local-llm-rn 1.0.0

package/cpp/ggml/src/ggml-musa/CMakeLists.txt

@@ -0,0 +1,125 @@
+if (NOT EXISTS $ENV{MUSA_PATH})
+    if (NOT EXISTS /opt/musa)
+        set(MUSA_PATH /usr/local/musa)
+    else()
+        set(MUSA_PATH /opt/musa)
+    endif()
+else()
+    set(MUSA_PATH $ENV{MUSA_PATH})
+endif()
+
+set(CMAKE_C_COMPILER "${MUSA_PATH}/bin/clang")
+set(CMAKE_C_EXTENSIONS OFF)
+set(CMAKE_CXX_COMPILER "${MUSA_PATH}/bin/clang++")
+set(CMAKE_CXX_EXTENSIONS OFF)
+
+list(APPEND CMAKE_MODULE_PATH "${MUSA_PATH}/cmake")
+
+find_package(MUSAToolkit)
+
+if (MUSAToolkit_FOUND)
+    message(STATUS "MUSA Toolkit found")
+
+    if (NOT DEFINED MUSA_ARCHITECTURES)
+        set(MUSA_ARCHITECTURES "21;22;31")
+    endif()
+    message(STATUS "Using MUSA architectures: ${MUSA_ARCHITECTURES}")
+
+    file(GLOB   GGML_HEADERS_MUSA "../ggml-cuda/*.cuh")
+    list(APPEND GGML_HEADERS_MUSA "../../include/ggml-cuda.h")
+    list(APPEND GGML_HEADERS_MUSA "../ggml-musa/mudnn.cuh")
+
+    file(GLOB   GGML_SOURCES_MUSA "../ggml-cuda/*.cu")
+    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-tile*.cu")
+    list(APPEND GGML_SOURCES_MUSA ${SRCS})
+    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-mma*.cu")
+    list(APPEND GGML_SOURCES_MUSA ${SRCS})
+    file(GLOB   SRCS "../ggml-cuda/template-instances/mmq*.cu")
+    list(APPEND GGML_SOURCES_MUSA ${SRCS})
+
+    if (GGML_MUSA_MUDNN_COPY)
+        file(GLOB   SRCS "../ggml-musa/*.cu")
+        list(APPEND GGML_SOURCES_MUSA ${SRCS})
+        add_compile_definitions(GGML_MUSA_MUDNN_COPY)
+    endif()
+
+    if (GGML_CUDA_FA_ALL_QUANTS)
+        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*.cu")
+        list(APPEND GGML_SOURCES_MUSA ${SRCS})
+        add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
+    else()
+        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu")
+        list(APPEND GGML_SOURCES_MUSA ${SRCS})
+        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu")
+        list(APPEND GGML_SOURCES_MUSA ${SRCS})
+        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*f16-f16.cu")
+        list(APPEND GGML_SOURCES_MUSA ${SRCS})
+    endif()
+
+    set_source_files_properties(${GGML_SOURCES_MUSA} PROPERTIES LANGUAGE CXX)
+    foreach(SOURCE ${GGML_SOURCES_MUSA})
+        set(COMPILE_FLAGS "-Od3 -fno-strict-aliasing -ffast-math -fsigned-char -x musa -mtgpu -fmusa-flush-denormals-to-zero")
+        foreach(ARCH ${MUSA_ARCHITECTURES})
+            set(COMPILE_FLAGS "${COMPILE_FLAGS} --cuda-gpu-arch=mp_${ARCH}")
+        endforeach()
+        set_property(SOURCE ${SOURCE} PROPERTY COMPILE_FLAGS ${COMPILE_FLAGS})
+    endforeach()
+
+    ggml_add_backend_library(ggml-musa
+                             ${GGML_HEADERS_MUSA}
+                             ${GGML_SOURCES_MUSA}
+                            )
+
+    # TODO: do not use CUDA definitions for MUSA
+    if (NOT GGML_BACKEND_DL)
+        target_compile_definitions(ggml PUBLIC GGML_USE_CUDA)
+    endif()
+
+    add_compile_definitions(GGML_USE_MUSA)
+    add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
+
+    if (GGML_MUSA_GRAPHS)
+        add_compile_definitions(GGML_MUSA_GRAPHS)
+    endif()
+
+    if (GGML_CUDA_FORCE_MMQ)
+        add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+    endif()
+
+    if (GGML_CUDA_FORCE_CUBLAS)
+        add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
+    endif()
+
+    if (GGML_CUDA_NO_VMM)
+        add_compile_definitions(GGML_CUDA_NO_VMM)
+    endif()
+
+    if (NOT GGML_CUDA_FA)
+        add_compile_definitions(GGML_CUDA_NO_FA)
+    endif()
+
+    if (GGML_CUDA_NO_PEER_COPY)
+        add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
+    endif()
+
+    if (GGML_STATIC)
+        target_link_libraries(ggml-musa PRIVATE MUSA::musart_static MUSA::mublas_static)
+        # TODO: mudnn has not provided static libraries yet
+        # if (GGML_MUSA_MUDNN_COPY)
+        #     target_link_libraries(ggml-musa PRIVATE mudnn_static)
+        # endif()
+    else()
+        target_link_libraries(ggml-musa PRIVATE MUSA::musart MUSA::mublas)
+        if (GGML_MUSA_MUDNN_COPY)
+            target_link_libraries(ggml-musa PRIVATE mudnn)
+        endif()
+    endif()
+
+    if (GGML_CUDA_NO_VMM)
+        # No VMM requested, no need to link directly with the musa driver lib (libmusa.so)
+    else()
+        target_link_libraries(ggml-musa PRIVATE MUSA::musa_driver)
+    endif()
+else()
+    message(FATAL_ERROR "MUSA Toolkit not found")
+endif()

package/cpp/ggml/src/ggml-musa/mudnn.cu

@@ -0,0 +1,112 @@
+#include <mutex>
+#include <mudnn.h>
+
+#include "mudnn.cuh"
+
+namespace mudnn = musa::dnn;
+
+// Returns a human-readable error string for mudnn::Status
+const char* mudnnGetErrorString(mudnn::Status err) {
+    switch (err) {
+        case mudnn::Status::SUCCESS:
+            return "Success";
+        case mudnn::Status::INVALID_PARAMETER:
+            return "Invalid parameter";
+        case mudnn::Status::NOT_INITIALIZED:
+            return "Not initialized";
+        case mudnn::Status::ALLOC_FAILED:
+            return "Allocation failed";
+        case mudnn::Status::NOT_SUPPORTED:
+            return "Not supported";
+        case mudnn::Status::INTERNAL_ERROR:
+            return "Internal error";
+        case mudnn::Status::ARCH_MISMATCH:
+            return "Architecture mismatch";
+        case mudnn::Status::EXECUTION_FAILED:
+            return "Execution failed";
+        default:
+            return "Unknown mudnn status";
+    }
+}
+
+// Error checking macro for MUDNN calls
+#define MUDNN_CHECK(err) CUDA_CHECK_GEN(err, mudnn::Status::SUCCESS, mudnnGetErrorString)
+
+namespace {
+    // Thread-safe cache for mudnn::Handle objects per device
+    std::unordered_map<int, std::unique_ptr<mudnn::Handle>> handle_cache;
+    std::mutex handle_cache_mutex;
+
+    mudnn::Handle* get_cached_handle(int device_id) {
+        std::lock_guard<std::mutex> lock(handle_cache_mutex);
+        auto it = handle_cache.find(device_id);
+        if (it != handle_cache.end()) {
+            return it->second.get();
+        }
+        auto handle = std::make_unique<mudnn::Handle>(device_id);
+        mudnn::Handle* handle_ptr = handle.get();
+        handle_cache[device_id] = std::move(handle);
+        return handle_ptr;
+    }
+}
+
+// Extracts dimensions and strides from a ggml_tensor
+int get_ggml_dims_and_strides(const ggml_tensor* tensor,
+                              std::vector<int64_t>& dims,
+                              std::vector<int64_t>& strides) {
+    const int ndims = ggml_n_dims(tensor);
+    const size_t element_size = ggml_element_size(tensor);
+
+    dims.resize(ndims);
+    strides.resize(ndims);
+
+    for (int i = 0; i < ndims; ++i) {
+        dims[i] = tensor->ne[i];
+        strides[i] = tensor->nb[i] / static_cast<int64_t>(element_size);
+    }
+    return ndims;
+}
+
+// Converts ggml_type to mudnn::Tensor::Type
+mudnn::Tensor::Type ggml_type_to_mudnn_type(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_F32:
+            return mudnn::Tensor::Type::FLOAT;
+        case GGML_TYPE_F16:
+            return mudnn::Tensor::Type::HALF;
+
+        // TODO: Add support for other types
+
+        default:
+            MUDNN_CHECK(mudnn::Status::NOT_SUPPORTED);
+    }
+
+    return mudnn::Tensor::Type::FLOAT; // Default fallback
+}
+
+// Asynchronous memory copy using mudnn::Unary::IDENTITY
+musaError_t mudnnMemcpyAsync(ggml_backend_cuda_context& ctx, const ggml_tensor* dst, const ggml_tensor* src) {
+    mudnn::Tensor tensor_dst, tensor_src;
+
+    MUDNN_CHECK(tensor_dst.SetType(ggml_type_to_mudnn_type(dst->type)));
+    MUDNN_CHECK(tensor_src.SetType(ggml_type_to_mudnn_type(src->type)));
+
+    std::vector<int64_t> dims, strides;
+    const int ndims = get_ggml_dims_and_strides(src, dims, strides);
+
+    MUDNN_CHECK(tensor_dst.SetNdInfo(ndims, dims.data(), strides.data()));
+    MUDNN_CHECK(tensor_src.SetNdInfo(ndims, dims.data(), strides.data()));
+    MUDNN_CHECK(tensor_dst.SetAddr(dst->data));
+    MUDNN_CHECK(tensor_src.SetAddr(src->data));
+
+    mudnn::Unary op;
+    MUDNN_CHECK(op.SetMode(mudnn::Unary::Mode::IDENTITY));
+    MUDNN_CHECK(op.SetAlpha(0.0f));
+    MUDNN_CHECK(op.SetBeta(0.0f));
+
+    mudnn::Handle* handle = get_cached_handle(ctx.device);
+    MUDNN_CHECK(handle->SetStream(ctx.stream()));
+    MUDNN_CHECK(op.Run(*handle, tensor_dst, tensor_src));
+
+    return musaSuccess;
+}

package/cpp/ggml/src/ggml-musa/mudnn.cuh

@@ -0,0 +1,12 @@
+#pragma once
+
+#include "ggml-cuda/common.cuh"
+#include "ggml.h"
+
+// Asynchronously copies data from src tensor to dst tensor using the provided context.
+// Returns a musaError_t indicating success or failure.
+musaError_t mudnnMemcpyAsync(
+    ggml_backend_cuda_context &ctx,
+    const ggml_tensor *dst,
+    const ggml_tensor *src
+);

package/cpp/ggml/src/ggml-opencl/CMakeLists.txt

@@ -0,0 +1,150 @@
+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 ${OpenCL_INCLUDE_DIRS})
+
+if (GGML_OPENCL_PROFILING)
+    message(STATUS "OpenCL profiling enabled (increases CPU overhead)")
+    add_compile_definitions(GGML_OPENCL_PROFILING)
+endif ()
+
+add_compile_definitions(GGML_OPENCL_SOA_Q)
+add_compile_definitions(GGML_OPENCL_TARGET_VERSION=${GGML_OPENCL_TARGET_VERSION})
+
+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(EMBED_KERNEL_SCRIPT "${CMAKE_CURRENT_SOURCE_DIR}/kernels/embed_kernel.py")
+    file(MAKE_DIRECTORY     "${CMAKE_CURRENT_BINARY_DIR}/autogenerated")
+
+    target_include_directories(${TARGET_NAME} PRIVATE "${CMAKE_CURRENT_BINARY_DIR}/autogenerated")
+endif ()
+
+function(ggml_opencl_add_kernel KNAME)
+    set(KERN_HDR ${CMAKE_CURRENT_BINARY_DIR}/autogenerated/${KNAME}.cl.h)
+    set(KERN_SRC ${CMAKE_CURRENT_SOURCE_DIR}/kernels/${KNAME}.cl)
+
+    if (GGML_OPENCL_EMBED_KERNELS)
+        message(STATUS "opencl: embedding kernel ${KNAME}")
+
+        # Python must be accessible from command line
+        add_custom_command(
+            OUTPUT ${KERN_HDR}
+            COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT} ${KERN_SRC} ${KERN_HDR}
+            DEPENDS ${KERN_SRC} ${EMBED_KERNEL_SCRIPT}
+            COMMENT "Generate ${KERN_HDR}"
+        )
+
+        target_sources(${TARGET_NAME} PRIVATE ${KERN_HDR})
+    else ()
+        message(STATUS "opencl: adding kernel ${KNAME}")
+        configure_file(${KERN_SRC} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${KNAME}.cl COPYONLY)
+    endif ()
+endfunction()
+
+set(GGML_OPENCL_KERNELS
+    add
+    add_id
+    argsort
+    tri
+    fill
+    clamp
+    cpy
+    cvt
+    diag_mask_inf
+    div
+    gelu
+    gemv_noshuffle_general
+    gemv_noshuffle
+    get_rows
+    glu
+    group_norm
+    solve_tri
+    im2col_f32
+    im2col_f16
+    mean
+    mul_mat_Ab_Bi_8x4
+    mul_mv_f16_f16
+    mul_mv_f16_f32_1row
+    mul_mv_f16_f32_l4
+    mul_mv_f16_f32
+    mul_mv_f32_f32
+    mul_mv_q4_0_f32
+    mul_mv_q4_0_f32_v
+    mul_mv_q4_0_f32_8x_flat
+    mul_mv_q4_0_f32_1d_8x_flat
+    mul_mv_q4_0_f32_1d_16x_flat
+    mul_mv_q4_1_f32
+    mul_mv_q4_1_f32_flat
+    mul_mv_q4_k_f32
+    mul_mv_q6_k_f32
+    mul_mv_q6_k_f32_flat
+    mul_mv_q8_0_f32
+    mul_mv_q8_0_f32_flat
+    mul_mv_mxfp4_f32
+    mul_mv_mxfp4_f32_flat
+    mul_mv_id_q4_0_f32_8x_flat
+    mul_mv_id_q8_0_f32
+    mul_mv_id_q8_0_f32_flat
+    mul_mv_id_mxfp4_f32
+    mul_mv_id_mxfp4_f32_flat
+    gemm_moe_mxfp4_f32
+    gemv_moe_mxfp4_f32
+    mul_mm_f32_f32_l4_lm
+    mul_mm_f16_f32_l4_lm
+    mul_mm_q4_0_f32_l4_lm
+    mul_mm_q4_1_f32_l4_lm
+    mul_mm_q8_0_f32_l4_lm
+    mul_mm_q6_k_f32_l4_lm
+    mul_mm_q8_0_f32_8x4
+    gemv_noshuffle_general_q8_0_f32
+    mul
+    norm
+    relu
+    rms_norm
+    rope
+    scale
+    set_rows
+    sigmoid
+    silu
+    softmax_4_f32
+    softmax_4_f16
+    softmax_f32
+    softmax_f16
+    sqr
+    sqrt
+    ssm_conv
+    sub
+    sum_rows
+    transpose
+    concat
+    tsembd
+    upscale
+    tanh
+    expm1
+    softplus
+    pad
+    repeat
+    mul_mat_f16_f32
+    mul_mm_f16_f32_kq_kqv
+    conv2d
+    conv2d_f16_f32
+    flash_attn_f32_f16
+    flash_attn_f16
+    flash_attn_f32
+)
+
+foreach (K ${GGML_OPENCL_KERNELS})
+    ggml_opencl_add_kernel(${K})
+endforeach()