vulkan: tune MMVQ for Intel Windows (#19988)

* scripts : improve get-wikitext-2.sh

Switch to sh, add curl fallback, and avoid redundant downloads

Signed-off-by: Adrien Gallouët <adrien@gallouet.fr>

* fix indent

Signed-off-by: Adrien Gallouët <angt@huggingface.co>

---------

Signed-off-by: Adrien Gallouët <adrien@gallouet.fr>
Signed-off-by: Adrien Gallouët <angt@huggingface.co>

.github/actions/windows-setup-rocm/action.yml

@@ -11,5 +11,5 @@ runs:
     - name: Setup ROCm
       uses: ./.github/actions/install-exe
       with:
-        url: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ inputs.version }}-WinSvr2022-For-HIP.exe
+        url: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ inputs.version }}-Win11-For-HIP.exe
         args: -install

.github/workflows/build-cache.yml

@@ -68,7 +68,7 @@ jobs:
 
     env:
       # Make sure this is in sync with build.yml
-      HIPSDK_INSTALLER_VERSION: "25.Q3"
+      HIPSDK_INSTALLER_VERSION: "26.Q1"
 
     steps:
       - name: Clone

.github/workflows/build.yml

@@ -1175,10 +1175,8 @@ jobs:
     runs-on: windows-2022
 
     env:
-      # The ROCm version must correspond to the version used in the HIP SDK.
-      ROCM_VERSION: "6.4.2"
       # Make sure this is in sync with build-cache.yml
-      HIPSDK_INSTALLER_VERSION: "25.Q3"
+      HIPSDK_INSTALLER_VERSION: "26.Q1"
 
     steps:
       - name: Clone
@@ -1188,7 +1186,7 @@ jobs:
       - name: Grab rocWMMA package
         id: grab_rocwmma
         run: |
-          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/${{ env.ROCM_VERSION }}/pool/main/r/rocwmma-dev/rocwmma-dev_1.7.0.60402-120~24.04_amd64.deb"
+          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/7.2/pool/main/r/rocwmma-dev/rocwmma-dev_2.2.0.70200-43~24.04_amd64.deb"
           7z x rocwmma.deb
           7z x data.tar
 
@@ -1231,7 +1229,7 @@ jobs:
           cmake -G "Unix Makefiles" -B build -S . `
             -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
             -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
-            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-${{ env.ROCM_VERSION }}/include/" `
+            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-7.2.0/include/" `
             -DCMAKE_BUILD_TYPE=Release `
             -DLLAMA_BUILD_BORINGSSL=ON `
             -DROCM_DIR="${env:HIP_PATH}" `

.github/workflows/gguf-publish.yml

@@ -21,7 +21,7 @@ on:
 jobs:
   deploy:
 
-    runs-on: ubuntu-slim
+    runs-on: ubuntu-latest
 
     steps:
     - uses: actions/checkout@v6

.github/workflows/release.yml

@@ -616,13 +616,13 @@ jobs:
     runs-on: windows-2022
 
     env:
-      HIPSDK_INSTALLER_VERSION: "25.Q3"
+      HIPSDK_INSTALLER_VERSION: "26.Q1"
 
     strategy:
       matrix:
         include:
           - name: "radeon"
-            gpu_targets: "gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
+            gpu_targets: "gfx1150;gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
 
     steps:
       - name: Clone
@@ -632,7 +632,7 @@ jobs:
       - name: Grab rocWMMA package
         id: grab_rocwmma
         run: |
-          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/7.0.1/pool/main/r/rocwmma-dev/rocwmma-dev_2.0.0.70001-42~24.04_amd64.deb"
+          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/7.2/pool/main/r/rocwmma-dev/rocwmma-dev_2.2.0.70200-43~24.04_amd64.deb"
           7z x rocwmma.deb
           7z x data.tar
 
@@ -655,7 +655,7 @@ jobs:
         run: |
           $ErrorActionPreference = "Stop"
           write-host "Downloading AMD HIP SDK Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ env.HIPSDK_INSTALLER_VERSION }}-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ env.HIPSDK_INSTALLER_VERSION }}-Win11-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
           write-host "Installing AMD HIP SDK"
           $proc = Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -PassThru
           $completed = $proc.WaitForExit(600000)
@@ -689,20 +689,20 @@ jobs:
           cmake -G "Unix Makefiles" -B build -S . `
             -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
             -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
-            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-7.0.1/include/ -Wno-ignored-attributes -Wno-nested-anon-types" `
+            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-7.2.0/include/ -Wno-ignored-attributes -Wno-nested-anon-types" `
             -DCMAKE_BUILD_TYPE=Release `
             -DGGML_BACKEND_DL=ON `
             -DGGML_NATIVE=OFF `
             -DGGML_CPU=OFF `
-            -DAMDGPU_TARGETS="${{ matrix.gpu_targets }}" `
+            -DGPU_TARGETS="${{ matrix.gpu_targets }}" `
             -DGGML_HIP_ROCWMMA_FATTN=ON `
             -DGGML_HIP=ON `
             -DLLAMA_BUILD_BORINGSSL=ON
           cmake --build build --target ggml-hip -j ${env:NUMBER_OF_PROCESSORS}
           md "build\bin\rocblas\library\"
           md "build\bin\hipblaslt\library"
-          cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
-          cp "${env:HIP_PATH}\bin\hipblaslt.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\libhipblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\libhipblaslt.dll" "build\bin\"
           cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
           cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"
           cp "${env:HIP_PATH}\bin\hipblaslt\library\*" "build\bin\hipblaslt\library\"

common/arg.cpp

@@ -1578,7 +1578,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_sparam());
     add_opt(common_arg(
-        {"--temp"}, "N",
+        {"--temp", "--temperature"}, "N",
         string_format("temperature (default: %.2f)", (double)params.sampling.temp),
         [](common_params & params, const std::string & value) {
             params.sampling.temp = std::stof(value);
@@ -1611,7 +1611,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_sparam());
     add_opt(common_arg(
-        {"--top-nsigma"}, "N",
+        {"--top-nsigma", "--top-n-sigma"}, "N",
         string_format("top-n-sigma sampling (default: %.2f, -1.0 = disabled)", params.sampling.top_n_sigma),
         [](common_params & params, const std::string & value) {
             params.sampling.top_n_sigma = std::stof(value);
@@ -1634,7 +1634,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_sparam());
     add_opt(common_arg(
-        {"--typical"}, "N",
+        {"--typical", "--typical-p"}, "N",
         string_format("locally typical sampling, parameter p (default: %.2f, 1.0 = disabled)", (double)params.sampling.typ_p),
         [](common_params & params, const std::string & value) {
             params.sampling.typ_p = std::stof(value);
@@ -2520,11 +2520,28 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ));
     add_opt(common_arg(
         {"-a", "--alias"}, "STRING",
-        "set alias for model name (to be used by REST API)",
+        "set model name aliases, comma-separated (to be used by API)",
         [](common_params & params, const std::string & value) {
-            params.model_alias = value;
+            for (auto & alias : string_split<std::string>(value, ',')) {
+                alias = string_strip(alias);
+                if (!alias.empty()) {
+                    params.model_alias.insert(alias);
+                }
+            }
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_ALIAS"));
+    add_opt(common_arg(
+        {"--tags"}, "STRING",
+        "set model tags, comma-separated (informational, not used for routing)",
+        [](common_params & params, const std::string & value) {
+            for (auto & tag : string_split<std::string>(value, ',')) {
+                tag = string_strip(tag);
+                if (!tag.empty()) {
+                    params.model_tags.insert(tag);
+                }
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_TAGS"));
     add_opt(common_arg(
         {"-m", "--model"}, "FNAME",
         ex == LLAMA_EXAMPLE_EXPORT_LORA

common/common.h

@@ -410,7 +410,8 @@ struct common_params {
 
     struct common_params_model model;
 
-    std::string model_alias          = ""; // model alias                                                   // NOLINT
+    std::set<std::string> model_alias;     // model aliases                                                 // NOLINT
+    std::set<std::string> model_tags;      // model tags (informational, not used for routing)              // NOLINT
     std::string hf_token             = ""; // HF token                                                      // NOLINT
     std::string prompt               = "";                                                                  // NOLINT
     std::string system_prompt        = "";                                                                  // NOLINT

common/jinja/runtime.cpp

@@ -721,6 +721,8 @@ value member_expression::execute_impl(context & ctx) {
         int64_t arr_size = 0;
         if (is_val<value_array>(object)) {
             arr_size = object->as_array().size();
+        } else if (is_val<value_string>(object)) {
+            arr_size = object->as_string().length();
         }
 
         if (is_stmt<slice_expression>(this->property)) {

convert_hf_to_gguf.py

@@ -116,7 +116,8 @@ class ModelBase:
                  split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False,
                  small_first_shard: bool = False, hparams: dict[str, Any] | None = None, remote_hf_model_id: str | None = None,
                  disable_mistral_community_chat_template: bool = False,
-                 sentence_transformers_dense_modules: bool = False):
+                 sentence_transformers_dense_modules: bool = False,
+                 fuse_gate_up_exps: bool = False):
         if type(self) is ModelBase or \
                 type(self) is TextModel or \
                 type(self) is MmprojModel:
@@ -135,6 +136,9 @@ class ModelBase:
         self.dry_run = dry_run
         self.remote_hf_model_id = remote_hf_model_id
         self.sentence_transformers_dense_modules = sentence_transformers_dense_modules
+        self.fuse_gate_up_exps = fuse_gate_up_exps
+        self._gate_exp_buffer: dict[int, Tensor] = {}
+        self._up_exp_buffer: dict[int, Tensor] = {}
         self.hparams = ModelBase.load_hparams(self.dir_model, self.is_mistral_format) if hparams is None else hparams
         self.model_tensors = self.index_tensors(remote_hf_model_id=remote_hf_model_id)
         self.metadata_override = metadata_override
@@ -512,8 +516,31 @@ class ModelBase:
         raise NotImplementedError("set_gguf_parameters() must be implemented in subclasses")
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid # unused
-        return [(self.map_tensor_name(name), data_torch)]
+        new_name = self.map_tensor_name(name)
+
+        # Handle gate/up expert tensor fusion if enabled
+        if self.fuse_gate_up_exps and bid is not None:
+            if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.FFN_GATE_EXP, bid):
+                self._gate_exp_buffer[bid] = data_torch
+            elif self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.FFN_UP_EXP, bid):
+                self._up_exp_buffer[bid] = data_torch
+
+            # Check if both gate and up are buffered for this layer
+            if bid in self._gate_exp_buffer and bid in self._up_exp_buffer:
+                gate_data = self._gate_exp_buffer.pop(bid)
+                up_data = self._up_exp_buffer.pop(bid)
+                # gate/up shape: (n_expert, n_ff, n_embd), concatenate to (n_expert, n_ff*2, n_embd)
+                fused_data = torch.cat([gate_data, up_data], dim=1)
+                fused_name = self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_UP_EXP, bid)
+                logger.info(f"Fused gate_exps and up_exps for layer {bid}")
+                return [(fused_name, fused_data)]
+
+            # If we buffered a gate/up tensor, wait for the other
+            if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.FFN_GATE_EXP, bid) or \
+               self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.FFN_UP_EXP, bid):
+                return []
+
+        return [(new_name, data_torch)]
 
     def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
         del name, new_name, bid, n_dims  # unused
@@ -1148,6 +1175,9 @@ class TextModel(ModelBase):
         if chkhsh == "27949a2493fc4a9f53f5b9b029c82689cfbe5d3a1929bb25e043089e28466de6":
             # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-de
             res = "jina-v2-de"
+        if chkhsh == "a023e9fdc5a11f034d3ef515b92350e56fb2af1f66c6b6811a4444ea9bf8763d":
+            # ref: https://huggingface.co/jinaai/jina-embeddings-v5-text-nano
+            res = "jina-v5-nano"
         if chkhsh == "c136ed14d01c2745d4f60a9596ae66800e2b61fa45643e72436041855ad4089d":
             # ref: https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct
             res = "smaug-bpe"
@@ -6125,6 +6155,32 @@ class NeoBert(BertModel):
         yield from super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register("EuroBertModel", "JinaEmbeddingsV5Model")
+class EuroBertModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.EUROBERT
+
+    def set_vocab(self):
+        self.gguf_writer.add_add_bos_token(False)
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        # EuroBert is bidirectional (encoder)
+        self.gguf_writer.add_causal_attention(False)
+
+        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+
+        self._try_set_pooling_type()
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Strip "model." prefix from tensor names
+        if name.startswith("model."):
+            name = name[6:]
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("XLMRobertaModel", "XLMRobertaForSequenceClassification")
 class XLMRobertaModel(BertModel):
     model_arch = gguf.MODEL_ARCH.BERT
@@ -11913,6 +11969,11 @@ def parse_args() -> argparse.Namespace:
               "Default these modules are not included.")
     )
 
+    parser.add_argument(
+        "--fuse-gate-up-exps", action="store_true",
+        help="Fuse gate_exps and up_exps tensors into a single gate_up_exps tensor for MoE models.",
+    )
+
     args = parser.parse_args()
     if not args.print_supported_models and args.model is None:
         parser.error("the following arguments are required: model")
@@ -12050,7 +12111,8 @@ def main() -> None:
                                      split_max_size=split_str_to_n_bytes(args.split_max_size), dry_run=args.dry_run,
                                      small_first_shard=args.no_tensor_first_split,
                                      remote_hf_model_id=hf_repo_id, disable_mistral_community_chat_template=disable_mistral_community_chat_template,
-                                     sentence_transformers_dense_modules=args.sentence_transformers_dense_modules
+                                     sentence_transformers_dense_modules=args.sentence_transformers_dense_modules,
+                                     fuse_gate_up_exps=args.fuse_gate_up_exps
                                      )
 
         if args.vocab_only:

convert_hf_to_gguf_update.py

@@ -107,6 +107,7 @@ models = [
     {"name": "jina-v2-en",       "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", }, # WPM!
     {"name": "jina-v2-es",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
     {"name": "jina-v2-de",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
+    {"name": "jina-v5-nano",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v5-text-nano", },
     {"name": "smaug-bpe",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct", },
     {"name": "poro-chat",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Poro-34B-chat", },
     {"name": "jina-v2-code",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-code", },

docs/backend/VirtGPU.md

@@ -152,7 +152,9 @@ Commands and data are serialized using a custom binary protocol with:
 - **VM-specific**: Only works in virtual machines with virtio-gpu support
 - **Host dependency**: Requires properly configured host-side backend
 - **Latency**: Small overhead from VM escaping for each operation
-
+- **Shared-memory size**: with the `libkrun` hypervisor, the RAM + VRAM
+  addressable memory is limited to 64 GB. So the maximum GPU memory
+  will be `64GB - RAM`, regardless of the hardware VRAM size.
 
 * This work is pending upstream changes in the VirglRenderer
   project.

docs/backend/ZenDNN.md

@@ -22,7 +22,7 @@
 
 **Llama.cpp + ZenDNN**
 
-The llama.cpp ZenDNN backend leverages AMD's optimized matrix multiplication primitives to accelerate inference on AMD CPUs. It utilizes ZenDNN's **LowOHA (Low Overhead Hardware Accelerated)** MatMul operator for efficient GEMM operations with minimal execution overhead, built-in weight caching, and direct access to backend libraries (AOCL BLIS, LibXSMM, OneDNN).
+The llama.cpp ZenDNN backend leverages AMD's optimized matrix multiplication primitives to accelerate inference on AMD CPUs. It utilizes ZenDNN's **LowOHA (Low Overhead Hardware Accelerated)** MatMul operator for efficient GEMM operations with minimal execution overhead, built-in weight caching, and direct access to backend libraries (AOCL DLP, LibXSMM, OneDNN).
 
 For more information about ZenDNN, visit: https://www.amd.com/en/developer/zendnn.html
 
@@ -32,7 +32,7 @@ For more information about ZenDNN, visit: https://www.amd.com/en/developer/zendn
 |:-------:|:-------:|:----------------------------------------------:|
 | Linux   | Support | Ubuntu 20.04, 22.04, 24.04                     |
 
-For the latest list of supported operating systems, see the [ZenDNN Supported OS](https://github.com/amd/ZenDNN/blob/zendnnl/README.md#15-supported-os).
+For the latest list of supported operating systems, see the [ZenDNN Supported OS](https://github.com/amd/ZenDNN/blob/a18adf8c605fb5f5e52cefd7eda08a7b18febbaf/README.md#15-supported-os).
 
 ## Hardware
 
@@ -44,9 +44,9 @@ ZenDNN is optimized for AMD EPYC™ processors and AMD Ryzen™ processors based
 
 | CPU Family                    | Status  | Notes                              |
 |:-----------------------------:|:-------:|:----------------------------------:|
-| AMD EPYC™ 9005 Series (Turin)| Support | 5th Gen - Zen 5 architecture       |
-| AMD EPYC™ 9004 Series (Genoa)| Support | 4th Gen - Zen 4 architecture       |
-| AMD EPYC™ 7003 Series (Milan)| Support | 3rd Gen - Zen 3 architecture       |
+| AMD EPYC™ 9005 Series (Turin) | Support | 5th Gen - Zen 5 architecture       |
+| AMD EPYC™ 9004 Series (Genoa) | Support | 4th Gen - Zen 4 architecture       |
+| AMD EPYC™ 7003 Series (Milan) | Support | 3rd Gen - Zen 3 architecture       |
 | AMD Ryzen™ AI MAX (Strix Halo)| Support | High-performance mobile processors |
 
 *Notes:*
@@ -61,7 +61,7 @@ The ZenDNN backend currently accelerates **matrix multiplication (MUL_MAT)** ope
 
 | Operation    | Status  | Notes                                          |
 |:-------------|:-------:|:----------------------------------------------:|
-| MUL_MAT      |    ✓    | Accelerated via ZenDNN LowOHA MatMul           |
+| MUL_MAT      | Support | Accelerated via ZenDNN LowOHA MatMul           |
 
 *Note:* Since only MUL_MAT is accelerated, models will benefit most from ZenDNN when matrix multiplications dominate the computational workload (which is typical for transformer-based LLMs).
 
@@ -104,7 +104,6 @@ If you want to build ZenDNN yourself or use a specific version:
 # Clone ZenDNN repository
 git clone https://github.com/amd/ZenDNN.git
 cd ZenDNN
-git checkout zendnnl
 
 # Build and install (requires CMake >= 3.25)
 mkdir build && cd build
@@ -114,7 +113,7 @@ cmake --build . --target all
 
 Default installation path: `ZenDNN/build/install`
 
-**For detailed build instructions**, refer to the [ZenDNN README](https://github.com/amd/ZenDNN/blob/zendnnl/README.md).
+**For detailed build instructions**, refer to the [ZenDNN README](https://github.com/amd/ZenDNN/blob/a18adf8c605fb5f5e52cefd7eda08a7b18febbaf/README.md).
 
 **Step 2: Build llama.cpp with custom ZenDNN path**
 
@@ -146,8 +145,7 @@ Run llama.cpp server with ZenDNN acceleration:
 
 ```sh
 # Set optimal configuration
-export OMP_NUM_THREADS=64  # Adjust to your CPU core count
-export ZENDNNL_MATMUL_ALGO=2  # Blocked AOCL BLIS for best performance
+export ZENDNNL_MATMUL_ALGO=1    # Blocked AOCL DLP algo for best performance
 
 # Start server
 ./build/bin/llama-server \
@@ -160,62 +158,26 @@ export ZENDNNL_MATMUL_ALGO=2  # Blocked AOCL BLIS for best performance
 Access the server at `http://localhost:8080`.
 
 **Performance tips**:
-- Set `OMP_NUM_THREADS` to match your physical core count
-- Use `ZENDNNL_MATMUL_ALGO=2` for optimal performance
+- Use `ZENDNNL_MATMUL_ALGO=1` for optimal performance
 - For NUMA systems: `numactl --cpunodebind=0 --membind=0 ./build/bin/llama-server ...`
 
 ## Environment Variable
 
-### Build Time
+For environment variables related to ZenDNN, refer to the [ZenDNN Environment Variables Documentation](https://github.com/amd/ZenDNN/blob/a18adf8c605fb5f5e52cefd7eda08a7b18febbaf/docs/runtime_env.md).
 
-| Name               | Value                                 | Function                                    |
-|--------------------|---------------------------------------|---------------------------------------------|
-| GGML_ZENDNN        | ON/OFF                                | Enable ZenDNN backend support               |
-| ZENDNN_ROOT        | Path to ZenDNN installation           | Set ZenDNN installation directory           |
-| GGML_OPENMP        | ON/OFF (recommended: ON)              | Enable OpenMP for multi-threading           |
+### Performance Optimization
 
-### Runtime
-
-| Name                    | Value                    | Function                                                          |
-|-------------------------|--------------------------|-------------------------------------------------------------------|
-| OMP_NUM_THREADS         | Number (e.g., 64)        | Set number of OpenMP threads (recommended: physical core count)   |
-| ZENDNNL_MATMUL_ALGO     | 0-5                      | Select MatMul backend algorithm (see Performance Optimization)    |
-| ZENDNNL_PROFILE_LOG_LEVEL | 0-4                    | Profiling log level (0=disabled, 4=verbose)                       |
-| ZENDNNL_ENABLE_PROFILER | 0 or 1                   | Enable detailed profiling (1=enabled)                             |
-| ZENDNNL_API_LOG_LEVEL   | 0-4                      | API log level (0=disabled, 4=verbose)                             |
-
-**Example**:
+ZenDNN's LowOHA MatMul supports multiple backend algorithms. For **best performance**, use the **Blocked AOCL DLP** algorithm:
 
 ```sh
-export OMP_NUM_THREADS=64
-export ZENDNNL_MATMUL_ALGO=2  # Use Blocked AOCL BLIS for best performance
-./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Test" -n 100
+export ZENDNNL_MATMUL_ALGO=1    # Blocked AOCL DLP algo (recommended)
 ```
 
-## Performance Optimization
-
-### MatMul Algorithm Selection
-
-ZenDNN's LowOHA MatMul supports multiple backend algorithms. For **best performance**, use the **Blocked AOCL BLIS** algorithm:
-
-```sh
-export ZENDNNL_MATMUL_ALGO=2  # Blocked AOCL BLIS (recommended)
-```
-
-**Available algorithms**:
-
-| Value | Algorithm              | Description                                    |
-|:-----:|:-----------------------|:----------------------------------------------|
-| 0     | Dynamic Dispatch       | Automatic backend selection (default)         |
-| 1     | AOCL BLIS              | AOCL BLIS backend                             |
-| 2     | AOCL BLIS Blocked      | **Blocked AOCL BLIS (recommended)**           |
-| 3     | OneDNN                 | OneDNN backend                                |
-| 4     | OneDNN Blocked         | Blocked OneDNN                                |
-| 5     | LibXSMM                | LibXSMM backend                               |
+For more details on available algorithms, see the [ZenDNN MatMul Algorithm Documentation](https://github.com/amd/ZenDNN/blob/a18adf8c605fb5f5e52cefd7eda08a7b18febbaf/docs/runtime_env.md#algorithm-details).
 
 ### Profiling and Debugging
 
-For detailed profiling and logging options, refer to the [ZenDNN Logging Documentation](https://github.com/amd/ZenDNN/blob/zendnnl/docs/logging.md).
+For detailed profiling and logging options, refer to the [ZenDNN Logging Documentation](https://github.com/amd/ZenDNN/blob/a18adf8c605fb5f5e52cefd7eda08a7b18febbaf/docs/logging.md).
 
 ## Known Issues
 
@@ -245,10 +207,9 @@ A: Currently, ZenDNN primarily supports FP32 and BF16 data types. Quantized mode
 
 A: Ensure:
 1. You're using an AMD EPYC or Ryzen processor (Zen 2 or newer)
-2. `OMP_NUM_THREADS` is set appropriately (physical core count)
-3. `ZENDNNL_MATMUL_ALGO=2` is set for best performance (Blocked AOCL BLIS)
-4. You're using a sufficiently large model (small models may not benefit as much)
-5. Enable profiling to verify ZenDNN MatMul is being called
+2. `ZENDNNL_MATMUL_ALGO=1` is set for best performance (Blocked AOCL DLP)
+3. You're using a sufficiently large model (small models may not benefit as much)
+4. Enable profiling to verify ZenDNN MatMul is being called
 
 ### **GitHub Contribution**:
 Please add the **[ZenDNN]** prefix/tag in issues/PRs titles to help the ZenDNN-team check/address them without delay.

ggml/include/ggml.h

@@ -730,10 +730,6 @@ extern "C" {
     GGML_API size_t  ggml_type_size(enum ggml_type type);             // size in bytes for all elements in a block
     GGML_API size_t  ggml_row_size (enum ggml_type type, int64_t ne); // size in bytes for all elements in a row
 
-    GGML_DEPRECATED(
-    GGML_API double ggml_type_sizef(enum ggml_type type), // ggml_type_size()/ggml_blck_size() as float
-    "use ggml_row_size() instead");
-
     GGML_API const char * ggml_type_name(enum ggml_type type);
     GGML_API const char * ggml_op_name  (enum ggml_op   op);
     GGML_API const char * ggml_op_symbol(enum ggml_op   op);

ggml/src/ggml-cpu/amx/amx.cpp

@@ -141,27 +141,50 @@ static size_t ggml_backend_amx_buffer_type_get_alignment(ggml_backend_buffer_typ
 namespace ggml::cpu::amx {
 class extra_buffer_type : ggml::cpu::extra_buffer_type {
     bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override {
-        // handle only 2d gemm for now
-        auto is_contiguous_2d = [](const struct ggml_tensor * t) {
-            return ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1;
-        };
-
-        if (op->op == GGML_OP_MUL_MAT && is_contiguous_2d(op->src[0]) &&  // src0 must be contiguous
-            is_contiguous_2d(op->src[1]) &&                               // src1 must be contiguous
-            op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_amx_buffer_type() &&
-            op->src[0]->ne[0] % (TILE_K * 2 * 32) == 0 && // TODO: not sure if correct (https://github.com/ggml-org/llama.cpp/pull/16315)
-            op->ne[0] % (TILE_N * 2) == 0 &&                              // out_features is 32x
-            (qtype_has_amx_kernels(op->src[0]->type) || (op->src[0]->type == GGML_TYPE_F16))) {
-            // src1 must be host buffer
-            if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
-                return false;
-            }
-            // src1 must be float32
-            if (op->src[1]->type == GGML_TYPE_F32) {
-                return true;
-            }
+        if (op->op != GGML_OP_MUL_MAT) {
+            return false;
         }
-        return false;
+        auto * src0 = op->src[0];
+        auto * src1 = op->src[1];
+
+        if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1)) {
+            return false;
+        }
+        if (!src0->buffer || src0->buffer->buft != ggml_backend_amx_buffer_type()) {
+            return false;
+        }
+        if (src1->buffer && !ggml_backend_buft_is_host(src1->buffer->buft)) {
+            return false;
+        }
+        if (op->ne[0] % (TILE_N * 2)) {
+            return false;
+        }
+        int alignment;
+        switch (src0->type) {
+            case GGML_TYPE_Q4_0:
+            case GGML_TYPE_Q4_1:
+            case GGML_TYPE_Q8_0:
+                alignment = TILE_K;
+                break;
+            case GGML_TYPE_Q4_K:
+            case GGML_TYPE_Q5_K:
+            case GGML_TYPE_Q6_K:
+            case GGML_TYPE_IQ4_XS:
+                alignment = 256; // QK_K
+                break;
+            case GGML_TYPE_F16:
+                alignment = 16;
+                break;
+            default:
+                return false;
+        }
+        if (src0->ne[0] % alignment) {
+            return false;
+        }
+        if (src1->type != GGML_TYPE_F32) {
+            return false;
+        }
+        return true;
     }
 
     ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override {

ggml/src/ggml-cpu/amx/mmq.cpp

@@ -1,4 +1,3 @@
-
 #if defined(__GNUC__)
 #pragma GCC diagnostic ignored "-Wpedantic"
 #pragma GCC diagnostic ignored "-Wunused-local-typedefs"
@@ -202,35 +201,27 @@ struct tile_config_t{
 //    advanced-matrix-extensions-intrinsics-functions.html
 //
 
-#define TC_CONFIG_TILE(i, r, cb) tc.rows[i] = r; tc.colsb[i] = cb
-void ggml_tile_config_init(void) {
-    static thread_local bool is_first_time = true;
+inline void ggml_tile_config_init(void) {
+    static thread_local bool done = false;
 
-    if (!is_first_time) {
+    if (done) {
         return;
     }
 
-    static thread_local tile_config_t tc;
-    tile_config_t current_tc;
-    _tile_storeconfig(&current_tc);
+    alignas(64) tile_config_t tc = {};
+    tc.palette_id = 1;
+    tc.start_row = 0;
+    tc.rows[0] = 8;   tc.colsb[0] = 64;
+    tc.rows[1] = 8;   tc.colsb[1] = 64;
+    tc.rows[2] = 16;  tc.colsb[2] = 32;
+    tc.rows[3] = 16;  tc.colsb[3] = 32;
+    tc.rows[4] = 16;  tc.colsb[4] = 64;
+    tc.rows[5] = 16;  tc.colsb[5] = 64;
+    tc.rows[6] = 16;  tc.colsb[6] = 64;
+    tc.rows[7] = 16;  tc.colsb[7] = 64;
 
-    // load only when config changes
-    if (tc.palette_id == 0 || (memcmp(&current_tc.colsb, &tc.colsb, sizeof(uint16_t) * 8) != 0 &&
-                               memcmp(&current_tc.rows, &tc.rows, sizeof(uint8_t) * 8) != 0)) {
-        tc.palette_id = 1;
-        tc.start_row = 0;
-        TC_CONFIG_TILE(TMM0, 8, 64);
-        TC_CONFIG_TILE(TMM1, 8, 64);
-        TC_CONFIG_TILE(TMM2, 16, 32);
-        TC_CONFIG_TILE(TMM3, 16, 32);
-        TC_CONFIG_TILE(TMM4, 16, 64);
-        TC_CONFIG_TILE(TMM5, 16, 64);
-        TC_CONFIG_TILE(TMM6, 16, 64);
-        TC_CONFIG_TILE(TMM7, 16, 64);
-        _tile_loadconfig(&tc);
-    }
-
-    is_first_time = false;
+    _tile_loadconfig(&tc);
+    done = true;
 }
 
 // we need an extra 16 * 4B (TILE_N * int32_t) for each NB/KB block for compensation.
@@ -268,33 +259,6 @@ int get_row_size(int K) {
     return row_size;
 }
 
-// vectorized dtype conversion
-inline float FP16_TO_FP32(ggml_half val) {
-    __m256i v = _mm256_setr_epi16(
-        val, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
-    __m512 o = _mm512_cvtph_ps(v);
-    return _mm512_cvtss_f32(o);
-}
-
-inline __m512 FP16_TO_FP32_VEC(ggml_half val) {
-    __m256i v = _mm256_set1_epi16(val);
-    return _mm512_cvtph_ps(v);
-}
-
-// horizontal reduce
-inline float _mm512_reduce_max_ps(const __m512 x) {
-    __m512 v = x;
-    __m512 v1 = _mm512_shuffle_f32x4(v, v, 0x4E);
-    v = _mm512_max_ps(v, v1);
-    v1 = _mm512_shuffle_f32x4(v, v, 0xB1);
-    v = _mm512_max_ps(v, v1);
-    v1 = _mm512_shuffle_ps(v, v, 0x4E);
-    v = _mm512_max_ps(v, v1);
-    v1 = _mm512_shuffle_ps(v, v, 0xB1);
-    v = _mm512_max_ps(v, v1);
-    return _mm512_cvtss_f32(v);
-}
-
 // transpose utils
 #define SHUFFLE_EPI32(a, b, mask) \
     _mm256_castps_si256(_mm256_shuffle_ps(_mm256_castsi256_ps(a), _mm256_castsi256_ps(b), mask))
@@ -1370,9 +1334,9 @@ struct tinygemm_kernel_avx<float, ggml_fp16_t, float, BLOCK_M, BLOCK_N, BLOCK_K>
 
 #define LAUNCH_TINYGEMM_KERNEL_AVX(MB_SIZE, NB_SIZE)                                \
     tinygemm_kernel_avx<float, type, float, MB_SIZE, NB_SIZE, blck_size>::apply(    \
-        K, (const float *)src1->data + mb_start * K,                                \
-        (const type *)src0->data + nb_start * K,                                    \
-        (float *)dst->data + mb_start * ldc + nb_start, ldc);
+        K, (const float *)src1->data + src1_offset + mb_start * K,                  \
+        (const type *)src0->data + src0_offset + nb_start * K,                      \
+        (float *)dst->data + dst_offset + mb_start * ldc + nb_start, ldc)
 
 
 // re-organize in the format {NB, KB, TILE_SIZE}:
@@ -2019,11 +1983,11 @@ struct tinygemm_kernel_vnni<block_q8_K, block_iq4_xs, float, BLOCK_M, BLOCK_N, B
     }
 };
 
-#define LAUNCH_TINYGEMM_KERNEL_VNNI(NB_SIZE)                                         \
-    tinygemm_kernel_vnni<vec_dot_type, type, float, 1, NB_SIZE, blck_size>::apply(   \
-        KB, (const char *)wdata + 0 * row_size_A,                                    \
-        (const char *)src0->data + PACKED_INDEX(nb * kTilesN, 0, KB, TILE_SIZE),     \
-        (float *) dst->data + 0 * N + nb_start, ldc)
+#define LAUNCH_TINYGEMM_KERNEL_VNNI(NB_SIZE)                                                   \
+    tinygemm_kernel_vnni<vec_dot_type, type, float, 1, NB_SIZE, blck_size>::apply(             \
+        KB, wdata_batch,                                                                       \
+        (const char *)src0->data + src0_offset + PACKED_INDEX(nb * kTilesN, 0, KB, TILE_SIZE), \
+        (float *) dst->data + dst_offset + nb_start, ldc)
 
 template <typename TA, typename TB, typename TC, int BLOCK_K,
           typename std::enable_if<!is_type_qkk<TB>::value, int>::type = 0>
@@ -2079,7 +2043,7 @@ void tinygemm_kernel_amx(int M, int N, int KB, const void * RESTRICT _A, const v
         _tile_stored(TMM5, Tile5(C_pre), TILE_N * sizeof(int32_t));
 
         if (need_unpack) {
-            unpack_B<TB>(Tile1, B_blk0);
+            unpack_B<TB>(Tile1, B_blk1);
             _tile_loadd(TMM1, Tile1, TILE_N * VNNI_BLK);
         } else {
             _tile_loadd(TMM1, B_blk1, TILE_N * VNNI_BLK);
@@ -2336,6 +2300,13 @@ void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * d
     });
 }
 
+// ne2 is passed explicitly to help compiler optimize repeated calls
+inline int64_t ggml_batch_offset(const ggml_tensor * t, int64_t batch_idx, int64_t ne2) {
+    const int64_t i2 = batch_idx % ne2;
+    const int64_t i3 = batch_idx / ne2;
+    return i3 * t->nb[3] + i2 * t->nb[2];
+}
+
 size_t ggml_backend_amx_desired_wsize(const struct ggml_tensor * dst) {
     struct ggml_tensor * src0 = dst->src[0];
 
@@ -2348,12 +2319,13 @@ size_t ggml_backend_amx_desired_wsize(const struct ggml_tensor * dst) {
 
     const int M = dst->ne[1];
     const int K = src0->ne[0];
+    const int64_t n_batch = dst->ne[2] * dst->ne[3];
 
     size_t desired_wsize = 0;
 
     GGML_DISPATCH_QTYPES(TYPE, [&] {
         const size_t row_size_A = K / blck_size * sizeof(vec_dot_type);
-        desired_wsize = M * row_size_A;
+        desired_wsize = n_batch * M * row_size_A;
     });
 
     return desired_wsize;
@@ -2365,7 +2337,7 @@ size_t ggml_backend_amx_desired_wsize(const struct ggml_tensor * dst) {
 // src1: input  in shape of {M, K}, float32
 // dst:  output in shape of {M, N}, float32
 //
-// the function performs: dst = src1 @ src0.T
+// the function performs: dst = src1 @ src0.T for each batch
 //
 void ggml_backend_amx_mul_mat(const ggml_compute_params * params, struct ggml_tensor * dst) {
     struct ggml_tensor * src0 = dst->src[0];
@@ -2382,17 +2354,26 @@ void ggml_backend_amx_mul_mat(const ggml_compute_params * params, struct ggml_te
     const int K = src0->ne[0];
     const int ldc = dst->nb[1] / dst->nb[0];
 
+    const int64_t ne2 = dst->ne[2];
+    const int64_t n_batch = ne2 * dst->ne[3];
+
     if (is_floating_type) {
         constexpr int BLOCK_M = 4;
         constexpr int BLOCK_N = 6;
         const int MB = div_up(M, BLOCK_M);
         const int NB = div_up(N, BLOCK_N);
 
-        parallel_for_ggml(params, MB * NB, [&](int begin, int end) {
+        parallel_for_ggml(params, n_batch * MB * NB, [&](int begin, int end) {
             GGML_DISPATCH_FLOATING_TYPES(TYPE, [&] {
                 for (int i = begin; i < end; ++i) {
-                    int mb = i / NB;
-                    int nb = i % NB;
+                    int batch_idx = i / (MB * NB);
+                    int remaining = i % (MB * NB);
+                    int mb = remaining / NB;
+                    int nb = remaining % NB;
+
+                    int64_t src0_offset = ggml_batch_offset(src0, batch_idx, ne2);
+                    int64_t src1_offset = ggml_batch_offset(src1, batch_idx, ne2);
+                    int64_t dst_offset  = ggml_batch_offset(dst,  batch_idx, ne2);
 
                     int mb_start = mb * BLOCK_M;
                     int mb_size = std::min(BLOCK_M, M - mb_start);
@@ -2424,10 +2405,10 @@ void ggml_backend_amx_mul_mat(const ggml_compute_params * params, struct ggml_te
     void * wdata = params->wdata;
 
     //TODO: performance improvement: merge quant A
-    if (params->ith == 0) {
+ // if (params->ith == 0) {
         GGML_DISPATCH_QTYPES(TYPE, [&] {
             const size_t row_size_A = K / blck_size * sizeof(vec_dot_type);
-            const size_t desired_wsize = M * row_size_A;
+            const size_t desired_wsize = n_batch * M * row_size_A;
             if (params->wsize < desired_wsize) {
                 GGML_ABORT("insufficient work space size");
             }
@@ -2436,12 +2417,19 @@ void ggml_backend_amx_mul_mat(const ggml_compute_params * params, struct ggml_te
             // Q4_K, Q5_K, Q6_K, IQ4_XS handles 8 TILE_K per blck_size
             GGML_ASSERT(TILE_K == blck_size || TILE_K * 8 == blck_size);
 
-            const float * A_data = static_cast<const float *>(src1->data);
-            for (int m = 0; m < M; ++m) {
-                from_float<vec_dot_type>(A_data + m * K, (char *)wdata + m * row_size_A, K);
-            }
+            parallel_for_ggml(params, n_batch, [&](int begin, int end) {
+                for (int batch_idx = begin; batch_idx < end; ++batch_idx) {
+                    int64_t src1_offset = ggml_batch_offset(src1, batch_idx, ne2);
+                    const float * A_data = (const float *)((const char *)src1->data + src1_offset);
+                    char * wdata_batch = (char *)wdata + batch_idx * M * row_size_A;
+
+                    for (int m = 0; m < M; ++m) {
+                        from_float<vec_dot_type>(A_data + m * K, wdata_batch + m * row_size_A, K);
+                    }
+                }
+            });
         });
-    }
+ // }
 
     ggml_barrier(params->threadpool);
 
@@ -2451,13 +2439,19 @@ void ggml_backend_amx_mul_mat(const ggml_compute_params * params, struct ggml_te
         constexpr int BLOCK_N = TILE_N * kTilesN;
         const int NB = div_up(N, BLOCK_N);
 
-        parallel_for_ggml(params, NB, [&](int begin, int end) {
+        parallel_for_ggml(params, n_batch * NB, [&](int begin, int end) {
             GGML_DISPATCH_QTYPES(TYPE, [&] {
                 const int KB = K / blck_size;
                 const int TILE_SIZE = get_tile_size<type>();
                 const int row_size_A = KB * sizeof(vec_dot_type);
                 for (int i = begin; i < end; ++i) {
-                    int nb = i;
+                    int batch_idx = i / NB;
+                    int nb = i % NB;
+
+                    int64_t src0_offset = ggml_batch_offset(src0, batch_idx, ne2);
+                    int64_t dst_offset  = ggml_batch_offset(dst,  batch_idx, ne2);
+                    const char * wdata_batch = (const char *)wdata + batch_idx * row_size_A;
+
                     int nb_start = nb * BLOCK_N;
                     int nb_size = std::min(BLOCK_N, N - nb_start); // 32, 64, 96
 
@@ -2481,7 +2475,7 @@ void ggml_backend_amx_mul_mat(const ggml_compute_params * params, struct ggml_te
     const int MB = div_up(M, BLOCK_M);
     const int NB = div_up(N, BLOCK_N);
 
-    parallel_for_ggml(params, MB * NB, [&](int begin, int end) {
+    parallel_for_ggml(params, n_batch * MB * NB, [&](int begin, int end) {
         // init tile config for each thread
         ggml_tile_config_init();
 
@@ -2491,8 +2485,14 @@ void ggml_backend_amx_mul_mat(const ggml_compute_params * params, struct ggml_te
             const int row_size_A = KB * sizeof(vec_dot_type);
 
             for (int i = begin; i < end; ++i) {
-                int mb = i / NB;
-                int nb = i % NB;
+                int batch_idx = i / (MB * NB);
+                int remaining = i % (MB * NB);
+                int mb = remaining / NB;
+                int nb = remaining % NB;
+
+                int64_t src0_offset = ggml_batch_offset(src0, batch_idx, ne2);
+                int64_t dst_offset  = ggml_batch_offset(dst,  batch_idx, ne2);
+                const char * wdata_batch = (const char *)wdata + batch_idx * M * row_size_A;
 
                 int mb_start = mb * BLOCK_M;
                 int mb_size = std::min(BLOCK_M, M - mb_start);
@@ -2501,9 +2501,9 @@ void ggml_backend_amx_mul_mat(const ggml_compute_params * params, struct ggml_te
 
                 tinygemm_kernel_amx<vec_dot_type, type, float, blck_size>(
                     mb_size, nb_size, KB,
-                    (const char *)wdata + mb_start * row_size_A,
-                    (const char *)src0->data + PACKED_INDEX(nb * 2, 0, KB, TILE_SIZE),
-                    (float *) dst->data + mb_start * N + nb_start, ldc);
+                    wdata_batch + mb_start * row_size_A,
+                    (const char *)src0->data + src0_offset + PACKED_INDEX(nb * 2, 0, KB, TILE_SIZE),
+                    (float *) dst->data + dst_offset + mb_start * N + nb_start, ldc);
             }
         });
     });

ggml/src/ggml-cpu/arch-fallback.h

@@ -48,6 +48,8 @@
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
+#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
+#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
 #define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
@@ -62,6 +64,8 @@
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
+#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
+#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
 #define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
 #elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64)
@@ -69,8 +73,10 @@
 #define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
+#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
+#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
 // repack.cpp
@@ -84,6 +90,7 @@
 #define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
 #define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
@@ -94,6 +101,7 @@
 #define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
 #define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
 #elif defined(__POWERPC__) || defined(__powerpc__)
@@ -120,6 +128,8 @@
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
+#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
+#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
 #define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
@@ -134,6 +144,8 @@
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
+#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
+#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
 #define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
 #elif defined(__loongarch64)
@@ -160,6 +172,8 @@
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
+#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
+#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
 #define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
@@ -174,6 +188,8 @@
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
+#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
+#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
 #define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
 #elif defined(__riscv)
@@ -201,6 +217,8 @@
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
+#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
+#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
 #define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
@@ -214,6 +232,8 @@
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
+#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
+#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
 #define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
 #elif defined(__s390x__)
@@ -246,6 +266,8 @@
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
+#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
+#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
 #define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
@@ -260,6 +282,8 @@
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
+#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
+#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
 #define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
 #elif defined(__wasm__)
@@ -294,6 +318,8 @@
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
+#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
+#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
 #define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
@@ -308,6 +334,8 @@
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
+#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
+#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
 #define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
 #endif

ggml/src/ggml-cpu/arch/arm/repack.cpp

@@ -498,6 +498,81 @@ void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
     ggml_gemv_iq4_nl_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
+void ggml_gemv_mxfp4_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    const int8x16_t kvalues = vld1q_s8(kvalues_mxfp4);
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    float * res_ptr = s;
+
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_mxfp4x4 * b_ptr = (const block_mxfp4x4 *) vx + (x * nb);
+
+        float32x4_t sumf = vdupq_n_f32(0);
+        for (int l = 0; l < nb; l++) {
+            uint8x16_t b_0 = vld1q_u8(b_ptr[l].qs + 0);
+            uint8x16_t b_1 = vld1q_u8(b_ptr[l].qs + 16);
+            uint8x16_t b_2 = vld1q_u8(b_ptr[l].qs + 32);
+            uint8x16_t b_3 = vld1q_u8(b_ptr[l].qs + 48);
+
+            int8x16_t b_0_hi = vqtbl1q_s8(kvalues, b_0 >> 4);
+            int8x16_t b_0_lo = vqtbl1q_s8(kvalues, b_0 & 0x0F);
+            int8x16_t b_1_hi = vqtbl1q_s8(kvalues, b_1 >> 4);
+            int8x16_t b_1_lo = vqtbl1q_s8(kvalues, b_1 & 0x0F);
+            int8x16_t b_2_hi = vqtbl1q_s8(kvalues, b_2 >> 4);
+            int8x16_t b_2_lo = vqtbl1q_s8(kvalues, b_2 & 0x0F);
+            int8x16_t b_3_hi = vqtbl1q_s8(kvalues, b_3 >> 4);
+            int8x16_t b_3_lo = vqtbl1q_s8(kvalues, b_3 & 0x0F);
+
+            int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 0);
+            int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16);
+
+            int32x4_t sumi = vdupq_n_s32(0);
+            sumi = vdotq_laneq_s32(sumi, b_0_lo, a_0, 0);
+            sumi = vdotq_laneq_s32(sumi, b_0_hi, a_1, 0);
+            sumi = vdotq_laneq_s32(sumi, b_1_lo, a_0, 1);
+            sumi = vdotq_laneq_s32(sumi, b_1_hi, a_1, 1);
+            sumi = vdotq_laneq_s32(sumi, b_2_lo, a_0, 2);
+            sumi = vdotq_laneq_s32(sumi, b_2_hi, a_1, 2);
+            sumi = vdotq_laneq_s32(sumi, b_3_lo, a_0, 3);
+            sumi = vdotq_laneq_s32(sumi, b_3_hi, a_1, 3);
+
+            float32x4_t a_d = vcvt_f32_f16(vld1_dup_f16((const float16_t *)&a_ptr[l].d));
+            float32x4_t b_d = {
+                GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[0]),
+                GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[1]),
+                GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[2]),
+                GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[3]),
+            };
+            float32x4_t d = a_d * b_d;
+
+            sumf = vmlaq_f32(sumf, d, vcvtq_f32_s32(sumi));
+        }
+
+        vst1q_f32(res_ptr + x * 4, sumf);
+    }
+    return;
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
+    ggml_gemv_mxfp4_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
+}
+
 void ggml_gemv_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     constexpr int qk = QK_K;
     const int     nb = n / qk;
@@ -3164,6 +3239,87 @@ void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
     ggml_gemm_iq4_nl_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
+void ggml_gemm_mxfp4_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    const int8x16_t kvalues = vld1q_s8(kvalues_mxfp4);
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_mxfp4x4 * b_ptr = (const block_mxfp4x4 *) vx + (x * nb);
+
+            float32x4_t sumf[4];
+            for (int m = 0; m < 4; m++) {
+                sumf[m] = vdupq_n_f32(0);
+            }
+
+            for (int l = 0; l < nb; l++) {
+                float32x4_t a_d = vcvt_f32_f16(vld1_f16((const float16_t *)a_ptr[l].d));
+                float32x4_t b_d = {
+                    GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[0]),
+                    GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[1]),
+                    GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[2]),
+                    GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[3]),
+                };
+
+                int32x4_t sumi_0 = vdupq_n_s32(0);
+                int32x4_t sumi_1 = vdupq_n_s32(0);
+                int32x4_t sumi_2 = vdupq_n_s32(0);
+                int32x4_t sumi_3 = vdupq_n_s32(0);
+
+                for (int k = 0; k < 4; k++) {
+                    int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 16 * k + 0);
+                    int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16 * k + 64);
+
+                    uint8x16_t b = vld1q_u8(b_ptr[l].qs + 16 * k);
+                    int8x16_t b_hi = vqtbl1q_s8(kvalues, b >> 4);
+                    int8x16_t b_lo = vqtbl1q_s8(kvalues, b & 0xF);
+
+                    sumi_0 = vdotq_laneq_s32(sumi_0, b_lo, a_0, 0);
+                    sumi_1 = vdotq_laneq_s32(sumi_1, b_lo, a_0, 1);
+                    sumi_2 = vdotq_laneq_s32(sumi_2, b_lo, a_0, 2);
+                    sumi_3 = vdotq_laneq_s32(sumi_3, b_lo, a_0, 3);
+                    sumi_0 = vdotq_laneq_s32(sumi_0, b_hi, a_1, 0);
+                    sumi_1 = vdotq_laneq_s32(sumi_1, b_hi, a_1, 1);
+                    sumi_2 = vdotq_laneq_s32(sumi_2, b_hi, a_1, 2);
+                    sumi_3 = vdotq_laneq_s32(sumi_3, b_hi, a_1, 3);
+                }
+
+                sumf[0] = vmlaq_f32(sumf[0], vmulq_laneq_f32(b_d, a_d, 0), vcvtq_f32_s32(sumi_0));
+                sumf[1] = vmlaq_f32(sumf[1], vmulq_laneq_f32(b_d, a_d, 1), vcvtq_f32_s32(sumi_1));
+                sumf[2] = vmlaq_f32(sumf[2], vmulq_laneq_f32(b_d, a_d, 2), vcvtq_f32_s32(sumi_2));
+                sumf[3] = vmlaq_f32(sumf[3], vmulq_laneq_f32(b_d, a_d, 3), vcvtq_f32_s32(sumi_3));
+            }
+
+            for (int m = 0; m < 4; m++) {
+                vst1q_f32(s + (y * 4 + m) * bs + x * 4, sumf[m]);
+            }
+        }
+    }
+    return;
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
+    ggml_gemm_mxfp4_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
+}
+
 void ggml_gemm_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     constexpr int qk = QK_K;
     const int     nb = n / qk;

ggml/src/ggml-cpu/arch/s390/quants.c

@@ -181,11 +181,11 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const int8x16_t v_yh = vec_xl(QK8_0/2, y[ib].qs);
 
         const int16x8_t v_xylso = vec_mulo(v_xls, v_yl);
-        const int16x8_t v_xylse = vec_mule(v_xls, v_yl);
+        const int16x8_t v_xyl = vec_meadd(v_xls, v_yl, v_xylso);
         const int16x8_t v_xyhso = vec_mulo(v_xhs, v_yh);
-        const int16x8_t v_xyhse = vec_mule(v_xhs, v_yh);
+        const int16x8_t v_xyh = vec_meadd(v_xhs, v_yh, v_xyhso);
 
-        int16x8_t v_xy_ = v_xylso + v_xylse + v_xyhso + v_xyhse; v_xy_ += vec_reve(v_xy_);
+        int16x8_t v_xy_ = v_xyl + v_xyh; v_xy_ += vec_reve(v_xy_);
 
         const float32x4_t v_xy = vec_float(vec_unpackh(v_xy_));
         const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
@@ -890,8 +890,7 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const int16x8_t v_minsh = (int16x8_t)vec_unpackh((uint8x16_t)v_mins8);
 
         const int32x4_t v_minso = vec_mulo(v_ysums, v_minsh);
-        const int32x4_t v_minse = vec_mule(v_ysums, v_minsh);
-        const int32x4_t v_mins = v_minso + v_minse;
+        const int32x4_t v_mins = vec_meadd(v_ysums, v_minsh, v_minso);
         sumf -= dmin * (v_mins[0] + v_mins[1] + v_mins[2] + v_mins[3]);
 
         const uint8_t * scales = (const uint8_t *)utmp;
@@ -1004,8 +1003,7 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const int16x8_t v_minsh = (int16x8_t)vec_unpackh(v_mins8);
 
         const int32x4_t v_minsho = vec_mulo(v_ysums, v_minsh);
-        const int32x4_t v_minshe = vec_mule(v_ysums, v_minsh);
-        const int32x4_t v_mins = vec_add(v_minsho, v_minshe);
+        const int32x4_t v_mins = vec_meadd(v_ysums, v_minsh, v_minsho);
         const int32_t mins = vec_hsum_i32x4(v_mins);
 
         const uint8_t * scales = (const uint8_t *)utmp;
@@ -1110,10 +1108,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const int16x8_t v_scaleh = vec_unpackl(v_scale);
 
         const int32x4_t v_minslo = vec_mulo(v_ysumsl, v_scalel);
-        const int32x4_t v_minsle = vec_mule(v_ysumsl, v_scalel);
+        const int32x4_t v_minsl = vec_meadd(v_ysumsl, v_scalel, v_minslo);
         const int32x4_t v_minsho = vec_mulo(v_ysumsh, v_scaleh);
-        const int32x4_t v_minshe = vec_mule(v_ysumsh, v_scaleh);
-        const int32x4_t v_mins = v_minslo + v_minsle + v_minsho + v_minshe;
+        const int32x4_t v_minsh = vec_meadd(v_ysumsh, v_scaleh, v_minsho);
+        const int32x4_t v_mins = vec_add(v_minsl, v_minsh);
 
         const int32_t mins = vec_hsum_i32x4(v_mins);
 

ggml/src/ggml-cpu/arch/x86/repack.cpp

@@ -522,7 +522,8 @@ template<typename block_tx8>
 static void gemv_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc, __m256i signextendlut) {
     static_assert(
             std::is_same_v<block_tx8, block_q4_0x8> ||
-            std::is_same_v<block_tx8, block_iq4_nlx8>,
+            std::is_same_v<block_tx8, block_iq4_nlx8> ||
+            std::is_same_v<block_tx8, block_mxfp4x8>,
             "Unsupported block type");
 
     const int qk = QK8_0;
@@ -580,6 +581,18 @@ static void gemv_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
                         std::is_same_v<block_tx8, block_q4_0x8> ||
                         std::is_same_v<block_tx8, block_iq4_nlx8>) {
                     col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, changemask);
+                } else if constexpr (std::is_same_v<block_tx8, block_mxfp4x8>) {
+                    // Load 8 E8M0 exponents and convert to float via LUT
+                    // Rearranged to match changemask order: 0,4,1,5,2,6,3,7
+                    col_scale_f32 = _mm256_set_ps(
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[7]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[3]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[6]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[2]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[5]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[1]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[4]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[0]));
                 }
 
                 // Load and convert to FP32 scale from block_q8_0
@@ -628,7 +641,8 @@ template<typename block_tx8>
 static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc, __m256i signextendlut) {
     static_assert(
             std::is_same_v<block_tx8, block_q4_0x8> ||
-            std::is_same_v<block_tx8, block_iq4_nlx8>,
+            std::is_same_v<block_tx8, block_iq4_nlx8> ||
+            std::is_same_v<block_tx8, block_mxfp4x8>,
             "Unsupported block type");
 
     const int qk = QK8_0;
@@ -749,6 +763,25 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
                         std::is_same_v<block_tx8, block_q4_0x8> ||
                         std::is_same_v<block_tx8, block_iq4_nlx8>) {
                     col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
+                } else if constexpr (std::is_same_v<block_tx8, block_mxfp4x8>) {
+                    //TODO: simd-ify
+                    col_scale_f32 = _mm512_set_ps(
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[7]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[6]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[5]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[4]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[3]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[2]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[1]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[0]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[7]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[6]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[5]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[4]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[3]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[2]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[1]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[0]));
                 }
 
                 // Process LHS in pairs of rows
@@ -941,6 +974,25 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
                         std::is_same_v<block_tx8, block_q4_0x8> ||
                         std::is_same_v<block_tx8, block_iq4_nlx8>) {
                     col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
+                } else if constexpr (std::is_same_v<block_tx8, block_mxfp4x8>) {
+                    //TODO: simd-ify
+                    col_scale_f32 = _mm512_set_ps(
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[7]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[6]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[5]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[4]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[3]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[2]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[1]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[0]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[7]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[6]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[5]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[4]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[3]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[2]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[1]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[0]));
                 }
 
                 // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
@@ -1123,6 +1175,16 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
                         std::is_same_v<block_tx8, block_q4_0x8> ||
                         std::is_same_v<block_tx8, block_iq4_nlx8>) {
                     col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
+                } else if constexpr (std::is_same_v<block_tx8, block_mxfp4x8>) {
+                    col_scale_f32 = _mm256_set_ps(
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[7]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[6]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[5]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[4]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[3]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[2]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[1]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[0]));
                 }
 
                 // Process LHS in groups of four
@@ -1283,6 +1345,16 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
                         std::is_same_v<block_tx8, block_q4_0x8> ||
                         std::is_same_v<block_tx8, block_iq4_nlx8>) {
                     col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
+                } else if constexpr (std::is_same_v<block_tx8, block_mxfp4x8>) {
+                    col_scale_f32 = _mm256_set_ps(
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[7]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[6]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[5]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[4]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[3]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[2]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[1]),
+                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[0]));
                 }
 
                 // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
@@ -1625,6 +1697,19 @@ void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
     ggml_gemv_iq4_nl_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
+void ggml_gemv_mxfp4_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+#if defined(__AVX2__)
+    __m256i signextendlut = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i*)kvalues_mxfp4));
+    signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
+
+    gemv_q4_b32_8x8_q8_0_lut_avx<block_mxfp4x8>(n, s, bs, vx, vy, nr, nc, signextendlut);
+
+    return;
+#endif
+
+    ggml_gemv_mxfp4_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
+}
+
 void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK_K;
     const int nb = n / qk;
@@ -3423,6 +3508,21 @@ void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
     ggml_gemm_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+void ggml_gemm_mxfp4_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+#if defined(__AVX2__) || defined(__AVX512F__)
+    {
+        __m256i signextendlut = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i*)kvalues_mxfp4));
+        signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
+
+        gemm_q4_b32_8x8_q8_0_lut_avx<block_mxfp4x8>(n, s, bs, vx, vy, nr, nc, signextendlut);
+
+        return;
+    }
+#endif // defined(__AVX2__) || defined(__AVX512F__)
+
+    ggml_gemm_mxfp4_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
+}
+
 void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK_K;
     const int nb = n / qk;

ggml/src/ggml-cpu/repack.cpp

@@ -1098,6 +1098,82 @@ void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
     }
 }
 
+void ggml_gemv_mxfp4_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert(nr == 1);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float sumf[4];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_mxfp4x4 * b_ptr = (const block_mxfp4x4 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                        const int v1 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
+                    }
+                    sumf[j] += sumi * GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+}
+
+void ggml_gemv_mxfp4_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert(nr == 1);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float sumf[8];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_mxfp4x8 * b_ptr = (const block_mxfp4x8 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                        const int v1 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
+                    }
+                    sumf[j] += sumi * GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+}
+
 void ggml_gemv_q8_0_4x4_q8_0_generic(int                        n,
                                      float * GGML_RESTRICT      s,
                                      size_t                     bs,
@@ -1726,6 +1802,94 @@ void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
     }
 }
 
+void ggml_gemm_mxfp4_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    float sumf[4][4];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_mxfp4x4 * b_ptr = (const block_mxfp4x4 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                                const int v1 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
+                            }
+                            sumf[m][j] += sumi * GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+}
+
+void ggml_gemm_mxfp4_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    float sumf[4][8];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_mxfp4x8 * b_ptr = (const block_mxfp4x8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                                const int v1 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
+                            }
+                            sumf[m][j] += sumi * GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+}
+
 void ggml_gemm_q8_0_4x4_q8_0_generic(int                        n,
                                      float * GGML_RESTRICT      s,
                                      size_t                     bs,
@@ -2510,6 +2674,121 @@ static int repack_iq4_nl_to_iq4_nl_8_bl(struct ggml_tensor * t, int interleave_b
     GGML_UNUSED(data_size);
 }
 
+
+static block_mxfp4x4 make_block_mxfp4x4(block_mxfp4 * in, unsigned int blck_size_interleave) {
+    block_mxfp4x4 out;
+
+    for (int i = 0; i < 4; i++) {
+        out.e[i] = in[i].e;
+    }
+
+    const int end = QK_MXFP4 * 2 / blck_size_interleave;
+
+    if (blck_size_interleave == 4) {
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 4;
+            int src_offset = (i / 4) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint32_t));
+        }
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    return out;
+}
+
+static int repack_mxfp4_to_mxfp4_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_MXFP4);
+    GGML_ASSERT(interleave_block == 4);
+
+    const block_mxfp4   * src = (const block_mxfp4   *)data;
+          block_mxfp4x4 * dst = (      block_mxfp4x4 *)t->data;
+
+    block_mxfp4 dst_tmp[4];
+
+    int nrow = ggml_nrows(t);
+    int nrows_interleaved = 4;
+    int nblocks = t->ne[0] / QK_MXFP4;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_mxfp4));
+
+    if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i = 0; i < nrows_interleaved; i++) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_mxfp4x4(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
+static block_mxfp4x8 make_block_mxfp4x8(block_mxfp4 * in, unsigned int blck_size_interleave) {
+    block_mxfp4x8 out;
+
+    for (int i = 0; i < 8; i++) {
+        out.e[i] = in[i].e;
+    }
+
+    const int end = QK_MXFP4 * 4 / blck_size_interleave;
+
+    if (blck_size_interleave == 8) {
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 8;
+            int src_offset = (i / 8) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint64_t));
+        }
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    return out;
+}
+
+static int repack_mxfp4_to_mxfp4_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_MXFP4);
+    GGML_ASSERT(interleave_block == 8);
+
+    const block_mxfp4   * src = (const block_mxfp4   *)data;
+          block_mxfp4x8 * dst = (      block_mxfp4x8 *)t->data;
+
+    block_mxfp4 dst_tmp[8];
+
+    int nrow = ggml_nrows(t);
+    int nrows_interleaved = 8;
+    int nblocks = t->ne[0] / QK_MXFP4;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_mxfp4));
+
+    if (t->ne[1] % nrows_interleaved != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i = 0; i < nrows_interleaved; i++) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_mxfp4x8(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
 namespace ggml::cpu::repack {
 // repack
 template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS>
@@ -2569,6 +2848,14 @@ template <> int repack<block_iq4_nl, 8, 8>(struct ggml_tensor * t, const void *
     return repack_iq4_nl_to_iq4_nl_8_bl(t, 8, data, data_size);
 }
 
+template <> int repack<block_mxfp4, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_mxfp4_to_mxfp4_4_bl(t, 4, data, data_size);
+}
+
+template <> int repack<block_mxfp4, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_mxfp4_to_mxfp4_8_bl(t, 8, data, data_size);
+}
+
 template <> int repack<block_q8_0, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
     return repack_q8_0_to_q8_0_4_bl(t, 4, data, data_size);
 }
@@ -2636,6 +2923,14 @@ template <> void gemv<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size
     ggml_gemv_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemv<block_mxfp4, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_mxfp4_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemv<block_mxfp4, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_mxfp4_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
 template <> void gemv<block_q8_0, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
     ggml_gemv_q8_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
 }
@@ -2703,6 +2998,14 @@ template <> void gemm<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size
     ggml_gemm_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemm<block_mxfp4, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_mxfp4_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemm<block_mxfp4, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_mxfp4_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
 template <> void gemm<block_q8_0, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
     ggml_gemm_q8_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
 }
@@ -3111,6 +3414,10 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
     static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0> iq4_nl_4x4_q8_0;
     static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0> iq4_nl_8x8_q8_0;
 
+    // instance for MXFP4
+    static const ggml::cpu::repack::tensor_traits<block_mxfp4, 4, 4, GGML_TYPE_Q8_0> mxfp4_4x4_q8_0;
+    static const ggml::cpu::repack::tensor_traits<block_mxfp4, 8, 8, GGML_TYPE_Q8_0> mxfp4_8x8_q8_0;
+
     // instance for Q8_0
     static const ggml::cpu::repack::tensor_traits<block_q8_0, 4, 4, GGML_TYPE_Q8_0> q8_0_4x4_q8_0;
     static const ggml::cpu::repack::tensor_traits<block_q8_0, 8, 4, GGML_TYPE_Q8_0> q8_0_4x8_q8_0;
@@ -3187,6 +3494,17 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
                 return &iq4_nl_4x4_q8_0;
             }
         }
+    } else if (cur->type == GGML_TYPE_MXFP4) {
+        if (ggml_cpu_has_avx2()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &mxfp4_8x8_q8_0;
+            }
+        }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+            if (cur->ne[1] % 4 == 0) {
+                return &mxfp4_4x4_q8_0;
+            }
+        }
     } else if (cur->type == GGML_TYPE_Q8_0) {
         if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
             if (cur->ne[1] % 4 == 0) {

ggml/src/ggml-cpu/repack.h

@@ -97,6 +97,19 @@ struct block_iq4_nlx8 {
 
 static_assert(sizeof(block_iq4_nlx8) == 8 * sizeof(ggml_half) + QK4_NL * 4, "wrong iq4_nlx8 block size/padding");
 
+struct block_mxfp4x4 {
+    uint8_t e[4];
+    uint8_t qs[QK_MXFP4 * 2];
+};
+static_assert(sizeof(block_mxfp4x4) == 4 + QK_MXFP4 * 2, "wrong mxfp4x4 block size/padding");
+
+struct block_mxfp4x8 {
+    uint8_t e[8];
+    uint8_t qs[QK_MXFP4 * 4];
+};
+static_assert(sizeof(block_mxfp4x8) == 8 + QK_MXFP4 * 4, "wrong mxfp4x8 block size/padding");
+
+
 #if defined(__cplusplus)
 extern "C" {
 #endif
@@ -117,6 +130,8 @@ void ggml_gemv_q6_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 void ggml_gemv_q6_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_mxfp4_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_mxfp4_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -129,6 +144,8 @@ void ggml_gemm_q6_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 void ggml_gemm_q6_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_mxfp4_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_mxfp4_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q8_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -151,6 +168,8 @@ void ggml_gemv_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
 void ggml_gemv_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_mxfp4_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_mxfp4_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -163,6 +182,8 @@ void ggml_gemm_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
 void ggml_gemm_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_mxfp4_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_mxfp4_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q8_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);

ggml/src/ggml-cuda/convert.cu

@@ -16,27 +16,27 @@ static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __
         return;
     }
 
-    const int64_t i01 = blockIdx.y;
+    for (int64_t i01 = blockIdx.y; i01 < ne01; i01 += gridDim.y) {
+        for (int64_t i0203 = blockIdx.z; i0203 < ne0203; i0203 += gridDim.z) {
+            const uint2 dm = fast_div_modulo((uint32_t)i0203, ne02);
+            const int64_t i02 = dm.y;
+            const int64_t i03 = dm.x;
 
-    for (int64_t i0203 = blockIdx.z; i0203 < ne0203; i0203 += gridDim.z) {
-        const uint2 dm = fast_div_modulo((uint32_t)i0203, ne02);
-        const int64_t i02 = dm.y;
-        const int64_t i03 = dm.x;
+            const int64_t ibx0 = i03*s03 + i02*s02 + i01*s01;
 
-        const int64_t ibx0 = i03*s03 + i02*s02 + i01*s01;
+            const int64_t ib = ibx0 + i00/qk; // block index
+            const int64_t iqs = (i00%qk)/qr; // quant index
+            const int64_t iybs = i00 - i00%qk; // y block start index
+            const int64_t y_offset = qr == 1 ? 1 : qk/2;
 
-        const int64_t ib = ibx0 + i00/qk; // block index
-        const int64_t iqs = (i00%qk)/qr; // quant index
-        const int64_t iybs = i00 - i00%qk; // y block start index
-        const int64_t y_offset = qr == 1 ? 1 : qk/2;
+            // dequantize
+            float2 v;
+            dequantize_kernel(vx, ib, iqs, v);
 
-        // dequantize
-        float2 v;
-        dequantize_kernel(vx, ib, iqs, v);
-
-        const int64_t iy0 = (i0203*ne01 + i01)*ne00 + iybs + iqs;
-        y[iy0 + 0]        = ggml_cuda_cast<dst_t>(v.x);
-        y[iy0 + y_offset] = ggml_cuda_cast<dst_t>(v.y);
+            const int64_t iy0 = (i0203*ne01 + i01)*ne00 + iybs + iqs;
+            y[iy0 + 0]        = ggml_cuda_cast<dst_t>(v.x);
+            y[iy0 + y_offset] = ggml_cuda_cast<dst_t>(v.y);
+        }
     }
 }
 
@@ -492,7 +492,7 @@ static void dequantize_block_cuda(const void * vx, dst_t * y,
         const int64_t s01, const int64_t s02, const int64_t s03, cudaStream_t stream) {
     const int64_t ne0203 = ne02*ne03;
     const uint3 ne02_fdv = init_fastdiv_values(ne02);
-    const dim3 num_blocks((ne00 + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE), ne01, (int)std::min(ne0203, (int64_t)65535));
+    const dim3 num_blocks((ne00 + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE), (int)std::min(ne01, (int64_t)65535), (int)std::min(ne0203, (int64_t)65535));
     dequantize_block<qk, qr, dequantize_kernel><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>
         (vx, y, ne00, ne01, ne0203, ne02_fdv, s01, s02, s03);
 }
@@ -628,18 +628,18 @@ static __global__ void convert_unary(
         return;
     }
 
-    const int64_t i01 = blockIdx.y;
-
     const src_t * x = (const src_t *) vx;
 
-    for (int64_t i0203 = blockIdx.z; i0203 < ne0203; i0203 += gridDim.z) {
-        const uint2 dm = fast_div_modulo((uint32_t)i0203, ne02);
-        const int64_t i02 = dm.y;
-        const int64_t i03 = dm.x;
+    for (int64_t i01 = blockIdx.y; i01 < ne01; i01 += gridDim.y) {
+        for (int64_t i0203 = blockIdx.z; i0203 < ne0203; i0203 += gridDim.z) {
+            const uint2 dm = fast_div_modulo((uint32_t)i0203, ne02);
+            const int64_t i02 = dm.y;
+            const int64_t i03 = dm.x;
 
-        const int64_t ix = i03*s03 + i02*s02 + i01*s01 + i00;
-        const int64_t iy = (i0203*ne01 + i01)*ne00 + i00;
-        y[iy] = ggml_cuda_cast<dst_t>(x[ix]);
+            const int64_t ix = i03*s03 + i02*s02 + i01*s01 + i00;
+            const int64_t iy = (i0203*ne01 + i01)*ne00 + i00;
+            y[iy] = ggml_cuda_cast<dst_t>(x[ix]);
+        }
     }
 }
 
@@ -649,7 +649,7 @@ static void convert_unary_cuda(const void * vx, dst_t * y,
         const int64_t s01, const int64_t s02, const int64_t s03, cudaStream_t stream) {
     const int64_t ne0203 = ne02*ne03;
     const uint3 ne02_fdv = init_fastdiv_values(ne02);
-    const dim3 num_blocks((ne00 + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE, ne01, (int)std::min(ne0203, (int64_t)65535));
+    const dim3 num_blocks((ne00 + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE, (int)std::min(ne01, (int64_t)65535), (int)std::min(ne0203, (int64_t)65535));
     convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>
         (vx, y, ne00, ne01, ne0203, ne02_fdv, s01, s02, s03);
 }

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

@@ -111,6 +111,44 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
     return ggml_cuda_fattn_mma_get_config_ampere(DKQ, DV, ncols);
 }
 
+static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_config_cdna(const int DKQ, const int DV, const int ncols) {
+    // Conservative configs for CDNA (MI100+): 64KB LDS, wavefront64, nstages=1 (no cp.async).
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 64,  64,  8, 128, 2, 128,  32,  32,  32, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 64,  64, 16, 128, 2,  64,  32,  32,  32, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 64,  64, 32, 128, 2,  64,  32,  32,  32, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 64,  64, 64, 256, 2,  64,  32,  32,  32, 1, true);
+
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 80,  80,  8, 128, 2, 128,  40,  40,  40, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 80,  80, 16, 128, 2,  64,  40,  40,  40, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 80,  80, 32, 128, 2,  64,  40,  40,  40, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 80,  80, 64, 256, 2,  64,  40,  40,  40, 1, true);
+
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 96,  96,  8, 128, 2, 128,  48,  48,  48, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 96,  96, 16, 128, 2,  64,  48,  48,  48, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 96,  96, 32, 128, 2,  64,  48,  48,  48, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 96,  96, 64, 256, 2,  64,  48,  48,  48, 1, true);
+
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(112, 112,  8, 128, 2, 128,  56,  56,  56, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(112, 112, 16, 128, 2,  64,  56,  56,  56, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(112, 112, 32, 128, 2,  64,  56,  56,  56, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(112, 112, 64, 256, 2,  64,  56,  56,  56, 1, true);
+
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(128, 128,  8, 128, 2, 128,  64,  64,  64, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(128, 128, 16, 128, 2,  64,  64,  64,  64, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(128, 128, 32, 128, 2,  64,  64,  64,  64, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(128, 128, 64, 256, 2,  64,  64,  64,  64, 1, true);
+
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256,  8,  64, 4,  64, 128, 128, 128, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 16,  64, 4,  32, 128, 128, 128, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  32, 128, 128, 128, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 256, 2,  32, 128, 128, 128, 1, true);
+
+    // Fallback for unsupported DKQ values (e.g. 576). Must return non-zero values to satisfy
+    // compile-time static_asserts even though the kernel guard prevents runtime execution.
+    // nthreads=256 gives nwarps=4 (warp_size=64) or 8 (warp_size=32), nbatch_fa=128 satisfies np*16 divisibility.
+    return fattn_mma_config(256, 1, 128, 4, 4, 4, 1, false);
+}
+
 static __host__ fattn_mma_config ggml_cuda_fattn_mma_get_config(const int DKQ, const int DV, const int ncols, const int cc) {
     if (ampere_mma_available(cc)) {
         return ggml_cuda_fattn_mma_get_config_ampere(DKQ, DV, ncols);
@@ -118,6 +156,9 @@ static __host__ fattn_mma_config ggml_cuda_fattn_mma_get_config(const int DKQ, c
     if (turing_mma_available(cc)) {
         return ggml_cuda_fattn_mma_get_config_turing(DKQ, DV, ncols);
     }
+    if (amd_mfma_available(cc)) {
+        return ggml_cuda_fattn_mma_get_config_cdna(DKQ, DV, ncols);
+    }
     if (amd_wmma_available(cc)) {
         return ggml_cuda_fattn_mma_get_config_rdna(DKQ, DV, ncols);
     }
@@ -130,6 +171,8 @@ static constexpr __device__ fattn_mma_config ggml_cuda_fattn_mma_get_config(cons
     return ggml_cuda_fattn_mma_get_config_ampere(DKQ, DV, ncols);
 #elif defined(TURING_MMA_AVAILABLE)
     return ggml_cuda_fattn_mma_get_config_turing(DKQ, DV, ncols);
+#elif defined(AMD_MFMA_AVAILABLE)
+    return ggml_cuda_fattn_mma_get_config_cdna(DKQ, DV, ncols);
 #elif defined(VOLTA_MMA_AVAILABLE)
     return ggml_cuda_fattn_mma_get_config_volta(DKQ, DV, ncols);
 #elif defined(AMD_WMMA_AVAILABLE)
@@ -205,15 +248,15 @@ static constexpr __device__ bool ggml_cuda_fattn_mma_get_Q_in_reg(const int DKQ,
 }
 
 static constexpr __device__ int get_cols_per_thread() {
-#if defined(AMD_WMMA_AVAILABLE)
-    return 1; // RDNA has a single column.
+#if defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
+    return 1; // AMD has a single column per thread.
 #else
     return 2; // This is specifically KQ columns, Volta only has a single VKQ column.
-#endif // defined(AMD_WMMA_AVAILABLE)
+#endif // defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
 }
 
 static __host__ int get_cols_per_warp(const int cc) {
-    if (turing_mma_available(cc) || amd_wmma_available(cc)) {
+    if (turing_mma_available(cc) || amd_wmma_available(cc) || amd_mfma_available(cc)) {
         return 16;
     } else {
         // Volta
@@ -241,6 +284,7 @@ static constexpr __device__ int ggml_cuda_fattn_mma_get_nstages(const int DKQ, c
 template<int stride_tile, int nwarps, int nbatch_fa, bool use_cp_async, bool oob_check>
 static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
         const half2 * const __restrict__ KV, half2 * const __restrict__ tile_KV, const int D2, const int stride_KV, const int i_sup) {
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     // K/V data is loaded with decreasing granularity for D for better memory bandwidth.
     // The minimum granularity with cp.async is 16 bytes, with synchronous data loading it's 4 bytes.
     if constexpr (use_cp_async) {
@@ -252,10 +296,10 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
         const unsigned int tile_KV_32 = ggml_cuda_cvta_generic_to_shared(tile_KV);
 
         auto load = [&] __device__ (auto n) {
-            const int stride_k = WARP_SIZE >> n;
-            const int k0_start = stride_k == WARP_SIZE ? 0 : chunks_per_row - chunks_per_row % (2*stride_k);
+            const int stride_k = warp_size >> n;
+            const int k0_start = stride_k == warp_size ? 0 : chunks_per_row - chunks_per_row % (2*stride_k);
             const int k0_stop  =                             chunks_per_row - chunks_per_row % (1*stride_k);
-            const int stride_i = WARP_SIZE / stride_k;
+            const int stride_i = warp_size / stride_k;
 
             if (k0_start == k0_stop) {
                 return;
@@ -263,7 +307,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
 
 #pragma unroll
             for (int i0 = 0; i0 < nbatch_fa; i0 += nwarps*stride_i) {
-                const int i = i0 + threadIdx.y*stride_i + (stride_k == WARP_SIZE ? 0 : threadIdx.x / stride_k);
+                const int i = i0 + threadIdx.y*stride_i + (stride_k == warp_size ? 0 : threadIdx.x / stride_k);
 
                 if (i0 + nwarps*stride_i > nbatch_fa && i >= nbatch_fa) {
                     break;
@@ -271,7 +315,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
 
 #pragma unroll
                 for (int k0 = k0_start; k0 < k0_stop; k0 += stride_k) {
-                    const int k = k0 + (stride_k == WARP_SIZE ? threadIdx.x : threadIdx.x % stride_k);
+                    const int k = k0 + (stride_k == warp_size ? threadIdx.x : threadIdx.x % stride_k);
 
                     cp_async_cg_16<preload>(tile_KV_32 + i*(stride_tile*sizeof(half2)) + k*16, KV + i*stride_KV + k*h2_per_chunk);
                 }
@@ -287,10 +331,10 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
     } else {
         // TODO use ggml_cuda_memcpy_1
         auto load = [&] __device__ (const int n) {
-            const int stride_k = WARP_SIZE >> n;
-            const int k0_start = stride_k == WARP_SIZE ? 0 : D2 - D2 % (2*stride_k);
+            const int stride_k = warp_size >> n;
+            const int k0_start = stride_k == warp_size ? 0 : D2 - D2 % (2*stride_k);
             const int k0_stop  =                             D2 - D2 % (1*stride_k);
-            const int stride_i = WARP_SIZE / stride_k;
+            const int stride_i = warp_size / stride_k;
 
             if (k0_start == k0_stop) {
                 return;
@@ -298,7 +342,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
 
 #pragma unroll
             for (int i0 = 0; i0 < nbatch_fa; i0 += nwarps*stride_i) {
-                const int i = i0 + threadIdx.y*stride_i + (stride_k == WARP_SIZE ? 0 : threadIdx.x / stride_k);
+                const int i = i0 + threadIdx.y*stride_i + (stride_k == warp_size ? 0 : threadIdx.x / stride_k);
 
                 if (i0 + nwarps*stride_i > nbatch_fa && i >= nbatch_fa) {
                     break;
@@ -306,7 +350,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
 
 #pragma unroll
                 for (int k0 = k0_start; k0 < k0_stop; k0 += stride_k) {
-                    const int k = k0 + (stride_k == WARP_SIZE ? threadIdx.x : threadIdx.x % stride_k);
+                    const int k = k0 + (stride_k == warp_size ? threadIdx.x : threadIdx.x % stride_k);
 
                     tile_KV[i*stride_tile + k] = !oob_check || i < i_sup ? KV[i*stride_KV + k] : make_half2(0.0f, 0.0f);
                 }
@@ -324,18 +368,19 @@ template<int ncols1, int nwarps, int nbatch_fa, bool use_cp_async, bool oob_chec
 static __device__ __forceinline__ void flash_attn_ext_f16_load_mask(
         const half * const __restrict__ mask_h, half * const __restrict__ tile_mask,
         const int stride_mask, const int i_sup, const int j0, const uint3 ne01) {
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     if constexpr (use_cp_async) {
-        static_assert(nbatch_fa <= 8*WARP_SIZE && nbatch_fa % 8 == 0, "bad nbatch_fa");
+        static_assert(nbatch_fa <= 8*warp_size && nbatch_fa % 8 == 0, "bad nbatch_fa");
         static_assert(!oob_check, "OOB check incompatible with cp_async");
         constexpr int preload = nbatch_fa >= 32 ? nbatch_fa * sizeof(half) : 64;
-        constexpr int cols_per_warp = 8*WARP_SIZE/nbatch_fa;
+        constexpr int cols_per_warp = 8*warp_size/nbatch_fa;
         constexpr int stride_j = nwarps * cols_per_warp;
 
         const unsigned int tile_mask_32 = ggml_cuda_cvta_generic_to_shared(tile_mask);
 
 #pragma unroll
         for (int j1 = 0; j1 < ncols1; j1 += stride_j) {
-            const int j_sram = j1 + threadIdx.y*cols_per_warp + threadIdx.x / (WARP_SIZE/cols_per_warp);
+            const int j_sram = j1 + threadIdx.y*cols_per_warp + threadIdx.x / (warp_size/cols_per_warp);
             const int j_vram = fastmodulo(j0 + j_sram, ne01);
 
             if (j1 + stride_j > ncols1 && j_sram >= ncols1) {
@@ -357,25 +402,25 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_mask(
             }
 
 #pragma unroll
-            for (int i0 = 0; i0 < nbatch_fa; i0 += WARP_SIZE) {
+            for (int i0 = 0; i0 < nbatch_fa; i0 += warp_size) {
                 const int i = i0 + threadIdx.x;
 
                 tile_mask[j_sram*(nbatch_fa + 8) + i] = i < i_sup ? mask_h[j_vram*stride_mask + i] : half(0.0f);
             }
         }
-    } else if constexpr (nbatch_fa < 2*WARP_SIZE) {
-        constexpr int cols_per_warp = 2*WARP_SIZE/nbatch_fa;
+    } else if constexpr (nbatch_fa < 2*warp_size) {
+        constexpr int cols_per_warp = 2*warp_size/nbatch_fa;
         constexpr int stride_j = nwarps * cols_per_warp;
 #pragma unroll
         for (int j1 = 0; j1 < ncols1; j1 += stride_j) {
-            const int j_sram = j1 + threadIdx.y*cols_per_warp + threadIdx.x / (WARP_SIZE/cols_per_warp);
+            const int j_sram = j1 + threadIdx.y*cols_per_warp + threadIdx.x / (warp_size/cols_per_warp);
             const int j_vram = fastmodulo(j0 + j_sram, ne01);
 
             if (j1 + stride_j > ncols1 && j_sram >= ncols1) {
                 break;
             }
 
-            const int i = threadIdx.x % (WARP_SIZE/cols_per_warp);
+            const int i = threadIdx.x % (warp_size/cols_per_warp);
 
             ggml_cuda_memcpy_1<sizeof(half2)>(tile_mask + j_sram*(nbatch_fa + 8) + 2*i, mask_h + j_vram*stride_mask + 2*i);
         }
@@ -390,7 +435,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_mask(
             }
 
 #pragma unroll
-            for (int i0 = 0; i0 < nbatch_fa; i0 += 2*WARP_SIZE) {
+            for (int i0 = 0; i0 < nbatch_fa; i0 += 2*warp_size) {
                 const int i = i0 + 2*threadIdx.x;
 
                 ggml_cuda_memcpy_1<sizeof(half2)>(tile_mask + j_sram*(nbatch_fa + 8) + i, mask_h + j_vram*stride_mask + i);
@@ -428,7 +473,8 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         const int jt,
         const int kb0,
         const int k_VKQ_sup) {
-#if defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4))
+#if defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4)) || defined(AMD_MFMA_AVAILABLE)
+    constexpr int  warp_size       = ggml_cuda_get_physical_warp_size();
     constexpr int  ncols           = ncols1 * ncols2;
     constexpr int  cols_per_warp   = T_B_KQ::I;
     constexpr int  cols_per_thread = get_cols_per_thread();
@@ -447,7 +493,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     const int k_VKQ_0 = kb0 * nbatch_fa;
 #if defined(TURING_MMA_AVAILABLE)
     T_C_KQ KQ_C[nbatch_fa/(np*(cols_per_warp == 8 ? T_C_KQ::I : T_C_KQ::J))];
-#elif defined(AMD_WMMA_AVAILABLE)
+#elif defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
     T_C_KQ KQ_C[nbatch_fa/(np*T_C_KQ::J)];
 #else // Volta
     T_C_KQ KQ_C[nbatch_fa/(np*T_C_KQ::J)];
@@ -500,13 +546,13 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                         mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[k_KQ_0/T_A_KQ::J]);
                     } else {
                         // Wide version of KQ_C is column-major
-#if defined(AMD_WMMA_AVAILABLE)
-                        // RDNA matrix C is column-major.
+#if defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
+                        // AMD matrix C is column-major.
                         mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[k_KQ_0/T_A_KQ::J]);
 #else
                         // swap A and B for CUDA.
                         mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], Q_B[k_KQ_0/T_A_KQ::J], K_A);
-#endif // defined(AMD_WMMA_AVAILABLE)
+#endif // defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
                     }
                 }
             }
@@ -526,13 +572,13 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                         mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
                     } else {
                         // Wide version of KQ_C is column-major
-#if defined(AMD_WMMA_AVAILABLE)
-                        // RDNA matrix C is column-major.
+#if defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
+                        // AMD matrix C is column-major.
                         mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
 #else
                         // swap A and B for CUDA.
                         mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], Q_B[0], K_A);
-#endif // defined(AMD_WMMA_AVAILABLE)
+#endif // defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
                     }
                 }
             }
@@ -585,12 +631,12 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
 #pragma unroll
             for (int l = 0; l < T_C_KQ::ne; ++l) {
                 if (!oob_check || k0 + (threadIdx.y % np)*T_C_KQ::I + T_C_KQ::get_i(l) < k_VKQ_sup) {
-#if defined(AMD_WMMA_AVAILABLE)
+#if defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
                     constexpr int KQ_idx = 0;
 #else
                     // Turing + Volta:
                     const int KQ_idx = l % 2;
-#endif // defined(AMD_WMMA_AVAILABLE)
+#endif // defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
                     KQ_max_new[KQ_idx] = fmaxf(KQ_max_new[KQ_idx], KQ_C[k0/(np*T_C_KQ::I)].x[l] + FATTN_KQ_MAX_OFFSET);
                 }
             }
@@ -601,7 +647,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         for (int col = 0; col < cols_per_thread; ++col) {
 #pragma unroll
             for (int offset = 16; offset >= 4; offset >>= 1) {
-                KQ_max_new[col] = fmaxf(KQ_max_new[col], __shfl_xor_sync(0xFFFFFFFF, KQ_max_new[col], offset, WARP_SIZE));
+                KQ_max_new[col] = fmaxf(KQ_max_new[col], __shfl_xor_sync(0xFFFFFFFF, KQ_max_new[col], offset, warp_size));
             }
         }
 
@@ -611,12 +657,12 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
 #pragma unroll
             for (int l = 0; l < T_C_KQ::ne; ++l) {
                 if (!oob_check || k0 + (threadIdx.y % np)*T_C_KQ::I + T_C_KQ::get_i(l) < k_VKQ_sup) {
-#if defined(AMD_WMMA_AVAILABLE)
+#if defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
                     constexpr int KQ_idx = 0;
 #else
                     // Turing + Volta:
                     const int KQ_idx = l % 2;
-#endif // defined(AMD_WMMA_AVAILABLE)
+#endif // defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
                     KQ_C[k0/(np*T_C_KQ::I)].x[l] = expf(KQ_C[k0/(np*T_C_KQ::I)].x[l] - KQ_max_new[KQ_idx]);
                     KQ_rowsum_add[KQ_idx] += KQ_C[k0/(np*T_C_KQ::I)].x[l];
                 } else {
@@ -649,12 +695,12 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
 #pragma unroll
             for (int l = 0; l < T_C_KQ::ne; ++l) {
                 if (!oob_check || k0 + (threadIdx.y % np)*T_C_KQ::J + T_C_KQ::get_j(l) < k_VKQ_sup) {
-#if defined(AMD_WMMA_AVAILABLE)
+#if defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
                     constexpr int KQ_idx = 0;
 #else
                     // Turing + Volta:
                     const int KQ_idx = (l/2) % 2;
-#endif // defined(AMD_WMMA_AVAILABLE)
+#endif // defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
                     KQ_max_new[KQ_idx] = fmaxf(KQ_max_new[KQ_idx], KQ_C[(k0/(np*T_C_KQ::J))].x[l] + FATTN_KQ_MAX_OFFSET);
                 }
             }
@@ -666,6 +712,10 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
             // Values per KQ column are spread across 4 threads:
             constexpr int offset_first = 2;
             constexpr int offset_last  = 1;
+#elif defined(AMD_MFMA_AVAILABLE)
+            // MFMA: 4 threads per Q column (threadIdx.x % 16 == col, spaced by 16).
+            constexpr int offset_first = 32;
+            constexpr int offset_last  = 16;
 #elif defined(AMD_WMMA_AVAILABLE)
             // Values per KQ column are spread across 2 threads:
             constexpr int offset_first = 16;
@@ -677,7 +727,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
 #endif // defined(TURING_MMA_AVAILABLE)
 #pragma unroll
             for (int offset = offset_first; offset >= offset_last; offset >>= 1) {
-                KQ_max_new[col] = fmaxf(KQ_max_new[col], __shfl_xor_sync(0xFFFFFFFF, KQ_max_new[col], offset, WARP_SIZE));
+                KQ_max_new[col] = fmaxf(KQ_max_new[col], __shfl_xor_sync(0xFFFFFFFF, KQ_max_new[col], offset, warp_size));
             }
         }
 
@@ -687,12 +737,12 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
 #pragma unroll
             for (int l = 0; l < T_C_KQ::ne; ++l) {
                 if (!oob_check || k0 + (threadIdx.y % np)*T_C_KQ::J + T_C_KQ::get_j(l) < k_VKQ_sup) {
-#if defined(AMD_WMMA_AVAILABLE)
+#if defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
                     constexpr int KQ_idx = 0;
 #else
                     // Turing + Volta:
                     const int KQ_idx = (l/2) % 2;
-#endif // defined(AMD_WMMA_AVAILABLE)
+#endif // defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
                     KQ_C[(k0/(np*T_C_KQ::J))].x[l] = expf(KQ_C[(k0/(np*T_C_KQ::J))].x[l] - KQ_max_new[KQ_idx]);
                     KQ_rowsum_add[KQ_idx] += KQ_C[(k0/(np*T_C_KQ::J))].x[l];
                 } else {
@@ -739,7 +789,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                 }
             }
         }
-#elif defined(AMD_WMMA_AVAILABLE)
+#elif defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
         const half2 KQ_max_scale_h2 = make_half2(
             KQ_max_scale[0], KQ_max_scale[0]);
 #pragma unroll
@@ -818,7 +868,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         }
         const half2 * tile_V_i = !V_is_K_view || i0_stop > 2*nbatch_K2 ? tile_V : tile_V + i0_start/2;
 
-#if defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+#if defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
         constexpr int i0_stride = cols_per_warp == 8 ? T_C_VKQ::I : 2*T_C_VKQ::J;
 #pragma unroll
         for (int i_VKQ_0 = i0_start; i_VKQ_0 < i0_stop; i_VKQ_0 += i0_stride) {
@@ -830,24 +880,38 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                 T_A_VKQ A; // Transposed in SRAM but not in registers, gets transposed on load.
 #if defined(LDMATRIX_TRANS_AVAILABLE)
                 load_ldmatrix_trans(A, tile_V_i + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2, stride_tile_V);
+#elif defined(AMD_MFMA_AVAILABLE)
+                // MFMA A register layout: A_mat[i=lane%16][k=4*(lane/16)+reg].
+                // Normal load gives A_mat[seq][dv] but we need A_mat[dv][seq] = V^T.
+                // Load with transposed addressing: 4 strided half loads.
+                {
+                    const half2 * xs0 = tile_V_i + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2;
+                    const half * xs0_h = (const half *) xs0;
+                    const int stride_h = stride_tile_V * 2; // stride in half units
+                    half * A_h = (half *) A.x;
+#pragma unroll
+                    for (int l = 0; l < 4; ++l) {
+                        A_h[l] = xs0_h[(4*(threadIdx.x / 16) + l) * stride_h + threadIdx.x % 16];
+                    }
+                }
 #else
                 // TODO: Try to transpose tile_V when loading gmem to smem.
                 // Use mma to transpose T_A_VKQ for RDNA.
                 T_A_VKQ A_trans;
                 load_ldmatrix(A_trans, tile_V_i + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2, stride_tile_V);
                 mma(A, A_trans, A_identity);
-#endif // defined(TURING_MMA_AVAILABLE)
+#endif // defined(LDMATRIX_TRANS_AVAILABLE)
                 if constexpr (T_B_KQ::I == 8) {
                     mma(VKQ_C[i_VKQ_0/i0_stride], A, B[k00/(np*T_A_VKQ::J)]);
                 } else {
                     // Wide version of VKQ_C is column-major.
-#if defined(AMD_WMMA_AVAILABLE)
-                    // RDNA matrix C is column-major.
+#if defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
+                    // AMD matrix C is column-major.
                     mma(VKQ_C[i_VKQ_0/i0_stride], A, B[k00/(np*T_A_VKQ::J)]);
 #else
                     // swap A and B for CUDA.
                     mma(VKQ_C[i_VKQ_0/i0_stride], B[k00/(np*T_A_VKQ::J)], A);
-#endif // defined(AMD_WMMA_AVAILABLE)
+#endif // defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
                 }
             }
         }
@@ -866,7 +930,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                 mma(VKQ_C[i_VKQ_0/i0_stride], B[k00/(np*T_A_VKQ::I)], A);
             }
         }
-#endif // defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+#endif // defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
 
         if constexpr (nstages <= 1) {
             __syncthreads(); // Only needed if tile_K == tile_V.
@@ -879,7 +943,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         tile_Q, tile_K, tile_V, tile_mask,
         Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0);
     NO_DEVICE_CODE;
-#endif // defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4))
+#endif // defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4)) || defined(AMD_MFMA_AVAILABLE)
 }
 
 #if defined(TURING_MMA_AVAILABLE)
@@ -899,7 +963,7 @@ template<> struct mma_tile_sizes<8> {
     using T_B_VKQ = tile< 8,  8, half2>; // column-major
     using T_C_VKQ = tile<16,  4, half2>; // row-major
 };
-#elif defined(AMD_WMMA_AVAILABLE)
+#elif defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
 template<int ncols> struct mma_tile_sizes {
     using T_A_KQ  = tile<16,  8, half2>; // row-major
     using T_B_KQ  = tile<16,  8, half2>; // column-major
@@ -944,9 +1008,10 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         const int zt_gqa,
         const int kb0_start,
         const int kb0_stop) {
-#if defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4))
+#if defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4)) || defined(AMD_MFMA_AVAILABLE)
     //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
 
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     constexpr int ncols = ncols1 * ncols2;
     using     T_A_KQ    = typename mma_tile_sizes<ncols>::T_A_KQ;
     using     T_B_KQ    = typename mma_tile_sizes<ncols>::T_B_KQ;
@@ -986,7 +1051,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
     T_B_KQ    Q_B[(Q_in_reg ? DKQ/(2*T_B_KQ::J) : 1)];
 #if defined(TURING_MMA_AVAILABLE)
     T_C_VKQ VKQ_C[cols_per_warp == 8 ? DV/T_C_VKQ::I : DV/(2*T_C_VKQ::J)];
-#elif defined(AMD_WMMA_AVAILABLE)
+#elif defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
     T_C_VKQ VKQ_C[                                     DV/(2*T_C_VKQ::J)];
 #else // Volta
     T_C_VKQ VKQ_C[                                     DV/(2*T_C_VKQ::J)];
@@ -1004,10 +1069,10 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
     // The loading is done with decreasing granularity for D for better memory bandwidth.
     const half2 scale_h2 = make_half2(scale, scale);
 #pragma unroll
-    for (int stride_k : {WARP_SIZE, WARP_SIZE/2, WARP_SIZE/4}) {
-        const int k0_start  = stride_k == WARP_SIZE ? 0 : DKQ/2 - (DKQ/2) % (2*stride_k);
+    for (int stride_k : {warp_size, warp_size/2, warp_size/4, warp_size/8}) {
+        const int k0_start  = stride_k == warp_size ? 0 : DKQ/2 - (DKQ/2) % (2*stride_k);
         const int k0_stop   =                             DKQ/2 - (DKQ/2) % (1*stride_k);
-        const int stride_jc = WARP_SIZE / stride_k;
+        const int stride_jc = warp_size / stride_k;
 
         if (k0_start == k0_stop) {
             continue;
@@ -1015,7 +1080,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
 
 #pragma unroll
         for (int jc0 = 0; jc0 < ncols; jc0 += nwarps*stride_jc) {
-            const int jc = jc0 + threadIdx.y*stride_jc + (stride_k == WARP_SIZE ? 0 : threadIdx.x / stride_k);
+            const int jc = jc0 + threadIdx.y*stride_jc + (stride_k == warp_size ? 0 : threadIdx.x / stride_k);
 
             if (jc0 + nwarps*stride_jc > ncols && jc >= ncols) {
                 break;
@@ -1027,7 +1092,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
             if ((ncols1 == 1 || jt*ncols1 + j < int(ne01.z)) && (ncols2 == 1 || zt_gqa*ncols2 + c < gqa_ratio)) {
 #pragma unroll
                 for (int k0 = k0_start; k0 < k0_stop; k0 += stride_k) {
-                    const int k = k0 + (stride_k == WARP_SIZE ? threadIdx.x : threadIdx.x % stride_k);
+                    const int k = k0 + (stride_k == warp_size ? threadIdx.x : threadIdx.x % stride_k);
 
                     const float2 tmp = Q_f2[(jt*ncols1 + j)*stride_Q1 + c*stride_Q2 + k];
                     tile_Q[jc*stride_tile_Q + k] = scale_h2 * make_half2(tmp.x, tmp.y);
@@ -1035,7 +1100,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
             } else {
 #pragma unroll
                 for (int k0 = k0_start; k0 < k0_stop; k0 += stride_k) {
-                    const int k = k0 + (stride_k == WARP_SIZE ? threadIdx.x : threadIdx.x % stride_k);
+                    const int k = k0 + (stride_k == warp_size ? threadIdx.x : threadIdx.x % stride_k);
 
                     tile_Q[jc*stride_tile_Q + k] = make_half2(0.0f, 0.0f);
                 }
@@ -1127,6 +1192,10 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         // The partial sums are spread across 8/4 threads.
         constexpr int offset_first = cols_per_warp == 8 ? 16 : 2;
         constexpr int offset_last  = cols_per_warp == 8 ?  4 : 1;
+#elif defined(AMD_MFMA_AVAILABLE)
+        // The partial sums are spread across 4 threads (wavefront64, 16 cols).
+        constexpr int offset_first = 32;
+        constexpr int offset_last  = 16;
 #elif defined(AMD_WMMA_AVAILABLE)
         // The partial sums are spread across 2 threads.
         constexpr int offset_first = 16;
@@ -1140,7 +1209,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         for (int col = 0; col < cols_per_thread; ++col) {
 #pragma unroll
             for (int offset = offset_first; offset >= offset_last; offset >>= 1) {
-                KQ_rowsum[col] += __shfl_xor_sync(0xFFFFFFFF, KQ_rowsum[col], offset, WARP_SIZE);
+                KQ_rowsum[col] += __shfl_xor_sync(0xFFFFFFFF, KQ_rowsum[col], offset, warp_size);
             }
         }
     }
@@ -1189,7 +1258,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
                 }
             }
         }
-#elif defined(AMD_WMMA_AVAILABLE)
+#elif defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
         const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[0], KQ_max_scale[0]);
 #pragma unroll
         for (int i = 0; i < (DV/2)/T_C_VKQ::J; ++i) {
@@ -1249,7 +1318,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         const int jc_cwm = threadIdx.y*cols_per_warp + T_C_VKQ::get_i(threadIdx.x % 4);
         const float2 KQ_cmr = make_float2(KQ_max[threadIdx.x % cols_per_thread], KQ_rowsum[threadIdx.x % cols_per_thread]);
         const bool thread_should_write = threadIdx.x % 4 < cols_per_thread;
-#elif defined(AMD_WMMA_AVAILABLE)
+#elif defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
         const int jc_cwm = threadIdx.y*cols_per_warp + T_C_VKQ::get_i(0);
         const float2 KQ_cmr = make_float2(KQ_max[0], KQ_rowsum[0]);
         const bool thread_should_write = threadIdx.x / 16 < cols_per_thread;
@@ -1283,14 +1352,14 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         // Warps with threadIdx.y % np != 0 must NOT return early.
         // All threads must return simultaneously to avoid race conditions with work on the next tile.
 
-        constexpr int nmeta = np*cols_per_warp >= WARP_SIZE ? np*cols_per_warp/WARP_SIZE : 1;
+        constexpr int nmeta = np*cols_per_warp >= warp_size ? np*cols_per_warp/warp_size : 1;
 
-        const int jc_meta = threadIdx.y*cols_per_warp + (np*cols_per_warp < WARP_SIZE ? threadIdx.x % (np*cols_per_warp) : threadIdx.x);
+        const int jc_meta = threadIdx.y*cols_per_warp + (np*cols_per_warp < warp_size ? threadIdx.x % (np*cols_per_warp) : threadIdx.x);
         float2 * const meta_ptr = ((float2 *) tile_Q) + jc_meta*(tile_stride/2) + nbatch_combine/2;
         float2 meta[nmeta];
 #pragma unroll
         for (int imeta = 0; imeta < nmeta; ++imeta) {
-            meta[imeta] = meta_ptr[imeta * WARP_SIZE * tile_stride/2];
+            meta[imeta] = meta_ptr[imeta * warp_size * tile_stride/2];
         }
 
         float KQ_cmn = meta[0].x; // KQ combine max new, max between all parallel warps.
@@ -1300,8 +1369,8 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         }
 #pragma unroll
         for (int offset = np*cols_per_warp/2; offset >= cols_per_warp; offset >>= 1) {
-            if (offset < WARP_SIZE) {
-                KQ_cmn = fmaxf(KQ_cmn, __shfl_xor_sync(0xFFFFFFFF, KQ_cmn, offset, WARP_SIZE));
+            if (offset < warp_size) {
+                KQ_cmn = fmaxf(KQ_cmn, __shfl_xor_sync(0xFFFFFFFF, KQ_cmn, offset, warp_size));
             }
         }
 
@@ -1318,8 +1387,8 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         }
 #pragma unroll
         for (int offset = np*cols_per_warp/2; offset >= cols_per_warp; offset >>= 1) {
-            if (offset < WARP_SIZE) {
-                KQ_crs += __shfl_xor_sync(0xFFFFFFFF, KQ_crs, offset, WARP_SIZE);
+            if (offset < warp_size) {
+                KQ_crs += __shfl_xor_sync(0xFFFFFFFF, KQ_crs, offset, warp_size);
             }
         }
 
@@ -1328,19 +1397,19 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         // Write back combined meta data:
 #pragma unroll
         for (int imeta = 0; imeta < nmeta; ++imeta) {
-            if (np*cols_per_warp >= WARP_SIZE || threadIdx.x < np*cols_per_warp) {
+            if (np*cols_per_warp >= warp_size || threadIdx.x < np*cols_per_warp) {
                 // Combined KQ max scale + rowsum.
-                meta_ptr[imeta * WARP_SIZE * tile_stride/2] = make_float2(KQ_cms[imeta], KQ_crs);
+                meta_ptr[imeta * warp_size * tile_stride/2] = make_float2(KQ_cms[imeta], KQ_crs);
             }
         }
 
         // Combined KQ max + rowsum.
-        static_assert(cols_per_warp <= WARP_SIZE);
-        if (needs_fixup && (cols_per_warp == WARP_SIZE || threadIdx.x < cols_per_warp)) {
+        static_assert(cols_per_warp <= warp_size);
+        if (needs_fixup && (cols_per_warp == warp_size || threadIdx.x < cols_per_warp)) {
             float2 * dstk_fixup_meta = dstk_fixup + blockIdx.x*ncols;
             dstk_fixup_meta[(threadIdx.y/np)*cols_per_warp + threadIdx.x] = make_float2(KQ_cmn, KQ_crs);
         }
-        if (is_fixup && (cols_per_warp == WARP_SIZE || threadIdx.x < cols_per_warp)) {
+        if (is_fixup && (cols_per_warp == warp_size || threadIdx.x < cols_per_warp)) {
             float2 * dstk_fixup_meta = dstk_fixup + (gridDim.x + blockIdx.x)*ncols;
             dstk_fixup_meta[(threadIdx.y/np)*cols_per_warp + threadIdx.x] = make_float2(KQ_cmn, KQ_crs);
         }
@@ -1388,10 +1457,10 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
             float2 * dstk_fixup_data = dstk_fixup + gridDim.x*(2*ncols) + blockIdx.x*(ncols*(DV/2));
 
 #pragma unroll
-            for (int stride_k : {WARP_SIZE, WARP_SIZE/2, WARP_SIZE/4}) {
-                const int k0_start  = stride_k == WARP_SIZE ? 0 : nbatch_combine - nbatch_combine % (2*stride_k);
+            for (int stride_k : {warp_size, warp_size/2, warp_size/4, warp_size/8}) {
+                const int k0_start  = stride_k == warp_size ? 0 : nbatch_combine - nbatch_combine % (2*stride_k);
                 const int k0_stop   =                             nbatch_combine - nbatch_combine % (1*stride_k);
-                const int stride_jc = WARP_SIZE / stride_k;
+                const int stride_jc = warp_size / stride_k;
 
                 if (k0_start == k0_stop) {
                     continue;
@@ -1399,7 +1468,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
 
 #pragma unroll
                 for (int jc0_dst = 0; jc0_dst < ncols; jc0_dst += (nwarps/np)*stride_jc) {
-                    const int jc_dst = jc0_dst + (threadIdx.y/np)*stride_jc + (stride_k == WARP_SIZE ? 0 : threadIdx.x / stride_k);
+                    const int jc_dst = jc0_dst + (threadIdx.y/np)*stride_jc + (stride_k == warp_size ? 0 : threadIdx.x / stride_k);
 
                     if (jc0_dst + (nwarps/np)*stride_jc > ncols && jc_dst >= ncols) {
                         break;
@@ -1417,7 +1486,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
                     const float * meta_j = (const float *) tile_Q + jc_tile_K*tile_stride + nbatch_combine;
 #pragma unroll
                     for (int k0 = k0_start; k0 < k0_stop; k0 += stride_k) {
-                        const int k = k0 + (stride_k == WARP_SIZE ? threadIdx.x : threadIdx.x % stride_k);
+                        const int k = k0 + (stride_k == warp_size ? threadIdx.x : threadIdx.x % stride_k);
 
                         float2 dstk_val = make_float2(0.0f, 0.0f);
 #pragma unroll
@@ -1453,7 +1522,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         stride_Q1, stride_Q2, stride_K, stride_V, stride_mask,
         jt, kb0_start, kb0_stop);
     NO_DEVICE_CODE;
-#endif // defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4))
+#endif // defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4)) || defined(AMD_MFMA_AVAILABLE)
 }
 
 template<int DKQ, int DV, int ncols1, int ncols2, bool use_logit_softcap, bool V_is_K_view>
@@ -1480,7 +1549,7 @@ static __global__ void flash_attn_ext_f16(
                             const int32_t nb21, const int32_t nb22, const int64_t nb23,
                             const int32_t ne31, const int32_t ne32, const int32_t ne33,
                             const int32_t nb31, const int32_t nb32, const int64_t nb33) {
-#if defined(FLASH_ATTN_AVAILABLE) && (defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4)))
+#if defined(FLASH_ATTN_AVAILABLE) && (defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4)) || defined(AMD_MFMA_AVAILABLE))
 
     // Skip unused kernel variants for faster compilation:
     if (use_logit_softcap && !(DKQ == 128 || DKQ == 256)) {
@@ -1508,10 +1577,18 @@ static __global__ void flash_attn_ext_f16(
     }
 #endif // defined(AMD_WMMA_AVAILABLE)
 
+#if defined(AMD_MFMA_AVAILABLE)
+    if (DKQ != 64 && DKQ != 80 && DKQ != 96 && DKQ != 112 && DKQ != 128) {
+        NO_DEVICE_CODE;
+        return;
+    }
+#endif // defined(AMD_MFMA_AVAILABLE)
+
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     constexpr int ncols     = ncols1 * ncols2;
     constexpr int nbatch_fa = ggml_cuda_fattn_mma_get_nbatch_fa(DKQ, DV, ncols);
     constexpr int nthreads  = ggml_cuda_fattn_mma_get_nthreads(DKQ, DV, ncols);
-    constexpr int nwarps    = nthreads / WARP_SIZE;
+    constexpr int nwarps    = nthreads / warp_size;
 
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
 
@@ -1624,7 +1701,7 @@ static __global__ void flash_attn_ext_f16(
               ne31, ne32, ne33,
               nb31, nb32, nb33);
     NO_DEVICE_CODE;
-#endif // defined(FLASH_ATTN_AVAILABLE) && (defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4)))
+#endif // defined(FLASH_ATTN_AVAILABLE) && (defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4)) || defined(AMD_MFMA_AVAILABLE))
 }
 
 template <int DKQ, int DV, int ncols1, int ncols2>
@@ -1644,7 +1721,8 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
     const int  nstages        = ggml_cuda_fattn_mma_get_nstages       (DKQ, DV, ncols1, ncols2, cc);
 
     const int cols_per_warp = std::min(ncols, get_cols_per_warp(cc));
-    const int nwarps        = nthreads / WARP_SIZE;
+    const int warp_size_host = ggml_cuda_info().devices[ctx.device].warp_size;
+    const int nwarps         = nthreads / warp_size_host;
 
     constexpr bool V_is_K_view = DKQ == 576; // Guaranteed by the kernel selection logic in fattn.cu
 
@@ -1694,7 +1772,7 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
     }
 
     launch_fattn<DV, ncols1, ncols2>
-        (ctx, dst, fattn_kernel, nwarps, nbytes_shared_total, nbatch_fa, true, true, true);
+        (ctx, dst, fattn_kernel, nwarps, nbytes_shared_total, nbatch_fa, true, true, true, warp_size_host);
 }
 
 

ggml/src/ggml-cuda/fattn.cu

@@ -440,6 +440,18 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
         return BEST_FATTN_KERNEL_MMA_F16;
     }
 
+    // Use MFMA flash attention for CDNA (MI100+):
+    if (amd_mfma_available(cc) && Q->ne[0] != 40 && Q->ne[0] != 72 && Q->ne[0] != 256 && Q->ne[0] != 576) {
+        const int64_t eff_nq = Q->ne[1] * (gqa_opt_applies ? gqa_ratio : 1);
+        // MMA vs tile crossover benchmarked on MI300X @ d32768:
+        //   hsk=64  (gqa=4): MMA wins at eff >= 128 (+11%)
+        //   hsk=128 (gqa=4): MMA wins at eff >= 128 (+4%)
+        if (eff_nq >= (GGML_CUDA_CC_IS_CDNA1(cc) && Q->ne[0] == 64 ? 64 : 128)) {
+            return BEST_FATTN_KERNEL_MMA_F16;
+        }
+        // Fall through to tile kernel for small effective batch sizes.
+    }
+
     // If there are no tensor cores available, use the generic tile kernel:
     if (can_use_vector_kernel) {
         if (!ggml_is_quantized(K->type) && !ggml_is_quantized(V->type)) {

ggml/src/ggml-cuda/mma.cuh

@@ -668,7 +668,7 @@ namespace ggml_cuda_mma {
 
         return ret;
     }
-#elif defined(AMD_WMMA_AVAILABLE)
+#elif defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
     template <int I, int J>
     static __device__ __forceinline__ tile<I, J/2, half2> get_half2(const tile<I, J, float> & tile_float) {
         tile<I, J/2, half2> ret;
@@ -964,6 +964,34 @@ namespace ggml_cuda_mma {
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // defined(RDNA4)
+#elif defined(AMD_MFMA_AVAILABLE)
+        // MFMA: FP16 input, FP32 accumulate, convert back to half2.
+        using halfx4_t = __attribute__((ext_vector_type(4))) _Float16;
+        using floatx4_t = __attribute__((ext_vector_type(4))) float;
+
+        // Convert existing half2 accumulator to float for MFMA:
+        floatx4_t acc_f32;
+        {
+            const halfx4_t acc_h = reinterpret_cast<const halfx4_t&>(D.x[0]);
+#pragma unroll
+            for (int i = 0; i < 4; ++i) {
+                acc_f32[i] = (float)acc_h[i];
+            }
+        }
+
+        const halfx4_t& a_frag = reinterpret_cast<const halfx4_t&>(A.x[0]);
+        const halfx4_t& b_frag = reinterpret_cast<const halfx4_t&>(B.x[0]);
+        acc_f32 = __builtin_amdgcn_mfma_f32_16x16x16f16(a_frag, b_frag, acc_f32, 0, 0, 0);
+
+        // Convert back to half2:
+        {
+            halfx4_t result_h;
+#pragma unroll
+            for (int i = 0; i < 4; ++i) {
+                result_h[i] = (_Float16)acc_f32[i];
+            }
+            reinterpret_cast<halfx4_t&>(D.x[0]) = result_h;
+        }
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;

ggml/src/ggml-sycl/add-id.cpp

@@ -55,7 +55,11 @@ void ggml_sycl_add_id(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
   const int32_t* src2_d = (const int32_t*)src2->data;
   float* dst_d = (float*)dst->data;
 
-  int threads = std::min((int)ne00, 768);  // cols
+  const unsigned int max_work_group_size = ggml_sycl_info().max_work_group_sizes[ctx.device];
+  assert(work_group_size % (WARP_SIZE * WARP_SIZE) == 0);
+
+  int threads = std::min((unsigned int)ne00, max_work_group_size);  // cols
+
   ctx.stream()->parallel_for(
       sycl::nd_range<3>(
           sycl::range<3>(1, ne02, ne01) * sycl::range<3>(1, 1, threads),

ggml/src/ggml-sycl/binbcast.cpp

@@ -11,8 +11,8 @@ static void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst,
         int ne0, int ne1, int ne2, int ne3,
         int ne10, int ne11, int ne12, int ne13,
         /*int s0, */ int s1,  int s2,  int s3,
-        /*int s00,*/ int s01, int s02, int s03,
-        /*int s10,*/ int s11, int s12, int s13,
+        int s00, int s01, int s02, int s03,
+        int s10, int s11, int s12, int s13,
         const sycl::nd_item<3> &item_ct1) {
     const int i0s = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
                     item_ct1.get_local_id(2);
@@ -44,7 +44,7 @@ static void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst,
     for (int i0 = i0s; i0 < ne0;
          i0 += item_ct1.get_local_range(2) * item_ct1.get_group_range(2)) {
         const int i10 = i0 % ne10;
-        dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
+        dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0*s00] : 0.0f, (float)src1_row[i10*s10]);
     }
 }
 
@@ -53,8 +53,8 @@ static void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t
         int ne0, int ne1, int ne2, int ne3,
         int ne10, int ne11, int ne12, int ne13,
         /*int s0, */ int s1,  int s2,  int s3,
-        /*int s00,*/ int s01, int s02, int s03,
-        /*int s10,*/ int s11, int s12, int s13,
+        int s00, int s01, int s02, int s03,
+        int s10, int s11, int s12, int s13,
         const sycl::nd_item<3> &item_ct1) {
 
     const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
@@ -82,7 +82,7 @@ static void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t
     dst_t * dst_row = dst + i_dst;
 
     const int i10 = i0 % ne10;
-    dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
+    dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0*s00] : 0.0f, (float)src1_row[i10*s10]);
 }
 
 
@@ -95,7 +95,8 @@ struct bin_bcast_sycl {
                     const int64_t ne3, const size_t nb00, const size_t nb01, const size_t nb02, const size_t nb03,
                     const size_t nb10, const size_t nb11, const size_t nb12, const size_t nb13, const size_t nb0,
                     const size_t nb1, const size_t nb2, const size_t nb3, const bool src0_is_contiguous,
-                    const bool src1_is_contiguous, const bool dst_is_contiguous, queue_ptr stream) {
+                    const bool src1_is_contiguous, const bool src0_is_permuted, const bool src1_is_permuted,
+                    queue_ptr stream) {
         int nr0 = ne10 / ne0;
         int nr1 = ne11/ne1;
         int nr2 = ne12/ne2;
@@ -123,7 +124,7 @@ struct bin_bcast_sycl {
             cnb[3] *= cne[3];
         };
 
-        if (src0_is_contiguous && src1_is_contiguous && dst_is_contiguous) {
+        if (src0_is_contiguous && src1_is_contiguous && !src0_is_permuted && !src1_is_permuted) {
             for (int i = 0; i < 4; i++) {
                 if (nr[i] != 1) {
                     break;
@@ -164,7 +165,7 @@ struct bin_bcast_sycl {
             size_t nb12 = cnb1[2];
             size_t nb13 = cnb1[3];
 
-            size_t s0 = nb0 / sizeof(dst_t);
+            // size_t s0 = nb0 / sizeof(dst_t);
             size_t s1 = nb1 / sizeof(dst_t);
             size_t s2 = nb2 / sizeof(dst_t);
             size_t s3 = nb3 / sizeof(dst_t);
@@ -196,9 +197,6 @@ struct bin_bcast_sycl {
             GGML_ASSERT(nb12 % sizeof(src1_t) == 0);
             GGML_ASSERT(nb13 % sizeof(src1_t) == 0);
 
-            GGML_ASSERT(s0 == 1);
-            GGML_ASSERT(s10 == 1);
-
             const int block_size = 128;
 
             int64_t hne0 = std::max(ne0/2LL, 1LL);
@@ -232,8 +230,8 @@ struct bin_bcast_sycl {
                         [=](sycl::nd_item<3> item_ct1) {
                             k_bin_bcast_unravel<bin_op>(
                                 src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3,
-                                ne10, ne11, ne12, ne13, s1, s2, s3, s01, s02,
-                                s03, s11, s12, s13, item_ct1);
+                                ne10, ne11, ne12, ne13, s1, s2, s3, s00, s01, s02,
+                                s03, s10, s11, s12, s13, item_ct1);
                         });
                 }
             } else {
@@ -251,7 +249,7 @@ struct bin_bcast_sycl {
                     [=](sycl::nd_item<3> item_ct1) {
                         k_bin_bcast<bin_op>(src0_dd, src1_dd, dst_dd, ne0, ne1,
                                             ne2, ne3, ne10, ne11, ne12, ne13,
-                                            s1, s2, s3, s01, s02, s03, s11, s12, s13,
+                                            s1, s2, s3, s00, s01, s02, s03, s10, s11, s12, s13,
                                             item_ct1);
                     });
             }
@@ -268,24 +266,27 @@ inline void ggml_sycl_op_bin_bcast(ggml_backend_sycl_context & ctx, const ggml_t
     if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
         op()((const float *) src0->data, (const float *) src1->data, (float *) dst->data, ne00, ne01, ne02, ne03, ne10,
              ne11, ne12, ne13, ne0, ne1, ne2, ne3, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb0, nb1, nb2, nb3,
-             ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_contiguous(dst), main_stream);
+             ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_permuted(src0), ggml_is_permuted(src1), main_stream);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
         op()((const sycl::half *) src0->data, (const sycl::half *) src1->data, (sycl::half *) dst->data, ne00, ne01,
              ne02, ne03, ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13,
-             nb0, nb1, nb2, nb3, ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_contiguous(dst),
+             nb0, nb1, nb2, nb3, ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_permuted(src0), ggml_is_permuted(src1),
              main_stream);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) {
         op()((const sycl::half *) src0->data, (const float *) src1->data, (sycl::half *) dst->data, ne00, ne01, ne02,
              ne03, ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb0, nb1,
-             nb2, nb3, ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_contiguous(dst), main_stream);
+             nb2, nb3, ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_permuted(src0), ggml_is_permuted(src1),
+             main_stream);
     } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32 && dst->type == GGML_TYPE_I32) {
         op()((const int32_t *) src0->data, (const int32_t *) src1->data, (int32_t *) dst->data, ne00, ne01, ne02, ne03,
              ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb0, nb1, nb2,
-             nb3, ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_contiguous(dst), main_stream);
+             nb3, ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_permuted(src0), ggml_is_permuted(src1),
+             main_stream);
     } else if (src0->type == GGML_TYPE_I16 && src1->type == GGML_TYPE_I16 && dst->type == GGML_TYPE_I16) {
         op()((const int16_t *) src0->data, (const int16_t *) src1->data, (int16_t *) dst->data, ne00, ne01, ne02, ne03,
              ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb0, nb1, nb2,
-             nb3, ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_contiguous(dst), main_stream);
+             nb3, ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_permuted(src0), ggml_is_permuted(src1),
+             main_stream);
     } else {
         fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__, ggml_type_name(dst->type),
                 ggml_type_name(src0->type), ggml_type_name(src1->type));

ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp

@@ -7,9 +7,21 @@
 
 #include <cstdint>
 
+static uint32_t validate_graph_operation(size_t cgraph_size, uint32_t shmem_res_id, const char * operation) {
+    if (cgraph_size == 0) {
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Zero-size computation graph\n", operation);
+        return 1;
+    }
+
+    // place-holder: validate that the size of shmem_res_id is <= cgraph_size
+    // need to add another method in the Virgl->APIR callback interface
+    GGML_UNUSED(shmem_res_id);
+
+    return 0;  // Valid
+}
+
 uint32_t backend_backend_graph_compute(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
     GGML_UNUSED(ctx);
-    GGML_UNUSED(enc);
 
     static bool async_backend_initialized = false;
     static bool async_backend;
@@ -34,10 +46,26 @@ uint32_t backend_backend_graph_compute(apir_encoder * enc, apir_decoder * dec, v
     size_t cgraph_size;
     apir_decode_size_t(dec, &cgraph_size);
 
+    if (validate_graph_operation(cgraph_size, shmem_res_id, __func__) != 0) {
+        apir_decoder_set_fatal(dec);
+        return 1;
+    }
+
     apir_decoder secondary_dec = apir_new_decoder((const char *) shmem_data, cgraph_size);
 
     ggml_cgraph * cgraph = apir_decode_ggml_cgraph(&secondary_dec, cgraph_size);
 
+    if (!cgraph || apir_decoder_get_fatal(&secondary_dec)) {
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Failed to deserialize computation graph\n", __func__);
+        return 1;
+    }
+
+    if (cgraph->n_nodes < 0 || cgraph->n_leafs < 0) {
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Invalid negative node/leaf count: nodes=%d leafs=%d\n", __func__,
+                       cgraph->n_nodes, cgraph->n_leafs);
+        return 1;
+    }
+
     ggml_status status;
 #if APIR_BACKEND_CHECK_SUPPORTS_OP == 1
     for (int idx = 0; idx < cgraph->n_nodes; idx++) {
@@ -45,7 +73,8 @@ uint32_t backend_backend_graph_compute(apir_encoder * enc, apir_decoder * dec, v
         if (dev->iface.supports_op(dev, op)) {
             continue;
         }
-        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Graph node %d (%s) not supported by the backend\n", idx, ggml_op_desc(op));
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Graph node %d (%s) not supported by the backend\n", __func__, idx,
+                       ggml_op_desc(op));
 
         status = GGML_STATUS_ABORTED;
         apir_encode_ggml_status(enc, &status);
@@ -53,9 +82,17 @@ uint32_t backend_backend_graph_compute(apir_encoder * enc, apir_decoder * dec, v
         return 0;
     }
 #endif
+
+    // Check if backend is properly initialized
+    if (!bck) {
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Backend not initialized (bck is null)\n", __func__);
+
+        return 1;
+    }
+
     status = bck->iface.graph_compute(bck, cgraph);
 
-    if (async_backend) {
+    if (async_backend && bck->iface.synchronize) {
         bck->iface.synchronize(bck);
     }
 

ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp

@@ -85,7 +85,19 @@ uint32_t backend_buffer_type_get_alloc_size(apir_encoder * enc, apir_decoder * d
 
     const ggml_tensor * op = apir_decode_ggml_tensor_inplace(dec);
 
-    size_t value = buft->iface.get_alloc_size(buft, op);
+    // Check for decode error
+    if (op == nullptr) {
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Failed to decode tensor\n", __func__);
+        apir_decoder_set_fatal(dec);
+        return 1;
+    }
+
+    size_t value;
+    if (buft->iface.get_alloc_size) {
+        value = buft->iface.get_alloc_size(buft, op);
+    } else {
+        value = ggml_nbytes(op);  // Default fallback
+    }
 
     apir_encode_size_t(enc, &value);
 

ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp

@@ -6,11 +6,26 @@
 
 #include <cstdint>
 
+static uint32_t validate_buffer_operation(size_t offset, size_t size, const char * operation) {
+    // Only check for critical integer overflow - no arbitrary size limits
+    if (offset > SIZE_MAX - size) {
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Integer overflow in offset+size: %zu + %zu\n", operation, offset, size);
+        return 1;
+    }
+
+    return 0;  // Valid
+}
+
 uint32_t backend_buffer_get_base(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
     GGML_UNUSED(ctx);
     ggml_backend_buffer_t buffer;
     buffer = apir_decode_ggml_buffer(dec);
 
+    if (!buffer || apir_decoder_get_fatal(dec)) {
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Invalid buffer handle from guest\n", __func__);
+        return 1;
+    }
+
     uintptr_t base = (uintptr_t) buffer->iface.get_base(buffer);
     apir_encode_uintptr_t(enc, &base);
 
@@ -24,6 +39,11 @@ uint32_t backend_buffer_set_tensor(apir_encoder * enc, apir_decoder * dec, virgl
     ggml_backend_buffer_t buffer;
     buffer = apir_decode_ggml_buffer(dec);
 
+    if (!buffer || apir_decoder_get_fatal(dec)) {
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Invalid buffer handle from guest\n", __func__);
+        return 1;
+    }
+
     ggml_tensor * tensor;
     // safe to remove the const qualifier here
     tensor = (ggml_tensor *) (uintptr_t) apir_decode_ggml_tensor(dec);
@@ -37,6 +57,10 @@ uint32_t backend_buffer_set_tensor(apir_encoder * enc, apir_decoder * dec, virgl
     size_t size;
     apir_decode_size_t(dec, &size);
 
+    if (validate_buffer_operation(offset, size, __func__) != 0) {
+        return 1;
+    }
+
     void * shmem_data = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id);
 
     if (!shmem_data) {
@@ -56,6 +80,11 @@ uint32_t backend_buffer_get_tensor(apir_encoder * enc, apir_decoder * dec, virgl
     ggml_backend_buffer_t buffer;
     buffer = apir_decode_ggml_buffer(dec);
 
+    if (!buffer || apir_decoder_get_fatal(dec)) {
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Invalid buffer handle from guest\n", __func__);
+        return 1;
+    }
+
     const ggml_tensor * tensor;
     // safe to remove the const qualifier here
     tensor = apir_decode_ggml_tensor(dec);
@@ -69,6 +98,10 @@ uint32_t backend_buffer_get_tensor(apir_encoder * enc, apir_decoder * dec, virgl
     size_t size;
     apir_decode_size_t(dec, &size);
 
+    if (validate_buffer_operation(offset, size, __func__) != 0) {
+        return 1;
+    }
+
     void * shmem_data = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id);
     if (!shmem_data) {
         GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Couldn't get the shmem addr from virgl\n", __func__);
@@ -86,6 +119,11 @@ uint32_t backend_buffer_cpy_tensor(apir_encoder * enc, apir_decoder * dec, virgl
     ggml_backend_buffer_t buffer;
     buffer = apir_decode_ggml_buffer(dec);
 
+    if (!buffer || apir_decoder_get_fatal(dec)) {
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Invalid buffer handle from guest\n", __func__);
+        return 1;
+    }
+
     const ggml_tensor * src;
     // safe to remove the const qualifier here
     src               = apir_decode_ggml_tensor(dec);
@@ -105,6 +143,11 @@ uint32_t backend_buffer_clear(apir_encoder * enc, apir_decoder * dec, virgl_apir
     ggml_backend_buffer_t buffer;
     buffer = apir_decode_ggml_buffer(dec);
 
+    if (!buffer || apir_decoder_get_fatal(dec)) {
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Invalid buffer handle from guest\n", __func__);
+        return 1;
+    }
+
     uint8_t value;
     apir_decode_uint8_t(dec, &value);
 
@@ -120,6 +163,11 @@ uint32_t backend_buffer_free_buffer(apir_encoder * enc, apir_decoder * dec, virg
     ggml_backend_buffer_t buffer;
     buffer = apir_decode_ggml_buffer(dec);
 
+    if (!buffer || apir_decoder_get_fatal(dec)) {
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Invalid buffer handle from guest\n", __func__);
+        return 1;
+    }
+
     if (!apir_untrack_backend_buffer(buffer)) {
         GGML_LOG_WARN(GGML_VIRTGPU_BCK "%s: unknown buffer %p\n", __func__, (void *) buffer);
         return 1;

ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp

@@ -1,6 +1,6 @@
 #include "backend-dispatched.h"
-#include "backend-virgl-apir.h"
 
+#include "backend-virgl-apir.h"
 #include "ggml-backend-impl.h"
 #include "ggml-backend.h"
 #include "ggml-impl.h"
@@ -28,19 +28,24 @@ uint32_t backend_dispatch_initialize(void * ggml_backend_reg_fct_p) {
         return APIR_BACKEND_INITIALIZE_BACKEND_REG_FAILED;
     }
 
-    if (!reg->iface.get_device_count(reg)) {
-        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: backend initialization failed: no device found\n", __func__);
+    size_t device_count = reg->iface.get_device_count(reg);
+    if (!device_count) {
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: no device found\n", __func__);
         return APIR_BACKEND_INITIALIZE_NO_DEVICE;
     }
 
     dev = reg->iface.get_device(reg, 0);
 
     if (!dev) {
-        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: backend initialization failed: no device received\n", __func__);
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: failed to get device\n", __func__);
         return APIR_BACKEND_INITIALIZE_NO_DEVICE;
     }
 
     bck = dev->iface.init_backend(dev, NULL);
+    if (!bck) {
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: backend initialization failed\n", __func__);
+        return APIR_BACKEND_INITIALIZE_BACKEND_INIT_FAILED;
+    }
 
     return APIR_BACKEND_INITIALIZE_SUCCESS;
 }

ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h

@@ -32,64 +32,6 @@ uint32_t backend_buffer_free_buffer(apir_encoder * enc, apir_decoder * dec, virg
 /* backend */
 uint32_t backend_backend_graph_compute(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
 
-static inline const char * backend_dispatch_command_name(ApirBackendCommandType type) {
-    switch (type) {
-        /* device */
-        case APIR_COMMAND_TYPE_DEVICE_GET_DEVICE_COUNT:
-            return "backend_device_get_device_count";
-        case APIR_COMMAND_TYPE_DEVICE_GET_COUNT:
-            return "backend_device_get_count";
-        case APIR_COMMAND_TYPE_DEVICE_GET_NAME:
-            return "backend_device_get_name";
-        case APIR_COMMAND_TYPE_DEVICE_GET_DESCRIPTION:
-            return "backend_device_get_description";
-        case APIR_COMMAND_TYPE_DEVICE_GET_TYPE:
-            return "backend_device_get_type";
-        case APIR_COMMAND_TYPE_DEVICE_GET_MEMORY:
-            return "backend_device_get_memory";
-        case APIR_COMMAND_TYPE_DEVICE_SUPPORTS_OP:
-            return "backend_device_supports_op";
-        case APIR_COMMAND_TYPE_DEVICE_GET_BUFFER_TYPE:
-            return "backend_device_get_buffer_type";
-        case APIR_COMMAND_TYPE_DEVICE_GET_PROPS:
-            return "backend_device_get_props";
-        case APIR_COMMAND_TYPE_DEVICE_BUFFER_FROM_PTR:
-            return "backend_device_buffer_from_ptr";
-        /* buffer-type */
-        case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_NAME:
-            return "backend_buffer_type_get_name";
-        case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALIGNMENT:
-            return "backend_buffer_type_get_alignment";
-        case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE:
-            return "backend_buffer_type_get_max_size";
-        case APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST:
-            return "backend_buffer_type_is_host (DEPRECATED)";
-        case APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER:
-            return "backend_buffer_type_alloc_buffer";
-        case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE:
-            return "backend_buffer_type_get_alloc_size";
-        /* buffer */
-        case APIR_COMMAND_TYPE_BUFFER_GET_BASE:
-            return "backend_buffer_get_base";
-        case APIR_COMMAND_TYPE_BUFFER_SET_TENSOR:
-            return "backend_buffer_set_tensor";
-        case APIR_COMMAND_TYPE_BUFFER_GET_TENSOR:
-            return "backend_buffer_get_tensor";
-        case APIR_COMMAND_TYPE_BUFFER_CPY_TENSOR:
-            return "backend_buffer_cpy_tensor";
-        case APIR_COMMAND_TYPE_BUFFER_CLEAR:
-            return "backend_buffer_clear";
-        case APIR_COMMAND_TYPE_BUFFER_FREE_BUFFER:
-            return "backend_buffer_free_buffer";
-        /* backend */
-        case APIR_COMMAND_TYPE_BACKEND_GRAPH_COMPUTE:
-            return "backend_backend_graph_compute";
-
-        default:
-            return "unknown";
-    }
-}
-
 extern "C" {
 static const backend_dispatch_t apir_backend_dispatch_table[APIR_BACKEND_DISPATCH_TABLE_COUNT] = {
 

ggml/src/ggml-virtgpu/backend/backend-dispatched.h

@@ -1,5 +1,6 @@
 #pragma once
 
+// clang-format off
 #include <cstdint>
 #include <cstddef>
 
@@ -10,6 +11,7 @@
 #include "shared/apir_backend.h"
 #include "shared/apir_cs.h"
 #include "shared/apir_cs_ggml.h"
+// clang-format on
 
 #define GGML_VIRTGPU_BCK "ggml-virtgpu-backend: "
 

ggml/src/ggml-virtgpu/backend/backend-virgl-apir.h

@@ -19,7 +19,7 @@ struct virgl_apir_callbacks {
 };
 
 extern "C" {
-ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct virgl_apir_callbacks *virgl_cbs);
+ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct virgl_apir_callbacks * virgl_cbs);
 void                      apir_backend_deinit(uint32_t virgl_ctx_id);
 uint32_t                  apir_backend_dispatcher(uint32_t               virgl_ctx_id,
                                                   virgl_apir_callbacks * virgl_cbs,

ggml/src/ggml-virtgpu/backend/backend.cpp

@@ -1,6 +1,5 @@
 #include "backend-dispatched.h"
 #include "backend-virgl-apir.h"
-
 #include "shared/api_remoting.h"
 #include "shared/apir_backend.h"
 #include "shared/apir_cs.h"
@@ -17,10 +16,10 @@
 #define GGML_DEFAULT_BACKEND_REG "ggml_backend_init"
 
 static void * backend_library_handle = NULL;
-static FILE * apir_logfile = NULL;
+static FILE * apir_logfile           = NULL;
 
 static void log_to_file_callback(enum ggml_log_level level, const char * text, void * user_data) {
-    FILE * logfile = (FILE *)user_data;
+    FILE * logfile = (FILE *) user_data;
     fprintf(logfile, "[%d] %s", level, text);
     fflush(logfile);
 }
@@ -48,9 +47,9 @@ void apir_backend_deinit(uint32_t virgl_ctx_id) {
 }
 
 #define APIR_GGML_LIBRARY_PATH_KEY "ggml.library.path"
-#define APIR_GGML_LIBRARY_REG_KEY "ggml.library.reg"
+#define APIR_GGML_LIBRARY_REG_KEY  "ggml.library.reg"
 
-ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct virgl_apir_callbacks *virgl_cbs) {
+ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct virgl_apir_callbacks * virgl_cbs) {
     const char * dlsym_error;
 
     const char * apir_log_to_file = getenv(APIR_LLAMA_CPP_LOG_TO_FILE_ENV);
@@ -63,15 +62,13 @@ ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct
         }
     }
 
-    const char * library_name = virgl_cbs->get_config(virgl_ctx_id, APIR_GGML_LIBRARY_PATH_KEY);
+    const char * library_name      = virgl_cbs->get_config(virgl_ctx_id, APIR_GGML_LIBRARY_PATH_KEY);
     const char * virgl_library_reg = virgl_cbs->get_config(virgl_ctx_id, APIR_GGML_LIBRARY_REG_KEY);
-    const char * library_reg = virgl_library_reg ? virgl_library_reg : GGML_DEFAULT_BACKEND_REG;
+    const char * library_reg       = virgl_library_reg ? virgl_library_reg : GGML_DEFAULT_BACKEND_REG;
 
     if (!library_name) {
-        GGML_LOG_ERROR(GGML_VIRTGPU_BCK
-                       "%s: cannot open the GGML library: env var '%s' not defined\n",
-                       __func__, APIR_LLAMA_CPP_GGML_LIBRARY_PATH_ENV);
-
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: cannot open the GGML library: env var '%s' not defined\n", __func__,
+                       APIR_LLAMA_CPP_GGML_LIBRARY_PATH_ENV);
 
         return APIR_LOAD_LIBRARY_ENV_VAR_MISSING;
     }
@@ -79,16 +76,14 @@ ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct
     backend_library_handle = dlopen(library_name, RTLD_LAZY);
 
     if (!backend_library_handle) {
-        GGML_LOG_ERROR(GGML_VIRTGPU_BCK
-                       "%s: cannot open the GGML library: %s\n", __func__, dlerror());
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: cannot open the GGML library: %s\n", __func__, dlerror());
 
         return APIR_LOAD_LIBRARY_CANNOT_OPEN;
     }
 
     if (!library_reg) {
-        GGML_LOG_ERROR(GGML_VIRTGPU_BCK
-                       "%s: cannot register the GGML library: env var '%s' not defined\n",
-                       __func__, APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV);
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: cannot register the GGML library: env var '%s' not defined\n", __func__,
+                       APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV);
 
         return APIR_LOAD_LIBRARY_ENV_VAR_MISSING;
     }
@@ -96,11 +91,9 @@ ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct
     void * ggml_backend_reg_fct = dlsym(backend_library_handle, library_reg);
     dlsym_error                 = dlerror();
     if (dlsym_error) {
-        GGML_LOG_ERROR(GGML_VIRTGPU_BCK
-                       "%s: cannot find the GGML backend registration symbol '%s' (from %s): %s\n",
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: cannot find the GGML backend registration symbol '%s' (from %s): %s\n",
                        __func__, library_reg, APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV, dlsym_error);
 
-
         return APIR_LOAD_LIBRARY_SYMBOL_MISSING;
     }
 
@@ -132,13 +125,12 @@ uint32_t apir_backend_dispatcher(uint32_t               virgl_ctx_id,
 
     virgl_apir_context ctx = {
         .ctx_id = virgl_ctx_id,
-        .iface = virgl_cbs,
+        .iface  = virgl_cbs,
     };
 
     if (cmd_type >= APIR_BACKEND_DISPATCH_TABLE_COUNT) {
-        GGML_LOG_ERROR(GGML_VIRTGPU_BCK
-                       "%s: Received an invalid dispatch index (%d >= %d)\n",
-                        __func__, cmd_type, APIR_BACKEND_DISPATCH_TABLE_COUNT);
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Received an invalid dispatch index (%d >= %d)\n", __func__, cmd_type,
+                       APIR_BACKEND_DISPATCH_TABLE_COUNT);
         return APIR_BACKEND_FORWARD_INDEX_INVALID;
     }
 

ggml/src/ggml-virtgpu/backend/shared/api_remoting.h

@@ -16,28 +16,32 @@ enum ApirCommandType {
     APIR_COMMAND_TYPE_LOADLIBRARY = 1,
     APIR_COMMAND_TYPE_FORWARD     = 2,
 
-    APIR_COMMAND_TYPE_LENGTH      = 3,
+    APIR_COMMAND_TYPE_LENGTH = 3,
 };
 
 typedef uint64_t ApirCommandFlags;
 
 enum ApirLoadLibraryReturnCode {
     APIR_LOAD_LIBRARY_SUCCESS                        = 0,
+    // these error codes are returned by the Virglrenderer APIR component
     APIR_LOAD_LIBRARY_HYPERCALL_INITIALIZATION_ERROR = 1,
     APIR_LOAD_LIBRARY_ALREADY_LOADED                 = 2,
     APIR_LOAD_LIBRARY_ENV_VAR_MISSING                = 3,
     APIR_LOAD_LIBRARY_CANNOT_OPEN                    = 4,
     APIR_LOAD_LIBRARY_SYMBOL_MISSING                 = 5,
-    APIR_LOAD_LIBRARY_INIT_BASE_INDEX                = 6,  // anything above this is a APIR backend library initialization return code
+    // any value greater than this is an APIR *backend library* initialization return code
+    APIR_LOAD_LIBRARY_INIT_BASE_INDEX                = 6,
 };
 
 enum ApirForwardReturnCode {
-    APIR_FORWARD_SUCCESS         = 0,
-    APIR_FORWARD_NO_DISPATCH_FCT = 1,
-    APIR_FORWARD_TIMEOUT         = 2,
-
-    APIR_FORWARD_BASE_INDEX      = 3,  // anything above this is a APIR backend library forward return code
-} ;
+    APIR_FORWARD_SUCCESS                = 0,
+    // these error codes are returned by the Virglrenderer APIR component
+    APIR_FORWARD_NO_DISPATCH_FCT        = 1,
+    APIR_FORWARD_TIMEOUT                = 2,
+    APIR_FORWARD_FAILED_TO_SYNC_STREAMS = 3,
+    // any value greater than this index an APIR *backend library* forward return code
+    APIR_FORWARD_BASE_INDEX             = 4,
+};
 
 __attribute__((unused)) static inline const char * apir_command_name(ApirCommandType type) {
     switch (type) {
@@ -82,6 +86,7 @@ __attribute__((unused)) static const char * apir_forward_error(ApirForwardReturn
     APIR_FORWARD_ERROR(APIR_FORWARD_SUCCESS);
     APIR_FORWARD_ERROR(APIR_FORWARD_NO_DISPATCH_FCT);
     APIR_FORWARD_ERROR(APIR_FORWARD_TIMEOUT);
+    APIR_FORWARD_ERROR(APIR_FORWARD_FAILED_TO_SYNC_STREAMS);
     APIR_FORWARD_ERROR(APIR_FORWARD_BASE_INDEX);
 
     return "Unknown APIR_COMMAND_TYPE_FORWARD error";

ggml/src/ggml-virtgpu/backend/shared/apir_backend.gen.h

@@ -34,3 +34,61 @@ typedef enum ApirBackendCommandType {
     // last command_type index + 1
     APIR_BACKEND_DISPATCH_TABLE_COUNT = 23,
 } ApirBackendCommandType;
+
+static inline const char * apir_dispatch_command_name(ApirBackendCommandType type) {
+    switch (type) {
+        /* device */
+        case APIR_COMMAND_TYPE_DEVICE_GET_DEVICE_COUNT:
+            return "device_get_device_count";
+        case APIR_COMMAND_TYPE_DEVICE_GET_COUNT:
+            return "device_get_count";
+        case APIR_COMMAND_TYPE_DEVICE_GET_NAME:
+            return "device_get_name";
+        case APIR_COMMAND_TYPE_DEVICE_GET_DESCRIPTION:
+            return "device_get_description";
+        case APIR_COMMAND_TYPE_DEVICE_GET_TYPE:
+            return "device_get_type";
+        case APIR_COMMAND_TYPE_DEVICE_GET_MEMORY:
+            return "device_get_memory";
+        case APIR_COMMAND_TYPE_DEVICE_SUPPORTS_OP:
+            return "device_supports_op";
+        case APIR_COMMAND_TYPE_DEVICE_GET_BUFFER_TYPE:
+            return "device_get_buffer_type";
+        case APIR_COMMAND_TYPE_DEVICE_GET_PROPS:
+            return "device_get_props";
+        case APIR_COMMAND_TYPE_DEVICE_BUFFER_FROM_PTR:
+            return "device_buffer_from_ptr";
+        /* buffer-type */
+        case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_NAME:
+            return "buffer_type_get_name";
+        case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALIGNMENT:
+            return "buffer_type_get_alignment";
+        case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE:
+            return "buffer_type_get_max_size";
+        case APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST:
+            return "buffer_type_is_host";
+        case APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER:
+            return "buffer_type_alloc_buffer";
+        case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE:
+            return "buffer_type_get_alloc_size";
+        /* buffer */
+        case APIR_COMMAND_TYPE_BUFFER_GET_BASE:
+            return "buffer_get_base";
+        case APIR_COMMAND_TYPE_BUFFER_SET_TENSOR:
+            return "buffer_set_tensor";
+        case APIR_COMMAND_TYPE_BUFFER_GET_TENSOR:
+            return "buffer_get_tensor";
+        case APIR_COMMAND_TYPE_BUFFER_CPY_TENSOR:
+            return "buffer_cpy_tensor";
+        case APIR_COMMAND_TYPE_BUFFER_CLEAR:
+            return "buffer_clear";
+        case APIR_COMMAND_TYPE_BUFFER_FREE_BUFFER:
+            return "buffer_free_buffer";
+        /* backend */
+        case APIR_COMMAND_TYPE_BACKEND_GRAPH_COMPUTE:
+            return "backend_graph_compute";
+
+        default:
+            return "unknown";
+    }
+}

ggml/src/ggml-virtgpu/backend/shared/apir_backend.h

@@ -14,7 +14,7 @@
 #define APIR_BACKEND_INITIALIZE_BACKEND_REG_FAILED          6
 #define APIR_BACKEND_INITIALIZE_ALREADY_INITED              7
 #define APIR_BACKEND_INITIALIZE_NO_DEVICE                   8
-
+#define APIR_BACKEND_INITIALIZE_BACKEND_INIT_FAILED         9
 
 // new entries here need to be added to the apir_backend_initialize_error function below
 
@@ -39,6 +39,10 @@ static const char * apir_backend_initialize_error(int code) {
     APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_MISSING_BACKEND_SYMBOLS);
     APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_MISSING_GGML_SYMBOLS);
     APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_BACKEND_FAILED);
+    APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_BACKEND_REG_FAILED);
+    APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_ALREADY_INITED);
+    APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_NO_DEVICE);
+    APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_BACKEND_INIT_FAILED);
 
     return "Unknown APIR_BACKEND_INITIALIZE error:/";
 

ggml/src/ggml-virtgpu/backend/shared/apir_cs.h

@@ -13,7 +13,6 @@ struct apir_encoder {
     const char * start;
     const char * end;
     bool         fatal;
-
 };
 
 struct apir_decoder {
@@ -28,8 +27,8 @@ struct apir_decoder {
 
 static apir_decoder apir_new_decoder(const char * ptr, size_t size) {
     apir_decoder dec = {
-        .cur = ptr,
-        .end = ptr + size,
+        .cur   = ptr,
+        .end   = ptr + size,
         .fatal = false,
     };
 
@@ -79,10 +78,7 @@ static inline bool apir_decoder_get_fatal(const apir_decoder * dec) {
  * encode peek
  */
 
-static inline bool apir_decoder_peek_internal(apir_decoder * dec,
-                                              size_t                size,
-                                              void *                val,
-                                              size_t                val_size) {
+static inline bool apir_decoder_peek_internal(apir_decoder * dec, size_t size, void * val, size_t val_size) {
     assert(val_size <= size);
 
     if (unlikely(size > (size_t) (dec->end - dec->cur))) {
@@ -332,8 +328,7 @@ static inline void apir_decode_char_array(apir_decoder * dec, char * val, size_t
 static inline void * apir_decoder_alloc_array(size_t size, size_t count) {
     size_t alloc_size;
     if (unlikely(__builtin_mul_overflow(size, count, &alloc_size))) {
-        GGML_LOG_ERROR("%s: overflow in array allocation of %zu * %zu bytes\n",
-                       __func__, size, count);
+        GGML_LOG_ERROR("%s: overflow in array allocation of %zu * %zu bytes\n", __func__, size, count);
         return NULL;
     }
 
@@ -352,20 +347,19 @@ static inline void apir_decode_bool_t(apir_decoder * dec, bool * val) {
 
 /* apir_buffer_type_host_handle_t */
 
-static inline void apir_encode_apir_buffer_type_host_handle_t(apir_encoder *                  enc,
+static inline void apir_encode_apir_buffer_type_host_handle_t(apir_encoder *                         enc,
                                                               const apir_buffer_type_host_handle_t * val) {
     apir_encode(enc, sizeof(apir_buffer_type_host_handle_t), val, sizeof(apir_buffer_type_host_handle_t));
 }
 
-static inline void apir_decode_apir_buffer_type_host_handle_t(apir_decoder *            dec,
+static inline void apir_decode_apir_buffer_type_host_handle_t(apir_decoder *                   dec,
                                                               apir_buffer_type_host_handle_t * val) {
     apir_decode(dec, sizeof(apir_buffer_type_host_handle_t), val, sizeof(apir_buffer_type_host_handle_t));
 }
 
 /* apir_buffer_host_handle_t */
 
-static inline void apir_encode_apir_buffer_host_handle_t(apir_encoder *             enc,
-                                                         const apir_buffer_host_handle_t * val) {
+static inline void apir_encode_apir_buffer_host_handle_t(apir_encoder * enc, const apir_buffer_host_handle_t * val) {
     apir_encode(enc, sizeof(apir_buffer_host_handle_t), val, sizeof(apir_buffer_host_handle_t));
 }
 

ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h

@@ -1,11 +1,10 @@
-#include "ggml-impl.h"
 #include "apir_cs.h"
 #include "apir_cs_rpc.h"
+#include "ggml-impl.h"
 
 // ggml_buffer_to_apir_host_handle(ggml_backend_buffer_t buffer);
 
-static inline void apir_encode_ggml_buffer_host_handle(apir_encoder *                    enc,
-                                                       const apir_buffer_host_handle_t * handle);
+static inline void apir_encode_ggml_buffer_host_handle(apir_encoder * enc, const apir_buffer_host_handle_t * handle);
 
 static inline ggml_backend_buffer_t apir_decode_ggml_buffer(apir_decoder * dec);
 
@@ -22,8 +21,7 @@ static inline apir_rpc_tensor * apir_decode_apir_rpc_tensor_inplace(apir_decoder
     return (apir_rpc_tensor *) (uintptr_t) apir_decoder_use_inplace(dec, apir_rpc_tensor_size);
 }
 
-static inline apir_rpc_tensor * apir_decode_apir_rpc_tensor_array_inplace(apir_decoder * dec,
-                                                                          uint32_t       n_tensors) {
+static inline apir_rpc_tensor * apir_decode_apir_rpc_tensor_array_inplace(apir_decoder * dec, uint32_t n_tensors) {
     size_t apir_rpc_tensor_size = sizeof(apir_rpc_tensor) * n_tensors;
 
     return (apir_rpc_tensor *) (uintptr_t) apir_decoder_use_inplace(dec, apir_rpc_tensor_size);
@@ -45,9 +43,9 @@ static inline const ggml_tensor * apir_decode_ggml_tensor(apir_decoder * dec) {
     }
 
     ggml_init_params params{
-        /*.mem_size   =*/ ggml_tensor_overhead(),
-        /*.mem_buffer =*/ NULL,
-        /*.no_alloc   =*/ true,
+        /*.mem_size   =*/ggml_tensor_overhead(),
+        /*.mem_buffer =*/NULL,
+        /*.no_alloc   =*/true,
     };
 
     ggml_context * ctx = ggml_init(params);
@@ -105,6 +103,19 @@ static inline ggml_backend_buffer_t apir_decode_ggml_buffer(apir_decoder * dec)
 
     apir_decoder_read(dec, buffer_ptr_size, &buffer, buffer_ptr_size);
 
+    // SECURITY: Validate buffer handle against tracked buffers to prevent
+    // guest VM from providing arbitrary host memory addresses
+    if (buffer) {
+        extern std::unordered_set<ggml_backend_buffer_t> backend_buffers;
+        if (backend_buffers.find(buffer) == backend_buffers.end()) {
+            GGML_LOG_WARN("ggml-virtgpu-backend: %s: Invalid buffer handle from guest: %p\n", __func__,
+                          (void *) buffer);
+            // Set fatal flag to prevent further processing with invalid handle
+            apir_decoder_set_fatal(dec);
+            return NULL;
+        }
+    }
+
     return buffer;
 }
 

ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h

@@ -1,3 +1,6 @@
+#pragma once
+
+// clang-format off
 #include "ggml.h"
 #include "ggml-backend-impl.h"
 
@@ -5,6 +8,7 @@
 #include <unordered_set>
 #include <vector>
 #include <cstdint>
+// clang-format on
 
 // ggml_tensor is serialized into apir_rpc_tensor
 struct apir_rpc_tensor {

ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp

@@ -34,6 +34,7 @@ static ggml_backend_buffer_t ggml_backend_remoting_buffer_type_alloc_buffer(ggml
 static const char * ggml_backend_remoting_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
     virtgpu * gpu = BUFT_TO_GPU(buft);
 
+    // Return the prefixed name that was built once during initialization
     return gpu->cached_buffer_type.name;
 }
 
@@ -53,9 +54,8 @@ static size_t ggml_backend_remoting_buffer_type_get_alloc_size(ggml_backend_buff
                                                                const ggml_tensor *        tensor) {
     virtgpu * gpu = BUFT_TO_GPU(buft);
 
-    if (tensor->buffer == NULL
-        || !tensor->buffer->context
-        || !buft->device->iface.supports_buft(buft->device, tensor->buffer->buft)) {
+    if (tensor->buffer == NULL || !tensor->buffer->context ||
+        !buft->device->iface.supports_buft(buft->device, tensor->buffer->buft)) {
         return ggml_nbytes(tensor);
     }
 

ggml/src/ggml-virtgpu/ggml-backend-device.cpp

@@ -3,6 +3,7 @@
 static const char * ggml_backend_remoting_device_get_name(ggml_backend_dev_t dev) {
     virtgpu * gpu = DEV_TO_GPU(dev);
 
+    // Return the prefixed name that was built once during initialization
     return gpu->cached_device_info.name;
 }
 
@@ -22,7 +23,7 @@ static enum ggml_backend_dev_type ggml_backend_remoting_device_get_type(ggml_bac
 static void ggml_backend_remoting_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
     virtgpu * gpu = DEV_TO_GPU(dev);
 
-    *free = gpu->cached_device_info.memory_free;
+    *free  = gpu->cached_device_info.memory_free;
     *total = gpu->cached_device_info.memory_total;
 }
 
@@ -72,7 +73,7 @@ static void ggml_backend_remoting_device_get_props(ggml_backend_dev_t dev, ggml_
 ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_type(ggml_backend_dev_t dev) {
     virtgpu * gpu = DEV_TO_GPU(dev);
 
-    static std::atomic<bool> initialized = false;
+    static std::atomic<bool>        initialized = false;
     static ggml_backend_buffer_type buft;
 
     if (!initialized) {
@@ -95,7 +96,7 @@ ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_type(ggml_bac
 static ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_from_ptr_type(ggml_backend_dev_t dev) {
     virtgpu * gpu = DEV_TO_GPU(dev);
 
-    static std::atomic<bool> initialized = false;
+    static std::atomic<bool>        initialized = false;
     static ggml_backend_buffer_type buft;
 
     if (!initialized) {

ggml/src/ggml-virtgpu/ggml-backend-reg.cpp

@@ -7,8 +7,8 @@
 void ggml_virtgpu_cleanup(virtgpu * gpu);
 
 static virtgpu * apir_initialize() {
-    static virtgpu *         gpu          = NULL;
-    static std::atomic<bool> initialized  = false;
+    static virtgpu *         gpu         = NULL;
+    static std::atomic<bool> initialized = false;
 
     if (initialized) {
         // fast track
@@ -31,29 +31,53 @@ static virtgpu * apir_initialize() {
         }
 
         // Pre-fetch and cache all device information, it will not change
-        gpu->cached_device_info.description  = apir_device_get_description(gpu);
+        gpu->cached_device_info.description = apir_device_get_description(gpu);
         if (!gpu->cached_device_info.description) {
             GGML_ABORT(GGML_VIRTGPU "%s: failed to initialize the virtgpu device description", __func__);
         }
-        gpu->cached_device_info.name         = apir_device_get_name(gpu);
-        if (!gpu->cached_device_info.name) {
-            GGML_ABORT(GGML_VIRTGPU "%s: failed to initialize the virtgpu device name", __func__);
-        }
         gpu->cached_device_info.device_count = apir_device_get_count(gpu);
         gpu->cached_device_info.type         = apir_device_get_type(gpu);
 
-        apir_device_get_memory(gpu,
-                              &gpu->cached_device_info.memory_free,
-                              &gpu->cached_device_info.memory_total);
+        {
+            // Get the remote name and create prefixed version
+            char * rmt_device_name = apir_device_get_name(gpu);
+            if (!rmt_device_name) {
+                GGML_ABORT(GGML_VIRTGPU "%s: failed to get the virtgpu device name", __func__);
+            }
+
+            size_t device_name_len       = strlen(rmt_device_name) + 11;  // "[virtgpu] " + null terminator
+            gpu->cached_device_info.name = (char *) malloc(device_name_len);
+            if (!gpu->cached_device_info.name) {
+                free(rmt_device_name);
+                GGML_ABORT(GGML_VIRTGPU "%s: failed to allocate memory for prefixed device name", __func__);
+            }
+            snprintf(gpu->cached_device_info.name, device_name_len, "[virtgpu] %s", rmt_device_name);
+            free(rmt_device_name);
+        }
+
+        apir_device_get_memory(gpu, &gpu->cached_device_info.memory_free, &gpu->cached_device_info.memory_total);
 
         apir_buffer_type_host_handle_t buft_host_handle = apir_device_get_buffer_type(gpu);
         gpu->cached_buffer_type.host_handle             = buft_host_handle;
-        gpu->cached_buffer_type.name                    = apir_buffer_type_get_name(gpu, buft_host_handle);
-        if (!gpu->cached_buffer_type.name) {
-            GGML_ABORT(GGML_VIRTGPU "%s: failed to initialize the virtgpu buffer type name", __func__);
+        {
+            // Get the remote name and create prefixed version
+            char * rmt_name = apir_buffer_type_get_name(gpu, buft_host_handle);
+            if (!rmt_name) {
+                GGML_ABORT(GGML_VIRTGPU "%s: failed to get the virtgpu buffer type name", __func__);
+            }
+
+            size_t prefixed_len          = strlen(rmt_name) + 11;  // "[virtgpu] " + null terminator
+            gpu->cached_buffer_type.name = (char *) malloc(prefixed_len);
+            if (!gpu->cached_buffer_type.name) {
+                free(rmt_name);
+                GGML_ABORT(GGML_VIRTGPU "%s: failed to allocate memory for prefixed buffer type name", __func__);
+            }
+            snprintf(gpu->cached_buffer_type.name, prefixed_len, "[virtgpu] %s", rmt_name);
+            free(rmt_name);
         }
-        gpu->cached_buffer_type.alignment               = apir_buffer_type_get_alignment(gpu, buft_host_handle);
-        gpu->cached_buffer_type.max_size                = apir_buffer_type_get_max_size(gpu, buft_host_handle);
+
+        gpu->cached_buffer_type.alignment = apir_buffer_type_get_alignment(gpu, buft_host_handle);
+        gpu->cached_buffer_type.max_size  = apir_buffer_type_get_max_size(gpu, buft_host_handle);
 
         initialized = true;
     }
@@ -98,7 +122,7 @@ static void ggml_backend_remoting_reg_init_devices(ggml_backend_reg_t reg) {
     static std::atomic<bool> initialized = false;
 
     if (initialized) {
-        return; // fast track
+        return;  // fast track
     }
 
     {

ggml/src/ggml-virtgpu/ggml-backend.cpp

@@ -1,5 +1,5 @@
-#include "ggml-remoting.h"
 #include "../../include/ggml-virtgpu.h"
+#include "ggml-remoting.h"
 
 static const char * ggml_backend_remoting_get_name(ggml_backend_t backend) {
     UNUSED(backend);

ggml/src/ggml-virtgpu/ggml-remoting.h

@@ -9,7 +9,7 @@
 #include <string>
 
 #define GGML_VIRTGPU_NAME "ggml-virtgpu"
-#define GGML_VIRTGPU "ggml-virtgpu: "
+#define GGML_VIRTGPU      "ggml-virtgpu: "
 
 // USE_ALWAYS_TRUE_SUPPORTS_OP: 1 is fast, 0 avoid micro-benchmark crashes
 

ggml/src/ggml-virtgpu/include/apir_hw.h

@@ -3,7 +3,7 @@
 #include <stdint.h>
 
 struct virgl_renderer_capset_apir {
-   uint32_t apir_version;
-   uint32_t supports_blob_resources;
-   uint32_t reserved[4];           // For future expansion
+    uint32_t apir_version;
+    uint32_t supports_blob_resources;
+    uint32_t reserved[4];  // For future expansion
 };

ggml/src/ggml-virtgpu/regenerate_remoting.py

@@ -145,8 +145,31 @@ class RemotingCodebaseGenerator:
         enum_lines.append(f"  APIR_BACKEND_DISPATCH_TABLE_COUNT = {total_count},")
         enum_lines.append("} ApirBackendCommandType;")
 
+        # Generate function name mapping
+        func_lines = []
+        func_lines.append("static inline const char * apir_dispatch_command_name(ApirBackendCommandType type) {")
+        func_lines.append("    switch (type) {")
+
+        current_group = None
+        for func in functions:
+            # Add comment for new group
+            if func['group_name'] != current_group:
+                func_lines.append(f"        /* {func['group_description']} */")
+                current_group = func['group_name']
+
+            # Generate clean function name without backend_ prefix
+            clean_name = f"{func['group_name']}_{func['function_name']}"
+            func_lines.append(f"        case {func['enum_name']}:")
+            func_lines.append(f"            return \"{clean_name}\";")
+
+        func_lines.append("")
+        func_lines.append("        default:")
+        func_lines.append("            return \"unknown\";")
+        func_lines.append("    }")
+        func_lines.append("}")
+
         # Full header template
-        header_content = NL.join(enum_lines) + "\n"
+        header_content = NL.join(enum_lines) + "\n\n" + NL.join(func_lines) + "\n"
 
         return header_content
 
@@ -170,19 +193,6 @@ class RemotingCodebaseGenerator:
 
             decl_lines.append(f"{signature} {func['backend_function']}({params});")
 
-        # Switch cases
-        switch_lines = []
-        current_group = None
-
-        for func in functions:
-            if func['group_name'] != current_group:
-                switch_lines.append(f"  /* {func['group_description']} */")
-                current_group = func['group_name']
-
-            deprecated = " (DEPRECATED)" if func['deprecated'] else ""
-
-            switch_lines.append(f"  case {func['enum_name']}: return \"{func['backend_function']}{deprecated}\";")
-
         # Dispatch table
         table_lines = []
         current_group = None
@@ -201,15 +211,6 @@ class RemotingCodebaseGenerator:
 
 {NL.join(decl_lines)}
 
-static inline const char *backend_dispatch_command_name(ApirBackendCommandType type)
-{{
-  switch (type) {{
-{NL.join(switch_lines)}
-
-  default: return "unknown";
-  }}
-}}
-
 extern "C" {{
 static const backend_dispatch_t apir_backend_dispatch_table[APIR_BACKEND_DISPATCH_TABLE_COUNT] = {{
   {NL.join(table_lines)}

ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp

@@ -17,8 +17,8 @@ ggml_status apir_backend_graph_compute(virtgpu * gpu, ggml_cgraph * cgraph) {
     size_t               cgraph_size = apir_serialize_ggml_cgraph(cgraph, cgraph_data);
 
     virtgpu_shmem   temp_shmem;  // Local storage for large buffers
-    virtgpu_shmem * shmem = &temp_shmem;
-    bool using_shared_shmem = false;
+    virtgpu_shmem * shmem              = &temp_shmem;
+    bool            using_shared_shmem = false;
 
     if (cgraph_size <= gpu->data_shmem.mmap_size) {
         // Lock mutex before using shared data_shmem buffer
@@ -26,7 +26,7 @@ ggml_status apir_backend_graph_compute(virtgpu * gpu, ggml_cgraph * cgraph) {
             GGML_ABORT(GGML_VIRTGPU "%s: Failed to lock data_shmem mutex", __func__);
         }
         using_shared_shmem = true;
-        shmem = &gpu->data_shmem;
+        shmem              = &gpu->data_shmem;
     } else if (virtgpu_shmem_create(gpu, cgraph_size, shmem)) {
         GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate the guest-host shared buffer", __func__);
     }

ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp

@@ -62,7 +62,9 @@ size_t apir_buffer_type_get_max_size(virtgpu * gpu, apir_buffer_type_host_handle
     return max_size;
 }
 
-apir_buffer_context_t apir_buffer_type_alloc_buffer(virtgpu * gpu, apir_buffer_type_host_handle_t host_handle, size_t size) {
+apir_buffer_context_t apir_buffer_type_alloc_buffer(virtgpu *                      gpu,
+                                                    apir_buffer_type_host_handle_t host_handle,
+                                                    size_t                         size) {
     apir_encoder *        encoder;
     apir_decoder *        decoder;
     ApirForwardReturnCode ret;
@@ -84,7 +86,9 @@ apir_buffer_context_t apir_buffer_type_alloc_buffer(virtgpu * gpu, apir_buffer_t
     return buffer_context;
 }
 
-size_t apir_buffer_type_get_alloc_size(virtgpu * gpu, apir_buffer_type_host_handle_t host_handle, const ggml_tensor * op) {
+size_t apir_buffer_type_get_alloc_size(virtgpu *                      gpu,
+                                       apir_buffer_type_host_handle_t host_handle,
+                                       const ggml_tensor *            op) {
     apir_encoder *        encoder;
     apir_decoder *        decoder;
     ApirForwardReturnCode ret;

ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp

@@ -35,8 +35,8 @@ void apir_buffer_set_tensor(virtgpu *               gpu,
     apir_encode_ggml_tensor(encoder, tensor);
 
     virtgpu_shmem   temp_shmem;  // Local storage for large buffers
-    virtgpu_shmem * shmem = &temp_shmem;
-    bool using_shared_shmem = false;
+    virtgpu_shmem * shmem              = &temp_shmem;
+    bool            using_shared_shmem = false;
 
     if (size <= gpu->data_shmem.mmap_size) {
         // Lock mutex before using shared data_shmem buffer
@@ -44,7 +44,7 @@ void apir_buffer_set_tensor(virtgpu *               gpu,
             GGML_ABORT(GGML_VIRTGPU "%s: Failed to lock data_shmem mutex", __func__);
         }
         using_shared_shmem = true;
-        shmem = &gpu->data_shmem;
+        shmem              = &gpu->data_shmem;
 
     } else if (virtgpu_shmem_create(gpu, size, shmem)) {
         GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate the guest-host shared buffer", __func__);
@@ -86,8 +86,8 @@ void apir_buffer_get_tensor(virtgpu *               gpu,
     apir_encode_ggml_tensor(encoder, tensor);
 
     virtgpu_shmem   temp_shmem;  // Local storage for large buffers
-    virtgpu_shmem * shmem = &temp_shmem;
-    bool using_shared_shmem = false;
+    virtgpu_shmem * shmem              = &temp_shmem;
+    bool            using_shared_shmem = false;
 
     if (size <= gpu->data_shmem.mmap_size) {
         // Lock mutex before using shared data_shmem buffer
@@ -95,7 +95,7 @@ void apir_buffer_get_tensor(virtgpu *               gpu,
             GGML_ABORT(GGML_VIRTGPU "%s: Failed to lock data_shmem mutex", __func__);
         }
         using_shared_shmem = true;
-        shmem = &gpu->data_shmem;
+        shmem              = &gpu->data_shmem;
 
     } else if (virtgpu_shmem_create(gpu, size, shmem)) {
         GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate the guest-host shared buffer", __func__);

ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp

@@ -26,7 +26,7 @@ char * apir_device_get_name(virtgpu * gpu) {
     REMOTE_CALL(gpu, encoder, decoder, ret);
 
     const size_t string_size = apir_decode_array_size_unchecked(decoder);
-    char            * string = (char *) apir_decoder_alloc_array(sizeof(char), string_size);
+    char *       string      = (char *) apir_decoder_alloc_array(sizeof(char), string_size);
     if (!string) {
         GGML_LOG_ERROR(GGML_VIRTGPU "%s: Could not allocate the device name buffer\n", __func__);
         return NULL;
@@ -173,7 +173,7 @@ apir_buffer_context_t apir_device_buffer_from_ptr(virtgpu * gpu, size_t size, si
     REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_BUFFER_FROM_PTR);
 
     if (virtgpu_shmem_create(gpu, size, &buffer_context.shmem)) {
-        GGML_ABORT(GGML_VIRTGPU "Couldn't allocate the guest-host shared buffer");
+        GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate %ldb of guest-host shared buffer", __func__, size);
     }
 
     apir_encode_virtgpu_shmem_res_id(encoder, buffer_context.shmem.res_id);

ggml/src/ggml-virtgpu/virtgpu-forward-impl.h

@@ -1,29 +1,36 @@
-#include "virtgpu.h"
+#pragma once
 
+// clang-format off
+#include "virtgpu.h"
 #include "ggml-remoting.h"
 #include "backend/shared/apir_backend.h"
 #include "backend/shared/apir_cs_ggml.h"
-
 #include "ggml-backend-impl.h"
+// clang-format on
 
-#define REMOTE_CALL_PREPARE(gpu_dev_name, encoder_name, apir_command_type__)                               \
-    do {                                                                                                   \
-        int32_t forward_flag = (int32_t) apir_command_type__;                                              \
-        encoder_name         = remote_call_prepare(gpu_dev_name, APIR_COMMAND_TYPE_FORWARD, forward_flag); \
-        if (!encoder_name) {                                                                               \
-            GGML_ABORT(GGML_VIRTGPU "%s: failed to prepare the remote call encoder", __func__);                       \
-        }                                                                                                  \
+#define REMOTE_CALL_PREPARE(gpu_dev_name, encoder_name, apir_command_type__)                                           \
+    int32_t      REMOTE_CALL_PREPARE_forward_flag = (int32_t) apir_command_type__;                                     \
+    const char * REMOTE_CALL_PREPARE_command_name = apir_dispatch_command_name(apir_command_type__);                   \
+    do {                                                                                                               \
+        encoder_name = remote_call_prepare(gpu_dev_name, APIR_COMMAND_TYPE_FORWARD, REMOTE_CALL_PREPARE_forward_flag); \
+        if (!encoder_name) {                                                                                           \
+            GGML_ABORT(GGML_VIRTGPU "%s: failed to prepare the remote call encoder", __func__);                        \
+        }                                                                                                              \
     } while (0)
 
-#define REMOTE_CALL(gpu_dev_name, encoder_name, decoder_name, ret_name)                                           \
-    do {                                                                                                          \
-        ret_name = (ApirForwardReturnCode) remote_call(gpu_dev_name, encoder_name, &decoder_name, 0, NULL);       \
-        if (!decoder_name) {                                                                                      \
-            GGML_ABORT(GGML_VIRTGPU "%s: failed to kick the remote call", __func__);                                         \
-        }                                                                                                         \
-        if (ret_name < APIR_FORWARD_BASE_INDEX) {                                                                 \
-            GGML_ABORT(GGML_VIRTGPU "%s: failed to forward the API call: %s: code %d", __func__,                             \
-                       apir_forward_error(ret_name), ret_name);                                                   \
-        }                                                                                                         \
-        ret_name = (ApirForwardReturnCode) (ret_name - APIR_FORWARD_BASE_INDEX);                                  \
+#define REMOTE_CALL(gpu_dev_name, encoder_name, decoder_name, ret_name)                                     \
+    do {                                                                                                    \
+        ret_name = (ApirForwardReturnCode) remote_call(gpu_dev_name, encoder_name, &decoder_name, 0, NULL); \
+        if (!decoder_name) {                                                                                \
+            GGML_ABORT(GGML_VIRTGPU "%s: failed to kick the remote call", __func__);                        \
+        }                                                                                                   \
+        if (ret_name < APIR_FORWARD_BASE_INDEX) {                                                           \
+            GGML_ABORT(GGML_VIRTGPU "%s: failed to forward the API call: %s: code %d", __func__,            \
+                       apir_forward_error(ret_name), ret_name);                                             \
+        }                                                                                                   \
+        ret_name = (ApirForwardReturnCode) (ret_name - APIR_FORWARD_BASE_INDEX);                            \
+        if (ret_name != 0) {                                                                                \
+            GGML_ABORT(GGML_VIRTGPU "backend function '%s' failed (return code: %d)",                       \
+                       REMOTE_CALL_PREPARE_command_name, ret_name);                                         \
+        }                                                                                                   \
     } while (0)

ggml/src/ggml-virtgpu/virtgpu-forward.gen.h

@@ -20,6 +20,7 @@ apir_buffer_context_t          apir_device_buffer_from_ptr(struct virtgpu * gpu,
 char *                apir_buffer_type_get_name(struct virtgpu * gpu, apir_buffer_type_host_handle_t host_handle);
 size_t                apir_buffer_type_get_alignment(struct virtgpu * gpu, apir_buffer_type_host_handle_t host_handle);
 size_t                apir_buffer_type_get_max_size(struct virtgpu * gpu, apir_buffer_type_host_handle_t host_handle);
+/* apir_buffer_type_is_host is deprecated. */
 apir_buffer_context_t apir_buffer_type_alloc_buffer(struct virtgpu *               gpu,
                                                     apir_buffer_type_host_handle_t host_handle,
                                                     size_t                         size);

ggml/src/ggml-virtgpu/virtgpu.cpp

@@ -53,9 +53,9 @@ static int virtgpu_handshake(virtgpu * gpu) {
 
     if (!decoder) {
         GGML_ABORT(GGML_VIRTGPU
-            "%s: failed to initiate the communication with the virglrenderer library. "
-            "Most likely, the wrong virglrenderer library was loaded in the hypervisor.",
-            __func__);
+                   "%s: failed to initiate the communication with the virglrenderer library. "
+                   "Most likely, the wrong virglrenderer library was loaded in the hypervisor.",
+                   __func__);
         return 1;
     }
 
@@ -65,8 +65,7 @@ static int virtgpu_handshake(virtgpu * gpu) {
     uint32_t host_minor;
 
     if (ret_magic != APIR_HANDSHAKE_MAGIC) {
-        GGML_ABORT(GGML_VIRTGPU
-                   "%s: handshake with the virglrenderer failed (code=%d | %s)", __func__, ret_magic,
+        GGML_ABORT(GGML_VIRTGPU "%s: handshake with the virglrenderer failed (code=%d | %s)", __func__, ret_magic,
                    apir_backend_initialize_error(ret_magic));
     } else {
         apir_decode_uint32_t(decoder, &host_major);
@@ -140,15 +139,13 @@ static ApirLoadLibraryReturnCode virtgpu_load_library(virtgpu * gpu) {
                        "Make sure virglrenderer is correctly configured by the hypervisor. (%s) ",
                        __func__, apir_load_library_error(ret));
         } else {
-            GGML_ABORT(GGML_VIRTGPU
-                       "%s: virglrenderer could not load the API Remoting backend library. (%s - code %d)", __func__,
-                       apir_load_library_error(ret), ret);
+            GGML_ABORT(GGML_VIRTGPU "%s: virglrenderer could not load the API Remoting backend library. (%s - code %d)",
+                       __func__, apir_load_library_error(ret), ret);
         }
         return ret;
     }
 
-    GGML_LOG_INFO(GGML_VIRTGPU
-                  "%s: virglrenderer successfully loaded the API Remoting backend library.\n", __func__);
+    GGML_LOG_INFO(GGML_VIRTGPU "%s: virglrenderer successfully loaded the API Remoting backend library.\n", __func__);
 
     ApirLoadLibraryReturnCode apir_ret = (ApirLoadLibraryReturnCode) (ret - APIR_LOAD_LIBRARY_INIT_BASE_INDEX);
 
@@ -158,10 +155,11 @@ static ApirLoadLibraryReturnCode virtgpu_load_library(virtgpu * gpu) {
                    "Make sure virglrenderer is correctly configured by the hypervisor. (%s)",
                    __func__, apir_load_library_error(apir_ret));
     } else if (apir_ret == APIR_LOAD_LIBRARY_SYMBOL_MISSING) {
-        GGML_ABORT(GGML_VIRTGPU
-                   "%s: the API Remoting backend library couldn't load the GGML backend library, some symbols are missing. "
-                   "Make sure virglrenderer is correctly configured by the hypervisor. (%s)",
-                   __func__, apir_load_library_error(apir_ret));
+        GGML_ABORT(
+            GGML_VIRTGPU
+            "%s: the API Remoting backend library couldn't load the GGML backend library, some symbols are missing. "
+            "Make sure virglrenderer is correctly configured by the hypervisor. (%s)",
+            __func__, apir_load_library_error(apir_ret));
     } else if (apir_ret < APIR_LOAD_LIBRARY_INIT_BASE_INDEX) {
         GGML_ABORT(GGML_VIRTGPU
                    "%s: the API Remoting backend library couldn't load the GGML backend library: apir code=%d | %s)",
@@ -169,8 +167,8 @@ static ApirLoadLibraryReturnCode virtgpu_load_library(virtgpu * gpu) {
     } else {
         uint32_t lib_ret = apir_ret - APIR_LOAD_LIBRARY_INIT_BASE_INDEX;
         GGML_ABORT(GGML_VIRTGPU
-                   "%s: the API Remoting backend library initialize its backend library: apir code=%d)", __func__,
-                   lib_ret);
+                   "%s: the API Remoting backend library failed to initialize its backend library: apir code=%d)",
+                   __func__, lib_ret);
     }
     return ret;
 }
@@ -184,55 +182,49 @@ virtgpu * create_virtgpu() {
     // Initialize mutex to protect shared data_shmem buffer
     if (mtx_init(&gpu->data_shmem_mutex, mtx_plain) != thrd_success) {
         delete gpu;
-        GGML_ABORT(GGML_VIRTGPU
-                   "%s: failed to initialize data_shmem mutex", __func__);
+        GGML_ABORT(GGML_VIRTGPU "%s: failed to initialize data_shmem mutex", __func__);
         return NULL;
     }
 
     if (virtgpu_open(gpu) != APIR_SUCCESS) {
-        GGML_LOG_ERROR(GGML_VIRTGPU
-                       "%s: failed to open the virtgpu device\n", __func__);
+        GGML_LOG_ERROR(GGML_VIRTGPU "%s: failed to open the virtgpu device\n", __func__);
         return NULL;
     }
 
     if (virtgpu_init_capset(gpu) != APIR_SUCCESS) {
         if (gpu->use_apir_capset) {
             GGML_ABORT(GGML_VIRTGPU
-                       "%s: failed to initialize the virtgpu APIR capset. Make sure that the virglrenderer library supports it.", __func__);
+                       "%s: failed to initialize the virtgpu APIR capset. Make sure that the virglrenderer library "
+                       "supports it.",
+                       __func__);
         } else {
-            GGML_ABORT(GGML_VIRTGPU
-                       "%s: failed to initialize the virtgpu Venus capset", __func__);
+            GGML_ABORT(GGML_VIRTGPU "%s: failed to initialize the virtgpu Venus capset", __func__);
         }
         return NULL;
     }
 
     if (virtgpu_init_context(gpu) != APIR_SUCCESS) {
-        GGML_ABORT(GGML_VIRTGPU
-                   "%s: failed to initialize the GPU context", __func__);
+        GGML_ABORT(GGML_VIRTGPU "%s: failed to initialize the GPU context", __func__);
         return NULL;
     }
 
     if (virtgpu_shmem_create(gpu, SHMEM_REPLY_SIZE, &gpu->reply_shmem)) {
-        GGML_ABORT(GGML_VIRTGPU
-                   "%s: failed to create the shared reply memory pages", __func__);
+        GGML_ABORT(GGML_VIRTGPU "%s: failed to create the shared reply memory pages", __func__);
         return NULL;
     }
 
     if (virtgpu_shmem_create(gpu, SHMEM_DATA_SIZE, &gpu->data_shmem)) {
-        GGML_ABORT(GGML_VIRTGPU
-                   "%s: failed to create the shared data memory pages", __func__);
+        GGML_ABORT(GGML_VIRTGPU "%s: failed to create the shared data memory pages", __func__);
         return NULL;
     }
 
     if (virtgpu_handshake(gpu)) {
-        GGML_ABORT(GGML_VIRTGPU
-                   "%s: failed to handshake with the virglrenderer library", __func__);
+        GGML_ABORT(GGML_VIRTGPU "%s: failed to handshake with the virglrenderer library", __func__);
         return NULL;
     }
 
     if (virtgpu_load_library(gpu) != APIR_LOAD_LIBRARY_SUCCESS) {
-        GGML_ABORT(GGML_VIRTGPU
-                   "%s: failed to load the backend library", __func__);
+        GGML_ABORT(GGML_VIRTGPU "%s: failed to load the backend library", __func__);
         return NULL;
     }
 
@@ -243,8 +235,7 @@ static virt_gpu_result_t virtgpu_open(virtgpu * gpu) {
     drmDevicePtr devs[8];
     int          count = drmGetDevices2(0, devs, ARRAY_SIZE(devs));
     if (count < 0) {
-        GGML_LOG_ERROR(GGML_VIRTGPU
-                       "%s: failed to enumerate DRM devices\n", __func__);
+        GGML_LOG_ERROR(GGML_VIRTGPU "%s: failed to enumerate DRM devices\n", __func__);
         return APIR_ERROR_INITIALIZATION_FAILED;
     }
 
@@ -266,19 +257,17 @@ static virt_gpu_result_t virtgpu_open_device(virtgpu * gpu, const drmDevicePtr d
 
     int fd = open(node_path, O_RDWR | O_CLOEXEC);
     if (fd < 0) {
-        GGML_ABORT(GGML_VIRTGPU
-                   "%s: failed to open %s", __func__, node_path);
+        GGML_ABORT(GGML_VIRTGPU "%s: failed to open %s", __func__, node_path);
         return APIR_ERROR_INITIALIZATION_FAILED;
     }
 
     drmVersionPtr version = drmGetVersion(fd);
     if (!version || strcmp(version->name, "virtio_gpu") || version->version_major != 0) {
         if (version) {
-            GGML_LOG_ERROR(GGML_VIRTGPU
-                           "%s: unknown DRM driver %s version %d\n", __func__, version->name, version->version_major);
+            GGML_LOG_ERROR(GGML_VIRTGPU "%s: unknown DRM driver %s version %d\n", __func__, version->name,
+                           version->version_major);
         } else {
-            GGML_LOG_ERROR(GGML_VIRTGPU
-                           "%s: failed to get DRM driver version\n", __func__);
+            GGML_LOG_ERROR(GGML_VIRTGPU "%s: failed to get DRM driver version\n", __func__);
         }
 
         if (version) {
@@ -322,9 +311,8 @@ static virt_gpu_result_t virtgpu_init_capset(virtgpu * gpu) {
         virtgpu_ioctl_get_caps(gpu, gpu->capset.id, gpu->capset.version, &gpu->capset.data, sizeof(gpu->capset.data));
 
     if (ret) {
-        GGML_LOG_ERROR(GGML_VIRTGPU
-                       "%s: failed to get APIR v%d capset: %s\n",
-                       __func__, gpu->capset.version, strerror(errno));
+        GGML_LOG_ERROR(GGML_VIRTGPU "%s: failed to get APIR v%d capset: %s\n", __func__, gpu->capset.version,
+                       strerror(errno));
         return APIR_ERROR_INITIALIZATION_FAILED;
     }
 
@@ -547,13 +535,10 @@ static void log_call_duration(long long call_duration_ns, const char * name) {
     double call_duration_s  = (double) call_duration_ns / 1e9;  // 1 second = 1e9 nanoseconds
 
     if (call_duration_s > 1) {
-        GGML_LOG_INFO(GGML_VIRTGPU
-                      "waited %.2fs for the %s host reply...\n", call_duration_s, name);
+        GGML_LOG_INFO(GGML_VIRTGPU "waited %.2fs for the %s host reply...\n", call_duration_s, name);
     } else if (call_duration_ms > 1) {
-        GGML_LOG_INFO(GGML_VIRTGPU
-                      "waited %.2fms for the %s host reply...\n", call_duration_ms, name);
+        GGML_LOG_INFO(GGML_VIRTGPU "waited %.2fms for the %s host reply...\n", call_duration_ms, name);
     } else {
-        GGML_LOG_INFO(GGML_VIRTGPU
-                      "waited %lldns for the %s host reply...\n", call_duration_ns, name);
+        GGML_LOG_INFO(GGML_VIRTGPU "waited %lldns for the %s host reply...\n", call_duration_ns, name);
     }
 }

ggml/src/ggml-virtgpu/virtgpu.h

@@ -1,5 +1,6 @@
 #pragma once
 
+// clang-format off
 #include "virtgpu-utils.h"
 #include "virtgpu-shm.h"
 #include "virtgpu-apir.h"
@@ -23,20 +24,21 @@
 #include "apir_hw.h"
 #include <drm/virtgpu_drm.h>
 #include "venus_hw.h"
+// clang-format on
 
 #ifndef VIRTGPU_DRM_CAPSET_APIR
 // Will be defined include/drm/virtgpu_drm.h when
 // https://gitlab.freedesktop.org/virgl/virglrenderer/-/merge_requests/1590/diffs
 // is merged
-#define VIRTGPU_DRM_CAPSET_APIR 10
+#    define VIRTGPU_DRM_CAPSET_APIR 10
 #endif
 
 // Mesa/Virlgrenderer Venus internal. Only necessary during the
 // Venus->APIR transition in Virglrenderer
 #define VENUS_COMMAND_TYPE_LENGTH 331
 
-#ifndef VIRTGPU_DRM_CAPSET_VENUS // only available with Linux >= v6.16
-#define VIRTGPU_DRM_CAPSET_VENUS 4
+#ifndef VIRTGPU_DRM_CAPSET_VENUS  // only available with Linux >= v6.16
+#    define VIRTGPU_DRM_CAPSET_VENUS 4
 #endif
 
 typedef uint32_t virgl_renderer_capset;

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

@@ -3,9 +3,13 @@
 #extension GL_EXT_control_flow_attributes : enable
 #extension GL_EXT_shader_16bit_storage : require
 
-#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
 #extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
 
+#ifdef FLOAT16
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#extension GL_EXT_shader_subgroup_extended_types_float16 : require
+#endif
+
 #extension GL_KHR_shader_subgroup_shuffle : enable
 #extension GL_KHR_shader_subgroup_vote : enable
 
@@ -15,8 +19,10 @@
 const uint32_t HSK_per_thread = HSK / D_split;
 const uint32_t HSV_per_thread = HSV / D_split;
 
-const uint32_t cols_per_iter = WorkGroupSize / D_split;
+const uint32_t rows_per_thread = Br / row_split;
+const uint32_t cols_per_iter = WorkGroupSize / D_split / row_split;
 const uint32_t cols_per_thread = Bc / cols_per_iter;
+const uint32_t num_subgroups = SubGroupSize == 0 ? 0 : WorkGroupSize / SubGroupSize;
 
 
 layout (binding = 0) readonly buffer Q {float data_q[];};
@@ -27,20 +33,22 @@ layout (binding = 2) readonly buffer V {float16_t data_v[];};
 layout (binding = 2) readonly buffer VV4 {f16vec4 data_vv4[];};
 layout (binding = 3) readonly buffer M {float16_t data_m[];};
 
-// Store the output when doing grouped query attention.
-// Rows index by Q's dimension 2, and the first N rows are valid.
-D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
-{
-    uint32_t offset = (iq2 + r) * HSV + c;
-    data_o[o_offset + offset] = D_TYPE(elem);
-    return elem;
-}
+// If SubGroupSize is set to 0 then only use shmem reductions
+const uint32_t tmpsh_size = (SubGroupSize > 0) ? (row_split == 1 ? num_subgroups * D_split : num_subgroups) : WorkGroupSize;
+shared float tmpsh[tmpsh_size];
+shared FLOAT_TYPEV4 tmpshv4[tmpsh_size];
 
-shared FLOAT_TYPE tmpsh[WorkGroupSize];
-shared vec4 tmpshv4[WorkGroupSize];
+const uint32_t masksh_stride = Br + 1;
+shared FLOAT_TYPE masksh[Bc * masksh_stride];
 
-shared float masksh[Bc][Br];
-shared vec4 Qf[Br][HSK / 4];
+const uint32_t qf_stride = HSK / 4 + 1;
+shared FLOAT_TYPEV4 Qf[Br * qf_stride];
+
+const uint32_t D = HSK > HSV ? HSK : HSV;
+const uint32_t kvsh_stride = D / 4 + 1;
+shared FLOAT_TYPEV4 kvsh[SHMEM_STAGING != 0 ? Bc * kvsh_stride : 1];
+
+shared vec4 occupancy_limiter[LIMIT_OCCUPANCY_SHMEM > 0 ? LIMIT_OCCUPANCY_SHMEM : 1];
 
 void main() {
 #ifdef NEEDS_INIT_IQ_SHMEM
@@ -50,8 +58,24 @@ void main() {
     init_indices();
 
     const uint32_t tid = gl_LocalInvocationIndex;
+    const uint32_t threads_per_rowgroup = gl_WorkGroupSize.x / row_split;
+    const uint32_t row_tid = gl_LocalInvocationIndex / threads_per_rowgroup;
+    const uint32_t rowgroup_tid = gl_LocalInvocationIndex % threads_per_rowgroup;
     const uint32_t d_tid = gl_LocalInvocationIndex % D_split;
-    const uint32_t col_tid = gl_LocalInvocationIndex / D_split;
+    const uint32_t col_tid = (gl_LocalInvocationIndex % threads_per_rowgroup) / D_split;
+
+    if (LIMIT_OCCUPANCY_SHMEM > 0) {
+        // This just exists to avoid the occupancy_limiter array getting optimized out
+        occupancy_limiter[tid] = vec4(tid);
+
+        barrier();
+
+        if (occupancy_limiter[tid] == vec4(99999.0)) {
+            data_ov4[0] = D_TYPEV4(occupancy_limiter[tid]);
+        }
+    }
+
+#define tile_row(r) (row_tid * rows_per_thread + (r))
 
     uint32_t q_offset = gqa_iq1*p.nb01 + (iq2*p.nb02 + iq3*p.nb03) / 4;
 
@@ -60,37 +84,37 @@ void main() {
         uint32_t r = (idx + tid) / (HSK / 4);
         if (r < Br && d < HSK / 4 &&
             i * Br + r < N) {
-            Qf[r][d] = vec4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d]) * p.scale;
+            Qf[r * qf_stride + d] = FLOAT_TYPEV4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d] * p.scale);
         }
     }
     barrier();
 
-    vec4 Of[Br][HSV_per_thread / 4];
+    FLOAT_TYPEV4 Of[rows_per_thread][HSV_per_thread / 4];
     [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
-        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-            Of[r][d] = vec4(0.0);
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            Of[r][d] = FLOAT_TYPEV4(0.0);
         }
     }
 
-    float Lf[Br], Mf[Br];
+    float Lf[rows_per_thread], Mf[rows_per_thread];
 
     // Use -FLT_MAX/2 rather than -inf to reduce the possibility of NaNs, e.g. when computing Mold-M.
     const float NEG_FLT_MAX_OVER_2 = uintBitsToFloat(0xFEFFFFFF);
 
-    [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
         Lf[r] = 0;
         Mf[r] = NEG_FLT_MAX_OVER_2;
     }
 
-    float slope[Br];
-    [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-        slope[r] = 1.0;
+    ACC_TYPE slope[rows_per_thread];
+    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+        slope[r] = ACC_TYPE(1.0);
     }
 
     // ALiBi
     if (p.max_bias > 0.0f) {
-        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-            slope[r] = perElemOpComputeSlope(r, col_tid, ACC_TYPE(0), iq2);
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            slope[r] = perElemOpComputeSlope(tile_row(r), col_tid, ACC_TYPE(0), iq2);
         }
     }
 
@@ -113,75 +137,141 @@ void main() {
 
     uint32_t mask_opt = 0;
     uint32_t mask_opt_idx = ~0;
+    uint32_t mask_opt_bits = 0;
 
     [[dont_unroll]]
     for (uint32_t j = start_j; j < end_j; ++j) {
+        if (MASK_ENABLE) {
+            if (USE_MASK_OPT && mask_opt_idx != j / 16) {
+                mask_opt_idx = j / 16;
+                mask_opt = data_mask_opt[mo_offset + mask_opt_idx];
+            }
+            mask_opt_bits = (mask_opt >> ((j % 16) * 2)) & 0x3;
+            if (mask_opt_bits == MASK_OPT_ALL_NEG_INF) {
+                // skip this block
+                continue;
+            }
+            // Only load if the block is not all zeros
+            if (mask_opt_bits != MASK_OPT_ALL_ZERO) {
+                bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;
 
-        if (USE_MASK_OPT && mask_opt_idx != j / 16) {
-            mask_opt_idx = j / 16;
-            mask_opt = data_mask_opt[mo_offset + mask_opt_idx];
+                float max_mask = NEG_FLT_MAX_OVER_2;
+                barrier();
+                [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
+                    uint32_t c = (idx + tid) % Bc;
+                    uint32_t r = (idx + tid) / Bc;
+                    if (idx + tid < Bc * Br) {
+                        if ((!KV_bounds_check || j * Bc + c < KV) && (!nem1_bounds_check || i * Br + r < p.nem1)) {
+                            FLOAT_TYPE m = FLOAT_TYPE(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]);
+                            masksh[c * masksh_stride + r] = m;
+                            max_mask = max(max_mask, float(m));
+                        } else {
+                            masksh[c * masksh_stride + r] = FLOAT_TYPE(0);
+                        }
+                    }
+                }
+                // skip the block if the mask is entirely -inf
+                bool all_less = subgroupAll(max_mask <= NEG_FLT_MAX_OVER_2);
+                barrier();
+                if (gl_SubgroupInvocationID == 0) {
+                    tmpsh[gl_SubgroupID] = all_less ? NEG_FLT_MAX_OVER_2 : 0.0f;
+                }
+                barrier();
+                [[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
+                    max_mask = max(max_mask, tmpsh[s]);
+                }
+                if (max_mask <= NEG_FLT_MAX_OVER_2) {
+                    continue;
+                }
+            }
         }
-        uint32_t mask_opt_bits = (mask_opt >> ((j % 16) * 2)) & 0x3;
-        if (mask_opt_bits == MASK_OPT_ALL_NEG_INF) {
-            // skip this block
-            continue;
-        }
-        // Only load if the block is not all zeros
-        if (MASK_ENABLE && mask_opt_bits != MASK_OPT_ALL_ZERO) {
-            bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;
 
-            float max_mask = NEG_FLT_MAX_OVER_2;
-            [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
-                uint32_t c = (idx + tid) % Bc;
-                uint32_t r = (idx + tid) / Bc;
-                if (idx + tid < Bc * Br) {
-                    if ((!KV_bounds_check || j * Bc + c < KV) && (!nem1_bounds_check || i * Br + r < p.nem1)) {
-                        float m = float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]);
-                        masksh[c][r] = m;
-                        max_mask = max(max_mask, m);
+        ACC_TYPE Sf[rows_per_thread][cols_per_thread];
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
+                Sf[r][c] = ACC_TYPE(0.0);
+            }
+        }
+
+        if (SHMEM_STAGING != 0) {
+            barrier();
+            [[unroll]] for (uint32_t idx = 0; idx < Bc * HSK / 4; idx += gl_WorkGroupSize.x) {
+                uint32_t d = (idx + tid) % (HSK / 4);
+                uint32_t c = (idx + tid) / (HSK / 4);
+                if (idx + gl_WorkGroupSize.x <= Bc * HSK / 4 || c < Bc) {
+                    FLOAT_TYPEV4 K_Tf = FLOAT_TYPEV4(0);
+                    if (!KV_bounds_check || j * Bc + c < KV) {
+#if BLOCK_SIZE > 1
+                        uint coord = (j * Bc + c) * k_stride * BLOCK_SIZE + 4 * d;
+                        uint ib = coord / BLOCK_SIZE;
+                        uint iqs = (coord % BLOCK_SIZE);
+                        K_Tf = dequantize4(ib, iqs, k_offset, BINDING_IDX_K);
+#else
+                        K_Tf = FLOAT_TYPEV4(data_kv4[k_offset / 4 + (j * Bc + c) * k_stride / 4 + d]);
+#endif
+                    }
+
+                    kvsh[c * kvsh_stride + d] = K_Tf;
+                }
+            }
+            barrier();
+        }
+
+        // More d iterations means Q register caching becomes relevant
+        // Few iterations means the additional registers needed are worse than the speed-up from caching
+        if (HSK_per_thread / 4 > 4) {
+            [[unroll]] for (uint32_t d = 0; d < HSK_per_thread / 4; ++d) {
+                FLOAT_TYPEV4 Q_cache[rows_per_thread];
+                [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                    Q_cache[r] = Qf[tile_row(r) * qf_stride + d * D_split + d_tid];
+                }
+
+                [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
+                    if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
+                        continue;
+                    }
+
+                    FLOAT_TYPEV4 K_Tf;
+                    if (SHMEM_STAGING != 0) {
+                        K_Tf = kvsh[(c * cols_per_iter + col_tid) * kvsh_stride + (d * D_split + d_tid)];
                     } else {
-                        masksh[c][r] = float(0);
+#if BLOCK_SIZE > 1
+                        uint coord = (j * Bc + c * cols_per_iter + col_tid) * k_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
+                        uint ib = coord / BLOCK_SIZE;
+                        uint iqs = (coord % BLOCK_SIZE);
+                        K_Tf = dequantize4(ib, iqs, k_offset, BINDING_IDX_K);
+#else
+                        K_Tf = FLOAT_TYPEV4(data_kv4[k_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * k_stride / 4 + d * D_split + d_tid]);
+#endif
+                    }
+                    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                        Sf[r][c] += ACC_TYPE(dot(Q_cache[r], K_Tf));
                     }
                 }
             }
-            // skip the block if the mask is entirely -inf
-            bool all_less = subgroupAll(max_mask <= NEG_FLT_MAX_OVER_2);
-            barrier();
-            if (gl_SubgroupInvocationID == 0) {
-                tmpsh[gl_SubgroupID] = all_less ? NEG_FLT_MAX_OVER_2 : 0.0f;
-            }
-            barrier();
-            [[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
-                max_mask = max(max_mask, tmpsh[s]);
-            }
-            if (max_mask <= NEG_FLT_MAX_OVER_2) {
-                continue;
-            }
-        }
-
-        float Sf[Br][cols_per_thread];
-        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+        } else {
             [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-                Sf[r][c] = 0.0;
-            }
-        }
+                if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
+                    continue;
+                }
 
-
-        [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-            if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
-                continue;
-            }
-            [[unroll]] for (uint32_t d = 0; d < HSK_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSK_per_thread / 4; ++d) {
+                    FLOAT_TYPEV4 K_Tf;
+                    if (SHMEM_STAGING != 0) {
+                        K_Tf = kvsh[(c * cols_per_iter + col_tid) * kvsh_stride + (d * D_split + d_tid)];
+                    } else {
 #if BLOCK_SIZE > 1
-                uint coord = (j * Bc + c * cols_per_iter + col_tid) * k_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
-                uint ib = coord / BLOCK_SIZE;
-                uint iqs = (coord % BLOCK_SIZE);
-                vec4 K_Tf = dequantize4(ib, iqs, k_offset, BINDING_IDX_K);
+                        uint coord = (j * Bc + c * cols_per_iter + col_tid) * k_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
+                        uint ib = coord / BLOCK_SIZE;
+                        uint iqs = (coord % BLOCK_SIZE);
+                        K_Tf = dequantize4(ib, iqs, k_offset, BINDING_IDX_K);
 #else
-                vec4 K_Tf = vec4(data_kv4[k_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * k_stride / 4 + d * D_split + d_tid]);
+                        K_Tf = FLOAT_TYPEV4(data_kv4[k_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * k_stride / 4 + d * D_split + d_tid]);
 #endif
-                [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-                    Sf[r][c] += dot(Qf[r][d * D_split + d_tid], K_Tf);
+                    }
+                    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                        Sf[r][c] += ACC_TYPE(dot(Qf[tile_row(r) * qf_stride + d * D_split + d_tid], K_Tf));
+                    }
                 }
             }
         }
@@ -189,89 +279,109 @@ void main() {
         [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
             // Compute sum across the D_split
             [[unroll]] for (uint s = D_split / 2; s > 0; s >>= 1) {
-                [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+                [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
                     Sf[r][c] += subgroupShuffleXor(Sf[r][c], s);
                 }
             }
         }
 
         if (LOGIT_SOFTCAP) {
-            [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
                 [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-                    Sf[r][c] = p.logit_softcap * tanh(Sf[r][c]);
+                    Sf[r][c] = ACC_TYPE(p.logit_softcap * tanh(Sf[r][c]));
                 }
             }
         }
 
         if (MASK_ENABLE && mask_opt_bits != MASK_OPT_ALL_ZERO) {
             [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-                [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-                    float mvf = masksh[c * cols_per_iter + col_tid][r];
+                [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                    FLOAT_TYPE mvf = masksh[(c * cols_per_iter + col_tid) * masksh_stride + tile_row(r)];
 
                     Sf[r][c] += slope[r]*mvf;
                 }
             }
-            barrier();
         }
 
-        float rowmaxf[Br], Pf[Br][cols_per_thread], rowsumf[Br], eMf[Br], Moldf[Br];
-        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-            rowmaxf[r] = NEG_FLT_MAX_OVER_2;
+        float eMf[rows_per_thread];
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            float rowmaxf = NEG_FLT_MAX_OVER_2;
             [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
                 if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
                     continue;
                 }
-                rowmaxf[r] = max(rowmaxf[r], Sf[r][c]);
+                rowmaxf = max(rowmaxf, float(Sf[r][c]));
             }
-            Moldf[r] = Mf[r];
+            float Moldf = Mf[r];
 
             // M = max(rowmax, Mold)
             // P = e^(S - M)
             // eM = e^(Mold - M)
-            Mf[r] = max(rowmaxf[r], Moldf[r]);
-            [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-                Pf[r][c] = exp(Sf[r][c] - Mf[r]);
-            }
-            eMf[r] = exp(Moldf[r] - Mf[r]);
-
-            // Compute sum across row of P
-            rowsumf[r] = 0.0;
-            [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-                if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
-                    continue;
-                }
-                rowsumf[r] += Pf[r][c];
-            }
-
-            Lf[r] = eMf[r]*Lf[r] + rowsumf[r];
+            Mf[r] = max(rowmaxf, Moldf);
+            eMf[r] = exp(Moldf - Mf[r]);
+            Lf[r] = eMf[r]*Lf[r];
         }
 
         [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
-            [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-                Of[r][d] = eMf[r] * Of[r][d];
+            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                Of[r][d] = FLOAT_TYPE(eMf[r]) * Of[r][d];
             }
         }
 
+        if (SHMEM_STAGING != 0) {
+            barrier();
+            [[unroll]] for (uint32_t idx = 0; idx < Bc * HSV / 4; idx += gl_WorkGroupSize.x) {
+                uint32_t d = (idx + tid) % (HSV / 4);
+                uint32_t c = (idx + tid) / (HSV / 4);
+                if (idx + gl_WorkGroupSize.x <= Bc * HSV / 4 || c < Bc) {
+                    FLOAT_TYPEV4 V_Tf = FLOAT_TYPEV4(0);
+                    if (!KV_bounds_check || j * Bc + c < KV) {
+#if BLOCK_SIZE > 1
+                        uint coord = (j * Bc + c) * v_stride * BLOCK_SIZE + 4 * d;
+                        uint ib = coord / BLOCK_SIZE;
+                        uint iqs = (coord % BLOCK_SIZE);
+                        V_Tf = dequantize4(ib, iqs, v_offset, BINDING_IDX_V);
+#else
+                        V_Tf = FLOAT_TYPEV4(data_vv4[v_offset / 4 + (j * Bc + c) * v_stride / 4 + d]);
+#endif
+                    }
+
+                    kvsh[c * kvsh_stride + d] = V_Tf;
+                }
+            }
+            barrier();
+        }
+
         [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
             if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
                 continue;
             }
+
+            FLOAT_TYPE Pf[rows_per_thread];
+            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                Pf[r] = FLOAT_TYPE(exp(float(Sf[r][c]) - Mf[r]));
+                Lf[r] += Pf[r];
+            }
+
             [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+                FLOAT_TYPEV4 Vf;
+                if (SHMEM_STAGING != 0) {
+                    Vf = kvsh[(c * cols_per_iter + col_tid) * kvsh_stride + (d * D_split + d_tid)];
+                } else {
 #if BLOCK_SIZE > 1
-                uint coord = (j * Bc + c * cols_per_iter + col_tid) * v_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
-                uint ib = coord / BLOCK_SIZE;
-                uint iqs = (coord % BLOCK_SIZE);
-                vec4 Vf = dequantize4(ib, iqs, v_offset, BINDING_IDX_V);
+                    uint coord = (j * Bc + c * cols_per_iter + col_tid) * v_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
+                    uint ib = coord / BLOCK_SIZE;
+                    uint iqs = (coord % BLOCK_SIZE);
+                    Vf = dequantize4(ib, iqs, v_offset, BINDING_IDX_V);
 #else
-                vec4 Vf = vec4(data_vv4[v_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * v_stride / 4 + d * D_split + d_tid]);
+                    Vf = FLOAT_TYPEV4(data_vv4[v_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * v_stride / 4 + d * D_split + d_tid]);
 #endif
-                [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-                    Of[r][d] += Pf[r][c] * Vf;
+                }
+                [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                    Of[r][d] += FLOAT_TYPEV4(Pf[r] * Vf);
                 }
             }
         }
-
-        barrier();
     }
 
     // prevent race on tmpsh
@@ -279,58 +389,115 @@ void main() {
 
     // reduce across threads
 
-    [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-        float rowmaxf, eMf;
+    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+        float rowmaxf = Mf[r];
 
-        tmpsh[tid] = Mf[r];
         // Compute max across the row
-        barrier();
-        [[unroll]] for (int s = int(gl_WorkGroupSize.x) / 2; s >= D_split; s >>= 1) {
-            if (tid < s) {
-                tmpsh[tid] = max(tmpsh[tid], tmpsh[tid + s]);
+        if (SubGroupSize > 0) {
+            [[unroll]] for (uint s = D_split; s < SubGroupSize; s *= 2) {
+                rowmaxf = max(rowmaxf, subgroupShuffleXor(rowmaxf, s));
             }
+            if (row_split == 1) {
+                // Reduce inside workgroup with shmem
+                barrier();
+                if (gl_SubgroupInvocationID == d_tid) {
+                    tmpsh[gl_SubgroupID * D_split + d_tid] = rowmaxf;
+                }
+                barrier();
+                rowmaxf = tmpsh[d_tid];
+                [[unroll]] for (uint32_t s = 1; s < num_subgroups; ++s) {
+                    rowmaxf = max(rowmaxf, tmpsh[s * D_split + d_tid]);
+                }
+            }
+        } else {
             barrier();
+            tmpsh[tid] = rowmaxf;
+            barrier();
+            [[unroll]] for (int s = int(threads_per_rowgroup) / 2; s >= D_split; s >>= 1) {
+                if (rowgroup_tid < s) {
+                    tmpsh[tid] = max(tmpsh[tid], tmpsh[tid ^ s]);
+                }
+                barrier();
+            }
+            rowmaxf = tmpsh[row_tid * threads_per_rowgroup + d_tid];
         }
-        rowmaxf = tmpsh[d_tid];
-        barrier();
 
         float Moldf = Mf[r];
 
         // M = max(rowmax, Mold)
         // eM = e^(Mold - M)
         Mf[r] = max(rowmaxf, Moldf);
-        eMf = exp(Moldf - Mf[r]);
+        float eMf = exp(Moldf - Mf[r]);
 
         Lf[r] = eMf*Lf[r];
 
-        tmpsh[tid] = Lf[r];
-
         // Compute sum across the row
-        barrier();
-        [[unroll]] for (int s = int(gl_WorkGroupSize.x) / 2; s >= D_split; s >>= 1) {
-            if (tid < s) {
-                tmpsh[tid] = tmpsh[tid] + tmpsh[tid + s];
+        if (SubGroupSize > 0) {
+            [[unroll]] for (uint s = D_split; s < SubGroupSize; s *= 2) {
+                Lf[r] += subgroupShuffleXor(Lf[r], s);
             }
+            if (row_split == 1) {
+                barrier();
+                if (gl_SubgroupInvocationID == d_tid) {
+                    tmpsh[gl_SubgroupID * D_split + d_tid] = Lf[r];
+                }
+                barrier();
+                Lf[r] = tmpsh[d_tid];
+                [[unroll]] for (uint32_t s = 1; s < num_subgroups; ++s) {
+                    Lf[r] += tmpsh[s * D_split + d_tid];
+                }
+            }
+        } else {
             barrier();
-        }
-        Lf[r] = tmpsh[d_tid];
-        barrier();
-
-        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
-
-            Of[r][d] = eMf * Of[r][d];
-            tmpshv4[tid] = Of[r][d];
-
+            tmpsh[tid] = Lf[r];
             barrier();
-            [[unroll]] for (int s = int(gl_WorkGroupSize.x) / 2; s >= D_split; s >>= 1) {
-                if (tid < s) {
-                    Of[r][d] += tmpshv4[tid + s];
-                    tmpshv4[tid] = Of[r][d];
+            [[unroll]] for (int s = int(threads_per_rowgroup) / 2; s >= D_split; s >>= 1) {
+                if (rowgroup_tid < s) {
+                    tmpsh[tid] = tmpsh[tid] + tmpsh[tid ^ s];
                 }
                 barrier();
             }
-            Of[r][d] = tmpshv4[d_tid];
-            barrier();
+            Lf[r] = tmpsh[row_tid * threads_per_rowgroup + d_tid];
+        }
+
+        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+            Of[r][d] = FLOAT_TYPE(eMf) * Of[r][d];
+
+            if (SubGroupSize > 0) {
+                [[unroll]] for (uint s = D_split; s < SubGroupSize; s *= 2) {
+                    if (!OLD_AMD_WINDOWS) {
+                        Of[r][d] += subgroupShuffleXor(Of[r][d], s);
+                    } else {
+                        // Something about f16vec4 subgroupShuffleXor is broken on AMD Windows RDNA2 and below.
+                        // Shuffle full vec4 as workaround.
+                        // See https://github.com/ggml-org/llama.cpp/issues/19881#issuecomment-3958643697
+                        Of[r][d] += FLOAT_TYPEV4(subgroupShuffleXor(vec4(Of[r][d]), s));
+                    }
+                }
+                if (row_split == 1) {
+                    barrier();
+                    if (gl_SubgroupInvocationID == d_tid) {
+                        tmpshv4[gl_SubgroupID * D_split + d_tid] = Of[r][d];
+                    }
+                    barrier();
+                    Of[r][d] = tmpshv4[d_tid];
+                    [[unroll]] for (uint32_t s = 1; s < num_subgroups; ++s) {
+                        Of[r][d] += tmpshv4[s * D_split + d_tid];
+                    }
+                }
+            } else {
+                barrier();
+                tmpshv4[tid] = Of[r][d];
+                barrier();
+                [[unroll]] for (int s = int(threads_per_rowgroup) / 2; s >= D_split; s >>= 1) {
+                    if (rowgroup_tid < s) {
+                        Of[r][d] += tmpshv4[tid ^ s];
+                        tmpshv4[tid] = Of[r][d];
+                    }
+                    barrier();
+                }
+                Of[r][d] = tmpshv4[row_tid * threads_per_rowgroup + d_tid];
+            }
         }
     }
 
@@ -338,33 +505,53 @@ void main() {
     // If there is split_k, then the split_k resolve shader does the final
     // division by L. Store the intermediate O value and per-row m and L values.
     if (p.k_num > 1) {
-        // note: O and Q have swapped coord 1,2.
-        uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
+        if (p.gqa_ratio > 1) {
+            // note: O and Q have swapped coord 1,2.
+            uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3)) / 4;
 
-        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-            if (r < N) {
-                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
-                    [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N);
+            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                const uint row = tile_row(r);
+                if (row < N) {
+                    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+                        gqaStore(row, d * D_split + d_tid, Of[r][d], o_offset, iq2, N);
                     }
                 }
             }
-        }
 
-        o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
-        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-            if (r < N) {
-                perElemOpStoreCol0(r, 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
-                perElemOpStoreCol0(r, 0u, ACC_TYPE(Mf[r]), o_offset + p.ne1, iq2, N);
+            o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
+            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                const uint row = tile_row(r);
+                if (row < N) {
+                    perElemOpStoreCol0(row, 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
+                    perElemOpStoreCol0(row, 0u, ACC_TYPE(Mf[r]), o_offset + p.ne1, iq2, N);
+                }
+            }
+        } else {
+            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                const uint row = tile_row(r);
+                const uint global_row = i * Br + row;
+
+                if (global_row < N) {
+                    uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (global_row + p.ne2 * iq3)) / 4;
+
+                    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+                        data_ov4[o_offset + iq2 * HSV/4 + d * D_split + d_tid] = D_TYPEV4(Of[r][d]);
+                    }
+                }
+
+                if (global_row < N && d_tid == 0 && col_tid == 0) {
+                    uint32_t lm_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (global_row + p.ne2 * iq3));
+                    data_o[lm_offset + iq2] = D_TYPE(Lf[r]);
+                    data_o[lm_offset + p.ne1 + iq2] = D_TYPE(Mf[r]);
+                }
             }
         }
-
         return;
     }
 
     if ((p.mask_n_head_log2 & SINK_ENABLE_BIT) != 0) {
-        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-            float sink = perElemOpGetSink(r, 0u, ACC_TYPE(0), iq2);
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            float sink = perElemOpGetSink(tile_row(r), 0u, ACC_TYPE(0), iq2);
 
             float ms = 1.0f;
             float vs = 1.0f;
@@ -373,7 +560,7 @@ void main() {
                 ms = exp(Mf[r] - sink);
 
                 [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
-                    Of[r][d] *= ms;
+                    Of[r][d] *= FLOAT_TYPE(ms);
                 }
             } else {
                 vs = exp(sink - Mf[r]);
@@ -383,39 +570,37 @@ void main() {
         }
     }
 
-    float Lfrcp[Br];
-    [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+    float Lfrcp[rows_per_thread];
+    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
         Lfrcp[r] = (Lf[r] == 0.0) ? 0.0 : (1.0 / Lf[r]);
     }
 
     [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
-        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-            Of[r][d] *= Lfrcp[r];
-#if defined(ACC_TYPE_MAX)
-            Of[r][d] = clamp(Of[r][d], -vec4(ACC_TYPE_MAX), vec4(ACC_TYPE_MAX));
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            Of[r][d] *= FLOAT_TYPE(Lfrcp[r]);
+#if defined(FLOAT_TYPE_MAX)
+            Of[r][d] = clamp(Of[r][d], -FLOAT_TYPE_MAX, FLOAT_TYPE_MAX);
 #endif
         }
     }
 
-    uint32_t o_offset = gqa_iq1*p.ne1*HSV + iq3*p.ne2*p.ne1*HSV;
+    uint32_t o_offset = (gqa_iq1*p.ne1*HSV + iq3*p.ne2*p.ne1*HSV) / 4;
 
     if (p.gqa_ratio > 1) {
-        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-            if (r < N) {
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            const uint row = tile_row(r);
+            if (row < N) {
                 [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
-                    [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N);
-                    }
+                    gqaStore(row, d * D_split + d_tid, Of[r][d], o_offset, iq2, N);
                 }
             }
         }
     } else {
-        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-            if (i * Br + r < N) {
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            const uint row = tile_row(r);
+            if (i * Br + row < N) {
                 [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
-                    [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        data_o[o_offset + iq2 * HSV + (i * Br + r) * p.ne1 * HSV + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
-                    }
+                    data_ov4[o_offset + (iq2 * HSV + (i * Br + row) * p.ne1 * HSV) / 4 + d * D_split + d_tid] = D_TYPEV4(Of[r][d]);
                 }
             }
         }

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl

@@ -1,20 +1,23 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
-layout (constant_id = 0) const uint32_t WorkGroupSize = 128;
-layout (constant_id = 1) const uint32_t Br = 1;
-layout (constant_id = 2) const uint32_t Bc = 32;
-layout (constant_id = 3) const uint32_t HSK = 32;
-layout (constant_id = 4) const uint32_t HSV = 32;
-layout (constant_id = 5) const uint32_t Clamp = 0;
-layout (constant_id = 6) const uint32_t D_split = 16;
-layout (constant_id = 7) const uint32_t SubGroupSize = 32;
-layout (constant_id = 8) const uint32_t K_LOAD_SHMEM = 0;
-layout (constant_id = 9) const uint32_t Flags = 0;
+layout (constant_id =  0) const uint32_t WorkGroupSize = 128;
+layout (constant_id =  1) const uint32_t Br = 1;
+layout (constant_id =  2) const uint32_t Bc = 32;
+layout (constant_id =  3) const uint32_t HSK = 32;
+layout (constant_id =  4) const uint32_t HSV = 32;
+layout (constant_id =  5) const uint32_t Clamp = 0;
+layout (constant_id =  6) const uint32_t D_split = 16;
+layout (constant_id =  7) const uint32_t row_split = 1;
+layout (constant_id =  8) const uint32_t SubGroupSize = 32;
+layout (constant_id =  9) const uint32_t SHMEM_STAGING = 0;
+layout (constant_id = 10) const uint32_t Flags = 0;
+layout (constant_id = 11) const uint32_t LIMIT_OCCUPANCY_SHMEM = 0;
 
-const bool USE_MASK_OPT  = (Flags & 1) != 0;
-const bool MASK_ENABLE   = (Flags & 2) != 0;
-const bool LOGIT_SOFTCAP = (Flags & 4) != 0;
+const bool USE_MASK_OPT    = (Flags & 1) != 0;
+const bool MASK_ENABLE     = (Flags & 2) != 0;
+const bool LOGIT_SOFTCAP   = (Flags & 4) != 0;
+const bool OLD_AMD_WINDOWS = (Flags & 8) != 0;
 
 // Round up head sizes to a multiple of 16, for coopmat1/coopmat2 paths
 const uint32_t HSK_pad = (HSK + 15) & ~15;
@@ -69,6 +72,7 @@ layout (push_constant) uniform parameter {
 layout (binding = 4) readonly buffer S {float data_s[];};
 
 layout (binding = 5) writeonly buffer O {D_TYPE data_o[];};
+layout (binding = 5) writeonly buffer OV4 {D_TYPEV4 data_ov4[];};
 
 layout (binding = 6) readonly buffer MO {uint32_t data_mask_opt[];};
 
@@ -94,12 +98,12 @@ layout (binding = 2) readonly buffer V_PACKED16 {A_TYPE_PACKED16 v_data_packed16
 #define BLOCK_SIZE 4
 #define BLOCK_BYTE_SIZE 16
 
-vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
+FLOAT_TYPEV4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
     // iqs is currently always zero in the flash attention shaders
     if (binding_idx == BINDING_IDX_K) {
-        return k_packed.k_data_packed[a_offset + ib];
+        return FLOAT_TYPEV4(k_packed.k_data_packed[a_offset + ib]);
     } else {
-        return v_packed.v_data_packed[a_offset + ib];
+        return FLOAT_TYPEV4(v_packed.v_data_packed[a_offset + ib]);
     }
 }
 #endif
@@ -107,7 +111,7 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
 #if defined(DATA_A_Q4_0)
 #define BLOCK_BYTE_SIZE 18
 
-vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
+FLOAT_TYPEV4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
     if (binding_idx == BINDING_IDX_K) {
         uint vui_lo = uint(k_packed.k_data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 0]);
         uint vui_hi = uint(k_packed.k_data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 1]);
@@ -115,7 +119,7 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
         vui_lo >>= shift;
         vui_hi >>= shift;
 
-        return float(k_packed.k_data_packed16[a_offset + ib].d) * (vec4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - 8.0f);
+        return FLOAT_TYPE(k_packed.k_data_packed16[a_offset + ib].d) * (FLOAT_TYPEV4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - FLOAT_TYPE(8.0f));
     } else {
         uint vui_lo = uint(v_packed.v_data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 0]);
         uint vui_hi = uint(v_packed.v_data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 1]);
@@ -123,24 +127,24 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
         vui_lo >>= shift;
         vui_hi >>= shift;
 
-        return float(v_packed.v_data_packed16[a_offset + ib].d) * (vec4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - 8.0f);
+        return FLOAT_TYPE(v_packed.v_data_packed16[a_offset + ib].d) * (FLOAT_TYPEV4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - FLOAT_TYPE(8.0f));
     }
 }
 #endif
 
 #if defined(DATA_A_Q8_0)
 #define BLOCK_BYTE_SIZE 34
-vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
+FLOAT_TYPEV4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
     if (binding_idx == BINDING_IDX_K) {
         const i8vec2 v0 = unpack8(int32_t(k_packed.k_data_packed16[a_offset + ib].qs[iqs / 2])).xy; // vec4 used due to #12147
         const i8vec2 v1 = unpack8(int32_t(k_packed.k_data_packed16[a_offset + ib].qs[iqs / 2 + 1])).xy;
 
-        return float(k_packed.k_data_packed16[a_offset + ib].d) * vec4(v0.x, v0.y, v1.x, v1.y);
+        return FLOAT_TYPE(k_packed.k_data_packed16[a_offset + ib].d) * FLOAT_TYPEV4(v0.x, v0.y, v1.x, v1.y);
     } else {
         const i8vec2 v0 = unpack8(int32_t(v_packed.v_data_packed16[a_offset + ib].qs[iqs / 2])).xy; // vec4 used due to #12147
         const i8vec2 v1 = unpack8(int32_t(v_packed.v_data_packed16[a_offset + ib].qs[iqs / 2 + 1])).xy;
 
-        return float(v_packed.v_data_packed16[a_offset + ib].d) * vec4(v0.x, v0.y, v1.x, v1.y);
+        return FLOAT_TYPE(v_packed.v_data_packed16[a_offset + ib].d) * FLOAT_TYPEV4(v0.x, v0.y, v1.x, v1.y);
     }
 }
 #endif
@@ -189,10 +193,16 @@ void init_indices()
     KV = p.KV;
 
     if (p.k_num > 1) {
-        i = 0;
-        // batch and split_k share gl_WorkGroupID.x
-        gqa_iq1 = gl_WorkGroupID.x / p.k_num;
-        split_k_index = gl_WorkGroupID.x % p.k_num;
+        if (p.gqa_ratio > 1) {
+            i = 0;
+            // batch and split_k share gl_WorkGroupID.x
+            gqa_iq1 = gl_WorkGroupID.x / p.k_num;
+            split_k_index = gl_WorkGroupID.x % p.k_num;
+        } else {
+            gqa_iq1 = 0;
+            split_k_index = gl_WorkGroupID.x % p.k_num;
+            i = gl_WorkGroupID.x / p.k_num;
+        }
     } else if (p.gqa_ratio > 1) {
         i = 0;
         gqa_iq1 = gl_WorkGroupID.x;
@@ -244,3 +254,11 @@ void init_indices()
 // Bias applied to softmax to stay in fp16 range.
 // Based on ggml-cuda issue https://github.com/ggml-org/llama.cpp/issues/18606
 const float FATTN_KQ_MAX_OFFSET = 3.0f*0.6931f;
+
+// Store the output when doing grouped query attention.
+// Rows index by Q's dimension 2, and the first N rows are valid.
+void gqaStore(const in uint32_t r, const in uint32_t c, const in FLOAT_TYPEV4 elems, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
+{
+    uint32_t offset = (iq2 + r) * HSV / 4 + c;
+    data_ov4[o_offset + offset] = D_TYPEV4(elems);
+}

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp

@@ -19,7 +19,6 @@
 const uint32_t MatBr = 16;
 const uint32_t MatBc = 16;
 
-const uint32_t row_split = Bc / MatBc;
 const uint32_t rows_per_thread = Br / row_split;
 const uint32_t cols_per_iter = gl_WorkGroupSize.x / row_split;
 const uint32_t cols_per_thread = Bc / cols_per_iter;
@@ -33,15 +32,6 @@ layout (binding = 2) readonly buffer V {float16_t data_v[];};
 layout (binding = 2) readonly buffer VV4 {f16vec4 data_vv4[];};
 layout (binding = 3) readonly buffer M {float16_t data_m[];};
 
-// Store the output when doing grouped query attention.
-// Rows index by Q's dimension 2, and the first N rows are valid.
-D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
-{
-    uint32_t offset = (iq2 + r) * HSV + c;
-    data_o[o_offset + offset] = D_TYPE(elem);
-    return elem;
-}
-
 shared float tmpsh[row_split];
 
 const uint32_t qstride = HSK_pad / 4 + 2; // in units of f16vec4
@@ -54,10 +44,14 @@ shared f16vec4 Psh[Bc * psh_stride];
 const uint32_t sfshstride = (HSK <= 128) ? (Br / 4 + 2) : Br / 4;
 shared ACC_TYPEV4 sfsh[Bc * sfshstride];
 
-const uint32_t kshstride = (K_LOAD_SHMEM != 0 ? HSK_pad : MatBr) / 4 + 2; // in units of f16vec4
+const uint32_t D_pad = HSK_pad > HSV_pad ? HSK_pad : HSV_pad;
+const uint32_t kvsh_stride = (SHMEM_STAGING != 0 ? D_pad : MatBr) / 4 + 2; // in units of f16vec4
 const uint v_cols = MatBc / 4 * row_split; // total cols, 4 vec4s per MatBc * number of subgroups
 const uint vsh_stride = v_cols;
-shared f16vec4 ksh[(kshstride >= vsh_stride) ? (Bc * kshstride) : (Bc * vsh_stride)];
+shared f16vec4 kvsh[(kvsh_stride >= vsh_stride) ? (Bc * kvsh_stride) : (Bc * vsh_stride)];
+
+const uint32_t osh_stride = row_split * MatBr / 4;
+shared f16vec4 pvsh[MatBc * osh_stride];
 
 shared ACC_TYPE slope[Br];
 
@@ -84,11 +78,6 @@ void main() {
                 Qf[i + tid] = f16vec4(0);
             }
         }
-        [[unroll]] for (uint i = 0; i < Bc * kshstride; i += gl_WorkGroupSize.x) {
-            if (i + tid < Bc * kshstride) {
-                ksh[i + tid] = f16vec4(0);
-            }
-        }
         barrier();
     }
 
@@ -104,10 +93,10 @@ void main() {
     }
     barrier();
 
-    ACC_TYPEV4 Of[rows_per_thread][d_per_thread];
+    f16vec4 Of[rows_per_thread][d_per_thread];
     [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
         [[unroll]] for (uint32_t d = 0; d < d_per_thread; ++d) {
-            Of[r][d] = ACC_TYPEV4(0.0);
+            Of[r][d] = f16vec4(0.0);
         }
     }
 
@@ -153,22 +142,22 @@ void main() {
 
     uint32_t mask_opt = 0;
     uint32_t mask_opt_idx = ~0;
+    uint32_t mask_opt_bits = 0;
+    f16vec4 mask_cache[Bc * Br / 4 / WorkGroupSize];
 
     [[dont_unroll]]
     for (uint32_t j = start_j; j < end_j; ++j) {
 
-        f16vec4 mask_cache[Bc * Br / 4 / WorkGroupSize];
         [[unroll]] for (uint32_t idx = 0; idx < mask_cache.length(); ++idx) {
             mask_cache[idx] = f16vec4(0);
         }
 
         if (MASK_ENABLE) {
-
             if (USE_MASK_OPT && mask_opt_idx != j / 16) {
                 mask_opt_idx = j / 16;
                 mask_opt = data_mask_opt[mo_offset + mask_opt_idx];
             }
-            uint32_t mask_opt_bits = (mask_opt >> ((j % 16) * 2)) & 0x3;
+            mask_opt_bits = (mask_opt >> ((j % 16) * 2)) & 0x3;
             if (mask_opt_bits == MASK_OPT_ALL_NEG_INF) {
                 // skip this block
                 continue;
@@ -231,24 +220,24 @@ void main() {
             }
         }
 
-        if (K_LOAD_SHMEM != 0) {
-            [[unroll]] for (uint32_t idx = 0; idx < Bc * HSK / 4; idx += gl_WorkGroupSize.x) {
-                uint32_t d = (idx + tid) % (HSK / 4);
-                uint32_t c = (idx + tid) / (HSK / 4);
-                if (c < Bc && d < HSK / 4) {
+        if (SHMEM_STAGING != 0) {
+            [[unroll]] for (uint32_t idx = 0; idx < Bc * HSK_pad / 4; idx += gl_WorkGroupSize.x) {
+                uint32_t d = (idx + tid) % (HSK_pad / 4);
+                uint32_t c = (idx + tid) / (HSK_pad / 4);
+                if (idx + gl_WorkGroupSize.x <= Bc * HSK_pad / 4 || c < Bc) {
                     f16vec4 K_Tf = f16vec4(0);
-                    if (!KV_bounds_check || j * Bc + c < KV) {
+                    if ((!KV_bounds_check || j * Bc + c < KV) && (HSK == HSK_pad || d < HSK / 4)) {
 #if BLOCK_SIZE > 1
                         uint coord = (j * Bc + c) * k_stride * BLOCK_SIZE + 4 * d;
                         uint ib = coord / BLOCK_SIZE;
                         uint iqs = (coord % BLOCK_SIZE);
-                        K_Tf = f16vec4(dequantize4(ib, iqs, k_offset, BINDING_IDX_K));
+                        K_Tf = dequantize4(ib, iqs, k_offset, BINDING_IDX_K);
 #else
                         K_Tf = f16vec4(data_kv4[k_offset / 4 + (j * Bc + c) * k_stride / 4 + d]);
 #endif
                     }
 
-                    ksh[c * kshstride + d] = K_Tf;
+                    kvsh[c * kvsh_stride + d] = K_Tf;
                 }
             }
             barrier();
@@ -262,7 +251,11 @@ void main() {
         coopmat<float16_t, gl_ScopeSubgroup, 16, MatBr, gl_MatrixUseB> QMat;
 
         [[unroll]] for (uint32_t d = 0; d < HSK_pad / 16; ++d) {
-            if (K_LOAD_SHMEM == 0) {
+            // If SHMEM_STAGING is set, a Bc * HSK_pad size tile of K is loaded to shmem
+            // If not, f16 K is loaded directly from global memory if aligned, otherwise
+            // staged through a Bc * MatBr size staging buffer.
+            // If K is not type f16, then it is always staged for dequantization.
+            if (SHMEM_STAGING == 0) {
 #if BLOCK_SIZE == 1
             if (KV_bounds_check || d * 16 + 16 > HSK) {
 #endif
@@ -277,13 +270,13 @@ void main() {
                         uint coord = (j * Bc + row) * k_stride * BLOCK_SIZE + d * 16 + col_vec * 4;
                         uint ib = coord / BLOCK_SIZE;
                         uint iqs = (coord % BLOCK_SIZE);
-                        K_Tf = f16vec4(dequantize4(ib, iqs, k_offset, BINDING_IDX_K));
+                        K_Tf = dequantize4(ib, iqs, k_offset, BINDING_IDX_K);
 #else
                         K_Tf = f16vec4(data_kv4[k_offset / 4 + (j * Bc + row) * k_stride / 4 + d * 16 / 4 + col_vec]);
 #endif
                     }
 
-                    ksh[row * kshstride + col_vec] = K_Tf;
+                    kvsh[row * kvsh_stride + col_vec] = K_Tf;
                 }
             }
             barrier();
@@ -295,8 +288,8 @@ void main() {
             if (KV_bounds_check || d * 16 + 16 > HSK)
 #endif
             {
-                uint coord = (gl_SubgroupID * MatBc) * kshstride;
-                coopMatLoad(KMat, ksh, coord, kshstride, gl_CooperativeMatrixLayoutRowMajor);
+                uint coord = (gl_SubgroupID * MatBc) * kvsh_stride;
+                coopMatLoad(KMat, kvsh, coord, kvsh_stride, gl_CooperativeMatrixLayoutRowMajor);
             }
 #if BLOCK_SIZE == 1
             else {
@@ -305,8 +298,8 @@ void main() {
             }
 #endif
             } else {
-                uint coord = (gl_SubgroupID * MatBc) * kshstride + d * 16 / 4;
-                coopMatLoad(KMat, ksh, coord, kshstride, gl_CooperativeMatrixLayoutRowMajor);
+                uint coord = (gl_SubgroupID * MatBc) * kvsh_stride + d * 16 / 4;
+                coopMatLoad(KMat, kvsh, coord, kvsh_stride, gl_CooperativeMatrixLayoutRowMajor);
             }
 
             coopMatLoad(QMat, Qf, d * 16 / 4, qstride, gl_CooperativeMatrixLayoutColumnMajor);
@@ -329,7 +322,7 @@ void main() {
             barrier();
         }
 
-        if (MASK_ENABLE) {
+        if (MASK_ENABLE && mask_opt_bits != MASK_OPT_ALL_ZERO) {
             [[unroll]] for (uint32_t idx = 0; idx < Bc * Br / 4; idx += gl_WorkGroupSize.x) {
                 uint32_t c = (idx + tid) / (Br / 4);
                 uint32_t r = (idx + tid) % (Br / 4);
@@ -374,7 +367,7 @@ void main() {
         [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
             const uint d_local = d0 / threads_per_rowgroup;
             [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-                Of[r][d_local] = ACC_TYPE(eMf[r]) * Of[r][d_local];
+                Of[r][d_local] = float16_t(eMf[r]) * Of[r][d_local];
             }
         }
 
@@ -397,19 +390,47 @@ void main() {
             }
         }
 
+        if (SHMEM_STAGING != 0) {
+            [[unroll]] for (uint32_t idx = 0; idx < Bc * HSV_pad / 4; idx += gl_WorkGroupSize.x) {
+                uint32_t d = (idx + tid) % (HSV_pad / 4);
+                uint32_t c = (idx + tid) / (HSV_pad / 4);
+                if (idx + gl_WorkGroupSize.x <= Bc * HSV_pad / 4 || c < Bc) {
+                    f16vec4 V_Tf = f16vec4(0);
+                    if ((!KV_bounds_check || j * Bc + c < KV) && (HSV == HSV_pad || d < HSV / 4)) {
+#if BLOCK_SIZE > 1
+                        uint coord = (j * Bc + c) * v_stride * BLOCK_SIZE + 4 * d;
+                        uint ib = coord / BLOCK_SIZE;
+                        uint iqs = (coord % BLOCK_SIZE);
+                        V_Tf = dequantize4(ib, iqs, v_offset, BINDING_IDX_V);
+#else
+                        V_Tf = f16vec4(data_vv4[v_offset / 4 + (j * Bc + c) * v_stride / 4 + d]);
+#endif
+                    }
+
+                    kvsh[c * kvsh_stride + d] = V_Tf;
+                }
+            }
+        }
+        barrier();
+
         const uint num_hsv_tiles = (HSV + MatBc * row_split - 1) / (MatBc * row_split); // round up
 
         // Each subgroup handles HSV/4 columns
         [[unroll]] for (uint32_t hsv_tile = 0; hsv_tile < num_hsv_tiles; ++hsv_tile) {
             const uint hsv_offset = (hsv_tile * row_split + gl_SubgroupID) * 16;
 
-            SfMat = coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator>(0);
+            coopmat<float16_t, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator> PVMat = coopmat<float16_t, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator>(0);
 
             // Preload V tiles for [Bc, 16 * num subgroups]
             const uint v_rows = Bc;
             const uint v_total = v_rows * v_cols;
             const uint v_loads_per_thread = v_total / gl_WorkGroupSize.x;
 
+            // If SHMEM_STAGING is set, a Bc * HSV_pad size tile of V is loaded to shmem.
+            // If not, f16 V is loaded directly from global memory if aligned, otherwise
+            // staged through a Bc * MatBr size staging buffer.
+            // If V is not type f16, then it is always staged for dequantization.
+            if (SHMEM_STAGING == 0) {
 #if BLOCK_SIZE == 1
             // For f16, only preload if not aligned
             if (KV_bounds_check) {
@@ -428,44 +449,52 @@ void main() {
 
                 if (!KV_bounds_check || (v_row < KV && v_col < HSV)) {
 #if BLOCK_SIZE > 1
-                    ksh[row * vsh_stride + col] = f16vec4(dequantize4(ib, iqs, v_offset, BINDING_IDX_V));
+                    kvsh[row * vsh_stride + col] = dequantize4(ib, iqs, v_offset, BINDING_IDX_V);
 #else
-                    ksh[row * vsh_stride + col] = data_vv4[(v_offset + v_row * v_stride + v_col) / 4];
+                    kvsh[row * vsh_stride + col] = data_vv4[(v_offset + v_row * v_stride + v_col) / 4];
 #endif
                 } else {
-                    ksh[row * vsh_stride + col] = f16vec4(0.0f);
+                    kvsh[row * vsh_stride + col] = f16vec4(0.0f);
                 }
             }
+
 #if BLOCK_SIZE == 1
             }
 #endif
-
+            }
             barrier();
 
-            [[unroll]] for (uint32_t bc_chunk = 0; bc_chunk < Bc / MatBc; ++bc_chunk) {
-                coopMatLoad(KMat, Psh, bc_chunk * MatBc * psh_stride, psh_stride, gl_CooperativeMatrixLayoutColumnMajor);
+            const uint o_offset = gl_SubgroupID * MatBr / 4;
 
+            if (hsv_offset < HSV_pad) {
+                [[unroll]] for (uint32_t bc_chunk = 0; bc_chunk < Bc / MatBc; ++bc_chunk) {
+                    coopMatLoad(KMat, Psh, bc_chunk * MatBc * psh_stride, psh_stride, gl_CooperativeMatrixLayoutColumnMajor);
+
+                    if (SHMEM_STAGING == 0) {
 #if BLOCK_SIZE == 1
-                if (!KV_bounds_check) {
-                    // F16 values can be loaded directly from global memory
-                    const uint v_tile_row = j * Bc + bc_chunk * MatBc;
-                    const uint v_tile_offset = v_offset / 4 + v_tile_row * v_stride / 4 + hsv_offset / 4;
-                    coopMatLoad(QMat, data_vv4, v_tile_offset, v_stride / 4, gl_CooperativeMatrixLayoutRowMajor);
-                } else
+                    if (!KV_bounds_check) {
+                        // F16 values can be loaded directly from global memory
+                        const uint v_tile_row = j * Bc + bc_chunk * MatBc;
+                        const uint v_tile_offset = v_offset / 4 + v_tile_row * v_stride / 4 + hsv_offset / 4;
+                        coopMatLoad(QMat, data_vv4, v_tile_offset, v_stride / 4, gl_CooperativeMatrixLayoutRowMajor);
+                    } else
 #endif
-                {
-                    const uint v_tile_offset = bc_chunk * MatBr * v_cols + gl_SubgroupID * (MatBc / 4);
-                    coopMatLoad(QMat, ksh, v_tile_offset, vsh_stride, gl_CooperativeMatrixLayoutRowMajor);
+                    {
+                        const uint v_tile_offset = bc_chunk * MatBr * v_cols + gl_SubgroupID * (MatBc / 4);
+                        coopMatLoad(QMat, kvsh, v_tile_offset, vsh_stride, gl_CooperativeMatrixLayoutRowMajor);
+                    }
+                    } else {
+                        const uint v_tile_offset = bc_chunk * MatBc * kvsh_stride + (hsv_tile * row_split + gl_SubgroupID) * (MatBc / 4);
+                        coopMatLoad(QMat, kvsh, v_tile_offset, kvsh_stride, gl_CooperativeMatrixLayoutRowMajor);
+                    }
+
+                    PVMat = coopMatMulAdd(KMat, QMat, PVMat);
                 }
 
-                SfMat = coopMatMulAdd(KMat, QMat, SfMat);
+                // Store PVMat to pvsh and load into Of
+                coopMatStore(PVMat, pvsh, o_offset, osh_stride, gl_CooperativeMatrixLayoutRowMajor);
             }
 
-            // Store SfMat to sfsh and load into Of
-            const uint osh_stride = row_split * MatBc / 4;
-            const uint o_offset = gl_SubgroupID * MatBc / 4;
-            coopMatStore(SfMat, sfsh, o_offset, osh_stride, gl_CooperativeMatrixLayoutRowMajor);
-
             barrier();
 
             const uint hsv_per_tile = row_split * MatBc;
@@ -484,7 +513,7 @@ void main() {
 
                     if (hsv_col >= hsv_base && hsv_col < hsv_base + hsv_per_tile && hsv_col < HSV) {
                         const uint local_hsv = (hsv_col - hsv_base) / 4;
-                        Of[r][d_local] += ACC_TYPEV4(sfsh[row * osh_stride + local_hsv]);
+                        Of[r][d_local] += pvsh[row * osh_stride + local_hsv];
                     }
                 }
             }
@@ -500,27 +529,48 @@ void main() {
     // If there is split_k, then the split_k resolve shader does the final
     // division by L. Store the intermediate O value and per-row m and L values.
     if (p.k_num > 1) {
-        // note: O and Q have swapped coord 1,2.
-        uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
+        if (p.gqa_ratio > 1) {
+            // note: O and Q have swapped coord 1,2.
+            uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3)) / 4;
 
-        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-            if (tile_row(r) < N) {
-                [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
-                    const uint d = d0 + col_tid;
-                    if (d >= HSV/4) break;
-                    const uint d_local = d0 / threads_per_rowgroup;
-                    [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        perElemOpGqaStore(tile_row(r), 4 * d + comp, float(Of[r][d_local][comp]), o_offset, iq2, N);
+            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                if (tile_row(r) < N) {
+                    [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
+                        const uint d = d0 + col_tid;
+                        if (d >= HSV/4) break;
+                        const uint d_local = d0 / threads_per_rowgroup;
+                        gqaStore(tile_row(r), d, Of[r][d_local], o_offset, iq2, N);
                     }
                 }
             }
-        }
 
-        o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
-        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-            if (tile_row(r) < N) {
-                perElemOpStoreCol0(tile_row(r), 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
-                perElemOpStoreCol0(tile_row(r), 0u, ACC_TYPE(Mf[r]), o_offset + p.ne1, iq2, N);
+            o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
+            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                if (tile_row(r) < N) {
+                    perElemOpStoreCol0(tile_row(r), 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
+                    perElemOpStoreCol0(tile_row(r), 0u, ACC_TYPE(Mf[r]), o_offset + p.ne1, iq2, N);
+                }
+            }
+        } else {
+            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                const uint row = tile_row(r);
+                const uint global_row = i * Br + row;
+
+                if (global_row < N) {
+                    uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (global_row + p.ne2 * iq3)) / 4;
+
+                    [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
+                        const uint d = d0 + col_tid;
+                        if (d >= HSV/4) break;
+                        data_ov4[o_offset + iq2 * HSV/4 + d] = D_TYPEV4(Of[r][d/threads_per_rowgroup]);
+                    }
+                }
+
+                if (global_row < N && col_tid == 0) {
+                    uint32_t lm_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (global_row + p.ne2 * iq3));
+                    data_o[lm_offset + iq2] = D_TYPE(Lf[r]);
+                    data_o[lm_offset + p.ne1 + iq2] = D_TYPE(Mf[r]);
+                }
             }
         }
 
@@ -539,7 +589,7 @@ void main() {
 
                 [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
                     const uint d_local = d0 / threads_per_rowgroup;
-                    Of[r][d_local] *= ACC_TYPE(ms);
+                    Of[r][d_local] *= float16_t(ms);
                 }
             } else {
                 vs = exp(sink - Mf[r]);
@@ -557,14 +607,14 @@ void main() {
     [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
         const uint d_local = d0 / threads_per_rowgroup;
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-            Of[r][d_local] *= ACC_TYPE(Lfrcp[r]);
-#if defined(ACC_TYPE_MAX)
-            Of[r][d_local] = clamp(Of[r][d_local], -ACC_TYPE_MAX, ACC_TYPE_MAX);
+            Of[r][d_local] *= float16_t(Lfrcp[r]);
+#if defined(FLOAT_TYPE_MAX)
+            Of[r][d_local] = clamp(Of[r][d_local], -FLOAT_TYPE_MAX, FLOAT_TYPE_MAX);
 #endif
         }
     }
 
-    uint32_t o_offset = gqa_iq1*p.ne1*HSV + iq3*p.ne2*p.ne1*HSV;
+    uint32_t o_offset = (gqa_iq1*p.ne1*HSV + iq3*p.ne2*p.ne1*HSV) / 4;
 
     if (p.gqa_ratio > 1) {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
@@ -573,9 +623,7 @@ void main() {
                     const uint d = d0 + col_tid;
                     if (d >= HSV / 4) break;
                     const uint d_local = d0 / threads_per_rowgroup;
-                    [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        perElemOpGqaStore(tile_row(r), 4 * d + comp, float(Of[r][d_local][comp]), o_offset, iq2, N);
-                    }
+                    gqaStore(tile_row(r), d, Of[r][d_local], o_offset, iq2, N);
                 }
             }
         }
@@ -586,9 +634,7 @@ void main() {
                     const uint d = d0 + col_tid;
                     if (d >= HSV / 4) break;
                     const uint d_local = d0 / threads_per_rowgroup;
-                    [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        data_o[o_offset + iq2 * HSV + (i * Br + tile_row(r)) * p.ne1 * HSV + 4 * d + comp] = D_TYPE(Of[r][d_local][comp]);
-                    }
+                    data_ov4[o_offset + (iq2 * HSV + (i * Br + tile_row(r)) * p.ne1 * HSV) / 4 + d] = D_TYPEV4(Of[r][d_local]);
                 }
             }
         }

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

@@ -72,6 +72,28 @@ D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TY
     return elem;
 }
 
+// Store O values for non-GQA split_k. Rows are tokens, not heads.
+D_TYPE perElemOpNonGqaSplitKStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t unused, const in uint32_t iq2, const in uint32_t N) {
+    uint32_t global_row = i * Br + r;
+    if (global_row < N && c < HSV) {
+        uint32_t o_off = HSV * p.ne1
+            * (split_k_index + p.k_num * (global_row + p.ne2 * iq3));
+        data_o[o_off + iq2 * HSV + c] = D_TYPE(elem);
+    }
+    return elem;
+}
+
+// Store L/M values for non-GQA split_k.
+ACC_TYPE perElemOpNonGqaSplitKStoreCol0(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t lm_base, const in uint32_t iq2, const in uint32_t N) {
+    uint32_t global_row = i * Br + r;
+    if (global_row < N && c == 0) {
+        uint32_t lm_off = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3
+            + p.ne1 * 2 * (split_k_index + p.k_num * (global_row + p.ne2 * iq3));
+        data_o[lm_off + lm_base + iq2] = D_TYPE(elem);
+    }
+    return elem;
+}
+
 void main() {
 #ifdef NEEDS_INIT_IQ_SHMEM
     init_iq_shmem(gl_WorkGroupSize);
@@ -290,13 +312,19 @@ void main() {
     if (p.k_num > 1) {
         coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(O);
 
-        // note: O and Q have swapped coord 1,2.
-        uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
-        coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
+        if (p.gqa_ratio > 1) {
+            // note: O and Q have swapped coord 1,2.
+            uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
+            coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
 
-        o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
-        coopMatPerElementNV(L, L, perElemOpStoreCol0, o_offset, iq2, N);
-        coopMatPerElementNV(M, M, perElemOpStoreCol0, o_offset + p.ne1, iq2, N);
+            o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
+            coopMatPerElementNV(L, L, perElemOpStoreCol0, o_offset, iq2, N);
+            coopMatPerElementNV(M, M, perElemOpStoreCol0, o_offset + p.ne1, iq2, N);
+        } else {
+            coopMatPerElementNV(O_D, O_D, perElemOpNonGqaSplitKStore, 0u, iq2, N);
+            coopMatPerElementNV(L, L, perElemOpNonGqaSplitKStoreCol0, 0u, iq2, N);
+            coopMatPerElementNV(M, M, perElemOpNonGqaSplitKStoreCol0, p.ne1, iq2, N);
+        }
         return;
     }
 

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

@@ -595,8 +595,6 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
 }
 
 void process_shaders() {
-    std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}};
-
     // matmul
     for (const MatMulIdType& matmul_id_type : {MatMulIdType::NONE, MatMulIdType::DEFAULT, MatMulIdType::SUBGROUP}) {
         // No coopmats
@@ -622,49 +620,63 @@ void process_shaders() {
         }
     }
 
-    // flash attention
-    for (const auto& f16acc : {false, true}) {
-        std::map<std::string, std::string> fa_base_dict = base_dict;
-        fa_base_dict["ACC_TYPE"] = f16acc ? "float16_t" : "float";
-        fa_base_dict["ACC_TYPEV4"] = f16acc ? "f16vec4" : "vec4";
-        if (f16acc) {
-            fa_base_dict["ACC_TYPE_MAX"] = "float16_t(65504.0)";
+    for (const bool& fp16 : {false, true}) {
+        std::map<std::string, std::string> base_dict;
+        if (fp16) {
+            base_dict = {{"FLOAT_TYPE", "float16_t"}, {"FLOAT_TYPEV4", "f16vec4"}, {"FLOAT16", "1"}, {"FLOAT_TYPE_MAX", "float16_t(65504.0)"}};
+        } else {
+            base_dict = {{"FLOAT_TYPE", "float"}, {"FLOAT_TYPEV4", "vec4"}};
         }
 
-        for (const auto& tname : type_names) {
-            if (tname == "bf16") continue;
-
-#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
-            if (tname == "f16") {
-                string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
-                    merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}}), true, false, true, f16acc);
-            } else {
-                std::string data_a_key = "DATA_A_" + to_uppercase(tname);
-                string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
-                    merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"DEQUANTFUNC", "dequantFunc"+to_uppercase(tname) }, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, false, true, f16acc);
+        // flash attention
+        for (const bool& f16acc : {false, true}) {
+            std::map<std::string, std::string> fa_base_dict = base_dict;
+            fa_base_dict["ACC_TYPE"] = fp16 && f16acc ? "float16_t" : "float";
+            fa_base_dict["ACC_TYPEV4"] = fp16 && f16acc ? "f16vec4" : "vec4";
+            if (fp16 && f16acc) {
+                fa_base_dict["ACC_TYPE_MAX"] = "float16_t(65504.0)";
             }
+
+            for (const auto& tname : type_names) {
+                if (tname == "bf16") continue;
+
+                if (fp16) {
+#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+                if (tname == "f16") {
+                    string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
+                        merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}}), fp16, false, true, f16acc);
+                } else {
+                    std::string data_a_key = "DATA_A_" + to_uppercase(tname);
+                    string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
+                        merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"DEQUANTFUNC", "dequantFunc"+to_uppercase(tname) }, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), fp16, false, true, f16acc);
+                }
 #endif
 #if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
-            if (tname == "f16") {
-                string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm1.comp",
-                    merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"COOPMAT", "1"}}), true, true, false, f16acc);
-            } else if (tname == "q4_0" || tname == "q8_0" || tname == "f32") {
-                std::string data_a_key = "DATA_A_" + to_uppercase(tname);
-                string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm1.comp",
-                    merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname)}, {"COOPMAT", "1"}}), true, true, false, f16acc);
-            }
+                if (tname == "f16") {
+                    string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm1.comp",
+                        merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"COOPMAT", "1"}}), fp16, true, false, f16acc);
+                } else if (tname == "q4_0" || tname == "q8_0" || tname == "f32") {
+                    std::string data_a_key = "DATA_A_" + to_uppercase(tname);
+                    string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm1.comp",
+                        merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname)}, {"COOPMAT", "1"}}), fp16, true, false, f16acc);
+                }
 #endif
-            if (tname == "f16") {
-                string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
-                    merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}}), true, false, false, f16acc);
-            } else if (tname == "q4_0" || tname == "q8_0" || tname == "f32") {
-                std::string data_a_key = "DATA_A_" + to_uppercase(tname);
-                string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
-                    merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, false, false, f16acc);
+                }
+
+                if (tname == "f16") {
+                    string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
+                        merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}}), fp16, false, false, f16acc);
+                } else if (tname == "q4_0" || tname == "q8_0" || tname == "f32") {
+                    std::string data_a_key = "DATA_A_" + to_uppercase(tname);
+                    string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
+                        merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), fp16, false, false, f16acc);
+                }
             }
         }
     }
 
+    std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}};
+
     for (const auto& tname : type_names) {
         // mul mat vec
         std::string data_a_key = "DATA_A_" + to_uppercase(tname);

ggml/src/ggml-zendnn/CMakeLists.txt

@@ -1,12 +1,19 @@
 ggml_add_backend_library(ggml-zendnn
                          ggml-zendnn.cpp)
 
-# Get ZenDNN path
 if (NOT DEFINED ZENDNN_ROOT OR ZENDNN_ROOT STREQUAL "")
     set(ZENDNN_ROOT "$ENV{ZENDNN_ROOT}")
 endif()
 
-# Check if path is still empty or OFF
+if (BUILD_SHARED_LIBS)
+    set(ZENDNN_SHARED_LIB ON)
+    set(ZENDNN_ARCHIVE_LIB OFF)
+else()
+    set(ZENDNN_SHARED_LIB OFF)
+    set(ZENDNN_ARCHIVE_LIB ON)
+endif()
+
+# Download and build ZenDNN if not provided
 if (NOT ZENDNN_ROOT OR ZENDNN_ROOT STREQUAL "" OR ZENDNN_ROOT STREQUAL "OFF")
     message(STATUS "ZENDNN_ROOT not set. Automatically downloading and building ZenDNN...")
     message(STATUS "This will take several minutes on first build...")
@@ -21,7 +28,7 @@ if (NOT ZENDNN_ROOT OR ZENDNN_ROOT STREQUAL "" OR ZENDNN_ROOT STREQUAL "OFF")
     ExternalProject_Add(
         zendnn
         GIT_REPOSITORY https://github.com/amd/ZenDNN.git
-        GIT_TAG 21ce8f7879c86bf3637f707fae6f29e0951db5fe
+        GIT_TAG a18adf8c605fb5f5e52cefd7eda08a7b18febbaf    # ZenDNN-2026-WW08
         PREFIX      ${ZENDNN_PREFIX}
         SOURCE_DIR  ${ZENDNN_SOURCE_DIR}
         BINARY_DIR  ${ZENDNN_BUILD_DIR}
@@ -32,7 +39,9 @@ if (NOT ZENDNN_ROOT OR ZENDNN_ROOT STREQUAL "" OR ZENDNN_ROOT STREQUAL "OFF")
             -DZENDNNL_BUILD_DOXYGEN=OFF
             -DZENDNNL_BUILD_GTEST=OFF
             -DZENDNNL_BUILD_BENCHDNN=OFF
-            # Enable ALL matmul algorithm backends
+            -DZENDNNL_DEPENDS_FBGEMM=OFF
+            -DZENDNNL_LIB_BUILD_ARCHIVE=${ZENDNN_ARCHIVE_LIB}
+            -DZENDNNL_LIB_BUILD_SHARED=${ZENDNN_SHARED_LIB}
             -DZENDNNL_DEPENDS_AOCLDLP=ON
             -DZENDNNL_DEPENDS_ONEDNN=ON
             -DZENDNNL_DEPENDS_LIBXSMM=ON
@@ -45,47 +54,37 @@ if (NOT ZENDNN_ROOT OR ZENDNN_ROOT STREQUAL "" OR ZENDNN_ROOT STREQUAL "OFF")
         LOG_INSTALL ON
     )
 
-    # Add dependency so ZenDNN builds before our library
     add_dependencies(ggml-zendnn zendnn)
-
-    # Set ZENDNN_ROOT to the installation directory
     set(ZENDNN_ROOT ${ZENDNN_INSTALL_DIR})
-
     message(STATUS "ZenDNN will be built to: ${ZENDNN_ROOT}")
 else()
     message(STATUS "Using custom ZenDNN installation at: ${ZENDNN_ROOT}")
 endif()
 
-# ZenDNN headers + libs
 target_include_directories(ggml-zendnn PRIVATE
     ${ZENDNN_ROOT}/zendnnl/include
-    ${ZENDNN_ROOT}/deps/aocldlp/include
-    ${ZENDNN_ROOT}/deps/aoclutils/include
     ${ZENDNN_ROOT}/deps/json/include
-    ${ZENDNN_ROOT}/deps/libxsmm/include
+    ${ZENDNN_ROOT}/deps/aoclutils/include
+    ${ZENDNN_ROOT}/deps/aocldlp/include
     ${ZENDNN_ROOT}/deps/onednn/include
-)
+    ${ZENDNN_ROOT}/deps/libxsmm/include)
 
-target_link_directories(ggml-zendnn PRIVATE
-    ${ZENDNN_ROOT}/zendnnl/lib
-    ${ZENDNN_ROOT}/deps/aocldlp/lib
-    ${ZENDNN_ROOT}/deps/aoclutils/lib
-    ${ZENDNN_ROOT}/deps/libxsmm/lib
-    ${ZENDNN_ROOT}/deps/onednn/lib
-)
+if (ZENDNN_SHARED_LIB)
+    target_link_directories(ggml-zendnn PRIVATE ${ZENDNN_ROOT}/zendnnl/lib)
+    target_link_libraries(ggml-zendnn PRIVATE zendnnl)
+elseif (ZENDNN_ARCHIVE_LIB)
+    target_link_libraries(ggml-zendnn PRIVATE
+        ${ZENDNN_ROOT}/zendnnl/lib/libzendnnl_archive.a
+        ${ZENDNN_ROOT}/deps/aoclutils/${CMAKE_INSTALL_LIBDIR}/libaoclutils.a
+        ${ZENDNN_ROOT}/deps/aoclutils/${CMAKE_INSTALL_LIBDIR}/libau_cpuid.a
+        ${ZENDNN_ROOT}/deps/aocldlp/lib/libaocl-dlp.a
+        ${ZENDNN_ROOT}/deps/onednn/${CMAKE_INSTALL_LIBDIR}/libdnnl.a
+        ${ZENDNN_ROOT}/deps/libxsmm/lib/libxsmm.a
+        ${ZENDNN_ROOT}/deps/libxsmm/lib/libxsmmext.a
+        ${ZENDNN_ROOT}/deps/libxsmm/lib/libxsmmnoblas.a)
+endif()
 
-target_link_libraries(ggml-zendnn PRIVATE
-    zendnnl_archive    # ZenDNN main
-    aocl-dlp           # AOCL libraries
-    aoclutils
-    au_cpuid
-    dnnl               # OneDNN
-    xsmm               # libxsmm small matrix math
-    xsmmext
-    xsmmnoblas
-    m
-    pthread
-)
+target_link_libraries(ggml-zendnn PRIVATE m pthread)
 
 if (GGML_OPENMP)
     target_link_libraries(ggml-zendnn PRIVATE OpenMP::OpenMP_CXX)

ggml/src/ggml-zendnn/ggml-zendnn.cpp

@@ -41,13 +41,13 @@ static bool ggml_zendnn_matmul(ggml_backend_zendnn_context * ctx, int64_t m, int
                                const TA * A, int64_t lda, const TB * B, int64_t ldb, TC * C,
                                int64_t ldc) {
 
-    zendnnl::lowoha::lowoha_params params;
+    zendnnl::lowoha::matmul::matmul_params params;
     params.dtypes.src = ggml_to_zendnn_type<TB>();
     params.dtypes.wei = ggml_to_zendnn_type<TA>();
     params.dtypes.dst = ggml_to_zendnn_type<TC>();
     params.num_threads = ctx->n_threads;
 
-    zendnnl::lowoha::status_t status = zendnnl::lowoha::matmul_direct(
+    zendnnl::error_handling::status_t status = zendnnl::lowoha::matmul::matmul_direct(
         'r', false, true,   // row-major, don't transpose B, transpose A (because it's column-major)
         n,                  // M: rows of B and C
         m,                  // N: cols of A^T and C
@@ -63,7 +63,7 @@ static bool ggml_zendnn_matmul(ggml_backend_zendnn_context * ctx, int64_t m, int
         params              // params
     );
 
-    if (status != zendnnl::lowoha::status_t::success) {
+    if (status != zendnnl::error_handling::status_t::success) {
         GGML_LOG_ERROR("%s, ZenDNN matmul failed: status=%d\n", __func__, static_cast<int>(status));
         return false;
     }

ggml/src/ggml.c

@@ -899,7 +899,8 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
 };
 
 const struct ggml_type_traits * ggml_get_type_traits(enum ggml_type type) {
-    GGML_ASSERT(type < GGML_TYPE_COUNT);
+    assert(type >= 0);
+    assert(type < GGML_TYPE_COUNT);
     return &type_traits[type];
 }
 
@@ -1265,27 +1266,33 @@ size_t ggml_nbytes_pad(const struct ggml_tensor * tensor) {
 }
 
 int64_t ggml_blck_size(enum ggml_type type) {
+    assert(type >= 0);
+    assert(type < GGML_TYPE_COUNT);
     return type_traits[type].blck_size;
 }
 
 size_t ggml_type_size(enum ggml_type type) {
+    assert(type >= 0);
+    assert(type < GGML_TYPE_COUNT);
     return type_traits[type].type_size;
 }
 
 size_t ggml_row_size(enum ggml_type type, int64_t ne) {
+    assert(type >= 0);
+    assert(type < GGML_TYPE_COUNT);
     assert(ne % ggml_blck_size(type) == 0);
     return ggml_type_size(type)*ne/ggml_blck_size(type);
 }
 
-double ggml_type_sizef(enum ggml_type type) {
-    return ((double)(type_traits[type].type_size))/type_traits[type].blck_size;
-}
-
 const char * ggml_type_name(enum ggml_type type) {
-    return type < GGML_TYPE_COUNT ? type_traits[type].type_name : "NONE";
+    assert(type >= 0);
+    assert(type < GGML_TYPE_COUNT);
+    return type_traits[type].type_name;
 }
 
 bool ggml_is_quantized(enum ggml_type type) {
+    assert(type >= 0);
+    assert(type < GGML_TYPE_COUNT);
     return type_traits[type].is_quantized;
 }
 
@@ -1629,11 +1636,23 @@ static struct ggml_object * ggml_new_object(struct ggml_context * ctx, enum ggml
     const size_t cur_end  = cur_offs + cur_size;
 
     // align to GGML_MEM_ALIGN
+    GGML_ASSERT(size <= SIZE_MAX - (GGML_MEM_ALIGN - 1));
     size_t size_needed = GGML_PAD(size, GGML_MEM_ALIGN);
 
     char * const mem_buffer = ctx->mem_buffer;
     struct ggml_object * const obj_new = (struct ggml_object *)(mem_buffer + cur_end);
 
+    // integer overflow checks
+    if (cur_end > SIZE_MAX - size_needed) {
+        GGML_LOG_WARN("%s: overflow detected in cur_end (%zu) + size_needed (%zu)\n", __func__, cur_end, size_needed);
+        return NULL;
+    }
+    if (cur_end + size_needed > SIZE_MAX - GGML_OBJECT_SIZE) {
+        GGML_LOG_WARN("%s: overflow detected in cur_end (%zu) + size_needed (%zu) + GGML_OBJECT_SIZE (%zu)\n", __func__,
+                cur_end, size_needed, (size_t) GGML_OBJECT_SIZE);
+        return NULL;
+    }
+
     if (cur_end + size_needed + GGML_OBJECT_SIZE > ctx->mem_size) {
         GGML_LOG_WARN("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n",
                 __func__, cur_end + size_needed + GGML_OBJECT_SIZE, ctx->mem_size);
@@ -1702,6 +1721,8 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         obj_alloc_size = data_size;
     }
 
+    GGML_ASSERT(GGML_TENSOR_SIZE <= SIZE_MAX - obj_alloc_size);
+
     struct ggml_object * const obj_new = ggml_new_object(ctx, GGML_OBJECT_TYPE_TENSOR, GGML_TENSOR_SIZE + obj_alloc_size);
     GGML_ASSERT(obj_new);
 

ggml/src/gguf.cpp

@@ -15,6 +15,17 @@
 #include <string>
 #include <vector>
 
+#define GGUF_MAX_STRING_LENGTH  (1024*1024*1024)
+#define GGUF_MAX_ARRAY_ELEMENTS (1024*1024*1024)
+
+#ifdef _WIN32
+#    define gguf_ftell _ftelli64
+#    define gguf_fseek _fseeki64
+#else
+#    define gguf_ftell ftello
+#    define gguf_fseek fseeko
+#endif
+
 template <typename T>
 struct type_to_gguf_type;
 
@@ -217,17 +228,64 @@ struct gguf_context {
 };
 
 struct gguf_reader {
-    FILE * file;
+    gguf_reader(FILE * file) : file(file) {
+        // read the remaining bytes once and update on each read
+        nbytes_remain = file_remain(file);
+    }
 
-    gguf_reader(FILE * file) : file(file) {}
+    // helper for remaining bytes in a file
+    static uint64_t file_remain(FILE * file) {
+        const int64_t cur = gguf_ftell(file);
+        if (cur < 0) {
+            return 0;
+        }
+        if (gguf_fseek(file, 0, SEEK_END) != 0) {
+            gguf_fseek(file, cur, SEEK_SET);
+
+            return 0;
+        }
+        const int64_t end = gguf_ftell(file);
+        if (end < 0) {
+            gguf_fseek(file, cur, SEEK_SET);
+
+            return 0;
+        }
+        gguf_fseek(file, cur, SEEK_SET);
+        return static_cast<uint64_t>(end - cur);
+    }
 
     template <typename T>
     bool read(T & dst) const {
-        return fread(&dst, 1, sizeof(dst), file) == sizeof(dst);
+        const size_t size = sizeof(dst);
+        if (nbytes_remain < size) {
+            return false;
+        }
+        const size_t nread = fread(&dst, 1, size, file);
+        nbytes_remain -= nread;
+        return nread == size;
     }
 
     template <typename T>
     bool read(std::vector<T> & dst, const size_t n) const {
+        if (n > GGUF_MAX_ARRAY_ELEMENTS) {
+            return false;
+        }
+        if constexpr (std::is_same<T, std::string>::value) {
+            // strings are prefixed with their length, so we need to account for that
+            if (n > SIZE_MAX / sizeof(uint64_t)) {
+                return false;
+            }
+            if (nbytes_remain < n * sizeof(uint64_t)) {
+                return false;
+            }
+        } else {
+            if (n > SIZE_MAX / sizeof(T)) {
+                return false;
+            }
+            if (nbytes_remain < n * sizeof(T)) {
+                return false;
+            }
+        }
         dst.resize(n);
         for (size_t i = 0; i < dst.size(); ++i) {
             if constexpr (std::is_same<T, bool>::value) {
@@ -277,13 +335,33 @@ struct gguf_reader {
         if (!read(size)) {
             return false;
         }
-        dst.resize(size);
-        return fread(dst.data(), 1, dst.length(), file) == dst.length();
+        if (size > GGUF_MAX_STRING_LENGTH) {
+            GGML_LOG_ERROR("%s: string length %" PRIu64 " exceeds maximum %" PRIu64 "\n", __func__, size, (uint64_t) GGUF_MAX_STRING_LENGTH);
+            return false;
+        }
+        if (size > nbytes_remain) {
+            GGML_LOG_ERROR("%s: string length %" PRIu64 " exceeds remaining file size %" PRIu64 " bytes\n", __func__, size, nbytes_remain);
+            return false;
+        }
+        dst.resize(static_cast<size_t>(size));
+        const size_t nread = fread(dst.data(), 1, size, file);
+        nbytes_remain -= nread;
+        return nread == size;
     }
 
     bool read(void * dst, const size_t size) const {
-        return fread(dst, 1, size, file) == size;
+        if (size > nbytes_remain) {
+            return false;
+        }
+        const size_t nread = fread(dst, 1, size, file);
+        nbytes_remain -= nread;
+        return nread == size;
     }
+
+private:
+    FILE * file;
+
+    mutable uint64_t nbytes_remain;
 };
 
 struct gguf_context * gguf_init_empty(void) {
@@ -568,8 +646,8 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
 
             // check that tensor type is within defined range
             if (info.t.type < 0 || info.t.type >= GGML_TYPE_COUNT) {
-                GGML_LOG_ERROR("%s: tensor '%s' has invalid ggml type %d (%s)\n",
-                    __func__, info.t.name, info.t.type, ggml_type_name(info.t.type));
+                GGML_LOG_ERROR("%s: tensor '%s' has invalid ggml type %d. should be in [0, %d)\n",
+                    __func__, info.t.name, info.t.type, GGML_TYPE_COUNT);
                 ok = false;
                 break;
             }
@@ -618,14 +696,14 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
     GGML_ASSERT(int64_t(ctx->info.size()) == n_tensors);
 
     // we require the data section to be aligned, so take into account any padding
-    if (fseek(file, GGML_PAD(ftell(file), ctx->alignment), SEEK_SET) != 0) {
+    if (gguf_fseek(file, GGML_PAD(gguf_ftell(file), ctx->alignment), SEEK_SET) != 0) {
         GGML_LOG_ERROR("%s: failed to seek to beginning of data section\n", __func__);
         gguf_free(ctx);
         return nullptr;
     }
 
     // store the current file offset - this is where the data section starts
-    ctx->offset = ftell(file);
+    ctx->offset = gguf_ftell(file);
 
     // compute the total size of the data section, taking into account the alignment
     {
@@ -657,10 +735,34 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
         //   the ggml_tensor structs to the appropriate locations in the binary blob
 
         // compute the exact size needed for the new ggml_context
-        const size_t mem_size =
-            params.no_alloc ?
-            (n_tensors    )*ggml_tensor_overhead() :
-            (n_tensors + 1)*ggml_tensor_overhead() + ctx->size;
+        size_t mem_size = 0;
+        if (params.no_alloc) {
+            if (n_tensors != 0 && SIZE_MAX / n_tensors < ggml_tensor_overhead()) {
+                GGML_LOG_ERROR("%s: memory size overflow while allocating ggml context\n", __func__);
+                gguf_free(ctx);
+                return nullptr;
+            }
+
+            const size_t overhead = n_tensors * ggml_tensor_overhead();
+
+            mem_size = overhead;
+        } else {
+            if ((n_tensors + 1) != 0 && SIZE_MAX / (n_tensors + 1) < ggml_tensor_overhead()) {
+                GGML_LOG_ERROR("%s: memory size overflow while allocating ggml context\n", __func__);
+                gguf_free(ctx);
+                return nullptr;
+            }
+
+            const size_t overhead = (n_tensors + 1) * ggml_tensor_overhead();
+
+            if (SIZE_MAX - overhead < ctx->size) {
+                GGML_LOG_ERROR("%s: memory size overflow while allocating ggml context\n", __func__);
+                gguf_free(ctx);
+                return nullptr;
+            }
+
+            mem_size = overhead + ctx->size;
+        }
 
         struct ggml_init_params pdata = {
             /*mem_size   =*/ mem_size,

gguf-py/gguf/constants.py

@@ -379,6 +379,7 @@ class MODEL_ARCH(IntEnum):
     NEO_BERT         = auto()
     JINA_BERT_V2     = auto()
     JINA_BERT_V3     = auto()
+    EUROBERT         = auto()
     BLOOM            = auto()
     STABLELM         = auto()
     QWEN             = auto()
@@ -531,6 +532,7 @@ class MODEL_TENSOR(IntEnum):
     FFN_GATE_EXP         = auto()
     FFN_DOWN_EXP         = auto()
     FFN_UP_EXP           = auto()
+    FFN_GATE_UP_EXP      = auto()
     FFN_GATE_SHEXP       = auto()
     FFN_DOWN_SHEXP       = auto()
     FFN_UP_SHEXP         = auto()
@@ -820,6 +822,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.NEO_BERT:         "neo-bert",
     MODEL_ARCH.JINA_BERT_V2:     "jina-bert-v2",
     MODEL_ARCH.JINA_BERT_V3:     "jina-bert-v3",
+    MODEL_ARCH.EUROBERT:         "eurobert",
     MODEL_ARCH.BLOOM:            "bloom",
     MODEL_ARCH.STABLELM:         "stablelm",
     MODEL_ARCH.QWEN:             "qwen",
@@ -978,6 +981,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.FFN_GATE_EXP:              "blk.{bid}.ffn_gate_exps",
     MODEL_TENSOR.FFN_DOWN_EXP:              "blk.{bid}.ffn_down_exps",
     MODEL_TENSOR.FFN_UP_EXP:                "blk.{bid}.ffn_up_exps",
+    MODEL_TENSOR.FFN_GATE_UP_EXP:           "blk.{bid}.ffn_gate_up_exps",
     MODEL_TENSOR.FFN_EXP_PROBS_B:           "blk.{bid}.exp_probs_b",
     MODEL_TENSOR.LAYER_OUT_NORM:            "blk.{bid}.layer_output_norm",
     MODEL_TENSOR.PER_LAYER_TOKEN_EMBD:      "per_layer_token_embd",           # gemma3n
@@ -1587,6 +1591,19 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_UP,
         MODEL_TENSOR.LAYER_OUT_NORM,
     ],
+    MODEL_ARCH.EUROBERT: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_DOWN,
+    ],
     MODEL_ARCH.MPT: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -1805,6 +1822,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_EXP,
         MODEL_TENSOR.FFN_UP_EXP,
         MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_GATE_UP_EXP,
         MODEL_TENSOR.SSM_A,
         MODEL_TENSOR.SSM_CONV1D,
         MODEL_TENSOR.SSM_DT,
@@ -1894,6 +1912,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_EXP,
         MODEL_TENSOR.FFN_UP_EXP,
         MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_GATE_UP_EXP,
         MODEL_TENSOR.SSM_A,
         MODEL_TENSOR.SSM_CONV1D,
         MODEL_TENSOR.SSM_DT,
@@ -2595,6 +2614,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_GATE_EXP,
         MODEL_TENSOR.FFN_DOWN_EXP,
         MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_GATE_UP_EXP,
         MODEL_TENSOR.FFN_GATE_SHEXP,
         MODEL_TENSOR.FFN_DOWN_SHEXP,
         MODEL_TENSOR.FFN_UP_SHEXP,

gguf-py/gguf/gguf_reader.py

@@ -175,6 +175,9 @@ class GGUFReader:
             if new_align.types != [GGUFValueType.UINT32]:
                 raise ValueError('Bad type for general.alignment field')
             self.alignment = new_align.parts[-1][0]
+            # Ensure alignment is a non-zero power of two
+            if self.alignment == 0 or (self.alignment & (self.alignment - 1)) != 0:
+                raise ValueError('Invalid alignment: must be a non-zero power of two')
         padding = offs % self.alignment
         if padding != 0:
             offs += self.alignment - padding
@@ -202,11 +205,11 @@ class GGUFReader:
 
     def _push_field(self, field: ReaderField, skip_sum: bool = False) -> int:
         if field.name in self.fields:
-            # TODO: add option to generate error on duplicate keys
-            # raise KeyError(f'Duplicate {field.name} already in list at offset {field.offset}')
+            # TODO: add option to make this a warning and accept duplicate keys like below
+            raise KeyError(f'Duplicate {field.name} already in list at offset {field.offset}')
 
-            logger.warning(f'Duplicate key {field.name} at offset {field.offset}')
-            self.fields[field.name + '_{}'.format(field.offset)] = field
+            # logger.warning(f'Duplicate key {field.name} at offset {field.offset}')
+            # self.fields[field.name + '_{}'.format(field.offset)] = field
         else:
             self.fields[field.name] = field
         return 0 if skip_sum else sum(int(part.nbytes) for part in field.parts)

gguf-py/gguf/gguf_writer.py

@@ -501,6 +501,8 @@ class GGUFWriter:
         self.add_uint32(Keys.General.QUANTIZATION_VERSION, quantization_version)
 
     def add_custom_alignment(self, alignment: int) -> None:
+        if alignment <= 0 or (alignment & (alignment - 1)) != 0:
+            raise ValueError('Invalid alignment: must be a non-zero power of two')
         self.data_alignment = alignment
         self.add_uint32(Keys.General.ALIGNMENT, alignment)
 

gguf-py/gguf/tensor_mapping.py

@@ -567,6 +567,10 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.chunk_experts.gate_proj",           # grovemoe
         ),
 
+        MODEL_TENSOR.FFN_GATE_UP_EXP: (
+            "model.layers.{bid}.mlp.experts.gate_up_proj",
+        ),
+
         # Feed-forward down
         MODEL_TENSOR.FFN_DOWN: (
             "gpt_neox.layers.{bid}.mlp.dense_4h_to_h",                # gptneox

gguf-py/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "gguf"
-version = "0.17.1"
+version = "0.18.0"
 description = "Read and write ML models in GGUF for GGML"
 authors = ["GGML <ggml@ggml.ai>"]
 packages = [

scripts/compare-logprobs.py

@@ -25,16 +25,12 @@ Example usage:
 """
 
 
-def generate_input_prompt(length: int) -> list[str]:
-    CORPUS = """
-    You are an advanced AI assistant capable of using tools to gather information, perform calculations, or execute tasks. Always think step by step before responding. If a user's query requires external data, computation, or actions beyond your internal knowledge, use the appropriate tools via function calls.
-
-    ### Tool Call Format:
-    When you need to use a tool, output the call in this exact XML format. Include the opening and closing tags. Do not escape arguments; they will be parsed as plain text.
-
-    You can make multiple calls in one go by placing them one after another.
-    """
-    words = [w.strip() for w in CORPUS.strip().split(" ")]
+def get_remote_corpus(url: str, length: int) -> list[str]:
+    response = requests.get(url)
+    response.raise_for_status()
+    corpus = response.text
+    words = [w.strip() for w in corpus.strip().split(" ")]
+    words = [w for w in words if "<" not in w] # make sure nothing looks like special tokens
     words = [w for w in words if len(w) > 0]  # filter out empty strings
     while len(words) < length:
         words += words
@@ -226,9 +222,9 @@ def parse_args() -> argparse.Namespace:
     )
     parser_dump.add_argument(
         "--file",
-        type=Path,
-        default=None,
-        help="File containing prompt to use instead of the default",
+        type=str,
+        default="https://raw.githubusercontent.com/ggml-org/llama.cpp/eaba92c3dcc980ebe753348855d4a5d75c069997/tools/server/README.md",
+        help="File containing prompt to use instead of the default (can also be an URL)",
     )
     parser_dump.add_argument(
         "--pattern",
@@ -259,17 +255,19 @@ def main():
 
     if args.verb == "dump":
         pattern = parse_pattern(args.pattern)
-        input_length = sum(n for _, n in pattern)
-        input_words = generate_input_prompt(input_length)
-        if args.file is not None:
-            with args.file.open("r") as f:
+        required_words = sum(n for _, n in pattern)
+        if args.file.startswith("http"):
+            input_words = get_remote_corpus(args.file, required_words)
+            logger.info(f"Fetched {len(input_words)} words from remote {args.file}")
+        else:
+            with open(args.file, "r") as f:
                 input_words = f.read().strip().split(" ")
-                if input_length < sum(n for _, n in pattern):
+                input_words = [w for w in input_words if len(w) > 0]  # filter out empty strings
+                if len(input_words) < required_words:
                     raise ValueError(
-                        f"Input file has only {input_length} words, but pattern requires at least {input_length} words."
+                        f"Input file has only {len(input_words)} words, but pattern requires at least {required_words} words."
                     )
-                input_length = len(input_words)
-        logger.info(f"Using {input_length} words")
+        logger.info(f"Using {len(input_words)} words")
         dump_logits(args.endpoint, args.output, input_words, pattern, args.api_key)
     elif args.verb == "compare":
         compare_logits(args.input1, args.input2, args.output)

scripts/get-wikitext-2.sh

@@ -1,11 +1,43 @@
-#!/usr/bin/env bash
+#!/bin/sh
+# vim: set ts=4 sw=4 et:
 
-wget https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
-unzip wikitext-2-raw-v1.zip
+ZIP="wikitext-2-raw-v1.zip"
+FILE="wikitext-2-raw/wiki.test.raw"
+URL="https://huggingface.co/datasets/ggml-org/ci/resolve/main/$ZIP"
 
-echo "Usage:"
-echo ""
-echo "  ./llama-perplexity -m model.gguf -f wikitext-2-raw/wiki.test.raw [other params]"
-echo ""
+die() {
+    printf "%s\n" "$@" >&2
+    exit 1
+}
 
-exit 0
+have_cmd() {
+    for cmd; do
+        command -v "$cmd" >/dev/null || return
+    done
+}
+
+dl() {
+    [ -f "$2" ] && return
+    if have_cmd wget; then
+        wget "$1" -O "$2"
+    elif have_cmd curl; then
+        curl -L "$1" -o "$2"
+    else
+        die "Please install wget or curl"
+    fi
+}
+
+have_cmd unzip || die "Please install unzip"
+
+if [ ! -f "$FILE" ]; then
+    dl "$URL" "$ZIP" || exit
+    unzip -o "$ZIP"  || exit
+    rm -f -- "$ZIP"
+fi
+
+cat <<EOF
+Usage:
+
+  llama-perplexity -m model.gguf -f $FILE [other params]
+
+EOF

scripts/sync_vendor.py

@@ -5,7 +5,7 @@ import os
 import sys
 import subprocess
 
-HTTPLIB_VERSION = "refs/tags/v0.34.0"
+HTTPLIB_VERSION = "refs/tags/v0.35.0"
 
 vendor = {
     "https://github.com/nlohmann/json/releases/latest/download/json.hpp":     "vendor/nlohmann/json.hpp",
@@ -14,8 +14,8 @@ vendor = {
     "https://raw.githubusercontent.com/nothings/stb/refs/heads/master/stb_image.h": "vendor/stb/stb_image.h",
 
     # not using latest tag to avoid this issue: https://github.com/ggml-org/llama.cpp/pull/17179#discussion_r2515877926
-    # "https://github.com/mackron/miniaudio/raw/refs/tags/0.11.23/miniaudio.h": "vendor/miniaudio/miniaudio.h",
-    "https://github.com/mackron/miniaudio/raw/669ed3e844524fcd883231b13095baee9f6de304/miniaudio.h": "vendor/miniaudio/miniaudio.h",
+    # "https://github.com/mackron/miniaudio/raw/refs/tags/0.11.24/miniaudio.h": "vendor/miniaudio/miniaudio.h",
+    "https://github.com/mackron/miniaudio/raw/13d161bc8d856ad61ae46b798bbeffc0f49808e8/miniaudio.h": "vendor/miniaudio/miniaudio.h",
 
     f"https://raw.githubusercontent.com/yhirose/cpp-httplib/{HTTPLIB_VERSION}/httplib.h": "httplib.h",
     f"https://raw.githubusercontent.com/yhirose/cpp-httplib/{HTTPLIB_VERSION}/split.py":  "split.py",

src/CMakeLists.txt

@@ -62,6 +62,7 @@ add_library(llama
             models/dream.cpp
             models/ernie4-5-moe.cpp
             models/ernie4-5.cpp
+            models/eurobert.cpp
             models/exaone-moe.cpp
             models/exaone.cpp
             models/exaone4.cpp

src/llama-arch.cpp

@@ -26,6 +26,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_NEO_BERT,         "neo-bert"         },
     { LLM_ARCH_JINA_BERT_V2,     "jina-bert-v2"     },
     { LLM_ARCH_JINA_BERT_V3,     "jina-bert-v3"     },
+    { LLM_ARCH_EUROBERT,         "eurobert"         },
     { LLM_ARCH_BLOOM,            "bloom"            },
     { LLM_ARCH_STABLELM,         "stablelm"         },
     { LLM_ARCH_QWEN,             "qwen"             },
@@ -348,6 +349,7 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
     { LLM_TENSOR_FFN_DOWN_EXP,                           "blk.%d.ffn_down.%d" },
     { LLM_TENSOR_FFN_UP_EXP,                             "blk.%d.ffn_up.%d" },
     { LLM_TENSOR_FFN_GATE_EXPS,                          "blk.%d.ffn_gate_exps" },
+    { LLM_TENSOR_FFN_GATE_UP_EXPS,                       "blk.%d.ffn_gate_up_exps" },
     { LLM_TENSOR_FFN_DOWN_EXPS,                          "blk.%d.ffn_down_exps" },
     { LLM_TENSOR_FFN_UP_EXPS,                            "blk.%d.ffn_up_exps" },
     { LLM_TENSOR_ATTN_POST_NORM,                         "blk.%d.post_attention_norm" },
@@ -819,6 +821,20 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_CLS,
                 LLM_TENSOR_CLS_OUT,
             };
+        case LLM_ARCH_EUROBERT:
+            return {
+                LLM_TENSOR_TOKEN_EMBD,
+                LLM_TENSOR_OUTPUT_NORM,
+                LLM_TENSOR_ATTN_NORM,
+                LLM_TENSOR_ATTN_Q,
+                LLM_TENSOR_ATTN_K,
+                LLM_TENSOR_ATTN_V,
+                LLM_TENSOR_ATTN_OUT,
+                LLM_TENSOR_FFN_NORM,
+                LLM_TENSOR_FFN_GATE,
+                LLM_TENSOR_FFN_UP,
+                LLM_TENSOR_FFN_DOWN,
+            };
         case LLM_ARCH_MODERN_BERT:
             return {
                 LLM_TENSOR_TOKEN_EMBD,
@@ -989,6 +1005,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_FFN_GATE_EXPS,
                 LLM_TENSOR_FFN_DOWN_EXPS,
                 LLM_TENSOR_FFN_UP_EXPS,
+                LLM_TENSOR_FFN_GATE_UP_EXPS,
                 LLM_TENSOR_FFN_GATE_INP_SHEXP,
                 LLM_TENSOR_FFN_GATE_SHEXP,
                 LLM_TENSOR_FFN_DOWN_SHEXP,
@@ -1046,6 +1063,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_FFN_GATE_EXPS,
                 LLM_TENSOR_FFN_DOWN_EXPS,
                 LLM_TENSOR_FFN_UP_EXPS,
+                LLM_TENSOR_FFN_GATE_UP_EXPS,
                 LLM_TENSOR_FFN_GATE_INP_SHEXP,
                 LLM_TENSOR_FFN_GATE_SHEXP,
                 LLM_TENSOR_FFN_DOWN_SHEXP,
@@ -1586,6 +1604,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_FFN_GATE_EXPS,
                 LLM_TENSOR_FFN_DOWN_EXPS,
                 LLM_TENSOR_FFN_UP_EXPS,
+                LLM_TENSOR_FFN_GATE_UP_EXPS,
                 LLM_TENSOR_FFN_GATE_INP_SHEXP,
                 LLM_TENSOR_FFN_GATE_SHEXP,
                 LLM_TENSOR_FFN_DOWN_SHEXP,
@@ -2670,6 +2689,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_FFN_DOWN_EXPS,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
     {LLM_TENSOR_FFN_GATE_EXPS,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
     {LLM_TENSOR_FFN_UP_EXPS,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
+    {LLM_TENSOR_FFN_GATE_UP_EXPS,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
     {LLM_TENSOR_FFN_DOWN_CHEXPS,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
     {LLM_TENSOR_FFN_GATE_CHEXPS,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
     {LLM_TENSOR_FFN_UP_CHEXPS,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},

src/llama-arch.h

@@ -30,6 +30,7 @@ enum llm_arch {
     LLM_ARCH_NEO_BERT,
     LLM_ARCH_JINA_BERT_V2,
     LLM_ARCH_JINA_BERT_V3,
+    LLM_ARCH_EUROBERT,
     LLM_ARCH_BLOOM,
     LLM_ARCH_STABLELM,
     LLM_ARCH_QWEN,
@@ -372,6 +373,7 @@ enum llm_tensor {
     LLM_TENSOR_FFN_DOWN_EXPS, // merged experts
     LLM_TENSOR_FFN_GATE_EXPS,
     LLM_TENSOR_FFN_UP_EXPS,
+    LLM_TENSOR_FFN_GATE_UP_EXPS,
     LLM_TENSOR_FFN_DOWN_SHEXP,
     LLM_TENSOR_FFN_GATE_SHEXP,
     LLM_TENSOR_FFN_UP_SHEXP,

src/llama-graph.cpp

@@ -1165,7 +1165,8 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
                float   w_scale,
          llama_expert_gating_func_type gating_op,
                  int   il,
-         ggml_tensor * probs_in) const {
+         ggml_tensor * probs_in,
+         ggml_tensor * gate_up_exps) const {
     return build_moe_ffn(
         cur,
         gate_inp,  /* gate_inp_b  */ nullptr,
@@ -1181,7 +1182,8 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         w_scale,
         gating_op,
         il,
-        probs_in
+        probs_in,
+        gate_up_exps
     );
 }
 
@@ -1204,7 +1206,9 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
                float   w_scale,
         llama_expert_gating_func_type gating_op,
                  int   il,
-         ggml_tensor * probs_in) const {
+         ggml_tensor * probs_in,
+         ggml_tensor * gate_up_exps,
+         ggml_tensor * gate_up_exps_b) const {
     const int64_t n_embd   = cur->ne[0];
     const int64_t n_tokens = cur->ne[1];
     const bool weight_before_ffn = arch == LLM_ARCH_LLAMA4; // for llama4, we apply the sigmoid-ed weights before the FFN
@@ -1343,26 +1347,48 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         cb(cur, "ffn_moe_weighted", il);
     }
 
-    ggml_tensor * up = build_lora_mm_id(up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
-    cb(up, "ffn_moe_up", il);
-
-    if (up_exps_b) {
-        up = ggml_add_id(ctx0, up, up_exps_b, selected_experts);
-        cb(up, "ffn_moe_up_biased", il);
-    }
-
+    ggml_tensor * up = nullptr;
     ggml_tensor * experts = nullptr;
-    if (gate_exps) {
-        cur = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
+
+    if (gate_up_exps) {
+        // merged gate_up path: one mul_mat_id, then split into gate and up views
+        ggml_tensor * gate_up = build_lora_mm_id(gate_up_exps, cur, selected_experts); // [n_ff*2, n_expert_used, n_tokens]
+        cb(gate_up, "ffn_moe_gate_up", il);
+
+        if (gate_up_exps_b) {
+            gate_up = ggml_add_id(ctx0, gate_up, gate_up_exps_b, selected_experts);
+            cb(gate_up, "ffn_moe_gate_up_biased", il);
+        }
+
+        const int64_t n_ff = gate_up->ne[0] / 2;
+        cur = ggml_view_3d(ctx0, gate_up, n_ff, gate_up->ne[1], gate_up->ne[2], gate_up->nb[1], gate_up->nb[2], 0);
         cb(cur, "ffn_moe_gate", il);
+        up  = ggml_view_3d(ctx0, gate_up, n_ff, gate_up->ne[1], gate_up->ne[2], gate_up->nb[1], gate_up->nb[2], n_ff * gate_up->nb[0]);
+        cb(up, "ffn_moe_up", il);
     } else {
-        cur = up;
+        // separate gate and up path
+        up = build_lora_mm_id(up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
+        cb(up, "ffn_moe_up", il);
+
+        if (up_exps_b) {
+            up = ggml_add_id(ctx0, up, up_exps_b, selected_experts);
+            cb(up, "ffn_moe_up_biased", il);
+        }
+
+        if (gate_exps) {
+            cur = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
+            cb(cur, "ffn_moe_gate", il);
+        } else {
+            cur = up;
+        }
+
+        if (gate_exps_b) {
+            cur = ggml_add_id(ctx0, cur, gate_exps_b, selected_experts);
+            cb(cur, "ffn_moe_gate_biased", il);
+        }
     }
 
-    if (gate_exps_b) {
-        cur = ggml_add_id(ctx0, cur, gate_exps_b, selected_experts);
-        cb(cur, "ffn_moe_gate_biased", il);
-    }
+    const bool has_gate = gate_exps || gate_up_exps;
 
     switch (type_op) {
         case LLM_FFN_SILU:
@@ -1385,7 +1411,9 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
                         break;
                     }
                 }
+            }
 
+            if (has_gate) {
                 cur = ggml_swiglu_split(ctx0, cur, up);
                 cb(cur, "ffn_moe_swiglu", il);
             } else {
@@ -1393,7 +1421,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
                 cb(cur, "ffn_moe_silu", il);
             } break;
         case LLM_FFN_GELU:
-            if (gate_exps) {
+            if (has_gate) {
                 cur = ggml_geglu_split(ctx0, cur, up);
                 cb(cur, "ffn_moe_geglu", il);
             } else {
@@ -1409,7 +1437,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
                 cb(cur, "ffn_moe_swiglu_oai", il);
             } break;
         case LLM_FFN_RELU:
-            if (gate_exps) {
+            if (has_gate) {
                 cur = ggml_reglu_split(ctx0, cur, up);
                 cb(cur, "ffn_moe_reglu", il);
             } else {
@@ -1417,7 +1445,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
                 cb(cur, "ffn_moe_relu", il);
             } break;
         case LLM_FFN_RELU_SQR:
-            if (gate_exps) {
+            if (has_gate) {
                 // TODO: add support for gated squared relu
                 GGML_ABORT("fatal error: gated squared relu not implemented");
             } else {

src/llama-graph.h

@@ -814,7 +814,8 @@ struct llm_graph_context {
                    float   w_scale,
             llama_expert_gating_func_type gating_op,
                      int   il,
-             ggml_tensor * probs_in = nullptr) const;
+             ggml_tensor * probs_in = nullptr,
+             ggml_tensor * gate_up_exps = nullptr) const;
 
     ggml_tensor * build_moe_ffn(
              ggml_tensor * cur,
@@ -835,7 +836,9 @@ struct llm_graph_context {
                    float   w_scale,
             llama_expert_gating_func_type gating_op,
                      int   il,
-             ggml_tensor * probs_in = nullptr) const;
+             ggml_tensor * probs_in = nullptr,
+             ggml_tensor * gate_up_exps = nullptr,
+             ggml_tensor * gate_up_exps_b = nullptr) const;
 
     //
     // inputs

src/llama-kv-cache.cpp

@@ -978,6 +978,9 @@ bool llama_kv_cache::get_can_shift() const {
     if (model.arch == LLM_ARCH_STEP35) {
         return false;
     }
+    if (hparams.n_pos_per_embd() > 1) {
+        return false;
+    }
     return true;
 }
 

src/llama-memory-recurrent.cpp

@@ -163,7 +163,7 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
             const auto & cell = cells[tail_id];
             // partial intersection is invalid if it includes the final pos
             if (0 < p0 && p0 <= cell.pos && p1 > cell.pos) {
-                //printf("[DEBUG] inside `llama_memory_recurrent::seq_rm`: partial intersection is invalid, so returning false\n");
+                //printf("[DEBUG] inside `llama_memory_recurrent::seq_rm`: partial intersection is invalid, so returning false, p0 = %d, cell.pos = %d, p1 = %d\n", p0, cell.pos, p1);
                 return false;
             }
             // invalidate tails which will be cleared

src/llama-model.cpp

@@ -123,6 +123,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_8B_A1B:        return "8B.A1B";
         case LLM_TYPE_16B_A1B:       return "16B.A1B";
         case LLM_TYPE_21B_A3B:       return "21B.A3B";
+        case LLM_TYPE_24B_A2B:       return "24B.A2B";
         case LLM_TYPE_30B_A3B:       return "30B.A3B";
         case LLM_TYPE_31B_A3_5B:     return "31B.A3.5B";
         case LLM_TYPE_35B_A3B:       return "35B.A3B";
@@ -978,6 +979,16 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     type = LLM_TYPE_250M;
                 }
             } break;
+        case LLM_ARCH_EUROBERT:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_ATTENTION_CAUSAL,            hparams.causal_attn);
+                ml.get_key(LLM_KV_POOLING_TYPE,                hparams.pooling_type);
+
+                if (hparams.n_layer == 12) {
+                    type = LLM_TYPE_SMALL;  // 0.2B
+                }
+            } break;
         case LLM_ARCH_BLOOM:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
@@ -2381,7 +2392,11 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     hparams.recurrent_layer_arr[il] = hparams.n_head_kv(il) == 0;
                 }
 
-                type = LLM_TYPE_8B_A1B;
+                switch (hparams.n_layer) {
+                    case 24: type = LLM_TYPE_8B_A1B;  break;
+                    case 40: type = LLM_TYPE_24B_A2B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
             } break;
         case LLM_ARCH_SMALLTHINKER:
             {
@@ -2965,6 +2980,15 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
         // TODO: move to a separate function
         const auto tn = LLM_TN(arch);
+
+        // helper: try merged gate_up_exps first, fall back to separate gate and up
+        auto create_tensor_gate_up_exps = [&](llama_layer & layer, int bid, int64_t n_embd_, int64_t n_ff_, int64_t n_expert_, int flags) {
+            layer.ffn_gate_up_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_UP_EXPS, "weight", bid), {n_embd_, n_ff_ * 2, n_expert_}, TENSOR_NOT_REQUIRED);
+            if (layer.ffn_gate_up_exps == nullptr) {
+                layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", bid), {n_embd_, n_ff_, n_expert_}, flags);
+                layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", bid), {n_embd_, n_ff_, n_expert_}, flags);
+            }
+        };
         switch (arch) {
             case LLM_ARCH_LLAMA:
             case LLM_ARCH_REFACT:
@@ -3565,6 +3589,29 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
                     }
                 } break;
+            case LLM_ARCH_EUROBERT:
+                {
+                    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);
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+                    }
+                } break;
             case LLM_ARCH_JINA_BERT_V2:
                 {
                     tok_embd  = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0); // word_embeddings
@@ -5183,9 +5230,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             }
 
                             // MoE branch
-                            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
                             layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
-                            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+                            create_tensor_gate_up_exps(layer, i, n_embd, n_ff_exp, n_expert, 0);
 
                             // Shared expert branch
                             layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
@@ -7387,9 +7433,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         }
 
                         layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), { n_embd, n_expert }, 0);
-                        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, 0);
                         layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert }, 0);
-                        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), { n_embd, n_ff_exp, n_expert }, 0);
+                        create_tensor_gate_up_exps(layer, i, n_embd, n_ff_exp, n_expert, 0);
 
                         // Shared experts
                         layer.ffn_gate_inp_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", i), { n_embd }, 0);
@@ -7453,9 +7498,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         }
 
                         layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), { n_embd, n_expert }, 0);
-                        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, 0);
                         layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert }, 0);
-                        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), { n_embd, n_ff_exp, n_expert }, 0);
+                        create_tensor_gate_up_exps(layer, i, n_embd, n_ff_exp, n_expert, 0);
 
                         // Shared experts
                         const int64_t n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff;
@@ -8176,6 +8220,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
         case LLM_ARCH_NOMIC_BERT:
         case LLM_ARCH_NOMIC_BERT_MOE:
         case LLM_ARCH_NEO_BERT:
+        case LLM_ARCH_EUROBERT:
         case LLM_ARCH_WAVTOKENIZER_DEC:
         case LLM_ARCH_MODERN_BERT:
         case LLM_ARCH_GEMMA_EMBEDDING:
@@ -8373,6 +8418,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_neo_bert>(*this, params);
             } break;
+        case LLM_ARCH_EUROBERT:
+            {
+                llm = std::make_unique<llm_build_eurobert>(*this, params);
+            } break;
         case LLM_ARCH_BLOOM:
             {
                 llm = std::make_unique<llm_build_bloom>(*this, params);
@@ -8999,6 +9048,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_MODERN_BERT:
         case LLM_ARCH_NOMIC_BERT:
         case LLM_ARCH_NOMIC_BERT_MOE:
+        case LLM_ARCH_EUROBERT:
         case LLM_ARCH_STABLELM:
         case LLM_ARCH_BITNET:
         case LLM_ARCH_QWEN:

src/llama-model.h

@@ -116,6 +116,7 @@ enum llm_type {
     LLM_TYPE_8B_A1B, // lfm2moe
     LLM_TYPE_16B_A1B,
     LLM_TYPE_21B_A3B, // Ernie MoE small
+    LLM_TYPE_24B_A2B, // lfm2moe
     LLM_TYPE_30B_A3B,
     LLM_TYPE_31B_A3_5B,
     LLM_TYPE_35B_A3B, // Qwen3.5
@@ -279,14 +280,16 @@ struct llama_layer {
     struct ggml_tensor * ffn_up_enc   = nullptr;
 
     // ff MoE
-    struct ggml_tensor * ffn_gate_inp    = nullptr;
-    struct ggml_tensor * ffn_gate_exps   = nullptr;
-    struct ggml_tensor * ffn_down_exps   = nullptr;
-    struct ggml_tensor * ffn_up_exps     = nullptr;
-    struct ggml_tensor * ffn_gate_inp_b  = nullptr;
-    struct ggml_tensor * ffn_gate_exps_b = nullptr;
-    struct ggml_tensor * ffn_down_exps_b = nullptr;
-    struct ggml_tensor * ffn_up_exps_b   = nullptr;
+    struct ggml_tensor * ffn_gate_inp      = nullptr;
+    struct ggml_tensor * ffn_gate_exps     = nullptr;
+    struct ggml_tensor * ffn_down_exps     = nullptr;
+    struct ggml_tensor * ffn_up_exps       = nullptr;
+    struct ggml_tensor * ffn_gate_up_exps  = nullptr;
+    struct ggml_tensor * ffn_gate_inp_b    = nullptr;
+    struct ggml_tensor * ffn_gate_exps_b   = nullptr;
+    struct ggml_tensor * ffn_down_exps_b   = nullptr;
+    struct ggml_tensor * ffn_up_exps_b     = nullptr;
+    struct ggml_tensor * ffn_gate_up_exps_b = nullptr;
 
     // ff shared expert (shexp)
     struct ggml_tensor * ffn_gate_inp_shexp = nullptr;

src/llama-vocab.cpp

@@ -1890,7 +1890,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     tokenizer_pre == "falcon-h1" ||
                     tokenizer_pre == "pixtral"  ||
                     tokenizer_pre == "midm-2.0" ||
-                    tokenizer_pre == "lfm2") {
+                    tokenizer_pre == "lfm2"     ||
+                    tokenizer_pre == "jina-v5-nano") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_LLAMA3;
                 ignore_merges = true;
                 add_bos = true;

src/models/deepseek2.cpp

@@ -218,7 +218,9 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
                 LLM_FFN_SILU, hparams.expert_weights_norm,
                 hparams.expert_weights_scale, hparams.expert_weights_scale,
                 (llama_expert_gating_func_type) hparams.expert_gating_func,
-                il);
+                il,
+                nullptr,
+                model.layers[il].ffn_gate_up_exps);
             cb(moe_out, "ffn_moe_out", il);
 
             // FFN shared expert

src/models/eurobert.cpp

@@ -0,0 +1,97 @@
+#include "models.h"
+
+llm_build_eurobert::llm_build_eurobert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    inpL = build_inp_embd(model.tok_embd);
+    cb(inpL, "inp_embd", -1);
+
+    auto * inp_attn = build_attn_inp_no_cache();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * cur = inpL;
+
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+
+        {
+            ggml_tensor * Qcur;
+            ggml_tensor * Kcur;
+            ggml_tensor * Vcur;
+
+            Qcur = build_lora_mm(model.layers[il].wq, cur);
+            Kcur = build_lora_mm(model.layers[il].wk, cur);
+            Vcur = build_lora_mm(model.layers[il].wv, cur);
+
+            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);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, nullptr,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+            cb(cur, "kqv_out", il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        cur = ggml_add(ctx0, cur, inpL);
+
+        ggml_tensor * ffn_inp = cur;
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up, NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_embd", -1);
+    res->t_embd = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/kimi-linear.cpp

@@ -116,6 +116,8 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
         cur = build_norm(inpL, layer.attn_norm, NULL, LLM_NORM_RMS, il);
         cb(cur, "attn_norm", il);
 
+        ggml_build_forward_expand(gf, cur);
+
         // Check layer type by checking which tensors exist
         // KDA layers have ssm_a_log tensor, MLA layers have wkv_a_mqa tensor
         bool is_kda = (layer.ssm_a != nullptr);

src/models/models.h

@@ -424,6 +424,10 @@ struct llm_build_neo_bert : public llm_graph_context {
     llm_build_neo_bert(const llama_model & model, const llm_graph_params & params);
 };
 
+struct llm_build_eurobert : public llm_graph_context {
+    llm_build_eurobert(const llama_model & model, const llm_graph_params & params);
+};
+
 template <bool iswa>
 struct llm_build_olmo2 : public llm_graph_context {
     llm_build_olmo2(const llama_model & model, const llm_graph_params & params);

src/models/qwen35.cpp

@@ -29,6 +29,8 @@ llm_build_qwen35::llm_build_qwen35(const llama_model & model, const llm_graph_pa
         cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
         cb(cur, "attn_norm", il);
 
+        ggml_build_forward_expand(gf, cur);
+
         // Determine layer type and build appropriate attention mechanism
         if (hparams.is_recurrent(il)) {
             // Linear attention layer (gated delta net)
@@ -269,7 +271,6 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
     cb(state_update_target, "state_update_target", il);
 
     ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
-    cb(conv_states_all, "conv_states_updated", il);
 
     ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
     state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);

src/models/qwen35moe.cpp

@@ -29,6 +29,8 @@ llm_build_qwen35moe::llm_build_qwen35moe(const llama_model & model, const llm_gr
         cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
         cb(cur, "attn_norm", il);
 
+        ggml_build_forward_expand(gf, cur);
+
         // Determine layer type and build appropriate attention mechanism
         if (hparams.is_recurrent(il)) {
             // Linear attention layer (gated delta net)
@@ -269,7 +271,6 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
     cb(state_update_target, "state_update_target", il);
 
     ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
-    cb(conv_states_all, "conv_states_updated", il);
 
     ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
     state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);
@@ -379,7 +380,8 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_ffn(ggml_tensor * cur, const int
             model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps,
             nullptr,
             n_expert, n_expert_used, LLM_FFN_SILU,
-            true, false, 0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il);
+            true, false, 0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il,
+            nullptr, model.layers[il].ffn_gate_up_exps);
     cb(moe_out, "ffn_moe_out", il);
 
     // Add shared experts if present - following Qwen3Next reference implementation

src/models/qwen3next.cpp

@@ -21,6 +21,8 @@ llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_gr
         cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
         cb(cur, "attn_norm", il);
 
+        ggml_build_forward_expand(gf, cur);
+
         // Determine layer type and build appropriate attention mechanism
         if (hparams.is_recurrent(il)) {
             // Linear attention layer (gated delta net)
@@ -354,7 +356,6 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     cb(state_update_target, "state_update_target", il);
 
     ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
-    cb(conv_states_all, "conv_states_updated", il);
 
     ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
     state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);
@@ -478,7 +479,8 @@ ggml_tensor * llm_build_qwen3next::build_layer_ffn(ggml_tensor * cur, const int
                 model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps,
                 nullptr,
                 n_expert, n_expert_used, LLM_FFN_SILU,
-                true, false, 0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il);
+                true, false, 0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il,
+                nullptr, model.layers[il].ffn_gate_up_exps);
         cb(moe_out, "ffn_moe_out", il);
 
         // Add shared experts if present - following Qwen3Next reference implementation

tests/CMakeLists.txt

@@ -152,7 +152,7 @@ if (NOT WIN32 OR NOT BUILD_SHARED_LIBS)
     llama_build_and_test(test-grammar-parser.cpp)
     llama_build_and_test(test-grammar-integration.cpp)
     llama_build_and_test(test-llama-grammar.cpp)
-    llama_build_and_test(test-chat.cpp)
+    llama_build_and_test(test-chat.cpp WORKING_DIRECTORY ${PROJECT_SOURCE_DIR})
     # TODO: disabled on loongarch64 because the ggml-ci node lacks Python 3.8
     if (NOT ${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
         llama_build_and_test(test-json-schema-to-grammar.cpp   WORKING_DIRECTORY ${PROJECT_SOURCE_DIR})
@@ -257,6 +257,21 @@ set(LLAMA_TEST_NAME test-mtmd-c-api)
 llama_build_and_test(test-mtmd-c-api.c)
 target_link_libraries(${LLAMA_TEST_NAME} PRIVATE mtmd)
 
+# GGUF model data fetcher library for tests that need real model metadata
+# Only compile when cpp-httplib has SSL support (CPPHTTPLIB_OPENSSL_SUPPORT)
+if (TARGET cpp-httplib)
+    get_target_property(_cpp_httplib_defs cpp-httplib INTERFACE_COMPILE_DEFINITIONS)
+    if (_cpp_httplib_defs MATCHES "CPPHTTPLIB_OPENSSL_SUPPORT")
+        add_library(gguf-model-data STATIC gguf-model-data.cpp)
+        target_link_libraries(gguf-model-data PRIVATE common cpp-httplib)
+        target_include_directories(gguf-model-data PUBLIC ${CMAKE_CURRENT_SOURCE_DIR})
+
+        add_executable(test-gguf-model-data test-gguf-model-data.cpp)
+        target_link_libraries(test-gguf-model-data PRIVATE gguf-model-data common)
+        llama_test(test-gguf-model-data LABEL "model")
+    endif()
+endif()
+
 # dummy executable - not installed
 get_filename_component(TEST_TARGET test-c.c NAME_WE)
 add_executable(${TEST_TARGET} test-c.c)

tests/gguf-model-data.cpp

@@ -0,0 +1,613 @@
+// GGUF binary parser adapted from the huggingface/gguf package.
+// Reference: https://github.com/huggingface/huggingface.js
+
+#include "gguf-model-data.h"
+
+#include "common.h"
+#include "gguf.h"
+
+#include <algorithm>
+#include <cstdio>
+#include <cstring>
+#include <filesystem>
+#include <fstream>
+
+#include "http.h"
+#define JSON_ASSERT GGML_ASSERT
+#include <nlohmann/json.hpp>
+
+// Equivalent of RangeView
+struct gguf_buf_reader {
+    const char * data;
+    size_t       size;
+    size_t       pos;
+
+    gguf_buf_reader(const std::vector<char> & buf) : data(buf.data()), size(buf.size()), pos(0) {}
+
+    bool has_n_bytes(size_t n) const {
+        return pos + n <= size;
+    }
+
+    template <typename T>
+    bool read_val(T & out) {
+        if (!has_n_bytes(sizeof(T))) {
+            return false;
+        }
+        memcpy(&out, data + pos, sizeof(T));
+        pos += sizeof(T);
+        return true;
+    }
+
+    bool read_str(std::string & out) {
+        uint64_t len;
+        if (!read_val(len)) {
+            return false;
+        }
+        if (!has_n_bytes((size_t)len)) {
+            return false;
+        }
+        out.assign(data + pos, (size_t)len);
+        pos += (size_t)len;
+        return true;
+    }
+
+    bool skip(size_t n) {
+        if (!has_n_bytes(n)) {
+            return false;
+        }
+        pos += n;
+        return true;
+    }
+};
+
+static size_t gguf_val_type_size(int32_t vtype) {
+    switch (vtype) {
+        case GGUF_TYPE_UINT8:   return 1;
+        case GGUF_TYPE_INT8:    return 1;
+        case GGUF_TYPE_UINT16:  return 2;
+        case GGUF_TYPE_INT16:   return 2;
+        case GGUF_TYPE_UINT32:  return 4;
+        case GGUF_TYPE_INT32:   return 4;
+        case GGUF_TYPE_FLOAT32: return 4;
+        case GGUF_TYPE_BOOL:    return 1;
+        case GGUF_TYPE_UINT64:  return 8;
+        case GGUF_TYPE_INT64:   return 8;
+        case GGUF_TYPE_FLOAT64: return 8;
+        default:                return 0; // string/array handled separately
+    }
+}
+
+// Equivalent of readMetadataValue(), skips unused values rather than storing
+static bool gguf_skip_value(gguf_buf_reader & r, int32_t vtype) {
+    if (vtype == GGUF_TYPE_STRING) {
+        std::string tmp;
+        return r.read_str(tmp);
+    }
+    if (vtype == GGUF_TYPE_ARRAY) {
+        int32_t elem_type;
+        uint64_t count;
+        if (!r.read_val(elem_type)) {
+            return false;
+        }
+        if (!r.read_val(count)) {
+            return false;
+        }
+        if (elem_type == GGUF_TYPE_STRING) {
+            for (uint64_t i = 0; i < count; i++) {
+                std::string tmp;
+                if (!r.read_str(tmp)) {
+                    return false;
+                }
+            }
+            return true;
+        }
+        if (elem_type == GGUF_TYPE_ARRAY) {
+            // nested arrays - recurse
+            for (uint64_t i = 0; i < count; i++) {
+                if (!gguf_skip_value(r, GGUF_TYPE_ARRAY)) {
+                    return false;
+                }
+            }
+            return true;
+        }
+        size_t elem_sz = gguf_val_type_size(elem_type);
+        if (elem_sz == 0) {
+            return false;
+        }
+        return r.skip((size_t)count * elem_sz);
+    }
+    size_t sz = gguf_val_type_size(vtype);
+    if (sz == 0) {
+        return false;
+    }
+    return r.skip(sz);
+}
+
+static bool gguf_read_uint32_val(gguf_buf_reader & r, int32_t vtype, uint32_t & out) {
+    if (vtype == GGUF_TYPE_UINT8) {
+        uint8_t v;
+        if (!r.read_val(v)) {
+            return false;
+        }
+        out = v;
+        return true;
+    }
+    if (vtype == GGUF_TYPE_INT8) {
+        int8_t v;
+        if (!r.read_val(v)) {
+            return false;
+        }
+        out = (uint32_t)v;
+        return true;
+    }
+    if (vtype == GGUF_TYPE_UINT16) {
+        uint16_t v;
+        if (!r.read_val(v)) {
+            return false;
+        }
+        out = v;
+        return true;
+    }
+    if (vtype == GGUF_TYPE_INT16) {
+        int16_t v;
+        if (!r.read_val(v)) {
+            return false;
+        }
+        out = (uint32_t)v;
+        return true;
+    }
+    if (vtype == GGUF_TYPE_UINT32) {
+        uint32_t v;
+        if (!r.read_val(v)) {
+            return false;
+        }
+        out = v;
+        return true;
+    }
+    if (vtype == GGUF_TYPE_INT32) {
+        int32_t v;
+        if (!r.read_val(v)) {
+            return false;
+        }
+        out = (uint32_t)v;
+        return true;
+    }
+    if (vtype == GGUF_TYPE_UINT64) {
+        uint64_t v;
+        if (!r.read_val(v)) {
+            return false;
+        }
+        out = (uint32_t)v;
+        return true;
+    }
+    if (vtype == GGUF_TYPE_INT64) {
+        int64_t v;
+        if (!r.read_val(v)) {
+            return false;
+        }
+        out = (uint32_t)v;
+        return true;
+    }
+    return false;
+}
+
+// Follows the same header -> KV -> tensor parsing sequence as gguf() huggingface/gguf
+static std::optional<gguf_remote_model> gguf_parse_meta(const std::vector<char> & buf) {
+    gguf_buf_reader r(buf);
+
+    // Header: magic(4) + version(4) + tensor_count(8) + kv_count(8) = 24 bytes minimum
+    uint32_t magic_raw;
+    if (!r.read_val(magic_raw)) {
+        return std::nullopt;
+    }
+    if (memcmp(&magic_raw, "GGUF", 4) != 0) {
+        fprintf(stderr, "gguf_parse_meta: invalid magic\n");
+        return std::nullopt;
+    }
+
+    uint32_t version;
+    if (!r.read_val(version)) {
+        return std::nullopt;
+    }
+    if (version < 2 || version > 3) {
+        fprintf(stderr, "gguf_parse_meta: unsupported version %u\n", version);
+        return std::nullopt;
+    }
+
+    int64_t tensor_count_raw;
+    int64_t kv_count_raw;
+    if (!r.read_val(tensor_count_raw)) {
+        return std::nullopt;
+    }
+    if (!r.read_val(kv_count_raw)) {
+        return std::nullopt;
+    }
+
+    uint64_t tensor_count = (uint64_t)tensor_count_raw;
+    uint64_t kv_count     = (uint64_t)kv_count_raw;
+
+    gguf_remote_model model;
+
+    std::string arch_prefix;
+
+    // Parse KV pairs
+    for (uint64_t i = 0; i < kv_count; i++) {
+        std::string key;
+        if (!r.read_str(key)) {
+            return std::nullopt;
+        }
+
+        int32_t vtype;
+        if (!r.read_val(vtype)) {
+            return std::nullopt;
+        }
+
+        if (key == "general.architecture" && vtype == GGUF_TYPE_STRING) {
+            if (!r.read_str(model.architecture)) {
+                return std::nullopt;
+            }
+            arch_prefix = model.architecture + ".";
+            continue;
+        }
+
+        // Extract split.count for proper handling of split files
+        if (key == "split.count") {
+            uint32_t v;
+            if (!gguf_read_uint32_val(r, vtype, v)) {
+                return std::nullopt;
+            }
+            model.n_split = (uint16_t)v;
+            continue;
+        }
+
+        // Extract split.tensors.count so we can verify we have all tensors
+        if (key == "split.tensors.count") {
+            uint32_t v;
+            if (!gguf_read_uint32_val(r, vtype, v)) {
+                return std::nullopt;
+            }
+            model.n_split_tensors = v;
+            continue;
+        }
+
+        if (!arch_prefix.empty()) {
+            uint32_t * target = nullptr;
+
+            if      (key == arch_prefix + "embedding_length")         { target = &model.n_embd; }
+            else if (key == arch_prefix + "feed_forward_length")      { target = &model.n_ff; }
+            else if (key == arch_prefix + "block_count")              { target = &model.n_layer; }
+            else if (key == arch_prefix + "attention.head_count")     { target = &model.n_head; }
+            else if (key == arch_prefix + "attention.head_count_kv")  { target = &model.n_head_kv; }
+            else if (key == arch_prefix + "expert_count")             { target = &model.n_expert; }
+            else if (key == arch_prefix + "attention.key_length")     { target = &model.n_embd_head_k; }
+            else if (key == arch_prefix + "attention.value_length")   { target = &model.n_embd_head_v; }
+
+            if (target) {
+                if (!gguf_read_uint32_val(r, vtype, *target)) {
+                    return std::nullopt;
+                }
+                continue;
+            }
+        }
+
+        if (!gguf_skip_value(r, vtype)) {
+            return std::nullopt;
+        }
+    }
+
+    // Parse tensor info entries
+    model.tensors.reserve((size_t)tensor_count);
+    for (uint64_t i = 0; i < tensor_count; i++) {
+        gguf_remote_tensor t;
+
+        if (!r.read_str(t.name)) {
+            return std::nullopt;
+        }
+        if (!r.read_val(t.n_dims)) {
+            return std::nullopt;
+        }
+
+        if (t.n_dims > 4) {
+            fprintf(stderr, "gguf_parse_meta: tensor '%s' has %u dims (max 4)\n", t.name.c_str(), t.n_dims);
+            return std::nullopt;
+        }
+
+        for (uint32_t d = 0; d < t.n_dims; d++) {
+            if (!r.read_val(t.ne[d])) {
+                return std::nullopt;
+            }
+        }
+
+        int32_t type_raw;
+        if (!r.read_val(type_raw)) {
+            return std::nullopt;
+        }
+        t.type = (ggml_type)type_raw;
+
+        uint64_t offset;
+        if (!r.read_val(offset)) {
+            return std::nullopt;
+        }
+
+        // Infer n_vocab from token_embd.weight
+        if (t.name == "token_embd.weight") {
+            model.n_vocab = (uint32_t)t.ne[1];
+        }
+
+        model.tensors.push_back(std::move(t));
+    }
+
+    return model;
+}
+
+// cache handling for local download
+static std::string get_default_cache_dir() {
+    return fs_get_cache_directory() + "gguf-headers/";
+}
+
+static std::string sanitize_for_path(const std::string & s) {
+    std::string out = s;
+    for (char & c : out) {
+        if (c == '/' || c == '\\' || c == ':') {
+            c = '_';
+        }
+    }
+    return out;
+}
+
+static bool read_file(const std::string & path, std::vector<char> & out) {
+    std::ifstream f(path, std::ios::binary | std::ios::ate);
+    if (!f.good()) {
+        return false;
+    }
+    auto sz = f.tellg();
+    if (sz <= 0) {
+        return false;
+    }
+    out.resize((size_t)sz);
+    f.seekg(0);
+    f.read(out.data(), sz);
+    return f.good();
+}
+
+static bool write_file(const std::string & path, const std::vector<char> & data) {
+    std::ofstream f(path, std::ios::binary | std::ios::trunc);
+    if (!f.good()) {
+        return false;
+    }
+    f.write(data.data(), (std::streamsize)data.size());
+    return f.good();
+}
+
+// HuggingFace file auto-detection and HTTP download
+static std::pair<long, std::vector<char>> gguf_http_get(
+        const std::string & url,
+        const httplib::Headers & headers = {},
+        int timeout_sec = 60) {
+    try {
+        auto [cli, parts] = common_http_client(url);
+
+        if (timeout_sec > 0) {
+            cli.set_read_timeout(timeout_sec, 0);
+            cli.set_write_timeout(timeout_sec, 0);
+        }
+        cli.set_connection_timeout(30, 0);
+
+        std::vector<char> body;
+        auto res = cli.Get(parts.path, headers,
+            [&](const char * data, size_t len) {
+                body.insert(body.end(), data, data + len);
+                return true;
+            }, nullptr);
+
+        if (!res) {
+            fprintf(stderr, "gguf_fetch: HTTP request failed for %s (error %d)\n",
+                    url.c_str(), (int)res.error());
+            return {-1, {}};
+        }
+        return {res->status, std::move(body)};
+    } catch (const std::exception & e) {
+        fprintf(stderr, "gguf_fetch: HTTP error: %s\n", e.what());
+        return {-1, {}};
+    }
+}
+
+// Find the filename for given repo/quant.
+// For split models, returns the first shard (the one containing "00001-of-")
+// split_prefix is set to the portion before "-00001-of-XXXXX.gguf" when a split file is found
+static std::string detect_gguf_filename(const std::string & repo, const std::string & quant,
+                                        std::string & split_prefix) {
+    split_prefix.clear();
+    std::string api_url = "https://huggingface.co/api/models/" + repo;
+
+    auto [code, body] = gguf_http_get(api_url, {}, 30);
+    if (code != 200 || body.empty()) {
+        fprintf(stderr, "gguf_fetch: failed to query HF API for %s (HTTP %ld)\n", repo.c_str(), code);
+        return "";
+    }
+
+    nlohmann::json j;
+    try {
+        j = nlohmann::json::parse(body.begin(), body.end());
+    } catch (...) {
+        fprintf(stderr, "gguf_fetch: failed to parse HF API response\n");
+        return "";
+    }
+
+    if (!j.contains("siblings") || !j["siblings"].is_array()) {
+        fprintf(stderr, "gguf_fetch: unexpected HF API response format\n");
+        return "";
+    }
+
+    std::vector<std::string> matches;
+    std::string quant_upper = quant;
+    for (char & c : quant_upper) { c = (char)toupper(c); }
+
+    for (const auto & sibling : j["siblings"]) {
+        if (!sibling.contains("rfilename")) { continue; }
+        std::string fname = sibling["rfilename"].get<std::string>();
+        if (fname.size() < 5 || fname.substr(fname.size() - 5) != ".gguf") {
+            continue;
+        }
+
+        std::string fname_upper = fname;
+        for (char & c : fname_upper) { c = (char)toupper(c); }
+        if (fname_upper.find(quant_upper) != std::string::npos) {
+            matches.push_back(fname);
+        }
+    }
+
+    if (matches.empty()) {
+        fprintf(stderr, "gguf_fetch: no .gguf files matching '%s' in %s\n", quant.c_str(), repo.c_str());
+        return "";
+    }
+
+    std::sort(matches.begin(), matches.end());
+
+    // Prefer non-split, non-supplementary file
+    for (const auto & m : matches) {
+        if (m.find("-of-") == std::string::npos && m.find("mmproj") == std::string::npos) {
+            return m;
+        }
+    }
+
+    // Return the first shard (00001-of-) and extract the prefix
+    for (const auto & m : matches) {
+        auto pos = m.find("-00001-of-");
+        if (pos != std::string::npos) {
+            split_prefix = m.substr(0, pos);
+            return m;
+        }
+    }
+
+    return matches[0];
+}
+
+static std::optional<gguf_remote_model> fetch_and_parse(
+        const std::string & repo,
+        const std::string & filename,
+        const std::string & cache_path) {
+    std::string url = "https://huggingface.co/" + repo + "/resolve/main/" + filename;
+
+    // Progressive download inspired by RangeView.fetchChunk()
+    // Start at 2MB, double each time, cap at 64MB
+    size_t chunk_size = 2 * 1024 * 1024;
+    const size_t max_chunk = 64 * 1024 * 1024;
+
+    while (chunk_size <= max_chunk) {
+        fprintf(stderr, "gguf_fetch: downloading %zu bytes from %s\n", chunk_size, filename.c_str());
+
+        char range_buf[64];
+        snprintf(range_buf, sizeof(range_buf), "bytes=0-%zu", chunk_size - 1);
+        httplib::Headers headers = {{"Range", range_buf}};
+
+        auto [code, body] = gguf_http_get(url, headers, 120);
+        if (code != 200 && code != 206) {
+            fprintf(stderr, "gguf_fetch: HTTP %ld fetching %s\n", code, url.c_str());
+            return std::nullopt;
+        }
+
+        if (body.empty()) {
+            fprintf(stderr, "gguf_fetch: empty response\n");
+            return std::nullopt;
+        }
+
+        auto result = gguf_parse_meta(body);
+        if (result.has_value()) {
+            write_file(cache_path, body);
+            return result;
+        }
+
+        if (code == 200) {
+            fprintf(stderr, "gguf_fetch: server returned full response but metadata parse failed\n");
+            return std::nullopt;
+        }
+
+        // Parse failed, try larger chunk
+        chunk_size *= 2;
+    }
+
+    fprintf(stderr, "gguf_fetch: metadata exceeds 64MB, giving up\n");
+    return std::nullopt;
+}
+
+// Try cache first, then fetch and parse a single GGUF shard.
+static std::optional<gguf_remote_model> fetch_or_cached(
+        const std::string & repo,
+        const std::string & filename,
+        const std::string & cdir,
+        const std::string & repo_part) {
+    std::string fname_part = sanitize_for_path(filename);
+    std::string cache_path = cdir + "/" + repo_part + "--" + fname_part + ".partial";
+
+    {
+        std::vector<char> cached;
+        if (std::filesystem::exists(cache_path) && read_file(cache_path, cached)) {
+            auto result = gguf_parse_meta(cached);
+            if (result.has_value()) {
+                fprintf(stderr, "gguf_fetch: loaded from cache: %s\n", cache_path.c_str());
+                return result;
+            }
+        }
+    }
+
+    fs_create_directory_with_parents(cdir);
+    return fetch_and_parse(repo, filename, cache_path);
+}
+
+std::optional<gguf_remote_model> gguf_fetch_model_meta(
+        const std::string & repo,
+        const std::string & quant,
+        const std::string & cache_dir) {
+    std::string cdir = cache_dir.empty() ? get_default_cache_dir() : cache_dir;
+    std::string repo_part = sanitize_for_path(repo);
+
+    std::string split_prefix;
+    std::string filename = detect_gguf_filename(repo, quant, split_prefix);
+    if (filename.empty()) {
+        return std::nullopt;
+    }
+
+    auto model_opt = fetch_or_cached(repo, filename, cdir, repo_part);
+    if (!model_opt.has_value()) {
+        fprintf(stderr, "gguf_fetch: failed to fetch %s\n", filename.c_str());
+        return std::nullopt;
+    }
+
+    auto & model = model_opt.value();
+
+    // If the model is split across multiple files we need to fetch the remaining shards metadata
+    if (model.n_split > 1) {
+        if (split_prefix.empty()) {
+            fprintf(stderr, "gguf_fetch: model reports %u splits but filename has no split pattern\n", model.n_split);
+            return std::nullopt;
+        }
+
+        fprintf(stderr, "gguf_fetch: split model with %u shards, fetching remaining %u...\n",
+                model.n_split, model.n_split - 1);
+
+        for (int i = 2; i <= model.n_split; i++) {
+            char num_buf[6], total_buf[6];
+            snprintf(num_buf,   sizeof(num_buf),   "%05d", i);
+            snprintf(total_buf, sizeof(total_buf), "%05d", (int)model.n_split);
+            std::string shard_name = split_prefix + "-" + num_buf + "-of-" + total_buf + ".gguf";
+
+            auto shard = fetch_or_cached(repo, shard_name, cdir, repo_part);
+            if (!shard.has_value()) {
+                fprintf(stderr, "gguf_fetch: failed to fetch shard %d: %s\n", i, shard_name.c_str());
+                return std::nullopt;
+            }
+
+            model.tensors.insert(model.tensors.end(),
+                std::make_move_iterator(shard->tensors.begin()),
+                std::make_move_iterator(shard->tensors.end()));
+        }
+
+        if (model.n_split_tensors > 0 && model.tensors.size() != model.n_split_tensors) {
+            fprintf(stderr, "gguf_fetch: WARNING: expected %u tensors from split.tensors.count, got %zu\n",
+                    model.n_split_tensors, model.tensors.size());
+        }
+    }
+
+    return model_opt;
+}

tests/gguf-model-data.h

@@ -0,0 +1,42 @@
+#pragma once
+
+#include "ggml.h"
+
+#include <cstdint>
+#include <optional>
+#include <string>
+#include <vector>
+
+struct gguf_remote_tensor {
+    std::string  name;
+    ggml_type    type    = GGML_TYPE_F32;
+    int64_t      ne[4]   = {1, 1, 1, 1};  // dimensions, unused dims = 1
+    uint32_t     n_dims  = 0;
+};
+
+struct gguf_remote_model {
+    // Selected KV metadata
+    std::string architecture;               // general.architecture
+    uint32_t    n_embd          = 0;        // <arch>.embedding_length
+    uint32_t    n_ff            = 0;        // <arch>.feed_forward_length
+    uint32_t    n_vocab         = 0;        // inferred from token_embd.weight ne[1]
+    uint32_t    n_layer         = 0;        // <arch>.block_count
+    uint32_t    n_head          = 0;        // <arch>.attention.head_count
+    uint32_t    n_head_kv       = 0;        // <arch>.attention.head_count_kv
+    uint32_t    n_expert        = 0;        // <arch>.expert_count (0 if absent)
+    uint32_t    n_embd_head_k   = 0;        // <arch>.attention.key_length
+    uint32_t    n_embd_head_v   = 0;        // <arch>.attention.value_length
+    uint16_t    n_split         = 0;        // split.count (0 = not split)
+    uint32_t    n_split_tensors = 0;        // split.tensors.count (0 if not split)
+
+    std::vector<gguf_remote_tensor> tensors;
+};
+
+// Fetch model metadata from HuggingFace with local caching.
+// repo: e.g., "ggml-org/Qwen3-32B-GGUF"
+// quant: e.g., "Q8_0" -- auto-detects filename (including first shard of split models)
+// Returns nullopt if download fails or network is unavailable.
+std::optional<gguf_remote_model> gguf_fetch_model_meta(
+    const std::string & repo,
+    const std::string & quant = "Q8_0",
+    const std::string & cache_dir = "");  // empty = default

tests/test-gguf-model-data.cpp

@@ -0,0 +1,121 @@
+#include "gguf-model-data.h"
+
+#include <cstdio>
+
+#define TEST_ASSERT(cond, msg) \
+    do { \
+        if (!(cond)) { \
+            fprintf(stderr, "FAIL: %s (line %d): %s\n", #cond, __LINE__, msg); \
+            return 1; \
+        } \
+    } while (0)
+
+int main() {
+    fprintf(stderr, "=== test-gguf-model-data ===\n");
+
+    // Fetch Qwen3-0.6B Q8_0 metadata
+    auto result = gguf_fetch_model_meta("ggml-org/Qwen3-0.6B-GGUF", "Q8_0");
+
+    if (!result.has_value()) {
+        fprintf(stderr, "SKIP: could not fetch model metadata (no network or HTTP disabled)\n");
+        return 0;
+    }
+
+    const auto & model = result.value();
+
+    fprintf(stderr, "Architecture:  %s\n", model.architecture.c_str());
+    fprintf(stderr, "n_embd:        %u\n", model.n_embd);
+    fprintf(stderr, "n_ff:          %u\n", model.n_ff);
+    fprintf(stderr, "n_vocab:       %u\n", model.n_vocab);
+    fprintf(stderr, "n_layer:       %u\n", model.n_layer);
+    fprintf(stderr, "n_head:        %u\n", model.n_head);
+    fprintf(stderr, "n_head_kv:     %u\n", model.n_head_kv);
+    fprintf(stderr, "n_expert:      %u\n", model.n_expert);
+    fprintf(stderr, "n_embd_head_k: %u\n", model.n_embd_head_k);
+    fprintf(stderr, "n_embd_head_v: %u\n", model.n_embd_head_v);
+    fprintf(stderr, "tensors:       %zu\n", model.tensors.size());
+
+    // Verify architecture
+    TEST_ASSERT(model.architecture == "qwen3", "expected architecture 'qwen3'");
+
+    // Verify key dimensions (Qwen3-0.6B)
+    TEST_ASSERT(model.n_layer == 28, "expected n_layer == 28");
+    TEST_ASSERT(model.n_embd == 1024, "expected n_embd == 1024");
+    TEST_ASSERT(model.n_head == 16, "expected n_head == 16");
+    TEST_ASSERT(model.n_head_kv == 8, "expected n_head_kv == 8");
+    TEST_ASSERT(model.n_expert == 0, "expected n_expert == 0 (not MoE)");
+    TEST_ASSERT(model.n_vocab == 151936, "expected n_vocab == 151936");
+
+    // Verify tensor count
+    TEST_ASSERT(model.tensors.size() == 311, "expected tensor count == 311");
+
+    // Verify known tensor names exist
+    bool found_attn_q = false;
+    bool found_token_embd = false;
+    bool found_output_norm = false;
+    for (const auto & t : model.tensors) {
+        if (t.name == "blk.0.attn_q.weight") {
+            found_attn_q = true;
+        }
+        if (t.name == "token_embd.weight") {
+            found_token_embd = true;
+        }
+        if (t.name == "output_norm.weight") {
+            found_output_norm = true;
+        }
+    }
+    TEST_ASSERT(found_attn_q, "expected tensor 'blk.0.attn_q.weight'");
+    TEST_ASSERT(found_token_embd, "expected tensor 'token_embd.weight'");
+    TEST_ASSERT(found_output_norm, "expected tensor 'output_norm.weight'");
+
+    // Verify token_embd.weight shape
+    for (const auto & t : model.tensors) {
+        if (t.name == "token_embd.weight") {
+            TEST_ASSERT(t.ne[0] == 1024, "expected token_embd.weight ne[0] == 1024");
+            TEST_ASSERT(t.n_dims == 2, "expected token_embd.weight to be 2D");
+            break;
+        }
+    }
+
+    // Test that second call uses cache (just call again, it should work)
+    auto result2 = gguf_fetch_model_meta("ggml-org/Qwen3-0.6B-GGUF", "Q8_0");
+    TEST_ASSERT(result2.has_value(), "cached fetch should succeed");
+    TEST_ASSERT(result2->tensors.size() == model.tensors.size(), "cached result should match");
+
+    // Test a split MoE model without specifying quant (should default to Q8_0)
+    auto result3 = gguf_fetch_model_meta("ggml-org/GLM-4.6V-GGUF");
+    if (!result3.has_value()) {
+        fprintf(stderr, "SKIP: could not fetch GLM-4.6V metadata (no network?)\n");
+        return 0;
+    }
+    const auto & model3 = result3.value();
+
+    fprintf(stderr, "Architecture:  %s\n", model3.architecture.c_str());
+    fprintf(stderr, "n_embd:        %u\n", model3.n_embd);
+    fprintf(stderr, "n_ff:          %u\n", model3.n_ff);
+    fprintf(stderr, "n_vocab:       %u\n", model3.n_vocab);
+    fprintf(stderr, "n_layer:       %u\n", model3.n_layer);
+    fprintf(stderr, "n_head:        %u\n", model3.n_head);
+    fprintf(stderr, "n_head_kv:     %u\n", model3.n_head_kv);
+    fprintf(stderr, "n_expert:      %u\n", model3.n_expert);
+    fprintf(stderr, "n_embd_head_k: %u\n", model3.n_embd_head_k);
+    fprintf(stderr, "n_embd_head_v: %u\n", model3.n_embd_head_v);
+    fprintf(stderr, "tensors:       %zu\n", model3.tensors.size());
+
+    // Verify architecture
+    TEST_ASSERT(model3.architecture == "glm4moe", "expected architecture 'glm4moe'");
+
+    // Verify key dimensions (GLM-4.6V)
+    TEST_ASSERT(model3.n_layer == 46, "expected n_layer == 46");
+    TEST_ASSERT(model3.n_embd == 4096, "expected n_embd == 4096");
+    TEST_ASSERT(model3.n_head == 96, "expected n_head == 96");
+    TEST_ASSERT(model3.n_head_kv == 8, "expected n_head_kv == 8");
+    TEST_ASSERT(model3.n_expert == 128, "expected n_expert == 128 (MoE)");
+    TEST_ASSERT(model3.n_vocab == 151552, "expected n_vocab == 151552");
+
+    // Verify tensor count
+    TEST_ASSERT(model3.tensors.size() == 780, "expected tensor count == 780");
+
+    fprintf(stderr, "=== ALL TESTS PASSED ===\n");
+    return 0;
+}

tests/test-gguf.cpp

@@ -48,6 +48,7 @@ enum handcrafted_file_type {
     HANDCRAFTED_DATA_NOT_ENOUGH_DATA       =  10 + offset_has_data,
     HANDCRAFTED_DATA_BAD_ALIGN             =  15 + offset_has_data,
     HANDCRAFTED_DATA_INCONSISTENT_ALIGN    =  20 + offset_has_data,
+    HANDCRAFTED_DATA_MEM_SIZE_OVERFLOW     =  30 + offset_has_data,
     HANDCRAFTED_DATA_SUCCESS               = 800 + offset_has_data,
     HANDCRAFTED_DATA_CUSTOM_ALIGN          = 810 + offset_has_data,
 };
@@ -84,6 +85,7 @@ static std::string handcrafted_file_type_name(const enum handcrafted_file_type h
         case HANDCRAFTED_DATA_NOT_ENOUGH_DATA:       return "DATA_NOT_ENOUGH_DATA";
         case HANDCRAFTED_DATA_BAD_ALIGN:             return "DATA_BAD_ALIGN";
         case HANDCRAFTED_DATA_INCONSISTENT_ALIGN:    return "DATA_INCONSISTENT_ALIGN";
+        case HANDCRAFTED_DATA_MEM_SIZE_OVERFLOW:     return "DATA_MEM_SIZE_OVERFLOW";
         case HANDCRAFTED_DATA_SUCCESS:               return "DATA_SUCCESS";
         case HANDCRAFTED_DATA_CUSTOM_ALIGN:          return "DATA_CUSTOM_ALIGN";
     }
@@ -196,6 +198,13 @@ static FILE * get_handcrafted_file(const unsigned int seed, const enum handcraft
         tensor_configs = get_tensor_configs(rng);
     }
 
+    if (hft == HANDCRAFTED_DATA_MEM_SIZE_OVERFLOW) {
+        tensor_configs.resize(2);
+
+        tensor_configs[0] = { GGML_TYPE_I8, { 0x7FFFFFFFFFFFFFC0, 1, 1, 1 } };
+        tensor_configs[1] = { GGML_TYPE_I8, { 0x7FFFFFFFFFFFFFC0, 1, 1, 1 } };
+    }
+
     if (hft == HANDCRAFTED_HEADER_BAD_N_TENSORS) {
         const uint64_t n_tensors = -1;
         helper_write(file, n_tensors);
@@ -397,7 +406,8 @@ static FILE * get_handcrafted_file(const unsigned int seed, const enum handcraft
         for (uint32_t i = 1; i < n_dims; ++i) {
             ne *= shape[i];
         }
-        offset += GGML_PAD(ggml_row_size(type, ne), alignment);
+
+        offset += GGML_PAD(ggml_row_size(type, ne), (uint64_t) alignment);
     }
 
     while (ftell(file) % alignment != 0) {
@@ -411,6 +421,9 @@ static FILE * get_handcrafted_file(const unsigned int seed, const enum handcraft
         if (hft == HANDCRAFTED_DATA_NOT_ENOUGH_DATA) {
             nbytes -= 1;
         }
+        if (hft == HANDCRAFTED_DATA_MEM_SIZE_OVERFLOW) {
+            nbytes = 32;
+        }
         for (uint64_t i = 0; i < nbytes; ++i) {
             const uint8_t random_byte = i % 256;
             helper_write(file, random_byte);
@@ -704,6 +717,7 @@ static std::pair<int, int> test_handcrafted_file(const unsigned int seed) {
         HANDCRAFTED_DATA_NOT_ENOUGH_DATA,
         HANDCRAFTED_DATA_BAD_ALIGN,
         HANDCRAFTED_DATA_INCONSISTENT_ALIGN,
+        HANDCRAFTED_DATA_MEM_SIZE_OVERFLOW,
         HANDCRAFTED_DATA_SUCCESS,
         HANDCRAFTED_DATA_CUSTOM_ALIGN,
     };

tests/test-tokenizer-0.sh

@@ -13,7 +13,12 @@ fi
 name=$1
 input=$2
 
-make -j tests/test-tokenizer-0
+# Build using CMake if binary doesn't exist
+if [ ! -f ./build/bin/test-tokenizer-0 ]; then
+    printf "Building test-tokenizer-0 with CMake...\n"
+    cmake -B build -DLLAMA_BUILD_TESTS=ON
+    cmake --build build --target test-tokenizer-0 -j
+fi
 
 printf "Testing %s on %s ...\n" $name $input
 
@@ -23,7 +28,7 @@ printf "Tokenizing using (py)  Python AutoTokenizer ...\n"
 python3 ./tests/test-tokenizer-0.py ./models/tokenizers/$name --fname-tok $input > /tmp/test-tokenizer-0-$name-py.log 2>&1
 
 printf "Tokenizing using (cpp) llama.cpp ...\n"
-./tests/test-tokenizer-0 ./models/ggml-vocab-$name.gguf $input > /tmp/test-tokenizer-0-$name-cpp.log 2>&1
+./build/bin/test-tokenizer-0 ./models/ggml-vocab-$name.gguf $input > /tmp/test-tokenizer-0-$name-cpp.log 2>&1
 
 cat /tmp/test-tokenizer-0-$name-py.log | grep "tokenized in"
 cat /tmp/test-tokenizer-0-$name-cpp.log | grep "tokenized in"

tools/cli/README.md

@@ -57,8 +57,8 @@
 | `-dt, --defrag-thold N` | KV cache defragmentation threshold (DEPRECATED)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
 | `-np, --parallel N` | number of parallel sequences to decode (default: 1)<br/>(env: LLAMA_ARG_N_PARALLEL) |
 | `--mlock` | force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
-| `--mmap, --no-mmap` | whether to memory-map model. Explicitly enabling mmap disables direct-io. (if mmap disabled, slower load but may reduce pageouts if not using mlock) (default: enabled)<br/>(env: LLAMA_ARG_MMAP) |
-| `-dio, --direct-io, -ndio, --no-direct-io` | use DirectIO if available. Takes precedence over --mmap (default: enabled)<br/>(env: LLAMA_ARG_DIO) |
+| `--mmap, --no-mmap` | whether to memory-map model. (if mmap disabled, slower load but may reduce pageouts if not using mlock) (default: enabled)<br/>(env: LLAMA_ARG_MMAP) |
+| `-dio, --direct-io, -ndio, --no-direct-io` | use DirectIO if available. (default: disabled)<br/>(env: LLAMA_ARG_DIO) |
 | `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggml-org/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
 | `-dev, --device <dev1,dev2,..>` | comma-separated list of devices to use for offloading (none = don't offload)<br/>use --list-devices to see a list of available devices<br/>(env: LLAMA_ARG_DEVICE) |
 | `--list-devices` | print list of available devices and exit |
@@ -109,14 +109,14 @@
 | `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) |
 | `--sampler-seq, --sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: edskypmxt) |
 | `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) |
-| `--temp N` | temperature (default: 0.80) |
+| `--temp, --temperature N` | temperature (default: 0.80) |
 | `--top-k N` | top-k sampling (default: 40, 0 = disabled)<br/>(env: LLAMA_ARG_TOP_K) |
 | `--top-p N` | top-p sampling (default: 0.95, 1.0 = disabled) |
 | `--min-p N` | min-p sampling (default: 0.05, 0.0 = disabled) |
-| `--top-nsigma N` | top-n-sigma sampling (default: -1.00, -1.0 = disabled) |
+| `--top-nsigma, --top-n-sigma N` | top-n-sigma sampling (default: -1.00, -1.0 = disabled) |
 | `--xtc-probability N` | xtc probability (default: 0.00, 0.0 = disabled) |
 | `--xtc-threshold N` | xtc threshold (default: 0.10, 1.0 = disabled) |
-| `--typical N` | locally typical sampling, parameter p (default: 1.00, 1.0 = disabled) |
+| `--typical, --typical-p N` | locally typical sampling, parameter p (default: 1.00, 1.0 = disabled) |
 | `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) |
 | `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.00, 1.0 = disabled) |
 | `--presence-penalty N` | repeat alpha presence penalty (default: 0.00, 0.0 = disabled) |

tools/completion/README.md

@@ -140,8 +140,8 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
 | `-dt, --defrag-thold N` | KV cache defragmentation threshold (DEPRECATED)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
 | `-np, --parallel N` | number of parallel sequences to decode (default: 1)<br/>(env: LLAMA_ARG_N_PARALLEL) |
 | `--mlock` | force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
-| `--mmap, --no-mmap` | whether to memory-map model. Explicitly enabling mmap disables direct-io. (if mmap disabled, slower load but may reduce pageouts if not using mlock) (default: enabled)<br/>(env: LLAMA_ARG_MMAP) |
-| `-dio, --direct-io, -ndio, --no-direct-io` | use DirectIO if available. Takes precedence over --mmap (default: enabled)<br/>(env: LLAMA_ARG_DIO) |
+| `--mmap, --no-mmap` | whether to memory-map model. (if mmap disabled, slower load but may reduce pageouts if not using mlock) (default: enabled)<br/>(env: LLAMA_ARG_MMAP) |
+| `-dio, --direct-io, -ndio, --no-direct-io` | use DirectIO if available. (default: disabled)<br/>(env: LLAMA_ARG_DIO) |
 | `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggml-org/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
 | `-dev, --device <dev1,dev2,..>` | comma-separated list of devices to use for offloading (none = don't offload)<br/>use --list-devices to see a list of available devices<br/>(env: LLAMA_ARG_DEVICE) |
 | `--list-devices` | print list of available devices and exit |
@@ -192,14 +192,14 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
 | `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) |
 | `--sampler-seq, --sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: edskypmxt) |
 | `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) |
-| `--temp N` | temperature (default: 0.80) |
+| `--temp, --temperature N` | temperature (default: 0.80) |
 | `--top-k N` | top-k sampling (default: 40, 0 = disabled)<br/>(env: LLAMA_ARG_TOP_K) |
 | `--top-p N` | top-p sampling (default: 0.95, 1.0 = disabled) |
 | `--min-p N` | min-p sampling (default: 0.05, 0.0 = disabled) |
-| `--top-nsigma N` | top-n-sigma sampling (default: -1.00, -1.0 = disabled) |
+| `--top-nsigma, --top-n-sigma N` | top-n-sigma sampling (default: -1.00, -1.0 = disabled) |
 | `--xtc-probability N` | xtc probability (default: 0.00, 0.0 = disabled) |
 | `--xtc-threshold N` | xtc threshold (default: 0.10, 1.0 = disabled) |
-| `--typical N` | locally typical sampling, parameter p (default: 1.00, 1.0 = disabled) |
+| `--typical, --typical-p N` | locally typical sampling, parameter p (default: 1.00, 1.0 = disabled) |
 | `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) |
 | `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.00, 1.0 = disabled) |
 | `--presence-penalty N` | repeat alpha presence penalty (default: 0.00, 0.0 = disabled) |

tools/imatrix/imatrix.cpp

@@ -912,7 +912,9 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params, c
 
     const bool add_bos = llama_vocab_get_add_bos(vocab);
 
-    GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
+    if (llama_pooling_type(ctx) != LLAMA_POOLING_TYPE_LAST) {
+        GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
+    }
 
     auto tim1 = std::chrono::high_resolution_clock::now();
     LOG_INF("%s: tokenizing the input ..\n", __func__);