SYCL: implement memset ggml backend buffer interface (#12580)

* SYCL: implement memset ggml backend buffer interface

* use GGML_ABORT macro

* Do not wait for all queues to finish for memset operation

.github/workflows/build.yml

@@ -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
 

build-xcframework.sh

@@ -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
 

ci/README.md

@@ -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 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.

ci/run.sh

@@ -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="-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

common/CMakeLists.txt

@@ -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

common/llguidance.cpp

@@ -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

convert_hf_to_gguf.py

@@ -529,6 +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}]")
@@ -538,13 +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.
@@ -703,6 +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")
@@ -1747,7 +1752,7 @@ class Mistral3Model(LlamaModel):
 
     # 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
@@ -3380,7 +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
@@ -3798,8 +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)
 
@@ -3809,6 +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):
@@ -5353,7 +5360,7 @@ 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,

convert_hf_to_gguf_update.py

@@ -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", },
 ]
 
 

docs/backend/CUDA-FEDORA.md

@@ -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.

docs/backend/SYCL.md

@@ -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

docs/build.md

@@ -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)

docs/install.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

examples/llava/clip.cpp

@@ -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;

examples/run/run.cpp

@@ -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) {

examples/server/server.cpp

@@ -830,6 +830,11 @@ 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;
     }
 };

examples/server/webui/src/components/ChatScreen.tsx

@@ -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,
+  };
+}

examples/server/webui/src/utils/llama-vscode.ts

@@ -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(() => {

examples/tts/tts.cpp

@@ -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);

ggml/src/CMakeLists.txt

@@ -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)

ggml/src/ggml-cpu/CMakeLists.txt

@@ -359,9 +359,9 @@ 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)

ggml/src/ggml-cpu/ggml-cpu.c

@@ -3110,17 +3110,17 @@ static void ggml_compute_forward_dup_same_cont(
     const int ith = params->ith; // thread index
     const int nth = params->nth; // number of threads
 
-    // parallelize by elements
-    const int ne = ggml_nelements(dst);
-    const int dr = (ne + nth - 1) / nth;
-    const int ie0 = dr * ith;
-    const int ie1 = MIN(ie0 + dr, ne);
+    // parallelize by blocks
+    const int nk = ggml_nelements(src0)/ggml_blck_size(src0->type);
+    const int dr = (nk + nth - 1) / nth;
+    const int k0 = dr * ith;
+    const int k1 = MIN(k0 + dr, nk);
 
-    if (ie0 < ie1) {
+    if (k0 < k1) {
         memcpy(
-            ((char *)  dst->data + ie0*nb0),
-            ((char *) src0->data + ie0*nb0),
-            (ie1 - ie0) * nb0);
+            ((char *)  dst->data + k0*nb0),
+            ((char *) src0->data + k0*nb0),
+            (k1 - k0) * nb0);
     }
 }
 
@@ -4055,7 +4055,6 @@ static void ggml_compute_forward_dup_f32(
 static void ggml_compute_forward_dup_bytes(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
-
     const struct ggml_tensor * src0 = dst->src[0];
 
     GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
@@ -4069,10 +4068,10 @@ static void ggml_compute_forward_dup_bytes(
     }
 
     const size_t type_size = ggml_type_size(src0->type);
+
     const int ith = params->ith; // thread index
     const int nth = params->nth; // number of threads
 
-
     // parallelize by rows
     const int nr = ne01;
     // number of rows per thread
@@ -4082,10 +4081,10 @@ static void ggml_compute_forward_dup_bytes(
     const int ir1 = MIN(ir0 + dr, nr);
 
     if (src0->type == dst->type &&
-        ne00 == ne0 &&
+        ggml_are_same_shape(src0, dst) &&
         nb00 == type_size && nb0 == type_size) {
         // copy by rows
-        const size_t rs = ne00 * type_size;
+        const size_t rs = ggml_row_size(src0->type, ne00);
         for (int64_t i03 = 0; i03 < ne03; i03++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
                 for (int64_t i01 = ir0; i01 < ir1; i01++) {
@@ -4140,17 +4139,20 @@ static void ggml_compute_forward_dup_bytes(
     }
 
     // dst counters
-
-    int64_t i10 = 0;
+    int64_t k10 = 0;
     int64_t i11 = 0;
     int64_t i12 = 0;
     int64_t i13 = 0;
 
+    // number of blocks in a row
+    const int64_t nk00 = ne00 / ggml_blck_size(src0->type);
+    const int64_t nk0  = ne0  / ggml_blck_size(dst->type);
+
     for (int64_t i03 = 0; i03 < ne03; i03++) {
         for (int64_t i02 = 0; i02 < ne02; i02++) {
-            i10 += ne00 * ir0;
-            while (i10 >= ne0) {
-                i10 -= ne0;
+            k10 += nk00 * ir0;
+            while (k10 >= nk0) {
+                k10 -= nk0;
                 if (++i11 == ne1) {
                     i11 = 0;
                     if (++i12 == ne2) {
@@ -4162,14 +4164,14 @@ static void ggml_compute_forward_dup_bytes(
                 }
             }
             for (int64_t i01 = ir0; i01 < ir1; i01++) {
-                for (int64_t i00 = 0; i00 < ne00; i00++) {
-                    const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
-                          char * dst_ptr  = ((char *)  dst->data + i10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
+                for (int64_t k00 = 0; k00 < nk00; k00++) {
+                    const char * src0_ptr = ((char *) src0->data + k00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                          char * dst_ptr  = ((char *)  dst->data + k10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
 
                     memcpy(dst_ptr, src0_ptr, type_size);
 
-                    if (++i10 == ne0) {
-                        i10 = 0;
+                    if (++k10 == nk0) {
+                        k10 = 0;
                         if (++i11 == ne1) {
                             i11 = 0;
                             if (++i12 == ne2) {
@@ -4182,9 +4184,9 @@ static void ggml_compute_forward_dup_bytes(
                     }
                 }
             }
-            i10 += ne00 * (ne01 - ir1);
-            while (i10 >= ne0) {
-                i10 -= ne0;
+            k10 += nk00 * (ne01 - ir1);
+            while (k10 >= nk0) {
+                k10 -= nk0;
                 if (++i11 == ne1) {
                     i11 = 0;
                     if (++i12 == ne2) {
@@ -14308,7 +14310,9 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
     }
 
     // extra_buffer op?
-    if (ggml_cpu_extra_compute_forward(params, tensor)) return;
+    if (ggml_cpu_extra_compute_forward(params, tensor)) {
+        return;
+    }
 
     switch (tensor->op) {
         case GGML_OP_DUP:

ggml/src/ggml-cpu/kleidiai/kernels.cpp

@@ -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,

ggml/src/ggml-cpu/kleidiai/kernels.h

@@ -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 {

ggml/src/ggml-cpu/kleidiai/kleidiai.cpp

@@ -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);

ggml/src/ggml-cuda/common.cuh

@@ -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) {
@@ -397,7 +411,7 @@ 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;

ggml/src/ggml-cuda/fattn-common.cuh

@@ -606,48 +606,47 @@ 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);
         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;
 }
 
 static void on_no_fattn_vec_case(const int D) {
@@ -671,12 +670,10 @@ 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 +745,14 @@ 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 +768,43 @@ 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 +836,7 @@ 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 +848,7 @@ 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 +857,14 @@ 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());
 }

ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -970,7 +970,8 @@ 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);
 }
 
 

ggml/src/ggml-cuda/fattn-tile-f16.cu

@@ -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];
@@ -271,16 +269,16 @@ 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
@@ -288,7 +286,7 @@ static __global__ void flash_attn_tile_ext_f16(
 #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 +294,17 @@ 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 +324,22 @@ 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);
     }
 }

ggml/src/ggml-cuda/fattn-tile-f32.cu

@@ -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];
@@ -269,17 +267,17 @@ 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
@@ -287,7 +285,7 @@ static __global__ void flash_attn_tile_ext_f32(
 #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 +293,17 @@ 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 +320,22 @@ 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);
     }
 }

ggml/src/ggml-cuda/fattn-vec-f16.cuh

@@ -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 +281,29 @@ 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;
 #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 +323,48 @@ 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);
     }
 }
 

ggml/src/ggml-cuda/fattn-vec-f32.cuh

@@ -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 +266,29 @@ 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
     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 +305,48 @@ 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);
     }
 }
 

ggml/src/ggml-cuda/fattn-wmma-f16.cu

@@ -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;
         }
 
@@ -428,7 +427,7 @@ 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;
@@ -462,60 +461,26 @@ 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) {

ggml/src/ggml-cuda/fattn.cu

@@ -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:

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

@@ -264,9 +264,9 @@ static ggml_cuda_device_info ggml_cuda_init() {
 #elif defined(GGML_USE_MUSA)
         // FIXME: Ensure compatibility with varying warp sizes across different MUSA archs.
         info.devices[id].warp_size = 32;
-        // TODO: refine the .cc to reflect MUSA's actual CC capabilities
         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
@@ -1188,11 +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) {
@@ -1216,7 +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(
@@ -1759,7 +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;
@@ -1836,7 +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);
     }
@@ -3230,6 +3232,9 @@ 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[0]->ne[3] != 1) {
+                return false;
+            }
             if (op->src[1]->type == GGML_TYPE_BF16 || op->src[2]->type == GGML_TYPE_BF16) {
                 return false;
             }

ggml/src/ggml-cuda/mmq.cu

@@ -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;
 }

ggml/src/ggml-cuda/mmq.cuh

@@ -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() {
@@ -2577,9 +2577,9 @@ 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)
@@ -2772,14 +2772,14 @@ 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;
@@ -2832,7 +2832,7 @@ void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cuda
     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;

ggml/src/ggml-cuda/mmvq.cu

@@ -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)) {

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

@@ -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

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

@@ -134,5 +134,6 @@
 #define cudaStreamCaptureModeRelaxed musaStreamCaptureModeRelaxed
 #define cudaStreamBeginCapture musaStreamBeginCapture
 #define cudaStreamEndCapture musaStreamEndCapture
+#define cudaOccupancyMaxActiveBlocksPerMultiprocessor musaOccupancyMaxActiveBlocksPerMultiprocessor
 
 typedef mt_bfloat16 nv_bfloat16;

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -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

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

@@ -2561,171 +2561,180 @@ static void ggml_metal_encode_node(
                     [encoder setThreadgroupMemoryLength:8192 atIndex:0];
                     [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
                 } else {
-                    int nth0 = 32;
-                    int nth1 = 1;
-                    int nrows = 1;
-                    //printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
-
                     id<MTLComputePipelineState> pipeline = nil;
 
+                    int nsg = 0; // number of simdgroups
+                    int nr0 = 0; // number of src0 rows per simdgroup
+                    int nr1 = 1; // number of src1 rows per threadgroup
+
+                    size_t smem = 0; // shared memory
+
                     // use custom matrix x vector kernel
                     switch (src0t) {
                         case GGML_TYPE_F32:
                             {
                                 GGML_ASSERT(src1t == GGML_TYPE_F32);
+                                nsg = 1;
+                                nr0 = 1;
+                                nr1 = 4;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32].pipeline;
-                                nrows = 4;
                             } break;
                         case GGML_TYPE_F16:
                             {
-                                nth0 = 32;
-                                nth1 = 1;
+                                nsg = 1;
+                                nr0 = 1;
                                 if (src1t == GGML_TYPE_F32) {
                                     if (ne11 * ne12 < 4) {
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW].pipeline;
                                     } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4].pipeline;
-                                        nrows = ne11;
+                                        nr1 = ne11;
                                     } else {
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32].pipeline;
-                                        nrows = 4;
+                                        nr1 = 4;
                                     }
                                 } else {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16].pipeline;
-                                    nrows = 4;
+                                    nr1 = 4;
                                 }
                             } break;
                         case GGML_TYPE_BF16:
                             {
-                                nth0 = 32;
-                                nth1 = 1;
+                                nsg = 1;
+                                nr0 = 1;
                                 if (src1t == GGML_TYPE_F32) {
                                     if (ne11 * ne12 < 4) {
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW].pipeline;
                                     } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4].pipeline;
-                                        nrows = ne11;
+                                        nr1 = ne11;
                                     } else {
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32].pipeline;
-                                        nrows = 4;
+                                        nr1 = 4;
                                     }
                                 } else {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16].pipeline;
-                                    nrows = 4;
+                                    nr1 = 4;
                                 }
                             } break;
                         case GGML_TYPE_Q4_0:
                             {
-                                nth0 = 8;
-                                nth1 = 8;
+                                nsg = N_SG_Q4_0;
+                                nr0 = N_R0_Q4_0;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32].pipeline;
                             } break;
                         case GGML_TYPE_Q4_1:
                             {
-                                nth0 = 8;
-                                nth1 = 8;
+                                nsg = N_SG_Q4_1;
+                                nr0 = N_R0_Q4_1;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32].pipeline;
                             } break;
                         case GGML_TYPE_Q5_0:
                             {
-                                nth0 = 8;
-                                nth1 = 8;
+                                nsg = N_SG_Q5_0;
+                                nr0 = N_R0_Q5_0;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32].pipeline;
                             } break;
                         case GGML_TYPE_Q5_1:
                             {
-                                nth0 = 8;
-                                nth1 = 8;
+                                nsg = N_SG_Q5_1;
+                                nr0 = N_R0_Q5_1;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32].pipeline;
                             } break;
                         case GGML_TYPE_Q8_0:
                             {
-                                nth0 = 8;
-                                nth1 = 8;
+                                nsg = N_SG_Q8_0;
+                                nr0 = N_R0_Q8_0;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32].pipeline;
                             } break;
                         case GGML_TYPE_Q2_K:
                             {
-                                nth0 = 2;
-                                nth1 = 32;
+                                nsg = N_SG_Q2_K;
+                                nr0 = N_R0_Q2_K;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32].pipeline;
                             } break;
                         case GGML_TYPE_Q3_K:
                             {
-                                nth0 = 2;
-                                nth1 = 32;
+                                nsg = N_SG_Q3_K;
+                                nr0 = N_R0_Q3_K;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32].pipeline;
                             } break;
                         case GGML_TYPE_Q4_K:
                             {
-                                nth0 = 4; //1;
-                                nth1 = 8; //32;
+                                nsg = N_SG_Q4_K;
+                                nr0 = N_R0_Q4_K;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32].pipeline;
                             } break;
                         case GGML_TYPE_Q5_K:
                             {
-                                nth0 = 2;
-                                nth1 = 32;
+                                nsg = N_SG_Q5_K;
+                                nr0 = N_R0_Q5_K;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32].pipeline;
                             } break;
                         case GGML_TYPE_Q6_K:
                             {
-                                nth0 = 2;
-                                nth1 = 32;
+                                nsg = N_SG_Q6_K;
+                                nr0 = N_R0_Q6_K;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32].pipeline;
                             } break;
                         case GGML_TYPE_IQ2_XXS:
                             {
-                                nth0 = 4;
-                                nth1 = 16;
+                                nsg = N_SG_IQ2_XXS;
+                                nr0 = N_R0_IQ2_XXS;
+                                smem = 256*8+128;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32].pipeline;
                             } break;
                         case GGML_TYPE_IQ2_XS:
                             {
-                                nth0 = 4;
-                                nth1 = 16;
+                                nsg = N_SG_IQ2_XS;
+                                nr0 = N_R0_IQ2_XS;
+                                smem = 512*8+128;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32].pipeline;
                             } break;
                         case GGML_TYPE_IQ3_XXS:
                             {
-                                nth0 = 4;
-                                nth1 = 16;
+                                nsg = N_SG_IQ3_XXS;
+                                nr0 = N_R0_IQ3_XXS;
+                                smem = 256*4+128;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32].pipeline;
                             } break;
                         case GGML_TYPE_IQ3_S:
                             {
-                                nth0 = 4;
-                                nth1 = 16;
+                                nsg = N_SG_IQ3_S;
+                                nr0 = N_R0_IQ3_S;
+                                smem = 512*4;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32].pipeline;
                             } break;
                         case GGML_TYPE_IQ2_S:
                             {
-                                nth0 = 4;
-                                nth1 = 16;
+                                nsg = N_SG_IQ2_S;
+                                nr0 = N_R0_IQ2_S;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32].pipeline;
                             } break;
                         case GGML_TYPE_IQ1_S:
                             {
-                                nth0 = 4;
-                                nth1 = 16;
+                                nsg = N_SG_IQ1_S;
+                                nr0 = N_R0_IQ1_S;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32].pipeline;
                             } break;
                         case GGML_TYPE_IQ1_M:
                             {
-                                nth0 = 4;
-                                nth1 = 16;
+                                nsg = N_SG_IQ1_M;
+                                nr0 = N_R0_IQ1_M;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32].pipeline;
                             } break;
                         case GGML_TYPE_IQ4_NL:
                             {
-                                nth0 = 4;
-                                nth1 = 16;
+                                nsg = N_SG_IQ4_NL;
+                                nr0 = N_R0_IQ4_NL;
+                                smem = 32*sizeof(float);
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32].pipeline;
                             } break;
                         case GGML_TYPE_IQ4_XS:
                             {
-                                nth0 = 4;
-                                nth1 = 16;
+                                nsg = N_SG_IQ4_XS;
+                                nr0 = N_R0_IQ4_XS;
+                                smem = 32*sizeof(float);
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32].pipeline;
                             } break;
                         default:
@@ -2762,41 +2771,10 @@ static void ggml_metal_encode_node(
                     [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
                     [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
 
-                    if (src0t == GGML_TYPE_Q4_0  || src0t == GGML_TYPE_Q4_1  || src0t == GGML_TYPE_Q5_0 ||
-                        src0t == GGML_TYPE_Q5_1  || src0t == GGML_TYPE_Q8_0  || src0t == GGML_TYPE_Q2_K ||
-                        src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
-                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                    }
-                    else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
-                        const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
-                        [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                    }
-                    else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) {
-                        const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
-                        [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                    }
-                    else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) {
-                        const int mem_size = 32*sizeof(float);
-                        [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                    }
-                    else if (src0t == GGML_TYPE_Q4_K) {
-                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                    }
-                    else if (src0t == GGML_TYPE_Q3_K) {
-                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                    }
-                    else if (src0t == GGML_TYPE_Q5_K) {
-                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                    }
-                    else if (src0t == GGML_TYPE_Q6_K) {
-                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                    } else {
-                        const int64_t ny = (ne11 + nrows - 1)/nrows;
-                        [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    if (smem > 0) {
+                        [encoder setThreadgroupMemoryLength:smem atIndex:0];
                     }
+                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nr0*nsg - 1)/(nr0*nsg), (ne11 + nr1 - 1)/nr1, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)];
                 }
             } break;
         case GGML_OP_MUL_MAT_ID:
@@ -2902,146 +2880,155 @@ static void ggml_metal_encode_node(
 
                     [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 31)/32, (ne01 + 63)/64, n_as) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
                 } else {
-                    int nth0 = 32;
-                    int nth1 = 1;
-                    int nrows = 1;
-                    //printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
-
                     id<MTLComputePipelineState> pipeline = nil;
 
+                    int nsg = 0; // number of simdgroups
+                    int nr0 = 0; // number of src0 rows per simdgroup
+                    int nr1 = 1; // number of src1 rows per threadgroup
+
+                    size_t smem = 0; // shared memory
+
                     // use custom matrix x vector kernel
                     switch (src0t) {
                         case GGML_TYPE_F32:
                             {
                                 GGML_ASSERT(src1t == GGML_TYPE_F32);
+                                nsg = 1;
+                                nr0 = 1;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32].pipeline;
                             } break;
                         case GGML_TYPE_F16:
                             {
                                 GGML_ASSERT(src1t == GGML_TYPE_F32);
-                                nth0 = 32;
-                                nth1 = 1;
+                                nsg = 1;
+                                nr0 = 1;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32].pipeline;
                             } break;
                         case GGML_TYPE_BF16:
                             {
                                 GGML_ASSERT(src1t == GGML_TYPE_F32);
-                                nth0 = 32;
-                                nth1 = 1;
+                                nsg = 1;
+                                nr0 = 1;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_BF16_F32].pipeline;
                             } break;
                         case GGML_TYPE_Q4_0:
                             {
-                                nth0 = 8;
-                                nth1 = 8;
+                                nsg = N_SG_Q4_0;
+                                nr0 = N_R0_Q4_0;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32].pipeline;
                             } break;
                         case GGML_TYPE_Q4_1:
                             {
-                                nth0 = 8;
-                                nth1 = 8;
+                                nsg = N_SG_Q4_1;
+                                nr0 = N_R0_Q4_1;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32].pipeline;
                             } break;
                         case GGML_TYPE_Q5_0:
                             {
-                                nth0 = 8;
-                                nth1 = 8;
+                                nsg = N_SG_Q5_0;
+                                nr0 = N_R0_Q5_0;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32].pipeline;
                             } break;
                         case GGML_TYPE_Q5_1:
                             {
-                                nth0 = 8;
-                                nth1 = 8;
+                                nsg = N_SG_Q5_1;
+                                nr0 = N_R0_Q5_1;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32].pipeline;
                             } break;
                         case GGML_TYPE_Q8_0:
                             {
-                                nth0 = 8;
-                                nth1 = 8;
+                                nsg = N_SG_Q8_0;
+                                nr0 = N_R0_Q8_0;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32].pipeline;
                             } break;
                         case GGML_TYPE_Q2_K:
                             {
-                                nth0 = 2;
-                                nth1 = 32;
+                                nsg = N_SG_Q2_K;
+                                nr0 = N_R0_Q2_K;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32].pipeline;
                             } break;
                         case GGML_TYPE_Q3_K:
                             {
-                                nth0 = 2;
-                                nth1 = 32;
+                                nsg = N_SG_Q3_K;
+                                nr0 = N_R0_Q3_K;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32].pipeline;
                             } break;
                         case GGML_TYPE_Q4_K:
                             {
-                                nth0 = 4; //1;
-                                nth1 = 8; //32;
+                                nsg = N_SG_Q4_K;
+                                nr0 = N_R0_Q4_K;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32].pipeline;
                             } break;
                         case GGML_TYPE_Q5_K:
                             {
-                                nth0 = 2;
-                                nth1 = 32;
+                                nsg = N_SG_Q5_K;
+                                nr0 = N_R0_Q5_K;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32].pipeline;
                             } break;
                         case GGML_TYPE_Q6_K:
                             {
-                                nth0 = 2;
-                                nth1 = 32;
+                                nsg = N_SG_Q6_K;
+                                nr0 = N_R0_Q6_K;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32].pipeline;
                             } break;
                         case GGML_TYPE_IQ2_XXS:
                             {
-                                nth0 = 4;
-                                nth1 = 16;
+                                nsg = N_SG_IQ2_XXS;
+                                nr0 = N_R0_IQ2_XXS;
+                                smem = 256*8+128;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32].pipeline;
                             } break;
                         case GGML_TYPE_IQ2_XS:
                             {
-                                nth0 = 4;
-                                nth1 = 16;
+                                nsg = N_SG_IQ2_XS;
+                                nr0 = N_R0_IQ2_XS;
+                                smem = 512*8+128;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32].pipeline;
                             } break;
                         case GGML_TYPE_IQ3_XXS:
                             {
-                                nth0 = 4;
-                                nth1 = 16;
+                                nsg = N_SG_IQ3_XXS;
+                                nr0 = N_R0_IQ3_XXS;
+                                smem = 256*4+128;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32].pipeline;
                             } break;
                         case GGML_TYPE_IQ3_S:
                             {
-                                nth0 = 4;
-                                nth1 = 16;
+                                nsg = N_SG_IQ3_S;
+                                nr0 = N_R0_IQ3_S;
+                                smem = 512*4;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32].pipeline;
                             } break;
                         case GGML_TYPE_IQ2_S:
                             {
-                                nth0 = 4;
-                                nth1 = 16;
+                                nsg = N_SG_IQ2_S;
+                                nr0 = N_R0_IQ2_S;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32].pipeline;
                             } break;
                         case GGML_TYPE_IQ1_S:
                             {
-                                nth0 = 4;
-                                nth1 = 16;
+                                nsg = N_SG_IQ1_S;
+                                nr0 = N_R0_IQ1_S;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32].pipeline;
                             } break;
                         case GGML_TYPE_IQ1_M:
                             {
-                                nth0 = 4;
-                                nth1 = 16;
+                                nsg = N_SG_IQ1_M;
+                                nr0 = N_R0_IQ1_M;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32].pipeline;
                             } break;
                         case GGML_TYPE_IQ4_NL:
                             {
-                                nth0 = 4;
-                                nth1 = 16;
+                                nsg = N_SG_IQ4_NL;
+                                nr0 = N_R0_IQ4_NL;
+                                smem = 32*sizeof(float);
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32].pipeline;
                             } break;
                         case GGML_TYPE_IQ4_XS:
                             {
-                                nth0 = 4;
-                                nth1 = 16;
+                                nsg = N_SG_IQ4_XS;
+                                nr0 = N_R0_IQ4_XS;
+                                smem = 32*sizeof(float);
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32].pipeline;
                             } break;
                         default:
@@ -3052,7 +3039,7 @@ static void ggml_metal_encode_node(
                     };
 
                     if (ggml_is_quantized(src0t)) {
-                        GGML_ASSERT(ne00 >= nth0*nth1);
+                        GGML_ASSERT(ne00 >= nsg*nr0);
                     }
 
                     ggml_metal_kargs_mul_mv_id args = {
@@ -3085,43 +3072,12 @@ static void ggml_metal_encode_node(
                     [encoder setBuffer:id_src2 offset:offs_src2 atIndex:4];
 
                     const int64_t _ne1 = 1;
-                    const int tgz = dst_rows;
+                    const int64_t ne123 = dst_rows;
 
-                    if (src0t == GGML_TYPE_Q4_0  || src0t == GGML_TYPE_Q4_1  || src0t == GGML_TYPE_Q5_0 ||
-                            src0t == GGML_TYPE_Q5_1  || src0t == GGML_TYPE_Q8_0  || src0t == GGML_TYPE_Q2_K ||
-                            src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
-                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                    }
-                    else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
-                        const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
-                        [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                    }
-                    else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) {
-                        const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
-                        [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                    }
-                    else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) {
-                        const int mem_size = 32*sizeof(float);
-                        [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                    }
-                    else if (src0t == GGML_TYPE_Q4_K) {
-                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                    }
-                    else if (src0t == GGML_TYPE_Q3_K) {
-                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                    }
-                    else if (src0t == GGML_TYPE_Q5_K) {
-                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                    }
-                    else if (src0t == GGML_TYPE_Q6_K) {
-                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                    } else {
-                        const int64_t ny = (_ne1 + nrows - 1)/nrows; // = _ne1
-                        [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    if (smem > 0) {
+                        [encoder setThreadgroupMemoryLength:smem atIndex:0];
                     }
+                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nr0*nsg - 1)/(nr0*nsg), (_ne1 + nr1 - 1)/nr1, ne123) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)];
                 }
             } break;
         case GGML_OP_GET_ROWS:

ggml/src/ggml-opencl/CMakeLists.txt

@@ -25,124 +25,46 @@ 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
+)
+
+foreach (K ${GGML_OPENCL_KERNELS})
+    ggml_opencl_add_kernel(${K})
+endforeach()

ggml/src/ggml-sycl/CMakeLists.txt

@@ -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)

ggml/src/ggml-sycl/common.hpp

@@ -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];
 

ggml/src/ggml-sycl/gemm.hpp

@@ -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);
     }

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

@@ -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);

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

@@ -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) {

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

@@ -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"
 

ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.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

ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp

@@ -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];

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

@@ -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) {

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp

@@ -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);

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp

@@ -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)]));

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

@@ -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) {

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp

@@ -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];
+        }
     }
 }

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

@@ -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);

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

@@ -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

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

@@ -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"}});
     }
 

gguf-py/gguf/constants.py

@@ -1113,6 +1113,7 @@ 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,

include/llama.h

@@ -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 {

src/llama-arch.cpp

@@ -778,6 +778,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
         {
             { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
             { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
             { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
             { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
             { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },

src/llama-context.cpp

@@ -294,10 +294,7 @@ llama_context::llama_context(
         // TODO: something cleaner
         const auto n_outputs_save = n_outputs;
 
-        // max number of outputs
-        n_outputs = n_tokens;
-
-        LLAMA_LOG_DEBUG("%s: n_tokens = %d, n_seqs = %d, n_outputs = %d\n", __func__, n_tokens, n_seqs, n_outputs);
+        LLAMA_LOG_DEBUG("%s: worst-case: n_tokens = %d, n_seqs = %d, n_outputs = %d\n", __func__, n_tokens, n_seqs, n_outputs);
 
         int n_splits_pp = -1;
         int n_nodes_pp  = -1;
@@ -313,8 +310,15 @@ llama_context::llama_context(
         // reserve pp graph first so that buffers are only allocated once
         {
             llama_ubatch ubatch_pp = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr};
+
+            // max number of outputs
+            n_outputs = ubatch_pp.n_tokens;
+
+            LLAMA_LOG_DEBUG("%s: reserving graph for n_tokens = %d, n_seqs = %d\n", __func__, ubatch_pp.n_tokens, ubatch_pp.n_seqs);
+
             auto * gf = graph_init();
             graph_build(ctx_compute.get(), gf, ubatch_pp, LLM_GRAPH_TYPE_DEFAULT);
+
             if (!ggml_backend_sched_reserve(sched.get(), gf)) {
                 throw std::runtime_error("failed to allocate compute pp buffers");
             }
@@ -326,11 +330,18 @@ llama_context::llama_context(
         // reserve with tg graph to get the number of splits and nodes
         {
             llama_ubatch ubatch_tg = { true, 1, 1, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr};
+
+            n_outputs = ubatch_tg.n_tokens;
+
+            LLAMA_LOG_DEBUG("%s: reserving graph for n_tokens = %d, n_seqs = %d\n", __func__, ubatch_tg.n_tokens, ubatch_tg.n_seqs);
+
             auto * gf = graph_init();
             graph_build(ctx_compute.get(), gf, ubatch_tg, LLM_GRAPH_TYPE_DEFAULT);
+
             if (!ggml_backend_sched_reserve(sched.get(), gf)) {
                 throw std::runtime_error("failed to allocate compute tg buffers");
             }
+
             n_splits_tg = ggml_backend_sched_get_n_splits(sched.get());
             n_nodes_tg  = ggml_graph_n_nodes(gf);
         }
@@ -338,8 +349,14 @@ llama_context::llama_context(
         // reserve again with pp graph to avoid ggml-alloc reallocations during inference
         {
             llama_ubatch ubatch_pp = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr};
+
+            n_outputs = ubatch_pp.n_tokens;
+
+            LLAMA_LOG_DEBUG("%s: reserving graph for n_tokens = %d, n_seqs = %d\n", __func__, ubatch_pp.n_tokens, ubatch_pp.n_seqs);
+
             auto * gf = graph_init();
             graph_build(ctx_compute.get(), gf, ubatch_pp, LLM_GRAPH_TYPE_DEFAULT);
+
             if (!ggml_backend_sched_reserve(sched.get(), gf)) {
                 throw std::runtime_error("failed to allocate compute pp buffers");
             }
@@ -1153,6 +1170,7 @@ int llama_context::encode(llama_batch & inp_batch) {
         // remember the sequence ids used during the encoding - needed for cross attention later
         cross.seq_ids_enc.resize(n_tokens);
         for (int32_t i = 0; i < n_tokens; i++) {
+            cross.seq_ids_enc[i].clear();
             for (int s = 0; s < ubatch.n_seq_id[i]; s++) {
                 llama_seq_id seq_id = ubatch.seq_id[i][s];
                 cross.seq_ids_enc[i].insert(seq_id);

src/llama-mmap.cpp

@@ -476,7 +476,7 @@ struct llama_mlock::impl {
 
         char* errmsg = std::strerror(errno);
         bool suggest = (errno == ENOMEM);
-#if defined(TARGET_OS_VISION) || defined(TARGET_OS_TV)
+#if defined(TARGET_OS_VISION) || defined(TARGET_OS_TV) || defined(_AIX)
         // visionOS/tvOS dont't support RLIMIT_MEMLOCK
         // Skip resource limit checks on visionOS/tvOS
         suggest = false;

src/llama-model.cpp

@@ -271,19 +271,32 @@ static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & de
         }
     }
 
-    // add extra buffer types
-    auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
-    auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
-    auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
-        ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
-    if (ggml_backend_dev_get_extra_bufts_fn) {
-        ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev);
-        while (extra_bufts && *extra_bufts) {
-            buft_list.emplace_back(cpu_dev, *extra_bufts);
-            ++extra_bufts;
+    bool has_gpu_device = false;
+    for (auto * dev : devices) {
+        if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_GPU) {
+            has_gpu_device = true;
+            break;
         }
     }
 
+    // add extra buffer types, only if no GPU device is present
+    // ref: https://github.com/ggml-org/llama.cpp/issues/12481#issuecomment-2743136094
+    if (!has_gpu_device) {
+        auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
+        auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
+            ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
+        if (ggml_backend_dev_get_extra_bufts_fn) {
+            ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev);
+            while (extra_bufts && *extra_bufts) {
+                buft_list.emplace_back(cpu_dev, *extra_bufts);
+                ++extra_bufts;
+            }
+        }
+    } else {
+        LLAMA_LOG_WARN("%s: disabling extra buffer types (i.e. repacking) since a GPU device is available\n", __func__);
+    }
+
     // add a host buffer type
     // storing the tensors in a host buffer is useful when the processing of large batches
     // is offloaded to a GPU device, since it reduces the time spent on data transfers
@@ -2210,9 +2223,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
 
                         // optional bias tensors
-                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, 0);
-                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, 0);
-                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, 0);
+                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
 
                         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
 
@@ -2329,7 +2342,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);
                         layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias",   i), { n_embd }, 0);
 
-                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, TENSOR_NOT_REQUIRED);
                         if (layer.wqkv == nullptr) {
                             layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, 0);
                             layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias",   i), {n_embd}, 0);
@@ -2558,7 +2571,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
                     // output
                     output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
-                    output      = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,   "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
 
                     for (int i = 0; i < n_layer; ++i) {
                         auto & layer = layers[i];
@@ -3215,16 +3233,16 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         auto & layer = layers[i];
 
                         layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
-                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                        layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED);
 
                         if (layer.wqkv == nullptr) {
                             layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
                             layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
                             layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
-                            layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd},     llama_model_loader::TENSOR_NOT_REQUIRED);
-                            layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                            layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                            layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+                            layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                            layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
                         }
 
                         layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
@@ -3335,12 +3353,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.time_mix_w2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_W2, "weight", i), {time_mix_extra_dim, n_embd, 5}, 0);
 
                         layer.time_mix_lerp_x = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_X, "weight", i), {n_embd, 1, 1}, 0);
-                        layer.time_mix_lerp_w = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_W, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_lerp_k = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_K, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_lerp_v = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_V, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_lerp_r = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_R, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_lerp_g = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_G, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 5}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_w = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_W, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_k = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_K, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_v = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_V, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_r = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_R, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_g = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_G, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 5}, TENSOR_NOT_REQUIRED);
                         GGML_ASSERT(!(layer.time_mix_lerp_fused == NULL && layer.time_mix_lerp_w == NULL));
 
                         layer.time_mix_first = create_tensor(tn(LLM_TENSOR_TIME_MIX_FIRST, "weight", i), {head_size, n_embd / head_size}, 0);
@@ -3370,7 +3388,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
                     output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
-                    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, TENSOR_NOT_REQUIRED);
                     output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0);
 
                     const int time_mix_extra_dim = hparams.time_mix_extra_dim;
@@ -3396,7 +3414,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.time_mix_lerp_x = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_X, "weight", i), {n_embd, 1, 1}, 0);
                         layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 5}, 0);
 
-                        layer.time_mix_first = create_tensor(tn(LLM_TENSOR_TIME_MIX_FIRST, "weight", i), {head_size, n_embd / head_size}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.time_mix_first = create_tensor(tn(LLM_TENSOR_TIME_MIX_FIRST, "weight", i), {head_size, n_embd / head_size}, TENSOR_NOT_REQUIRED);
                         layer.time_mix_decay = create_tensor(tn(LLM_TENSOR_TIME_MIX_DECAY, "weight", i), {n_embd}, 0);
                         layer.time_mix_decay_w1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_DECAY_W1, "weight", i), {n_embd, time_decay_extra_dim}, 0);
                         layer.time_mix_decay_w2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_DECAY_W2, "weight", i), {time_decay_extra_dim, attn_hidden_size}, 0);
@@ -3405,9 +3423,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.time_mix_receptance = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "weight", i), {attn_hidden_size, n_embd}, 0);
                         layer.time_mix_gate = create_tensor(tn(LLM_TENSOR_TIME_MIX_GATE, "weight", i), {attn_hidden_size, n_embd}, 0);
                         // optional bias tensors
-                        layer.time_mix_key_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_KEY, "bias", i), {attn_key_value_size}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_value_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_VALUE, "bias", i), {attn_key_value_size}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_receptance_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "bias", i), {attn_hidden_size}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.time_mix_key_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_KEY, "bias", i), {attn_key_value_size}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_value_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_VALUE, "bias", i), {attn_key_value_size}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_receptance_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "bias", i), {attn_hidden_size}, TENSOR_NOT_REQUIRED);
 
                         layer.time_mix_output = create_tensor(tn(LLM_TENSOR_TIME_MIX_OUTPUT, "weight", i), {n_embd, attn_hidden_size}, 0);
 
@@ -3528,8 +3546,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             layer.time_mix_v2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V2, "weight", i), {n_lora_value_res_mix, n_embd}, 0);
                         }
 
-                        layer.time_mix_g1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G1, "weight", i), {n_embd, n_lora_gate}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_g2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G2, "weight", i), {n_lora_gate, n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.time_mix_g1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G1, "weight", i), {n_embd, n_lora_gate}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_g2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G2, "weight", i), {n_lora_gate, n_embd}, TENSOR_NOT_REQUIRED);
 
                         try {
                             layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 6}, 0);
@@ -3546,8 +3564,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.time_mix_value = create_tensor(tn(LLM_TENSOR_TIME_MIX_VALUE, "weight", i), {attn_hidden_size, n_embd}, 0);
                         layer.time_mix_receptance = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "weight", i), {attn_hidden_size, n_embd}, 0);
 
-                        layer.time_mix_ln = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_ln_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.time_mix_ln = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_ln_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
                         layer.time_mix_output = create_tensor(tn(LLM_TENSOR_TIME_MIX_OUTPUT, "weight", i), {n_embd, attn_hidden_size}, 0);
 
                         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
@@ -6193,16 +6211,25 @@ struct llm_build_qwen2moe : public llm_graph_context {
             {
                 // compute Q and K and RoPE them
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
                 cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
 
                 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
                 cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
 
                 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
                 cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
 
                 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
                 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);

src/llama-vocab.cpp

@@ -400,6 +400,12 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                     "[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))*((?=[\\p{L}])([^A-Z]))+(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))+((?=[\\p{L}])([^A-Z]))*(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_SUPERBPE:
+                regex_exprs = {
+                    "\\p{N}+",
+                    "(?=(\\d{3})+(?!\\d))",
+                };
+                break;
             default:
                 // default regex for BPE tokenization pre-processing
                 regex_exprs = {
@@ -1604,6 +1610,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 tokenizer_pre == "gpt-4o") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_GPT4O;
                 clean_spaces = false;
+            } else if (
+                tokenizer_pre == "superbpe") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_SUPERBPE;
+                clean_spaces = false;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }

tests/test-backend-ops.cpp

@@ -259,6 +259,10 @@ static std::string var_to_str(ggml_type type) {
     return ggml_type_name(type);
 }
 
+static std::string var_to_str(ggml_prec prec) {
+    return prec == GGML_PREC_F32 ? "f32" : "def";
+}
+
 static std::string var_to_str(ggml_op_pool pool) {
     switch (pool) {
         case GGML_OP_POOL_AVG:  return "avg";
@@ -1459,11 +1463,13 @@ struct test_cpy : public test_case {
     const ggml_type type_src;
     const ggml_type type_dst;
     const std::array<int64_t, 4> ne;
-    const std::array<int64_t, 4> permute;
+    const std::array<int64_t, 4> permute_src;
+    const std::array<int64_t, 4> permute_dst;
     bool _src_use_permute;
+    bool _dst_use_permute;
 
     std::string vars() override {
-        return VARS_TO_STR4(type_src, type_dst, ne, permute);
+        return VARS_TO_STR5(type_src, type_dst, ne, permute_src, permute_dst);
     }
 
     double max_nmse_err() override {
@@ -1476,9 +1482,11 @@ struct test_cpy : public test_case {
 
     test_cpy(ggml_type type_src = GGML_TYPE_F32, ggml_type type_dst = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {10, 10, 10, 1},
-            std::array<int64_t, 4> permute = {0, 0, 0, 0})
-        : type_src(type_src), type_dst(type_dst), ne(ne), permute(permute),
-          _src_use_permute(permute[0] + permute[1] + permute[2] + permute[3] > 0) {}
+            std::array<int64_t, 4> permute_src = {0, 0, 0, 0},
+            std::array<int64_t, 4> permute_dst = {0, 0, 0, 0})
+        : type_src(type_src), type_dst(type_dst), ne(ne), permute_src(permute_src), permute_dst(permute_dst),
+          _src_use_permute(permute_src[0] + permute_src[1] + permute_src[2] + permute_src[3] > 0),
+          _dst_use_permute(permute_dst[0] + permute_dst[1] + permute_dst[2] + permute_dst[3] > 0) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * src = ggml_new_tensor(ctx, type_src, 4, ne.data());
@@ -1486,13 +1494,18 @@ struct test_cpy : public test_case {
         ggml_set_name(src, "src");
 
         if (_src_use_permute) {
-            src = ggml_permute(ctx, src, permute[0], permute[1], permute[2], permute[3]);
+            src = ggml_permute(ctx, src, permute_src[0], permute_src[1], permute_src[2], permute_src[3]);
             ggml_set_name(src, "src_permuted");
         }
 
-        ggml_tensor* dst = ggml_new_tensor(ctx, type_dst, 4, src->ne);
+        ggml_tensor * dst = ggml_new_tensor(ctx, type_dst, 4, src->ne);
         ggml_set_name(dst, "dst");
 
+        if (_dst_use_permute) {
+            dst = ggml_permute(ctx, dst, permute_dst[0], permute_dst[1], permute_dst[2], permute_dst[3]);
+            ggml_set_name(dst, "dst_permuted");
+        }
+
         ggml_tensor * out = ggml_cpy(ctx, src, dst);
         ggml_set_name(out, "out");
 
@@ -1960,9 +1973,10 @@ struct test_mul_mat : public test_case {
     const std::array<int64_t, 2> bs;  // dims 3 and 4
     const std::array<int64_t, 2> nr;  // repeat in dims 3 and 4
     const std::array<int64_t, 4> per; // permutation of dimensions
+    const bool v; // whether a is a non-contiguous view
 
     std::string vars() override {
-        return VARS_TO_STR8(type_a, type_b, m, n, k, bs, nr, per);
+        return VARS_TO_STR9(type_a, type_b, m, n, k, bs, nr, per, v);
     }
 
     double max_nmse_err() override {
@@ -1982,8 +1996,9 @@ struct test_mul_mat : public test_case {
             int64_t m = 32, int64_t n = 32, int64_t k = 32,
             std::array<int64_t, 2> bs = {10, 10},
             std::array<int64_t, 2> nr = {2, 2},
-            std::array<int64_t, 4> per = {0, 1, 2, 3})
-        : type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr), per(per) {}
+            std::array<int64_t, 4> per = {0, 1, 2, 3},
+            bool v = false)
+        : type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr), per(per), v(v) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         // C^T = A * B^T: (k, m) * (k, n) => (m, n)
@@ -1993,6 +2008,7 @@ struct test_mul_mat : public test_case {
         const int npermuted = (per[0] != 0) + (per[1] != 1) + (per[2] != 2) + (per[3] != 3);
         if (npermuted > 0) {
             GGML_ASSERT(npermuted == 2);
+            GGML_ASSERT(!v); // not handled
             GGML_ASSERT(!ggml_is_quantized(type_a) || per[0] == 0);
             GGML_ASSERT(!ggml_is_quantized(type_b) || per[0] == 0);
 
@@ -2016,7 +2032,13 @@ struct test_mul_mat : public test_case {
             ggml_set_name(a, "a_permuted");
             ggml_set_name(b, "b_permuted");
         } else {
-            a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0],       bs[1]);
+
+            if (v) {
+                a = ggml_new_tensor_4d(ctx, type_a, k*2, m, bs[0], bs[1]);
+                a = ggml_view_4d(ctx, a, k, m, bs[0], bs[1], a->nb[1], a->nb[2], a->nb[3], 0);
+            } else {
+                a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0],       bs[1]);
+            }
             b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
             if (!ggml_is_quantized(type_a)) {
                 if (bs[1] == 1 && nr[1] == 1) {
@@ -3206,11 +3228,12 @@ struct test_flash_attn_ext : public test_case {
     const float max_bias; // ALiBi
     const float logit_softcap; // Gemma 2
 
+    const ggml_prec prec;
     const ggml_type type_KV;
     std::array<int32_t, 4> permute;
 
     std::string vars() override {
-        return VARS_TO_STR10(hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, type_KV, permute);
+        return VARS_TO_STR11(hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, permute);
     }
 
     double max_nmse_err() override {
@@ -3225,9 +3248,9 @@ struct test_flash_attn_ext : public test_case {
     }
 
     test_flash_attn_ext(int64_t hs = 128, int64_t nh = 32, int64_t nr = 1, int64_t kv = 96, int64_t nb = 8,
-                        bool mask = true, float max_bias = 0.0f, float logit_softcap = 0.0f, ggml_type type_KV = GGML_TYPE_F16,
-                        std::array<int32_t, 4> permute = {0, 1, 2, 3})
-        : hs(hs), nh(nh), nr(nr), kv(kv), nb(nb), mask(mask), max_bias(max_bias), logit_softcap(logit_softcap), type_KV(type_KV), permute(permute) {}
+                        bool mask = true, float max_bias = 0.0f, float logit_softcap = 0.0f, ggml_prec prec = GGML_PREC_F32,
+                        ggml_type type_KV = GGML_TYPE_F16, std::array<int32_t, 4> permute = {0, 1, 2, 3})
+        : hs(hs), nh(nh), nr(nr), kv(kv), nb(nb), mask(mask), max_bias(max_bias), logit_softcap(logit_softcap), prec(prec), type_KV(type_KV), permute(permute) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         const int64_t hs_padded = GGML_PAD(hs, ggml_blck_size(type_KV));
@@ -3261,6 +3284,7 @@ struct test_flash_attn_ext : public test_case {
         }
 
         ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, m, 1.0f/sqrtf(hs), max_bias, logit_softcap);
+        ggml_flash_attn_ext_set_prec(out, prec);
         ggml_set_name(out, "out");
 
         return out;
@@ -3989,14 +4013,25 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         test_cases.emplace_back(new test_set(GGML_TYPE_I32, GGML_TYPE_I32, {6, 5, 4, 3}, dim));
     }
 
-    for (ggml_type type_src : {GGML_TYPE_F16, GGML_TYPE_F32}) {
+    // same-type copy
+    for (ggml_type type : all_types) {
+        const auto nk = ggml_blck_size(type);
+
+        for (int k = 1; k < 4; ++k) {
+            test_cases.emplace_back(new test_cpy(type, type, {k*nk, 2, 3, 4}));
+            test_cases.emplace_back(new test_cpy(type, type, {k*nk, 2, 3, 4}, {0, 2, 1, 3}));
+            test_cases.emplace_back(new test_cpy(type, type, {k*nk, 2, 3, 4}, {0, 3, 1, 2}, {0, 2, 1, 3}));
+        }
+    }
+
+    for (ggml_type type_src : {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_F32}) {
         for (ggml_type type_dst : all_types) {
             test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 4, 4, 4}));
             test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 2, 3, 4}, {0, 2, 1, 3})); // cpy by rows
         }
     }
-    for (ggml_type type_dst : {GGML_TYPE_F32}) {
-        for (ggml_type type_src : all_types) {
+    for (ggml_type type_src : all_types) {
+        for (ggml_type type_dst : {GGML_TYPE_F32}) {
             test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 4, 4, 4}));
             test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 2, 3, 4}, {0, 2, 1, 3})); // cpy by rows
         }
@@ -4169,6 +4204,19 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32,  83, 2,   64, { 8,  1}, {4, 1}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32,  64, 45, 128, { 8,  1}, {4, 1}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128, 45,  64, { 8,  1}, {4, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056, 1, 193, {1,  1}, {4, 1}, {0, 2, 1, 3}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056, 1, 67,  {1,  1}, {4, 1}, {0, 2, 1, 3}));
+
+    for (auto bs : {1,2,4,8}) {
+        for (auto nr : {1,4}) {
+            for (uint32_t m = 0; m < 2; ++m) {
+                for (uint32_t k = 0; k < 2; ++k) {
+                    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056 + m, 1, 128 + k,  {bs,  1}, {nr, 1}, {0, 2, 1, 3}));
+                    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128 + m,  1, 1056 + k, {bs,  1}, {nr, 1}, {0, 1, 2, 3}, true));
+                }
+            }
+        }
+    }
 
     // sycl backend will limit task global_range < MAX_INT
     // test case for f16-type-convert-to-fp32 kernel with large k under fp32 compute dtype (occurs in stable-diffusion)
@@ -4376,11 +4424,16 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                             for (int kv : { 512, 1024, }) {
                                 if (nr != 1 && kv != 512) continue;
                                 for (int nb : { 1, 3, 32, 35, }) {
-                                    for (ggml_type type_KV : {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {
-                                        test_cases.emplace_back(new test_flash_attn_ext(hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, type_KV));
-                                        // run fewer test cases permuted
-                                        if (mask == true && max_bias == 0.0f && logit_softcap == 0 && kv == 512) {
-                                            test_cases.emplace_back(new test_flash_attn_ext(hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, type_KV, {0, 2, 1, 3}));
+                                    for (ggml_prec prec : {GGML_PREC_F32, GGML_PREC_DEFAULT}) {
+                                        if (hs != 128 && prec == GGML_PREC_DEFAULT) continue;
+                                        for (ggml_type type_KV : {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {
+                                            test_cases.emplace_back(new test_flash_attn_ext(
+                                                hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV));
+                                            // run fewer test cases permuted
+                                            if (mask == true && max_bias == 0.0f && logit_softcap == 0 && kv == 512) {
+                                                test_cases.emplace_back(new test_flash_attn_ext(
+                                                    hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, {0, 2, 1, 3}));
+                                            }
                                         }
                                     }
                                 }
@@ -4433,6 +4486,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {1024, 10, 1, 1}));
     test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {32000, 512, 1, 1}));
 
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 16416, 1, 128, {8,  1}, {4, 1}, {0, 2, 1, 3}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128, 1, 16416, {8,  1}, {4, 1}, {0, 1, 2, 3}, true));
+
     for (int bs : {1, 2, 3, 4, 5, 8, 512}) {
         for (ggml_type type_a : all_types) {
             for (ggml_type type_b : {GGML_TYPE_F32}) {

tests/test-grammar-llguidance.cpp

@@ -1086,6 +1086,65 @@ static void test_json_schema() {
         });
 }
 
+static void one_hot(llama_token_data_array & tok_arr, llama_token selected) {
+    auto n_vocab = tok_arr.size;
+
+    tok_arr.selected = -1;
+    tok_arr.sorted   = false;
+    for (llama_token token_id = 0; token_id < (llama_token) n_vocab; token_id++) {
+        tok_arr.data[token_id].id    = token_id;
+        tok_arr.data[token_id].logit = 0.0f;
+    }
+
+    tok_arr.data[selected].logit = 100.0f;
+}
+
+static void test_sampler_chain(void) {
+    auto sparams            = llama_sampler_chain_default_params();
+    sparams.no_perf         = false;
+    llama_sampler * sampler = llama_sampler_chain_init(sparams);
+
+    const auto grammar_data = R"(%llguidance {}
+start: /[A-Z ]*/)";
+
+    llama_sampler_chain_add(sampler, llama_sampler_init_llg(vocab, "lark", grammar_data));
+    llama_sampler_chain_add(sampler, llama_sampler_init_dist(42));
+
+    auto input  = "ALL YOUR BASE ARE BELONG TO US";
+    auto tokens = common_tokenize(vocab, input, false, false);
+
+    auto n_vocab = llama_vocab_n_tokens(vocab);
+
+    std::vector<llama_token_data> cur;
+    cur.reserve(n_vocab);
+    for (llama_token token_id = 0; token_id < (llama_token) n_vocab; token_id++) {
+        cur.emplace_back(llama_token_data{ token_id, 0.0f, 0.0f });
+    }
+    auto tok_arr = llama_token_data_array{ cur.data(), cur.size(), -1, false };
+
+    for (const auto token : tokens) {
+        one_hot(tok_arr, token);
+
+        fprintf(stderr, "applying token: %d\n", token);
+        llama_sampler_apply(sampler, &tok_arr);
+
+        auto idx = tok_arr.selected;
+        fprintf(stderr, " -> %d %f\n", cur[idx].id, cur[idx].logit);
+        assert(cur[tok_arr.selected].id == token);
+        llama_sampler_accept(sampler, token);
+    }
+
+    auto tok_eos = llama_vocab_eot(vocab);
+    if (tok_eos == LLAMA_TOKEN_NULL) {
+        tok_eos = llama_vocab_eos(vocab);
+    }
+
+    one_hot(tok_arr, tok_eos);
+
+    llama_sampler_apply(sampler, &tok_arr);
+    assert(cur[tok_arr.selected].id == tok_eos);
+}
+
 int main(int argc, const char ** argv) {
     fprintf(stdout, "Running llguidance integration tests...\n");
 
@@ -1135,6 +1194,9 @@ int main(int argc, const char ** argv) {
     test_special_chars();
     test_quantifiers();
     test_json_schema();
+
+    test_sampler_chain();
+
     fprintf(stdout, "All tests passed.\n");
     return 0;
 }