Using dp4a ptx intrinsics for an improved Mul8MAT perf [By Alcpz]

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

.github/workflows/build.yml

@@ -860,7 +860,8 @@ jobs:
           mkdir build
           cd build
           cmake .. -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON
-          cmake --build . --config Release -j $((${env:NUMBER_OF_PROCESSORS} - 1))
+          cmake --build . --config Release -j $((${env:NUMBER_OF_PROCESSORS} - 1)) -t ggml
+          cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}
 
       - name: Determine tag name
         id: tag

.gitignore

@@ -50,6 +50,7 @@ build*
 !docs/build.md
 /libllama.so
 /llama-*
+/vulkan-shaders-gen
 android-ndk-*
 arm_neon.h
 cmake-build-*

Makefile

@@ -528,10 +528,21 @@ ifndef GGML_NO_ACCELERATE
 	endif
 endif # GGML_NO_ACCELERATE
 
+ifdef GGML_MUSA
+	CC := clang
+	CXX := clang++
+	GGML_CUDA := 1
+	MK_CPPFLAGS += -DGGML_USE_MUSA
+endif
+
 ifndef GGML_NO_OPENMP
 	MK_CPPFLAGS += -DGGML_USE_OPENMP
 	MK_CFLAGS   += -fopenmp
 	MK_CXXFLAGS += -fopenmp
+	ifdef GGML_MUSA
+		MK_CPPFLAGS += -I/usr/lib/llvm-10/include/openmp
+		MK_LDFLAGS  += -L/usr/lib/llvm-10/lib
+	endif # GGML_MUSA
 endif # GGML_NO_OPENMP
 
 ifdef GGML_OPENBLAS
@@ -582,15 +593,27 @@ else
 endif # GGML_CUDA_FA_ALL_QUANTS
 
 ifdef GGML_CUDA
-	ifneq ('', '$(wildcard /opt/cuda)')
-		CUDA_PATH ?= /opt/cuda
-	else
-		CUDA_PATH ?= /usr/local/cuda
-	endif
+	ifdef GGML_MUSA
+		ifneq ('', '$(wildcard /opt/musa)')
+			CUDA_PATH ?= /opt/musa
+		else
+			CUDA_PATH ?= /usr/local/musa
+		endif
 
-	MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS
-	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
-	MK_NVCCFLAGS += -use_fast_math
+		MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include
+		MK_LDFLAGS   += -lmusa -lmublas -lmusart -lpthread -ldl -lrt -L$(CUDA_PATH)/lib -L/usr/lib64
+		MK_NVCCFLAGS += -x musa -mtgpu --cuda-gpu-arch=mp_22
+	else
+		ifneq ('', '$(wildcard /opt/cuda)')
+			CUDA_PATH ?= /opt/cuda
+		else
+			CUDA_PATH ?= /usr/local/cuda
+		endif
+
+		MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS
+		MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
+		MK_NVCCFLAGS += -use_fast_math
+	endif # GGML_MUSA
 
 	OBJ_GGML += ggml/src/ggml-cuda.o
 	OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
@@ -600,9 +623,11 @@ ifdef LLAMA_FATAL_WARNINGS
 	MK_NVCCFLAGS += -Werror all-warnings
 endif # LLAMA_FATAL_WARNINGS
 
+ifndef GGML_MUSA
 ifndef JETSON_EOL_MODULE_DETECT
 	MK_NVCCFLAGS += --forward-unknown-to-host-compiler
 endif # JETSON_EOL_MODULE_DETECT
+endif # GGML_MUSA
 
 ifdef LLAMA_DEBUG
 	MK_NVCCFLAGS += -lineinfo
@@ -615,8 +640,12 @@ endif # GGML_CUDA_DEBUG
 ifdef GGML_CUDA_NVCC
 	NVCC = $(CCACHE) $(GGML_CUDA_NVCC)
 else
-	NVCC = $(CCACHE) nvcc
-endif #GGML_CUDA_NVCC
+	ifdef GGML_MUSA
+		NVCC = $(CCACHE) mcc
+	else
+		NVCC = $(CCACHE) nvcc
+	endif # GGML_MUSA
+endif # GGML_CUDA_NVCC
 
 ifdef CUDA_DOCKER_ARCH
 	MK_NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
@@ -687,9 +716,15 @@ define NVCC_COMPILE
 	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUDA -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endef # NVCC_COMPILE
 else
+	ifdef GGML_MUSA
+define NVCC_COMPILE
+	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -c $< -o $@
+endef # NVCC_COMPILE
+	else
 define NVCC_COMPILE
 	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endef # NVCC_COMPILE
+	endif # GGML_MUSA
 endif # JETSON_EOL_MODULE_DETECT
 
 ggml/src/ggml-cuda/%.o: \
@@ -944,6 +979,7 @@ $(info I CXX:       $(shell $(CXX)  --version | head -n 1))
 ifdef GGML_CUDA
 $(info I NVCC:      $(shell $(NVCC) --version | tail -n 1))
 CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])')
+ifndef GGML_MUSA
 ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
 
 ifndef CUDA_DOCKER_ARCH
@@ -953,6 +989,7 @@ endif # CUDA_POWER_ARCH
 endif # CUDA_DOCKER_ARCH
 
 endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
+endif # GGML_MUSA
 endif # GGML_CUDA
 $(info )
 

README.md

@@ -409,6 +409,7 @@ Please refer to [Build llama.cpp locally](./docs/build.md)
 | [BLAS](./docs/build.md#blas-build) | All |
 | [BLIS](./docs/backend/BLIS.md) | All |
 | [SYCL](./docs/backend/SYCL.md) | Intel and Nvidia GPU |
+| [MUSA](./docs/build.md#musa) | Moore Threads GPU |
 | [CUDA](./docs/build.md#cuda) | Nvidia GPU |
 | [hipBLAS](./docs/build.md#hipblas) | AMD GPU |
 | [Vulkan](./docs/build.md#vulkan) | GPU |

docs/build.md

@@ -192,6 +192,19 @@ The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/c
 | GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                         |
 | GGML_CUDA_FA_ALL_QUANTS       | Boolean                | false   | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer.                                                                                                  |
 
+### MUSA
+
+- Using `make`:
+  ```bash
+  make GGML_MUSA=1
+  ```
+- Using `CMake`:
+
+  ```bash
+  cmake -B build -DGGML_MUSA=ON
+  cmake --build build --config Release
+  ```
+
 ### hipBLAS
 
 This provides BLAS acceleration on HIP-supported AMD GPUs.

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

@@ -47,7 +47,7 @@ int main(int argc, char ** argv) {
     // save state (rng, logits, embedding and kv_cache) to file
     {
         std::vector<uint8_t> state_mem(llama_state_get_size(ctx));
-        const size_t written = llama_state_get_data(ctx, state_mem.data());
+        const size_t written = llama_state_get_data(ctx, state_mem.data(), state_mem.size());
 
         FILE *fp_write = fopen("dump_state.bin", "wb");
         fwrite(state_mem.data(), 1, written, fp_write);
@@ -99,13 +99,16 @@ int main(int argc, char ** argv) {
 
     // load state (rng, logits, embedding and kv_cache) from file
     {
-        std::vector<uint8_t> state_mem(llama_state_get_size(ctx2));
+        std::vector<uint8_t> state_mem;
 
         FILE * fp_read = fopen("dump_state.bin", "rb");
+        fseek(fp_read, 0, SEEK_END);
+        state_mem.resize(ftell(fp_read));
+        fseek(fp_read, 0, SEEK_SET);
         const size_t read = fread(state_mem.data(), 1, state_mem.size(), fp_read);
         fclose(fp_read);
 
-        if (read != llama_state_set_data(ctx2, state_mem.data())) {
+        if (read != llama_state_set_data(ctx2, state_mem.data(), state_mem.size())) {
             fprintf(stderr, "\n%s : failed to read state\n", __func__);
             llama_free(ctx2);
             llama_free_model(model);
@@ -159,13 +162,16 @@ int main(int argc, char ** argv) {
 
     // load state (rng, logits, embedding and kv_cache) from file
     {
-        std::vector<uint8_t> state_mem(llama_state_get_size(ctx3));
+        std::vector<uint8_t> state_mem;
 
         FILE * fp_read = fopen("dump_state.bin", "rb");
+        fseek(fp_read, 0, SEEK_END);
+        state_mem.resize(ftell(fp_read));
+        fseek(fp_read, 0, SEEK_SET);
         const size_t read = fread(state_mem.data(), 1, state_mem.size(), fp_read);
         fclose(fp_read);
 
-        if (read != llama_state_set_data(ctx3, state_mem.data())) {
+        if (read != llama_state_set_data(ctx3, state_mem.data(), state_mem.size())) {
             fprintf(stderr, "\n%s : failed to read state\n", __func__);
             llama_free(ctx3);
             llama_free_model(model);
@@ -182,7 +188,7 @@ int main(int argc, char ** argv) {
     {
         // save kv of seq 0
         std::vector<uint8_t> seq_store(llama_state_seq_get_size(ctx3, 0));
-        const size_t ncopy = llama_state_seq_get_data(ctx3, seq_store.data(), 0);
+        const size_t ncopy = llama_state_seq_get_data(ctx3, seq_store.data(), seq_store.size(), 0);
         if (ncopy != seq_store.size()) {
             fprintf(stderr, "\n%s : seq copy data length %zd does not match expected length %zd\n", __func__, ncopy, seq_store.size());
             llama_free(ctx3);
@@ -196,7 +202,7 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s : kv cache cleared\n", __func__);
 
         // restore kv into seq 1
-        const size_t nset = llama_state_seq_set_data(ctx3, seq_store.data(), 1);
+        const size_t nset = llama_state_seq_set_data(ctx3, seq_store.data(), seq_store.size(), 1);
         if (nset != seq_store.size()) {
             fprintf(stderr, "\n%s : seq set data length %zd does not match expected length %zd\n", __func__, nset, seq_store.size());
             llama_free(ctx3);

ggml/CMakeLists.txt

@@ -113,6 +113,7 @@ set(GGML_BLAS_VENDOR ${GGML_BLAS_VENDOR_DEFAULT} CACHE STRING
 option(GGML_LLAMAFILE                       "ggml: use LLAMAFILE"                             OFF)
 
 option(GGML_CUDA                            "ggml: use CUDA"                                  OFF)
+option(GGML_MUSA                            "ggml: use MUSA"                                  OFF)
 option(GGML_CUDA_FORCE_DMMV                 "ggml: use dmmv instead of mmvq CUDA kernels"     OFF)
 option(GGML_CUDA_FORCE_MMQ                  "ggml: use mmq kernels instead of cuBLAS"         OFF)
 option(GGML_CUDA_FORCE_CUBLAS               "ggml: always use cuBLAS instead of mmq kernels"  OFF)

ggml/include/ggml-cuda.h

@@ -6,6 +6,9 @@
 #ifdef GGML_USE_HIPBLAS
 #define GGML_CUDA_NAME "ROCm"
 #define GGML_CUBLAS_NAME "hipBLAS"
+#elif defined(GGML_USE_MUSA)
+#define GGML_CUDA_NAME "MUSA"
+#define GGML_CUBLAS_NAME "muBLAS"
 #else
 #define GGML_CUDA_NAME "CUDA"
 #define GGML_CUBLAS_NAME "cuBLAS"

ggml/src/CMakeLists.txt

@@ -139,6 +139,17 @@ if (GGML_METAL)
         )
 endif()
 
+if (GGML_MUSA)
+    set(CMAKE_C_COMPILER clang)
+    set(CMAKE_C_EXTENSIONS OFF)
+    set(CMAKE_CXX_COMPILER clang++)
+    set(CMAKE_CXX_EXTENSIONS OFF)
+
+    set(GGML_CUDA ON)
+
+    list(APPEND GGML_CDEF_PUBLIC GGML_USE_MUSA)
+endif()
+
 if (GGML_OPENMP)
     find_package(OpenMP)
     if (OpenMP_FOUND)
@@ -147,6 +158,11 @@ if (GGML_OPENMP)
         add_compile_definitions(GGML_USE_OPENMP)
 
         set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
+
+        if (GGML_MUSA)
+            set(GGML_EXTRA_INCLUDES ${GGML_EXTRA_INCLUDES} "/usr/lib/llvm-10/include/openmp")
+            set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} "/usr/lib/llvm-10/lib/libomp.so")
+        endif()
     else()
         message(WARNING "OpenMP not found")
     endif()
@@ -249,7 +265,13 @@ endif()
 if (GGML_CUDA)
     cmake_minimum_required(VERSION 3.18)  # for CMAKE_CUDA_ARCHITECTURES
 
-    find_package(CUDAToolkit)
+    if (GGML_MUSA)
+        list(APPEND CMAKE_MODULE_PATH "/usr/local/musa/cmake/")
+        find_package(MUSAToolkit)
+        set(CUDAToolkit_FOUND ${MUSAToolkit_FOUND})
+    else()
+        find_package(CUDAToolkit)
+    endif()
 
     if (CUDAToolkit_FOUND)
         message(STATUS "CUDA found")
@@ -268,7 +290,11 @@ if (GGML_CUDA)
         endif()
         message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
 
-        enable_language(CUDA)
+        if (GGML_MUSA)
+            set(CMAKE_CUDA_COMPILER ${MUSAToolkit_MCC_EXECUTABLE})
+        else()
+            enable_language(CUDA)
+        endif()
 
         file(GLOB   GGML_HEADERS_CUDA "ggml-cuda/*.cuh")
         list(APPEND GGML_HEADERS_CUDA "../include/ggml-cuda.h")
@@ -332,21 +358,40 @@ if (GGML_CUDA)
             add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
         endif()
 
+        if (GGML_MUSA)
+            set_source_files_properties(${GGML_SOURCES_CUDA} PROPERTIES LANGUAGE CXX)
+            foreach(SOURCE ${GGML_SOURCES_CUDA})
+                set_property(SOURCE ${SOURCE} PROPERTY COMPILE_FLAGS "-x musa -mtgpu --cuda-gpu-arch=mp_22")
+            endforeach()
+        endif()
+
         if (GGML_STATIC)
             if (WIN32)
                 # As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
                 set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
             else ()
-                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+                if (GGML_MUSA)
+                    set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} MUSA::musart_static MUSA::mublas_static)
+                else()
+                    set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+                endif()
             endif()
         else()
-            set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
+            if (GGML_MUSA)
+                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} MUSA::musart MUSA::mublas)
+            else()
+                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
+            endif()
         endif()
 
         if (GGML_CUDA_NO_VMM)
             # No VMM requested, no need to link directly with the cuda driver lib (libcuda.so)
         else()
-            set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cuda_driver) # required by cuDeviceGetAttribute(), cuMemGetAllocationGranularity(...), ...
+            if (GGML_MUSA)
+                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} MUSA::musa_driver) # required by muDeviceGetAttribute(), muMemGetAllocationGranularity(...), ...
+            else()
+                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cuda_driver) # required by cuDeviceGetAttribute(), cuMemGetAllocationGranularity(...), ...
+            endif()
         endif()
     else()
         message(WARNING "CUDA not found")
@@ -857,8 +902,10 @@ function(get_flags CCID CCVER)
         set(C_FLAGS   -Wdouble-promotion)
         set(CXX_FLAGS -Wno-array-bounds)
 
-        if (CCVER VERSION_GREATER_EQUAL 7.1.0)
-            list(APPEND CXX_FLAGS -Wno-format-truncation)
+        if (NOT GGML_MUSA)
+            if (CCVER VERSION_GREATER_EQUAL 7.1.0)
+                list(APPEND CXX_FLAGS -Wno-format-truncation)
+            endif()
         endif()
         if (CCVER VERSION_GREATER_EQUAL 8.1.0)
             list(APPEND CXX_FLAGS -Wextra-semi)
@@ -1264,6 +1311,7 @@ endif()
 target_compile_definitions(ggml PUBLIC  ${GGML_CDEF_PUBLIC})
 target_include_directories(ggml PUBLIC ../include)
 target_include_directories(ggml PRIVATE . ${GGML_EXTRA_INCLUDES})
+target_link_directories(ggml PRIVATE ${GGML_EXTRA_LIBDIRS})
 target_compile_features   (ggml PRIVATE c_std_11) # don't bump
 
 target_link_libraries(ggml PRIVATE Threads::Threads ${GGML_EXTRA_LIBS})

ggml/src/ggml-common.h

@@ -19,7 +19,11 @@ typedef half2 ggml_half2;
 
 #define GGML_COMMON_DECL
 #elif defined(GGML_COMMON_DECL_CUDA)
+#if defined(GGML_COMMON_DECL_MUSA)
+#include <musa_fp16.h>
+#else
 #include <cuda_fp16.h>
+#endif
 #include <cstdint>
 
 typedef half  ggml_half;
@@ -415,7 +419,7 @@ static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_
 #define GGML_TABLE_END() };
 
 #define GGML_COMMON_IMPL
-#elif defined(GGML_COMMON_IMPL_CUDA) || defined(GGML_COMMON_IMPL_HIP)
+#elif defined(GGML_COMMON_IMPL_CUDA) || defined(GGML_COMMON_IMPL_HIP) || defined(GGML_COMMON_IMPL_MUSA)
 #include <cstdint>
 
 #define GGML_TABLE_BEGIN(type, name, size) static const __device__ type name[size] = {

ggml/src/ggml-cuda.cu

@@ -167,7 +167,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
     for (int id = 0; id < info.device_count; ++id) {
         int device_vmm = 0;
 
-#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM) && !defined(GGML_USE_MUSA)
         CUdevice device;
         CU_CHECK(cuDeviceGet(&device, id));
         CU_CHECK(cuDeviceGetAttribute(&device_vmm, CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED, device));
@@ -179,7 +179,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
             alloc_prop.location.id = id;
             CU_CHECK(cuMemGetAllocationGranularity(&info.devices[id].vmm_granularity, &alloc_prop, CU_MEM_ALLOC_GRANULARITY_RECOMMENDED));
         }
-#endif // !defined(GGML_USE_HIPBLAS)
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM) && !defined(GGML_USE_MUSA)
         info.devices[id].vmm = !!device_vmm;
 
         cudaDeviceProp prop;
@@ -315,7 +315,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
 };
 
 // pool with virtual memory
-#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM) && !defined(GGML_USE_MUSA)
 struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
     static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB
 
@@ -409,14 +409,14 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
         GGML_ASSERT(ptr == (void *) (pool_addr + pool_used));
     }
 };
-#endif // !defined(GGML_USE_HIPBLAS)
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM) && !defined(GGML_USE_MUSA)
 
 std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int device) {
-#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM) && !defined(GGML_USE_MUSA)
     if (ggml_cuda_info().devices[device].vmm) {
         return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_vmm(device));
     }
-#endif
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM) && !defined(GGML_USE_MUSA)
     return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_leg(device));
 }
 
@@ -1341,7 +1341,7 @@ static void ggml_cuda_set_peer_access(const int n_tokens, int main_device) {
 static cudaError_t ggml_cuda_Memcpy2DPeerAsync(
     void * dst, int dstDevice, size_t dpitch, void * src, int srcDevice, size_t spitch, size_t width, size_t height, cudaStream_t stream) {
 
-#if !defined(GGML_USE_HIPBLAS)
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
     // cudaMemcpy2DAsync may fail with copies between vmm pools of different devices
     cudaMemcpy3DPeerParms p = {};
     p.dstDevice = dstDevice;
@@ -1355,7 +1355,7 @@ static cudaError_t ggml_cuda_Memcpy2DPeerAsync(
     GGML_UNUSED(dstDevice);
     GGML_UNUSED(srcDevice);
     return cudaMemcpy2DAsync(dst, dpitch, src, spitch, width, height, cudaMemcpyDeviceToDevice, stream);
-#endif // !defined(GGML_USE_HIPBLAS)
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
 }
 
 static void ggml_cuda_op_mul_mat(
@@ -1828,6 +1828,9 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
         }
     }
 #else
+#ifdef GGML_USE_MUSA
+    GGML_ASSERT(false);
+#else // !GGML_USE_MUSA
     if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
         // there is no broadcast and src0, src1 are contiguous across dims 2, 3
         // use cublasGemmStridedBatchedEx
@@ -1870,6 +1873,7 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
                 cu_compute_type,
                 CUBLAS_GEMM_DEFAULT_TENSOR_OP));
     }
+#endif // GGML_USE_MUSA
 #endif
 
     if (dst->op_params[0] == GGML_PREC_DEFAULT) {
@@ -3027,7 +3031,7 @@ GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size
         return false;
     }
 
-#if CUDART_VERSION >= 11100
+#if CUDART_VERSION >= 11100 || defined(GGML_USE_MUSA)
     cudaError_t err = cudaHostRegister(buffer, size, cudaHostRegisterPortable | cudaHostRegisterReadOnly);
     if (err != cudaSuccess) {
         // clear the error

ggml/src/ggml-cuda/common.cuh

@@ -12,6 +12,10 @@
 #else
 #define GGML_COMMON_DECL_CUDA
 #define GGML_COMMON_IMPL_CUDA
+#if defined(GGML_USE_MUSA)
+#define GGML_COMMON_DECL_MUSA
+#define GGML_COMMON_IMPL_MUSA
+#endif
 #endif
 #include "ggml-common.h"
 
@@ -23,111 +27,11 @@
 #include <vector>
 
 #if defined(GGML_USE_HIPBLAS)
-#include <hip/hip_runtime.h>
-#include <hipblas/hipblas.h>
-#include <hip/hip_fp16.h>
-#ifdef __HIP_PLATFORM_AMD__
-// for rocblas_initialize()
-#include "rocblas/rocblas.h"
-#endif // __HIP_PLATFORM_AMD__
-#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F
-#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
-#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
-#define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
-#define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
-#define CUBLAS_OP_N HIPBLAS_OP_N
-#define CUBLAS_OP_T HIPBLAS_OP_T
-#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
-#define CUBLAS_TF32_TENSOR_OP_MATH 0
-#define CUDA_R_16F  HIPBLAS_R_16F
-#define CUDA_R_32F  HIPBLAS_R_32F
-#define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
-#define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6
-#define cublasCreate hipblasCreate
-#define cublasDestroy hipblasDestroy
-#define cublasGemmEx hipblasGemmEx
-#define cublasGemmBatchedEx hipblasGemmBatchedEx
-#define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx
-#define cublasHandle_t hipblasHandle_t
-#define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS
-#define cublasSetStream hipblasSetStream
-#define cublasSgemm hipblasSgemm
-#define cublasStatus_t hipblasStatus_t
-#define cudaDataType_t hipblasDatatype_t //deprecated, new hipblasDatatype not in 5.6
-#define cudaDeviceCanAccessPeer hipDeviceCanAccessPeer
-#define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
-#define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
-#define cudaDeviceProp hipDeviceProp_t
-#define cudaDeviceSynchronize hipDeviceSynchronize
-#define cudaError_t hipError_t
-#define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled
-#define cudaErrorPeerAccessNotEnabled hipErrorPeerAccessNotEnabled
-#define cudaEventCreateWithFlags hipEventCreateWithFlags
-#define cudaEventDisableTiming hipEventDisableTiming
-#define cudaEventRecord hipEventRecord
-#define cudaEventSynchronize hipEventSynchronize
-#define cudaEvent_t hipEvent_t
-#define cudaEventDestroy hipEventDestroy
-#define cudaFree hipFree
-#define cudaFreeHost hipHostFree
-#define cudaGetDevice hipGetDevice
-#define cudaGetDeviceCount hipGetDeviceCount
-#define cudaGetDeviceProperties hipGetDeviceProperties
-#define cudaGetErrorString hipGetErrorString
-#define cudaGetLastError hipGetLastError
-#define cudaHostRegister hipHostRegister
-#define cudaHostRegisterPortable hipHostRegisterPortable
-#define cudaHostRegisterReadOnly hipHostRegisterReadOnly
-#define cudaHostUnregister hipHostUnregister
-#define cudaLaunchHostFunc hipLaunchHostFunc
-#define cudaMalloc hipMalloc
-#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
-#define cudaMemcpy hipMemcpy
-#define cudaMemcpyAsync hipMemcpyAsync
-#define cudaMemcpyPeerAsync hipMemcpyPeerAsync
-#define cudaMemcpy2DAsync hipMemcpy2DAsync
-#define cudaMemcpyDeviceToDevice hipMemcpyDeviceToDevice
-#define cudaMemcpyDeviceToHost hipMemcpyDeviceToHost
-#define cudaMemcpyHostToDevice hipMemcpyHostToDevice
-#define cudaMemcpyKind hipMemcpyKind
-#define cudaMemset hipMemset
-#define cudaMemsetAsync hipMemsetAsync
-#define cudaMemGetInfo hipMemGetInfo
-#define cudaOccupancyMaxPotentialBlockSize hipOccupancyMaxPotentialBlockSize
-#define cudaSetDevice hipSetDevice
-#define cudaStreamCreateWithFlags hipStreamCreateWithFlags
-#define cudaStreamDestroy hipStreamDestroy
-#define cudaStreamFireAndForget hipStreamFireAndForget
-#define cudaStreamNonBlocking hipStreamNonBlocking
-#define cudaStreamPerThread hipStreamPerThread
-#define cudaStreamSynchronize hipStreamSynchronize
-#define cudaStreamWaitEvent(stream, event, flags) hipStreamWaitEvent(stream, event, flags)
-#define cudaStream_t hipStream_t
-#define cudaSuccess hipSuccess
-#define __trap() do { abort(); __builtin_unreachable(); } while(0)
-#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
-#define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED
-#define CUBLAS_STATUS_ALLOC_FAILED HIPBLAS_STATUS_ALLOC_FAILED
-#define CUBLAS_STATUS_INVALID_VALUE HIPBLAS_STATUS_INVALID_VALUE
-#define CUBLAS_STATUS_ARCH_MISMATCH HIPBLAS_STATUS_ARCH_MISMATCH
-#define CUBLAS_STATUS_MAPPING_ERROR HIPBLAS_STATUS_MAPPING_ERROR
-#define CUBLAS_STATUS_EXECUTION_FAILED HIPBLAS_STATUS_EXECUTION_FAILED
-#define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR
-#define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED
+#include "vendors/hip.h"
+#elif defined(GGML_USE_MUSA)
+#include "vendors/musa.h"
 #else
-#include <cuda_runtime.h>
-#include <cuda.h>
-#include <cublas_v2.h>
-#include <cuda_fp16.h>
-
-#if CUDART_VERSION < 11020
-#define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED
-#define CUBLAS_TF32_TENSOR_OP_MATH CUBLAS_TENSOR_OP_MATH
-#define CUBLAS_COMPUTE_16F CUDA_R_16F
-#define CUBLAS_COMPUTE_32F CUDA_R_32F
-#define cublasComputeType_t cudaDataType_t
-#endif // CUDART_VERSION < 11020
-
+#include "vendors/cuda.h"
 #endif // defined(GGML_USE_HIPBLAS)
 
 #define STRINGIZE_IMPL(...) #__VA_ARGS__
@@ -168,7 +72,7 @@ void ggml_cuda_error(const char * stmt, const char * func, const char * file, in
 
 #define CUDA_CHECK(err) CUDA_CHECK_GEN(err, cudaSuccess, cudaGetErrorString)
 
-#if CUDART_VERSION >= 12000
+#if CUDART_VERSION >= 12000 || defined(GGML_USE_MUSA)
     static const char * cublas_get_error_str(const cublasStatus_t err) {
         return cublasGetStatusString(err);
     }
@@ -200,7 +104,7 @@ static const char * cu_get_error_str(CUresult err) {
 #define CU_CHECK(err) CUDA_CHECK_GEN(err, CUDA_SUCCESS, cu_get_error_str)
 #endif
 
-#if CUDART_VERSION >= 11100
+#if CUDART_VERSION >= 11100 || defined(GGML_USE_MUSA)
 #define GGML_CUDA_ASSUME(x) __builtin_assume(x)
 #else
 #define GGML_CUDA_ASSUME(x)
@@ -212,93 +116,7 @@ typedef half2 dfloat2;
 #else
 typedef float dfloat; // dequantize float
 typedef float2 dfloat2;
-#endif //GGML_CUDA_F16
-
-#if defined(GGML_USE_HIPBLAS)
-#define __CUDA_ARCH__ 1300
-
-#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
-    defined(__gfx1150__) || defined(__gfx1151__)
-#define RDNA3
-#endif
-
-#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || defined(__gfx1033__) || \
-    defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || defined(__gfx1037__)
-#define RDNA2
-#endif
-
-#if defined(__gfx1010__) || defined(__gfx1012__)
-#define RDNA1
-#endif
-
-#ifndef __has_builtin
-    #define __has_builtin(x) 0
-#endif
-
-typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
-typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
-static __device__ __forceinline__ int __vsubss4(const int a, const int b) {
-    const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
-    const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
-#if __has_builtin(__builtin_elementwise_sub_sat)
-    const int8x4_t c = __builtin_elementwise_sub_sat(va, vb);
-    return reinterpret_cast<const int &>(c);
-#else
-    int8x4_t c;
-    int16_t tmp;
-#pragma unroll
-    for (int i = 0; i < 4; i++) {
-        tmp = va[i] - vb[i];
-        if(tmp > std::numeric_limits<int8_t>::max()) tmp = std::numeric_limits<int8_t>::max();
-        if(tmp < std::numeric_limits<int8_t>::min()) tmp = std::numeric_limits<int8_t>::min();
-        c[i] = tmp;
-    }
-    return reinterpret_cast<int &>(c);
-#endif // __has_builtin(__builtin_elementwise_sub_sat)
-}
-
-static __device__ __forceinline__ int __vsub4(const int a, const int b) {
-    return __vsubss4(a, b);
-}
-
-static __device__ __forceinline__ unsigned int __vcmpeq4(unsigned int a, unsigned int b) {
-    const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
-    const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
-    unsigned int c;
-    uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
-#pragma unroll
-    for (int i = 0; i < 4; ++i) {
-        vc[i] = va[i] == vb[i] ? 0xff : 0x00;
-    }
-    return c;
-}
-
-static __device__ __forceinline__ unsigned int __vcmpne4(unsigned int a, unsigned int b) {
-    const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
-    const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
-    unsigned int c;
-    uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
-#pragma unroll
-    for (int i = 0; i < 4; ++i) {
-        vc[i] = va[i] == vb[i] ? 0x00 : 0xff;
-    }
-    return c;
-}
-
-#if defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
-// __shfl_xor() for half2 was added in ROCm 5.6
-static __device__ __forceinline__ half2 __shfl_xor(half2 var, int laneMask, int width) {
-    typedef union half2_b32 {
-        half2 val;
-        int   b32;
-    } half2_b32_t;
-    half2_b32_t tmp;
-    tmp.val = var;
-    tmp.b32 = __shfl_xor(tmp.b32, laneMask, width);
-    return tmp.val;
-}
-#endif // defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
-#endif // defined(GGML_USE_HIPBLAS)
+#endif // GGML_CUDA_F16
 
 #if (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
 #define FP16_AVAILABLE
@@ -455,7 +273,7 @@ static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half
     const uint32_t mask_high = 0xFFFF0000 * (float(__high2half(a)) > float(__high2half(b)));
     return mask_low | mask_high;
 }
-#endif // CUDART_VERSION < 12000
+#endif // CUDART_VERSION < CUDART_HMASK
 
 static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, int c) {
 #if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)

ggml/src/ggml-cuda/vendors/cuda.h

@@ -0,0 +1,14 @@
+#pragma once
+
+#include <cuda_runtime.h>
+#include <cuda.h>
+#include <cublas_v2.h>
+#include <cuda_fp16.h>
+
+#if CUDART_VERSION < 11020
+#define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED
+#define CUBLAS_TF32_TENSOR_OP_MATH CUBLAS_TENSOR_OP_MATH
+#define CUBLAS_COMPUTE_16F CUDA_R_16F
+#define CUBLAS_COMPUTE_32F CUDA_R_32F
+#define cublasComputeType_t cudaDataType_t
+#endif // CUDART_VERSION < 11020

ggml/src/ggml-cuda/vendors/hip.h

@@ -0,0 +1,177 @@
+#pragma once
+
+#include <hip/hip_runtime.h>
+#include <hipblas/hipblas.h>
+#include <hip/hip_fp16.h>
+#ifdef __HIP_PLATFORM_AMD__
+// for rocblas_initialize()
+#include "rocblas/rocblas.h"
+#endif // __HIP_PLATFORM_AMD__
+#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F
+#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
+#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
+#define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
+#define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
+#define CUBLAS_OP_N HIPBLAS_OP_N
+#define CUBLAS_OP_T HIPBLAS_OP_T
+#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
+#define CUBLAS_TF32_TENSOR_OP_MATH 0
+#define CUDA_R_16F  HIPBLAS_R_16F
+#define CUDA_R_32F  HIPBLAS_R_32F
+#define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
+#define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6
+#define cublasCreate hipblasCreate
+#define cublasDestroy hipblasDestroy
+#define cublasGemmEx hipblasGemmEx
+#define cublasGemmBatchedEx hipblasGemmBatchedEx
+#define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx
+#define cublasHandle_t hipblasHandle_t
+#define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS
+#define cublasSetStream hipblasSetStream
+#define cublasSgemm hipblasSgemm
+#define cublasStatus_t hipblasStatus_t
+#define cudaDataType_t hipblasDatatype_t //deprecated, new hipblasDatatype not in 5.6
+#define cudaDeviceCanAccessPeer hipDeviceCanAccessPeer
+#define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
+#define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
+#define cudaDeviceProp hipDeviceProp_t
+#define cudaDeviceSynchronize hipDeviceSynchronize
+#define cudaError_t hipError_t
+#define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled
+#define cudaErrorPeerAccessNotEnabled hipErrorPeerAccessNotEnabled
+#define cudaEventCreateWithFlags hipEventCreateWithFlags
+#define cudaEventDisableTiming hipEventDisableTiming
+#define cudaEventRecord hipEventRecord
+#define cudaEventSynchronize hipEventSynchronize
+#define cudaEvent_t hipEvent_t
+#define cudaEventDestroy hipEventDestroy
+#define cudaFree hipFree
+#define cudaFreeHost hipHostFree
+#define cudaGetDevice hipGetDevice
+#define cudaGetDeviceCount hipGetDeviceCount
+#define cudaGetDeviceProperties hipGetDeviceProperties
+#define cudaGetErrorString hipGetErrorString
+#define cudaGetLastError hipGetLastError
+#define cudaHostRegister hipHostRegister
+#define cudaHostRegisterPortable hipHostRegisterPortable
+#define cudaHostRegisterReadOnly hipHostRegisterReadOnly
+#define cudaHostUnregister hipHostUnregister
+#define cudaLaunchHostFunc hipLaunchHostFunc
+#define cudaMalloc hipMalloc
+#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
+#define cudaMemcpy hipMemcpy
+#define cudaMemcpyAsync hipMemcpyAsync
+#define cudaMemcpyPeerAsync hipMemcpyPeerAsync
+#define cudaMemcpy2DAsync hipMemcpy2DAsync
+#define cudaMemcpyDeviceToDevice hipMemcpyDeviceToDevice
+#define cudaMemcpyDeviceToHost hipMemcpyDeviceToHost
+#define cudaMemcpyHostToDevice hipMemcpyHostToDevice
+#define cudaMemcpyKind hipMemcpyKind
+#define cudaMemset hipMemset
+#define cudaMemsetAsync hipMemsetAsync
+#define cudaMemGetInfo hipMemGetInfo
+#define cudaOccupancyMaxPotentialBlockSize hipOccupancyMaxPotentialBlockSize
+#define cudaSetDevice hipSetDevice
+#define cudaStreamCreateWithFlags hipStreamCreateWithFlags
+#define cudaStreamDestroy hipStreamDestroy
+#define cudaStreamFireAndForget hipStreamFireAndForget
+#define cudaStreamNonBlocking hipStreamNonBlocking
+#define cudaStreamPerThread hipStreamPerThread
+#define cudaStreamSynchronize hipStreamSynchronize
+#define cudaStreamWaitEvent(stream, event, flags) hipStreamWaitEvent(stream, event, flags)
+#define cudaStream_t hipStream_t
+#define cudaSuccess hipSuccess
+#define __trap() do { abort(); __builtin_unreachable(); } while(0)
+#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
+#define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED
+#define CUBLAS_STATUS_ALLOC_FAILED HIPBLAS_STATUS_ALLOC_FAILED
+#define CUBLAS_STATUS_INVALID_VALUE HIPBLAS_STATUS_INVALID_VALUE
+#define CUBLAS_STATUS_ARCH_MISMATCH HIPBLAS_STATUS_ARCH_MISMATCH
+#define CUBLAS_STATUS_MAPPING_ERROR HIPBLAS_STATUS_MAPPING_ERROR
+#define CUBLAS_STATUS_EXECUTION_FAILED HIPBLAS_STATUS_EXECUTION_FAILED
+#define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR
+#define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED
+
+#define __CUDA_ARCH__ 1300
+
+#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
+    defined(__gfx1150__) || defined(__gfx1151__)
+#define RDNA3
+#endif
+
+#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || defined(__gfx1033__) || \
+    defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || defined(__gfx1037__)
+#define RDNA2
+#endif
+
+#if defined(__gfx1010__) || defined(__gfx1012__)
+#define RDNA1
+#endif
+
+#ifndef __has_builtin
+    #define __has_builtin(x) 0
+#endif
+
+typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
+typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
+static __device__ __forceinline__ int __vsubss4(const int a, const int b) {
+    const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
+    const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
+#if __has_builtin(__builtin_elementwise_sub_sat)
+    const int8x4_t c = __builtin_elementwise_sub_sat(va, vb);
+    return reinterpret_cast<const int &>(c);
+#else
+    int8x4_t c;
+    int16_t tmp;
+#pragma unroll
+    for (int i = 0; i < 4; i++) {
+        tmp = va[i] - vb[i];
+        if(tmp > std::numeric_limits<int8_t>::max()) tmp = std::numeric_limits<int8_t>::max();
+        if(tmp < std::numeric_limits<int8_t>::min()) tmp = std::numeric_limits<int8_t>::min();
+        c[i] = tmp;
+    }
+    return reinterpret_cast<int &>(c);
+#endif // __has_builtin(__builtin_elementwise_sub_sat)
+}
+
+static __device__ __forceinline__ int __vsub4(const int a, const int b) {
+    return __vsubss4(a, b);
+}
+
+static __device__ __forceinline__ unsigned int __vcmpeq4(unsigned int a, unsigned int b) {
+    const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
+    const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
+    unsigned int c;
+    uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
+#pragma unroll
+    for (int i = 0; i < 4; ++i) {
+        vc[i] = va[i] == vb[i] ? 0xff : 0x00;
+    }
+    return c;
+}
+
+static __device__ __forceinline__ unsigned int __vcmpne4(unsigned int a, unsigned int b) {
+    const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
+    const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
+    unsigned int c;
+    uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
+#pragma unroll
+    for (int i = 0; i < 4; ++i) {
+        vc[i] = va[i] == vb[i] ? 0x00 : 0xff;
+    }
+    return c;
+}
+
+#if defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
+// __shfl_xor() for half2 was added in ROCm 5.6
+static __device__ __forceinline__ half2 __shfl_xor(half2 var, int laneMask, int width) {
+    typedef union half2_b32 {
+        half2 val;
+        int   b32;
+    } half2_b32_t;
+    half2_b32_t tmp;
+    tmp.val = var;
+    tmp.b32 = __shfl_xor(tmp.b32, laneMask, width);
+    return tmp.val;
+}
+#endif // defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000

ggml/src/ggml-cuda/vendors/musa.h

@@ -0,0 +1,171 @@
+#pragma once
+
+#include <musa_runtime.h>
+#include <musa.h>
+#include <mublas.h>
+#include <musa_fp16.h>
+#define CUBLAS_COMPUTE_16F CUDA_R_16F
+#define CUBLAS_COMPUTE_32F CUDA_R_32F
+#define CUBLAS_COMPUTE_32F_FAST_16F MUBLAS_COMPUTE_32F_FAST_16F
+#define CUBLAS_GEMM_DEFAULT MUBLAS_GEMM_DEFAULT
+#define CUBLAS_GEMM_DEFAULT_TENSOR_OP MUBLAS_GEMM_DEFAULT
+#define CUBLAS_OP_N MUBLAS_OP_N
+#define CUBLAS_OP_T MUBLAS_OP_T
+#define CUBLAS_STATUS_SUCCESS MUBLAS_STATUS_SUCCESS
+#define CUBLAS_TF32_TENSOR_OP_MATH MUBLAS_MATH_MODE_DEFAULT
+#define CUDA_R_16F  MUSA_R_16F
+#define CUDA_R_32F  MUSA_R_32F
+#define cublasComputeType_t cudaDataType_t
+#define cublasCreate mublasCreate
+#define cublasDestroy mublasDestroy
+#define cublasGemmEx mublasGemmEx
+#define cublasGemmBatchedEx mublasGemmBatchedEx
+#define cublasGemmStridedBatchedEx mublasGemmStridedBatchedEx
+#define cublasHandle_t mublasHandle_t
+#define cublasSetMathMode mublasSetMathMode
+#define cublasSetStream mublasSetStream
+#define cublasSgemm mublasSgemm
+#define cublasStatus_t mublasStatus_t
+#define cublasGetStatusString mublasStatus_to_string
+#define cudaDataType_t musaDataType_t
+#define cudaDeviceCanAccessPeer musaDeviceCanAccessPeer
+#define cudaDeviceDisablePeerAccess musaDeviceDisablePeerAccess
+#define cudaDeviceEnablePeerAccess musaDeviceEnablePeerAccess
+#define cudaDeviceProp musaDeviceProp
+#define cudaDeviceSynchronize musaDeviceSynchronize
+#define cudaError_t musaError_t
+#define cudaErrorPeerAccessAlreadyEnabled musaErrorPeerAccessAlreadyEnabled
+#define cudaErrorPeerAccessNotEnabled musaErrorPeerAccessNotEnabled
+#define cudaEventCreateWithFlags musaEventCreateWithFlags
+#define cudaEventDisableTiming musaEventDisableTiming
+#define cudaEventRecord musaEventRecord
+#define cudaEventSynchronize musaEventSynchronize
+#define cudaEvent_t musaEvent_t
+#define cudaEventDestroy musaEventDestroy
+#define cudaFree musaFree
+#define cudaFreeHost musaFreeHost
+#define cudaGetDevice musaGetDevice
+#define cudaGetDeviceCount musaGetDeviceCount
+#define cudaGetDeviceProperties musaGetDeviceProperties
+#define cudaGetErrorString musaGetErrorString
+#define cudaGetLastError musaGetLastError
+#define cudaHostRegister musaHostRegister
+#define cudaHostRegisterPortable musaHostRegisterPortable
+#define cudaHostRegisterReadOnly musaHostRegisterReadOnly
+#define cudaHostUnregister musaHostUnregister
+#define cudaLaunchHostFunc musaLaunchHostFunc
+#define cudaMalloc musaMalloc
+#define cudaMallocHost musaMallocHost
+#define cudaMemcpy musaMemcpy
+#define cudaMemcpyAsync musaMemcpyAsync
+#define cudaMemcpyPeerAsync musaMemcpyPeerAsync
+#define cudaMemcpy2DAsync musaMemcpy2DAsync
+#define cudaMemcpyDeviceToDevice musaMemcpyDeviceToDevice
+#define cudaMemcpyDeviceToHost musaMemcpyDeviceToHost
+#define cudaMemcpyHostToDevice musaMemcpyHostToDevice
+#define cudaMemcpyKind musaMemcpyKind
+#define cudaMemset musaMemset
+#define cudaMemsetAsync musaMemsetAsync
+#define cudaMemGetInfo musaMemGetInfo
+#define cudaOccupancyMaxPotentialBlockSize musaOccupancyMaxPotentialBlockSize
+#define cudaSetDevice musaSetDevice
+#define cudaStreamCreateWithFlags musaStreamCreateWithFlags
+#define cudaStreamDestroy musaStreamDestroy
+#define cudaStreamFireAndForget musaStreamFireAndForget
+#define cudaStreamNonBlocking musaStreamNonBlocking
+#define cudaStreamPerThread musaStreamPerThread
+#define cudaStreamSynchronize musaStreamSynchronize
+#define cudaStreamWaitEvent musaStreamWaitEvent
+#define cudaStream_t musaStream_t
+#define cudaSuccess musaSuccess
+
+// Additional mappings for MUSA virtual memory pool
+#define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED MU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED
+#define CU_MEM_ACCESS_FLAGS_PROT_READWRITE MU_MEM_ACCESS_FLAGS_PROT_READWRITE
+#define CU_MEM_ALLOC_GRANULARITY_RECOMMENDED MU_MEM_ALLOC_GRANULARITY_RECOMMENDED
+#define CU_MEM_ALLOCATION_TYPE_PINNED MU_MEM_ALLOCATION_TYPE_PINNED
+#define CU_MEM_LOCATION_TYPE_DEVICE MU_MEM_LOCATION_TYPE_DEVICE
+#define CUdevice MUdevice
+#define CUdeviceptr MUdeviceptr
+#define CUmemAccessDesc MUmemAccessDesc
+#define CUmemAllocationProp MUmemAllocationProp
+#define CUmemGenericAllocationHandle MUmemGenericAllocationHandle
+#define cuDeviceGet muDeviceGet
+#define cuDeviceGetAttribute muDeviceGetAttribute
+#define cuMemAddressFree muMemAddressFree
+#define cuMemAddressReserve muMemAddressReserve
+#define cuMemCreate muMemCreate
+#define cuMemGetAllocationGranularity muMemGetAllocationGranularity
+#define cuMemMap muMemMap
+#define cuMemRelease muMemRelease
+#define cuMemSetAccess muMemSetAccess
+#define cuMemUnmap muMemUnmap
+#define cudaFuncAttributeMaxDynamicSharedMemorySize musaFuncAttributeMaxDynamicSharedMemorySize
+#define cudaFuncSetAttribute musaFuncSetAttribute
+#define cudaMemcpy3DPeerParms musaMemcpy3DPeerParms
+#define make_cudaExtent make_musaExtent
+#define make_cudaPitchedPtr make_musaPitchedPtr
+
+// Additional mappings for MUSA graphs
+#define CUDA_SUCCESS MUSA_SUCCESS
+#define CUresult MUresult
+#define cuGetErrorString muGetErrorString
+#define cudaErrorGraphExecUpdateFailure musaErrorGraphExecUpdateFailure
+#define cudaErrorInvalidDeviceFunction musaErrorInvalidDeviceFunction
+#define cudaGraphDestroy musaGraphDestroy
+#define cudaGraphExecDestroy musaGraphExecDestroy
+#define cudaGraphExec_t musaGraphExec_t
+#define cudaGraphExecUpdate musaGraphExecUpdate
+#define cudaGraphExecUpdateResultInfo musaGraphExecUpdateResult
+#define cudaGraphGetNodes musaGraphGetNodes
+#define cudaGraphInstantiate musaGraphInstantiate
+#define cudaGraphKernelNodeGetParams musaGraphKernelNodeGetParams
+#define cudaGraphKernelNodeSetParams musaGraphKernelNodeSetParams
+#define cudaGraphLaunch musaGraphLaunch
+#define cudaGraphNodeGetType musaGraphNodeGetType
+#define cudaGraphNode_t musaGraphNode_t
+#define cudaGraphNodeType musaGraphNodeType
+#define cudaGraphNodeTypeKernel musaGraphNodeTypeKernel
+#define cudaGraph_t musaGraph_t
+#define cudaKernelNodeParams musaKernelNodeParams
+#define cudaStreamCaptureModeRelaxed musaStreamCaptureModeRelaxed
+#define cudaStreamEndCapture musaStreamEndCapture
+
+// XXX: Clang builtins mapping
+#define __vsub4   __vsub4_musa
+#define __vcmpeq4 __vcmpeq4_musa
+#define __vcmpne4 __vcmpne4_musa
+
+#ifndef __has_builtin
+    #define __has_builtin(x) 0
+#endif
+
+typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
+
+static __device__ __forceinline__ int __vsub4_musa(const int a, const int b) {
+    return __vsubss4(a, b);
+}
+
+static __device__ __forceinline__ unsigned int __vcmpeq4_musa(unsigned int a, unsigned int b) {
+    const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
+    const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
+    unsigned int c;
+    uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
+#pragma unroll
+    for (int i = 0; i < 4; ++i) {
+        vc[i] = va[i] == vb[i] ? 0xff : 0x00;
+    }
+    return c;
+}
+
+static __device__ __forceinline__ unsigned int __vcmpne4_musa(unsigned int a, unsigned int b) {
+    const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
+    const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
+    unsigned int c;
+    uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
+#pragma unroll
+    for (int i = 0; i < 4; ++i) {
+        vc[i] = va[i] == vb[i] ? 0x00 : 0xff;
+    }
+    return c;
+}

ggml/src/ggml-sycl.cpp

@@ -3981,6 +3981,9 @@ bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tens
     ggml_sycl_func_t func;
 
     switch (tensor->op) {
+        case GGML_OP_CONV_TRANSPOSE_1D:
+            func = ggml_sycl_op_conv_transpose_1d;
+            break;
         case GGML_OP_REPEAT:
             func = ggml_sycl_repeat;
             break;
@@ -5090,6 +5093,15 @@ GGML_CALL static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t back
 
 GGML_CALL static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
     switch (op->op) {
+        case GGML_OP_CONV_TRANSPOSE_1D:
+            {
+                ggml_type src0_type = op->src[0]->type;
+                ggml_type src1_type = op->src[1]->type;
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
+                    return true;
+                }
+                return false;
+            } break;
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
                 case GGML_UNARY_OP_GELU:

ggml/src/ggml-sycl/backend.hpp

@@ -15,6 +15,7 @@
 
 #include "concat.hpp"
 #include "common.hpp"
+#include "conv.hpp"
 #include "convert.hpp"
 #include "dequantize.hpp"
 #include "dmmv.hpp"

ggml/src/ggml-sycl/conv.cpp

@@ -0,0 +1,99 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#include "conv.hpp"
+
+static  void conv_transpose_1d_kernel(
+        const int s0, const int output_size,
+        const int src0_ne0, const int src0_ne1, const int src0_ne2,
+        const int src1_ne0, const int dst_ne0,
+        const float * src0, const float * src1,  float * dst,
+        const sycl::nd_item<3> &item_ct1) {
+    int global_index = item_ct1.get_local_id(2) +
+                       item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    if (global_index >= output_size) {
+        return;
+    }
+
+    int out_index = global_index / dst_ne0;
+
+    float accumulator = 0;
+
+    for (int c = 0; c < src0_ne2; c++) {
+        int idx = global_index % dst_ne0;
+
+        int kernel_offset = (src0_ne0 * src0_ne1 * c) + (out_index * src0_ne0);
+        int input_offset = src1_ne0 * c;
+
+        for (int i = 0; i < src1_ne0; i++) {
+            if (!(idx >= i*s0 && idx < i*s0 + src0_ne0)) {
+                continue;
+            }
+            int weight_idx = idx - i*s0;
+
+            float kernel_weight = src0[kernel_offset + weight_idx];
+            float input_value =  src1[input_offset+i];
+
+            accumulator += kernel_weight * input_value;
+        }
+    }
+    dst[global_index] = accumulator;
+}
+
+static void conv_transpose_1d_f32_f32_sycl(
+    const int s0, const int output_size,
+    const int src0_ne0, const int src0_ne1, const int src0_ne2,
+    const int src1_ne0, const int dst_ne0,
+    const float *src0, const float *src1, float *dst,
+    const queue_ptr& stream) {
+
+    const int num_blocks = (output_size + SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE - 1) / SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE;
+    const sycl::range<3> block_dims(1, 1, SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE);
+    const sycl::range<3> block_nums(1, 1, num_blocks);
+    stream->parallel_for(
+        sycl::nd_range<3>(
+            block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) {
+            conv_transpose_1d_kernel(
+                s0, output_size,
+                src0_ne0, src0_ne1, src0_ne2,
+                src1_ne0, dst_ne0,
+                src0, src1, dst, item_ct1);
+        });
+}
+
+void ggml_sycl_op_conv_transpose_1d(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+    const ggml_tensor *src1, ggml_tensor *dst) {
+    const float * src0_d = (const float *)src0->data;
+    const float * src1_d = (const float *)src1->data;
+
+    float * dst_d = (float *)dst->data;
+    dpct::queue_ptr stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src1));
+
+    const int32_t * opts = (const int32_t *)dst->op_params;
+
+    const int s0 = opts[0];
+
+    const int64_t output_size = ggml_nelements(dst);
+
+    conv_transpose_1d_f32_f32_sycl(s0, output_size,
+        src0->ne[0], src0->ne[1], src0->ne[2],
+        src1->ne[0], dst->ne[0],
+        src0_d, src1_d, dst_d, stream);
+}
+

ggml/src/ggml-sycl/conv.hpp

@@ -0,0 +1,21 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_CONV_HPP
+#define GGML_SYCL_CONV_HPP
+
+#include "common.hpp"
+
+void ggml_sycl_op_conv_transpose_1d(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+  const ggml_tensor *src1, ggml_tensor *dst);
+
+#endif // GGML_SYCL_CONV_HPP

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

@@ -1834,6 +1834,20 @@ namespace dpct
     template <typename T1, typename T2, typename T3>
     inline auto dp4a(T1 a, T2 b, T3 c)
     {
+#if defined(__SYCL_DEVICE_ONLY__) && defined(__NVPTX__) &&                     \
+    defined(__SYCL_CUDA_ARCH__) && __SYCL_CUDA_ARCH__ >= 610
+        dot_product_acc_t<T1, T2> res;
+        if constexpr (std::is_same_v<dot_product_acc_t<T1, T2>, uint32_t>) {
+            asm volatile("dp4a.u32.u32 %0, %1, %2, %3;"
+                         : "=r"(res)
+                         : "r"(a), "r"(b), "r"(c));
+        } else {
+            asm volatile("dp4a.s32.s32 %0, %1, %2, %3;"
+                         : "=r"(res)
+                         : "r"(a), "r"(b), "r"(c));
+        }
+        return res;
+#else
         dot_product_acc_t<T1, T2> res = c;
         auto va = extract_and_sign_or_zero_extend4(a);
         auto vb = extract_and_sign_or_zero_extend4(b);
@@ -1842,6 +1856,7 @@ namespace dpct
         res += va[2] * vb[2];
         res += va[3] * vb[3];
         return res;
+#endif
     }
 
     struct sub_sat

ggml/src/ggml-sycl/presets.hpp

@@ -41,6 +41,7 @@
 #define SYCL_ACC_BLOCK_SIZE 256
 #define SYCL_IM2COL_BLOCK_SIZE 256
 #define SYCL_POOL2D_BLOCK_SIZE 256
+#define SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE 256
 
 // dmmv = dequantize_mul_mat_vec
 #ifndef GGML_SYCL_DMMV_X

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

@@ -30,6 +30,20 @@
 
 #define ASYNCIO_CONCURRENCY 64
 
+// define prototypes
+void execute_command(const std::string& command, std::string& stdout_str, std::string& stderr_str);
+bool directory_exists(const std::string& path);
+bool create_directory(const std::string& path);
+std::string to_uppercase(const std::string& input);
+bool string_ends_with(const std::string& str, const std::string& suffix);
+std::string join_paths(const std::string& path1, const std::string& path2);
+std::string basename(const std::string &path);
+void string_to_spv(const std::string& _name, const std::string& in_fname, const std::map<std::string, std::string>& defines, bool fp16);
+std::map<std::string, std::string> merge_maps(const std::map<std::string, std::string>& a, const std::map<std::string, std::string>& b);
+void matmul_shaders(std::vector<std::future<void>>& tasks, bool fp16, bool matmul_id);
+void process_shaders(std::vector<std::future<void>>& tasks);
+void write_output_files();
+
 std::mutex lock;
 std::vector<std::pair<std::string, std::string>> shader_fnames;
 
@@ -38,7 +52,7 @@ std::string input_dir = "vulkan-shaders";
 std::string output_dir = "/tmp";
 std::string target_hpp = "ggml-vulkan-shaders.hpp";
 std::string target_cpp = "ggml-vulkan-shaders.cpp";
-bool no_clean = false;
+bool clean = true;
 
 const std::vector<std::string> type_names = {
     "f32",
@@ -464,8 +478,9 @@ void write_output_files() {
         }
         fprintf(src, "\n};\n\n");
 
-        if (!no_clean) {
+        if (clean) {
             std::remove(path.c_str());
+            // fprintf(stderr, "Removed: %s\n", path.c_str());
         }
     }
 
@@ -481,6 +496,18 @@ int main(int argc, char** argv) {
         }
     }
 
+    if (argc <= 1 || args.find("--help") != args.end()) {
+        std::cout << "Usage:\n"
+                     "\tvulkan-shaders-gen [options]\n\n"
+                     "Options:\n"
+                     "\t--glslc <path>        Path to glslc executable (default: /usr/bin/glslc)\n"
+                     "\t--input-dir           Directory containing shader sources (required)\n"
+                     "\t--output-dir          Output directory for generated SPIR-V files and optional C++ headers\n"
+                     "\t--target-hpp <path>   Path to generate a header file with shader declarations in C++ format\n"
+                     "\t--target-cpp <path>   Path to generate a source code file implementing the declared shaders (optional)\n"
+                     "\t--no-clean            Keep temporary SPIR-V files after build (default: remove them)\n";
+        return EXIT_SUCCESS;
+    }
     if (args.find("--glslc") != args.end()) {
         GLSLC = args["--glslc"]; // Path to glslc
     }
@@ -497,7 +524,7 @@ int main(int argc, char** argv) {
         target_cpp = args["--target-cpp"]; // Path to generated cpp file
     }
     if (args.find("--no-clean") != args.end()) {
-        no_clean = true; // Keep temporary SPIR-V files in output-dir after build
+        clean = false; // Keep temporary SPIR-V files in output-dir after build
     }
 
     if (!directory_exists(input_dir)) {

include/llama.h

@@ -33,17 +33,15 @@
 
 #define LLAMA_DEFAULT_SEED 0xFFFFFFFF
 
-#define LLAMA_MAX_RNG_STATE (64*1024)
-
 #define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla'
 #define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
 #define LLAMA_FILE_MAGIC_GGSQ 0x67677371u // 'ggsq'
 
 #define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
-#define LLAMA_SESSION_VERSION 7
+#define LLAMA_SESSION_VERSION 8
 
 #define LLAMA_STATE_SEQ_MAGIC   LLAMA_FILE_MAGIC_GGSQ
-#define LLAMA_STATE_SEQ_VERSION 1
+#define LLAMA_STATE_SEQ_VERSION 2
 
 #ifdef __cplusplus
 extern "C" {
@@ -691,10 +689,11 @@ extern "C" {
     // State / sessions
     //
 
-    // Returns the maximum size in bytes of the state (rng, logits, embedding
-    // and kv_cache) - will often be smaller after compacting tokens
-    LLAMA_API size_t llama_state_get_size(const struct llama_context * ctx);
-    LLAMA_API DEPRECATED(size_t llama_get_state_size(const struct llama_context * ctx),
+    // Returns the *actual* size in bytes of the state
+    // (rng, logits, embedding and kv_cache)
+    // Only use when saving the state, not when restoring it, otherwise the size may be too small.
+    LLAMA_API size_t llama_state_get_size(struct llama_context * ctx);
+    LLAMA_API DEPRECATED(size_t llama_get_state_size(struct llama_context * ctx),
         "use llama_state_get_size instead");
 
     // Copies the state to the specified destination address.
@@ -702,7 +701,8 @@ extern "C" {
     // Returns the number of bytes copied
     LLAMA_API size_t llama_state_get_data(
             struct llama_context * ctx,
-                         uint8_t * dst);
+                         uint8_t * dst,
+                          size_t   size);
     LLAMA_API DEPRECATED(size_t llama_copy_state_data(
             struct llama_context * ctx,
                          uint8_t * dst),
@@ -712,7 +712,8 @@ extern "C" {
     // Returns the number of bytes read
     LLAMA_API size_t llama_state_set_data(
             struct llama_context * ctx,
-                   const uint8_t * src);
+                   const uint8_t * src,
+                          size_t   size);
     LLAMA_API DEPRECATED(size_t llama_set_state_data(
             struct llama_context * ctx,
                    const uint8_t * src),
@@ -754,6 +755,7 @@ extern "C" {
     LLAMA_API size_t llama_state_seq_get_data(
             struct llama_context * ctx,
                          uint8_t * dst,
+                          size_t   size,
                     llama_seq_id   seq_id);
 
     // Copy the sequence data (originally copied with `llama_state_seq_get_data`) into the specified sequence
@@ -763,6 +765,7 @@ extern "C" {
     LLAMA_API size_t llama_state_seq_set_data(
             struct llama_context * ctx,
                    const uint8_t * src,
+                          size_t   size,
                     llama_seq_id   dest_seq_id);
 
     LLAMA_API size_t llama_state_seq_save_file(

scripts/sync-ggml-am.sh

@@ -102,6 +102,8 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
     # cmake/FindSIMD.cmake    -> ggml/cmake/FindSIMD.cmake
     #
     # src/ggml.c              -> ggml/src/ggml.c
+    # src/ggml-aarch64.c      -> ggml/src/ggml-aarch64.c
+    # src/ggml-aarch64.h      -> ggml/src/ggml-aarch64.h
     # src/ggml-alloc.c        -> ggml/src/ggml-alloc.c
     # src/ggml-backend-impl.h -> ggml/src/ggml-backend-impl.h
     # src/ggml-backend.c      -> ggml/src/ggml-backend.c
@@ -117,6 +119,7 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
     # src/ggml-sycl/*         -> ggml/src/ggml-sycl/
     # src/ggml-sycl.cpp       -> ggml/src/ggml-sycl.cpp
     # src/ggml-vulkan.cpp     -> ggml/src/ggml-vulkan.cpp
+    # src/vulkan-shaders/*    -> ggml/src/vulkan-shaders/
     #
     # include/ggml.h         -> ggml/include/ggml.h
     # include/ggml-alloc.h   -> ggml/include/ggml-alloc.h
@@ -143,6 +146,8 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
         -e 's/([[:space:]]|[ab]\/)src\/CMakeLists.txt/\1ggml\/src\/CMakeLists.txt/g' \
         -e 's/([[:space:]]|[ab]\/)cmake\/FindSIMD.cmake/\1ggml\/cmake\/FindSIMD.cmake/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml\.c/\1ggml\/src\/ggml.c/g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-aarch64\.c/\1ggml\/src\/ggml-aarch64.c/g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-aarch64\.h/\1ggml\/src\/ggml-aarch64.h/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-alloc\.c/\1ggml\/src\/ggml-alloc.c/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-backend-impl\.h/\1ggml\/src\/ggml-backend-impl.h/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-backend\.c/\1ggml\/src\/ggml-backend.c/g' \
@@ -158,6 +163,7 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
         -e 's/([[:space:]]|[ab]\/)src\/ggml-sycl\//\1ggml\/src\/ggml-sycl\//g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-sycl\.cpp/\1ggml\/src\/ggml-sycl.cpp/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-vulkan\.cpp/\1ggml\/src\/ggml-vulkan.cpp/g' \
+        -e 's/([[:space:]]|[ab]\/)src\/vulkan-shaders\//\1ggml\/src\/vulkan-shaders\//g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml\.h/\1ggml\/include\/ggml.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-alloc\.h/\1ggml\/include\/ggml-alloc.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-backend\.h/\1ggml\/include\/ggml-backend.h/g' \

scripts/sync-ggml.sh

@@ -5,6 +5,8 @@ cp -rpv ../ggml/src/CMakeLists.txt   ./ggml/src/CMakeLists.txt
 cp -rpv ../ggml/cmake/FindSIMD.cmake ./ggml/cmake/FindSIMD.cmake
 
 cp -rpv ../ggml/src/ggml.c              ./ggml/src/ggml.c
+cp -rpv ../ggml/src/ggml-aarch64.c      ./ggml/src/ggml-aarch64.c
+cp -rpv ../ggml/src/ggml-aarch64.h      ./ggml/src/ggml-aarch64.h
 cp -rpv ../ggml/src/ggml-alloc.c        ./ggml/src/ggml-alloc.c
 cp -rpv ../ggml/src/ggml-backend-impl.h ./ggml/src/ggml-backend-impl.h
 cp -rpv ../ggml/src/ggml-backend.c      ./ggml/src/ggml-backend.c
@@ -21,6 +23,7 @@ cp -rpv ../ggml/src/ggml-rpc.cpp        ./ggml/src/ggml-rpc.cpp
 cp -rpv ../ggml/src/ggml-sycl/*         ./ggml/src/ggml-sycl/
 cp -rpv ../ggml/src/ggml-sycl.cpp       ./ggml/src/ggml-sycl.cpp
 cp -rpv ../ggml/src/ggml-vulkan.cpp     ./ggml/src/ggml-vulkan.cpp
+cp -rpv ../ggml/src/vulkan-shaders/*    ./ggml/src/vulkan-shaders/
 
 cp -rpv ../ggml/include/ggml.h         ./ggml/include/ggml.h
 cp -rpv ../ggml/include/ggml-alloc.h   ./ggml/include/ggml-alloc.h