 @@ public:
         const auto a_in_md = dnnl::memory::desc(a_dims, at, a_trans ? tag::ba : tag::ab);
         const auto b_in_md = dnnl::memory::desc(b_dims, bt, b_trans ? tag::ba : tag::ab);
         const auto c_md = dnnl::memory::desc(c_dims, ct, tag::ab);
-        auto a_mem = dnnl::memory(a_in_md, eng, (void*)a);
+        auto a_mem = dnnl::memory(a_in_md, eng, const_cast<void*>(a));
-        auto b_mem = dnnl::memory(b_in_md, eng, (void*)b);
+        auto b_mem = dnnl::memory(b_in_md, eng, const_cast<void*>(b));
         auto matmul_pd = dnnl::matmul::primitive_desc(eng, a_in_md, b_in_md, c_md);
         auto c_mem = dnnl::memory(matmul_pd.dst_desc(), eng, c);
 
@@ -79,8 +79,8 @@ public:
         const auto a_in_md = dnnl::memory::desc(a_dims, at, a_trans ? tag::ba : tag::ab);
         const auto b_in_md = dnnl::memory::desc(b_dims, bt, b_trans ? tag::ba : tag::ab);
         const auto c_md = dnnl::memory::desc(c_dims, ct, tag::ab);
-        auto a_mem = dnnl::memory(a_in_md, eng, (void*)a);
+        auto a_mem = dnnl::memory(a_in_md, eng, const_cast<void*>(a));
-        auto b_mem = dnnl::memory(b_in_md, eng, (void*)b);
+        auto b_mem = dnnl::memory(b_in_md, eng, const_cast<void*>(b));
         auto matmul_pd = dnnl::matmul::primitive_desc(eng, a_in_md, b_in_md, c_md);
         auto c_mem = dnnl::memory(matmul_pd.dst_desc(), eng, c);
 

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -47,7 +47,7 @@ static ggml_sycl_device_info ggml_sycl_init() {
 
     info.device_count = dpct::dev_mgr::instance().device_count();
     if (info.device_count == 0) {
-        GGML_LOG_ERROR("%s: failed to initialize " GGML_SYCL_NAME ": %s\n", __func__);
+        GGML_LOG_ERROR("%s: failed to initialize: %s\n", GGML_SYCL_NAME, __func__);
         return info;
     }
 
@@ -64,7 +64,7 @@ static ggml_sycl_device_info ggml_sycl_init() {
 #else
     GGML_LOG_INFO("%s: SYCL_USE_XMX: no\n", __func__);
 #endif
-    GGML_LOG_INFO("%s: found %d " GGML_SYCL_NAME " devices:\n", __func__, info.device_count);
+    GGML_LOG_INFO("%s: found %d %s devices:\n", __func__, info.device_count, GGML_SYCL_NAME);
 
     for (int i = 0; i < info.device_count; ++i) {
         info.devices[i].vmm = 0;
@@ -137,7 +137,6 @@ void ggml_backend_sycl_print_sycl_devices() {
 
     for (int id = 0; id < device_count; ++id) {
       sycl::device device = dpct::dev_mgr::instance().get_device(id);
-      sycl::backend backend = device.get_backend();
       std::string backend_type = get_device_backend_and_type(device);
       int type_id = DeviceNums[backend_type]++;
       std::stringstream device_type;
@@ -420,14 +419,12 @@ ggml_backend_sycl_buffer_cpy_tensor(ggml_backend_buffer_t buffer,
         return true;
     }
     return false;
-}
+    GGML_UNUSED(buffer);
-catch (sycl::exception const &exc) {
+} catch (const sycl::exception & exc) {
-  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+    std::cerr << exc.what() << "Exception caught at file:" << __FILE__ << ", line:" << __LINE__ << std::endl;
-            << ", line:" << __LINE__ << std::endl;
     std::exit(1);
 }
 
-
 static void ggml_backend_sycl_buffer_clear(ggml_backend_buffer_t buffer,
                                            uint8_t value) try {
      ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
@@ -1092,10 +1089,7 @@ struct ggml_sycl_pool_leg : public ggml_sycl_pool {
     ggml_sycl_buffer buffer_pool[MAX_SYCL_BUFFERS] = {};
     size_t pool_size = 0;
 
-    explicit ggml_sycl_pool_leg(queue_ptr qptr_, int device_) :
+    explicit ggml_sycl_pool_leg(queue_ptr qptr_, int device_) : device(device_), qptr(qptr_) {}
-        qptr(qptr_),
-        device(device_) {
-    }
 
     ~ggml_sycl_pool_leg() {
         for (int i = 0; i < MAX_SYCL_BUFFERS; ++i) {
@@ -1238,7 +1232,7 @@ static void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy,
         zeros[i] = 0.f;
         qzeros[i] = 0;
     }
-    const TC xi = ix < kx ? *(TC *)&x[iy * kx + ix] : zeros;
+    const TC xi = ix < kx ? *(const TC *)&x[iy * kx + ix] : zeros;
     float sum = xi[0];
     float amax = sycl::fabs(xi[0]);
 #pragma unroll
@@ -1799,6 +1793,9 @@ static  void pool2d_nchw_kernel(
         switch (op) {
             case GGML_OP_POOL_AVG: res = 0; break;
             case GGML_OP_POOL_MAX: res = -FLT_MAX; break;
+            default:
+                res      = (To) sycl::nan(uint32_t(0));
+                break;
         }
 
         for (int i = bh; i < eh; i += 1) {
@@ -1817,6 +1814,9 @@ static  void pool2d_nchw_kernel(
                 switch (op) {
                     case GGML_OP_POOL_AVG: res += (cur / (kh * kw)); break;
                     case GGML_OP_POOL_MAX: res = sycl::max(res, (To)cur); break;
+                    default:
+                        res = (To) sycl::nan(uint32_t(0));
+                        break;
                 }
             }
         }
@@ -1855,7 +1855,8 @@ static void get_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor *sr
                                  s3, nb01, nb02, nb03, s10, s11, s12, item_ct1);
                          });
 
-    (void) dst;
+    GGML_UNUSED(dst);
+    GGML_UNUSED(ctx);
 }
 
 template <typename src0_t>
@@ -1893,10 +1894,10 @@ static void get_rows_sycl_float(ggml_backend_sycl_context & ctx, const ggml_tens
             });
     }
 
-    (void) dst;
+    GGML_UNUSED(dst);
+    GGML_UNUSED(ctx);
 }
 
-
 static void quantize_row_q8_1_sycl(const float *x, void *vy, const int kx,
                                    const int ky, const int kx_padded,
                                    queue_ptr stream) {
@@ -2464,8 +2465,8 @@ static void ggml_sycl_op_repeat(ggml_backend_sycl_context & ctx, const ggml_tens
 
     ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_repeat>>(ctx, dst, src0, dst, nullptr, src0_d, dst_d, main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) src1_d;
+    GGML_UNUSED(src1_d);
 }
 
 
@@ -2484,17 +2485,18 @@ inline void ggml_sycl_op_mul_mat_sycl(
     const int64_t ne00 = src0->ne[0];
     const int64_t ne10 = src1->ne[0];
 
-    const int64_t ne0 = dst->ne[0];
 
     const int64_t row_diff = row_high - row_low;
 
     int id;
     SYCL_CHECK(
         CHECK_TRY_ERROR(id = get_current_device_id()));
-
+#if !GGML_SYCL_DNNL
+    const int64_t ne0 = dst->ne[0];
     // the main device has a larger memory buffer to hold the results from all GPUs
     // ldc == nrows of the matrix that cuBLAS writes into
     int ldc = id == ctx.device ? ne0 : row_diff;
+#endif
 
 #ifdef GGML_SYCL_F16
     bool use_fp16 = true;  // TODO(Yu) SYCL capability check
@@ -2531,9 +2533,9 @@ inline void ggml_sycl_op_mul_mat_sycl(
                                          : src1_as_f16.get();
         ggml_sycl_pool_alloc<sycl::half> dst_f16(ctx.pool(), row_diff * src1_ncols);
 
+#if !GGML_SYCL_DNNL
         const sycl::half alpha_f16 = 1.0f;
         const sycl::half beta_f16  = 0.0f;
-#if !GGML_SYCL_DNNL
         SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm(
             *stream, oneapi::mkl::transpose::trans,
             oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10,
@@ -2570,9 +2572,9 @@ inline void ggml_sycl_op_mul_mat_sycl(
         const float * src0_ddf_i = src0->type == GGML_TYPE_F32 ? (const float *) src0_dd_i : src0_ddq_as_f32.get();
         const float * src1_ddf1_i = src1->type == GGML_TYPE_F32 ? (const float *) src1_ddf_i : src1_ddq_as_f32.get();
 
+#if !GGML_SYCL_DNNL
         const float alpha = 1.0f;
         const float beta  = 0.0f;
-#if !GGML_SYCL_DNNL
 #    ifdef GGML_SYCL_NVIDIA
         SYCL_CHECK(CHECK_TRY_ERROR(oneapi::mkl::blas::column_major::gemm(
             oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ *stream }, oneapi::mkl::transpose::trans,
@@ -2590,9 +2592,9 @@ inline void ggml_sycl_op_mul_mat_sycl(
             src0_ddf_i, DnnlGemmWrapper::to_dt<float>(), dst_dd_i, DnnlGemmWrapper::to_dt<float>());
 #endif
     }
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_ddq_i;
+    GGML_UNUSED(src1_ddq_i);
-    (void) src1_padded_row_size;
+    GGML_UNUSED(src1_padded_row_size);
 }
 catch (sycl::exception const &exc) {
   std::cerr << exc.what() << "Exception caught at file:" << __FILE__
@@ -2638,8 +2640,9 @@ static void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, const ggml_tens
                                item_ct1);
         });
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_sum(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -2654,9 +2657,10 @@ inline void ggml_sycl_op_sum(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     sum_rows_f32_sycl(src0_dd, dst_dd, ne, 1, main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_sum_rows(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -2673,9 +2677,10 @@ inline void ggml_sycl_op_sum_rows(ggml_backend_sycl_context & ctx, const ggml_te
 
     sum_rows_f32_sycl(src0_dd, dst_dd, ncols, nrows, main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_argsort(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -2694,9 +2699,10 @@ inline void ggml_sycl_op_argsort(ggml_backend_sycl_context & ctx, const ggml_ten
 
     argsort_f32_i32_sycl(src0_dd, (int *)dst_dd, ncols, nrows, order, main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_argmax(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -2713,9 +2719,10 @@ inline void ggml_sycl_op_argmax(ggml_backend_sycl_context & ctx, const ggml_tens
 
     argmax_f32_i32_sycl(src0_dd, (int *)dst_dd, ncols, nrows, main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_diag_mask_inf(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -2735,9 +2742,10 @@ inline void ggml_sycl_op_diag_mask_inf(ggml_backend_sycl_context & ctx, const gg
 
     diag_mask_inf_f32_sycl(src0_dd, dst_dd, ne00, nrows0, ne01, n_past, main_stream);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_scale(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
@@ -2758,9 +2766,10 @@ inline void ggml_sycl_op_scale(ggml_backend_sycl_context & ctx, const ggml_tenso
     */
     SYCL_CHECK(0);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
@@ -2783,9 +2792,10 @@ inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, const ggml_tenso
     */
     SYCL_CHECK(0);
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 static void ggml_sycl_set_peer_access(const int n_tokens, int main_device) {
@@ -2862,7 +2872,6 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
 
     ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
     ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
-    ggml_tensor_extra_gpu *  dst_extra = (ggml_tensor_extra_gpu *)  dst->extra;
 
     const bool src0_is_contiguous = ggml_is_contiguous(src0);
     const bool src1_is_contiguous = ggml_is_contiguous(src1);
@@ -3289,7 +3298,6 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx,
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
-    const int64_t ne_dst = ggml_nelements(dst);
 
     SYCL_CHECK(ggml_sycl_set_device(ctx.device));
     queue_ptr main_stream = ctx.stream();;
@@ -3397,6 +3405,7 @@ catch (sycl::exception const &exc) {
 
 inline bool ggml_sycl_supports_mmq(enum ggml_type type) {
     // TODO: accuracy issues in MMQ
+    GGML_UNUSED(type);
     return false;
 }
 
@@ -3772,7 +3781,7 @@ static void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor *sr
         GGML_ABORT("fatal error");
     }
 
-    (void) dst;
+    GGML_UNUSED(dst);
 }
 catch (sycl::exception const &exc) {
   std::cerr << exc.what() << "Exception caught at file:" << __FILE__
@@ -3783,7 +3792,7 @@ catch (sycl::exception const &exc) {
 static void ggml_sycl_dup(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     // TODO: why do we pass dst as src1 here?
     ggml_sycl_cpy(ctx, src0, dst, nullptr);
-    (void) src1;
+    GGML_UNUSED(src1);
 }
 
 static void ggml_sycl_diag_mask_inf(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -3828,13 +3837,16 @@ static void ggml_sycl_argmax(ggml_backend_sycl_context & ctx, const ggml_tensor
 }
 
 static void ggml_sycl_nop(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    (void) src0;
+    GGML_UNUSED(src0);
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
+    GGML_UNUSED(ctx);
 }
 
 void ggml_sycl_set_main_device(const int main_device) try {
-    if (dpct::get_current_device_id() == main_device) return;
+    if (dpct::get_current_device_id() == static_cast<unsigned int> (main_device)) {
+        return;
+    }
     check_allow_gpu_index(main_device);
     dpct::select_device(main_device);
 
@@ -4202,6 +4214,7 @@ try
 {
     ggml_backend_sycl_context *sycl_ctx =
         (ggml_backend_sycl_context *)backend->context;
+
     sycl::event *sycl_event = static_cast<sycl::event *>(event->context);
 
     const queue_ptr &stream = sycl_ctx->stream(sycl_ctx->device, 0);
@@ -4216,7 +4229,7 @@ catch (sycl::exception const &exc)
 }
 
 static void ggml_backend_sycl_event_wait(ggml_backend_t backend, ggml_backend_event_t event) try {
-    ggml_backend_sycl_context* sycl_ctx = static_cast<ggml_backend_sycl_context*>(backend->context);
+
     sycl::event* sycl_event = static_cast<sycl::event*>(event->context);
 
     if (ggml_backend_is_sycl(backend)) {
@@ -4475,7 +4488,16 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_SOFT_MAX:
             return true;
         case GGML_OP_ROPE:
+            {
+                const int mode = ((const int32_t *) op->op_params)[2];
+                if (mode & GGML_ROPE_TYPE_MROPE) {
+                    return false;
+                }
+                if (mode & GGML_ROPE_TYPE_VISION) {
+                    return false;
+                }
                 return ggml_is_contiguous(op->src[0]);
+            }
         case GGML_OP_IM2COL:
             // TODO: add support for the new F32 operations
             return op->src[0]->type == GGML_TYPE_F16;
@@ -4624,6 +4646,7 @@ static void *ggml_backend_sycl_reg_get_proc_address(ggml_backend_reg_t reg, cons
     // SYCL doesn't support registering host memory, left here for reference
     // "ggml_backend_register_host_buffer"
     // "ggml_backend_unregister_host_buffer"
+    GGML_UNUSED(name);
     return nullptr;
 }
 

ggml/src/ggml-sycl/im2col.cpp

@@ -120,6 +120,7 @@ void ggml_sycl_op_im2col(
         im2col_sycl(src1_dd, (float *)dst_dd, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, main_stream);
     }
 
-    (void) src0;
+    GGML_UNUSED(src0);
-    (void) src0_dd;
+    GGML_UNUSED(src0_dd);
+    GGML_UNUSED(ctx);
 }

ggml/src/ggml-sycl/mmq.cpp

@@ -813,7 +813,7 @@ load_tiles_q4_K(const void *__restrict__ vx, int *__restrict__ x_ql,
         x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx);
     }
 
-    const int blocks_per_tile_x_row = WARP_SIZE / QI4_K; // == 1 if QK_K == 256
+    constexpr int blocks_per_tile_x_row = QI4_K > WARP_SIZE ? 1 : WARP_SIZE / QI4_K; // == 1 if QK_K == 256
     const int kbxd = k % blocks_per_tile_x_row;          // == 0 if QK_K == 256
 
 #pragma unroll
@@ -961,7 +961,7 @@ load_tiles_q5_K(const void *__restrict__ vx, int *__restrict__ x_ql,
         x_ql[i * (2*WARP_SIZE + 1) + kq1] = ql1 | qh1;
     }
 
-    const int blocks_per_tile_x_row = WARP_SIZE / QI5_K; // == 1 if QK_K == 256
+    constexpr int blocks_per_tile_x_row = QI5_K > WARP_SIZE ? 1 : WARP_SIZE / QI5_K; // == 1 if QK_K == 256
     const int kbxd = k % blocks_per_tile_x_row;          // == 0 if QK_K == 256
 
 #pragma unroll
@@ -1109,7 +1109,7 @@ load_tiles_q6_K(const void *__restrict__ vx, int *__restrict__ x_ql,
                                                  dpct::sub_sat());
     }
 
-    const int blocks_per_tile_x_row = WARP_SIZE / QI6_K; // == 1 if QK_K == 256
+    constexpr int blocks_per_tile_x_row = QI6_K > WARP_SIZE ? 1 : WARP_SIZE / QI6_K; // == 1 if QK_K == 256
     const int kbxd = k % blocks_per_tile_x_row;          // == 0 if QK_K == 256
     float * x_dmf = (float *) x_dm;
 
@@ -3020,9 +3020,9 @@ void ggml_sycl_op_mul_mat_q(
             break;
     }
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_ddf_i;
+    GGML_UNUSED(src1_ddf_i);
 }
 catch (sycl::exception const &exc) {
   std::cerr << exc.what() << "Exception caught at file:" << __FILE__

ggml/src/ggml-sycl/mmvq.cpp

@@ -753,11 +753,7 @@ static void mul_mat_vec_iq2_xs_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
     {
-
+        stream->submit([&](sycl::handler & cgh) {
-        stream->submit([&](sycl::handler &cgh) {
-            auto iq2xs_grid_ptr_ct1 = &iq2xs_grid[0];
-            auto ksigns64_ptr_ct1 = &ksigns64[0];
-
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
@@ -780,9 +776,6 @@ static void mul_mat_vec_iq2_s_q8_1_sycl(const void *vx, const void *vy,
     {
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq2xs_grid_ptr_ct1 = &iq2xs_grid[0];
-            auto ksigns64_ptr_ct1 = &ksigns64[0];
-
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
@@ -805,9 +798,6 @@ static void mul_mat_vec_iq3_xxs_q8_1_sycl(const void *vx, const void *vy,
     {
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq3xxs_grid_ptr_ct1 = &iq3xxs_grid[0];
-            auto ksigns64_ptr_ct1 = &ksigns64[0];
-
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
@@ -830,8 +820,6 @@ static void mul_mat_vec_iq3_s_q8_1_sycl(const void *vx, const void *vy,
     {
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq3s_grid_ptr_ct1 = &iq3s_grid[0];
-
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
@@ -854,9 +842,6 @@ static void mul_mat_vec_iq1_s_q8_1_sycl(const void *vx, const void *vy,
     {
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq1s_grid_ptr_ct1 = &iq1s_grid_gpu[0];
-            auto ksigns64_ptr_ct1 = &ksigns64[0];
-
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
@@ -954,7 +939,7 @@ void ggml_sycl_op_mul_mat_vec_q(
     const size_t q8_1_bs = QK8_1;
     // the main device has a larger memory buffer to hold the results from all GPUs
     // nrows_dst == nrows of the matrix that the kernel writes into
-    const int64_t nrows_dst = id == ctx.device ? ne00 : row_diff;
+
     for (int i = 0; i < src1_ncols; i++)
     {
         const size_t src1_ddq_i_offset = i * src1_padded_col_size * q8_1_ts / q8_1_bs;
@@ -1023,7 +1008,8 @@ void ggml_sycl_op_mul_mat_vec_q(
             break;
         }
     }
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_ddf_i;
+    GGML_UNUSED(src1_ddf_i);
+    GGML_UNUSED(ctx);
 }

ggml/src/ggml-sycl/norm.cpp

@@ -31,7 +31,7 @@ static void norm_f32(const float* x, float* dst, const int ncols, const float ep
         */
         item_ct1.barrier(sycl::access::fence_space::local_space);
         mean_var = 0.f;
-        int nreduce = nwarps / WARP_SIZE;
+        size_t nreduce = nwarps / WARP_SIZE;
         for (size_t i = 0; i < nreduce; i += 1)
         {
             mean_var += s_sum[lane_id + i * WARP_SIZE];
@@ -55,7 +55,7 @@ static void group_norm_f32(const float* x, float* dst, const int group_size, con
     const int nthreads = item_ct1.get_local_range(2);
     const int nwarps = nthreads / WARP_SIZE;
     start += item_ct1.get_local_id(2);
-    int nreduce = nwarps / WARP_SIZE;
+    size_t nreduce = nwarps / WARP_SIZE;
 
     if (end >= ne_elements) {
         end = ne_elements;
@@ -163,7 +163,7 @@ static void rms_norm_f32(const float* x, float* dst, const int ncols, const floa
         converged control flow. You may need to adjust the code.
         */
         item_ct1.barrier(sycl::access::fence_space::local_space);
-        int nreduce = nwarps / WARP_SIZE;
+        size_t nreduce = nwarps / WARP_SIZE;
         tmp = 0.f;
         for (size_t i = 0; i < nreduce; i += 1)
         {
@@ -352,6 +352,7 @@ void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, const ggml_tensor*
     (void)src1;
     (void)dst;
     (void)src1_dd;
+    GGML_UNUSED(ctx);
 }
 
 void ggml_sycl_op_rms_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0,

ggml/src/ggml-sycl/rope.cpp

@@ -269,7 +269,8 @@ void ggml_sycl_op_rope(
         }
     }
 
-    (void) src1;
+    GGML_UNUSED(src1);
-    (void) dst;
+    GGML_UNUSED(dst);
-    (void) src1_dd;
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }

ggml/src/ggml-sycl/softmax.cpp

@@ -16,7 +16,7 @@ static void soft_max_f32(const float * x, const float * mask, float * dst, const
     const int lane_id = item_ct1.get_local_id(2) % WARP_SIZE;
     const int nthreads = block_size;
     const int nwarps = nthreads / WARP_SIZE;
-    int nreduce = nwarps / WARP_SIZE;
+    size_t nreduce = nwarps / WARP_SIZE;
     float slope = 1.0f;
 
     // ALiBi
@@ -53,9 +53,10 @@ static void soft_max_f32(const float * x, const float * mask, float * dst, const
     if (block_size > WARP_SIZE) {
         if (warp_id == 0) {
             buf[lane_id] = -INFINITY;
-            for (size_t i = 1; i < nreduce; i += 1)
+            for (size_t i = 1; i < nreduce; i += 1) {
                 buf[lane_id + i * WARP_SIZE] = -INFINITY;
             }
+        }
         item_ct1.barrier(sycl::access::fence_space::local_space);
 
         if (lane_id == 0) {
@@ -63,8 +64,7 @@ static void soft_max_f32(const float * x, const float * mask, float * dst, const
         }
         item_ct1.barrier(sycl::access::fence_space::local_space);
         max_val = buf[lane_id];
-        for (size_t i = 1; i < nreduce; i += 1)
+        for (size_t i = 1; i < nreduce; i += 1) {
-        {
             max_val = std::max(max_val, buf[lane_id + i * WARP_SIZE]);
         }
         max_val = warp_reduce_max(max_val, item_ct1);
@@ -89,9 +89,10 @@ static void soft_max_f32(const float * x, const float * mask, float * dst, const
         item_ct1.barrier(sycl::access::fence_space::local_space);
         if (warp_id == 0) {
             buf[lane_id] = 0.f;
-            for (size_t i = 1; i < nreduce; i += 1)
+            for (size_t i = 1; i < nreduce; i += 1) {
                 buf[lane_id + i * WARP_SIZE] = 0.f;
             }
+        }
         item_ct1.barrier(sycl::access::fence_space::local_space);
 
         if (lane_id == 0) {
@@ -100,8 +101,7 @@ static void soft_max_f32(const float * x, const float * mask, float * dst, const
         item_ct1.barrier(sycl::access::fence_space::local_space);
 
         tmp = buf[lane_id];
-        for (size_t i = 1; i < nreduce; i += 1)
+        for (size_t i = 1; i < nreduce; i += 1) {
-        {
             tmp += buf[lane_id + i * WARP_SIZE];
         }
         tmp = warp_reduce_sum(tmp, item_ct1);

ggml/src/ggml-sycl/tsembd.cpp

@@ -68,4 +68,5 @@ void ggml_sycl_op_timestep_embedding(ggml_backend_sycl_context & ctx, const ggml
     const int max_period = dst->op_params[1];
 
     timestep_embedding_f32_sycl(src0_d, dst_d, src0->ne[0], dst->nb[1], dim, max_period, stream);
+    GGML_UNUSED(src1);
 }

ggml/src/ggml-sycl/wkv6.cpp

@@ -59,7 +59,7 @@ static void rwkv_wkv_f32_kernel(
         float y = 0;
 
         // Process in chunks of 4 for better vectorization
-        sycl::float4 k4, r4, tf4, td4, s4, kv4;
+        sycl::float4 k4, r4, tf4, td4, s4;
         #pragma unroll
         for (int j = 0; j < head_size; j += 4) {
             // Load data in vec4 chunks
@@ -135,4 +135,7 @@ void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context& ctx, const ggml_tensor* s
                 );
             });
     });
+
+    GGML_UNUSED(src0);
+    GGML_UNUSED(src1);
 }

ggml/src/ggml-threading.h

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

ggml/src/ggml-vulkan/CMakeLists.txt

@@ -81,7 +81,7 @@ if (Vulkan_FOUND)
             --target-cpp ${_ggml_vk_source}
             --no-clean
 
-        DEPENDS ${_ggml_vk_shader_deps}
+        DEPENDS ${_ggml_vk_shader_deps} ${_ggml_vk_genshaders_cmd}
         COMMENT "Generate vulkan shaders"
     )
 

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

@@ -44,12 +44,6 @@
 
 #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);                                    \
@@ -168,7 +162,12 @@ struct vk_device_struct {
     uint32_t subgroup_size;
     uint32_t shader_core_count;
     bool uma;
-    bool coopmat2;
+    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 coopmat_support;
     bool coopmat_acc_f32_support;
@@ -176,6 +175,7 @@ struct vk_device_struct {
     uint32_t coopmat_m;
     uint32_t coopmat_n;
     uint32_t coopmat_k;
+    bool coopmat2;
 
     size_t idx;
 
@@ -754,8 +754,12 @@ 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) {
+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,
-    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 << ")");
+                                         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 << ")");
     GGML_ASSERT(parameter_count > 0);
     GGML_ASSERT(wg_denoms[0] > 0 && wg_denoms[1] > 0 && wg_denoms[2] > 0); // NOLINT
 
@@ -814,14 +818,28 @@ 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(
-            vk::PipelineShaderStageCreateFlags(),
+            pipeline_shader_stage_create_flags,
             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);
 
@@ -1501,7 +1519,9 @@ 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) {
+    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) {
         {
             // wait until fewer than N compiles are in progress
             uint32_t N = std::max(1u, std::thread::hardware_concurrency());
@@ -1511,7 +1531,8 @@ 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));
+        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));
     };
 
 #if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
@@ -1617,28 +1638,33 @@ static void ggml_vk_load_shaders(vk_device& device) {
         // 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);   \
+            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);   \
         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);   \
+            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);   \
         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);   \
+            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);   \
         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);   \
+            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);   \
         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);   \
+            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);   \
         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);   \
+            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);   \
 
         // 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_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, );
@@ -1651,6 +1677,20 @@ static void ggml_vk_load_shaders(vk_device& device) {
             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, );
+        }
 
         // 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) {
@@ -1658,6 +1698,7 @@ 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);
@@ -1670,7 +1711,22 @@ static void ggml_vk_load_shaders(vk_device& device) {
                 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);
             }
+        }
+#undef CREATE_MM2
 #undef CREATE_MM
     } else if (device->fp16) {
         // Create 6 variants, {s,m,l}x{unaligned,aligned}
@@ -1688,6 +1744,11 @@ 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, );
@@ -1725,6 +1786,7 @@ 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}
@@ -1779,53 +1841,58 @@ 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
     }
 
     // mul mat vec
-    // computing two rows per workgroup is a benefit for Q4_0 -> Q5_1, but not for Q8_0.
+
+    // AMD GCN and Intel graphics cards perform best when the number of rows per shader is doubled
+    uint32_t rm = 1;
+    if ((device->vendor_id == VK_VENDOR_ID_AMD && device->subgroup_min_size == 64 && device->subgroup_max_size == 64) || device->vendor_id == VK_VENDOR_ID_INTEL)
+        rm = 2;
+
+    // computing additional rows per workgroup is a benefit for Q4_0 -> Q5_1, but not for Q8_0.
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f32_f32",  mul_mat_vec_f32_f32_f32_len,  mul_mat_vec_f32_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32_f32",  mul_mat_vec_f16_f32_f32_len,  mul_mat_vec_f16_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_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_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_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_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    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_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*rm, 1, 1}, {device->subgroup_size, 2*rm}, 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_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*rm, 1, 1}, {device->subgroup_size, 2*rm}, 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_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*rm, 1, 1}, {device->subgroup_size, 1*rm}, 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_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}, {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}, {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}, {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}, {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}, {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}, {subgroup_size_16}, 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);
+    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*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
 
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f16_f32",  mul_mat_vec_f32_f16_f32_len,  mul_mat_vec_f32_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f16_f32",  mul_mat_vec_f16_f16_f32_len,  mul_mat_vec_f16_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_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_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_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_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    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_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*rm, 1, 1}, {device->subgroup_size, 2*rm}, 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_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*rm, 1, 1}, {device->subgroup_size, 2*rm}, 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_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*rm, 1, 1}, {device->subgroup_size, 1*rm}, 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_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}, {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}, {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}, {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}, {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}, {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}, {subgroup_size_16}, 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);
+    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*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
 
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32",  mul_mat_vec_id_f32_f32_len,  mul_mat_vec_id_f32_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32",  mul_mat_vec_id_f16_f32_len,  mul_mat_vec_id_f16_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_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_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_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_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    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_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*rm, 1, 1}, {device->subgroup_size, 2*rm}, 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_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*rm, 1, 1}, {device->subgroup_size, 2*rm}, 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_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*rm, 1, 1}, {device->subgroup_size, 1*rm}, 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_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}, {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}, {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}, {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}, {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}, {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}, {subgroup_size_16}, 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);
+    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*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
 
     // dequant shaders
     ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_F32 ], "f32_to_f16",   dequant_f32_len,  dequant_f32_data,  "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
@@ -1922,17 +1989,26 @@ 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_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_f32, "rope_neox_f32", rope_neox_f32_len, rope_neox_f32_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_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_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);
 
@@ -1994,6 +2070,8 @@ 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;
@@ -2013,11 +2091,15 @@ 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 *)&driver_props;
+        VkBaseOutStructure * last_struct = (VkBaseOutStructure *)&vk12_props;
 
         if (maintenance4_support) {
             last_struct->pNext = (VkBaseOutStructure *)&props4;
@@ -2031,6 +2113,10 @@ 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;
@@ -2063,12 +2149,13 @@ 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 || (props2.properties.vendorID == VK_VENDOR_ID_AMD && driver_props.driverID == vk::DriverId::eAmdProprietary)) {
+        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))) {
             // Intel drivers don't support coopmat properly yet
             // Only RADV supports coopmat properly on AMD
             device->coopmat_support = false;
@@ -2124,6 +2211,17 @@ 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;
@@ -2151,6 +2249,20 @@ static vk_device ggml_vk_get_device(size_t idx) {
 
         device->pipeline_robustness = pl_robustness_features.pipelineRobustness;
 
+        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_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_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) {
@@ -2300,7 +2412,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
                 }
             }
 
-            if (device->coopmat_m == 0) {
+            if (device->coopmat_m == 0 || !device->coopmat_acc_f32_support) {
                 // 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;
@@ -2425,13 +2537,15 @@ static void ggml_vk_print_gpu_info(size_t idx) {
         } else if (strcmp("VK_KHR_cooperative_matrix", properties.extensionName) == 0 &&
                    !getenv("GGML_VK_DISABLE_COOPMAT")) {
             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)) {
+    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))) {
         // Intel drivers don't support coopmat properly yet
         // Only RADV supports coopmat properly on AMD
         coopmat_support = false;
@@ -2718,7 +2832,7 @@ 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) {
+    if (prec == GGML_PREC_DEFAULT && ctx->device->fp16 && !(ctx->device->coopmat_support && !ctx->device->coopmat_acc_f16_support)) {
         if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
             return ctx->device->pipeline_matmul_f16_f32.f16acc;
         }
@@ -2793,7 +2907,7 @@ 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) {
+    if (prec == GGML_PREC_DEFAULT && ctx->device->fp16 && !(ctx->device->coopmat_support && !ctx->device->coopmat_acc_f16_support)) {
         if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
             return ctx->device->pipeline_matmul_id_f16_f32.f16acc;
         }
@@ -7573,7 +7687,16 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_REPEAT:
             return ggml_type_size(op->type) == sizeof(float) && ggml_type_size(op->src[0]->type) == sizeof(float);
         case GGML_OP_ROPE:
+            {
+                const int mode = ((const int32_t *) op->op_params)[2];
+                if (mode & GGML_ROPE_TYPE_MROPE) {
+                    return false;
+                }
+                if (mode & GGML_ROPE_TYPE_VISION) {
+                    return false;
+                }
                 return ggml_is_contiguous(op->src[0]);
+            }
         case GGML_OP_NONE:
         case GGML_OP_RESHAPE:
         case GGML_OP_VIEW:

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

@@ -25,92 +25,94 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
 
 #if defined(DATA_A_Q4_0)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a[a_offset + ib].d);
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
-    return (vec2(vui & 0xF, vui >> 4) - 8.0f) * d;
+    return (vec2(vui & 0xF, vui >> 4) - 8.0f);
 }
 vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a_packed16[a_offset + ib].d);
     const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
-    return (vec4(vui & 0xF, (vui >> 4) & 0xF, (vui >> 8) & 0xF, (vui >> 12) & 0xF) - 8.0f) * d;
+    return (vec4(vui & 0xF, (vui >> 4) & 0xF, (vui >> 8) & 0xF, vui >> 12) - 8.0f);
 }
 #endif
 
 #if defined(DATA_A_Q4_1)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a[a_offset + ib].d);
-    const float m = float(data_a[a_offset + ib].m);
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
-    return vec2(vui & 0xF, vui >> 4) * d + m;
+    return vec2(vui & 0xF, vui >> 4);
 }
 vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a_packed16[a_offset + ib].d);
-    const float m = float(data_a_packed16[a_offset + ib].m);
     const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
-    return vec4(vui & 0xF, (vui >> 4) & 0xF, (vui >> 8) & 0xF, (vui >> 12) & 0xF) * d + m;
+    return vec4(vui & 0xF, (vui >> 4) & 0xF, (vui >> 8) & 0xF, vui >> 12);
 }
 #endif
 
 #if defined(DATA_A_Q5_0)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a[a_offset + ib].d);
     const uint uint_qh = uint(data_a[a_offset + ib].qh[1]) << 16 | data_a[a_offset + ib].qh[0];
     const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
-    return (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f) * d;
+    return (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f);
 }
 vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a_packed16[a_offset + ib].d);
     const uint uint_qh = uint(data_a_packed16[a_offset + ib].qh[1]) << 16 | data_a_packed16[a_offset + ib].qh[0];
     const ivec2 qh0 = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
     const ivec2 qh1 = ivec2(((uint_qh >> (iqs + 1)) << 4) & 0x10, (uint_qh >> (iqs + 13)) & 0x10);
     const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
-    return (vec4(((vui >> 0) & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, ((vui >> 12) & 0xF) | qh1.y) - 16.0f) * d;
+    return (vec4((vui & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, (vui >> 12) | qh1.y) - 16.0f);
 }
 #endif
 
 #if defined(DATA_A_Q5_1)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a[a_offset + ib].d);
-    const float m = float(data_a[a_offset + ib].m);
     const uint uint_qh = data_a[a_offset + ib].qh;
     const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
-    return vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) * d + m;
+    return vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y);
 }
 vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a_packed16[a_offset + ib].d);
-    const float m = float(data_a_packed16[a_offset + ib].m);
     const uint uint_qh = data_a_packed16[a_offset + ib].qh;
     const ivec2 qh0 = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
     const ivec2 qh1 = ivec2(((uint_qh >> (iqs + 1)) << 4) & 0x10, (uint_qh >> (iqs + 13)) & 0x10);
     const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
-    return vec4(((vui >> 0) & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, ((vui >> 12) & 0xF) | qh1.y) * d + m;
+    return vec4((vui & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, (vui >> 12) | qh1.y);
 }
 #endif
 
 #if defined(DATA_A_Q8_0)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a[a_offset + ib].d);
+    return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1]));
-    return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1])) * d;
 }
 vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a_packed16[a_offset + ib].d);
     uint32_t v0 = data_a_packed16[a_offset + ib].qs[iqs/2];
     uint32_t v1 = data_a_packed16[a_offset + ib].qs[iqs/2 + 1];
-    return vec4(int8_t(v0 & 0xFF), int8_t((v0 >> 8) & 0xFF), int8_t(v1 & 0xFF), int8_t((v1 >> 8) & 0xFF)) * d;
+    return vec4(int8_t(v0 & 0xFF), int8_t(v0 >> 8), int8_t(v1 & 0xFF), int8_t(v1 >> 8));
 }
 #endif
 
 #if defined(DATA_A_IQ4_NL)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a[a_offset + ib].d);
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
-    return vec2(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[vui >> 4]) * d;
+    return vec2(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[vui >> 4]);
 }
 vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a_packed16[a_offset + ib].d);
     const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
-    return vec4(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[(vui >> 4) & 0xF], kvalues_iq4nl[(vui >> 8) & 0xF], kvalues_iq4nl[(vui >> 12) & 0xF]) * d;
+    return vec4(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[(vui >> 4) & 0xF], kvalues_iq4nl[(vui >> 8) & 0xF], kvalues_iq4nl[vui >> 12]);
+}
+#endif
+
+#if defined(DATA_A_F32) || defined(DATA_A_F16)
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(0, 0);
+}
+#endif
+
+#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ4_NL)
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(float(data_a[a_offset + ib].d), 0);
+}
+#endif
+
+#if defined(DATA_A_Q4_1) || defined(DATA_A_Q5_1)
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(float(data_a[a_offset + ib].d), float(data_a[a_offset + ib].m));
 }
 #endif

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 i = gl_WorkGroupID.x * 256 + wgy;
+        const uint ib = gl_WorkGroupID.x * 256 + wgy;
-        if (i >= p.M * p.K / QUANT_K) {
+        if (ib >= p.M * p.K / QUANT_K) {
             return;
         }
 
@@ -20,37 +20,49 @@ void main() {
         const uint is = 2 * il;
         const uint n = 4;
 
-        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].d.y);
 
-        const uint y_idx = i * QUANT_K + 64 * il + n * ir;
+        const uint y_idx = ib * QUANT_K + 64 * il + n * ir;
         const uint qs_idx = 32*il + n * ir;
 
-        uint8_t sc;
+        uint scidx0 = (is < 4) ? is : (is + 4);
-        uint8_t m;
+        uint scidx1 = (is < 4) ? is : (is - 4);
-        if (is < 4) {
+        uint scidxmask1 = (is < 4) ? 0x30 : 0xC0;
-            sc = uint8_t(data_a[i].scales[is] & 63);
+        uint scidxshift1 = (is < 4) ? 0 : 2;
-            m  = uint8_t(data_a[i].scales[is + 4] & 63);
+        uint mbidx0 = is + 4;
-        } else {
+        uint mbidx1 = (is < 4) ? is + 4 : is;
-            sc = uint8_t((data_a[i].scales[is + 4] & 0xF) | ((data_a[i].scales[is - 4] >> 6) << 4));
+        uint mbidxmask0 = (is < 4) ? 0xF : 0xF0;
-            m  = uint8_t((data_a[i].scales[is + 4] >>  4) | ((data_a[i].scales[is    ] >> 6) << 4));
+        uint mbidxshift0 = (is < 4) ? 0 : 4;
-        }
+        uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
-        const FLOAT_TYPE d1 = dall * sc;
+        uint mbidxshift1 = (is < 4) ? 0 : 2;
-        const FLOAT_TYPE m1 = dmin * m;
+
+        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));
+
+        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));
 
-        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 * m;
+        const FLOAT_TYPE m2 = dmin * mbyte;
 
         [[unroll]] for (uint l = 0; l < n; ++l) {
-            data_b[y_idx + l     ] = D_TYPE(d1 * FLOAT_TYPE(data_a[i].qs[qs_idx + l] & 0xF) - m1);
+            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[i].qs[qs_idx + l] >>  4) - m2);
+            data_b[y_idx + l + 32] = D_TYPE(d2 * FLOAT_TYPE(data_a[ib].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 i = gl_WorkGroupID.x * 256 + wgy;
+        const uint ib = gl_WorkGroupID.x * 256 + wgy;
-        if (i >= p.M * p.K / QUANT_K) {
+        if (ib >= p.M * p.K / QUANT_K) {
             return;
         }
 
@@ -19,40 +19,52 @@ void main() {
         const uint ir = tid % 16;
         const uint is = 2 * il;
 
-        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].d.y);
 
-        const uint y_idx = i * QUANT_K + 64 * il + 2 * ir;
+        const uint y_idx = ib * QUANT_K + 64 * il + 2 * ir;
         const uint qs_idx = 32*il + 2 * ir;
         const uint qh_idx = 2 * ir;
 
-        uint8_t sc;
+        uint scidx0 = (is < 4) ? is : (is + 4);
-        uint8_t m;
+        uint scidx1 = (is < 4) ? is : (is - 4);
-        if (is < 4) {
+        uint scidxmask1 = (is < 4) ? 0x30 : 0xC0;
-            sc = uint8_t(data_a[i].scales[is] & 63);
+        uint scidxshift1 = (is < 4) ? 0 : 2;
-            m  = uint8_t(data_a[i].scales[is + 4] & 63);
+        uint mbidx0 = is + 4;
-        } else {
+        uint mbidx1 = (is < 4) ? is + 4 : is;
-            sc = uint8_t((data_a[i].scales[is + 4] & 0xF) | ((data_a[i].scales[is - 4] >> 6) << 4));
+        uint mbidxmask0 = (is < 4) ? 0xF : 0xF0;
-            m  = uint8_t((data_a[i].scales[is + 4] >>  4) | ((data_a[i].scales[is    ] >> 6) << 4));
+        uint mbidxshift0 = (is < 4) ? 0 : 4;
-        }
+        uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
-        const FLOAT_TYPE d1 = dall * sc;
+        uint mbidxshift1 = (is < 4) ? 0 : 2;
-        const FLOAT_TYPE m1 = dmin * m;
+
+        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));
+
+        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));
 
-        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 * m;
+        const FLOAT_TYPE m2 = dmin * mbyte;
 
         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[i].qs[qs_idx    ] & 0xF) + (((data_a[i].qh[qh_idx    ] & hm1) != 0) ? 16 : 0)) - m1);
+        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[i].qs[qs_idx + 1] & 0xF) + (((data_a[i].qh[qh_idx + 1] & 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[i].qs[qs_idx    ]  >> 4) + (((data_a[i].qh[qh_idx    ] & hm2) != 0) ? 16 : 0)) - m2);
+        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[i].qs[qs_idx + 1]  >> 4) + (((data_a[i].qh[qh_idx + 1] & 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);
     }
 }

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

@@ -31,6 +31,8 @@ void main() {
     const uint y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
 
     vec2 v = dequantize(ib, iqs, 0);
+    const vec2 dm = get_dm(ib, 0);
+    v = v * dm.x + dm.y;
 
     data_d[d_offset + iybs + iqs           ] = D_TYPE(v.x);
     data_d[d_offset + iybs + iqs + y_offset] = D_TYPE(v.y);

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

@@ -1,6 +1,11 @@
 #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.comp

@@ -31,27 +31,13 @@ void iter(inout FLOAT_TYPE temp[NUM_ROWS], const uint first_row, const uint num_
 
 #if K_PER_ITER == 8
 #if QUANT_R == 2
-    B_TYPE_VEC4 bv02 = data_b_v4[(b_offset + iybs + iqs) / 4];
+    const B_TYPE_VEC4 bv02 = data_b_v4[(b_offset + iybs + iqs) / 4];
-    B_TYPE_VEC4 bv13 = data_b_v4[(b_offset + iybs + iqs + y_offset) / 4];
+    const B_TYPE_VEC4 bv13 = data_b_v4[(b_offset + iybs + iqs + y_offset) / 4];
-    FLOAT_TYPE b0 = FLOAT_TYPE(bv02.x);
+    const vec4 bv0 = vec4(bv02.x, bv13.x, bv02.y, bv13.y);
-    FLOAT_TYPE b1 = FLOAT_TYPE(bv13.x);
+    const vec4 bv1 = vec4(bv02.z, bv13.z, bv02.w, bv13.w);
-    FLOAT_TYPE b2 = FLOAT_TYPE(bv02.y);
-    FLOAT_TYPE b3 = FLOAT_TYPE(bv13.y);
-    FLOAT_TYPE b4 = FLOAT_TYPE(bv02.z);
-    FLOAT_TYPE b5 = FLOAT_TYPE(bv13.z);
-    FLOAT_TYPE b6 = FLOAT_TYPE(bv02.w);
-    FLOAT_TYPE b7 = FLOAT_TYPE(bv13.w);
 #else
-    B_TYPE_VEC4 bv0 = data_b_v4[(b_offset + iybs + iqs) / 4];
+    const vec4 bv0 = vec4(data_b_v4[(b_offset + iybs + iqs) / 4]);
-    B_TYPE_VEC4 bv1 = data_b_v4[(b_offset + iybs + iqs) / 4 + 1];
+    const vec4 bv1 = vec4(data_b_v4[(b_offset + iybs + iqs) / 4 + 1]);
-    FLOAT_TYPE b0 = FLOAT_TYPE(bv0.x);
-    FLOAT_TYPE b1 = FLOAT_TYPE(bv0.y);
-    FLOAT_TYPE b2 = FLOAT_TYPE(bv0.z);
-    FLOAT_TYPE b3 = FLOAT_TYPE(bv0.w);
-    FLOAT_TYPE b4 = FLOAT_TYPE(bv1.x);
-    FLOAT_TYPE b5 = FLOAT_TYPE(bv1.y);
-    FLOAT_TYPE b6 = FLOAT_TYPE(bv1.z);
-    FLOAT_TYPE b7 = FLOAT_TYPE(bv1.w);
 #endif
 #else
     // Check if the second of the pair of elements is OOB, and don't fetch B or
@@ -67,22 +53,29 @@ void iter(inout FLOAT_TYPE temp[NUM_ROWS], const uint first_row, const uint num_
         b1 = FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]);
     }
 #endif
+    uint ibi = first_row*p.ncols;
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        const uint ib = ((first_row + n)*p.ncols + col)/QUANT_K; // block index
+        const uint ib = (ibi + col)/QUANT_K; // block index
+        ibi += p.ncols;
 
 #if K_PER_ITER == 8
-        const vec4 v = dequantize4(ib, iqs, a_offset);
+        vec4 v = dequantize4(ib, iqs, a_offset);
-        const vec4 v2 = dequantize4(ib, iqs+(4/QUANT_R), a_offset);
+        vec4 v2 = dequantize4(ib, iqs+(4/QUANT_R), a_offset);
+
+        const vec2 dm = get_dm(ib, a_offset);
+        if (dm.y != 0) { // quant has min component
+            v = v * dm.x + dm.y;
+            v2 = v2 * dm.x + dm.y;
+        }
 
         // matrix multiplication
-        temp[n] = fma(FLOAT_TYPE(v.x), b0, temp[n]);
+        FLOAT_TYPE rowtmp = dot(bv0, v);
-        temp[n] = fma(FLOAT_TYPE(v.y), b1, temp[n]);
+        rowtmp += dot(bv1, v2);
-        temp[n] = fma(FLOAT_TYPE(v.z), b2, temp[n]);
+
-        temp[n] = fma(FLOAT_TYPE(v.w), b3, temp[n]);
+        if (dm.y == 0)
-        temp[n] = fma(FLOAT_TYPE(v2.x), b4, temp[n]);
+            rowtmp *= dm.x;
-        temp[n] = fma(FLOAT_TYPE(v2.y), b5, temp[n]);
+
-        temp[n] = fma(FLOAT_TYPE(v2.z), b6, temp[n]);
+        temp[n] += rowtmp;
-        temp[n] = fma(FLOAT_TYPE(v2.w), b7, temp[n]);
 #else
         const vec2 v = dequantize(ib, iqs, a_offset);
 

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

@@ -2,8 +2,6 @@
 #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

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