ggml : add SSE 4.2 and x64 base variant for CPUs without AVX (#12871)

* ggml : add SSE 4.2 variant for CPUs without AVX

* ggml : add x64 base ABI variant

.github/workflows/build.yml

@@ -601,8 +601,9 @@ jobs:
             -DGGML_SYCL_F16=ON
           cmake --build build --config Release -j $(nproc)
 
-  build-linux-cross:
-    uses: ./.github/workflows/build-linux-cross.yml
+# Disabled for now due to sporadic issue syncing.
+#  build-linux-cross:
+#    uses: ./.github/workflows/build-linux-cross.yml
 
   macOS-latest-cmake-ios:
     runs-on: macos-latest
@@ -1766,16 +1767,17 @@ jobs:
     if: ${{ github.event_name != 'pull_request' || contains(github.event.pull_request.labels.*.name, 'Ascend NPU') }}
     defaults:
       run:
-       shell: bash -el {0}
-    runs-on: ubuntu-24.04-arm
+        shell: bash -el {0}
     strategy:
       matrix:
+        arch: [x86, aarch64]
         cann:
           - '8.1.RC1.alpha001-910b-openeuler22.03-py3.10'
         device:
           - 'ascend910b3'
         build:
           - 'Release'
+    runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
     container: ascendai/cann:${{ matrix.cann }}
     steps:
       - name: Checkout

Makefile

@@ -780,10 +780,6 @@ ifdef GGML_HIP
 
 	MK_CPPFLAGS += -DGGML_USE_HIP -DGGML_USE_CUDA
 
-ifdef GGML_HIP_UMA
-	MK_CPPFLAGS += -DGGML_HIP_UMA
-endif # GGML_HIP_UMA
-
 	MK_LDFLAGS += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
 	MK_LDFLAGS += -L$(ROCM_PATH)/lib64 -Wl,-rpath=$(ROCM_PATH)/lib64
 	MK_LDFLAGS += -lhipblas -lamdhip64 -lrocblas

README.md

@@ -16,6 +16,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ## Hot topics
 
+- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli` and `gemma3-cli` https://github.com/ggml-org/llama.cpp/pull/13012, `libllava` will be deprecated
 - **How to use [MTLResidencySet](https://developer.apple.com/documentation/metal/mtlresidencyset?language=objc) to keep the GPU memory active?** https://github.com/ggml-org/llama.cpp/pull/11427
 - **VS Code extension for FIM completions:** https://github.com/ggml-org/llama.vscode
 - Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639

common/arg.cpp

@@ -2726,7 +2726,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.chat_template = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CHAT_TEMPLATE"));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_LLAVA}).set_env("LLAMA_ARG_CHAT_TEMPLATE"));
     add_opt(common_arg(
         {"--chat-template-file"}, "JINJA_TEMPLATE_FILE",
         string_format(

convert_lora_to_gguf.py

@@ -24,7 +24,7 @@ if 'NO_LOCAL_GGUF' not in os.environ:
 import gguf
 
 # reuse model definitions from convert_hf_to_gguf.py
-from convert_hf_to_gguf import LazyTorchTensor, Model
+from convert_hf_to_gguf import LazyTorchTensor, ModelBase
 
 logger = logging.getLogger("lora-to-gguf")
 
@@ -340,11 +340,11 @@ if __name__ == '__main__':
             sys.exit(1)
     else:
         logger.info(f"Loading base model: {dir_base_model.name}")
-        hparams = Model.load_hparams(dir_base_model)
+        hparams = ModelBase.load_hparams(dir_base_model)
 
     with torch.inference_mode():
         try:
-            model_class = Model.from_model_architecture(hparams["architectures"][0])
+            model_class = ModelBase.from_model_architecture(hparams["architectures"][0])
         except NotImplementedError:
             logger.error(f"Model {hparams['architectures'][0]} is not supported")
             sys.exit(1)

docs/build.md

@@ -259,8 +259,6 @@ You can download it from your Linux distro's package manager or from here: [ROCm
       cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
       && cmake --build build --config Release -- -j 16
   ```
-  On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DGGML_HIP_UMA=ON`.
-  However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
 
   To enhance flash attention performance on RDNA3+ or CDNA architectures, you can utilize the rocWMMA library by enabling the `-DGGML_HIP_ROCWMMA_FATTN=ON` option. This requires rocWMMA headers to be installed on the build system.
 
@@ -296,6 +294,10 @@ You can download it from your Linux distro's package manager or from here: [ROCm
 The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
 If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.
 
+### Unified Memory
+
+On Linux it is possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1`. However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
+
 ## Vulkan
 
 **Windows**

examples/llava/CMakeLists.txt

@@ -61,19 +61,9 @@ if(TARGET BUILD_INFO)
     add_dependencies(mtmd BUILD_INFO)
 endif()
 
-set(TARGET llama-llava-cli)
-add_executable(${TARGET} llava-cli.cpp)
-set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-llava-cli)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_17)
-
-set(TARGET llama-minicpmv-cli)
-add_executable(${TARGET} minicpmv-cli.cpp)
-set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-minicpmv-cli)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_17)
+add_executable(llama-llava-cli    deprecation-warning.cpp)
+add_executable(llama-gemma3-cli   deprecation-warning.cpp)
+add_executable(llama-minicpmv-cli deprecation-warning.cpp)
 
 set(TARGET llama-qwen2vl-cli)
 add_executable(${TARGET} qwen2vl-cli.cpp)
@@ -82,9 +72,9 @@ install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
-set(TARGET llama-gemma3-cli)
-add_executable(${TARGET} gemma3-cli.cpp)
-set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-gemma3-cli)
+set(TARGET llama-mtmd-cli)
+add_executable(${TARGET} mtmd-cli.cpp)
+set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-mtmd-cli)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common mtmd ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/llava/clip-impl.h

@@ -50,7 +50,6 @@
 // tensor name constants
 //
 
-#define TN_TOKEN_EMBD      "%s.token_embd.weight"
 #define TN_POS_EMBD        "%s.position_embd.weight"
 #define TN_CLASS_EMBD      "v.class_embd"
 #define TN_PATCH_EMBD      "v.patch_embd.weight"  // not rename tensor with ".0" postfix for backwrad compat
@@ -66,8 +65,6 @@
 #define TN_LN_2            "%s.blk.%d.ln2.%s"
 #define TN_LN_PRE          "%s.pre_ln.%s"
 #define TN_LN_POST         "%s.post_ln.%s"
-#define TN_TEXT_PROJ       "text_projection.weight"
-#define TN_VIS_PROJ        "visual_projection.weight"
 #define TN_LLAVA_PROJ      "mm.%d.%s"
 #define TN_MVLM_PROJ_MLP   "mm.model.mlp.%d.%s"
 #define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"

examples/llava/clip.h

@@ -30,12 +30,13 @@ struct clip_image_size {
     int height;
 };
 
+struct clip_image_f32;
 struct clip_image_u8_batch;
 struct clip_image_f32_batch;
 
 struct clip_context_params {
     bool use_gpu;
-    ggml_log_level verbosity;
+    enum ggml_log_level verbosity;
 };
 
 // deprecated, use clip_init
@@ -84,7 +85,7 @@ CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch * batch);
 CLIP_API size_t clip_image_f32_batch_n_images(const struct clip_image_f32_batch * batch); // equivalent to batch->size()
 CLIP_API size_t clip_image_f32_batch_nx(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->nx
 CLIP_API size_t clip_image_f32_batch_ny(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->ny
-CLIP_API clip_image_f32 * clip_image_f32_get_img(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->data
+CLIP_API struct clip_image_f32 * clip_image_f32_get_img(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->data
 
 /**
  * Build image from pixels decoded by other libraries instead of stb_image.h for better performance.

examples/llava/deprecation-warning.cpp

@@ -0,0 +1,22 @@
+#include <cstdio>
+#include <string>
+
+int main(int argc, char** argv) {
+    std::string filename = "main";
+    if (argc >= 1) {
+        filename = argv[0];
+    }
+
+    // Get only the program name from the full path
+    size_t pos = filename.find_last_of("/\\");
+    if (pos != std::string::npos) {
+        filename = filename.substr(pos+1);
+    }
+
+    fprintf(stdout, "\n");
+    fprintf(stdout, "WARNING: The binary '%s' is deprecated.\n", filename.c_str());
+    fprintf(stdout, "Please use 'llama-mtmd-cli' instead.\n");
+    fprintf(stdout, "\n");
+
+    return EXIT_FAILURE;
+}

examples/llava/llava-cli.cpp

@@ -1,332 +0,0 @@
-#include "arg.h"
-#include "base64.hpp"
-#include "log.h"
-#include "common.h"
-#include "sampling.h"
-#include "clip.h"
-#include "llava.h"
-#include "llama.h"
-#include "ggml.h"
-
-#include <cstdio>
-#include <cstdlib>
-#include <cstring>
-#include <vector>
-
-static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past) {
-    int N = (int) tokens.size();
-    for (int i = 0; i < N; i += n_batch) {
-        int n_eval = (int) tokens.size() - i;
-        if (n_eval > n_batch) {
-            n_eval = n_batch;
-        }
-        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval))) {
-            LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
-            return false;
-        }
-        *n_past += n_eval;
-    }
-    return true;
-}
-
-static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
-    std::vector<llama_token> tokens;
-    tokens.push_back(id);
-    return eval_tokens(ctx_llama, tokens, 1, n_past);
-}
-
-static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
-    std::string              str2     = str;
-    std::vector<llama_token> embd_inp = common_tokenize(ctx_llama, str2, add_bos, true);
-    eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
-    return true;
-}
-
-static const char * sample(struct common_sampler * smpl,
-                           struct llama_context * ctx_llama,
-                           int * n_past) {
-    const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
-    common_sampler_accept(smpl, id, true);
-
-    const llama_model * model = llama_get_model(ctx_llama);
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
-    static std::string ret;
-    if (llama_vocab_is_eog(vocab, id)) {
-        ret = "</s>";
-    } else {
-        ret = common_token_to_piece(ctx_llama, id);
-    }
-    eval_id(ctx_llama, id, n_past);
-    return ret.c_str();
-}
-
-static const char* IMG_BASE64_TAG_BEGIN = "<img src=\"data:image/jpeg;base64,";
-static const char* IMG_BASE64_TAG_END = "\">";
-
-static void find_image_tag_in_prompt(const std::string& prompt, size_t& begin_out, size_t& end_out) {
-    begin_out = prompt.find(IMG_BASE64_TAG_BEGIN);
-    end_out = prompt.find(IMG_BASE64_TAG_END, (begin_out == std::string::npos) ? 0UL : begin_out);
-}
-
-static bool prompt_contains_image(const std::string& prompt) {
-    size_t begin, end;
-    find_image_tag_in_prompt(prompt, begin, end);
-    return (begin != std::string::npos);
-}
-
-// replaces the base64 image tag in the prompt with `replacement`
-static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip_ctx * ctx_clip, int n_threads, const std::string& prompt) {
-    size_t img_base64_str_start, img_base64_str_end;
-    find_image_tag_in_prompt(prompt, img_base64_str_start, img_base64_str_end);
-    if (img_base64_str_start == std::string::npos || img_base64_str_end == std::string::npos) {
-        LOG_ERR("%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
-        return NULL;
-    }
-
-    auto base64_bytes_start = img_base64_str_start + strlen(IMG_BASE64_TAG_BEGIN);
-    auto base64_bytes_count = img_base64_str_end - base64_bytes_start;
-    auto base64_str = prompt.substr(base64_bytes_start, base64_bytes_count );
-
-    auto required_bytes = base64::required_encode_size(base64_str.size());
-    auto img_bytes = std::vector<unsigned char>(required_bytes);
-    base64::decode(base64_str.begin(), base64_str.end(), img_bytes.begin());
-
-    auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, img_bytes.data(), img_bytes.size());
-    if (!embed) {
-        LOG_ERR("%s: could not load image from base64 string.\n", __func__);
-        return NULL;
-    }
-
-    return embed;
-}
-
-static std::string remove_image_from_prompt(const std::string& prompt, const char * replacement = "") {
-    size_t begin, end;
-    find_image_tag_in_prompt(prompt, begin, end);
-    if (begin == std::string::npos || end == std::string::npos) {
-        return prompt;
-    }
-    auto pre = prompt.substr(0, begin);
-    auto post = prompt.substr(end + strlen(IMG_BASE64_TAG_END));
-    return pre + replacement + post;
-}
-
-struct llava_context {
-    struct clip_ctx * ctx_clip = NULL;
-    struct llama_context * ctx_llama = NULL;
-    struct llama_model * model = NULL;
-};
-
-static void print_usage(int, char ** argv) {
-    LOG("\n example usage:\n");
-    LOG("\n     %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
-    LOG("\n note: a lower temperature value like 0.1 is recommended for better quality.\n");
-}
-
-static struct llava_image_embed * load_image(llava_context * ctx_llava, common_params * params, const std::string & fname) {
-
-    // load and preprocess the image
-    llava_image_embed * embed = NULL;
-    auto prompt = params->prompt;
-    if (prompt_contains_image(prompt)) {
-        if (!params->image.empty()) {
-            LOG_INF("using base64 encoded image instead of command line image path\n");
-        }
-        embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->cpuparams.n_threads, prompt);
-        if (!embed) {
-            LOG_ERR("%s: can't load image from prompt\n", __func__);
-            return NULL;
-        }
-        params->prompt = remove_image_from_prompt(prompt);
-    } else {
-        embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->cpuparams.n_threads, fname.c_str());
-        if (!embed) {
-            fprintf(stderr, "%s: is %s really an image file?\n", __func__, fname.c_str());
-            return NULL;
-        }
-    }
-
-    return embed;
-}
-
-static void process_prompt(struct llava_context * ctx_llava, struct llava_image_embed * image_embed, common_params * params, const std::string & prompt) {
-    int n_past = 0;
-
-    const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict;
-
-    std::string system_prompt, user_prompt;
-    size_t image_pos = prompt.find("<image>");
-    if (image_pos != std::string::npos) {
-        // new templating mode: Provide the full prompt including system message and use <image> as a placeholder for the image
-        system_prompt = prompt.substr(0, image_pos);
-        user_prompt = prompt.substr(image_pos + std::string("<image>").length());
-        LOG_INF("system_prompt: %s\n", system_prompt.c_str());
-        if (params->verbose_prompt) {
-            auto tmp = common_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
-            for (int i = 0; i < (int) tmp.size(); i++) {
-                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
-            }
-        }
-        LOG_INF("user_prompt: %s\n", user_prompt.c_str());
-        if (params->verbose_prompt) {
-            auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
-            for (int i = 0; i < (int) tmp.size(); i++) {
-                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
-            }
-        }
-    } else {
-        // llava-1.5 native mode
-        system_prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:";
-        user_prompt = prompt + "\nASSISTANT:";
-        if (params->verbose_prompt) {
-            auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
-            for (int i = 0; i < (int) tmp.size(); i++) {
-                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
-            }
-        }
-    }
-
-    eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, true);
-    llava_eval_image_embed(ctx_llava->ctx_llama, image_embed, params->n_batch, &n_past);
-    eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false);
-
-    // generate the response
-
-    LOG("\n");
-
-    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sampling);
-    if (!smpl) {
-        LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__);
-        exit(1);
-    }
-
-    std::string response = "";
-    for (int i = 0; i < max_tgt_len; i++) {
-        const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past);
-        response += tmp;
-        if (strcmp(tmp, "</s>") == 0) break;
-        if (strstr(tmp, "###")) break; // Yi-VL behavior
-        LOG("%s", tmp);
-        if (strstr(response.c_str(), "<|im_end|>")) break; // Yi-34B llava-1.6 - for some reason those decode not as the correct token (tokenizer works)
-        if (strstr(response.c_str(), "<|im_start|>")) break; // Yi-34B llava-1.6
-        if (strstr(response.c_str(), "USER:")) break; // mistral llava-1.6
-
-        fflush(stdout);
-    }
-
-    common_sampler_free(smpl);
-    LOG("\n");
-}
-
-static struct llama_model * llava_init(common_params * params) {
-    llama_backend_init();
-    llama_numa_init(params->numa);
-
-    llama_model_params model_params = common_model_params_to_llama(*params);
-
-    llama_model * model = llama_model_load_from_file(params->model.path.c_str(), model_params);
-    if (model == NULL) {
-        LOG_ERR("%s: unable to load model\n" , __func__);
-        return NULL;
-    }
-    return model;
-}
-
-static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
-    const char * clip_path = params->mmproj.path.c_str();
-
-    auto prompt = params->prompt;
-    if (prompt.empty()) {
-        prompt = "describe the image in detail.";
-    }
-
-    auto ctx_clip = clip_model_load(clip_path, GGML_LOG_LEVEL_INFO);
-
-    llama_context_params ctx_params = common_context_params_to_llama(*params);
-    ctx_params.n_ctx           = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings
-
-    llama_context * ctx_llama = llama_init_from_model(model, ctx_params);
-
-    if (ctx_llama == NULL) {
-        LOG_ERR("%s: failed to create the llama_context\n" , __func__);
-        return NULL;
-    }
-
-    auto * ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
-
-    ctx_llava->ctx_llama = ctx_llama;
-    ctx_llava->ctx_clip = ctx_clip;
-    ctx_llava->model = model;
-    return ctx_llava;
-}
-
-static void llava_free(struct llava_context * ctx_llava) {
-    if (ctx_llava->ctx_clip) {
-        clip_free(ctx_llava->ctx_clip);
-        ctx_llava->ctx_clip = NULL;
-    }
-
-    llama_free(ctx_llava->ctx_llama);
-    llama_model_free(ctx_llava->model);
-    llama_backend_free();
-}
-
-int main(int argc, char ** argv) {
-    ggml_time_init();
-
-    common_params params;
-
-    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, print_usage)) {
-        return 1;
-    }
-
-    common_init();
-
-    if (params.mmproj.path.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
-        print_usage(argc, argv);
-        return 1;
-    }
-
-    auto * model = llava_init(&params);
-    if (model == NULL) {
-        fprintf(stderr, "%s: error: failed to init llava model\n", __func__);
-        return 1;
-    }
-
-    if (prompt_contains_image(params.prompt)) {
-        auto * ctx_llava = llava_init_context(&params, model);
-
-        auto * image_embed = load_image(ctx_llava, &params, "");
-
-        // process the prompt
-        process_prompt(ctx_llava, image_embed, &params, params.prompt);
-
-        llama_perf_context_print(ctx_llava->ctx_llama);
-        llava_image_embed_free(image_embed);
-        ctx_llava->model = NULL;
-        llava_free(ctx_llava);
-    } else {
-        for (auto & image : params.image) {
-            auto * ctx_llava = llava_init_context(&params, model);
-
-            auto * image_embed = load_image(ctx_llava, &params, image);
-            if (!image_embed) {
-                LOG_ERR("%s: failed to load image %s. Terminating\n\n", __func__, image.c_str());
-                return 1;
-            }
-
-            // process the prompt
-            process_prompt(ctx_llava, image_embed, &params, params.prompt);
-
-            llama_perf_context_print(ctx_llava->ctx_llama);
-            llava_image_embed_free(image_embed);
-            ctx_llava->model = NULL;
-            llava_free(ctx_llava);
-        }
-    }
-
-    llama_model_free(model);
-
-    return 0;
-}

examples/llava/minicpmv-cli.cpp

@@ -1,354 +0,0 @@
-#include "arg.h"
-#include "log.h"
-#include "common.h"
-#include "sampling.h"
-#include "clip.h"
-#include "llava.h"
-#include "llama.h"
-#include "ggml.h"
-
-#include <algorithm>
-#include <cstdio>
-#include <cstdlib>
-#include <cstring>
-#include <vector>
-#include <iostream> // TODO: remove me
-
-struct llava_context {
-    struct clip_ctx * ctx_clip = NULL;
-    struct llama_context * ctx_llama = NULL;
-    struct llama_model * model = NULL;
-};
-
-static void show_additional_info(int /*argc*/, char ** argv) {
-    LOG("\nexample usage:\n\n%s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
-    LOG("\nnote: a lower temperature value like 0.1 is recommended for better quality.\n");
-}
-
-static struct llama_model * llava_init(common_params * params) {
-    llama_backend_init();
-    llama_numa_init(params->numa);
-
-    llama_model_params model_params = common_model_params_to_llama(*params);
-
-    llama_model * model = llama_model_load_from_file(params->model.path.c_str(), model_params);
-    if (model == NULL) {
-        LOG_ERR("%s: unable to load model\n" , __func__);
-        return NULL;
-    }
-    return model;
-}
-
-static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
-    auto prompt = params->prompt;
-    if (prompt.empty()) {
-        prompt = "describe the image in detail.";
-    }
-
-    llama_context_params ctx_params = common_context_params_to_llama(*params);
-    if (params->n_ctx < 2048) {
-        // warn user here, "Image processing requires at least 2048 context, setting context to 2048"
-        LOG_WRN("%s: Image processing requires at least 2048 context, setting context to 2048\n" , __func__);
-        ctx_params.n_ctx = 2048;
-    } else {
-        ctx_params.n_ctx = params->n_ctx;
-    }
-
-    llama_context * ctx_llama = llama_init_from_model(model, ctx_params);
-
-    if (ctx_llama == NULL) {
-        LOG_ERR("%s: failed to create the llama_context\n" , __func__);
-        return NULL;
-    }
-
-    auto * ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
-
-    ctx_llava->ctx_llama = ctx_llama;
-    ctx_llava->model = model;
-    return ctx_llava;
-}
-
-static void llava_free(struct llava_context * ctx_llava) {
-    if (ctx_llava->ctx_clip) {
-        clip_free(ctx_llava->ctx_clip);
-        ctx_llava->ctx_clip = NULL;
-    }
-
-    llama_free(ctx_llava->ctx_llama);
-    llama_model_free(ctx_llava->model);
-    llama_backend_free();
-}
-
-static struct clip_ctx * clip_init_context(common_params * params) {
-    const char * clip_path = params->mmproj.path.c_str();
-
-    auto prompt = params->prompt;
-    if (prompt.empty()) {
-        prompt = "describe the image in detail.";
-    }
-    struct clip_context_params clip_params = {
-        /* use_gpu */   params->n_gpu_layers != 0,
-        /* verbosity */ GGML_LOG_LEVEL_INFO, // TODO: make this configurable
-    };
-    auto * ctx_clip = clip_init(clip_path, clip_params);
-    return ctx_clip;
-}
-
-static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past) {
-    int N = (int) tokens.size();
-    for (int i = 0; i < N; i += n_batch) {
-        int n_eval = (int) tokens.size() - i;
-        if (n_eval > n_batch) {
-            n_eval = n_batch;
-        }
-        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval))) {
-            LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
-            return false;
-        }
-        *n_past += n_eval;
-    }
-    return true;
-}
-
-static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
-    std::vector<llama_token> tokens;
-    tokens.push_back(id);
-    return eval_tokens(ctx_llama, tokens, 1, n_past);
-}
-
-static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
-    std::string              str2     = str;
-    std::vector<llama_token> embd_inp = common_tokenize(ctx_llama, str2, add_bos, true);
-    return eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
-}
-
-static void process_eval_image_embed(struct llava_context * ctx_llava, const struct llava_image_embed * embeds, int n_batch, int * n_past, int idx) {
-    float * image_embed = (float *)malloc(clip_embd_nbytes(ctx_llava->ctx_clip));
-    std::memcpy(image_embed, embeds->embed + idx * clip_n_patches(ctx_llava->ctx_clip) * clip_n_mmproj_embd(ctx_llava->ctx_clip), clip_embd_nbytes(ctx_llava->ctx_clip));
-
-    auto * slice_embed = (llava_image_embed*)malloc(sizeof(llava_image_embed));
-    slice_embed->embed = image_embed;
-    slice_embed->n_image_pos = clip_n_patches(ctx_llava->ctx_clip);
-    llava_eval_image_embed(ctx_llava->ctx_llama, slice_embed, n_batch, n_past);
-    llava_image_embed_free(slice_embed);
-}
-
-static void process_image(struct llava_context * ctx_llava, struct llava_image_embed * embeds, common_params * params, int &n_past) {
-    std::string system_prompt;
-    int idx = 0;
-    int num_image_embeds = embeds->n_image_pos / clip_n_patches(ctx_llava->ctx_clip);
-    int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip);
-    if (has_minicpmv_projector == 2) {
-        system_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n";
-    }
-    else if (has_minicpmv_projector == 3) {
-        system_prompt = "<|im_start|>user\n";
-    }
-    else if (has_minicpmv_projector == 4) {
-        system_prompt = "<|im_start|>user\n";
-    }
-    LOG_INF("%s: image token past: %d\n", __func__, n_past);
-    eval_string(ctx_llava->ctx_llama, (system_prompt+"<image>").c_str(), params->n_batch, &n_past, false);
-    process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
-    eval_string(ctx_llava->ctx_llama, std::string("</image>").c_str(), params->n_batch, &n_past, false);
-    if (num_image_embeds > 1) {
-        if (has_minicpmv_projector == 2) {
-            size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip);
-            eval_string(ctx_llava->ctx_llama, std::string("<slice>").c_str(), params->n_batch, &n_past, false);
-            for (size_t i = 0; i < (num_image_embeds-1)/num_image_embeds_col; ++i) {
-                for (size_t j = 0; j < num_image_embeds_col; ++j) {
-                    eval_string(ctx_llava->ctx_llama, std::string("<image>").c_str(), params->n_batch, &n_past, false);
-                    process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
-                    eval_string(ctx_llava->ctx_llama, std::string("</image>").c_str(), params->n_batch, &n_past, false);
-                    if (j == num_image_embeds_col - 1) {
-                        eval_string(ctx_llava->ctx_llama, std::string("\n").c_str(), params->n_batch, &n_past, false);
-                    }
-                }
-            }
-            eval_string(ctx_llava->ctx_llama, std::string("</slice>").c_str(), params->n_batch, &n_past, false);
-        }
-        else if (has_minicpmv_projector == 3 || has_minicpmv_projector == 4) {
-            size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip);
-            for (size_t i = 0; i < (num_image_embeds-1)/num_image_embeds_col; ++i) {
-                for (size_t j = 0; j < num_image_embeds_col; ++j) {
-                    eval_string(ctx_llava->ctx_llama, std::string("<slice>").c_str(), params->n_batch, &n_past, false);
-                    process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
-                    eval_string(ctx_llava->ctx_llama, std::string("</slice>").c_str(), params->n_batch, &n_past, false);
-                    if (j == num_image_embeds_col - 1) {
-                        eval_string(ctx_llava->ctx_llama, std::string("\n").c_str(), params->n_batch, &n_past, false);
-                    }
-                }
-            }
-        }
-    }
-    LOG_INF("%s: image token past: %d\n", __func__, n_past);
-}
-
-static const char * sample(struct common_sampler * smpl,
-                           struct llama_context * ctx_llama,
-                           int * n_past) {
-    const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
-    common_sampler_accept(smpl, id, true);
-
-    const llama_model * model = llama_get_model(ctx_llama);
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
-    static std::string ret;
-    if (llama_vocab_is_eog(vocab, id)) {
-        ret = "</s>";
-    } else {
-        ret = common_token_to_piece(ctx_llama, id);
-    }
-    eval_id(ctx_llama, id, n_past);
-    return ret.c_str();
-}
-
-static struct llava_context * minicpmv_init(common_params * params, const std::string & fname, int &n_past){
-    auto * ctx_clip = clip_init_context(params);
-    auto * embeds = llava_image_embed_make_with_filename(ctx_clip, params->cpuparams.n_threads, fname.c_str());
-    if (!embeds) {
-        LOG_ERR("failed to load image %s. Terminating\n\n", fname.c_str());
-        return NULL;
-    }
-
-    // process the prompt
-    if (params->prompt.empty() && params->interactive == false) {
-        LOG_ERR("prompt should be given or interactive mode should be on");
-        return NULL;
-    }
-
-    auto * model = llava_init(params);
-    if (model == NULL) {
-        fprintf(stderr, "%s: error: failed to init minicpmv model\n", __func__);
-        return NULL;
-    }
-    const int64_t t_llava_init_start_us = ggml_time_us();
-    auto * ctx_llava = llava_init_context(params, model);
-    ctx_llava->ctx_clip = ctx_clip;
-    const int64_t t_llava_init_end_us = ggml_time_us();
-    float t_llava_init_ms = (t_llava_init_end_us - t_llava_init_start_us) / 1000.0;
-    LOG_INF("%s: llava init in %8.2f ms.\n", __func__, t_llava_init_ms);
-
-    const int64_t t_process_image_start_us = ggml_time_us();
-    process_image(ctx_llava, embeds, params, n_past);
-    const int64_t t_process_image_end_us = ggml_time_us();
-    float t_process_image_ms = (t_process_image_end_us - t_process_image_start_us) / 1000.0;
-    LOG_INF("%s: llama process image in %8.2f ms.\n", __func__, t_process_image_ms);
-
-    llava_image_embed_free(embeds);
-    return ctx_llava;
-}
-
-static struct common_sampler * llama_init(struct llava_context * ctx_llava, common_params * params, const std::string & prompt, int & n_past, bool is_first = false){
-    std::string user_prompt = prompt;
-    int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip);
-    if (!is_first) {
-        if (has_minicpmv_projector == 2) {
-            user_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n" + prompt;
-        }
-        else if (has_minicpmv_projector == 3) {
-            user_prompt = "<|im_start|>user\n" + prompt;
-        }
-        else if (has_minicpmv_projector == 4) {
-            user_prompt = "<|im_start|>user\n" + prompt;
-        }
-    }
-
-    eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false);
-    if (has_minicpmv_projector == 2) {
-        eval_string(ctx_llava->ctx_llama, "<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n", params->n_batch, &n_past, false);
-    }
-    else if (has_minicpmv_projector == 3) {
-        eval_string(ctx_llava->ctx_llama, "<|im_end|><|im_start|>assistant\n", params->n_batch, &n_past, false);
-    }
-    else if (has_minicpmv_projector == 4) {
-        eval_string(ctx_llava->ctx_llama, "<|im_end|><|im_start|>assistant\n", params->n_batch, &n_past, false);
-    }
-
-    // generate the response
-
-    LOG_INF("\n");
-
-    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sampling);
-    return smpl;
-}
-
-static const char * llama_loop(struct llava_context * ctx_llava,struct common_sampler * smpl, int &n_past){
-
-    const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past);
-    return tmp;
-}
-
-int main(int argc, char ** argv) {
-    ggml_time_init();
-
-    common_params params;
-
-    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) {
-        return 1;
-    }
-
-    common_init();
-
-    if (params.mmproj.path.empty() || (params.image.empty())) {
-        show_additional_info(argc, argv);
-        return 1;
-    }
-
-    for (auto & image : params.image) {
-        int n_past = 0;
-        auto * ctx_llava = minicpmv_init(&params, image, n_past);
-
-        if (!params.prompt.empty()) {
-            LOG("<user>%s\n", params.prompt.c_str());
-            LOG("<assistant>");
-            auto * smpl = llama_init(ctx_llava, &params, params.prompt, n_past, true);
-            const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
-            std::string response;
-            bool have_tmp = false;
-            for (int i = 0; i < max_tgt_len; i++) {
-                const auto * tmp = llama_loop(ctx_llava, smpl, n_past);
-                response += tmp;
-                if (strcmp(tmp, "</s>") == 0){
-                    if (!have_tmp) {
-                        continue;
-                    }
-                    break;
-                }
-                if (strstr(tmp, "###")) break; // Yi-VL behavior
-                have_tmp = true;
-                printf("%s", tmp);
-                if (strstr(response.c_str(), "<user>")) break; // minicpm-v
-
-                fflush(stdout);
-            }
-            common_sampler_free(smpl);
-        }else {
-            while (true) {
-                LOG("<user>");
-                std::string prompt;
-                std::getline(std::cin, prompt);
-                LOG("<assistant>");
-                auto * smpl = llama_init(ctx_llava, &params, prompt, n_past, true);
-                const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
-                std::string response;
-                for (int i = 0; i < max_tgt_len; i++) {
-                    const auto * tmp = llama_loop(ctx_llava, smpl, n_past);
-                    response += tmp;
-                    if (strcmp(tmp, "</s>") == 0) break;
-                    printf("%s", tmp);// mistral llava-1.6
-                    if (strstr(response.c_str(), "<user>")) break; // minicpm-v
-                    fflush(stdout);
-                }
-                common_sampler_free(smpl);
-            }
-        }
-        printf("\n");
-        llama_perf_context_print(ctx_llava->ctx_llama);
-
-        ctx_llava->model = NULL;
-        llava_free(ctx_llava);
-    }
-
-    return 0;
-}

examples/llava/mtmd-cli.cpp

@@ -28,15 +28,16 @@ static bool g_is_generating = false;
 
 /**
  * Please note that this is NOT a production-ready stuff.
- * It is a playground for trying Gemma 3 vision capabilities.
+ * It is a playground for trying multimodal support in llama.cpp.
  * For contributors: please keep this code simple and easy to understand.
  */
 
 static void show_additional_info(int /*argc*/, char ** argv) {
     LOG(
-        "Experimental CLI for using Gemma 3 vision model\n\n"
+        "Experimental CLI for multimodal\n\n"
         "Usage: %s [options] -m <model> --mmproj <mmproj> --image <image> -p <prompt>\n\n"
         "  -m and --mmproj are required\n"
+        "  -hf user/repo can replace both -m and --mmproj in most cases\n"
         "  --image and -p are optional, if NOT provided, the CLI will run in chat mode\n",
         argv[0]
     );
@@ -56,7 +57,7 @@ static void sigint_handler(int signo) {
 }
 #endif
 
-struct gemma3_context {
+struct mtmd_cli_context {
     mtmd_context_ptr ctx_vision;
     common_init_result llama_init;
 
@@ -70,18 +71,38 @@ struct gemma3_context {
     // so here we don't need to keep track of chat history
     common_chat_templates_ptr tmpls;
 
+    // support for legacy templates (models not having EOT token)
+    llama_tokens antiprompt_tokens;
+
     int n_threads    = 1;
     llama_pos n_past = 0;
 
-    gemma3_context(common_params & params) : llama_init(common_init_from_params(params)) {
+    mtmd_cli_context(common_params & params) : llama_init(common_init_from_params(params)) {
         model = llama_init.model.get();
         lctx = llama_init.context.get();
         vocab = llama_model_get_vocab(model);
         n_threads = params.cpuparams.n_threads;
         batch = llama_batch_init(params.n_batch, 0, 1);
         n_batch = params.n_batch;
+
+        if (!llama_model_chat_template(model, nullptr) && params.chat_template.empty()) {
+            LOG_ERR("Model does not have chat template.\n");
+            LOG_ERR("  For old llava models, you may need to use '--chat-template vicuna'\n");
+            LOG_ERR("  For MobileVLM models, use '--chat-template deepseek'\n");
+            exit(1);
+        }
+
         tmpls = common_chat_templates_init(model, params.chat_template);
+        LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(tmpls.get(), params.use_jinja).c_str());
+
         init_vision_context(params);
+
+        // load antiprompt tokens for legacy templates
+        if (params.chat_template == "vicuna") {
+            antiprompt_tokens = common_tokenize(lctx, "ASSISTANT:", false, true);
+        } else if (params.chat_template == "deepseek") {
+            antiprompt_tokens = common_tokenize(lctx, "###", false, true);
+        }
     }
 
     void init_vision_context(common_params & params) {
@@ -97,6 +118,17 @@ struct gemma3_context {
             exit(1);
         }
     }
+
+    bool check_antiprompt(const llama_tokens & generated_tokens) {
+        if (antiprompt_tokens.empty() || generated_tokens.size() < antiprompt_tokens.size()) {
+            return false;
+        }
+        return std::equal(
+            generated_tokens.end() - antiprompt_tokens.size(),
+            generated_tokens.end(),
+            antiprompt_tokens.begin()
+        );
+    }
 };
 
 struct decode_embd_batch {
@@ -132,7 +164,8 @@ struct decode_embd_batch {
     }
 };
 
-static int generate_response(gemma3_context & ctx, common_sampler * smpl, int n_predict) {
+static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int n_predict) {
+    llama_tokens generated_tokens;
     for (int i = 0; i < n_predict; i++) {
         if (i > n_predict || !g_is_generating) {
             printf("\n");
@@ -140,9 +173,10 @@ static int generate_response(gemma3_context & ctx, common_sampler * smpl, int n_
         }
 
         llama_token token_id = common_sampler_sample(smpl, ctx.lctx, -1);
+        generated_tokens.push_back(token_id);
         common_sampler_accept(smpl, token_id, true);
 
-        if (llama_vocab_is_eog(ctx.vocab, token_id)) {
+        if (llama_vocab_is_eog(ctx.vocab, token_id) || ctx.check_antiprompt(generated_tokens)) {
             printf("\n");
             break; // end of generation
         }
@@ -161,7 +195,7 @@ static int generate_response(gemma3_context & ctx, common_sampler * smpl, int n_
     return 0;
 }
 
-static int eval_message(gemma3_context & ctx, common_chat_msg & msg, std::vector<std::string> & images_fname, bool add_bos = false) {
+static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg, std::vector<std::string> & images_fname, bool add_bos = false) {
     std::vector<mtmd_bitmap> bitmaps;
 
     common_chat_templates_inputs tmpl_inputs;
@@ -184,18 +218,19 @@ static int eval_message(gemma3_context & ctx, common_chat_msg & msg, std::vector
     text.text          = formatted_chat.prompt;
     text.add_special   = add_bos;
     text.parse_special = true;
-    mtmd_input_chunks_ptr chunks(mtmd_tokenize(ctx.ctx_vision.get(), text, bitmaps));
-    if (chunks == nullptr) {
-        LOG_ERR("Unable to tokenize prompt\n");
+    mtmd_input_chunks chunks;
+    int32_t res = mtmd_tokenize(ctx.ctx_vision.get(), chunks, text, bitmaps);
+    if (res != 0) {
+        LOG_ERR("Unable to tokenize prompt, res = %d\n", res);
         return 1;
     }
 
-    if (mtmd_helper_eval(ctx.ctx_vision.get(), ctx.lctx, chunks.get(), ctx.n_past, 0, ctx.n_batch)) {
+    if (mtmd_helper_eval(ctx.ctx_vision.get(), ctx.lctx, chunks, ctx.n_past, 0, ctx.n_batch)) {
         LOG_ERR("Unable to eval prompt\n");
         return 1;
     }
 
-    ctx.n_past += mtmd_helper_get_n_tokens(chunks.get());
+    ctx.n_past += mtmd_helper_get_n_tokens(chunks);
 
     return 0;
 }
@@ -217,7 +252,7 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    gemma3_context ctx(params);
+    mtmd_cli_context ctx(params);
     printf("%s: %s\n", __func__, params.model.path.c_str());
 
     bool is_single_turn = !params.prompt.empty() && !params.image.empty();
@@ -317,6 +352,6 @@ int main(int argc, char ** argv) {
             is_first_msg = false;
         }
     }
-
+    llama_perf_context_print(ctx.lctx);
     return 0;
 }

examples/llava/mtmd.cpp

@@ -12,19 +12,43 @@
 #include <limits>
 #include <vector>
 
+// slice template, used by some llava-uhd models to correctly place the special tokens around image embeddings
+// models not having it (llava-1.6) will process embeddings without any special tokens in-between
+enum mtmd_slice_tmpl {
+    MTMD_SLICE_TMPL_NONE,
+    MTMD_SLICE_TMPL_MINICPMV_2_5,
+    MTMD_SLICE_TMPL_MINICPMV_2_6,
+    // TODO @ngxson : add support for idefics (SmolVLM)
+};
+
 struct mtmd_context {
     struct clip_ctx * ctx_clip;
     const struct llama_model * text_model;
     std::vector<float> image_embd_v; // image embedding vector
+
     bool print_timings;
     int n_threads;
     std::string image_marker;
 
+    // for minicpmv, we need special tokens in-between slices
+    mtmd_slice_tmpl slice_tmpl    = MTMD_SLICE_TMPL_NONE;
+    llama_token tok_ov_img_start  = LLAMA_TOKEN_NULL; // overview image
+    llama_token tok_ov_img_end    = LLAMA_TOKEN_NULL; // overview image
+    llama_token tok_slices_start  = LLAMA_TOKEN_NULL; // start of all slices
+    llama_token tok_slices_end    = LLAMA_TOKEN_NULL; // end of all slices
+    llama_token tok_sli_img_start = LLAMA_TOKEN_NULL; // single slice
+    llama_token tok_sli_img_end   = LLAMA_TOKEN_NULL; // single slice
+    llama_token tok_row_end       = LLAMA_TOKEN_NULL; // end of row
+
     // TODO @ngxson : add timings
 
     mtmd_context(const char * mmproj_fname,
                    const llama_model * text_model,
-                   const mtmd_context_params & ctx_params) : print_timings(ctx_params.print_timings), n_threads(ctx_params.n_threads), image_marker(ctx_params.image_marker) {
+                   const mtmd_context_params & ctx_params) :
+        print_timings(ctx_params.print_timings),
+        n_threads    (ctx_params.n_threads),
+        image_marker (ctx_params.image_marker)
+    {
         clip_context_params ctx_clip_params;
         ctx_clip_params.use_gpu   = ctx_params.use_gpu;
         ctx_clip_params.verbosity = ctx_params.verbosity;
@@ -33,11 +57,66 @@ struct mtmd_context {
             throw std::runtime_error(string_format("Failed to load CLIP model from %s\n", mmproj_fname));
         }
         this->text_model = text_model;
+
+        GGML_ASSERT(!clip_is_qwen2vl(ctx_clip) && "Qwen2VL model is not supported yet, use llama-qwen2vl-cli instead");
+
+        int minicpmv_version = clip_is_minicpmv(ctx_clip);
+        if (minicpmv_version == 2) {
+            // minicpmv 2.5 format:
+            // <image> (overview) </image><slice><image> (slice) </image><image> (slice) </image>\n ... </slice>
+            slice_tmpl        = MTMD_SLICE_TMPL_MINICPMV_2_5;
+            tok_ov_img_start  = lookup_token("<image>");
+            tok_ov_img_end    = lookup_token("</image>");
+            tok_slices_start  = lookup_token("<slice>");
+            tok_slices_end    = lookup_token("</slice>");
+            tok_sli_img_start = tok_ov_img_start;
+            tok_sli_img_end   = tok_ov_img_end;
+            tok_row_end       = lookup_token("\n");
+
+        } else if (minicpmv_version == 3 || minicpmv_version == 4) {
+            // minicpmv 2.6 format:
+            // <image> (overview) </image><slice> (slice) </slice><slice> (slice) </slice>\n ...
+            slice_tmpl        = MTMD_SLICE_TMPL_MINICPMV_2_6;
+            tok_ov_img_start  = lookup_token("<image>");
+            tok_ov_img_end    = lookup_token("</image>");
+            tok_sli_img_start = lookup_token("<slice>");
+            tok_sli_img_end   = lookup_token("</slice>");
+            tok_row_end       = lookup_token("\n");
+
+        } else if (minicpmv_version != 0) {
+            GGML_ASSERT(false && "unsupported minicpmv version");
+        }
     }
 
     ~mtmd_context() {
         clip_free(ctx_clip);
     }
+
+private:
+    llama_token lookup_token(const std::string & token_text) {
+        const llama_vocab * vocab = llama_model_get_vocab(text_model);
+        const int n_vocab = llama_vocab_n_tokens(vocab);
+        for (int i = 0; i < n_vocab; i++) {
+            if (token_to_piece(vocab, i, true) == token_text) {
+                return i;
+            }
+        }
+        return LLAMA_TOKEN_NULL;
+    }
+
+    std::string token_to_piece(const llama_vocab * vocab, llama_token token, bool special) {
+        std::string piece;
+        piece.resize(piece.capacity());  // using string internal cache, 15 bytes + '\n'
+        const int n_chars = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
+        if (n_chars < 0) {
+            piece.resize(-n_chars);
+            int check = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
+            GGML_ASSERT(check == -n_chars);
+        } else {
+            piece.resize(n_chars);
+        }
+        return piece;
+    }
 };
 
 struct mtmd_image_tokens_data {
@@ -49,6 +128,7 @@ struct mtmd_image_tokens {
     uint32_t ny; // number of tokens in y direction
     uint32_t n_tokens() const { return nx * ny; }
     clip_image_f32_batch batch_f32; // preprocessed image patches
+    std::string id; // optional user-defined ID, useful for KV cache tracking
 };
 
 mtmd_context * mtmd_init_from_file(const char * mmproj_fname,
@@ -88,29 +168,66 @@ static std::vector<llama_token> mtmd_tokenize_text_internal(
     return result;
 }
 
-mtmd_input_chunks * mtmd_tokenize(mtmd_context * ctx,
-                                const mtmd_input_text & text,
-                                const std::vector<mtmd_bitmap> & bitmaps) {
-    mtmd_input_chunks * output = new mtmd_input_chunks;
+int32_t mtmd_tokenize(mtmd_context * ctx,
+                        std::vector<mtmd_input_chunk> & output,
+                        const mtmd_input_text & text,
+                        const std::vector<mtmd_bitmap> & bitmaps) {
     auto vocab = llama_model_get_vocab(ctx->text_model);
 
     std::string prompt_modified(text.text);
     std::string marker_modified(ctx->image_marker);
-    projector_type proj_type = clip_get_projector_type(ctx->ctx_clip);
     // a bit hacky here, but works for now
     // for some models, we need to add prefix and suffix to the image embeddings
-    if (proj_type == PROJECTOR_TYPE_GEMMA3) {
+    if (clip_is_gemma3(ctx->ctx_clip)) {
+        // gemma 3
         // <start_of_image> ... (image embeddings) ... <end_of_image>
         marker_modified = "<start_of_image>" + ctx->image_marker + "<end_of_image>";
         string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
     }
 
-    std::vector<std::string> parts = string_split_str(text.text, ctx->image_marker);
-    output->clear();
-    output->reserve(parts.size());
+    // llava-1.5, llava-1.6, Yi-VL, Yi-34B, granite: don't need to add prefix and suffix
+    // for glm-edge, we don't need to add because the tokens are already in the returned embeddings
+
+    // TODO @ngxson : glm-edge : remove BOI / EOI tokens embeddings, decode them as normal tokens
+
+    std::vector<std::string> parts = string_split_str(prompt_modified, ctx->image_marker);
+    output.clear();
+    output.reserve(parts.size());
 
     size_t i_img = 0;
 
+    // utility for adding raw tokens
+    auto add_text_chunk = [&output](std::vector<llama_token> && tokens) {
+        mtmd_input_chunk chunk{
+            MTMD_INPUT_CHUNK_TYPE_TEXT,
+            std::move(tokens),
+            {},
+        };
+        output.emplace_back(std::move(chunk));
+    };
+
+    // utility for splitting batch of multiple images into chunks of batch having single images
+    auto split_batch_to_chunk = [&ctx](clip_image_f32_batch && batch_f32, const std::string & id) {
+        std::vector<mtmd_input_chunk> chunks;
+
+        for (auto & entry : batch_f32.entries) {
+            mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
+            image_tokens->nx = clip_n_patches(ctx->ctx_clip);
+            image_tokens->ny = 1;
+            image_tokens->batch_f32.entries.push_back(std::move(entry));
+            image_tokens->id = id;
+
+            mtmd_input_chunk chunk{
+                MTMD_INPUT_CHUNK_TYPE_IMAGE,
+                {},
+                std::move(image_tokens),
+            };
+            chunks.emplace_back(std::move(chunk));
+        }
+
+        return chunks;
+    };
+
     for (const auto & part : parts) {
         //printf("tokenizing part: %s\n", part.c_str());
         bool add_bos = &parts.front() == &part;
@@ -123,66 +240,156 @@ mtmd_input_chunks * mtmd_tokenize(mtmd_context * ctx,
             std::move(tokens),
             {},
         };
-        output->emplace_back(std::move(chunk));
+        output.emplace_back(std::move(chunk));
 
         if (&parts.back() != &part) {
             // add image token to middle of 2 parts
 
             if (i_img >= bitmaps.size()) {
                 LOG_ERR("%s: error: not enough images for %d parts\n", __func__, (int)parts.size());
-                return nullptr;
+                return 1;
             }
 
-            // shim layer
+            // convert mtmd_bitmap to clip_image_u8
             clip_image_u8_ptr img_u8(clip_image_u8_init());
             img_u8->nx = bitmaps[i_img].nx;
             img_u8->ny = bitmaps[i_img].ny;
             img_u8->buf.resize(bitmaps[i_img].data.size());
             std::memcpy(img_u8->buf.data(), bitmaps[i_img].data.data(), img_u8->nx * img_u8->ny * 3);
+            clip_image_size img_u8_size{img_u8->nx, img_u8->ny};
 
             // preprocess image
             clip_image_f32_batch batch_f32;
             bool ok = clip_image_preprocess(ctx->ctx_clip, img_u8.get(), &batch_f32);
             if (!ok) {
                 LOG_ERR("Unable to preprocess image\n");
-                return nullptr;
+                return 2;
             }
 
-            mtmd_image_tokens * image_tokens = new mtmd_image_tokens;
-            image_tokens->nx = clip_n_patches(ctx->ctx_clip); // TODO @ngxson : use clip_n_patches_by_image
-            image_tokens->ny = 1; // TODO
-            image_tokens->batch_f32 = std::move(batch_f32);
+            if (ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_5 || ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_6) {
+                // split batch into chunks of single images
+                auto chunks = split_batch_to_chunk(std::move(batch_f32), bitmaps[i_img].id);
+                GGML_ASSERT(chunks.size() > 0);
 
-            mtmd_input_chunk chunk{
-                MTMD_INPUT_CHUNK_TYPE_IMAGE,
-                {},
-                image_tokens,
-            };
-            output->emplace_back(std::move(chunk));
-            i_img++;
+                // add overview image
+                add_text_chunk({ctx->tok_ov_img_start});
+                output.emplace_back(std::move(chunks.front()));
+                chunks.erase(chunks.begin());
+                add_text_chunk({ctx->tok_ov_img_end});
+
+                // add slices
+                if (!chunks.empty()) {
+                    clip_add_load_image_size(ctx->ctx_clip, &img_u8_size);
+                    int n_col = clip_uhd_num_image_embeds_col(ctx->ctx_clip);
+                    int n_row = (int)chunks.size() / n_col;
+                    GGML_ASSERT(n_row * n_col == (int)chunks.size());
+                    if (ctx->tok_slices_start != LLAMA_TOKEN_NULL) {
+                        add_text_chunk({ctx->tok_slices_start});
+                    }
+                    for (int y = 0; y < n_row; y++) {
+                        for (int x = 0; x < n_col; x++) {
+                            if (ctx->tok_sli_img_start != LLAMA_TOKEN_NULL) {
+                                add_text_chunk({ctx->tok_sli_img_start});
+                            }
+                            output.emplace_back(std::move(chunks[y * n_col + x]));
+                            if (ctx->tok_sli_img_end != LLAMA_TOKEN_NULL) {
+                                add_text_chunk({ctx->tok_sli_img_end});
+                            }
+                        }
+                        if (ctx->tok_row_end != LLAMA_TOKEN_NULL && y != n_row - 1) {
+                            add_text_chunk({ctx->tok_row_end});
+                        }
+                    }
+                    if (ctx->tok_slices_end != LLAMA_TOKEN_NULL) {
+                        add_text_chunk({ctx->tok_slices_end});
+                    }
+                }
+
+            } else {
+                mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
+                image_tokens->nx = clip_n_patches(ctx->ctx_clip) * batch_f32.entries.size(); // TODO @ngxson : use clip_n_patches_by_image
+                image_tokens->ny = 1; // TODO
+                image_tokens->batch_f32 = std::move(batch_f32);
+                image_tokens->id = bitmaps[i_img].id; // optional
+
+                LOG_DBG("image_tokens->nx = %d\n", image_tokens->nx);
+                LOG_DBG("image_tokens->ny = %d\n", image_tokens->ny);
+                LOG_DBG("batch_f32 size = %d\n", (int)image_tokens->batch_f32.entries.size());
+
+                if (clip_is_glm(ctx->ctx_clip)) {
+                    // glm-edge
+                    image_tokens->nx += 2; // add 2 for the begin_of_image and end_of_image token embeddings
+                }
+
+                mtmd_input_chunk chunk{
+                    MTMD_INPUT_CHUNK_TYPE_IMAGE,
+                    {},
+                    std::move(image_tokens),
+                };
+                output.emplace_back(std::move(chunk));
+            }
+
+            i_img++; // move to next image
         }
     }
 
-    return output;
+    return 0;
 }
 
-void mtmd_input_chunks_free(mtmd_input_chunks * chunks) {
-    for (auto & chunk : *chunks) {
-        if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE && chunk.tokens_image) {
-            delete chunk.tokens_image;
-        }
+void mtmd_image_tokens_free(mtmd_image_tokens * image_tokens) {
+    if (image_tokens) {
+        delete image_tokens;
     }
-    delete chunks;
+}
+
+size_t mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * image_tokens) {
+    return image_tokens->n_tokens();
+}
+
+size_t mtmd_image_tokens_get_nx(const mtmd_image_tokens * image_tokens) {
+    return image_tokens->nx;
+}
+
+size_t mtmd_image_tokens_get_ny(const mtmd_image_tokens * image_tokens) {
+    return image_tokens->ny;
+}
+
+std::string mtmd_image_tokens_get_id(const mtmd_image_tokens * image_tokens) {
+    return image_tokens->id;
 }
 
 int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens) {
     int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip);
     ctx->image_embd_v.resize(image_tokens->n_tokens() * n_mmproj_embd);
-    bool ok = clip_image_batch_encode(
-        ctx->ctx_clip,
-        ctx->n_threads,
-        &image_tokens->batch_f32,
-        ctx->image_embd_v.data());
+    bool ok = false;
+
+    // only effective for minicpmv and qwen2vl, other models will ignore load_image_size
+    {
+        clip_image_size slice_size{
+            image_tokens->batch_f32.entries[0]->nx,
+            image_tokens->batch_f32.entries[0]->ny};
+        clip_add_load_image_size(ctx->ctx_clip, &slice_size);
+    }
+
+    if (clip_is_llava(ctx->ctx_clip) || clip_is_minicpmv(ctx->ctx_clip) || clip_is_glm(ctx->ctx_clip)) {
+        // TODO @ngxson : llava does not support batched encoding ; this should be fixed inside clip_image_batch_encode()
+        const auto & entries = image_tokens->batch_f32.entries;
+        for (size_t i = 0; i < entries.size(); i++) {
+            int n_tokens_per_image = clip_n_patches(ctx->ctx_clip);
+            ok = clip_image_encode(
+                ctx->ctx_clip,
+                ctx->n_threads,
+                entries[i].get(),
+                ctx->image_embd_v.data() + i*n_mmproj_embd*n_tokens_per_image);
+        }
+    } else {
+        ok = clip_image_batch_encode(
+            ctx->ctx_clip,
+            ctx->n_threads,
+            &image_tokens->batch_f32,
+            ctx->image_embd_v.data());
+    }
+
     return ok ? 0 : 1;
 }
 
@@ -190,9 +397,9 @@ float * mtmd_get_output_embd(mtmd_context * ctx) {
     return ctx->image_embd_v.data();
 }
 
-size_t mtmd_helper_get_n_tokens(mtmd_input_chunks * chunks) {
+size_t mtmd_helper_get_n_tokens(mtmd_input_chunks & chunks) {
     size_t n_tokens = 0;
-    for (auto & chunk : *chunks) {
+    for (auto & chunk : chunks) {
         if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
             n_tokens += chunk.tokens_text.size();
         } else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
@@ -241,35 +448,38 @@ struct decode_embd_batch {
 
 int32_t mtmd_helper_eval(mtmd_context * ctx,
         llama_context * lctx,
-        mtmd_input_chunks * chunks,
+        mtmd_input_chunks & chunks,
         llama_pos pos0,
         llama_seq_id seq_id,
         int32_t n_batch) {
     int32_t ret;
     llama_pos n_past = pos0;
     llama_batch text_batch = llama_batch_init(n_batch, 0, 1);
+    int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip);
 
-    for (auto & chunk : *chunks) {
-        bool is_last = &chunk == &chunks->back();
+    for (auto & chunk : chunks) {
+        bool is_last = &chunk == &chunks.back();
         if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
-            // TODO @ngxson : may need to split into smaller batches
             text_batch.n_tokens = chunk.tokens_text.size();
-            for (size_t i = 0; i < chunk.tokens_text.size(); i++) {
-                text_batch.token   [i]    = chunk.tokens_text[i];
-                text_batch.pos     [i]    = n_past++;
-                text_batch.n_seq_id[i]    = 1;
-                text_batch.seq_id  [i][0] = seq_id;
-                text_batch.logits  [i]    = false;
-            }
-            if (is_last) {
-                // always get logits for last input chunk
-                text_batch.logits[text_batch.n_tokens - 1] = true;
-            }
-            ret = llama_decode(lctx, text_batch);
-            if (ret != 0) {
-                LOG_ERR("failed to decode text\n");
-                llama_batch_free(text_batch);
-                return ret;
+            size_t i = 0;
+            while (i < chunk.tokens_text.size()) { // split into batches
+                for (; i < chunk.tokens_text.size() && text_batch.n_tokens < n_batch; i++) {
+                    text_batch.token   [i]    = chunk.tokens_text[i];
+                    text_batch.pos     [i]    = n_past++;
+                    text_batch.n_seq_id[i]    = 1;
+                    text_batch.seq_id  [i][0] = seq_id;
+                    text_batch.logits  [i]    = false;
+                }
+                if (is_last) {
+                    // always get logits for last input chunk
+                    text_batch.logits[text_batch.n_tokens - 1] = true;
+                }
+                ret = llama_decode(lctx, text_batch);
+                if (ret != 0) {
+                    LOG_ERR("failed to decode text\n");
+                    llama_batch_free(text_batch);
+                    return ret;
+                }
             }
 
         } else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
@@ -277,33 +487,56 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
             GGML_ASSERT(chunk.tokens_image != nullptr);
             int64_t t0 = ggml_time_ms();
             if (ctx->print_timings) {
-                LOG_INF("encoding image...\n");
+                LOG_INF("encoding image or slice...\n");
             }
-            ret = mtmd_encode(ctx, chunk.tokens_image);
+            ret = mtmd_encode(ctx, chunk.tokens_image.get());
             if (ret != 0) {
                 LOG_ERR("failed to encode image\n");
                 llama_batch_free(text_batch);
                 return ret;
             }
             if (ctx->print_timings) {
-                LOG_INF("image encoded in %" PRId64 " ms\n", ggml_time_ms() - t0);
+                LOG_INF("image/slice encoded in %" PRId64 " ms\n", ggml_time_ms() - t0);
             }
 
-            int32_t n_tokens = chunk.tokens_image->n_tokens();
+            int32_t n_tokens = mtmd_image_tokens_get_n_tokens(chunk.tokens_image.get());
+            int32_t i_batch = 0;
+            int32_t n_img_batches = GGML_PAD(n_tokens, n_batch) / n_batch;
             float * embd = mtmd_get_output_embd(ctx);
-            decode_embd_batch batch_img(embd, n_tokens, n_past, 0);
-            int64_t t1 = ggml_time_ms();
-            ret = llama_decode(lctx, batch_img.batch);
-            if (ret != 0) {
-                LOG_ERR("failed to decode image\n");
-                llama_batch_free(text_batch);
-                return ret;
-            }
-            if (ctx->print_timings) {
-                LOG_INF("image decoded in %" PRId64 " ms\n", ggml_time_ms() - t1);
+
+            if (mtmd_decode_use_non_causal(ctx)) {
+                llama_set_causal_attn(lctx, false);
+                // TODO @ngxson : need to make sure only one image is processed at a time, and n_ubatch must be enough to hold the image
             }
 
-            n_past += n_tokens;
+            while (i_batch < n_img_batches) { // split into batches
+                int32_t pos_offset = i_batch*n_batch;
+                int32_t n_tokens_batch = std::min(n_batch, n_tokens - pos_offset);
+                float * embd_batch = embd + pos_offset*n_mmproj_embd;
+                decode_embd_batch batch_img(embd_batch, n_tokens_batch, n_past, 0);
+
+                printf("decoding image batch %d/%d, n_tokens_batch = %d\n", i_batch+1, n_img_batches, n_tokens_batch);
+
+                int64_t t1 = ggml_time_ms();
+                ret = llama_decode(lctx, batch_img.batch);
+                if (ret != 0) {
+                    LOG_ERR("failed to decode image\n");
+                    llama_set_causal_attn(lctx, true); // restore causal attn
+                    llama_batch_free(text_batch);
+                    return ret;
+                }
+
+                if (ctx->print_timings) {
+                    LOG_INF("image decoded (batch %d/%d) in %" PRId64 " ms\n", i_batch+1, n_img_batches, ggml_time_ms() - t1);
+                }
+
+                i_batch++;
+                n_past += n_tokens_batch;
+            }
+
+            if (mtmd_decode_use_non_causal(ctx)) {
+                llama_set_causal_attn(lctx, true);
+            }
 
         } else {
             GGML_ASSERT(false && "chunk type not supported");
@@ -339,3 +572,15 @@ int32_t mtmd_helper_bitmap_init_from_file(const char * fname, mtmd_bitmap & outp
     std::memcpy(output.data.data(), data, output.nx * output.ny * 3);
     return 0;
 }
+
+bool mtmd_decode_use_non_causal(mtmd_context * ctx) {
+    projector_type proj_type = clip_get_projector_type(ctx->ctx_clip);
+    if (proj_type == PROJECTOR_TYPE_GEMMA3) {
+        return true;
+    }
+    return false;
+}
+
+void mtmd_image_tokens_deleter::operator()(mtmd_image_tokens * val) {
+    mtmd_image_tokens_free(val);
+}

examples/llava/mtmd.h

@@ -39,12 +39,18 @@ struct mtmd_bitmap {
     uint32_t nx;
     uint32_t ny;
     std::vector<unsigned char> data;
+    std::string id; // optional user-defined id, for ex: can be set to image hash, useful for KV cache tracking
 };
 
+struct mtmd_image_tokens_deleter {
+    void operator()(mtmd_image_tokens * val); // forward declaration
+};
+using mtmd_image_tokens_ptr = std::unique_ptr<mtmd_image_tokens, mtmd_image_tokens_deleter>;
+
 struct mtmd_input_chunk {
     mtmd_input_chunk_type type;
     std::vector<llama_token> tokens_text;
-    mtmd_image_tokens * tokens_image = nullptr;
+    mtmd_image_tokens_ptr tokens_image;
 };
 
 using mtmd_input_chunks = std::vector<mtmd_input_chunk>;
@@ -82,12 +88,21 @@ MTMD_API void mtmd_free(mtmd_context * ctx);
 //   3. "<end_of_image>\ndescribe it in detail."
 // number of bitmaps must be equal to the number of image markers in the prompt
 // this function is thread-safe (shared ctx)
-MTMD_API mtmd_input_chunks * mtmd_tokenize(mtmd_context * ctx,
+// return values:
+//   0 on success
+//   1 on number of images not matching the number of markers
+//   2 on image preprocessing error
+MTMD_API int32_t mtmd_tokenize(mtmd_context * ctx,
+                                std::vector<mtmd_input_chunk> & output,
                                 const mtmd_input_text & text,
                                 const std::vector<mtmd_bitmap> & bitmaps);
 
-// free image chunk data
-MTMD_API void mtmd_input_chunks_free(mtmd_input_chunks * chunks);
+// access mtmd_image_tokens
+MTMD_API size_t      mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * image_tokens);
+MTMD_API size_t      mtmd_image_tokens_get_nx(const mtmd_image_tokens * image_tokens);
+MTMD_API size_t      mtmd_image_tokens_get_ny(const mtmd_image_tokens * image_tokens);
+MTMD_API std::string mtmd_image_tokens_get_id(const mtmd_image_tokens * image_tokens);
+MTMD_API void        mtmd_image_tokens_free(mtmd_image_tokens * image_tokens);
 
 // returns 0 on success
 MTMD_API int32_t mtmd_encode(mtmd_context * ctx,
@@ -96,12 +111,17 @@ MTMD_API int32_t mtmd_encode(mtmd_context * ctx,
 // get output embeddings from the last encode pass
 MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx);
 
+// whether we need to set non-causal mask before llama_decode
+MTMD_API bool mtmd_decode_use_non_causal(mtmd_context * ctx);
+
+
+
 //
 // helper functions (can be implemented based on other functions)
 //
 
 // helper to count the total number of tokens from a list of chunks, useful to keep track of n_past
-MTMD_API size_t mtmd_helper_get_n_tokens(mtmd_input_chunks * chunks);
+MTMD_API size_t mtmd_helper_get_n_tokens(mtmd_input_chunks & chunks);
 
 // helper function that automatically:
 // 1. run llama_decode() on text chunks
@@ -110,7 +130,7 @@ MTMD_API size_t mtmd_helper_get_n_tokens(mtmd_input_chunks * chunks);
 // otherwise, returns 0 on success
 MTMD_API int32_t mtmd_helper_eval(mtmd_context * ctx,
                                 llama_context * lctx,
-                                mtmd_input_chunks * chunks,
+                                mtmd_input_chunks & chunks,
                                 llama_pos pos0,
                                 llama_seq_id seq_id,
                                 int32_t n_batch);
@@ -132,11 +152,6 @@ struct mtmd_context_deleter {
 };
 using mtmd_context_ptr = std::unique_ptr<mtmd_context, mtmd_context_deleter>;
 
-struct mtmd_input_chunks_deleter {
-    void operator()(mtmd_input_chunks * val) { mtmd_input_chunks_free(val); }
-};
-using mtmd_input_chunks_ptr = std::unique_ptr<mtmd_input_chunks, mtmd_input_chunks_deleter>;
-
 #else
 
 static_assert(false && "C header is not yet supported by this library");

examples/llava/tests.sh

@@ -17,26 +17,30 @@ cd $PROJ_ROOT
 
 arr_bin=()
 arr_hf=()
+arr_tmpl=() # chat template
 
 add_test() {
     local bin=$1
     local hf=$2
+    local tmpl=${3:-""} # default to empty string if not provided
     arr_bin+=("$bin")
     arr_hf+=("$hf")
+    arr_tmpl+=("$tmpl")
 }
 
-add_test "llama-gemma3-cli"   "ggml-org/gemma-3-4b-it-GGUF:Q4_K_M"
-add_test "llama-llava-cli"    "cmp-nct/Yi-VL-6B-GGUF:Q5_K"
-add_test "llama-llava-cli"    "guinmoon/MobileVLM-3B-GGUF:Q4_K_M"
-add_test "llama-llava-cli"    "THUDM/glm-edge-v-5b-gguf:Q4_K_M"
-add_test "llama-llava-cli"    "second-state/Llava-v1.5-7B-GGUF:Q2_K"
-add_test "llama-llava-cli"    "cjpais/llava-1.6-mistral-7b-gguf:Q3_K"
-add_test "llama-llava-cli"    "ibm-research/granite-vision-3.2-2b-GGUF:Q4_K_M"
-add_test "llama-minicpmv-cli" "second-state/MiniCPM-Llama3-V-2_5-GGUF:Q2_K" # model from openbmb is corrupted
-add_test "llama-minicpmv-cli" "openbmb/MiniCPM-V-2_6-gguf:Q2_K"
-add_test "llama-minicpmv-cli" "openbmb/MiniCPM-o-2_6-gguf:Q4_0"
+add_test "llama-mtmd-cli"  "ggml-org/gemma-3-4b-it-GGUF:Q4_K_M"
+add_test "llama-mtmd-cli"  "guinmoon/MobileVLM-3B-GGUF:Q4_K_M"               "deepseek"
+add_test "llama-mtmd-cli"  "THUDM/glm-edge-v-5b-gguf:Q4_K_M"
+add_test "llama-mtmd-cli"  "second-state/Llava-v1.5-7B-GGUF:Q2_K"            "vicuna"
+add_test "llama-mtmd-cli"  "cjpais/llava-1.6-mistral-7b-gguf:Q3_K"           "vicuna"
+add_test "llama-mtmd-cli"  "ibm-research/granite-vision-3.2-2b-GGUF:Q4_K_M"
+add_test "llama-mtmd-cli"  "second-state/MiniCPM-Llama3-V-2_5-GGUF:Q2_K"  # model from openbmb is corrupted
+add_test "llama-mtmd-cli"  "openbmb/MiniCPM-V-2_6-gguf:Q2_K"
+add_test "llama-mtmd-cli"  "openbmb/MiniCPM-o-2_6-gguf:Q4_0"
 add_test "llama-qwen2vl-cli"  "bartowski/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M"
 
+# add_test "llama-mtmd-cli"  "cmp-nct/Yi-VL-6B-GGUF:Q5_K"  # this model has broken chat template, not usable
+
 ###############
 
 cmake --build build -j --target "${arr_bin[@]}"
@@ -46,12 +50,20 @@ arr_res=()
 for i in "${!arr_bin[@]}"; do
     bin="${arr_bin[$i]}"
     hf="${arr_hf[$i]}"
+    tmpl="${arr_tmpl[$i]}"
 
     echo "Running test with binary: $bin and HF model: $hf"
     echo ""
     echo ""
 
-    output=$("$PROJ_ROOT/build/bin/$bin" -hf "$hf" --image $SCRIPT_DIR/test-1.jpeg -p "what is the publisher name of the newspaper?" --temp 0 2>&1 | tee /dev/tty)
+    output=$(\
+        "$PROJ_ROOT/build/bin/$bin" \
+        -hf "$hf" \
+        --image $SCRIPT_DIR/test-1.jpeg \
+        -p "what is the publisher name of the newspaper?" \
+        --temp 0 -n 128 \
+        ${tmpl:+--chat-template "$tmpl"} \
+        2>&1 | tee /dev/tty)
 
     echo "$output" > $SCRIPT_DIR/output/$bin-$(echo "$hf" | tr '/' '-').log
 

examples/main/main.cpp

@@ -865,9 +865,22 @@ int main(int argc, char ** argv) {
                 console::set_display(console::reset);
                 display = true;
 
-                // Add tokens to embd only if the input buffer is non-empty
-                // Entering a empty line lets the user pass control back
-                if (buffer.length() > 1) {
+                if (buffer.empty()) { // Ctrl+D on empty line exits
+                    LOG("EOF by user\n");
+                    break;
+                }
+
+                if (buffer.back() == '\n') {
+                    // Implement #587:
+                    // If the user wants the text to end in a newline,
+                    // this should be accomplished by explicitly adding a newline by using \ followed by return,
+                    // then returning control by pressing return again.
+                    buffer.pop_back();
+                }
+
+                if (buffer.empty()) { // Enter key on empty line lets the user pass control back
+                    LOG_DBG("empty line, passing control back\n");
+                } else { // Add tokens to embd only if the input buffer is non-empty
                     // append input suffix if any
                     if (!params.input_suffix.empty() && !params.conversation_mode) {
                         LOG_DBG("appending input suffix: '%s'\n", params.input_suffix.c_str());
@@ -915,8 +928,6 @@ int main(int argc, char ** argv) {
 
                     n_remain -= line_inp.size();
                     LOG_DBG("n_remain: %d\n", n_remain);
-                } else {
-                    LOG_DBG("empty line, passing control back\n");
                 }
 
                 input_echo = false; // do not echo this again

examples/rpc/rpc-server.cpp

@@ -297,7 +297,10 @@ int main(int argc, char * argv[]) {
         }
         cache_dir = cache_dir_str.c_str();
     }
-    printf("Starting RPC server\n");
+    printf("Starting RPC server v%d.%d.%d\n",
+           RPC_PROTO_MAJOR_VERSION,
+           RPC_PROTO_MINOR_VERSION,
+           RPC_PROTO_PATCH_VERSION);
     printf("  endpoint       : %s\n", endpoint.c_str());
     printf("  local cache    : %s\n", cache_dir ? cache_dir : "n/a");
     printf("  backend memory : %zu MB\n", free_mem / (1024 * 1024));

examples/server/server.cpp

@@ -1552,29 +1552,30 @@ struct server_queue {
     std::condition_variable condition_tasks;
 
     // callback functions
-    std::function<void(server_task)> callback_new_task;
-    std::function<void(void)>        callback_update_slots;
+    std::function<void(server_task &&)> callback_new_task;
+    std::function<void(void)>           callback_update_slots;
 
     // Add a new task to the end of the queue
-    int post(server_task task, bool front = false) {
+    int post(server_task && task, bool front = false) {
         std::unique_lock<std::mutex> lock(mutex_tasks);
         GGML_ASSERT(task.id != -1);
         // if this is cancel task make sure to clean up pending tasks
         if (task.type == SERVER_TASK_TYPE_CANCEL) {
             cleanup_pending_task(task.id_target);
         }
-        QUE_DBG("new task, id = %d, front = %d\n", task.id, front);
+        const int task_id = task.id;
+        QUE_DBG("new task, id = %d, front = %d\n", task_id, front);
         if (front) {
             queue_tasks.push_front(std::move(task));
         } else {
             queue_tasks.push_back(std::move(task));
         }
         condition_tasks.notify_one();
-        return task.id;
+        return task_id;
     }
 
     // multi-task version of post()
-    int post(std::vector<server_task> & tasks, bool front = false) {
+    int post(std::vector<server_task> && tasks, bool front = false) {
         std::unique_lock<std::mutex> lock(mutex_tasks);
         for (auto & task : tasks) {
             if (task.id == -1) {
@@ -1596,7 +1597,7 @@ struct server_queue {
     }
 
     // Add a new task, but defer until one slot is available
-    void defer(server_task task) {
+    void defer(server_task && task) {
         std::unique_lock<std::mutex> lock(mutex_tasks);
         QUE_DBG("defer task, id = %d\n", task.id);
         queue_tasks_deferred.push_back(std::move(task));
@@ -1611,7 +1612,7 @@ struct server_queue {
     }
 
     // Register function to process a new task
-    void on_new_task(std::function<void(server_task)> callback) {
+    void on_new_task(std::function<void(server_task &&)> callback) {
         callback_new_task = std::move(callback);
     }
 
@@ -1660,7 +1661,7 @@ struct server_queue {
                     lock.unlock();
                     break;
                 }
-                server_task task = queue_tasks.front();
+                server_task task = std::move(queue_tasks.front());
                 queue_tasks.pop_front();
                 lock.unlock();
 
@@ -2004,7 +2005,7 @@ struct server_context {
 
             slot.reset();
 
-            slots.push_back(slot);
+            slots.push_back(std::move(slot));
         }
 
         default_generation_settings_for_props = slots[0].to_json();
@@ -2105,7 +2106,7 @@ struct server_context {
         return true;
     }
 
-    bool launch_slot_with_task(server_slot & slot, const server_task & task) {
+    bool launch_slot_with_task(server_slot & slot, server_task && task) {
         slot.reset();
         slot.id_task       = task.id;
         slot.index         = task.index;
@@ -2113,10 +2114,10 @@ struct server_context {
         slot.params        = std::move(task.params);
         slot.prompt_tokens = std::move(task.prompt_tokens);
 
-        if (!are_lora_equal(task.params.lora, slot.lora)) {
+        if (!are_lora_equal(slot.params.lora, slot.lora)) {
             // if lora is changed, we cannot reuse cached tokens
             slot.cache_tokens.clear();
-            slot.lora = task.params.lora;
+            slot.lora = slot.params.lora;
         }
 
         bool can_detokenize = can_be_detokenized(ctx, slot.prompt_tokens);
@@ -2547,10 +2548,10 @@ struct server_context {
             server_task task(SERVER_TASK_TYPE_CANCEL);
             task.id_target = id_task;
             queue_results.remove_waiting_task_id(id_task);
-            cancel_tasks.push_back(task);
+            cancel_tasks.push_back(std::move(task));
         }
         // push to beginning of the queue, so it has highest priority
-        queue_tasks.post(cancel_tasks, true);
+        queue_tasks.post(std::move(cancel_tasks), true);
     }
 
     // receive the results from task(s)
@@ -2637,7 +2638,7 @@ struct server_context {
     // Functions to process the task
     //
 
-    void process_single_task(server_task task) {
+    void process_single_task(server_task && task) {
         switch (task.type) {
             case SERVER_TASK_TYPE_COMPLETION:
             case SERVER_TASK_TYPE_INFILL:
@@ -2651,17 +2652,17 @@ struct server_context {
                     if (slot == nullptr) {
                         // if no slot is available, we defer this task for processing later
                         SRV_DBG("no slot is available, defer task, id_task = %d\n", task.id);
-                        queue_tasks.defer(task);
+                        queue_tasks.defer(std::move(task));
                         break;
                     }
                     if (slot->is_processing()) {
                         // if requested slot is unavailable, we defer this task for processing later
                         SRV_DBG("requested slot is unavailable, defer task, id_task = %d\n", task.id);
-                        queue_tasks.defer(task);
+                        queue_tasks.defer(std::move(task));
                         break;
                     }
 
-                    if (!launch_slot_with_task(*slot, task)) {
+                    if (!launch_slot_with_task(*slot, std::move(task))) {
                         SRV_ERR("failed to launch slot with task, id_task = %d\n", task.id);
                         break;
                     }
@@ -2740,7 +2741,7 @@ struct server_context {
                     if (slot->is_processing()) {
                         // if requested slot is unavailable, we defer this task for processing later
                         SRV_DBG("requested slot is unavailable, defer task, id_task = %d\n", task.id);
-                        queue_tasks.defer(task);
+                        queue_tasks.defer(std::move(task));
                         break;
                     }
 
@@ -2776,7 +2777,7 @@ struct server_context {
                     if (slot->is_processing()) {
                         // if requested slot is unavailable, we defer this task for processing later
                         SRV_DBG("requested slot is unavailable, defer task, id_task = %d\n", task.id);
-                        queue_tasks.defer(task);
+                        queue_tasks.defer(std::move(task));
                         break;
                     }
 
@@ -2819,7 +2820,7 @@ struct server_context {
                     if (slot->is_processing()) {
                         // if requested slot is unavailable, we defer this task for processing later
                         SRV_DBG("requested slot is unavailable, defer task, id_task = %d\n", task.id);
-                        queue_tasks.defer(task);
+                        queue_tasks.defer(std::move(task));
                         break;
                     }
 
@@ -2871,7 +2872,7 @@ struct server_context {
 
             server_task task(SERVER_TASK_TYPE_NEXT_RESPONSE);
             task.id = queue_tasks.get_new_id();
-            queue_tasks.post(task);
+            queue_tasks.post(std::move(task));
         }
 
         // apply context-shift if needed
@@ -3633,14 +3634,17 @@ int main(int argc, char ** argv) {
         }
 
         // request slots data using task queue
-        server_task task(SERVER_TASK_TYPE_METRICS);
-        task.id = ctx_server.queue_tasks.get_new_id();
-        ctx_server.queue_results.add_waiting_task_id(task.id);
-        ctx_server.queue_tasks.post(task, true); // high-priority task
+        int task_id = ctx_server.queue_tasks.get_new_id();
+        {
+            server_task task(SERVER_TASK_TYPE_METRICS);
+            task.id = task_id;
+            ctx_server.queue_results.add_waiting_task_id(task_id);
+            ctx_server.queue_tasks.post(std::move(task), true); // high-priority task
+        }
 
         // get the result
-        server_task_result_ptr result = ctx_server.queue_results.recv(task.id);
-        ctx_server.queue_results.remove_waiting_task_id(task.id);
+        server_task_result_ptr result = ctx_server.queue_results.recv(task_id);
+        ctx_server.queue_results.remove_waiting_task_id(task_id);
 
         if (result->is_error()) {
             res_error(res, result->to_json());
@@ -3669,16 +3673,17 @@ int main(int argc, char ** argv) {
         }
 
         // request slots data using task queue
-        server_task task(SERVER_TASK_TYPE_METRICS);
-        task.id = ctx_server.queue_tasks.get_new_id();
-        task.metrics_reset_bucket = true;
-
-        ctx_server.queue_results.add_waiting_task_id(task.id);
-        ctx_server.queue_tasks.post(task, true); // high-priority task
+        int task_id = ctx_server.queue_tasks.get_new_id();
+        {
+            server_task task(SERVER_TASK_TYPE_METRICS);
+            task.id = task_id;
+            ctx_server.queue_results.add_waiting_task_id(task_id);
+            ctx_server.queue_tasks.post(std::move(task), true); // high-priority task
+        }
 
         // get the result
-        server_task_result_ptr result = ctx_server.queue_results.recv(task.id);
-        ctx_server.queue_results.remove_waiting_task_id(task.id);
+        server_task_result_ptr result = ctx_server.queue_results.recv(task_id);
+        ctx_server.queue_results.remove_waiting_task_id(task_id);
 
         if (result->is_error()) {
             res_error(res, result->to_json());
@@ -3775,17 +3780,20 @@ int main(int argc, char ** argv) {
         }
         std::string filepath = params.slot_save_path + filename;
 
-        server_task task(SERVER_TASK_TYPE_SLOT_SAVE);
-        task.id = ctx_server.queue_tasks.get_new_id();
-        task.slot_action.slot_id  = id_slot;
-        task.slot_action.filename = filename;
-        task.slot_action.filepath = filepath;
+        int task_id = ctx_server.queue_tasks.get_new_id();
+        {
+            server_task task(SERVER_TASK_TYPE_SLOT_SAVE);
+            task.id = task_id;
+            task.slot_action.slot_id  = id_slot;
+            task.slot_action.filename = filename;
+            task.slot_action.filepath = filepath;
 
-        ctx_server.queue_results.add_waiting_task_id(task.id);
-        ctx_server.queue_tasks.post(task);
+            ctx_server.queue_results.add_waiting_task_id(task_id);
+            ctx_server.queue_tasks.post(std::move(task));
+        }
 
-        server_task_result_ptr result = ctx_server.queue_results.recv(task.id);
-        ctx_server.queue_results.remove_waiting_task_id(task.id);
+        server_task_result_ptr result = ctx_server.queue_results.recv(task_id);
+        ctx_server.queue_results.remove_waiting_task_id(task_id);
 
         if (result->is_error()) {
             res_error(res, result->to_json());
@@ -3804,17 +3812,20 @@ int main(int argc, char ** argv) {
         }
         std::string filepath = params.slot_save_path + filename;
 
-        server_task task(SERVER_TASK_TYPE_SLOT_RESTORE);
-        task.id = ctx_server.queue_tasks.get_new_id();
-        task.slot_action.slot_id  = id_slot;
-        task.slot_action.filename = filename;
-        task.slot_action.filepath = filepath;
+        int task_id = ctx_server.queue_tasks.get_new_id();
+        {
+            server_task task(SERVER_TASK_TYPE_SLOT_RESTORE);
+            task.id = task_id;
+            task.slot_action.slot_id  = id_slot;
+            task.slot_action.filename = filename;
+            task.slot_action.filepath = filepath;
 
-        ctx_server.queue_results.add_waiting_task_id(task.id);
-        ctx_server.queue_tasks.post(task);
+            ctx_server.queue_results.add_waiting_task_id(task_id);
+            ctx_server.queue_tasks.post(std::move(task));
+        }
 
-        server_task_result_ptr result = ctx_server.queue_results.recv(task.id);
-        ctx_server.queue_results.remove_waiting_task_id(task.id);
+        server_task_result_ptr result = ctx_server.queue_results.recv(task_id);
+        ctx_server.queue_results.remove_waiting_task_id(task_id);
 
         if (result->is_error()) {
             res_error(res, result->to_json());
@@ -3826,15 +3837,18 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_slots_erase = [&ctx_server, &res_error, &res_ok](const httplib::Request & /* req */, httplib::Response & res, int id_slot) {
-        server_task task(SERVER_TASK_TYPE_SLOT_ERASE);
-        task.id = ctx_server.queue_tasks.get_new_id();
-        task.slot_action.slot_id = id_slot;
+        int task_id = ctx_server.queue_tasks.get_new_id();
+        {
+            server_task task(SERVER_TASK_TYPE_SLOT_ERASE);
+            task.id = task_id;
+            task.slot_action.slot_id = id_slot;
 
-        ctx_server.queue_results.add_waiting_task_id(task.id);
-        ctx_server.queue_tasks.post(task);
+            ctx_server.queue_results.add_waiting_task_id(task_id);
+            ctx_server.queue_tasks.post(std::move(task));
+        }
 
-        server_task_result_ptr result = ctx_server.queue_results.recv(task.id);
-        ctx_server.queue_results.remove_waiting_task_id(task.id);
+        server_task_result_ptr result = ctx_server.queue_results.recv(task_id);
+        ctx_server.queue_results.remove_waiting_task_id(task_id);
 
         if (result->is_error()) {
             res_error(res, result->to_json());
@@ -3938,9 +3952,10 @@ int main(int argc, char ** argv) {
         }
 
         auto completion_id = gen_chatcmplid();
-        std::vector<server_task> tasks;
-
+        std::unordered_set<int> task_ids;
         try {
+            std::vector<server_task> tasks;
+
             const auto & prompt = data.at("prompt");
             // TODO: this log can become very long, put it behind a flag or think about a more compact format
             //SRV_DBG("Prompt: %s\n", prompt.is_string() ? prompt.get<std::string>().c_str() : prompt.dump(2).c_str());
@@ -3955,9 +3970,9 @@ int main(int argc, char ** argv) {
 
                 task.prompt_tokens    = std::move(tokenized_prompts[i]);
                 task.params           = server_task::params_from_json_cmpl(
-                                            ctx_server.ctx,
-                                            ctx_server.params_base,
-                                            data);
+                        ctx_server.ctx,
+                        ctx_server.params_base,
+                        data);
                 task.id_selected_slot = json_value(data, "id_slot", -1);
 
                 // OAI-compat
@@ -3965,18 +3980,18 @@ int main(int argc, char ** argv) {
                 task.params.oaicompat_cmpl_id         = completion_id;
                 // oaicompat_model is already populated by params_from_json_cmpl
 
-                tasks.push_back(task);
+                tasks.push_back(std::move(task));
             }
+
+            task_ids = server_task::get_list_id(tasks);
+            ctx_server.queue_results.add_waiting_tasks(tasks);
+            ctx_server.queue_tasks.post(std::move(tasks));
         } catch (const std::exception & e) {
             res_error(res, format_error_response(e.what(), ERROR_TYPE_INVALID_REQUEST));
             return;
         }
 
-        ctx_server.queue_results.add_waiting_tasks(tasks);
-        ctx_server.queue_tasks.post(tasks);
-
         bool stream = json_value(data, "stream", false);
-        const auto task_ids = server_task::get_list_id(tasks);
 
         if (!stream) {
             ctx_server.receive_multi_results(task_ids, [&](std::vector<server_task_result_ptr> & results) {
@@ -4268,6 +4283,7 @@ int main(int argc, char ** argv) {
         // create and queue the task
         json responses = json::array();
         bool error = false;
+        std::unordered_set<int> task_ids;
         {
             std::vector<server_task> tasks;
             for (size_t i = 0; i < tokenized_prompts.size(); i++) {
@@ -4280,28 +4296,27 @@ int main(int argc, char ** argv) {
                 // OAI-compat
                 task.params.oaicompat = oaicompat;
 
-                tasks.push_back(task);
+                tasks.push_back(std::move(task));
             }
 
+            task_ids = server_task::get_list_id(tasks);
             ctx_server.queue_results.add_waiting_tasks(tasks);
-            ctx_server.queue_tasks.post(tasks);
-
-            // get the result
-            std::unordered_set<int> task_ids = server_task::get_list_id(tasks);
-
-            ctx_server.receive_multi_results(task_ids, [&](std::vector<server_task_result_ptr> & results) {
-                for (auto & res : results) {
-                    GGML_ASSERT(dynamic_cast<server_task_result_embd*>(res.get()) != nullptr);
-                    responses.push_back(res->to_json());
-                }
-            }, [&](const json & error_data) {
-                res_error(res, error_data);
-                error = true;
-            }, req.is_connection_closed);
-
-            ctx_server.queue_results.remove_waiting_task_ids(task_ids);
+            ctx_server.queue_tasks.post(std::move(tasks));
         }
 
+        // get the result
+        ctx_server.receive_multi_results(task_ids, [&](std::vector<server_task_result_ptr> & results) {
+            for (auto & res : results) {
+                GGML_ASSERT(dynamic_cast<server_task_result_embd*>(res.get()) != nullptr);
+                responses.push_back(res->to_json());
+            }
+        }, [&](const json & error_data) {
+            res_error(res, error_data);
+            error = true;
+        }, req.is_connection_closed);
+
+        ctx_server.queue_results.remove_waiting_task_ids(task_ids);
+
         if (error) {
             return;
         }
@@ -4367,6 +4382,7 @@ int main(int argc, char ** argv) {
         // create and queue the task
         json responses = json::array();
         bool error = false;
+        std::unordered_set<int> task_ids;
         {
             std::vector<server_task> tasks;
             std::vector<llama_tokens> tokenized_docs = tokenize_input_prompts(ctx_server.vocab, documents, /* add_special */ false, true);
@@ -4376,26 +4392,24 @@ int main(int argc, char ** argv) {
                 task.id            = ctx_server.queue_tasks.get_new_id();
                 task.index         = i;
                 task.prompt_tokens = format_rerank(ctx_server.vocab, tokenized_query, tokenized_docs[i]);
-                tasks.push_back(task);
+                tasks.push_back(std::move(task));
             }
 
+            task_ids = server_task::get_list_id(tasks);
             ctx_server.queue_results.add_waiting_tasks(tasks);
-            ctx_server.queue_tasks.post(tasks);
-
-            // get the result
-            std::unordered_set<int> task_ids = server_task::get_list_id(tasks);
-
-            ctx_server.receive_multi_results(task_ids, [&](std::vector<server_task_result_ptr> & results) {
-                for (auto & res : results) {
-                    GGML_ASSERT(dynamic_cast<server_task_result_rerank*>(res.get()) != nullptr);
-                    responses.push_back(res->to_json());
-                }
-            }, [&](const json & error_data) {
-                res_error(res, error_data);
-                error = true;
-            }, req.is_connection_closed);
+            ctx_server.queue_tasks.post(std::move(tasks));
         }
 
+        ctx_server.receive_multi_results(task_ids, [&](std::vector<server_task_result_ptr> & results) {
+            for (auto & res : results) {
+                GGML_ASSERT(dynamic_cast<server_task_result_rerank*>(res.get()) != nullptr);
+                responses.push_back(res->to_json());
+            }
+        }, [&](const json & error_data) {
+            res_error(res, error_data);
+            error = true;
+        }, req.is_connection_closed);
+
         if (error) {
             return;
         }
@@ -4431,14 +4445,19 @@ int main(int argc, char ** argv) {
             res_error(res, format_error_response("Request body must be an array", ERROR_TYPE_INVALID_REQUEST));
             return;
         }
-        server_task task(SERVER_TASK_TYPE_SET_LORA);
-        task.id = ctx_server.queue_tasks.get_new_id();
-        task.set_lora = parse_lora_request(ctx_server.params_base.lora_adapters, body);
-        ctx_server.queue_results.add_waiting_task_id(task.id);
-        ctx_server.queue_tasks.post(task);
 
-        server_task_result_ptr result = ctx_server.queue_results.recv(task.id);
-        ctx_server.queue_results.remove_waiting_task_id(task.id);
+        int task_id = ctx_server.queue_tasks.get_new_id();
+        {
+            server_task task(SERVER_TASK_TYPE_SET_LORA);
+            task.id = task_id;
+            task.set_lora = parse_lora_request(ctx_server.params_base.lora_adapters, body);
+            ctx_server.queue_results.add_waiting_task_id(task_id);
+            ctx_server.queue_tasks.post(std::move(task));
+        }
+
+        // get the result
+        server_task_result_ptr result = ctx_server.queue_results.recv(task_id);
+        ctx_server.queue_results.remove_waiting_task_id(task_id);
 
         if (result->is_error()) {
             res_error(res, result->to_json());
@@ -4582,8 +4601,8 @@ int main(int argc, char ** argv) {
         common_chat_templates_source(ctx_server.chat_templates.get()),
         common_chat_format_example(ctx_server.chat_templates.get(), ctx_server.params_base.use_jinja).c_str());
 
-    ctx_server.queue_tasks.on_new_task([&ctx_server](const server_task & task) {
-        ctx_server.process_single_task(task);
+    ctx_server.queue_tasks.on_new_task([&ctx_server](server_task && task) {
+        ctx_server.process_single_task(std::move(task));
     });
 
     ctx_server.queue_tasks.on_update_slots([&ctx_server]() {

examples/sycl/build.sh

@@ -8,10 +8,10 @@ cd build
 source /opt/intel/oneapi/setvars.sh
 
 #for FP16
-#cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON # faster for long-prompt inference
+#cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON -DLLAMA_CURL=OFF # faster for long-prompt inference
 
 #for FP32
-cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=OFF
 
 #build example/main
 #cmake --build . --config Release --target main

ggml/CMakeLists.txt

@@ -107,6 +107,7 @@ message(DEBUG "INS_ENB             : ${INS_ENB}")
 option(GGML_CPU_HBM          "ggml: use memkind for CPU HBM" OFF)
 option(GGML_CPU_AARCH64      "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
 option(GGML_CPU_KLEIDIAI     "ggml: use KleidiAI optimized kernels if applicable" OFF)
+option(GGML_SSE42            "ggml: enable SSE 4.2"          ${INS_ENB})
 option(GGML_AVX              "ggml: enable AVX"              ${INS_ENB})
 option(GGML_AVX_VNNI         "ggml: enable AVX-VNNI"         OFF)
 option(GGML_AVX2             "ggml: enable AVX2"             ${INS_ENB})
@@ -170,7 +171,6 @@ option(GGML_HIP                             "ggml: use HIP"
 option(GGML_HIP_GRAPHS                      "ggml: use HIP graph, experimental, slow"         OFF)
 option(GGML_HIP_NO_VMM                      "ggml: do not try to use HIP VMM"                 ON)
 option(GGML_HIP_ROCWMMA_FATTN               "ggml: enable rocWMMA for FlashAttention"         OFF)
-option(GGML_HIP_UMA                         "ggml: use HIP unified memory architecture"       OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
 option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
 option(GGML_VULKAN_DEBUG                    "ggml: enable Vulkan debug output"                OFF)

ggml/include/ggml-rpc.h

@@ -7,6 +7,9 @@
 extern "C" {
 #endif
 
+#define RPC_PROTO_MAJOR_VERSION    1
+#define RPC_PROTO_MINOR_VERSION    0
+#define RPC_PROTO_PATCH_VERSION    0
 #define GGML_RPC_MAX_SERVERS       16
 
 // backend API

ggml/src/CMakeLists.txt

@@ -267,6 +267,7 @@ function(ggml_add_cpu_backend_variant tag_name)
     set(GGML_CPU_TAG_NAME ${tag_name})
     # other: OPENMP LLAMAFILE CPU_HBM
     foreach (feat NATIVE
+                  SSE42
                   AVX AVX2 BMI2 AVX_VNNI FMA F16C
                   AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16
                   AMX_TILE AMX_INT8 AMX_BF16)
@@ -286,14 +287,16 @@ if (GGML_CPU_ALL_VARIANTS)
     if (NOT GGML_BACKEND_DL)
         message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS requires GGML_BACKEND_DL")
     endif()
-    ggml_add_cpu_backend_variant(sandybridge    AVX)
-    ggml_add_cpu_backend_variant(haswell        AVX F16C AVX2 BMI2 FMA)
-    ggml_add_cpu_backend_variant(skylakex       AVX F16C AVX2 BMI2 FMA AVX512)
-    ggml_add_cpu_backend_variant(icelake        AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
-    ggml_add_cpu_backend_variant(alderlake      AVX F16C AVX2 BMI2 FMA AVX_VNNI)
+    ggml_add_cpu_backend_variant(x64)
+    ggml_add_cpu_backend_variant(sse42        SSE42)
+    ggml_add_cpu_backend_variant(sandybridge  SSE42 AVX)
+    ggml_add_cpu_backend_variant(haswell      SSE42 AVX F16C AVX2 BMI2 FMA)
+    ggml_add_cpu_backend_variant(skylakex     SSE42 AVX F16C AVX2 BMI2 FMA AVX512)
+    ggml_add_cpu_backend_variant(icelake      SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
+    ggml_add_cpu_backend_variant(alderlake    SSE42 AVX F16C AVX2 BMI2 FMA AVX_VNNI)
     if (NOT MSVC)
         # MSVC doesn't support AMX
-        ggml_add_cpu_backend_variant(sapphirerapids AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
+        ggml_add_cpu_backend_variant(sapphirerapids SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
     endif()
 elseif (GGML_CPU)
     ggml_add_cpu_backend_variant_impl("")

ggml/src/ggml-cann/aclnn_ops.h

@@ -23,6 +23,7 @@
 #ifndef CANN_ACLNN_OPS
 #define CANN_ACLNN_OPS
 
+#include <functional>
 #include <aclnnop/aclnn_abs.h>
 #include <aclnnop/aclnn_neg.h>
 #include <aclnnop/aclnn_exp.h>
@@ -713,6 +714,270 @@ void ggml_cann_count_equal(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  */
 void ggml_cann_step(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
+/*
+ * @brief A generic wrapper for ACL resources with custom deleter support.
+ */
+using any_acl_resource = std::unique_ptr<void, std::function<void(void*)>>;
+
+/**
+ * @brief Trait structure used to define how to destroy a given ACL resource type.
+ *
+ * @tparam T ACL resource type.
+ */
+template<typename T>
+struct acl_resource_traits;
+
+/**
+ * @brief Specialization for aclTensor, defines how to destroy an aclTensor resource.
+ */
+template<>
+struct acl_resource_traits<aclTensor> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyTensor(static_cast<aclTensor*>(p)));
+    }
+};
+
+/**
+ * @brief Specialization for aclIntArray, defines how to destroy an aclIntArray resource.
+ */
+template<>
+struct acl_resource_traits<aclIntArray> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyIntArray(static_cast<aclIntArray*>(p)));
+    }
+};
+
+/**
+ * @brief Specialization for aclScalar, defines how to destroy an aclScalar resource.
+ */
+template<>
+struct acl_resource_traits<aclScalar> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyScalar(static_cast<aclScalar*>(p)));
+    }
+};
+
+/**
+ * @brief Specialization for aclTensorList, defines how to destroy an aclTensorList resource.
+ */
+template<>
+struct acl_resource_traits<aclTensorList> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyTensorList(static_cast<aclTensorList*>(p)));
+    }
+};
+
+/**
+ * @brief Creates a generic ACL resource wrapper with proper destruction logic.
+ *
+ * @tparam T ACL resource type.
+ * @param ptr Raw pointer to ACL resource.
+ * @return any_acl_resource Smart pointer that handles destruction.
+ */
+template<typename T>
+any_acl_resource make_acl_resource(T* ptr) {
+    return any_acl_resource(
+        static_cast<void*>(ptr),
+        [](void* p) {
+            acl_resource_traits<T>::destroy(p);
+        }
+    );
+}
+
+/**
+ * @brief Registers multiple ACL resources into a vector for lifetime management.
+ *
+ * @tparam Args Variadic list of ACL resource types.
+ * @param vec Target vector to hold ACL resources.
+ * @param args Raw pointers to ACL resources.
+ */
+template<typename... Args>
+void register_acl_resources(std::vector<any_acl_resource>& vec, Args*... args) {
+    (vec.emplace_back(make_acl_resource(args)), ...);
+}
+
+/**
+ * @brief Task class that wraps the execution of an aclnn function call.
+ */
+class aclnn_task : public cann_task {
+    public:
+        aclnn_task(aclnn_func_t aclnn_func, void * workspace_addr,
+                   uint64_t workspace_size, aclOpExecutor * executor,
+                   aclrtStream stream) :
+            aclnn_func_(aclnn_func),
+            workspace_addr_(workspace_addr),
+            workspace_size_(workspace_size),
+            executor_(executor),
+            stream_(stream) {}
+        virtual void run_task() override {
+            ACL_CHECK(aclnn_func_(workspace_addr_, workspace_size_, executor_, stream_));
+        }
+    private:
+        aclnn_func_t aclnn_func_;
+        void *          workspace_addr_;
+        uint64_t        workspace_size_;
+        aclOpExecutor * executor_;
+        aclrtStream     stream_;
+};
+
+/**
+ * @brief Task class that releases ACL resources after usage.
+ */
+class release_resource_task : public cann_task {
+public:
+    release_resource_task(std::vector<any_acl_resource>&& resources){
+        resource_ = std::move(resources);
+    }
+
+    virtual void run_task() override {
+        resource_.clear();
+    }
+private:
+    std::vector<any_acl_resource> resource_;
+};
+
+/**
+ * @brief Task class for performing asynchronous memory copy operations.
+ */
+class async_memcpy_task : public cann_task {
+public:
+    async_memcpy_task(void* dst, const void* src, size_t size,
+                      aclrtMemcpyKind kind, aclrtStream stream)
+        : dst_(dst), src_(src), size_(size), kind_(kind), stream_(stream) {}
+
+    virtual void run_task() override {
+        ACL_CHECK(aclrtMemcpyAsync(dst_, size_, src_, size_, kind_, stream_));
+    }
+private:
+    void* dst_;
+    const void* src_;
+    size_t size_;
+    aclrtMemcpyKind kind_;
+    aclrtStream stream_;
+};
+
+/**
+ * @brief Task class for performing asynchronous memory set operations.
+ */
+class async_memset_task : public cann_task {
+    public:
+    async_memset_task(void* buffer, size_t size, int32_t value, aclrtStream stream)
+            : buffer_(buffer), size_(size), value_(value), stream_(stream) {}
+
+        virtual void run_task() override {
+            ACL_CHECK(aclrtMemsetAsync(buffer_, size_, value_, size_, stream_));
+        }
+    private:
+        void* buffer_;
+        size_t size_;
+        int32_t value_;
+        aclrtStream stream_;
+};
+
+/**
+ * @brief Launches an asynchronous task using the memory allocator.
+ *
+ * This macro submit an asynchronous task on the specified stream.
+ * The task uses memory allocated by the allocator. It is guaranteed
+ * that the memory will not be accessed by other tasks until this task
+ * completes, due to the sequential execution order within the same stream.
+ *
+ * @param OP_NAME aclnn operator name.
+ * @param args Additional arguments required by the task.
+ *
+ * @note
+ * Memory from the allocator will be "freed" immediately and can be
+ * reallocated to other pointers. However, it won't be accessed by any
+ * other task before this asynchronous task ends, because all tasks in the
+ * same stream are executed in queue order.
+ */
+
+#define GGML_CANN_CALL_ACLNN_OP(CTX, OP_NAME, ...)                                          \
+    do {                                                                                    \
+        uint64_t        workspaceSize = 0;                                                  \
+        aclOpExecutor * executor;                                                           \
+        void *          workspaceAddr = nullptr;                                            \
+        ACL_CHECK(aclnn##OP_NAME##GetWorkspaceSize(__VA_ARGS__, &workspaceSize, &executor));\
+        /* workspace should alloced in main thread to keep malloc order when using vmm. */  \
+        if (workspaceSize > 0) {                                                            \
+            ggml_cann_pool_alloc workspace_allocator(CTX.pool(), workspaceSize);            \
+            workspaceAddr = workspace_allocator.get();                                      \
+        }                                                                                   \
+        if (CTX.async_mode) {                                                               \
+            auto task =                                                                     \
+                std::make_unique<aclnn_task>(aclnn##OP_NAME, workspaceAddr, workspaceSize,  \
+                    executor, CTX.stream()); \
+            CTX.task_queue.submit_task(std::move(task));                                    \
+        } else {                                                                            \
+            ACL_CHECK(aclnn##OP_NAME(workspaceAddr, workspaceSize, executor, CTX.stream()));\
+        }                                                                                   \
+    } while (0)
+
+/**
+ * @brief Registers and releases multiple ACL resources, optionally deferring the release
+ *        using a task.
+ *
+ * @tparam Args Types of the ACL resources.
+ * @param ctx Backend context which manages task submission and async mode.
+ * @param args Pointers to ACL resources to be released.
+ */
+template <typename... Args>
+void ggml_cann_release_resources(ggml_backend_cann_context & ctx, Args &&... args) {
+    std::vector<any_acl_resource> resources;
+    register_acl_resources(resources, std::forward<Args>(args)...);
+    if(ctx.async_mode) {
+        auto task = std::make_unique<release_resource_task>(std::move(resources));
+        ctx.task_queue.submit_task(std::move(task));
+    }
+}
+
+/**
+ * @brief Performs an asynchronous memory copy operation, optionally deferred via task submission.
+ *
+ * @param ctx Backend context containing stream and async configuration.
+ * @param dst Destination memory address.
+ * @param src Source memory address.
+ * @param len Size of memory to copy (in bytes).
+ * @param kind Type of memory copy (host-to-device, device-to-host, etc).
+ */
+inline void ggml_cann_async_memcpy(ggml_backend_cann_context & ctx, void * dst,
+                                   const void * src, size_t len, aclrtMemcpyKind kind) {
+    if (ctx.async_mode) {
+        auto task = std::make_unique<async_memcpy_task>(dst, const_cast<void *>(src), len, kind, ctx.stream());
+        ctx.task_queue.submit_task(std::move(task));
+    } else {
+        ACL_CHECK(aclrtMemcpyAsync(dst, len, src, len, kind, ctx.stream()));
+    }
+}
+
+inline void ggml_cann_async_memcpy(ggml_backend_cann_context * ctx, void * dst,
+                                   const void * src, size_t len, aclrtMemcpyKind kind) {
+    if (ctx->async_mode) {
+        auto task = std::make_unique<async_memcpy_task>(dst, const_cast<void *>(src), len, kind, ctx->stream());
+        ctx->task_queue.submit_task(std::move(task));
+    } else {
+        ACL_CHECK(aclrtMemcpyAsync(dst, len, src, len, kind, ctx->stream()));
+    }
+}
+
+/**
+ * @brief Performs an asynchronous memory set operation, optionally deferred via task submission.
+ *
+ * @param ctx Backend context containing stream and async configuration.
+ * @param buffer Memory buffer to be set.
+ * @param size Size of the memory buffer (in bytes).
+ * @param value Value to set in the buffer.
+ */
+inline void ggml_cann_async_memset(ggml_backend_cann_context & ctx, void * buffer,
+                                   size_t size, int value) {
+    if (ctx.async_mode) {
+        auto task = std::make_unique<async_memset_task>(buffer, size, value, ctx.stream());
+        ctx.task_queue.submit_task(std::move(task));
+    } else {
+        ACL_CHECK(aclrtMemsetAsync(buffer, size, value, size, ctx.stream()));
+    }
+}
+
 /**
  * @brief Applies a element-wise operation to two input tensors using the CANN
  * backend.
@@ -742,42 +1007,9 @@ void ggml_cann_binary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     bcast_shape(src0, src1, dst, &acl_src0, &acl_src1, &acl_dst);
     binary_op(ctx, acl_src0, acl_src1, acl_dst);
 
-    ACL_CHECK(aclDestroyTensor(acl_src0));
-    ACL_CHECK(aclDestroyTensor(acl_src1));
-    ACL_CHECK(aclDestroyTensor(acl_dst));
+    ggml_cann_release_resources(ctx, acl_src0, acl_src1, acl_dst);
 }
 
-/**
- * @brief Launches an asynchronous task using the memory allocator.
- *
- * This macro submit an asynchronous task on the specified stream.
- * The task uses memory allocated by the allocator. It is guaranteed
- * that the memory will not be accessed by other tasks until this task
- * completes, due to the sequential execution order within the same stream.
- *
- * @param OP_NAME aclnn operator name.
- * @param args Additional arguments required by the task.
- *
- * @note
- * Memory from the allocator will be "freed" immediately and can be
- * reallocated to other pointers. However, it won't be accessed by any
- * other task before this asynchronous task ends, because all tasks in the
- * same stream are executed in queue order.
- */
-#define GGML_CANN_CALL_ACLNN_OP(OP_NAME, ...)                                                \
-    do {                                                                                     \
-        uint64_t        workspaceSize = 0;                                                   \
-        aclOpExecutor * executor;                                                            \
-        void *          workspaceAddr = nullptr;                                             \
-                                                                                             \
-        ACL_CHECK(aclnn##OP_NAME##GetWorkspaceSize(__VA_ARGS__, &workspaceSize, &executor)); \
-                                                                                             \
-        if (workspaceSize > 0) {                                                             \
-            ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);             \
-            workspaceAddr = workspace_allocator.get();                                       \
-        }                                                                                    \
-        ACL_CHECK(aclnn##OP_NAME(workspaceAddr, workspaceSize, executor, ctx.stream()));     \
-    } while (0)
 
 /**
  * @brief Applies a unary operation to an input tensor using the CANN backend.
@@ -799,9 +1031,7 @@ template <void unary_op(ggml_backend_cann_context&, aclTensor*, aclTensor*)>
     aclTensor* acl_dst = ggml_cann_create_tensor(dst);
 
     unary_op(ctx, acl_src, acl_dst);
-
-    ACL_CHECK(aclDestroyTensor(acl_src));
-    ACL_CHECK(aclDestroyTensor(acl_dst));
+    ggml_cann_release_resources(ctx, acl_src, acl_dst);
 }
 
 /**
@@ -832,7 +1062,7 @@ void ggml_cann_unary_op(
  *
  * Internally, the lambda will call:
  * @code
- * GGML_CANN_CALL_ACLNN_OP(OP_NAME, acl_src, acl_dst);
+ * GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);
  * @endcode
  *
  * @param OP_NAME The name of the ACL unary operator to invoke via GGML_CANN_CALL_ACLNN_OP.
@@ -840,14 +1070,14 @@ void ggml_cann_unary_op(
  * @see ggml_cann_unary_op
  * @see GGML_CANN_CALL_ACLNN_OP
  */
-#define GGML_CANN_CALL_UNARY_OP(OP_NAME)                         \
-    do {                                                         \
-        auto lambda = [](ggml_backend_cann_context& ctx,         \
-            aclTensor* acl_src,                                  \
-            aclTensor* acl_dst) {                                \
-            GGML_CANN_CALL_ACLNN_OP(OP_NAME, acl_src, acl_dst);  \
-        };                                                       \
-        ggml_cann_unary_op(lambda, ctx, dst);                    \
-    }                                                            \
+#define GGML_CANN_CALL_UNARY_OP(OP_NAME)                              \
+    do {                                                              \
+        auto lambda = [](ggml_backend_cann_context& ctx,              \
+            aclTensor* acl_src,                                       \
+            aclTensor* acl_dst) {                                     \
+            GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);  \
+        };                                                            \
+        ggml_cann_unary_op(lambda, ctx, dst);                         \
+    }                                                                 \
     while (0)
 #endif  // CANN_ACLNN_OPS

ggml/src/ggml-cann/common.h

@@ -31,9 +31,16 @@
 #include <memory>
 #include <string>
 #include <vector>
+#include <atomic>
+#include <condition_variable>
+#include <mutex>
+#include <thread>
+#include <unistd.h>
+#include <functional>
 
 #include "../include/ggml-cann.h"
 #include "../include/ggml.h"
+#include "../ggml-impl.h"
 
 #define MATRIX_ROW_PADDING 512
 #define GGML_CANN_MAX_STREAMS 8
@@ -205,6 +212,127 @@ struct ggml_cann_pool_alloc {
     ggml_cann_pool_alloc& operator=(ggml_cann_pool_alloc&&) = delete;
 };
 
+/**
+ * @brief Function pointer type for ACLNN operator calls.
+ */
+using aclnn_func_t = aclnnStatus (*)(void*, uint64_t, aclOpExecutor*, aclrtStream);
+
+/**
+ * @brief Base class for all CANN tasks to be submitted to the task queue.
+ *
+ * Users should override the run_task() method with actual task logic.
+ */
+class cann_task {
+public:
+    virtual void run_task() {}
+};
+
+/**
+ * @brief A lock-free ring-buffer based task queue for asynchronously executing cann_task instances.
+ */
+class cann_task_queue {
+public:
+    /**
+     * @brief Constructs a task queue with a fixed power-of-two capacity for a specific device.
+     *
+     * @param capacity Queue capacity. Must be a power of 2.
+     * @param device Target device ID (used for context setting).
+     */
+    explicit cann_task_queue(size_t capacity, int32_t device)
+        : buffer_(capacity), capacity_(capacity), head_(0), tail_(0),
+          running_(false), device_(device) {
+        GGML_ASSERT((capacity & (capacity - 1)) == 0 && "capacity must be power of 2");
+        mask_ = capacity_ - 1;
+    }
+
+    /**
+     * @brief Attempts to enqueue a task into the queue.
+     *
+     * @param item Unique pointer to the task.
+     * @return true if the task was successfully enqueued, false if the queue was full.
+     */
+    bool enqueue(std::unique_ptr<cann_task>&& item) {
+        size_t next_tail = (tail_ + 1) & mask_;
+
+        if (next_tail == head_) {
+            return false;
+        }
+
+        buffer_[tail_] = std::move(item);
+        std::atomic_thread_fence(std::memory_order_release);
+        tail_ = next_tail;
+
+        return true;
+    }
+
+    /**
+     * @brief Submits a task to the queue, and starts the worker thread if not already running.
+     *
+     * @param task Task to be submitted.
+     */
+    void submit_task(std::unique_ptr<cann_task>&& task) {
+        while(!enqueue(std::move(task))) {
+            std::this_thread::yield();
+            continue;
+        }
+
+        if (!running_) {
+            running_ = true;
+            thread_ = std::thread(&cann_task_queue::execute, this);
+        }
+
+    }
+
+    /**
+     * @brief Waits until the queue is completely empty and no tasks are being processed.
+     */
+    void wait() {
+        while (running_ && head_ != tail_) {
+            std::this_thread::yield();
+            continue;
+        }
+    }
+
+    /**
+     * @brief Stops the task queue and joins the worker thread.
+     */
+    void stop() {
+        running_ = false;
+        if (thread_.joinable()) {
+            thread_.join();
+        }
+    }
+
+private:
+    /**
+     * @brief Worker thread function that continuously dequeues and executes tasks.
+     */
+    void execute() {
+        ggml_cann_set_device(device_);
+
+        while (running_) {
+            if(head_ == tail_) {
+                std::this_thread::yield();
+                continue;
+            }
+
+            std::atomic_thread_fence(std::memory_order_acquire);
+            buffer_[head_]->run_task();
+            buffer_[head_].reset();
+            head_ = (head_ + 1) & mask_;
+        }
+    }
+
+    std::vector<std::unique_ptr<cann_task>> buffer_;
+    const size_t capacity_;
+    size_t mask_;
+    size_t head_;
+    size_t tail_;
+    bool running_;
+    std::thread thread_;
+    int32_t device_;
+};
+
 /**
  * @brief Context for managing CANN backend operations.
  */
@@ -213,6 +341,8 @@ struct ggml_backend_cann_context {
     std::string name;                /**< Name of the device. */
     std::string description;         /**< Description of the device. */
     aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
+    cann_task_queue task_queue;
+    bool async_mode;
 
     aclrtStream streams[GGML_CANN_MAX_STREAMS] = {nullptr}; /**< Array of streams for the device. */
 
@@ -221,9 +351,12 @@ struct ggml_backend_cann_context {
      * @param device Device ID.
      */
     explicit ggml_backend_cann_context(int device)
-        : device(device), name("CANN" + std::to_string(device)) {
+        : device(device), name("CANN" + std::to_string(device)), task_queue(1024, device) {
         ggml_cann_set_device(device);
         description = aclrtGetSocName();
+        async_mode = (getenv("GGML_CANN_ASYNC_MODE") != nullptr);
+        GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__,
+            device, async_mode ? "ON" : "OFF");
     }
 
     /**
@@ -231,6 +364,7 @@ struct ggml_backend_cann_context {
      */
     ~ggml_backend_cann_context() {
         ggml_cann_set_device(device);
+        task_queue.stop();
         if (copy_event != nullptr) {
             ACL_CHECK(aclrtDestroyEvent(copy_event));
         }

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

@@ -29,6 +29,8 @@
 #include <cstdio>
 #include <cstring>
 #include <mutex>
+#include <queue>
+#include <chrono>
 
 #include "ggml-impl.h"
 #include "ggml-backend-impl.h"
@@ -119,9 +121,10 @@ static ggml_cann_device_info ggml_cann_init() {
         prop.location.type = ACL_MEM_LOCATION_TYPE_DEVICE;
         prop.location.id = id;
         prop.reserve = 0;
-        ACL_CHECK(aclrtMemGetAllocationGranularity(
+        err = aclrtMemGetAllocationGranularity(
             &prop, ACL_RT_MEM_ALLOC_GRANULARITY_RECOMMENDED,
-            &info.devices[id].vmm_granularity));
+            &info.devices[id].vmm_granularity);
+        info.devices[id].vmm = err == ACL_SUCCESS;
 
         size_t free, total;
         ggml_backend_cann_get_device_memory(id, &free, &total);
@@ -148,11 +151,223 @@ const ggml_cann_device_info& ggml_cann_info() {
 
 //#define DEBUG_CANN_MALLOC
 /**
- * @brief A pool of CANN buffers(legacy).
+ * @brief A pool of CANN buffers(priority segment buffer).
  *
  * This class manages a pool of CANN buffers for a specific device.
  */
-struct ggml_cann_pool_leg : public ggml_cann_pool {
+struct ggml_cann_pool_buf_prio : public ggml_cann_pool {
+    /**
+     * @brief The maximum reuse margin for a buffer.
+     */
+    static const size_t max_reuse_margin = 1ull << 22;  // 4MB
+
+    /**
+     * @brief The minimum free margin for a buffer.
+     */
+    static const size_t min_free_margin = 1ull << 20;   // 1MB
+
+    /**
+     * @brief The alignment for buffer allocation.
+     */
+    static const size_t alignment = 128;
+
+    /**
+     * @brief The device ID associated with this buffer pool.
+     */
+    int device;
+
+    /**
+     * @brief Whether to disable clean during buffer allocation.
+     */
+    bool disable_clean = false;
+
+    /**
+     * @brief Structure representing a CANN buffer.
+     */
+    struct ggml_cann_buffer {
+        void* ptr = nullptr;  ///< Pointer to the buffer.
+        size_t size = 0;      ///< Size of the buffer.
+        std::chrono::steady_clock::time_point last_used;  ///< Last used time.
+
+        bool operator>(const ggml_cann_buffer& other) const {
+            return size > other.size;
+        }
+    };
+
+    /**
+     * @brief Array of CANN buffers in the pool.
+     */
+    std::unordered_map<void*, size_t> buffer_pool;
+    std::priority_queue<ggml_cann_buffer,
+                        std::vector<ggml_cann_buffer>,
+                        std::greater<>> free_buffers ;
+
+    /**
+     * @brief Total size of all buffers in the pool.
+     */
+    size_t pool_size = 0;
+
+    /**
+     * @brief Constructor to initialize the buffer pool for a specific device.
+     *
+     * @param device The device ID to associate with this buffer pool.
+     */
+    explicit ggml_cann_pool_buf_prio(int device) : device(device) {
+        disable_clean = getenv("GGML_CANN_DISABLE_BUF_POOL_CLEAN") != nullptr;
+    }
+
+    /**
+     * @brief Destructor to free all buffers in the pool.
+     */
+    ~ggml_cann_pool_buf_prio() {
+        ggml_cann_set_device(device);
+        for (auto& [b_ptr, b_size] : buffer_pool) {
+            aclrtFree(b_ptr);
+            pool_size -= b_size;
+        }
+        buffer_pool.clear();
+        GGML_ASSERT(pool_size == 0);
+    }
+
+    /**
+     * @brief Allocate a buffer of the given size.
+     *
+     * @param size The size of the buffer to allocate.
+     * @param actual_size A pointer to a variable to receive the actual size of
+     * the allocated buffer.
+     * @return A pointer to the allocated buffer.
+     */
+    void* alloc(size_t size, size_t* actual_size) override {
+        size = GGML_PAD(size, alignment);
+        if (size == 0) {
+            size = alignment;
+        }
+
+        void* ptr = nullptr;
+        auto now = std::chrono::steady_clock::now();
+
+        std::vector<ggml_cann_buffer> free_buffers_rest;
+        free_buffers_rest.reserve(free_buffers.size());
+        while (!free_buffers.empty()) {
+            auto b = free_buffers.top();
+            free_buffers.pop();
+
+            if (b.size >= size) {
+                // reuse the buffer if the size is enough
+                const size_t margin = b.size - size;
+                if (margin <= max_reuse_margin) {
+                    *actual_size = b.size;
+                    ptr = b.ptr;
+#ifdef DEBUG_CANN_MALLOC
+                    GGML_LOG_INFO(
+                        "cann pool[%d]: reused   %p, "
+                        "pool_size = %5u MB, "
+                        "size = %5u MB, "
+                        "margin = %5u MB\n",
+                        device, b.ptr,
+                        (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+                        (uint32_t)(GGML_PAD(size, 1048576) / 1048576),
+                        (uint32_t)(GGML_PAD(margin, 1048576) / 1048576));
+#endif
+                    break;
+                }
+            }
+
+            bool should_clean = !disable_clean &&
+                                b.size > min_free_margin &&
+                                std::chrono::duration_cast<std::chrono::milliseconds>(now - b.last_used).count() > 100;
+            if (should_clean) {
+                // free the buffer if the size is needed to be freed
+                ACL_CHECK(aclrtFree(b.ptr));
+                pool_size -= b.size;
+                buffer_pool.erase(b.ptr);
+#ifdef DEBUG_CANN_MALLOC
+                GGML_LOG_INFO(
+                    "cann pool[%d]: clean    %p, "
+                    "pool_size = %5u MB, "
+                    "size = %5u MB\n",
+                    device, b.ptr,
+                    (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+                    (uint32_t)(GGML_PAD(b.size, 1048576) / 1048576));
+#endif
+                continue;
+            }
+            free_buffers_rest.push_back(b);
+        }
+        for (ggml_cann_buffer &b : free_buffers_rest) {
+            free_buffers.push(std::move(b));
+        }
+
+#ifdef DEBUG_CANN_MALLOC
+        GGML_LOG_INFO("cann pool[%d] free pool_size = %5u MB\n\n", device, (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576));
+#endif
+        if (ptr != nullptr) {
+            return ptr;
+        }
+
+        // allocate a new buffer if no buffer can be reused
+        ggml_cann_set_device(device);
+        ACL_CHECK(aclrtMalloc(&ptr, size, ACL_MEM_MALLOC_HUGE_FIRST));
+        *actual_size = size;
+        pool_size += size;
+#ifdef DEBUG_CANN_MALLOC
+        GGML_LOG_INFO(
+            "cann pool[%d]: allocate %p, "
+            "pool_size = %5u MB, "
+            "size = %5u MB\n",
+            device, ptr, (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+            (uint32_t)(GGML_PAD(size, 1048576) / 1048576));
+#endif
+        buffer_pool.emplace(ptr, size);
+        return ptr;
+    }
+
+    /**
+     * @brief Free a buffer and return it to the pool.
+     *
+     * @param ptr Pointer to the buffer to free.
+     * @param size Size of the buffer to free.
+     */
+    void free(void* ptr, size_t size) override {
+        GGML_UNUSED(size);
+        auto it = buffer_pool.find(ptr);
+        if (it == buffer_pool.end()) {
+            GGML_ABORT("cann pool[%d]: buffer %p not found in pool\n", device, ptr);
+        }
+
+        auto now = std::chrono::steady_clock::now();
+        free_buffers.emplace(ggml_cann_buffer{ptr, it->second, now});
+#ifdef DEBUG_CANN_MALLOC
+        GGML_LOG_INFO(
+            "cann pool[%d]: return   %p, "
+            "pool_size = %5u MB\n",
+            device, ptr,
+            (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576));
+#endif
+    }
+};
+
+/**
+ * @brief A pool of CANN buffers(segment buffer).
+ *
+ * This class manages a pool of CANN buffers for a specific device.
+ */
+struct ggml_cann_pool_buf : public ggml_cann_pool {
+    /**
+     * @brief The maximum reuse margin for a buffer.
+     */
+    static const size_t max_reuse_margin = 1ull << 22;  // 4MB
+
+    /**
+     * @brief The minimum free margin for a buffer.
+     */
+    static const size_t min_free_margin = 1ull << 20;   // 1MB
+
+    /**
+     * @brief The alignment for buffer allocation.
+     */
+    static const size_t alignment = 128;
+
     /**
      * @brief The maximum number of buffers in the pool.
      */
@@ -163,12 +378,19 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
      */
     int device;
 
+    /**
+     * @brief Whether to disable clean during buffer allocation.
+     */
+    bool disable_clean = false;
+
     /**
      * @brief Structure representing a CANN buffer.
      */
     struct ggml_cann_buffer {
         void* ptr = nullptr;  ///< Pointer to the buffer memory.
         size_t size = 0;      ///< Size of the buffer.
+        bool used = false;    ///< Whether the buffer is currently in use.
+        std::chrono::steady_clock::time_point last_used;  ///< Last used time.
     };
 
     /**
@@ -186,17 +408,19 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
      *
      * @param device The device ID to associate with this buffer pool.
      */
-    explicit ggml_cann_pool_leg(int device) : device(device) {}
+    explicit ggml_cann_pool_buf(int device) : device(device) {
+        disable_clean = getenv("GGML_CANN_DISABLE_BUF_POOL_CLEAN") != nullptr;
+    }
 
     /**
      * @brief Destructor to free all buffers in the pool.
      */
-    ~ggml_cann_pool_leg() {
+    ~ggml_cann_pool_buf() {
         ggml_cann_set_device(device);
         for (int i = 0; i < MAX_BUFFERS; ++i) {
             ggml_cann_buffer& b = buffer_pool[i];
             if (b.ptr != nullptr) {
-                ACL_CHECK(aclrtFree(b.ptr));
+                aclrtFree(b.ptr);
                 pool_size -= b.size;
             }
         }
@@ -212,63 +436,93 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
      * @return A pointer to the allocated buffer.
      */
     void* alloc(size_t size, size_t* actual_size) override {
-        const size_t alignment = 128;
         size = GGML_PAD(size, alignment);
         if (size == 0) {
             size = alignment;
         }
-#ifdef DEBUG_CANN_MALLOC
-        int nnz = 0;
-        size_t max_size = 0;
-#endif
-        size_t best_diff = 1ull << 36;
-        int ibest = -1;
-        for (int i = 0; i < MAX_BUFFERS; ++i) {
+
+        void* ptr = nullptr;
+        auto now = std::chrono::steady_clock::now();
+
+        int i = 0;
+        for (; i < MAX_BUFFERS; ++i) {
             ggml_cann_buffer& b = buffer_pool[i];
-            if (b.ptr != nullptr) {
+            if (b.ptr == nullptr) {
+                break;
+            }
+            if (b.used) {
+                continue;
+            }
+            if (b.size >= size) {
+                // reuse the buffer if the size is enough
+                const size_t margin = b.size - size;
+                if (margin <= max_reuse_margin) {
+                    *actual_size = b.size;
+                    b.used = true;
+                    ptr = b.ptr;
 #ifdef DEBUG_CANN_MALLOC
-                ++nnz;
-                if (b.size > max_size) max_size = b.size;
+                    GGML_LOG_INFO(
+                        "cann pool[%d]: reused   %p, "
+                        "pool_size = %5u MB, "
+                        "size = %5u MB, "
+                        "margin = %5u MB\n",
+                        device, b.ptr,
+                        (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+                        (uint32_t)(GGML_PAD(size, 1048576) / 1048576),
+                        (uint32_t)(GGML_PAD(margin, 1048576) / 1048576));
 #endif
-                if (b.size >= size) {
-                    size_t diff = b.size - size;
-                    if (diff < best_diff) {
-                        best_diff = diff;
-                        ibest = i;
-                        if (!best_diff) {
-                            void* ptr = b.ptr;
-                            *actual_size = b.size;
-                            b.ptr = nullptr;
-                            b.size = 0;
-                            return ptr;
-                        }
-                    }
+                    break;
                 }
             }
+
+            bool should_clean = !disable_clean &&
+                                b.size > min_free_margin &&
+                                std::chrono::duration_cast<std::chrono::milliseconds>(now - b.last_used).count() > 100;
+            if (should_clean) {
+                // free the buffer if the size is needed to be freed
+                ACL_CHECK(aclrtFree(b.ptr));
+                pool_size -= b.size;
+#ifdef DEBUG_CANN_MALLOC
+                GGML_LOG_INFO(
+                    "cann pool[%d]: clean    %p, "
+                    "pool_size = %5u MB, "
+                    "size = %5u MB\n",
+                    device, b.ptr,
+                    (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+                    (uint32_t)(GGML_PAD(b.size, 1048576) / 1048576));
+#endif
+                b.ptr = nullptr;
+            }
         }
-        if (ibest >= 0) {
-            ggml_cann_buffer& b = buffer_pool[ibest];
-            void* ptr = b.ptr;
-            *actual_size = b.size;
-            b.ptr = nullptr;
-            b.size = 0;
+        if (ptr != nullptr) {
             return ptr;
         }
-        void* ptr;
-        ggml_cann_set_device(device);
-        ACL_CHECK(
-            aclrtMalloc(&ptr, size, ACL_MEM_MALLOC_HUGE_FIRST));
-        *actual_size = size;
-        pool_size += size;
+
+        if (i < MAX_BUFFERS) {
+            // allocate a new buffer if no buffer can be reused
+            ggml_cann_buffer& b = buffer_pool[i];
+            ggml_cann_set_device(device);
+            ACL_CHECK(aclrtMalloc(&b.ptr, size, ACL_MEM_MALLOC_HUGE_FIRST));
+            pool_size += size;
+            *actual_size = size;
+            b.size = size;
+            b.used = true;
+            if (i >= MAX_BUFFERS - 8) {
+                GGML_LOG_WARN("cann pool[%d]: slots almost full\n", device);
+            }
 #ifdef DEBUG_CANN_MALLOC
-        GGML_LOG_INFO(
-            "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, "
-            "requested %u MB\n",
-            __func__, device, nnz, (uint32_t)(max_size / 1024 / 1024),
-            (uint32_t)(pool_size / 1024 / 1024),
-            (uint32_t)(size / 1024 / 1024));
+            GGML_LOG_INFO(
+                "cann pool[%d]: allocate %p, "
+                "pool_size = %5u MB, "
+                "size = %5u MB\n",
+                device, b.ptr,
+                (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+                (uint32_t)(GGML_PAD(b.size, 1048576) / 1048576));
 #endif
-        return ptr;
+            return b.ptr;
+        }
+
+        GGML_ABORT("cann pool[%d]: slots full\n", device);
     }
 
     /**
@@ -278,18 +532,24 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
      * @param size Size of the buffer to free.
      */
     void free(void* ptr, size_t size) override {
+        GGML_UNUSED(size);
         for (int i = 0; i < MAX_BUFFERS; ++i) {
             ggml_cann_buffer& b = buffer_pool[i];
-            if (b.ptr == nullptr) {
-                b.ptr = ptr;
-                b.size = size;
-                return;
+            if (b.ptr != ptr) {
+                continue;
             }
+            b.used = false;
+            b.last_used = std::chrono::steady_clock::now();
+#ifdef DEBUG_CANN_MALLOC
+            GGML_LOG_INFO(
+                "cann pool[%d]: return   %p, "
+                "pool_size = %5u MB\n",
+                device, b.ptr,
+                (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576));
+#endif
+            return;
         }
-        // memory should always buffered. these memory may still needed by
-        // tasks in stream.
-        // TODO, fix me.
-        GGML_ABORT("Cann buffer pool full, increase MAX_CANN_BUFFERS\n");
+        GGML_ABORT("cann pool[%d]: slots full\n", device);
     }
 };
 
@@ -347,8 +607,7 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
      * @param device The device ID to associate with this buffer pool.
      */
     explicit ggml_cann_pool_vmm(int device)
-        : device(device),
-          granularity(ggml_cann_info().devices[device].vmm_granularity) {
+    : device(device) {
         auto dev = ggml_cann_info().devices[device];
         granularity = dev.vmm_granularity;
         max_size = dev.total_vram;
@@ -471,7 +730,18 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
  */
 std::unique_ptr<ggml_cann_pool> ggml_backend_cann_context::new_pool_for_device(
     int device) {
-    return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_vmm(device));
+    bool disable_vmm = (getenv("GGML_CANN_DISABLE_VMM_POOL") != nullptr);
+    if (!disable_vmm && ggml_cann_info().devices[device].vmm) {
+        GGML_LOG_INFO("%s: device %d use vmm pool\n", __func__, device);
+        return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_vmm(device));
+    }
+    bool enable_buf_prio = (getenv("GGML_CANN_ENABLE_BUF_PRIO_POOL") != nullptr);
+    if (enable_buf_prio) {
+        GGML_LOG_INFO("%s: device %d use buffer pool with priority queue\n", __func__, device);
+        return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_buf_prio(device));
+    }
+    GGML_LOG_INFO("%s: device %d use buffer pool\n", __func__, device);
+    return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_buf(device));
 }
 
 // cann buffer
@@ -1020,8 +1290,11 @@ ggml_backend_cann_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
 
     ggml_cann_set_device(buft_ctx->device);
 
-    size = std::max(size, (size_t)1);
-
+    const size_t alignment = 128;
+    size = GGML_PAD(size, alignment);
+    if (size == 0) {
+        size = alignment;
+    }
     void* dev_ptr;
     aclError err = aclrtMalloc(&dev_ptr, size, ACL_MEM_MALLOC_HUGE_FIRST);
     if (err != ACL_SUCCESS) {
@@ -1333,7 +1606,7 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
                     auto lambda = [](ggml_backend_cann_context& ctx,
                         aclTensor* acl_src,
                         aclTensor* acl_dst) {
-                        GGML_CANN_CALL_ACLNN_OP(GeluV2, acl_src, 0, acl_dst);
+                        GGML_CANN_CALL_ACLNN_OP(ctx, GeluV2, acl_src, 0, acl_dst);
                     };
                     ggml_cann_unary_op(lambda, ctx, dst);
                 } break;
@@ -1516,12 +1789,11 @@ static void ggml_backend_cann_free(ggml_backend_t backend) {
     delete backend;
 }
 
+
 /**
  * @brief Sets tensor data asynchronously in the CANN backend.
  *
- * This function asynchronously sets tensor data in the CANN backend. Depending
- * on the tensor type, it may perform data transformations before copying data
- * to the device.
+ * This function asynchronously sets tensor data in the CANN backend.
  *
  * @param backend Pointer to the CANN backend structure.
  * @param tensor Pointer to the tensor structure to set data for.
@@ -1536,23 +1808,28 @@ static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend,
                                                size_t size) {
     ggml_backend_cann_context *cann_ctx =
         (ggml_backend_cann_context *)backend->context;
+    ggml_backend_buffer_t buf =
+        tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
-    if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpyAsync((char *)tensor->data + offset, size, data,
-                                   size, ACL_MEMCPY_HOST_TO_DEVICE,
-                                   cann_ctx->stream()));
-    } else {
-        void *transform_buffer = malloc(size);
-        ggml_backend_cann_transform(tensor, data, transform_buffer);
+    GGML_ASSERT(buf->buft == ggml_backend_cann_buffer_type(cann_ctx->device) &&
+        "unsupported buffer type");
+    GGML_ASSERT(!ggml_is_quantized(tensor->type));
 
-        ACL_CHECK(aclrtMemcpyAsync(
-            (char *)tensor->data + offset, size, transform_buffer, size,
-            ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
-        ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
-        free(transform_buffer);
-    }
+    ggml_cann_async_memcpy(cann_ctx, (char *)tensor->data + offset, data, size,
+        ACL_MEMCPY_HOST_TO_DEVICE);
 }
 
+/**
+ * @brief Gets tensor data asynchronously in the CANN backend.
+ *
+ * This function asynchronously gets tensor data in the CANN backend.
+ *
+ * @param backend Pointer to the CANN backend structure.
+ * @param tensor Pointer to the tensor structure to get data from.
+ * @param data Pointer to the host data to copy from the tensor.
+ * @param offset Offset in bytes within the host data.
+ * @param size Size of the data to copy in bytes.
+ */
 static void ggml_backend_cann_get_tensor_async(
     ggml_backend_t backend, const ggml_tensor *tensor, void *data,
     size_t offset, size_t size) {
@@ -1563,20 +1840,11 @@ static void ggml_backend_cann_get_tensor_async(
 
     GGML_ASSERT(buf->buft == ggml_backend_cann_buffer_type(cann_ctx->device) &&
                 "unsupported buffer type");
+    GGML_ASSERT(!ggml_is_quantized(tensor->type));
+
+    ggml_cann_async_memcpy(cann_ctx, data, (char *)tensor->data + offset, size,
+        ACL_MEMCPY_DEVICE_TO_HOST);
 
-    if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpyAsync(data, size, (char *)tensor->data + offset,
-                                   size, ACL_MEMCPY_DEVICE_TO_HOST,
-                                   cann_ctx->stream()));
-    } else {
-        void *transform_buffer = malloc(size);
-        ACL_CHECK(aclrtMemcpyAsync(
-            transform_buffer, size, (char *)tensor->data + offset, size,
-            ACL_MEMCPY_DEVICE_TO_HOST, cann_ctx->stream()));
-        ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
-        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
-        free(transform_buffer);
-    }
 }
 
 /**
@@ -1636,6 +1904,8 @@ static bool ggml_backend_cann_cpy_tensor_async(
         ggml_cann_set_device(cann_ctx_src->device);
         ACL_CHECK(aclrtDeviceEnablePeerAccess(cann_ctx_dst->device, 0));
 
+        // wait for task_queue empty to keep task order.
+        cann_ctx_src->task_queue.wait();
         ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size,
                                    ACL_MEMCPY_DEVICE_TO_DEVICE,
                                    cann_ctx_src->stream()));
@@ -1663,9 +1933,8 @@ static bool ggml_backend_cann_cpy_tensor_async(
 static void ggml_backend_cann_synchronize(ggml_backend_t backend) {
     ggml_backend_cann_context* cann_ctx =
         (ggml_backend_cann_context*)backend->context;
-
+    cann_ctx->task_queue.wait();
     ggml_cann_set_device(cann_ctx->device);
-
     ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
 }
 
@@ -1749,6 +2018,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                     return true;
                 case GGML_TYPE_Q8_0:
                 case GGML_TYPE_Q4_0:
+#ifdef ASCEND_310P
+                    // Q4 && Q8 per group is not suppor on 310p device
+                    return false;
+#endif
                     // only support contiguous for quantized types.
                     return ggml_is_contiguous(op->src[0]) &&
                             ggml_is_contiguous(op->src[1]);
@@ -1816,6 +2089,9 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 return false;
             }
 
+            if(!ggml_is_contiguous(op->src[0])){
+                return false;
+            }
             return true;
         }
         case GGML_OP_UPSCALE: {
@@ -1831,6 +2107,12 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         }
         case GGML_OP_POOL_2D: {
             const int32_t * opts = (const int32_t *) op->op_params;
+#ifdef ASCEND_310P
+            enum ggml_op_pool opt = static_cast<ggml_op_pool>(opts[0]);
+            if(opt == GGML_OP_POOL_MAX){
+                return false;
+            }
+#endif
             const int       k0   = opts[1];
             const int       k1   = opts[2];
             const int       p0   = opts[5];

ggml/src/ggml-cpu/CMakeLists.txt

@@ -222,7 +222,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             elseif (GGML_AVX)
                 list(APPEND ARCH_FLAGS /arch:AVX)
                 list(APPEND ARCH_DEFINITIONS GGML_AVX)
-            else ()
+            elseif (GGML_SSE42)
                 list(APPEND ARCH_FLAGS /arch:SSE4.2)
                 list(APPEND ARCH_DEFINITIONS GGML_SSE42)
             endif()
@@ -237,8 +237,10 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             if (GGML_NATIVE)
                 list(APPEND ARCH_FLAGS -march=native)
             else ()
-                list(APPEND ARCH_FLAGS -msse4.2)
-                list(APPEND ARCH_DEFINITIONS GGML_SSE42)
+                if (GGML_SSE42)
+                    list(APPEND ARCH_FLAGS -msse4.2)
+                    list(APPEND ARCH_DEFINITIONS GGML_SSE42)
+                endif()
                 if (GGML_F16C)
                     list(APPEND ARCH_FLAGS -mf16c)
                     list(APPEND ARCH_DEFINITIONS GGML_F16C)

ggml/src/ggml-cpu/cpu-feats-x86.cpp

@@ -263,7 +263,7 @@ void test_x86_is() {
 static int ggml_backend_cpu_x86_score() {
     // FIXME: this does not check for OS support
 
-    int score = 0;
+    int score = 1;
     cpuid_x86 is;
 
 #ifdef GGML_FMA

ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp

@@ -4044,7 +4044,7 @@ static void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, c
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-#if defined(__AVX2__)
+#if defined(__AVX2__) || defined(__AVX512F__)
     const block_q4_Kx8 * b_ptr_start = (const block_q4_Kx8 * ) vx;
     const block_q8_Kx4 * a_ptr_start = (const block_q8_Kx4 * ) vy;
     int64_t b_nb = n / QK_K;
@@ -4054,8 +4054,748 @@ static void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, c
     const __m256i m4b = _mm256_set1_epi8(0x0F);
     // Permute mask used for easier vector processing at later stages
     __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4);
-
+    int64_t xstart = 0;
     int anr = nr - nr % 16;; // Used to align nr with boundary of 16
+#ifdef __AVX512F__
+    int anc = nc - nc % 16; // Used to align nc with boundary of 16
+    // Mask to mask out nibbles from packed bytes expanded to 512 bit length
+    const __m512i m4bexpanded = _mm512_set1_epi8(0x0F);
+    //Take group of four block_q8_Kx4 structures at each pass of the loop and perform dot product operation
+    for (; y < anr / 4; y += 4) {
+
+        const block_q8_Kx4 * a_ptrs[4];
+
+        a_ptrs[0] = a_ptr_start + (y * nb);
+        for (int i = 0; i < 3; ++i) {
+            a_ptrs[i + 1] = a_ptrs[i] + nb;
+        }
+
+        // Take group of eight block_q4_kx8 structures at each pass of the loop and perform dot product operation
+        for (int64_t x = 0; x < anc / 8; x += 2) {
+
+            const block_q4_Kx8 * b_ptr_0 = b_ptr_start + ((x) * b_nb);
+            const block_q4_Kx8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
+
+            // Master FP accumulators
+            __m512 acc_rows[16];
+            for (int i = 0; i < 16; i++) {
+                acc_rows[i] = _mm512_setzero_ps();
+            }
+
+            __m512 acc_min_rows[16];
+            for (int i = 0; i < 16; i++) {
+                acc_min_rows[i] = _mm512_setzero_ps();
+            }
+
+            // For super block
+            for (int64_t b = 0; b < nb; b++) {
+                // Scale values - Load the sixteen scale values from two block_q4_kx8 structures
+                const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
+
+                // dmin values - Load the sixteen dmin values from two block_q4_kx8 structures
+                const __m512 col_dmin_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].dmin, b_ptr_1[b].dmin);
+
+                // Loop to iterate over the eight sub blocks of a super block - two sub blocks are processed per iteration
+                for (int sb = 0; sb < QK_K / 64; sb++) {
+
+                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + sb * 256));
+                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 32 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 64 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 96 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 128 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 160 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 192 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 224 + sb * 256));
+
+                    const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 32 + sb * 256));
+                    const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 64 + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 96 + sb * 256));
+                    const __m256i rhs_raw_mat_89AB_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 128 + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 160 + sb * 256));
+                    const __m256i rhs_raw_mat_89AB_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 192 + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 224 + sb * 256));
+
+                    const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+                    const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+                    const __m256i rhs_raw_mat_0145_2 = _mm256_blend_epi32(rhs_raw_mat_0123_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_2, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_2, requiredOrder), rhs_raw_mat_4567_2, 240);
+                    const __m256i rhs_raw_mat_0145_3 = _mm256_blend_epi32(rhs_raw_mat_0123_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_3, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_3, requiredOrder), rhs_raw_mat_4567_3, 240);
+
+                    const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240);
+                    const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240);
+                    const __m256i rhs_raw_mat_89CD_2 = _mm256_blend_epi32(rhs_raw_mat_89AB_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_2, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_2, requiredOrder), rhs_raw_mat_CDEF_2, 240);
+                    const __m256i rhs_raw_mat_89CD_3 = _mm256_blend_epi32(rhs_raw_mat_89AB_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_3, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_3, requiredOrder), rhs_raw_mat_CDEF_3, 240);
+
+                    const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
+                    const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
+
+                    const __m512i rhs_raw_mat_014589CD_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_2), rhs_raw_mat_89CD_2, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_2), rhs_raw_mat_ABEF_2, 1);
+                    const __m512i rhs_raw_mat_014589CD_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_3), rhs_raw_mat_89CD_3, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_3), rhs_raw_mat_ABEF_3, 1);
+
+                    //4-bit -> 8-bit
+                    const __m512i rhs_mat_014589CD_00 = _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded); //B00(0-7) B01(0-7) B04(0-7) B05(0-7) B08(0-7) B09(0-7) B0C(0-7) B0D(0-7)
+                    const __m512i rhs_mat_2367ABEF_00 = _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded); //B02(0-7) B03(0-7) B06(0-7) B07(0-7) B0A(0-7) B0B(0-7) B0E(0-7) B0F(0-7)
+                    const __m512i rhs_mat_014589CD_01 = _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded); //B00(8-15) B01(8-15) B04(8-15) B05(8-15) B08(8-15) B09(8-15) B0C(8-15) B0D(8-15)
+                    const __m512i rhs_mat_2367ABEF_01 = _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded); //B02(8-15) B03(8-15) B06(8-15) B07(8-15) B0A(8-15) B0B(8-15) B0E(8-15) B0F(8-15)
+
+                    const __m512i rhs_mat_014589CD_02 = _mm512_and_si512(rhs_raw_mat_014589CD_2, m4bexpanded); //B00(16-23) B01(16-23) B04(16-23) B05(16-23) B08(16-23) B09(16-23) B0C(16-23) B0D(16-23)
+                    const __m512i rhs_mat_2367ABEF_02 = _mm512_and_si512(rhs_raw_mat_2367ABEF_2, m4bexpanded); //B02(16-23) B03(16-23) B06(16-23) B07(16-23) B0A(16-23) B0B(16-23) B0E(16-23) B0F(16-23)
+                    const __m512i rhs_mat_014589CD_03 = _mm512_and_si512(rhs_raw_mat_014589CD_3, m4bexpanded); //B00(24-31) B01(24-31) B04(24-31) B05(24-31) B08(24-31) B09(24-31) B0C(24-31) B0D(24-31)
+                    const __m512i rhs_mat_2367ABEF_03 = _mm512_and_si512(rhs_raw_mat_2367ABEF_3, m4bexpanded); //B02(24-31) B03(24-31) B06(24-31) B07(24-31) B0A(24-31) B0B(24-31) B0E(24-31) B0F(24-31)
+
+                    const __m512i rhs_mat_014589CD_10 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded); //B10(0-7) B11(0-7) B14(0-7) B15(0-7) B18(0-7) B19(0-7) B1C(0-7) B1D(0-7)
+                    const __m512i rhs_mat_2367ABEF_10 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded); //B12(0-7) B13(0-7) B16(0-7) B17(0-7) B1A(0-7) B1B(0-7) B1E(0-7) B1F(0-7)
+                    const __m512i rhs_mat_014589CD_11 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded); //B10(8-15) B11(8-15) B14(8-15) B15(8-15) B18(8-15) B19(8-15) B1C(8-15) B1D(8-15)
+                    const __m512i rhs_mat_2367ABEF_11 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded); //B12(8-15) B13(8-15) B16(8-15) B17(8-15) B1A(8-15) B1B(8-15) B1E(8-15) B1F(8-15)
+
+                    const __m512i rhs_mat_014589CD_12 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2, 4), m4bexpanded); //B10(16-23) B11(16-23) B14(16-23) B15(16-23) B18(16-23) B19(16-23) B1C(16-23) B1D(16-23)
+                    const __m512i rhs_mat_2367ABEF_12 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2, 4), m4bexpanded); //B12(16-23) B13(16-23) B16(16-23) B17(16-23) B1A(16-23) B1B(16-23) B1E(16-23) B1F(16-23)
+                    const __m512i rhs_mat_014589CD_13 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3, 4), m4bexpanded); //B10(24-31) B11(24-31) B14(24-31) B15(24-31) B18(24-31) B19(24-31) B1C(24-31) B1D(24-31)
+                    const __m512i rhs_mat_2367ABEF_13 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3, 4), m4bexpanded); //B12(24-31) B13(24-31) B16(24-31) B17(24-31) B1A(24-31) B1B(24-31) B1E(24-31) B1F(24-31)
+
+                    // Shuffle pattern one - right side input
+                    const __m512i rhs_mat_014589CD_00_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_00, (_MM_PERM_ENUM)136); //B00(0-3) B01(0-3) B00(0-3) B01(0-3) B04(0-3) B05(0-3) B04(0-3) B05(0-3) B08(0-3) B09(0-3) B08(0-3) B09(0-3) B0C(0-3) B0D(0-3) B0C(0-3) B0D(0-3)
+                    const __m512i rhs_mat_2367ABEF_00_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_00, (_MM_PERM_ENUM)136); //B02(0-3) B03(0-3) B02(0-3) B03(0-3) B06(0-3) B07(0-3) B06(0-3) B07(0-3) B0A(0-3) B0B(0-3) B0A(0-3) B0B(0-3) B0E(0-3) B0F(0-3) B0E(0-3) B0F(0-3)
+                    const __m512i rhs_mat_014589CD_01_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_01, (_MM_PERM_ENUM)136); //B00(8-11) B01(8-11) B00(8-11) B01(8-11) B04(8-11) B05(8-11) B04(8-11) B05(8-11) B08(8-11) B09(8-11) B08(8-11) B09(8-11) B0C(8-11) B0D(8-11) B0C(8-11) B0D(8-11)
+                    const __m512i rhs_mat_2367ABEF_01_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_01, (_MM_PERM_ENUM)136); //B02(8-11) B03(8-11) B02(8-11) B03(8-11) B06(8-11) B07(8-11) B06(8-11) B07(8-11) B0A(8-11) B0B(8-11) B0A(8-11) B0B(8-11) B0E(8-11) B0F(8-11) B0E(8-11) B0F(8-11)
+                    const __m512i rhs_mat_014589CD_02_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_02, (_MM_PERM_ENUM)136); //B00(16-19) B01(16-19) B00(16-19) B01(16-19) B04(16-19) B05(16-19) B04(16-19) B05(16-19) B08(16-19) B09(16-19) B08(16-19) B09(16-19) B0C(16-19) B0D(16-19) B0C(16-19) B0D(16-19)
+                    const __m512i rhs_mat_2367ABEF_02_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_02, (_MM_PERM_ENUM)136); //B02(16-19) B03(16-19) B02(16-19) B03(16-19) B06(16-19) B07(16-19) B06(16-19) B07(16-19) B0A(16-19) B0B(16-19) B0A(16-19) B0B(16-19) B0E(16-19) B0F(16-19) B0E(16-19) B0F(16-19)
+                    const __m512i rhs_mat_014589CD_03_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_03, (_MM_PERM_ENUM)136); //B00(24-27) B01(24-27) B00(24-27) B01(24-27) B04(24-27) B05(24-27) B04(24-27) B05(24-27) B08(24-27) B09(24-27) B08(24-27) B09(24-27) B0C(24-27) B0D(24-27) B0C(24-27) B0D(24-27)
+                    const __m512i rhs_mat_2367ABEF_03_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_03, (_MM_PERM_ENUM)136); //B02(24-27) B03(24-27) B02(24-27) B03(24-27) B06(24-27) B07(24-27) B06(24-27) B07(24-27) B0A(24-27) B0B(24-27) B0A(24-27) B0B(24-27) B0E(24-27) B0F(24-27) B0E(24-27) B0F(24-27)
+
+                    const __m512i rhs_mat_014589CD_10_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_10, (_MM_PERM_ENUM)136); //B10(0-3) B11(0-3) B10(0-3) B11(0-3) B14(0-3) B15(0-3) B14(0-3) B15(0-3) B18(0-3) B19(0-3) B18(0-3) B19(0-3) B1C(0-3) B1D(0-3) B1C(0-3) B1D(0-3)
+                    const __m512i rhs_mat_2367ABEF_10_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_10, (_MM_PERM_ENUM)136); //B12(0-3) B13(0-3) B12(0-3) B13(0-3) B16(0-3) B17(0-3) B16(0-3) B17(0-3) B1A(0-3) B1B(0-3) B1A(0-3) B1B(0-3) B1E(0-3) B1F(0-3) B1E(0-3) B1F(0-3)
+                    const __m512i rhs_mat_014589CD_11_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_11, (_MM_PERM_ENUM)136); //B10(8-11) B11(8-11) B10(8-11) B11(8-11) B14(8-11) B15(8-11) B14(8-11) B15(8-11) B18(8-11) B19(8-11) B18(8-11) B19(8-11) B1C(8-11) B1D(8-11) B1C(8-11) B1D(8-11)
+                    const __m512i rhs_mat_2367ABEF_11_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_11, (_MM_PERM_ENUM)136); //B12(8-11) B13(8-11) B12(8-11) B13(8-11) B16(8-11) B17(8-11) B16(8-11) B17(8-11) B1A(8-11) B1B(8-11) B1A(8-11) B1B(8-11) B1E(8-11) B1F(8-11) B1E(8-11) B1F(8-11)
+                    const __m512i rhs_mat_014589CD_12_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_12, (_MM_PERM_ENUM)136); //B10(16-19) B11(16-19) B10(16-19) B11(16-19) B14(16-19) B15(16-19) B14(16-19) B15(16-19) B18(16-19) B19(16-19) B18(16-19) B19(16-19) B1C(16-19) B1D(16-19) B1C(16-19) B1D(16-19)
+                    const __m512i rhs_mat_2367ABEF_12_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_12, (_MM_PERM_ENUM)136); //B12(16-19) B13(16-19) B12(16-19) B13(16-19) B16(16-19) B17(16-19) B16(16-19) B17(16-19) B1A(16-19) B1B(16-19) B1A(16-19) B1B(16-19) B1E(16-19) B1F(16-19) B1E(16-19) B1F(16-19)
+                    const __m512i rhs_mat_014589CD_13_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_13, (_MM_PERM_ENUM)136); //B10(24-27) B11(24-27) B10(24-27) B11(24-27) B14(24-27) B15(24-27) B14(24-27) B15(24-27) B18(24-27) B19(24-27) B18(24-27) B19(24-27) B1C(24-27) B1D(24-27) B1C(24-27) B1D(24-27)
+                    const __m512i rhs_mat_2367ABEF_13_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_13, (_MM_PERM_ENUM)136); //B12(24-27) B13(24-27) B12(24-27) B13(24-27) B16(24-27) B17(24-27) B16(24-27) B17(24-27) B1A(24-27) B1B(24-27) B1A(24-27) B1B(24-27) B1E(24-27) B1F(24-27) B1E(24-27) B1F(24-27)
+
+                    // Shuffle pattern two - right side input
+                    const __m512i rhs_mat_014589CD_00_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_00, (_MM_PERM_ENUM)221); //B00(4-7) B01(4-7) B00(4-7) B01(4-7) B04(4-7) B05(4-7) B04(4-7) B05(4-7) B08(4-7) B09(4-7) B08(4-7) B09(4-7) B0C(4-7) B0D(4-7) B0C(4-7) B0D(4-7)
+                    const __m512i rhs_mat_2367ABEF_00_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_00, (_MM_PERM_ENUM)221); //B02(4-7) B03(4-7) B02(4-7) B03(4-7) B06(4-7) B07(4-7) B06(4-7) B07(4-7) B0A(4-7) B0B(4-7) B0A(4-7) B0B(4-7) B0E(4-7) B0F(4-7) B0E(4-7) B0F(4-7)
+                    const __m512i rhs_mat_014589CD_01_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_01, (_MM_PERM_ENUM)221); //B00(12-15) B01(12-15) B00(12-15) B01(12-15) B04(12-15) B05(12-15) B04(12-15) B05(12-15) B08(12-15) B09(12-15) B08(12-15) B09(12-15) B0C(12-15) B0D(12-15) B0C(12-15) B0D(12-15)
+                    const __m512i rhs_mat_2367ABEF_01_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_01, (_MM_PERM_ENUM)221); //B02(12-15) B03(12-15) B02(12-15) B03(12-15) B06(12-15) B07(12-15) B06(12-15) B07(12-15) B0A(12-15) B0B(12-15) B0A(12-15) B0B(12-15) B0E(12-15) B0F(12-15) B0E(12-15) B0F(12-15)
+                    const __m512i rhs_mat_014589CD_02_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_02, (_MM_PERM_ENUM)221); //B00(20-23) B01(20-23) B00(20-23) B01(20-23) B04(20-23) B05(20-23) B04(20-23) B05(20-23) B08(20-23) B09(20-23) B08(20-23) B09(20-23) B0C(20-23) B0D(20-23) B0C(20-23) B0D(20-23)
+                    const __m512i rhs_mat_2367ABEF_02_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_02, (_MM_PERM_ENUM)221); //B02(20-23) B03(20-23) B02(20-23) B03(20-23) B06(20-23) B07(20-23) B06(20-23) B07(20-23) B0A(20-23) B0B(20-23) B0A(20-23) B0B(20-23) B0E(20-23) B0F(20-23) B0E(20-23) B0F(20-23)
+                    const __m512i rhs_mat_014589CD_03_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_03, (_MM_PERM_ENUM)221); //B00(28-31) B01(28-31) B00(28-31) B01(28-31) B04(28-31) B05(28-31) B04(28-31) B05(28-31) B08(28-31) B09(28-31) B08(28-31) B09(28-31) B0C(28-31) B0D(28-31) B0C(28-31) 0BD(28-31)
+                    const __m512i rhs_mat_2367ABEF_03_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_03, (_MM_PERM_ENUM)221); //B02(28-31) B03(28-31) B02(28-31) B03(28-31) B06(28-31) B07(28-31) B06(28-31) B07(28-31) B0A(28-31) B0B(28-31) B0A(28-31) B0B(28-31) B0E(28-31) B0F(28-31) B0E(28-31) B0F(28-31)
+
+                    const __m512i rhs_mat_014589CD_10_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_10, (_MM_PERM_ENUM)221); //B10(4-7) B11(4-7) B10(4-7) B11(4-7) B14(4-7) B15(4-7) B14(4-7) B15(4-7) B18(4-7) B19(4-7) B18(4-7) B19(4-7) B1C(4-7) B1D(4-7) B1C(4-7) B1D(4-7)
+                    const __m512i rhs_mat_2367ABEF_10_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_10, (_MM_PERM_ENUM)221); //B12(4-7) B13(4-7) B12(4-7) B13(4-7) B16(4-7) B17(4-7) B16(4-7) B17(4-7) B1A(4-7) B1B(4-7) B1A(4-7) B1B(4-7) B1E(4-7) B1F(4-7) B1E(4-7) B1F(4-7)
+                    const __m512i rhs_mat_014589CD_11_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_11, (_MM_PERM_ENUM)221); //B10(12-15) B11(12-15) B10(12-15) B11(12-15) B14(12-15) B15(12-15) B14(12-15) B15(12-15) B18(12-15) B19(12-15) B18(12-15) B19(12-15) B1C(12-15) B1D(12-15) B1C(12-15) B1D(12-15)
+                    const __m512i rhs_mat_2367ABEF_11_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_11, (_MM_PERM_ENUM)221); //B12(12-15) B13(12-15) B12(12-15) B13(12-15) B16(12-15) B17(12-15) B16(12-15) B17(12-15) B1A(12-15) B1B(12-15) B1A(12-15) B1B(12-15) B1E(12-15) B1F(12-15) B1E(12-15) B1F(12-15)
+                    const __m512i rhs_mat_014589CD_12_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_12, (_MM_PERM_ENUM)221); //B10(20-23) B11(20-23) B10(20-23) B11(20-23) B14(20-23) B15(20-23) B14(20-23) B15(20-23) B18(20-23) B19(20-23) B18(20-23) B19(20-23) B1C(20-23) B1D(20-23) B1C(20-23) B1D(20-23)
+                    const __m512i rhs_mat_2367ABEF_12_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_12, (_MM_PERM_ENUM)221); //B12(20-23) B13(20-23) B12(20-23) B13(20-23) B16(20-23) B17(20-23) B16(20-23) B17(20-23) B1A(20-23) B1B(20-23) B1A(20-23) B1B(20-23) B1E(20-23) B1F(20-23) B1E(20-23) B1F(20-23)
+                    const __m512i rhs_mat_014589CD_13_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_13, (_MM_PERM_ENUM)221); //B10(28-31) B11(28-31) B10(28-31) B11(28-31) B14(28-31) B15(28-31) B14(28-31) B15(28-31) B18(28-31) B19(28-31) B18(28-31) B19(28-31) B1C(28-31) B1D(28-31) B1C(28-31) B1D(28-31)
+                    const __m512i rhs_mat_2367ABEF_13_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_13, (_MM_PERM_ENUM)221); //B12(28-31) B13(28-31) B12(28-31) B13(28-31) B16(28-31) B17(28-31) B16(28-31) B17(28-31) B1A(28-31) B1B(28-31) B1A(28-31) B1B(28-31) B1E(28-31) B1F(28-31) B1E(28-31) B1F(28-31)
+
+                    uint32_t utmp_00[4], utmp_01[4], utmp_10[4], utmp_11[4];
+
+                    // Scales and Mins of corresponding sub blocks from different Q4_K structures are stored together
+                    // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop
+                    memcpy(utmp_00, b_ptr_0[b].scales + 24 * sb, 12);
+                    utmp_00[3] = ((utmp_00[2] >> 4) & kmask2) | (((utmp_00[1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_00 = utmp_00[1] & kmask1;
+                    utmp_00[1] = (utmp_00[2] & kmask2) | (((utmp_00[0] >> 6) & kmask3) << 4);
+                    utmp_00[2] = uaux_00;
+                    utmp_00[0] &= kmask1;
+
+                    // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop
+                    memcpy(utmp_01, b_ptr_0[b].scales + 12 + sb * 24, 12);
+                    utmp_01[3] = ((utmp_01[2] >> 4) & kmask2) | (((utmp_01[1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_01 = utmp_01[1] & kmask1;
+                    utmp_01[1] = (utmp_01[2] & kmask2) | (((utmp_01[0] >> 6) & kmask3) << 4);
+                    utmp_01[2] = uaux_01;
+                    utmp_01[0] &= kmask1;
+
+                    memcpy(utmp_10, b_ptr_1[b].scales + sb * 24, 12);
+                    utmp_10[3] = ((utmp_10[2] >> 4) & kmask2) | (((utmp_10[1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_10 = utmp_10[1] & kmask1;
+                    utmp_10[1] = (utmp_10[2] & kmask2) | (((utmp_10[0] >> 6) & kmask3) << 4);
+                    utmp_10[2] = uaux_10;
+                    utmp_10[0] &= kmask1;
+
+                    // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop
+                    memcpy(utmp_11, b_ptr_1[b].scales + 12 + sb * 24, 12);
+                    utmp_11[3] = ((utmp_11[2] >> 4) & kmask2) | (((utmp_11[1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_11 = utmp_11[1] & kmask1;
+                    utmp_11[1] = (utmp_11[2] & kmask2) | (((utmp_11[0] >> 6) & kmask3) << 4);
+                    utmp_11[2] = uaux_11;
+                    utmp_11[0] &= kmask1;
+
+                    // Scales of first sub block in the sb loop
+                    const __m256i mins_and_scales_0 = _mm256_set_epi32(utmp_10[3], utmp_10[2], utmp_10[1], utmp_10[0], utmp_00[3], utmp_00[2], utmp_00[1], utmp_00[0]);
+                    const __m512i scales_0 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(mins_and_scales_0, mins_and_scales_0));
+
+                    // Scales of second sub block in the sb loop
+                    const __m256i mins_and_scales_1 = _mm256_set_epi32(utmp_11[3], utmp_11[2], utmp_11[1], utmp_11[0], utmp_01[3], utmp_01[2], utmp_01[1], utmp_01[0]);
+                    const __m512i scales_1 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(mins_and_scales_1, mins_and_scales_1));
+
+                    // Mins of first and second sub block of Q4_K block are arranged side by side
+                    const __m512i mins_01 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(_mm256_shuffle_epi32(mins_and_scales_0, 78), _mm256_shuffle_epi32(mins_and_scales_1, 78)));
+
+                    const __m512i scale_014589CD_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)238);
+
+                    const __m512i scale_014589CD_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)238);
+
+                    for (int rp = 0; rp < 4; rp++) {
+
+                        // Load the four block_q8_k quantized values interleaved with each other in chunks of eight bytes - A0,A1,A2,A3
+                        // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector
+                        __m256i lhs_mat_ymm_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 256 * sb)));
+                        __m256i lhs_mat_ymm_01_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 0);
+                        __m256i lhs_mat_ymm_23_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 17);
+                        __m256i lhs_mat_ymm_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 32 + 256 * sb)));
+                        __m256i lhs_mat_ymm_01_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 0);
+                        __m256i lhs_mat_ymm_23_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 17);
+                        __m256i lhs_mat_ymm_0123_02 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 64 + 256 * sb)));
+                        __m256i lhs_mat_ymm_01_02 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_02, lhs_mat_ymm_0123_02, 0);
+                        __m256i lhs_mat_ymm_23_02 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_02, lhs_mat_ymm_0123_02, 17);
+                        __m256i lhs_mat_ymm_0123_03 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 96 + 256 * sb)));
+                        __m256i lhs_mat_ymm_01_03 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_03, lhs_mat_ymm_0123_03, 0);
+                        __m256i lhs_mat_ymm_23_03 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_03, lhs_mat_ymm_0123_03, 17);
+                        __m256i lhs_mat_ymm_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 128 + 256 * sb)));
+                        __m256i lhs_mat_ymm_01_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 0);
+                        __m256i lhs_mat_ymm_23_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 17);
+                        __m256i lhs_mat_ymm_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 160 + 256 * sb)));
+                        __m256i lhs_mat_ymm_01_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 0);
+                        __m256i lhs_mat_ymm_23_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 17);
+                        __m256i lhs_mat_ymm_0123_12 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 192 + 256 * sb)));
+                        __m256i lhs_mat_ymm_01_12 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_12, lhs_mat_ymm_0123_12, 0);
+                        __m256i lhs_mat_ymm_23_12 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_12, lhs_mat_ymm_0123_12, 17);
+                        __m256i lhs_mat_ymm_0123_13 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 224 + 256 * sb)));
+                        __m256i lhs_mat_ymm_01_13 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_13, lhs_mat_ymm_0123_13, 0);
+                        __m256i lhs_mat_ymm_23_13 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_13, lhs_mat_ymm_0123_13, 17);
+
+                        __m512i lhs_mat_01_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_00), lhs_mat_ymm_01_00, 1);
+                        __m512i lhs_mat_23_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_00), lhs_mat_ymm_23_00, 1);
+                        __m512i lhs_mat_01_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_01), lhs_mat_ymm_01_01, 1);
+                        __m512i lhs_mat_23_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_01), lhs_mat_ymm_23_01, 1);
+                        __m512i lhs_mat_01_02 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_02), lhs_mat_ymm_01_02, 1);
+                        __m512i lhs_mat_23_02 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_02), lhs_mat_ymm_23_02, 1);
+                        __m512i lhs_mat_01_03 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_03), lhs_mat_ymm_01_03, 1);
+                        __m512i lhs_mat_23_03 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_03), lhs_mat_ymm_23_03, 1);
+
+                        __m512i lhs_mat_01_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_10), lhs_mat_ymm_01_10, 1);
+                        __m512i lhs_mat_23_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_10), lhs_mat_ymm_23_10, 1);
+                        __m512i lhs_mat_01_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_11), lhs_mat_ymm_01_11, 1);
+                        __m512i lhs_mat_23_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_11), lhs_mat_ymm_23_11, 1);
+                        __m512i lhs_mat_01_12 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_12), lhs_mat_ymm_01_12, 1);
+                        __m512i lhs_mat_23_12 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_12), lhs_mat_ymm_23_12, 1);
+                        __m512i lhs_mat_01_13 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_13), lhs_mat_ymm_01_13, 1);
+                        __m512i lhs_mat_23_13 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_13), lhs_mat_ymm_23_13, 1);
+
+                        // Bsums are loaded - four bsums are loaded (for two sub blocks) for the different Q8_K blocks
+                        __m256i lhs_bsums_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].bsums + 16 * sb)));
+                        __m256i lhs_bsums_hsum_ymm_0123_01 = _mm256_castsi128_si256(_mm_hadd_epi16(_mm256_castsi256_si128(lhs_bsums_0123_01), _mm256_extractf128_si256(lhs_bsums_0123_01, 1)));
+                        lhs_bsums_hsum_ymm_0123_01 = _mm256_permute2x128_si256(lhs_bsums_hsum_ymm_0123_01, lhs_bsums_hsum_ymm_0123_01, 0);
+                        __m512i lhs_bsums_hsum_0123_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_hsum_ymm_0123_01), lhs_bsums_hsum_ymm_0123_01, 1);
+
+                        // Shuffle pattern one - left side input
+                        const __m512i lhs_mat_01_00_sp1 = _mm512_shuffle_epi32(lhs_mat_01_00, (_MM_PERM_ENUM)160); //A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3)
+                        const __m512i lhs_mat_23_00_sp1 = _mm512_shuffle_epi32(lhs_mat_23_00, (_MM_PERM_ENUM)160); //A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3)
+                        const __m512i lhs_mat_01_01_sp1 = _mm512_shuffle_epi32(lhs_mat_01_01, (_MM_PERM_ENUM)160); //A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11)
+                        const __m512i lhs_mat_23_01_sp1 = _mm512_shuffle_epi32(lhs_mat_23_01, (_MM_PERM_ENUM)160); //A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11)
+                        const __m512i lhs_mat_01_02_sp1 = _mm512_shuffle_epi32(lhs_mat_01_02, (_MM_PERM_ENUM)160); //A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19)
+                        const __m512i lhs_mat_23_02_sp1 = _mm512_shuffle_epi32(lhs_mat_23_02, (_MM_PERM_ENUM)160); //A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19)
+                        const __m512i lhs_mat_01_03_sp1 = _mm512_shuffle_epi32(lhs_mat_01_03, (_MM_PERM_ENUM)160); //A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27)
+                        const __m512i lhs_mat_23_03_sp1 = _mm512_shuffle_epi32(lhs_mat_23_03, (_MM_PERM_ENUM)160); //A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27)
+
+                        const __m512i lhs_mat_01_10_sp1 = _mm512_shuffle_epi32(lhs_mat_01_10, (_MM_PERM_ENUM)160); //A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3)
+                        const __m512i lhs_mat_23_10_sp1 = _mm512_shuffle_epi32(lhs_mat_23_10, (_MM_PERM_ENUM)160); //A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3)
+                        const __m512i lhs_mat_01_11_sp1 = _mm512_shuffle_epi32(lhs_mat_01_11, (_MM_PERM_ENUM)160); //A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11)
+                        const __m512i lhs_mat_23_11_sp1 = _mm512_shuffle_epi32(lhs_mat_23_11, (_MM_PERM_ENUM)160); //A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11)
+                        const __m512i lhs_mat_01_12_sp1 = _mm512_shuffle_epi32(lhs_mat_01_12, (_MM_PERM_ENUM)160); //A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19)
+                        const __m512i lhs_mat_23_12_sp1 = _mm512_shuffle_epi32(lhs_mat_23_12, (_MM_PERM_ENUM)160); //A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19)
+                        const __m512i lhs_mat_01_13_sp1 = _mm512_shuffle_epi32(lhs_mat_01_13, (_MM_PERM_ENUM)160); //A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27)
+                        const __m512i lhs_mat_23_13_sp1 = _mm512_shuffle_epi32(lhs_mat_23_13, (_MM_PERM_ENUM)160); //A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27)
+
+                        const __m512i lhs_mat_01_00_sp2 = _mm512_shuffle_epi32(lhs_mat_01_00, (_MM_PERM_ENUM)245); //A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7)
+                        const __m512i lhs_mat_23_00_sp2 = _mm512_shuffle_epi32(lhs_mat_23_00, (_MM_PERM_ENUM)245); //A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7)
+                        const __m512i lhs_mat_01_01_sp2 = _mm512_shuffle_epi32(lhs_mat_01_01, (_MM_PERM_ENUM)245); //A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15)
+                        const __m512i lhs_mat_23_01_sp2 = _mm512_shuffle_epi32(lhs_mat_23_01, (_MM_PERM_ENUM)245); //A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15)
+                        const __m512i lhs_mat_01_02_sp2 = _mm512_shuffle_epi32(lhs_mat_01_02, (_MM_PERM_ENUM)245); //A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23)
+                        const __m512i lhs_mat_23_02_sp2 = _mm512_shuffle_epi32(lhs_mat_23_02, (_MM_PERM_ENUM)245); //A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23)
+                        const __m512i lhs_mat_01_03_sp2 = _mm512_shuffle_epi32(lhs_mat_01_03, (_MM_PERM_ENUM)245); //A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31)
+                        const __m512i lhs_mat_23_03_sp2 = _mm512_shuffle_epi32(lhs_mat_23_03, (_MM_PERM_ENUM)245); //A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31)
+
+                        const __m512i lhs_mat_01_10_sp2 = _mm512_shuffle_epi32(lhs_mat_01_10, (_MM_PERM_ENUM)245); //A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7)
+                        const __m512i lhs_mat_23_10_sp2 = _mm512_shuffle_epi32(lhs_mat_23_10, (_MM_PERM_ENUM)245); //A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7)
+                        const __m512i lhs_mat_01_11_sp2 = _mm512_shuffle_epi32(lhs_mat_01_11, (_MM_PERM_ENUM)245); //A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15)
+                        const __m512i lhs_mat_23_11_sp2 = _mm512_shuffle_epi32(lhs_mat_23_11, (_MM_PERM_ENUM)245); //A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15)
+                        const __m512i lhs_mat_01_12_sp2 = _mm512_shuffle_epi32(lhs_mat_01_12, (_MM_PERM_ENUM)245); //A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23)
+                        const __m512i lhs_mat_23_12_sp2 = _mm512_shuffle_epi32(lhs_mat_23_12, (_MM_PERM_ENUM)245); //A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23)
+                        const __m512i lhs_mat_01_13_sp2 = _mm512_shuffle_epi32(lhs_mat_01_13, (_MM_PERM_ENUM)245); //A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31)
+                        const __m512i lhs_mat_23_13_sp2 = _mm512_shuffle_epi32(lhs_mat_23_13, (_MM_PERM_ENUM)245); //A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(28-31)
+
+                        // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+                        __m512i iacc_mat_00_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp1, lhs_mat_01_03_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp1, lhs_mat_01_02_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_01_01_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_01_00_sp1));
+                        __m512i iacc_mat_01_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp1, lhs_mat_01_03_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp1, lhs_mat_01_02_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_01_01_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_01_00_sp1));
+                        __m512i iacc_mat_10_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp1, lhs_mat_23_03_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp1, lhs_mat_23_02_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_23_01_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_23_00_sp1));
+                        __m512i iacc_mat_11_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp1, lhs_mat_23_03_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp1, lhs_mat_23_02_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_23_01_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_23_00_sp1));
+                        __m512i iacc_mat_00_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp1, lhs_mat_01_13_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp1, lhs_mat_01_12_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_01_11_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_01_10_sp1));
+                        __m512i iacc_mat_01_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp1, lhs_mat_01_13_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp1, lhs_mat_01_12_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_01_11_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_01_10_sp1));
+                        __m512i iacc_mat_10_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp1, lhs_mat_23_13_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp1, lhs_mat_23_12_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_23_11_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_23_10_sp1));
+                        __m512i iacc_mat_11_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp1, lhs_mat_23_13_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp1, lhs_mat_23_12_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_23_11_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_23_10_sp1));
+
+                        __m512i iacc_mat_00_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp2, lhs_mat_01_03_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp2, lhs_mat_01_02_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_01_01_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_01_00_sp2));
+                        __m512i iacc_mat_01_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp2, lhs_mat_01_03_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp2, lhs_mat_01_02_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_01_01_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_01_00_sp2));
+                        __m512i iacc_mat_10_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp2, lhs_mat_23_03_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp2, lhs_mat_23_02_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_23_01_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_23_00_sp2));
+                        __m512i iacc_mat_11_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp2, lhs_mat_23_03_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp2, lhs_mat_23_02_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_23_01_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_23_00_sp2));
+                        __m512i iacc_mat_00_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp2, lhs_mat_01_13_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp2, lhs_mat_01_12_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_01_11_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_01_10_sp2));
+                        __m512i iacc_mat_01_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp2, lhs_mat_01_13_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp2, lhs_mat_01_12_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_01_11_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_01_10_sp2));
+                        __m512i iacc_mat_10_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp2, lhs_mat_23_13_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp2, lhs_mat_23_12_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_23_11_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_23_10_sp2));
+                        __m512i iacc_mat_11_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp2, lhs_mat_23_13_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp2, lhs_mat_23_12_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_23_11_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_23_10_sp2));
+
+                        // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+                        __m512i iacc_mat_00_0 = _mm512_add_epi16(iacc_mat_00_0_sp1, iacc_mat_00_0_sp2);
+                        __m512i iacc_mat_01_0 = _mm512_add_epi16(iacc_mat_01_0_sp1, iacc_mat_01_0_sp2);
+                        __m512i iacc_mat_10_0 = _mm512_add_epi16(iacc_mat_10_0_sp1, iacc_mat_10_0_sp2);
+                        __m512i iacc_mat_11_0 = _mm512_add_epi16(iacc_mat_11_0_sp1, iacc_mat_11_0_sp2);
+
+                        __m512i iacc_mat_00_1 = _mm512_add_epi16(iacc_mat_00_1_sp1, iacc_mat_00_1_sp2);
+                        __m512i iacc_mat_01_1 = _mm512_add_epi16(iacc_mat_01_1_sp1, iacc_mat_01_1_sp2);
+                        __m512i iacc_mat_10_1 = _mm512_add_epi16(iacc_mat_10_1_sp1, iacc_mat_10_1_sp2);
+                        __m512i iacc_mat_11_1 = _mm512_add_epi16(iacc_mat_11_1_sp1, iacc_mat_11_1_sp2);
+
+                        iacc_mat_00_0 = _mm512_madd_epi16(iacc_mat_00_0, scale_014589CD_0);
+                        iacc_mat_01_0 = _mm512_madd_epi16(iacc_mat_01_0, scale_2367ABEF_0);
+                        iacc_mat_10_0 = _mm512_madd_epi16(iacc_mat_10_0, scale_014589CD_0);
+                        iacc_mat_11_0 = _mm512_madd_epi16(iacc_mat_11_0, scale_2367ABEF_0);
+
+                        iacc_mat_00_1 = _mm512_madd_epi16(iacc_mat_00_1, scale_014589CD_1);
+                        iacc_mat_01_1 = _mm512_madd_epi16(iacc_mat_01_1, scale_2367ABEF_1);
+                        iacc_mat_10_1 = _mm512_madd_epi16(iacc_mat_10_1, scale_014589CD_1);
+                        iacc_mat_11_1 = _mm512_madd_epi16(iacc_mat_11_1, scale_2367ABEF_1);
+
+                        // Straighten out to make 4 row vectors (4 for each sub block which are accumulated together in the next step)
+                        __m512i iacc_row_0_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00_0, _mm512_shuffle_epi32(iacc_mat_01_0, (_MM_PERM_ENUM)78));
+                        __m512i iacc_row_1_0 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00_0, (_MM_PERM_ENUM)78), iacc_mat_01_0);
+                        __m512i iacc_row_2_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10_0, _mm512_shuffle_epi32(iacc_mat_11_0, (_MM_PERM_ENUM)78));
+                        __m512i iacc_row_3_0 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10_0, (_MM_PERM_ENUM)78), iacc_mat_11_0);
+                        __m512i iacc_row_0_1 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00_1, _mm512_shuffle_epi32(iacc_mat_01_1, (_MM_PERM_ENUM)78));
+                        __m512i iacc_row_1_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00_1, (_MM_PERM_ENUM)78), iacc_mat_01_1);
+                        __m512i iacc_row_2_1 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10_1, _mm512_shuffle_epi32(iacc_mat_11_1, (_MM_PERM_ENUM)78));
+                        __m512i iacc_row_3_1 = _mm512_mask_blend_epi32(0xCCCC,_mm512_shuffle_epi32(iacc_mat_10_1, (_MM_PERM_ENUM)78), iacc_mat_11_1);
+
+                        __m512i iacc_row_0 = _mm512_add_epi32(iacc_row_0_0, iacc_row_0_1);
+                        __m512i iacc_row_1 = _mm512_add_epi32(iacc_row_1_0, iacc_row_1_1);
+                        __m512i iacc_row_2 = _mm512_add_epi32(iacc_row_2_0, iacc_row_2_1);
+                        __m512i iacc_row_3 = _mm512_add_epi32(iacc_row_3_0, iacc_row_3_1);
+
+                        // Load the scale(d) values for all the 4 Q8_k blocks and repeat it across lanes
+                        const __m128 row_scale_f32_sse = _mm_load_ps(a_ptrs[rp][b].d);
+                        const __m256 row_scale_f32_ymm = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
+                        const __m512 row_scale_f32 = _mm512_insertf32x8(_mm512_castps256_ps512(row_scale_f32_ymm), row_scale_f32_ymm, 1);
+
+                        // Multiply with appropiate scales and accumulate (for both d and dmin) below
+                        acc_rows[rp * 4] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
+                        acc_rows[rp * 4  + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
+                        acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
+                        acc_rows[rp * 4 + 3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]);
+
+                        __m512i iacc_row_min_0 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)0), mins_01);
+                        __m512i iacc_row_min_1 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)85), mins_01);
+                        __m512i iacc_row_min_2 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)170), mins_01);
+                        __m512i iacc_row_min_3 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)255), mins_01);
+
+                        acc_min_rows[rp * 4] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_0), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_min_rows[rp * 4]);
+                        acc_min_rows[rp * 4 + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_1), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_min_rows[rp * 4 + 1]);
+                        acc_min_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_2), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_min_rows[rp * 4 + 2]);
+                        acc_min_rows[rp * 4 + 3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_3), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_min_rows[rp * 4 + 3]);
+                    }
+                }
+            }
+            // Store the accumulated values
+            for (int i = 0; i < 16; i++) {
+                _mm512_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm512_sub_ps(acc_rows[i], acc_min_rows[i]));
+            }
+        }
+    }
+
+    for (; y < nr / 4; y++) {
+
+        const block_q8_Kx4 * a_ptr = a_ptr_start + (y * nb);
+
+        // Take group of eight block_q4_kx8 structures at each pass of the loop and perform dot product operation
+        for (int64_t x = 0; x < anc / 8; x += 2) {
+
+            const block_q4_Kx8 * b_ptr_0 = b_ptr_start + ((x) * b_nb);
+            const block_q4_Kx8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
+
+            // Master FP accumulators
+            __m512 acc_rows[4];
+            for (int i = 0; i < 4; i++) {
+                acc_rows[i] = _mm512_setzero_ps();
+            }
+
+            __m512 acc_min_rows[4];
+            for (int i = 0; i < 4; i++) {
+                acc_min_rows[i] = _mm512_setzero_ps();
+            }
+
+            // For super block
+            for (int64_t b = 0; b < nb; b++) {
+                // Scale values - Load the sixteen scale values from two block_q4_kx8 structures
+                const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
+
+                // dmin values - Load the sixteen dmin values from two block_q4_kx8 structures
+                const __m512 col_dmin_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].dmin, b_ptr_1[b].dmin);
+
+                // Loop to iterate over the eight sub blocks of a super block - two sub blocks are processed per iteration
+                for (int sb = 0; sb < QK_K / 64; sb++) {
+
+                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + sb * 256));
+                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 32 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 64 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 96 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 128 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 160 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 192 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 224 + sb * 256));
+
+                    const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 32 + sb * 256));
+                    const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 64 + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 96 + sb * 256));
+                    const __m256i rhs_raw_mat_89AB_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 128 + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 160 + sb * 256));
+                    const __m256i rhs_raw_mat_89AB_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 192 + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 224 + sb * 256));
+
+                    const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+                    const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+                    const __m256i rhs_raw_mat_0145_2 = _mm256_blend_epi32(rhs_raw_mat_0123_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_2, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_2, requiredOrder), rhs_raw_mat_4567_2, 240);
+                    const __m256i rhs_raw_mat_0145_3 = _mm256_blend_epi32(rhs_raw_mat_0123_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_3, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_3, requiredOrder), rhs_raw_mat_4567_3, 240);
+
+                    const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240);
+                    const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240);
+                    const __m256i rhs_raw_mat_89CD_2 = _mm256_blend_epi32(rhs_raw_mat_89AB_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_2, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_2, requiredOrder), rhs_raw_mat_CDEF_2, 240);
+                    const __m256i rhs_raw_mat_89CD_3 = _mm256_blend_epi32(rhs_raw_mat_89AB_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_3, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_3, requiredOrder), rhs_raw_mat_CDEF_3, 240);
+
+                    const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
+                    const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
+
+                    const __m512i rhs_raw_mat_014589CD_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_2), rhs_raw_mat_89CD_2, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_2), rhs_raw_mat_ABEF_2, 1);
+                    const __m512i rhs_raw_mat_014589CD_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_3), rhs_raw_mat_89CD_3, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_3), rhs_raw_mat_ABEF_3, 1);
+
+                    //4-bit -> 8-bit
+                    const __m512i rhs_mat_014589CD_00 = _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded); //B00(0-7) B01(0-7) B04(0-7) B05(0-7) B08(0-7) B09(0-7) B0C(0-7) B0D(0-7)
+                    const __m512i rhs_mat_2367ABEF_00 = _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded); //B02(0-7) B03(0-7) B06(0-7) B07(0-7) B0A(0-7) B0B(0-7) B0E(0-7) B0F(0-7)
+                    const __m512i rhs_mat_014589CD_01 = _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded); //B00(8-15) B01(8-15) B04(8-15) B05(8-15) B08(8-15) B09(8-15) B0C(8-15) B0D(8-15)
+                    const __m512i rhs_mat_2367ABEF_01 = _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded); //B02(8-15) B03(8-15) B06(8-15) B07(8-15) B0A(8-15) B0B(8-15) B0E(8-15) B0F(8-15)
+
+                    const __m512i rhs_mat_014589CD_02 = _mm512_and_si512(rhs_raw_mat_014589CD_2, m4bexpanded); //B00(16-23) B01(16-23) B04(16-23) B05(16-23) B08(16-23) B09(16-23) B0C(16-23) B0D(16-23)
+                    const __m512i rhs_mat_2367ABEF_02 = _mm512_and_si512(rhs_raw_mat_2367ABEF_2, m4bexpanded); //B02(16-23) B03(16-23) B06(16-23) B07(16-23) B0A(16-23) B0B(16-23) B0E(16-23) B0F(16-23)
+                    const __m512i rhs_mat_014589CD_03 = _mm512_and_si512(rhs_raw_mat_014589CD_3, m4bexpanded); //B00(24-31) B01(24-31) B04(24-31) B05(24-31) B08(24-31) B09(24-31) B0C(24-31) B0D(24-31)
+                    const __m512i rhs_mat_2367ABEF_03 = _mm512_and_si512(rhs_raw_mat_2367ABEF_3, m4bexpanded); //B02(24-31) B03(24-31) B06(24-31) B07(24-31) B0A(24-31) B0B(24-31) B0E(24-31) B0F(24-31)
+
+                    const __m512i rhs_mat_014589CD_10 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded); //B10(0-7) B11(0-7) B14(0-7) B15(0-7) B18(0-7) B19(0-7) B1C(0-7) B1D(0-7)
+                    const __m512i rhs_mat_2367ABEF_10 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded); //B12(0-7) B13(0-7) B16(0-7) B17(0-7) B1A(0-7) B1B(0-7) B1E(0-7) B1F(0-7)
+                    const __m512i rhs_mat_014589CD_11 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded); //B10(8-15) B11(8-15) B14(8-15) B15(8-15) B18(8-15) B19(8-15) B1C(8-15) B1D(8-15)
+                    const __m512i rhs_mat_2367ABEF_11 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded); //B12(8-15) B13(8-15) B16(8-15) B17(8-15) B1A(8-15) B1B(8-15) B1E(8-15) B1F(8-15)
+
+                    const __m512i rhs_mat_014589CD_12 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2, 4), m4bexpanded); //B10(16-23) B11(16-23) B14(16-23) B15(16-23) B18(16-23) B19(16-23) B1C(16-23) B1D(16-23)
+                    const __m512i rhs_mat_2367ABEF_12 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2, 4), m4bexpanded); //B12(16-23) B13(16-23) B16(16-23) B17(16-23) B1A(16-23) B1B(16-23) B1E(16-23) B1F(16-23)
+                    const __m512i rhs_mat_014589CD_13 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3, 4), m4bexpanded); //B10(24-31) B11(24-31) B14(24-31) B15(24-31) B18(24-31) B19(24-31) B1C(24-31) B1D(24-31)
+                    const __m512i rhs_mat_2367ABEF_13 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3, 4), m4bexpanded); //B12(24-31) B13(24-31) B16(24-31) B17(24-31) B1A(24-31) B1B(24-31) B1E(24-31) B1F(24-31)
+
+                    // Shuffle pattern one - right side input
+                    const __m512i rhs_mat_014589CD_00_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_00, (_MM_PERM_ENUM)136); //B00(0-3) B01(0-3) B00(0-3) B01(0-3) B04(0-3) B05(0-3) B04(0-3) B05(0-3) B08(0-3) B09(0-3) B08(0-3) B09(0-3) B0C(0-3) B0D(0-3) B0C(0-3) B0D(0-3)
+                    const __m512i rhs_mat_2367ABEF_00_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_00, (_MM_PERM_ENUM)136); //B02(0-3) B03(0-3) B02(0-3) B03(0-3) B06(0-3) B07(0-3) B06(0-3) B07(0-3) B0A(0-3) B0B(0-3) B0A(0-3) B0B(0-3) B0E(0-3) B0F(0-3) B0E(0-3) B0F(0-3)
+                    const __m512i rhs_mat_014589CD_01_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_01, (_MM_PERM_ENUM)136); //B00(8-11) B01(8-11) B00(8-11) B01(8-11) B04(8-11) B05(8-11) B04(8-11) B05(8-11) B08(8-11) B09(8-11) B08(8-11) B09(8-11) B0C(8-11) B0D(8-11) B0C(8-11) B0D(8-11)
+                    const __m512i rhs_mat_2367ABEF_01_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_01, (_MM_PERM_ENUM)136); //B02(8-11) B03(8-11) B02(8-11) B03(8-11) B06(8-11) B07(8-11) B06(8-11) B07(8-11) B0A(8-11) B0B(8-11) B0A(8-11) B0B(8-11) B0E(8-11) B0F(8-11) B0E(8-11) B0F(8-11)
+                    const __m512i rhs_mat_014589CD_02_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_02, (_MM_PERM_ENUM)136); //B00(16-19) B01(16-19) B00(16-19) B01(16-19) B04(16-19) B05(16-19) B04(16-19) B05(16-19) B08(16-19) B09(16-19) B08(16-19) B09(16-19) B0C(16-19) B0D(16-19) B0C(16-19) B0D(16-19)
+                    const __m512i rhs_mat_2367ABEF_02_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_02, (_MM_PERM_ENUM)136); //B02(16-19) B03(16-19) B02(16-19) B03(16-19) B06(16-19) B07(16-19) B06(16-19) B07(16-19) B0A(16-19) B0B(16-19) B0A(16-19) B0B(16-19) B0E(16-19) B0F(16-19) B0E(16-19) B0F(16-19)
+                    const __m512i rhs_mat_014589CD_03_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_03, (_MM_PERM_ENUM)136); //B00(24-27) B01(24-27) B00(24-27) B01(24-27) B04(24-27) B05(24-27) B04(24-27) B05(24-27) B08(24-27) B09(24-27) B08(24-27) B09(24-27) B0C(24-27) B0D(24-27) B0C(24-27) B0D(24-27)
+                    const __m512i rhs_mat_2367ABEF_03_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_03, (_MM_PERM_ENUM)136); //B02(24-27) B03(24-27) B02(24-27) B03(24-27) B06(24-27) B07(24-27) B06(24-27) B07(24-27) B0A(24-27) B0B(24-27) B0A(24-27) B0B(24-27) B0E(24-27) B0F(24-27) B0E(24-27) B0F(24-27)
+
+                    const __m512i rhs_mat_014589CD_10_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_10, (_MM_PERM_ENUM)136); //B10(0-3) B11(0-3) B10(0-3) B11(0-3) B14(0-3) B15(0-3) B14(0-3) B15(0-3) B18(0-3) B19(0-3) B18(0-3) B19(0-3) B1C(0-3) B1D(0-3) B1C(0-3) B1D(0-3)
+                    const __m512i rhs_mat_2367ABEF_10_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_10, (_MM_PERM_ENUM)136); //B12(0-3) B13(0-3) B12(0-3) B13(0-3) B16(0-3) B17(0-3) B16(0-3) B17(0-3) B1A(0-3) B1B(0-3) B1A(0-3) B1B(0-3) B1E(0-3) B1F(0-3) B1E(0-3) B1F(0-3)
+                    const __m512i rhs_mat_014589CD_11_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_11, (_MM_PERM_ENUM)136); //B10(8-11) B11(8-11) B10(8-11) B11(8-11) B14(8-11) B15(8-11) B14(8-11) B15(8-11) B18(8-11) B19(8-11) B18(8-11) B19(8-11) B1C(8-11) B1D(8-11) B1C(8-11) B1D(8-11)
+                    const __m512i rhs_mat_2367ABEF_11_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_11, (_MM_PERM_ENUM)136); //B12(8-11) B13(8-11) B12(8-11) B13(8-11) B16(8-11) B17(8-11) B16(8-11) B17(8-11) B1A(8-11) B1B(8-11) B1A(8-11) B1B(8-11) B1E(8-11) B1F(8-11) B1E(8-11) B1F(8-11)
+                    const __m512i rhs_mat_014589CD_12_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_12, (_MM_PERM_ENUM)136); //B10(16-19) B11(16-19) B10(16-19) B11(16-19) B14(16-19) B15(16-19) B14(16-19) B15(16-19) B18(16-19) B19(16-19) B18(16-19) B19(16-19) B1C(16-19) B1D(16-19) B1C(16-19) B1D(16-19)
+                    const __m512i rhs_mat_2367ABEF_12_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_12, (_MM_PERM_ENUM)136); //B12(16-19) B13(16-19) B12(16-19) B13(16-19) B16(16-19) B17(16-19) B16(16-19) B17(16-19) B1A(16-19) B1B(16-19) B1A(16-19) B1B(16-19) B1E(16-19) B1F(16-19) B1E(16-19) B1F(16-19)
+                    const __m512i rhs_mat_014589CD_13_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_13, (_MM_PERM_ENUM)136); //B10(24-27) B11(24-27) B10(24-27) B11(24-27) B14(24-27) B15(24-27) B14(24-27) B15(24-27) B18(24-27) B19(24-27) B18(24-27) B19(24-27) B1C(24-27) B1D(24-27) B1C(24-27) B1D(24-27)
+                    const __m512i rhs_mat_2367ABEF_13_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_13, (_MM_PERM_ENUM)136); //B12(24-27) B13(24-27) B12(24-27) B13(24-27) B16(24-27) B17(24-27) B16(24-27) B17(24-27) B1A(24-27) B1B(24-27) B1A(24-27) B1B(24-27) B1E(24-27) B1F(24-27) B1E(24-27) B1F(24-27)
+
+                    // Shuffle pattern two - right side input
+                    const __m512i rhs_mat_014589CD_00_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_00, (_MM_PERM_ENUM)221); //B00(4-7) B01(4-7) B00(4-7) B01(4-7) B04(4-7) B05(4-7) B04(4-7) B05(4-7) B08(4-7) B09(4-7) B08(4-7) B09(4-7) B0C(4-7) B0D(4-7) B0C(4-7) B0D(4-7)
+                    const __m512i rhs_mat_2367ABEF_00_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_00, (_MM_PERM_ENUM)221); //B02(4-7) B03(4-7) B02(4-7) B03(4-7) B06(4-7) B07(4-7) B06(4-7) B07(4-7) B0A(4-7) B0B(4-7) B0A(4-7) B0B(4-7) B0E(4-7) B0F(4-7) B0E(4-7) B0F(4-7)
+                    const __m512i rhs_mat_014589CD_01_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_01, (_MM_PERM_ENUM)221); //B00(12-15) B01(12-15) B00(12-15) B01(12-15) B04(12-15) B05(12-15) B04(12-15) B05(12-15) B08(12-15) B09(12-15) B08(12-15) B09(12-15) B0C(12-15) B0D(12-15) B0C(12-15) B0D(12-15)
+                    const __m512i rhs_mat_2367ABEF_01_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_01, (_MM_PERM_ENUM)221); //B02(12-15) B03(12-15) B02(12-15) B03(12-15) B06(12-15) B07(12-15) B06(12-15) B07(12-15) B0A(12-15) B0B(12-15) B0A(12-15) B0B(12-15) B0E(12-15) B0F(12-15) B0E(12-15) B0F(12-15)
+                    const __m512i rhs_mat_014589CD_02_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_02, (_MM_PERM_ENUM)221); //B00(20-23) B01(20-23) B00(20-23) B01(20-23) B04(20-23) B05(20-23) B04(20-23) B05(20-23) B08(20-23) B09(20-23) B08(20-23) B09(20-23) B0C(20-23) B0D(20-23) B0C(20-23) B0D(20-23)
+                    const __m512i rhs_mat_2367ABEF_02_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_02, (_MM_PERM_ENUM)221); //B02(20-23) B03(20-23) B02(20-23) B03(20-23) B06(20-23) B07(20-23) B06(20-23) B07(20-23) B0A(20-23) B0B(20-23) B0A(20-23) B0B(20-23) B0E(20-23) B0F(20-23) B0E(20-23) B0F(20-23)
+                    const __m512i rhs_mat_014589CD_03_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_03, (_MM_PERM_ENUM)221); //B00(28-31) B01(28-31) B00(28-31) B01(28-31) B04(28-31) B05(28-31) B04(28-31) B05(28-31) B08(28-31) B09(28-31) B08(28-31) B09(28-31) B0C(28-31) B0D(28-31) B0C(28-31) 0BD(28-31)
+                    const __m512i rhs_mat_2367ABEF_03_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_03, (_MM_PERM_ENUM)221); //B02(28-31) B03(28-31) B02(28-31) B03(28-31) B06(28-31) B07(28-31) B06(28-31) B07(28-31) B0A(28-31) B0B(28-31) B0A(28-31) B0B(28-31) B0E(28-31) B0F(28-31) B0E(28-31) B0F(28-31)
+
+                    const __m512i rhs_mat_014589CD_10_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_10, (_MM_PERM_ENUM)221); //B10(4-7) B11(4-7) B10(4-7) B11(4-7) B14(4-7) B15(4-7) B14(4-7) B15(4-7) B18(4-7) B19(4-7) B18(4-7) B19(4-7) B1C(4-7) B1D(4-7) B1C(4-7) B1D(4-7)
+                    const __m512i rhs_mat_2367ABEF_10_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_10, (_MM_PERM_ENUM)221); //B12(4-7) B13(4-7) B12(4-7) B13(4-7) B16(4-7) B17(4-7) B16(4-7) B17(4-7) B1A(4-7) B1B(4-7) B1A(4-7) B1B(4-7) B1E(4-7) B1F(4-7) B1E(4-7) B1F(4-7)
+                    const __m512i rhs_mat_014589CD_11_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_11, (_MM_PERM_ENUM)221); //B10(12-15) B11(12-15) B10(12-15) B11(12-15) B14(12-15) B15(12-15) B14(12-15) B15(12-15) B18(12-15) B19(12-15) B18(12-15) B19(12-15) B1C(12-15) B1D(12-15) B1C(12-15) B1D(12-15)
+                    const __m512i rhs_mat_2367ABEF_11_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_11, (_MM_PERM_ENUM)221); //B12(12-15) B13(12-15) B12(12-15) B13(12-15) B16(12-15) B17(12-15) B16(12-15) B17(12-15) B1A(12-15) B1B(12-15) B1A(12-15) B1B(12-15) B1E(12-15) B1F(12-15) B1E(12-15) B1F(12-15)
+                    const __m512i rhs_mat_014589CD_12_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_12, (_MM_PERM_ENUM)221); //B10(20-23) B11(20-23) B10(20-23) B11(20-23) B14(20-23) B15(20-23) B14(20-23) B15(20-23) B18(20-23) B19(20-23) B18(20-23) B19(20-23) B1C(20-23) B1D(20-23) B1C(20-23) B1D(20-23)
+                    const __m512i rhs_mat_2367ABEF_12_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_12, (_MM_PERM_ENUM)221); //B12(20-23) B13(20-23) B12(20-23) B13(20-23) B16(20-23) B17(20-23) B16(20-23) B17(20-23) B1A(20-23) B1B(20-23) B1A(20-23) B1B(20-23) B1E(20-23) B1F(20-23) B1E(20-23) B1F(20-23)
+                    const __m512i rhs_mat_014589CD_13_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_13, (_MM_PERM_ENUM)221); //B10(28-31) B11(28-31) B10(28-31) B11(28-31) B14(28-31) B15(28-31) B14(28-31) B15(28-31) B18(28-31) B19(28-31) B18(28-31) B19(28-31) B1C(28-31) B1D(28-31) B1C(28-31) B1D(28-31)
+                    const __m512i rhs_mat_2367ABEF_13_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_13, (_MM_PERM_ENUM)221); //B12(28-31) B13(28-31) B12(28-31) B13(28-31) B16(28-31) B17(28-31) B16(28-31) B17(28-31) B1A(28-31) B1B(28-31) B1A(28-31) B1B(28-31) B1E(28-31) B1F(28-31) B1E(28-31) B1F(28-31)
+
+                    uint32_t utmp_00[4], utmp_01[4], utmp_10[4], utmp_11[4];
+
+                    // Scales and Mins of corresponding sub blocks from different Q4_K structures are stored together
+                    // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop
+                    memcpy(utmp_00, b_ptr_0[b].scales + 24 * sb, 12);
+                    utmp_00[3] = ((utmp_00[2] >> 4) & kmask2) | (((utmp_00[1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_00 = utmp_00[1] & kmask1;
+                    utmp_00[1] = (utmp_00[2] & kmask2) | (((utmp_00[0] >> 6) & kmask3) << 4);
+                    utmp_00[2] = uaux_00;
+                    utmp_00[0] &= kmask1;
+
+                    // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop
+                    memcpy(utmp_01, b_ptr_0[b].scales + 12 + sb * 24, 12);
+                    utmp_01[3] = ((utmp_01[2] >> 4) & kmask2) | (((utmp_01[1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_01 = utmp_01[1] & kmask1;
+                    utmp_01[1] = (utmp_01[2] & kmask2) | (((utmp_01[0] >> 6) & kmask3) << 4);
+                    utmp_01[2] = uaux_01;
+                    utmp_01[0] &= kmask1;
+
+                    // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop
+                    memcpy(utmp_10, b_ptr_1[b].scales + sb * 24, 12);
+                    utmp_10[3] = ((utmp_10[2] >> 4) & kmask2) | (((utmp_10[1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_10 = utmp_10[1] & kmask1;
+                    utmp_10[1] = (utmp_10[2] & kmask2) | (((utmp_10[0] >> 6) & kmask3) << 4);
+                    utmp_10[2] = uaux_10;
+                    utmp_10[0] &= kmask1;
+
+                    // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop
+                    memcpy(utmp_11, b_ptr_1[b].scales + 12 + sb * 24, 12);
+                    utmp_11[3] = ((utmp_11[2] >> 4) & kmask2) | (((utmp_11[1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_11 = utmp_11[1] & kmask1;
+                    utmp_11[1] = (utmp_11[2] & kmask2) | (((utmp_11[0] >> 6) & kmask3) << 4);
+                    utmp_11[2] = uaux_11;
+                    utmp_11[0] &= kmask1;
+
+                    // Scales of first sub block in the sb loop
+                    const __m256i mins_and_scales_0 = _mm256_set_epi32(utmp_10[3], utmp_10[2], utmp_10[1], utmp_10[0], utmp_00[3], utmp_00[2], utmp_00[1], utmp_00[0]);
+                    const __m512i scales_0 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(mins_and_scales_0, mins_and_scales_0));
+
+                    // Scales of second sub block in the sb loop
+                    const __m256i mins_and_scales_1 = _mm256_set_epi32(utmp_11[3], utmp_11[2], utmp_11[1], utmp_11[0], utmp_01[3], utmp_01[2], utmp_01[1], utmp_01[0]);
+                    const __m512i scales_1 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(mins_and_scales_1, mins_and_scales_1));
+
+                    // Mins of first and second sub block of Q4_K block are arranged side by side
+                    const __m512i mins_01 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(_mm256_shuffle_epi32(mins_and_scales_0, 78), _mm256_shuffle_epi32(mins_and_scales_1, 78)));
+
+                    const __m512i scale_014589CD_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)238);
+
+                    const __m512i scale_014589CD_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)238);
+
+                    // Load the four block_q8_k quantized values interleaved with each other in chunks of eight bytes - A0,A1,A2,A3
+                    // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
+                    __m256i lhs_mat_ymm_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 256 * sb)));
+                    __m256i lhs_mat_ymm_01_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 0);
+                    __m256i lhs_mat_ymm_23_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 17);
+                    __m256i lhs_mat_ymm_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 32 + 256 * sb)));
+                    __m256i lhs_mat_ymm_01_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 0);
+                    __m256i lhs_mat_ymm_23_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 17);
+                    __m256i lhs_mat_ymm_0123_02 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 64 + 256 * sb)));
+                    __m256i lhs_mat_ymm_01_02 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_02, lhs_mat_ymm_0123_02, 0);
+                    __m256i lhs_mat_ymm_23_02 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_02, lhs_mat_ymm_0123_02, 17);
+                    __m256i lhs_mat_ymm_0123_03 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 96 + 256 * sb)));
+                    __m256i lhs_mat_ymm_01_03 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_03, lhs_mat_ymm_0123_03, 0);
+                    __m256i lhs_mat_ymm_23_03 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_03, lhs_mat_ymm_0123_03, 17);
+                    __m256i lhs_mat_ymm_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 128 + 256 * sb)));
+                    __m256i lhs_mat_ymm_01_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 0);
+                    __m256i lhs_mat_ymm_23_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 17);
+                    __m256i lhs_mat_ymm_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 160 + 256 * sb)));
+                    __m256i lhs_mat_ymm_01_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 0);
+                    __m256i lhs_mat_ymm_23_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 17);
+                    __m256i lhs_mat_ymm_0123_12 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 192 + 256 * sb)));
+                    __m256i lhs_mat_ymm_01_12 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_12, lhs_mat_ymm_0123_12, 0);
+                    __m256i lhs_mat_ymm_23_12 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_12, lhs_mat_ymm_0123_12, 17);
+                    __m256i lhs_mat_ymm_0123_13 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 224 + 256 * sb)));
+                    __m256i lhs_mat_ymm_01_13 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_13, lhs_mat_ymm_0123_13, 0);
+                    __m256i lhs_mat_ymm_23_13 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_13, lhs_mat_ymm_0123_13, 17);
+
+                    //Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into a 512 bit vector
+                    __m512i lhs_mat_01_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_00), lhs_mat_ymm_01_00, 1);
+                    __m512i lhs_mat_23_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_00), lhs_mat_ymm_23_00, 1);
+                    __m512i lhs_mat_01_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_01), lhs_mat_ymm_01_01, 1);
+                    __m512i lhs_mat_23_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_01), lhs_mat_ymm_23_01, 1);
+                    __m512i lhs_mat_01_02 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_02), lhs_mat_ymm_01_02, 1);
+                    __m512i lhs_mat_23_02 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_02), lhs_mat_ymm_23_02, 1);
+                    __m512i lhs_mat_01_03 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_03), lhs_mat_ymm_01_03, 1);
+                    __m512i lhs_mat_23_03 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_03), lhs_mat_ymm_23_03, 1);
+
+                    __m512i lhs_mat_01_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_10), lhs_mat_ymm_01_10, 1);
+                    __m512i lhs_mat_23_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_10), lhs_mat_ymm_23_10, 1);
+                    __m512i lhs_mat_01_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_11), lhs_mat_ymm_01_11, 1);
+                    __m512i lhs_mat_23_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_11), lhs_mat_ymm_23_11, 1);
+                    __m512i lhs_mat_01_12 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_12), lhs_mat_ymm_01_12, 1);
+                    __m512i lhs_mat_23_12 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_12), lhs_mat_ymm_23_12, 1);
+                    __m512i lhs_mat_01_13 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_13), lhs_mat_ymm_01_13, 1);
+                    __m512i lhs_mat_23_13 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_13), lhs_mat_ymm_23_13, 1);
+
+                    // Bsums are loaded - four bsums are loaded (for two sub blocks) for the different Q8_K blocks
+                    __m256i lhs_bsums_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].bsums + 16 * sb)));
+                    __m256i lhs_bsums_hsum_ymm_0123_01 = _mm256_castsi128_si256(_mm_hadd_epi16(_mm256_castsi256_si128(lhs_bsums_0123_01), _mm256_extractf128_si256(lhs_bsums_0123_01, 1)));
+                    lhs_bsums_hsum_ymm_0123_01 = _mm256_permute2x128_si256(lhs_bsums_hsum_ymm_0123_01, lhs_bsums_hsum_ymm_0123_01, 0);
+                    __m512i lhs_bsums_hsum_0123_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_hsum_ymm_0123_01), lhs_bsums_hsum_ymm_0123_01, 1);
+
+                    // Shuffle pattern one - left side input
+                    const __m512i lhs_mat_01_00_sp1 = _mm512_shuffle_epi32(lhs_mat_01_00, (_MM_PERM_ENUM)160); //A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3)
+                    const __m512i lhs_mat_23_00_sp1 = _mm512_shuffle_epi32(lhs_mat_23_00, (_MM_PERM_ENUM)160); //A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3)
+                    const __m512i lhs_mat_01_01_sp1 = _mm512_shuffle_epi32(lhs_mat_01_01, (_MM_PERM_ENUM)160); //A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11)
+                    const __m512i lhs_mat_23_01_sp1 = _mm512_shuffle_epi32(lhs_mat_23_01, (_MM_PERM_ENUM)160); //A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11)
+                    const __m512i lhs_mat_01_02_sp1 = _mm512_shuffle_epi32(lhs_mat_01_02, (_MM_PERM_ENUM)160); //A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19)
+                    const __m512i lhs_mat_23_02_sp1 = _mm512_shuffle_epi32(lhs_mat_23_02, (_MM_PERM_ENUM)160); //A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19)
+                    const __m512i lhs_mat_01_03_sp1 = _mm512_shuffle_epi32(lhs_mat_01_03, (_MM_PERM_ENUM)160); //A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27)
+                    const __m512i lhs_mat_23_03_sp1 = _mm512_shuffle_epi32(lhs_mat_23_03, (_MM_PERM_ENUM)160); //A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27)
+
+                    const __m512i lhs_mat_01_10_sp1 = _mm512_shuffle_epi32(lhs_mat_01_10, (_MM_PERM_ENUM)160); //A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3)
+                    const __m512i lhs_mat_23_10_sp1 = _mm512_shuffle_epi32(lhs_mat_23_10, (_MM_PERM_ENUM)160); //A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3)
+                    const __m512i lhs_mat_01_11_sp1 = _mm512_shuffle_epi32(lhs_mat_01_11, (_MM_PERM_ENUM)160); //A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11)
+                    const __m512i lhs_mat_23_11_sp1 = _mm512_shuffle_epi32(lhs_mat_23_11, (_MM_PERM_ENUM)160); //A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11)
+                    const __m512i lhs_mat_01_12_sp1 = _mm512_shuffle_epi32(lhs_mat_01_12, (_MM_PERM_ENUM)160); //A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19)
+                    const __m512i lhs_mat_23_12_sp1 = _mm512_shuffle_epi32(lhs_mat_23_12, (_MM_PERM_ENUM)160); //A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19)
+                    const __m512i lhs_mat_01_13_sp1 = _mm512_shuffle_epi32(lhs_mat_01_13, (_MM_PERM_ENUM)160); //A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27)
+                    const __m512i lhs_mat_23_13_sp1 = _mm512_shuffle_epi32(lhs_mat_23_13, (_MM_PERM_ENUM)160); //A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27)
+
+                    const __m512i lhs_mat_01_00_sp2 = _mm512_shuffle_epi32(lhs_mat_01_00, (_MM_PERM_ENUM)245); //A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7)
+                    const __m512i lhs_mat_23_00_sp2 = _mm512_shuffle_epi32(lhs_mat_23_00, (_MM_PERM_ENUM)245); //A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7)
+                    const __m512i lhs_mat_01_01_sp2 = _mm512_shuffle_epi32(lhs_mat_01_01, (_MM_PERM_ENUM)245); //A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15)
+                    const __m512i lhs_mat_23_01_sp2 = _mm512_shuffle_epi32(lhs_mat_23_01, (_MM_PERM_ENUM)245); //A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15)
+                    const __m512i lhs_mat_01_02_sp2 = _mm512_shuffle_epi32(lhs_mat_01_02, (_MM_PERM_ENUM)245); //A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23)
+                    const __m512i lhs_mat_23_02_sp2 = _mm512_shuffle_epi32(lhs_mat_23_02, (_MM_PERM_ENUM)245); //A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23)
+                    const __m512i lhs_mat_01_03_sp2 = _mm512_shuffle_epi32(lhs_mat_01_03, (_MM_PERM_ENUM)245); //A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31)
+                    const __m512i lhs_mat_23_03_sp2 = _mm512_shuffle_epi32(lhs_mat_23_03, (_MM_PERM_ENUM)245); //A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31)
+
+                    const __m512i lhs_mat_01_10_sp2 = _mm512_shuffle_epi32(lhs_mat_01_10, (_MM_PERM_ENUM)245); //A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7)
+                    const __m512i lhs_mat_23_10_sp2 = _mm512_shuffle_epi32(lhs_mat_23_10, (_MM_PERM_ENUM)245); //A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7)
+                    const __m512i lhs_mat_01_11_sp2 = _mm512_shuffle_epi32(lhs_mat_01_11, (_MM_PERM_ENUM)245); //A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15)
+                    const __m512i lhs_mat_23_11_sp2 = _mm512_shuffle_epi32(lhs_mat_23_11, (_MM_PERM_ENUM)245); //A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15)
+                    const __m512i lhs_mat_01_12_sp2 = _mm512_shuffle_epi32(lhs_mat_01_12, (_MM_PERM_ENUM)245); //A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23)
+                    const __m512i lhs_mat_23_12_sp2 = _mm512_shuffle_epi32(lhs_mat_23_12, (_MM_PERM_ENUM)245); //A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23)
+                    const __m512i lhs_mat_01_13_sp2 = _mm512_shuffle_epi32(lhs_mat_01_13, (_MM_PERM_ENUM)245); //A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31)
+                    const __m512i lhs_mat_23_13_sp2 = _mm512_shuffle_epi32(lhs_mat_23_13, (_MM_PERM_ENUM)245); //A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(28-31)
+
+                    // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+                    __m512i iacc_mat_00_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp1, lhs_mat_01_03_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp1, lhs_mat_01_02_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_01_01_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_01_00_sp1));
+                    __m512i iacc_mat_01_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp1, lhs_mat_01_03_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp1, lhs_mat_01_02_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_01_01_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_01_00_sp1));
+                    __m512i iacc_mat_10_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp1, lhs_mat_23_03_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp1, lhs_mat_23_02_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_23_01_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_23_00_sp1));
+                    __m512i iacc_mat_11_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp1, lhs_mat_23_03_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp1, lhs_mat_23_02_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_23_01_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_23_00_sp1));
+                    __m512i iacc_mat_00_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp1, lhs_mat_01_13_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp1, lhs_mat_01_12_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_01_11_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_01_10_sp1));
+                    __m512i iacc_mat_01_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp1, lhs_mat_01_13_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp1, lhs_mat_01_12_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_01_11_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_01_10_sp1));
+                    __m512i iacc_mat_10_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp1, lhs_mat_23_13_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp1, lhs_mat_23_12_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_23_11_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_23_10_sp1));
+                    __m512i iacc_mat_11_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp1, lhs_mat_23_13_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp1, lhs_mat_23_12_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_23_11_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_23_10_sp1));
+
+                    __m512i iacc_mat_00_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp2, lhs_mat_01_03_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp2, lhs_mat_01_02_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_01_01_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_01_00_sp2));
+                    __m512i iacc_mat_01_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp2, lhs_mat_01_03_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp2, lhs_mat_01_02_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_01_01_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_01_00_sp2));
+                    __m512i iacc_mat_10_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp2, lhs_mat_23_03_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp2, lhs_mat_23_02_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_23_01_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_23_00_sp2));
+                    __m512i iacc_mat_11_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp2, lhs_mat_23_03_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp2, lhs_mat_23_02_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_23_01_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_23_00_sp2));
+                    __m512i iacc_mat_00_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp2, lhs_mat_01_13_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp2, lhs_mat_01_12_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_01_11_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_01_10_sp2));
+                    __m512i iacc_mat_01_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp2, lhs_mat_01_13_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp2, lhs_mat_01_12_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_01_11_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_01_10_sp2));
+                    __m512i iacc_mat_10_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp2, lhs_mat_23_13_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp2, lhs_mat_23_12_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_23_11_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_23_10_sp2));
+                    __m512i iacc_mat_11_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp2, lhs_mat_23_13_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp2, lhs_mat_23_12_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_23_11_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_23_10_sp2));
+
+                    // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+                    __m512i iacc_mat_00_0 = _mm512_add_epi16(iacc_mat_00_0_sp1, iacc_mat_00_0_sp2);
+                    __m512i iacc_mat_01_0 = _mm512_add_epi16(iacc_mat_01_0_sp1, iacc_mat_01_0_sp2);
+                    __m512i iacc_mat_10_0 = _mm512_add_epi16(iacc_mat_10_0_sp1, iacc_mat_10_0_sp2);
+                    __m512i iacc_mat_11_0 = _mm512_add_epi16(iacc_mat_11_0_sp1, iacc_mat_11_0_sp2);
+
+                    __m512i iacc_mat_00_1 = _mm512_add_epi16(iacc_mat_00_1_sp1, iacc_mat_00_1_sp2);
+                    __m512i iacc_mat_01_1 = _mm512_add_epi16(iacc_mat_01_1_sp1, iacc_mat_01_1_sp2);
+                    __m512i iacc_mat_10_1 = _mm512_add_epi16(iacc_mat_10_1_sp1, iacc_mat_10_1_sp2);
+                    __m512i iacc_mat_11_1 = _mm512_add_epi16(iacc_mat_11_1_sp1, iacc_mat_11_1_sp2);
+
+                    iacc_mat_00_0 = _mm512_madd_epi16(iacc_mat_00_0, scale_014589CD_0);
+                    iacc_mat_01_0 = _mm512_madd_epi16(iacc_mat_01_0, scale_2367ABEF_0);
+                    iacc_mat_10_0 = _mm512_madd_epi16(iacc_mat_10_0, scale_014589CD_0);
+                    iacc_mat_11_0 = _mm512_madd_epi16(iacc_mat_11_0, scale_2367ABEF_0);
+
+                    iacc_mat_00_1 = _mm512_madd_epi16(iacc_mat_00_1, scale_014589CD_1);
+                    iacc_mat_01_1 = _mm512_madd_epi16(iacc_mat_01_1, scale_2367ABEF_1);
+                    iacc_mat_10_1 = _mm512_madd_epi16(iacc_mat_10_1, scale_014589CD_1);
+                    iacc_mat_11_1 = _mm512_madd_epi16(iacc_mat_11_1, scale_2367ABEF_1);
+
+                    // Straighten out to make 4 row vectors (4 for each sub block which are accumulated together in the next step)
+                    __m512i iacc_row_0_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00_0, _mm512_shuffle_epi32(iacc_mat_01_0, (_MM_PERM_ENUM)78));
+                    __m512i iacc_row_1_0 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00_0, (_MM_PERM_ENUM)78), iacc_mat_01_0);
+                    __m512i iacc_row_2_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10_0, _mm512_shuffle_epi32(iacc_mat_11_0, (_MM_PERM_ENUM)78));
+                    __m512i iacc_row_3_0 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10_0, (_MM_PERM_ENUM)78), iacc_mat_11_0);
+                    __m512i iacc_row_0_1 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00_1, _mm512_shuffle_epi32(iacc_mat_01_1, (_MM_PERM_ENUM)78));
+                    __m512i iacc_row_1_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00_1, (_MM_PERM_ENUM)78), iacc_mat_01_1);
+                    __m512i iacc_row_2_1 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10_1, _mm512_shuffle_epi32(iacc_mat_11_1, (_MM_PERM_ENUM)78));
+                    __m512i iacc_row_3_1 = _mm512_mask_blend_epi32(0xCCCC,_mm512_shuffle_epi32(iacc_mat_10_1, (_MM_PERM_ENUM)78), iacc_mat_11_1);
+
+                    __m512i iacc_row_0 = _mm512_add_epi32(iacc_row_0_0, iacc_row_0_1);
+                    __m512i iacc_row_1 = _mm512_add_epi32(iacc_row_1_0, iacc_row_1_1);
+                    __m512i iacc_row_2 = _mm512_add_epi32(iacc_row_2_0, iacc_row_2_1);
+                    __m512i iacc_row_3 = _mm512_add_epi32(iacc_row_3_0, iacc_row_3_1);
+
+                    // Load the scale(d) values for all the 4 Q8_k blocks and repeat it across lanes
+                    const __m128 row_scale_f32_sse = _mm_load_ps(a_ptr[b].d);
+                    const __m256 row_scale_f32_ymm = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
+                    const __m512 row_scale_f32 = _mm512_insertf32x8(_mm512_castps256_ps512(row_scale_f32_ymm), row_scale_f32_ymm, 1);
+
+                    // Multiply with appropiate scales and accumulate (for both d and dmin) below
+                    acc_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
+                    acc_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
+                    acc_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
+                    acc_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]);
+
+                    __m512i iacc_row_min_0 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)0), mins_01);
+                    __m512i iacc_row_min_1 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)85), mins_01);
+                    __m512i iacc_row_min_2 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)170), mins_01);
+                    __m512i iacc_row_min_3 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)255), mins_01);
+
+                    acc_min_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_0), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_min_rows[0]);
+                    acc_min_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_1), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_min_rows[1]);
+                    acc_min_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_2), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_min_rows[2]);
+                    acc_min_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_3), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_min_rows[3]);
+                }
+            }
+            // Store accumlated values
+            for (int i = 0; i < 4; i++) {
+                _mm512_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm512_sub_ps(acc_rows[i], acc_min_rows[i]));
+            }
+        }
+    }
+    if (anc != nc) {
+        xstart = anc/8;
+        y = 0;
+    }
+#endif //AVX512F
+
     // Take group of four block_q8_Kx4 structures at each pass of the loop and perform dot product operation
     for (; y < anr / 4; y += 4) {
 
@@ -4067,7 +4807,7 @@ static void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, c
         }
 
         // Take group of eight block_q4_kx8 structures at each pass of the loop and perform dot product operation
-        for (int64_t x = 0; x < nc / 8; x++) {
+        for (int64_t x = xstart; x < nc / 8; x++) {
 
             const block_q4_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
 
@@ -4401,7 +5141,7 @@ static void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, c
 
         const block_q8_Kx4 * a_ptr = a_ptr_start + (y * nb);
 
-        for (int64_t x = 0; x < nc / 8; x++) {
+        for (int64_t x = xstart; x < nc / 8; x++) {
 
             const block_q4_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
 

ggml/src/ggml-cuda/cpy.cu

@@ -551,7 +551,7 @@ static void ggml_cpy_f16_f16_cuda(
         (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
-void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1) {
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1, bool disable_indirection_for_this_node) {
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
 
@@ -588,7 +588,7 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
     char ** dest_ptrs_d = nullptr;
     int graph_cpynode_index = -1;
 #if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS)
-    if(ctx.cuda_graph->use_cpy_indirection) {
+    if(ctx.cuda_graph->use_cpy_indirection && !disable_indirection_for_this_node) {
         dest_ptrs_d = ctx.cuda_graph->dest_ptrs_d;
         graph_cpynode_index = ctx.cuda_graph->graph_cpynode_index;
     }
@@ -636,7 +636,7 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
                 ggml_type_name(src0->type), ggml_type_name(src1->type));
     }
 #if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS)
-    if(ctx.cuda_graph->use_cpy_indirection) {
+    if(ctx.cuda_graph->use_cpy_indirection && !disable_indirection_for_this_node) {
         ctx.cuda_graph->graph_cpynode_index = graph_cpynode_index;
     }
 #endif
@@ -645,7 +645,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
 
 void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
-    ggml_cuda_cpy(ctx, src0, dst);
+    bool disable_indirection = true;
+    ggml_cuda_cpy(ctx, src0, dst, disable_indirection);
 }
 
 void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {

ggml/src/ggml-cuda/cpy.cuh

@@ -2,7 +2,7 @@
 
 #define CUDA_CPY_BLOCK_SIZE 64
 
-void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1);
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1,  bool disable_indirection = false);
 
 void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 

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

@@ -96,31 +96,32 @@ int ggml_cuda_get_device() {
 
 static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
     ggml_cuda_set_device(device);
-#if defined(GGML_USE_HIP) && defined(GGML_HIP_UMA)
-    auto res = hipMallocManaged(ptr, size);
-    if (res == hipSuccess) {
-        // if error we "need" to know why...
-        CUDA_CHECK(hipMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device));
-    }
-    return res;
-#else
-
-#if !defined(GGML_USE_HIP)
     cudaError_t err;
     if (getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr)
     {
         err = cudaMallocManaged(ptr, size);
+#if defined(GGML_USE_HIP)
+        if (err == hipSuccess) {
+            CUDA_CHECK(cudaMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device));
+        }
+
+        // fall back to cudaMalloc if not supported (e.g. on Windows)
+        if (err == hipErrorNotSupported) {
+            static bool warned_unsupported = false;
+            if (!warned_unsupported) {
+                GGML_LOG_WARN("hipMallocManaged unsupported, falling back to hipMalloc.\n");
+                warned_unsupported = true;
+            }
+
+            err = cudaMalloc(ptr, size);
+        }
+#endif // defined(GGML_USE_HIP)
     }
     else
     {
         err = cudaMalloc(ptr, size);
     }
     return err;
-#else
-    return cudaMalloc(ptr, size);
-#endif // !defined(GGML_USE_HIP)
-
-#endif
 }
 
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
@@ -2488,7 +2489,7 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
 #endif
         }
 
-        if (node->op == GGML_OP_MUL_MAT_ID || node->op == GGML_OP_CONT || node->op == GGML_OP_DUP) {
+        if (node->op == GGML_OP_MUL_MAT_ID) {
             use_cuda_graph = false; // This node type is not supported by CUDA graph capture
 #ifndef NDEBUG
             GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported node type\n", __func__);
@@ -3236,6 +3237,10 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             if (op->src[0]->ne[0] == 192) {
                 return false;
             }
+            if (op->src[0]->ne[0] == 576) {
+                // DeepSeek MLA
+                return false;
+            }
             if (op->src[0]->ne[3] != 1) {
                 return false;
             }

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

@@ -71,6 +71,8 @@
 #define cudaLaunchHostFunc hipLaunchHostFunc
 #define cudaMalloc hipMalloc
 #define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
+#define cudaMallocManaged hipMallocManaged
+#define cudaMemAdvise hipMemAdvise
 #define cudaMemcpy hipMemcpy
 #define cudaMemcpyAsync hipMemcpyAsync
 #define cudaMemcpyPeerAsync hipMemcpyPeerAsync

ggml/src/ggml-hip/CMakeLists.txt

@@ -89,10 +89,6 @@ endif()
 
 add_compile_definitions(GGML_USE_HIP)
 
-if (GGML_HIP_UMA)
-    add_compile_definitions(GGML_HIP_UMA)
-endif()
-
 if (GGML_CUDA_FORCE_MMQ)
     add_compile_definitions(GGML_CUDA_FORCE_MMQ)
 endif()

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

@@ -354,6 +354,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H192,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H96,
@@ -362,6 +363,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H192,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H96,
@@ -370,6 +372,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H192,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H96,
@@ -378,6 +381,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H192,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H96,
@@ -386,6 +390,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H192,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H96,
@@ -394,6 +399,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H192,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H96,
@@ -402,6 +408,14 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H192,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H96,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H128,
@@ -430,6 +444,13 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H256,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H256,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_SET_I32,
     GGML_METAL_KERNEL_TYPE_SET_F32,
     GGML_METAL_KERNEL_TYPE_CPY_F32_F32,
@@ -460,6 +481,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_SQRT,
     GGML_METAL_KERNEL_TYPE_SIN,
     GGML_METAL_KERNEL_TYPE_COS,
+    GGML_METAL_KERNEL_TYPE_NEG,
     GGML_METAL_KERNEL_TYPE_SUM_ROWS,
     GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,
     GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,
@@ -1011,6 +1033,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H192,         flash_attn_ext_f16_h192,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK192_HV128,  flash_attn_ext_f16_hk192_hv128,  has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,         flash_attn_ext_f16_h256,         has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK576_HV512,  flash_attn_ext_f16_hk576_hv512,  has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H64,         flash_attn_ext_bf16_h64,         has_simdgroup_mm && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H80,         flash_attn_ext_bf16_h80,         has_simdgroup_mm && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H96,         flash_attn_ext_bf16_h96,         has_simdgroup_mm && use_bfloat);
@@ -1019,6 +1042,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H192,        flash_attn_ext_bf16_h192,        has_simdgroup_mm && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK192_HV128, flash_attn_ext_bf16_hk192_hv128, has_simdgroup_mm && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H256,        flash_attn_ext_bf16_h256,        has_simdgroup_mm && use_bfloat);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK576_HV512, flash_attn_ext_bf16_hk576_hv512, has_simdgroup_mm && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H64,         flash_attn_ext_q4_0_h64,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H80,         flash_attn_ext_q4_0_h80,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H96,         flash_attn_ext_q4_0_h96,         has_simdgroup_mm);
@@ -1027,6 +1051,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H192,        flash_attn_ext_q4_0_h192,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK192_HV128, flash_attn_ext_q4_0_hk192_hv128, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H256,        flash_attn_ext_q4_0_h256,        has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK576_HV512, flash_attn_ext_q4_0_hk576_hv512, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H64,         flash_attn_ext_q4_1_h64,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H80,         flash_attn_ext_q4_1_h80,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H96,         flash_attn_ext_q4_1_h96,         has_simdgroup_mm);
@@ -1035,6 +1060,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H192,        flash_attn_ext_q4_1_h192,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK192_HV128, flash_attn_ext_q4_1_hk192_hv128, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H256,        flash_attn_ext_q4_1_h256,        has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK576_HV512, flash_attn_ext_q4_1_hk576_hv512, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H64,         flash_attn_ext_q5_0_h64,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H80,         flash_attn_ext_q5_0_h80,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H96,         flash_attn_ext_q5_0_h96,         has_simdgroup_mm);
@@ -1043,6 +1069,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H192,        flash_attn_ext_q5_0_h192,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK192_HV128, flash_attn_ext_q5_0_hk192_hv128, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H256,        flash_attn_ext_q5_0_h256,        has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK576_HV512, flash_attn_ext_q5_0_hk576_hv512, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H64,         flash_attn_ext_q5_1_h64,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H80,         flash_attn_ext_q5_1_h80,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H96,         flash_attn_ext_q5_1_h96,         has_simdgroup_mm);
@@ -1051,6 +1078,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H192,        flash_attn_ext_q5_1_h192,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK192_HV128, flash_attn_ext_q5_1_hk192_hv128, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H256,        flash_attn_ext_q5_1_h256,        has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK576_HV512, flash_attn_ext_q5_1_hk576_hv512, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H64,         flash_attn_ext_q8_0_h64,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H80,         flash_attn_ext_q8_0_h80,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H96,         flash_attn_ext_q8_0_h96,         has_simdgroup_mm);
@@ -1059,6 +1087,14 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H192,        flash_attn_ext_q8_0_h192,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128, flash_attn_ext_q8_0_hk192_hv128, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256,        flash_attn_ext_q8_0_h256,        has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK576_HV512, flash_attn_ext_q8_0_hk576_hv512, has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H96,      flash_attn_ext_vec_f16_h96,      has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H96,     flash_attn_ext_vec_bf16_h96,     has_simdgroup_reduction && use_bfloat);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H96,     flash_attn_ext_vec_q4_0_h96,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H96,     flash_attn_ext_vec_q4_1_h96,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H96,     flash_attn_ext_vec_q5_0_h96,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H96,     flash_attn_ext_vec_q5_1_h96,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H96,     flash_attn_ext_vec_q8_0_h96,     has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128,     flash_attn_ext_vec_f16_h128,     has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H128,    flash_attn_ext_vec_bf16_h128,    has_simdgroup_reduction && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H128,    flash_attn_ext_vec_q4_0_h128,    has_simdgroup_reduction);
@@ -1087,6 +1123,13 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H256,    flash_attn_ext_vec_q5_0_h256,    has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H256,    flash_attn_ext_vec_q5_1_h256,    has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H256,    flash_attn_ext_vec_q8_0_h256,    has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_HK576_HV512,     flash_attn_ext_vec_f16_hk576_hv512,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_HK576_HV512,    flash_attn_ext_vec_bf16_hk576_hv512,    has_simdgroup_reduction && use_bfloat);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_HK576_HV512,    flash_attn_ext_vec_q4_0_hk576_hv512,    has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_HK576_HV512,    flash_attn_ext_vec_q4_1_hk576_hv512,    has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_HK576_HV512,    flash_attn_ext_vec_q5_0_hk576_hv512,    has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_HK576_HV512,    flash_attn_ext_vec_q5_1_hk576_hv512,    has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_HK576_HV512,    flash_attn_ext_vec_q8_0_hk576_hv512,    has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_F32,                         set_f32,                         true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_I32,                         set_i32,                         true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32,                     cpy_f32_f32,                     true);
@@ -1117,6 +1160,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQRT,                            sqrt,                            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIN,                             sin,                             true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_COS,                             cos,                             true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NEG,                             neg,                             true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                        sum_rows,                        true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGMAX,                          argmax,                          true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,                 pool_2d_avg_f32,                 true);
@@ -1278,6 +1322,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_ELU:
+                case GGML_UNARY_OP_NEG:
                     return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
                 default:
                     return false;
@@ -1351,6 +1396,11 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
                 // TODO: not sure if it is worth adding kernels for this size
                 return false;
             }
+            if (op->src[0]->ne[0] == 576) {
+                // DeepSeek sizes
+                // TODO: disabled for now, until optmized
+                return false;
+            }
             if (op->src[1]->type != op->src[2]->type) {
                 return false;
             }
@@ -1963,6 +2013,18 @@ static void ggml_metal_encode_node(
 
                     [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                 } break;
+                case GGML_UNARY_OP_NEG:
+                {
+                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_NEG].pipeline;
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
                 default:
                 {
                     GGML_LOG_WARN("%s: node %3d, op = %8s not implemented\n", __func__, idx, ggml_op_name(dst->op));
@@ -3843,12 +3905,14 @@ static void ggml_metal_encode_node(
                 // TODO: add vec kernels for (ne00%64 == 0) and maybe also for (ne00%32 == 0)
                 //       for now avoiding mainly to keep the number of templates/kernels a bit lower
                 //       these are now trivial to add after: https://github.com/ggml-org/llama.cpp/pull/12612
-                if (ne01 >= 4 || (ne00%128 != 0 && ne00 != 192)) {
+                if (ne01 >= 4 || (ne00%128 != 0 && ne00 != 96 && ne00 != 192 && ne00 != 576)) {
                     switch (src1->type) {
                         case GGML_TYPE_F16:
                             {
                                 if (ne00 == 192 && ne20 == 128) {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK192_HV128].pipeline;
+                                } else if (ne00 == 576 && ne20 == 512) {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64 ].pipeline; break;
@@ -3871,6 +3935,8 @@ static void ggml_metal_encode_node(
                             {
                                 if (ne00 == 192 && ne20 == 128) {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK192_HV128].pipeline;
+                                } else if (ne00 == 576 && ne20 == 512) {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H64 ].pipeline; break;
@@ -3893,6 +3959,8 @@ static void ggml_metal_encode_node(
                             {
                                 if (ne00 == 192 && ne20 == 128) {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK192_HV128].pipeline;
+                                } else if (ne00 == 576 && ne20 == 512) {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H64 ].pipeline; break;
@@ -3915,6 +3983,8 @@ static void ggml_metal_encode_node(
                             {
                                 if (ne00 == 192 && ne20 == 128) {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK192_HV128].pipeline;
+                                } else if (ne00 == 576 && ne20 == 512) {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H64 ].pipeline; break;
@@ -3937,6 +4007,8 @@ static void ggml_metal_encode_node(
                             {
                                 if (ne00 == 192 && ne20 == 128) {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK192_HV128].pipeline;
+                                } else if (ne00 == 576 && ne20 == 512) {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H64 ].pipeline; break;
@@ -3959,6 +4031,8 @@ static void ggml_metal_encode_node(
                             {
                                 if (ne00 == 192 && ne20 == 128) {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK192_HV128].pipeline;
+                                } else if (ne00 == 576 && ne20 == 512) {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H64 ].pipeline; break;
@@ -3981,6 +4055,8 @@ static void ggml_metal_encode_node(
                             {
                                 if (ne00 == 192 && ne20 == 128) {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128].pipeline;
+                                } else if (ne00 == 576 && ne20 == 512) {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H64 ].pipeline; break;
@@ -4010,6 +4086,24 @@ static void ggml_metal_encode_node(
                     use_vec_kernel = true;
 
                     switch (ne00) {
+                        case 96:
+                            {
+                                switch (src1->type) {
+                                    case GGML_TYPE_F16:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H96].pipeline; break;
+                                    case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H96].pipeline; break;
+                                    case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H96].pipeline; break;
+                                    case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H96].pipeline; break;
+                                    case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H96].pipeline; break;
+                                    case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H96].pipeline; break;
+                                    case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H96].pipeline; break;
+                                    default:
+                                        {
+                                            GGML_LOG_ERROR("unsupported type: %d\n", src1->type);
+                                            GGML_LOG_ERROR("add template specialization for this type\n");
+                                            GGML_ABORT("add template specialization for this type");
+                                        }
+                                }
+                            } break;
                         case 128:
                             {
                                 switch (src1->type) {
@@ -4082,12 +4176,36 @@ static void ggml_metal_encode_node(
                                         }
                                 }
                             } break;
+                        case 576:
+                            {
+                                if (ne20 == 512) {
+                                    switch (src1->type) {
+                                        case GGML_TYPE_F16:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_HK576_HV512].pipeline; break;
+                                        case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_HK576_HV512].pipeline; break;
+                                        case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_HK576_HV512].pipeline; break;
+                                        case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_HK576_HV512].pipeline; break;
+                                        case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_HK576_HV512].pipeline; break;
+                                        case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_HK576_HV512].pipeline; break;
+                                        case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_HK576_HV512].pipeline; break;
+                                        default:
+                                            {
+                                                GGML_LOG_ERROR("unsupported type: %d\n", src1->type);
+                                                GGML_LOG_ERROR("add template specialization for this type\n");
+                                                GGML_ABORT("add template specialization for this type");
+                                            }
+                                    }
+                                } else {
+                                    GGML_LOG_ERROR("unsupported size: %lld\n", ne20);
+                                    GGML_LOG_ERROR("add template specialization for this size\n");
+                                    GGML_ABORT("add template specialization for this size");
+                                }
+                            } break;
                         default:
-                                  {
-                                      GGML_LOG_ERROR("unsupported size: %lld\n", ne00);
-                                      GGML_LOG_ERROR("add template specialization for this size\n");
-                                      GGML_ABORT("add template specialization for this size");
-                                  }
+                            {
+                                GGML_LOG_ERROR("unsupported size: %lld\n", ne00);
+                                GGML_LOG_ERROR("add template specialization for this size\n");
+                                GGML_ABORT("add template specialization for this size");
+                            }
                     }
                 }
 

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

@@ -949,6 +949,13 @@ kernel void kernel_cos(
     dst[tpig] = cos(src0[tpig]);
 }
 
+kernel void kernel_neg(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = -src0[tpig];
+}
+
 kernel void kernel_sum_rows(
         device const float * src0,
         device       float * dst,
@@ -3546,6 +3553,7 @@ template [[host_name("kernel_flash_attn_ext_f16_h128")]]         kernel flash_at
 template [[host_name("kernel_flash_attn_ext_f16_h192")]]         kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  192, 192>;
 template [[host_name("kernel_flash_attn_ext_f16_hk192_hv128")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  192, 128>;
 template [[host_name("kernel_flash_attn_ext_f16_h256")]]         kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  256, 256>;
+template [[host_name("kernel_flash_attn_ext_f16_hk576_hv512")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  576, 512>;
 
 #if defined(GGML_METAL_USE_BF16)
 template [[host_name("kernel_flash_attn_ext_bf16_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 64,  64>;
@@ -3556,6 +3564,7 @@ template [[host_name("kernel_flash_attn_ext_bf16_h128")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_bf16_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_bf16_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_bf16_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_bf16_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 576, 512>;
 #endif
 
 template [[host_name("kernel_flash_attn_ext_q4_0_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 64,  64>;
@@ -3566,6 +3575,7 @@ template [[host_name("kernel_flash_attn_ext_q4_0_h128")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q4_0_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_q4_0_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q4_0_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_q4_0_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 576, 512>;
 
 template [[host_name("kernel_flash_attn_ext_q4_1_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_q4_1_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 80,  80>;
@@ -3575,6 +3585,7 @@ template [[host_name("kernel_flash_attn_ext_q4_1_h128")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q4_1_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_q4_1_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q4_1_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_q4_1_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 576, 512>;
 
 template [[host_name("kernel_flash_attn_ext_q5_0_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_q5_0_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 80,  80>;
@@ -3584,6 +3595,7 @@ template [[host_name("kernel_flash_attn_ext_q5_0_h128")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q5_0_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_q5_0_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q5_0_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_q5_0_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 576, 512>;
 
 template [[host_name("kernel_flash_attn_ext_q5_1_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_q5_1_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 80,  80>;
@@ -3593,6 +3605,7 @@ template [[host_name("kernel_flash_attn_ext_q5_1_h128")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q5_1_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_q5_1_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q5_1_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_q5_1_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 576, 512>;
 
 template [[host_name("kernel_flash_attn_ext_q8_0_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_q8_0_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 80,  80>;
@@ -3602,6 +3615,7 @@ template [[host_name("kernel_flash_attn_ext_q8_0_h128")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q8_0_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_q8_0_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q8_0_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_q8_0_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 576, 512>;
 
 #undef FA_TYPES
 
@@ -3959,6 +3973,16 @@ kernel void kernel_flash_attn_ext_vec(
 
 typedef decltype(kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 128, 128, 4>) flash_attn_ext_vec_t;
 
+template [[host_name("kernel_flash_attn_ext_vec_f16_h96")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,             1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  96, 96, 4>;
+#if defined(GGML_METAL_USE_BF16)
+template [[host_name("kernel_flash_attn_ext_vec_bf16_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,           1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 96, 96, 4>;
+#endif
+template [[host_name("kernel_flash_attn_ext_vec_q4_0_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0,        8, dequantize_q4_0_t4, block_q4_0,  8, dequantize_q4_0_t4, 96, 96, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q4_1_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_1,        8, dequantize_q4_1_t4, block_q4_1,  8, dequantize_q4_1_t4, 96, 96, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_0_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_0,        8, dequantize_q5_0_t4, block_q5_0,  8, dequantize_q5_0_t4, 96, 96, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_1_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1,        8, dequantize_q5_1_t4, block_q5_1,  8, dequantize_q5_1_t4, 96, 96, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q8_0_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0,        8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 96, 96, 4>;
+
 template [[host_name("kernel_flash_attn_ext_vec_f16_h128")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,             1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  128, 128, 4>;
 #if defined(GGML_METAL_USE_BF16)
 template [[host_name("kernel_flash_attn_ext_vec_bf16_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,           1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 128, 128, 4>;
@@ -3999,6 +4023,16 @@ template [[host_name("kernel_flash_attn_ext_vec_q5_0_h256")]] kernel flash_attn_
 template [[host_name("kernel_flash_attn_ext_vec_q5_1_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1,        8, dequantize_q5_1_t4, block_q5_1,  8, dequantize_q5_1_t4, 256, 256, 4>;
 template [[host_name("kernel_flash_attn_ext_vec_q8_0_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0,        8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 256, 256, 4>;
 
+template [[host_name("kernel_flash_attn_ext_vec_f16_hk576_hv512")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,      1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  576, 512, 2>;
+#if defined(GGML_METAL_USE_BF16)
+template [[host_name("kernel_flash_attn_ext_vec_bf16_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,    1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 576, 512, 2>;
+#endif
+template [[host_name("kernel_flash_attn_ext_vec_q4_0_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0, 8, dequantize_q4_0_t4, block_q4_0,  8, dequantize_q4_0_t4, 576, 512, 2>;
+template [[host_name("kernel_flash_attn_ext_vec_q4_1_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_1, 8, dequantize_q4_1_t4, block_q4_1,  8, dequantize_q4_1_t4, 576, 512, 2>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_0_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_0, 8, dequantize_q5_0_t4, block_q5_0,  8, dequantize_q5_0_t4, 576, 512, 2>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_1_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1, 8, dequantize_q5_1_t4, block_q5_1,  8, dequantize_q5_1_t4, 576, 512, 2>;
+template [[host_name("kernel_flash_attn_ext_vec_q8_0_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0, 8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 576, 512, 2>;
+
 #undef FA_TYPES
 
 template<typename T>

ggml/src/ggml-opencl/CMakeLists.txt

@@ -54,16 +54,41 @@ function(ggml_opencl_add_kernel KNAME)
 endfunction()
 
 set(GGML_OPENCL_KERNELS
-    ggml-opencl
-    ggml-opencl_mm
-    ggml-opencl_cvt
-    ggml-opencl_gemv_noshuffle
-    ggml-opencl_gemv_noshuffle_general
-    ggml-opencl_mul_mat_Ab_Bi_8x4
-    ggml-opencl_transpose_16
-    ggml-opencl_transpose_32
-    ggml-opencl_transpose_32_16
-    ggml-opencl_im2col
+    add
+    clamp
+    cpy
+    cvt
+    diag_mask_inf
+    gelu
+    gemv_noshuffle_general
+    gemv_noshuffle
+    get_rows
+    im2col_f32
+    im2col_f16
+    mul_mat_Ab_Bi_8x4
+    mul_mv_f16_f16
+    mul_mv_f16_f32_1row
+    mul_mv_f16_f32_l4
+    mul_mv_f16_f32
+    mul_mv_f32_f32
+    mul_mv_q4_0_f32
+    mul_mv_q4_0_f32_v
+    mul_mv_q4_0_f32_8x_flat
+    mul_mv_q4_0_f32_1d_8x_flat
+    mul_mv_q4_0_f32_1d_16x_flat
+    mul_mv_q6_k
+    mul
+    norm
+    relu
+    rms_norm
+    rope
+    scale
+    silu
+    softmax_4_f32
+    softmax_4_f16
+    softmax_f32
+    softmax_f16
+    transpose
 )
 
 foreach (K ${GGML_OPENCL_KERNELS})

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

@@ -0,0 +1,83 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// add
+//------------------------------------------------------------------------------
+
+// general-purpose kernel for addition of two tensors
+// pros: works for non-contiguous tensors, supports broadcast across dims 1, 2 and 3
+// cons: not very efficient
+kernel void kernel_add(
+        global char * src0,
+        ulong  offset0,
+        global char * src1,
+        ulong  offset1,
+        global char * dst,
+        ulong  offsetd,
+        int   ne00,
+        int   ne01,
+        int   ne02,
+        int   ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int   ne10,
+        int   ne11,
+        int   ne12,
+        int   ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int   ne0,
+        int   ne1,
+        int   ne2,
+        int   ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst = dst + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03 % ne13;
+    int i12 = i02 % ne12;
+    int i11 = i01 % ne11;
+
+    global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    global char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i10 = i0 % ne10;
+        *((global float *)(dst_ptr + i0*nb0)) = *((global float *)(src0_ptr + i0*nb00)) + *((global float *)(src1_ptr + i10*nb10));
+    }
+}
+
+// assumption: src1 is a row
+// broadcast src1 into src0
+kernel void kernel_add_row(
+        global float4 * src0,
+        ulong  offset0,
+        global float4 * src1,
+        ulong  offset1,
+        global float4 * dst,
+        ulong  offsetd,
+        int ne
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    // This performs better than using %.
+    uint gid = get_global_id(0);
+    uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
+    dst[gid] = src0[gid] + src1[idx1];
+}

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

@@ -0,0 +1,20 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// clamp
+//------------------------------------------------------------------------------
+kernel void kernel_clamp(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        float min,
+        float max
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = src0[get_global_id(0)] < min ?
+        min :
+        (src0[get_global_id(0)] > max ? max : src0[get_global_id(0)]);
+}

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

@@ -0,0 +1,184 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// cpy
+//------------------------------------------------------------------------------
+
+kernel void kernel_cpy_f16_f16(
+        global half * src0,
+        ulong offset0,
+        global half * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = (global half*)((global char*)src0 + offset0);
+    dst = (global half*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    int i3 = n / (ne2*ne1*ne0);
+    int i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    int i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    int i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    global half * dst_data = (global half *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        global const half * src = (global half *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+        dst_data[i00] = src[0];
+    }
+}
+
+kernel void kernel_cpy_f16_f32(
+        global half * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+
+    src0 = (global half*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    int i3 = n / (ne2*ne1*ne0);
+    int i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    int i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    int i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    global float * dst_data = (global float *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        global half * src = (global half *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+        dst_data[i00] = src[0];
+    }
+}
+
+kernel void kernel_cpy_f32_f16(
+        global float * src0,
+        ulong offset0,
+        global half * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global half*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    int i3 = n / (ne2*ne1*ne0);
+    int i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    int i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    int i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    global half * dst_data = (global half *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        global const float * src = (global float *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        dst_data[i00] = src[0];
+    }
+}
+
+kernel void kernel_cpy_f32_f32(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    int i3 = n / (ne2*ne1*ne0);
+    int i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    int i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    int i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    global float * dst_data = (global float *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        global const float * src = (global float *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        dst_data[i00] = src[0];
+    }
+}

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

@@ -1,39 +1,20 @@
 //------------------------------------------------------------------------------
-// This file is contains additional kernels for data conversion.
+// This file is contains kernels for data conversion.
 // These kernels are used when loading the model, so its performance is less
 // important.
 //------------------------------------------------------------------------------
-#ifdef cl_khr_fp16
 #pragma OPENCL EXTENSION cl_khr_fp16 : enable
-#elif defined(cl_amd_fp16)
-#pragma OPENCL EXTENSION cl_amd_fp16 : enable
-#else
-#error "Half precision floating point not supportedby OpenCL implementation on your device."
-#endif
-
-#ifdef cl_khr_subgroups
-#pragma OPENCL EXTENSION cl_khr_subgroups : enable
-#elif defined(cl_intel_subgroups)
-#pragma OPENCL EXTENSION cl_intel_subgroups : enable
-#else
-#error "Subgroup not supported on your device."
-#endif
 
 #ifdef cl_intel_required_subgroup_size
-// Always use subgroup size of 32 on Intel.
 #pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
 #define INTEL_GPU 1
 #define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
 #define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
 #elif defined(cl_qcom_reqd_sub_group_size)
-// Always use subgroups size of 64 on Adreno.
 #pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
 #define ADRENO_GPU 1
 #define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
 #define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
-#else
-// TODO: do not know how to choose subgroup size on other GPUs.
-#error "Selecting subgroup size is not supported on your device."
 #endif
 
 #define QK4_0                   32
@@ -66,13 +47,44 @@ struct block_q4_0
 };
 
 //------------------------------------------------------------------------------
-// mul_vec_q_n_f32_flat_noshuffle
-//
-// This variation uses flat arrays (struct of arrays, SOA) representation for
-// quant tensors. It also uses non shuffled bit order for weights.
-//
-// The shuffled version is kept in the original file because moving it here
-// seems to result in worse performance for adreno.
+// kernel_convert_block_q4_0
+// Convert the block_q4_0 format to 2 separate arrays (AOS -> SOA).
+// This kernel does not deshuffle the bits.
+//------------------------------------------------------------------------------
+kernel void kernel_convert_block_q4_0(
+    global struct block_q4_0 * src0,
+    global uchar * dst_q,
+    global half  * dst_d
+) {
+    global struct block_q4_0 * b = (global struct block_q4_0 *) src0 + get_global_id(0);
+    global uchar * q = (global uchar *) dst_q + QK4_0/2*get_global_id(0);
+    global half  * d = (global half *) dst_d + get_global_id(0);
+
+    *d = b->d;
+
+    for (int i = 0; i < QK4_0/2; ++i) {
+        q[i] = b->qs[i];
+    }
+}
+
+kernel void kernel_restore_block_q4_0(
+    global uchar * src_q,
+    global half  * src_d,
+    global struct block_q4_0 * dst
+) {
+    global struct block_q4_0 * b = (global struct block_q4_0 *) dst + get_global_id(0);
+    global uchar * q = (global uchar *) src_q + QK4_0/2*get_global_id(0);
+    global half  * d = (global half *) src_d + get_global_id(0);
+
+    b->d = *d;
+    for (int i = 0; i < QK4_0/2; ++i) {
+        b->qs[i] = q[i];
+    }
+}
+
+//------------------------------------------------------------------------------
+// kernel_convert_block_q4_0_noshuffle
+// Flatten q4_0 weights and unshuffle the bits
 //------------------------------------------------------------------------------
 
 kernel void kernel_convert_block_q4_0_noshuffle(

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

@@ -0,0 +1,58 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// diag_mask_inf kernels
+//------------------------------------------------------------------------------
+kernel void kernel_diag_mask_inf(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int n_past
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i02 = get_global_id(2);
+    int i01 = get_global_id(1);
+    int i00 = get_global_id(0);
+
+    if (i00 > n_past + i01) {
+        dst[i02*ne01*ne00 + i01*ne00 + i00] = -INFINITY;
+    } else {
+        dst[i02*ne01*ne00 + i01*ne00 + i00] = src0[i02*ne01*ne00 + i01*ne00 + i00];
+    }
+}
+
+kernel void kernel_diag_mask_inf_8(
+        global float4 * src0,
+        ulong offset0,
+        global float4 * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int n_past
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    int i = 2*get_global_id(0);
+
+    dst[i+0] = src0[i+0];
+    dst[i+1] = src0[i+1];
+    int i4 = 4*i;
+    int i02 = i4/(ne00*ne01); i4 -= i02*ne00*ne01;
+    int i01 = i4/(ne00);      i4 -= i01*ne00;
+    int i00 = i4;
+    for (int k = 3; k >= 0; --k) {
+        if (i00 + 4 + k <= n_past + i01) {
+            break;
+        }
+        (&dst[i+1])[k] = -INFINITY;
+        if (i00 + k > n_past + i01) {
+            (&dst[i])[k] = -INFINITY;
+        }
+    }
+}

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

@@ -0,0 +1,62 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// gelu
+//------------------------------------------------------------------------------
+#define GELU_COEF_A     0.044715f
+#define GELU_QUICK_COEF -1.702f
+#define SQRT_2_OVER_PI  0.79788456080286535587989211986876f
+
+kernel void kernel_gelu(
+    global float * src0,
+    ulong offset0,
+    global float * dst,
+    ulong offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    float x = src0[get_global_id(0)];
+
+    dst[get_global_id(0)] = 0.5f*x*(1.0f + tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
+}
+
+kernel void kernel_gelu_4(
+    global float4 * src0,
+    ulong offset0,
+    global float4 * dst,
+    ulong offsetd
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    float4 x = src0[get_global_id(0)];
+
+    dst[get_global_id(0)] = 0.5f*x*(1.0f + tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
+}
+
+kernel void kernel_gelu_quick(
+    global float * src0,
+    ulong offset0,
+    global float * dst,
+    ulong offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    float x = src0[get_global_id(0)];
+    dst[get_global_id(0)] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x)));
+}
+
+kernel void kernel_gelu_quick_4(
+    global float4 * src0,
+    ulong offset0,
+    global float4 * dst,
+    ulong offsetd
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    float4 x = src0[get_global_id(0)];
+    dst[get_global_id(0)] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x)));
+}

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

@@ -0,0 +1,163 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+#define QK4_0                   32
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+
+//------------------------------------------------------------------------------
+// dequantize_q4_0_f32, dequantize_q4_0_f16
+//------------------------------------------------------------------------------
+void dequantize_q4_0_f32(global struct block_q4_0 * xb, short il, float16 * reg) {
+    global ushort * qs = ((global ushort *)xb + 1);
+    float d1 = il ? (xb->d / 16.h) : xb->d;
+    float d2 = d1 / 256.f;
+    float md = -8.h * xb->d;
+    ushort mask0 = il ? 0x00F0 : 0x000F;
+    ushort mask1 = mask0 << 8;
+
+    reg->s0 = d1 * (qs[0] & mask0) + md;
+    reg->s1 = d2 * (qs[0] & mask1) + md;
+
+    reg->s2 = d1 * (qs[1] & mask0) + md;
+    reg->s3 = d2 * (qs[1] & mask1) + md;
+
+    reg->s4 = d1 * (qs[2] & mask0) + md;
+    reg->s5 = d2 * (qs[2] & mask1) + md;
+
+    reg->s6 = d1 * (qs[3] & mask0) + md;
+    reg->s7 = d2 * (qs[3] & mask1) + md;
+
+    reg->s8 = d1 * (qs[4] & mask0) + md;
+    reg->s9 = d2 * (qs[4] & mask1) + md;
+
+    reg->sa = d1 * (qs[5] & mask0) + md;
+    reg->sb = d2 * (qs[5] & mask1) + md;
+
+    reg->sc = d1 * (qs[6] & mask0) + md;
+    reg->sd = d2 * (qs[6] & mask1) + md;
+
+    reg->se = d1 * (qs[7] & mask0) + md;
+    reg->sf = d2 * (qs[7] & mask1) + md;
+}
+
+
+//------------------------------------------------------------------------------
+// get_rows
+//------------------------------------------------------------------------------
+kernel void kernel_get_rows_f32(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        ulong nb01,
+        ulong nb02,
+        int ne10,
+        ulong nb10,
+        ulong nb11,
+        ulong nb1,
+        ulong nb2
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i10 = get_group_id(0);
+    int i11 = get_group_id(1);
+
+    int r = ((global int *) ((global char *) src1 + i11*nb11 + i10*nb10))[0];
+
+    int i02 = i11;
+
+    for (int ind = get_local_id(0); ind < ne00; ind += get_local_size(0)) {
+        ((global float *) ((global char *) dst + i11*nb2 + i10*nb1))[ind] =
+            ((global float *) ((global char *) src0 + r*nb01 + i02*nb02))[ind];
+    }
+}
+
+kernel void kernel_get_rows_f16(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        ulong nb01,
+        ulong nb02,
+        int ne10,
+        ulong nb10,
+        ulong nb11,
+        ulong nb1,
+        ulong nb2
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i10 = get_group_id(0);
+    int i11 = get_group_id(1);
+
+    int r = ((global int32_t *) ((global char *) src1 + i11*nb11 + i10*nb10))[0];
+
+    int i02 = i11;
+
+    for (int ind = get_local_id(0); ind < ne00; ind += get_local_size(0)) {
+        ((global float *) ((global char *) dst + i11*nb2 + i10*nb1))[ind] =
+            ((global half *) ((global char *) src0 + r*nb01 + i02*nb02))[ind];
+    }
+}
+
+kernel void kernel_get_rows_q4_0(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        ulong nb01,
+        ulong nb02,
+        int ne10,
+        ulong nb10,
+        ulong nb11,
+        ulong nb1,
+        ulong nb2
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    const int NL = 2;
+
+    int i10 = get_group_id(0);
+    int i11 = get_group_id(1);
+
+    int r = ((global int32_t *) ((global char *) src1 + i11*nb11 + i10*nb10))[0];
+
+    int i02 = i11;
+
+    for (int ind = get_local_id(0); ind < ne00/16; ind += get_local_size(0)) {
+        float16 temp;
+        dequantize_q4_0_f32(
+            ((global struct block_q4_0 *) ((global char *) src0 + r*nb01 + i02*nb02)) + ind/NL, ind%NL, &temp);
+        *(((global float16 *) ((global char *) dst + i11*nb2 + i10*nb1)) + ind) = temp;
+    }
+}

ggml/src/ggml-opencl/kernels/ggml-opencl_im2col.cl

@@ -1,146 +0,0 @@
-#ifdef cl_khr_fp16
-#pragma OPENCL EXTENSION cl_khr_fp16 : enable
-#elif defined(cl_amd_fp16)
-#pragma OPENCL EXTENSION cl_amd_fp16 : enable
-#else
-#error "Half precision floating point not supportedby OpenCL implementation on your device."
-#endif
-
-#ifdef cl_khr_subgroups
-#pragma OPENCL EXTENSION cl_khr_subgroups : enable
-#elif defined(cl_intel_subgroups)
-#pragma OPENCL EXTENSION cl_intel_subgroups : enable
-#else
-#error "Subgroup not supported on your device."
-#endif
-
-#ifdef cl_intel_required_subgroup_size
-// Always use subgroup size of 32 on Intel.
-#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
-#define INTEL_GPU 1
-#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
-#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
-#elif defined(cl_qcom_reqd_sub_group_size)
-// Always use subgroups size of 64 on Adreno.
-#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
-#define ADRENO_GPU 1
-#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
-#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
-#else
-// TODO: do not know how to choose subgroup size on other GPUs.
-#error "Selecting subgroup size is not supported on your device."
-#endif
-
-kernel void kernel_im2col_f32(
-        global float * src1,
-        ulong offset1,
-        global float * dst,
-        ulong offsetd,
-        ulong batch_offset,
-        ulong delta_offset,
-        long IW,
-        long IH,
-        long IC,
-        long OW,
-        long OH,
-        long KW,
-        long KH,
-        long pelements,
-        long CHW,
-        int  s0,
-        int  s1,
-        int  p0,
-        int  p1,
-        int  d0,
-        int  d1
-) {
-    // threadIdx.x + blockIdx.x * blockDim.x
-    long i = get_global_id(0);
-    if (i >= pelements) {
-        return;
-    }
-
-    src1 = (global float*)((global char*)src1 + offset1);
-    dst = (global float*)((global char*)dst + offsetd);
-
-    long  ksize = OW * (KH > 1 ? KW : 1);
-    long  kx = i / ksize;
-    long  kd = kx * ksize;
-    long  ky = (i - kd) / OW;
-    long  ix = i % OW;
-
-    long  oh = get_group_id(1);
-    long  batch = get_group_id(2) / IC;
-    long  ic = get_group_id(2) % IC;
-
-    long iiw = ix * s0 + kx * d0 - p0;
-    long iih = oh * s1 + ky * d1 - p1;
-
-    long offset_dst =
-        ((batch * OH + oh) * OW + ix) * CHW +
-        (ic * (KW * KH) + ky * KW + kx);
-
-    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
-        dst[offset_dst] = 0.0f;
-    } else {
-        long offset_src = ic * delta_offset + batch * batch_offset;
-        dst[offset_dst] = src1[offset_src + iih * IW + iiw];
-    }
-}
-
-kernel void kernel_im2col_f16(
-        global float * src1,
-        ulong offset1,
-        global half  * dst,
-        ulong offsetd,
-        ulong batch_offset,
-        ulong delta_offset,
-        long IW,
-        long IH,
-        long IC,
-        long OW,
-        long OH,
-        long KW,
-        long KH,
-        long pelements,
-        long CHW,
-        int  s0,
-        int  s1,
-        int  p0,
-        int  p1,
-        int  d0,
-        int  d1
-) {
-    long i = get_global_id(0);
-
-    if (i >= pelements) {
-        return;
-    }
-
-    src1 = (global float*)((global char*)src1 + offset1);
-    dst = (global half*)((global char*)dst + offsetd);
-
-    long  ksize = OW * (KH > 1 ? KW : 1);
-    long  kx = i / ksize;
-    long  kd = kx * ksize;
-    long  ky = (i - kd) / OW;
-    long  ix = i % OW;
-
-    long  oh = get_group_id(1);
-    long  batch = get_group_id(2) / IC;
-    long  ic = get_group_id(2) % IC;
-
-    long iiw = ix * s0 + kx * d0 - p0;
-    long iih = oh * s1 + ky * d1 - p1;
-
-    long offset_dst =
-        ((batch * OH + oh) * OW + ix) * CHW +
-        (ic * (KW * KH) + ky * KW + kx);
-
-    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
-        dst[offset_dst] = 0.0f;
-    } else {
-        long offset_src = ic * delta_offset + batch * batch_offset;
-        dst[offset_dst] = src1[offset_src + iih * IW + iiw];
-    }
-}

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

@@ -1,26 +0,0 @@
-// 16-bit transpose, loading/storing a 4x4 tile of elements
-
-#pragma OPENCL EXTENSION cl_khr_fp16 : enable
-
-kernel void kernel_transpose_16(
-    __read_only image1d_buffer_t input,
-    __write_only image1d_buffer_t output,
-    const uint rows,
-    const uint cols
-) {
-
-    const int i = get_global_id(0);
-    const int j = get_global_id(1);
-    const int i_2 = i<<2;
-    const int j_2 = j<<2;
-
-    half4 temp0 = read_imageh(input, (j_2+0)*cols+i);
-    half4 temp1 = read_imageh(input, (j_2+1)*cols+i);
-    half4 temp2 = read_imageh(input, (j_2+2)*cols+i);
-    half4 temp3 = read_imageh(input, (j_2+3)*cols+i);
-
-    write_imageh(output, (i_2+0)*rows+j, (half4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0));
-    write_imageh(output, (i_2+1)*rows+j, (half4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1));
-    write_imageh(output, (i_2+2)*rows+j, (half4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2));
-    write_imageh(output, (i_2+3)*rows+j, (half4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
-}

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

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

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

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

ggml/src/ggml-opencl/kernels/im2col_f16.cl

@@ -0,0 +1,57 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+kernel void kernel_im2col_f16(
+        global float * src1,
+        ulong offset1,
+        global half  * dst,
+        ulong offsetd,
+        ulong batch_offset,
+        ulong delta_offset,
+        long IW,
+        long IH,
+        long IC,
+        long OW,
+        long OH,
+        long KW,
+        long KH,
+        long pelements,
+        long CHW,
+        int  s0,
+        int  s1,
+        int  p0,
+        int  p1,
+        int  d0,
+        int  d1
+) {
+    long i = get_global_id(0);
+    if (i >= pelements) {
+        return;
+    }
+
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global half*)((global char*)dst + offsetd);
+
+    long  ksize = OW * (KH > 1 ? KW : 1);
+    long  kx = i / ksize;
+    long  kd = kx * ksize;
+    long  ky = (i - kd) / OW;
+    long  ix = i % OW;
+
+    long  oh = get_group_id(1);
+    long  batch = get_group_id(2) / IC;
+    long  ic = get_group_id(2) % IC;
+
+    long iiw = ix * s0 + kx * d0 - p0;
+    long iih = oh * s1 + ky * d1 - p1;
+
+    long offset_dst =
+        ((batch * OH + oh) * OW + ix) * CHW +
+        (ic * (KW * KH) + ky * KW + kx);
+
+    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+        dst[offset_dst] = 0.0f;
+    } else {
+        long offset_src = ic * delta_offset + batch * batch_offset;
+        dst[offset_dst] = src1[offset_src + iih * IW + iiw];
+    }
+}

ggml/src/ggml-opencl/kernels/im2col_f32.cl

@@ -0,0 +1,57 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+kernel void kernel_im2col_f32(
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        ulong batch_offset,
+        ulong delta_offset,
+        long IW,
+        long IH,
+        long IC,
+        long OW,
+        long OH,
+        long KW,
+        long KH,
+        long pelements,
+        long CHW,
+        int  s0,
+        int  s1,
+        int  p0,
+        int  p1,
+        int  d0,
+        int  d1
+) {
+    long i = get_global_id(0);
+    if (i >= pelements) {
+        return;
+    }
+
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    long  ksize = OW * (KH > 1 ? KW : 1);
+    long  kx = i / ksize;
+    long  kd = kx * ksize;
+    long  ky = (i - kd) / OW;
+    long  ix = i % OW;
+
+    long  oh = get_group_id(1);
+    long  batch = get_group_id(2) / IC;
+    long  ic = get_group_id(2) % IC;
+
+    long iiw = ix * s0 + kx * d0 - p0;
+    long iih = oh * s1 + ky * d1 - p1;
+
+    long offset_dst =
+        ((batch * OH + oh) * OW + ix) * CHW +
+        (ic * (KW * KH) + ky * KW + kx);
+
+    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+        dst[offset_dst] = 0.0f;
+    } else {
+        long offset_src = ic * delta_offset + batch * batch_offset;
+        dst[offset_dst] = src1[offset_src + iih * IW + iiw];
+    }
+}

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

@@ -0,0 +1,79 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// mul
+//------------------------------------------------------------------------------
+kernel void kernel_mul(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03 % ne13;
+    int i12 = i02 % ne12;
+    int i11 = i01 % ne11;
+
+    global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    global char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i10 = i0 % ne10;
+        *((global float *)(dst_ptr + i0*nb0)) = *((global float *)(src0_ptr + i0*nb00)) * *((global float *)(src1_ptr + i10*nb10));
+    }
+}
+
+// assumption: src1 is a row
+// broadcast src1 into src0
+kernel void kernel_mul_row(
+        global float4 * src0,
+        ulong offset0,
+        global float4 * src1,
+        ulong offset1,
+        global float4 * dst,
+        ulong offsetd,
+        int ne
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    // This performs better than using %.
+    uint gid = get_global_id(0);
+    uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
+    dst[gid] = src0[gid] * src1[idx1];
+}

ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl

@@ -0,0 +1,118 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define N_F16_F16 4
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f16(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3)
+{
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int r0 = get_group_id(0);
+    int rb = get_group_id(1)*N_F16_F16;
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+
+    global half * x = (global half *) (src0 + offset_src0);
+
+    if (ne00 < 128) {
+        for (int row = 0; row < N_F16_F16; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global half * y = (global half *) (src1 + offset_src1);
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00; i += get_max_sub_group_size()) {
+                sumf += (half) x[i] * (half) y[i];
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    } else {
+        global half4 * x4 = (global half4 *)x;
+        for (int row = 0; row < N_F16_F16; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global half  * y  = (global half  *) (src1 + offset_src1);
+            global half4 * y4 = (global half4 *) y;
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) {
+                sumf += (half) x4[i].s0 * y4[i].s0;
+                sumf += (half) x4[i].s1 * y4[i].s1;
+                sumf += (half) x4[i].s2 * y4[i].s2;
+                sumf += (half) x4[i].s3 * y4[i].s3;
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                for (int i = 4*(ne00/4); i < ne00; ++i) {
+                    all_sum += (half) x[i] * y[i];
+                }
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl

@@ -0,0 +1,118 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define N_F16_F32 4
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int r0 = get_group_id(0);
+    int rb = get_group_id(1)*N_F16_F32;
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+
+    global half * x = (global half *) (src0 + offset_src0);
+
+    if (ne00 < 128) {
+        for (int row = 0; row < N_F16_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global float * y = (global float *) (src1 + offset_src1);
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00; i += get_max_sub_group_size()) {
+                sumf += convert_float(x[i]) * y[i];
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    } else {
+        global half4 * x4 = (global half4 *)x;
+        for (int row = 0; row < N_F16_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global float  * y  = (global float  *) (src1 + offset_src1);
+            global float4 * y4 = (global float4 *) y;
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) {
+                sumf += convert_float(x4[i].s0) * y4[i].s0;
+                sumf += convert_float(x4[i].s1) * y4[i].s1;
+                sumf += convert_float(x4[i].s2) * y4[i].s2;
+                sumf += convert_float(x4[i].s3) * y4[i].s3;
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                for (int i = 4*(ne00/4); i < ne00; ++i) {
+                    all_sum += (float) x[i] * y[i];
+                }
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl

@@ -0,0 +1,94 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32_1row(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global half  * x = (global half  *) (src0 + offset_src0);
+    global float * y = (global float *) (src1 + offset_src1);
+
+    float sumf = 0;
+    if (ne00 < 128) {
+        for (int i = get_sub_group_local_id(); i < ne00; i += get_max_sub_group_size()) {
+            sumf += (float) x[i] * (float) y[i];
+        }
+        float all_sum = sub_group_reduce_add(sumf);
+        if (get_sub_group_local_id() == 0) {
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+        }
+    } else {
+        global half4  * x4 = (global half4  *) x;
+        global float4 * y4 = (global float4 *) y;
+        for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) {
+            sumf += (float) x4[i].s0 * y4[i].s0;
+            sumf += (float) x4[i].s1 * y4[i].s1;
+            sumf += (float) x4[i].s2 * y4[i].s2;
+            sumf += (float) x4[i].s3 * y4[i].s3;
+        }
+        float all_sum = sub_group_reduce_add(sumf);
+        if (get_sub_group_local_id() == 0) {
+            for (int i = 4*(ne00/4); i < ne00; ++i) {
+                all_sum += (float) x[i] * y[i];
+            }
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+        }
+    }
+
+}

ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl

@@ -0,0 +1,84 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+// Assumes row size (ne00) is a multiple of 4
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32_l4(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int nrows = ne11;
+    int r0 = get_group_id(0);
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+
+    global half4 * x4 = (global half4 *) (src0 + offset_src0);
+
+    for (int r1 = 0; r1 < nrows; ++r1) {
+        ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+        global float4 * y4 = (global float4 *) (src1 + offset_src1);
+
+        float sumf = 0;
+        for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) {
+            sumf += convert_float(x4[i].s0) * y4[i].s0;
+            sumf += convert_float(x4[i].s1) * y4[i].s1;
+            sumf += convert_float(x4[i].s2) * y4[i].s2;
+            sumf += convert_float(x4[i].s3) * y4[i].s3;
+        }
+
+        float all_sum = sub_group_reduce_add(sumf);
+        if (get_sub_group_local_id() == 0) {
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_f32_f32.cl

@@ -0,0 +1,118 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define N_F32_F32 4
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f32_f32(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int r0 = get_group_id(0);
+    int rb = get_group_id(1)*N_F32_F32;
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+
+    global float * x = (global float *) (src0 + offset_src0);
+
+    if (ne00 < 128) {
+        for (int row = 0; row < N_F32_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global float * y = (global float *) (src1 + offset_src1);
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00; i += get_max_sub_group_size()) {
+                sumf += (float) x[i] * (float) y[i];
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    } else {
+        global float4 * x4 = (global float4 *)x;
+        for (int row = 0; row < N_F32_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global float  * y  = (global float  *) (src1 + offset_src1);
+            global float4 * y4 = (global float4 *) y;
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) {
+                sumf += (float) x4[i].s0 * y4[i].s0;
+                sumf += (float) x4[i].s1 * y4[i].s1;
+                sumf += (float) x4[i].s2 * y4[i].s2;
+                sumf += (float) x4[i].s3 * y4[i].s3;
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                for (int i = 4*(ne00/4); i < ne00; ++i) {
+                    all_sum += (float) x[i] * y[i];
+                }
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32.cl

@@ -0,0 +1,192 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+#define QR4_0                   2
+#define QK4_1                   32
+#define QR4_1                   2
+#define QK5_0                   32
+#define QR5_0                   2
+#define QK5_1                   32
+#define QR5_1                   2
+#define QK8_0                   32
+#define QR8_0                   1
+#define QK_K                    256
+#define K_QUANTS_PER_ITERATION  2
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+//------------------------------------------------------------------------------
+// mul_vec_q_n_f32
+//------------------------------------------------------------------------------
+// function for calculate inner product between half a q4_0 block and 16 floats (yl), sumy is SUM(yl[i])
+// il indicates where the q4 quants begin (0 or QK4_0/4)
+// we assume that the yl's have been multiplied with the appropriate scale factor
+// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
+inline float block_q_4_0_dot_y(
+        global struct block_q4_0 * qb_curr,
+        float sumy,
+        private float * yl,
+        int il
+) {
+    float d = qb_curr->d;
+    float2 acc = 0.f;
+    global ushort * qs = ((global ushort *)qb_curr + 1 + il/2);
+    for (int i = 0; i < 8; i+=2) {
+        acc.s0 += yl[i + 0] * (qs[i / 2] & 0x000F)
+                + yl[i + 1] * (qs[i / 2] & 0x0F00);
+        acc.s1 += yl[i + 8] * (qs[i / 2] & 0x00F0)
+                + yl[i + 9] * (qs[i / 2] & 0xF000);
+    }
+    return d * (sumy * -8.f + acc.s0 + acc.s1);
+}
+
+#ifdef INTEL_GPU
+#define N_DST 4 // each SIMD group works on 4 rows
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // assuming SIMD group size is 16
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_vec_q_n_f32(
+        global void * src0,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+
+    const ulong nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    // (r0 * N_SIMDGROUP + get_sub_group_id()) is essenatially the linear global
+    // id of a SIMD group in the grid.
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset0 = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    global struct block_q4_0 * x = (global struct block_q4_0 *) src0 + offset0;
+    global float             * y = (global float             *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[16];       // src1 vector cache
+    float sumf[N_DST]={0.f};
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix * QK4_0 + il;
+
+    // each thread in a SIMD group deals with half a block.
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0;
+        for (int i = 0; i < 8; i += 2) {
+            sumy += yb[i] + yb[i+1];
+            yl[i+0] = yb[i+ 0];
+            yl[i+1] = yb[i+ 1]/256.f;
+            sumy += yb[i+16] + yb[i+17];
+            yl[i+8] = yb[i+16]/16.f;
+            yl[i+9] = yb[i+17]/4096.f;
+        }
+
+        for (int row = 0; row < N_DST; row++) {
+            sumf[row] += block_q_4_0_dot_y(x+ib+row*nb, sumy, yl, il);
+        }
+
+        // One thread in a SIMD group (i.e., subgroup) handles a half block,
+        // hence then entire SIMD group handles SIMDWIDTH/2 blocks.
+        // y points to the activation matrix (of type float). Therefore for
+        // one thread, the # of blocks y should advance is SIMDWIDTH/2 (because
+        // SIMDWIDTH/2 blocks are processed by a SIMD group) - in terms of
+        // floats, it is QK4_0 * (SIMDWIDTH/2), where QK4_0 is the block size.
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    // The above does not work for Adreno - it produces incorrect results for
+    // row = 1, 2, 3 and only row = 0 gives the correct result.
+    // If N_DST is changed, the below array must be initialized accordingly.
+    // This also seems to perform better on Intel.
+    float tot[N_DST] = {
+        sub_group_reduce_add(sumf[0]), sub_group_reduce_add(sumf[1]),
+        sub_group_reduce_add(sumf[2]), sub_group_reduce_add(sumf[3])};
+    for (int row = 0; row < N_DST; ++row) {
+        if (get_sub_group_local_id() == 0 && first_row + row < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot[row];
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_q4_0_f32(
+        global void * src0,
+        ulong offset0,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    mul_vec_q_n_f32(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}

ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_16x_flat.cl

@@ -0,0 +1,307 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+#define QR4_0                   2
+#define QK4_1                   32
+#define QR4_1                   2
+#define QK5_0                   32
+#define QR5_0                   2
+#define QK5_1                   32
+#define QR5_1                   2
+#define QK8_0                   32
+#define QR8_0                   1
+#define QK_K                    256
+#define K_QUANTS_PER_ITERATION  2
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+inline float mm_block_q_4_0_dot_y_flat(
+        global uchar * x,
+        global half  * dh,
+        float sumy,
+        float16 yl,
+        int il
+) {
+    float           d   = *dh;
+    global ushort * qs  = ((global ushort *)x + il/2);
+    float           acc = 0.f;
+
+    acc += yl.s0 * (qs[0] & 0x000F);
+    acc += yl.s1 * (qs[0] & 0x0F00);
+    acc += yl.s8 * (qs[0] & 0x00F0);
+    acc += yl.s9 * (qs[0] & 0xF000);
+
+    acc += yl.s2 * (qs[1] & 0x000F);
+    acc += yl.s3 * (qs[1] & 0x0F00);
+    acc += yl.sa * (qs[1] & 0x00F0);
+    acc += yl.sb * (qs[1] & 0xF000);
+
+    acc += yl.s4 * (qs[2] & 0x000F);
+    acc += yl.s5 * (qs[2] & 0x0F00);
+    acc += yl.sc * (qs[2] & 0x00F0);
+    acc += yl.sd * (qs[2] & 0xF000);
+
+    acc += yl.s6 * (qs[3] & 0x000F);
+    acc += yl.s7 * (qs[3] & 0x0F00);
+    acc += yl.se * (qs[3] & 0x00F0);
+    acc += yl.sf * (qs[3] & 0xF000);
+
+    return d * (sumy * -8.f + acc);
+}
+
+#ifdef INTEL_GPU
+#define N_DST 16 // each SIMD group works on 8 rows (in weights matrix)
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // assuming SIMD group size is 16
+#elif defined (ADRENO_GPU)
+#define N_DST 16
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+//
+// This variant performs 1d blocking with 16x output.
+// Eeach simdgroup outputs 16 values on `n0` dim (row in the output matrix).
+//
+inline void mul_mat_q_n_f32_1d_16x_flat(
+        global uchar * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    const int nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    // (r0 * N_SIMDGROUP + get_sub_group_id()) is the linear global id of
+    // a SIMD group in the grid. Each SIMD group produces N_DST values in the
+    // result, hence uses nb blocks, i.e., the offset becomes first_row*nb.
+    // Currently with llama2 7B, im is always 0.
+    // TODO: how to handle im/gqa*(nb*ne0)?
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    // The number of scales is the same as the number of blocks.
+    ulong offset0_d = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    // Each block contains QK4_0/2 uchars, hence offset for qs is as follows.
+    ulong offset0_q = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02)) * QK4_0/2;
+
+    global uchar * x = (global uchar *) src0_q + offset0_q;
+    global half  * d = (global half  *) src0_d + offset0_d;
+    global float * y = (global float *) src1   + r1*ne10 + im*ne00*ne1;
+
+    float16 yl;
+    float16 sumf = (float16)(0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f,
+                             0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f);
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix*QK4_0 + il;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0.f;
+
+        sumy += yb[0];
+        sumy += yb[1];
+        sumy += yb[2];
+        sumy += yb[3];
+        sumy += yb[4];
+        sumy += yb[5];
+        sumy += yb[6];
+        sumy += yb[7];
+
+        sumy += yb[16];
+        sumy += yb[17];
+        sumy += yb[18];
+        sumy += yb[19];
+        sumy += yb[20];
+        sumy += yb[21];
+        sumy += yb[22];
+        sumy += yb[23];
+
+        yl.s0 = yb[0];
+        yl.s1 = yb[1]/256.f;
+
+        yl.s2 = yb[2];
+        yl.s3 = yb[3]/256.f;
+
+        yl.s4 = yb[4];
+        yl.s5 = yb[5]/256.f;
+
+        yl.s6 = yb[6];
+        yl.s7 = yb[7]/256.f;
+
+        yl.s8 = yb[16]/16.f;
+        yl.s9 = yb[17]/4096.f;
+
+        yl.sa = yb[18]/16.f;
+        yl.sb = yb[19]/4096.f;
+
+        yl.sc = yb[20]/16.f;
+        yl.sd = yb[21]/4096.f;
+
+        yl.se = yb[22]/16.f;
+        yl.sf = yb[23]/4096.f;
+
+        sumf.s0 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  0*nb*QK4_0/2, d + ib +  0*nb, sumy, yl, il);
+        sumf.s1 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  1*nb*QK4_0/2, d + ib +  1*nb, sumy, yl, il);
+        sumf.s2 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  2*nb*QK4_0/2, d + ib +  2*nb, sumy, yl, il);
+        sumf.s3 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  3*nb*QK4_0/2, d + ib +  3*nb, sumy, yl, il);
+
+        sumf.s4 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  4*nb*QK4_0/2, d + ib +  4*nb, sumy, yl, il);
+        sumf.s5 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  5*nb*QK4_0/2, d + ib +  5*nb, sumy, yl, il);
+        sumf.s6 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  6*nb*QK4_0/2, d + ib +  6*nb, sumy, yl, il);
+        sumf.s7 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  7*nb*QK4_0/2, d + ib +  7*nb, sumy, yl, il);
+
+        sumf.s8 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  8*nb*QK4_0/2, d + ib +  8*nb, sumy, yl, il);
+        sumf.s9 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  9*nb*QK4_0/2, d + ib +  9*nb, sumy, yl, il);
+        sumf.sa += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 10*nb*QK4_0/2, d + ib + 10*nb, sumy, yl, il);
+        sumf.sb += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 11*nb*QK4_0/2, d + ib + 11*nb, sumy, yl, il);
+
+        sumf.sc += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 12*nb*QK4_0/2, d + ib + 12*nb, sumy, yl, il);
+        sumf.sd += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 13*nb*QK4_0/2, d + ib + 13*nb, sumy, yl, il);
+        sumf.se += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 14*nb*QK4_0/2, d + ib + 14*nb, sumy, yl, il);
+        sumf.sf += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 15*nb*QK4_0/2, d + ib + 15*nb, sumy, yl, il);
+
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    float16 tot = (float16)(
+        sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3),
+        sub_group_reduce_add(sumf.s4), sub_group_reduce_add(sumf.s5),
+        sub_group_reduce_add(sumf.s6), sub_group_reduce_add(sumf.s7),
+
+        sub_group_reduce_add(sumf.s8), sub_group_reduce_add(sumf.s9),
+        sub_group_reduce_add(sumf.sa), sub_group_reduce_add(sumf.sb),
+        sub_group_reduce_add(sumf.sc), sub_group_reduce_add(sumf.sd),
+        sub_group_reduce_add(sumf.se), sub_group_reduce_add(sumf.sf)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+
+        if (first_row + 4 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 4] = tot.s4;
+        }
+        if (first_row + 5 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 5] = tot.s5;
+        }
+        if (first_row + 6 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 6] = tot.s6;
+        }
+        if (first_row + 7 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 7] = tot.s7;
+        }
+
+        if (first_row + 8 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 8] = tot.s8;
+        }
+        if (first_row + 9 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 9] = tot.s9;
+        }
+        if (first_row + 10 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 10] = tot.sa;
+        }
+        if (first_row + 11 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 11] = tot.sb;
+        }
+
+        if (first_row + 12 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 12] = tot.sc;
+        }
+        if (first_row + 13 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 13] = tot.sd;
+        }
+        if (first_row + 14 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 14] = tot.se;
+        }
+        if (first_row + 15 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 15] = tot.sf;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_q4_0_f32_1d_16x_flat(
+        global uchar * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    mul_mat_q_n_f32_1d_16x_flat(src0_q, src0_d, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}

ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_8x_flat.cl

@@ -0,0 +1,265 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+#define QR4_0                   2
+#define QK4_1                   32
+#define QR4_1                   2
+#define QK5_0                   32
+#define QR5_0                   2
+#define QK5_1                   32
+#define QR5_1                   2
+#define QK8_0                   32
+#define QR8_0                   1
+#define QK_K                    256
+#define K_QUANTS_PER_ITERATION  2
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+inline float mm_block_q_4_0_dot_y_flat(
+        global uchar * x,
+        global half  * dh,
+        float sumy,
+        float16 yl,
+        int il
+) {
+    float           d   = *dh;
+    global ushort * qs  = ((global ushort *)x + il/2);
+    float           acc = 0.f;
+
+    acc += yl.s0 * (qs[0] & 0x000F);
+    acc += yl.s1 * (qs[0] & 0x0F00);
+    acc += yl.s8 * (qs[0] & 0x00F0);
+    acc += yl.s9 * (qs[0] & 0xF000);
+
+    acc += yl.s2 * (qs[1] & 0x000F);
+    acc += yl.s3 * (qs[1] & 0x0F00);
+    acc += yl.sa * (qs[1] & 0x00F0);
+    acc += yl.sb * (qs[1] & 0xF000);
+
+    acc += yl.s4 * (qs[2] & 0x000F);
+    acc += yl.s5 * (qs[2] & 0x0F00);
+    acc += yl.sc * (qs[2] & 0x00F0);
+    acc += yl.sd * (qs[2] & 0xF000);
+
+    acc += yl.s6 * (qs[3] & 0x000F);
+    acc += yl.s7 * (qs[3] & 0x0F00);
+    acc += yl.se * (qs[3] & 0x00F0);
+    acc += yl.sf * (qs[3] & 0xF000);
+
+    return d * (sumy * -8.f + acc);
+}
+
+#ifdef INTEL_GPU
+#define N_DST 8 // each SIMD group works on 8 rows (in weights matrix)
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // assuming SIMD group size is 16
+#elif defined (ADRENO_GPU)
+#define N_DST 8
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+//
+// This variant performs 1d blocking with 8x output.
+// Eeach simdgroup outputs 8 values on `n0` dim (row in the output matrix).
+//
+inline void mul_mat_q_n_f32_1d_8x_flat(
+        global uchar * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    const int nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    // (r0 * N_SIMDGROUP + get_sub_group_id()) is the linear global id of
+    // a SIMD group in the grid. Each SIMD group produces N_DST values in the
+    // result, hence uses nb blocks, i.e., the offset becomes first_row*nb.
+    // Currently with llama2 7B, im is always 0.
+    // TODO: how to handle im/gqa*(nb*ne0)?
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    // The number of scales is the same as the number of blocks.
+    ulong offset0_d = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    // Each block contains QK4_0/2 uchars, hence offset for qs is as follows.
+    ulong offset0_q = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02)) * QK4_0/2;
+
+    global uchar * x = (global uchar *) src0_q + offset0_q;
+    global half  * d = (global half  *) src0_d + offset0_d;
+    global float * y = (global float *) src1   + r1*ne10 + im*ne00*ne1;
+
+    float16 yl;
+    float8 sumf = (float8)(0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f);
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix*QK4_0 + il;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0.f;
+
+        sumy += yb[0];
+        sumy += yb[1];
+        sumy += yb[2];
+        sumy += yb[3];
+        sumy += yb[4];
+        sumy += yb[5];
+        sumy += yb[6];
+        sumy += yb[7];
+
+        sumy += yb[16];
+        sumy += yb[17];
+        sumy += yb[18];
+        sumy += yb[19];
+        sumy += yb[20];
+        sumy += yb[21];
+        sumy += yb[22];
+        sumy += yb[23];
+
+        yl.s0 = yb[0];
+        yl.s1 = yb[1]/256.f;
+
+        yl.s2 = yb[2];
+        yl.s3 = yb[3]/256.f;
+
+        yl.s4 = yb[4];
+        yl.s5 = yb[5]/256.f;
+
+        yl.s6 = yb[6];
+        yl.s7 = yb[7]/256.f;
+
+        yl.s8 = yb[16]/16.f;
+        yl.s9 = yb[17]/4096.f;
+
+        yl.sa = yb[18]/16.f;
+        yl.sb = yb[19]/4096.f;
+
+        yl.sc = yb[20]/16.f;
+        yl.sd = yb[21]/4096.f;
+
+        yl.se = yb[22]/16.f;
+        yl.sf = yb[23]/4096.f;
+
+        sumf.s0 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 0*nb*QK4_0/2, d + ib + 0*nb, sumy, yl, il);
+        sumf.s1 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 1*nb*QK4_0/2, d + ib + 1*nb, sumy, yl, il);
+        sumf.s2 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 2*nb*QK4_0/2, d + ib + 2*nb, sumy, yl, il);
+        sumf.s3 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 3*nb*QK4_0/2, d + ib + 3*nb, sumy, yl, il);
+
+        sumf.s4 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 4*nb*QK4_0/2, d + ib + 4*nb, sumy, yl, il);
+        sumf.s5 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 5*nb*QK4_0/2, d + ib + 5*nb, sumy, yl, il);
+        sumf.s6 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 6*nb*QK4_0/2, d + ib + 6*nb, sumy, yl, il);
+        sumf.s7 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 7*nb*QK4_0/2, d + ib + 7*nb, sumy, yl, il);
+
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    float8 tot = (float8)(
+        sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3),
+        sub_group_reduce_add(sumf.s4), sub_group_reduce_add(sumf.s5),
+        sub_group_reduce_add(sumf.s6), sub_group_reduce_add(sumf.s7)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+
+        if (first_row + 4 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 4] = tot.s4;
+        }
+        if (first_row + 5 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 5] = tot.s5;
+        }
+        if (first_row + 6 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 6] = tot.s6;
+        }
+        if (first_row + 7 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 7] = tot.s7;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_q4_0_f32_1d_8x_flat(
+        global uchar * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    mul_mat_q_n_f32_1d_8x_flat(src0_q, src0_d, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}

ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_8x_flat.cl

@@ -0,0 +1,272 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+#define QR4_0                   2
+#define QK4_1                   32
+#define QR4_1                   2
+#define QK5_0                   32
+#define QR5_0                   2
+#define QK5_1                   32
+#define QR5_1                   2
+#define QK8_0                   32
+#define QR8_0                   1
+#define QK_K                    256
+#define K_QUANTS_PER_ITERATION  2
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+// This function requires the original shuffled weights.
+// As a reminder, the original weights are shuffled so that (q[0], q[16]) are
+// packed together in a byte, so are (q[1], q[17]) and so on.
+inline float block_q_4_0_dot_y_flat(
+        global uchar * x,
+        global half  * dh,
+        float sumy,
+        float16 yl,
+        int il
+) {
+    float           d   = *dh;
+    global ushort * qs  = ((global ushort *)x + il/2);
+    float           acc = 0.f;
+
+    acc += yl.s0 * (qs[0] & 0x000F);
+    acc += yl.s1 * (qs[0] & 0x0F00);
+    acc += yl.s8 * (qs[0] & 0x00F0);
+    acc += yl.s9 * (qs[0] & 0xF000);
+
+    acc += yl.s2 * (qs[1] & 0x000F);
+    acc += yl.s3 * (qs[1] & 0x0F00);
+    acc += yl.sa * (qs[1] & 0x00F0);
+    acc += yl.sb * (qs[1] & 0xF000);
+
+    acc += yl.s4 * (qs[2] & 0x000F);
+    acc += yl.s5 * (qs[2] & 0x0F00);
+    acc += yl.sc * (qs[2] & 0x00F0);
+    acc += yl.sd * (qs[2] & 0xF000);
+
+    acc += yl.s6 * (qs[3] & 0x000F);
+    acc += yl.s7 * (qs[3] & 0x0F00);
+    acc += yl.se * (qs[3] & 0x00F0);
+    acc += yl.sf * (qs[3] & 0xF000);
+
+    return d * (sumy * -8.f + acc);
+}
+
+//
+// This variant outputs 8 values.
+//
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 8 // each SIMD group works on 8 rows
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // assuming SIMD group size is 32
+#elif defined (ADRENO_GPU)
+#define N_DST 8
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_vec_q_n_f32_8x_flat(
+        global uchar * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    const ulong nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    // (r0 * N_SIMDGROUP + get_sub_group_id()) is the linear global id of
+    // a SIMD group in the grid. Each SIMD group produces N_DST values in the
+    // result, hence uses nb blocks, i.e., the offset becomes first_row*nb.
+    // Currently with llama2 7B, im is always 0.
+    // TODO: how to handle im/gqa*(nb*ne0)?
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    // The number of scales is the same as the number of blocks.
+    ulong offset0_d = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    // Each block contains QK4_0/2 uchars, hence offset for qs is as follows.
+    ulong offset0_q = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02)) * QK4_0/2;
+
+    global uchar * x = (global uchar *) src0_q + offset0_q;
+    global half  * d = (global half  *) src0_d + offset0_d;
+    global float * y = (global float *) src1   + r1*ne10 + im*ne00*ne1;
+
+    float16 yl;
+    float8 sumf = 0.f;
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix*QK4_0 + il;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0.f;
+
+        sumy += yb[0];
+        sumy += yb[1];
+        sumy += yb[2];
+        sumy += yb[3];
+        sumy += yb[4];
+        sumy += yb[5];
+        sumy += yb[6];
+        sumy += yb[7];
+
+        sumy += yb[16];
+        sumy += yb[17];
+        sumy += yb[18];
+        sumy += yb[19];
+        sumy += yb[20];
+        sumy += yb[21];
+        sumy += yb[22];
+        sumy += yb[23];
+
+        yl.s0 = yb[0];
+        yl.s1 = yb[1]/256.f;
+
+        yl.s2 = yb[2];
+        yl.s3 = yb[3]/256.f;
+
+        yl.s4 = yb[4];
+        yl.s5 = yb[5]/256.f;
+
+        yl.s6 = yb[6];
+        yl.s7 = yb[7]/256.f;
+
+        yl.s8 = yb[16]/16.f;
+        yl.s9 = yb[17]/4096.f;
+
+        yl.sa = yb[18]/16.f;
+        yl.sb = yb[19]/4096.f;
+
+        yl.sc = yb[20]/16.f;
+        yl.sd = yb[21]/4096.f;
+
+        yl.se = yb[22]/16.f;
+        yl.sf = yb[23]/4096.f;
+
+        sumf.s0 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 0*nb*QK4_0/2, d + ib + 0*nb, sumy, yl, il);
+        sumf.s1 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 1*nb*QK4_0/2, d + ib + 1*nb, sumy, yl, il);
+        sumf.s2 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 2*nb*QK4_0/2, d + ib + 2*nb, sumy, yl, il);
+        sumf.s3 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 3*nb*QK4_0/2, d + ib + 3*nb, sumy, yl, il);
+
+        sumf.s4 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 4*nb*QK4_0/2, d + ib + 4*nb, sumy, yl, il);
+        sumf.s5 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 5*nb*QK4_0/2, d + ib + 5*nb, sumy, yl, il);
+        sumf.s6 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 6*nb*QK4_0/2, d + ib + 6*nb, sumy, yl, il);
+        sumf.s7 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 7*nb*QK4_0/2, d + ib + 7*nb, sumy, yl, il);
+
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    float8 tot = (float8)(
+        sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3),
+        sub_group_reduce_add(sumf.s4), sub_group_reduce_add(sumf.s5),
+        sub_group_reduce_add(sumf.s6), sub_group_reduce_add(sumf.s7)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+
+        if (first_row + 4 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 4] = tot.s4;
+        }
+        if (first_row + 5 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 5] = tot.s5;
+        }
+        if (first_row + 6 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 6] = tot.s6;
+        }
+        if (first_row + 7 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 7] = tot.s7;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_q4_0_f32_8x_flat(
+        global uchar * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    mul_vec_q_n_f32_8x_flat(src0_q, src0_d, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}

ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_v.cl

@@ -0,0 +1,254 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+#define QR4_0                   2
+#define QK4_1                   32
+#define QR4_1                   2
+#define QK5_0                   32
+#define QR5_0                   2
+#define QK5_1                   32
+#define QR5_1                   2
+#define QK8_0                   32
+#define QR8_0                   1
+#define QK_K                    256
+#define K_QUANTS_PER_ITERATION  2
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+//
+// This variant unrolls the loops and uses vector types instead of pointers.
+// It improves performance on Adreno but not so much on Intel.
+//
+inline float block_q_4_0_dot_y_v(
+        global struct block_q4_0 * qb_curr,
+        float sumy,
+        float16 yl,
+        int il
+) {
+    float d = qb_curr->d;
+    float acc = 0.f;
+    global ushort * qs = ((global ushort *)qb_curr + 1 + il/2);
+
+    acc += yl.s0 * (qs[0] & 0x000F);
+    acc += yl.s1 * (qs[0] & 0x0F00);
+    acc += yl.s8 * (qs[0] & 0x00F0);
+    acc += yl.s9 * (qs[0] & 0xF000);
+
+    acc += yl.s2 * (qs[1] & 0x000F);
+    acc += yl.s3 * (qs[1] & 0x0F00);
+    acc += yl.sa * (qs[1] & 0x00F0);
+    acc += yl.sb * (qs[1] & 0xF000);
+
+    acc += yl.s4 * (qs[2] & 0x000F);
+    acc += yl.s5 * (qs[2] & 0x0F00);
+    acc += yl.sc * (qs[2] & 0x00F0);
+    acc += yl.sd * (qs[2] & 0xF000);
+
+    acc += yl.s6 * (qs[3] & 0x000F);
+    acc += yl.s7 * (qs[3] & 0x0F00);
+    acc += yl.se * (qs[3] & 0x00F0);
+    acc += yl.sf * (qs[3] & 0xF000);
+
+    return d * (sumy * -8.f + acc);
+}
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 4 // each SIMD group works on 4 rows
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // assuming SIMD group size is 16
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_vec_q_n_f32_v(
+        global void * src0,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    const ulong nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    // (r0 * N_SIMDGROUP + get_sub_group_id()) is essenatially the linear global
+    // id of a SIMD group in the grid.
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset0 = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    global struct block_q4_0 * x = (global struct block_q4_0 *) src0 + offset0;
+    global float             * y = (global float             *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float16 yl;       // src1 vector cache
+    float4 sumf = (float4)(0.f, 0.f, 0.f, 0.f);
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix * QK4_0 + il;
+
+    // each thread in a SIMD group deals with half a block.
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0;
+
+        sumy += yb[0];
+        sumy += yb[1];
+        sumy += yb[2];
+        sumy += yb[3];
+        sumy += yb[4];
+        sumy += yb[5];
+        sumy += yb[6];
+        sumy += yb[7];
+
+        sumy += yb[16];
+        sumy += yb[17];
+        sumy += yb[18];
+        sumy += yb[19];
+        sumy += yb[20];
+        sumy += yb[21];
+        sumy += yb[22];
+        sumy += yb[23];
+
+
+        yl.s0 = yb[0];
+        yl.s1 = yb[1]/256.f;
+
+        yl.s2 = yb[2];
+        yl.s3 = yb[3]/256.f;
+
+        yl.s4 = yb[4];
+        yl.s5 = yb[5]/256.f;
+
+        yl.s6 = yb[6];
+        yl.s7 = yb[7]/256.f;
+
+        yl.s8 = yb[16]/16.f;
+        yl.s9 = yb[17]/4096.f;
+
+        yl.sa = yb[18]/16.f;
+        yl.sb = yb[19]/4096.f;
+
+        yl.sc = yb[20]/16.f;
+        yl.sd = yb[21]/4096.f;
+
+        yl.se = yb[22]/16.f;
+        yl.sf = yb[23]/4096.f;
+
+        sumf.s0 += block_q_4_0_dot_y_v(x+ib+0*nb, sumy, yl, il);
+        sumf.s1 += block_q_4_0_dot_y_v(x+ib+1*nb, sumy, yl, il);
+        sumf.s2 += block_q_4_0_dot_y_v(x+ib+2*nb, sumy, yl, il);
+        sumf.s3 += block_q_4_0_dot_y_v(x+ib+3*nb, sumy, yl, il);
+
+        // One thread in a SIMD group (i.e., subgroup) handles a half block,
+        // hence then entire SIMD group handles SIMDWIDTH/2 blocks.
+        // y points to the activation matrix (of type float). Therefore for
+        // one thread, the # of blocks y should advance is SIMDWIDTH/2 (because
+        // SIMDWIDTH/2 blocks are processed by a SIMD group) - in terms of
+        // floats, it is QK4_0 * (SIMDWIDTH/2), where QK4_0 is the block size.
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    // The above does not work for Adreno - it produces incorrect results for
+    // row = 1, 2, 3 and only row = 0 gives the correct result.
+    // If N_DST is changed, the below array must be initialized accordingly.
+    // This also seems to perform better on Intel.
+    float4 tot = (float4)(
+        sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_q4_0_f32_v(
+        global void * src0,
+        ulong offset0,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    mul_vec_q_n_f32_v(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}

ggml/src/ggml-opencl/kernels/mul_mv_q6_k.cl

@@ -0,0 +1,190 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+#define QR4_0                   2
+#define QK4_1                   32
+#define QR4_1                   2
+#define QK5_0                   32
+#define QR5_0                   2
+#define QK5_1                   32
+#define QR5_1                   2
+#define QK8_0                   32
+#define QR8_0                   1
+#define QK_K                    256
+#define K_QUANTS_PER_ITERATION  2
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q6_K
+//------------------------------------------------------------------------------
+// 6-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elements each
+// Effectively 6.5625 bits per weight
+typedef struct {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    half d;             // super-block scale
+} block_q6_K;
+
+//------------------------------------------------------------------------------
+// kernel_mul_mv_q6_K_f32
+//------------------------------------------------------------------------------
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 1 // number of rows each SIMD group works on
+#define N_SIMDGROUP 2 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // SIMD group size
+#elif defined (ADRENO_GPU)
+#define N_DST 1
+#define N_SIMDGROUP 2
+#define N_SIMDWIDTH 64
+#endif
+
+#define BLOCK_STRIDE (N_SIMDWIDTH/16) // number of blocks each subgroup processes
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_q6_K_f32(
+        global void * src0,
+        ulong offset0,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    uchar kmask1 = 0x03;
+    uchar kmask2 = 0x0C;
+    uchar kmask3 = 0x30;
+    uchar kmask4 = 0xC0;
+
+    int nb = ne00/QK_K;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int row = N_SIMDGROUP * r0 + get_sub_group_id();
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    global block_q6_K * x = (global block_q6_K *) src0 + row*nb + offset_src0;
+    global float      * yy = (global float     *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float sumf = 0;
+
+    // For Q6_K quantization, 16 values forms a subblock, 16 subblock forms a
+    // block. Values in a subblock shares a scale that is quantized with 8 bits;
+    // the entire block shares a single floating point scale.
+    // For work distribution, each thread processes a subblock (16 weights), hence
+    // 16 threads process a (super) block -- a subgroup thus handles SIMDWIDTH/16
+    // (super) blocks -- this is the block stride.
+    // The 16 threads that process a (super) block are split into 2 portions, each has
+    // 8 threads; each portion works on 8 subblocks.
+    // For subgroup of 16 threads, the entire subgroup works on a single (super) block
+    // before moving to the next (super) block. Thread0 - thread7 work on the
+    // first 8 subblocks; thread8 - thread15 works on the last 8 subblocks.
+    // Thread0 - thread3 work on subblocks 0, 2, 4, 6; thread4 - thread7 work on
+    // subblocks 1, 3, 5, 7. Each thread does not work on an entire subblock, but
+    // works on a total of 16 weight values.
+    int tid  = get_sub_group_local_id()/BLOCK_STRIDE; // first block_stride groups have tid=0
+    int ix   = get_sub_group_local_id()%BLOCK_STRIDE; // first block is 0..block_stride-1
+    int ip   = tid/8;   // first or second half of (super) block (0 or 1)
+    int il   = tid%8;   // each half has 8 parts, one per scale
+    int n    = 4;       // 4 scales at a time (and 4 sums)
+    int l0   = n*il;    // offset into half-block, 0..28
+    int is   = 8*ip + l0/16; // 0, 1, 8, 9
+
+    int y_offset = 128*ip + l0;
+    int q_offset_l = 64*ip + l0;
+    int q_offset_h = 32*ip + l0;
+
+    for (int i = ix; i < nb; i += BLOCK_STRIDE) {
+
+        global uint8_t * q1 = x[i].ql + q_offset_l;
+        global uint8_t * q2 = q1 + QK_K/8;
+        global uint8_t * qh = x[i].qh + q_offset_h;
+        global int8_t  * sc = x[i].scales + is;
+
+        global float * y = yy + i * QK_K + y_offset;
+
+        float dall = x[i].d;
+
+        float4 sums = {0.f, 0.f, 0.f, 0.f};
+
+        sums.s0 += y[0+ 0] * ((float)((q1[0] & 0xF) | ((qh[0] & kmask1) << 4)) - 32.f);
+        sums.s1 += y[0+32] * ((float)((q2[0] & 0xF) | ((qh[0] & kmask2) << 2)) - 32.f);
+        sums.s2 += y[0+64] * ((float)((q1[0]  >> 4) | ((qh[0] & kmask3) << 0)) - 32.f);
+        sums.s3 += y[0+96] * ((float)((q2[0]  >> 4) | ((qh[0] & kmask4) >> 2)) - 32.f);
+
+        sums.s0 += y[1+ 0] * ((float)((q1[1] & 0xF) | ((qh[1] & kmask1) << 4)) - 32.f);
+        sums.s1 += y[1+32] * ((float)((q2[1] & 0xF) | ((qh[1] & kmask2) << 2)) - 32.f);
+        sums.s2 += y[1+64] * ((float)((q1[1]  >> 4) | ((qh[1] & kmask3) << 0)) - 32.f);
+        sums.s3 += y[1+96] * ((float)((q2[1]  >> 4) | ((qh[1] & kmask4) >> 2)) - 32.f);
+
+        sums.s0 += y[2+ 0] * ((float)((q1[2] & 0xF) | ((qh[2] & kmask1) << 4)) - 32.f);
+        sums.s1 += y[2+32] * ((float)((q2[2] & 0xF) | ((qh[2] & kmask2) << 2)) - 32.f);
+        sums.s2 += y[2+64] * ((float)((q1[2]  >> 4) | ((qh[2] & kmask3) << 0)) - 32.f);
+        sums.s3 += y[2+96] * ((float)((q2[2]  >> 4) | ((qh[2] & kmask4) >> 2)) - 32.f);
+
+        sums.s0 += y[3+ 0] * ((float)((q1[3] & 0xF) | ((qh[3] & kmask1) << 4)) - 32.f);
+        sums.s1 += y[3+32] * ((float)((q2[3] & 0xF) | ((qh[3] & kmask2) << 2)) - 32.f);
+        sums.s2 += y[3+64] * ((float)((q1[3]  >> 4) | ((qh[3] & kmask3) << 0)) - 32.f);
+        sums.s3 += y[3+96] * ((float)((q2[3]  >> 4) | ((qh[3] & kmask4) >> 2)) - 32.f);
+
+        sumf += dall * (sums.s0 * sc[0] + sums.s1 * sc[2] + sums.s2 * sc[4] + sums.s3 * sc[6]);
+    }
+
+    float tot = sub_group_reduce_add(sumf);
+    if (get_sub_group_local_id() == 0) {
+        dst[r1*ne0 + im*ne0*ne1 + row] = tot;
+    }
+}

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

@@ -0,0 +1,81 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+//------------------------------------------------------------------------------
+// norm
+//------------------------------------------------------------------------------
+kernel void kernel_norm(
+        global void * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        float eps,
+        local float * sum
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    dst = (global void*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    global float * x = (global float *) ((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01);
+
+    // MEAN
+    // parallel sum
+    sum[get_local_id(0)] = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        sum[get_local_id(0)] += x[i00];
+    }
+    // reduce
+    barrier(CLK_LOCAL_MEM_FENCE);
+    for (uint i = get_local_size(0)/2; i > 0; i /= 2) {
+        if (get_local_id(0) < i) {
+            sum[get_local_id(0)] += sum[get_local_id(0) + i];
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    float mean  = sum[0] / ne00;
+
+    // recenter and VARIANCE
+    barrier(CLK_LOCAL_MEM_FENCE);
+    global float * y = dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    sum[get_local_id(0)] = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        y[i00] = x[i00] - mean;
+        sum[get_local_id(0)] += y[i00] * y[i00];
+    }
+
+    // reduce
+    barrier(CLK_LOCAL_MEM_FENCE);
+    for (uint i = get_local_size(0)/2; i > 0; i /= 2) {
+        if (get_local_id(0) < i) {
+            sum[get_local_id(0)] += sum[get_local_id(0) + i];
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    float variance = sum[0] / ne00;
+
+    float scale = 1.0f/sqrt(variance + eps);
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        y[i00] = y[i00] * scale;
+    }
+}

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

@@ -0,0 +1,16 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// relu
+//------------------------------------------------------------------------------
+kernel void kernel_relu(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = fmax(0.0f, src0[get_global_id(0)]);
+}

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

@@ -0,0 +1,96 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+//------------------------------------------------------------------------------
+// rms_norm
+//------------------------------------------------------------------------------
+// This kernel depends on subgroup size.
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_32
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_rms_norm(
+        global void * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        float eps,
+        local float * sum // Note, the size depends on number of subgroups
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    global float4 * x = (global float4 *) ((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01);
+    global float * x_scalar = (global float *) x;
+    float4 sumf = 0;
+    float all_sum = 0;
+
+    // parallel sum
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        sumf += x[i00] * x[i00];
+    }
+    all_sum = sumf.s0 + sumf.s1 + sumf.s2 + sumf.s3;
+    all_sum = sub_group_reduce_add(all_sum);
+    if (get_sub_group_local_id() == 0) {
+        sum[get_sub_group_id()] = all_sum;
+    }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+    // broadcast
+    for (uint i = get_local_size(0) / get_max_sub_group_size() / 2; i > 0; i /= 2) {
+       if (get_local_id(0) < i) {
+           sum[get_local_id(0)] += sum[get_local_id(0) + i];
+       }
+    }
+    if (get_local_id(0) == 0) {
+        for (int i = 4 * (ne00 / 4); i < ne00; i++) {
+            sum[0] += x_scalar[i];
+        }
+        sum[0] /= ne00;
+    }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    const float mean  = sum[0];
+    const float scale = 1.0f/sqrt(mean + eps);
+
+    global float4 * y = (global float4 *) (dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+    global float * y_scalar = (global float *) y;
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        y[i00] = x[i00] * scale;
+    }
+    if (get_local_id(0) == 0) {
+        for (int i00 = 4 * (ne00 / 4); i00 < ne00; i00++) {
+            y_scalar[i00] = x_scalar[i00] * scale;
+        }
+    }
+}

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

@@ -0,0 +1,721 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// kernel_rope
+//------------------------------------------------------------------------------
+float rope_yarn_ramp(float low, float high, int i0) {
+    const float y = (i0 / 2 - low) / max(0.001f, high - low);
+    return 1.0f - min(1.0f, max(0.0f, y));
+}
+
+// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
+// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
+float2 rope_yarn(
+    float theta_extrap, float freq_scale, float2 corr_dims, int i0, float ext_factor, float mscale
+) {
+    // Get n-d rotational scaling corrected for extrapolation
+    float theta_interp = freq_scale * theta_extrap;
+    float theta = theta_interp;
+    if (ext_factor != 0.0f) {
+        float ramp_mix = rope_yarn_ramp(corr_dims.s0, corr_dims.s1, i0) * ext_factor;
+        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+        // Get n-d magnitude scaling corrected for interpolation
+        mscale *= 1.0f + 0.1f * log(1.0f / freq_scale);
+    }
+    return (float2)(cos(theta) * mscale, sin(theta) * mscale);
+}
+
+// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
+// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
+float rope_yarn_corr_factor(int n_dims, int n_ctx_orig, float n_rot, float base) {
+    return n_dims * log(n_ctx_orig / (n_rot * 2 * M_PI_F)) / (2 * log(base));
+}
+
+float2 rope_yarn_corr_dims(
+    int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow
+) {
+    // start and end correction dims
+    return (float2)(
+        max(0.0f,         floor(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_fast, freq_base))),
+        min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_slow, freq_base)))
+    );
+}
+
+kernel void kernel_rope_norm_f32(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    float theta_base = (float) pos[i2];
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        if (i0 < n_dims) {
+            int ic = i0/2;
+
+            float theta = theta_base * pow(freq_base, inv_ndims*i0);
+
+            float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+            global float * src       = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global float * dst_data  = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            float x0 = src[0];
+            float x1 = src[1];
+
+            dst_data[0] = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+            dst_data[1] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+        } else {
+            global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+kernel void kernel_rope_norm_f16(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    float theta_base = (float) pos[i2];
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        if (i0 < n_dims) {
+            int ic = i0/2;
+
+            float theta = theta_base * pow(freq_base, inv_ndims*i0);
+
+            float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+            global half * src       = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global half * dst_data  = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            float x0 = src[0];
+            float x1 = src[1];
+
+            dst_data[0] = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+            dst_data[1] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+        } else {
+            global half * src      = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global half * dst_data = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+kernel void kernel_rope_neox_f32(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    float theta_base = (float) pos[i2];
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        if (i0 < n_dims) {
+            int ic = i0/2;
+
+            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+
+            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+            global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+            global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[n_dims/2];
+
+            dst_data[0]        = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+            dst_data[n_dims/2] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+        } else {
+            global float * const src = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global float * dst_data  = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+kernel void kernel_rope_neox_f16(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    float theta_base = (float) pos[i2];
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        if (i0 < n_dims) {
+            int ic = i0/2;
+
+            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+
+            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+            global half * src       = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+            global half * dst_data  = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[n_dims/2];
+
+            dst_data[0]        = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+            dst_data[n_dims/2] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+        } else {
+            global half * const src = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global half * dst_data  = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+kernel void kernel_rope_multi_f32(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow,
+        int4 sections
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    const int sect_dims = sections.s0 + sections.s1 + sections.s2 + sections.s3;
+    const int sec_w = sections.s1 + sections.s0;
+
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        if (i0 < n_dims) {
+            int ic = i0/2;
+
+            const int sector = (i0 / 2) % sect_dims;
+            float theta_base = 0.0f;
+
+            if (sector < sections.s0) {
+                theta_base = pos[i2];
+            }
+            else if (sector >= sections.s0 && sector < sec_w) {
+                theta_base = pos[i2 + ne2 * 1];
+            }
+            else if (sector >= sec_w && sector < sec_w + sections.s2) {
+                theta_base = pos[i2 + ne2 * 2];
+            }
+            else if (sector >= sec_w + sections.s2) {
+                theta_base = pos[i2 + ne2 * 3];
+            }
+
+            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+
+            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+            global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+            global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[n_dims/2];
+
+            dst_data[0]        = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+            dst_data[n_dims/2] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+        } else {
+            global float * const src = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global float * dst_data  = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+kernel void kernel_rope_multi_f16(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global half * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow,
+        int4 sections
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    const int sect_dims = sections.s0 + sections.s1 + sections.s2 + sections.s3;
+    const int sec_w = sections.s1 + sections.s0;
+
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        if (i0 < n_dims) {
+            int ic = i0/2;
+
+            const int sector = (i0 / 2) % sect_dims;
+            float theta_base = 0.0f;
+
+            if (sector < sections.s0) {
+                theta_base = pos[i2];
+            }
+            else if (sector >= sections.s0 && sector < sec_w) {
+                theta_base = pos[i2 + ne2 * 1];
+            }
+            else if (sector >= sec_w && sector < sec_w + sections.s2) {
+                theta_base = pos[i2 + ne2 * 2];
+            }
+            else if (sector >= sec_w + sections.s2) {
+                theta_base = pos[i2 + ne2 * 3];
+            }
+
+            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+
+            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+            global half * src      = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+            global half * dst_data = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[n_dims/2];
+
+            dst_data[0]        = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+            dst_data[n_dims/2] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+        } else {
+            global half * const src = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global half * dst_data  = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+kernel void kernel_rope_vision_f32(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow,
+        int4 sections
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    const int sect_dims = sections.s0 + sections.s1;
+    const int sec_w = sections.s1 + sections.s0;
+
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        int ic = i0/2;
+
+        const int sector = (i0/2) % sect_dims;
+        float theta_base = 0.0f;
+
+        if (sector < sections.s0) {
+            const int p = sector;
+            theta_base = pos[i2] * pow(freq_base, inv_ndims*2.0f*p);
+        } else if (sector >= sections.s0 && sector < sec_w) {
+            const int p = sector - sections.s0;
+            theta_base = pos[i2 + ne2] * pow(freq_base, inv_ndims*2.0f*p);
+        }
+
+        const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+        float2 cos_sin_theta = rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+        global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+        global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+        const float x0 = src[0];
+        const float x1 = src[n_dims];
+
+        dst_data[0]      = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+        dst_data[n_dims] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+    }
+}
+
+kernel void kernel_rope_vision_f16(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global half * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow,
+        int4 sections
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    const int sect_dims = sections.s0 + sections.s1;
+    const int sec_w = sections.s1 + sections.s0;
+
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        int ic = i0/2;
+
+        const int sector = (i0/2) % sect_dims;
+        float theta_base = 0.0f;
+
+        if (sector < sections.s0) {
+            const int p = sector;
+            theta_base = pos[i2] * pow(freq_base, inv_ndims*2.0f*p);
+        } else if (sector >= sections.s0 && sector < sec_w) {
+            const int p = sector - sections.s0;
+            theta_base = pos[i2 + ne2] * pow(freq_base, inv_ndims*2.0f*p);
+        }
+
+        const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+        float2 cos_sin_theta = rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+        global half * src      = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+        global half * dst_data = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+        const float x0 = src[0];
+        const float x1 = src[n_dims];
+
+        dst_data[0]      = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+        dst_data[n_dims] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+    }
+}

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

@@ -0,0 +1,16 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// scale
+//------------------------------------------------------------------------------
+kernel void kernel_scale(
+        global float4 * src0,
+        ulong offset0,
+        global float4 * dst,
+        ulong offsetd,
+        float scale
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst = (global float4*)((global char*)dst + offsetd);
+    dst[get_global_id(0)] = src0[get_global_id(0)] * scale;
+}

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

@@ -0,0 +1,30 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// silu
+//------------------------------------------------------------------------------
+kernel void kernel_silu(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    float x = src0[get_global_id(0)];
+    dst[get_global_id(0)] = x / (1.0f + exp(-x));
+}
+
+kernel void kernel_silu_4(
+        global float4 * src0,
+        ulong offset0,
+        global float4 * dst,
+        ulong offsetd
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    float4 x = src0[get_global_id(0)];
+    dst[get_global_id(0)] = x / (1.0f + exp(-x));
+}

ggml/src/ggml-opencl/kernels/softmax_4_f16.cl

@@ -0,0 +1,87 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_soft_max_4_f16(
+        global float * src0,
+        ulong offset0,
+        global half * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        float scale,
+        float max_bias,
+        float m0,
+        float m1,
+        int n_head_log2
+) {
+    src0 = (global float *)((global char *)src0 + offset0);
+    src1 = (global half *)((global char *)src1 + offset1);
+    dst = (global float *)((global char *)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    global float4 * psrc4 = (global float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+    global half4  * pmask = (global char *)src1 != (global char *)src0 ? (global half4 *)(src1 + i01*ne00) : 0;
+    global float4 * pdst4 = (global float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        int h = i02;
+
+        float base = h < n_head_log2 ? m0 : m1;
+        int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float4 lmax4 = -INFINITY;
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        lmax4 = fmax(lmax4, psrc4[i00]*scale + slope*(pmask ? convert_float4(pmask[i00]) : 0.0f));
+    }
+    float lmax = fmax(fmax(lmax4.s0, lmax4.s1), fmax(lmax4.s2, lmax4.s3));
+
+    const float max = sub_group_reduce_max(lmax);
+
+    // parallel sum
+    float4 lsum4 = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        const float4 exp_psrc4 = exp((psrc4[i00]*scale + slope*(pmask ? convert_float4(pmask[i00]) : 0.0f)) - max);
+        lsum4 += exp_psrc4;
+        pdst4[i00] = exp_psrc4;
+    }
+    float lsum = lsum4.s0 + lsum4.s1 + lsum4.s2 + lsum4.s3;
+
+    const float sum = sub_group_reduce_add(lsum);
+
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        pdst4[i00] /= sum;
+    }
+}

ggml/src/ggml-opencl/kernels/softmax_4_f32.cl

@@ -0,0 +1,87 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_soft_max_4(
+        global float * src0,
+        ulong offset0,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        float scale,
+        float max_bias,
+        float m0,
+        float m1,
+        int n_head_log2
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    global float4 * psrc4 = (global float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+    global float4 * pmask = src1 != src0 ? (global float4 *)(src1 + i01*ne00) : 0;
+    global float4 * pdst4 = (global float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        int h = i02;
+
+        float base = h < n_head_log2 ? m0 : m1;
+        int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float4 lmax4 = -INFINITY;
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        lmax4 = fmax(lmax4, psrc4[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f));
+    }
+    float lmax = fmax(fmax(lmax4.s0, lmax4.s1), fmax(lmax4.s2, lmax4.s3));
+
+    const float max = sub_group_reduce_max(lmax);
+
+    // parallel sum
+    float4 lsum4 = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        const float4 exp_psrc4 = exp((psrc4[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)) - max);
+        lsum4 += exp_psrc4;
+        pdst4[i00] = exp_psrc4;
+    }
+    float lsum = lsum4.s0 + lsum4.s1 + lsum4.s2 + lsum4.s3;
+
+    const float sum = sub_group_reduce_add(lsum);
+
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        pdst4[i00] /= sum;
+    }
+}

ggml/src/ggml-opencl/kernels/softmax_f16.cl

@@ -0,0 +1,86 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_soft_max_f16(
+        global float * src0,
+        ulong offset0,
+        global half * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        float scale,
+        float max_bias,
+        float m0,
+        float m1,
+        int n_head_log2
+) {
+    src0 = (global float *)((global char *)src0 + offset0);
+    src1 = (global half *)((global char *)src1 + offset1);
+    dst = (global float *)((global char *)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    global float * psrc0 = src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    global half  * pmask = (global char *)src1 != (global char *)src0 ? src1 + i01*ne00 : 0;
+    global float * pdst  = dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        int h = i02;
+
+        float base = h < n_head_log2 ? m0 : m1;
+        int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float lmax = -INFINITY;
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        lmax = fmax(lmax, psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f));
+    }
+    float max = sub_group_reduce_max(lmax);
+
+    // parallel sum
+    float lsum = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)) - max);
+        lsum += exp_psrc0;
+        // Remember the result of exp here. exp is expensive, so we really do not
+        // wish to compute it twice.
+        pdst[i00] = exp_psrc0;
+    }
+
+    const float sum = sub_group_reduce_add(lsum);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        pdst[i00] /= sum;
+    }
+}

ggml/src/ggml-opencl/kernels/softmax_f32.cl

@@ -0,0 +1,86 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_soft_max(
+        global float * src0,
+        ulong offset0,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        float scale,
+        float max_bias,
+        float m0,
+        float m1,
+        int n_head_log2
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    global float * psrc0 = src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    global float * pmask = src1 != src0 ? src1 + i01*ne00 : 0;
+    global float * pdst  = dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        int h = i02;
+
+        float base = h < n_head_log2 ? m0 : m1;
+        int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float lmax = -INFINITY;
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        lmax = fmax(lmax, psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f));
+    }
+    float max = sub_group_reduce_max(lmax);
+
+    // parallel sum
+    float lsum = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)) - max);
+        lsum += exp_psrc0;
+        // Remember the result of exp here. exp is expensive, so we really do not
+        // wish to compute it twice.
+        pdst[i00] = exp_psrc0;
+    }
+
+    const float sum = sub_group_reduce_add(lsum);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        pdst[i00] /= sum;
+    }
+}

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

@@ -0,0 +1,84 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+// 16-bit transpose, loading/storing a 4x4 tile of elements
+kernel void kernel_transpose_16(
+    __read_only image1d_buffer_t input,
+    __write_only image1d_buffer_t output,
+    const uint rows,
+    const uint cols
+) {
+
+    const int i = get_global_id(0);
+    const int j = get_global_id(1);
+    const int i_2 = i<<2;
+    const int j_2 = j<<2;
+
+    half4 temp0 = read_imageh(input, (j_2+0)*cols+i);
+    half4 temp1 = read_imageh(input, (j_2+1)*cols+i);
+    half4 temp2 = read_imageh(input, (j_2+2)*cols+i);
+    half4 temp3 = read_imageh(input, (j_2+3)*cols+i);
+
+    write_imageh(output, (i_2+0)*rows+j, (half4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0));
+    write_imageh(output, (i_2+1)*rows+j, (half4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1));
+    write_imageh(output, (i_2+2)*rows+j, (half4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2));
+    write_imageh(output, (i_2+3)*rows+j, (half4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
+}
+
+// 32-bit transpose, loading/storing a 4x4 tile of elements
+kernel void kernel_transpose_32(
+    __read_only image1d_buffer_t input,
+    __write_only image1d_buffer_t output,
+    const uint rows,
+    const uint cols
+) {
+
+    const int i = get_global_id(0);
+    const int j = get_global_id(1);
+    const int i_2 = i<<2;
+    const int j_2 = j<<2;
+
+    float4 temp0 = read_imagef(input, (j_2+0)*cols+i);
+    float4 temp1 = read_imagef(input, (j_2+1)*cols+i);
+    float4 temp2 = read_imagef(input, (j_2+2)*cols+i);
+    float4 temp3 = read_imagef(input, (j_2+3)*cols+i);
+
+    write_imagef(output, (i_2+0)*rows+j, (float4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0));
+    write_imagef(output, (i_2+1)*rows+j, (float4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1));
+    write_imagef(output, (i_2+2)*rows+j, (float4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2));
+    write_imagef(output, (i_2+3)*rows+j, (float4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
+
+}
+
+// 32-bit transpose, loading/storing a 4x4 tile of elements
+// Only used for activations
+// converts to FP16
+// also adds zero padding for non multiple of 8 prompt lengths
+kernel void kernel_transpose_32_16(__read_only image1d_buffer_t input, __write_only image1d_buffer_t output, const uint rows, const uint cols, const uint padded_rows) {
+
+    const int i = get_global_id(0);
+    const int j = get_global_id(1);
+    const int i_2 = i<<2;
+    const int j_2 = j<<2;
+    half4 temp0 = {0,0,0,0}; // initialize outputs to 0
+    half4 temp1 = {0,0,0,0};
+    half4 temp2 = {0,0,0,0};
+    half4 temp3 = {0,0,0,0};
+
+    if((j_2+0)*cols+i*4+3 < rows*cols*16){ // only load from a valid location. Otherwise keep register data as 0
+        temp0 = read_imageh(input, (j_2+0)*cols+i);
+    }
+    if((j_2+1)*cols+i*4+3 < rows*cols*16){
+        temp1 = read_imageh(input, (j_2+1)*cols+i);
+    }
+    if((j_2+2)*cols+i*4+3 < rows*cols*16){
+        temp2 = read_imageh(input, (j_2+2)*cols+i);
+    }
+    if((j_2+3)*cols+i*4+3 < rows*cols*16){
+        temp3 = read_imageh(input, (j_2+3)*cols+i);
+    }
+
+    write_imageh(output, (i_2+0)*padded_rows+j, (half4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0)); // no conditionals for output, includes zero padding
+    write_imageh(output, (i_2+1)*padded_rows+j, (half4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1));
+    write_imageh(output, (i_2+2)*padded_rows+j, (half4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2));
+    write_imageh(output, (i_2+3)*padded_rows+j, (half4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
+}

ggml/src/ggml-rpc/ggml-rpc.cpp

@@ -1,6 +1,7 @@
 #include "ggml-rpc.h"
 #include "ggml-impl.h"
 #include "ggml-backend-impl.h"
+#include "ggml-cpp.h"
 
 #include <cinttypes>
 #include <string>
@@ -91,12 +92,19 @@ enum rpc_cmd {
     RPC_CMD_GET_DEVICE_MEMORY,
     RPC_CMD_INIT_TENSOR,
     RPC_CMD_GET_ALLOC_SIZE,
+    RPC_CMD_HELLO,
     RPC_CMD_COUNT,
 };
 
 // Try RPC_CMD_SET_TENSOR_HASH first when data size is larger than this threshold
 const size_t HASH_THRESHOLD = 10 * 1024 * 1024;
 
+struct rpc_msg_hello_rsp {
+    uint8_t major;
+    uint8_t minor;
+    uint8_t patch;
+};
+
 struct rpc_msg_get_alloc_size_req {
     rpc_tensor tensor;
 };
@@ -399,6 +407,20 @@ static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cm
 
 // RPC client-side implementation
 
+static bool check_server_version(const std::shared_ptr<socket_t> & sock) {
+    rpc_msg_hello_rsp response;
+    bool status = send_rpc_cmd(sock, RPC_CMD_HELLO, nullptr, 0, &response, sizeof(response));
+    GGML_ASSERT(status);
+    if (response.major != RPC_PROTO_MAJOR_VERSION || response.minor > RPC_PROTO_MINOR_VERSION) {
+        fprintf(stderr, "RPC server version mismatch: %d.%d.%d\n", response.major, response.minor, response.patch);
+        return false;
+    }
+    if (response.minor != RPC_PROTO_MINOR_VERSION || response.patch != RPC_PROTO_PATCH_VERSION) {
+        fprintf(stderr, "WARNING: RPC server version mismatch: %d.%d.%d\n", response.major, response.minor, response.patch);
+    }
+    return true;
+}
+
 static std::shared_ptr<socket_t> get_socket(const std::string & endpoint) {
     static std::mutex mutex;
     std::lock_guard<std::mutex> lock(mutex);
@@ -432,6 +454,9 @@ static std::shared_ptr<socket_t> get_socket(const std::string & endpoint) {
     if (sock == nullptr) {
         return nullptr;
     }
+    if (!check_server_version(sock)) {
+        return nullptr;
+    }
     GGML_PRINT_DEBUG("[%s] connected to %s, sockfd=%d\n", __func__, endpoint.c_str(), sock->fd);
     sockets[endpoint] = sock;
     return sock;
@@ -817,6 +842,7 @@ public:
     }
     ~rpc_server();
 
+    void hello(rpc_msg_hello_rsp & response);
     void alloc_buffer(const rpc_msg_alloc_buffer_req & request, rpc_msg_alloc_buffer_rsp & response);
     void get_alignment(rpc_msg_get_alignment_rsp & response);
     void get_max_size(rpc_msg_get_max_size_rsp & response);
@@ -845,6 +871,13 @@ private:
     std::unordered_set<ggml_backend_buffer_t> buffers;
 };
 
+void rpc_server::hello(rpc_msg_hello_rsp & response) {
+    response.major = RPC_PROTO_MAJOR_VERSION;
+    response.minor = RPC_PROTO_MINOR_VERSION;
+    response.patch = RPC_PROTO_PATCH_VERSION;
+    GGML_PRINT_DEBUG("[%s] version: %d.%d.%d\n", __func__, response.major, response.minor, response.patch);
+}
+
 bool rpc_server::get_alloc_size(const rpc_msg_get_alloc_size_req & request, rpc_msg_get_alloc_size_rsp & response) {
     ggml_backend_buffer_type_t buft;
     struct ggml_init_params params {
@@ -853,12 +886,13 @@ bool rpc_server::get_alloc_size(const rpc_msg_get_alloc_size_req & request, rpc_
         /*.no_alloc   =*/ true,
     };
 
-    struct ggml_context * ctx = ggml_init(params);
+    ggml_context_ptr ctx_ptr { ggml_init(params) };
+    GGML_ASSERT(ctx_ptr != nullptr);
+    ggml_context * ctx = ctx_ptr.get();
     ggml_tensor * tensor = deserialize_tensor(ctx, &request.tensor);
 
     if (tensor == nullptr) {
         GGML_LOG_ERROR("Null tensor pointer passed to server get_alloc_size function.\n");
-        ggml_free(ctx);
         return false;
     }
 
@@ -871,7 +905,6 @@ bool rpc_server::get_alloc_size(const rpc_msg_get_alloc_size_req & request, rpc_
 
     response.alloc_size = ggml_backend_buft_get_alloc_size(buft,tensor);
 
-    ggml_free(ctx);
     return true;
 }
 
@@ -985,11 +1018,12 @@ bool rpc_server::set_tensor(const std::vector<uint8_t> & input) {
         /*.mem_buffer =*/ NULL,
         /*.no_alloc   =*/ true,
     };
-    struct ggml_context * ctx = ggml_init(params);
+    ggml_context_ptr ctx_ptr { ggml_init(params) };
+    GGML_ASSERT(ctx_ptr != nullptr);
+    ggml_context * ctx = ctx_ptr.get();
     ggml_tensor * tensor = deserialize_tensor(ctx, in_tensor);
     if (tensor == nullptr) {
         GGML_LOG_ERROR("[%s] error deserializing tensor\n", __func__);
-        ggml_free(ctx);
         return false;
     }
     GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %zu\n", __func__, (void*)tensor->buffer, tensor->data, offset, size);
@@ -1016,7 +1050,6 @@ bool rpc_server::set_tensor(const std::vector<uint8_t> & input) {
         printf("[%s] saved to '%s'\n", __func__, cache_file.c_str());
     }
     ggml_backend_tensor_set(tensor, data, offset, size);
-    ggml_free(ctx);
     return true;
 }
 
@@ -1060,11 +1093,12 @@ bool rpc_server::set_tensor_hash(const std::vector<uint8_t> & input, rpc_msg_set
         /*.mem_buffer =*/ NULL,
         /*.no_alloc   =*/ true,
     };
-    struct ggml_context * ctx = ggml_init(params);
+    ggml_context_ptr ctx_ptr { ggml_init(params) };
+    GGML_ASSERT(ctx_ptr != nullptr);
+    ggml_context * ctx = ctx_ptr.get();
     ggml_tensor * tensor = deserialize_tensor(ctx, in_tensor);
     if (tensor == nullptr) {
         GGML_LOG_ERROR("[%s] error deserializing tensor\n", __func__);
-        ggml_free(ctx);
         return false;
     }
     GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %zu, hash: %" PRIx64 "\n", __func__, (void*)tensor->buffer, tensor->data, offset, size, *hash);
@@ -1080,7 +1114,6 @@ bool rpc_server::set_tensor_hash(const std::vector<uint8_t> & input, rpc_msg_set
     }
     ggml_backend_tensor_set(tensor, cached_file.data(), offset, size);
     response.result = 1;
-    ggml_free(ctx);
     return true;
 }
 
@@ -1090,11 +1123,12 @@ bool rpc_server::init_tensor(const rpc_msg_init_tensor_req & request) {
         /*.mem_buffer =*/ NULL,
         /*.no_alloc   =*/ true,
     };
-    struct ggml_context * ctx = ggml_init(params);
+    ggml_context_ptr ctx_ptr { ggml_init(params) };
+    GGML_ASSERT(ctx_ptr != nullptr);
+    ggml_context * ctx = ctx_ptr.get();
     ggml_tensor * tensor = deserialize_tensor(ctx, &request.tensor);
     if (tensor == nullptr) {
         GGML_LOG_ERROR("Null tensor pointer passed to server init_tensor function.\n");
-        ggml_free(ctx);
         return false;
     }
 
@@ -1110,11 +1144,9 @@ bool rpc_server::init_tensor(const rpc_msg_init_tensor_req & request) {
         // This pointer can either be passed around client/server, or probably better stored server-side and kept track of.
         // Currently unimplemented.
         GGML_LOG_ERROR("tensor->extra populated by the backend, this is currently unsupported.\n");
-        ggml_free(ctx);
         return false;
     }
 
-    ggml_free(ctx);
     return true;
 }
 
@@ -1124,11 +1156,12 @@ bool rpc_server::get_tensor(const rpc_msg_get_tensor_req & request, std::vector<
         /*.mem_buffer =*/ NULL,
         /*.no_alloc   =*/ true,
     };
-    struct ggml_context * ctx = ggml_init(params);
+    ggml_context_ptr ctx_ptr { ggml_init(params) };
+    GGML_ASSERT(ctx_ptr != nullptr);
+    ggml_context * ctx = ctx_ptr.get();
     ggml_tensor * tensor = deserialize_tensor(ctx, &request.tensor);
     if (tensor == nullptr) {
         GGML_LOG_ERROR("[%s] error deserializing tensor\n", __func__);
-        ggml_free(ctx);
         return false;
     }
     GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %" PRIu64 "\n", __func__, (void*)tensor->buffer, tensor->data, request.offset, request.size);
@@ -1147,7 +1180,6 @@ bool rpc_server::get_tensor(const rpc_msg_get_tensor_req & request, std::vector<
 
     response.resize(request.size, 0);
     ggml_backend_tensor_get(tensor, response.data(), request.offset, request.size);
-    ggml_free(ctx);
     return true;
 }
 
@@ -1157,12 +1189,14 @@ bool rpc_server::copy_tensor(const rpc_msg_copy_tensor_req & request, rpc_msg_co
         /*.mem_buffer =*/ NULL,
         /*.no_alloc   =*/ true,
     };
-    struct ggml_context * ctx = ggml_init(params);
+    ggml_context_ptr ctx_ptr { ggml_init(params) };
+    GGML_ASSERT(ctx_ptr != nullptr);
+    ggml_context * ctx = ctx_ptr.get();
+
     ggml_tensor * src = deserialize_tensor(ctx, &request.src);
     ggml_tensor * dst = deserialize_tensor(ctx, &request.dst);
     if (src == nullptr || dst == nullptr) {
         GGML_LOG_ERROR("[%s] error deserializing tensors\n", __func__);
-        ggml_free(ctx);
         return false;
     }
 
@@ -1180,7 +1214,6 @@ bool rpc_server::copy_tensor(const rpc_msg_copy_tensor_req & request, rpc_msg_co
                          dst_data + src_size,
                          dst_base,
                          dst_base + dst_buf_sz);
-        ggml_free(ctx);
         return false;
     }
 
@@ -1188,7 +1221,6 @@ bool rpc_server::copy_tensor(const rpc_msg_copy_tensor_req & request, rpc_msg_co
                      __func__, (void*) src->buffer, (void*) dst->buffer);
 
     response.result = ggml_backend_buffer_copy_tensor(src, dst);
-    ggml_free(ctx);
     return true;
 }
 
@@ -1242,7 +1274,9 @@ bool rpc_server::graph_compute(const std::vector<uint8_t> & input, rpc_msg_graph
         /*.mem_buffer =*/ NULL,
         /*.no_alloc   =*/ true,
     };
-    struct ggml_context * ctx = ggml_init(params);
+    ggml_context_ptr ctx_ptr { ggml_init(params) };
+    GGML_ASSERT(ctx_ptr != nullptr);
+    ggml_context * ctx = ctx_ptr.get();
     struct ggml_cgraph * graph = ggml_new_graph_custom(ctx, n_nodes, false);
     graph->n_nodes = n_nodes;
     std::unordered_map<uint64_t, const rpc_tensor*> tensor_ptrs;
@@ -1257,7 +1291,6 @@ bool rpc_server::graph_compute(const std::vector<uint8_t> & input, rpc_msg_graph
     }
     ggml_status status = ggml_backend_graph_compute(backend, graph);
     response.result = status;
-    ggml_free(ctx);
     return true;
 }
 
@@ -1270,8 +1303,24 @@ rpc_server::~rpc_server() {
 static void rpc_serve_client(ggml_backend_t backend, const char * cache_dir,
                              sockfd_t sockfd, size_t free_mem, size_t total_mem) {
     rpc_server server(backend, cache_dir);
+    uint8_t cmd;
+    if (!recv_data(sockfd, &cmd, 1)) {
+        return;
+    }
+    // the first command sent by the client must be HELLO
+    if (cmd != RPC_CMD_HELLO) {
+        fprintf(stderr, "Expected HELLO command, update client\n");
+        return;
+    }
+    if (!recv_msg(sockfd, nullptr, 0)) {
+        return;
+    }
+    rpc_msg_hello_rsp response;
+    server.hello(response);
+    if (!send_msg(sockfd, &response, sizeof(response))) {
+        return;
+    }
     while (true) {
-        uint8_t cmd;
         if (!recv_data(sockfd, &cmd, 1)) {
             break;
         }
@@ -1281,6 +1330,10 @@ static void rpc_serve_client(ggml_backend_t backend, const char * cache_dir,
             break;
         }
         switch (cmd) {
+            case RPC_CMD_HELLO: {
+                // HELLO command is handled above
+                return;
+            }
             case RPC_CMD_ALLOC_BUFFER: {
                 rpc_msg_alloc_buffer_req request;
                 if (!recv_msg(sockfd, &request, sizeof(request))) {

ggml/src/ggml-sycl/backend.hpp

@@ -13,6 +13,7 @@
 #ifndef GGML_SYCL_BACKEND_HPP
 #define GGML_SYCL_BACKEND_HPP
 
+#include "binbcast.hpp"
 #include "concat.hpp"
 #include "common.hpp"
 #include "conv.hpp"

ggml/src/ggml-sycl/binbcast.cpp

@@ -0,0 +1,350 @@
+#include "binbcast.hpp"
+
+#include <cstddef>
+#include <cstdint>
+#include <sycl/sycl.hpp>
+
+#include "ggml.h"
+
+template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
+static void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst,
+        int ne0, int ne1, int ne2, int ne3,
+        int ne10, int ne11, int ne12, int ne13,
+        /*int s0, */ int s1,  int s2,  int s3,
+        /*int s00,*/ int s01, int s02, int s03,
+        /*int s10,*/ int s11, int s12, int s13,
+        const sycl::nd_item<3> &item_ct1) {
+    const int i0s = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                    item_ct1.get_local_id(2);
+    const int i1 = (item_ct1.get_local_range(1) * item_ct1.get_group(1) +
+                    item_ct1.get_local_id(1));
+    const int i2 = (item_ct1.get_local_range(0) * item_ct1.get_group(0) +
+                    item_ct1.get_local_id(0)) /
+                   ne3;
+    const int i3 = (item_ct1.get_local_range(0) * item_ct1.get_group(0) +
+                    item_ct1.get_local_id(0)) %
+                   ne3;
+
+    if (i0s >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
+        return;
+    }
+
+    const int i11 = i1 % ne11;
+    const int i12 = i2 % ne12;
+    const int i13 = i3 % ne13;
+
+    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
+    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
+    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
+
+    const src0_t * src0_row = src0 + i_src0;
+    const src1_t * src1_row = src1 + i_src1;
+    dst_t * dst_row = dst + i_dst;
+
+    for (int i0 = i0s; i0 < ne0;
+         i0 += item_ct1.get_local_range(2) * item_ct1.get_group_range(2)) {
+        const int i10 = i0 % ne10;
+        dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
+    }
+}
+
+template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
+static void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t * dst,
+        int ne0, int ne1, int ne2, int ne3,
+        int ne10, int ne11, int ne12, int ne13,
+        /*int s0, */ int s1,  int s2,  int s3,
+        /*int s00,*/ int s01, int s02, int s03,
+        /*int s10,*/ int s11, int s12, int s13,
+        const sycl::nd_item<3> &item_ct1) {
+
+    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                  item_ct1.get_local_id(2);
+
+    const int i3 = i/(ne2*ne1*ne0);
+    const int i2 = (i/(ne1*ne0)) % ne2;
+    const int i1 = (i/ne0) % ne1;
+    const int i0 = i % ne0;
+
+    if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
+        return;
+    }
+
+    const int i11 = i1 % ne11;
+    const int i12 = i2 % ne12;
+    const int i13 = i3 % ne13;
+
+    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
+    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
+    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
+
+    const src0_t * src0_row = src0 + i_src0;
+    const src1_t * src1_row = src1 + i_src1;
+    dst_t * dst_row = dst + i_dst;
+
+    const int i10 = i0 % ne10;
+    dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
+}
+
+
+template<float (*bin_op)(const float, const float)>
+struct bin_bcast_sycl {
+    template <typename src0_t, typename src1_t, typename dst_t>
+    void operator()(const src0_t * src0_dd, const src1_t * src1_dd, dst_t * dst_dd, const int64_t ne00,
+                    const int64_t ne01, const int64_t ne02, const int64_t ne03, const int64_t ne10, const int64_t ne11,
+                    const int64_t ne12, const int64_t ne13, const int64_t ne0, const int64_t ne1, const int64_t ne2,
+                    const int64_t ne3, const size_t nb00, const size_t nb01, const size_t nb02, const size_t nb03,
+                    const size_t nb10, const size_t nb11, const size_t nb12, const size_t nb13, const size_t nb0,
+                    const size_t nb1, const size_t nb2, const size_t nb3, const bool src0_is_contiguous,
+                    const bool src1_is_contiguous, const bool dst_is_contiguous, queue_ptr stream) {
+        int nr0 = ne10 / ne0;
+        int nr1 = ne11/ne1;
+        int nr2 = ne12/ne2;
+        int nr3 = ne13/ne3;
+
+        int nr[4] = { nr0, nr1, nr2, nr3 };
+
+        // collapse dimensions until first broadcast dimension
+        int64_t cne[] = {ne0, ne1, ne2, ne3};
+        int64_t cne0[] = {ne00, ne01, ne02, ne03};
+        int64_t cne1[] = {ne10, ne11, ne12, ne13};
+        size_t cnb[] = {nb0, nb1, nb2, nb3};
+        size_t cnb0[] = {nb00, nb01, nb02, nb03};
+        size_t cnb1[] = {nb10, nb11, nb12, nb13};
+        auto collapse = [](int64_t cne[]) {
+            cne[0] *= cne[1];
+            cne[1] = cne[2];
+            cne[2] = cne[3];
+            cne[3] = 1;
+        };
+
+        auto collapse_nb = [](size_t cnb[], int64_t cne[]) {
+            cnb[1] *= cne[1];
+            cnb[2] *= cne[2];
+            cnb[3] *= cne[3];
+        };
+
+        if (src0_is_contiguous && src1_is_contiguous && dst_is_contiguous) {
+            for (int i = 0; i < 4; i++) {
+                if (nr[i] != 1) {
+                    break;
+                }
+                if (i > 0) {
+                    collapse_nb(cnb, cne);
+                    collapse_nb(cnb0, cne0);
+                    collapse_nb(cnb1, cne1);
+                    collapse(cne);
+                    collapse(cne0);
+                    collapse(cne1);
+                }
+            }
+        }
+        {
+            int64_t ne0 = cne[0];
+            int64_t ne1 = cne[1];
+            int64_t ne2 = cne[2];
+            int64_t ne3 = cne[3];
+
+            int64_t ne10 = cne1[0];
+            int64_t ne11 = cne1[1];
+            int64_t ne12 = cne1[2];
+            int64_t ne13 = cne1[3];
+
+            size_t nb0 = cnb[0];
+            size_t nb1 = cnb[1];
+            size_t nb2 = cnb[2];
+            size_t nb3 = cnb[3];
+
+            size_t nb00 = cnb0[0];
+            size_t nb01 = cnb0[1];
+            size_t nb02 = cnb0[2];
+            size_t nb03 = cnb0[3];
+
+            size_t nb10 = cnb1[0];
+            size_t nb11 = cnb1[1];
+            size_t nb12 = cnb1[2];
+            size_t nb13 = cnb1[3];
+
+            size_t s0 = nb0 / sizeof(dst_t);
+            size_t s1 = nb1 / sizeof(dst_t);
+            size_t s2 = nb2 / sizeof(dst_t);
+            size_t s3 = nb3 / sizeof(dst_t);
+
+            size_t s10 = nb10 / sizeof(src1_t);
+            size_t s11 = nb11 / sizeof(src1_t);
+            size_t s12 = nb12 / sizeof(src1_t);
+            size_t s13 = nb13 / sizeof(src1_t);
+
+            size_t s00 = nb00 / sizeof(src0_t);
+            size_t s01 = nb01 / sizeof(src0_t);
+            size_t s02 = nb02 / sizeof(src0_t);
+            size_t s03 = nb03 / sizeof(src0_t);
+
+            GGML_UNUSED(s00);
+
+            GGML_ASSERT(nb0 % sizeof(dst_t) == 0);
+            GGML_ASSERT(nb1 % sizeof(dst_t) == 0);
+            GGML_ASSERT(nb2 % sizeof(dst_t) == 0);
+            GGML_ASSERT(nb3 % sizeof(dst_t) == 0);
+
+            GGML_ASSERT(nb00 % sizeof(src0_t) == 0);
+            GGML_ASSERT(nb01 % sizeof(src0_t) == 0);
+            GGML_ASSERT(nb02 % sizeof(src0_t) == 0);
+            GGML_ASSERT(nb03 % sizeof(src0_t) == 0);
+
+            GGML_ASSERT(nb10 % sizeof(src1_t) == 0);
+            GGML_ASSERT(nb11 % sizeof(src1_t) == 0);
+            GGML_ASSERT(nb12 % sizeof(src1_t) == 0);
+            GGML_ASSERT(nb13 % sizeof(src1_t) == 0);
+
+            GGML_ASSERT(s0 == 1);
+            GGML_ASSERT(s10 == 1);
+
+            const int block_size = 128;
+
+            int64_t hne0 = std::max(ne0/2LL, 1LL);
+
+            sycl::range<3> block_dims(1, 1, 1);
+            block_dims[2] = std::min<unsigned int>(hne0, block_size);
+            block_dims[1] = std::min<unsigned int>(
+                ne1, block_size / (unsigned int)block_dims[2]);
+            block_dims[0] = std::min(
+                std::min<unsigned int>(
+                    ne2 * ne3, block_size / (unsigned int)block_dims[2] /
+                                   (unsigned int)block_dims[1]),
+                64U);
+
+            sycl::range<3> block_nums(
+                (ne2 * ne3 + block_dims[0] - 1) / block_dims[0],
+                (ne1 + block_dims[1] - 1) / block_dims[1],
+                (hne0 + block_dims[2] - 1) / block_dims[2]);
+
+            if (block_nums[0] > 65535) {
+                // this is the maximum number of blocks in z direction, fallback to 1D grid kernel
+                int block_num = (ne0*ne1*ne2*ne3 + block_size - 1) / block_size;
+                {
+                    dpct::has_capability_or_fail(stream->get_device(),
+                                                 {sycl::aspect::fp16});
+
+                    stream->parallel_for(
+                        sycl::nd_range<3>(sycl::range<3>(1, 1, block_num) *
+                                              sycl::range<3>(1, 1, block_size),
+                                          sycl::range<3>(1, 1, block_size)),
+                        [=](sycl::nd_item<3> item_ct1) {
+                            k_bin_bcast_unravel<bin_op>(
+                                src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3,
+                                ne10, ne11, ne12, ne13, s1, s2, s3, s01, s02,
+                                s03, s11, s12, s13, item_ct1);
+                        });
+                }
+            } else {
+                /*
+                DPCT1049:16: The work-group size passed to the SYCL kernel may
+                exceed the limit. To get the device limit, query
+                info::device::max_work_group_size. Adjust the work-group size if
+                needed.
+                */
+                dpct::has_capability_or_fail(stream->get_device(),
+                                             {sycl::aspect::fp16});
+
+                stream->parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
+                        k_bin_bcast<bin_op>(src0_dd, src1_dd, dst_dd, ne0, ne1,
+                                            ne2, ne3, ne10, ne11, ne12, ne13,
+                                            s1, s2, s3, s01, s02, s03, s11, s12, s13,
+                                            item_ct1);
+                    });
+            }
+        }
+    }
+};
+
+template <class op>
+inline void ggml_sycl_op_bin_bcast(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1,
+                                   ggml_tensor * dst) {
+    dpct::queue_ptr main_stream = ctx.stream();
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+        op()((const float *) src0->data, (const float *) src1->data, (float *) dst->data, ne00, ne01, ne02, ne03, ne10,
+             ne11, ne12, ne13, ne0, ne1, ne2, ne3, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb0, nb1, nb2, nb3,
+             ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_contiguous(dst), main_stream);
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
+        op()((const sycl::half *) src0->data, (const sycl::half *) src1->data, (sycl::half *) dst->data, ne00, ne01,
+             ne02, ne03, ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13,
+             nb0, nb1, nb2, nb3, ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_contiguous(dst),
+             main_stream);
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) {
+        op()((const sycl::half *) src0->data, (const float *) src1->data, (sycl::half *) dst->data, ne00, ne01, ne02,
+             ne03, ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb0, nb1,
+             nb2, nb3, ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_contiguous(dst), main_stream);
+    } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32 && dst->type == GGML_TYPE_I32) {
+        op()((const int32_t *) src0->data, (const int32_t *) src1->data, (int32_t *) dst->data, ne00, ne01, ne02, ne03,
+             ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb0, nb1, nb2,
+             nb3, ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_contiguous(dst), main_stream);
+    } else if (src0->type == GGML_TYPE_I16 && src1->type == GGML_TYPE_I16 && dst->type == GGML_TYPE_I16) {
+        op()((const int16_t *) src0->data, (const int16_t *) src1->data, (int16_t *) dst->data, ne00, ne01, ne02, ne03,
+             ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb0, nb1, nb2,
+             nb3, ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_contiguous(dst), main_stream);
+    } else {
+        fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__, ggml_type_name(dst->type),
+                ggml_type_name(src0->type), ggml_type_name(src1->type));
+        GGML_ABORT("fatal error");
+    }
+}
+
+inline void ggml_sycl_op_add(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
+
+    ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_add>>(ctx, dst->src[0], dst->src[1], dst);
+}
+
+inline void ggml_sycl_op_sub(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
+
+    ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_sub>>(ctx, dst->src[0], dst->src[1], dst);
+}
+
+inline void ggml_sycl_op_mul(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
+
+    ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_mul>>(ctx, dst->src[0], dst->src[1], dst);
+}
+
+inline void ggml_sycl_op_div(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
+
+    ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_div>>(ctx, dst->src[0], dst->src[1], dst);
+}
+
+inline void ggml_sycl_op_repeat(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
+    ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_repeat>>(ctx, dst, dst->src[0], dst);
+}
+
+
+void ggml_sycl_add(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_add(ctx, dst);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+void ggml_sycl_sub(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_sub(ctx, dst);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+void ggml_sycl_mul(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_mul(ctx, dst);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+void ggml_sycl_div(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_div(ctx, dst);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+void ggml_sycl_repeat(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_repeat(ctx, dst);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+

ggml/src/ggml-sycl/binbcast.hpp

@@ -0,0 +1,39 @@
+#ifndef GGML_SYCL_BINBCAST_HPP
+#define GGML_SYCL_BINBCAST_HPP
+#include "common.hpp"
+
+
+static __dpct_inline__ float op_repeat(const float a, const float b) {
+    return b;
+    GGML_UNUSED(a);
+}
+
+static __dpct_inline__ float op_add(const float a, const float b) {
+    return a + b;
+}
+
+static __dpct_inline__ float op_sub(const float a, const float b) {
+    return a - b;
+}
+
+static __dpct_inline__ float op_mul(const float a, const float b) {
+    return a * b;
+}
+
+static __dpct_inline__ float op_div(const float a, const float b) {
+    return a / b;
+}
+
+void ggml_sycl_add(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+void ggml_sycl_sub(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+void ggml_sycl_mul(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+void ggml_sycl_div(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+void ggml_sycl_repeat(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+
+#endif //GGML_SYCL_BINBCAST_HPP
+

ggml/src/ggml-sycl/common.hpp

@@ -494,286 +494,5 @@ static __dpct_inline__ Tp* get_pointer(sycl::local_accessor<Tp, dim> acc) {
 
 int64_t downsample_sycl_global_range(int64_t accumulate_block_num, int64_t block_size);
 
-template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
-static void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst,
-        int ne0, int ne1, int ne2, int ne3,
-        int ne10, int ne11, int ne12, int ne13,
-        /*int s0, */ int s1,  int s2,  int s3,
-        /*int s00,*/ int s01, int s02, int s03,
-        /*int s10,*/ int s11, int s12, int s13,
-        const sycl::nd_item<3> &item_ct1) {
-    const int i0s = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                    item_ct1.get_local_id(2);
-    const int i1 = (item_ct1.get_local_range(1) * item_ct1.get_group(1) +
-                    item_ct1.get_local_id(1));
-    const int i2 = (item_ct1.get_local_range(0) * item_ct1.get_group(0) +
-                    item_ct1.get_local_id(0)) /
-                   ne3;
-    const int i3 = (item_ct1.get_local_range(0) * item_ct1.get_group(0) +
-                    item_ct1.get_local_id(0)) %
-                   ne3;
-
-    if (i0s >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
-        return;
-    }
-
-    const int i11 = i1 % ne11;
-    const int i12 = i2 % ne12;
-    const int i13 = i3 % ne13;
-
-    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
-    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
-    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
-
-    const src0_t * src0_row = src0 + i_src0;
-    const src1_t * src1_row = src1 + i_src1;
-    dst_t * dst_row = dst + i_dst;
-
-    for (int i0 = i0s; i0 < ne0;
-         i0 += item_ct1.get_local_range(2) * item_ct1.get_group_range(2)) {
-        const int i10 = i0 % ne10;
-        dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
-    }
-}
-
-template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
-static void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t * dst,
-        int ne0, int ne1, int ne2, int ne3,
-        int ne10, int ne11, int ne12, int ne13,
-        /*int s0, */ int s1,  int s2,  int s3,
-        /*int s00,*/ int s01, int s02, int s03,
-        /*int s10,*/ int s11, int s12, int s13,
-        const sycl::nd_item<3> &item_ct1) {
-
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
-
-    const int i3 = i/(ne2*ne1*ne0);
-    const int i2 = (i/(ne1*ne0)) % ne2;
-    const int i1 = (i/ne0) % ne1;
-    const int i0 = i % ne0;
-
-    if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
-        return;
-    }
-
-    const int i11 = i1 % ne11;
-    const int i12 = i2 % ne12;
-    const int i13 = i3 % ne13;
-
-    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
-    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
-    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
-
-    const src0_t * src0_row = src0 + i_src0;
-    const src1_t * src1_row = src1 + i_src1;
-    dst_t * dst_row = dst + i_dst;
-
-    const int i10 = i0 % ne10;
-    dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
-}
-
-
-template<float (*bin_op)(const float, const float)>
-struct bin_bcast_sycl {
-    template <typename src0_t, typename src1_t, typename dst_t>
-    void operator()(ggml_backend_sycl_context & ctx,
-                    const struct ggml_tensor *src0,
-                    const struct ggml_tensor *src1, struct ggml_tensor *dst,
-                    const src0_t *src0_dd, const src1_t *src1_dd, dst_t *dst_dd,
-                    queue_ptr stream) {
-
-        GGML_TENSOR_BINARY_OP_LOCALS
-
-        int nr0 = ne10/ne0;
-        int nr1 = ne11/ne1;
-        int nr2 = ne12/ne2;
-        int nr3 = ne13/ne3;
-
-        int nr[4] = { nr0, nr1, nr2, nr3 };
-
-        // collapse dimensions until first broadcast dimension
-        int64_t cne[] = {ne0, ne1, ne2, ne3};
-        int64_t cne0[] = {ne00, ne01, ne02, ne03};
-        int64_t cne1[] = {ne10, ne11, ne12, ne13};
-        size_t cnb[] = {nb0, nb1, nb2, nb3};
-        size_t cnb0[] = {nb00, nb01, nb02, nb03};
-        size_t cnb1[] = {nb10, nb11, nb12, nb13};
-        auto collapse = [](int64_t cne[]) {
-            cne[0] *= cne[1];
-            cne[1] = cne[2];
-            cne[2] = cne[3];
-            cne[3] = 1;
-        };
-
-        auto collapse_nb = [](size_t cnb[], int64_t cne[]) {
-            cnb[1] *= cne[1];
-            cnb[2] *= cne[2];
-            cnb[3] *= cne[3];
-        };
-
-        if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
-            for (int i = 0; i < 4; i++) {
-                if (nr[i] != 1) {
-                    break;
-                }
-                if (i > 0) {
-                    collapse_nb(cnb, cne);
-                    collapse_nb(cnb0, cne0);
-                    collapse_nb(cnb1, cne1);
-                    collapse(cne);
-                    collapse(cne0);
-                    collapse(cne1);
-                }
-            }
-        }
-        {
-            int64_t ne0 = cne[0];
-            int64_t ne1 = cne[1];
-            int64_t ne2 = cne[2];
-            int64_t ne3 = cne[3];
-
-            int64_t ne10 = cne1[0];
-            int64_t ne11 = cne1[1];
-            int64_t ne12 = cne1[2];
-            int64_t ne13 = cne1[3];
-
-            size_t nb0 = cnb[0];
-            size_t nb1 = cnb[1];
-            size_t nb2 = cnb[2];
-            size_t nb3 = cnb[3];
-
-            size_t nb00 = cnb0[0];
-            size_t nb01 = cnb0[1];
-            size_t nb02 = cnb0[2];
-            size_t nb03 = cnb0[3];
-
-            size_t nb10 = cnb1[0];
-            size_t nb11 = cnb1[1];
-            size_t nb12 = cnb1[2];
-            size_t nb13 = cnb1[3];
-
-            size_t s0 = nb0 / sizeof(dst_t);
-            size_t s1 = nb1 / sizeof(dst_t);
-            size_t s2 = nb2 / sizeof(dst_t);
-            size_t s3 = nb3 / sizeof(dst_t);
-
-            size_t s10 = nb10 / sizeof(src1_t);
-            size_t s11 = nb11 / sizeof(src1_t);
-            size_t s12 = nb12 / sizeof(src1_t);
-            size_t s13 = nb13 / sizeof(src1_t);
-
-            size_t s00 = nb00 / sizeof(src0_t);
-            size_t s01 = nb01 / sizeof(src0_t);
-            size_t s02 = nb02 / sizeof(src0_t);
-            size_t s03 = nb03 / sizeof(src0_t);
-
-            GGML_UNUSED(s00);
-
-            GGML_ASSERT(nb0 % sizeof(dst_t) == 0);
-            GGML_ASSERT(nb1 % sizeof(dst_t) == 0);
-            GGML_ASSERT(nb2 % sizeof(dst_t) == 0);
-            GGML_ASSERT(nb3 % sizeof(dst_t) == 0);
-
-            GGML_ASSERT(nb00 % sizeof(src0_t) == 0);
-            GGML_ASSERT(nb01 % sizeof(src0_t) == 0);
-            GGML_ASSERT(nb02 % sizeof(src0_t) == 0);
-            GGML_ASSERT(nb03 % sizeof(src0_t) == 0);
-
-            GGML_ASSERT(nb10 % sizeof(src1_t) == 0);
-            GGML_ASSERT(nb11 % sizeof(src1_t) == 0);
-            GGML_ASSERT(nb12 % sizeof(src1_t) == 0);
-            GGML_ASSERT(nb13 % sizeof(src1_t) == 0);
-
-            GGML_ASSERT(s0 == 1);
-            GGML_ASSERT(s10 == 1);
-
-            const int block_size = 128;
-
-            int64_t hne0 = std::max(ne0/2LL, 1LL);
-
-            sycl::range<3> block_dims(1, 1, 1);
-            block_dims[2] = std::min<unsigned int>(hne0, block_size);
-            block_dims[1] = std::min<unsigned int>(
-                ne1, block_size / (unsigned int)block_dims[2]);
-            block_dims[0] = std::min(
-                std::min<unsigned int>(
-                    ne2 * ne3, block_size / (unsigned int)block_dims[2] /
-                                   (unsigned int)block_dims[1]),
-                64U);
-
-            sycl::range<3> block_nums(
-                (ne2 * ne3 + block_dims[0] - 1) / block_dims[0],
-                (ne1 + block_dims[1] - 1) / block_dims[1],
-                (hne0 + block_dims[2] - 1) / block_dims[2]);
-
-            if (block_nums[0] > 65535) {
-                // this is the maximum number of blocks in z direction, fallback to 1D grid kernel
-                int block_num = (ne0*ne1*ne2*ne3 + block_size - 1) / block_size;
-                {
-                    dpct::has_capability_or_fail(stream->get_device(),
-                                                 {sycl::aspect::fp16});
-
-                    stream->parallel_for(
-                        sycl::nd_range<3>(sycl::range<3>(1, 1, block_num) *
-                                              sycl::range<3>(1, 1, block_size),
-                                          sycl::range<3>(1, 1, block_size)),
-                        [=](sycl::nd_item<3> item_ct1) {
-                            k_bin_bcast_unravel<bin_op>(
-                                src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3,
-                                ne10, ne11, ne12, ne13, s1, s2, s3, s01, s02,
-                                s03, s11, s12, s13, item_ct1);
-                        });
-                }
-            } else {
-                /*
-                DPCT1049:16: The work-group size passed to the SYCL kernel may
-                exceed the limit. To get the device limit, query
-                info::device::max_work_group_size. Adjust the work-group size if
-                needed.
-                */
-                dpct::has_capability_or_fail(stream->get_device(),
-                                             {sycl::aspect::fp16});
-
-                stream->parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
-                        k_bin_bcast<bin_op>(src0_dd, src1_dd, dst_dd, ne0, ne1,
-                                            ne2, ne3, ne10, ne11, ne12, ne13,
-                                            s1, s2, s3, s01, s02, s03, s11, s12, s13,
-                                            item_ct1);
-                    });
-            }
-        }
-        GGML_UNUSED(ctx);
-    }
-};
-
-template <class op>
-inline void ggml_sycl_op_bin_bcast(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
-                                   const ggml_tensor *src1, ggml_tensor *dst) {
-    dpct::queue_ptr main_stream = ctx.stream();
-
-    if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-        op()(ctx, src0, src1, dst, (const float *)src0->data, (const float *)src1->data, (float *)dst->data, main_stream);
-    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
-        op()(ctx, src0, src1, dst, (const sycl::half *)src0->data, (const float *)src1->data,
-             (sycl::half *)dst->data, main_stream);
-    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
-        op()(ctx, src0, src1, dst, (const sycl::half *)src0->data, (const float *)src1->data, (float *)dst->data,
-             main_stream);
-    } else if (src0->type == GGML_TYPE_I32 && dst->type == GGML_TYPE_I32) {
-        op()(ctx, src0, src1, dst, (const int32_t *)src0->data, (const int32_t *)src1->data, (int32_t *)dst->data,
-             main_stream);
-    } else if (src0->type == GGML_TYPE_I16 && dst->type == GGML_TYPE_I16) {
-        op()(ctx, src0, src1, dst, (const int16_t *)src0->data, (const int16_t *)src1->data, (int16_t *)dst->data,
-             main_stream);
-    } else {
-        fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__,
-            ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type));
-        GGML_ABORT("fatal error");
-    }
-}
-
 bool gpu_has_xmx(sycl::device &dev);
 #endif // GGML_SYCL_COMMON_HPP

ggml/src/ggml-sycl/element_wise.cpp

@@ -1261,27 +1261,6 @@ inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, ggml_tensor *dst)
 }
 
 
-inline void ggml_sycl_op_add(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
-
-    ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_add>>(ctx, dst->src[0], dst->src[1], dst);
-}
-
-inline void ggml_sycl_op_sub(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
-
-    ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_sub>>(ctx, dst->src[0], dst->src[1], dst);
-}
-
-inline void ggml_sycl_op_mul(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
-
-    ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_mul>>(ctx, dst->src[0], dst->src[1], dst);
-}
-
-inline void ggml_sycl_op_div(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
-
-    ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_div>>(ctx, dst->src[0], dst->src[1], dst);
-}
-
-
 void ggml_sycl_sqrt(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type));
     ggml_sycl_op_sqrt(ctx, dst);
@@ -1409,29 +1388,3 @@ void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     GGML_SYCL_DEBUG("call %s done\n", __func__);
 }
 
-
-
-
-void ggml_sycl_add(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    GGML_SYCL_DEBUG("call %s\n", __func__);
-    ggml_sycl_op_add(ctx, dst);
-    GGML_SYCL_DEBUG("call %s done\n", __func__);
-}
-
-void ggml_sycl_sub(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    GGML_SYCL_DEBUG("call %s\n", __func__);
-    ggml_sycl_op_sub(ctx, dst);
-    GGML_SYCL_DEBUG("call %s done\n", __func__);
-}
-
-void ggml_sycl_mul(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    GGML_SYCL_DEBUG("call %s\n", __func__);
-    ggml_sycl_op_mul(ctx, dst);
-    GGML_SYCL_DEBUG("call %s done\n", __func__);
-}
-
-void ggml_sycl_div(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    GGML_SYCL_DEBUG("call %s\n", __func__);
-    ggml_sycl_op_div(ctx, dst);
-    GGML_SYCL_DEBUG("call %s done\n", __func__);
-}

ggml/src/ggml-sycl/element_wise.hpp

@@ -10,27 +10,6 @@ T neg_infinity() {
     return -std::numeric_limits<T>::infinity();
 }
 
-static __dpct_inline__ float op_repeat(const float a, const float b) {
-    return b;
-    GGML_UNUSED(a);
-}
-
-static __dpct_inline__ float op_add(const float a, const float b) {
-    return a + b;
-}
-
-static __dpct_inline__ float op_sub(const float a, const float b) {
-    return a - b;
-}
-
-static __dpct_inline__ float op_mul(const float a, const float b) {
-    return a * b;
-}
-
-static __dpct_inline__ float op_div(const float a, const float b) {
-    return a / b;
-}
-
 template<typename T>
 struct typed_data {
     const T * src;
@@ -87,14 +66,5 @@ void ggml_sycl_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
 void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
-// ---------
-
-void ggml_sycl_add(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
-
-void ggml_sycl_sub(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
-
-void ggml_sycl_mul(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
-
-void ggml_sycl_div(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
-
 #endif // GGML_SYCL_ELEMENTWISE_HPP
+

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

@@ -1967,11 +1967,6 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
 
-static void ggml_sycl_op_repeat(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
-    ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_repeat>>(ctx, dst, dst->src[0], dst);
-}
-
-
 inline void ggml_sycl_op_mul_mat_sycl(
     ggml_backend_sycl_context & ctx,
     const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst,
@@ -2600,12 +2595,6 @@ catch (sycl::exception const &exc) {
 }
 
 
-static void ggml_sycl_repeat(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    GGML_SYCL_DEBUG("call %s\n", __func__);
-    ggml_sycl_op_repeat(ctx, dst);
-    GGML_SYCL_DEBUG("call %s done\n", __func__);
-}
-
 static void ggml_sycl_get_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     GGML_SYCL_DEBUG("call %s\n", __func__);
     ggml_sycl_op_get_rows(ctx, dst);
@@ -3179,11 +3168,6 @@ static void ggml_sycl_diag_mask_inf(ggml_backend_sycl_context & ctx, ggml_tensor
     ggml_sycl_op_diag_mask_inf(ctx, dst);
 }
 
-static void ggml_sycl_rope(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    GGML_ASSERT(ggml_is_contiguous(dst->src[0])); // TODO: this restriction is temporary until non-cont support is implemented
-    ggml_sycl_op_rope(ctx, dst);
-}
-
 static void ggml_sycl_pool2d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     ggml_sycl_op_pool2d(ctx, dst);
 }
@@ -3972,7 +3956,6 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_ARGMAX:
         case GGML_OP_NONE:
         case GGML_OP_RESHAPE:
-        case GGML_OP_REPEAT:
         case GGML_OP_VIEW:
         case GGML_OP_PERMUTE:
         case GGML_OP_TRANSPOSE:
@@ -3982,7 +3965,8 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_SUB:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
-            return (op->src[0]->type == GGML_TYPE_F32);
+        case GGML_OP_REPEAT:
+            return true;
         case GGML_OP_SQR:
         case GGML_OP_SQRT:
         case GGML_OP_SIN:
@@ -4009,17 +3993,14 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_ROPE:
             {
                 const int mode = ((const int32_t *) op->op_params)[2];
-                if (mode & GGML_ROPE_TYPE_MROPE) {
+                // mode is not used as a bitmask in practice, the various rope type modes are independent implementations
+                if (mode == GGML_ROPE_TYPE_MROPE) {
                     return false;
                 }
-                if (mode & GGML_ROPE_TYPE_VISION) {
-                    return false;
-                }
-                return ggml_is_contiguous(op->src[0]);
+                return true;
             }
         case GGML_OP_IM2COL:
-            // TODO: add support for the new F32 operations
-            return op->src[0]->type == GGML_TYPE_F16;
+            return true;
         case GGML_OP_UPSCALE:
             return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST;
         case GGML_OP_POOL_2D:

ggml/src/ggml-sycl/im2col.cpp

@@ -12,110 +12,125 @@
 
 #include "im2col.hpp"
 
+#include <sycl/sycl.hpp>
+#include <type_traits>  // For std::is_same_v
+
+#include "ggml.h"
+
 template <typename T>
-static void im2col_kernel(
-        const float *x, T *dst, int64_t batch_offset, int64_t offset_delta,
-        int64_t IC, int64_t IW, int64_t IH, int64_t OH, int64_t OW, int64_t KW, int64_t KH,
-        int64_t pelements, int64_t CHW, int s0, int s1, int p0, int p1, int d0, int d1,
-        const sycl::nd_item<3> &item_ct1) {
+static void im2col_kernel(const float * x, T * dst, int64_t batch_offset, int64_t offset_delta, int64_t IC, int64_t IW,
+                          int64_t IH, int64_t OH, int64_t OW, int64_t KW, int64_t KH, int64_t pelements, int64_t CHW,
+                          int s0, int s1, int p0, int p1, int d0, int d1, const sycl::nd_item<3> & item_ct1) {
     const int64_t work_group_size = item_ct1.get_local_range(2);
-    const int64_t global_id = item_ct1.get_local_id(2) + work_group_size * item_ct1.get_group(2);
+    const int64_t global_id       = item_ct1.get_local_id(2) + (work_group_size * item_ct1.get_group(2));
 
     // make each work-item deal with more elements since sycl global range can not exceed max int
-    for (int64_t i = global_id; i < pelements; i += work_group_size * item_ct1.get_group_range(2)) {
-
+    for (int64_t i = global_id; i < pelements; i += (work_group_size * item_ct1.get_group_range(2))) {
         const int64_t ksize = OW * (KH > 1 ? KW : 1);
-        const int64_t kx = i / ksize;
-        const int64_t kd = kx * ksize;
-        const int64_t ky = (i - kd) / OW;
-        const int64_t ix = i % OW;
+        const int64_t kx    = i / ksize;
+        const int64_t kd    = kx * ksize;
+        const int64_t ky    = (i - kd) / OW;
+        const int64_t ix    = i % OW;
 
-        const int64_t  oh = item_ct1.get_group(1);
-        const int64_t  batch = item_ct1.get_group(0) / IC;
-        const int64_t  ic = item_ct1.get_group(0) % IC;
+        const int64_t oh    = item_ct1.get_group(1);
+        const int64_t batch = item_ct1.get_group(0) / IC;
+        const int64_t ic    = item_ct1.get_group(0) % IC;
 
-        const int64_t iiw = ix * s0 + kx * d0 - p0;
-        const int64_t iih = oh * s1 + ky * d1 - p1;
+        const int64_t iiw = (ix * s0) + (kx * d0) - p0;
+        const int64_t iih = (oh * s1) + (ky * d1) - p1;
 
-        const int64_t offset_dst =
-            ((batch * OH + oh) * OW + ix) * CHW +
-            (ic * (KW * KH) + ky * KW + kx);
+        const int64_t offset_dst = (((batch * OH + oh) * OW + ix) * CHW) + (ic * (KW * KH) + ky * KW + kx);
 
-        if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
-            dst[offset_dst] =
-                sycl::vec<float, 1>(0.0f)
-                    .convert<sycl::half, sycl::rounding_mode::automatic>()[0];
-        } else {
-            const int64_t offset_src = ic * offset_delta + batch * batch_offset;
-            dst[offset_dst] =
-                sycl::vec<float, 1>(x[offset_src + iih * IW + iiw])
-                    .convert<sycl::half, sycl::rounding_mode::automatic>()[0];
+        const int64_t offset_src_base = (ic * offset_delta) + (batch * batch_offset);
+        const int64_t offset_src      = offset_src_base + (iih * IW) + iiw;
+
+        const bool  out_of_bounds = (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW);
+        const float src_val       = out_of_bounds ? 0.0f : x[offset_src];
+
+        if constexpr (std::is_same_v<T, sycl::half>) {
+            dst[offset_dst] = sycl::half(src_val);
+        } else if constexpr (std::is_same_v<T, float>) {
+            dst[offset_dst] = src_val;
         }
     }
 }
 
 template <typename T>
-static void im2col_sycl(
-        const float *x, T *dst, int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW,
-        int64_t KH, int64_t IC, int64_t batch, int64_t batch_offset, int64_t offset_delta,
-        int s0, int s1, int p0, int p1, int d0, int d1,
-        queue_ptr stream) {
+static void im2col_sycl_internal(const float * x, T * dst, int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW,
+                                 int64_t KH, int64_t IC, int64_t batch, int64_t batch_offset, int64_t offset_delta,
+                                 int s0, int s1, int p0, int p1, int d0, int d1, queue_ptr stream) {
     const int64_t parallel_elements = OW * KW * KH;
-    const int64_t num_blocks = (parallel_elements + SYCL_IM2COL_BLOCK_SIZE - 1) / SYCL_IM2COL_BLOCK_SIZE;
+    const int64_t num_blocks        = (parallel_elements + SYCL_IM2COL_BLOCK_SIZE - 1) / SYCL_IM2COL_BLOCK_SIZE;
 
     // decrease global range when it exceeds the max int
     int64_t local_size = downsample_sycl_global_range(batch * IC * OH * num_blocks, SYCL_IM2COL_BLOCK_SIZE);
+
     sycl::range<3> block_nums(batch * IC, OH, num_blocks);
     sycl::range<3> local_range(1, 1, local_size);
 
-    {
-        dpct::has_capability_or_fail(stream->get_device(),
-                                     {sycl::aspect::fp16});
+    const int64_t CHW = IC * KH * KW;
 
-        stream->parallel_for(
-            sycl::nd_range<3>(block_nums * local_range, local_range),
-            [=](sycl::nd_item<3> item_ct1) {
-                im2col_kernel(x, dst, batch_offset, offset_delta, IC, IW, IH, OH, OW, KW, KH,
-                               parallel_elements, (IC * KH * KW), s0, s1, p0,
-                               p1, d0, d1, item_ct1);
-            });
-    }
+    stream->parallel_for(sycl::nd_range<3>(block_nums * local_range, local_range), [=](sycl::nd_item<3> item_ct1) {
+        im2col_kernel<T>(x, dst, batch_offset, offset_delta, IC, IW, IH, OH, OW, KW, KH, parallel_elements, CHW, s0, s1,
+                         p0, p1, d0, d1, item_ct1);
+    });
 }
 
-void ggml_sycl_op_im2col(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
+static void im2col_sycl_f16(const float * x, sycl::half * dst, int64_t IW, int64_t IH, int64_t OW, int64_t OH,
+                            int64_t KW, int64_t KH, int64_t IC, int64_t batch, int64_t batch_offset,
+                            int64_t offset_delta, int s0, int s1, int p0, int p1, int d0, int d1, queue_ptr stream) {
+    if (!stream->get_device().has(sycl::aspect::fp16)) {
+        throw sycl::exception(sycl::make_error_code(sycl::errc::kernel_not_supported),
+                              "Device does not support half precision (fp16) operations!");
+    }
+    im2col_sycl_internal<sycl::half>(x, dst, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, offset_delta, s0, s1, p0,
+                                     p1, d0, d1, stream);
+}
+
+static void im2col_sycl_f32(const float * x, float * dst, int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW,
+                            int64_t KH, int64_t IC, int64_t batch, int64_t batch_offset, int64_t offset_delta, int s0,
+                            int s1, int p0, int p1, int d0, int d1, queue_ptr stream) {
+    im2col_sycl_internal<float>(x, dst, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, offset_delta, s0, s1, p0, p1,
+                                d0, d1, stream);
+}
+
+void ggml_sycl_op_im2col(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
 
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
     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 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;
+    const bool is_2D = ((const int32_t *) (dst->op_params))[6] == 1;
 
     const int64_t IC = src1->ne[is_2D ? 2 : 1];
     const int64_t IH = is_2D ? src1->ne[1] : 1;
-    const int64_t IW =         src1->ne[0];
+    const int64_t IW = src1->ne[0];
 
     const int64_t KH = is_2D ? src0->ne[1] : 1;
-    const int64_t KW =         src0->ne[0];
+    const int64_t KW = src0->ne[0];
 
     const int64_t OH = is_2D ? dst->ne[2] : 1;
-    const int64_t OW =         dst->ne[1];
+    const int64_t OW = dst->ne[1];
 
-    const size_t delta_offset = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
-    const int64_t batch = src1->ne[3];
-    const size_t batch_offset = src1->nb[3] / 4; // nb is byte offset, src is type float32
+    const size_t  delta_offset = src1->nb[is_2D ? 2 : 1] / sizeof(float);
+    const int64_t batch        = src1->ne[is_2D ? 3 : 2];
+    const size_t  batch_offset = src1->nb[is_2D ? 3 : 2] / sizeof(float);
+
+    queue_ptr stream = ctx.stream();
 
     if (dst->type == GGML_TYPE_F16) {
-        im2col_sycl((const float *) src1->data, (sycl::half *)dst->data, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, ctx.stream());
+        im2col_sycl_f16((const float *) src1->data, (sycl::half *) dst->data, IW, IH, OW, OH, KW, KH, IC, batch,
+                        batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, stream);
     } else {
-        im2col_sycl((const float *) src1->data, (float *)dst->data, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, ctx.stream());
+        im2col_sycl_f32((const float *) src1->data, (float *) dst->data, IW, IH, OW, OH, KW, KH, IC, batch,
+                        batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, stream);
     }
 }

ggml/src/ggml-sycl/rope.cpp

@@ -1,9 +1,15 @@
 #include "rope.hpp"
+#include "ggml-sycl/common.hpp"
+#include "ggml.h"
 
 struct rope_corr_dims {
     float v[2];
 };
 
+struct mrope_sections {
+    int v[4];
+};
+
 static float rope_yarn_ramp(const float low, const float high, const int i0) {
     const float y = (i0 / 2 - low) / sycl::max(0.001f, high - low);
     return 1.0f - sycl::min(1.0f, sycl::max(0.0f, y));
@@ -28,23 +34,21 @@ static void rope_yarn(
     *sin_theta = sycl::sin(theta) * mscale;
 }
 
-template<typename T, bool has_ff>
-static void rope_norm(
-    const T * x, T * dst, int ne0, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows,
-    float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, const float * freq_factors,
-    const sycl::nd_item<3> &item_ct1) {
-    const int i0 = 2 * (item_ct1.get_local_range(1) * item_ct1.get_group(1) +
-                         item_ct1.get_local_id(1));
+template <typename T, bool has_ff>
+static void rope_norm(const T * x, T * dst, const int ne0, const int ne1, const int s1, const int s2, const int n_dims,
+                      const int32_t * pos, float freq_scale, float ext_factor, float attn_factor,
+                      const rope_corr_dims corr_dims, const float theta_scale, const float * freq_factors,
+                      const sycl::nd_item<3> & item_ct1) {
+    const int i0 = 2 * (item_ct1.get_local_range(1) * item_ct1.get_group(1) + item_ct1.get_local_id(1));
 
     if (i0 >= ne0) {
         return;
     }
 
-    const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                    item_ct1.get_local_id(2);
+    const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2);
 
     if (i0 >= n_dims) {
-        const int i = row*ne0 + i0;
+        const int i = row * ne0 + i0;
 
         dst[i + 0] = x[i + 0];
         dst[i + 1] = x[i + 1];
@@ -52,42 +56,43 @@ static void rope_norm(
         return;
     }
 
-    const int i = row*ne0 + i0;
-    const int i2 = row/p_delta_rows;
+    const int row0     = row % ne1;
+    const int channel0 = row / ne1;
 
-    const float theta_base = pos[i2] * sycl::pow(theta_scale, i0 / 2.0f);
+    const int i  = row * ne0 + i0;
+    const int i2 = channel0 * s2 + row0 * s1 + i0;
 
-    const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
+    const float theta_base = pos[channel0] * sycl::pow(theta_scale, i0 / 2.0f);
+
+    const float freq_factor = has_ff ? freq_factors[i0 / 2] : 1.0f;
 
     float cos_theta;
     float sin_theta;
 
-    rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
+    rope_yarn(theta_base / freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
 
-    const float x0 = x[i + 0];
-    const float x1 = x[i + 1];
+    const float x0 = x[i2 + 0];
+    const float x1 = x[i2 + 1];
 
-    dst[i + 0] = x0*cos_theta - x1*sin_theta;
-    dst[i + 1] = x0*sin_theta + x1*cos_theta;
+    dst[i + 0] = x0 * cos_theta - x1 * sin_theta;
+    dst[i + 1] = x0 * sin_theta + x1 * cos_theta;
 }
 
-template<typename T, bool has_ff>
-static void rope_neox(
-    const T * x, T * dst, int ne0, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows,
-    float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, const float * freq_factors,
-    const sycl::nd_item<3> &item_ct1) {
-    const int i0 = 2 * (item_ct1.get_local_range(1) * item_ct1.get_group(1) +
-                         item_ct1.get_local_id(1));
+template <typename T, bool has_ff>
+static void rope_neox(const T * x, T * dst, const int ne0, const int ne1, const int s1, const int s2, const int n_dims,
+                      const int32_t * pos, const float freq_scale, const float ext_factor, const float attn_factor,
+                      const rope_corr_dims corr_dims, const float theta_scale, const float * freq_factors,
+                      const sycl::nd_item<3> & item_ct1) {
+    const int i0 = 2 * (item_ct1.get_local_range(1) * item_ct1.get_group(1) + item_ct1.get_local_id(1));
 
     if (i0 >= ne0) {
         return;
     }
 
-    const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                    item_ct1.get_local_id(2);
+    const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2);
 
     if (i0 >= n_dims) {
-        const int i = row*ne0 + i0;
+        const int i = row * ne0 + i0;
 
         dst[i + 0] = x[i + 0];
         dst[i + 1] = x[i + 1];
@@ -95,38 +100,83 @@ static void rope_neox(
         return;
     }
 
-    const int i  = row*ne0 + i0/2;
-    const int i2 = row/p_delta_rows;
+    const int row0     = row % ne1;
+    const int channel0 = row / ne1;
 
-    const float theta_base = pos[i2] * sycl::pow(theta_scale, i0 / 2.0f);
+    const int i  = row * ne0 + i0 / 2;
+    const int i2 = channel0 * s2 + row0 * s1 + i0 / 2;
 
-    const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
+    const float theta_base = pos[channel0] * sycl::pow(theta_scale, i0 / 2.0f);
+
+    const float freq_factor = has_ff ? freq_factors[i0 / 2] : 1.0f;
 
     float cos_theta;
     float sin_theta;
 
-    rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
+    rope_yarn(theta_base / freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
 
-    const float x0 = x[i + 0];
-    const float x1 = x[i + n_dims/2];
+    const float x0 = x[i2 + 0];
+    const float x1 = x[i2 + n_dims / 2];
 
-    dst[i + 0]        = x0*cos_theta - x1*sin_theta;
-    dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta;
+    dst[i + 0]          = x0 * cos_theta - x1 * sin_theta;
+    dst[i + n_dims / 2] = x0 * sin_theta + x1 * cos_theta;
+}
+
+template <typename T, bool has_ff>
+static void rope_vision(const T * x, T * dst, const int ne0, const int ne1, const int ne2, const size_t s1,
+                        const size_t s2, const int n_dims, const int32_t * pos, const float freq_scale,
+                        const float ext_factor, const float attn_factor, const rope_corr_dims corr_dims,
+                        const float theta_scale, const float * freq_factors, const mrope_sections sections,
+                        const sycl::nd_item<3> & item_ct1) {
+    // get index pos
+    const int i0 = 2 * (item_ct1.get_group(1) * item_ct1.get_local_range(1) + item_ct1.get_local_id(1));
+    if (i0 >= ne0) {
+        return;
+    }
+    const int    row_dst   = (item_ct1.get_group(2) * item_ct1.get_local_range(2)) + item_ct1.get_local_id(2);
+    const int    row_x     = row_dst % ne1;
+    const int    channel_x = row_dst / ne1;
+    const int    idst      = (row_dst * ne0) + (i0 / 2);
+    const size_t ix        = ((size_t) channel_x * s2) + ((size_t) row_x * s1) + (i0 / 2);
+
+    const int sect_dims = sections.v[0] + sections.v[1];
+    const int sector    = (i0 / 2) % sect_dims;
+
+    float theta_base = 0.0f;
+    if (sector < sections.v[0]) {
+        const int p = sector;
+        theta_base  = pos[channel_x] * sycl::pow(theta_scale, (float) p);
+    } else {
+        // Simplified from CUDA backend code: if (sector >= sections.v[0] && sector < sec_w) which is just sector >= sections.v[0]
+        const int p = sector - sections.v[0];
+        theta_base  = pos[channel_x + ne2] * sycl::pow(theta_scale, (float) p);
+    }
+
+    const float freq_factor = has_ff ? freq_factors[i0 / 2] : 1.0f;
+    float       cos_theta;
+    float       sin_theta;
+    rope_yarn(theta_base / freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
+    const float x0 = x[ix + 0];
+    const float x1 = x[ix + n_dims];
+
+    // store results in dst
+    dst[idst + 0]      = x0 * cos_theta - x1 * sin_theta;
+    dst[idst + n_dims] = x0 * sin_theta + x1 * cos_theta;
 }
 
 template <typename T>
-static void rope_norm_sycl(
-    const T *x, T *dst, int ne0, int n_dims, int nr, const int32_t *pos, float freq_scale, int p_delta_rows,
-    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, queue_ptr stream) {
+static void rope_norm_sycl(const T * x, T * dst, const int ne0, const int ne1, const int s1, const int s2,
+                           const int n_dims, int nr, const int32_t * pos, const float freq_scale, const float freq_base,
+                           const float ext_factor, const float attn_factor, const rope_corr_dims corr_dims,
+                           const float * freq_factors, queue_ptr stream) {
     GGML_ASSERT(ne0 % 2 == 0);
     const sycl::range<3> block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1);
-    const int num_blocks_x = (ne0 + 2*SYCL_ROPE_BLOCK_SIZE - 1) / (2*SYCL_ROPE_BLOCK_SIZE);
+    const int            num_blocks_x = (ne0 + 2 * SYCL_ROPE_BLOCK_SIZE - 1) / (2 * SYCL_ROPE_BLOCK_SIZE);
     const sycl::range<3> block_nums(1, num_blocks_x, nr);
 
-    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+    const float theta_scale = powf(freq_base, -2.0f / n_dims);
 
-    dpct::has_capability_or_fail(stream->get_device(),
-                                     {sycl::aspect::fp16});
+    dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
     if (freq_factors == nullptr) {
         /*
@@ -134,79 +184,102 @@ static void rope_norm_sycl(
         the limit. To get the device limit, query
         info::device::max_work_group_size. Adjust the work-group size if needed.
         */
-        stream->parallel_for(
-            sycl::nd_range<3>(block_nums * block_dims, block_dims),
-            [=](sycl::nd_item<3> item_ct1) {
-                rope_norm<T, false>(x, dst, ne0, n_dims, pos, freq_scale, p_delta_rows,
-                               ext_factor, attn_factor, corr_dims, theta_scale, freq_factors,
-                               item_ct1);
-            });
+        stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+            rope_norm<T, false>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                                theta_scale, freq_factors, item_ct1);
+        });
     } else {
         /*
         DPCT1049:41: The work-group size passed to the SYCL kernel may exceed
         the limit. To get the device limit, query
         info::device::max_work_group_size. Adjust the work-group size if needed.
         */
-        stream->parallel_for(
-            sycl::nd_range<3>(block_nums * block_dims, block_dims),
-            [=](sycl::nd_item<3> item_ct1) {
-                rope_norm<T, true>(x, dst, ne0, n_dims, pos, freq_scale, p_delta_rows,
-                              ext_factor, attn_factor, corr_dims, theta_scale, freq_factors,
-                              item_ct1);
-            });
+        stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+            rope_norm<T, true>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                               theta_scale, freq_factors, item_ct1);
+        });
     }
 }
 
 template <typename T>
-static void rope_neox_sycl(
-    const T *x, T *dst, int ne0, int n_dims, int nr, const int32_t *pos, float freq_scale, int p_delta_rows,
-    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, queue_ptr stream) {
+static void rope_neox_sycl(const T * x, T * dst, const int ne0, const int ne1, const int s1, const int s2,
+                           const int n_dims, const int nr, const int32_t * pos, const float freq_scale,
+                           const float freq_base, const float ext_factor, const float attn_factor,
+                           const rope_corr_dims corr_dims, const float * freq_factors, queue_ptr stream) {
     GGML_ASSERT(ne0 % 2 == 0);
     const sycl::range<3> block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1);
-    const int num_blocks_x = (ne0 + 2*SYCL_ROPE_BLOCK_SIZE - 1) / (2*SYCL_ROPE_BLOCK_SIZE);
+    const int            num_blocks_x = (ne0 + 2 * SYCL_ROPE_BLOCK_SIZE - 1) / (2 * SYCL_ROPE_BLOCK_SIZE);
     const sycl::range<3> block_nums(1, num_blocks_x, nr);
 
-    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+    const float theta_scale = powf(freq_base, -2.0f / n_dims);
 
-    dpct::has_capability_or_fail(stream->get_device(),
-                                    {sycl::aspect::fp16});
+    dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
     if (freq_factors == nullptr) {
-        stream->parallel_for(
-            sycl::nd_range<3>(block_nums * block_dims, block_dims),
-            [=](sycl::nd_item<3> item_ct1) {
-                rope_neox<T, false>(x, dst, ne0, n_dims, pos, freq_scale,
-                                    p_delta_rows, ext_factor, attn_factor,
-                                    corr_dims, theta_scale, freq_factors,
-                                    item_ct1);
-            });
+        stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+            rope_neox<T, false>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                                theta_scale, freq_factors, item_ct1);
+        });
     } else {
-        stream->parallel_for(
-            sycl::nd_range<3>(block_nums * block_dims, block_dims),
-            [=](sycl::nd_item<3> item_ct1) {
-                rope_neox<T, true>(x, dst, ne0, n_dims, pos, freq_scale,
-                                    p_delta_rows, ext_factor, attn_factor,
-                                    corr_dims, theta_scale, freq_factors,
-                                    item_ct1);
-            });
+        stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+            rope_neox<T, true>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                               theta_scale, freq_factors, item_ct1);
+        });
     }
 }
 
-void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
+// rope vision
+template <typename T>
+static void rope_vision_sycl(const T * x, T * dst, const int ne0, const int ne1, const int ne2, const size_t s1,
+                             const size_t s2, const int n_dims, const int nr, const int32_t * pos,
+                             const float freq_scale, const float freq_base, const float ext_factor,
+                             const float attn_factor, const rope_corr_dims corr_dims, const float * freq_factors,
+                             const mrope_sections sections, queue_ptr stream) {
+    GGML_ASSERT(ne0 % 2 == 0);
+    const sycl::range<3>    block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1);
+    const int               n_blocks_y = (ne0 + 2 * SYCL_ROPE_BLOCK_SIZE - 1) / (2 * SYCL_ROPE_BLOCK_SIZE);
+    const sycl::range<3>    grid_dims(1, n_blocks_y, nr);
+    const sycl::nd_range<3> nd_range(grid_dims * block_dims, block_dims);
+
+    const float theta_scale = std::pow(freq_base, -2.0f / n_dims);
+    // Add FP16 capability check if T could be sycl::half
+    if constexpr (std::is_same_v<T, sycl::half>) {
+        dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
+    }
+    // launch kernel
+    if (freq_factors == nullptr) {
+        stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
+            rope_vision<T, false>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
+                                  corr_dims, theta_scale, freq_factors, sections, item_ct1);
+        });
+    } else {
+        stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
+            rope_vision<T, true>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
+                                 corr_dims, theta_scale, freq_factors, sections, item_ct1);
+        });
+    }
+}
+
+inline void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
 
     GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
     GGML_ASSERT( dst->type == GGML_TYPE_F32 ||  dst->type == GGML_TYPE_F16);
     GGML_ASSERT(dst->src[0]->type == dst->type);
-
-    const int64_t ne00 = dst->src[0]->ne[0];
-    const int64_t ne01 = dst->src[0]->ne[1];
+    const int64_t ne00 = dst->src[0]->ne[0]; // head dims
+    const int64_t ne01 = dst->src[0]->ne[1]; // num heads
+    const int64_t ne02 = dst->src[0]->ne[2]; // num heads
     const int64_t nr = ggml_nrows(dst->src[0]);
 
+    const size_t s01 = dst->src[0]->nb[1] / ggml_type_size(dst->src[0]->type);
+    const size_t s02 = dst->src[0]->nb[2] / ggml_type_size(dst->src[0]->type);
+
+
     //const int n_past      = ((int32_t *) dst->op_params)[0];
     const int n_dims      = ((int32_t *) dst->op_params)[1];
     const int mode        = ((int32_t *) dst->op_params)[2];
     //const int n_ctx       = ((int32_t *) dst->op_params)[3];
     const int n_ctx_orig  = ((int32_t *) dst->op_params)[4];
+    mrope_sections sections;
 
     // RoPE alteration for extended context
     float freq_base;
@@ -222,8 +295,10 @@ void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
     memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
     memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
     memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
+    memcpy(&sections.v,  (int32_t *) dst->op_params + 11, sizeof(int)*4);
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+    const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
 
     const int32_t * pos = (const int32_t *) dst->src[1]->data;
 
@@ -240,32 +315,48 @@ void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
 
     // compute
     if (is_neox) {
+        GGML_SYCL_DEBUG("%s: neox path\n", __func__);
         if (dst->src[0]->type == GGML_TYPE_F32) {
-            rope_neox_sycl(
-                (const float *)dst->src[0]->data, (float *)dst->data, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
-                attn_factor, corr_dims, freq_factors, main_stream
-            );
+            rope_neox_sycl((const float *) dst->src[0]->data, (float *) dst->data, ne00, ne01, s01, s02, n_dims, nr,
+                           pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, main_stream);
         } else if (dst->src[0]->type == GGML_TYPE_F16) {
-            rope_neox_sycl(
-                (const sycl::half *)dst->src[0]->data, (sycl::half *)dst->data, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
-                attn_factor, corr_dims, freq_factors, main_stream
-            );
+            rope_neox_sycl((const sycl::half *) dst->src[0]->data, (sycl::half *) dst->data, ne00, ne01, s01, s02,
+                           n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims, freq_factors,
+                           main_stream);
         } else {
             GGML_ABORT("fatal error");
         }
+    } else if (is_vision) {
+        GGML_SYCL_DEBUG("%s: vision path\n", __func__);
+        if (dst->src[0]->type == GGML_TYPE_F16) {
+            rope_vision_sycl((const sycl::half *) dst->src[0]->data, (sycl::half *) dst->data, ne00, ne01, ne02, s01,
+                             s02, n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
+                             freq_factors, sections, main_stream);
+        } else if (dst->src[0]->type == GGML_TYPE_F32) {
+            rope_vision_sycl((const float *) dst->src[0]->data, (float *) dst->data, ne00, ne01, ne02, s01, s02, n_dims,
+                             nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections,
+                             main_stream);
+        } else {
+            GGML_ABORT("Fatal error: Tensor type unsupported!");
+        }
     } else {
+        GGML_SYCL_DEBUG("%s: norm path\n", __func__);
         if (dst->src[0]->type == GGML_TYPE_F32) {
-            rope_norm_sycl(
-                (const float *)dst->src[0]->data, (float *)dst->data, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
-                attn_factor, corr_dims, freq_factors, main_stream
-            );
+            rope_norm_sycl((const float *) dst->src[0]->data, (float *) dst->data, ne00, ne01, s01, s02, n_dims, nr,
+                           pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, main_stream);
         } else if (dst->src[0]->type == GGML_TYPE_F16) {
-            rope_norm_sycl(
-                (const sycl::half *)dst->src[0]->data, (sycl::half *)dst->data, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
-                attn_factor, corr_dims, freq_factors, main_stream
-            );
+            rope_norm_sycl((const sycl::half *) dst->src[0]->data, (sycl::half *) dst->data, ne00, ne01, s01, s02,
+                           n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims, freq_factors,
+                           main_stream);
         } else {
             GGML_ABORT("fatal error");
         }
     }
 }
+
+void ggml_sycl_rope(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_rope(ctx, dst);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+

ggml/src/ggml-sycl/rope.hpp

@@ -15,6 +15,6 @@
 
 #include "common.hpp"
 
-void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst);
+void ggml_sycl_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst);
 
 #endif // GGML_SYCL_ROPE_HPP

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

@@ -2397,7 +2397,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_rms_norm_f32, "rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_rms_norm_f32, "rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_rms_norm_back_f32, "rms_norm_back_f32", rms_norm_back_f32_len, rms_norm_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_l2_norm_f32, "l2_norm_f32", l2_norm_f32_len, l2_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
 
@@ -5531,7 +5531,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     uint32_t workgroups_y = (uint32_t)neq2;
     uint32_t workgroups_z = (uint32_t)neq3;
 
-    if (N == 1 && qk_ratio > 1 && is_pow2(qk_ratio) && gqa_ratio <= flash_attention_num_small_rows &&
+    if (N == 1 && qk_ratio > 1 && gqa_ratio <= flash_attention_num_small_rows &&
         qk_ratio * nek2 == neq2 && nek2 == nev2 && neq3 == 1 && nek3 == 1 && nev3 == 1) {
         // grouped query attention - make the N dimension equal to gqa_ratio, reduce
         // workgroups proportionally in y dimension. The shader will detect gqa_ratio > 1
@@ -5544,8 +5544,8 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     uint32_t split_kv = KV;
     uint32_t split_k = 1;
 
-    if (gqa_ratio > 1 && ctx->device->shader_core_count > 0) {
-        GGML_ASSERT(workgroups_x == 1);
+    // Try to use split_k when KV is large enough to be worth the overhead
+    if (workgroups_x == 1 && ctx->device->shader_core_count > 0 && KV >= 512) {
         // Try to run two workgroups per SM.
         split_k = ctx->device->shader_core_count * 2 / workgroups_y;
         if (split_k > 1) {
@@ -6006,6 +6006,7 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
     case GGML_OP_REPEAT:
     case GGML_OP_REPEAT_BACK:
     case GGML_OP_ROPE:
+    case GGML_OP_RMS_NORM:
         return true;
     default:
         return false;
@@ -6216,7 +6217,6 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
 
     switch (op) {
     case GGML_OP_NORM:
-    case GGML_OP_RMS_NORM:
     case GGML_OP_RMS_NORM_BACK:
     case GGML_OP_L2_NORM:
     case GGML_OP_SOFT_MAX:
@@ -6233,6 +6233,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
                 elements = { nr, 1, 1 };
             }
         } break;
+    case GGML_OP_RMS_NORM:
+        elements = { (uint32_t)ne01, (uint32_t)ne02, (uint32_t)ne03 };
+        break;
+
     case GGML_OP_SUM:
         // We use GGML_OP_SUM_ROWS with 1 row.
         elements = { 1, 1, 1 };
@@ -6883,7 +6887,17 @@ static void ggml_vk_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx
 
 static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     float * op_params = (float *)dst->op_params;
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_RMS_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_RMS_NORM, {
+        (uint32_t)ggml_nelements(src0),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        op_params[0], 0.0f,
+        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    }, dryrun);
 }
 
 static void ggml_vk_rms_norm_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -9261,6 +9275,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 case 112:
                 case 128:
                 case 256:
+                case 575: // DeepSeek MLA
                     break;
                 default:
                     return false;
@@ -9387,10 +9402,10 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_VIEW:
         case GGML_OP_PERMUTE:
         case GGML_OP_TRANSPOSE:
+        case GGML_OP_RMS_NORM:
             return true;
         case GGML_OP_NORM:
         case GGML_OP_GROUP_NORM:
-        case GGML_OP_RMS_NORM:
         case GGML_OP_L2_NORM:
             return ggml_is_contiguous(op->src[0]);
         case GGML_OP_ADD:

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp

@@ -131,7 +131,7 @@ ACC_TYPE perElemOpStoreCol0(const in uint32_t r, const in uint32_t c, const in A
 // Load the slope matrix, indexed by Q's dimension 2.
 ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t iq2)
 {
-    const uint32_t h = iq2 + (r & (p.gqa_ratio - 1));
+    const uint32_t h = iq2 + (r % p.gqa_ratio);
 
     const ACC_TYPE base = ACC_TYPE(h < p.n_head_log2 ? p.m0 : p.m1);
     const int      exph = int(h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1);

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

@@ -1,6 +1,6 @@
 #version 450
 
-#include "generic_head.comp"
+#include "generic_unary_head.comp"
 #include "types.comp"
 
 #extension GL_EXT_control_flow_attributes : enable
@@ -8,19 +8,29 @@
 
 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_TYPE sum[BLOCK_SIZE];
 
 void main() {
-    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
-    const uint tid = gl_LocalInvocationID.x;
+    const uint ncols     = p.ne00;
+    const uint nrows     = gl_NumWorkGroups.x;
+    const uint nchannels = gl_NumWorkGroups.y;
+
+    const uint row       = gl_WorkGroupID.x;
+    const uint channel   = gl_WorkGroupID.y;
+    const uint samp      = gl_WorkGroupID.z;
+    const uint tid       = gl_LocalInvocationID.x;
+
+    const uint stride_row       = p.nb01;
+    const uint stride_channel   = p.nb02;
+    const uint stride_sample    = p.nb03;
+
+    uint32_t a_offset = samp*stride_sample + channel*stride_channel + row*stride_row + get_aoffset();
+    uint32_t d_offset = ((samp*nchannels + channel)*nrows + row)*ncols + get_doffset();
 
     sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp
 
-    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
-        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[row*p.KX + col]);
+    [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
+        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[a_offset + col]);
         sum[tid] += xi * xi;
     }
 
@@ -33,10 +43,10 @@ void main() {
         barrier();
     }
 
-    const FLOAT_TYPE mean = sum[0] / FLOAT_TYPE(p.KX);
+    const FLOAT_TYPE mean = sum[0] / FLOAT_TYPE(ncols);
     const FLOAT_TYPE scale = inversesqrt(mean + FLOAT_TYPE(p.param1));
 
-    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
-        data_d[row*p.KX + col] = D_TYPE(scale * FLOAT_TYPE(data_a[row*p.KX + col]));
+    [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
+        data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]));
     }
 }

gguf-py/README.md

@@ -11,6 +11,11 @@ as an example for its usage.
 pip install gguf
 ```
 
+Optionally, you can install gguf with the extra 'gui' to enable the visual GGUF editor.
+```sh
+pip install gguf[gui]
+```
+
 ## API Examples/Simple Tools
 
 [examples/writer.py](https://github.com/ggml-org/llama.cpp/blob/master/gguf-py/examples/writer.py) — Generates `example.gguf` in the current directory to demonstrate generating a GGUF file. Note that this file cannot be used as a model.
@@ -25,6 +30,8 @@ pip install gguf
 
 [gguf/scripts/gguf_new_metadata.py](https://github.com/ggml-org/llama.cpp/blob/master/gguf-py/gguf/scripts/gguf_new_metadata.py) — Copies a GGUF file with added/modified/removed metadata values.
 
+[gguf/scripts/gguf_editor_gui.py](https://github.com/ggml-org/llama.cpp/blob/master/gguf-py/gguf/scripts/gguf_editor_gui.py) — Allows for viewing, editing, adding, or removing metadata values within a GGUF file as well as viewing its tensors with a Qt interface.
+
 ## Development
 Maintainers who participate in development of this package are advised to install it in editable mode:
 

gguf-py/gguf/constants.py

@@ -139,6 +139,8 @@ class Keys:
         REL_BUCKETS_COUNT            = "{arch}.attention.relative_buckets_count"
         SLIDING_WINDOW               = "{arch}.attention.sliding_window"
         SCALE                        = "{arch}.attention.scale"
+        KEY_LENGTH_MLA               = "{arch}.attention.key_length_mla"
+        VALUE_LENGTH_MLA             = "{arch}.attention.value_length_mla"
 
     class Rope:
         DIMENSION_COUNT         = "{arch}.rope.dimension_count"
@@ -216,17 +218,37 @@ class Keys:
         TYPE       = "adapter.type"
         LORA_ALPHA = "adapter.lora.alpha"
 
+    class ClipVision:
+        PROJECTOR_TYPE      = "clip.projector_type"
+        HAS_VISION_ENCODER  = "clip.has_vision_encoder"
+        HAS_LLAVA_PROJECTOR = "clip.has_llava_projector"
+        IMAGE_SIZE          = "clip.vision.image_size"
+        PATCH_SIZE          = "clip.vision.patch_size"
+        EMBEDDING_LENGTH    = "clip.vision.embedding_length"
+        FEED_FORWARD_LENGTH = "clip.vision.feed_forward_length"
+        PROJECTION_DIM      = "clip.vision.projection_dim"
+        BLOCK_COUNT         = "clip.vision.block_count"
+        IMAGE_MEAN          = "clip.vision.image_mean"
+        IMAGE_STD           = "clip.vision.image_std"
+        USE_GELU            = "clip.use_gelu"
+
+        class Attention:
+            HEAD_COUNT      = "clip.vision.attention.head_count"
+            LAYERNORM_EPS   = "clip.vision.attention.layer_norm_epsilon"
+
 #
 # recommended mapping of model tensor names for storage in gguf
 #
 
 
 class GGUFType:
-    MODEL   = "model"
-    ADAPTER = "adapter"
+    MODEL       = "model"
+    ADAPTER     = "adapter"
+    CLIP_VISION = "clip-vision"
 
 
 class MODEL_ARCH(IntEnum):
+    CLIP_VISION      = auto() # dummy arch for clip.cpp
     LLAMA            = auto()
     LLAMA4           = auto()
     DECI             = auto()
@@ -295,6 +317,16 @@ class MODEL_ARCH(IntEnum):
     BAILINGMOE       = auto()
 
 
+class VISION_PROJECTOR_TYPE(IntEnum):
+    MLP       = auto()
+    LDP       = auto()
+    LDPV2     = auto()
+    RESAMPLER = auto()
+    GLM_EDGE  = auto()
+    MERGER    = auto()
+    GEMMA3    = auto()
+
+
 class MODEL_TENSOR(IntEnum):
     TOKEN_EMBD           = auto()
     TOKEN_EMBD_NORM      = auto()
@@ -382,6 +414,8 @@ class MODEL_TENSOR(IntEnum):
     ATTN_Q_B             = auto()
     ATTN_KV_A_MQA        = auto()
     ATTN_KV_B            = auto()
+    ATTN_K_B             = auto()
+    ATTN_V_B             = auto()
     ATTN_Q_A_NORM        = auto()
     ATTN_KV_A_NORM       = auto()
     FFN_SUB_NORM         = auto()
@@ -432,9 +466,41 @@ class MODEL_TENSOR(IntEnum):
     POSNET_ATTN_K        = auto()
     POSNET_ATTN_V        = auto()
     POSNET_ATTN_OUT      = auto()
+    # vision
+    V_MMPROJ             = auto()
+    V_MMPROJ_FC          = auto()
+    V_MMPROJ_MLP         = auto()
+    V_MMPROJ_PEG         = auto()
+    V_ENC_EMBD_CLS       = auto()
+    V_ENC_EMBD_PATCH     = auto()
+    V_ENC_EMBD_POS       = auto()
+    V_ENC_ATTN_Q         = auto()
+    V_ENC_ATTN_K         = auto()
+    V_ENC_ATTN_V         = auto()
+    V_ENC_INPUT_NORM     = auto()
+    V_ENC_OUTPUT         = auto()
+    V_ENC_OUTPUT_NORM    = auto()
+    V_ENC_FFN_UP         = auto()
+    V_ENC_FFN_DOWN       = auto()
+    V_PRE_NORM           = auto()
+    V_POST_NORM          = auto()
+    V_MM_INP_PROJ        = auto() # gemma3
+    V_MM_SOFT_EMB_NORM   = auto() # gemma3
+    V_RESMPL_POS_EMBD_K  = auto() # minicpmv
+    V_RESMPL_ATTN_Q      = auto() # minicpmv
+    V_RESMPL_ATTN_K      = auto() # minicpmv
+    V_RESMPL_ATTN_V      = auto() # minicpmv
+    V_RESMPL_ATTN_OUT    = auto() # minicpmv
+    V_RESMPL_KV          = auto() # minicpmv
+    V_RESMPL_KV_NORM     = auto() # minicpmv
+    V_RESMPL_POST_NORM   = auto() # minicpmv
+    V_RESMPL_Q_NORM      = auto() # minicpmv
+    V_RESMPL_PROJ        = auto() # minicpmv
+    V_RESMPL_QUERY       = auto() # minicpmv
 
 
 MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
+    MODEL_ARCH.CLIP_VISION:      "clip", # dummy arch for clip.cpp
     MODEL_ARCH.LLAMA:            "llama",
     MODEL_ARCH.LLAMA4:           "llama4",
     MODEL_ARCH.DECI:             "deci",
@@ -503,6 +569,16 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.BAILINGMOE:       "bailingmoe",
 }
 
+VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
+    VISION_PROJECTOR_TYPE.MLP:       "mlp",
+    VISION_PROJECTOR_TYPE.LDP:       "ldp",
+    VISION_PROJECTOR_TYPE.LDPV2:     "ldpv2",
+    VISION_PROJECTOR_TYPE.RESAMPLER: "resampler",
+    VISION_PROJECTOR_TYPE.GLM_EDGE:  "adapter",
+    VISION_PROJECTOR_TYPE.MERGER:    "qwen2vl_merger",
+    VISION_PROJECTOR_TYPE.GEMMA3:    "gemma3",
+}
+
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.TOKEN_EMBD:                "token_embd",
     MODEL_TENSOR.TOKEN_EMBD_NORM:           "token_embd_norm",
@@ -590,6 +666,8 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.ATTN_Q_B:                  "blk.{bid}.attn_q_b",
     MODEL_TENSOR.ATTN_KV_A_MQA:             "blk.{bid}.attn_kv_a_mqa",
     MODEL_TENSOR.ATTN_KV_B:                 "blk.{bid}.attn_kv_b",
+    MODEL_TENSOR.ATTN_K_B:                  "blk.{bid}.attn_k_b",
+    MODEL_TENSOR.ATTN_V_B:                  "blk.{bid}.attn_v_b",
     MODEL_TENSOR.ATTN_Q_A_NORM:             "blk.{bid}.attn_q_a_norm",
     MODEL_TENSOR.ATTN_KV_A_NORM:            "blk.{bid}.attn_kv_a_norm",
     MODEL_TENSOR.ATTN_SUB_NORM:             "blk.{bid}.attn_sub_norm",
@@ -640,9 +718,72 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.POSNET_ATTN_K:             "posnet.{bid}.attn_k",
     MODEL_TENSOR.POSNET_ATTN_V:             "posnet.{bid}.attn_v",
     MODEL_TENSOR.POSNET_ATTN_OUT:           "posnet.{bid}.attn_output",
+    # vision
+    MODEL_TENSOR.V_MMPROJ:                  "mm.{bid}",
+    MODEL_TENSOR.V_MMPROJ_FC:               "mm.model.fc",
+    MODEL_TENSOR.V_MMPROJ_MLP:              "mm.model.mlp.{bid}",
+    MODEL_TENSOR.V_MMPROJ_PEG:              "mm.model.peg.{bid}",
+    MODEL_TENSOR.V_ENC_EMBD_CLS:            "v.class_embd",
+    MODEL_TENSOR.V_ENC_EMBD_PATCH:          "v.patch_embd",
+    MODEL_TENSOR.V_ENC_EMBD_POS:            "v.position_embd",
+    MODEL_TENSOR.V_ENC_ATTN_Q:              "v.blk.{bid}.attn_q",
+    MODEL_TENSOR.V_ENC_ATTN_K:              "v.blk.{bid}.attn_k",
+    MODEL_TENSOR.V_ENC_ATTN_V:              "v.blk.{bid}.attn_v",
+    MODEL_TENSOR.V_ENC_INPUT_NORM:          "v.blk.{bid}.ln1",
+    MODEL_TENSOR.V_ENC_OUTPUT:              "v.blk.{bid}.attn_out",
+    MODEL_TENSOR.V_ENC_OUTPUT_NORM:         "v.blk.{bid}.ln2",
+    MODEL_TENSOR.V_ENC_FFN_UP:              "v.blk.{bid}.ffn_up",
+    MODEL_TENSOR.V_ENC_FFN_DOWN:            "v.blk.{bid}.ffn_down",
+    MODEL_TENSOR.V_PRE_NORM:                "v.pre_ln",
+    MODEL_TENSOR.V_POST_NORM:               "v.post_ln",
+    MODEL_TENSOR.V_MM_INP_PROJ:             "mm.input_projection",
+    MODEL_TENSOR.V_MM_SOFT_EMB_NORM:        "mm.soft_emb_norm",
+    MODEL_TENSOR.V_RESMPL_POS_EMBD_K:       "resampler.pos_embd_k",
+    MODEL_TENSOR.V_RESMPL_ATTN_Q:           "resampler.attn.q",
+    MODEL_TENSOR.V_RESMPL_ATTN_K:           "resampler.attn.k",
+    MODEL_TENSOR.V_RESMPL_ATTN_V:           "resampler.attn.v",
+    MODEL_TENSOR.V_RESMPL_ATTN_OUT:         "resampler.attn.out",
+    MODEL_TENSOR.V_RESMPL_KV:               "resampler.kv",
+    MODEL_TENSOR.V_RESMPL_KV_NORM:          "resampler.ln_kv",
+    MODEL_TENSOR.V_RESMPL_POST_NORM:        "resampler.ln_post",
+    MODEL_TENSOR.V_RESMPL_Q_NORM:           "resampler.ln_q",
+    MODEL_TENSOR.V_RESMPL_PROJ:             "resampler.proj",
+    MODEL_TENSOR.V_RESMPL_QUERY:            "resampler.query",
 }
 
 MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
+    MODEL_ARCH.CLIP_VISION: [
+        MODEL_TENSOR.V_MMPROJ,
+        MODEL_TENSOR.V_MMPROJ_FC,
+        MODEL_TENSOR.V_MMPROJ_MLP,
+        MODEL_TENSOR.V_MMPROJ_PEG,
+        MODEL_TENSOR.V_ENC_EMBD_CLS,
+        MODEL_TENSOR.V_ENC_EMBD_PATCH,
+        MODEL_TENSOR.V_ENC_EMBD_POS,
+        MODEL_TENSOR.V_ENC_ATTN_Q,
+        MODEL_TENSOR.V_ENC_ATTN_K,
+        MODEL_TENSOR.V_ENC_ATTN_V,
+        MODEL_TENSOR.V_ENC_INPUT_NORM,
+        MODEL_TENSOR.V_ENC_OUTPUT,
+        MODEL_TENSOR.V_ENC_OUTPUT_NORM,
+        MODEL_TENSOR.V_ENC_FFN_UP,
+        MODEL_TENSOR.V_ENC_FFN_DOWN,
+        MODEL_TENSOR.V_PRE_NORM,
+        MODEL_TENSOR.V_POST_NORM,
+        MODEL_TENSOR.V_MM_INP_PROJ,
+        MODEL_TENSOR.V_MM_SOFT_EMB_NORM,
+        MODEL_TENSOR.V_RESMPL_POS_EMBD_K,
+        MODEL_TENSOR.V_RESMPL_ATTN_Q,
+        MODEL_TENSOR.V_RESMPL_ATTN_K,
+        MODEL_TENSOR.V_RESMPL_ATTN_V,
+        MODEL_TENSOR.V_RESMPL_ATTN_OUT,
+        MODEL_TENSOR.V_RESMPL_KV,
+        MODEL_TENSOR.V_RESMPL_KV_NORM,
+        MODEL_TENSOR.V_RESMPL_POST_NORM,
+        MODEL_TENSOR.V_RESMPL_Q_NORM,
+        MODEL_TENSOR.V_RESMPL_PROJ,
+        MODEL_TENSOR.V_RESMPL_QUERY,
+    ],
     MODEL_ARCH.LLAMA: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -1517,6 +1658,8 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.ATTN_Q_B,
         MODEL_TENSOR.ATTN_KV_A_MQA,
         MODEL_TENSOR.ATTN_KV_B,
+        MODEL_TENSOR.ATTN_K_B,
+        MODEL_TENSOR.ATTN_V_B,
         MODEL_TENSOR.ATTN_Q_A_NORM,
         MODEL_TENSOR.ATTN_KV_A_NORM,
         MODEL_TENSOR.ATTN_OUT,

gguf-py/gguf/gguf_writer.py

@@ -689,6 +689,12 @@ class GGUFWriter:
     def add_value_length(self, length: int) -> None:
         self.add_uint32(Keys.Attention.VALUE_LENGTH.format(arch=self.arch), length)
 
+    def add_key_length_mla(self, length: int) -> None:
+        self.add_uint32(Keys.Attention.KEY_LENGTH_MLA.format(arch=self.arch), length)
+
+    def add_value_length_mla(self, length: int) -> None:
+        self.add_uint32(Keys.Attention.VALUE_LENGTH_MLA.format(arch=self.arch), length)
+
     def add_max_alibi_bias(self, bias: float) -> None:
         self.add_float32(Keys.Attention.MAX_ALIBI_BIAS.format(arch=self.arch), bias)
 

gguf-py/gguf/scripts/__init__.py

@@ -4,3 +4,4 @@ from .gguf_convert_endian import main as gguf_convert_endian_entrypoint
 from .gguf_dump import main as gguf_dump_entrypoint
 from .gguf_set_metadata import main as gguf_set_metadata_entrypoint
 from .gguf_new_metadata import main as gguf_new_metadata_entrypoint
+from .gguf_editor_gui import main as gguf_editor_gui_entrypoint