llama-chat : Add Yandex instruct model template support (#12621 )

* add yandex template * update yandex chat template * fix tests * adjust chat template * fix style * fix tool macro in template * add clarify comment --------- Co-authored-by: Sergei Vorobev <serv01@yandex-team.ru> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com>
musa: fix all warnings, re-enable -DLLAMA_FATAL_WARNINGS=ON in ci and update doc (#12611 )
2025-03-30 20:12:03 +02:00 · 2025-03-30 10:59:38 +02:00 · 2025-03-30 08:33:31 +03:00 · 2025-03-30 08:33:31 +03:00 · 2025-03-30 08:33:31 +03:00 · 2025-03-30 08:33:31 +03:00
120 changed files with 8874 additions and 5599 deletions
@@ -676,6 +676,35 @@ jobs:
            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO

+  macOS-latest-cmake-visionos:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          cmake -B build -G Xcode \
+            -DGGML_METAL_USE_BF16=ON \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=visionOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=1.0 \
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
+
  macOS-latest-swift:
    runs-on: macos-latest

@@ -432,8 +432,8 @@ cmake -B build-visionos -G Xcode \
    -DCMAKE_SYSTEM_NAME=visionOS \
    -DCMAKE_OSX_SYSROOT=xros \
    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xros \
-    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 -Du_int=unsigned\ int -Du_char=unsigned\ char -Du_short=unsigned\ short ${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 -Du_int=unsigned\ int -Du_char=unsigned\ char -Du_short=unsigned\ short ${COMMON_CXX_FLAGS}" \
+    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
    -S .
 cmake --build build-visionos --config Release -- -quiet

@@ -445,8 +445,8 @@ cmake -B build-visionos-sim -G Xcode \
    -DCMAKE_SYSTEM_NAME=visionOS \
    -DCMAKE_OSX_SYSROOT=xrsimulator \
    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xrsimulator \
-    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 -Du_int=unsigned\ int -Du_char=unsigned\ char -Du_short=unsigned\ short ${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 -Du_int=unsigned\ int -Du_char=unsigned\ char -Du_short=unsigned\ short ${COMMON_CXX_FLAGS}" \
+    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
    -S .
 cmake --build build-visionos-sim --config Release -- -quiet

@@ -26,4 +26,43 @@ GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 # with SYCL support
 source /opt/intel/oneapi/setvars.sh
 GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+
+# with MUSA support
+GG_BUILD_MUSA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 ```
+
+## Running MUSA CI in a Docker Container
+
+Assuming `$PWD` is the root of the `llama.cpp` repository, follow these steps to set up and run MUSA CI in a Docker container:
+
+### 1. Create a local directory to store cached models, configuration files and venv:
+
+```bash
+mkdir -p $HOME/llama.cpp/ci-cache
+```
+
+### 2. Create a local directory to store CI run results:
+
+```bash
+mkdir -p $HOME/llama.cpp/ci-results
+```
+
+### 3. Start a Docker container and run the CI:
+
+```bash
+docker run --privileged -it \
+    -v $HOME/llama.cpp/ci-cache:/ci-cache \
+    -v $HOME/llama.cpp/ci-results:/ci-results \
+    -v $PWD:/ws -w /ws \
+    mthreads/musa:rc3.1.1-devel-ubuntu22.04
+```
+
+Inside the container, execute the following commands:
+
+```bash
+apt update -y && apt install -y bc cmake ccache git python3.10-venv time unzip wget
+git config --global --add safe.directory /ws
+GG_BUILD_MUSA=1 bash ./ci/run.sh /ci-results /ci-cache
+```
+
+This setup ensures that the CI runs within an isolated Docker environment while maintaining cached files and results across runs.
@@ -16,6 +16,9 @@
 # # with VULKAN support
 # GG_BUILD_VULKAN=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 #
+# # with MUSA support
+# GG_BUILD_MUSA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+#

 if [ -z "$2" ]; then
    echo "usage: $0 <output-dir> <mnt-dir>"
@@ -52,13 +55,22 @@ if [ ! -z ${GG_BUILD_SYCL} ]; then
        echo "source /opt/intel/oneapi/setvars.sh"
        exit 1
    fi
-
+    # Use only main GPU
+    export ONEAPI_DEVICE_SELECTOR="level_zero:0"
+    # Enable sysman for correct memory reporting
+    export ZES_ENABLE_SYSMAN=1
    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=1 -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON"
 fi

 if [ ! -z ${GG_BUILD_VULKAN} ]; then
    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_VULKAN=1"
 fi
+
+if [ ! -z ${GG_BUILD_MUSA} ]; then
+    # Use qy1 by default (MTT S80)
+    MUSA_ARCH=${MUSA_ARCH:-21}
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_MUSA=ON -DMUSA_ARCHITECTURES=${MUSA_ARCH}"
+fi
 ## helpers

 # download a file if it does not exist or if it is outdated
@@ -808,7 +820,7 @@ export LLAMA_LOG_PREFIX=1
 export LLAMA_LOG_TIMESTAMPS=1

 if [ -z ${GG_BUILD_LOW_PERF} ]; then
-    # Create symlink: ./llama.cpp/models-mnt -> $MNT/models/models-mnt
+    # Create symlink: ./llama.cpp/models-mnt -> $MNT/models
    rm -rf ${SRC}/models-mnt
    mnt_models=${MNT}/models
    mkdir -p ${mnt_models}
@@ -826,8 +838,10 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
 fi

 ret=0
-
-test $ret -eq 0 && gg_run ctest_debug
+if [ -z ${GG_BUILD_SYCL} ]; then
+    # SYCL build breaks with debug build flags
+    test $ret -eq 0 && gg_run ctest_debug
+fi
 test $ret -eq 0 && gg_run ctest_release

 if [ -z ${GG_BUILD_LOW_PERF} ]; then
@@ -835,7 +849,9 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
    test $ret -eq 0 && gg_run rerank_tiny

    if [ -z ${GG_BUILD_CLOUD} ] || [ ${GG_BUILD_EXTRA_TESTS_0} ]; then
-        test $ret -eq 0 && gg_run test_scripts_debug
+        if [ -z ${GG_BUILD_SYCL} ]; then
+            test $ret -eq 0 && gg_run test_scripts_debug
+        fi
        test $ret -eq 0 && gg_run test_scripts_release
    fi

@@ -846,7 +862,9 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
            test $ret -eq 0 && gg_run pythia_2_8b
            #test $ret -eq 0 && gg_run open_llama_7b_v2
        fi
-        test $ret -eq 0 && gg_run ctest_with_model_debug
+        if [ -z ${GG_BUILD_SYCL} ]; then
+            test $ret -eq 0 && gg_run ctest_with_model_debug
+        fi
        test $ret -eq 0 && gg_run ctest_with_model_release
    fi
 fi
@@ -114,8 +114,8 @@ if (LLAMA_LLGUIDANCE)

    ExternalProject_Add(llguidance_ext
        GIT_REPOSITORY https://github.com/guidance-ai/llguidance
-        # v0.6.12:
-        GIT_TAG ced1c9023d47ec194fa977932d35ce65c2ebfc09
+        # v0.7.10:
+        GIT_TAG 0309d2a6bf40abda35344a362edc71e06d5009f8
        PREFIX ${CMAKE_BINARY_DIR}/llguidance
        SOURCE_DIR ${LLGUIDANCE_SRC}
        BUILD_IN_SOURCE TRUE
@@ -1979,7 +1979,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
    add_opt(common_arg(
        {"--host"}, "HOST",
-        string_format("ip address to listen (default: %s)", params.hostname.c_str()),
+        string_format("ip address to listen, or bind to an UNIX socket if the address ends with .sock (default: %s)", params.hostname.c_str()),
        [](common_params & params, const std::string & value) {
            params.hostname = value;
        }
@@ -11,25 +11,24 @@ struct llama_sampler_llg {
    std::string         grammar_kind;
    std::string         grammar_data;
    LlgTokenizer *      tokenizer;
-    LlgConstraint *     grammar;
-    LlgMaskResult       llg_res;
-    bool                has_llg_res;
+    LlgMatcher *        grammar;
 };

-static LlgConstraint * llama_sampler_llg_new(LlgTokenizer * tokenizer, const char * grammar_kind,
-                                             const char * grammar_data) {
+static LlgMatcher * llama_sampler_llg_new(LlgTokenizer * tokenizer, const char * grammar_kind,
+                                          const char * grammar_data) {
    LlgConstraintInit cinit;
    llg_constraint_init_set_defaults(&cinit, tokenizer);
    const char * log_level = getenv("LLGUIDANCE_LOG_LEVEL");
    if (log_level && *log_level) {
        cinit.log_stderr_level = atoi(log_level);
    }
-    auto c = llg_new_constraint_any(&cinit, grammar_kind, grammar_data);
-    if (llg_get_error(c)) {
-        LOG_ERR("llg error: %s\n", llg_get_error(c));
-        llg_free_constraint(c);
+    auto c = llg_new_matcher(&cinit, grammar_kind, grammar_data);
+    if (llg_matcher_get_error(c)) {
+        LOG_ERR("llg error: %s\n", llg_matcher_get_error(c));
+        llg_free_matcher(c);
        return nullptr;
    }
+
    return c;
 }

@@ -40,39 +39,29 @@ static const char * llama_sampler_llg_name(const llama_sampler * /*smpl*/) {
 static void llama_sampler_llg_accept_impl(llama_sampler * smpl, llama_token token) {
    auto * ctx = (llama_sampler_llg *) smpl->ctx;
    if (ctx->grammar) {
-        LlgCommitResult res;
-        llg_commit_token(ctx->grammar, token, &res);
-        ctx->has_llg_res = false;
+        llg_matcher_consume_token(ctx->grammar, token);
    }
 }

 static void llama_sampler_llg_apply(llama_sampler * smpl, llama_token_data_array * cur_p) {
    auto * ctx = (llama_sampler_llg *) smpl->ctx;
    if (ctx->grammar) {
-        if (!ctx->has_llg_res) {
-            if (llg_compute_mask(ctx->grammar, &ctx->llg_res) == 0) {
-                ctx->has_llg_res = true;
+        const uint32_t * mask = llg_matcher_get_mask(ctx->grammar);
+        if (mask == nullptr) {
+            if (llg_matcher_compute_mask(ctx->grammar) == 0) {
+                mask = llg_matcher_get_mask(ctx->grammar);
            } else {
-                LOG_ERR("llg error: %s\n", llg_get_error(ctx->grammar));
-                llg_free_constraint(ctx->grammar);
+                LOG_ERR("llg error: %s\n", llg_matcher_get_error(ctx->grammar));
+                llg_free_matcher(ctx->grammar);
                ctx->grammar = nullptr;
+                return;
            }
        }
-        if (ctx->has_llg_res) {
-            if (ctx->llg_res.is_stop) {
-                for (size_t i = 0; i < cur_p->size; ++i) {
-                    if (!llama_vocab_is_eog(ctx->vocab, cur_p->data[i].id)) {
-                        cur_p->data[i].logit = -INFINITY;
-                    }
-                }
-            } else {
-                const uint32_t * mask = ctx->llg_res.sample_mask;
-                for (size_t i = 0; i < cur_p->size; ++i) {
-                    auto token = cur_p->data[i].id;
-                    if ((mask[token / 32] & (1 << (token % 32))) == 0) {
-                        cur_p->data[i].logit = -INFINITY;
-                    }
-                }
+
+        for (size_t i = 0; i < cur_p->size; ++i) {
+            auto token = cur_p->data[i].id;
+            if ((mask[token / 32] & (1 << (token % 32))) == 0) {
+                cur_p->data[i].logit = -INFINITY;
            }
        }
    }
@@ -80,14 +69,9 @@ static void llama_sampler_llg_apply(llama_sampler * smpl, llama_token_data_array

 static void llama_sampler_llg_reset(llama_sampler * smpl) {
    auto * ctx = (llama_sampler_llg *) smpl->ctx;
-    if (!ctx->grammar) {
-        return;
+    if (ctx->grammar) {
+        llg_matcher_reset(ctx->grammar);
    }
-
-    auto * grammar_new = llama_sampler_llg_new(ctx->tokenizer, ctx->grammar_kind.c_str(), ctx->grammar_data.c_str());
-    llg_free_constraint(ctx->grammar);
-    ctx->grammar     = grammar_new;
-    ctx->has_llg_res = false;
 }

 static llama_sampler * llama_sampler_llg_clone(const llama_sampler * smpl) {
@@ -102,7 +86,7 @@ static llama_sampler * llama_sampler_llg_clone(const llama_sampler * smpl) {
        if (ctx->grammar) {
            result_ctx->grammar_kind = ctx->grammar_kind;
            result_ctx->grammar_data = ctx->grammar_data;
-            result_ctx->grammar      = llg_clone_constraint(ctx->grammar);
+            result_ctx->grammar      = llg_clone_matcher(ctx->grammar);
            result_ctx->tokenizer    = llg_clone_tokenizer(ctx->tokenizer);
        }
    }
@@ -114,7 +98,7 @@ static void llama_sampler_llg_free(llama_sampler * smpl) {
    const auto * ctx = (llama_sampler_llg *) smpl->ctx;

    if (ctx->grammar) {
-        llg_free_constraint(ctx->grammar);
+        llg_free_matcher(ctx->grammar);
        llg_free_tokenizer(ctx->tokenizer);
    }

@@ -239,9 +223,11 @@ llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab, const char * g
            /* .grammar_data = */ grammar_data,
            /* .tokenizer    = */ tokenizer,
            /* .grammar      = */ llama_sampler_llg_new(tokenizer, grammar_kind, grammar_data),
-            /* .llg_res      = */ {},
-            /* .has_llg_res  = */ false,
        };
+        if (ctx->grammar) {
+            GGML_ASSERT(((size_t) llama_vocab_n_tokens(vocab) + 31) / 32 * 4 ==
+                        llg_matcher_get_mask_byte_size(ctx->grammar));
+        }
    } else {
        *ctx = {
            /* .vocab        = */ vocab,
@@ -249,15 +235,12 @@ llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab, const char * g
            /* .grammar_data = */ {},
            /* .tokenizer    = */ nullptr,
            /* .grammar      = */ nullptr,
-            /* .llg_res      = */ {},
-            /* .has_llg_res  = */ false,
        };
    }

    return llama_sampler_init(
        /* .iface = */ &llama_sampler_llg_i,
-        /* .ctx   = */ ctx
-    );
+        /* .ctx   = */ ctx);
 }

 #else
@@ -208,6 +208,9 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                                                        trigger_patterns_c.data(), trigger_patterns_c.size(),
                                                        trigger_tokens.data(), trigger_tokens.size())
             :      llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root");
+        if (!grmr) {
+            return nullptr;
+        }
    }

    auto * result = new common_sampler {
@@ -180,7 +180,8 @@ class Model:
            extra = sorted(tensor_names_from_parts.difference(self.tensor_names))
            missing_files = sorted(set(weight_map[n] for n in missing if n in weight_map))
            if len(extra) == 0 and len(missing_files) > 0:
-                raise ValueError(f"Missing or incomplete model files: {missing_files}")
+                raise ValueError(f"Missing or incomplete model files: {missing_files}\n"
+                                 f"Missing tensors: {missing}")
            else:
                raise ValueError("Mismatch between weight map and model parts for tensor names:\n"
                                 f"Missing tensors: {missing}\n"
@@ -528,6 +529,8 @@ class Model:
        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}
        added_vocab = tokenizer.get_added_vocab()

+        added_tokens_decoder = tokenizer.added_tokens_decoder
+
        for i in range(vocab_size):
            if i not in reverse_vocab:
                tokens.append(f"[PAD{i}]")
@@ -537,13 +540,13 @@ class Model:
                if token in added_vocab:
                    # The tokenizer in llama.cpp assumes the CONTROL and USER_DEFINED tokens are pre-normalized.
                    # To avoid unexpected issues - we make sure to normalize non-normalized tokens
-                    if not tokenizer.added_tokens_decoder[i].normalized:
+                    if not added_tokens_decoder[i].normalized:
                        previous_token = token
                        token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))
                        if previous_token != token:
                            logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")

-                    if tokenizer.added_tokens_decoder[i].special or self.does_token_look_special(token):
+                    if added_tokens_decoder[i].special or self.does_token_look_special(token):
                        toktypes.append(gguf.TokenType.CONTROL)
                    else:
                        # NOTE: this was added for Gemma.
@@ -702,6 +705,9 @@ class Model:
        if chkhsh == "ccc2ef013c104be7bae2965776d611e1d7a8a2a9c547dd93a682c9a9fc80352e":
            # ref: https://huggingface.co/Xenova/gpt-4o
            res = "gpt-4o"
+        if chkhsh == "7dec86086fcc38b66b7bc1575a160ae21cf705be7718b9d5598190d7c12db76f":
+            # ref: https://huggingface.co/UW/OLMo2-8B-SuperBPE-t180k
+            res = "superbpe"

        if res is None:
            logger.warning("\n")
@@ -1099,13 +1105,6 @@ class BloomModel(Model):

        tensors.append((self.map_tensor_name(name), data_torch))

-        if name == "word_embeddings.weight":
-            assert self.tensor_names is not None
-
-            # TODO: tie them at runtime, don't duplicate in the model file
-            if all(s not in self.tensor_names for s in ("lm_head.weight", "output.weight")):
-                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch))
-
        return tensors


@@ -1747,6 +1746,25 @@ class LlamaModel(Model):
                raise ValueError(f"Unprocessed experts: {experts}")


+@Model.register("Mistral3ForConditionalGeneration")
+class Mistral3Model(LlamaModel):
+    model_arch = gguf.MODEL_ARCH.LLAMA
+
+    # we need to merge the text_config into the root level of hparams
+    def __init__(self, *args, **kwargs):
+        hparams = kwargs["hparams"] if "hparams" in kwargs else Model.load_hparams(args[0])
+        if "text_config" in hparams:
+            hparams = {**hparams, **hparams["text_config"]}
+            kwargs["hparams"] = hparams
+        super().__init__(*args, **kwargs)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        name = name.replace("language_model.", "")
+        if "multi_modal_projector" in name or "vision_tower" in name:
+            return []
+        return super().modify_tensors(data_torch, name, bid)
+
+
@Model.register("DeciLMForCausalLM")
 class DeciModel(Model):
    model_arch = gguf.MODEL_ARCH.DECI
@@ -2251,7 +2269,7 @@ class Qwen2Model(Model):
                self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])


-@Model.register("Qwen2VLForConditionalGeneration")
+@Model.register("Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration")
 class Qwen2VLModel(Model):
    model_arch = gguf.MODEL_ARCH.QWEN2VL

@@ -2404,10 +2422,6 @@ class GPT2Model(Model):

        tensors.append((new_name, data_torch))

-        # note: GPT2 output is tied to (same as) wte in original model
-        if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD):
-            tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch))
-
        return tensors


@@ -2737,21 +2751,26 @@ class CodeShellModel(Model):
        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
        self.gguf_writer.add_rope_scaling_factor(1.0)

+    _has_tok_embd = False
+
    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        del bid  # unused

+        output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT)
+        tok_embd_name = self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD)
+
        new_name = self.map_tensor_name(name)

-        tensors: list[tuple[str, Tensor]] = [(new_name, data_torch)]
+        # assuming token_embd.weight is seen before output.weight
+        if not self._has_tok_embd and new_name == self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT):
+            # even though the tensor file(s) does not contain the word embeddings they are still in the weight map
+            if self.tensor_names and "transformer.wte.weight" in self.tensor_names:
+                logger.debug(f"{tok_embd_name} not found before {output_name}, assuming they are tied")
+                self.tensor_names.remove("transformer.wte.weight")
+        elif new_name == tok_embd_name:
+            self._has_tok_embd = True

-        if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD):
-            assert self.tensor_names is not None
-
-            if all(s not in self.tensor_names for s in ("lm_head.weight", "output.weight")):
-                # copy tok_embd.weight to output.weight
-                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch))
-
-        return tensors
+        return [(new_name, data_torch)]


@Model.register("InternLM2ForCausalLM")
@@ -3366,7 +3385,7 @@ class Gemma3Model(Model):

    # we need to merge the text_config into the root level of hparams
    def __init__(self, *args, **kwargs):
-        hparams = Model.load_hparams(kwargs["dir_model"])
+        hparams = kwargs["hparams"] if "hparams" in kwargs else Model.load_hparams(args[0])
        if "text_config" in hparams:
            hparams = {**hparams, **hparams["text_config"]}
            kwargs["hparams"] = hparams
@@ -3784,8 +3803,6 @@ class MambaModel(Model):
    _tok_embd = None

    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
        output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT)
        tok_embd_name = self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD)

@@ -3795,6 +3812,10 @@ class MambaModel(Model):
            logger.debug("A_log --> A ==> " + new_name)
            data_torch = -torch.exp(data_torch)

+        # [4 1 8192 1] -> [4 8192 1 1]
+        if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.SSM_CONV1D, bid):
+            data_torch = data_torch.squeeze()
+
        # assuming token_embd.weight is seen before output.weight
        if self._tok_embd is not None and new_name == output_name:
            if torch.equal(self._tok_embd, data_torch):
@@ -4398,6 +4419,29 @@ class DeepseekV2Model(Model):
                raise ValueError(f"Unprocessed experts: {experts}")


+@Model.register("PLMForCausalLM")
+class PLMModel(Model):
+    model_arch = gguf.MODEL_ARCH.PLM
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        self.gguf_writer.add_kv_lora_rank(hparams["kv_lora_rank"])
+        self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
+        self.gguf_writer.add_value_length(hparams["v_head_dim"])
+        self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+
@Model.register("T5WithLMHeadModel")
@Model.register("T5ForConditionalGeneration")
@Model.register("MT5ForConditionalGeneration")
@@ -5339,7 +5383,7 @@ def main() -> None:
            logger.error(f"Model {model_architecture} is not supported")
            sys.exit(1)

-        model_instance = model_class(dir_model=dir_model, ftype=output_type, fname_out=fname_out,
+        model_instance = model_class(dir_model, output_type, fname_out,
                                     is_big_endian=args.bigendian, use_temp_file=args.use_temp_file,
                                     eager=args.no_lazy,
                                     metadata_override=args.metadata, model_name=args.model_name,
@@ -110,6 +110,7 @@ models = [
    {"name": "deepseek-v3",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-V3"},
    {"name": "deepseek-r1-qwen", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"},
    {"name": "gpt-4o",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Xenova/gpt-4o", },
+    {"name": "superbpe",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/UW/OLMo2-8B-SuperBPE-t180k", },
 ]


@@ -14,9 +14,7 @@ In this guide we setup [Nvidia CUDA](https://docs.nvidia.com/cuda/) in a toolbox
 - [Creating a Fedora Toolbox Environment](#creating-a-fedora-toolbox-environment)
 - [Installing Essential Development Tools](#installing-essential-development-tools)
 - [Adding the CUDA Repository](#adding-the-cuda-repository)
- [Installing `nvidia-driver-libs`](#installing-nvidia-driver-libs)
- [Manually Resolving Package Conflicts](#manually-resolving-package-conflicts)
- [Finalizing the Installation of `nvidia-driver-libs`](#finalizing-the-installation-of-nvidia-driver-libs)
+- [Installing Nvidia Driver Libraries](#installing-nvidia-driver-libraries)
 - [Installing the CUDA Meta-Package](#installing-the-cuda-meta-package)
 - [Configuring the Environment](#configuring-the-environment)
 - [Verifying the Installation](#verifying-the-installation)
@@ -67,7 +65,7 @@ This guide focuses on Fedora hosts, but with small adjustments, it can work for
   sudo dnf distro-sync
   ```

-2. **Install the Default Text Editor (Optional):**
+2. **Install **Vim** the default text editor (Optional):**

   ```bash
   sudo dnf install vim-default-editor --allowerasing
@@ -97,36 +95,48 @@ After adding the repository, synchronize the package manager again:
 sudo dnf distro-sync
 ```

-## Installing `nvidia-driver-libs` and `nvidia-driver-cuda-libs`
+## Installing Nvidia Driver Libraries

-We need to detect if the host is supplying the [NVIDIA driver libraries into the toolbox](https://github.com/containers/toolbox/blob/main/src/pkg/nvidia/nvidia.go).
+First, we need to detect if the host is supplying the [NVIDIA driver libraries into the toolbox](https://github.com/containers/toolbox/blob/main/src/pkg/nvidia/nvidia.go):

 ```bash
 ls -la /usr/lib64/libcuda.so.1
 ```

+### If *`libcuda.so.1`* is missing:
+
+```
+ls: cannot access '/usr/lib64/libcuda.so.1': No such file or directory
+```
+
 **Explanation:**
+The host dose not supply the CUDA drivers, **install them now:**

- `nvidia-driver-libs` and `nvidia-driver-cuda-libs` contains necessary NVIDIA driver libraries required by CUDA,
-  on hosts with NVIDIA drivers installed the Fedora Container will supply the host libraries.
-
-### Install Nvidia Driver Libraries on Guest (if `libcuda.so.1` was NOT found).
+#### Install the Nvidia Driver Libraries on Guest:

 ```bash
-sudo dnf install nvidia-driver-libs nvidia-driver-cuda-libs
+sudo dnf install nvidia-driver-cuda nvidia-driver-libs nvidia-driver-cuda-libs nvidia-persistenced
 ```

-### Manually Updating the RPM database for host-supplied NVIDIA drivers (if `libcuda.so.1` was found).
+### If *`libcuda.so.1`* exists:
+```
+lrwxrwxrwx. 1 root root 21 Mar 24 11:26 /usr/lib64/libcuda.so.1 -> libcuda.so.570.133.07
+```

-If the installation fails due to conflicts, we'll manually download and install the required packages, excluding conflicting files.
+**Explanation:**
+The host is supply the CUDA drivers, **we need to update the guest RPM Database accordingly:**

-#### 1. Download `nvidia-driver-libs` and `nvidia-driver-cuda-libs` RPM's (with dependencies)
+#### Update the Toolbox RPM Database to include the Host-Supplied Libraries:
+
+Note: we do not actually install the libraries, we just update the DB so that the guest system knows they are supplied by the host.
+
+##### 1. Download `nvidia-` parts that are supplied by the host RPM's (with dependencies)

 ```bash
-sudo dnf download --destdir=/tmp/nvidia-driver-libs --resolve --arch x86_64 nvidia-driver-libs nvidia-driver-cuda-libs
+sudo dnf download --destdir=/tmp/nvidia-driver-libs --resolve --arch x86_64 nvidia-driver-cuda nvidia-driver-libs nvidia-driver-cuda-libs nvidia-persistenced
 ```

-#### 2. Update the RPM database to assume the installation of these packages.
+##### 2. Update the RPM database to assume the installation of these packages.

 ```bash
 sudo rpm --install --verbose --hash --justdb /tmp/nvidia-driver-libs/*
@@ -134,23 +144,26 @@ sudo rpm --install --verbose --hash --justdb /tmp/nvidia-driver-libs/*

 **Note:**

- The `--justdb` option only updates the RPM database, without touching the filesystem.
+- The `--justdb` option only updates the RPM database, without touching the filesystem elsewhere.

-#### Finalizing the Installation of `nvidia-driver-libs` and `nvidia-driver-cuda-libs`
+##### Check that the RPM Database has been correctly updated:
+
+**Note:** This is the same command as in the *"Install the Nvidia Driver Libraries on Guest"* for if *`libcuda.so.1`* was missing.

-After manually installing the dependencies, run:

 ```bash
-sudo dnf install nvidia-driver-libs nvidia-driver-cuda-libs
+sudo dnf install nvidia-driver-cuda nvidia-driver-libs nvidia-driver-cuda-libs nvidia-persistenced
 ```

-You should receive a message indicating the package is already installed:
+*(this time it will not install anything, as the database things that these packages are already installed)*

 ```
 Updating and loading repositories:
 Repositories loaded.
-Package "nvidia-driver-libs-3:570.86.10-1.fc41.x86_64" is already installed.
-Package "nvidia-driver-cuda-libs-3:570.86.10-1.fc41.x86_64" is already installed.
+Package "nvidia-driver-cuda-3:570.124.06-1.fc41.x86_64" is already installed.
+Package "nvidia-driver-libs-3:570.124.06-1.fc41.x86_64" is already installed.
+Package "nvidia-driver-cuda-libs-3:570.124.06-1.fc41.x86_64" is already installed.
+Package "nvidia-persistenced-3:570.124.06-1.fc41.x86_64" is already installed.

 Nothing to do.
 ```
@@ -207,9 +220,9 @@ You should see output similar to:
 ```
 nvcc: NVIDIA (R) Cuda compiler driver
 Copyright (c) 2005-2025 NVIDIA Corporation
-Built on Wed_Jan_15_19:20:09_PST_2025
-Cuda compilation tools, release 12.8, V12.8.61
-Build cuda_12.8.r12.8/compiler.35404655_0
+Built on Fri_Feb_21_20:23:50_PST_2025
+Cuda compilation tools, release 12.8, V12.8.93
+Build cuda_12.8.r12.8/compiler.35583870_0
 ```

 This output confirms that the CUDA compiler is accessible and indicates the installed version.
@@ -237,6 +237,15 @@ cmake -B buildWithCublas -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENAB
 cmake --build buildWithCublas --config Release
 ```

+**oneDNN**: The current oneDNN releases *(shipped with the oneAPI base-toolkit)* do not include the NVIDIA backend. Therefore, oneDNN must be compiled from source to enable the NVIDIA target:
+
+```sh
+git clone https://github.com/oneapi-src/oneDNN.git
+cd oneDNN
+cmake -GNinja -Bbuild-nvidia -DDNNL_CPU_RUNTIME=DPCPP -DDNNL_GPU_RUNTIME=DPCPP -DDNNL_GPU_VENDOR=NVIDIA -DONEDNN_BUILD_GRAPH=OFF -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+cmake --build build-nvidia --config Release
+```
+
 - **Adding support to AMD GPUs**

 **oneAPI Plugin**: In order to enable SYCL support on AMD GPUs, please install the [Codeplay oneAPI Plugin for AMD GPUs](https://developer.codeplay.com/products/oneapi/amd/download). As with Nvidia GPUs, the user should also make sure the plugin version matches the installed base toolkit.
@@ -327,10 +336,10 @@ export CPLUS_INCLUDE_DIR=/path/to/oneMKL/include:$CPLUS_INCLUDE_DIR
 GGML_SYCL_DEVICE_ARCH=sm_80 # Example architecture

 # Option 1: Use FP32 (recommended for better performance in most cases)
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DDNNL_DIR=/path/to/oneDNN/build-nvidia/install/lib/cmake/dnnl

 # Option 2: Use FP16
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON -DDNNL_DIR=/path/to/oneDNN/build-nvidia/install/lib/cmake/dnnl

 # build all binary
 cmake --build build --config Release -j -v
@@ -132,12 +132,14 @@ You may find the official downloads here: [NVIDIA developer site](https://develo


 #### Compile and run inside a Fedora Toolbox Container
-We also have a [guide](./cuda-fedora.md) for setting up CUDA toolkit in a Fedora [toolbox container](https://containertoolbx.org/).
+We also have a [guide](./backend/CUDA-FEDORA.md) for setting up CUDA toolkit in a Fedora [toolbox container](https://containertoolbx.org/).

 **Recommended for:**
-
- ***Particularly*** *convenient* for users of [Atomic Desktops for Fedora](https://fedoraproject.org/atomic-desktops/); such as: [Silverblue](https://fedoraproject.org/atomic-desktops/silverblue/) and [Kinoite](https://fedoraproject.org/atomic-desktops/kinoite/).
- Toolbox is installed by default: [Fedora Workstation](https://fedoraproject.org/workstation/) or [Fedora KDE Plasma Desktop](https://fedoraproject.org/spins/kde).
+- ***Necessary*** for users of [Atomic Desktops for Fedora](https://fedoraproject.org/atomic-desktops/); such as: [Silverblue](https://fedoraproject.org/atomic-desktops/silverblue/) and [Kinoite](https://fedoraproject.org/atomic-desktops/kinoite/).
+  - (there are no supported CUDA packages for these systems)
+- ***Necessary*** for users that have a host that is not a: [Supported Nvidia CUDA Release Platform](https://developer.nvidia.com/cuda-downloads).
+  - (for example, you may have [Fedora 42 Beta](https://fedoramagazine.org/announcing-fedora-linux-42-beta/) as your your host operating system)
+- ***Convenient*** For those running [Fedora Workstation](https://fedoraproject.org/workstation/) or [Fedora KDE Plasma Desktop](https://fedoraproject.org/spins/kde), and want to keep their host system clean.
 - *Optionally* toolbox packages are available: [Arch Linux](https://archlinux.org/), [Red Hat Enterprise Linux >= 8.5](https://www.redhat.com/en/technologies/linux-platforms/enterprise-linux), or [Ubuntu](https://ubuntu.com/download)


@@ -189,7 +191,7 @@ The following compilation options are also available to tweak performance:

 | Option                        | Legal values           | Default | Description                                                                                                                                                                                                                                                                             |
 |-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
-| GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, RDNA3). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower.                       |
+| GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, CDNA and RDNA3+). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower.                       |
 | GGML_CUDA_FORCE_CUBLAS        | Boolean                | false   | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models                                                                                                                                                                                       |
 | GGML_CUDA_F16                 | Boolean                | false   | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs.                                                           |
 | GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                         |
@@ -216,6 +218,7 @@ By default, all supported compute capabilities are enabled. To customize this be

 ```bash
 cmake -B build -DGGML_MUSA=ON -DMUSA_ARCHITECTURES="21"
+cmake --build build --config Release
 ```

 This configuration enables only compute capability `2.1` (MTT S80) during compilation, which can help reduce compilation time.
@@ -433,6 +436,26 @@ llama_new_context_with_model:       CANN compute buffer size =  1260.81 MiB

 For detailed info, such as model/device supports, CANN install, please refer to [llama.cpp for CANN](./backend/CANN.md).

+## Arm® KleidiAI™
+KleidiAI is a library of optimized microkernels for AI workloads, specifically designed for Arm CPUs. These microkernels enhance performance and can be enabled for use by the CPU backend.
+
+To enable KleidiAI, go to the llama.cpp directory and build using CMake
+```bash
+cmake -B build -DGGML_CPU_KLEIDIAI=ON
+cmake --build build --config Release
+```
+You can verify that KleidiAI is being used by running
+```bash
+./build/bin/llama-cli -m PATH_TO_MODEL -p "What is a car?"
+```
+If KleidiAI is enabled, the ouput will contain a line similar to:
+```
+load_tensors: CPU_KLEIDIAI model buffer size =  3474.00 MiB
+```
+KleidiAI's microkernels implement optimized tensor operations using Arm CPU features such as dotprod, int8mm and SME. llama.cpp selects the most efficient kernel based on runtime CPU feature detection. However, on platforms that support SME, you must manually enable SME microkernels by setting the environment variable `GGML_KLEIDIAI_SME=1`.
+
+Depending on your build target, other higher priority backends may be enabled by default. To ensure the CPU backend is used, you must disable the higher priority backends either at compile time, e.g. -DGGML_METAL=OFF, or during run-time using the command line option `--device none`.
+
 ## Android

 To read documentation for how to build on Android, [click here](./android.md)
@@ -9,6 +9,13 @@ brew install llama.cpp
 ```
 The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggml-org/llama.cpp/discussions/7668

+## MacPorts
+
+```sh
+sudo port install llama.cpp
+```
+see also: https://ports.macports.org/port/llama.cpp/details/
+
 ## Nix

 On Mac and Linux, the Nix package manager can be used via
@@ -2989,7 +2989,10 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
    assert(itype < GGML_TYPE_COUNT);
    ggml_type type = static_cast<ggml_type>(itype);

-    auto * ctx_clip = clip_model_load(fname_inp, 2);
+    auto * ctx_clip = clip_init(fname_inp, clip_context_params{
+        /* use_gpu */   false,
+        /* verbosity */ 2,
+    });

    const auto & ctx_src = ctx_clip->ctx_gguf;
    const auto & ctx_data = ctx_clip->ctx_data;
@@ -1,2 +1,4 @@
-add_executable(rpc-server rpc-server.cpp)
-target_link_libraries(rpc-server PRIVATE ggml llama)
+set(TARGET rpc-server)
+add_executable(${TARGET} rpc-server.cpp)
+target_link_libraries(${TARGET} PRIVATE ggml)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
@@ -72,3 +72,14 @@ $ bin/llama-cli -m ../models/tinyllama-1b/ggml-model-f16.gguf -p "Hello, my name

 This way you can offload model layers to both local and remote devices.

+### Local cache
+
+The RPC server can use a local cache to store large tensors and avoid transferring them over the network.
+This can speed up model loading significantly, especially when using large models.
+To enable the cache, use the `-c` option:
+
+```bash
+$ bin/rpc-server -c
+```
+
+By default, the cache is stored in the `$HOME/.cache/llama.cpp/rpc` directory and can be controlled via the `LLAMA_CACHE` environment variable.
@@ -1,3 +1,7 @@
+#if defined(_MSC_VER)
+#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
+#endif
+
 #include "ggml-cpu.h"

 #ifdef GGML_USE_CUDA
@@ -18,26 +22,142 @@

 #include "ggml-rpc.h"
 #ifdef _WIN32
+#  define DIRECTORY_SEPARATOR '\\'
+#  include <locale>
 #  include <windows.h>
+#  include <fcntl.h>
+#  include <io.h>
 #else
+#  define DIRECTORY_SEPARATOR '/'
 #  include <unistd.h>
+#  include <sys/stat.h>
 #endif
+#include <codecvt>
 #include <string>
 #include <stdio.h>
+#include <vector>
+#include <filesystem>
+
+namespace fs = std::filesystem;
+
+// NOTE: this is copied from common.cpp to avoid linking with libcommon
+// returns true if successful, false otherwise
+static bool fs_create_directory_with_parents(const std::string & path) {
+#ifdef _WIN32
+    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
+    std::wstring wpath = converter.from_bytes(path);
+
+    // if the path already exists, check whether it's a directory
+    const DWORD attributes = GetFileAttributesW(wpath.c_str());
+    if ((attributes != INVALID_FILE_ATTRIBUTES) && (attributes & FILE_ATTRIBUTE_DIRECTORY)) {
+        return true;
+    }
+
+    size_t pos_slash = 0;
+
+    // process path from front to back, procedurally creating directories
+    while ((pos_slash = path.find('\\', pos_slash)) != std::string::npos) {
+        const std::wstring subpath = wpath.substr(0, pos_slash);
+        const wchar_t * test = subpath.c_str();
+
+        const bool success = CreateDirectoryW(test, NULL);
+        if (!success) {
+            const DWORD error = GetLastError();
+
+            // if the path already exists, ensure that it's a directory
+            if (error == ERROR_ALREADY_EXISTS) {
+                const DWORD attributes = GetFileAttributesW(subpath.c_str());
+                if (attributes == INVALID_FILE_ATTRIBUTES || !(attributes & FILE_ATTRIBUTE_DIRECTORY)) {
+                    return false;
+                }
+            } else {
+                return false;
+            }
+        }
+
+        pos_slash += 1;
+    }
+
+    return true;
+#else
+    // if the path already exists, check whether it's a directory
+    struct stat info;
+    if (stat(path.c_str(), &info) == 0) {
+        return S_ISDIR(info.st_mode);
+    }
+
+    size_t pos_slash = 1; // skip leading slashes for directory creation
+
+    // process path from front to back, procedurally creating directories
+    while ((pos_slash = path.find('/', pos_slash)) != std::string::npos) {
+        const std::string subpath = path.substr(0, pos_slash);
+        struct stat info;
+
+        // if the path already exists, ensure that it's a directory
+        if (stat(subpath.c_str(), &info) == 0) {
+            if (!S_ISDIR(info.st_mode)) {
+                return false;
+            }
+        } else {
+            // create parent directories
+            const int ret = mkdir(subpath.c_str(), 0755);
+            if (ret != 0) {
+                return false;
+            }
+        }
+
+        pos_slash += 1;
+    }
+
+    return true;
+#endif // _WIN32
+}
+
+// NOTE: this is copied from common.cpp to avoid linking with libcommon
+static std::string fs_get_cache_directory() {
+    std::string cache_directory = "";
+    auto ensure_trailing_slash = [](std::string p) {
+        // Make sure to add trailing slash
+        if (p.back() != DIRECTORY_SEPARATOR) {
+            p += DIRECTORY_SEPARATOR;
+        }
+        return p;
+    };
+    if (getenv("LLAMA_CACHE")) {
+        cache_directory = std::getenv("LLAMA_CACHE");
+    } else {
+#ifdef __linux__
+        if (std::getenv("XDG_CACHE_HOME")) {
+            cache_directory = std::getenv("XDG_CACHE_HOME");
+        } else {
+            cache_directory = std::getenv("HOME") + std::string("/.cache/");
+        }
+#elif defined(__APPLE__)
+        cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
+#elif defined(_WIN32)
+        cache_directory = std::getenv("LOCALAPPDATA");
+#endif // __linux__
+        cache_directory = ensure_trailing_slash(cache_directory);
+        cache_directory += "llama.cpp";
+    }
+    return ensure_trailing_slash(cache_directory);
+}

 struct rpc_server_params {
    std::string host        = "127.0.0.1";
    int         port        = 50052;
    size_t      backend_mem = 0;
+    bool        use_cache   = false;
 };

 static void print_usage(int /*argc*/, char ** argv, rpc_server_params params) {
    fprintf(stderr, "Usage: %s [options]\n\n", argv[0]);
    fprintf(stderr, "options:\n");
-    fprintf(stderr, "  -h, --help            show this help message and exit\n");
-    fprintf(stderr, "  -H HOST, --host HOST  host to bind to (default: %s)\n", params.host.c_str());
-    fprintf(stderr, "  -p PORT, --port PORT  port to bind to (default: %d)\n", params.port);
-    fprintf(stderr, "  -m MEM, --mem MEM     backend memory size (in MB)\n");
+    fprintf(stderr, "  -h, --help                show this help message and exit\n");
+    fprintf(stderr, "  -H HOST, --host HOST      host to bind to (default: %s)\n", params.host.c_str());
+    fprintf(stderr, "  -p PORT, --port PORT      port to bind to (default: %d)\n", params.port);
+    fprintf(stderr, "  -m MEM,  --mem MEM        backend memory size (in MB)\n");
+    fprintf(stderr, "  -c,      --cache          enable local file cache\n");
    fprintf(stderr, "\n");
 }

@@ -58,6 +178,8 @@ static bool rpc_server_params_parse(int argc, char ** argv, rpc_server_params &
            if (params.port <= 0 || params.port > 65535) {
                return false;
            }
+        } else if (arg == "-c" || arg == "--cache") {
+            params.use_cache = true;
        } else if (arg == "-m" || arg == "--mem") {
            if (++i >= argc) {
                return false;
@@ -164,8 +286,20 @@ int main(int argc, char * argv[]) {
    } else {
        get_backend_memory(&free_mem, &total_mem);
    }
-    printf("Starting RPC server on %s, backend memory: %zu MB\n", endpoint.c_str(), free_mem / (1024 * 1024));
-    ggml_backend_rpc_start_server(backend, endpoint.c_str(), free_mem, total_mem);
+    const char * cache_dir = nullptr;
+    std::string cache_dir_str = fs_get_cache_directory() + "rpc/";
+    if (params.use_cache) {
+        if (!fs_create_directory_with_parents(cache_dir_str)) {
+            fprintf(stderr, "Failed to create cache directory: %s\n", cache_dir_str.c_str());
+            return 1;
+        }
+        cache_dir = cache_dir_str.c_str();
+    }
+    printf("Starting RPC server\n");
+    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));
+    ggml_backend_rpc_start_server(backend, endpoint.c_str(), cache_dir, free_mem, total_mem);
    ggml_backend_free(backend);
    return 0;
 }
@@ -38,24 +38,6 @@
 }
 #endif

-GGML_ATTRIBUTE_FORMAT(1, 2)
-static std::string fmt(const char * fmt, ...) {
-    va_list ap;
-    va_list ap2;
-    va_start(ap, fmt);
-    va_copy(ap2, ap);
-    const int size = vsnprintf(NULL, 0, fmt, ap);
-    GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
-    std::string buf;
-    buf.resize(size);
-    const int size2 = vsnprintf(const_cast<char *>(buf.data()), buf.size() + 1, fmt, ap2);
-    GGML_ASSERT(size2 == size);
-    va_end(ap2);
-    va_end(ap);
-
-    return buf;
-}
-
 GGML_ATTRIBUTE_FORMAT(1, 2)
 static int printe(const char * fmt, ...) {
    va_list args;
@@ -525,11 +507,11 @@ class HttpClient {
        int secs = static_cast<int>(seconds) % 60;

        if (hrs > 0) {
-            return fmt("%dh %02dm %02ds", hrs, mins, secs);
+            return string_format("%dh %02dm %02ds", hrs, mins, secs);
        } else if (mins > 0) {
-            return fmt("%dm %02ds", mins, secs);
+            return string_format("%dm %02ds", mins, secs);
        } else {
-            return fmt("%ds", secs);
+            return string_format("%ds", secs);
        }
    }

@@ -544,7 +526,7 @@ class HttpClient {
            }
        }

-        return fmt("%.2f %s", dbl_size, suffix[i]);
+        return string_format("%.2f %s", dbl_size, suffix[i]);
    }

    static int update_progress(void * ptr, curl_off_t total_to_download, curl_off_t now_downloaded, curl_off_t,
@@ -578,7 +560,9 @@ class HttpClient {
        return (now_downloaded_plus_file_size * 100) / total_to_download;
    }

-    static std::string generate_progress_prefix(curl_off_t percentage) { return fmt("%3ld%% |", static_cast<long int>(percentage)); }
+    static std::string generate_progress_prefix(curl_off_t percentage) {
+        return string_format("%3ld%% |", static_cast<long int>(percentage));
+    }

    static double calculate_speed(curl_off_t now_downloaded, const std::chrono::steady_clock::time_point & start_time) {
        const auto                          now             = std::chrono::steady_clock::now();
@@ -589,9 +573,9 @@ class HttpClient {
    static std::string generate_progress_suffix(curl_off_t now_downloaded_plus_file_size, curl_off_t total_to_download,
                                                double speed, double estimated_time) {
        const int width = 10;
-        return fmt("%*s/%*s%*s/s%*s", width, human_readable_size(now_downloaded_plus_file_size).c_str(), width,
-                   human_readable_size(total_to_download).c_str(), width, human_readable_size(speed).c_str(), width,
-                   human_readable_time(estimated_time).c_str());
+        return string_format("%*s/%*s%*s/s%*s", width, human_readable_size(now_downloaded_plus_file_size).c_str(),
+                             width, human_readable_size(total_to_download).c_str(), width,
+                             human_readable_size(speed).c_str(), width, human_readable_time(estimated_time).c_str());
    }

    static int calculate_progress_bar_width(const std::string & progress_prefix, const std::string & progress_suffix) {
@@ -489,8 +489,12 @@ struct result_timings {
    double predicted_per_token_ms;
    double predicted_per_second;

+    // Optional speculative metrics - only included when > 0
+    int32_t draft_n = 0;
+    int32_t draft_n_accepted = 0;
+
    json to_json() const {
-        return {
+        json base = {
            {"prompt_n",               prompt_n},
            {"prompt_ms",              prompt_ms},
            {"prompt_per_token_ms",    prompt_per_token_ms},
@@ -501,6 +505,13 @@ struct result_timings {
            {"predicted_per_token_ms", predicted_per_token_ms},
            {"predicted_per_second",   predicted_per_second},
        };
+
+        if (draft_n > 0) {
+            base["draft_n"] = draft_n;
+            base["draft_n_accepted"] = draft_n_accepted;
+        }
+
+        return base;
    }
 };

@@ -830,6 +841,11 @@ struct server_task_result_cmpl_final : server_task_result {
            ret.push_back({"timings", timings.to_json()});
        }

+        // extra fields for debugging purposes
+        if (verbose) {
+            ret["__verbose"] = to_json_non_oaicompat();
+        }
+
        return ret;
    }
 };
@@ -1294,6 +1310,10 @@ struct server_slot {

    std::function<void(int)> callback_on_release;

+    // Speculative decoding stats
+    int32_t n_draft_total = 0;      // Total draft tokens generated
+    int32_t n_draft_accepted = 0;   // Draft tokens actually accepted
+
    void reset() {
        SLT_DBG(*this, "%s", "\n");

@@ -1310,6 +1330,10 @@ struct server_slot {

        generated_tokens.clear();
        generated_token_probs.clear();
+
+        // clear speculative decoding stats
+        n_draft_total = 0;
+        n_draft_accepted = 0;
    }

    bool is_non_causal() const {
@@ -1376,6 +1400,12 @@ struct server_slot {
        timings.predicted_per_token_ms = t_token_generation / n_decoded;
        timings.predicted_per_second = 1e3 / t_token_generation * n_decoded;

+        // Add speculative metrics
+        if (n_draft_total > 0) {
+            timings.draft_n = n_draft_total;
+            timings.draft_n_accepted = n_draft_accepted;
+        }
+
        return timings;
    }

@@ -1423,6 +1453,15 @@ struct server_slot {
                t_prompt_processing, n_prompt_tokens_processed, t_prompt, n_prompt_second,
                t_token_generation, n_decoded, t_gen, n_gen_second,
                t_prompt_processing + t_token_generation, n_prompt_tokens_processed + n_decoded);
+
+        if (n_draft_total > 0) {
+            const float draft_ratio = (float) n_draft_accepted / n_draft_total;
+            SLT_INF(*this,
+                    "\n"
+                    "draft acceptance rate = %0.5f (%5d accepted / %5d generated)\n",
+                    draft_ratio, n_draft_accepted, n_draft_total
+            );
+        }
    }

    json to_json() const {
@@ -3285,6 +3324,9 @@ struct server_context {

                llama_tokens draft = common_speculative_gen_draft(slot.spec, params_spec, slot.cache_tokens, id);

+                // keep track of total number of tokens generated in the draft
+                slot.n_draft_total += draft.size();
+
                // ignore small drafts
                if (slot.params.speculative.n_min > (int) draft.size()) {
                    SLT_DBG(slot, "ignoring small draft: %d < %d\n", (int) draft.size(), slot.params.speculative.n_min);
@@ -3310,6 +3352,9 @@ struct server_context {
                slot.n_past    += ids.size();
                slot.n_decoded += ids.size();

+                // update how many tokens out of draft was accepted
+                slot.n_draft_accepted += ids.size() - 1;
+
                slot.cache_tokens.push_back(id);
                slot.cache_tokens.insert(slot.cache_tokens.end(), ids.begin(), ids.end() - 1);

@@ -4454,15 +4499,24 @@ int main(int argc, char ** argv) {
        llama_backend_free();
    };

-    // bind HTTP listen port
    bool was_bound = false;
-    if (params.port == 0) {
-        int bound_port = svr->bind_to_any_port(params.hostname);
-        if ((was_bound = (bound_port >= 0))) {
-            params.port = bound_port;
-        }
+    if (string_ends_with(std::string(params.hostname), ".sock")) {
+        LOG_INF("%s: setting address family to AF_UNIX\n", __func__);
+        svr->set_address_family(AF_UNIX);
+        // bind_to_port requires a second arg, any value other than 0 should
+        // simply get ignored
+        was_bound = svr->bind_to_port(params.hostname, 8080);
    } else {
-        was_bound = svr->bind_to_port(params.hostname, params.port);
+        LOG_INF("%s: binding port with default address family\n", __func__);
+        // bind HTTP listen port
+        if (params.port == 0) {
+            int bound_port = svr->bind_to_any_port(params.hostname);
+            if ((was_bound = (bound_port >= 0))) {
+                params.port = bound_port;
+            }
+        } else {
+            was_bound = svr->bind_to_port(params.hostname, params.port);
+        }
    }

    if (!was_bound) {
@@ -99,13 +99,9 @@ export default function ChatScreen() {
    canvasData,
    replaceMessageAndGenerate,
  } = useAppContext();
-  const [inputMsg, setInputMsg] = useState(prefilledMsg.content());
-  const inputRef = useRef<HTMLTextAreaElement>(null);
+  const textarea = useOptimizedTextarea(prefilledMsg.content());

-  const { extraContext, clearExtraContext } = useVSCodeContext(
-    inputRef,
-    setInputMsg
-  );
+  const { extraContext, clearExtraContext } = useVSCodeContext(textarea);
  // TODO: improve this when we have "upload file" feature
  const currExtra: Message['extra'] = extraContext ? [extraContext] : undefined;

@@ -135,9 +131,10 @@ export default function ChatScreen() {
  };

  const sendNewMessage = async () => {
-    if (inputMsg.trim().length === 0 || isGenerating(currConvId ?? '')) return;
-    const lastInpMsg = inputMsg;
-    setInputMsg('');
+    const lastInpMsg = textarea.value();
+    if (lastInpMsg.trim().length === 0 || isGenerating(currConvId ?? ''))
+      return;
+    textarea.setValue('');
    scrollToBottom(false);
    setCurrNodeId(-1);
    // get the last message node
@@ -146,13 +143,13 @@ export default function ChatScreen() {
      !(await sendMessage(
        currConvId,
        lastMsgNodeId,
-        inputMsg,
+        lastInpMsg,
        currExtra,
        onChunk
      ))
    ) {
      // restore the input message if failed
-      setInputMsg(lastInpMsg);
+      textarea.setValue(lastInpMsg);
    }
    // OK
    clearExtraContext();
@@ -195,16 +192,13 @@ export default function ChatScreen() {
      // send the prefilled message if needed
      sendNewMessage();
    } else {
-      // otherwise, focus on the input and move the cursor to the end
-      if (inputRef.current) {
-        inputRef.current.focus();
-        inputRef.current.selectionStart = inputRef.current.value.length;
-      }
+      // otherwise, focus on the input
+      textarea.focus();
    }
    prefilledMsg.clear();
    // no need to keep track of sendNewMessage
    // eslint-disable-next-line react-hooks/exhaustive-deps
-  }, [inputRef]);
+  }, [textarea.ref]);

  // due to some timing issues of StorageUtils.appendMsg(), we need to make sure the pendingMsg is not duplicated upon rendering (i.e. appears once in the saved conversation and once in the pendingMsg)
  const pendingMsgDisplay: MessageDisplay[] =
@@ -258,9 +252,7 @@ export default function ChatScreen() {
          <textarea
            className="textarea textarea-bordered w-full"
            placeholder="Type a message (Shift+Enter to add a new line)"
-            ref={inputRef}
-            value={inputMsg}
-            onChange={(e) => setInputMsg(e.target.value)}
+            ref={textarea.ref}
            onKeyDown={(e) => {
              if (e.nativeEvent.isComposing || e.keyCode === 229) return;
              if (e.key === 'Enter' && e.shiftKey) return;
@@ -280,11 +272,7 @@ export default function ChatScreen() {
              Stop
            </button>
          ) : (
-            <button
-              className="btn btn-primary ml-2"
-              onClick={sendNewMessage}
-              disabled={inputMsg.trim().length === 0}
-            >
+            <button className="btn btn-primary ml-2" onClick={sendNewMessage}>
              Send
            </button>
          )}
@@ -298,3 +286,43 @@ export default function ChatScreen() {
    </div>
  );
 }
+
+export interface OptimizedTextareaValue {
+  value: () => string;
+  setValue: (value: string) => void;
+  focus: () => void;
+  ref: React.RefObject<HTMLTextAreaElement>;
+}
+
+// This is a workaround to prevent the textarea from re-rendering when the inner content changes
+// See https://github.com/ggml-org/llama.cpp/pull/12299
+function useOptimizedTextarea(initValue: string): OptimizedTextareaValue {
+  const [savedInitValue, setSavedInitValue] = useState<string>(initValue);
+  const textareaRef = useRef<HTMLTextAreaElement>(null);
+
+  useEffect(() => {
+    if (textareaRef.current && savedInitValue) {
+      textareaRef.current.value = savedInitValue;
+      setSavedInitValue('');
+    }
+  }, [textareaRef, savedInitValue, setSavedInitValue]);
+
+  return {
+    value: () => {
+      return textareaRef.current?.value ?? savedInitValue;
+    },
+    setValue: (value: string) => {
+      if (textareaRef.current) {
+        textareaRef.current.value = value;
+      }
+    },
+    focus: () => {
+      if (textareaRef.current) {
+        // focus and move the cursor to the end
+        textareaRef.current.focus();
+        textareaRef.current.selectionStart = textareaRef.current.value.length;
+      }
+    },
+    ref: textareaRef,
+  };
+}
@@ -1,5 +1,6 @@
 import { useEffect, useState } from 'react';
 import { MessageExtraContext } from './types';
+import { OptimizedTextareaValue } from '../components/ChatScreen';

 // Extra context when using llama.cpp WebUI from llama-vscode, inside an iframe
 // Ref: https://github.com/ggml-org/llama.cpp/pull/11940
@@ -14,10 +15,7 @@ interface SetTextEvData {
 * window.postMessage({ command: 'setText', text: 'Spot the syntax error', context: 'def test()\n  return 123' }, '*');
 */

-export const useVSCodeContext = (
-  inputRef: React.RefObject<HTMLTextAreaElement>,
-  setInputMsg: (text: string) => void
-) => {
+export const useVSCodeContext = (textarea: OptimizedTextareaValue) => {
  const [extraContext, setExtraContext] = useState<MessageExtraContext | null>(
    null
  );
@@ -27,20 +25,20 @@ export const useVSCodeContext = (
    const handleMessage = (event: MessageEvent) => {
      if (event.data?.command === 'setText') {
        const data: SetTextEvData = event.data;
-        setInputMsg(data?.text);
+        textarea.setValue(data?.text);
        if (data?.context && data.context.length > 0) {
          setExtraContext({
            type: 'context',
            content: data.context,
          });
        }
-        inputRef.current?.focus();
+        textarea.focus();
      }
    };

    window.addEventListener('message', handleMessage);
    return () => window.removeEventListener('message', handleMessage);
-  }, [inputRef, setInputMsg]);
+  }, [textarea]);

  // Add a keydown listener that sends the "escapePressed" message to the parent window
  useEffect(() => {
@@ -331,11 +331,11 @@ int main(int argc, char ** argv) {
                        }

                        active_seqs.erase(s);
-                        for(int i = 0; i < n_seq_dft; i++) {
+                        for (int i = 0; i < n_seq_dft; i++) {
                            if (i == s) {
                                continue;
                            }
-                            if (drafts[i].tokens[i_dft] == drafts[s].tokens[i_dft]) {
+                            if (drafts[i].active && drafts[i].tokens[i_dft] == drafts[s].tokens[i_dft]) {
                                // synchronize active status for sequences with the same drafted token
                                drafts[i].active = drafts[i].active && accept;
                                if (!drafts[i].active) {
@@ -571,6 +571,10 @@ int main(int argc, char ** argv) {
    model_ttc = llama_init_ttc.model.get();
    ctx_ttc   = llama_init_ttc.context.get();

+    if (model_ttc == nullptr || ctx_ttc == nullptr) {
+        return ENOENT;
+    }
+
    const llama_vocab * vocab = llama_model_get_vocab(model_ttc);

    // TODO: refactor in a common struct
@@ -586,6 +590,10 @@ int main(int argc, char ** argv) {
    model_cts = llama_init_cts.model.get();
    ctx_cts   = llama_init_cts.context.get();

+    if (model_cts == nullptr || ctx_cts == nullptr) {
+        return ENOENT;
+    }
+
    std::vector<common_sampler *> smpl(n_parallel);
    for (int i = 0; i < n_parallel; ++i) {
        params.sampling.no_perf = (i != 0);
@@ -100,6 +100,10 @@ else()
    set(INS_ENB ON)
 endif()

+message(DEBUG "GGML_NATIVE         : ${GGML_NATIVE}")
+message(DEBUG "GGML_NATIVE_DEFAULT : ${GGML_NATIVE_DEFAULT}")
+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)
@@ -123,10 +127,12 @@ endif()
 option(GGML_LASX             "ggml: enable lasx"             ON)
 option(GGML_LSX              "ggml: enable lsx"              ON)
 option(GGML_RVV              "ggml: enable rvv"              ON)
+option(GGML_RV_ZFH           "ggml: enable riscv zfh"        OFF)
 option(GGML_VXE              "ggml: enable vxe"              ON)

 option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
-set(GGML_CPU_ARM_ARCH "" CACHE STRING "ggml: CPU architecture for ARM")
+set(GGML_CPU_ARM_ARCH        "" CACHE STRING "ggml: CPU architecture for ARM")
+set(GGML_CPU_POWERPC_CPUTYPE "" CACHE STRING "ggml: CPU type for PowerPC")


 if (WIN32)
@@ -0,0 +1,22 @@
+find_package(Git)
+
+# the commit's SHA1
+execute_process(COMMAND
+    "${GIT_EXECUTABLE}" describe --match=NeVeRmAtCh --always --abbrev=8
+    WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
+    OUTPUT_VARIABLE GIT_SHA1
+    ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+
+# the date of the commit
+execute_process(COMMAND
+    "${GIT_EXECUTABLE}" log -1 --format=%ad --date=local
+    WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
+    OUTPUT_VARIABLE GIT_DATE
+    ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+
+# the subject of the commit
+execute_process(COMMAND
+    "${GIT_EXECUTABLE}" log -1 --format=%s
+    WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
+    OUTPUT_VARIABLE GIT_COMMIT_SUBJECT
+    ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
@@ -5,7 +5,7 @@

 set_and_check(GGML_INCLUDE_DIR "@PACKAGE_GGML_INCLUDE_INSTALL_DIR@")
 set_and_check(GGML_LIB_DIR "@PACKAGE_GGML_LIB_INSTALL_DIR@")
-set_and_check(GGML_BIN_DIR "@PACKAGE_GGML_BIN_INSTALL_DIR@")
+#set_and_check(GGML_BIN_DIR "@PACKAGE_GGML_BIN_INSTALL_DIR@")

 find_package(Threads REQUIRED)

@@ -17,7 +17,9 @@ GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const c

 GGML_BACKEND_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);

-GGML_BACKEND_API void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
+GGML_BACKEND_API void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint,
+                                                    const char * cache_dir,
+                                                    size_t free_mem, size_t total_mem);

 GGML_BACKEND_API ggml_backend_reg_t ggml_backend_rpc_reg(void);

@@ -1791,11 +1791,11 @@ extern "C" {

 #define GGML_KQ_MASK_PAD 64

-    // q:    [n_embd, n_batch,     n_head,    1]
-    // k:    [n_embd, n_kv,        n_head_kv, 1]
-    // v:    [n_embd, n_kv,        n_head_kv, 1] !! not transposed !!
-    // mask: [n_kv,   n_batch_pad, 1,         1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
-    // res:  [n_embd, n_head,      n_batch,   1] !! permuted !!
+    // q:    [n_embd_k, n_batch,     n_head,    1]
+    // k:    [n_embd_k, n_kv,        n_head_kv, 1]
+    // v:    [n_embd_v, n_kv,        n_head_kv, 1] !! not transposed !!
+    // mask: [n_kv,     n_batch_pad, 1,         1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
+    // res:  [n_embd_v, n_head,      n_batch,   1] !! permuted !!
    GGML_API struct ggml_tensor * ggml_flash_attn_ext(
            struct ggml_context * ctx,
            struct ggml_tensor  * q,
@@ -65,7 +65,7 @@ if (GGML_LTO)
    endif()
 endif()

-if (GGML_CCACHE)
+if (GGML_CCACHE AND NOT CMAKE_C_COMPILER_LAUNCHER AND NOT CMAKE_CXX_COMPILER_LAUNCHER)
    find_program(GGML_CCACHE_FOUND ccache)
    find_program(GGML_SCCACHE_FOUND sccache)

@@ -76,7 +76,11 @@ if (GGML_CCACHE)
            set(GGML_CCACHE_VARIANT sccache)
        endif()
        # TODO: should not be set globally
-        set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE "${GGML_CCACHE_VARIANT}")
+        if (GGML_SYCL AND GGML_CCACHE_FOUND AND WIN32)
+            set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE "ccache compiler_type=icl")
+        else ()
+            set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE "${GGML_CCACHE_VARIANT}")
+        endif ()
        set(ENV{CCACHE_SLOPPINESS} time_macros)
        message(STATUS "${GGML_CCACHE_VARIANT} found, compilation results will be cached. Disable with GGML_CCACHE=OFF.")
    else()
@@ -325,6 +329,10 @@ if (CMAKE_SYSTEM_NAME MATCHES "Android")
    target_link_libraries(ggml-base PRIVATE dl)
 endif()

+if(CMAKE_SYSTEM_NAME MATCHES "visionOS")
+    target_compile_definitions(ggml-base PUBLIC _DARWIN_C_SOURCE)
+endif()
+
 if (BUILD_SHARED_LIBS)
    foreach (target ggml-base ggml)
        set_target_properties(${target} PROPERTIES POSITION_INDEPENDENT_CODE ON)
@@ -1,168 +0,0 @@
---
-Language:        Cpp
-# BasedOnStyle:  Google
-AccessModifierOffset: -1
-AlignAfterOpenBracket: Align
-AlignConsecutiveMacros: false
-AlignConsecutiveAssignments: false
-AlignConsecutiveDeclarations: false
-AlignEscapedNewlines: Left
-AlignOperands:   true
-AlignTrailingComments: true
-AllowAllArgumentsOnNextLine: true
-AllowAllConstructorInitializersOnNextLine: true
-AllowAllParametersOfDeclarationOnNextLine: true
-AllowShortBlocksOnASingleLine: Never
-AllowShortCaseLabelsOnASingleLine: false
-AllowShortFunctionsOnASingleLine: All
-AllowShortLambdasOnASingleLine: All
-AllowShortIfStatementsOnASingleLine: WithoutElse
-AllowShortLoopsOnASingleLine: true
-AlwaysBreakAfterDefinitionReturnType: None
-AlwaysBreakAfterReturnType: None
-AlwaysBreakBeforeMultilineStrings: true
-AlwaysBreakTemplateDeclarations: Yes
-BinPackArguments: true
-BinPackParameters: true
-BraceWrapping:
-  AfterCaseLabel:  false
-  AfterClass:      false
-  AfterControlStatement: false
-  AfterEnum:       false
-  AfterFunction:   false
-  AfterNamespace:  false
-  AfterObjCDeclaration: false
-  AfterStruct:     false
-  AfterUnion:      false
-  AfterExternBlock: false
-  BeforeCatch:     false
-  BeforeElse:      false
-  IndentBraces:    false
-  SplitEmptyFunction: true
-  SplitEmptyRecord: true
-  SplitEmptyNamespace: true
-BreakBeforeBinaryOperators: None
-BreakBeforeBraces: Attach
-BreakBeforeInheritanceComma: false
-BreakInheritanceList: BeforeColon
-BreakBeforeTernaryOperators: true
-BreakConstructorInitializersBeforeComma: false
-BreakConstructorInitializers: BeforeColon
-BreakAfterJavaFieldAnnotations: false
-BreakStringLiterals: true
-ColumnLimit:     80
-CommentPragmas:  '^ IWYU pragma:'
-CompactNamespaces: false
-ConstructorInitializerAllOnOneLineOrOnePerLine: true
-ConstructorInitializerIndentWidth: 4
-ContinuationIndentWidth: 4
-Cpp11BracedListStyle: true
-DeriveLineEnding: true
-DerivePointerAlignment: true
-DisableFormat:   false
-ExperimentalAutoDetectBinPacking: false
-FixNamespaceComments: true
-ForEachMacros:
-  - foreach
-  - Q_FOREACH
-  - BOOST_FOREACH
-IncludeBlocks:   Regroup
-IncludeCategories:
-  - Regex:           '^<ext/.*\.h>'
-    Priority:        2
-    SortPriority:    0
-  - Regex:           '^<.*\.h>'
-    Priority:        1
-    SortPriority:    0
-  - Regex:           '^<.*'
-    Priority:        2
-    SortPriority:    0
-  - Regex:           '.*'
-    Priority:        3
-    SortPriority:    0
-IncludeIsMainRegex: '([-_](test|unittest))?$'
-IncludeIsMainSourceRegex: ''
-IndentCaseLabels: true
-IndentGotoLabels: true
-IndentPPDirectives: None
-IndentWidth:     4
-IndentWrappedFunctionNames: false
-JavaScriptQuotes: Leave
-JavaScriptWrapImports: true
-KeepEmptyLinesAtTheStartOfBlocks: false
-MacroBlockBegin: ''
-MacroBlockEnd:   ''
-MaxEmptyLinesToKeep: 1
-NamespaceIndentation: None
-ObjCBinPackProtocolList: Never
-ObjCBlockIndentWidth: 2
-ObjCSpaceAfterProperty: false
-ObjCSpaceBeforeProtocolList: true
-PenaltyBreakAssignment: 2
-PenaltyBreakBeforeFirstCallParameter: 1
-PenaltyBreakComment: 300
-PenaltyBreakFirstLessLess: 120
-PenaltyBreakString: 1000
-PenaltyBreakTemplateDeclaration: 10
-PenaltyExcessCharacter: 1000000
-PenaltyReturnTypeOnItsOwnLine: 200
-PointerAlignment: Left
-RawStringFormats:
-  - Language:        Cpp
-    Delimiters:
-      - cc
-      - CC
-      - cpp
-      - Cpp
-      - CPP
-      - 'c++'
-      - 'C++'
-    CanonicalDelimiter: ''
-    BasedOnStyle:    google
-  - Language:        TextProto
-    Delimiters:
-      - pb
-      - PB
-      - proto
-      - PROTO
-    EnclosingFunctions:
-      - EqualsProto
-      - EquivToProto
-      - PARSE_PARTIAL_TEXT_PROTO
-      - PARSE_TEST_PROTO
-      - PARSE_TEXT_PROTO
-      - ParseTextOrDie
-      - ParseTextProtoOrDie
-    CanonicalDelimiter: ''
-    BasedOnStyle:    google
-ReflowComments:  true
-SortIncludes:    true
-SortUsingDeclarations: true
-SpaceAfterCStyleCast: false
-SpaceAfterLogicalNot: false
-SpaceAfterTemplateKeyword: true
-SpaceBeforeAssignmentOperators: true
-SpaceBeforeCpp11BracedList: false
-SpaceBeforeCtorInitializerColon: true
-SpaceBeforeInheritanceColon: true
-SpaceBeforeParens: ControlStatements
-SpaceBeforeRangeBasedForLoopColon: true
-SpaceInEmptyBlock: false
-SpaceInEmptyParentheses: false
-SpacesBeforeTrailingComments: 2
-SpacesInAngles:  false
-SpacesInConditionalStatement: false
-SpacesInContainerLiterals: true
-SpacesInCStyleCastParentheses: false
-SpacesInParentheses: false
-SpacesInSquareBrackets: false
-SpaceBeforeSquareBrackets: false
-Standard:        Auto
-StatementMacros:
-  - Q_UNUSED
-  - QT_REQUIRE_VERSION
-TabWidth:        8
-UseCRLF:         false
-UseTab:          Never
-...
-
@@ -158,6 +158,12 @@ typedef sycl::half2 ggml_half2;

 #endif // GGML_COMMON_DECL_CUDA || GGML_COMMON_DECL_HIP

+#ifdef _MSC_VER
+#define GGML_EXTENSION
+#else // _MSC_VER
+#define GGML_EXTENSION __extension__
+#endif // _MSC_VER
+
 #define QK4_0 32
 typedef struct {
    ggml_half d;           // delta
@@ -167,7 +173,7 @@ static_assert(sizeof(block_q4_0) == sizeof(ggml_half) + QK4_0 / 2, "wrong q4_0 b

 #define QK4_1 32
 typedef struct {
-    union {
+    GGML_EXTENSION union {
        struct {
            ggml_half d; // delta
            ggml_half m; // min
@@ -188,7 +194,7 @@ static_assert(sizeof(block_q5_0) == sizeof(ggml_half) + sizeof(uint32_t) + QK5_0

 #define QK5_1 32
 typedef struct {
-    union {
+    GGML_EXTENSION union {
        struct {
            ggml_half d; // delta
            ggml_half m; // min
@@ -209,7 +215,7 @@ static_assert(sizeof(block_q8_0) == sizeof(ggml_half) + QK8_0, "wrong q8_0 block

 #define QK8_1 32
 typedef struct {
-    union {
+    GGML_EXTENSION union {
        struct {
            ggml_half d; // delta
            ggml_half s; // d * sum(qs[i])
@@ -250,7 +256,7 @@ static_assert(sizeof(block_tq2_0) == sizeof(ggml_half) + QK_K / 4, "wrong tq2_0
 typedef struct {
    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
    uint8_t qs[QK_K/4];      // quants
-    union {
+    GGML_EXTENSION union {
        struct {
            ggml_half d;    // super-block scale for quantized scales
            ggml_half dmin; // super-block scale for quantized mins
@@ -277,7 +283,7 @@ static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 12
 // weight is represented as x = a * q + b
 // Effectively 4.5 bits per weight
 typedef struct {
-    union {
+    GGML_EXTENSION union {
        struct {
            ggml_half d;    // super-block scale for quantized scales
            ggml_half dmin; // super-block scale for quantized mins
@@ -294,7 +300,7 @@ static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2,
 // weight is represented as x = a * q + b
 // Effectively 5.5 bits per weight
 typedef struct {
-    union {
+    GGML_EXTENSION union {
        struct {
            ggml_half d;    // super-block scale for quantized scales
            ggml_half dmin; // super-block scale for quantized mins
@@ -23,6 +23,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
        ggml-cpu/amx/mmq.cpp
        ggml-cpu/amx/mmq.h
        ggml-cpu/ggml-cpu-impl.h
+        ggml-cpu/common.h
+        ggml-cpu/binary-ops.h
+        ggml-cpu/binary-ops.cpp
+        ggml-cpu/unary-ops.h
+        ggml-cpu/unary-ops.cpp
        )

    target_compile_features(${GGML_CPU_NAME} PRIVATE c_std_11 cxx_std_17)
@@ -289,23 +294,29 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
        endif()
    elseif ("${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "ppc64le " OR "${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "powerpc ")
        message(STATUS "PowerPC detected")
-        if(${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
-            file(READ "/proc/cpuinfo" POWER10_M)
-        elseif(${CMAKE_SYSTEM_PROCESSOR} MATCHES "powerpc")
-            execute_process(COMMAND bash -c "prtconf |grep 'Implementation' | head -n 1" OUTPUT_VARIABLE POWER10_M)
-        endif()
+        if (GGML_NATIVE)
+            if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
+                file(READ "/proc/cpuinfo" POWER10_M)
+            elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "powerpc")
+                execute_process(COMMAND bash -c "prtconf |grep 'Implementation' | head -n 1" OUTPUT_VARIABLE POWER10_M)
+            endif()

-        string(REGEX MATCHALL "POWER *([0-9]+)" MATCHED_STRING "${POWER10_M}")
-        string(REGEX REPLACE "POWER *([0-9]+)" "\\1" EXTRACTED_NUMBER "${MATCHED_STRING}")
+            string(REGEX MATCHALL "POWER *([0-9]+)" MATCHED_STRING "${POWER10_M}")
+            string(REGEX REPLACE "POWER *([0-9]+)" "\\1" EXTRACTED_NUMBER "${MATCHED_STRING}")

-        if (EXTRACTED_NUMBER GREATER_EQUAL 10)
-            list(APPEND ARCH_FLAGS -mcpu=power10 -mpowerpc64)
-        elseif (EXTRACTED_NUMBER EQUAL 9)
-            list(APPEND ARCH_FLAGS -mcpu=power9 -mpowerpc64)
-        elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
-            list(APPEND ARCH_FLAGS -mcpu=powerpc64le -mtune=native)
+            if (EXTRACTED_NUMBER GREATER_EQUAL 10)
+                list(APPEND ARCH_FLAGS -mcpu=power10 -mpowerpc64)
+            elseif (EXTRACTED_NUMBER EQUAL 9)
+                list(APPEND ARCH_FLAGS -mcpu=power9 -mpowerpc64)
+            elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
+                list(APPEND ARCH_FLAGS -mcpu=powerpc64le -mtune=native)
+            else()
+                list(APPEND ARCH_FLAGS -mcpu=native -mtune=native -mpowerpc64)
+            endif()
        else()
-            list(APPEND ARCH_FLAGS -mcpu=native -mtune=native -mpowerpc64)
+            if (GGML_CPU_POWERPC_CPUTYPE)
+                list(APPEND ARCH_FLAGS -mcpu=${GGML_CPU_POWERPC_CPUTYPE})
+            endif()
        endif()
    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
        message(STATUS "loongarch64 detected")
@@ -320,7 +331,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "riscv64")
        message(STATUS "RISC-V detected")
        if (GGML_RVV)
-            list(APPEND ARCH_FLAGS -march=rv64gcv -mabi=lp64d)
+            if (GGML_RV_ZFH)
+                list(APPEND ARCH_FLAGS -march=rv64gcv_zfhmin -DGGML_RV_ZFH -mabi=lp64d)
+            else()
+                list(APPEND ARCH_FLAGS -march=rv64gcv -mabi=lp64d)
+            endif()
        endif()
    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "s390x")
        message(STATUS "s390x detected")
@@ -359,9 +374,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)

        # Fetch KleidiAI sources:
        include(FetchContent)
-        set(KLEIDIAI_COMMIT_TAG "v1.3.0")
+        set(KLEIDIAI_COMMIT_TAG "v1.5.0")
        set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
-        set(KLEIDIAI_ARCHIVE_MD5  "060bd2dc64642b091f461cc8dd7426d9")
+        set(KLEIDIAI_ARCHIVE_MD5  "ea22e1aefb800e9bc8c74d91633cc58e")

        if (POLICY CMP0135)
            cmake_policy(SET CMP0135 NEW)
@@ -0,0 +1,158 @@
+#include "binary-ops.h"
+
+#if defined(GGML_USE_ACCELERATE)
+#include <Accelerate/Accelerate.h>
+
+using vDSP_fn_t = void (*)(const float *, vDSP_Stride, const float *, vDSP_Stride, float *, vDSP_Stride, vDSP_Length);
+#endif
+
+static inline float op_add(float a, float b) {
+    return a + b;
+}
+
+static inline float op_sub(float a, float b) {
+    return a - b;
+}
+
+static inline float op_mul(float a, float b) {
+    return a * b;
+}
+
+static inline float op_div(float a, float b) {
+    return a / b;
+}
+
+template <float (*op)(float, float), typename src0_t, typename src1_t, typename dst_t>
+static inline void vec_binary_op_contiguous(const int64_t n, dst_t * z, const src0_t * x, const src1_t * y) {
+    constexpr auto src0_to_f32 = type_conversion_table<src0_t>::to_f32;
+    constexpr auto src1_to_f32 = type_conversion_table<src1_t>::to_f32;
+    constexpr auto f32_to_dst  = type_conversion_table<dst_t >::from_f32;
+
+    for (int i = 0; i < n; i++) {
+        z[i] = f32_to_dst(op(src0_to_f32(x[i]), src1_to_f32(y[i])));
+    }
+}
+
+template <float (*op)(float, float), typename src0_t, typename src1_t, typename dst_t>
+static inline void vec_binary_op_non_contiguous(const int64_t n, const int64_t ne10, const int64_t nb10, dst_t * z, const src0_t * x, const src1_t * y) {
+    constexpr auto src0_to_f32 = type_conversion_table<src0_t>::to_f32;
+    constexpr auto src1_to_f32 = type_conversion_table<src1_t>::to_f32;
+    constexpr auto f32_to_dst  = type_conversion_table<dst_t >::from_f32;
+
+    for (int i = 0; i < n; i++) {
+        int i10 = i % ne10;
+        const src1_t * y_ptr = (const src1_t *)((const char *)y + i10*nb10);
+        z[i] = f32_to_dst(op(src0_to_f32(x[i]), src1_to_f32(*y_ptr)));
+    }
+}
+
+template <float (*op)(float, float), typename src0_t, typename src1_t, typename dst_t>
+static void apply_binary_op(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    GGML_ASSERT( nb0 == sizeof(dst_t));
+    GGML_ASSERT(nb00 == sizeof(src0_t));
+
+    const auto [ir0, ir1] = get_thread_range(params, src0);
+    const bool is_src1_contiguous = (nb10 == sizeof(src1_t));
+
+    if (!is_src1_contiguous) { // broadcast not implemented yet for non-contiguous
+        GGML_ASSERT(ggml_are_same_shape(src0, src1));
+    }
+
+#ifdef GGML_USE_ACCELERATE
+    vDSP_fn_t vDSP_op = nullptr;
+    // TODO - avoid the f32-only check using type 'trait' lookup tables and row-based src-to-float conversion functions
+    if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+        if (op == op_add) {
+            vDSP_op = vDSP_vadd;
+        } else if (op == op_sub) {
+            vDSP_op = vDSP_vsub;
+        } else if (op == op_mul) {
+            vDSP_op = vDSP_vmul;
+        } else if (op == op_div) {
+            vDSP_op = vDSP_vdiv;
+        }
+    }
+#endif
+
+    for (int64_t ir = ir0; ir < ir1; ++ir) {
+        const int64_t i03 = ir/(ne02*ne01);
+        const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+        const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
+
+        const int64_t i13 = i03 % ne13;
+        const int64_t i12 = i02 % ne12;
+        const int64_t i11 = i01 % ne11;
+
+        dst_t        * dst_ptr  = (dst_t  *)       ((char *)       dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
+        const src0_t * src0_ptr = (const src0_t *) ((const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
+        const src1_t * src1_ptr = (const src1_t *) ((const char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);
+
+        if (is_src1_contiguous) {
+            // src1 is broadcastable across src0 and dst in i1, i2, i3
+            const int64_t nr0 = ne00 / ne10;
+
+            for (int64_t r = 0; r < nr0; ++r) {
+#ifdef GGML_USE_ACCELERATE
+                if constexpr (std::is_same_v<src0_t, float> && std::is_same_v<src1_t, float> && std::is_same_v<dst_t, float>) {
+                    if (vDSP_op != nullptr) {
+                        vDSP_op(src1_ptr, 1, src0_ptr + r*ne10, 1, dst_ptr + r*ne10, 1, ne10);
+                        continue;
+                    }
+                }
+#endif
+                vec_binary_op_contiguous<op>(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr);
+            }
+        } else {
+            vec_binary_op_non_contiguous<op>(ne0, ne10, nb10, dst_ptr, src0_ptr, src1_ptr);
+        }
+    }
+}
+
+// TODO: Use the 'traits' lookup table (for type conversion fns), instead of a mass of 'if' conditions with long templates
+template <float (*op)(float, float)>
+static void binary_op(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    /*  */ if (src0->type == GGML_TYPE_F32  && src1->type == GGML_TYPE_F32  && dst->type == GGML_TYPE_F32) { // all f32
+        apply_binary_op<op, float, float, float>(params, dst);
+    } else if (src0->type == GGML_TYPE_F16  && src1->type == GGML_TYPE_F16  && dst->type == GGML_TYPE_F16) { // all f16
+        apply_binary_op<op, ggml_fp16_t, ggml_fp16_t, ggml_fp16_t>(params, dst);
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_BF16) { // all bf16
+        apply_binary_op<op, ggml_bf16_t, ggml_bf16_t, ggml_bf16_t>(params, dst);
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32  && dst->type == GGML_TYPE_BF16) {
+        apply_binary_op<op, ggml_bf16_t, float, ggml_bf16_t>(params, dst);
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32  && dst->type == GGML_TYPE_F32) {
+        apply_binary_op<op, ggml_bf16_t, float, float>(params, dst);
+    } else if (src0->type == GGML_TYPE_F16  && src1->type == GGML_TYPE_F32  && dst->type == GGML_TYPE_F16) {
+        apply_binary_op<op, ggml_fp16_t, float, ggml_fp16_t>(params, dst);
+    } else if (src0->type == GGML_TYPE_F16  && src1->type == GGML_TYPE_F32  && dst->type == GGML_TYPE_F32) {
+        apply_binary_op<op, ggml_fp16_t, float, float>(params, dst);
+    } else {
+        GGML_ABORT("%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));
+    }
+}
+
+void ggml_compute_forward_add_non_quantized(const ggml_compute_params * params, ggml_tensor * dst) {
+    binary_op<op_add>(params, dst);
+}
+
+void ggml_compute_forward_sub(const ggml_compute_params * params, ggml_tensor * dst) {
+    binary_op<op_sub>(params, dst);
+}
+
+void ggml_compute_forward_mul(const ggml_compute_params * params, ggml_tensor * dst) {
+    binary_op<op_mul>(params, dst);
+}
+
+void ggml_compute_forward_div(const ggml_compute_params * params, ggml_tensor * dst) {
+    binary_op<op_div>(params, dst);
+}
@@ -0,0 +1,16 @@
+#pragma once
+
+#include "common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void ggml_compute_forward_add_non_quantized(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_sub(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_mul(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_div(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+
+#ifdef __cplusplus
+}
+#endif
@@ -0,0 +1,72 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-cpu-traits.h"
+#include "ggml-cpu-impl.h"
+#include "ggml-impl.h"
+
+#ifdef __cplusplus
+
+#include <utility>
+
+// convenience functions/macros for use in template calls
+// note: these won't be required after the 'traits' lookup table is used.
+static inline ggml_fp16_t f32_to_f16(float x) {
+    return GGML_FP32_TO_FP16(x);
+}
+
+static inline float f16_to_f32(ggml_fp16_t x) {
+    return GGML_FP16_TO_FP32(x);
+}
+
+static inline ggml_bf16_t f32_to_bf16(float x) {
+    return GGML_FP32_TO_BF16(x);
+}
+
+static inline float bf16_to_f32(ggml_bf16_t x) {
+    return GGML_BF16_TO_FP32(x);
+}
+
+static inline float f32_to_f32(float x) {
+    return x;
+}
+
+// TODO - merge this into the traits table, after using row-based conversions
+template <class T>
+struct type_conversion_table;
+
+template <>
+struct type_conversion_table<ggml_fp16_t> {
+    static constexpr float (*to_f32)(ggml_fp16_t) = f16_to_f32;
+    static constexpr ggml_fp16_t (*from_f32)(float) = f32_to_f16;
+};
+
+template <>
+struct type_conversion_table<float> {
+    static constexpr float (*to_f32)(float) = f32_to_f32;
+    static constexpr float (*from_f32)(float) = f32_to_f32;
+};
+
+template <>
+struct type_conversion_table<ggml_bf16_t> {
+    static constexpr float (*to_f32)(ggml_bf16_t) = bf16_to_f32;
+    static constexpr ggml_bf16_t (*from_f32)(float) = f32_to_bf16;
+};
+
+static std::pair<int64_t, int64_t> get_thread_range(const struct ggml_compute_params * params, const struct ggml_tensor * src0) {
+    const int64_t ith = params->ith;
+    const int64_t nth = params->nth;
+
+    const int64_t nr  = ggml_nrows(src0);
+
+    // rows per thread
+    const int64_t dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int64_t ir0 = dr*ith;
+    const int64_t ir1 = MIN(ir0 + dr, nr);
+
+    return {ir0, ir1};
+}
+
+#endif
@@ -51,11 +51,10 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
            /* .run_kernel            = */ kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
        },
        /* .lhs_info = */ {
-            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
+            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32_neon,
+            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32_neon,
            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32_neon,
            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32_neon,
-            /* .require_aligned_m_idx = */ true,
        },
        /* .rhs_info = */ {
            /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
@@ -100,7 +99,6 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
-            /* .require_aligned_m_idx = */ false,
        },
        /* .rhs_info = */ {
            /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
@@ -144,7 +142,6 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
-            /* .require_aligned_m_idx = */ false,
        },
        /* .rhs_info = */ {
            /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
@@ -189,7 +186,6 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
-            /* .require_aligned_m_idx = */ false,
        },
        /* .rhs_info = */ {
            /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
@@ -233,7 +229,6 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
-            /* .require_aligned_m_idx = */ false,
        },
        /* .rhs_info = */ {
            /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
@@ -40,7 +40,6 @@ struct lhs_packing_info {
    size_t (*packed_size)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr);
    void (*pack_func)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const float* lhs,
                      size_t lhs_stride, void* lhs_packed);
-    bool require_aligned_m_idx;
 };

 struct rhs_packing_info {
@@ -124,8 +124,7 @@ class tensor_traits : public ggml::cpu::tensor_traits {
            size_t sr = kernel->get_sr();

            // Calculate number of columns to be processed per thread
-            const bool use_multithread = lhs_info->require_aligned_m_idx && m <= mr ? false : true;
-            const size_t num_m_per_thread = use_multithread ? kai_roundup(m, nth) / nth : m;
+            const size_t num_m_per_thread = kai_roundup(m, mr * nth) / nth;
            const size_t m_start = ith * num_m_per_thread;
            size_t m_to_process = num_m_per_thread;
            if ((m_start + m_to_process) > m) {
@@ -135,11 +134,11 @@ class tensor_traits : public ggml::cpu::tensor_traits {
            if(m_start < m) {
                // Transform LHS
                const size_t src_stride        = src1->nb[1];
-                const float * src_ptr          = reinterpret_cast<const float *>(lhs + lhs_info->get_offset(0, dst->src[1]->nb[1]));
+                const float * src_ptr          = reinterpret_cast<const float *>(lhs + lhs_info->get_offset(m_start, dst->src[1]->nb[1]));
                const size_t lhs_packed_offset = lhs_info->get_packed_offset(m_start, k, QK4_0, mr, kr, sr);
                void * lhs_packed_ptr          = static_cast<void *>(lhs_packed + lhs_packed_offset);

-                lhs_info->pack_func(m_to_process, k, QK4_0, mr, kr, sr, m_start, src_ptr, src_stride, lhs_packed_ptr);
+                lhs_info->pack_func(m_to_process, k, QK4_0, mr, kr, sr, 0, src_ptr, src_stride, lhs_packed_ptr);
            }

            ggml_barrier(params->threadpool);
@@ -55,6 +55,7 @@

 #include <atomic>
 #include <array>
+#include <type_traits>

 #ifdef _MSC_VER
 #define NOINLINE __declspec(noinline)
@@ -1092,13 +1093,403 @@ class tinyBLAS_Q0_PPC {
       }
    }

-    template<typename VA, typename VB>
-    void packNormal(const TA* a, int64_t lda, int rows, int cols, VA* vec, bool flip) {
+    template<typename VA, typename VB, int size>
+    void packNormalInt4(const TA* a, int64_t lda, int rows, int cols, VA* vec, std::array<int, size>& comparray) {
        int64_t i, j;
        TA *aoffset = NULL;
        VA *vecOffset = NULL;
        TA *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL;
        TA *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL;
+        VB c1[2] = {0}, c2[2] = {0}, c3[2] = {0}, c4[2] = {0};
+        VB c5[2] = {0}, c6[2] = {0}, c7[2] = {0}, c8[2] = {0};
+        VB t1, t2, t3, t4, t5, t6, t7, t8;
+        const vector signed char lowMask = vec_splats((signed char)0xF);
+        const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+        const vector signed char v8 = vec_splats((signed char)0x8);
+        aoffset = const_cast<TA*>(a);
+        vecOffset = vec;
+        vector unsigned char swiz1 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
+        vector unsigned char swiz2 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31};
+        vector unsigned char swiz3 = {0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27};
+        vector unsigned char swiz4 = {4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31};
+        vector signed int vsum = {0};
+        vector signed int vsum2 = {0};
+
+        j = (rows >> 3);
+        if (j > 0) {
+            do {
+                aoffset1 = aoffset;
+                aoffset2 = aoffset1 + lda;
+                aoffset3 = aoffset2 + lda;
+                aoffset4 = aoffset3 + lda;
+                aoffset5 = aoffset4 + lda;
+                aoffset6 = aoffset5 + lda;
+                aoffset7 = aoffset6 + lda;
+                aoffset8 = aoffset7 + lda;
+                aoffset += 8 * lda;
+
+                i = (cols >> 2);
+                if (i > 0) {
+                    do {
+                        c1[1] = reinterpret_cast<VB>(vec_xl(0, aoffset1->qs));
+                        c2[1] = reinterpret_cast<VB>(vec_xl(0, aoffset2->qs));
+                        c3[1] = reinterpret_cast<VB>(vec_xl(0, aoffset3->qs));
+                        c4[1] = reinterpret_cast<VB>(vec_xl(0, aoffset4->qs));
+                        c5[1] = reinterpret_cast<VB>(vec_xl(0, aoffset5->qs));
+                        c6[1] = reinterpret_cast<VB>(vec_xl(0, aoffset6->qs));
+                        c7[1] = reinterpret_cast<VB>(vec_xl(0, aoffset7->qs));
+                        c8[1] = reinterpret_cast<VB>(vec_xl(0, aoffset8->qs));
+
+                        c1[0] = vec_and(c1[1], lowMask);
+                        c1[1] = vec_sr(c1[1], v4);
+                        c1[0] = vec_sub(c1[0], v8);
+                        c1[1] = vec_sub(c1[1], v8);
+                        vsum = vec_sum4s(c1[0], vsum);
+                        vsum2 = vec_sum4s(c1[1], vsum2);
+                        vsum = vec_add(vsum, vsum2);
+                        comparray[0] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                        vsum = vec_splats(0);
+                        vsum2 = vec_splats(0);
+
+                        c2[0] = vec_and(c2[1], lowMask);
+                        c2[1] = vec_sr(c2[1], v4);
+                        c2[0] = vec_sub(c2[0], v8);
+                        c2[1] = vec_sub(c2[1], v8);
+                        vsum = vec_sum4s(c2[0], vsum);
+                        vsum2 = vec_sum4s(c2[1], vsum2);
+                        vsum = vec_add(vsum, vsum2);
+                        comparray[1] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                        vsum = vec_splats(0);
+                        vsum2 = vec_splats(0);
+
+                        c3[0] = vec_and(c3[1], lowMask);
+                        c3[1] = vec_sr(c3[1], v4);
+                        c3[0] = vec_sub(c3[0], v8);
+                        c3[1] = vec_sub(c3[1], v8);
+                        vsum = vec_sum4s(c3[0], vsum);
+                        vsum2 = vec_sum4s(c3[1], vsum2);
+                        vsum = vec_add(vsum, vsum2);
+                        comparray[2] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                        vsum = vec_splats(0);
+                        vsum2 = vec_splats(0);
+
+                        c4[0] = vec_and(c4[1], lowMask);
+                        c4[1] = vec_sr(c4[1], v4);
+                        c4[0] = vec_sub(c4[0], v8);
+                        c4[1] = vec_sub(c4[1], v8);
+                        vsum = vec_sum4s(c4[0], vsum);
+                        vsum2 = vec_sum4s(c4[1], vsum2);
+                        vsum = vec_add(vsum, vsum2);
+                        comparray[3] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                        vsum = vec_splats(0);
+                        vsum2 = vec_splats(0);
+
+                        c5[0] = vec_and(c5[1], lowMask);
+                        c5[1] = vec_sr(c5[1], v4);
+                        c5[0] = vec_sub(c5[0], v8);
+                        c5[1] = vec_sub(c5[1], v8);
+                        vsum = vec_sum4s(c5[0], vsum);
+                        vsum2 = vec_sum4s(c5[1], vsum2);
+                        vsum = vec_add(vsum, vsum2);
+                        comparray[4] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                        vsum = vec_splats(0);
+                        vsum2 = vec_splats(0);
+
+                        c6[0] = vec_and(c6[1], lowMask);
+                        c6[1] = vec_sr(c6[1], v4);
+                        c6[0] = vec_sub(c6[0], v8);
+                        c6[1] = vec_sub(c6[1], v8);
+                        vsum = vec_sum4s(c6[0], vsum);
+                        vsum2 = vec_sum4s(c6[1], vsum2);
+                        vsum = vec_add(vsum, vsum2);
+                        comparray[5] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                        vsum = vec_splats(0);
+                        vsum2 = vec_splats(0);
+
+                        c7[0] = vec_and(c7[1], lowMask);
+                        c7[1] = vec_sr(c7[1], v4);
+                        c7[0] = vec_sub(c7[0], v8);
+                        c7[1] = vec_sub(c7[1], v8);
+                        vsum = vec_sum4s(c7[0], vsum);
+                        vsum2 = vec_sum4s(c7[1], vsum2);
+                        vsum = vec_add(vsum, vsum2);
+                        comparray[6] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                        vsum = vec_splats(0);
+                        vsum2 = vec_splats(0);
+
+                        c8[0] = vec_and(c8[1], lowMask);
+                        c8[1] = vec_sr(c8[1], v4);
+                        c8[0] = vec_sub(c8[0], v8);
+                        c8[1] = vec_sub(c8[1], v8);
+                        vsum = vec_sum4s(c8[0], vsum);
+                        vsum2 = vec_sum4s(c8[1], vsum2);
+                        vsum = vec_add(vsum, vsum2);
+                        comparray[7] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                        vsum = vec_splats(0);
+                        vsum2 = vec_splats(0);
+
+                        t1 = vec_perm(c1[0], c2[0], swiz1);
+                        t2 = vec_perm(c1[0], c2[0], swiz2);
+                        t3 = vec_perm(c3[0], c4[0], swiz1);
+                        t4 = vec_perm(c3[0], c4[0], swiz2);
+                        t5 = vec_perm(t1, t3, swiz3);
+                        t6 = vec_perm(t1, t3, swiz4);
+                        t7 = vec_perm(t2, t4, swiz3);
+                        t8 = vec_perm(t2, t4, swiz4);
+                        vec_xst(t5, 0, vecOffset);
+                        vec_xst(t6, 0, vecOffset+16);
+                        vec_xst(t7, 0, vecOffset+32);
+                        vec_xst(t8, 0, vecOffset+48);
+
+                        t1 = vec_perm(c1[1], c2[1], swiz1);
+                        t2 = vec_perm(c1[1], c2[1], swiz2);
+                        t3 = vec_perm(c3[1], c4[1], swiz1);
+                        t4 = vec_perm(c3[1], c4[1], swiz2);
+                        t5 = vec_perm(t1, t3, swiz3);
+                        t6 = vec_perm(t1, t3, swiz4);
+                        t7 = vec_perm(t2, t4, swiz3);
+                        t8 = vec_perm(t2, t4, swiz4);
+                        vec_xst(t5, 0, vecOffset+64);
+                        vec_xst(t6, 0, vecOffset+80);
+                        vec_xst(t7, 0, vecOffset+96);
+                        vec_xst(t8, 0, vecOffset+112);
+
+                        t1 = vec_perm(c5[0], c6[0], swiz1);
+                        t2 = vec_perm(c5[0], c6[0], swiz2);
+                        t3 = vec_perm(c7[0], c8[0], swiz1);
+                        t4 = vec_perm(c7[0], c8[0], swiz2);
+                        t5 = vec_perm(t1, t3, swiz3);
+                        t6 = vec_perm(t1, t3, swiz4);
+                        t7 = vec_perm(t2, t4, swiz3);
+                        t8 = vec_perm(t2, t4, swiz4);
+                        vec_xst(t5, 0, vecOffset+128);
+                        vec_xst(t6, 0, vecOffset+144);
+                        vec_xst(t7, 0, vecOffset+160);
+                        vec_xst(t8, 0, vecOffset+176);
+
+                        t1 = vec_perm(c5[1], c6[1], swiz1);
+                        t2 = vec_perm(c5[1], c6[1], swiz2);
+                        t3 = vec_perm(c7[1], c8[1], swiz1);
+                        t4 = vec_perm(c7[1], c8[1], swiz2);
+                        t5 = vec_perm(t1, t3, swiz3);
+                        t6 = vec_perm(t1, t3, swiz4);
+                        t7 = vec_perm(t2, t4, swiz3);
+                        t8 = vec_perm(t2, t4, swiz4);
+                        vec_xst(t5, 0, vecOffset+192);
+                        vec_xst(t6, 0, vecOffset+208);
+                        vec_xst(t7, 0, vecOffset+224);
+                        vec_xst(t8, 0, vecOffset+240);
+
+                        aoffset1 += lda;
+                        aoffset2 += lda;
+                        aoffset3 += lda;
+                        aoffset4 += lda;
+                        aoffset5 += lda;
+                        aoffset6 += lda;
+                        aoffset7 += lda;
+                        aoffset8 += lda;
+                        vecOffset += 256;
+                        i--;
+                    } while (i > 0);
+                }
+                j--;
+            } while (j > 0);
+        }
+
+        if (rows & 4) {
+            aoffset1 = aoffset;
+            aoffset2 = aoffset1 + lda;
+            aoffset3 = aoffset2 + lda;
+            aoffset4 = aoffset3 + lda;
+            aoffset += 4 * lda;
+
+            i = (cols >> 2);
+            if (i > 0) {
+                do {
+                    c1[1] = reinterpret_cast<VB>(vec_xl(0, aoffset1->qs));
+                    c2[1] = reinterpret_cast<VB>(vec_xl(0, aoffset2->qs));
+                    c3[1] = reinterpret_cast<VB>(vec_xl(0, aoffset3->qs));
+                    c4[1] = reinterpret_cast<VB>(vec_xl(0, aoffset4->qs));
+
+                    c1[0] = vec_and(c1[1], lowMask);
+                    c1[1] = vec_sr(c1[1], v4);
+                    c1[0] = vec_sub(c1[0], v8);
+                    c1[1] = vec_sub(c1[1], v8);
+                    vsum = vec_sum4s(c1[0], vsum);
+                    vsum2 = vec_sum4s(c1[1], vsum2);
+                    vsum = vec_add(vsum, vsum2);
+                    comparray[0] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                    vsum = vec_splats(0);
+                    vsum2 = vec_splats(0);
+
+                    c2[0] = vec_and(c2[1], lowMask);
+                    c2[1] = vec_sr(c2[1], v4);
+                    c2[0] = vec_sub(c2[0], v8);
+                    c2[1] = vec_sub(c2[1], v8);
+                    vsum = vec_sum4s(c2[0], vsum);
+                    vsum2 = vec_sum4s(c2[1], vsum2);
+                    vsum = vec_add(vsum, vsum2);
+                    comparray[1] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                    vsum = vec_splats(0);
+                    vsum2 = vec_splats(0);
+
+                    c3[0] = vec_and(c3[1], lowMask);
+                    c3[1] = vec_sr(c3[1], v4);
+                    c3[0] = vec_sub(c3[0], v8);
+                    c3[1] = vec_sub(c3[1], v8);
+                    vsum = vec_sum4s(c3[0], vsum);
+                    vsum2 = vec_sum4s(c3[1], vsum2);
+                    vsum = vec_add(vsum, vsum2);
+                    comparray[2] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                    vsum = vec_splats(0);
+                    vsum2 = vec_splats(0);
+
+                    c4[0] = vec_and(c4[1], lowMask);
+                    c4[1] = vec_sr(c4[1], v4);
+                    c4[0] = vec_sub(c4[0], v8);
+                    c4[1] = vec_sub(c4[1], v8);
+                    vsum = vec_sum4s(c4[0], vsum);
+                    vsum2 = vec_sum4s(c4[1], vsum2);
+                    vsum = vec_add(vsum, vsum2);
+                    comparray[3] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                    vsum = vec_splats(0);
+                    vsum2 = vec_splats( 0);
+
+                    t1 = vec_perm(c1[0], c2[0], swiz1);
+                    t2 = vec_perm(c1[0], c2[0], swiz2);
+                    t3 = vec_perm(c3[0], c4[0], swiz1);
+                    t4 = vec_perm(c3[0], c4[0], swiz2);
+                    t5 = vec_perm(t1, t3, swiz3);
+                    t6 = vec_perm(t1, t3, swiz4);
+                    t7 = vec_perm(t2, t4, swiz3);
+                    t8 = vec_perm(t2, t4, swiz4);
+                    vec_xst(t5, 0, vecOffset);
+                    vec_xst(t6, 0, vecOffset+16);
+                    vec_xst(t7, 0, vecOffset+32);
+                    vec_xst(t8, 0, vecOffset+48);
+
+                    t1 = vec_perm(c1[1], c2[1], swiz1);
+                    t2 = vec_perm(c1[1], c2[1], swiz2);
+                    t3 = vec_perm(c3[1], c4[1], swiz1);
+                    t4 = vec_perm(c3[1], c4[1], swiz2);
+                    t5 = vec_perm(t1, t3, swiz3);
+                    t6 = vec_perm(t1, t3, swiz4);
+                    t7 = vec_perm(t2, t4, swiz3);
+                    t8 = vec_perm(t2, t4, swiz4);
+                    vec_xst(t5, 0, vecOffset+64);
+                    vec_xst(t6, 0, vecOffset+80);
+                    vec_xst(t7, 0, vecOffset+96);
+                    vec_xst(t8, 0, vecOffset+112);
+
+                    aoffset1 += lda;
+                    aoffset2 += lda;
+                    aoffset3 += lda;
+                    aoffset4 += lda;
+                    vecOffset += 128;
+                    i--;
+                } while (i > 0);
+            }
+        }
+
+        if (rows & 3) {
+            aoffset1 = aoffset;
+            aoffset2 = aoffset1 + lda;
+            aoffset3 = aoffset2 + lda;
+            i = (cols >> 2);
+            if (i > 0) {
+                do {
+                    switch(rows) {
+                        case 3: c3[1] = reinterpret_cast<VB>(vec_xl(0, aoffset3->qs));
+                        case 2: c2[1] = reinterpret_cast<VB>(vec_xl(0, aoffset2->qs));
+                        case 1: c1[1] = reinterpret_cast<VB>(vec_xl(0, aoffset1->qs));
+                            break;
+                    }
+                    c1[0] = vec_and(c1[1], lowMask);
+                    c1[1] = vec_sr(c1[1], v4);
+                    c1[0] = vec_sub(c1[0], v8);
+                    c1[1] = vec_sub(c1[1], v8);
+                    vsum = vec_sum4s(c1[0], vsum);
+                    vsum2 = vec_sum4s(c1[1], vsum2);
+                    vsum = vec_add(vsum, vsum2);
+                    comparray[0] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                    vsum = vec_splats(0);
+                    vsum2 = vec_splats(0);
+
+                    c2[0] = vec_and(c2[1], lowMask);
+                    c2[1] = vec_sr(c2[1], v4);
+                    c2[0] = vec_sub(c2[0], v8);
+                    c2[1] = vec_sub(c2[1], v8);
+                    vsum = vec_sum4s(c2[0], vsum);
+                    vsum2 = vec_sum4s(c2[1], vsum2);
+                    vsum = vec_add(vsum, vsum2);
+                    comparray[1] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                    vsum = vec_splats(0);
+                    vsum2 = vec_splats(0);
+
+                    c3[0] = vec_and(c3[1], lowMask);
+                    c3[1] = vec_sr(c3[1], v4);
+                    c3[0] = vec_sub(c3[0], v8);
+                    c3[1] = vec_sub(c3[1], v8);
+                    vsum = vec_sum4s(c3[0], vsum);
+                    vsum2 = vec_sum4s(c3[1], vsum2);
+                    vsum = vec_add(vsum, vsum2);
+                    comparray[2] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                    vsum = vec_splats(0);
+                    vsum2 = vec_splats(0);
+
+                    c4[0] = vec_and(c4[1], lowMask);
+                    c4[1] = vec_sr(c4[1], v4);
+                    c4[0] = vec_sub(c4[0], v8);
+                    c4[1] = vec_sub(c4[1], v8);
+                    vsum = vec_sum4s(c4[0], vsum);
+                    vsum2 = vec_sum4s(c4[1], vsum2);
+                    vsum = vec_add(vsum, vsum2);
+                    comparray[3] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                    vsum = vec_splats(0);
+                    vsum2 = vec_splats(0);
+
+                    t1 = vec_perm(c1[0], c2[0], swiz1);
+                    t2 = vec_perm(c1[0], c2[0], swiz2);
+                    t3 = vec_perm(c3[0], c4[0], swiz1);
+                    t4 = vec_perm(c3[0], c4[0], swiz2);
+                    t5 = vec_perm(t1, t3, swiz3);
+                    t6 = vec_perm(t1, t3, swiz4);
+                    t7 = vec_perm(t2, t4, swiz3);
+                    t8 = vec_perm(t2, t4, swiz4);
+                    vec_xst(t5, 0, vecOffset);
+                    vec_xst(t6, 0, vecOffset+16);
+                    vec_xst(t7, 0, vecOffset+32);
+                    vec_xst(t8, 0, vecOffset+48);
+
+                    t1 = vec_perm(c1[1], c2[1], swiz1);
+                    t2 = vec_perm(c1[1], c2[1], swiz2);
+                    t3 = vec_perm(c3[1], c4[1], swiz1);
+                    t4 = vec_perm(c3[1], c4[1], swiz2);
+                    t5 = vec_perm(t1, t3, swiz3);
+                    t6 = vec_perm(t1, t3, swiz4);
+                    t7 = vec_perm(t2, t4, swiz3);
+                    t8 = vec_perm(t2, t4, swiz4);
+                    vec_xst(t5, 0, vecOffset+64);
+                    vec_xst(t6, 0, vecOffset+80);
+                    vec_xst(t7, 0, vecOffset+96);
+                    vec_xst(t8, 0, vecOffset+112);
+                    aoffset1 += lda;
+                    aoffset2 += lda;
+                    aoffset3 += lda;
+                    vecOffset += 128;
+                    i--;
+                } while(i > 0);
+            }
+        }
+    }
+
+    template<typename VA, typename VB>
+    void packNormal(const TB* a, int64_t lda, int rows, int cols, VA* vec, bool flip) {
+        int64_t i, j;
+        TB *aoffset = NULL;
+        VA *vecOffset = NULL;
+        TB *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL;
+        TB *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL;
        __vector_pair C1, C2, C3, C4, C5, C6, C7, C8;
        VB c1[2] = {0}, c2[2] = {0}, c3[2] = {0}, c4[2]={0};
        VB c5[2] = {0}, c6[2] = {0}, c7[2] = {0}, c8[2]={0};
@@ -1111,24 +1502,24 @@ class tinyBLAS_Q0_PPC {
        vector unsigned char swiz3 = {0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27};
        vector unsigned char swiz4 = {4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31};

-        aoffset = const_cast<TA*>(a);
+        aoffset = const_cast<TB*>(a);
        vecOffset = vec;
        j = (rows >> 3);
        if (j > 0) {
            do {
-            aoffset1 = aoffset;
-            aoffset2 = aoffset1 + lda;
-            aoffset3 = aoffset2 + lda;
-            aoffset4 = aoffset3 + lda;
-            aoffset5 = aoffset4 + lda;
-            aoffset6 = aoffset5 + lda;
-            aoffset7 = aoffset6 + lda;
-            aoffset8 = aoffset7 + lda;
-            aoffset += 8 * lda;
+                aoffset1 = aoffset;
+                aoffset2 = aoffset1 + lda;
+                aoffset3 = aoffset2 + lda;
+                aoffset4 = aoffset3 + lda;
+                aoffset5 = aoffset4 + lda;
+                aoffset6 = aoffset5 + lda;
+                aoffset7 = aoffset6 + lda;
+                aoffset8 = aoffset7 + lda;
+                aoffset += 8 * lda;

-            i = (cols >> 3);
-            if (i > 0) {
-               do {
+                i = (cols >> 3);
+                if (i > 0) {
+                do {
                    C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1->qs);
                    C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2->qs);
                    C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3->qs);
@@ -1156,10 +1547,10 @@ class tinyBLAS_Q0_PPC {
                    t7 = vec_perm(t2, t4, swiz3);
                    t8 = vec_perm(t2, t4, swiz4);
                    if (flip == true) {
-                       t5 = vec_xor(t5, xor_vector);
-                       t6 = vec_xor(t6, xor_vector);
-                       t7 = vec_xor(t7, xor_vector);
-                       t8 = vec_xor(t8, xor_vector);
+                        t5 = vec_xor(t5, xor_vector);
+                        t6 = vec_xor(t6, xor_vector);
+                        t7 = vec_xor(t7, xor_vector);
+                        t8 = vec_xor(t8, xor_vector);
                    }
                    vec_xst(t5, 0, vecOffset);
                    vec_xst(t6, 0, vecOffset+16);
@@ -1175,10 +1566,10 @@ class tinyBLAS_Q0_PPC {
                    t7 = vec_perm(t2, t4, swiz3);
                    t8 = vec_perm(t2, t4, swiz4);
                    if (flip == true) {
-                       t5 = vec_xor(t5, xor_vector);
-                       t6 = vec_xor(t6, xor_vector);
-                       t7 = vec_xor(t7, xor_vector);
-                       t8 = vec_xor(t8, xor_vector);
+                        t5 = vec_xor(t5, xor_vector);
+                        t6 = vec_xor(t6, xor_vector);
+                        t7 = vec_xor(t7, xor_vector);
+                        t8 = vec_xor(t8, xor_vector);
                    }
                    vec_xst(t5, 0, vecOffset+64);
                    vec_xst(t6, 0, vecOffset+80);
@@ -1194,10 +1585,10 @@ class tinyBLAS_Q0_PPC {
                    t7 = vec_perm(t2, t4, swiz3);
                    t8 = vec_perm(t2, t4, swiz4);
                    if (flip == true) {
-                       t5 = vec_xor(t5, xor_vector);
-                       t6 = vec_xor(t6, xor_vector);
-                       t7 = vec_xor(t7, xor_vector);
-                       t8 = vec_xor(t8, xor_vector);
+                        t5 = vec_xor(t5, xor_vector);
+                        t6 = vec_xor(t6, xor_vector);
+                        t7 = vec_xor(t7, xor_vector);
+                        t8 = vec_xor(t8, xor_vector);
                    }
                    vec_xst(t5, 0, vecOffset+128);
                    vec_xst(t6, 0, vecOffset+144);
@@ -1213,10 +1604,10 @@ class tinyBLAS_Q0_PPC {
                    t7 = vec_perm(t2, t4, swiz3);
                    t8 = vec_perm(t2, t4, swiz4);
                    if (flip == true) {
-                       t5 = vec_xor(t5, xor_vector);
-                       t6 = vec_xor(t6, xor_vector);
-                       t7 = vec_xor(t7, xor_vector);
-                       t8 = vec_xor(t8, xor_vector);
+                        t5 = vec_xor(t5, xor_vector);
+                        t6 = vec_xor(t6, xor_vector);
+                        t7 = vec_xor(t7, xor_vector);
+                        t8 = vec_xor(t8, xor_vector);
                    }
                    vec_xst(t5, 0, vecOffset+192);
                    vec_xst(t6, 0, vecOffset+208);
@@ -1240,11 +1631,11 @@ class tinyBLAS_Q0_PPC {
    }

    if (rows & 4) {
-            aoffset1 = aoffset;
-            aoffset2 = aoffset1 + lda;
-            aoffset3 = aoffset2 + lda;
-            aoffset4 = aoffset3 + lda;
-            aoffset += 4 * lda;
+        aoffset1 = aoffset;
+        aoffset2 = aoffset1 + lda;
+        aoffset3 = aoffset2 + lda;
+        aoffset4 = aoffset3 + lda;
+        aoffset += 4 * lda;

        i = (cols >> 3);
            if (i > 0) {
@@ -1311,7 +1702,7 @@ class tinyBLAS_Q0_PPC {
            aoffset2 = aoffset1 + lda;
            aoffset3 = aoffset2 + lda;
            i = (cols >> 3);
-        if (i > 0) {
+            if (i > 0) {
                do {
                    switch(rows) {
                        case 3: C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3->qs);
@@ -1527,13 +1918,18 @@ class tinyBLAS_Q0_PPC {
    void KERNEL_4x8(int64_t ii, int64_t jj) {
        vec_t vec_A[8], vec_B[16] = {0};
        acc_t acc_0, acc_1;
-        std::array<int, 4> comparray;
+        std::array<int, 4> comparray {};
        vector float fin_res[8] = {0};
        vector float vs[8] = {0};
+        bool isAblock_q4 = std::is_same_v<TA, block_q4_0>;
        for (int l = 0; l < k; l++) {
            __builtin_mma_xxsetaccz(&acc_0);
            __builtin_mma_xxsetaccz(&acc_1);
-            packNormal<int8_t, vector signed char>((A+(ii*lda)+l), lda, 4, 8, (int8_t*)vec_A, false);
+            if (std::is_same_v<TA, block_q4_0>) {
+               packNormalInt4<int8_t, vector signed char, 4>((A+(ii*lda)+l), lda, 4, 4, (int8_t*)vec_A, comparray);
+            } else {
+               packNormal<int8_t, vector signed char>((const TB*)(A+(ii*lda)+l), lda, 4, 8, (int8_t*)vec_A, false);
+            }
            packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B, true);
            for(int x = 0; x < 8; x++) {
                __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
@@ -1545,15 +1941,17 @@ class tinyBLAS_Q0_PPC {
                    *((float*)&vs[I+4]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J+4)*ldb)+l)->d));
                }
            }
-            auto aoffset = A+(ii*lda)+l;
-            for (int i = 0; i < 4; i++) {
-                comparray[i] = 0;
-                int ca = 0;
-                const int8_t *at = aoffset->qs;
-                for (int j = 0; j < 32; j++)
-                    ca += (int)*at++;
-                comparray[i] = ca;
-                aoffset += lda;
+            if (!isAblock_q4) {
+                auto aoffset = A+(ii*lda)+l;
+                for (int i = 0; i < 4; i++) {
+                    comparray[i] = 0;
+                    int ca = 0;
+                    auto *at = aoffset->qs;
+                    for (int j = 0; j < 32; j++)
+                        ca += (int)*at++;
+                    comparray[i] = ca;
+                    aoffset += lda;
+                }
            }
            compute<4>(&acc_0, 0, 0, comparray, vs, fin_res);
            compute<4>(&acc_1, 0, 4, comparray, vs, fin_res);
@@ -1565,13 +1963,18 @@ class tinyBLAS_Q0_PPC {
    void KERNEL_8x4(int64_t ii, int64_t jj) {
        vec_t vec_A[16], vec_B[8] = {0};
        acc_t acc_0, acc_1;
-        std::array<int, 8> comparray;
+        std::array<int, 8> comparray {};
        vector float fin_res[8] = {0};
        vector float vs[8] = {0};
+        bool isAblock_q4 = std::is_same_v<TA, block_q4_0>;
        for (int l = 0; l < k; l++) {
            __builtin_mma_xxsetaccz(&acc_0);
            __builtin_mma_xxsetaccz(&acc_1);
-            packNormal<int8_t, vector signed char>((A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false);
+            if (std::is_same_v<TA, block_q4_0>) {
+               packNormalInt4<int8_t, vector signed char, 8>((A+(ii*lda)+l), lda, 8, 4, (int8_t*)vec_A, comparray);
+            } else {
+               packNormal<int8_t, vector signed char>((const TB*)(A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false);
+            }
            packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, 4, 8, (uint8_t*)vec_B, true);
            for(int x = 0; x < 8; x++) {
                __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
@@ -1582,15 +1985,17 @@ class tinyBLAS_Q0_PPC {
                    *((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d));
                }
            }
-            auto aoffset = A+(ii*lda)+l;
-            for (int i = 0; i < 8; i++) {
-                comparray[i] = 0;
-                int ca = 0;
-                const int8_t *at = aoffset->qs;
-                for (int j = 0; j < 32; j++)
-                    ca += (int)*at++;
-                comparray[i] = ca;
-                aoffset += lda;
+            if (!isAblock_q4) {
+                auto aoffset = A+(ii*lda)+l;
+                for (int i = 0; i < 8; i++) {
+                    comparray[i] = 0;
+                    int ca = 0;
+                    auto *at = aoffset->qs;
+                    for (int j = 0; j < 32; j++)
+                        ca += (int)*at++;
+                    comparray[i] = ca;
+                    aoffset += lda;
+                }
            }
            compute<8>(&acc_0, 0, 0, comparray, vs, fin_res);
            compute<8>(&acc_1, 4, 4, comparray, vs, fin_res);
@@ -1602,15 +2007,20 @@ class tinyBLAS_Q0_PPC {
    void KERNEL_8x8(int64_t ii, int64_t jj) {
        vec_t vec_A[16], vec_B[16] = {0};
        acc_t acc_0, acc_1, acc_2, acc_3;
-        std::array<int, 8> comparray;
+        std::array<int, 8> comparray {};
        vector float fin_res[16] = {0};
        vector float vs[16] = {0};
+        bool isAblock_q4 = std::is_same_v<TA, block_q4_0>;
        for (int l = 0; l < k; l++) {
            __builtin_mma_xxsetaccz(&acc_0);
            __builtin_mma_xxsetaccz(&acc_1);
            __builtin_mma_xxsetaccz(&acc_2);
            __builtin_mma_xxsetaccz(&acc_3);
-            packNormal<int8_t, vector signed char>((A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false);
+            if (std::is_same_v<TA, block_q4_0>) {
+               packNormalInt4<int8_t, vector signed char, 8>((A+(ii*lda)+l), lda, 8, 4, (int8_t*)vec_A, comparray);
+            } else {
+               packNormal<int8_t, vector signed char>((const TB*)(A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false);
+            }
            packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B, true);
            for(int x = 0; x < 8; x++) {
                __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
@@ -1624,15 +2034,17 @@ class tinyBLAS_Q0_PPC {
                    *((float*)&vs[I+8]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J+4)*ldb)+l)->d));
                }
            }
-            auto aoffset = A+(ii*lda)+l;
-            for (int i = 0; i < 8; i++) {
-                comparray[i] = 0;
-                int ca = 0;
-                const int8_t *at = aoffset->qs;
-                for (int j = 0; j < 32; j++)
-                    ca += (int)*at++;
-                comparray[i] = ca;
-                aoffset += lda;
+            if (!isAblock_q4) {
+                auto aoffset = A+(ii*lda)+l;
+                for (int i = 0; i < 8; i++) {
+                    comparray[i] = 0;
+                    int ca = 0;
+                    auto *at = aoffset->qs;
+                    for (int j = 0; j < 32; j++)
+                        ca += (int)*at++;
+                    comparray[i] = ca;
+                    aoffset += lda;
+                }
            }
            compute<8>(&acc_0, 0, 0, comparray, vs, fin_res);
            compute<8>(&acc_1, 4, 4, comparray, vs, fin_res);
@@ -1653,16 +2065,17 @@ class tinyBLAS_Q0_PPC {
        int64_t duty = (tiles + nth - 1) / nth;
        int64_t start = duty * ith;
        int64_t end = start + duty;
-        vec_t vec_A[8], vec_B[8] = {0};
+        vec_t vec_A[8] = {0}, vec_B[8] = {0};
        vector signed int vec_C[4];
        acc_t acc_0;
+        bool isAblock_q4 = std::is_same_v<TA, block_q4_0>;

        if (end > tiles)
            end = tiles;
        for (int64_t job = start; job < end; ++job) {
            int64_t ii = m0 + job / xtiles * RM;
            int64_t jj = n0 + job % xtiles * RN;
-            std::array<int, RM> comparray;
+            std::array<int, 4> comparray{};
            vector float res[4] = {0};
            vector float fin_res[4] = {0};
            vector float vs[4] = {0};
@@ -1673,7 +2086,11 @@ class tinyBLAS_Q0_PPC {
                __builtin_prefetch((A+(ii*lda)+(l+1))->qs, 0, 1); // prefetch one loop ahead
                __builtin_prefetch((B+(jj*ldb)+(l+1))->qs, 0, 1); // prefetch one loop ahead
                __builtin_mma_xxsetaccz(&acc_0);
-                packNormal<int8_t, vector signed char>((A+(ii*lda)+l), lda, RM, 8, (int8_t*)vec_A, false);
+                if (isAblock_q4) {
+                   packNormalInt4<int8_t, vector signed char, 4>((A+(ii*lda)+l), lda, RM, 4, (int8_t*)vec_A, comparray);
+                } else {
+                   packNormal<int8_t, vector signed char>((const TB*)(A+(ii*lda)+l), lda, RM, 8, (int8_t*)vec_A, false);
+                }
                packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, RN, 8, (uint8_t*)vec_B, true);
                for(int x = 0; x < 8; x+=4) {
                    __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
@@ -1687,17 +2104,18 @@ class tinyBLAS_Q0_PPC {
                    }
                }
                __builtin_mma_disassemble_acc(vec_C, &acc_0);
-                auto aoffset = A+(ii*lda)+l;
-                for (int i = 0; i < RM; i++) {
-                    comparray[i] = 0;
-                    int ca = 0;
-                    const int8_t *at = aoffset->qs;
-                    for (int j = 0; j < 32; j++)
-                        ca += (int)*at++;
-                    comparray[i] = ca;
-                    aoffset += lda;
+                if (!isAblock_q4) {
+                    auto aoffset = A+(ii*lda)+l;
+                    for (int i = 0; i < RM; i++) {
+                        comparray[i] = 0;
+                        int ca = 0;
+                        auto *at = aoffset->qs;
+                        for (int j = 0; j < 32; j++)
+                            ca += (int)*at++;
+                        comparray[i] = ca;
+                        aoffset += lda;
+                    }
                }
-
                for (int i = 0; i < RM; i++) {
                    CA[i] = vec_splats((float)(((double)comparray[i]) * -128.0));
                    res[i] = vec_add(vec_ctf(vec_C[i], 0), CA[i]);
@@ -2013,6 +2431,7 @@ class tinyBLAS_PPC {
            }
        }
    }
+
    void KERNEL_4x4(int64_t ii, int64_t jj) {
        vec_t vec_A[4], vec_B[4], vec_C[4];
        acc_t acc_0;
@@ -2259,15 +2678,27 @@ class tinyBLAS_PPC {
            vec_t vec_C[4];
            acc_t acc_0;
            __builtin_mma_xxsetaccz(&acc_0);
-            vec_t vec_A[4], vec_B[4];
+            vec_t vec_A[4] {0}, vec_B[4] = {0};
            for (int l=0; l<k; l+=4) {
-                if (RN >= 4 && RM == 1) {
+                /* 'GEMV Forwarding' concept is used in first two conditional loops.
+                 * when one of the matrix has a single row/column, the elements are
+                 * broadcasted, instead of using packing routine to prepack the
+                 * matrix elements.
+                 */
+                if (RM == 1) {
                    TA* a = const_cast<TA*>(A+(ii)*lda+l);
-                    packTranspose<vector float>(B+(jj*ldb)+l, ldb, 4, 4, (TA*)vec_B);
+                    packTranspose<vector float>(B+(jj*ldb)+l, ldb, RN, 4, (TA*)vec_B);
                    vec_A[0] = (vec_t)vec_xl(0,a);
                    vec_A[1] = (vec_t)vec_splats(*((TA*)&vec_A+1));
                    vec_A[2] = (vec_t)vec_splats(*((TA*)&vec_A+2));
                    vec_A[3] = (vec_t)vec_splats(*((TA*)&vec_A+3));
+                } else if (RN == 1) {
+                    packTranspose<vector float>(A+(ii*lda)+l, lda, RM, 4, (TA*)vec_A);
+                    TB* b = const_cast<TB*>(B+(jj)*ldb+l);
+                    vec_B[0] = (vec_t)vec_xl(0,b);
+                    vec_B[1] = (vec_t)vec_splats(*((TB*)&vec_B+1));
+                    vec_B[2] = (vec_t)vec_splats(*((TB*)&vec_B+2));
+                    vec_B[3] = (vec_t)vec_splats(*((TB*)&vec_B+3));
                } else {
                    packTranspose<vector float>(A+(ii*lda)+l, lda, RM, 4, (TA*)vec_A);
                    packTranspose<vector float>(B+(jj*ldb)+l, ldb, RN, 4, (TA*)vec_B);
@@ -2371,8 +2802,10 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
    assert(params->ith < params->nth);

    // only enable sgemm for prompt processing
+#if !defined(__MMA__)
    if (n < 2)
        return false;
+#endif

    if (Ctype != GGML_TYPE_F32)
        return false;
@@ -2503,8 +2936,8 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
            params->ith, params->nth};
        tb.matmul(m, n);
        return true;
-
 #elif defined(__MMA__)
+    //TO-DO: Remove this condition once gemv forwarding is enabled.
        if (n < 8 && n != 4)
           return false;
        if (m < 8 && m != 4)
@@ -2516,7 +2949,6 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
            params->ith, params->nth};
        tb.matmul(m, n);
        return true;
-
 #else
        return false;
 #endif
@@ -2541,6 +2973,19 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
            params->ith, params->nth};
        tb.matmul(m, n);
        return true;
+#elif defined(__MMA__)
+    //TO-DO: Remove this condition once gemv forwarding is enabled.
+        if (n < 8 && n != 4)
+           return false;
+        if (m < 8 && m != 4)
+           return false;
+        tinyBLAS_Q0_PPC<block_q4_0, block_q8_0, float> tb{
+            k, (const block_q4_0 *)A, lda,
+            (const block_q8_0 *)B, ldb,
+            (float *)C, ldc,
+            params->ith, params->nth};
+        tb.matmul(m, n);
+        return true;
 #else
        return false;
 #endif
@@ -0,0 +1,186 @@
+#include "unary-ops.h"
+
+static inline float op_abs(float x) {
+    return fabsf(x);
+}
+
+static inline float op_sgn(float x) {
+    return (x > 0.f) ? 1.f : ((x < 0.f) ? -1.f : 0.f);
+}
+
+static inline float op_neg(float x) {
+    return -x;
+}
+
+static inline float op_step(float x) {
+    return (x > 0.f) ? 1.f : 0.f;
+}
+
+static inline float op_tanh(float x) {
+    return tanhf(x);
+}
+
+static inline float op_elu(float x) {
+    return (x > 0.f) ? x : expm1f(x);
+}
+
+static inline float op_relu(float x) {
+    return (x > 0.f) ? x : 0.f;
+}
+
+static inline float op_sigmoid(float x) {
+    return 1.f / (1.f + expf(-x));
+}
+
+static inline float op_hardsigmoid(float x) {
+    return fminf(1.0f, fmaxf(0.0f, (x + 3.0f) / 6.0f));
+}
+
+static inline float op_exp(float x) {
+    return expf(x);
+}
+
+static inline float op_hardswish(float x) {
+    return x * fminf(1.0f, fmaxf(0.0f, (x + 3.0f) / 6.0f));
+}
+
+static inline float op_sqr(float x) {
+    return x * x;
+}
+
+static inline float op_sqrt(float x) {
+    return sqrtf(x);
+}
+
+static inline float op_sin(float x) {
+    return sinf(x);
+}
+
+static inline float op_cos(float x) {
+    return cosf(x);
+}
+
+static inline float op_log(float x) {
+    return logf(x);
+}
+
+template <float (*op)(float), typename src0_t, typename dst_t>
+static inline void vec_unary_op(int64_t n, dst_t * y, const src0_t * x) {
+    constexpr auto src0_to_f32 = type_conversion_table<src0_t>::to_f32;
+    constexpr auto f32_to_dst  = type_conversion_table<dst_t >::from_f32;
+
+    for (int i = 0; i < n; i++) {
+        y[i] = f32_to_dst(op(src0_to_f32(x[i])));
+    }
+}
+
+template <float (*op)(float), typename src0_t, typename dst_t>
+static void apply_unary_op(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0) && ggml_is_contiguous_1(dst) && ggml_are_same_shape(src0, dst));
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    GGML_ASSERT( nb0 == sizeof(dst_t));
+    GGML_ASSERT(nb00 == sizeof(src0_t));
+
+    const auto [ir0, ir1] = get_thread_range(params, src0);
+
+    for (int64_t ir = ir0; ir < ir1; ++ir) {
+        const int64_t i03 = ir/(ne02*ne01);
+        const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+        const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
+
+        dst_t        * dst_ptr  = (dst_t  *)       ((char *)       dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
+        const src0_t * src0_ptr = (const src0_t *) ((const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
+
+        vec_unary_op<op>(ne0, dst_ptr, src0_ptr);
+    }
+}
+
+// TODO: Use the 'traits' lookup table (for type conversion fns), instead of a mass of 'if' conditions with long templates
+template <float (*op)(float)>
+static void unary_op(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    /*  */ if (src0->type == GGML_TYPE_F32  && dst->type == GGML_TYPE_F32) { // all f32
+        apply_unary_op<op, float, float>(params, dst);
+    } else if (src0->type == GGML_TYPE_F16  && dst->type == GGML_TYPE_F16) { // all f16
+        apply_unary_op<op, ggml_fp16_t, ggml_fp16_t>(params, dst);
+    } else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_BF16) { // all bf16
+        apply_unary_op<op, ggml_bf16_t, ggml_bf16_t>(params, dst);
+    } else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_F32) {
+        apply_unary_op<op, ggml_bf16_t, float>(params, dst);
+    } else if (src0->type == GGML_TYPE_F16  && dst->type == GGML_TYPE_F32) {
+        apply_unary_op<op, ggml_fp16_t, float>(params, dst);
+    } else {
+        fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s\n", __func__,
+            ggml_type_name(dst->type), ggml_type_name(src0->type));
+        GGML_ABORT("fatal error");
+    }
+}
+
+void ggml_compute_forward_abs(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_abs>(params, dst);
+}
+
+void ggml_compute_forward_sgn(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_sgn>(params, dst);
+}
+
+void ggml_compute_forward_neg(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_neg>(params, dst);
+}
+
+void ggml_compute_forward_step(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_step>(params, dst);
+}
+
+void ggml_compute_forward_tanh(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_tanh>(params, dst);
+}
+
+void ggml_compute_forward_elu(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_elu>(params, dst);
+}
+
+void ggml_compute_forward_relu(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_relu>(params, dst);
+}
+
+void ggml_compute_forward_sigmoid(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_sigmoid>(params, dst);
+}
+
+void ggml_compute_forward_hardsigmoid(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_hardsigmoid>(params, dst);
+}
+
+void ggml_compute_forward_exp(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_exp>(params, dst);
+}
+
+void ggml_compute_forward_hardswish(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_hardswish>(params, dst);
+}
+
+void ggml_compute_forward_sqr(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_sqr>(params, dst);
+}
+
+void ggml_compute_forward_sqrt(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_sqrt>(params, dst);
+}
+
+void ggml_compute_forward_sin(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_sin>(params, dst);
+}
+
+void ggml_compute_forward_cos(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_cos>(params, dst);
+}
+
+void ggml_compute_forward_log(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_log>(params, dst);
+}
@@ -0,0 +1,28 @@
+#pragma once
+
+#include "common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void ggml_compute_forward_abs(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_sgn(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_neg(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_step(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_tanh(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_elu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_relu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_sigmoid(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_hardsigmoid(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_exp(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_hardswish(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_sqr(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_sqrt(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_sin(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_cos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_log(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+
+#ifdef __cplusplus
+}
+#endif
@@ -41,15 +41,18 @@
 #define CUDART_HMAX   11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed)
 #define CUDART_HMASK  12000 // CUDA 12.0, min. ver. for half2 -> uint mask comparisons

-#define GGML_CUDA_CC_PASCAL       600
-#define GGML_CUDA_CC_DP4A         610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
-#define GGML_CUDA_CC_VOLTA        700
-#define GGML_CUDA_CC_TURING       750
-#define GGML_CUDA_CC_AMPERE       800
-#define GGML_CUDA_CC_ADA_LOVELACE 890
-#define GGML_CUDA_CC_OFFSET_AMD   0x1000000
+#define GGML_CUDA_CC_PASCAL          600
+#define GGML_CUDA_CC_DP4A            610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
+#define GGML_CUDA_CC_VOLTA           700
+#define GGML_CUDA_CC_TURING          750
+#define GGML_CUDA_CC_AMPERE          800
+#define GGML_CUDA_CC_ADA_LOVELACE    890
+#define GGML_CUDA_CC_OFFSET_AMD      0x1000000
+#define GGML_CUDA_CC_OFFSET_MTHREADS 0x0100000
+#define GGML_CUDA_CC_IS_NVIDIA(cc)   (cc < GGML_CUDA_CC_OFFSET_MTHREADS)

-// GCN/CNDA, wave size is 64
+// AMD
+// GCN/CDNA, wave size is 64
 #define GGML_CUDA_CC_GCN4       (GGML_CUDA_CC_OFFSET_AMD + 0x803)  // Tonga, Fiji, Polaris, minimum for fast fp16
 #define GGML_CUDA_CC_VEGA       (GGML_CUDA_CC_OFFSET_AMD + 0x900)  // Vega56/64, minimum for fp16 dual issue
 #define GGML_CUDA_CC_VEGA20     (GGML_CUDA_CC_OFFSET_AMD + 0x906)  // MI50/Radeon VII, minimum for dp4a
@@ -57,21 +60,32 @@
 #define GGML_CUDA_CC_CDNA2      (GGML_CUDA_CC_OFFSET_AMD + 0x910)  // MI210, minimum acc register renameing
 #define GGML_CUDA_CC_CDNA3      (GGML_CUDA_CC_OFFSET_AMD + 0x942)  // MI300

-// RNDA removes MFMA, dp4a, xnack, acc registers, wave size is 32
+// RDNA removes MFMA, dp4a, xnack, acc registers, wave size is 32
 #define GGML_CUDA_CC_RDNA1      (GGML_CUDA_CC_OFFSET_AMD + 0x1010) // RX 5000
 #define GGML_CUDA_CC_RDNA2      (GGML_CUDA_CC_OFFSET_AMD + 0x1030) // RX 6000, minimum for dp4a
 #define GGML_CUDA_CC_RDNA3      (GGML_CUDA_CC_OFFSET_AMD + 0x1100) // RX 7000, minimum for WMMA
+#define GGML_CUDA_CC_RDNA4      (GGML_CUDA_CC_OFFSET_AMD + 0x1200) // RX 9000

 #define GGML_CUDA_CC_IS_AMD(cc)   (cc >= GGML_CUDA_CC_OFFSET_AMD)
 #define GGML_CUDA_CC_IS_RDNA(cc)  (cc >= GGML_CUDA_CC_RDNA1)
 #define GGML_CUDA_CC_IS_RDNA1(cc) (cc >= GGML_CUDA_CC_RDNA1 && cc < GGML_CUDA_CC_RDNA2)
 #define GGML_CUDA_CC_IS_RDNA2(cc) (cc >= GGML_CUDA_CC_RDNA2 && cc < GGML_CUDA_CC_RDNA3)
-#define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3)
+#define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3 && cc < GGML_CUDA_CC_RDNA4)
+#define GGML_CUDA_CC_IS_RDNA4(cc) (cc >= GGML_CUDA_CC_RDNA4)
 #define GGML_CUDA_CC_IS_GCN(cc)   (cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA)
 #define GGML_CUDA_CC_IS_CDNA(cc)  (cc >= GGML_CUDA_CC_CDNA && cc < GGML_CUDA_CC_RDNA1)

-#define GGML_CUDA_CC_QY1        210
-#define GGML_CUDA_CC_QY2        220
+// Moore Threads
+#define GGML_CUDA_MUSA_ARCH_IS_QY1 (__MUSA_ARCH__ <= 210)
+
+#define GGML_CUDA_CC_QY1  (GGML_MUSA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
+#define GGML_CUDA_CC_QY2  (GGML_MUSA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
+#define GGML_CUDA_CC_NG   (GGML_MUSA_CC_OFFSET_MTHREADS + 0x310) // TBD
+
+#define GGML_CUDA_CC_IS_MTHREADS(cc) (cc >= GGML_CUDA_CC_OFFSET_MTHREADS && cc < GGML_CUDA_CC_OFFSET_AMD)
+#define GGML_CUDA_CC_IS_QY1(cc)      (cc >= GGML_CUDA_CC_QY1 && cc < GGML_CUDA_CC_QY2)
+#define GGML_CUDA_CC_IS_QY2(cc)      (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_NEXT)
+#define GGML_CUDA_CC_IS_NG(cc)       (cc >= GGML_CUDA_CC_NG)

 #ifdef __CUDA_ARCH_LIST__
 constexpr bool ggml_cuda_has_arch_impl(int) {
@@ -197,9 +211,9 @@ typedef float2 dfloat2;
 #define FP16_MMA_AVAILABLE
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA

-#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3))
+#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4))
 #define FP16_MMA_AVAILABLE
-#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3))
+#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4))

 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
 #define NEW_MMA_AVAILABLE
@@ -209,21 +223,21 @@ typedef float2 dfloat2;
 #define CP_ASYNC_AVAILABLE
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE

-#if !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1)
+#if !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && GGML_CUDA_MUSA_ARCH_IS_QY1)
 #define FLASH_ATTN_AVAILABLE
-#endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1)
+#endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && GGML_CUDA_MUSA_ARCH_IS_QY1)

 static bool fp16_available(const int cc) {
    return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL;
 }

 static bool fast_fp16_available(const int cc) {
-    return fp16_available(cc) && cc != 610;
+    return (GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && cc != 610) || GGML_CUDA_CC_IS_AMD(cc);
 }

 // To be used for feature selection of external libraries, e.g. cuBLAS.
 static bool fast_fp16_hardware_available(const int cc) {
-    return cc >= GGML_CUDA_CC_PASCAL && cc != 610;
+    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_PASCAL && cc != 610) || GGML_CUDA_CC_IS_AMD(cc);
 }

 // Any FP16 tensor core instructions are available for ggml code.
@@ -231,20 +245,20 @@ static bool fp16_mma_available(const int cc) {
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN)
    return false;
 #else
-    return cc < GGML_CUDA_CC_OFFSET_AMD && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA ||
-        GGML_CUDA_CC_IS_CDNA(cc) || cc >= GGML_CUDA_CC_RDNA3;
+    return (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) ||
+        GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc);
 #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN)
 }

 // To be used for feature selection of external libraries, e.g. cuBLAS.
 static bool fp16_mma_hardware_available(const int cc) {
-    return cc < GGML_CUDA_CC_OFFSET_AMD && cc >= GGML_CUDA_CC_VOLTA ||
-        GGML_CUDA_CC_IS_CDNA(cc) || cc >= GGML_CUDA_CC_RDNA3;
+    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) ||
+        GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc);
 }

 // Volta technically had FP16 tensor cores but they work very differently compared to Turing and later.
 static bool new_mma_available(const int cc) {
-    return cc < GGML_CUDA_CC_OFFSET_AMD && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_TURING;
+    return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_TURING;
 }

 static bool cp_async_available(const int cc) {
@@ -274,6 +288,10 @@ static __device__ void no_device_code(
    __trap();

    GGML_UNUSED(no_device_code); // suppress unused function warning
+
+#if defined(GGML_USE_MUSA)
+    __builtin_unreachable();
+#endif // defined(GGML_USE_MUSA)
 }

 #ifdef __CUDA_ARCH__
@@ -397,7 +415,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 #if defined(CDNA) || defined(RDNA2) || defined(__gfx906__)
    c = __builtin_amdgcn_sdot4(a, b, c, false);
-#elif defined(RDNA3)
+#elif defined(RDNA3) || defined(RDNA4)
    c = __builtin_amdgcn_sudot4( true, a, true, b, c, false);
 #elif defined(RDNA1) || defined(__gfx900__)
    int tmp1;
@@ -38,7 +38,7 @@ static __global__ void concat_f32_dim1(const float * x, const float * y, float *
        blockIdx.y * ne0 +
        blockIdx.z * ne0 * gridDim.y;

-    if (blockIdx.y < ne01) { // src0
+    if (blockIdx.y < (unsigned)ne01) { // src0
        int offset_src =
            nidx +
            blockIdx.y * ne0 +
@@ -64,7 +64,7 @@ static __global__ void concat_f32_dim2(const float * x, const float * y, float *
        blockIdx.y * ne0 +
        blockIdx.z * ne0 * gridDim.y;

-    if (blockIdx.z < ne02) { // src0
+    if (blockIdx.z < (unsigned)ne02) { // src0
        int offset_src =
            nidx +
            blockIdx.y * ne0 +
@@ -34,6 +34,10 @@ static  __global__ void conv_transpose_1d_kernel(
        }
    }
    dst[global_index] = accumulator;
+    GGML_UNUSED(p0); GGML_UNUSED(d0); GGML_UNUSED(src0_ne3);
+    GGML_UNUSED(src1_ne3); GGML_UNUSED(dst_ne3);
+    GGML_UNUSED(src1_ne1); GGML_UNUSED(dst_ne1);
+    GGML_UNUSED(src1_ne2); GGML_UNUSED(dst_ne2);
 }

 static void conv_transpose_1d_f32_f32_cuda(
@@ -75,8 +79,6 @@ void ggml_cuda_op_conv_transpose_1d(ggml_backend_cuda_context & ctx, ggml_tensor
    const int p0 = 0;//opts[3];
    const int d0 = 1;//opts[4];

-    const int64_t kernel_size = ggml_nelements(src0);
-    const int64_t input_size = ggml_nelements(src1);
    const int64_t output_size = ggml_nelements(dst);

    conv_transpose_1d_f32_f32_cuda(s0, p0, d0, output_size,
@@ -577,7 +577,7 @@ static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __res
        return;
    }

-    const src_t * x = (src_t *) vx;
+    const src_t * x = (const src_t *) vx;

    y[i] = x[i];
 }
@@ -315,14 +315,14 @@ static __device__ __forceinline__ void quantize_q8_1_to_shared(

    float vals[sizeof(int)] = {0.0f};
 #pragma unroll
-    for (int l = 0; l < sizeof(int); ++l) {
+    for (int l = 0; l < int(sizeof(int)); ++l) {
        vals[l] = scale * x[4*threadIdx.x + l];
    }

    float amax = fabsf(vals[0]);
    float sum  = vals[0];
 #pragma unroll
-    for (int l = 1; l < sizeof(int); ++l) {
+    for (int l = 1; l < int(sizeof(int)); ++l) {
        amax = fmaxf(amax, fabsf(vals[l]));
        sum += vals[l];
    }
@@ -338,7 +338,7 @@ static __device__ __forceinline__ void quantize_q8_1_to_shared(

    if (d != 0.0f) {
 #pragma unroll
-        for (int l = 0; l < sizeof(int); ++l) {
+        for (int l = 0; l < int(sizeof(int)); ++l) {
            q8[l] = roundf(vals[l] / d);
        }
    }
@@ -606,50 +606,50 @@ static __global__ void flash_attn_stream_k_fixup(
    *dst = dst_val / rowsum;
 }

-template<int D, int parallel_blocks> // D == head size
+template<int D> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_combine_results(
        const float  * __restrict__ VKQ_parts,
        const float2 * __restrict__ VKQ_meta,
-        float * __restrict__ dst) {
-    VKQ_parts += parallel_blocks*D * gridDim.y*blockIdx.x;
-    VKQ_meta  += parallel_blocks   * gridDim.y*blockIdx.x;
-    dst       +=                 D * gridDim.y*blockIdx.x;
+        float * __restrict__ dst,
+        const int parallel_blocks) {
+    VKQ_parts += parallel_blocks*D * gridDim.z*blockIdx.x;
+    VKQ_meta  += parallel_blocks   * gridDim.z*blockIdx.x;
+    dst       +=                 D * gridDim.z*blockIdx.x;

    const int tid = threadIdx.x;
    __builtin_assume(tid < D);

-    __shared__ float2 meta[parallel_blocks];
+    extern __shared__ float2 meta[];
    if (tid < 2*parallel_blocks) {
-        ((float *) meta)[threadIdx.x] = ((const float *)VKQ_meta) [blockIdx.y*(2*parallel_blocks) + tid];
+        ((float *) meta)[threadIdx.x] = ((const float *)VKQ_meta) [blockIdx.z*(2*parallel_blocks) + tid];
    }

    __syncthreads();

    float kqmax = meta[0].x;
-#pragma unroll
    for (int l = 1; l < parallel_blocks; ++l) {
        kqmax = max(kqmax, meta[l].x);
    }

    float VKQ_numerator   = 0.0f;
    float VKQ_denominator = 0.0f;
-#pragma unroll
    for (int l = 0; l < parallel_blocks; ++l) {
        const float diff = meta[l].x - kqmax;
-        const float KQ_max_scale = expf(diff);
+        float KQ_max_scale = expf(diff);
        const uint32_t ftz_mask = 0xFFFFFFFF * (diff > SOFTMAX_FTZ_THRESHOLD);
        *((uint32_t *) &KQ_max_scale) &= ftz_mask;

-        VKQ_numerator   += KQ_max_scale * VKQ_parts[l*gridDim.y*D + blockIdx.y*D + tid];
+        VKQ_numerator   += KQ_max_scale * VKQ_parts[l*gridDim.z*D + blockIdx.z*D + tid];
        VKQ_denominator += KQ_max_scale * meta[l].y;
    }

-    dst[blockIdx.y*D + tid] = VKQ_numerator / VKQ_denominator;
+    dst[blockIdx.z*D + tid] = VKQ_numerator / VKQ_denominator;
 }

+[[noreturn]]
 static void on_no_fattn_vec_case(const int D) {
    if (D == 64) {
        fprintf(stderr, "Unsupported KV type combination for head_size 64.\n");
@@ -671,12 +671,10 @@ static void on_no_fattn_vec_case(const int D) {
    }
 }

-// parallel_blocks == 0 is stream-k decomposition
-template <int D, int ncols1, int ncols2, int parallel_blocks, int KQ_stride>
+template <int D, int ncols1, int ncols2, int KQ_stride>
 void launch_fattn(
-    ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel,
-    const int nwarps, const size_t nbytes_shared, const bool need_f16_K, const bool need_f16_V,
-    const int warp_size = WARP_SIZE
+    ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel, const int nwarps, const size_t nbytes_shared,
+    const int KQ_row_granularity, const bool need_f16_K, const bool need_f16_V, const bool stream_k, const int warp_size = WARP_SIZE
 ) {
    constexpr int ncols = ncols1 * ncols2;

@@ -748,12 +746,14 @@ void launch_fattn(
        nb23 = nb23*bs*sizeof(half)/ts;
    }

+    int parallel_blocks = 1;
+
    const int ntiles_x = ((Q->ne[1] + ncols1 - 1) / ncols1);
    const int ntiles_total = ntiles_x * (Q->ne[2] / ncols2) * Q->ne[3];

    const dim3 block_dim(warp_size, nwarps, 1);
    dim3 blocks_num;
-    if (parallel_blocks == 0) {
+    if (stream_k) {
        // For short contexts it can be faster to have the SMs work on whole tiles because this lets us skip the fixup.
        const int max_blocks = 2*nsm;
        const int tiles_nwaves = (ntiles_total + max_blocks - 1) / max_blocks;
@@ -769,9 +769,43 @@ void launch_fattn(

        dst_tmp_meta.alloc(blocks_num.x*ncols * (2*2 + D) * sizeof(float));
    } else {
-        blocks_num.x = parallel_blocks*ntiles_x;
-        blocks_num.y = Q->ne[2];
-        blocks_num.z = Q->ne[3];
+        GGML_ASSERT(K->ne[1] % KQ_row_granularity == 0);
+        const int ntiles_KQ = K->ne[1] / KQ_row_granularity; // Max. number of parallel blocks limited by tensor size.
+
+        int max_blocks_per_sm = 1; // Max. number of active blocks limited by occupancy.
+        CUDA_CHECK(cudaOccupancyMaxActiveBlocksPerMultiprocessor(&max_blocks_per_sm, fattn_kernel, block_dim.x * block_dim.y * block_dim.z, nbytes_shared));
+
+        // parallel_blocks should be at least large enough to achieve max. occupancy for a single wave:
+        parallel_blocks = std::max((nsm * max_blocks_per_sm) / ntiles_total, 1);
+
+        // parallel_blocks must not be larger than what the tensor size allows:
+        parallel_blocks = std::min(parallel_blocks, ntiles_KQ);
+
+        // If ntiles_total % blocks_per_wave != 0 then some efficiency is lost due to tail effects.
+        // Test whether parallel_blocks can be set to a higher value for better efficiency.
+        const int blocks_per_wave = nsm * max_blocks_per_sm;
+        int nwaves_best = 0;
+        int efficiency_percent_best = 0;
+        for (int parallel_blocks_test = parallel_blocks; parallel_blocks_test <= ntiles_KQ; ++parallel_blocks_test) {
+            const int nblocks_total = ntiles_total * parallel_blocks_test;
+            const int nwaves = (nblocks_total + blocks_per_wave - 1) / blocks_per_wave;
+            const int efficiency_percent = 100 * nblocks_total / (nwaves*blocks_per_wave);
+
+            // Stop trying configurations with more waves if we already have good efficiency to avoid excessive overhead.
+            if (efficiency_percent_best >= 90 && nwaves > nwaves_best) {
+                break;
+            }
+
+            if (efficiency_percent > efficiency_percent_best) {
+                nwaves_best = nwaves;
+                efficiency_percent_best = efficiency_percent;
+                parallel_blocks = parallel_blocks_test;
+            }
+        }
+
+        blocks_num.x = ntiles_x;
+        blocks_num.y = parallel_blocks;
+        blocks_num.z = Q->ne[2]*Q->ne[3];

        if (parallel_blocks > 1) {
            dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
@@ -803,7 +837,7 @@ void launch_fattn(
        K_data,
        V_data,
        mask ? ((const char *) mask->data) : nullptr,
-        (parallel_blocks) > 1 ? dst_tmp.ptr : (float *) KQV->data, dst_tmp_meta.ptr,
+        !stream_k && parallel_blocks > 1 ? dst_tmp.ptr : (float *) KQV->data, dst_tmp_meta.ptr,
        scale, max_bias, m0, m1, n_head_log2, logit_softcap,
        Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
        K->ne[0], K->ne[1], K->ne[2], K->ne[3],
@@ -815,7 +849,7 @@ void launch_fattn(
    );
    CUDA_CHECK(cudaGetLastError());

-    if constexpr (parallel_blocks == 0) {
+    if (stream_k) {
        if (ntiles_total % blocks_num.x != 0) { // Fixup is only needed if the SMs work on fractional tiles.
            const dim3 block_dim_combine(D, 1, 1);
            const dim3 blocks_num_combine = {blocks_num.x, ncols1, ncols2};
@@ -824,13 +858,14 @@ void launch_fattn(
                <<<blocks_num_combine, block_dim_combine, 0, main_stream>>>
                ((float *) KQV->data, dst_tmp_meta.ptr, Q->ne[1], Q->ne[2], K->ne[1]);
        }
-    } else if constexpr (parallel_blocks > 1) {
+    } else if (parallel_blocks > 1) {
        const dim3 block_dim_combine(D, 1, 1);
-        const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
+        const dim3 blocks_num_combine(Q->ne[1], 1, blocks_num.z);
+        const size_t nbytes_shared_combine = parallel_blocks*sizeof(float2);

-        flash_attn_combine_results<D, parallel_blocks>
-            <<<blocks_num_combine, block_dim_combine, 0, main_stream>>>
-            (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
+        flash_attn_combine_results<D>
+            <<<blocks_num_combine, block_dim_combine, nbytes_shared_combine, main_stream>>>
+            (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data, parallel_blocks);
    }
    CUDA_CHECK(cudaGetLastError());
 }
@@ -406,6 +406,15 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
 #endif // CP_ASYNC_AVAILABLE

 #else
+    GGML_UNUSED(Q_f2); GGML_UNUSED(K_h2); GGML_UNUSED(V_h2);
+    GGML_UNUSED(mask_h2); GGML_UNUSED(dstk); GGML_UNUSED(dstk_fixup);
+    GGML_UNUSED(scale); GGML_UNUSED(slope); GGML_UNUSED(logit_softcap);
+    GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(stride_KV);
+    GGML_UNUSED(stride_mask); GGML_UNUSED(jt); GGML_UNUSED(tile_K);
+    GGML_UNUSED(stride_mask); GGML_UNUSED(jt); GGML_UNUSED(tile_K);
+    GGML_UNUSED(tile_V); GGML_UNUSED(tile_mask); GGML_UNUSED(Q_B);
+    GGML_UNUSED(VKQ_C); GGML_UNUSED(KQ_max); GGML_UNUSED(KQ_rowsum);
+    GGML_UNUSED(kb0);
    NO_DEVICE_CODE;
 #endif // NEW_MMA_AVAILABLE
 }
@@ -797,6 +806,12 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
        __syncthreads();
    }
 #else
+    GGML_UNUSED(Q_f2); GGML_UNUSED(K_h2); GGML_UNUSED(V_h2);
+    GGML_UNUSED(mask_h2); GGML_UNUSED(dstk); GGML_UNUSED(dstk_fixup);
+    GGML_UNUSED(scale); GGML_UNUSED(slope); GGML_UNUSED(logit_softcap);
+    GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(stride_Q1);
+    GGML_UNUSED(stride_Q2); GGML_UNUSED(stride_KV); GGML_UNUSED(stride_mask);
+    GGML_UNUSED(jt); GGML_UNUSED(kb0_start); GGML_UNUSED(kb0_stop);
    NO_DEVICE_CODE;
 #endif // NEW_MMA_AVAILABLE
 }
@@ -931,6 +946,16 @@ static __global__ void flash_attn_ext_f16(
        (Q_f2, K_h2, V_h2, mask_h2, dstk, dst_meta, scale, slope, logit_softcap,
         ne01, ne02, stride_Q1, stride_Q2, stride_KV, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel);
 #else
+    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
+    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
+    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
+    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap); GGML_UNUSED(ne00);
+    GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03); GGML_UNUSED(ne10);
+    GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
+    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13); GGML_UNUSED(nb21);
+    GGML_UNUSED(nb22); GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
+    GGML_UNUSED(ne2); GGML_UNUSED(ne3);
    NO_DEVICE_CODE;
 #endif // defined(FLASH_ATTN_AVAILABLE) && defined(NEW_MMA_AVAILABLE)
 }
@@ -970,7 +995,8 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
        fattn_kernel = flash_attn_ext_f16<D, ncols1, ncols2, nwarps, KQ_per_iter, ntiles, use_logit_softcap>;
    }

-    launch_fattn<D, ncols1, ncols2, 0, KQ_per_iter>(ctx, dst, fattn_kernel, nwarps, nbytes_shared_total, true, true);
+    launch_fattn<D, ncols1, ncols2, KQ_per_iter>
+        (ctx, dst, fattn_kernel, nwarps, nbytes_shared_total, FATTN_KQ_STRIDE, true, true, true);
 }


@@ -984,38 +1010,38 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
    extern DECL_FATTN_MMA_F16_CASE(D, (ncols)/4, 4); \
    extern DECL_FATTN_MMA_F16_CASE(D, (ncols)/8, 8); \

-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64,   8);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80,   8);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96,   8);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112,   8);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128,   8);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256,   8);
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64,   8)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80,   8)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96,   8)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112,   8)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128,   8)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256,   8)

-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64,  16);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80,  16);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96,  16);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112,  16);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128,  16);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256,  16);
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64,  16)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80,  16)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96,  16)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112,  16)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128,  16)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256,  16)

-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64,  32);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80,  32);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96,  32);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112,  32);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128,  32);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256,  32);
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64,  32)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80,  32)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96,  32)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112,  32)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128,  32)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256,  32)

-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64,  64);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80,  64);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96,  64);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112,  64);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128,  64);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256,  64);
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64,  64)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80,  64)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96,  64)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112,  64)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128,  64)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256,  64)

 // Kernels with ncols == 128 are only 4% faster due to register pressure.
-// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64, 128);
-// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80, 128);
-// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96, 128);
-// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 128);
-// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128, 128);
-// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 128); // Needs too much shared memory.
+// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64, 128)
+// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80, 128)
+// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96, 128)
+// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 128)
+// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128, 128)
+// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 128) // Needs too much shared memory.
@@ -4,7 +4,7 @@

 #define FATTN_KQ_STRIDE_TILE_F16 64

-template<int D, int ncols, int nwarps, int parallel_blocks, bool use_logit_softcap> // D == head size
+template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(nwarps*WARP_SIZE, 1)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
@@ -58,18 +58,17 @@ static __global__ void flash_attn_tile_ext_f16(

    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.

-    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
-    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+    const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.

    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
-    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
-    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.z              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.z / gqa_ratio));
+    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
    const half   * maskh = (const half   *)  mask + ne11*ic0;

    const int stride_KV2 = nb11 / sizeof(half2);

-    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
    const half  slopeh = __float2half(slopef);

    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
@@ -105,8 +104,7 @@ static __global__ void flash_attn_tile_ext_f16(

    __syncthreads();

-    const int k_start = parallel_blocks == 1 ? 0 : ip*FATTN_KQ_STRIDE_TILE_F16;
-    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE_TILE_F16) {
+    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE_TILE_F16; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE_TILE_F16) {
        // Calculate KQ tile and keep track of new maximum KQ values:

        half kqmax_new[ncols/nwarps];
@@ -271,24 +269,36 @@ static __global__ void flash_attn_tile_ext_f16(
            const int i0 = i00 + 2*threadIdx.x;

            half2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
-            if (parallel_blocks == 1) {
+            if (gridDim.y == 1) {
                dst_val /= __half2half2(kqsum_j);
            }
-            const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
-            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 0] =  __low2float(dst_val);
-            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 1] = __high2float(dst_val);
+            const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
+            dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 0] =  __low2float(dst_val);
+            dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 1] = __high2float(dst_val);
        }

-        if (parallel_blocks != 1 && threadIdx.x == 0) {
-            dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
+        if (gridDim.y != 1 && threadIdx.x == 0) {
+            dst_meta[((ic0 + j_VKQ)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
        }
    }
 #else
-   NO_DEVICE_CODE;
+    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
+    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
+    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
+    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
+    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
+    GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
+    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+    GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
+    GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
+    GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
+    GGML_UNUSED(ne2); GGML_UNUSED(ne3);
+    NO_DEVICE_CODE;
 #endif // defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)
 }

-template <int cols_per_block, int parallel_blocks, bool use_logit_softcap>
+template <int cols_per_block, bool use_logit_softcap>
 void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const ggml_tensor * Q = dst->src[0];
    switch (Q->ne[0]) {
@@ -296,15 +306,17 @@ void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
            constexpr int    D             = 64;
            constexpr int    nwarps        = 8;
            constexpr size_t nbytes_shared = 0;
-            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, use_logit_softcap>;
+            launch_fattn<D, cols_per_block, 1, -1>
+                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F16, true, true, false);
        } break;
        case 128: {
            constexpr int    D             = 128;
            constexpr int    nwarps        = 8;
            constexpr size_t nbytes_shared = 0;
-            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, use_logit_softcap>;
+            launch_fattn<D, cols_per_block, 1, -1>
+                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F16, true, true, false);
        } break;
        default: {
            GGML_ABORT("FlashAttention without tensor cores only supports head sizes 64 and 128.");
@@ -324,37 +336,22 @@ void ggml_cuda_flash_attn_ext_tile_f16(ggml_backend_cuda_context & ctx, ggml_ten

    if (Q->ne[1] <= 16) {
        constexpr int cols_per_block = 16;
-        constexpr int parallel_blocks = 4;
        if (logit_softcap == 0.0f) {
            constexpr bool use_logit_softcap = false;
-            launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+            launch_fattn_tile_f16_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
        } else {
            constexpr bool use_logit_softcap = true;
-            launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
-        }
-        return;
-    }
-
-    if (Q->ne[1] <= 32) {
-        constexpr int cols_per_block = 32;
-        constexpr int parallel_blocks = 4;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
-        } else {
-            constexpr bool use_logit_softcap = true;
-            launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+            launch_fattn_tile_f16_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
        }
        return;
    }

    constexpr int cols_per_block = 32;
-    constexpr int parallel_blocks = 1;
    if (logit_softcap == 0.0f) {
        constexpr bool use_logit_softcap = false;
-        launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        launch_fattn_tile_f16_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
    } else {
        constexpr bool use_logit_softcap = true;
-        launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        launch_fattn_tile_f16_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
    }
 }
@@ -4,7 +4,7 @@

 #define FATTN_KQ_STRIDE_TILE_F32 32

-template<int D, int ncols, int nwarps, int parallel_blocks, bool use_logit_softcap> // D == head size
+template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(nwarps*WARP_SIZE, 1)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
@@ -58,18 +58,17 @@ static __global__ void flash_attn_tile_ext_f32(

    // In this kernel Q, K, V are matrices while i, j, k are matrix indices.

-    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
-    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+    const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.

    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
-    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
-    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.z              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.z / gqa_ratio));
+    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
    const half   * maskh = (const half   *)  mask + ne11*ic0;

    const int stride_KV2 = nb11 / sizeof(half2);

-    const float slope = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const float slope = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);

    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");

@@ -103,8 +102,7 @@ static __global__ void flash_attn_tile_ext_f32(

    __syncthreads();

-    const int k_start = parallel_blocks == 1 ? 0 : ip*FATTN_KQ_STRIDE_TILE_F32;
-    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE_TILE_F32) {
+    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE_TILE_F32; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE_TILE_F32) {
        // Calculate KQ tile and keep track of new maximum KQ values:

        float kqmax_new[ncols/nwarps];
@@ -269,25 +267,37 @@ static __global__ void flash_attn_tile_ext_f32(
            const int i0 = i00 + 2*threadIdx.x;

            float2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
-            if (parallel_blocks == 1) {
+            if (gridDim.y == 1) {
                dst_val.x /= kqsum_j;
                dst_val.y /= kqsum_j;
            }
-            const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
-            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 0] = dst_val.x;
-            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 1] = dst_val.y;
+            const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
+            dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 0] = dst_val.x;
+            dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 1] = dst_val.y;
        }

-        if (parallel_blocks != 1 && threadIdx.x == 0) {
-            dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
+        if (gridDim.y != 1 && threadIdx.x == 0) {
+            dst_meta[((ic0 + j_VKQ)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
        }
    }
 #else
+    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
+    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
+    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
+    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
+    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
+    GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
+    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+    GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
+    GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
+    GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
+    GGML_UNUSED(ne2); GGML_UNUSED(ne3);
    NO_DEVICE_CODE;
 #endif // FLASH_ATTN_AVAILABLE
 }

-template <int cols_per_block, int parallel_blocks, bool use_logit_softcap>
+template <int cols_per_block, bool use_logit_softcap>
 void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const ggml_tensor * Q = dst->src[0];
    switch (Q->ne[0]) {
@@ -295,15 +305,17 @@ void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
            constexpr int    D             = 64;
            constexpr int    nwarps        = 8;
            constexpr size_t nbytes_shared = 0;
-            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, use_logit_softcap>;
+            launch_fattn<D, cols_per_block, 1, -1>
+                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F32, true, true, false);
        } break;
        case 128: {
            constexpr int    D             = 128;
            constexpr int    nwarps        = 8;
            constexpr size_t nbytes_shared = 0;
-            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, use_logit_softcap>;
+            launch_fattn<D, cols_per_block, 1, -1>
+                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F32, true, true, false);
        } break;
        default: {
            GGML_ABORT("FlashAttention without tensor cores only supports head sizes 64 and 128.");
@@ -320,37 +332,22 @@ void ggml_cuda_flash_attn_ext_tile_f32(ggml_backend_cuda_context & ctx, ggml_ten

    if (Q->ne[1] <= 16) {
        constexpr int cols_per_block = 16;
-        constexpr int parallel_blocks = 4;
        if (logit_softcap == 0.0f) {
            constexpr bool use_logit_softcap = false;
-            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+            launch_fattn_tile_f32_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
        } else {
            constexpr bool use_logit_softcap = true;
-            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
-        }
-        return;
-    }
-
-    if (Q->ne[1] <= 32) {
-        constexpr int cols_per_block = 32;
-        constexpr int parallel_blocks = 4;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
-        } else {
-            constexpr bool use_logit_softcap = true;
-            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+            launch_fattn_tile_f32_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
        }
        return;
    }

    constexpr int cols_per_block = 32;
-    constexpr int parallel_blocks = 1;
    if (logit_softcap == 0.0f) {
        constexpr bool use_logit_softcap = false;
-        launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        launch_fattn_tile_f32_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
    } else {
        constexpr bool use_logit_softcap = true;
-        launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        launch_fattn_tile_f32_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
    }
 }
@@ -1,7 +1,7 @@
 #include "common.cuh"
 #include "fattn-common.cuh"

-template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
+template<int D, int ncols, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
@@ -55,17 +55,16 @@ static __global__ void flash_attn_vec_ext_f16(
    constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16;
    constexpr dequantize_1_f16_t dequantize_1_v = get_dequantize_1_f16(type_V);

-    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
-    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+    const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.

    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    Q += nb02* blockIdx.y              + nb01*ic0;
-    K += nb12*(blockIdx.y / gqa_ratio);
-    V += nb22*(blockIdx.y / gqa_ratio);
+    Q += nb02* blockIdx.z              + nb01*ic0;
+    K += nb12*(blockIdx.z / gqa_ratio);
+    V += nb22*(blockIdx.z / gqa_ratio);

    const half * maskh = (const half   *)  mask + ne11*ic0;

-    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
    const half  slopeh = __float2half(slopef);

    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
@@ -172,8 +171,7 @@ static __global__ void flash_attn_vec_ext_f16(

    half2 VKQ[ncols] = {{0.0f, 0.0f}};

-    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
-    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
+    for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*D) {
        // Calculate KQ tile and keep track of new maximum KQ values:

        // For unknown reasons using a half array of size 1 for kqmax_new causes a performance regression,
@@ -283,29 +281,41 @@ static __global__ void flash_attn_vec_ext_f16(
        kqsum[j_VKQ] = warp_reduce_sum((float)kqsum[j_VKQ]);

        half dst_val = (__low2half(VKQ[j_VKQ]) + __high2half(VKQ[j_VKQ]));
-        if (parallel_blocks == 1) {
+        if (gridDim.y == 1) {
            dst_val /= kqsum[j_VKQ];
        }
-        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
-        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
+        const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
+        dst[j_dst*D*gridDim.z + D*blockIdx.z + tid] = dst_val;
    }

-    if (parallel_blocks != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
-        dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]);
+    if (gridDim.y != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
+        dst_meta[((ic0 + tid)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]);
    }
 #else
-   NO_DEVICE_CODE;
+    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
+    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
+    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
+    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
+    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
+    GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
+    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+    GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
+    GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
+    GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
+    GGML_UNUSED(ne2); GGML_UNUSED(ne3);
+    NO_DEVICE_CODE;
 #endif // defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)
 }

-template <int D, int cols_per_block, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
+template <int D, int cols_per_block, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
 void ggml_cuda_flash_attn_ext_vec_f16_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    constexpr int nwarps = D/WARP_SIZE;
-    fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>;
+    fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, type_K, type_V, use_logit_softcap>;
    constexpr bool need_f16_K = D != 128;
    constexpr bool need_f16_V = D != 128 && D != 64;
    constexpr size_t nbytes_shared = 0;
-    launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, need_f16_K, need_f16_V);
+    launch_fattn<D, cols_per_block, 1, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
 }

 template <int D, ggml_type type_K, ggml_type type_V>
@@ -325,65 +335,48 @@ void ggml_cuda_flash_attn_ext_vec_f16_case(ggml_backend_cuda_context & ctx, ggml
    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));

    if (Q->ne[1] == 1) {
-        constexpr int cols_per_block  = 1;
-        constexpr int parallel_blocks = 4;
+        constexpr int cols_per_block = 1;
        if (logit_softcap == 0.0f) {
            constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
        } else {
            constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
        }
        return;
    }

    if (Q->ne[1] == 2) {
-        constexpr int cols_per_block  = 2;
-        constexpr int parallel_blocks = 4;
+        constexpr int cols_per_block = 2;
        if (logit_softcap == 0.0f) {
            constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
        } else {
            constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
        }
        return;
    }

    if (Q->ne[1] <= 4) {
-        constexpr int cols_per_block  = 4;
-        constexpr int parallel_blocks = 4;
+        constexpr int cols_per_block = 4;
        if (logit_softcap == 0.0f) {
            constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
        } else {
            constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
        }
        return;
    }

-    if (Q->ne[1] <= 8) {
-        constexpr int cols_per_block  = 8;
-        constexpr int parallel_blocks = 4;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
-        } else {
-            constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
-        }
-        return;
-    }
-
-    constexpr int cols_per_block  = 8;
-    constexpr int parallel_blocks = 1;
+    constexpr int cols_per_block = 8;
    if (logit_softcap == 0.0f) {
        constexpr bool use_logit_softcap = false;
-        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
    } else {
        constexpr bool use_logit_softcap = true;
-        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
    }
 }

@@ -1,7 +1,7 @@
 #include "common.cuh"
 #include "fattn-common.cuh"

-template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
+template<int D, int ncols, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
@@ -55,16 +55,15 @@ static __global__ void flash_attn_vec_ext_f32(
    constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16;
    constexpr dequantize_1_f32_t dequantize_1_v = get_dequantize_1_f32(type_V);

-    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
-    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+    const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.

    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    Q += nb02* blockIdx.y              + nb01*ic0;
-    K += nb12*(blockIdx.y / gqa_ratio);
-    V += nb22*(blockIdx.y / gqa_ratio); // K and V have same shape
+    Q += nb02* blockIdx.z              + nb01*ic0;
+    K += nb12*(blockIdx.z / gqa_ratio);
+    V += nb22*(blockIdx.z / gqa_ratio); // K and V have same shape
    const half * maskh = (const half   *)  mask + ne11*ic0;

-    const float slope = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const float slope = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);

    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
    constexpr int nwarps = D / WARP_SIZE;
@@ -167,8 +166,7 @@ static __global__ void flash_attn_vec_ext_f32(

    float VKQ[ncols] = {0.0f};

-    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
-    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
+    for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*D) {
        // Calculate KQ tile and keep track of new maximum KQ values:

        float kqmax_new_arr[ncols];
@@ -268,29 +266,39 @@ static __global__ void flash_attn_vec_ext_f32(
        kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);

        float dst_val = VKQ[j_VKQ];
-        if (parallel_blocks == 1) {
+        if (gridDim.y == 1) {
            dst_val /= kqsum[j_VKQ];
        }
-        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
-        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
+        const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
+        dst[j_dst*D*gridDim.z + D*blockIdx.z + tid] = dst_val;
    }

-    if (parallel_blocks != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
-        dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]);
+    if (gridDim.y != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
+        dst_meta[((ic0 + tid)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]);
    }
 #else
+    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
+    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
+    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
+    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap); GGML_UNUSED(ne00);
+    GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03); GGML_UNUSED(ne10);
+    GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
+    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13); GGML_UNUSED(nb21);
+    GGML_UNUSED(nb22); GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
+    GGML_UNUSED(ne2); GGML_UNUSED(ne3);
    NO_DEVICE_CODE;
 #endif // FLASH_ATTN_AVAILABLE
 }

-template <int D, int cols_per_block, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
+template <int D, int cols_per_block, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
 void ggml_cuda_flash_attn_ext_vec_f32_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    constexpr int nwarps = D/WARP_SIZE;
-    fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f32<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>;
+    fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f32<D, cols_per_block, type_K, type_V, use_logit_softcap>;
    constexpr bool need_f16_K = D != 128;
    constexpr bool need_f16_V = D != 128 && D != 64;
    constexpr size_t nbytes_shared = 0;
-    launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, need_f16_K, need_f16_V);
+    launch_fattn<D, cols_per_block, 1, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
 }

 template <int D, ggml_type type_K, ggml_type type_V>
@@ -307,65 +315,48 @@ void ggml_cuda_flash_attn_ext_vec_f32_case(ggml_backend_cuda_context & ctx, ggml
    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));

    if (Q->ne[1] == 1) {
-        constexpr int cols_per_block  = 1;
-        constexpr int parallel_blocks = 4;
+        constexpr int cols_per_block = 1;
        if (logit_softcap == 0.0f) {
            constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
        } else {
            constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
        }
        return;
    }

    if (Q->ne[1] == 2) {
-        constexpr int cols_per_block  = 2;
-        constexpr int parallel_blocks = 4;
+        constexpr int cols_per_block = 2;
        if (logit_softcap == 0.0f) {
            constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
        } else {
            constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
        }
        return;
    }

    if (Q->ne[1] <= 4) {
-        constexpr int cols_per_block  = 4;
-        constexpr int parallel_blocks = 4;
+        constexpr int cols_per_block = 4;
        if (logit_softcap == 0.0f) {
            constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
        } else {
            constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
        }
        return;
    }

-    if (Q->ne[1] <= 8) {
-        constexpr int cols_per_block  = 8;
-        constexpr int parallel_blocks = 4;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
-        } else {
-            constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
-        }
-        return;
-    }
-
-    constexpr int cols_per_block  = 8;
-    constexpr int parallel_blocks = 1;
+    constexpr int cols_per_block = 8;
    if (logit_softcap == 0.0f) {
        constexpr bool use_logit_softcap = false;
-        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
    } else {
        constexpr bool use_logit_softcap = true;
-        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
    }
 }

@@ -18,7 +18,7 @@ namespace wmma = rocwmma;
 #endif // FP16_MMA_AVAILABLE

 // D == head size, VKQ_stride == num VKQ rows calculated in parallel:
-template<int D, int ncols, int nwarps, int VKQ_stride, int parallel_blocks, typename KQ_acc_t, bool use_logit_softcap>
+template<int D, int ncols, int nwarps, int VKQ_stride, typename KQ_acc_t, bool use_logit_softcap>
 __launch_bounds__(nwarps*ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void flash_attn_ext_f16(
        const char * __restrict__ Q,
@@ -67,8 +67,7 @@ static __global__ void flash_attn_ext_f16(

    constexpr int warp_size = ggml_cuda_get_physical_warp_size();

-    const int ic0 = ncols*(blockIdx.x / parallel_blocks); // Index of the first Q/QKV column to work on.
-    const int ip  =        blockIdx.x % parallel_blocks;  // Index in group of blocks running for the same column in parallel.
+    const int ic0 = ncols*blockIdx.x; // Index of the first Q/QKV column to work on.

    static_assert(D <= FATTN_KQ_STRIDE, "D must be <= FATTN_KQ_STRIDE.");
    static_assert(ncols == 8 || ncols % 16 == 0, "ncols must be 8 or a multiple of 16.");
@@ -91,16 +90,16 @@ static __global__ void flash_attn_ext_f16(
    constexpr int kqar = sizeof(KQ_acc_t)/sizeof(half);

    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float * Q_f   = (const float *) (Q + nb02* blockIdx.y              + nb01*ic0);
-    const half  * K_h   = (const half  *) (K + nb12*(blockIdx.y / gqa_ratio));
-    const half  * V_h   = (const half  *) (V + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const float * Q_f   = (const float *) (Q + nb02* blockIdx.z              + nb01*ic0);
+    const half  * K_h   = (const half  *) (K + nb12*(blockIdx.z / gqa_ratio));
+    const half  * V_h   = (const half  *) (V + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
    const half  * maskh = (const half  *)  mask + (nb31/sizeof(half))* ic0;
    const half2 * mask2 = (const half2 *)  mask + (nb31/sizeof(half))*(ic0/2);

    const int stride_Q  = nb01 / sizeof(float);
    const int stride_KV = nb11 / sizeof(half);

-    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
    const half  slopeh = __float2half(slopef);
    const half2 slope2 = make_half2(slopef, slopef);

@@ -176,7 +175,7 @@ static __global__ void flash_attn_ext_f16(
    __syncthreads();

    // Iterate over ne11 == previous tokens:
-    for (int k_VKQ_0 = ip*FATTN_KQ_STRIDE; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE) {
+    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE) {
        // Calculate tile of KQ:
 #pragma unroll
        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE; i_KQ_0 += KQ_stride_tc) {
@@ -395,7 +394,7 @@ static __global__ void flash_attn_ext_f16(
        if (ic0 + j_VKQ >= ne01) {
            return;
        }
-        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+        const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;

        float KQ_rowsum_j;
        if (std::is_same<KQ_acc_t, float>::value) {
@@ -411,13 +410,13 @@ static __global__ void flash_attn_ext_f16(
                break;
            }
            float dst_val = VKQ[j_VKQ*D_padded + i];
-            if (parallel_blocks == 1) {
+            if (gridDim.y == 1) {
                dst_val /= KQ_rowsum_j;
            }
-            dst[j_dst*gridDim.y*D + blockIdx.y*D + i] = dst_val;
+            dst[j_dst*gridDim.z*D + blockIdx.z*D + i] = dst_val;
        }

-        if (parallel_blocks == 1 || threadIdx.x != 0) {
+        if (gridDim.y == 1 || threadIdx.x != 0) {
            continue;
        }

@@ -428,10 +427,20 @@ static __global__ void flash_attn_ext_f16(
            dst_meta_val.x = __low2float(KQ_max_h2[j0/nwarps]);
        }
        dst_meta_val.y = KQ_rowsum_j;
-        dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = dst_meta_val;
+        dst_meta[((ic0 + j_VKQ)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = dst_meta_val;
    }
 #else
-   NO_DEVICE_CODE;
+    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
+    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
+    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
+    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
+    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
+    GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
+    GGML_UNUSED(ne31); GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+    GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
+    GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
+    GGML_UNUSED(ne0); GGML_UNUSED(ne1); GGML_UNUSED(ne2); GGML_UNUSED(ne3);
+    NO_DEVICE_CODE;
 #endif // defined(FLASH_ATTN_AVAILABLE) && (__CUDA_ARCH__ == GGML_CUDA_CC_VOLTA || (defined(GGML_HIP_ROCWMMA_FATTN) && defined(FP16_MMA_AVAILABLE)))
 }

@@ -462,60 +471,26 @@ static_assert(get_VKQ_stride( 80, 4, 16) ==  16, "Test failed.");
 template <int D, int cols_per_block, typename KQ_acc_t>
 void ggml_cuda_flash_attn_ext_wmma_f16_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const ggml_tensor * KQV = dst;
-    const ggml_tensor * Q   = dst->src[0];

    constexpr int nwarps = 4;

    constexpr int frag_m = cols_per_block == 8 && D % 32 == 0 ? 32 : 16;
-    const int blocks_num_pb1 = ((Q->ne[1] + cols_per_block - 1) / cols_per_block)*Q->ne[2]*Q->ne[3];
-    const int nsm = ggml_cuda_info().devices[ggml_cuda_get_device()].nsm;
    const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size;

    float logit_softcap;
    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));

-    if (4*blocks_num_pb1 < 2*nsm) {
-        constexpr int parallel_blocks = 4;
-        fattn_kernel_t fattn_kernel;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            fattn_kernel = flash_attn_ext_f16<
-                D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
-        } else {
-            constexpr bool use_logit_softcap = true;
-            fattn_kernel = flash_attn_ext_f16<
-                D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
-        }
-        launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, 0, true, true, warp_size);
-        return;
-    }
-    if (2*blocks_num_pb1 < 2*nsm) {
-        constexpr int parallel_blocks = 2;
-        fattn_kernel_t fattn_kernel;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            fattn_kernel = flash_attn_ext_f16<
-                D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
-        } else {
-            constexpr bool use_logit_softcap = true;
-            fattn_kernel = flash_attn_ext_f16<
-                D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
-        }
-        launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, 0, true, true, warp_size);
-        return;
-    }
-    constexpr int parallel_blocks = 1;
    fattn_kernel_t fattn_kernel;
    if (logit_softcap == 0.0f) {
        constexpr bool use_logit_softcap = false;
        fattn_kernel = flash_attn_ext_f16<
-            D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
+            D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), KQ_acc_t, use_logit_softcap>;
    } else {
        constexpr bool use_logit_softcap = true;
        fattn_kernel = flash_attn_ext_f16<
-            D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
+            D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), KQ_acc_t, use_logit_softcap>;
    }
-    launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, 0, true, true, warp_size);
+    launch_fattn<D, cols_per_block, 1, -1>(ctx, dst, fattn_kernel, nwarps, 0, FATTN_KQ_STRIDE, true, true, false, warp_size);
 }

 void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -253,7 +253,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
    const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size;
    const enum ggml_prec prec = ggml_flash_attn_ext_get_prec(KQV);

-    if (cc >= GGML_CUDA_CC_OFFSET_AMD) {
+    if (GGML_CUDA_CC_IS_AMD(cc)) {
 #if defined(GGML_HIP_ROCWMMA_FATTN)
        if (fp16_mma_available(cc)) {
            ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
@@ -281,13 +281,13 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst

    if (!fp16_mma_available(cc)) {
        if (prec == GGML_PREC_DEFAULT) {
-            if (Q->ne[1] <= 8) {
+            if (Q->ne[1] <= 8 || Q->ne[0] == 256) {
                ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
            } else {
                ggml_cuda_flash_attn_ext_tile_f16(ctx, dst);
            }
        } else {
-            if (Q->ne[1] <= 8) {
+            if (Q->ne[1] <= 8 || Q->ne[0] == 256) {
                ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
            } else {
                ggml_cuda_flash_attn_ext_tile_f32(ctx, dst);
@@ -296,17 +296,17 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
        return;
    }

-    const int gqa_ratio = Q->ne[2] / K->ne[2];
-    const bool mma_fast_for_bs1 = fp16_mma_available(cc) && gqa_ratio % 2 == 0 &&
-        K->type == GGML_TYPE_F16 && V->type == GGML_TYPE_F16 && mask;
-    if (Q->ne[1] == 1 && Q->ne[0] % (2*warp_size) == 0 && !mma_fast_for_bs1) {
+    const bool gqa_opt_applies = ((Q->ne[2] / K->ne[2]) % 2 == 0) && mask; // The mma-based kernels have GQA-specific optimizations
+    const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16;
+    const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && cc < GGML_CUDA_CC_ADA_LOVELACE && !mma_needs_data_conversion;
+    const bool can_use_vector_kernel = (Q->ne[0] % (2*warp_size) == 0) && (prec == GGML_PREC_DEFAULT || Q->ne[0] <= 128);
+    if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1) {
        if (prec == GGML_PREC_DEFAULT) {
            ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
-            return;
-        } else if(Q->ne[0] <= 128) {
+        } else {
            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
-            return;
        }
+        return;
    }

    // The MMA implementation needs Turing or newer, use the old WMMA code for Volta:
@@ -262,9 +262,11 @@ static ggml_cuda_device_info ggml_cuda_init() {
                      id, prop.name, prop.gcnArchName, info.devices[id].cc & 0xffff,
                      device_vmm ? "yes" : "no", prop.warpSize);
 #elif defined(GGML_USE_MUSA)
-        // TODO: refine the .cc to reflect MUSA's actual CC capabilities
+        // FIXME: Ensure compatibility with varying warp sizes across different MUSA archs.
+        info.devices[id].warp_size = 32;
        info.devices[id].smpbo = prop.sharedMemPerBlockOptin;
-        info.devices[id].cc = 100*prop.major + 10*prop.minor;
+        info.devices[id].cc = GGML_CUDA_CC_OFFSET_MTHREADS + prop.major * 0x100;
+        info.devices[id].cc += prop.minor * 0x10;
        GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s\n",
                        id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
 #else
@@ -1186,11 +1188,11 @@ static void ggml_cuda_op_mul_mat_cublas(
    // ldc == nrows of the matrix that cuBLAS writes into
    int64_t ldc = id == ctx.device ? ne0 : row_diff;

-    const int compute_capability = ggml_cuda_info().devices[id].cc;
+    const int cc = ggml_cuda_info().devices[id].cc;

    const bool use_fp16 = (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT;

-    if (compute_capability >= GGML_CUDA_CC_VOLTA && use_fp16) {
+    if (((GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) || GGML_CUDA_CC_IS_AMD(cc)) && use_fp16) {
        // convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32
        ggml_cuda_pool_alloc<half> src0_as_f16(ctx.pool(id));
        if (src0->type != GGML_TYPE_F16) {
@@ -1214,7 +1216,7 @@ static void ggml_cuda_op_mul_mat_cublas(

        CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream));

-        if (GGML_CUDA_CC_IS_CDNA(compute_capability)) {
+        if (GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
            const float alpha = 1.0f;
            const float beta = 0.0f;
            CUBLAS_CHECK(
@@ -1757,7 +1759,9 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
        beta  = &beta_f32;
    }

-    if (GGML_CUDA_CC_IS_CDNA(ggml_cuda_info().devices[ctx.device].cc)) {
+    int id = ggml_cuda_get_device();
+    const int cc = ggml_cuda_info().devices[id].cc;
+    if (GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
        cu_compute_type = CUBLAS_COMPUTE_32F;
        alpha = &alpha_f32;
        beta  = &beta_f32;
@@ -1834,7 +1838,7 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
    }
 #endif

-    if (dst->op_params[0] == GGML_PREC_DEFAULT) {
+    if (dst->op_params[0] == GGML_PREC_DEFAULT && cu_data_type == CUDA_R_16F) {
        const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
        to_fp32_cuda(dst_f16.get(), dst_ddf, ne_dst, main_stream);
    }
@@ -3228,6 +3232,16 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
 #ifndef FLASH_ATTN_AVAILABLE
            return false;
 #endif // FLASH_ATTN_AVAILABLE
+            if (op->src[1]->ne[0] != op->src[2]->ne[0]) {
+                // different head sizes of K and V are not supported yet
+                return false;
+            }
+            if (op->src[0]->ne[0] == 192) {
+                return false;
+            }
+            if (op->src[0]->ne[3] != 1) {
+                return false;
+            }
            if (op->src[1]->type == GGML_TYPE_BF16 || op->src[2]->type == GGML_TYPE_BF16) {
                return false;
            }
@@ -26,6 +26,7 @@ static __device__ __forceinline__ int ggml_cuda_movmatrix(const int x) {
    asm("movmatrix.sync.aligned.m8n8.trans.b16 %0, %1;"
        : "=r"(ret) : "r"(x));
 #else
+    GGML_UNUSED(x);
    NO_DEVICE_CODE;
 #endif // defined(NEW_MMA_AVAILABLE)
    return ret;
@@ -178,6 +179,7 @@ namespace ggml_cuda_mma {
            : "l"(xs));
 #else
        load_generic(xs0, stride);
+        GGML_UNUSED(t);
 #endif // NEW_MMA_AVAILABLE
    }

@@ -27,8 +27,8 @@ void ggml_cuda_op_mul_mat_q(
    // The stream-k decomposition is only faster for recent NVIDIA GPUs.
    // Also its fixup needs to allocate a temporary buffer in the memory pool.
    // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
-    const bool use_stream_k = ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA &&
-        cc < GGML_CUDA_CC_OFFSET_AMD && src1_ncols == ne11;
+    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) &&
+        ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA && src1_ncols == ne11;
    const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst, use_stream_k};

    switch (src0->type) {
@@ -145,9 +145,9 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
    return true;
 #endif //GGML_CUDA_FORCE_MMQ

-    if (cc < GGML_CUDA_CC_OFFSET_AMD) {
+    if (GGML_CUDA_CC_IS_NVIDIA(cc)) {
        return !fp16_mma_hardware_available(cc) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
    }

-    return (!GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+    return (!GGML_CUDA_CC_IS_RDNA4(cc) && !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
 }
@@ -90,7 +90,7 @@ struct tile_x_sizes {

 static int get_mmq_x_max_host(const int cc) {
    return new_mma_available(cc) ? 128 :
-        ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA && cc < GGML_CUDA_CC_OFFSET_AMD ?
+        GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA ?
 #ifdef GGML_CUDA_FORCE_MMQ
            128                     : 64;
 #else
@@ -123,8 +123,8 @@ static constexpr __device__ int get_mmq_x_max_device() {
 }

 static int get_mmq_y_host(const int cc) {
-    return cc >= GGML_CUDA_CC_OFFSET_AMD ? (GGML_CUDA_CC_IS_RDNA1(cc) ? 64 : 128) :
-        (ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA ? 128 : 64);
+    return GGML_CUDA_CC_IS_AMD(cc) ? (GGML_CUDA_CC_IS_RDNA1(cc) ? 64 : 128) :
+        ((GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) ? 128 : 64);
 }

 static constexpr __device__ int get_mmq_y_device() {
@@ -945,7 +945,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma(
        }
    }
 #else
-    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum);
+    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum); GGML_UNUSED(k00);
    NO_DEVICE_CODE;
 #endif // NEW_MMA_AVAILABLE
 }
@@ -1024,7 +1024,7 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_dp4a(
    }

 #pragma unroll
-    for (int k01 = 0; k01 < WARP_SIZE; k01 += QR2_K*VDR_Q2_K_Q8_1_MMQ) {
+    for (int k01 = 0; k01 < WARP_SIZE/2; k01 += QR2_K*VDR_Q2_K_Q8_1_MMQ) {
        const int k0 = k00 + k01;

 #pragma unroll
@@ -1035,19 +1035,34 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_dp4a(
            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
                const int i = i0 + threadIdx.x;

-                if (k01 < WARP_SIZE/2) {
-                    constexpr int ns = 2;
-                    sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q2_K_q8_1_impl_mmq<ns>(
-                        &x_qs[i*(2*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01],
-                        &x_dm[i*(WARP_SIZE + 1) + k0/4], k01 < WARP_SIZE/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y,
-                        &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]);
-                } else {
-                    constexpr int ns = 1;
-                    sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q2_K_q8_1_impl_mmq<ns>(
-                        &x_qs[i*(2*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01],
-                        &x_dm[i*(WARP_SIZE + 1) + k0/4], k01 < WARP_SIZE/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y,
-                        &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]);
-                }
+                constexpr int ns = 2;
+                sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q2_K_q8_1_impl_mmq<ns>(
+                    &x_qs[i*(2*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01],
+                    &x_dm[i*(WARP_SIZE + 1) + k0/4], k01 < WARP_SIZE/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y,
+                    &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]);
+            }
+        }
+    }
+
+    // Some compilers fail to unroll the loop over k01 if there is a conditional statement for ns in the inner loop.
+    // As a workaround 2 separate loops are used instead.
+#pragma unroll
+    for (int k01 = WARP_SIZE/2; k01 < WARP_SIZE; k01 += QR2_K*VDR_Q2_K_Q8_1_MMQ) {
+        const int k0 = k00 + k01;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                constexpr int ns = 1;
+                sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q2_K_q8_1_impl_mmq<ns>(
+                    &x_qs[i*(2*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01],
+                    &x_dm[i*(WARP_SIZE + 1) + k0/4], k01 < WARP_SIZE/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y,
+                    &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]);
            }
        }
    }
@@ -1176,7 +1191,7 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
        }
    }
 #else
-    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum);
+    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum); GGML_UNUSED(k00);
    NO_DEVICE_CODE;
 #endif // NEW_MMA_AVAILABLE
 }
@@ -1253,7 +1268,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
        const float d = bxi->d;

 #pragma unroll
-        for (int l = 0; l < sizeof(int); ++l) {
+        for (int l = 0; l < int(sizeof(int)); ++l) {
            x_df[i*MMQ_MMA_TILE_X_K_Q3_K + sizeof(int)*(threadIdx.x % (WARP_SIZE/8)) + l] = d*sc8[l];
        }
 #else
@@ -1376,7 +1391,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
        const half2 dm = bxi->dm * make_half2(1.0f, -1.0f);

 #pragma unroll
-        for (int l = 0; l < sizeof(int); ++l) {
+        for (int l = 0; l < int(sizeof(int)); ++l) {
            x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + sizeof(int)*ksc + l] = dm*make_half2(sc8[l], m8[l]);
        }
    }
@@ -1517,7 +1532,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
        const half2 dm = bxi->dm * make_half2(1.0f, -1.0f);

 #pragma unroll
-        for (int l = 0; l < sizeof(int); ++l) {
+        for (int l = 0; l < int(sizeof(int)); ++l) {
            x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + sizeof(int)*ksc + l] = dm*make_half2(sc8[l], m8[l]);
        }
    }
@@ -1810,7 +1825,7 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma(
        }
    }
 #else
-    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum);
+    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum); GGML_UNUSED(k00);
    NO_DEVICE_CODE;
 #endif // NEW_MMA_AVAILABLE
 }
@@ -2570,6 +2585,8 @@ static __device__ void mul_mat_q_process_tile(
    } else {
        write_back(sum, dst + jt*mmq_x*ne0 + it*mmq_y, ne0, tile_x_max_i, tile_y_max_j);
    }
+
+    GGML_UNUSED(ne00); GGML_UNUSED(ne10);
 }


@@ -2577,9 +2594,9 @@ static __device__ void mul_mat_q_process_tile(

 template <ggml_type type, int mmq_x, int nwarps, bool need_check>
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
+#if defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
    __launch_bounds__(WARP_SIZE*nwarps, 2)
-#endif // defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
+#endif // defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
 #else
 #if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
    __launch_bounds__(WARP_SIZE*nwarps, 1)
@@ -2695,7 +2712,7 @@ static __global__ void mul_mat_q_stream_k_fixup(
        const int it = (kbc_stop - jt*(blocks_per_ne00*nty)) / blocks_per_ne00;

        // Skip fixup tile if it's unrelated to the output tile assigned to this CUDA block:
-        if (it != blockIdx.x || jt != blockIdx.y) {
+        if ((unsigned)it != blockIdx.x || (unsigned)jt != blockIdx.y) {
            continue;
        }

@@ -2772,14 +2789,14 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a

    const int shmem = mmq_get_shmem<type>(mmq_x, mmq_y, cc);

-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
    static bool shmem_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
    if (!shmem_limit_raised[id]) {
        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, false>, cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, true>,  cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
        shmem_limit_raised[id] = true;
    }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)

    const int nty = (args.ne01 + mmq_y - 1) / mmq_y;
    const int ntx = (args.ne11 + mmq_x - 1) / mmq_x;
@@ -2825,14 +2842,13 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
 template <ggml_type type>
 void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
    const int id    = ggml_cuda_get_device();
-    const int nsm   = ggml_cuda_info().devices[id].nsm;
    const int cc    = ggml_cuda_info().devices[id].cc;
    const int smpbo = ggml_cuda_info().devices[id].smpbo;

    const int mmq_x_max = get_mmq_x_max_host(cc);
    const int mmq_y = get_mmq_y_host(cc);
    const int block_num_y = (args.ne01 + mmq_y - 1) / mmq_y;
-    const bool use_stream_k = ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA && cc < GGML_CUDA_CC_OFFSET_AMD;
+    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA;

    int mmq_x_best  = 0;
    int nparts_best = INT_MAX;
@@ -29,7 +29,7 @@ static __global__ void mul_mat_vec(
        __syncthreads();
    }

-    float sumf;
+    float sumf = 0.0f;

    if constexpr (std::is_same<T, half>::value) {
        const half2 * x2 = (const half2 *) x;
@@ -54,7 +54,7 @@ enum mmvq_parameter_table_id {
 };

 static constexpr __device__ mmvq_parameter_table_id get_device_table_id() {
-#if defined(RDNA2) || defined(RDNA3)
+#if defined(RDNA2) || defined(RDNA3) || defined(RDNA4)
    return MMVQ_PARAMETERS_RDNA2;
 #elif defined(GCN) || defined(CDNA)
    return MMVQ_PARAMETERS_GCN;
@@ -64,7 +64,7 @@ static constexpr __device__ mmvq_parameter_table_id get_device_table_id() {
 }

 static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
-    if (GGML_CUDA_CC_IS_RDNA2(cc) || GGML_CUDA_CC_IS_RDNA3(cc)) {
+    if (GGML_CUDA_CC_IS_RDNA2(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
        return MMVQ_PARAMETERS_RDNA2;
    }
    if (GGML_CUDA_CC_IS_GCN(cc) || GGML_CUDA_CC_IS_CDNA(cc)) {
@@ -151,7 +151,7 @@ static __global__ void mul_mat_vec_q(
    constexpr int blocks_per_iter = vdr * nwarps*warp_size / qi;

    // partial sum for each thread
-    float tmp[ncols_y][rows_per_cuda_block] = {0.0f};
+    float tmp[ncols_y][rows_per_cuda_block] = {{0.0f}};

    const block_q8_1 * y = (const block_q8_1 *) vy;

@@ -197,10 +197,12 @@ static __global__ void mul_mat_vec_q(
            tmp[j][i] = warp_reduce_sum<warp_size>(tmp[j][i]);
        }

-        if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || row0 + threadIdx.x < nrows_dst)) {
+        if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || row0 + threadIdx.x < (unsigned)nrows_dst)) {
            dst[j*nrows_dst + row0 + threadIdx.x] = tmp[j][threadIdx.x];
        }
    }
+
+    GGML_UNUSED(nrows_x);
 }

 static std::pair<dim3, dim3> calc_launch_params(const int ncols_y, const int nrows_x, const int warp_size, const mmvq_parameter_table_id table_id) {
@@ -14,7 +14,7 @@ static __global__ void pad_f32(const float * x, float * dst, const int ne0, cons
        nidx +
        blockIdx.y * ne0 +
        blockIdx.z * ne0 * gridDim.y;
-    if (nidx < ne00 && blockIdx.y < ne01 && blockIdx.z < ne02*ne03) {
+    if (nidx < ne00 && blockIdx.y < (unsigned)ne01 && blockIdx.z < (unsigned)(ne02*ne03)) {
        int offset_src =
            nidx +
            blockIdx.y * ne00 +
@@ -19,7 +19,7 @@ static __global__ void upscale_f32(const float * x, float * dst,
    int i02 = i12 / sf2;
    int i03 = i13 / sf3;

-    dst[index] = *(float *)((char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00);
+    dst[index] = *( (const float *)((const char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00) );
 }

 static void upscale_f32_cuda(const float * x, float * dst,
@@ -129,6 +129,7 @@
 #define cudaGraph_t hipGraph_t
 #define cudaStream_t hipStream_t
 #define cudaSuccess hipSuccess
+#define cudaOccupancyMaxActiveBlocksPerMultiprocessor hipOccupancyMaxActiveBlocksPerMultiprocessor
 #define __trap() do { abort(); __builtin_unreachable(); } while(0)
 #define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
 #define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED
@@ -150,6 +151,10 @@
 #define CDNA
 #endif

+#if defined(__GFX12__)
+#define RDNA4
+#endif
+
 #if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
    defined(__gfx1150__) || defined(__gfx1151__)
 #define RDNA3
@@ -134,5 +134,6 @@
 #define cudaStreamCaptureModeRelaxed musaStreamCaptureModeRelaxed
 #define cudaStreamBeginCapture musaStreamBeginCapture
 #define cudaStreamEndCapture musaStreamEndCapture
+#define cudaOccupancyMaxActiveBlocksPerMultiprocessor musaOccupancyMaxActiveBlocksPerMultiprocessor

 typedef mt_bfloat16 nv_bfloat16;
@@ -381,6 +381,35 @@ GGML_API void ggml_aligned_free(void * ptr, size_t size);
        return r;
    }

+#elif defined(__riscv) && defined(GGML_RV_ZFH)
+
+    static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+        float f;
+        __asm__(
+            "fmv.h.x %[f], %[h]\n\t"
+            "fcvt.s.h %[f], %[f]"
+            : [f] "=&f" (f)
+            : [h] "r" (h)
+        );
+        return f;
+    }
+
+    static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+        ggml_fp16_t res;
+        __asm__(
+            "fcvt.h.s %[f], %[f]\n\t"
+            "fmv.x.h %[h], %[f]"
+            : [h] "=&r" (res)
+            : [f] "f" (f)
+        );
+        return res;
+    }
+
+    #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+    #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+    #define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
+    #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+
 #else

    // FP16 <-> FP32
@@ -1,6 +1,70 @@
 #ifndef GGML_METAL_IMPL
 #define GGML_METAL_IMPL

+// kernel parameters for mat-vec threadgroups
+//
+// N_R0: number of src0 rows to process per simdgroup
+// N_SG: number of simdgroups per threadgroup
+//
+// TODO: for optimal performance, become function of the device and work size
+
+#define N_R0_Q4_0 4
+#define N_SG_Q4_0 2
+
+#define N_R0_Q4_1 4
+#define N_SG_Q4_1 2
+
+#define N_R0_Q5_0 4
+#define N_SG_Q5_0 2
+
+#define N_R0_Q5_1 4
+#define N_SG_Q5_1 2
+
+#define N_R0_Q8_0 4
+#define N_SG_Q8_0 2
+
+#define N_R0_Q2_K 4
+#define N_SG_Q2_K 2
+
+#define N_R0_Q3_K 2
+#define N_SG_Q3_K 2
+
+#define N_R0_Q4_K 4
+#define N_SG_Q4_K 2
+
+#define N_R0_Q5_K 2
+#define N_SG_Q5_K 2
+
+#define N_R0_Q6_K 1
+#define N_SG_Q6_K 2
+
+#define N_R0_IQ1_S 4
+#define N_SG_IQ1_S 2
+
+#define N_R0_IQ1_M 4
+#define N_SG_IQ1_M 2
+
+#define N_R0_IQ2_XXS 4
+#define N_SG_IQ2_XXS 2
+
+#define N_R0_IQ2_XS 4
+#define N_SG_IQ2_XS 2
+
+#define N_R0_IQ2_S 4
+#define N_SG_IQ2_S 2
+
+#define N_R0_IQ3_XXS 4
+#define N_SG_IQ3_XXS 2
+
+#define N_R0_IQ3_S 4
+#define N_SG_IQ3_S 2
+
+#define N_R0_IQ4_NL 2
+#define N_SG_IQ4_NL 2
+
+#define N_R0_IQ4_XS 2
+#define N_SG_IQ4_XS 2
+
 // kernel argument structs
 //
 // - element counters (e.g. ne00) typically use int32_t to reduce register usage
@@ -155,9 +219,12 @@ typedef struct {
    int32_t  ne11;
    int32_t  ne_12_2; // assume K and V are same shape
    int32_t  ne_12_3;
-    uint64_t nb_12_1;
-    uint64_t nb_12_2;
-    uint64_t nb_12_3;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    uint64_t nb21;
+    uint64_t nb22;
+    uint64_t nb23;
    uint64_t nb31;
    int32_t  ne1;
    int32_t  ne2;
@@ -25,124 +25,47 @@ endif ()
 if (GGML_OPENCL_EMBED_KERNELS)
    add_compile_definitions(GGML_OPENCL_EMBED_KERNELS)

-    set(OPENCL_CL_SOURCE_EMBED         "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl.cl.h")
-    set(OPENCL_MM_CL_SOURCE_EMBED      "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_mm.cl.h")
-    set(OPENCL_CVT_CL_SOURCE_EMBED     "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_cvt.cl.h")
+    set(EMBED_KERNEL_SCRIPT "${CMAKE_CURRENT_SOURCE_DIR}/kernels/embed_kernel.py")
+    file(MAKE_DIRECTORY     "${CMAKE_CURRENT_BINARY_DIR}/autogenerated")

-    set(OPENCL_GEMV_NOSHUFFLE_SOURCE_EMBED             "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_gemv_noshuffle.cl.h")
-    set(OPENCL_GEMV_NOSHUFFLE_GENERAL_SOURCE_EMBED     "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_gemv_noshuffle_general.cl.h")
-    set(OPENCL_MUL_MAT_Ab_Bi_8x4_SOURCE_EMBED          "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_mul_mat_Ab_Bi_8x4.cl.h")
-    set(OPENCL_TRANSPOSE_16_SOURCE_EMBED               "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_transpose_16.cl.h")
-    set(OPENCL_TRANSPOSE_32_SOURCE_EMBED               "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_transpose_32.cl.h")
-    set(OPENCL_TRANSPOSE_32_16_SOURCE_EMBED            "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_transpose_32_16.cl.h")
-
-    set(EMBED_KERNEL_SCRIPT             "${CMAKE_CURRENT_SOURCE_DIR}/kernels/embed_kernel.py")
-    file(MAKE_DIRECTORY                 "${CMAKE_BINARY_DIR}/autogenerated")
-
-    include_directories("${CMAKE_BINARY_DIR}/autogenerated")
-
-    # Python must be accessible from command line
-    add_custom_command(
-        OUTPUT ${OPENCL_CL_SOURCE_EMBED}
-        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
-            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl.cl
-            ${OPENCL_CL_SOURCE_EMBED}
-        DEPENDS kernels/ggml-opencl.cl ${EMBED_KERNEL_SCRIPT}
-        COMMENT "Generate ggml-opencl.cl.h"
-    )
-
-    add_custom_command(
-        OUTPUT ${OPENCL_MM_CL_SOURCE_EMBED}
-        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
-            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_mm.cl
-            ${OPENCL_MM_CL_SOURCE_EMBED}
-        DEPENDS kernels/ggml-opencl_mm.cl ${EMBED_KERNEL_SCRIPT}
-        COMMENT "Generate ggml-opencl_mm.cl.h"
-    )
-
-    add_custom_command(
-        OUTPUT ${OPENCL_CVT_CL_SOURCE_EMBED}
-        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
-            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_cvt.cl
-            ${OPENCL_CVT_CL_SOURCE_EMBED}
-        DEPENDS kernels/ggml-opencl_cvt.cl ${EMBED_KERNEL_SCRIPT}
-        COMMENT "Generate ggml-opencl_cvt.cl.h"
-    )
-
-    add_custom_command(
-        OUTPUT ${OPENCL_GEMV_NOSHUFFLE_SOURCE_EMBED}
-        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
-            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_gemv_noshuffle.cl
-            ${OPENCL_GEMV_NOSHUFFLE_SOURCE_EMBED}
-        DEPENDS kernels/ggml-opencl_gemv_noshuffle.cl ${EMBED_KERNEL_SCRIPT}
-        COMMENT "Generate ggml-opencl_gemv_noshuffle.cl.h"
-    )
-
-    add_custom_command(
-        OUTPUT ${OPENCL_GEMV_NOSHUFFLE_GENERAL_SOURCE_EMBED}
-        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
-            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_gemv_noshuffle_general.cl
-            ${OPENCL_GEMV_NOSHUFFLE_GENERAL_SOURCE_EMBED}
-        DEPENDS kernels/ggml-opencl_gemv_noshuffle_general.cl ${EMBED_KERNEL_SCRIPT}
-        COMMENT "Generate ggml-opencl_gemv_noshuffle_general.cl.h"
-    )
-
-    add_custom_command(
-        OUTPUT ${OPENCL_MUL_MAT_Ab_Bi_8x4_SOURCE_EMBED}
-        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
-            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_mul_mat_Ab_Bi_8x4.cl
-            ${OPENCL_MUL_MAT_Ab_Bi_8x4_SOURCE_EMBED}
-        DEPENDS kernels/ggml-opencl_mul_mat_Ab_Bi_8x4.cl ${EMBED_KERNEL_SCRIPT}
-        COMMENT "Generate ggml-opencl_mul_mat_Ab_Bi_8x4.cl.cl.h"
-    )
-
-    add_custom_command(
-        OUTPUT ${OPENCL_TRANSPOSE_16_SOURCE_EMBED}
-        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
-            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_transpose_16.cl
-            ${OPENCL_TRANSPOSE_16_SOURCE_EMBED}
-        DEPENDS kernels/ggml-opencl_transpose_16.cl ${EMBED_KERNEL_SCRIPT}
-        COMMENT "Generate ggml-opencl_transpose_16.cl.h"
-    )
-
-    add_custom_command(
-        OUTPUT ${OPENCL_TRANSPOSE_32_SOURCE_EMBED}
-        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
-            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_transpose_32.cl
-            ${OPENCL_TRANSPOSE_32_SOURCE_EMBED}
-        DEPENDS kernels/ggml-opencl_transpose_32.cl ${EMBED_KERNEL_SCRIPT}
-        COMMENT "Generate ggml-opencl_transpose_32.cl.h"
-    )
-
-    add_custom_command(
-        OUTPUT ${OPENCL_TRANSPOSE_32_16_SOURCE_EMBED}
-        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
-            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_transpose_32_16.cl
-            ${OPENCL_TRANSPOSE_32_16_SOURCE_EMBED}
-        DEPENDS kernels/ggml-opencl_transpose_32_16.cl ${EMBED_KERNEL_SCRIPT}
-        COMMENT "Generate ggml-opencl_transpose_32_16.cl.h"
-    )
-
-    target_sources(${TARGET_NAME} PRIVATE
-                   ${OPENCL_CL_SOURCE_EMBED}
-                   ${OPENCL_MM_CL_SOURCE_EMBED}
-                   ${OPENCL_CVT_CL_SOURCE_EMBED}
-                   ${OPENCL_GEMV_NOSHUFFLE_SOURCE_EMBED}
-                   ${OPENCL_GEMV_NOSHUFFLE_GENERAL_SOURCE_EMBED}
-                   ${OPENCL_MUL_MAT_Ab_Bi_8x4_SOURCE_EMBED}
-                   ${OPENCL_TRANSPOSE_16_SOURCE_EMBED}
-                   ${OPENCL_TRANSPOSE_32_SOURCE_EMBED}
-                   ${OPENCL_TRANSPOSE_32_16_SOURCE_EMBED})
-else ()
-    # copy ggml-opencl.cl to bin directory
-    configure_file(kernels/ggml-opencl.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl.cl COPYONLY)
-    configure_file(kernels/ggml-opencl_mm.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_mm.cl COPYONLY)
-    configure_file(kernels/ggml-opencl_cvt.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_cvt.cl COPYONLY)
-
-    configure_file(kernels/ggml-opencl_gemv_noshuffle.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_gemv_noshuffle.cl COPYONLY)
-    configure_file(kernels/ggml-opencl_gemv_noshuffle_general.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_gemv_noshuffle_general.cl COPYONLY)
-    configure_file(kernels/ggml-opencl_mul_mat_Ab_Bi_8x4.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_mul_mat_Ab_Bi_8x4.cl COPYONLY)
-    configure_file(kernels/ggml-opencl_transpose_16.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_transpose_16.cl COPYONLY)
-    configure_file(kernels/ggml-opencl_transpose_32.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_transpose_32.cl COPYONLY)
-    configure_file(kernels/ggml-opencl_transpose_32_16.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_transpose_32_16.cl COPYONLY)
+    target_include_directories(${TARGET_NAME} PRIVATE "${CMAKE_CURRENT_BINARY_DIR}/autogenerated")
 endif ()
+
+function(ggml_opencl_add_kernel KNAME)
+    set(KERN_HDR ${CMAKE_CURRENT_BINARY_DIR}/autogenerated/${KNAME}.cl.h)
+    set(KERN_SRC ${CMAKE_CURRENT_SOURCE_DIR}/kernels/${KNAME}.cl)
+
+    if (GGML_OPENCL_EMBED_KERNELS)
+        message(STATUS "opencl: embedding kernel ${KNAME}")
+
+        # Python must be accessible from command line
+        add_custom_command(
+            OUTPUT ${KERN_HDR}
+            COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT} ${KERN_SRC} ${KERN_HDR}
+            DEPENDS ${KERN_SRC} ${EMBED_KERNEL_SCRIPT}
+            COMMENT "Generate ${KERN_HDR}"
+        )
+
+        target_sources(${TARGET_NAME} PRIVATE ${KERN_HDR})
+    else ()
+        message(STATUS "opencl: adding kernel ${KNAME}")
+        configure_file(${KERN_SRC} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${KNAME}.cl COPYONLY)
+    endif ()
+endfunction()
+
+set(GGML_OPENCL_KERNELS
+    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
+)
+
+foreach (K ${GGML_OPENCL_KERNELS})
+    ggml_opencl_add_kernel(${K})
+endforeach()
@@ -224,12 +224,14 @@ struct ggml_backend_opencl_context {
    cl_program program;
    cl_program program_1;
    cl_program program_2;
+    cl_program program_im2col;

    cl_kernel kernel_add, kernel_add_row;
    cl_kernel kernel_mul, kernel_mul_row;
    cl_kernel kernel_scale;
    cl_kernel kernel_silu, kernel_silu_4;
    cl_kernel kernel_gelu, kernel_gelu_4;
+    cl_kernel kernel_gelu_quick, kernel_gelu_quick_4;
    cl_kernel kernel_relu;
    cl_kernel kernel_clamp;
    cl_kernel kernel_norm;
@@ -239,6 +241,7 @@ struct ggml_backend_opencl_context {
    cl_kernel kernel_soft_max_f16, kernel_soft_max_4_f16;
    cl_kernel kernel_get_rows_f32, kernel_get_rows_f16, kernel_get_rows_q4_0;
    cl_kernel kernel_rope_norm_f32, kernel_rope_norm_f16, kernel_rope_neox_f32, kernel_rope_neox_f16;
+    cl_kernel kernel_rope_multi_f32, kernel_rope_multi_f16, kernel_rope_vision_f32, kernel_rope_vision_f16;
    cl_kernel kernel_cpy_f16_f16, kernel_cpy_f16_f32, kernel_cpy_f32_f16, kernel_cpy_f32_f32;
    cl_kernel kernel_mul_mat_f32_f32;
    cl_kernel kernel_mul_mat_f16_f16;
@@ -252,6 +255,7 @@ struct ggml_backend_opencl_context {
              kernel_mul_mat_q4_0_f32_flat_img_v0;
    cl_kernel kernel_mul_mat_q4_0_f32_1d_8x_flat, kernel_mul_mat_q4_0_f32_1d_16x_flat;
    cl_kernel kernel_mul_mv_q6_K_f32;
+    cl_kernel kernel_im2col_f32, kernel_im2col_f16;

 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
    // Transpose kernels
@@ -297,8 +301,27 @@ static int ggml_backend_opencl_n_devices = 0;
 struct ProfilingInfo {
    std::string op_name;
    std::string kernel_name;
-    // Kernel execution time in nanoseconds.
-    cl_ulong duration_ns;
+
+    cl_kernel kernel;
+    cl_event evt;
+
+    cl_ulong cmd_queued;
+    cl_ulong cmd_submit;
+    cl_ulong cmd_start;
+    cl_ulong cmd_end;
+    cl_ulong overhead_start;
+    cl_ulong overhead_end;
+    // For the times below, see spec for clGetEventProfilingInfo
+    // The time kernel spent in cmd queue - SUBMIT - QUEUED
+    cl_ulong cmd_queued_duration_ns;
+    // The time kernel spent for submission - START - SUBMIT
+    cl_ulong cmd_submit_duration_ns;
+    // Kernel execution time in nanoseconds - END - START
+    cl_ulong cmd_duration_ns;
+    // The time for the kernel to complete - COMPLETE - END
+    cl_ulong cmd_complete_duration_ns;
+    // Total time to finish the kernel - COMPELTE - QUEUED
+    cl_ulong cmd_total_duration_ns;
    // Global and local work sizes.
    size_t global_size[3];
    size_t local_size[3];
@@ -689,6 +712,8 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
    CL_CHECK((backend_ctx->kernel_silu_4             = clCreateKernel(backend_ctx->program, "kernel_silu_4", &err), err));
    CL_CHECK((backend_ctx->kernel_gelu               = clCreateKernel(backend_ctx->program, "kernel_gelu", &err), err));
    CL_CHECK((backend_ctx->kernel_gelu_4             = clCreateKernel(backend_ctx->program, "kernel_gelu_4", &err), err));
+    CL_CHECK((backend_ctx->kernel_gelu_quick         = clCreateKernel(backend_ctx->program, "kernel_gelu_quick", &err), err));
+    CL_CHECK((backend_ctx->kernel_gelu_quick_4       = clCreateKernel(backend_ctx->program, "kernel_gelu_quick_4", &err), err));
    CL_CHECK((backend_ctx->kernel_relu               = clCreateKernel(backend_ctx->program, "kernel_relu", &err), err));
    CL_CHECK((backend_ctx->kernel_clamp              = clCreateKernel(backend_ctx->program, "kernel_clamp", &err), err));
    CL_CHECK((backend_ctx->kernel_norm               = clCreateKernel(backend_ctx->program, "kernel_norm", &err), err));
@@ -703,6 +728,10 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
    CL_CHECK((backend_ctx->kernel_rope_norm_f16      = clCreateKernel(backend_ctx->program, "kernel_rope_norm_f16", &err), err));
    CL_CHECK((backend_ctx->kernel_rope_neox_f32      = clCreateKernel(backend_ctx->program, "kernel_rope_neox_f32", &err), err));
    CL_CHECK((backend_ctx->kernel_rope_neox_f16      = clCreateKernel(backend_ctx->program, "kernel_rope_neox_f16", &err), err));
+    CL_CHECK((backend_ctx->kernel_rope_multi_f32     = clCreateKernel(backend_ctx->program, "kernel_rope_multi_f32", &err), err));
+    CL_CHECK((backend_ctx->kernel_rope_multi_f16     = clCreateKernel(backend_ctx->program, "kernel_rope_multi_f16", &err), err));
+    CL_CHECK((backend_ctx->kernel_rope_vision_f32    = clCreateKernel(backend_ctx->program, "kernel_rope_vision_f32", &err), err));
+    CL_CHECK((backend_ctx->kernel_rope_vision_f16    = clCreateKernel(backend_ctx->program, "kernel_rope_vision_f16", &err), err));
    CL_CHECK((backend_ctx->kernel_cpy_f16_f16        = clCreateKernel(backend_ctx->program, "kernel_cpy_f16_f16", &err), err));
    CL_CHECK((backend_ctx->kernel_cpy_f16_f32        = clCreateKernel(backend_ctx->program, "kernel_cpy_f16_f32", &err), err));
    CL_CHECK((backend_ctx->kernel_cpy_f32_f16        = clCreateKernel(backend_ctx->program, "kernel_cpy_f32_f16", &err), err));
@@ -750,6 +779,19 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {

    CL_CHECK((backend_ctx->kernel_convert_block_q4_0_noshuffle     = clCreateKernel(backend_ctx->program_2, "kernel_convert_block_q4_0_noshuffle", &err), err));

+    // im2col kernels
+#ifdef GGML_OPENCL_EMBED_KERNELS
+    const std::string kernel_src_im2col {
+        #include "ggml-opencl_im2col.cl.h"
+    };
+#else
+    const std::string kernel_src_im2col = read_file("ggml-opencl_im2col.cl");
+#endif
+    backend_ctx->program_im2col = build_program_from_source(context, device, kernel_src_im2col.c_str(), compile_opts);
+
+    CL_CHECK((backend_ctx->kernel_im2col_f32 = clCreateKernel(backend_ctx->program_im2col, "kernel_im2col_f32", &err), err));
+    CL_CHECK((backend_ctx->kernel_im2col_f16 = clCreateKernel(backend_ctx->program_im2col, "kernel_im2col_f16", &err), err));
+
    // Kernels for Adreno
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -903,12 +945,56 @@ static void ggml_cl2_free(void) {
        return;
    }

+    // Populate profiling info
+    for (ProfilingInfo & info : g_profiling_info) {
+        cl_ulong cmd_queued;
+        cl_ulong cmd_submit;
+        cl_ulong cmd_start;
+        cl_ulong cmd_end;
+        cl_ulong cmd_complete;
+
+        CL_CHECK(clWaitForEvents(1, &info.evt));
+        CL_CHECK(clGetEventProfilingInfo(
+            info.evt, CL_PROFILING_COMMAND_QUEUED, sizeof(cl_ulong), &cmd_queued, NULL));
+        CL_CHECK(clGetEventProfilingInfo(
+            info.evt, CL_PROFILING_COMMAND_SUBMIT, sizeof(cl_ulong), &cmd_submit, NULL));
+        CL_CHECK(clGetEventProfilingInfo(
+            info.evt, CL_PROFILING_COMMAND_START, sizeof(cl_ulong), &cmd_start, NULL));
+        CL_CHECK(clGetEventProfilingInfo(
+            info.evt, CL_PROFILING_COMMAND_END, sizeof(cl_ulong), &cmd_end, NULL));
+        CL_CHECK(clGetEventProfilingInfo(
+            info.evt, CL_PROFILING_COMMAND_COMPLETE, sizeof(cl_ulong), &cmd_complete, NULL));
+        CL_CHECK(clReleaseEvent(info.evt));
+
+        char kernel_name[512];
+        CL_CHECK(clGetKernelInfo(info.kernel, CL_KERNEL_FUNCTION_NAME,
+            sizeof(kernel_name), kernel_name, NULL));
+        info.kernel_name = kernel_name;
+
+        info.cmd_queued = cmd_queued;
+        info.cmd_submit = cmd_submit;
+        info.cmd_start  = cmd_start;
+        info.cmd_end    = cmd_end;
+
+        info.cmd_queued_duration_ns     = cmd_submit    - cmd_queued;
+        info.cmd_submit_duration_ns     = cmd_start     - cmd_submit;
+        info.cmd_duration_ns            = cmd_end       - cmd_start;
+        info.cmd_complete_duration_ns   = cmd_complete  - cmd_end;
+        info.cmd_total_duration_ns      = cmd_complete  - cmd_queued;
+    }
+
+    // Dump a csv
    float total_kernel_time = 0;
-    fprintf(fperf, "op name, kernel name, duration (ms), global size, local size, output size\n");
+    fprintf(fperf, "op name, kernel name, queued duration (ms), submit duration(ms), exec duration (ms), complete duration (ms), total duration (ms), global size, local size, output size\n");
    for (const ProfilingInfo & info : g_profiling_info) {
-        total_kernel_time += info.duration_ns/1.e6f;
-        fprintf(fperf, "%s,%s,%f,%zux%zux%zu,%zux%zux%zu,%zux%zux%zux%zu\n",
-            info.op_name.c_str(), info.kernel_name.c_str(), info.duration_ns/1.e6f,
+        total_kernel_time += info.cmd_duration_ns/1.e6f;
+        fprintf(fperf, "%s,%s,%f,%f,%f,%f,%f,%zux%zux%zu,%zux%zux%zu,%zux%zux%zux%zu\n",
+            info.op_name.c_str(), info.kernel_name.c_str(),
+            info.cmd_queued_duration_ns/1.e6f,
+            info.cmd_submit_duration_ns/1.e6f,
+            info.cmd_duration_ns/1.e6f,
+            info.cmd_complete_duration_ns/1.e6f,
+            info.cmd_total_duration_ns/1.e6f,
            info.global_size[0], info.global_size[1], info.global_size[2],
            info.local_size[0], info.local_size[2], info.local_size[2],
            info.output_size[0], info.output_size[1], info.output_size[2], info.output_size[3]);
@@ -916,6 +1002,27 @@ static void ggml_cl2_free(void) {
    fclose(fperf);

    GGML_LOG_INFO("ggml_opencl: total kernel time: %f\n", total_kernel_time);
+
+    // Dump a simple chrome trace
+    FILE* ftrace = fopen("cl_trace.json", "w");
+    if (!ftrace) {
+        GGML_LOG_ERROR("Failed to open cl_trace.json\n");
+        return;
+    }
+
+    fprintf(ftrace, "[\n");
+    for (const ProfilingInfo & info : g_profiling_info) {
+        fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"B\", \"ts\": %lu, \"pid\": \"\", \"tid\": \"Host\"},\n",
+            info.kernel_name.c_str(), info.cmd_queued/1000);
+        fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"E\", \"ts\": %lu, \"pid\": \"\", \"tid\": \"Host\"},\n",
+            info.kernel_name.c_str(), info.cmd_submit/1000);
+
+        fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"B\", \"ts\": %lu, \"pid\": \"\", \"tid\": \"Device\"},\n",
+            info.kernel_name.c_str(), info.cmd_start/1000);
+        fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"E\", \"ts\": %lu, \"pid\": \"\", \"tid\": \"Device\"},\n",
+            info.kernel_name.c_str(), info.cmd_end/1000);
+    }
+    fclose(ftrace);
 #endif
 }

@@ -1103,6 +1210,7 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                case GGML_UNARY_OP_GELU:
                case GGML_UNARY_OP_SILU:
                case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_GELU_QUICK:
                   return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
                default:
                    return false;
@@ -1132,14 +1240,26 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
            return op->ne[3] == 1;
        case GGML_OP_ROPE: {
            const int mode = ((const int32_t *) op->op_params)[2];
-            if (mode & GGML_ROPE_TYPE_MROPE) {
+            const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
+            const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
+            if (is_mrope && !is_vision) {
+                if (op->src[0]->type == GGML_TYPE_F32 ||
+                    op->src[0]->type == GGML_TYPE_F16) {
+                    return true;
+                }
                return false;
            }
-            if (mode & GGML_ROPE_TYPE_VISION) {
+            if (is_vision) {
+                if (op->src[0]->type == GGML_TYPE_F32 ||
+                    op->src[0]->type == GGML_TYPE_F16) {
+                    return true;
+                }
                return false;
            }
            return true;
        }
+        case GGML_OP_IM2COL:
+            return true;
        default:
            return false;
    }
@@ -2062,25 +2182,14 @@ static void dump_tensor(ggml_backend_t backend, const struct ggml_tensor * tenso
 // Profiling utility
 //------------------------------------------------------------------------------
 #ifdef GGML_OPENCL_PROFILING
-void populateProfilingInfo(
+static void populateProfilingInfo(
        ProfilingInfo& info, cl_event evt, cl_kernel kernel,
        size_t global_size[3], size_t local_size[3],
        const ggml_tensor * tensor) {
-    cl_ulong start;
-    cl_ulong end;
-    CL_CHECK(clWaitForEvents(1, &evt));
-    CL_CHECK(clGetEventProfilingInfo(
-        evt, CL_PROFILING_COMMAND_START, sizeof(cl_ulong), &start, NULL));
-    CL_CHECK(clGetEventProfilingInfo(
-        evt, CL_PROFILING_COMMAND_END, sizeof(cl_ulong), &end, NULL));
+    info.op_name     = tensor->name;
+    info.kernel      = kernel;
+    info.evt         = evt;

-    char kernel_name[512];
-    CL_CHECK(clGetKernelInfo(kernel, CL_KERNEL_FUNCTION_NAME,
-        sizeof(kernel_name), kernel_name, NULL));
-
-    info.duration_ns = end - start;
-    info.op_name = tensor->name;
-    info.kernel_name = kernel_name;
    info.local_size[0]  = local_size[0];
    info.local_size[1]  = local_size[1];
    info.local_size[2]  = local_size[2];
@@ -2509,6 +2618,53 @@ static void ggml_cl_gelu(ggml_backend_t backend, const ggml_tensor * src0, const
 #endif
 }

+static void ggml_cl_gelu_quick(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+
+    int n = ggml_nelements(dst);
+
+    if (n % 4 == 0) {
+        kernel = backend_ctx->kernel_gelu_quick_4;
+        n /= 4;
+    } else {
+        kernel = backend_ctx->kernel_gelu_quick;
+    }
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+
+    size_t global_work_size[] = {(size_t)n, 1, 1};
+    size_t local_work_size[] = {64, 1, 1};
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt);
+
+    g_profiling_info.emplace_back();
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst);
+#else
+    clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL);
+#endif
+}
+
 static void ggml_cl_silu(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    GGML_ASSERT(src0);
    GGML_ASSERT(src0->extra);
@@ -3907,6 +4063,7 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
    float attn_factor;
    float beta_fast;
    float beta_slow;
+    int32_t sections[4];

    memcpy(&freq_base,   (int32_t *) dst->op_params + 5, sizeof(float));
    memcpy(&freq_scale,  (int32_t *) dst->op_params + 6, sizeof(float));
@@ -3914,23 +4071,23 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
    memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float));
    memcpy(&beta_fast,   (int32_t *) dst->op_params + 9, sizeof(float));
    memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
+    memcpy(&sections,    (int32_t *) dst->op_params + 11, sizeof(int32_t)*4);

    const bool is_neox = mode & 2;
+    const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
+    const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
+
+    if (is_mrope) {
+        GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0);
+    }
+
+    if (is_vision) {
+        GGML_ASSERT(n_dims == ne00/2);
+    }

    cl_kernel kernel;

-    if (!is_neox) {
-        switch (src0->type) {
-            case GGML_TYPE_F32:
-                kernel = backend_ctx->kernel_rope_norm_f32;
-                break;
-            case GGML_TYPE_F16:
-                kernel = backend_ctx->kernel_rope_norm_f16;
-                break;
-            default:
-                GGML_ASSERT(false);
-        };
-    } else {
+    if (is_neox) {
        switch (src0->type) {
            case GGML_TYPE_F32:
                kernel = backend_ctx->kernel_rope_neox_f32;
@@ -3941,6 +4098,39 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
            default:
                GGML_ASSERT(false);
        };
+    } else if (is_mrope && !is_vision) {
+        switch (src0->type) {
+            case GGML_TYPE_F32:
+                kernel = backend_ctx->kernel_rope_multi_f32;
+                break;
+            case GGML_TYPE_F16:
+                kernel = backend_ctx->kernel_rope_multi_f16;
+                break;
+            default:
+                GGML_ASSERT(false);
+        };
+    } else if (is_vision) {
+        switch (src0->type) {
+            case GGML_TYPE_F32:
+                kernel = backend_ctx->kernel_rope_vision_f32;
+                break;
+            case GGML_TYPE_F16:
+                kernel = backend_ctx->kernel_rope_vision_f16;
+                break;
+            default:
+                GGML_ASSERT(false);
+        }
+    } else {
+        switch (src0->type) {
+            case GGML_TYPE_F32:
+                kernel = backend_ctx->kernel_rope_norm_f32;
+                break;
+            case GGML_TYPE_F16:
+                kernel = backend_ctx->kernel_rope_norm_f16;
+                break;
+            default:
+                GGML_ASSERT(false);
+        };
    }

    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
@@ -3976,6 +4166,9 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
    CL_CHECK(clSetKernelArg(kernel, 30, sizeof(float),    &attn_factor));
    CL_CHECK(clSetKernelArg(kernel, 31, sizeof(float),    &beta_fast));
    CL_CHECK(clSetKernelArg(kernel, 32, sizeof(float),    &beta_slow));
+    if (is_mrope || is_vision) {
+        CL_CHECK(clSetKernelArg(kernel, 33, sizeof(int32_t)*4, &sections));
+    }

    size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
    size_t local_work_size[] = {(size_t)nth, 1, 1};
@@ -3991,6 +4184,98 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
 #endif
 }

+static void ggml_cl_im2col(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    // src0 - filter, src1 - input
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    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 cl_long IC = src1->ne[is_2D ? 2 : 1];
+    const cl_long IH = is_2D ? src1->ne[1] : 1;
+    const cl_long IW =         src1->ne[0];
+
+    const cl_long KH = is_2D ? src0->ne[1] : 1;
+    const cl_long KW =         src0->ne[0];
+
+    const cl_long OH = is_2D ? dst->ne[2] : 1;
+    const cl_long OW =         dst->ne[1];
+
+    // nb is byte offset, src is type float32
+    const cl_ulong delta_offset = src1->nb[is_2D ? 2 : 1]/4;
+    const cl_long  batch        = src1->ne[is_2D ? 3 : 2];
+    const cl_ulong batch_offset = src1->nb[is_2D ? 3 : 2]/4;
+
+    const cl_long pelements = OW*KW*KH;
+    const cl_long CHW       = IC*KH*KW;
+
+    cl_kernel kernel;
+
+    if(dst->type == GGML_TYPE_F16) {
+        kernel = backend_ctx->kernel_im2col_f16;
+    } else {
+        kernel = backend_ctx->kernel_im2col_f32;
+    }
+
+    CL_CHECK(clSetKernelArg(kernel,   0, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,   1, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel,   2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,   3, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,   4, sizeof(cl_ulong), &batch_offset));
+    CL_CHECK(clSetKernelArg(kernel,   5, sizeof(cl_ulong), &delta_offset));
+    CL_CHECK(clSetKernelArg(kernel,   6, sizeof(cl_long),  &IW));
+    CL_CHECK(clSetKernelArg(kernel,   7, sizeof(cl_long),  &IH));
+    CL_CHECK(clSetKernelArg(kernel,   8, sizeof(cl_long),  &IC));
+    CL_CHECK(clSetKernelArg(kernel,   9, sizeof(cl_long),  &OW));
+    CL_CHECK(clSetKernelArg(kernel,  10, sizeof(cl_long),  &OH));
+    CL_CHECK(clSetKernelArg(kernel,  11, sizeof(cl_long),  &KW));
+    CL_CHECK(clSetKernelArg(kernel,  12, sizeof(cl_long),  &KH));
+    CL_CHECK(clSetKernelArg(kernel,  13, sizeof(cl_long),  &pelements));
+    CL_CHECK(clSetKernelArg(kernel,  14, sizeof(cl_long),  &CHW));
+    CL_CHECK(clSetKernelArg(kernel,  15, sizeof(int),      &s0));
+    CL_CHECK(clSetKernelArg(kernel,  16, sizeof(int),      &s1));
+    CL_CHECK(clSetKernelArg(kernel,  17, sizeof(int),      &p0));
+    CL_CHECK(clSetKernelArg(kernel,  18, sizeof(int),      &p1));
+    CL_CHECK(clSetKernelArg(kernel,  19, sizeof(int),      &d0));
+    CL_CHECK(clSetKernelArg(kernel,  20, sizeof(int),      &d1));
+
+    const int num_blocks = (pelements + 256 - 1) / 256;
+    size_t global_work_size[] = {(size_t)num_blocks*256, (size_t)OH, (size_t)batch*IC};
+    size_t local_work_size[] = {256, 1, 1};
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+
+    g_profiling_info.emplace_back();
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst);
+#else
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL));
+#endif
+}
+
 //------------------------------------------------------------------------------
 // Op offloading
 //------------------------------------------------------------------------------
@@ -4049,6 +4334,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
                    }
                    func = ggml_cl_gelu;
                    break;
+                case GGML_UNARY_OP_GELU_QUICK:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cl_gelu_quick;
+                    break;
                case GGML_UNARY_OP_SILU:
                    if (!any_on_device) {
                        return false;
@@ -4121,6 +4412,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
            }
            func = ggml_cl_rope;
            break;
+        case GGML_OP_IM2COL:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_im2col;
+            break;
        default:
            return false;
    }
@@ -404,6 +404,7 @@ kernel void kernel_scale(
 // gelu
 //------------------------------------------------------------------------------
 #define GELU_COEF_A     0.044715f
+#define GELU_QUICK_COEF -1.702f
 #define SQRT_2_OVER_PI  0.79788456080286535587989211986876f

 kernel void kernel_gelu(
@@ -434,6 +435,32 @@ kernel void kernel_gelu_4(
    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)));
+}
+
 //------------------------------------------------------------------------------
 // silu
 //------------------------------------------------------------------------------
@@ -1325,6 +1352,368 @@ kernel void kernel_rope_neox_f16(
    }
 }

+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;
+    }
+}
+
 //------------------------------------------------------------------------------
 // cpy
 //------------------------------------------------------------------------------
@@ -0,0 +1,146 @@
+#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];
+    }
+}
@@ -26,6 +26,10 @@
 #  include <unistd.h>
 #endif
 #include <cstring>
+#include <fstream>
+#include <filesystem>
+
+namespace fs = std::filesystem;

 #ifdef _WIN32
 typedef SOCKET sockfd_t;
@@ -80,6 +84,7 @@ enum rpc_cmd {
    RPC_CMD_FREE_BUFFER,
    RPC_CMD_BUFFER_CLEAR,
    RPC_CMD_SET_TENSOR,
+    RPC_CMD_SET_TENSOR_HASH,
    RPC_CMD_GET_TENSOR,
    RPC_CMD_COPY_TENSOR,
    RPC_CMD_GRAPH_COMPUTE,
@@ -89,6 +94,9 @@ enum rpc_cmd {
    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_get_alloc_size_req {
    rpc_tensor tensor;
 };
@@ -135,6 +143,10 @@ struct rpc_msg_buffer_clear_req {
    uint8_t value;
 };

+struct rpc_msg_set_tensor_hash_rsp {
+    uint8_t result;
+};
+
 struct rpc_msg_get_tensor_req {
    rpc_tensor tensor;
    uint64_t offset;
@@ -187,6 +199,18 @@ struct ggml_backend_rpc_buffer_context {

 // RPC helper functions

+// Computes FNV-1a hash of the data
+static uint64_t fnv_hash(const uint8_t * data, size_t len) {
+    const uint64_t fnv_prime = 0x100000001b3ULL;
+    uint64_t hash = 0xcbf29ce484222325ULL;
+
+    for (size_t i = 0; i < len; ++i) {
+        hash ^= data[i];
+        hash *= fnv_prime;
+    }
+    return hash;
+}
+
 static std::shared_ptr<socket_t> make_socket(sockfd_t fd) {
 #ifdef _WIN32
    if (fd == INVALID_SOCKET) {
@@ -483,10 +507,26 @@ static enum ggml_status ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_

 static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
    ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
-    // input serialization format: | rpc_tensor | offset (8 bytes) | data (size bytes) |
+    rpc_tensor rpc_tensor = serialize_tensor(tensor);
+    if (size > HASH_THRESHOLD) {
+        // input serialization format: | rpc_tensor | offset (8 bytes) | hash (8 bytes)
+        size_t input_size = sizeof(rpc_tensor) + sizeof(uint64_t) + sizeof(uint64_t);
+        std::vector<uint8_t> input(input_size, 0);
+        uint64_t hash = fnv_hash((const uint8_t*)data, size);
+        memcpy(input.data(), &rpc_tensor, sizeof(rpc_tensor));
+        memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset));
+        memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), &hash, sizeof(hash));
+        rpc_msg_set_tensor_hash_rsp response;
+        bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR_HASH, input.data(), input.size(), &response, sizeof(response));
+        GGML_ASSERT(status);
+        if (response.result) {
+            // the server has the same data, no need to send it
+            return;
+        }
+    }
+    // input serialization format: | rpc_tensor | offset (8 bytes) | data (size bytes)
    size_t input_size = sizeof(rpc_tensor) + sizeof(uint64_t) + size;
    std::vector<uint8_t> input(input_size, 0);
-    rpc_tensor rpc_tensor = serialize_tensor(tensor);
    memcpy(input.data(), &rpc_tensor, sizeof(rpc_tensor));
    memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset));
    memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), data, size);
@@ -772,7 +812,9 @@ void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, si

 class rpc_server {
 public:
-    rpc_server(ggml_backend_t backend) : backend(backend) {}
+    rpc_server(ggml_backend_t backend, const char * cache_dir)
+        : backend(backend), cache_dir(cache_dir) {
+    }
    ~rpc_server();

    void alloc_buffer(const rpc_msg_alloc_buffer_req & request, rpc_msg_alloc_buffer_rsp & response);
@@ -782,6 +824,7 @@ public:
    bool free_buffer(const rpc_msg_free_buffer_req & request);
    bool buffer_clear(const rpc_msg_buffer_clear_req & request);
    bool set_tensor(const std::vector<uint8_t> & input);
+    bool set_tensor_hash(const std::vector<uint8_t> & input, rpc_msg_set_tensor_hash_rsp & response);
    bool get_tensor(const rpc_msg_get_tensor_req & request, std::vector<uint8_t> & response);
    bool copy_tensor(const rpc_msg_copy_tensor_req & request, rpc_msg_copy_tensor_rsp & response);
    bool graph_compute(const std::vector<uint8_t> & input, rpc_msg_graph_compute_rsp & response);
@@ -789,6 +832,7 @@ public:
    bool get_alloc_size(const rpc_msg_get_alloc_size_req & request, rpc_msg_get_alloc_size_rsp & response);

 private:
+    bool get_cached_file(uint64_t hash, std::vector<uint8_t> & data);
    ggml_tensor * deserialize_tensor(struct ggml_context * ctx, const rpc_tensor * tensor);
    ggml_tensor * create_node(uint64_t id,
                              struct ggml_context * ctx,
@@ -797,6 +841,7 @@ private:


    ggml_backend_t backend;
+    const char * cache_dir;
    std::unordered_set<ggml_backend_buffer_t> buffers;
 };

@@ -960,11 +1005,85 @@ bool rpc_server::set_tensor(const std::vector<uint8_t> & input) {
    }

    const void * data = input.data() + sizeof(rpc_tensor) + sizeof(offset);
+    if (cache_dir && size > HASH_THRESHOLD) {
+        uint64_t hash = fnv_hash((const uint8_t*)data, size);
+        char hash_str[17];
+        snprintf(hash_str, sizeof(hash_str), "%016" PRIx64, hash);
+        // save to cache_dir/hash_str
+        fs::path cache_file = fs::path(cache_dir) / hash_str;
+        std::ofstream ofs(cache_file, std::ios::binary);
+        ofs.write((const char *)data, size);
+        printf("[%s] saved to '%s'\n", __func__, cache_file.c_str());
+    }
    ggml_backend_tensor_set(tensor, data, offset, size);
    ggml_free(ctx);
    return true;
 }

+bool rpc_server::get_cached_file(uint64_t hash, std::vector<uint8_t> & data) {
+    if (!cache_dir) {
+        return false;
+    }
+    char hash_str[17];
+    snprintf(hash_str, sizeof(hash_str), "%016" PRIx64, hash);
+    fs::path cache_file = fs::path(cache_dir) / hash_str;
+    if (!fs::exists(cache_file)) {
+        return false;
+    }
+    std::ifstream ifs(cache_file, std::ios::binary);
+    ifs.seekg(0, std::ios::end);
+    size_t size = ifs.tellg();
+    ifs.seekg(0, std::ios::beg);
+    data.resize(size);
+    ifs.read((char *)data.data(), size);
+    return true;
+}
+
+bool rpc_server::set_tensor_hash(const std::vector<uint8_t> & input, rpc_msg_set_tensor_hash_rsp & response)
+{
+    // serialization format: | rpc_tensor | offset (8 bytes) | hash (8 bytes) |
+    if (input.size() != sizeof(rpc_tensor) + 16) {
+        return false;
+    }
+    const rpc_tensor * in_tensor = (const rpc_tensor *)input.data();
+    uint64_t offset;
+    memcpy(&offset, input.data() + sizeof(rpc_tensor), sizeof(offset));
+    const uint64_t * hash = (const uint64_t *)(input.data() + sizeof(rpc_tensor) + sizeof(offset));
+    std::vector<uint8_t> cached_file;
+    if (!get_cached_file(*hash, cached_file)) {
+        response.result = 0;
+        return true;
+    }
+    size_t size = cached_file.size();
+    struct ggml_init_params params {
+        /*.mem_size   =*/ ggml_tensor_overhead(),
+        /*.mem_buffer =*/ NULL,
+        /*.no_alloc   =*/ true,
+    };
+    struct ggml_context * ctx = ggml_init(params);
+    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);
+
+    // sanitize tensor->data
+    {
+        const size_t p0 = (size_t) ggml_backend_buffer_get_base(tensor->buffer);
+        const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer);
+
+        if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) {
+            GGML_ABORT("[%s] tensor->data out of bounds\n", __func__);
+        }
+    }
+    ggml_backend_tensor_set(tensor, cached_file.data(), offset, size);
+    response.result = 1;
+    ggml_free(ctx);
+    return true;
+}
+
 bool rpc_server::init_tensor(const rpc_msg_init_tensor_req & request) {
    struct ggml_init_params params {
        /*.mem_size   =*/ ggml_tensor_overhead(),
@@ -1148,8 +1267,9 @@ rpc_server::~rpc_server() {
    }
 }

-static void rpc_serve_client(ggml_backend_t backend, sockfd_t sockfd, size_t free_mem, size_t total_mem) {
-    rpc_server server(backend);
+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);
    while (true) {
        uint8_t cmd;
        if (!recv_data(sockfd, &cmd, 1)) {
@@ -1260,6 +1380,20 @@ static void rpc_serve_client(ggml_backend_t backend, sockfd_t sockfd, size_t fre
                }
                break;
            }
+            case RPC_CMD_SET_TENSOR_HASH: {
+                std::vector<uint8_t> input;
+                if (!recv_msg(sockfd, input)) {
+                    return;
+                }
+                rpc_msg_set_tensor_hash_rsp response;
+                if (!server.set_tensor_hash(input, response)) {
+                    return;
+                }
+                if (!send_msg(sockfd, &response, sizeof(response))) {
+                    return;
+                }
+                break;
+            }
            case RPC_CMD_INIT_TENSOR: {
                rpc_msg_init_tensor_req request;
                if (!recv_msg(sockfd, &request,sizeof(request))) {
@@ -1335,7 +1469,9 @@ static void rpc_serve_client(ggml_backend_t backend, sockfd_t sockfd, size_t fre
    }
 }

-void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem) {
+void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint,
+                                   const char * cache_dir,
+                                   size_t free_mem, size_t total_mem) {
    std::string host;
    int port;
    if (!parse_endpoint(endpoint, host, port)) {
@@ -1364,7 +1500,7 @@ void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint
        }
        printf("Accepted client connection, free_mem=%zu, total_mem=%zu\n", free_mem, total_mem);
        fflush(stdout);
-        rpc_serve_client(backend, client_socket->fd, free_mem, total_mem);
+        rpc_serve_client(backend, cache_dir, client_socket->fd, free_mem, total_mem);
        printf("Client connection closed\n");
        fflush(stdout);
    }
@@ -23,6 +23,38 @@ ggml_add_backend_library(ggml-sycl
                         ../../include/ggml-sycl.h
                        )

+find_package(DNNL)
+set(GGML_SYCL_DNNL 0)
+if(DNNL_FOUND)
+    if (DEFINED ENV{ONEAPI_ROOT} AND NOT DEFINED DNNL_GPU_VENDOR)
+        # Assuming oneDNN packaged with oneapi release is used which
+        # supports only intel target
+        set(DNNL_GPU_VENDOR "INTEL")
+        if(NOT "${GGML_SYCL_TARGET}" STREQUAL "INTEL")
+            message(WARNING "oneDNN builds bundled with oneapi release only support INTEL target")
+        endif()
+    endif()
+
+    # Verify oneDNN was compiled for the same target as llama
+    if("${GGML_SYCL_TARGET}" STREQUAL "${DNNL_GPU_VENDOR}")
+        target_link_libraries(ggml-sycl PRIVATE DNNL::dnnl)
+        set(GGML_SYCL_DNNL 1)
+        get_target_property(CONFIGS DNNL::dnnl IMPORTED_CONFIGURATIONS)
+        foreach(CONFIG ${CONFIGS})
+            get_target_property(DNNL_LIB DNNL::dnnl IMPORTED_LOCATION_${CONFIG})
+            message(STATUS "Found oneDNN: ${DNNL_LIB}")
+        endforeach()
+    else()
+        message(WARNING
+            "oneDNN must be compiled for the same target as llama.cpp.
+             llama.cpp: ${GGML_SYCL_TARGET}, oneDNN: ${DNNL_GPU_VENDOR}.
+             Disabling oneDNN support.")
+    endif()
+else()
+    message(STATUS "oneDNN not found, disabling oneDNN support")
+endif()
+target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_DNNL=${GGML_SYCL_DNNL})
+
 if (GGML_SYCL_F16)
    if (GGML_SYCL_TARGET STREQUAL "AMD")
        message(WARNING "AMD target does not entirely support FP16 in the SYCL backend.")
@@ -48,18 +80,6 @@ file(GLOB   GGML_HEADERS_SYCL "*.hpp")
 file(GLOB   GGML_SOURCES_SYCL "*.cpp")
 target_sources(ggml-sycl PRIVATE ${GGML_HEADERS_SYCL} ${GGML_SOURCES_SYCL})

-find_package(DNNL)
-message("-- DNNL found:" ${DNNL_FOUND})
-
-if (GGML_SYCL_TARGET STREQUAL "INTEL")
-    add_compile_definitions(GGML_SYCL_DNNL=${DNNL_FOUND})
-else()
-    add_compile_definitions(GGML_SYCL_DNNL=0)
-endif()
-
-if (${DNNL_FOUND} AND GGML_SYCL_TARGET STREQUAL "INTEL")
-    target_link_libraries(ggml-sycl PRIVATE DNNL::dnnl)
-endif()

 if (WIN32)
    find_package(IntelSYCL REQUIRED)
@@ -170,7 +170,6 @@ static size_t g_scratch_offset = 0;
 int get_current_device_id();

 inline dpct::err0 ggml_sycl_set_device(const int device) try {
-
  int current_device_id;
  SYCL_CHECK(CHECK_TRY_ERROR(current_device_id = get_current_device_id()));

@@ -242,6 +241,14 @@ struct ggml_sycl_pool_alloc {
        }
    }

+    T * realloc(size_t size) {
+        GGML_ASSERT(pool != nullptr);
+        if (ptr)
+            pool->free(ptr, actual_size);
+        ptr = (T *) pool->alloc(size * sizeof(T), &this->actual_size);
+        return ptr;
+    }
+
    // size is in number of elements
    T * alloc(size_t size) {
        GGML_ASSERT(pool != nullptr);
@@ -371,10 +378,29 @@ struct ggml_backend_sycl_context {
    dnnl::stream stream_dnnl() {
        return stream_dnnl(device, 0);
    }
+    dnnl::memory get_scratchpad_mem(const dnnl::memory::desc & scratchpad_md,
+                                    const dnnl::engine & eng, const queue_ptr q) {
+        ggml_sycl_pool_alloc<uint8_t> * pool;
+        auto it = scratchpad_map.find(q);
+        if (it == scratchpad_map.end()) {
+            scratchpad_map[q] = std::make_unique<ggml_sycl_pool_alloc<uint8_t>>(this->pool());
+            pool = scratchpad_map[q].get();
+        } else {
+            pool = it->second.get();
+        }
+
+        size_t scratchpad_size = scratchpad_md.get_size();
+        if (scratchpad_size > pool->actual_size) {
+            pool->realloc(scratchpad_size);
+        }
+        void * mem_ptr = pool->get();
+        return dnnl::memory(scratchpad_md, eng, mem_ptr);
+    }
 #endif

    // pool
    std::unique_ptr<ggml_sycl_pool> pools[GGML_SYCL_MAX_DEVICES];
+    std::unordered_map<sycl::queue *, std::unique_ptr<ggml_sycl_pool_alloc<uint8_t>>> scratchpad_map;

    std::unique_ptr<ggml_sycl_pool> host_pools[GGML_SYCL_MAX_DEVICES];

@@ -13,9 +13,6 @@
 #ifndef GGML_SYCL_GEMM_HPP
 #define GGML_SYCL_GEMM_HPP

-#include <fstream>
-#include <iostream>
-
 #include "ggml-sycl.h"

 #if GGML_SYCL_DNNL
@@ -35,62 +32,34 @@ public:
        else static_assert(0);
    }

-    static inline void row_gemm(sycl::queue& q, bool a_trans,
-        bool b_trans, int m, int n, int k,
-        const void* a, dt at, const void* b, dt bt, void* c, dt ct)
-    {
-        // Get the device associated with the queue
-        sycl::device dev = q.get_device();
-        // Get the context associated with the queue
-        sycl::context ctx = q.get_context();
-        const dnnl::engine eng = dnnl::sycl_interop::make_engine(dev, ctx);
-        const dnnl::stream stream = dnnl::sycl_interop::make_stream(eng, q);
+    static inline void row_gemm(ggml_backend_sycl_context & ctx, bool a_trans, bool b_trans, int m, int n, int k,
+                                const void * a, dt at, const void * b, dt bt, void * c, dt ct, const queue_ptr & q) {
+        auto stream = ctx.stream_dnnl(q);
+        auto eng = ctx.engine_dnnl(q);
        dnnl::memory::dims a_dims = { m, k };
        dnnl::memory::dims b_dims = { k, n };
        dnnl::memory::dims c_dims = { m, n };
        const auto a_in_md = dnnl::memory::desc(a_dims, at, a_trans ? tag::ba : tag::ab);
        const auto b_in_md = dnnl::memory::desc(b_dims, bt, b_trans ? tag::ba : tag::ab);
-        const auto c_md = dnnl::memory::desc(c_dims, ct, tag::ab);
+        const auto c_md    = dnnl::memory::desc(c_dims, ct, tag::ab);
+
+        dnnl::primitive_attr primitive_attr;
+        primitive_attr.set_scratchpad_mode(dnnl::scratchpad_mode::user);
+
        auto a_mem = dnnl::memory(a_in_md, eng, const_cast<void*>(a));
        auto b_mem = dnnl::memory(b_in_md, eng, const_cast<void*>(b));
-        auto matmul_pd = dnnl::matmul::primitive_desc(eng, a_in_md, b_in_md, c_md);
+        auto matmul_pd = dnnl::matmul::primitive_desc(eng, a_in_md, b_in_md, c_md, primitive_attr);
        auto c_mem = dnnl::memory(matmul_pd.dst_desc(), eng, c);

-        // Create the primitive.
+        auto scratchpad_md = matmul_pd.scratchpad_desc();
+        auto scratchpad_mem = ctx.get_scratchpad_mem(scratchpad_md, eng, q);
        auto matmul_prim = dnnl::matmul(matmul_pd);
-        // Primitive arguments.
-        std::unordered_map<int, dnnl::memory> matmul_args;
-        matmul_args.insert({ DNNL_ARG_SRC, a_mem });
-        matmul_args.insert({ DNNL_ARG_WEIGHTS, b_mem });
-        matmul_args.insert({ DNNL_ARG_DST, c_mem });
-
-        matmul_prim.execute(stream, matmul_args);
-    }
-
-
-    static inline void row_gemm(const dnnl::stream& stream, bool a_trans,
-        bool b_trans, int m, int n, int k,
-        const void* a, dt at, const void* b, dt bt, void* c, dt ct)
-    {
-        auto const eng = stream.get_engine();
-        dnnl::memory::dims a_dims = { m, k };
-        dnnl::memory::dims b_dims = { k, n };
-        dnnl::memory::dims c_dims = { m, n };
-        const auto a_in_md = dnnl::memory::desc(a_dims, at, a_trans ? tag::ba : tag::ab);
-        const auto b_in_md = dnnl::memory::desc(b_dims, bt, b_trans ? tag::ba : tag::ab);
-        const auto c_md = dnnl::memory::desc(c_dims, ct, tag::ab);
-        auto a_mem = dnnl::memory(a_in_md, eng, const_cast<void*>(a));
-        auto b_mem = dnnl::memory(b_in_md, eng, const_cast<void*>(b));
-        auto matmul_pd = dnnl::matmul::primitive_desc(eng, a_in_md, b_in_md, c_md);
-        auto c_mem = dnnl::memory(matmul_pd.dst_desc(), eng, c);
-
-        // Create the primitive.
-        auto matmul_prim = dnnl::matmul(matmul_pd);
-        // Primitive arguments.
+
        std::unordered_map<int, dnnl::memory> matmul_args;
        matmul_args.insert({ DNNL_ARG_SRC, a_mem });
        matmul_args.insert({ DNNL_ARG_WEIGHTS, b_mem });
        matmul_args.insert({ DNNL_ARG_DST, c_mem });
+        matmul_args.insert({ DNNL_ARG_SCRATCHPAD, scratchpad_mem });

        matmul_prim.execute(stream, matmul_args);
    }
@@ -37,6 +37,7 @@
 #include "ggml-backend-impl.h"

 #include "ggml-sycl/backend.hpp"
+#include "ggml-sycl/common.hpp"
 #include "ggml-sycl/presets.hpp"
 #include "ggml-sycl/gemm.hpp"
 #include "ggml-sycl/sycl_hw.hpp"
@@ -191,7 +192,7 @@ static void ggml_check_sycl() try {

    if (!initialized) {
        g_ggml_sycl_debug = get_sycl_env("GGML_SYCL_DEBUG", 0);
-        g_ggml_sycl_disable_optimize= get_sycl_env("GGML_SYCL_DISABLE_OPT", 0);
+        g_ggml_sycl_disable_optimize= get_sycl_env("GGML_SYCL_DISABLE_OPT", 1);
        g_ggml_sycl_disable_graph = get_sycl_env("GGML_SYCL_DISABLE_GRAPH", 1);
        GGML_SYCL_DEBUG("[SYCL] call ggml_check_sycl\n");
        GGML_LOG_INFO("Running with Environment Variables:\n");
@@ -490,6 +491,23 @@ catch (sycl::exception const &exc) {
  std::exit(1);
 }

+static void ggml_backend_sycl_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value,
+                                                   size_t offset, size_t size) {
+    GGML_SYCL_DEBUG(" [SYCL] call %s\n", __func__);
+    ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *) buffer->context;
+    SYCL_CHECK(ggml_sycl_set_device(ctx->device));
+    auto stream = &(dpct::dev_mgr::instance().get_device(ctx->device).default_queue());
+    if (size == 0) {
+        return;  // Nothing to do
+    }
+    if (tensor->data == nullptr) {
+        GGML_ABORT("Error: Tensor data pointer is null.\n");
+    }
+    void * target_ptr = static_cast<char *>(tensor->data) + offset;
+    SYCL_CHECK(CHECK_TRY_ERROR((*stream).memset(target_ptr, value, size)));
+    SYCL_CHECK(CHECK_TRY_ERROR((*stream).wait()));
+}
+
 static void ggml_backend_sycl_buffer_reset(ggml_backend_buffer_t buffer) {
    GGML_SYCL_DEBUG("[SYCL] call %s\n", __func__);
    if (buffer == nullptr) {
@@ -510,7 +528,7 @@ static const ggml_backend_buffer_i ggml_backend_sycl_buffer_interface = {
    /* .free_buffer     = */ ggml_backend_sycl_buffer_free_buffer,
    /* .get_base        = */ ggml_backend_sycl_buffer_get_base,
    /* .init_tensor     = */ ggml_backend_sycl_buffer_init_tensor,
-    /* .memset_tensor   = */ NULL,
+    /* .memset_tensor   = */ ggml_backend_sycl_buffer_memset_tensor,
    /* .set_tensor      = */ ggml_backend_sycl_buffer_set_tensor,
    /* .get_tensor      = */ ggml_backend_sycl_buffer_get_tensor,
    /* .cpy_tensor      = */ ggml_backend_sycl_buffer_cpy_tensor,
@@ -2058,9 +2076,9 @@ inline void ggml_sycl_op_mul_mat_sycl(
        const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16, dst);
        to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream);
 #else
-        auto dnnl_stream = ctx.stream_dnnl(stream);
-        DnnlGemmWrapper::row_gemm(dnnl_stream, false, true, src1_ncols, row_diff, ne10, src1_ptr, DnnlGemmWrapper::to_dt<sycl::half>(),
-            src0_ptr, DnnlGemmWrapper::to_dt<sycl::half>(), dst_f16.get(), DnnlGemmWrapper::to_dt<sycl::half>());
+        DnnlGemmWrapper::row_gemm(ctx, false, true, src1_ncols, row_diff, ne10, src1_ptr,
+                                  DnnlGemmWrapper::to_dt<sycl::half>(), src0_ptr, DnnlGemmWrapper::to_dt<sycl::half>(),
+                                  dst_f16.get(), DnnlGemmWrapper::to_dt<sycl::half>(), stream);
        const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16, dst);
        to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff* src1_ncols, stream);
 #endif
@@ -2099,9 +2117,9 @@ inline void ggml_sycl_op_mul_mat_sycl(
            dst_dd_i, ldc)));
 #    endif
 #else
-        auto dnnl_stream = ctx.stream_dnnl(stream);
-         DnnlGemmWrapper::row_gemm(dnnl_stream, false, true, src1_ncols, row_diff, ne10, src1_ddf1_i, DnnlGemmWrapper::to_dt<float>(),
-            src0_ddf_i, DnnlGemmWrapper::to_dt<float>(), dst_dd_i, DnnlGemmWrapper::to_dt<float>());
+        DnnlGemmWrapper::row_gemm(ctx, false, true, src1_ncols, row_diff, ne10, src1_ddf1_i,
+                                  DnnlGemmWrapper::to_dt<float>(), src0_ddf_i, DnnlGemmWrapper::to_dt<float>(),
+                                  dst_dd_i, DnnlGemmWrapper::to_dt<float>(), stream);
 #endif
    }
    GGML_UNUSED(dst);
@@ -23,32 +23,40 @@ if (Vulkan_FOUND)
                             ../../include/ggml-vulkan.h
                            )

-    # Compile a test shader to determine whether GL_KHR_cooperative_matrix is supported.
-    # If it's not, there will be an error to stderr.
-    # If it's supported, set a define to indicate that we should compile those shaders
-    execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_coopmat_support.comp"
-                    OUTPUT_VARIABLE glslc_output
-                    ERROR_VARIABLE glslc_error)
+    if(NOT DEFINED GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
+        # Compile a test shader to determine whether GL_KHR_cooperative_matrix is supported.
+        # If it's not, there will be an error to stderr.
+        # If it's supported, set a define to indicate that we should compile those shaders
+        execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_coopmat_support.comp"
+                        OUTPUT_VARIABLE glslc_output
+                        ERROR_VARIABLE glslc_error)

-    if (${glslc_error} MATCHES ".*extension not supported: GL_KHR_cooperative_matrix.*")
-        message(STATUS "GL_KHR_cooperative_matrix not supported by glslc")
-    else()
-        message(STATUS "GL_KHR_cooperative_matrix supported by glslc")
-        add_compile_definitions(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
+        if (${glslc_error} MATCHES ".*extension not supported: GL_KHR_cooperative_matrix.*")
+            message(STATUS "GL_KHR_cooperative_matrix not supported by glslc")
+            set(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT OFF CACHE INTERNAL "Whether coopmat is supported by glslc")
+        else()
+            message(STATUS "GL_KHR_cooperative_matrix supported by glslc")
+            add_compile_definitions(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
+            set(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT ON CACHE INTERNAL "Whether coopmat is supported by glslc")
+        endif()
    endif()

-    # Compile a test shader to determine whether GL_NV_cooperative_matrix2 is supported.
-    # If it's not, there will be an error to stderr.
-    # If it's supported, set a define to indicate that we should compile those shaders
-    execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_coopmat2_support.comp"
-                    OUTPUT_VARIABLE glslc_output
-                    ERROR_VARIABLE glslc_error)
+    if(NOT DEFINED GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+        # Compile a test shader to determine whether GL_NV_cooperative_matrix2 is supported.
+        # If it's not, there will be an error to stderr.
+        # If it's supported, set a define to indicate that we should compile those shaders
+        execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_coopmat2_support.comp"
+                        OUTPUT_VARIABLE glslc_output
+                        ERROR_VARIABLE glslc_error)

-    if (${glslc_error} MATCHES ".*extension not supported: GL_NV_cooperative_matrix2.*")
-        message(STATUS "GL_NV_cooperative_matrix2 not supported by glslc")
-    else()
-        message(STATUS "GL_NV_cooperative_matrix2 supported by glslc")
-        add_compile_definitions(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+        if (${glslc_error} MATCHES ".*extension not supported: GL_NV_cooperative_matrix2.*")
+            message(STATUS "GL_NV_cooperative_matrix2 not supported by glslc")
+            set(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT OFF CACHE INTERNAL "Whether coopmat2 is supported by glslc")
+        else()
+            message(STATUS "GL_NV_cooperative_matrix2 supported by glslc")
+            add_compile_definitions(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+            set(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT ON CACHE INTERNAL "Whether coopmat2 is supported by glslc")
+        endif()
    endif()

    target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan)
@@ -119,6 +127,8 @@ if (Vulkan_FOUND)
            SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders
            CMAKE_ARGS -DCMAKE_TOOLCHAIN_FILE=${HOST_CMAKE_TOOLCHAIN_FILE}
                    -DCMAKE_INSTALL_PREFIX=${CMAKE_BINARY_DIR}
+                    -DGGML_VULKAN_COOPMAT_GLSLC_SUPPORT=${GGML_VULKAN_COOPMAT_GLSLC_SUPPORT}
+                    -DGGML_VULKAN_COOPMAT2_GLSLC_SUPPORT=${GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT}
            BUILD_COMMAND ${CMAKE_COMMAND} --build .
            INSTALL_COMMAND ${CMAKE_COMMAND} --install .
            INSTALL_DIR ${CMAKE_BINARY_DIR}
@@ -149,6 +149,7 @@ class vk_perf_logger;
 static void ggml_vk_destroy_buffer(vk_buffer& buf);

 static constexpr uint32_t mul_mat_vec_max_cols = 8;
+static constexpr uint32_t p021_max_gqa_ratio = 8;

 enum vk_device_architecture {
    OTHER,
@@ -231,6 +232,7 @@ struct vk_device_struct {
    bool uma;
    bool prefer_host_memory;
    bool float_controls_rte_fp16;
+    bool subgroup_add;

    bool subgroup_size_control;
    uint32_t subgroup_min_size;
@@ -277,7 +279,7 @@ struct vk_device_struct {
    vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_COUNT][mul_mat_vec_max_cols];
    vk_pipeline pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_COUNT];

-    vk_pipeline pipeline_mul_mat_vec_p021_f16_f32;
+    vk_pipeline pipeline_mul_mat_vec_p021_f16_f32[p021_max_gqa_ratio];
    vk_pipeline pipeline_mul_mat_vec_nc_f16_f32;
    vk_pipeline pipeline_get_rows[GGML_TYPE_COUNT];
    vk_pipeline pipeline_get_rows_f32[GGML_TYPE_COUNT];
@@ -2265,7 +2267,13 @@ static void ggml_vk_load_shaders(vk_device& device) {

    ggml_vk_create_pipeline(device, device->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256 * 4, 1, 1}, {}, 1);

-    ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32, "mul_mat_vec_p021_f16_f32", mul_mat_vec_p021_f16_f32_len, mul_mat_vec_p021_f16_f32_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {}, 1);
+    for (uint32_t i = 0; i < p021_max_gqa_ratio; ++i) {
+        if (device->subgroup_add && device->subgroup_require_full_support) {
+            ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_subgroup_add_len, mul_mat_vec_p021_f16_f32_subgroup_add_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true, true);
+        } else {
+            ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_len,              mul_mat_vec_p021_f16_f32_data,              "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true);
+        }
+    }
    ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 7 * sizeof(uint32_t), {1, 1, 1}, {}, 1);

    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);
@@ -2281,13 +2289,21 @@ static void ggml_vk_load_shaders(vk_device& device) {
    ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f32, "contig_cpy_f32_f32", contig_cpy_f32_f32_len, contig_cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f16, "contig_cpy_f32_f16", contig_cpy_f32_f16_len, contig_cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f16_f16, "contig_cpy_f16_f16", contig_cpy_f16_f16_len, contig_cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
-
-    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_len, cpy_f32_q4_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_len, cpy_f32_q4_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_len, cpy_f32_q5_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_0), 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_len, cpy_f32_q5_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_1), 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_len, cpy_f32_q8_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_len, cpy_f32_iq4_nl_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1);
+    if (device->float_controls_rte_fp16) {
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_rte_len, cpy_f32_q4_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_rte_len, cpy_f32_q4_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_rte_len, cpy_f32_q5_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_0), 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_rte_len, cpy_f32_q5_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_1), 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_rte_len, cpy_f32_q8_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_rte_len, cpy_f32_iq4_nl_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1);
+    } else {
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_len, cpy_f32_q4_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_len, cpy_f32_q4_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_len, cpy_f32_q5_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_0), 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_len, cpy_f32_q5_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_1), 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_len, cpy_f32_q8_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_len, cpy_f32_iq4_nl_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1);
+    }

    ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q4_0], "cpy_q4_0_f32", cpy_q4_0_f32_len, cpy_q4_0_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q4_1], "cpy_q4_1_f32", cpy_q4_1_f32_len, cpy_q4_1_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1);
@@ -2471,13 +2487,15 @@ static vk_device ggml_vk_get_device(size_t idx) {
        vk::PhysicalDeviceDriverProperties driver_props;
        vk::PhysicalDeviceShaderSMBuiltinsPropertiesNV sm_props;
        vk::PhysicalDeviceShaderCoreProperties2AMD amd_shader_core_properties2_props;
+        vk::PhysicalDeviceVulkan11Properties vk11_props;
        vk::PhysicalDeviceVulkan12Properties vk12_props;
        vk::PhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_size_control_props;

        props2.pNext = &props3;
        props3.pNext = &subgroup_props;
        subgroup_props.pNext = &driver_props;
-        driver_props.pNext = &vk12_props;
+        driver_props.pNext = &vk11_props;
+        vk11_props.pNext = &vk12_props;

        VkBaseOutStructure * last_struct = (VkBaseOutStructure *)&vk12_props;

@@ -2541,6 +2559,9 @@ static vk_device ggml_vk_get_device(size_t idx) {
        }
        device->float_controls_rte_fp16 = vk12_props.shaderRoundingModeRTEFloat16;

+        device->subgroup_add = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
+                               (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eArithmetic);
+
        const bool force_disable_f16 = getenv("GGML_VK_DISABLE_F16") != nullptr;

        device->fp16 = !force_disable_f16 && fp16_storage && fp16_compute;
@@ -4627,9 +4648,15 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
    const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
    const uint64_t d_sz = sizeof(float) * d_ne;

+    // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    uint32_t gqa_ratio = (uint32_t)ne12 / (uint32_t)ne02;
+    if (gqa_ratio > 8 || gqa_ratio == 0 || ne12 != ne02 * gqa_ratio) {
+        gqa_ratio = 1;
+    }
+
    if (dryrun) {
        // Request descriptor sets
-        ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_mul_mat_vec_p021_f16_f32, 1);
+        ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1], 1);
        return;
    }

@@ -4653,8 +4680,15 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c

    // compute
    const std::array<uint32_t, 6> pc = { (uint32_t)ne00, (uint32_t)ne01, (uint32_t)ne02, (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
+
+    uint32_t workgroups_z = (uint32_t)ne12;
+    // When gqa_ratio > 1, each invocation does multiple rows and we can launch fewer workgroups
+    if (gqa_ratio > 1) {
+        workgroups_z /= gqa_ratio;
+    }
+
    ggml_vk_sync_buffers(subctx);
-    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, 6 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
+    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1], { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, 6 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, workgroups_z });
 }

 static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -8436,8 +8470,12 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
    VK_LOG_DEBUG("ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)");
    ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;

+    uint64_t total_mat_mul_bytes = 0;
    for (int i = 0; i < cgraph->n_nodes; i++) {
        ggml_vk_build_graph(ctx, cgraph->nodes[i], i, nullptr, 0, true, false, false);
+        if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) {
+            total_mat_mul_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]);
+        }
    }
    if (ctx->device->need_compiles) {
        ggml_vk_load_shaders(ctx->device);
@@ -8458,17 +8496,27 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
    bool first_node_in_batch = true; // true if next node will be first node in a batch
    int submit_node_idx = 0; // index to first node in a batch

-    // Submit work every nodes_per_submit nodes to overlap CPU cmdbuffer generation with GPU execution.
-    // Start with a smaller count to get work submitted right away, and increase it after each submit.
-    int nodes_per_submit = 20;
+    // Submit after enough work has accumulated, to overlap CPU cmdbuffer generation with GPU execution.
+    // Estimate the amount of matmul work by looking at the weight matrix size, and submit every 100MB
+    // (and scaled down based on model size, so smaller models submit earlier).
+    // Also submit at least every 100 nodes, in case there are workloads without as much matmul.
+    int nodes_per_submit = 100;
    int submitted_nodes = 0;
    int submit_count = 0;
+    uint64_t mul_mat_bytes = 0;
+    uint64_t mul_mat_bytes_per_submit = std::min(uint64_t(100*1000*1000), total_mat_mul_bytes / 40u);
    for (int i = 0; i < cgraph->n_nodes; i++) {
        if (first_node_in_batch) {
            submit_node_idx = i;
        }

-        bool submit = (submitted_nodes >= nodes_per_submit) || (i == last_node);
+        if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) {
+            mul_mat_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]);
+        }
+
+        bool submit = (submitted_nodes >= nodes_per_submit) ||
+                      (mul_mat_bytes >= mul_mat_bytes_per_submit) ||
+                      (i == last_node);

        bool enqueued = ggml_vk_build_graph(ctx, cgraph->nodes[i], i, cgraph->nodes[submit_node_idx], submit_node_idx, false, i == last_node, submit);

@@ -8485,13 +8533,9 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
        if (submit) {
            first_node_in_batch = true;
            submitted_nodes = 0;
-            switch (submit_count) {
-            case 0:
-                nodes_per_submit = 50;
-                break;
-            default:
-                nodes_per_submit = 100;
-                break;
+            mul_mat_bytes = 0;
+            if (submit_count < 3) {
+                mul_mat_bytes_per_submit *= 2;
            }
            submit_count++;
        }
@@ -8720,6 +8764,10 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                default:
                    return false;
                }
+                if (op->src[1]->ne[0] != op->src[2]->ne[0]) {
+                    // different head sizes of K and V are not supported yet
+                    return false;
+                }
                if (op->src[0]->type != GGML_TYPE_F32) {
                    return false;
                }
@@ -1,5 +1,11 @@
 find_package (Threads REQUIRED)

+if (GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
+    add_compile_definitions(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
+endif()
+if (GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+    add_compile_definitions(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+endif()
 set(TARGET vulkan-shaders-gen)
 add_executable(${TARGET} vulkan-shaders-gen.cpp)
 install(TARGETS ${TARGET} RUNTIME)
@@ -1,5 +1,10 @@
 #version 450

+#if RTE16
+#extension GL_EXT_spirv_intrinsics : enable
+spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits
+#endif // RTE16
+
 #include "types.comp"
 #include "generic_unary_head.comp"

@@ -82,8 +82,8 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
    return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1]));
 }
 vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
-    const i8vec2 v0 = unpack8(data_a_packed16[a_offset + ib].qs[iqs/2]);
-    const i8vec2 v1 = unpack8(data_a_packed16[a_offset + ib].qs[iqs/2 + 1]);
+    const i8vec2 v0 = unpack8(int32_t(data_a_packed16[a_offset + ib].qs[iqs/2])).xy; // vec4 used due to #12147
+    const i8vec2 v1 = unpack8(int32_t(data_a_packed16[a_offset + ib].qs[iqs/2 + 1])).xy;
    return vec4(v0.x, v0.y, v1.x, v1.y);
 }
 #endif
@@ -311,8 +311,8 @@ float16_t dequantFuncIQ1_S(const in decodeBufIQ1_S bl, const in uint blockCoords
    const float16_t d = bl.block.d;
    const uint idx = coordInBlock[1];

-    const uint ib32 = idx / 32;
-    const uint ib8 = idx / 8;
+    const uint ib32 = (idx & 0xE0) >> 5;
+    const uint ib8 = (idx & 0xF8) >> 3;

    const uint qh = bl.block.qh[ib32];
    const uint qs = bl.block.qs[ib8];
@@ -330,14 +330,20 @@ layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ1
   block_iq1_m block;
 };

+layout(buffer_reference, std430, buffer_reference_align = 8) buffer decodeBufIQ1_M_packed64 {
+   block_iq1_m_packed64 block;
+};
+
 float16_t dequantFuncIQ1_M(const in decodeBufIQ1_M bl, const in uint blockCoords[2], const in uint coordInBlock[2])
 {
-    const u16vec4 scales = u16vec4(bl.block.scales[0], bl.block.scales[1], bl.block.scales[2], bl.block.scales[3]) >> 12;
-    const float16_t d = uint16BitsToHalf(scales.x | (scales.y << 4) | (scales.z << 8) | (scales.w << 12));
+    decodeBufIQ1_M_packed64 bl64 = decodeBufIQ1_M_packed64(bl);
    const uint idx = coordInBlock[1];

-    const uint ib8 = idx / 8;
-    const uint ib16 = idx / 16;
+    uvec2 scales = unpack32(bl64.block.scales);
+    const float16_t d = uint16BitsToHalf(uint16_t(((scales.x & 0xF000) >> 12) | ((scales.x & 0xF0000000) >> 24) | ((scales.y & 0xF000) >> 4) | ((scales.y & 0xF0000000) >> 16)));
+
+    const uint ib8 = (idx & 0xF8) >> 3;
+    const uint ib16 = (idx & 0xF0) >> 4;
    const int i8 = int(idx % 8);
    const uint sc = bl.block.scales[ib8 / 8];
    const uint qs = bl.block.qs[ib8];
@@ -105,6 +105,16 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
    int unroll_count = 4;
    uint unrolled_iters = num_iters & ~(unroll_count - 1);

+#if K_PER_ITER == 2
+    // If the K dimension is odd, we need lastiter==true on the last iteration
+    // so OOB is computed correctly. Skip some unrolling to make that happen.
+    if ((p.ncols & 1) != 0 &&
+        unrolled_iters == num_iters &&
+        unrolled_iters > 0) {
+        unrolled_iters -= unroll_count;
+    }
+#endif
+
    uint i = 0;
    while (i < unrolled_iters) {
        // Manually partially unroll the loop
@@ -113,8 +123,18 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
            i++;
        }
    }
+
    unroll_count = 2;
    unrolled_iters = num_iters & ~(unroll_count - 1);
+
+#if K_PER_ITER == 2
+    if ((p.ncols & 1) != 0 &&
+        unrolled_iters == num_iters &&
+        unrolled_iters > 0) {
+        unrolled_iters -= unroll_count;
+    }
+#endif
+
    while (i < unrolled_iters) {
        // Manually partially unroll the loop
        [[unroll]] for (uint k = 0; k < unroll_count; ++k) {
@@ -19,8 +19,8 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
        const float db = d * (0.5 + scale) * 0.25;

        const uint qh = data_a[ibi].qh[ib32];
-        const u8vec2 qs16 = unpack8(data_a_packed16[ibi].qs[itid]);
-        const u8vec2 sign16 = unpack8(data_a_packed16[ibi].qs[QUANT_K / 16 + itid]);
+        const u8vec2 qs16 = unpack8(uint32_t(data_a_packed16[ibi].qs[itid])).xy; // vec4 used due to #12147
+        const u8vec2 sign16 = unpack8(uint32_t(data_a_packed16[ibi].qs[QUANT_K / 16 + itid])).xy;
        [[unroll]] for (uint l = 0; l < 2; ++l) {
            const uint8_t sign = sign16[l];
            const uint qs = qs16[l] | ((qh << (8 - nibble_shift - 2 * l)) & 0x300);
@@ -21,7 +21,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32,
            sum[j] = 0.0;
        }
        [[unroll]] for (uint l = 0; l < 4; ++l) {
-            const u8vec2 qs = unpack8(data_a_packed16[ibi].qs[4 * ib32 + l]);
+            const u8vec2 qs = unpack8(uint32_t(data_a_packed16[ibi].qs[4 * ib32 + l])).xy; // vec4 used due to #12147
            const uint sign = data_a[ibi].signs[4 * ib32 + l];
            const vec4 grid0 = vec4(unpack8(iq3s_grid[qs.x | ((qh << (8 - 2*l)) & 0x100)]));
            const vec4 grid1 = vec4(unpack8(iq3s_grid[qs.y | ((qh << (7 - 2*l)) & 0x100)]));
@@ -12,6 +12,9 @@ layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
 layout (binding = 2) writeonly buffer D {D_TYPE dst[];};

+layout (binding = 0) readonly buffer AV4 {A_TYPE_VEC4 data_a_v4[];};
+layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
+
 layout (push_constant) uniform parameter
 {
    uint ncols_x;
@@ -37,25 +40,66 @@ void main() {

    const uint idst = channel*nrows_dst + row_dst;

-    tmp[tid] = 0.0f;
+    FLOAT_TYPE temp = 0.0f;

-    for (uint col_x0 = 0; col_x0 < p.ncols_x; col_x0 += BLOCK_SIZE) {
-        const uint col_x = col_x0 + tid;
+    // Detect alignment for vector loads
+    bool is_aligned = (p.ncols_x % 4) == 0 && (p.row_stride_x % 4) == 0 && (p.channel_stride_x % 4) == 0;

-        if (col_x >= p.ncols_x) {
-            break;
+    for (uint col_x0 = 0; col_x0 < p.ncols_x;) {
+
+        // Unroll 2x and do vec4 loads if aligned
+        const uint unroll_count = 2;
+        if (col_x0 + unroll_count * 4 * BLOCK_SIZE <= p.ncols_x && is_aligned) {
+            [[unroll]] for (uint i = 0; i < unroll_count; ++i) {
+                const uint col_x = col_x0 + 4*tid;
+
+                const uint row_y = col_x;
+
+                const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
+                const uint iy = channel*nrows_y + row_y;
+
+                const vec4 av4 = vec4(data_a_v4[ix / 4]);
+                const vec4 bv4 = vec4(data_b_v4[iy / 4]);
+
+                temp += dot(av4, bv4);
+
+                col_x0 += 4*BLOCK_SIZE;
+            }
+        // do vec4 loads if aligned
+        } else if (col_x0 + 4*BLOCK_SIZE <= p.ncols_x && is_aligned) {
+            const uint col_x = col_x0 + 4*tid;
+
+            const uint row_y = col_x;
+
+            const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
+            const uint iy = channel*nrows_y + row_y;
+
+            const vec4 av4 = vec4(data_a_v4[ix / 4]);
+            const vec4 bv4 = vec4(data_b_v4[iy / 4]);
+
+            temp += dot(av4, bv4);
+
+            col_x0 += 4*BLOCK_SIZE;
+        } else {
+            const uint col_x = col_x0 + tid;
+            if (col_x >= p.ncols_x) {
+                break;
+            }
+
+            const uint row_y = col_x;
+
+            const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
+            const uint iy = channel*nrows_y + row_y;
+
+            const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);
+
+            temp = fma(xi, FLOAT_TYPE(data_b[iy]), temp);
+            col_x0 += BLOCK_SIZE;
        }
-
-        const uint row_y = col_x;
-
-        const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
-        const uint iy = channel*nrows_y + row_y;
-
-        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);
-
-        tmp[tid] = fma(xi, FLOAT_TYPE(data_b[iy]), tmp[tid]);
    }

+    tmp[tid] = temp;
+
    // sum up partial sums and write back result
    barrier();
    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
@@ -2,16 +2,25 @@

 #extension GL_EXT_control_flow_attributes : enable
 #extension GL_EXT_shader_16bit_storage : require
+#if USE_SUBGROUP_ADD
+#extension GL_KHR_shader_subgroup_arithmetic : enable
+#endif

-#define BLOCK_SIZE 32
 #define FLOAT_TYPE float

-layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;

 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
 layout (binding = 2) writeonly buffer D {D_TYPE dst[];};

+layout (binding = 0) readonly buffer AV4 {A_TYPE_VEC4 data_a_v4[];};
+layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
+
+layout(constant_id = 0) const int BLOCK_SIZE = 32;
+// gqa_ratio is in the range [1,8]
+layout(constant_id = 1) const uint gqa_ratio = 1;
+
 layout (push_constant) uniform parameter
 {
    uint ncols_x;
@@ -22,52 +31,124 @@ layout (push_constant) uniform parameter
    uint d_offset;
 } p;

-shared FLOAT_TYPE tmp[BLOCK_SIZE];
+#if !USE_SUBGROUP_ADD
+shared FLOAT_TYPE tmp[8][BLOCK_SIZE];
+#endif

 void main() {
    const uint tid = gl_LocalInvocationID.x;
    const uint row_x = gl_GlobalInvocationID.y;
-    const uint channel = gl_GlobalInvocationID.z;
-    const uint channel_x = channel / (p.nchannels_y / p.nchannels_x);
+
+    uint channel, channel_x;
+
+    // When gqa_ratio > 1, each invocation does multiple rows.
+    // The row in the A matrix is starting from channel / gqa_ratio and the
+    // rows in the B matrix are [channel, channel+gqa_ratio).
+    // When gpa_ratio is 1, each invocation does one row.
+    if (gqa_ratio > 1) {
+        channel_x = gl_GlobalInvocationID.z;
+        channel = channel_x * gqa_ratio;
+    } else {
+        channel = gl_GlobalInvocationID.z;
+        channel_x = channel / (p.nchannels_y / p.nchannels_x);;
+    }

    const uint nrows_y = p.ncols_x;
    const uint nrows_dst = p.nrows_x;
    const uint row_dst = row_x;

-    tmp[tid] = FLOAT_TYPE(0.0f);
-
-    for (uint col_x0 = 0; col_x0 < p.ncols_x; col_x0 += BLOCK_SIZE) {
-        const uint col_x = col_x0 + tid;
-
-        if (col_x >= p.ncols_x) {
-            break;
-        }
-
-        // x is transposed and permuted
-        const uint ix = row_x*p.nchannels_x*p.ncols_x + channel_x*p.ncols_x + col_x;
-        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);
-
-        const uint row_y = col_x;
-
-        // y is not transposed but permuted
-        const uint iy = channel*nrows_y + row_y;
-
-        tmp[tid] = fma(xi, FLOAT_TYPE(data_b[iy]), tmp[tid]);
+    FLOAT_TYPE temp[8];
+    [[unroll]] for (uint i = 0; i < 8; ++i) {
+        temp[i] = FLOAT_TYPE(0.0f);
    }

-    // dst is not transposed and not permuted
-    const uint idst = channel*nrows_dst + row_dst;
+    // Detect alignment for vector loads
+    bool is_aligned = (p.ncols_x % 4) == 0 && (p.nchannels_x % 4) == 0 && (nrows_y % 4) == 0;

+    for (uint col_x0 = 0; col_x0 < p.ncols_x; col_x0 += BLOCK_SIZE) {
+
+        // Use vec4 loads if aligned
+        if (col_x0 + 4*BLOCK_SIZE <= p.ncols_x && is_aligned) {
+
+            uint col_x = col_x0 + 4*tid;
+            const uint row_y = col_x;
+
+            // x is transposed and permuted
+            const uint ix = row_x*p.nchannels_x*p.ncols_x + channel_x*p.ncols_x + col_x;
+            const vec4 av4 = vec4(data_a_v4[ix / 4]);
+
+            [[unroll]] for (uint c = 0; c < gqa_ratio; ++c) {
+                // y is not transposed but permuted
+                const uint iy = (channel + c)*nrows_y + row_y;
+
+                vec4 bv4 = data_b_v4[iy / 4];
+                temp[c] += dot(av4, bv4);
+            }
+
+            col_x0 += 3*BLOCK_SIZE;
+        } else {
+            const uint col_x = col_x0 + tid;
+
+            if (col_x >= p.ncols_x) {
+                break;
+            }
+
+            // x is transposed and permuted
+            const uint ix = row_x*p.nchannels_x*p.ncols_x + channel_x*p.ncols_x + col_x;
+            const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);
+
+            const uint row_y = col_x;
+
+            [[unroll]] for (uint c = 0; c < gqa_ratio; ++c) {
+                // y is not transposed but permuted
+                const uint iy = (channel + c)*nrows_y + row_y;
+
+                temp[c] = fma(xi, FLOAT_TYPE(data_b[iy]), temp[c]);
+            }
+        }
+    }
+
+#if USE_SUBGROUP_ADD
+    // reduce vec4 at a time
+    vec4 t = vec4(temp[0], temp[1], temp[2], temp[3]);
+    t = subgroupAdd(t);
+    temp[0] = t[0];
+    temp[1] = t[1];
+    temp[2] = t[2];
+    temp[3] = t[3];
+    if (gqa_ratio > 4) {
+        t = vec4(temp[4], temp[5], temp[6], temp[7]);
+        t = subgroupAdd(t);
+        temp[4] = t[0];
+        temp[5] = t[1];
+        temp[6] = t[2];
+        temp[7] = t[3];
+    }
+#else
+    [[unroll]] for (uint c = 0; c < gqa_ratio; ++c) {
+        tmp[c][tid] = temp[c];
+    }
    // sum up partial sums and write back result
    barrier();
    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
+            [[unroll]] for (uint c = 0; c < gqa_ratio; ++c) {
+                temp[c] += tmp[c][tid + s];
+                tmp[c][tid] = temp[c];
+            }
        }
        barrier();
    }
+    [[unroll]] for (uint c = 0; c < gqa_ratio; ++c) {
+        temp[c] = tmp[c][tid];
+    }
+#endif

    if (tid == 0) {
-        dst[idst] = tmp[0];
+        [[unroll]] for (uint c = 0; c < gqa_ratio; ++c) {
+            // dst is not transposed and not permuted
+            const uint idst = (channel + c)*nrows_dst + row_dst;
+            dst[idst] = temp[c];
+        }
    }
 }
@@ -336,8 +336,8 @@ void main() {
            const uint iqs = idx & 0x07;

            const float d = float(data_a_packed16[ib].d);
-            const i8vec2 v0 = unpack8(data_a_packed16[ib].qs[2*iqs]);
-            const i8vec2 v1 = unpack8(data_a_packed16[ib].qs[2*iqs + 1]);
+            const i8vec2 v0 = unpack8(int32_t(data_a_packed16[ib].qs[2*iqs])).xy; // vec4 used due to #12147
+            const i8vec2 v1 = unpack8(int32_t(data_a_packed16[ib].qs[2*iqs + 1])).xy;
            const vec4 v = vec4(v0.x, v0.y, v1.x, v1.y) * d;

            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
@@ -544,7 +544,7 @@ void main() {
            const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4));
            const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
            const uint grid = iq2xxs_grid[qs][(idx % 4) / 2] >> (16 * (idx & 1));
-            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
+            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147

            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
@@ -564,7 +564,7 @@ void main() {
            const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4));
            const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
            const uint grid = iq2xs_grid[qs & 511][(idx % 4) / 2] >> (16 * (idx & 1));
-            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
+            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147

            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
@@ -586,7 +586,7 @@ void main() {
            const float db = d * 0.25 * (0.5 + scale);
            const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
            const uint16_t grid = unpack16(iq2s_grid[qs | ((qh << (8 - qhshift)) & 0x300)][(idx & 2) >> 1])[idx & 1];
-            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid));
+            const vec2 v = db * vec2(sign01) * vec2(unpack8(uint32_t(grid)).xy); // vec4 used due to #12147

            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
@@ -611,7 +611,7 @@ void main() {
            const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4));
            const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
            const uint grid = iq3xxs_grid[qs] >> (16 * (idx & 1));
-            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
+            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147

            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
@@ -631,7 +631,7 @@ void main() {
            const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(sign << 1, sign)));
            const float db = d * (1 + 2 * ((scale >> (4 * (iqh & 1))) & 0xf));
            const uint32_t grid = iq3s_grid[qs | ((qh << (8 - (iqs % 8))) & 256)] >> (16 * (idx % 2));
-            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
+            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147

            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
@@ -2,6 +2,7 @@
 #if !defined(GGML_TYPES_COMP)
 #define GGML_TYPES_COMP

+#extension GL_EXT_shader_explicit_arithmetic_types_int64 : require
 #extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
 #extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
 #extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
@@ -312,6 +313,12 @@ struct block_iq1_m {
    uint16_t scales[QUANT_K_IQ1_M/64];
 };

+struct block_iq1_m_packed64 {
+    uint64_t  qs[QUANT_K_IQ1_M/8/8];
+    uint64_t  qh[QUANT_K_IQ1_M/16/8];
+    uint64_t scales;
+};
+
 #if defined(DATA_A_IQ1_S)
 #define QUANT_K QUANT_K_IQ1_S
 #define QUANT_R QUANT_R_IQ1_S
@@ -426,8 +426,9 @@ void process_shaders() {
        }
    }

-    string_to_spv("mul_mat_vec_p021_f16_f32", "mul_mat_vec_p021.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
-    string_to_spv("mul_mat_vec_nc_f16_f32", "mul_mat_vec_nc.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
+    string_to_spv("mul_mat_vec_p021_f16_f32_subgroup_add", "mul_mat_vec_p021.comp", {{"A_TYPE", "float16_t"}, {"A_TYPE_VEC4", "f16vec4"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}});
+    string_to_spv("mul_mat_vec_p021_f16_f32",              "mul_mat_vec_p021.comp", {{"A_TYPE", "float16_t"}, {"A_TYPE_VEC4", "f16vec4"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}});
+    string_to_spv("mul_mat_vec_nc_f16_f32", "mul_mat_vec_nc.comp", {{"A_TYPE", "float16_t"}, {"A_TYPE_VEC4", "f16vec4"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}});

    // Norms
    string_to_spv("norm_f32", "norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
@@ -445,6 +446,7 @@ void process_shaders() {

    for (std::string t : {"q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
        string_to_spv("cpy_f32_" + t, "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+        string_to_spv("cpy_f32_" + t + "_rte", "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}});
        string_to_spv("cpy_" + t + "_f32", "copy_from_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
    }

@@ -4369,7 +4369,7 @@ struct ggml_tensor * ggml_flash_attn_ext(
    }

    // permute(0, 2, 1, 3)
-    int64_t ne[4] = { q->ne[0], q->ne[2], q->ne[1], q->ne[3] };
+    int64_t ne[4] = { v->ne[0], q->ne[2], q->ne[1], q->ne[3] };
    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);

    float params[] = { scale, max_bias, logit_softcap };
@@ -286,6 +286,7 @@ class MODEL_ARCH(IntEnum):
    GRANITE_MOE      = auto()
    CHAMELEON        = auto()
    WAVTOKENIZER_DEC = auto()
+    PLM              = auto()


 class MODEL_TENSOR(IntEnum):
@@ -488,6 +489,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
    MODEL_ARCH.GRANITE_MOE:      "granitemoe",
    MODEL_ARCH.CHAMELEON:        "chameleon",
    MODEL_ARCH.WAVTOKENIZER_DEC: "wavtokenizer-dec",
+    MODEL_ARCH.PLM:              "plm",
 }

 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -1113,6 +1115,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
    ],
    MODEL_ARCH.GEMMA3: [
        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_NORM,
@@ -1463,6 +1466,20 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.FFN_UP_SHEXP,
        MODEL_TENSOR.FFN_EXP_PROBS_B,
    ],
+    MODEL_ARCH.PLM: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_KV_A_MQA,
+        MODEL_TENSOR.ATTN_KV_A_NORM,
+        MODEL_TENSOR.ATTN_KV_B,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_DOWN,
+    ],
    MODEL_ARCH.CHATGLM : [
        MODEL_TENSOR.TOKEN_EMBD,
        MODEL_TENSOR.ROPE_FREQS,
@@ -154,7 +154,12 @@ class SpecialVocab:
            return True
        with open(tokenizer_config_file, encoding = 'utf-8') as f:
            tokenizer_config = json.load(f)
-        chat_template = tokenizer_config.get('chat_template')
+        chat_template_alt = None
+        chat_template_file = path / 'chat_template.json'
+        if chat_template_file.is_file():
+            with open(chat_template_file, encoding = 'utf-8') as f:
+                chat_template_alt = json.load(f).get('chat_template')
+        chat_template = tokenizer_config.get('chat_template', chat_template_alt)
        if chat_template is None or isinstance(chat_template, (str, list)):
            self.chat_template = chat_template
        else:
@@ -107,6 +107,7 @@ extern "C" {
        LLAMA_VOCAB_PRE_TYPE_MINERVA        = 27,
        LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM  = 28,
        LLAMA_VOCAB_PRE_TYPE_GPT4O          = 29,
+        LLAMA_VOCAB_PRE_TYPE_SUPERBPE       = 30,
    };

    enum llama_rope_type {
@@ -1264,6 +1265,10 @@ extern "C" {
                               float   tau,
                               float   eta);

+    /// @details Intializes a GBNF grammar, see grammars/README.md for details.
+    /// @param vocab The vocabulary that this grammar will be used with.
+    /// @param grammar_str The production rules for the grammar, encoded as a string. Returns an empty grammar if empty. Returns NULL if parsing of grammar_str fails.
+    /// @param grammar_root The name of the start symbol for the grammar.
    LLAMA_API struct llama_sampler * llama_sampler_init_grammar(
            const struct llama_vocab * vocab,
                          const char * grammar_str,
--- a/Show More
+++ b/Show More