correct cmd name

cann: fix buffer_num and runtime speed slowly error (#8865 )
readme : add ramalama to the availables UI (#8811 )
2026-07-16 17:35:58 +02:00 · 2024-08-06 00:15:33 +08:00 · 2024-08-05 21:10:37 +08:00 · 2024-08-05 15:45:01 +03:00 · 2024-08-05 15:43:40 +03:00 · 2024-08-05 21:15:28 +10:00
90 changed files with 3363 additions and 1587 deletions
@@ -3,7 +3,7 @@ ARG UBUNTU_VERSION=22.04
 FROM ubuntu:$UBUNTU_VERSION AS build

 RUN apt-get update && \
-    apt-get install -y build-essential git libcurl4-openssl-dev curl
+    apt-get install -y build-essential git libcurl4-openssl-dev

 WORKDIR /app

@@ -16,7 +16,7 @@ RUN make -j$(nproc) llama-server
 FROM ubuntu:$UBUNTU_VERSION AS runtime

 RUN apt-get update && \
-    apt-get install -y libcurl4-openssl-dev libgomp1
+    apt-get install -y libcurl4-openssl-dev libgomp1 curl

 COPY --from=build /app/llama-server /llama-server

@@ -126,16 +126,9 @@ let
    ++ optionals useMetalKit [ MetalKit ];

  cudaBuildInputs = with cudaPackages; [
-    cuda_cccl.dev # <nv/target>
-
-    # A temporary hack for reducing the closure size, remove once cudaPackages
-    # have stopped using lndir: https://github.com/NixOS/nixpkgs/issues/271792
-    cuda_cudart.dev
-    cuda_cudart.lib
-    cuda_cudart.static
-    libcublas.dev
-    libcublas.lib
-    libcublas.static
+    cuda_cudart
+    cuda_cccl # <nv/target>
+    libcublas
  ];

  rocmBuildInputs = with rocmPackages; [
@@ -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
@@ -50,6 +50,7 @@ build*
 !docs/build.md
 /libllama.so
 /llama-*
+/vulkan-shaders-gen
 android-ndk-*
 arm_neon.h
 cmake-build-*
@@ -139,7 +139,8 @@ set(LLAMA_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location o
 # determining _precisely_ which defines are necessary for the llama-config
 # package.
 #
-get_directory_property(GGML_DIR_DEFINES DIRECTORY ggml/src COMPILE_DEFINITIONS)
+get_target_property(GGML_DIRECTORY ggml SOURCE_DIR)
+get_directory_property(GGML_DIR_DEFINES DIRECTORY ${GGML_DIRECTORY} COMPILE_DEFINITIONS)
 get_target_property(GGML_TARGET_DEFINES ggml COMPILE_DEFINITIONS)
 set(GGML_TRANSIENT_DEFINES ${GGML_TARGET_DEFINES} ${GGML_DIR_DEFINES})
 get_target_property(GGML_LINK_LIBRARIES ggml LINK_LIBRARIES)
@@ -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 )

@@ -1568,42 +1605,41 @@ llama-q8dot: pocs/vdot/q8dot.cpp ggml/src/ggml.o \
 # Mark legacy binary targets as .PHONY so that they are always checked.
 .PHONY: main quantize perplexity embedding server

+# Define the object file target
+examples/deprecation-warning/deprecation-warning.o: examples/deprecation-warning/deprecation-warning.cpp
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
 # NOTE: We currently will always build the deprecation-warning `main` and `server` binaries to help users migrate.
 #  Eventually we will want to remove these target from building all the time.
-main: examples/deprecation-warning/deprecation-warning.cpp
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+main: examples/deprecation-warning/deprecation-warning.o
+	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
 	@echo "NOTICE: The 'main' binary is deprecated. Please use 'llama-cli' instead."

-server: examples/deprecation-warning/deprecation-warning.cpp
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+server: examples/deprecation-warning/deprecation-warning.o
+	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
 	@echo "NOTICE: The 'server' binary is deprecated. Please use 'llama-server' instead."

-quantize: examples/deprecation-warning/deprecation-warning.cpp
+quantize: examples/deprecation-warning/deprecation-warning.o
 ifneq (,$(wildcard quantize))
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
 	@echo "#########"
 	@echo "WARNING: The 'quantize' binary is deprecated. Please use 'llama-quantize' instead."
 	@echo "  Remove the 'quantize' binary to remove this warning."
 	@echo "#########"
 endif

-perplexity: examples/deprecation-warning/deprecation-warning.cpp
+perplexity: examples/deprecation-warning/deprecation-warning.o
 ifneq (,$(wildcard perplexity))
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
 	@echo "#########"
 	@echo "WARNING: The 'perplexity' binary is deprecated. Please use 'llama-perplexity' instead."
 	@echo "  Remove the 'perplexity' binary to remove this warning."
 	@echo "#########"
 endif

-embedding: examples/deprecation-warning/deprecation-warning.cpp
+embedding: examples/deprecation-warning/deprecation-warning.o
 ifneq (,$(wildcard embedding))
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
 	@echo "#########"
 	@echo "WARNING: The 'embedding' binary is deprecated. Please use 'llama-embedding' instead."
 	@echo "  Remove the 'embedding' binary to remove this warning."
@@ -95,8 +95,16 @@ Typically finetunes of the base models below are supported as well.
 - [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
 - [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
 - [x] [OLMo](https://allenai.org/olmo)
+- [x] [Granite models](https://huggingface.co/collections/ibm-granite/granite-code-models-6624c5cec322e4c148c8b330)
 - [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia)
+- [x] [Snowflake-Arctic MoE](https://huggingface.co/collections/Snowflake/arctic-66290090abe542894a5ac520)
+- [x] [Smaug](https://huggingface.co/models?search=Smaug)
+- [x] [Poro 34B](https://huggingface.co/LumiOpen/Poro-34B)
+- [x] [Bitnet b1.58 models](https://huggingface.co/1bitLLM)
+- [x] [Flan T5](https://huggingface.co/models?search=flan-t5)
+- [x] [Open Elm models](https://huggingface.co/collections/apple/openelm-instruct-models-6619ad295d7ae9f868b759ca)
 - [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b)
+- [x] [SmolLM](https://huggingface.co/collections/HuggingFaceTB/smollm-6695016cad7167254ce15966)

 (instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md))

@@ -145,6 +153,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [Faraday](https://faraday.dev/) (proprietary)
 - [LMStudio](https://lmstudio.ai/) (proprietary)
 - [Layla](https://play.google.com/store/apps/details?id=com.laylalite) (proprietary)
+- [ramalama](https://github.com/containers/ramalama) (MIT)
 - [LocalAI](https://github.com/mudler/LocalAI) (MIT)
 - [LostRuins/koboldcpp](https://github.com/LostRuins/koboldcpp) (AGPL)
 - [Mozilla-Ocho/llamafile](https://github.com/Mozilla-Ocho/llamafile)
@@ -409,6 +418,7 @@ 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 |
@@ -1634,7 +1634,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
    options.push_back({ "server",      "       --host HOST",            "ip address to listen (default: %s)", params.hostname.c_str() });
    options.push_back({ "server",      "       --port PORT",            "port to listen (default: %d)", params.port });
    options.push_back({ "server",      "       --path PATH",            "path to serve static files from (default: %s)", params.public_path.c_str() });
-    options.push_back({ "server",      "       --embedding(s)",         "enable embedding endpoint (default: %s)", params.embedding ? "enabled" : "disabled" });
+    options.push_back({ "server",      "       --embedding(s)",         "restrict to only support embedding use case; use only with dedicated embedding models (default: %s)", params.embedding ? "enabled" : "disabled" });
    options.push_back({ "server",      "       --api-key KEY",          "API key to use for authentication (default: none)" });
    options.push_back({ "server",      "       --api-key-file FNAME",   "path to file containing API keys (default: none)" });
    options.push_back({ "server",      "       --ssl-key-file FNAME",   "path to file a PEM-encoded SSL private key" });
@@ -316,7 +316,7 @@ class Model:
                if self.ftype != gguf.LlamaFileType.ALL_F32 and extra_f16 and not extra_f32:
                    if self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
                        data = gguf.quantize_bf16(data)
-                        assert data.dtype == np.int16
+                        assert data.dtype == np.uint16
                        data_qtype = gguf.GGMLQuantizationType.BF16

                    elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0 and gguf.can_quantize_to_q8_0(data):
@@ -178,7 +178,11 @@ For Jetson user, if you have Jetson Orin, you can try this: [Offical Support](ht
  cmake --build build --config Release
  ```

-The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used. The following compilation options are also available to tweak performance:
+The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used.
+
+The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted. In Windows this setting is available in the NVIDIA control panel as `System Memory Fallback`.
+
+The following compilation options are also available to tweak performance:

 | Option                        | Legal values           | Default | Description                                                                                                                                                                                                                                                                             |
 |-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
@@ -192,6 +196,19 @@ 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.
@@ -1,7 +1,6 @@
 #include "ggml.h"
 #include "train.h"

-#include <vector>
 #include <cassert>
 #include <cstdlib>
 #include <cstring>
@@ -69,7 +69,7 @@ int main(int argc, char ** argv) {
    llama_context_params ctx_params = llama_context_params_from_gpt_params(params);

    // ensure enough sequences are available
-    ctx_params.n_seq_max = *std::max_element(n_pl.begin(), n_pl.end());
+    ctx_params.n_seq_max = n_pl.empty() ? 1 : *std::max_element(n_pl.begin(), n_pl.end());

    llama_context * ctx = llama_new_context_with_model(model, ctx_params);

@@ -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);
@@ -247,7 +247,7 @@ server:
         --host HOST              ip address to listen (default: 127.0.0.1)
         --port PORT              port to listen (default: 8080)
         --path PATH              path to serve static files from (default: )
-         --embedding(s)           enable embedding endpoint (default: disabled)
+         --embedding(s)           restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)
         --api-key KEY            API key to use for authentication (default: none)
         --api-key-file FNAME     path to file containing API keys (default: none)
         --ssl-key-file FNAME     path to file a PEM-encoded SSL private key
@@ -900,7 +900,7 @@ struct server_context {

        slot.params.stream             = json_value(data, "stream",            false);
        slot.params.cache_prompt       = json_value(data, "cache_prompt",      false);
-        slot.params.n_predict          = json_value(data, "n_predict",         default_params.n_predict);
+        slot.params.n_predict          = json_value(data, "n_predict",         json_value(data, "max_tokens", default_params.n_predict));
        slot.sparams.top_k             = json_value(data, "top_k",             default_sparams.top_k);
        slot.sparams.top_p             = json_value(data, "top_p",             default_sparams.top_p);
        slot.sparams.min_p             = json_value(data, "min_p",             default_sparams.min_p);
@@ -355,24 +355,6 @@ static json oaicompat_completion_params_parse(

    llama_params["__oaicompat"] = true;

-    // Map OpenAI parameters to llama.cpp parameters
-    //
-    // For parameters that are defined by the OpenAI documentation (e.g.
-    // temperature), we explicitly specify OpenAI's intended default; we
-    // need to do that because sometimes OpenAI disagrees with llama.cpp
-    //
-    // https://platform.openai.com/docs/api-reference/chat/create
-    llama_sampling_params default_sparams;
-    llama_params["model"]             = json_value(body,   "model",             std::string("unknown"));
-    llama_params["frequency_penalty"] = json_value(body,   "frequency_penalty", 0.0);
-    llama_params["logit_bias"]        = json_value(body,   "logit_bias",        json::object());
-    llama_params["n_predict"]         = json_value(body,   "max_tokens",        -1);
-    llama_params["presence_penalty"]  = json_value(body,   "presence_penalty",  0.0);
-    llama_params["seed"]              = json_value(body,   "seed",              LLAMA_DEFAULT_SEED);
-    llama_params["stream"]            = json_value(body,   "stream",            false);
-    llama_params["temperature"]       = json_value(body,   "temperature",       1.0);
-    llama_params["top_p"]             = json_value(body,   "top_p",             1.0);
-
    // Apply chat template to the list of messages
    llama_params["prompt"] = format_chat(model, chat_template, body.at("messages"));

@@ -6,4 +6,4 @@ set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
@call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force


-.\build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0
+.\build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0
@@ -5,11 +5,11 @@
        "nixpkgs-lib": "nixpkgs-lib"
      },
      "locked": {
-        "lastModified": 1719994518,
-        "narHash": "sha256-pQMhCCHyQGRzdfAkdJ4cIWiw+JNuWsTX7f0ZYSyz0VY=",
+        "lastModified": 1722555600,
+        "narHash": "sha256-XOQkdLafnb/p9ij77byFQjDf5m5QYl9b2REiVClC+x4=",
        "owner": "hercules-ci",
        "repo": "flake-parts",
-        "rev": "9227223f6d922fee3c7b190b2cc238a99527bbb7",
+        "rev": "8471fe90ad337a8074e957b69ca4d0089218391d",
        "type": "github"
      },
      "original": {
@@ -20,11 +20,11 @@
    },
    "nixpkgs": {
      "locked": {
-        "lastModified": 1721379653,
-        "narHash": "sha256-8MUgifkJ7lkZs3u99UDZMB4kbOxvMEXQZ31FO3SopZ0=",
+        "lastModified": 1722421184,
+        "narHash": "sha256-/DJBI6trCeVnasdjUo9pbnodCLZcFqnVZiLUfqLH4jA=",
        "owner": "NixOS",
        "repo": "nixpkgs",
-        "rev": "1d9c2c9b3e71b9ee663d11c5d298727dace8d374",
+        "rev": "9f918d616c5321ad374ae6cb5ea89c9e04bf3e58",
        "type": "github"
      },
      "original": {
@@ -36,14 +36,14 @@
    },
    "nixpkgs-lib": {
      "locked": {
-        "lastModified": 1719876945,
-        "narHash": "sha256-Fm2rDDs86sHy0/1jxTOKB1118Q0O3Uc7EC0iXvXKpbI=",
+        "lastModified": 1722555339,
+        "narHash": "sha256-uFf2QeW7eAHlYXuDktm9c25OxOyCoUOQmh5SZ9amE5Q=",
        "type": "tarball",
-        "url": "https://github.com/NixOS/nixpkgs/archive/5daf0514482af3f97abaefc78a6606365c9108e2.tar.gz"
+        "url": "https://github.com/NixOS/nixpkgs/archive/a5d394176e64ab29c852d03346c1fc9b0b7d33eb.tar.gz"
      },
      "original": {
        "type": "tarball",
-        "url": "https://github.com/NixOS/nixpkgs/archive/5daf0514482af3f97abaefc78a6606365c9108e2.tar.gz"
+        "url": "https://github.com/NixOS/nixpkgs/archive/a5d394176e64ab29c852d03346c1fc9b0b7d33eb.tar.gz"
      }
    },
    "root": {
@@ -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)
@@ -206,6 +207,7 @@ set(GGML_PUBLIC_HEADERS
    include/ggml-alloc.h
    include/ggml-backend.h
    include/ggml-blas.h
+    include/ggml-cann.h
    include/ggml-cuda.h
    include/ggml.h
    include/ggml-kompute.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"
@@ -349,6 +349,7 @@ extern "C" {
    GGML_API ggml_bf16_t ggml_fp32_to_bf16(float);
    GGML_API float       ggml_bf16_to_fp32(ggml_bf16_t);  // consider just doing << 16
    GGML_API void        ggml_bf16_to_fp32_row(const ggml_bf16_t *, float *, int64_t);
+    GGML_API void        ggml_fp32_to_bf16_row_ref(const float *, ggml_bf16_t *, int64_t);
    GGML_API void        ggml_fp32_to_bf16_row(const float *, ggml_bf16_t *, int64_t);

    struct ggml_object;
@@ -1455,7 +1456,6 @@ extern "C" {
    // if mode & 2 == 1, GPT-NeoX style
    //
    // b is an int32 vector with size a->ne[2], it contains the positions
-    // c is freq factors (e.g. phi3-128k), (optional)
    GGML_API struct ggml_tensor * ggml_rope(
            struct ggml_context * ctx,
            struct ggml_tensor  * a,
@@ -1472,6 +1472,7 @@ extern "C" {
            int                   mode);

    // custom RoPE
+    // c is freq factors (e.g. phi3-128k), (optional)
    GGML_API struct ggml_tensor * ggml_rope_ext(
            struct ggml_context * ctx,
            struct ggml_tensor  * a,
@@ -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")
@@ -804,11 +849,6 @@ if (GGML_CANN)
                ${CANN_INSTALL_DIR}/acllib/include
            )

-            # TODO: find libs
-            link_directories(
-                ${CANN_INSTALL_DIR}/lib64
-            )
-
            add_subdirectory(ggml-cann/kernels)
            list(APPEND CANN_LIBRARIES
                ascendcl
@@ -827,6 +867,7 @@ if (GGML_CANN)

            set(GGML_EXTRA_LIBS     ${GGML_EXTRA_LIBS}     ${CANN_LIBRARIES} )
            set(GGML_EXTRA_INCLUDES ${GGML_EXTRA_INCLUDES} ${CANN_INCLUDE_DIRS})
+            set(GGML_EXTRA_LIBDIRS  ${GGML_EXTRA_LIBDIRS}  ${CANN_INSTALL_DIR}/lib64)
            list(APPEND GGML_CDEF_PUBLIC GGML_USE_CANN)
        endif()
    else()
@@ -857,8 +898,10 @@ 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 +1307,7 @@ 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})
@@ -384,8 +384,8 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void *
    UNUSED(blocklen);

 #if defined(__ARM_FEATURE_SVE)
-    if (svcntw() == 8) {
-        GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) &&
+    if (ggml_sve_cnt_b == QK8_0) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
                    "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
    }
 #endif
@@ -496,8 +496,8 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void *
    UNUSED(blocklen);

 #if defined(__ARM_FEATURE_SVE)
-    if (svcntw() == 8) {
-        GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) &&
+    if (ggml_sve_cnt_b == QK8_0) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
                    "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
    }
 #endif
@@ -614,7 +614,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
    UNUSED(blocklen);

 #if defined(__ARM_FEATURE_SVE) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
-    if (svcntw() == 8) {
+    if (ggml_sve_cnt_b == QK8_0) {
        const void * b_ptr = vx;
        const void * a_ptr = vy;
        float * res_ptr = s;
@@ -680,12 +680,12 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
        return;
    }
    else if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
-        GGML_ASSERT((ggml_cpu_has_sve() && (svcntw() == 8)) &&
+        GGML_ASSERT((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
                    "__ARM_FEATURE_SVE for vector size of 256-bits not defined, use the Q4_0_4_8 quantization format for optimal "
                    "performance");
    }
    else if (ggml_cpu_has_neon()) {
-        GGML_ASSERT(((ggml_cpu_has_sve() && (svcntw() == 8)) || ggml_cpu_has_matmul_int8()) &&
+        GGML_ASSERT(((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) || ggml_cpu_has_matmul_int8()) &&
                    "__ARM_FEATURE_SVE for vector size of 256-bits and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 "
                    "quantization format for optimal performance");
    }
@@ -745,8 +745,8 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void *
    UNUSED(blocklen);

 #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
-    if (svcntw() == 8) {
-        GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) &&
+    if (ggml_sve_cnt_b == QK8_0) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
                    "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
    }
 #endif
@@ -1266,8 +1266,8 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void *
    UNUSED(blocklen);

 #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
-    if (svcntw() == 8) {
-        GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) &&
+    if (ggml_sve_cnt_b == QK8_0) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
                    "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
    }
 #endif
@@ -1728,7 +1728,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
    UNUSED(blocklen);

 #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
-    if (svcntw() == 8) {
+    if (ggml_sve_cnt_b == QK8_0) {
        const void * b_ptr = vx;
        const void * a_ptr = vy;
        float * res_ptr = s;
@@ -2139,12 +2139,12 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
        return;
    }
    else if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
-        GGML_ASSERT((ggml_cpu_has_sve() && (svcntw() == 8)) &&
+        GGML_ASSERT((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
                    "__ARM_FEATURE_SVE for vector size of 256-bits not defined, use the Q4_0_4_8 quantization format for optimal "
                    "performance");
    }
    else if (ggml_cpu_has_neon()) {
-        GGML_ASSERT(((ggml_cpu_has_sve() && (svcntw() == 8)) || ggml_cpu_has_matmul_int8()) &&
+        GGML_ASSERT(((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) || ggml_cpu_has_matmul_int8()) &&
                    "__ARM_FEATURE_SVE for vector size of 256-bits and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 "
                    "quantization format for optimal performance");
    }
@@ -627,7 +627,6 @@ GGML_CALL static void* ggml_backend_cann_buffer_get_base(
 GGML_CALL static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor,
                                                       const void* src,
                                                       void* dst) {
-    GGML_ASSERT(tensor->op == GGML_OP_NONE);

    int64_t n_elems = ggml_nelements(tensor);
    int64_t groups = n_elems / QK4_0;
@@ -679,7 +678,6 @@ GGML_CALL static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor,
 */
 GGML_CALL static void ggml_backend_cann_transform_back_q4_0(
    const ggml_tensor* tensor, void* src, void* dst) {
-    GGML_ASSERT(tensor->op == GGML_OP_NONE);

    int64_t n_elems = ggml_nelements(tensor);
    int64_t groups = n_elems / QK4_0;
@@ -1666,10 +1664,13 @@ GGML_CALL static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
            }
        case GGML_OP_MUL_MAT: {
            switch (op->src[0]->type) {
-                // case GGML_TYPE_Q4_0:
                case GGML_TYPE_F16:
                case GGML_TYPE_F32:
                case GGML_TYPE_Q8_0:
+                    // TODO: fix me
+                    // Current groupsize should not be greater than k-1 in
+                    // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize().
+                case GGML_TYPE_Q4_0:
                    return true;
                default:
                    return false;
@@ -1694,6 +1695,7 @@ GGML_CALL static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
                case GGML_TYPE_F32:
                case GGML_TYPE_F16:
                case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q4_0:
                    return true;
                default:
                    return false;
@@ -37,6 +37,10 @@ aclDataType ggml_cann_type_mapping(ggml_type type) {
            return ACL_INT16;
        case GGML_TYPE_I32:
            return ACL_INT32;
+        case GGML_TYPE_Q4_0:
+            return ACL_INT4;
+        case GGML_TYPE_Q8_0:
+            return ACL_INT8;
        default:
            return ACL_DT_UNDEFINED;
    }
@@ -89,33 +93,6 @@ bool ggml_cann_need_bcast(const ggml_tensor* t0, const ggml_tensor* t1) {
    return false;
 }

-aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
-                                   size_t type_size, int64_t* ne, size_t* nb,
-                                   int64_t dims, aclFormat format,
-                                   size_t offset) {
-    int64_t tmp_ne[GGML_MAX_DIMS * 2];
-    int64_t tmp_stride[GGML_MAX_DIMS * 2];
-
-    memcpy(tmp_ne, ne, dims * sizeof(int64_t));
-    for (int i = 0; i < dims; i++) {
-        tmp_stride[i] = nb[i] / type_size;
-    }
-
-    std::reverse(tmp_ne, tmp_ne + dims);
-    std::reverse(tmp_stride, tmp_stride + dims);
-
-    int64_t acl_storage_len = 0;
-    for (int i = 0; i < dims; i++) {
-        acl_storage_len += (ne[i] - 1) * nb[i];
-    }
-
-    aclTensor* acl_tensor =
-        aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size,
-                        format, &acl_storage_len, 1, data_ptr);
-
-    return acl_tensor;
-}
-
 int64_t ggml_cann_get_bcast_shape(const ggml_tensor* src0,
                                  const ggml_tensor* src1,
                                  int64_t* bcast_src0_ne,
@@ -23,6 +23,9 @@
 #ifndef CANN_ACL_TENSOR_H
 #define CANN_ACL_TENSOR_H

+#include <algorithm>
+#include <cstring>
+
 #include <aclnn/aclnn_base.h>
 #include "common.h"

@@ -65,7 +68,8 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne = null
                             size_t offset = 0);

 /**
- * @brief   Creates an ACL tensor from provided parameters.
+ * @brief   Template for creating an ACL tensor from provided parameters. typename TYPE
+ *          should be size_t or float.
 *
 * @details This function creates an ACL tensor using the provided data pointer,
 *          data type, dimensions, strides, format, offset, and additional parameters.
@@ -83,10 +87,34 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne = null
 * @param   offset      Offset in bytes for the ACL tensor data. Defaults to 0.
 * @return  Pointer to the created ACL tensor.
 */
+template<typename TYPE>
 aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
-                             size_t type_size, int64_t* ne, size_t* nb,
-                             int64_t dims, aclFormat format = ACL_FORMAT_ND,
-                             size_t offset = 0);
+                                   TYPE type_size, int64_t* ne, TYPE* nb,
+                                   int64_t dims,
+                                   aclFormat format = ACL_FORMAT_ND,
+                                   size_t offset = 0) {
+    int64_t tmp_ne[GGML_MAX_DIMS * 2];
+    int64_t tmp_stride[GGML_MAX_DIMS * 2];
+
+    memcpy(tmp_ne, ne, dims * sizeof(int64_t));
+    for (int i = 0; i < dims; i++) {
+        tmp_stride[i] = nb[i] / type_size;
+    }
+
+    std::reverse(tmp_ne, tmp_ne + dims);
+    std::reverse(tmp_stride, tmp_stride + dims);
+
+    int64_t acl_storage_len = 0;
+    for (int i = 0; i < dims; i++) {
+        acl_storage_len += (ne[i] - 1) * nb[i];
+    }
+
+    aclTensor* acl_tensor =
+        aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size,
+                        format, &acl_storage_len, 1, data_ptr);
+
+    return acl_tensor;
+}

 /**
 * @brief   Checks if tensors require broadcasting based on their shapes.
@@ -910,6 +910,13 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                ((ggml_tensor*)dst->extra)->ne);
            return;
        }
+        if (dst->type == GGML_TYPE_Q4_0) {
+            aclrtlaunch_ascendc_quantize_f16_to_q4_0(
+                24, ctx.stream(), src->data, dst->data,
+                ((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb,
+                ((ggml_tensor*)dst->extra)->ne);
+            return;
+        }
        if (dst->type == GGML_TYPE_F16) {
            if (ggml_are_same_shape(src, dst)) {
                cann_copy(ctx, acl_src, acl_dst);
@@ -971,6 +978,13 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                ((ggml_tensor*)dst->extra)->ne);
            return;
        }
+        if (dst->type == GGML_TYPE_Q4_0) {
+            aclrtlaunch_ascendc_quantize_f32_to_q4_0(
+                24, ctx.stream(), src->data, dst->data,
+                ((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb,
+                ((ggml_tensor*)dst->extra)->ne);
+            return;
+        }
        if (dst->type == GGML_TYPE_F32) {
            if (ggml_are_same_shape(src, dst)) {
                cann_copy(ctx, acl_src, acl_dst);
@@ -1312,6 +1326,111 @@ aclnnStatus aclnnIm2col(void* workspace, uint64_t workspaceSize,
 #ifdef __cplusplus
 }
 #endif
+
+static void ggml_cann_im2col_2d_post_process(ggml_backend_cann_context& ctx,
+                                             ggml_tensor* dst,
+                                             ggml_tensor* src1,
+                                             aclTensor* tmp_cast_tensor,
+                                             aclTensor* tmp_im2col_tensor) {
+    // Permute: [N, IC * KH * KW, OW * OH] -> [N, OW * OH, IC * KH * KW]
+    int64_t dst_ne[] = {dst->ne[0], dst->ne[1] * dst->ne[2], dst->ne[3]};
+    size_t dst_nb[] = {dst->nb[0], dst->nb[1], dst->nb[3]};
+    aclTensor* acl_dst =
+        ggml_cann_create_tensor(dst, dst_ne, dst_nb, GGML_MAX_DIMS - 1);
+
+    int64_t permute_dim[] = {0, 2, 1};
+    if (src1->type != dst->type) {
+        aclnn_permute(ctx, tmp_cast_tensor, acl_dst, permute_dim, 3);
+    } else {
+        aclnn_permute(ctx, tmp_im2col_tensor, acl_dst, permute_dim, 3);
+    }
+
+    // release
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
+
+static void ggml_cann_im2col_1d_post_process(
+    ggml_backend_cann_context& ctx, ggml_tensor* dst, ggml_tensor* src1,
+    aclTensor* tmp_cast_tensor, aclTensor* tmp_im2col_tensor,
+    const std::vector<int64_t>& im2col_op_params) {
+    // get params
+    const int64_t KH = im2col_op_params[0];
+    const int64_t KW = im2col_op_params[1];
+    const int64_t IW = im2col_op_params[2];
+    const int64_t IC = im2col_op_params[3];
+    const int64_t N = im2col_op_params[4];
+    const int64_t OH = im2col_op_params[5];
+    const int64_t OW = im2col_op_params[6];
+    const int64_t s0 = im2col_op_params[7];
+    const int64_t p0 = im2col_op_params[8];
+    const int64_t d0 = im2col_op_params[9];
+    const int64_t n_bytes_factor = im2col_op_params[10];
+
+    // Permute: [N, IC * KH * KW, OW * OH] ->
+    // [N, OW * OH * n_bytes_factor, IC * KH * KW]
+    aclTensor* tmp_permute_tensor = nullptr;
+    ggml_cann_pool_alloc tmp_permute_allocator(ctx.pool());
+    tmp_permute_allocator.alloc(ggml_nbytes(dst) * n_bytes_factor);
+    void* tmp_permute_buffer = tmp_permute_allocator.get();
+
+    int64_t tmp_permute_ne[] = {IC * KH * KW, OW * OH * n_bytes_factor, N};
+    size_t tmp_permute_nb[GGML_MAX_DIMS - 1];
+    tmp_permute_nb[0] = ggml_type_size(dst->type);
+    for (int i = 1; i < GGML_MAX_DIMS - 1; i++) {
+        tmp_permute_nb[i] = tmp_permute_nb[i - 1] * tmp_permute_ne[i - 1];
+    }
+
+    tmp_permute_tensor = ggml_cann_create_tensor(
+        tmp_permute_buffer, ggml_cann_type_mapping(dst->type),
+        ggml_type_size(dst->type), tmp_permute_ne, tmp_permute_nb,
+        GGML_MAX_DIMS - 1, ACL_FORMAT_ND);
+
+    int64_t permute_dim[] = {0, 2, 1};
+    if (src1->type != dst->type) {
+        aclnn_permute(ctx, tmp_cast_tensor, tmp_permute_tensor, permute_dim, 3);
+    } else {
+        aclnn_permute(ctx, tmp_im2col_tensor, tmp_permute_tensor, permute_dim,
+                      3);
+    }
+
+    // number of times the kernel moves in W dimension
+    const int n_step_w = (IW + 2 * p0 - d0 * (KW - 1) - 1) / s0 + 1;
+    size_t offset;
+    void *cur_dst_buffer = dst->data, *cur_permute_buffer = tmp_permute_buffer;
+
+    // memory copy with offset to restore 1D im2col from 2d
+    if (IC > 1) {
+        offset = IC * KH * KW * n_step_w * ggml_type_size(dst->type);
+        size_t size_cpy = KH * KW * ggml_type_size(dst->type);
+
+        for (int c = 0; c < IC; c++) {
+            cur_permute_buffer = (char*)tmp_permute_buffer + offset +
+                                 KH * KW * c * ggml_type_size(dst->type);
+            cur_dst_buffer = (char*)dst->data +
+                             c * KH * KW * n_step_w * ggml_type_size(dst->type);
+
+            for (int i = 0; i < n_step_w; i++) {
+                ACL_CHECK(aclrtMemcpyAsync(
+                    cur_dst_buffer, size_cpy, cur_permute_buffer, size_cpy,
+                    ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
+                cur_dst_buffer =
+                    (char*)cur_dst_buffer + KH * KW * ggml_type_size(dst->type);
+                cur_permute_buffer = (char*)cur_permute_buffer +
+                                     KH * KW * IC * ggml_type_size(dst->type);
+            }
+        }
+    } else {
+        offset = KH * KW * n_step_w *
+                 ggml_type_size(dst->type);  // equal to ggml_nbytes(dst)
+        ACL_CHECK(aclrtMemcpyAsync(dst->data, offset,
+                                   (char*)tmp_permute_buffer + offset, offset,
+                                   ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
+    }
+
+    // release
+    ACL_CHECK(aclDestroyTensor(tmp_permute_tensor));
+}
+
 void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
    ggml_tensor* src0 = dst->src[0];  // kernel
    ggml_tensor* src1 = dst->src[1];  // input
@@ -1320,21 +1439,23 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
    GGML_ASSERT(src1->type == GGML_TYPE_F32);
    GGML_ASSERT(dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);

-    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
-    const int32_t s1 = ((const int32_t*)(dst->op_params))[1];
-    const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
-    const int32_t p1 = ((const int32_t*)(dst->op_params))[3];
-    const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
-    const int32_t d1 = ((const int32_t*)(dst->op_params))[5];
-    const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
-
    GGML_TENSOR_BINARY_OP_LOCALS;

-    const int64_t N = is_2D ? ne13 : ne12;
-    const int64_t IC = is_2D ? ne12 : ne11;
+    // aclnnIm2col only works on 2D. set s1, p1, d1 to 1 to perform 2D
+    // im2col and do post-processing to restore it to 1D.
+    const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
+    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
+    const int32_t s1 = is_2D ? ((const int32_t*)(dst->op_params))[1] : 1;
+    const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
+    const int32_t p1 = is_2D ? ((const int32_t*)(dst->op_params))[3] : 1;
+    const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
+    const int32_t d1 = is_2D ? ((const int32_t*)(dst->op_params))[5] : 1;

-    const int64_t KH = is_2D ? ne01 : 1;
+    const int64_t N = ne13;
+    const int64_t IC = ne12;
+    const int64_t KH = ne01;
    const int64_t KW = ne00;
+    const int64_t IW = ne10;

    const int64_t OH = is_2D ? ne2 : 1;
    const int64_t OW = ne1;
@@ -1342,9 +1463,12 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
    GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
    GGML_ASSERT(nb10 == sizeof(float));

-    // im2col: [N,C,H,W] -> [N, IC * KH * KW, OW * OH]
+    // memory allocated increased to 3x when is_2D == false
+    const int64_t n_bytes_factor = is_2D ? 1 : 3;
+
+    // im2col: [N,C,H,W] -> [N, IC * KH * KW, OW * OH * n_bytes_factor]
    aclTensor* acl_src1 = ggml_cann_create_tensor(src1);
-    int64_t tmp_im2col_ne[] = {OW * OH, IC * KH * KW, N};
+    int64_t tmp_im2col_ne[] = {OW * OH * n_bytes_factor, IC * KH * KW, N};
    size_t tmp_im2col_nb[GGML_MAX_DIMS - 1];

    tmp_im2col_nb[0] = ggml_type_size(src1->type);
@@ -1356,8 +1480,10 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
    // If dst is f16, tmp_buffer is f32, we need alloc src.typesize *
    // dst.elemcount.
    ggml_cann_pool_alloc im2col_allocator(
-        ctx.pool(), ggml_nelements(dst) * ggml_element_size(src1));
+        ctx.pool(),
+        ggml_nelements(dst) * ggml_element_size(src1) * n_bytes_factor);
    void* tmp_im2col_buffer = im2col_allocator.get();
+
    aclTensor* tmp_im2col_tensor = ggml_cann_create_tensor(
        tmp_im2col_buffer, ggml_cann_type_mapping(src1->type),
        ggml_type_size(src1->type), tmp_im2col_ne, tmp_im2col_nb,
@@ -1380,8 +1506,9 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                                          paddings, strides, tmp_im2col_tensor,
                                          &workspaceSize, &executor));

+    ggml_cann_pool_alloc workspace_allocator(ctx.pool());
    if (workspaceSize > 0) {
-        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspace_allocator.alloc(workspaceSize);
        workspaceAddr = workspace_allocator.get();
    }

@@ -1391,9 +1518,10 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
    // Cast if dst is f16.
    aclTensor* tmp_cast_tensor = nullptr;
    ggml_cann_pool_alloc tmp_cast_allocator(ctx.pool());
+    void* tmp_cast_buffer = nullptr;
    if (src1->type != dst->type) {
-        tmp_cast_allocator.alloc(ggml_nbytes(dst));
-        void* tmp_cast_buffer = tmp_cast_allocator.get();
+        tmp_cast_allocator.alloc(ggml_nbytes(dst) * n_bytes_factor);
+        tmp_cast_buffer = tmp_cast_allocator.get();
        size_t temp_cast_nb[GGML_MAX_DIMS - 1];
        temp_cast_nb[0] = ggml_type_size(dst->type);
        for (int i = 1; i < GGML_MAX_DIMS - 1; i++) {
@@ -1408,24 +1536,21 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                   ggml_cann_type_mapping(dst->type));
    }

-    // Permute: [N, IC * KH * KW, OW * OH] -> [N, OW * OH, IC * KH * KW]
-    int64_t dst_ne[] = {dst->ne[0], dst->ne[1] * dst->ne[2], dst->ne[3]};
-    size_t dst_nb[] = {dst->nb[0], dst->nb[1], dst->nb[3]};
-    aclTensor* acl_dst =
-        ggml_cann_create_tensor(dst, dst_ne, dst_nb, GGML_MAX_DIMS - 1);
-
-    int64_t permute_dim[] = {0, 2, 1};
-    if (src1->type != dst->type) {
-        aclnn_permute(ctx, tmp_cast_tensor, acl_dst, permute_dim, 3);
+    // post-processing
+    if (is_2D) {
+        ggml_cann_im2col_2d_post_process(ctx, dst, src1, tmp_cast_tensor,
+                                         tmp_im2col_tensor);
    } else {
-        aclnn_permute(ctx, tmp_im2col_tensor, acl_dst, permute_dim, 3);
+        std::vector<int64_t> im2col_op_params = {
+            KH, KW, IW, IC, N, OH, OW, s0, p0, d0, n_bytes_factor};
+        ggml_cann_im2col_1d_post_process(ctx, dst, src1, tmp_cast_tensor,
+                                         tmp_im2col_tensor, im2col_op_params);
    }

    // release
    ACL_CHECK(aclDestroyTensor(acl_src1));
    ACL_CHECK(aclDestroyTensor(tmp_im2col_tensor));
    ACL_CHECK(aclDestroyTensor(tmp_cast_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_dst));
    ACL_CHECK(aclDestroyIntArray(kernel_size));
    ACL_CHECK(aclDestroyIntArray(dilations));
    ACL_CHECK(aclDestroyIntArray(paddings));
@@ -2352,21 +2477,33 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx,
 * @param dst The destination tensor where the result of the matrix
 * multiplication will be stored.
 */
-static void ggml_cann_mul_mat_q8_0(ggml_backend_cann_context& ctx,
-                                   ggml_tensor* dst) {
+static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
+                                   ggml_tensor* dst,
+                                   const enum ggml_type type) {
    ggml_tensor* src0 = dst->src[0];  // weight
    ggml_tensor* src1 = dst->src[1];  // input

    // The shape of the weight is NCHW. Matrix multiplication uses HW dims. HC
    // is regarded as batch. weight need transpose.
    int64_t weight_ne[] = {src0->ne[1], src0->ne[0]};
-    size_t weight_elem_size = sizeof(uint8_t);
-    size_t weight_nb[] = {weight_elem_size * src0->ne[0], weight_elem_size};
+    float weight_elem_size;
+    if (type == GGML_TYPE_Q4_0) {
+        weight_elem_size = float(sizeof(uint8_t)) / 2;
+    }
+    else if (type == GGML_TYPE_Q8_0) {
+        weight_elem_size = float(sizeof(uint8_t));
+    }
+    else {
+        GGML_ABORT("Only support Q4_0 and Q8_0 MUL_MAT");
+    }
+    float weight_nb[] = {weight_elem_size * src0->ne[0], weight_elem_size};
+
    // size of one matrix is element_size * height * width.
    size_t weight_stride = weight_elem_size * src0->ne[0] * src0->ne[1];
    size_t weight_size = weight_stride * src0->ne[2] * src0->ne[3];

    // scale stored at the end of weight. Also need transpose.
+    GGML_ASSERT(QK4_0 == QK8_0);
    int64_t scale_ne[] = {src0->ne[1], src0->ne[0] / QK8_0};
    size_t scale_elem_size = sizeof(uint16_t);
    size_t scale_nb[] = {src0->ne[0] / QK8_0 * scale_elem_size,
@@ -2381,10 +2518,10 @@ static void ggml_cann_mul_mat_q8_0(ggml_backend_cann_context& ctx,
    size_t input_nb[] = {input_elem_size, input_elem_size * src1->ne[0]};
    size_t input_stride = input_elem_size * src1->ne[0] * src1->ne[1];

+    ggml_cann_pool_alloc input_alloctor(ctx.pool());
    if (src1->type != GGML_TYPE_F16) {
        aclTensor* acl_src1_tensor = ggml_cann_create_tensor(src1);
-        ggml_cann_pool_alloc input_alloctor(
-            ctx.pool(), ggml_nelements(src1) * input_elem_size);
+        input_alloctor.alloc(ggml_nelements(src1) * input_elem_size);
        input_buffer = input_alloctor.get();

        int64_t* input_cast_ne = src1->ne;
@@ -2430,8 +2567,9 @@ static void ggml_cann_mul_mat_q8_0(ggml_backend_cann_context& ctx,
                (char*)input_buffer + batch1 * input_stride, ACL_FLOAT16,
                input_elem_size, input_ne, input_nb, 2);
            aclTensor* acl_weight_tensor = ggml_cann_create_tensor(
-                (char*)src0->data + batch0 * weight_stride, ACL_INT8,
-                weight_elem_size, weight_ne, weight_nb, 2);
+                (char*)src0->data + batch0 * weight_stride,
+                ggml_cann_type_mapping(type), weight_elem_size, weight_ne,
+                weight_nb, 2);
            aclTensor* acl_scale_tensor = ggml_cann_create_tensor(
                scale_offset + batch0 * scale_stride, ACL_FLOAT16,
                scale_elem_size, scale_ne, scale_nb, 2);
@@ -2485,11 +2623,9 @@ void ggml_cann_mul_mat(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
        case GGML_TYPE_F16:
            ggml_cann_mat_mul_fp(ctx, dst);
            break;
-        // case GGML_TYPE_Q4_0:
-        //     ggml_cann_mul_mat_q4_0(ctx, dst);
-        //     break;
+        case GGML_TYPE_Q4_0:
        case GGML_TYPE_Q8_0:
-            ggml_cann_mul_mat_q8_0(ctx, dst);
+            ggml_cann_mul_mat_quant(ctx, dst, type);
            break;
        default:
            GGML_ABORT("fatal error");
@@ -9,6 +9,7 @@ file(GLOB SRC_FILES
    get_row_q8_0.cpp
    quantize_f32_q8_0.cpp
    quantize_f16_q8_0.cpp
+    quantize_float_to_q4_0.cpp
    dup.cpp
 )

@@ -29,4 +30,4 @@ ascendc_library(ascendc_kernels STATIC
    ${SRC_FILES}
 )

-#ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)
+# ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)
@@ -8,6 +8,8 @@

 #include "aclrtlaunch_ascendc_quantize_f32_q8_0.h"
 #include "aclrtlaunch_ascendc_quantize_f16_q8_0.h"
+#include "aclrtlaunch_ascendc_quantize_f16_to_q4_0.h"
+#include "aclrtlaunch_ascendc_quantize_f32_to_q4_0.h"

 #include "aclrtlaunch_ascendc_dup_by_rows_fp16.h"
 #include "aclrtlaunch_ascendc_dup_by_rows_fp32.h"
@@ -0,0 +1,278 @@
+#include "kernel_operator.h"
+
+using namespace AscendC;
+
+#define BUFFER_NUM 2
+#define Group_Size 32
+
+template <typename SRC_T>
+class QUANTIZE_FLOAT_TO_Q4_0 {
+   public:
+    __aicore__ inline QUANTIZE_FLOAT_TO_Q4_0() {}
+    __aicore__ inline void init(GM_ADDR input, GM_ADDR output,
+                                int64_t *input_ne_ub, size_t *input_nb_ub,
+                                int64_t *output_ne_ub) {
+        // TODO: fix test_case CPY(type_src=f16,type_dst=q4_0,ne=[256,4,4,4],
+        //                         permute=[0,0,0,0]):
+        // [CPY] NMSE = 0.000008343 > 0.000001000 FAIL
+        int64_t op_block_num = GetBlockNum();
+        int64_t op_block_idx = GetBlockIdx();
+
+        // input stride of data elements
+        for (int i = 0; i < 4; i++) {
+            input_ne[i] = input_ne_ub[i];
+            input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
+            output_ne[i] = output_ne_ub[i];
+        }
+
+        // output stride of data elements
+        output_stride[0] = 1;
+        for (int i = 1; i < 4; i++) {
+            output_stride[i] = output_stride[i - 1] * output_ne[i - 1];
+        }
+
+        // scale saved one by one after data:. [group1_scale, group2_scale, ...]
+        scale_ne = input_ne;
+        scale_stride[0] = 1;
+        scale_stride[1] = input_ne[0] / Group_Size;
+        for (int i = 2; i < 4; i++) {
+            scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
+        }
+
+        // split input tensor by rows.
+        uint64_t nr = input_ne[1] * input_ne[2] * input_ne[3];
+        dr = nr / op_block_num;
+
+        uint64_t tails = nr % op_block_num;
+        if (op_block_idx < tails) {
+            dr += 1;
+            ir = dr * op_block_idx;
+        } else {
+            ir = dr * op_block_idx + tails;
+        }
+
+        group_size_in_row = scale_stride[1];
+        int64_t scale_offset = output_ne[0] * output_ne[1] * output_ne[2] *
+                              output_ne[3] * sizeof(uint8_t) / 2;
+
+        input_gm.SetGlobalBuffer((__gm__ SRC_T *)input);
+        output_gm.SetGlobalBuffer((__gm__ int8_t *)output);
+        scale_gm.SetGlobalBuffer((__gm__ half *)(output + scale_offset + ir *
+                                                 group_size_in_row *
+                                                 sizeof(half)));
+
+        pipe.InitBuffer(input_queue, BUFFER_NUM, Group_Size * sizeof(SRC_T));
+        pipe.InitBuffer(output_queue, BUFFER_NUM,
+                            Group_Size * sizeof(int8_t) / 2);
+        pipe.InitBuffer(cast_queue , 1, Group_Size * sizeof(float));
+        pipe.InitBuffer(work_queue, 1, Group_Size * sizeof(float));
+        pipe.InitBuffer(max_queue, 1, Group_Size * sizeof(float));
+        pipe.InitBuffer(min_queue, 1, Group_Size * sizeof(float));
+        pipe.InitBuffer(scale_queue, 1, Group_Size / 2 * sizeof(half));
+        pipe.InitBuffer(int8_queue, 1, Group_Size * sizeof(int8_t));
+        pipe.InitBuffer(half_queue, 1, Group_Size * sizeof(half));
+    }
+
+    __aicore__ inline void copy_in(uint32_t offset) {
+        LocalTensor<SRC_T> input_local = input_queue.AllocTensor<SRC_T>();
+        DataCopy(input_local, input_gm[offset], Group_Size);
+        input_queue.EnQue(input_local);
+    }
+
+    __aicore__ inline void copy_out(uint32_t offset) {
+        // reinterpretcast Group_Size(32) * int4b_t to Group_Size / 2 * int8_t,
+        // and using DataCopyPad to avoid 32 bits align.
+        LocalTensor<int4b_t> output_local = output_queue.DeQue<int4b_t>();
+        LocalTensor<int8_t> output_int8_local =
+                                    output_local.ReinterpretCast<int8_t>();
+
+        DataCopyExtParams dataCopyParams;
+        dataCopyParams.blockCount = 1;
+        dataCopyParams.blockLen = Group_Size / 2  * sizeof(int8_t);
+        DataCopyPad(output_gm[offset], output_int8_local, dataCopyParams);
+
+        output_queue.FreeTensor(output_local);
+    }
+
+    __aicore__ inline void input_to_cast(LocalTensor<float> cast_local,
+                                         LocalTensor<float> input_local) {
+        DataCopy(cast_local, input_local, Group_Size);
+    }
+
+    __aicore__ inline void input_to_cast(LocalTensor<float> cast_local,
+                                         LocalTensor<half> input_local) {
+        Cast(cast_local, input_local, RoundMode::CAST_NONE, Group_Size);
+    }
+
+    __aicore__ inline half calculate_group(int64_t row, int64_t group) {
+        const int64_t i3 = row / (input_ne[1] * input_ne[2]);
+        const int64_t i2 = (row - i3 * input_ne[1] * input_ne[2]) / input_ne[1];
+        const int64_t i1 =
+            row - i3 * input_ne[1] * input_ne[2] - i2 * input_ne[1];
+
+        const int64_t input_offset = i1 * input_stride[1] +
+                                     i2 * input_stride[2] +
+                                     i3 * input_stride[3] + Group_Size * group;
+
+        // output_offset is stride for output_gm which datatype is int8_t and
+        // divided by 2 is needed for int4b_t.
+        const int64_t output_offset = (i1 * output_stride[1] +
+                                       i2 * output_stride[2] +
+                                       i3 * output_stride[3] +
+                                       Group_Size * group) / 2;
+        copy_in(input_offset);
+
+        LocalTensor<SRC_T> input_local = input_queue.DeQue<SRC_T>();
+        LocalTensor<int4b_t> output_local = output_queue.AllocTensor<int4b_t>();
+        LocalTensor<float> cast_local = cast_queue.AllocTensor<float>();
+        LocalTensor<float> work_local = work_queue.AllocTensor<float>();
+        LocalTensor<float> max_local = max_queue.AllocTensor<float>();
+        LocalTensor<float> min_local = min_queue.AllocTensor<float>();
+        LocalTensor<int8_t> int8_local = int8_queue.AllocTensor<int8_t>();
+        LocalTensor<half> half_local = half_queue.AllocTensor<half>();
+
+        input_to_cast(cast_local, input_local);
+
+        ReduceMax(max_local, cast_local, work_local, Group_Size);
+        ReduceMin(min_local, cast_local, work_local, Group_Size);
+        const float max_value = max_local.GetValue(0);
+        const float min_value = min_local.GetValue(0);
+        float d = max_value;
+        if (min_value < 0 && (-1 * min_value) > max_value) {
+            d = min_value;
+        }
+
+        d = d / (-8);
+        if (d != 0) {
+            Muls(cast_local, cast_local, 1.0f / d, Group_Size);
+        }
+
+        // range: [-8,8] -> [0.5,16.5] -> [0,16] -> [0,15] -> [-8,7]
+        float scalar = 8.5f;
+        Adds(cast_local, cast_local, scalar, Group_Size);
+        Cast(cast_local, cast_local, RoundMode::CAST_FLOOR, Group_Size);
+        scalar = 15.0f;
+        Mins(cast_local, cast_local, scalar, Group_Size);
+        scalar = -8.0f;
+        Adds(cast_local, cast_local, scalar, Group_Size);
+
+        // float->half->int4b
+        Cast(half_local, cast_local, RoundMode::CAST_NONE, Group_Size);
+        Cast(output_local, half_local, RoundMode::CAST_NONE, Group_Size);
+
+        output_queue.EnQue(output_local);
+        copy_out(output_offset);
+
+        input_queue.FreeTensor(input_local);
+        work_queue.FreeTensor(work_local);
+        max_queue.FreeTensor(max_local);
+        min_queue.FreeTensor(min_local);
+        int8_queue.FreeTensor(int8_local);
+        half_queue.FreeTensor(half_local);
+        cast_queue.FreeTensor(cast_local);
+        return (half)d;
+    }
+
+    __aicore__ inline void calculate() {
+        LocalTensor<half> scale_local = scale_queue.AllocTensor<half>();
+        uint32_t scale_local_offset = 0;
+        uint32_t scale_global_offset = 0;
+        for (int64_t i = ir; i < ir + dr; i++) {
+            for (int64_t j = 0; j < group_size_in_row; j++) {
+                half scale = calculate_group(i, j);
+                scale_local.SetValue(scale_local_offset++, scale);
+                // Copy Group_Size/2 length data each time.
+                if (scale_local_offset == Group_Size / 2) {
+                    scale_local_offset = 0;
+                    // TODO: OPTIMIZE ME
+                    pipe_barrier(PIPE_ALL);
+                    DataCopy(scale_gm[scale_global_offset], scale_local,
+                                      Group_Size / 2);
+                    pipe_barrier(PIPE_ALL);
+                    scale_global_offset += Group_Size / 2;
+                }
+            }
+        }
+
+        if (scale_local_offset != 0) {
+            pipe_barrier(PIPE_ALL);
+            DataCopyExtParams dataCopyParams;
+            dataCopyParams.blockCount = 1;
+            dataCopyParams.blockLen = scale_local_offset * sizeof(half);
+            DataCopyPad(scale_gm[scale_global_offset], scale_local,
+                        dataCopyParams);
+            pipe_barrier(PIPE_ALL);
+        }
+        scale_queue.FreeTensor(scale_local);
+    }
+
+   private:
+    int64_t input_ne[4];
+    size_t input_stride[4];
+
+    int64_t *scale_ne;
+    size_t scale_stride[4];
+
+    int64_t output_ne[4];
+    size_t output_stride[4];
+
+    int64_t group_size_in_row;
+
+    int64_t ir;
+    int64_t dr;
+
+    TPipe pipe;
+    GlobalTensor<SRC_T> input_gm;
+    GlobalTensor<half> scale_gm;
+    GlobalTensor<int8_t> output_gm;
+    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
+    TQue<QuePosition::VECIN, BUFFER_NUM> work_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> max_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> min_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> scale_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> cast_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> int8_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> half_queue;
+};
+
+template <typename T>
+__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
+    auto gm_ptr = (__gm__ uint8_t *)gm;
+    auto ub_ptr = (uint8_t *)(ub);
+    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
+        *ub_ptr = *gm_ptr;
+    }
+}
+
+extern "C" __global__ __aicore__ void ascendc_quantize_f16_to_q4_0(
+    GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+    GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+    int64_t input_ne_ub[4];
+    size_t input_nb_ub[4];
+    int64_t output_ne_ub[4];
+
+    copy_to_ub(input_ne_gm, input_ne_ub, 32);
+    copy_to_ub(input_nb_gm, input_nb_ub, 32);
+    copy_to_ub(output_ne_gm, output_ne_ub, 32);
+
+    QUANTIZE_FLOAT_TO_Q4_0<half> op;
+    op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
+    op.calculate();
+}
+
+extern "C" __global__ __aicore__ void ascendc_quantize_f32_to_q4_0(
+    GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+    GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+    int64_t input_ne_ub[4];
+    size_t input_nb_ub[4];
+    int64_t output_ne_ub[4];
+
+    copy_to_ub(input_ne_gm, input_ne_ub, 32);
+    copy_to_ub(input_nb_gm, input_nb_ub, 32);
+    copy_to_ub(output_ne_gm, output_ne_ub, 32);
+
+    QUANTIZE_FLOAT_TO_Q4_0<float> op;
+    op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
+    op.calculate();
+}
@@ -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] = {
@@ -130,7 +130,22 @@ static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device)
    }
    return res;
 #else
+
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
+    cudaError_t err;
+    if (getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr)
+    {
+        err = cudaMallocManaged(ptr, size);
+    }
+    else
+    {
+        err = cudaMalloc(ptr, size);
+    }
+    return err;
+#else
    return cudaMalloc(ptr, size);
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
+
 #endif
 }

@@ -167,7 +182,7 @@ 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 +194,7 @@ 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 +330,7 @@ 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 +424,14 @@ 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 +1356,7 @@ 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 +1370,7 @@ 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 +1843,9 @@ 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 +1888,7 @@ 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) {
@@ -1881,10 +1900,9 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
 static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    const bool split = ggml_backend_buffer_is_cuda_split(src0->buffer);

-    bool use_dequantize_mul_mat_vec = (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16)
+    bool use_dequantize_mul_mat_vec = ggml_cuda_dmmv_type_supported(src0->type)
        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
-        && src0->ne[0] % GGML_CUDA_DMMV_X == 0 && src0->ne[0] >= GGML_CUDA_DMMV_X*2
-        && src1->ne[1] == 1;
+        && src0->ne[0] % (GGML_CUDA_DMMV_X*2) == 0 && src1->ne[1] == 1;
    bool          use_mul_mat_vec_q =  ggml_is_quantized(src0->type)
        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
        && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
@@ -3027,7 +3045,7 @@ 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
@@ -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__)
@@ -500,7 +500,7 @@ static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, cons
 }

 static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
    // the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead
    const dim3 block_nums(block_num_y, 1, 1);
@@ -510,7 +510,7 @@ static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y,
 }

 static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
    const dim3 block_nums(block_num_y, 1, 1);
    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
@@ -519,7 +519,7 @@ static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y,
 }

 static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
    const dim3 block_nums(block_num_y, 1, 1);
    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
@@ -528,7 +528,7 @@ static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y,
 }

 static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
    const dim3 block_nums(block_num_y, 1, 1);
    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
@@ -537,7 +537,7 @@ static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y,
 }

 static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
    const dim3 block_nums(block_num_y, 1, 1);
    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
@@ -588,7 +588,7 @@ static void dequantize_mul_mat_vec_q6_K_cuda(const void * vx, const float * y, f
 }

 static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
    const dim3 block_nums(block_num_y, 1, 1);
    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
@@ -672,3 +672,12 @@ void ggml_cuda_op_dequantize_mul_mat_vec(
    GGML_UNUSED(src1_ncols);
    GGML_UNUSED(src1_padded_row_size);
 }
+
+bool ggml_cuda_dmmv_type_supported(ggml_type src0_type) {
+    return src0_type == GGML_TYPE_Q4_0 || src0_type == GGML_TYPE_Q4_1 ||
+        src0_type == GGML_TYPE_Q5_0 || src0_type == GGML_TYPE_Q5_1 ||
+        src0_type == GGML_TYPE_Q8_0 || src0_type == GGML_TYPE_Q2_K ||
+        src0_type == GGML_TYPE_Q3_K || src0_type == GGML_TYPE_Q4_K ||
+        src0_type == GGML_TYPE_Q5_K || src0_type == GGML_TYPE_Q6_K ||
+        src0_type == GGML_TYPE_F16;
+}
@@ -16,3 +16,5 @@ void ggml_cuda_op_dequantize_mul_mat_vec(
    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
    const int64_t src1_padded_row_size, cudaStream_t stream);
+
+bool ggml_cuda_dmmv_type_supported(ggml_type src0_type);
@@ -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
@@ -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
@@ -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;
+}
@@ -80,8 +80,9 @@ static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
 /**
 * Converts float32 to brain16.
 *
- * This function is binary identical to AMD Zen4 VCVTNEPS2BF16.
- * Subnormals shall be flushed to zero, and NANs will be quiet.
+ * This is binary identical with Google Brain float conversion.
+ * Floats shall round to nearest even, and NANs shall be quiet.
+ * Subnormals aren't flushed to zero, except perhaps when used.
 * This code should vectorize nicely if using modern compilers.
 */
 static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
@@ -95,10 +96,6 @@ static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
        h.bits = (u.i >> 16) | 64; /* force to quiet */
        return h;
    }
-    if (!(u.i & 0x7f800000)) { /* subnormal */
-        h.bits = (u.i & 0x80000000) >> 16; /* flush to zero */
-        return h;
-    }
    h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
    return h;
 }
@@ -146,6 +143,7 @@ extern "C" {

 #if defined(__ARM_FEATURE_SVE)
 #include <arm_sve.h>
+#include <sys/prctl.h>
 #endif

 // 16-bit float
@@ -3818,7 +3818,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
    float sumf = 0;

 #if defined(__ARM_FEATURE_SVE)
-    if (svcntb() == QK8_0) {
+    if (ggml_sve_cnt_b == QK8_0) {
        const svbool_t ptrueh = svptrue_pat_b8(SV_VL16);
        const svbool_t ptruel = svnot_b_z(svptrue_b8(), ptrueh);

@@ -5303,7 +5303,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
    float sumf = 0;

 #if defined(__ARM_FEATURE_SVE)
-    if (svcntb() == QK8_0) {
+    if (ggml_sve_cnt_b == QK8_0) {
        svfloat32_t sumv0 = svdup_n_f32(0.0f);
        svfloat32_t sumv1 = svdup_n_f32(0.0f);

@@ -6449,22 +6449,22 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
            // compute mask for subtraction
            vuint8m1_t qh_m0 = __riscv_vand_vx_u8m1(vqh, m, vl);
            vbool8_t vmask_0 = __riscv_vmseq_vx_u8m1_b8(qh_m0, 0, vl);
-            vint8m1_t q3_m0 = __riscv_vsub_vx_i8m1_m(vmask_0, q3_0, 0x4, vl);
+            vint8m1_t q3_m0 = __riscv_vsub_vx_i8m1_mu(vmask_0, q3_0, q3_0, 0x4, vl);
            m <<= 1;

            vuint8m1_t qh_m1 = __riscv_vand_vx_u8m1(vqh, m, vl);
            vbool8_t vmask_1 = __riscv_vmseq_vx_u8m1_b8(qh_m1, 0, vl);
-            vint8m1_t q3_m1 = __riscv_vsub_vx_i8m1_m(vmask_1, q3_1, 0x4, vl);
+            vint8m1_t q3_m1 = __riscv_vsub_vx_i8m1_mu(vmask_1, q3_1, q3_1, 0x4, vl);
            m <<= 1;

            vuint8m1_t qh_m2 = __riscv_vand_vx_u8m1(vqh, m, vl);
            vbool8_t vmask_2 = __riscv_vmseq_vx_u8m1_b8(qh_m2, 0, vl);
-            vint8m1_t q3_m2 = __riscv_vsub_vx_i8m1_m(vmask_2, q3_2, 0x4, vl);
+            vint8m1_t q3_m2 = __riscv_vsub_vx_i8m1_mu(vmask_2, q3_2, q3_2, 0x4, vl);
            m <<= 1;

            vuint8m1_t qh_m3 = __riscv_vand_vx_u8m1(vqh, m, vl);
            vbool8_t vmask_3 = __riscv_vmseq_vx_u8m1_b8(qh_m3, 0, vl);
-            vint8m1_t q3_m3 = __riscv_vsub_vx_i8m1_m(vmask_3, q3_3, 0x4, vl);
+            vint8m1_t q3_m3 = __riscv_vsub_vx_i8m1_mu(vmask_3, q3_3, q3_3, 0x4, vl);
            m <<= 1;

            // load Q8 and take product with Q3
@@ -7720,13 +7720,13 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
            vint8m1_t q5_a = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q5_x, 0x0F, vl));
            vuint8m1_t qh_m1 = __riscv_vand_vx_u8m1(vqh, m, vl);
            vbool8_t vmask_1 = __riscv_vmsne_vx_u8m1_b8(qh_m1, 0, vl);
-            vint8m1_t q5_m1 = __riscv_vadd_vx_i8m1_m(vmask_1, q5_a, 16, vl);
+            vint8m1_t q5_m1 = __riscv_vadd_vx_i8m1_mu(vmask_1, q5_a, q5_a, 16, vl);
            m <<= 1;

            vint8m1_t q5_l = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(q5_x, 0x04, vl));
            vuint8m1_t qh_m2 = __riscv_vand_vx_u8m1(vqh, m, vl);
            vbool8_t vmask_2 = __riscv_vmsne_vx_u8m1_b8(qh_m2, 0, vl);
-            vint8m1_t q5_m2 = __riscv_vadd_vx_i8m1_m(vmask_2, q5_l, 16, vl);
+            vint8m1_t q5_m2 = __riscv_vadd_vx_i8m1_mu(vmask_2, q5_l, q5_l, 16, vl);
            m <<= 1;

            vint16m2_t v0 = __riscv_vwmul_vv_i16m2(q5_m1, q8_y1, vl);
@@ -127,6 +127,10 @@ void iq2xs_free_impl(enum ggml_type type);
 void iq3xs_init_impl(int grid_size);
 void iq3xs_free_impl(int grid_size);

+#if defined(__ARM_FEATURE_SVE)
+extern int ggml_sve_cnt_b;
+#endif
+
 #ifdef __cplusplus
 }
 #endif
@@ -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;
@@ -4105,6 +4108,9 @@ bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tens
        case GGML_OP_ARGSORT:
            func = ggml_sycl_argsort;
            break;
+        case GGML_OP_TIMESTEP_EMBEDDING:
+            func = ggml_sycl_op_timestep_embedding;
+            break;
        default:
            return false;
    }
@@ -5090,6 +5096,15 @@ 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:
@@ -5213,6 +5228,7 @@ GGML_CALL static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, cons
        case GGML_OP_UPSCALE:
        case GGML_OP_PAD:
        case GGML_OP_LEAKY_RELU:
+        case GGML_OP_TIMESTEP_EMBEDDING:
            return true;
        default:
            return false;
@@ -15,6 +15,7 @@

 #include "concat.hpp"
 #include "common.hpp"
+#include "conv.hpp"
 #include "convert.hpp"
 #include "dequantize.hpp"
 #include "dmmv.hpp"
@@ -23,5 +24,6 @@
 #include "rope.hpp"
 #include "norm.hpp"
 #include "softmax.hpp"
+#include "tsembd.hpp"

 #endif // GGML_SYCL_BACKEND_HPP
@@ -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);
+}
+
@@ -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
@@ -902,7 +902,7 @@ static void mul_mat_vec_iq4_nl_q8_1_sycl(const void *vx, const void *vy,
                sycl::nd_range<3>(block_nums * block_dims, block_dims),
                [=](sycl::nd_item<3> item_ct1)
                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q_iq4_nl_q8_1<QK4_NL, QI4_NL, block_iq4_nl, 1>(
+                        mul_mat_vec_q_iq4_nl_q8_1<QK4_NL, QI4_NL, block_iq4_nl, 2>(
                            vx, vy, dst, ncols, nrows, item_ct1);
                    });
        });
@@ -41,6 +41,8 @@
 #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
+#define SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE 256

 // dmmv = dequantize_mul_mat_vec
 #ifndef GGML_SYCL_DMMV_X
@@ -0,0 +1,71 @@
+//
+// 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 "tsembd.hpp"
+
+static void timestep_embedding_f32(
+        const float * timesteps, float * dst, const int nb1,
+        const int dim, const int max_period, const sycl::nd_item<3> &item_ct1) {
+    // item_ct1.get_group(1)(blockIDx.y): idx of timesteps->ne[0]
+    // item_ct1.get_group(2) (blockIDx.x): idx of ((dim + 1) / 2) / BLOCK_SIZE
+    int i = item_ct1.get_group(1);
+    int j = item_ct1.get_local_id(2) + item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    float * embed_data = (float *)((char *)dst +  i*nb1);
+
+    if (dim % 2 != 0 && j == ((dim + 1) / 2)) {
+        embed_data[dim] = 0.f;
+    }
+
+    int half = dim / 2;
+    if (j >= half) {
+        return;
+    }
+
+    float timestep = timesteps[i];
+    float freq = (float)sycl::native::exp(-(sycl::log((float)max_period)) * j / half);
+    float arg = timestep * freq;
+    embed_data[j] = sycl::cos(arg);
+    embed_data[j + half] = sycl::sin(arg);
+}
+
+static void timestep_embedding_f32_sycl(
+        const float * x, float * dst, const int ne00, const int nb1,
+        const int dim, const int max_period, const queue_ptr& stream) {
+    // As the kernel returns when thread.idx is larger than dim/2, the half_ceil does not need to pad
+    int half_ceil = dim / 2;
+    int num_blocks = (half_ceil + SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE - 1) / SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE;
+    sycl::range<3> block_dims(1, 1, SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE);
+    sycl::range<3> gridDim(1, ne00, num_blocks);
+    stream->parallel_for(
+        sycl::nd_range<3>(
+            gridDim * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) {
+            timestep_embedding_f32(
+                x, dst, nb1, dim, max_period, item_ct1
+            );
+        });
+}
+
+void ggml_sycl_op_timestep_embedding(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+    const ggml_tensor *src1, ggml_tensor * dst) {
+    const float * src0_d = (const float *)src0->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);
+
+    const int dim = dst->op_params[0];
+    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);
+}
@@ -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_TSEMBD_HPP
+#define GGML_SYCL_TSEMBD_HPP
+
+#include "common.hpp"
+
+void ggml_sycl_op_timestep_embedding(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+    const ggml_tensor *src1, ggml_tensor * dst);
+
+#endif // GGML_SYCL_TSEMBD_HPP
@@ -37,6 +37,9 @@
 #include <unistd.h>
 #endif

+#if defined(__ARM_FEATURE_SVE)
+int ggml_sve_cnt_b = 0;
+#endif
 #if defined(__ARM_FEATURE_SVE) || defined(__ARM_FEATURE_MATMUL_INT8)
 #undef GGML_USE_LLAMAFILE
 #endif
@@ -141,7 +144,51 @@ typedef pthread_t ggml_thread_t;

 #include <sys/wait.h>

-#if defined(__linux__)
+#if defined(__ANDROID__)
+#include <unwind.h>
+#include <dlfcn.h>
+#include <stdio.h>
+
+struct backtrace_state {
+    void ** current;
+    void ** end;
+};
+
+static _Unwind_Reason_Code unwind_callback(struct _Unwind_Context* context, void* arg) {
+    struct backtrace_state * state = (struct backtrace_state *)arg;
+    uintptr_t pc = _Unwind_GetIP(context);
+    if (pc) {
+        if (state->current == state->end) {
+            return _URC_END_OF_STACK;
+        } else {
+            *state->current++ = (void*)pc;
+        }
+    }
+    return _URC_NO_REASON;
+}
+
+static void ggml_print_backtrace_symbols(void) {
+    const int max = 100;
+    void* buffer[max];
+
+    struct backtrace_state state = {buffer, buffer + max};
+    _Unwind_Backtrace(unwind_callback, &state);
+
+    int count = state.current - buffer;
+
+    for (int idx = 0; idx < count; ++idx) {
+        const void * addr = buffer[idx];
+        const char * symbol = "";
+
+        Dl_info info;
+        if (dladdr(addr, &info) && info.dli_sname) {
+            symbol = info.dli_sname;
+        }
+
+        fprintf(stderr, "%d: %p %s\n", idx, addr, symbol);
+    }
+}
+#elif defined(__linux__) && defined(__GLIBC__)
 #include <execinfo.h>
 static void ggml_print_backtrace_symbols(void) {
    void * trace[100];
@@ -436,9 +483,16 @@ void ggml_bf16_to_fp32_row(const ggml_bf16_t * x, float * y, int64_t n) {
    }
 }

+void ggml_fp32_to_bf16_row_ref(const float * x, ggml_bf16_t * y, int64_t n) {
+    for (int i = 0; i < n; i++) {
+        y[i] = ggml_compute_fp32_to_bf16(x[i]);
+    }
+}
+
 void ggml_fp32_to_bf16_row(const float * x, ggml_bf16_t * y, int64_t n) {
  int i = 0;
 #if defined(__AVX512BF16__)
+  // subnormals are flushed to zero on this platform
  for (; i + 32 <= n; i += 32) {
        _mm512_storeu_si512(
            (__m512i *)(y + i),
@@ -918,7 +972,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
        .is_quantized             = false,
        .to_float                 = (ggml_to_float_t) ggml_bf16_to_fp32_row,
        .from_float               = (ggml_from_float_t) ggml_fp32_to_bf16_row,
-        .from_float_ref           = (ggml_from_float_t) ggml_fp32_to_bf16_row,
+        .from_float_ref           = (ggml_from_float_t) ggml_fp32_to_bf16_row_ref,
        .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_bf16,
        .vec_dot_type             = GGML_TYPE_BF16,
        .nrows                    = 1,
@@ -2258,7 +2312,7 @@ inline static void ggml_vec_abs_f32  (const int n, float * y, const float * x) {
 inline static void ggml_vec_sgn_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : ((x[i] < 0.f) ? -1.f : 0.f); }
 inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : 0.f; }
 inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = tanhf(x[i]);  }
-inline static void ggml_vec_elu_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expf(x[i])-1; }
+inline static void ggml_vec_elu_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expm1f(x[i]); }
 inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; }
 inline static void ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); }
 inline static void ggml_vec_sigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = 1.f / (1.f + expf(-x[i])); }
@@ -3507,6 +3561,12 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {

    GGML_ASSERT_ALIGNED(ctx->mem_buffer);

+#if defined(__ARM_FEATURE_SVE)
+    if (!ggml_sve_cnt_b) {
+        ggml_sve_cnt_b = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
+    }
+#endif
+
    GGML_PRINT_DEBUG("%s: context initialized\n", __func__);

    ggml_critical_section_end();
@@ -20606,7 +20666,7 @@ size_t ggml_quantize_chunk(
        case GGML_TYPE_BF16:
            {
                size_t elemsize = sizeof(ggml_bf16_t);
-                ggml_fp32_to_bf16_row(src + start, (ggml_bf16_t *)dst + start, n);
+                ggml_fp32_to_bf16_row_ref(src + start, (ggml_bf16_t *)dst + start, n);
                result = n * elemsize;
            } break;
        case GGML_TYPE_F32:
@@ -4,9 +4,11 @@
 #include "generic_binary_head.comp"

 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
        return;
    }

-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) + FLOAT_TYPE(data_b[src1_idx(gl_GlobalInvocationID.x)]));
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) + FLOAT_TYPE(data_b[src1_idx(idx)]));
 }
@@ -4,10 +4,12 @@
 #include "generic_unary_head.comp"

 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
        return;
    }

-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]);
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
 }
@@ -0,0 +1,35 @@
+#version 450
+
+#include "types.comp"
+#include "generic_binary_head.comp"
+
+void main() {
+    const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+    const int dim = p.param3;
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint i3 = idx / (p.ne22*p.ne21*p.ne20);
+    const uint i3_offset = i3 * p.ne22*p.ne21*p.ne20;
+    const uint i2 = (idx - i3_offset) / (p.ne21*p.ne20);
+    const uint i2_offset = i2*p.ne21*p.ne20;
+    const uint i1 = (idx - i3_offset - i2_offset) / p.ne20;
+    const uint i0 = idx - i3_offset - i2_offset - i1*p.ne20;
+
+    uint o[4] = {0, 0, 0, 0};
+    o[dim] = dim == 0 ? p.ne00 : (dim == 1 ? p.ne01 : (dim == 2 ? p.ne02 : p.ne03));
+
+    const uint src0_idx = i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0*p.nb00;
+    const uint src1_idx = (i3 - o[3])*p.nb13 + (i2 - o[2])*p.nb12 + (i1 - o[1])*p.nb11 + (i0 - o[0])*p.nb10;
+    const uint dst_idx = i3*p.nb23 + i2*p.nb22 + i1*p.nb21 + i0*p.nb20;
+
+    const bool is_src0 = i0 < p.ne00 && i1 < p.ne01 && i2 < p.ne02 && i3 < p.ne03;
+
+#ifndef OPTIMIZATION_ERROR_WORKAROUND
+    data_d[p.d_offset + dst_idx] = D_TYPE(is_src0 ? data_a[src0_idx] : data_b[src1_idx]);
+#else
+    data_d[p.d_offset + dst_idx] = is_src0 ? data_a[src0_idx] : data_b[src1_idx];
+#endif
+}
@@ -4,13 +4,15 @@
 #include "generic_unary_head.comp"

 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
        return;
    }

 #ifndef OPTIMIZATION_ERROR_WORKAROUND
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]);
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(data_a[src0_idx(idx)]);
 #else
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = data_a[src0_idx(gl_GlobalInvocationID.x)];
+    data_d[p.d_offset + dst_idx(idx)] = data_a[src0_idx(idx)];
 #endif
 }
@@ -4,9 +4,11 @@
 #include "generic_binary_head.comp"

 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
        return;
    }

-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) / FLOAT_TYPE(data_b[src1_idx(gl_GlobalInvocationID.x)]));
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) / FLOAT_TYPE(data_b[src1_idx(idx)]));
 }
@@ -13,7 +13,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 void main() {
    const float GELU_COEF_A    = 0.044715f;
    const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
-    const uint i = gl_GlobalInvocationID.x;
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;

    if (i >= p.KX) {
        return;
@@ -0,0 +1,23 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const float GELU_QUICK_COEF = -1.702f;
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float x = float(data_a[i]);
+    data_d[i] = D_TYPE(x * (1.0f / (1.0f + exp(GELU_QUICK_COEF * x))));
+}
@@ -7,7 +7,7 @@ layout (push_constant) uniform parameter
    uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
    uint ne20; uint ne21; uint ne22; uint ne23; uint nb20; uint nb21; uint nb22; uint nb23;
    uint d_offset;
-    float param1; float param2;
+    float param1; float param2; int param3;
 } p;

 layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
@@ -16,6 +16,10 @@ layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
 layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};

+uint get_idx() {
+    return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+}
+
 uint src0_idx(uint idx) {
    const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
    const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
@@ -14,6 +14,10 @@ layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};

+uint get_idx() {
+    return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+}
+
 uint src0_idx(uint idx) {
    const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
    const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
@@ -0,0 +1,66 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+#define BLOCK_SIZE 512
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+shared float tmp[BLOCK_SIZE];
+
+void main() {
+    const uint group_size = p.KX;
+    const float eps = p.param1;
+
+    const uint tid = gl_LocalInvocationID.x;
+    const uint start = gl_WorkGroupID.x * group_size + tid;
+    const uint end = start + group_size;
+
+    tmp[tid] = 0.0f;
+
+    // Calculate mean
+    [[unroll]] for (uint col = start; col < end; col += BLOCK_SIZE) {
+        tmp[tid] += float(data_a[col]);
+    }
+
+    // tmp up partial tmps and write back result
+    barrier();
+    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+
+    const float mean = tmp[0] / group_size;
+    barrier();
+    tmp[tid] = 0.0f;
+
+    // Calculate variance
+    [[unroll]] for (uint col = start; col < end; col += BLOCK_SIZE) {
+        const float xi = float(data_a[col]) - mean;
+        data_d[col] = D_TYPE(xi);
+        tmp[tid] += xi * xi;
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+
+    const float variance = tmp[0] / group_size;
+    const float scale = inversesqrt(variance + eps);
+
+    [[unroll]] for (uint col = start; col < end; col += BLOCK_SIZE) {
+        data_d[col] *= D_TYPE(scale);
+    }
+}
@@ -0,0 +1,57 @@
+#version 450
+
+#extension GL_EXT_shader_16bit_storage : require
+
+layout (push_constant) uniform parameter
+{
+    uint batch_offset; uint offset_delta;
+    uint IC;
+    uint IW; uint IH;
+    uint OW; uint OH;
+    uint KW; uint KH;
+    uint pelements;
+    uint CHW;
+    int s0; int s1;
+    int p0; int p1;
+    int d0; int d1;
+} p;
+
+#include "types.comp"
+
+#define BLOCK_SIZE 256
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.x;
+    if (i >= p.pelements) {
+        return;
+    }
+
+    const uint ksize = p.OW * (p.KH > 1 ? p.KW : 1);
+    const uint kx = i / ksize;
+    const uint kd = kx * ksize;
+    const uint ky = (i - kd) / p.OW;
+    const uint ix = i % p.OW;
+
+    const uint oh = gl_GlobalInvocationID.y;
+    const uint batch = gl_GlobalInvocationID.z / p.IC;
+    const uint ic = gl_GlobalInvocationID.z % p.IC;
+
+    const uint iiw = ix * p.s0 + kx * p.d0 - p.p0;
+    const uint iih = oh * p.s1 + ky * p.d1 - p.p1;
+
+    const uint offset_dst =
+        ((batch * p.OH + oh) * p.OW + ix) * p.CHW +
+        (ic * (p.KW * p.KH) + ky * p.KW + kx);
+
+    if (iih < 0 || iih >= p.IH || iiw < 0 || iiw >= p.IW) {
+        data_d[offset_dst] = D_TYPE(0.0f);
+    } else {
+        const uint offset_src = ic * p.offset_delta + batch * p.batch_offset;
+        data_d[offset_dst] = D_TYPE(data_a[offset_src + iih * p.IW + iiw]);
+    }
+}
@@ -0,0 +1,22 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float val = float(data_a[i]);
+    data_d[i] = D_TYPE(max(val, 0.0f) + min(val, 0.0f) * p.param1);
+}
@@ -4,9 +4,11 @@
 #include "generic_binary_head.comp"

 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
        return;
    }

-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) * FLOAT_TYPE(data_b[src1_idx(gl_GlobalInvocationID.x)]));
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) * FLOAT_TYPE(data_b[src1_idx(idx)]));
 }
@@ -16,6 +16,13 @@ void main() {
    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
    const uint tid = gl_LocalInvocationID.x;

+    // There are not enough cols to use all threads
+    if (tid >= p.ncols) {
+        return;
+    }
+
+    const uint block_size = min(p.ncols, BLOCK_SIZE);
+
    uint a_offset, b_offset, d_offset;
    get_offsets(a_offset, b_offset, d_offset);

@@ -23,8 +30,8 @@ void main() {

    tmp[tid] = FLOAT_TYPE(0.0f);

-    [[unroll]] for (uint i = 0; i < p.ncols/BLOCK_SIZE; i += 2) {
-        const uint col = i*BLOCK_SIZE + 2*tid;
+    [[unroll]] for (uint i = 0; i < p.ncols/block_size; i += 2) {
+        const uint col = i*block_size + 2*tid;
        const uint ib = (row*p.ncols + col)/QUANT_K; // block index
        const uint iqs = (col%QUANT_K)/QUANT_R; // quant index
        const uint iybs = col - col%QUANT_K; // y block start index
@@ -38,7 +45,7 @@ void main() {

    // sum up partial sums and write back result
    barrier();
-    [[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
+    [[unroll]] for (uint s = block_size/2; s > 0; s >>= 1) {
        if (tid < s) {
            tmp[tid] += tmp[tid + s];
        }
@@ -14,7 +14,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 shared vec2 sum[BLOCK_SIZE];

 void main() {
-    const uint row = gl_WorkGroupID.x;
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
    const uint tid = gl_LocalInvocationID.x;

    sum[tid] = vec2(0.0f, 0.0f);
@@ -0,0 +1,26 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+void main() {
+    const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint i3 = idx / (p.ne12*p.ne11*p.ne10);
+    const uint i3_offset = i3 * p.ne12*p.ne11*p.ne10;
+    const uint i2 = (idx - i3_offset) / (p.ne11*p.ne10);
+    const uint i2_offset = i2*p.ne11*p.ne10;
+    const uint i1 = (idx - i3_offset - i2_offset) / p.ne10;
+    const uint i0 = idx - i3_offset - i2_offset - i1*p.ne10;
+
+    const uint src0_idx = i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0*p.nb00;
+    const uint dst_idx = i3*p.nb13 + i2*p.nb12 + i1*p.nb11 + i0*p.nb10;
+
+    const bool is_src0 = i0 < p.ne00 && i1 < p.ne01 && i2 < p.ne02 && i3 < p.ne03;
+
+    data_d[p.d_offset + dst_idx] = D_TYPE(is_src0 ? data_a[src0_idx] : 0.0f);
+}
@@ -11,7 +11,7 @@ layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
 layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};

 void main() {
-    const uint i = gl_GlobalInvocationID.x;
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;

    if (i >= p.KX) {
        return;
@@ -14,7 +14,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 shared FLOAT_TYPE sum[BLOCK_SIZE];

 void main() {
-    const uint row = gl_WorkGroupID.x;
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
    const uint tid = gl_LocalInvocationID.x;

    sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp
@@ -4,9 +4,11 @@
 #include "generic_unary_head.comp"

 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
        return;
    }

-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) * FLOAT_TYPE(p.param1));
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) * FLOAT_TYPE(p.param1));
 }
@@ -11,7 +11,7 @@ layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
 layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};

 void main() {
-    const uint i = gl_GlobalInvocationID.x;
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;

    if (i >= p.KX) {
        return;
@@ -28,7 +28,7 @@ shared FLOAT_TYPE vals[BLOCK_SIZE];

 void main() {
    const uint tid = gl_LocalInvocationID.x;
-    const uint rowx = gl_WorkGroupID.x;
+    const uint rowx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
    const uint rowy = rowx % p.KY;

    float slope = 1.0f;
@@ -4,10 +4,12 @@
 #include "generic_unary_head.comp"

 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
        return;
    }

-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]);
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(val * val);
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(val * val);
 }
@@ -14,7 +14,7 @@ layout (constant_id = 0) const uint BLOCK_SIZE = 32;
 shared FLOAT_TYPE tmp[BLOCK_SIZE];

 void main() {
-    const uint row = gl_WorkGroupID.x;
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
    const uint col = gl_LocalInvocationID.x;

    tmp[col] = FLOAT_TYPE(0.0f);
@@ -0,0 +1,21 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    data_d[i] = D_TYPE(tanh(data_a[i]));
+}
@@ -0,0 +1,41 @@
+#version 450
+
+#extension GL_EXT_shader_16bit_storage : require
+
+layout (push_constant) uniform parameter
+{
+    uint nb1;
+    uint dim;
+    uint max_period;
+} p;
+
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+#define BLOCK_SIZE 256
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_WorkGroupID.y;
+    const uint j = gl_GlobalInvocationID.x;
+    const uint d_offset = i * p.nb1;
+
+    if (p.dim % 2 != 0 && j == ((p.dim + 1) / 2)) {
+        data_d[d_offset + p.dim] = 0.f;
+    }
+
+    const uint half_dim = p.dim / 2;
+    if (j >= half_dim) {
+        return;
+    }
+
+    const float timestep = float(data_a[i]);
+    const float freq = float(exp(-log(p.max_period) * j / half_dim));
+    const float arg = timestep * freq;
+    data_d[d_offset + j] = D_TYPE(cos(arg));
+    data_d[d_offset + j + half_dim] = D_TYPE(sin(arg));
+}
@@ -6,7 +6,7 @@
 #define QUANT_K 1
 #define QUANT_R 1

-#ifndef LOAD_VEC_A
+#if !defined(LOAD_VEC_A) || LOAD_VEC_A == 1
 #define A_TYPE float
 #elif LOAD_VEC_A == 4
 #define A_TYPE vec4
@@ -19,7 +19,7 @@
 #define QUANT_K 1
 #define QUANT_R 1

-#ifndef LOAD_VEC_A
+#if !defined(LOAD_VEC_A) || LOAD_VEC_A == 1
 #define A_TYPE float16_t
 #elif LOAD_VEC_A == 4
 #define A_TYPE f16vec4
@@ -0,0 +1,36 @@
+#version 450
+
+layout (push_constant) uniform parameter
+{
+    uint ne; uint d_offset;
+    uint nb00; uint nb01; uint nb02; uint nb03;
+    uint ne10; uint ne11; uint ne12; uint ne13;
+    float sf0; float sf1; float sf2; float sf3;
+} p;
+
+#include "types.comp"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint i10 = idx % p.ne10;
+    const uint i11 = (idx / p.ne10) % p.ne11;
+    const uint i12 = (idx / (p.ne10 * p.ne11)) % p.ne12;
+    const uint i13 = (idx / (p.ne10 * p.ne11 * p.ne12)) % p.ne13;
+
+    const uint i00 = uint(i10 / p.sf0);
+    const uint i01 = uint(i11 / p.sf1);
+    const uint i02 = uint(i12 / p.sf2);
+    const uint i03 = uint(i13 / p.sf3);
+
+    data_d[p.d_offset + idx] = D_TYPE(data_a[i03 * p.nb03 + i02 * p.nb02 + i01 * p.nb01 + i00 * p.nb00]);
+}
@@ -179,11 +179,7 @@ bool string_ends_with(const std::string& str, const std::string& suffix) {
    return std::equal(suffix.rbegin(), suffix.rend(), str.rbegin());
 }

-#ifdef _WIN32
-    static const char path_separator = '\\';
-#else
-    static const char path_separator = '/';
-#endif
+static const char path_separator = '/';

 std::string join_paths(const std::string& path1, const std::string& path2) {
    return path1 + path_separator + path2;
@@ -198,7 +194,11 @@ void string_to_spv(const std::string& _name, const std::string& in_fname, const
    std::string out_fname = join_paths(output_dir, name + ".spv");
    std::string in_path = join_paths(input_dir, in_fname);

-    std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", in_path, "-o", out_fname};
+    #ifdef _WIN32
+        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", "\"" + in_path + "\"", "-o", "\"" + out_fname + "\""};
+    #else
+        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", in_path, "-o",  out_fname};
+    #endif
    for (const auto& define : defines) {
        cmd.push_back("-D" + define.first + "=" + define.second);
    }
@@ -269,9 +269,12 @@ void matmul_shaders(std::vector<std::future<void>>& tasks, bool fp16, bool matmu

    for (const auto& tname : type_names) {
        std::string data_a_key = "DATA_A_" + to_uppercase(tname);
+        // For unaligned, load one at a time for f32/f16, or two at a time for quants
+        std::string load_vec_a_unaligned = (tname == "f32" || tname == "f16") ? "1" : "2";
+        // For aligned matmul loads
        std::string load_vec_a = (tname == "f32" || tname == "f16") ? load_vec : "2";
        tasks.push_back(std::async(std::launch::async, [=] {
-            string_to_spv(shader_name + "_" + tname + "_f32", "mul_mm.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16);
+            string_to_spv(shader_name + "_" + tname + "_f32", "mul_mm.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16);
        }));
        tasks.push_back(std::async(std::launch::async, [=] {
            string_to_spv(shader_name + "_" + tname + "_f32_aligned", "mul_mm.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}}), fp16);
@@ -340,6 +343,9 @@ void process_shaders(std::vector<std::future<void>>& tasks) {
    tasks.push_back(std::async(std::launch::async, [=] {
        string_to_spv("norm_f32", "norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
    }));
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("group_norm_f32", "group_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+    }));
    tasks.push_back(std::async(std::launch::async, [=] {
        string_to_spv("rms_norm_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
    }));
@@ -357,6 +363,9 @@ void process_shaders(std::vector<std::future<void>>& tasks) {
    tasks.push_back(std::async(std::launch::async, [] {
        string_to_spv("add_f32", "add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("add_f16_f32_f16", "add.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"FLOAT_TYPE", "float"}});
+    }));

    tasks.push_back(std::async(std::launch::async, [] {
        string_to_spv("split_k_reduce", "mul_mat_split_k_reduce.comp", {});
@@ -382,15 +391,42 @@ void process_shaders(std::vector<std::future<void>>& tasks) {
        string_to_spv("clamp_f32", "clamp.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
    }));

+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("pad_f32", "pad.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("concat_f32", "concat.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("concat_f16", "concat.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("concat_i32", "concat.comp", {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("upscale_f32", "upscale.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+
    tasks.push_back(std::async(std::launch::async, [] {
        string_to_spv("gelu_f32", "gelu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("gelu_quick_f32", "gelu_quick.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
    tasks.push_back(std::async(std::launch::async, [] {
        string_to_spv("silu_f32", "silu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
    }));
    tasks.push_back(std::async(std::launch::async, [] {
        string_to_spv("relu_f32", "relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("leaky_relu_f32", "leaky_relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("tanh_f32", "tanh.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));

    tasks.push_back(std::async(std::launch::async, [] {
        string_to_spv("diag_mask_inf_f32", "diag_mask_inf.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
@@ -424,6 +460,17 @@ void process_shaders(std::vector<std::future<void>>& tasks) {
    tasks.push_back(std::async(std::launch::async, [=] {
        string_to_spv("sum_rows_f32", "sum_rows.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
    }));
+
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("im2col_f32", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+    }));
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("im2col_f32_f16", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}}));
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("timestep_embedding_f32", "timestep_embedding.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+    }));
 }

 void write_output_files() {
@@ -435,10 +482,16 @@ void write_output_files() {

    for (const auto& pair : shader_fnames) {
        const std::string& name = pair.first;
-        const std::string& path = pair.second;
+        #ifdef _WIN32
+            std::string path = pair.second;
+            std::replace(path.begin(), path.end(), '/', '\\' );
+        #else
+            const std::string& path = pair.second;
+        #endif
+
        FILE* spv = fopen(path.c_str(), "rb");
        if (!spv) {
-            std::cerr << "Error opening SPIR-V file: " << path << "\n";
+            std::cerr << "Error opening SPIR-V file: " << path << " (" << strerror(errno) << ")\n";
            continue;
        }

@@ -450,7 +503,7 @@ void write_output_files() {
        size_t read_size = fread(data.data(), 1, size, spv);
        fclose(spv);
        if (read_size != size) {
-            std::cerr << "Error reading SPIR-V file: " << path << "\n";
+            std::cerr << "Error reading SPIR-V file: " << path << " (" << strerror(errno) << ")\n";
            continue;
        }

@@ -161,6 +161,7 @@ class Keys:
        SUFFIX_ID            = "tokenizer.ggml.suffix_token_id"
        MIDDLE_ID            = "tokenizer.ggml.middle_token_id"
        EOT_ID               = "tokenizer.ggml.eot_token_id"
+        EOM_ID               = "tokenizer.ggml.eom_token_id"

    class Adapter:
        TYPE       = "adapter.type"
@@ -1327,3 +1328,4 @@ KEY_TOKENIZER_PRIFIX_ID  = Keys.Tokenizer.PREFIX_ID
 KEY_TOKENIZER_SUFFIX_ID  = Keys.Tokenizer.SUFFIX_ID
 KEY_TOKENIZER_MIDDLE_ID  = Keys.Tokenizer.MIDDLE_ID
 KEY_TOKENIZER_EOT_ID     = Keys.Tokenizer.EOT_ID
+KEY_TOKENIZER_EOM_ID     = Keys.Tokenizer.EOM_ID
@@ -312,6 +312,8 @@ class GGUFWriter:
        self.add_key_value(key, val, GGUFValueType.STRING)

    def add_array(self, key: str, val: Sequence[Any]) -> None:
+        if len(val) == 0:
+            return
        self.add_key_value(key, val, GGUFValueType.ARRAY)

    @staticmethod
@@ -826,6 +828,9 @@ class GGUFWriter:
    def add_eot_token_id(self, id: int) -> None:
        self.add_uint32(Keys.Tokenizer.EOT_ID, id)

+    def add_eom_token_id(self, id: int) -> None:
+        self.add_uint32(Keys.Tokenizer.EOM_ID, id)
+
    def _pack(self, fmt: str, value: Any, skip_pack_prefix: bool = False) -> bytes:
        pack_prefix = ''
        if not skip_pack_prefix:
@@ -845,7 +850,14 @@ class GGUFWriter:
            encoded_val = val.encode("utf-8") if isinstance(val, str) else val
            kv_data += self._pack("Q", len(encoded_val))
            kv_data += encoded_val
-        elif vtype == GGUFValueType.ARRAY and isinstance(val, Sequence) and val:
+        elif vtype == GGUFValueType.ARRAY:
+
+            if not isinstance(val, Sequence):
+                raise ValueError("Invalid GGUF metadata array, expecting sequence")
+
+            if len(val) == 0:
+                raise ValueError("Invalid GGUF metadata array. Empty array")
+
            if isinstance(val, bytes):
                ltype = GGUFValueType.UINT8
            else:
@@ -284,20 +284,67 @@ class Metadata:
        ########################
        if model_card is not None:

-            if "model_name" in model_card and metadata.name is None:
-                # Not part of huggingface model card standard but notice some model creator using it
-                # such as TheBloke in 'TheBloke/Mistral-7B-Instruct-v0.2-GGUF'
-                metadata.name = model_card.get("model_name")
+            def use_model_card_metadata(metadata_key: str, model_card_key: str):
+                if model_card_key in model_card and getattr(metadata, metadata_key, None) is None:
+                    setattr(metadata, metadata_key, model_card.get(model_card_key))

-            if "model_creator" in model_card and metadata.author is None:
-                # Not part of huggingface model card standard but notice some model creator using it
-                # such as TheBloke in 'TheBloke/Mistral-7B-Instruct-v0.2-GGUF'
-                metadata.author = model_card.get("model_creator")
+            def use_array_model_card_metadata(metadata_key: str, model_card_key: str):
+                # Note: Will append rather than replace if already exist
+                tags_value = model_card.get(model_card_key, None)
+                if tags_value is None:
+                    return

-            if "model_type" in model_card and metadata.basename is None:
-                # Not part of huggingface model card standard but notice some model creator using it
-                # such as TheBloke in 'TheBloke/Mistral-7B-Instruct-v0.2-GGUF'
-                metadata.basename = model_card.get("model_type")
+                current_value = getattr(metadata, metadata_key, None)
+                if current_value is None:
+                    current_value = []
+
+                if isinstance(tags_value, str):
+                    current_value.append(tags_value)
+                elif isinstance(tags_value, list):
+                    current_value.extend(tags_value)
+
+                setattr(metadata, metadata_key, current_value)
+
+            # LLAMA.cpp's direct internal convention
+            # (Definitely not part of hugging face formal/informal standard)
+            #########################################
+            use_model_card_metadata("name", "name")
+            use_model_card_metadata("author", "author")
+            use_model_card_metadata("version", "version")
+            use_model_card_metadata("organization", "organization")
+            use_model_card_metadata("description", "description")
+            use_model_card_metadata("finetune", "finetune")
+            use_model_card_metadata("basename", "basename")
+            use_model_card_metadata("size_label", "size_label")
+            use_model_card_metadata("source_url", "url")
+            use_model_card_metadata("source_doi", "doi")
+            use_model_card_metadata("source_uuid", "uuid")
+            use_model_card_metadata("source_repo_url", "repo_url")
+
+            # LLAMA.cpp's huggingface style convention
+            # (Definitely not part of hugging face formal/informal standard... but with model_ appended to match their style)
+            ###########################################
+            use_model_card_metadata("name", "model_name")
+            use_model_card_metadata("author", "model_author")
+            use_model_card_metadata("version", "model_version")
+            use_model_card_metadata("organization", "model_organization")
+            use_model_card_metadata("description", "model_description")
+            use_model_card_metadata("finetune", "model_finetune")
+            use_model_card_metadata("basename", "model_basename")
+            use_model_card_metadata("size_label", "model_size_label")
+            use_model_card_metadata("source_url", "model_url")
+            use_model_card_metadata("source_doi", "model_doi")
+            use_model_card_metadata("source_uuid", "model_uuid")
+            use_model_card_metadata("source_repo_url", "model_repo_url")
+
+            # Hugging Face Direct Convention
+            #################################
+
+            # Not part of huggingface model card standard but notice some model creator using it
+            # such as TheBloke in 'TheBloke/Mistral-7B-Instruct-v0.2-GGUF'
+            use_model_card_metadata("name", "model_name")
+            use_model_card_metadata("author", "model_creator")
+            use_model_card_metadata("basename", "model_type")

            if "base_model" in model_card:
                # This represents the parent models that this is based on
@@ -329,58 +376,18 @@ class Metadata:
                        base_model["repo_url"] = f"https://huggingface.co/{org_component}/{model_full_name_component}"
                    metadata.base_models.append(base_model)

-            if "license" in model_card and metadata.license is None:
-                metadata.license = model_card.get("license")
+            use_model_card_metadata("license", "license")
+            use_model_card_metadata("license_name", "license_name")
+            use_model_card_metadata("license_link", "license_link")

-            if "license_name" in model_card and metadata.license_name is None:
-                metadata.license_name = model_card.get("license_name")
+            use_array_model_card_metadata("tags", "tags")
+            use_array_model_card_metadata("tags", "pipeline_tag")

-            if "license_link" in model_card and metadata.license_link is None:
-                metadata.license_link = model_card.get("license_link")
+            use_array_model_card_metadata("languages", "languages")
+            use_array_model_card_metadata("languages", "language")

-            tags_value = model_card.get("tags", None)
-            if tags_value is not None:
-
-                if metadata.tags is None:
-                    metadata.tags = []
-
-                if isinstance(tags_value, str):
-                    metadata.tags.append(tags_value)
-                elif isinstance(tags_value, list):
-                    metadata.tags.extend(tags_value)
-
-            pipeline_tags_value = model_card.get("pipeline_tag", None)
-            if pipeline_tags_value is not None:
-
-                if metadata.tags is None:
-                    metadata.tags = []
-
-                if isinstance(pipeline_tags_value, str):
-                    metadata.tags.append(pipeline_tags_value)
-                elif isinstance(pipeline_tags_value, list):
-                    metadata.tags.extend(pipeline_tags_value)
-
-            language_value = model_card.get("languages", model_card.get("language", None))
-            if language_value is not None:
-
-                if metadata.languages is None:
-                    metadata.languages = []
-
-                if isinstance(language_value, str):
-                    metadata.languages.append(language_value)
-                elif isinstance(language_value, list):
-                    metadata.languages.extend(language_value)
-
-            dataset_value = model_card.get("datasets", model_card.get("dataset", None))
-            if dataset_value is not None:
-
-                if metadata.datasets is None:
-                    metadata.datasets = []
-
-                if isinstance(dataset_value, str):
-                    metadata.datasets.append(dataset_value)
-                elif isinstance(dataset_value, list):
-                    metadata.datasets.extend(dataset_value)
+            use_array_model_card_metadata("datasets", "datasets")
+            use_array_model_card_metadata("datasets", "dataset")

        # Hugging Face Parameter Heuristics
        ####################################
@@ -25,14 +25,12 @@ def quant_shape_from_byte_shape(shape: Sequence[int], quant_type: GGMLQuantizati

 # same as ggml_compute_fp32_to_bf16 in ggml-impl.h
 def __compute_fp32_to_bf16(n: np.ndarray) -> np.ndarray:
-    n = n.astype(np.float32, copy=False).view(np.int32)
+    n = n.astype(np.float32, copy=False).view(np.uint32)
    # force nan to quiet
-    n = np.where((n & 0x7fffffff) > 0x7f800000, (n & 0xffff0000) | (64 << 16), n)
-    # flush subnormals to zero
-    n = np.where((n & 0x7f800000) == 0, n & 0x80000000, n)
+    n = np.where((n & 0x7fffffff) > 0x7f800000, (n & np.uint32(0xffff0000)) | np.uint32(64 << 16), n)
    # round to nearest even
-    n = (n + (0x7fff + ((n >> 16) & 1))) >> 16
-    return n.astype(np.int16)
+    n = (np.uint64(n) + (0x7fff + ((n >> 16) & 1))) >> 16
+    return n.astype(np.uint16)


 # This is faster than np.vectorize and np.apply_along_axis because it works on more than one row at a time
@@ -49,10 +47,10 @@ def __apply_over_grouped_rows(func: Callable[[np.ndarray], np.ndarray], arr: np.


 def __quantize_bf16_array(n: np.ndarray) -> np.ndarray:
-    return __apply_over_grouped_rows(__compute_fp32_to_bf16, arr=n, otype=np.int16, oshape=n.shape)
+    return __apply_over_grouped_rows(__compute_fp32_to_bf16, arr=n, otype=np.uint16, oshape=n.shape)


-__quantize_bf16_lazy = LazyNumpyTensor._wrap_fn(__quantize_bf16_array, meta_noop=np.int16)
+__quantize_bf16_lazy = LazyNumpyTensor._wrap_fn(__quantize_bf16_array, meta_noop=np.uint16)


 def quantize_bf16(n: np.ndarray):
@@ -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(
@@ -64,6 +64,7 @@ while read c; do
        src/ggml*.cu \
        src/ggml-cuda/* \
        src/ggml-sycl/* \
+        src/vulkan-shaders/* \
        include/ggml*.h \
        tests/test-opt.cpp \
        tests/test-grad0.cpp \
@@ -1 +1 @@
-31d544f87835a55602883fe09156bb85a4c163d8
+18703ad600cc68dbdb04d57434c876989a841d12
@@ -1444,7 +1444,8 @@ llama_token_attr llama_token_get_attr_impl(const struct llama_vocab & vocab, lla
 bool llama_token_is_eog_impl(const struct llama_vocab & vocab, llama_token token) {
    return token != -1 && (
        token == llama_token_eos_impl(vocab) ||
-        token == llama_token_eot_impl(vocab)
+        token == llama_token_eot_impl(vocab) ||
+        token == llama_token_eom_impl(vocab)
    );
 }

@@ -1500,6 +1501,10 @@ llama_token llama_token_eot_impl(const struct llama_vocab & vocab) {
    return vocab.special_eot_id;
 }

+llama_token llama_token_eom_impl(const struct llama_vocab & vocab) {
+    return vocab.special_eom_id;
+}
+
 int32_t llama_tokenize_impl(
    const struct llama_vocab & vocab,
                  const char * text,
@@ -45,6 +45,7 @@ struct llama_vocab {
    id special_suffix_id = -1;
    id special_middle_id = -1;
    id special_eot_id    = -1; // TODO: move above after "eos_id", and here add "file separator" token
+    id special_eom_id    = -1;

    // tokenizer flags
    bool tokenizer_add_space_prefix = false;
@@ -101,6 +102,7 @@ llama_token llama_token_prefix_impl(const struct llama_vocab & vocab);
 llama_token llama_token_middle_impl(const struct llama_vocab & vocab);
 llama_token llama_token_suffix_impl(const struct llama_vocab & vocab);
 llama_token llama_token_eot_impl   (const struct llama_vocab & vocab);
+llama_token llama_token_eom_impl   (const struct llama_vocab & vocab);

 int32_t llama_tokenize_impl(
        const struct llama_vocab & vocab,
@@ -804,8 +804,7 @@ struct test_cpy : public test_case {

    test_cpy(ggml_type type_src = GGML_TYPE_F32, ggml_type type_dst = GGML_TYPE_F32,
            std::array<int64_t, 4> ne = {10, 10, 10, 1},
-            std::array<int64_t, 4> permute = {0, 0, 0, 0},
-            bool _dst_use_permute = false)
+            std::array<int64_t, 4> permute = {0, 0, 0, 0})
        : type_src(type_src), type_dst(type_dst), ne(ne), permute(permute),
          _src_use_permute(permute[0] + permute[1] + permute[2] + permute[3] > 0) {}

@@ -2140,6 +2139,9 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op

    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32));
    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16));
+    // test cases for 1D im2col
+    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false));
+    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false));

    test_cases.emplace_back(new test_conv_transpose_1d());
    test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 3, 0, 1));
@@ -2269,7 +2271,10 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op

    for (ggml_type type_a : other_types) {
        for (ggml_type type_b : {GGML_TYPE_F32}) {
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, { 1,  1}, {1, 1}));
+            if (ggml_blck_size(type_a) != 256) {
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, ggml_blck_size(type_a), {1,  1}, {1, 1}));
+            }
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {1,  1}, {1, 1}));
        }
    }
Author	SHA1	Message	Date
Neo Zhang	16dab13bde	correct cmd name	2024-08-06 00:15:33 +08:00
wangshuai09	bc0f887e15	cann: fix buffer_num and runtime speed slowly error (#8865 )	2024-08-05 21:10:37 +08:00
Eric Curtin	b42978e7e4	readme : add ramalama to the availables UI (#8811 ) ramalama is a repo agnostic boring CLI tool that supports pulling from ollama, huggingface and oci registries. Signed-off-by: Eric Curtin <ecurtin@redhat.com>	2024-08-05 15:45:01 +03:00
Justine Tunney	b9dfc25ca3	ggml : fix overflows in elu function (#8866 ) It's helpful to use expm1f(x), because expf(x)-1 will result in overflow for 25% of single-precision floating point numbers.	2024-08-05 15:43:40 +03:00
Brian	1ef14b3007	py: Add more authorship metadata from model card (#8810 ) * py: add more authorship metadata from model card * fixup! py: add more authorship metadata from model card	2024-08-05 21:15:28 +10:00
fairydreaming	d3f0c7166a	Stop the generation when <\|eom_id\|> token is encountered - needed for Llama 3.1 tool call support (#8858 ) * gguf-py, llama : add constants and methods related to Llama-3.1 <\|eom_id\|> token * llama : find Llama-3.1 <\|eom_id\|> token id during vocab loading * llama-vocab : add Llama-3.1 <\|eom_id\|> token to the set of tokens stopping the generation --------- Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>	2024-08-05 09:38:01 +02:00
stduhpf	e31a4f6797	cmake: fix paths for vulkan shaders compilation on Windows (#8573 ) * Vulkan-shaders: attempt fix compilation on windows * fix miss-matched parenthesis	2024-08-05 08:18:27 +02:00
BarfingLemurs	400ae6f65f	readme : update model list (#8851 )	2024-08-05 08:54:10 +03:00
Georgi Gerganov	f1ea5146d7	llama : better replace_all (#8852 )	2024-08-05 08:53:39 +03:00
0cc4m	064cdc265f	vulkan : fix Qantized Mat-Vec Mul on AMD GPUs for ncols < 64 (#8855 ) * Fix Vulkan mul mat vec invalid results when ncols < warp size * Only run backend ops mul mat vec block size test if block size not already covered	2024-08-05 08:52:55 +03:00
Georgi Gerganov	5587e57a76	sync : ggml ggml-ci	2024-08-05 08:50:57 +03:00
0cc4m	a3738b2fa7	vulkan : implement Stable Diffusion operators (ggml/904) * Fix Vulkan repeat op * Implement Vulkan concat op * Delete old Vulkan shader generator * Implement Vulkan im2col op * Implement Vulkan unary gelu_quick op * Implement Vulkan group_norm op * Implement Vulkan timestep_embedding op * Implement Vulkan upscale op * Fix Vulkan vk_context tensor extra index issue * Fix Vulkan matmul shader parameter bug * Properly fix Vulkan matmul shader parameter bug * Add Vulkan ADD f16 + f32 -> f16 operator support * Implement Vulkan tanh op * Fix Vulkan group count too large Validation error on non-Nvidia GPUs * Throw error when too much memory is requested * Fix another Vulkan group count too large Validation error on non-Nvidia GPUs * Fix matmul MMQ condition * Implement Vulkan pad op * Fix Vulkan crash when tensor is used multiple times in a compute graph * Add Vulkan CONCAT f16 + f16 -> f16 op * Add Vulkan LEAKY_RELU op	2024-08-05 08:50:57 +03:00
Daniel Bevenius	655858ace0	ggml : move c parameter comment to ggml_rope_ext (ggml/901) This commit moves the comment for the c parameter from ggml_rope to ggml_rope_ext. The comment is currently incorrect as ggml_rope does not have a c parameter (freq_factors tensor). Signed-off-by: Daniel Bevenius <daniel.bevenius@gmail.com>	2024-08-05 08:50:57 +03:00
wangshuai09	c02b0a8a4d	cann: support q4_0 model (#8822 )	2024-08-05 12:22:30 +08:00
Brandon Squizzato	0d6fb52be0	Install curl in runtime layer (#8693 )	2024-08-04 20:17:16 +02:00
ardfork	978ba3d83d	Server: Don't ignore llama.cpp params (#8754 ) * Don't ignore llama.cpp params * Add fallback for max_tokens	2024-08-04 20:16:23 +02:00
Brian Cunnie	ecf6b7f23e	batched-bench : handle empty `-npl` (#8839 ) * [example] batched-bench "segmentation fault" When `llama-batched-bench` is invoked _without_ setting `-npl`, "number of parallel prompts", it segfaults. The segfault is caused by invoking `max_element()` on a zero-length vector, `n_pl` This commit addresses that by first checking to see if the number of parallel prompts is zero, and if so sets the maximum sequence size to 1; otherwise, sets it to the original, the result of `max_element()`. Fixes, when running `lldb build/bin/llama-batched-bench -- -m models/Meta-Llama-3-8B.gguf` ``` * thread #1, queue = 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=1, address=0x0) frame #0: 0x000000010000366c llama-batched-bench`main(argc=3, argv=0x000000016fdff268) at batched-bench.cpp:72:28 69 llama_context_params ctx_params = llama_context_params_from_gpt_params(params); 70 71 // ensure enough sequences are available -> 72 ctx_params.n_seq_max = std::max_element(n_pl.begin(), n_pl.end()); ``` Update examples/batched-bench/batched-bench.cpp Co-authored-by: compilade <git@compilade.net> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: compilade <git@compilade.net>	2024-08-04 13:55:03 +03:00
Daniel Bevenius	01aae2b497	baby-llama : remove duplicate vector include	2024-08-04 13:24:59 +03:00
Georgi Gerganov	4b77ea95f5	flake.lock: Update (#8847 )	2024-08-03 19:53:20 -07:00
jdomke	76614f352e	ggml : reading the runtime sve config of the cpu (#8709 ) * ggml : reading the runtime sve config of the cpu * change to one time init to prevent performance drop * prefix variable to avoid possible conflicts * revert xxhash fix and add brackets --------- Co-authored-by: domke <673751-domke@users.noreply.gitlab.com>	2024-08-03 18:34:41 +02:00
Sigbjørn Skjæret	b72c20b85c	Fix conversion of unnormalized BF16->BF16 weights (#7843 ) * add truncate_bf16 * truncate intermediate fp32 if converting bf16 to bf16 * fix masking in __compute_fp32_to_bf16 * np.int16 no longer used * missing cast and additional numpy 2.x fix * ggml-impl : do not flush bf16 subnormals to zero * ggml : add reference fp32 to bf16 conversion The fast version is no longer equivalent for all platforms because of the handling of subnormal values. * gguf-py : remove flush to zero for bf16 subnormals * gguf-py : remove float32 truncation to bf16 Rounding achieves the same thing in the cases where this was used. * missed prototype update in merge * merge cleanup --------- Co-authored-by: Francis Couture-Harpin <git@compilade.net>	2024-08-02 15:11:39 -04:00
Mengqing Cao	e09a800f9a	cann: Fix ggml_cann_im2col for 1D im2col (#8819 ) * fix ggml_cann_im2col for 1D im2col * fix build warning	2024-08-02 16:50:53 +08:00
Ouadie EL FAROUKI	0fbbd88458	[SYCL] Fixing wrong VDR iq4nl value (#8812 )	2024-08-02 08:55:17 +08:00
matteo	afbb4c1322	ggml-cuda: Adding support for unified memory (#8035 ) * Adding support for unified memory * adding again the documentation about unified memory * refactoring: Moved the unified memory code in the correct location. * Fixed compilation error when using hipblas * cleaning up the documentation * Updating the documentation Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * adding one more case where the PR should not be enabled --------- Co-authored-by: matteo serva <matteo.serva@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>	2024-08-01 23:28:28 +02:00
Alex O'Connell	b7a08fd5e0	Build: Only include execinfo.h on linux systems that support it (#8783 ) * Only enable backtrace on GLIBC linux systems * fix missing file from copy * use glibc macro instead of defining a custom one	2024-08-01 18:53:46 +02:00
slaren	7a11eb3a26	cuda : fix dmmv cols requirement to 2GGML_CUDA_DMMV_X (#8800 ) cuda : fix dmmv cols requirement to 2GGML_CUDA_DMMV_X update asserts * only use dmmv for supported types * add test	2024-08-01 15:26:22 +02:00
wangshuai09	c8a0090922	cann: support q8_0 for Ascend backend (#8805 )	2024-08-01 10:39:05 +08:00
Igor Okulist	afbbcf3c04	server : update llama-server embedding flag documentation (#8779 ) Fixes #8763	2024-07-31 19:59:09 -04:00
Clint Herron	ed9d2854c9	Build: Fix potential race condition (#8781 ) * Fix potential race condition as pointed out by @fairydreaming in #8776 * Reference the .o rather than rebuilding every time. * Adding in CXXFLAGS and LDFLAGS * Removing unnecessary linker flags.	2024-07-31 15:51:06 -04:00
pculliton	398ede5efe	Adding Gemma 2 2B configs (#8784 ) * Adding Gemma 2 2B configs Updates to Q scaling and Gemma 2 model sizes to match v2 2B model. * Update src/llama.cpp Co-authored-by: slaren <slarengh@gmail.com> --------- Co-authored-by: slaren <slarengh@gmail.com>	2024-07-31 17:12:10 +02:00
Borislav Stanimirov	44d28ddd5c	cmake : fix use of external ggml (#8787 )	2024-07-31 15:40:08 +02:00
Someone	268c566006	nix: cuda: rely on propagatedBuildInputs (#8772 ) Listing individual outputs no longer necessary to reduce the runtime closure size after https://github.com/NixOS/nixpkgs/pull/323056.	2024-07-30 13:35:30 -07:00
Brian	7e72aa74fd	py: add_array() will not add to kv store if value is an empty array (#8774 ) * gguf_writer.py: add_array() should not add to kv store if empty * Apply suggestions from code review I was wondering if there was a specific reason for `if val` but good to hear we can safely use `len(val == 0` Co-authored-by: compilade <git@compilade.net> --------- Co-authored-by: compilade <git@compilade.net>	2024-07-31 00:57:03 +10:00
l3utterfly	7c27a19b2e	added android implementation of ggml_print_backtrace_symbols (#8751 ) * added android implementation of ggml_print_backtrace_symbols * Update ggml/src/ggml.c Co-authored-by: slaren <slarengh@gmail.com> * Update ggml/src/ggml.c Co-authored-by: slaren <slarengh@gmail.com> * Update ggml/src/ggml.c Co-authored-by: slaren <slarengh@gmail.com> * Update ggml/src/ggml.c Co-authored-by: slaren <slarengh@gmail.com> * Update ggml/src/ggml.c Co-authored-by: slaren <slarengh@gmail.com> --------- Co-authored-by: slaren <slarengh@gmail.com>	2024-07-30 16:40:18 +02:00
Georgi Gerganov	140074bb86	flake.lock: Update (#8729 )	2024-07-30 05:58:57 -07:00
wangshuai09	6e2b6000e5	cann: update cmake (#8765 )	2024-07-30 12:37:35 +02:00
zhentaoyu	c887d8b017	[SYCL] Add `TIMESTEP_EMBEDDING` OP (#8707 ) Signed-off-by: zhentaoyu <zhentao.yu@intel.com>	2024-07-30 14:56:51 +08:00
CarterLi999	75af08c475	ggml: bugfix: fix the inactive elements is agnostic for risc-v vector (#8748 ) In these codes, we want to retain the value that they previously held when mask[i] is false. So we should use undisturbed. With the default agnostic policy of rvv intrinsic, these values can be held or be written with 1s. Co-authored-by: carter.li <carter.li@starfivetech.com>	2024-07-29 18:38:34 +02:00
R0CKSTAR	439b3fc75a	cuda : organize vendor-specific headers into vendors directory (#8746 ) Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>	2024-07-29 14:56:12 +02:00
Meng, Hengyu	0832de7236	[SYCL] add conv support (#8688 )	2024-07-29 10:50:27 +08:00
Johannes Gäßler	6eeaeba126	cmake: use 1 more thread for non-ggml in CI (#8740 )	2024-07-28 22:32:44 +02:00
Austin	4730faca61	chore : Fix vulkan related compiler warnings, add help text, improve CLI options (#8477 ) * chore: Fix compiler warnings, add help text, improve CLI options * Add prototypes for function definitions * Invert logic of --no-clean option to be more intuitive * Provide a new help prompt with clear instructions * chore : Add ignore rule for vulkan shader generator Signed-off-by: teleprint-me <77757836+teleprint-me@users.noreply.github.com> * Update ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp Co-authored-by: 0cc4m <picard12@live.de> * chore : Remove void and apply C++ style empty parameters * chore : Remove void and apply C++ style empty parameters --------- Signed-off-by: teleprint-me <77757836+teleprint-me@users.noreply.github.com> Co-authored-by: 0cc4m <picard12@live.de>	2024-07-28 09:52:42 +02:00
compilade	4c676c85e5	llama : refactor session file management (#8699 ) * llama : refactor session file management * llama : saving and restoring state checks for overflow The size of the buffers should now be given to the functions working with them, otherwise a truncated file could cause out of bound reads. * llama : stream from session file instead of copying into a big buffer Loading session files should no longer cause a memory usage spike. * llama : llama_state_get_size returns the actual size instead of max This is a breaking change, but makes that function much easier to keep up to date, and it also makes it reflect the behavior of llama_state_seq_get_size. * llama : share code between whole and seq_id-specific state saving Both session file types now use a more similar format. * llama : no longer store all hparams in session files Instead, the model arch name is stored. The layer count and the embedding dimensions of the KV cache are still verified when loading. Storing all the hparams is not necessary. * llama : fix uint64_t format type * llama : various integer type cast and format string fixes Some platforms use "%lu" and others "%llu" for uint64_t. Not sure how to handle that, so casting to size_t when displaying errors. * llama : remove _context suffix for llama_data_context * llama : fix session file loading llama_state_get_size cannot be used to get the max size anymore. * llama : more graceful error handling of invalid session files * llama : remove LLAMA_MAX_RNG_STATE It's no longer necessary to limit the size of the RNG state, because the max size of session files is not estimated anymore. * llama : cast seq_id in comparison with unsigned n_seq_max	2024-07-28 00:42:05 -04:00
R0CKSTAR	e54c35e4fb	feat: Support Moore Threads GPU (#8383 ) * Update doc for MUSA Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com> * Add GGML_MUSA in Makefile Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com> * Add GGML_MUSA in CMake Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com> * CUDA => MUSA Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com> * MUSA adds support for __vsubss4 Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com> * Fix CI build failure Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com> --------- Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>	2024-07-28 01:41:25 +02:00