opencl: flash attention improvement (#25069)

* opencl: rework FA kernel for f16 and f32

* opencl: flash-attention prefill prepass kernels

- flash_attn_kv_pad_f16    pads the tail KV tile to a BLOCK_N multiple
- flash_attn_mask_pad_f16  pads the matching mask tile
- flash_attn_blk_f16       classifies each KV tile per query block as
                           fully masked / mixed / fully unmasked, so
                           the main kernel can skip fully-masked tiles
                           and the mask lookup for fully-unmasked ones

* opencl: FA kernels for q4_0 and q8_0

* opencl: `set_rows` for f32 to q8_0/q4_0

* opencl: dequant kernels for q4_0 and q8_0

* opencl: add FA tile tuning table with override

* opencl: wire host side for FA

* opencl: q4_0 MoE tensors are also SOA'ed

* opencl: cosmetic fix

* opencl: refactor, also clarify some code paths in comments

* opencl: fix inifity for `-cl-finite-math-only`

---------

Co-authored-by: Li He <lih@qti.qualcomm.com>

.devops/cann.Dockerfile

@@ -145,7 +145,7 @@ ENTRYPOINT ["/app/tools.sh"]
 # ==============================================================================
 FROM base AS light
 
-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app
 
 ENTRYPOINT [ "/app/llama-cli" ]
 
@@ -156,7 +156,7 @@ FROM base AS server
 
 ENV LLAMA_ARG_HOST=0.0.0.0
 
-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app
 
 HEALTHCHECK --interval=5m CMD [ "curl", "-f", "http://localhost:8080/health" ]
 

.devops/cpu.Dockerfile

@@ -104,7 +104,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app
 
 WORKDIR /app
 
@@ -115,7 +115,7 @@ FROM base AS server
 
 ENV LLAMA_ARG_HOST=0.0.0.0
 
-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app
 
 WORKDIR /app
 

.devops/cuda.Dockerfile

@@ -113,7 +113,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app
 
 WORKDIR /app
 
@@ -124,7 +124,7 @@ FROM base AS server
 
 ENV LLAMA_ARG_HOST=0.0.0.0
 
-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app
 
 WORKDIR /app
 

.devops/intel.Dockerfile

@@ -141,7 +141,7 @@ ENTRYPOINT ["/app/tools.sh"]
 FROM base AS light
 
 COPY --from=build /app/lib/ /app
-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app
 
 WORKDIR /app
 
@@ -153,7 +153,7 @@ FROM base AS server
 ENV LLAMA_ARG_HOST=0.0.0.0
 
 COPY --from=build /app/lib/ /app
-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app
 
 WORKDIR /app
 

.devops/musa.Dockerfile

@@ -115,7 +115,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app
 
 WORKDIR /app
 
@@ -126,7 +126,7 @@ FROM base AS server
 
 ENV LLAMA_ARG_HOST=0.0.0.0
 
-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app
 
 WORKDIR /app
 

.devops/openvino.Dockerfile

@@ -1,12 +1,12 @@
-ARG OPENVINO_VERSION_MAJOR=2026.2
-ARG OPENVINO_VERSION_FULL=2026.2.0.21903.52ddc073857
+ARG OPENVINO_VERSION_MAJOR=2026.2.1
+ARG OPENVINO_VERSION_FULL=2026.2.1.21919.ede283a88e3
 ARG UBUNTU_VERSION=24.04
 
 # Intel GPU driver versions. https://github.com/intel/compute-runtime/releases
-ARG IGC_VERSION=v2.34.4
-ARG IGC_VERSION_FULL=2_2.34.4+21428
-ARG COMPUTE_RUNTIME_VERSION=26.18.38308.1
-ARG COMPUTE_RUNTIME_VERSION_FULL=26.18.38308.1-0
+ARG IGC_VERSION=v2.36.3
+ARG IGC_VERSION_FULL=2_2.36.3+21719
+ARG COMPUTE_RUNTIME_VERSION=26.22.38646.4
+ARG COMPUTE_RUNTIME_VERSION_FULL=26.22.38646.4-0
 ARG IGDGMM_VERSION=22.10.0
 
 # Intel NPU driver versions. https://github.com/intel/linux-npu-driver/releases
@@ -214,7 +214,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app/
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app/
 
 WORKDIR /app
 
@@ -225,7 +225,7 @@ FROM base AS server
 
 ENV LLAMA_ARG_HOST=0.0.0.0
 
-COPY --from=build /app/full/llama-server /app/
+COPY --from=build /app/full/llama /app/full/llama-server /app/
 
 WORKDIR /app
 

.devops/rocm.Dockerfile

@@ -127,7 +127,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app
 
 WORKDIR /app
 
@@ -138,7 +138,7 @@ FROM base AS server
 
 ENV LLAMA_ARG_HOST=0.0.0.0
 
-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app
 
 WORKDIR /app
 

.devops/s390x.Dockerfile

@@ -124,7 +124,7 @@ WORKDIR /llama.cpp/bin
 
 # Copy llama.cpp binaries and libraries
 COPY --from=collector /llama.cpp/bin/*.so /llama.cpp/bin
-COPY --from=collector /llama.cpp/bin/llama-cli /llama.cpp/bin/llama-completion /llama.cpp/bin
+COPY --from=collector /llama.cpp/bin/llama /llama.cpp/bin/llama-cli /llama.cpp/bin/llama-completion /llama.cpp/bin
 
 ENTRYPOINT [ "/llama.cpp/bin/llama-cli" ]
 
@@ -138,7 +138,7 @@ WORKDIR /llama.cpp/bin
 
 # Copy llama.cpp binaries and libraries
 COPY --from=collector /llama.cpp/bin/*.so /llama.cpp/bin
-COPY --from=collector /llama.cpp/bin/llama-server /llama.cpp/bin
+COPY --from=collector /llama.cpp/bin/llama /llama.cpp/bin/llama-server /llama.cpp/bin
 
 EXPOSE 8080
 

.devops/vulkan.Dockerfile

@@ -107,7 +107,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app
 
 WORKDIR /app
 
@@ -118,7 +118,7 @@ FROM base AS server
 
 ENV LLAMA_ARG_HOST=0.0.0.0
 
-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app
 
 WORKDIR /app
 

.devops/zendnn.Dockerfile

@@ -97,7 +97,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app
 
 WORKDIR /app
 
@@ -108,7 +108,7 @@ FROM base AS server
 
 ENV LLAMA_ARG_HOST=0.0.0.0
 
-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app
 
 WORKDIR /app
 

.github/labeler.yml

@@ -35,8 +35,20 @@ AMD ZenDNN:
 documentation:
     - changed-files:
         - any-glob-to-any-file:
+            - "**/*.md"
             - docs/**
             - media/**
+examples:
+    - all:
+        - changed-files:
+            - any-glob-to-any-file:
+                - app/**
+                - examples/**
+                - tools/**
+            - all-globs-to-all-files:
+                - '!tools/server/**'
+                - '!tools/mtmd/**'
+                - '!tools/ui/**'
 testing:
     - changed-files:
         - any-glob-to-any-file:
@@ -47,28 +59,12 @@ build:
             - cmake/**
             - CMakeLists.txt
             - CMakePresets.json
-examples:
-    - changed-files:
-        - any-glob-to-any-file:
-            - examples/**
-            - tools/**
 devops:
     - changed-files:
         - any-glob-to-any-file:
             - .devops/**
             - .github/**
             - ci/**
-python:
-    - changed-files:
-        - any-glob-to-any-file:
-            - "**/*.py"
-            - requirements/**
-            - gguf-py/**
-            - .flake8
-script:
-    - changed-files:
-        - any-glob-to-any-file:
-            - scripts/**
 android:
     - changed-files:
         - any-glob-to-any-file:
@@ -81,9 +77,20 @@ server:
     - changed-files:
         - any-glob-to-any-file:
             - tools/server/**
-
-
-
+mtmd:
+    - changed-files:
+        - any-glob-to-any-file:
+            - tools/mtmd/**
+conversion:
+    - changed-files:
+        - any-glob-to-any-file:
+            - conversion/**
+            - convert_*.py
+            - gguf-py/**
+vendor:
+    - changed-files:
+        - any-glob-to-any-file:
+            - vendor/**
 ggml:
     - changed-files:
         - any-glob-to-any-file:

.github/workflows/build-cache.yml

@@ -68,8 +68,8 @@ jobs:
 
     env:
       # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2"
-      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
 
     steps:
       - name: Clone
@@ -96,8 +96,8 @@ jobs:
 
     env:
       # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2"
-      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
 
     steps:
       - name: Clone

.github/workflows/build-openvino.yml

@@ -39,8 +39,8 @@ jobs:
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2"
-      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
 
     steps:
       - name: Clone
@@ -96,8 +96,8 @@ jobs:
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2"
-      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
 
     steps:
       - name: Clone

.github/workflows/build-self-hosted.yml

@@ -266,8 +266,8 @@ jobs:
 
     env:
       # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2"
-      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
 
     steps:
       - name: Clone

.github/workflows/release.yml

@@ -446,8 +446,8 @@ jobs:
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2"
-      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
 
     steps:
       - name: Set OpenVINO version output
@@ -506,8 +506,11 @@ jobs:
           cmake -B build/ReleaseOV -G Ninja \
             -DCMAKE_BUILD_TYPE=Release \
             -DGGML_OPENVINO=ON \
-            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }}
-          cmake --build build/ReleaseOV --config Release -j $(nproc)
+            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
+            ${{ env.CMAKE_ARGS }}
+          cmake --build build/ReleaseOV --config Release --parallel
 
       - name: ccache-clear
         uses: ./.github/actions/ccache-clear
@@ -521,8 +524,26 @@ jobs:
       - name: Pack artifacts
         id: pack_artifacts
         run: |
-          cp LICENSE ./build/ReleaseOV/bin/
-          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz --transform "s,^\.,llama-${{ steps.tag.outputs.name }}," -C ./build/ReleaseOV/bin .
+          dest=./build/ReleaseOV/bin
+          OPENVINO_ROOT=./openvino_toolkit
+          ov_lib="$OPENVINO_ROOT/runtime/lib/intel64"
+
+          # Bundle OpenVINO runtime libs + TBB. Binaries built with RPATH=$ORIGIN
+          # load these siblings without setupvars.sh / LD_LIBRARY_PATH.
+          cp -P "$ov_lib"/libopenvino.so* \
+                "$ov_lib"/libopenvino_c.so* \
+                "$ov_lib"/libopenvino_*_plugin.so \
+                "$ov_lib"/libopenvino_intel_npu_compiler*.so \
+                "$OPENVINO_ROOT"/runtime/3rdparty/tbb/lib/*.so* \
+                "$dest"
+          cp -P /usr/lib/x86_64-linux-gnu/libOpenCL.so.1* "$dest" 2>/dev/null || true
+          cp "$ov_lib"/cache.json "$dest" 2>/dev/null || true
+
+          # OpenVINO licensing
+          cp -r "$OPENVINO_ROOT"/docs/licensing "$dest"/openvino-licensing
+
+          cp LICENSE "$dest"
+          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz --transform "s,^\.,llama-${{ steps.tag.outputs.name }}," -C "$dest" .
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v6
@@ -531,6 +552,9 @@ jobs:
           name: llama-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz
 
   windows-openvino:
+    needs: [check-release]
+    if: ${{ needs.check-release.outputs.should_release == 'true' }}
+
     runs-on: windows-2022
 
     outputs:
@@ -538,8 +562,8 @@ jobs:
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2"
-      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
 
     steps:
       - name: Set OpenVINO version output
@@ -607,7 +631,9 @@ jobs:
             -A x64 ^
             -DCMAKE_BUILD_TYPE=Release ^
             -DGGML_OPENVINO=ON ^
-            -DCMAKE_TOOLCHAIN_FILE=C:\vcpkg\scripts\buildsystems\vcpkg.cmake
+            -DLLAMA_BUILD_BORINGSSL=ON ^
+            -DCMAKE_TOOLCHAIN_FILE=C:\vcpkg\scripts\buildsystems\vcpkg.cmake ^
+            ${{ env.CMAKE_ARGS }}
 
           cmake --build build\ReleaseOV --config Release -- /m
 
@@ -624,8 +650,29 @@ jobs:
         id: pack_artifacts
         shell: powershell
         run: |
-          Copy-Item LICENSE .\build\ReleaseOV\bin\
-          7z a -snl llama-${{ steps.tag.outputs.name }}-bin-win-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.zip .\build\ReleaseOV\bin\*
+          # Locate the extracted OpenVINO toolkit root (same pattern as the Build step).
+          $OPENVINO_ROOT = (Get-ChildItem -Directory openvino_toolkit | Select-Object -First 1).FullName
+          if (-not $OPENVINO_ROOT) {
+            Write-Error "OpenVINO toolkit folder not found under .\openvino_toolkit"
+            exit 1
+          }
+
+          $dest = ".\build\ReleaseOV\bin\Release"
+
+          $ovBin = Join-Path $OPENVINO_ROOT 'runtime\bin\intel64\Release'
+          Copy-Item -Path (Join-Path $ovBin '*.dll')       -Destination $dest -Force
+          Copy-Item -Path (Join-Path $ovBin 'cache.json')  -Destination $dest -Force
+
+          $tbbBin = Join-Path $OPENVINO_ROOT 'runtime\3rdparty\tbb\bin'
+          Copy-Item -Path (Join-Path $tbbBin 'tbb*.dll') -Destination $dest -Force
+
+          # OpenVINO licensing
+          $licensingDest = Join-Path $dest 'openvino-licensing'
+          New-Item -ItemType Directory -Force -Path $licensingDest | Out-Null
+          Copy-Item -Path (Join-Path $OPENVINO_ROOT 'docs\licensing\*') -Destination $licensingDest -Recurse -Force
+
+          Copy-Item LICENSE $dest
+          7z a -snl llama-${{ steps.tag.outputs.name }}-bin-win-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.zip $dest\*
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v6

CMakeLists.txt

@@ -222,6 +222,16 @@ if (LLAMA_BUILD_APP)
     add_subdirectory(app)
 endif()
 
+# Standalone libmtmd build without pulling in the rest of the tools/ tree.
+# Useful when packaging just the mtmd library for language bindings (e.g. an
+# Apple XCFramework, or a WASM build). When the full tools build is enabled,
+# mtmd is already built by the tools/ subdirectory above; this hook only fires
+# when LLAMA_BUILD_TOOLS is OFF to avoid double-adding the target.
+option(LLAMA_BUILD_MTMD "llama: build tools/mtmd library standalone" OFF)
+if (LLAMA_BUILD_MTMD AND NOT (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TOOLS))
+    add_subdirectory(tools/mtmd)
+endif()
+
 #
 # install
 #

SECURITY.md

@@ -80,7 +80,7 @@ To protect sensitive data from potential leaks or unauthorized access, it is cru
 ### Untrusted environments or networks
 
 If you can't run your models in a secure and isolated environment or if it must be exposed to an untrusted network, make sure to take the following security precautions:
-* Do not use the RPC backend, [rpc-server](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) and [llama-server](https://github.com/ggml-org/llama.cpp/tree/master/tools/server) functionality (see https://github.com/ggml-org/llama.cpp/pull/13061).
+* Do not use the RPC backend, [ggml-rpc-server](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) and [llama-server](https://github.com/ggml-org/llama.cpp/tree/master/tools/server) functionality (see https://github.com/ggml-org/llama.cpp/pull/13061).
 * Confirm the hash of any downloaded artifact (e.g. pre-trained model weights) matches a known-good value.
 * Encrypt your data if sending it over the network.
 

app/CMakeLists.txt

@@ -1,6 +1,6 @@
 set(TARGET llama-app)
 
-add_executable(${TARGET} llama.cpp)
+add_executable(${TARGET} llama.cpp download.cpp)
 set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama)
 
 target_link_libraries(${TARGET} PRIVATE

app/download.cpp

@@ -0,0 +1,71 @@
+#include "arg.h"
+#include "common.h"
+#include "download.h"
+#include "log.h"
+
+#include <cstdio>
+#include <filesystem>
+
+static void print_usage(int /*argc*/, char ** argv) {
+    printf(
+        "\nexamples:\n"
+        "  %s -hf ggml-org/gemma-3-4b-it-qat-GGUF\n"
+        "  %s -hf ggml-org/gemma-3-4b-it-qat-GGUF:Q4_K_M\n"
+        "  %s -hf ggml-org/models -hff model.gguf\n"
+        "  %s -mu https://example.com/model.gguf -m model.gguf\n"
+        "\n",
+        argv[0], argv[0], argv[0], argv[0]
+    );
+}
+
+int llama_download(int argc, char ** argv);
+
+int llama_download(int argc, char ** argv) {
+    common_init();
+
+    common_params params;
+    params.verbosity = LOG_LEVEL_ERROR;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_DOWNLOAD, print_usage)) {
+        return 1;
+    }
+
+    const bool has_source = !params.model.hf_repo.empty() || !params.model.url.empty() ||
+                            !params.model.path.empty()    || !params.model.docker_repo.empty();
+    if (!has_source) {
+        fprintf(stderr, "error: no model source specified (use --hf-repo, --model-url, --model or --docker-repo)\n");
+        return 1;
+    }
+
+    try {
+        common_models_handler handler = common_models_handler_init(params, LLAMA_EXAMPLE_DOWNLOAD);
+        common_models_handler_apply(handler, params);
+    } catch (const std::exception & e) {
+        fprintf(stderr, "error: %s\n", e.what());
+        return 1;
+    }
+
+    if (!params.models_preset.empty()) {
+        // -hf pointed at a preset repo: print the preset path and stop
+        printf("%s\n", params.models_preset.c_str());
+        return 0;
+    }
+    if (params.model.path.empty()) {
+        fprintf(stderr, "error: model download failed\n");
+        return 1;
+    }
+    if (!std::filesystem::exists(params.model.path)) {
+        fprintf(stderr, "error: model file does not exist: %s\n", params.model.path.c_str());
+        return 1;
+    }
+
+    printf("%s\n", params.model.path.c_str());
+    if (!params.mmproj.path.empty()) {
+        printf("%s\n", params.mmproj.path.c_str());
+    }
+    if (!params.speculative.draft.mparams.path.empty()) {
+        printf("%s\n", params.speculative.draft.mparams.path.c_str());
+    }
+
+    return 0;
+}

app/llama.cpp

@@ -19,6 +19,7 @@ int llama_batched_bench(int argc, char ** argv);
 int llama_fit_params(int argc, char ** argv);
 int llama_quantize(int argc, char ** argv);
 int llama_perplexity(int argc, char ** argv);
+int llama_download(int argc, char ** argv);
 
 // Self-update is only supported for binaries built with llama-install.sh
 static int llama_update(int argc, char ** argv) {
@@ -49,6 +50,7 @@ struct command {
     std::vector<std::string> aliases;
     bool hidden;
     int (*func)(int, char **);
+    bool flags = false; // allow --name
 };
 
 #ifdef LLAMA_INSTALL_BUILD
@@ -61,15 +63,16 @@ static const command cmds[] = {
     {"serve",         "HTTP API server",                                    {"server"},   false,         llama_server       },
     {"cli",           "Command-line interactive interface",                 {"client"},   false,         llama_cli          },
     {"update",        "Update llama to the latest release",                 {},           UPDATE_HIDDEN, llama_update       },
+    {"download",      "Download a model",                                   {"get"},      false,         llama_download     },
     {"completion",    "Text completion",                                    {"complete"}, true,          llama_completion   },
     {"bench",         "Benchmark prompt processing and text generation",    {},           true,          llama_bench        },
     {"batched-bench", "Benchmark batched decoding performance",             {},           true,          llama_batched_bench},
     {"fit-params",    "Compute parameters to fit a model in device memory", {},           true,          llama_fit_params   },
     {"quantize",      "Quantize a model",                                   {},           true,          llama_quantize     },
     {"perplexity",    "Compute model perplexity and KL divergence",         {},           true,          llama_perplexity   },
-    {"version",       "Show version",                                       {},           false,         version            },
-    {"licenses",      "Show third-party licenses",                          {"credits"},  false,         licenses           },
-    {"help",          "Show available commands",                            {},           false,         help               },
+    {"version",       "Show version",                                       {},           false,         version,           true },
+    {"licenses",      "Show third-party licenses",                          {"credits"},  false,         licenses,          true },
+    {"help",          "Show available commands",                            {},           false,         help,              true },
 };
 
 #undef UPDATE_HIDDEN
@@ -106,7 +109,10 @@ static int help(int argc, char ** argv) {
     return 0;
 }
 
-static bool matches(const std::string & arg, const command & cmd) {
+static bool matches(std::string arg, const command & cmd) {
+    if (cmd.flags && arg.size() > 2 && arg[0] == '-' && arg[1] == '-') {
+        arg.erase(0, 2);
+    }
     if (arg == cmd.name) {
         return true;
     }

build-xcframework.sh

@@ -13,6 +13,7 @@ LLAMA_BUILD_EXAMPLES=OFF
 LLAMA_BUILD_TOOLS=OFF
 LLAMA_BUILD_TESTS=OFF
 LLAMA_BUILD_SERVER=OFF
+LLAMA_BUILD_MTMD=ON
 GGML_METAL=ON
 GGML_METAL_EMBED_LIBRARY=ON
 GGML_BLAS_DEFAULT=ON
@@ -39,6 +40,7 @@ COMMON_CMAKE_ARGS=(
     -DLLAMA_BUILD_TOOLS=${LLAMA_BUILD_TOOLS}
     -DLLAMA_BUILD_TESTS=${LLAMA_BUILD_TESTS}
     -DLLAMA_BUILD_SERVER=${LLAMA_BUILD_SERVER}
+    -DLLAMA_BUILD_MTMD=${LLAMA_BUILD_MTMD}
     -DGGML_METAL_EMBED_LIBRARY=${GGML_METAL_EMBED_LIBRARY}
     -DGGML_BLAS_DEFAULT=${GGML_BLAS_DEFAULT}
     -DGGML_METAL=${GGML_METAL}
@@ -126,6 +128,8 @@ setup_framework_structure() {
     cp ggml/include/ggml-cpu.h     ${header_path}
     cp ggml/include/ggml-blas.h    ${header_path}
     cp ggml/include/gguf.h         ${header_path}
+    cp tools/mtmd/mtmd.h           ${header_path}
+    cp tools/mtmd/mtmd-helper.h    ${header_path}
 
     # Create module map (common for all platforms)
     cat > ${module_path}module.modulemap << EOF
@@ -247,6 +251,7 @@ combine_static_libraries() {
         "${base_dir}/${build_dir}/ggml/src/${release_dir}/libggml-cpu.a"
         "${base_dir}/${build_dir}/ggml/src/ggml-metal/${release_dir}/libggml-metal.a"
         "${base_dir}/${build_dir}/ggml/src/ggml-blas/${release_dir}/libggml-blas.a"
+        "${base_dir}/${build_dir}/tools/mtmd/${release_dir}/libmtmd.a"
     )
 
     # Create temporary directory for processing
@@ -410,6 +415,7 @@ cmake -B build-ios-sim -G Xcode \
     -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
     -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
     -DLLAMA_OPENSSL=OFF \
+    -DMTMD_VIDEO=OFF \
     -S .
 cmake --build build-ios-sim --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
 
@@ -424,6 +430,7 @@ cmake -B build-ios-device -G Xcode \
     -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
     -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
     -DLLAMA_OPENSSL=OFF \
+    -DMTMD_VIDEO=OFF \
     -S .
 cmake --build build-ios-device --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
 
@@ -450,6 +457,7 @@ cmake -B build-visionos -G Xcode \
     -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
     -DLLAMA_OPENSSL=OFF \
     -DLLAMA_BUILD_SERVER=OFF \
+    -DMTMD_VIDEO=OFF \
     -S .
 cmake --build build-visionos --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
 
@@ -465,6 +473,7 @@ cmake -B build-visionos-sim -G Xcode \
     -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
     -DLLAMA_OPENSSL=OFF \
     -DLLAMA_BUILD_SERVER=OFF \
+    -DMTMD_VIDEO=OFF \
     -S .
 cmake --build build-visionos-sim --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
 
@@ -481,6 +490,7 @@ cmake -B build-tvos-sim -G Xcode \
     -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
     -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
     -DLLAMA_OPENSSL=OFF \
+    -DMTMD_VIDEO=OFF \
     -S .
 cmake --build build-tvos-sim --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
 
@@ -496,6 +506,7 @@ cmake -B build-tvos-device -G Xcode \
     -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
     -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
     -DLLAMA_OPENSSL=OFF \
+    -DMTMD_VIDEO=OFF \
     -S .
 cmake --build build-tvos-device --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
 

common/arg.cpp

@@ -297,60 +297,6 @@ struct handle_model_result {
     std::string preset_path;
 };
 
-static handle_model_result common_params_handle_model(struct common_params_model & model,
-                                                      const common_download_opts & opts) {
-    handle_model_result result;
-
-    // TODO @ngxson : refactor this into a new common_model_download_context
-
-    if (!model.docker_repo.empty()) {
-        model.path = common_docker_resolve_model(model.docker_repo);
-    } else if (!model.hf_repo.empty()) {
-        // If -m was used with -hf, treat the model "path" as the hf_file to download
-        if (model.hf_file.empty() && !model.path.empty()) {
-            model.hf_file = model.path;
-            model.path = "";
-        }
-        common_download_opts hf_opts = opts;
-        auto download_result = common_download_model(model, hf_opts);
-
-        if (!download_result.preset_path.empty()) {
-            result.found_preset = true;
-            result.preset_path = download_result.preset_path;
-            return result; // skip everything else if preset.ini is used
-        }
-
-        if (download_result.model_path.empty()) {
-            throw std::runtime_error("failed to download model from Hugging Face");
-        }
-
-        model.path = download_result.model_path;
-
-        if (!download_result.mmproj_path.empty()) {
-            result.found_mmproj = true;
-            result.mmproj.path  = download_result.mmproj_path;
-        }
-
-        if (!download_result.mtp_path.empty()) {
-            result.found_mtp = true;
-            result.mtp.path  = download_result.mtp_path;
-        }
-    } else if (!model.url.empty()) {
-        if (model.path.empty()) {
-            auto f = string_split<std::string>(model.url, '#').front();
-            f = string_split<std::string>(f, '?').front();
-            model.path = fs_get_cache_file(string_split<std::string>(f, '/').back());
-        }
-
-        auto download_result = common_download_model(model, opts);
-        if (download_result.model_path.empty()) {
-            throw std::runtime_error("failed to download model from " + model.url);
-        }
-    }
-
-    return result;
-}
-
 const std::vector<ggml_type> kv_cache_types = {
     GGML_TYPE_F32,
     GGML_TYPE_F16,
@@ -395,77 +341,241 @@ static bool parse_bool_value(const std::string & value) {
 }
 
 //
-// CLI argument parsing functions
+// common_models_handler
 //
 
-bool common_params_handle_models(common_params & params, llama_example curr_ex, const common_params_handle_models_params & handle_params) {
-    const bool spec_type_draft_mtp = std::find(params.speculative.types.begin(),
-                                         params.speculative.types.end(),
-                                         COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
+static std::string get_default_local_path(const std::string & url) {
+    auto f = string_split<std::string>(url, '#').front();
+    f = string_split<std::string>(f, '?').front();
+    return fs_get_cache_file(string_split<std::string>(f, '/').back());
+}
 
+common_models_handler common_models_handler_init(const common_params & params, llama_example curr_ex) {
+    common_download_hf_plan plan;
+    common_download_hf_plan plan_spec;
+    common_download_hf_plan plan_voc;
     common_download_opts opts;
+
+    const bool spec_type_draft_mtp = std::find(params.speculative.types.begin(),
+                                        params.speculative.types.end(),
+                                        COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
+
+    // only download mmproj if the current example is using it
+    bool use_mmproj = false;
+    for (const auto & ex : mmproj_examples) {
+        if (curr_ex == ex) {
+            use_mmproj = true;
+            break;
+        }
+    }
+
     opts.bearer_token    = params.hf_token;
     opts.offline         = params.offline;
-    opts.skip_download   = params.skip_download;
     opts.download_mtp    = spec_type_draft_mtp;
-    opts.download_mmproj = !params.no_mmproj && params.mmproj.path.empty() && params.mmproj.url.empty();
-    opts.preset_only     = handle_params.preset_only;
+    opts.download_mmproj = use_mmproj && !params.no_mmproj
+                        && params.mmproj.path.empty() && params.mmproj.url.empty();
 
-    if (handle_params.callback) {
-        opts.callback = handle_params.callback;
+    if (!params.model.hf_repo.empty()) {
+        plan = common_download_get_hf_plan(params.model, opts);
     }
 
-    // sub-models (draft, mmproj, vocoder) are explicitly specified by the user,
-    // so we should not auto-discover mtp/mmproj siblings for them
-    common_download_opts sub_opts = opts;
-    sub_opts.download_mtp    = false;
-    sub_opts.download_mmproj = false;
+    if (!params.speculative.draft.mparams.hf_repo.empty()) {
+        plan_spec = common_download_get_hf_plan(params.speculative.draft.mparams, opts);
+    }
 
-    try {
-        auto res = common_params_handle_model(params.model, opts);
-        if (res.found_preset) {
-            if (!params.models_preset.empty()) {
-                throw std::invalid_argument("cannot use both --models-preset and -hf with a preset.ini file");
+    if (!params.vocoder.model.hf_repo.empty()) {
+        plan_voc = common_download_get_hf_plan(params.vocoder.model, opts);
+    }
+
+    return common_models_handler{plan, plan_spec, plan_voc, opts};
+}
+
+bool common_models_handler_is_preset_repo(const common_models_handler & handler) {
+    return !handler.plan.preset.url.empty();
+}
+
+static std::vector<common_download_task> build_url_tasks(const common_params_model & model, common_download_opts opts) {
+    auto parts = common_download_get_all_parts(model.url);
+    std::vector<common_download_task> tasks;
+
+    // single-part: download straight to model.path if the user gave one (-m), else the cache default
+    if (parts.size() == 1) {
+        common_download_task task;
+        task.url        = parts[0];
+        task.local_path = model.path.empty() ? get_default_local_path(parts[0]) : model.path;
+        task.opts       = opts;
+        tasks.push_back(std::move(task));
+        return tasks;
+    }
+
+    // multi-part: place each part under the user's -m directory (if given), else the cache default
+    std::string base_dir;
+    if (!model.path.empty()) {
+        auto pos = model.path.rfind('/');
+        base_dir = pos == std::string::npos ? std::string(".") : model.path.substr(0, pos);
+    }
+
+    for (const auto & part : parts) {
+        common_download_task task;
+        task.url  = part;
+        task.opts = opts;
+
+        std::string local = get_default_local_path(part);
+        if (!base_dir.empty()) {
+            auto pos = local.rfind('/');
+            std::string name = pos == std::string::npos ? local : local.substr(pos + 1);
+            local = base_dir + "/" + name;
+        }
+        task.local_path = local;
+        tasks.push_back(std::move(task));
+    }
+    return tasks;
+}
+
+void common_models_handler_apply(common_models_handler & handler, common_params & params, common_download_callback * callback) {
+    std::vector<common_download_task> tasks;
+
+    auto & plan      = handler.plan;
+    auto & plan_spec = handler.plan_spec;
+    auto & plan_voc  = handler.plan_voc;
+
+    auto opts = handler.opts; // copy
+    opts.callback = callback;
+
+    // handle plain "url" if needed
+    auto handle_url = [&](common_params_model & model) {
+        if (!model.url.empty()) {
+            if (model.path.empty()) {
+                model.path = get_default_local_path(model.url);
             }
+        }
+    };
+    handle_url(params.model);
+    handle_url(params.mmproj);
+    handle_url(params.vocoder.model);
+    handle_url(params.speculative.draft.mparams);
+
+    // optionally, if docker repo is set, resolve it
+    if (!params.model.docker_repo.empty()) {
+        params.model.url  = common_docker_resolve_model(params.model.docker_repo);
+        params.model.path = get_default_local_path(params.model.url);
+    }
+
+    // handle plain "url" tasks (non-hf)
+    if (!params.model.url.empty()) {
+        auto url_tasks = build_url_tasks(params.model, opts);
+        // the first part is what gets loaded, so point params.model.path at it
+        if (!url_tasks.empty()) {
+            std::string first_path = url_tasks.front().local_path;
+            url_tasks.front().on_done = [&]() { params.model.path = first_path; };
+        }
+        for (auto & task : url_tasks) {
+            tasks.push_back(std::move(task));
+        }
+    }
+    if (!params.mmproj.url.empty()) {
+        common_download_task task;
+        task.url        = params.mmproj.url;
+        task.local_path = params.mmproj.path;
+        task.opts       = opts;
+        tasks.push_back(task);
+    }
+    if (!params.vocoder.model.url.empty()) {
+        common_download_task task;
+        task.url        = params.vocoder.model.url;
+        task.local_path = params.vocoder.model.path;
+        task.opts       = opts;
+        tasks.push_back(task);
+    }
+    if (!params.speculative.draft.mparams.url.empty()) {
+        common_download_task task;
+        task.url        = params.speculative.draft.mparams.url;
+        task.local_path = params.speculative.draft.mparams.path;
+        task.opts       = opts;
+        tasks.push_back(task);
+    }
+
+    // handle hf_plan tasks
+    auto add_tasks = [&opts, &tasks](const hf_cache::hf_files & model_files, common_params_model & model) {
+        for (size_t i = 0; i < model_files.size(); ++i) {
+            auto & model_file = model_files[i];
+            bool is_first = (i == 0);
+            tasks.emplace_back(model_file, opts, [&, is_first]() {
+                if (is_first) {
+                    // only use first part as model path
+                    model.path = hf_cache::finalize_file(model_file);
+                } else {
+                    hf_cache::finalize_file(model_file);
+                }
+            });
+        }
+    };
+    if (!plan.model_files.empty()) {
+        add_tasks(plan.model_files, params.model);
+    }
+    if (!plan.mmproj.local_path.empty()) {
+        tasks.emplace_back(plan.mmproj, opts, [&]() {
+            params.mmproj.path = hf_cache::finalize_file(plan.mmproj);
+        });
+    }
+    if (!plan.mtp.local_path.empty()) {
+        tasks.emplace_back(plan.mtp, opts, [&]() {
+            // only fall back to the discovered MTP head when no draft was explicitly provided
+            if (params.speculative.draft.mparams.empty()) {
+                params.speculative.draft.mparams.path = hf_cache::finalize_file(plan.mtp);
+            } else {
+                hf_cache::finalize_file(plan.mtp);
+            }
+        });
+    }
+    if (!plan.preset.local_path.empty()) {
+        tasks.emplace_back(plan.preset, opts, [&]() {
             // if HF repo is a preset repo, we simply run server in router mode with the preset.ini file
             params.models_preset_hf = params.model.hf_repo; // only for showing a warning
-            params.models_preset    = res.preset_path;
+            params.models_preset    = hf_cache::finalize_file(plan.preset);
             params.model = common_params_model{}; // make sure to clear model, so server starts in router mode
-            return true;
-        }
+        });
+    }
 
-        if (params.no_mmproj) {
-            params.mmproj = {};
-        } else if (res.found_mmproj && params.mmproj.path.empty() && params.mmproj.url.empty()) {
-            // optionally, handle mmproj model when -hf is specified
-            params.mmproj = res.mmproj;
-        }
-        // only download mmproj if the current example is using it
-        for (const auto & ex : mmproj_examples) {
-            if (curr_ex == ex) {
-                common_params_handle_model(params.mmproj, sub_opts);
-                break;
+    // handle plan_spec (e.g. --spec-draft-hf)
+    if (!plan_spec.model_files.empty()) {
+        add_tasks(plan_spec.model_files, params.speculative.draft.mparams);
+    }
+
+    // handle vocoder plan (e.g. --hf-repo-v)
+    if (!plan_voc.model_files.empty()) {
+        add_tasks(plan_voc.model_files, params.vocoder.model);
+    }
+
+    // run all tasks in parallel
+    if (!params.offline) {
+        // if duplicated files are found, only download once (but still call on_done for each task)
+        std::unordered_map<std::string, common_download_task *> unique_tasks;
+        for (auto & task : tasks) {
+            auto it = unique_tasks.find(task.local_path);
+            if (it == unique_tasks.end()) {
+                unique_tasks[task.local_path] = &task;
             }
         }
-
-        // when --spec-type mtp is set and no draft model was provided explicitly,
-        // fall back to the MTP head discovered alongside the -hf model
-        if (spec_type_draft_mtp && res.found_mtp &&
-            params.speculative.draft.mparams.path.empty() &&
-            params.speculative.draft.mparams.hf_repo.empty() &&
-            params.speculative.draft.mparams.url.empty()) {
-            params.speculative.draft.mparams.path = res.mtp.path;
+        std::vector<common_download_task> unique_tasks_vec;
+        for (auto & pair : unique_tasks) {
+            unique_tasks_vec.push_back(*pair.second);
+        }
+        common_download_run_tasks(unique_tasks_vec);
+    }
+
+    // download successful, update params with the downloaded paths
+    for (const auto & task : tasks) {
+        if (task.on_done) {
+            task.on_done();
         }
-        common_params_handle_model(params.speculative.draft.mparams, sub_opts);
-        common_params_handle_model(params.vocoder.model,             sub_opts);
-        return true;
-    } catch (const common_skip_download_exception &) {
-        return false;
-    } catch (const std::exception &) {
-        throw;
     }
 }
 
+//
+// CLI argument parsing functions
+//
+
 static bool common_params_parse_ex(int argc, char ** argv, common_params_context & ctx_arg) {
     common_params & params = ctx_arg.params;
 
@@ -594,12 +704,15 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
     const bool skip_model_download =
         // server will call common_params_handle_models() later, so we skip it here
         ctx_arg.ex == LLAMA_EXAMPLE_SERVER ||
+        // download calls common_params_handle_models() itself and prints the paths
+        ctx_arg.ex == LLAMA_EXAMPLE_DOWNLOAD ||
         // export_graph_ops loads only metadata
         ctx_arg.ex == LLAMA_EXAMPLE_EXPORT_GRAPH_OPS;
 
     if (!skip_model_download) {
         // handle model and download
-        common_params_handle_models(params, ctx_arg.ex, {});
+        common_models_handler handler = common_models_handler_init(params, ctx_arg.ex);
+        common_models_handler_apply(handler, params);
 
         // model is required (except for server)
         // TODO @ngxson : maybe show a list of available models in CLI in this case
@@ -671,15 +784,19 @@ static void common_params_print_usage(common_params_context & ctx_arg) {
             common_options.push_back(&opt);
         }
     }
-    printf("----- common params -----\n\n");
-    print_options(common_options);
-    printf("\n\n----- sampling params -----\n\n");
-    print_options(sampling_options);
-    printf("\n\n----- speculative params -----\n\n");
-    print_options(spec_options);
-    // TODO: maybe convert enum llama_example to string
-    printf("\n\n----- example-specific params -----\n\n");
-    print_options(specific_options);
+    bool first = true;
+    auto print_section = [&](const char * header, std::vector<common_arg *> & options) {
+        if (options.empty()) {
+            return;
+        }
+        printf("%s----- %s -----\n\n", first ? "" : "\n\n", header);
+        first = false;
+        print_options(options);
+    };
+    print_section("common params",           common_options);
+    print_section("sampling params",         sampling_options);
+    print_section("speculative params",      spec_options);
+    print_section("example-specific params", specific_options);
 }
 
 static void common_params_print_completion(common_params_context & ctx_arg) {
@@ -1079,7 +1196,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
      * - if both {LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_*,} are set, we will prioritize the LLAMA_EXAMPLE_* matching current example
      */
     auto add_opt = [&](common_arg arg) {
-        if ((arg.in_example(ex) || arg.in_example(LLAMA_EXAMPLE_COMMON)) && !arg.is_exclude(ex)) {
+        // download only exposes the handful of args explicitly tagged for it
+        const bool inherit_common = ex != LLAMA_EXAMPLE_DOWNLOAD;
+        if ((arg.in_example(ex) || (inherit_common && arg.in_example(LLAMA_EXAMPLE_COMMON))) && !arg.is_exclude(ex)) {
             ctx_arg.options.push_back(std::move(arg));
         }
     };
@@ -1090,7 +1209,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.usage = true;
         }
-    ));
+    ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD}));
     add_opt(common_arg(
         {"--version"},
         "show version and build info",
@@ -2212,7 +2331,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, bool value) {
             params.no_mmproj = !value;
         }
-    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_MMPROJ_AUTO"));
+    ).set_examples({LLAMA_EXAMPLE_MTMD, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI, LLAMA_EXAMPLE_DOWNLOAD}).set_env("LLAMA_ARG_MMPROJ_AUTO"));
     add_opt(common_arg(
         {"--mmproj-offload"},
         {"--no-mmproj-offload"},
@@ -2611,14 +2730,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.model.path = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}).set_env("LLAMA_ARG_MODEL"));
+    ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA, LLAMA_EXAMPLE_DOWNLOAD}).set_env("LLAMA_ARG_MODEL"));
     add_opt(common_arg(
         {"-mu", "--model-url"}, "MODEL_URL",
         "model download url (default: unused)",
         [](common_params & params, const std::string & value) {
             params.model.url = value;
         }
-    ).set_env("LLAMA_ARG_MODEL_URL"));
+    ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD}).set_env("LLAMA_ARG_MODEL_URL"));
     add_opt(common_arg(
         { "-dr", "--docker-repo" }, "[<repo>/]<model>[:quant]",
         "Docker Hub model repository. repo is optional, default to ai/. quant is optional, default to :latest.\n"
@@ -2627,7 +2746,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.model.docker_repo = value;
         }
-    ).set_env("LLAMA_ARG_DOCKER_REPO"));
+    ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD}).set_env("LLAMA_ARG_DOCKER_REPO"));
     add_opt(common_arg(
         {"-hf", "-hfr", "--hf-repo"}, "<user>/<model>[:quant]",
         "Hugging Face model repository; quant is optional, case-insensitive, default to Q4_K_M, or falls back to the first file in the repo if Q4_K_M doesn't exist.\n"
@@ -2637,14 +2756,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.model.hf_repo = value;
         }
-    ).set_env("LLAMA_ARG_HF_REPO"));
+    ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD}).set_env("LLAMA_ARG_HF_REPO"));
     add_opt(common_arg(
         {"-hff", "--hf-file"}, "FILE",
         "Hugging Face model file. If specified, it will override the quant in --hf-repo (default: unused)",
         [](common_params & params, const std::string & value) {
             params.model.hf_file = value;
         }
-    ).set_env("LLAMA_ARG_HF_FILE"));
+    ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD}).set_env("LLAMA_ARG_HF_FILE"));
     add_opt(common_arg(
         {"-hfv", "-hfrv", "--hf-repo-v"}, "<user>/<model>[:quant]",
         "Hugging Face model repository for the vocoder model (default: unused)",
@@ -2665,7 +2784,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.hf_token = value;
         }
-    ).set_env("HF_TOKEN"));
+    ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD}).set_env("HF_TOKEN"));
+    add_opt(common_arg(
+        {"--mtp"},
+        "also download the multi-token prediction (MTP) head, if available (default: unused)",
+        [](common_params & params) {
+            params.speculative.types.push_back(COMMON_SPECULATIVE_TYPE_DRAFT_MTP);
+        }
+    ).set_examples({LLAMA_EXAMPLE_DOWNLOAD}));
     add_opt(common_arg(
         {"--context-file"}, "FNAME",
         "file to load context from (use comma-separated values to specify multiple files)",
@@ -3622,6 +3748,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         "draft model for speculative decoding (default: unused)",
         [](common_params & params, const std::string & value) {
             params.speculative.draft.mparams.path = value;
+            params.speculative.draft.mparams.hf_file = value; // will be used if --spec-draft-hf is set
         }
     ).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_DRAFT_MODEL"));
     add_opt(common_arg(

common/arg.h

@@ -8,6 +8,7 @@
 #include <string>
 #include <vector>
 #include <cstring>
+#include <memory>
 
 // pseudo-env variable to identify preset-only arguments
 #define COMMON_ARG_PRESET_LOAD_ON_STARTUP "__PRESET_LOAD_ON_STARTUP"
@@ -130,19 +131,21 @@ bool common_params_to_map(int argc, char ** argv, llama_example ex, std::map<com
 // see: https://github.com/ggml-org/llama.cpp/issues/18163
 void common_params_add_preset_options(std::vector<common_arg> & args);
 
-struct common_params_handle_models_params {
-    common_download_callback * callback = nullptr;
-    bool preset_only = false; // if true, only check & download remote preset (for router mode)
+struct common_models_handler {
+    common_download_hf_plan plan;
+    common_download_hf_plan plan_spec;
+    common_download_hf_plan plan_voc;
+    common_download_opts opts;
 };
 
-// populate model paths (main model, mmproj, etc) from -hf if necessary
-// return true if the model is ready to use
-// throw an exception if there is an error that prevents the model from being used (e.g. network error, model not found, etc)
-// if params.skip_download is true, no downloads will be attempted. return false if the model is invalid or missing (e.g. ETag check failed)
-bool common_params_handle_models(
-    common_params & params,
-    llama_example curr_ex,
-    const common_params_handle_models_params & handle_params);
+// initialize downloading opts and hf_plan if needed, but does not download anything yet
+common_models_handler common_models_handler_init(const common_params & params, llama_example curr_ex);
+
+// check if the model is a preset repo (i.e. has a preset file)
+bool common_models_handler_is_preset_repo(const common_models_handler & handler);
+
+// download and update params with the downloaded model path
+void common_models_handler_apply(common_models_handler & handler, common_params & params, common_download_callback * callback = nullptr);
 
 // initialize argument parser context - used by test-arg-parser and preset
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);

common/chat.cpp

@@ -2758,5 +2758,9 @@ common_chat_msg common_chat_peg_parse(const common_peg_arena &          src_pars
 std::map<std::string, bool> common_chat_templates_get_caps(const common_chat_templates * chat_templates) {
     GGML_ASSERT(chat_templates != nullptr);
     GGML_ASSERT(chat_templates->template_default != nullptr);
+    if (chat_templates->template_tool_use != nullptr) {
+        // take the more expressive template when available
+        return chat_templates->template_tool_use->caps.to_map();
+    }
     return chat_templates->template_default->caps.to_map();
 }

common/common.h

@@ -96,6 +96,7 @@ enum llama_example {
     LLAMA_EXAMPLE_FIT_PARAMS,
     LLAMA_EXAMPLE_RESULTS,
     LLAMA_EXAMPLE_EXPORT_GRAPH_OPS,
+    LLAMA_EXAMPLE_DOWNLOAD,
 
     LLAMA_EXAMPLE_COUNT,
 };
@@ -290,13 +291,13 @@ struct common_params_sampling {
 };
 
 struct common_params_model {
-    std::string path        = ""; // model local path                                       // NOLINT
-    std::string url         = ""; // model url to download                                  // NOLINT
-    std::string hf_repo     = ""; // HF repo                                                // NOLINT
-    std::string hf_file     = ""; // HF file                                                // NOLINT
-    std::string docker_repo = ""; // Docker repo                                            // NOLINT
+    std::string path        = ""; // model local path
+    std::string url         = ""; // model url to download
+    std::string hf_repo     = ""; // HF repo
+    std::string hf_file     = ""; // HF file
+    std::string docker_repo = ""; // Docker repo
 
-    std::string get_name() {
+    std::string get_name() const {
         if (!hf_repo.empty()) {
             return hf_repo;
         }
@@ -305,6 +306,10 @@ struct common_params_model {
         }
         return path;
     }
+
+    bool empty() const {
+        return get_name().empty();
+    }
 };
 
 // draft-model-based speculative decoding parameters
@@ -367,7 +372,7 @@ struct common_params_speculative {
     common_params_speculative_ngram_cache ngram_cache;
 
     bool has_dft() const {
-        return !draft.mparams.path.empty() || !draft.mparams.hf_repo.empty();
+        return !draft.mparams.empty();
     }
 
     uint32_t need_n_rs_seq() const {
@@ -519,7 +524,6 @@ struct common_params {
     int32_t control_vector_layer_start = -1; // layer range for control vector
     int32_t control_vector_layer_end   = -1; // layer range for control vector
     bool    offline                    = false;
-    bool    skip_download              = false; // skip model file downloading
 
     int32_t ppl_stride      = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
     int32_t ppl_output_type = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line

common/download.cpp

@@ -292,10 +292,6 @@ static int common_download_file_single_online(const std::string & url,
 
     const bool file_exists = std::filesystem::exists(path);
 
-    if (!file_exists && opts.skip_download) {
-        return -2; // file is missing and download is disabled
-    }
-
     if (file_exists && skip_etag) {
         LOG_DBG("%s: using cached file: %s\n", __func__, path.c_str());
         return 304; // 304 Not Modified - fake cached response
@@ -362,9 +358,6 @@ static int common_download_file_single_online(const std::string & url,
             return 304; // 304 Not Modified - fake cached response
         }
         // pass this point, the file exists but is different from the server version, so we need to redownload it
-        if (opts.skip_download) {
-            return -2; // special code to indicate that the download was skipped due to etag mismatch
-        }
         if (remove(path.c_str()) != 0) {
             LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
             return -1;
@@ -691,19 +684,8 @@ static void list_available_gguf_files(const hf_cache::hf_files & files) {
     }
 }
 
-struct hf_plan {
-    hf_cache::hf_file primary;
-    hf_cache::hf_files model_files;
-    hf_cache::hf_file mmproj;
-    hf_cache::hf_file mtp;
-    hf_cache::hf_file preset; // if set, only this file is downloaded
-};
-
-static hf_plan get_hf_plan(const common_params_model  & model,
-                           const common_download_opts & opts,
-                           bool download_mmproj,
-                           bool download_mtp) {
-    hf_plan plan;
+common_download_hf_plan common_download_get_hf_plan(const common_params_model & model, const common_download_opts & opts) {
+    common_download_hf_plan plan;
     hf_cache::hf_files all;
 
     auto [repo, tag] = common_download_split_repo_tag(model.hf_repo);
@@ -752,127 +734,49 @@ static hf_plan get_hf_plan(const common_params_model  & model,
     plan.primary = primary;
     plan.model_files = get_split_files(all, primary);
 
-    if (download_mmproj) {
+    if (opts.download_mmproj) {
         plan.mmproj = find_best_mmproj(all, primary.path);
     }
-
-    if (download_mtp) {
+    if (opts.download_mtp) {
         plan.mtp = find_best_mtp(all, primary.path);
     }
 
     return plan;
 }
 
-struct download_task {
-    std::string url;
-    std::string path;
-};
-
-static std::vector<download_task> get_url_tasks(const common_params_model & model) {
-    auto split = get_gguf_split_info(model.url);
-
-    if (split.count <= 1) {
-        return {{model.url, model.path}};
-    }
-
-    auto filename = split.prefix;
-    if (auto pos = split.prefix.rfind('/'); pos != std::string::npos) {
-        filename = split.prefix.substr(pos + 1);
-    }
-
-    auto parent_path = std::filesystem::path(model.path).parent_path();
-    auto prefix_path = (parent_path / filename).string();
-
-    std::vector<download_task> tasks;
-    for (int i = 1; i <= split.count; i++) {
-        auto suffix = string_format("-%05d-of-%05d.gguf", i, split.count);
-        tasks.push_back({split.prefix + suffix, prefix_path + suffix});
-    }
-    return tasks;
-}
-
-common_download_model_result common_download_model(const common_params_model  & model,
-                                                   const common_download_opts & opts) {
-    common_download_model_result result;
-    std::vector<download_task> tasks;
-    hf_plan hf;
-
-    bool download_mmproj = opts.download_mmproj;
-    bool download_mtp = opts.download_mtp;
-    bool preset_only = opts.preset_only;
-    bool is_hf = !model.hf_repo.empty();
-
-    if (is_hf) {
-        hf = get_hf_plan(model, opts, download_mmproj, download_mtp);
-        if (!hf.preset.path.empty()) {
-            // if preset.ini exists, only download that file alone
-            tasks.push_back({hf.preset.url, hf.preset.local_path});
-        } else if (!preset_only) {
-            // only add other files if we're NOT in preset-only mode (normal run, non-router)
-            for (const auto & f : hf.model_files) {
-                tasks.push_back({f.url, f.local_path});
-            }
-            if (!hf.mmproj.path.empty()) {
-                tasks.push_back({hf.mmproj.url, hf.mmproj.local_path});
-            }
-            if (!hf.mtp.path.empty()) {
-                tasks.push_back({hf.mtp.url, hf.mtp.local_path});
-            }
-        }
-    } else if (!model.url.empty()) {
-        tasks = get_url_tasks(model);
-    } else {
-        result.model_path = model.path;
-        return result;
-    }
-
-    if (tasks.empty()) {
-        return result;
-    }
-
+void common_download_run_tasks(const std::vector<common_download_task> & tasks) {
     std::vector<std::future<int>> futures;
     for (const auto & task : tasks) {
         futures.push_back(std::async(std::launch::async,
-            [&task, &opts, is_hf]() {
-                return common_download_file_single(task.url, task.path, opts, is_hf);
+            [&task]() {
+                return common_download_file_single(task.url, task.local_path, task.opts, task.is_hf);
             }
         ));
     }
 
-    for (auto & f : futures) {
-        int status = f.get();
-        if (status == -2 && opts.skip_download) {
-            throw common_skip_download_exception();
-        }
+    for (size_t i = 0; i < futures.size(); ++i) {
+        std::string url = tasks[i].url;
+        int status = futures[i].get();
         bool is_ok = is_http_status_ok(status);
         if (!is_ok) {
-            return {};
+            throw std::runtime_error(string_format("Download '%s' failed with status code: %d", url.c_str(), status));
         }
     }
+}
 
-    if (is_hf) {
-        if (!hf.preset.path.empty()) {
-            // if preset.ini is used, do not set other paths
-            result.preset_path = hf_cache::finalize_file(hf.preset);
-        } else {
-            for (const auto & f : hf.model_files) {
-                hf_cache::finalize_file(f);
-            }
-            result.model_path = hf.primary.final_path;
+std::vector<std::string> common_download_get_all_parts(const std::string & url) {
+    auto split = get_gguf_split_info(url);
 
-            if (!hf.mmproj.path.empty()) {
-                result.mmproj_path = hf_cache::finalize_file(hf.mmproj);
-            }
-
-            if (!hf.mtp.path.empty()) {
-                result.mtp_path = hf_cache::finalize_file(hf.mtp);
-            }
-        }
-    } else {
-        result.model_path = model.path;
+    if (split.count <= 1) {
+        return {url};
     }
 
-    return result;
+    std::vector<std::string> parts;
+    for (int i = 1; i <= split.count; i++) {
+        auto suffix = string_format("-%05d-of-%05d.gguf", i, split.count);
+        parts.push_back(split.prefix + suffix);
+    }
+    return parts;
 }
 
 //

common/download.h

@@ -1,7 +1,10 @@
 #pragma once
 
+#include "hf-cache.h"
+
 #include <string>
 #include <vector>
+#include <functional>
 
 struct common_params_model;
 
@@ -47,67 +50,40 @@ struct common_cached_model_info {
     }
 };
 
-// Options for common_download_model and common_download_file_single
+// Options for common_download_file_single
 struct common_download_opts {
     std::string bearer_token;
     common_header_list headers;
     bool offline = false;
-    bool skip_download = false; // if true, only validation is performed, common_skip_download_exception may be thrown if the file is missing or invalid
     bool download_mmproj = false;
     bool download_mtp = false;
-    bool preset_only = false; // if true, only check & download remote preset (for router mode)
     common_download_callback * callback = nullptr;
 };
 
-// Result of common_download_model
-struct common_download_model_result {
-    std::string model_path;
-    std::string mmproj_path;
-    std::string mtp_path;
-    std::string preset_path;
+struct common_download_task {
+    common_download_opts opts;
+    std::string url;
+    std::string local_path;
+    std::function<void()> on_done;
+    bool is_hf = false;
+
+    common_download_task() = default;
+    common_download_task(hf_cache::hf_file f,
+            const common_download_opts & opts,
+            std::function<void()> on_done = nullptr)
+        : opts(opts), url(f.url), local_path(f.local_path), on_done(on_done), is_hf(true) {}
 };
 
-// throw if the file is missing or invalid (e.g. ETag check failed)
-struct common_skip_download_exception : public std::runtime_error {
-    common_skip_download_exception() : std::runtime_error("skip download") {}
-};
+void common_download_run_tasks(const std::vector<common_download_task> & tasks);
 
-// Download model from HuggingFace repo or URL
-//
-// input (via model struct):
-// - model.hf_repo: HF repo with optional tag, see common_download_split_repo_tag
-// - model.hf_file: specific file in the repo (requires hf_repo)
-// - model.url: simple download (used if hf_repo is empty)
-// - model.path: local file path
-//
-// tag matching (for HF repos without model.hf_file):
-// - if tag is specified, searches for GGUF matching that quantization
-// - if no tag, searches for Q4_K_M, then Q4_0, then first available GGUF
-//
-// split GGUF: multi-part files like "model-00001-of-00003.gguf" are automatically
-// detected and all parts are downloaded
-//
-// caching:
-// - HF repos: uses HuggingFace cache
-// - URLs: uses ETag-based caching
-//
-// when opts.offline=true, no network requests are made
-// when download_mmproj=true, searches for mmproj in same directory as model or any parent directory
-// then with the closest quantization bits
-// when download_mtp=true, applies the same sibling search for an MTP-head GGUF
-//
-// returns result with model_path, mmproj_path and mtp_path (empty when not found / on failure)
-common_download_model_result common_download_model(
-    const common_params_model & model,
-    const common_download_opts & opts = {}
-);
+// if url is a multi-part GGUF file, returns all parts, otherwise returns the single file
+std::vector<std::string> common_download_get_all_parts(const std::string & url);
 
 // returns list of cached models
 std::vector<common_cached_model_info> common_list_cached_models();
 
 // download single file from url to local path
 // returns status code or -1 on error
-// returns -2 if the download was skipped due to ETag mismatch (file outdated, skip_download=true)
 // skip_etag: if true, don't read/write .etag files (for HF cache where filename is the hash)
 int common_download_file_single(const std::string & url,
                                 const std::string & path,
@@ -124,3 +100,12 @@ std::string common_docker_resolve_model(const std::string & docker);
 // - if tag is present, removes only files matching that tag (and orphaned blobs)
 // returns true if anything was removed
 bool common_download_remove(const std::string & hf_repo_with_tag);
+
+struct common_download_hf_plan {
+    hf_cache::hf_file primary;
+    hf_cache::hf_files model_files;
+    hf_cache::hf_file mmproj;
+    hf_cache::hf_file mtp;
+    hf_cache::hf_file preset; // if set, only this file is downloaded
+};
+common_download_hf_plan common_download_get_hf_plan(const common_params_model & model, const common_download_opts & opts);

conversion/__init__.py

@@ -136,6 +136,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "LlamaModel": "llama",
     "Eagle3DraftModel": "llama",
     "Eagle3Speculator": "llama",
+    "Eagle3LlamaForCausalLM": "llama",
     "LlamaForCausalLMEagle3": "llama",
     "LlavaForConditionalGeneration": "llama",
     "LlavaStableLMEpochForCausalLM": "stablelm",

conversion/llama.py

@@ -23,6 +23,7 @@ from .base import ModelBase, TextModel, gguf, logger
     "LlavaForConditionalGeneration",
     "VoxtralForConditionalGeneration",
     "LlamaForCausalLMEagle3",
+    "Eagle3LlamaForCausalLM",
     "Eagle3Speculator",
     "Eagle3DraftModel",
     "IQuestCoderForCausalLM",

conversion/mamba.py

@@ -114,7 +114,8 @@ class Mamba2Model(TextModel):
             hparams["text_config"] = hparams["llm_config"]
         super().__init__(dir_model, *args, hparams=hparams, **kwargs)
         self.d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
-        self.d_inner = self.find_hparam(["mamba_d_ssm", "intermediate_size", "d_inner"], optional=True) or 2 * self.d_model
+        self.expand = self.find_hparam(["mamba_expand", "expand"], optional=True) or 2
+        self.d_inner = self.find_hparam(["mamba_d_ssm", "intermediate_size", "d_inner"], optional=True) or self.expand * self.d_model
         self.n_group = self.find_hparam(["n_groups"], optional=True) or 1
 
     def set_vocab(self):
@@ -144,11 +145,9 @@ class Mamba2Model(TextModel):
 
         rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5
 
-        # Fail early for models which don't have a block expansion factor of 2
-        # TODO: does this really matter?
         # skip the assertion for FalconH1 Model
         if self.model_arch != gguf.MODEL_ARCH.FALCON_H1:
-            assert self.d_inner == 2 * self.d_model
+            assert self.d_inner == self.expand * self.d_model
             assert self.d_inner % head_dim == 0
 
         self.gguf_writer.add_context_length(2**20)  # arbitrary value; for those who use the default

docs/backend/OPENVINO.md

@@ -237,8 +237,8 @@ chmod +x ubuntu-llamacpp-ov-install.sh
 # ============================================
 set -euo pipefail
 
-OPENVINO_VERSION_MAJOR="2026.2"
-OPENVINO_VERSION_FULL="2026.2.0.21903.52ddc073857"
+OPENVINO_VERSION_MAJOR="2026.2.1"
+OPENVINO_VERSION_FULL="2026.2.1.21919.ede283a88e3"
 
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 OPENVINO_INSTALL_DIR="/opt/intel/openvino_${OPENVINO_VERSION_MAJOR}"
@@ -334,7 +334,7 @@ echo "  ./build/ReleaseOV/bin/llama-cli -m model.gguf"
 ```
 
 > [!NOTE]
-> The script pins OpenVINO `2026.2` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release.
+> The script pins OpenVINO `2026.2.1` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release.
 
 </details>
 
@@ -364,8 +364,8 @@ REM ============================================
 REM llama.cpp OpenVINO Build Script (Ninja)
 REM ============================================
 
-set "OPENVINO_VERSION_MAJOR=2026.2"
-set "OPENVINO_VERSION_FULL=2026.2.0.21903.52ddc073857"
+set "OPENVINO_VERSION_MAJOR=2026.2.1"
+set "OPENVINO_VERSION_FULL=2026.2.1.21919.ede283a88e3"
 
 set "SCRIPT_DIR=%~dp0"
 set "VCPKG_DIR=C:\vcpkg"
@@ -547,7 +547,7 @@ endlocal
 ```
 
 > [!NOTE]
-> The script pins OpenVINO `2026.2` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release. From any new shell, source the matching `setupvars` script via the junction — `call "C:\Intel\openvino\setupvars.bat"` from `cmd`, or `& "C:\Intel\openvino\setupvars.ps1"` from PowerShell. If `winget` cannot register Visual Studio Build Tools on first run, install them once manually and re-run the script from an elevated **Developer Command Prompt for VS 2022**.
+> The script pins OpenVINO `2026.2.1` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release. From any new shell, source the matching `setupvars` script via the junction — `call "C:\Intel\openvino\setupvars.bat"` from `cmd`, or `& "C:\Intel\openvino\setupvars.ps1"` from PowerShell. If `winget` cannot register Visual Studio Build Tools on first run, install them once manually and re-run the script from an elevated **Developer Command Prompt for VS 2022**.
 
 </details>
 

docs/backend/SYCL.md

@@ -413,6 +413,15 @@ In two device selection modes, the default SYCL backend is level_zero, you can c
 |------------------|----------------------------------------|
 | Single device    | --split-mode none --main-gpu DEVICE_ID |
 | Multiple devices | --split-mode layer (default)           |
+| Multiple devices | --split-mode tensor (tensor parallelism) |
+
+`--split-mode tensor` (tensor parallelism) shards each layer across the selected
+GPUs. It requires flash attention, which is auto-enabled when `--flash-attn` is
+left at its default `auto`, so `--split-mode tensor` works out of the box.
+Passing `--flash-attn off` together with `--split-mode tensor` is rejected at
+context creation. The default `f16` KV cache is recommended. Tensor parallelism
+is currently optimized for 2 GPUs; other device counts fall back to a generic
+all-reduce.
 
 Examples:
 
@@ -715,6 +724,15 @@ In two device selection modes, the default SYCL backend is level_zero, you can c
 |------------------|----------------------------------------|
 | Single device    | --split-mode none --main-gpu DEVICE_ID |
 | Multiple devices | --split-mode layer (default)           |
+| Multiple devices | --split-mode tensor (tensor parallelism) |
+
+`--split-mode tensor` (tensor parallelism) shards each layer across the selected
+GPUs. It requires flash attention, which is auto-enabled when `--flash-attn` is
+left at its default `auto`, so `--split-mode tensor` works out of the box.
+Passing `--flash-attn off` together with `--split-mode tensor` is rejected at
+context creation. The default `f16` KV cache is recommended. Tensor parallelism
+is currently optimized for 2 GPUs; other device counts fall back to a generic
+all-reduce.
 
 Examples:
 

docs/speculative.md

@@ -13,6 +13,45 @@ The `llama-server` application supports several implementations of speculative d
 A much smaller model (called the _draft model_) generates drafts.
 A draft model is the most used approach in speculative decoding.
 
+### EAGLE-3 (`draft-eagle3`)
+
+EAGLE-3 uses a small draft model that reads the target model's hidden states to predict the next tokens, so it
+reaches higher acceptance than a standalone draft model of the same size. The draft is a one-layer transformer
+trained for a specific target model; it shares the target model's tokenizer and, optionally, uses a reduced draft
+vocabulary with its own `lm_head`, which is mapped back using a `d2t` table.
+
+Convert the EAGLE-3 checkpoint with `--target-model-dir` so it inherits the target's tokenizer and the layer
+indices to read. Both the SpecForge `LlamaForCausalLMEagle3` and the vLLM/AngelSlim `Eagle3LlamaForCausalLM`
+checkpoint formats are supported (for example [`AngelSlim/Qwen3-4B_eagle3`](https://huggingface.co/AngelSlim/Qwen3-4B_eagle3)
+for `Qwen/Qwen3-4B`):
+
+```bash
+python convert_hf_to_gguf.py AngelSlim/Qwen3-4B_eagle3 \
+    --target-model-dir Qwen/Qwen3-4B --outtype bf16 --outfile Qwen3-4B-eagle3.gguf
+
+llama-server -m Qwen3-4B.gguf -md Qwen3-4B-eagle3.gguf --spec-type draft-eagle3
+```
+
+Supported EAGLE-3 draft models include:
+
+- [yuhuili/EAGLE3-LLaMA3.1-Instruct-8B](https://huggingface.co/yuhuili/EAGLE3-LLaMA3.1-Instruct-8B)
+- [yuhuili/EAGLE3-LLaMA3.3-Instruct-70B](https://huggingface.co/yuhuili/EAGLE3-LLaMA3.3-Instruct-70B)
+- [RedHatAI/gemma-4-31B-it-speculator.eagle3](https://huggingface.co/RedHatAI/gemma-4-31B-it-speculator.eagle3)
+- [RedHatAI/gemma-4-26B-A4B-it-speculator.eagle3](https://huggingface.co/RedHatAI/gemma-4-26B-A4B-it-speculator.eagle3)
+- [Tengyunw/qwen3_8b_eagle3](https://huggingface.co/Tengyunw/qwen3_8b_eagle3)
+- [Tengyunw/qwen3_30b_moe_eagle3](https://huggingface.co/Tengyunw/qwen3_30b_moe_eagle3)
+- [AngelSlim/Qwen3-1.7B_eagle3](https://huggingface.co/AngelSlim/Qwen3-1.7B_eagle3)
+- [AngelSlim/Qwen3-4B_eagle3](https://huggingface.co/AngelSlim/Qwen3-4B_eagle3)
+- [AngelSlim/Qwen3-8B_eagle3](https://huggingface.co/AngelSlim/Qwen3-8B_eagle3)
+- [AngelSlim/Qwen3-14B_eagle3](https://huggingface.co/AngelSlim/Qwen3-14B_eagle3)
+- [AngelSlim/Qwen3-32B_eagle3](https://huggingface.co/AngelSlim/Qwen3-32B_eagle3)
+- [AngelSlim/Qwen3-a3B_eagle3](https://huggingface.co/AngelSlim/Qwen3-a3B_eagle3)
+- [RedHatAI/gpt-oss-20b-speculator.eagle3](https://huggingface.co/RedHatAI/gpt-oss-20b-speculator.eagle3)
+- [lmsys/EAGLE3-gpt-oss-120b-bf16](https://huggingface.co/lmsys/EAGLE3-gpt-oss-120b-bf16)
+- [nvidia/gpt-oss-120b-Eagle3-long-context](https://huggingface.co/nvidia/gpt-oss-120b-Eagle3-long-context)
+
+For the full and up-to-date list of supported models, see #18039.
+
 ### n-gram Cache (`ngram-cache`)
 
 An n-gram is a sequence of n tokens. The n-gram cache implementation maintains statistics about short n-gram sequences.
@@ -108,7 +147,7 @@ If a draft model is combined with a draftless decoding the draftless decoding ha
 ### General Speculative Parameters
 
 ```
---spec-type [none|draft-simple|draft-mtp|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v|ngram-mod]
+--spec-type [none|draft-simple|draft-eagle3|draft-mtp|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v|ngram-mod]
                                         comma-separated list of types of speculative decoding to use
                                         (default: none)
                                         (env: LLAMA_ARG_SPEC_TYPE)
@@ -247,6 +286,7 @@ Specifies a comma-separated list of speculative decoding types to use.
 |------|-------------|
 | `none` | No speculative decoding (default) |
 | `draft-simple` | Use a simple draft model for speculation |
+| `draft-eagle3` | Use an EAGLE-3 draft model that reads the target's hidden states |
 | `draft-mtp` | Use Multi Token Prediction (MTP) heads from the main model |
 | `ngram-cache` | Use n-gram cache lookup |
 | `ngram-simple` | Use simple n-gram pattern matching |

ggml/CMakeLists.txt

@@ -5,7 +5,7 @@ project("ggml" C CXX ASM)
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
 set(GGML_VERSION_MINOR 15)
-set(GGML_VERSION_PATCH 2)
+set(GGML_VERSION_PATCH 3)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
 
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")

ggml/include/ggml-sycl.h

@@ -27,6 +27,14 @@ GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int de
 // split tensor buffer that splits matrices by rows across multiple devices
 GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
 
+// Tensor parallelism (--split-mode tensor): comm_init/free/allreduce_tensor
+// trio queried by the meta-backend via ggml_backend_reg_get_proc_address.
+// See typedefs in ggml/include/ggml-backend.h. Mirrors the CUDA backend's
+// pattern (ggml_backend_cuda_comm_*).
+GGML_BACKEND_API void * ggml_backend_sycl_comm_init(ggml_backend_t * backends, size_t n_backends);
+GGML_BACKEND_API void   ggml_backend_sycl_comm_free(void * comm_ctx);
+GGML_BACKEND_API bool   ggml_backend_sycl_comm_allreduce_tensor(void * comm_ctx, struct ggml_tensor ** tensors);
+
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
 GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
 

ggml/src/ggml-backend.cpp

@@ -1551,6 +1551,8 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
         int split_backend_id = split->backend_id;
         ggml_backend_t split_backend = sched->backends[split_backend_id];
 
+        ggml_backend_synchronize(split_backend);
+
         // copy the input tensors to the split backend
         for (int input_id = 0; input_id < split->n_inputs; input_id++) {
             ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[input_id]);
@@ -1561,15 +1563,15 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                 // inputs from the user must be copied immediately to prevent the user overwriting the data before the copy is done
                 if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
                     ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
-                } else {
+                } else if (!split_backend->iface.cpy_tensor_async) {
                     ggml_backend_synchronize(split_backend);
                 }
-                ggml_backend_tensor_copy(input, input_cpy);
+                ggml_backend_tensor_copy_async(input_backend, split_backend, input, input_cpy);
             } else {
                 // wait for the split backend to finish using the input before overwriting it
                 if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
                     ggml_backend_event_wait(split_backend, sched->events[split_backend_id][sched->cur_copy]);
-                } else {
+                } else if (!split_backend->iface.cpy_tensor_async) {
                     ggml_backend_synchronize(split_backend);
                 }
 
@@ -1674,6 +1676,8 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
             }
         }
 
+        ggml_backend_synchronize(split_backend);
+
         if (!sched->callback_eval) {
             enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &split->graph);
             if (ec != GGML_STATUS_SUCCESS) {

ggml/src/ggml-cpu/vec.cpp

@@ -75,12 +75,12 @@ void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * G
             ay1 = GGML_F32_VEC_LOAD(y + i);
             sum1 = GGML_F32_VEC_FMA(sum1, ax1, ay1);
         }
-        // maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only
+        // maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmla on available elements only
         if (np2 < n) {
             svbool_t pg = svwhilelt_b32(np2, n);
             ax1 = svld1_f32(pg, x + np2);
             ay1 = svld1_f32(pg, y + np2);
-            sum1 = svmad_f32_m(pg, ax1, ay1, sum1);
+            sum1 = svmla_f32_m(pg, sum1, ax1, ay1);
         }
         // reduce sum1,sum2 to sum1
         GGML_F32_VEC_REDUCE(sumf, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8);

ggml/src/ggml-cuda/binbcast.cu

@@ -34,26 +34,26 @@ template <float (*bin_op)(const float, const float),
 static __global__ void k_bin_bcast(const src0_t *         src0,
                                    const src1_t *         src1,
                                    dst_t *                dst,
-                                   const int              ne0,
-                                   const int              ne1,
-                                   const int              ne2,
+                                   const uint32_t         ne0,
+                                   const uint32_t         ne1,
+                                   const uint32_t         ne2,
                                    const uint3            ne3,
                                    const uint3            ne10,
                                    const uint3            ne11,
                                    const uint3            ne12,
                                    const uint3            ne13,
-                                 /*const int              s0,*/
-                                   const int              s1,
-                                   const int              s2,
-                                   const int              s3,
-                                   const int              s00,
-                                   const int              s01,
-                                   const int              s02,
-                                   const int              s03,
-                                   const int              s10,
-                                   const int              s11,
-                                   const int              s12,
-                                   const int              s13,
+                                 /*const uint32_t         s0,*/
+                                   const uint32_t         s1,
+                                   const uint32_t         s2,
+                                   const uint32_t         s3,
+                                   const uint32_t         s00,
+                                   const uint32_t         s01,
+                                   const uint32_t         s02,
+                                   const uint32_t         s03,
+                                   const uint32_t         s10,
+                                   const uint32_t         s11,
+                                   const uint32_t         s12,
+                                   const uint32_t         s13,
                                    src1_ptrs... src1s) {
     ggml_cuda_pdl_lc();
     const uint32_t i0s = blockDim.x * blockIdx.x + threadIdx.x;
@@ -61,7 +61,7 @@ static __global__ void k_bin_bcast(const src0_t *         src0,
     const uint32_t i2  = fastdiv((blockDim.z * blockIdx.z + threadIdx.z), ne3);
     const uint32_t i3  = (blockDim.z * blockIdx.z + threadIdx.z) - (i2 * ne3.z);
 
-    if (i0s >= (uint32_t)ne0 || i1 >= (uint32_t)ne1 || i2 >= (uint32_t)ne2 || i3 >= ne3.z) {
+    if (i0s >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3.z) {
         return;
     }
 
@@ -69,25 +69,32 @@ static __global__ void k_bin_bcast(const src0_t *         src0,
     const uint32_t i12 = fastmodulo(i2, ne12);
     const uint32_t i13 = fastmodulo(i3, ne13);
 
-    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
-    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
-    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
+    const size_t i_src0 = size_t( i3)*s03 + size_t( i2)*s02 + size_t( i1)*s01;
+    const size_t i_src1 = size_t(i13)*s13 + size_t(i12)*s12 + size_t(i11)*s11;
+    const size_t i_dst  = size_t( i3)*s3  + size_t( i2)*s2  + size_t( i1)*s1;
 
     const src0_t * src0_row = src0 ? (src0 + i_src0) : nullptr;
     dst_t * dst_row = dst + i_dst;
 
+    const uint32_t s0 = blockDim.x * gridDim.x;
+
     ggml_cuda_pdl_sync();
-    for (int i0 = i0s; i0 < ne0; i0 += blockDim.x * gridDim.x) {
+    for (uint32_t i0 = i0s; i0 < ne0; i0 += s0) {
         const uint32_t i10 = fastmodulo(i0, ne10);
 
-        float result = src0_row ? (float) src0_row[i0*s00] : 0.0f;
+        float result = src0_row ? (float) src0_row[size_t(i0)*s00] : 0.0f;
         if constexpr (sizeof...(src1_ptrs) > 0) {
-            result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10*s10])));
+            result = (..., (result = bin_op(result, (float)src1s[i_src1 + size_t(i10)*s10])));
         } else {
-            result = bin_op(result, (float)src1[i_src1 + i10*s10]);
+            result = bin_op(result, (float)src1[i_src1 + size_t(i10)*s10]);
         }
 
         dst_row[i0] = (dst_t) result;
+
+        // protect i0 from overflow
+        if (ne0 - i0 <= s0) {
+           break;
+        }
     }
 }
 
@@ -110,19 +117,19 @@ static __global__ void k_bin_bcast_unravel(const src0_t *         src0,
                                            const uint3            ne12,
                                            const uint3            ne13,
                                          /*const int              s0,*/
-                                           const int              s1,
-                                           const int              s2,
-                                           const int              s3,
-                                           const int              s00,
-                                           const int              s01,
-                                           const int              s02,
-                                           const int              s03,
-                                           const int              s10,
-                                           const int              s11,
-                                           const int              s12,
-                                           const int              s13,
+                                           const uint32_t         s1,
+                                           const uint32_t         s2,
+                                           const uint32_t         s3,
+                                           const uint32_t         s00,
+                                           const uint32_t         s01,
+                                           const uint32_t         s02,
+                                           const uint32_t         s03,
+                                           const uint32_t         s10,
+                                           const uint32_t         s11,
+                                           const uint32_t         s12,
+                                           const uint32_t         s13,
                                            src1_ptrs... src1s) {
-    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+    const uint32_t i  = blockDim.x*blockIdx.x + threadIdx.x;
 
     const uint32_t i3 = fastdiv(i, prod_012);
     const uint32_t i2 = fastdiv(i - i3 * prod_012.z, prod_01);
@@ -133,25 +140,25 @@ static __global__ void k_bin_bcast_unravel(const src0_t *         src0,
         return;
     }
 
-    const int i11 = fastmodulo(i1, ne11);
-    const int i12 = fastmodulo(i2, ne12);
-    const int i13 = fastmodulo(i3, ne13);
+    const uint32_t i11 = fastmodulo(i1, ne11);
+    const uint32_t i12 = fastmodulo(i2, ne12);
+    const uint32_t i13 = fastmodulo(i3, ne13);
 
-    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
-    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
-    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
+    const size_t i_src0 = size_t( i3)*s03 + size_t( i2)*s02 + size_t( i1)*s01;
+    const size_t i_src1 = size_t(i13)*s13 + size_t(i12)*s12 + size_t(i11)*s11;
+    const size_t i_dst  = size_t( i3)*s3  + size_t( i2)*s2  + size_t( i1)*s1;
 
     const src0_t * src0_row = src0 ? (src0 + i_src0) : nullptr;
     dst_t * dst_row = dst + i_dst;
 
-    const int i10 = fastmodulo(i0, ne10);
+    const uint32_t i10 = fastmodulo(i0, ne10);
 
     ggml_cuda_pdl_sync();
-    float result = src0_row ? (float) src0_row[i0*s00] : 0.0f;
+    float result = src0_row ? (float) src0_row[size_t(i0)*s00] : 0.0f;
     if constexpr (sizeof...(src1_ptrs) > 0) {
-        result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10*s10])));
+        result = (..., (result = bin_op(result, (float)src1s[i_src1 + size_t(i10)*s10])));
     } else {
-        result = bin_op(result, (float)src1[i_src1 + i10*s10]);
+        result = bin_op(result, (float)src1[i_src1 + size_t(i10)*s10]);
     }
 
     dst_row[i0] = (dst_t) result;
@@ -248,6 +255,31 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
         size_t s02 = nb02 / sizeof(src0_t);
         size_t s03 = nb03 / sizeof(src0_t);
 
+        GGML_ASSERT(ne0 <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(ne1 <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(ne2 <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(ne3 <= std::numeric_limits<uint32_t>::max());
+
+      //GGML_ASSERT(s0  <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(s1  <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(s2  <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(s3  <= std::numeric_limits<uint32_t>::max());
+
+        GGML_ASSERT(s00 <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(s01 <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(s02 <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(s03 <= std::numeric_limits<uint32_t>::max());
+
+        GGML_ASSERT(s10 <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(s11 <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(s12 <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(s13 <= std::numeric_limits<uint32_t>::max());
+
+        GGML_ASSERT(cne1[0] <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(cne1[1] <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(cne1[2] <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(cne1[3] <= std::numeric_limits<uint32_t>::max());
+
         GGML_ASSERT(nb0 % sizeof(dst_t) == 0);
         GGML_ASSERT(nb1 % sizeof(dst_t) == 0);
         GGML_ASSERT(nb2 % sizeof(dst_t) == 0);
@@ -263,6 +295,8 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
         GGML_ASSERT(nb12 % sizeof(src1_t) == 0);
         GGML_ASSERT(nb13 % sizeof(src1_t) == 0);
 
+        GGML_ASSERT(ne2 * ne3 <= std::numeric_limits<unsigned int>::max());
+
         const int block_size = 128;
 
         int64_t hne0 = std::max(ne0 / 2LL, 1LL);
@@ -281,7 +315,13 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
         const uint3 ne13 = init_fastdiv_values((uint32_t) cne1[3]);
 
         if (block_nums.z > 65535 || block_nums.y > 65535) {
-            int         block_num  = (ne0 * ne1 * ne2 * ne3 + block_size - 1) / block_size;
+            int64_t     block_num   = (ne0 * ne1 * ne2 * ne3 + block_size - 1) / block_size;
+
+            GGML_ASSERT(block_num              <= std::numeric_limits<uint32_t>::max());
+            GGML_ASSERT(block_num * block_size <= std::numeric_limits<uint32_t>::max());
+            GGML_ASSERT(ne0 * ne1              <= std::numeric_limits<uint32_t>::max());
+            GGML_ASSERT(ne0 * ne1 * ne2        <= std::numeric_limits<uint32_t>::max());
+
             const uint3 prod_012    = init_fastdiv_values((uint32_t) (ne0 * ne1 * ne2));
             const uint3 prod_01     = init_fastdiv_values((uint32_t) (ne0 * ne1));
             const uint3 ne0_fastdiv = init_fastdiv_values((uint32_t) ne0);
@@ -298,6 +338,10 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
                     s10, s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
             }
         } else {
+            GGML_ASSERT(int64_t(block_nums.x) * block_dims.x <= std::numeric_limits<uint32_t>::max());
+            GGML_ASSERT(int64_t(block_nums.y) * block_dims.y <= std::numeric_limits<uint32_t>::max());
+            GGML_ASSERT(int64_t(block_nums.z) * block_dims.z <= std::numeric_limits<uint32_t>::max());
+
             const uint3 ne3_fastdiv = init_fastdiv_values((uint32_t) ne3);
             {
                 const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);

ggml/src/ggml-cuda/cpy.cu

@@ -53,10 +53,10 @@ static __global__ void cpy_scalar_transpose(const char * cx, char * cdst, const
     const int64_t nmat = ne / (ne00 * ne01);
     const int64_t n = ne00 * ne01;
 
-    const int x = blockIdx.x * CUDA_CPY_TILE_DIM_2D + threadIdx.x;
-    const int y = blockIdx.y * CUDA_CPY_TILE_DIM_2D + threadIdx.y;
-    const int tx = blockIdx.y * CUDA_CPY_TILE_DIM_2D + threadIdx.x;  // transpose block offset
-    const int ty = blockIdx.x * CUDA_CPY_TILE_DIM_2D + threadIdx.y;
+    const int64_t x  = (int64_t) blockIdx.x * CUDA_CPY_TILE_DIM_2D + threadIdx.x;
+    const int64_t y  = (int64_t) blockIdx.y * CUDA_CPY_TILE_DIM_2D + threadIdx.y;
+    const int64_t tx = (int64_t) blockIdx.y * CUDA_CPY_TILE_DIM_2D + threadIdx.x;  // transpose block offset
+    const int64_t ty = (int64_t) blockIdx.x * CUDA_CPY_TILE_DIM_2D + threadIdx.y;
 
     __shared__ float tile[2][CUDA_CPY_TILE_DIM_2D][CUDA_CPY_TILE_DIM_2D+1];
     int cur_tile_buf = 0;
@@ -197,7 +197,7 @@ static void ggml_cpy_scalar_contiguous_cuda(
 cudaStream_t stream) {
 
     const int64_t num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
     const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params((dim3)num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream);
     ggml_cuda_kernel_launch(cpy_scalar_contiguous<src_t, dst_t>, launch_params, cx, cdst, ne);
 }
@@ -208,6 +208,14 @@ static void ggml_cpy_scalar_cuda(
     const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t nb00, const int64_t nb01, const int64_t nb02,
     const int64_t nb03, const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13, cudaStream_t stream) {
 
+    const auto launch_scalar_generic = [&]() {
+        const int64_t num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
+        GGML_ASSERT(num_blocks <= INT_MAX);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params((dim3)num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream);
+        ggml_cuda_kernel_launch(cpy_scalar<cpy_1_scalar<src_t, dst_t>>, launch_params,
+            cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+    };
+
     if (transposed) {
         GGML_ASSERT(ne == ne00*ne01*ne02);  // ne[3] is 1 assumed
         int64_t ne00n, ne01n, ne02n;
@@ -224,20 +232,18 @@ static void ggml_cpy_scalar_cuda(
         int64_t grid_x = (ne01n + CUDA_CPY_TILE_DIM_2D - 1) / CUDA_CPY_TILE_DIM_2D;
         int64_t grid_y = (ne00n + CUDA_CPY_TILE_DIM_2D - 1) / CUDA_CPY_TILE_DIM_2D;
         int64_t grid_z = (ne/(ne01n*ne00n) + CUDA_CPY_BLOCK_NM - 1) / CUDA_CPY_BLOCK_NM;
-        GGML_ASSERT(grid_x < UINT_MAX);
-        GGML_ASSERT(grid_y < USHRT_MAX);
-        GGML_ASSERT(grid_z < USHRT_MAX);
-        dim3 dimGrid(grid_x, grid_y, grid_z);
-        dim3 dimBlock(CUDA_CPY_TILE_DIM_2D, CUDA_CPY_BLOCK_ROWS, 1);
-        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(dimGrid, dimBlock, 0, stream);
-        ggml_cuda_kernel_launch(cpy_scalar_transpose<dst_t>, launch_params,
-            cx, cdst, ne, ne00n, ne01n, ne02n, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        GGML_ASSERT(grid_x <= INT_MAX);
+        if (grid_y > USHRT_MAX || grid_z > USHRT_MAX) {
+            launch_scalar_generic();
+        } else {
+            dim3 dimGrid(grid_x, grid_y, grid_z);
+            dim3 dimBlock(CUDA_CPY_TILE_DIM_2D, CUDA_CPY_BLOCK_ROWS, 1);
+            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(dimGrid, dimBlock, 0, stream);
+            ggml_cuda_kernel_launch(cpy_scalar_transpose<dst_t>, launch_params,
+                cx, cdst, ne, ne00n, ne01n, ne02n, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        }
     } else {
-        const int64_t num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-        GGML_ASSERT(num_blocks < UINT_MAX);
-        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params((dim3)num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream);
-        ggml_cuda_kernel_launch(cpy_scalar<cpy_1_scalar<src_t, dst_t>>, launch_params,
-            cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        launch_scalar_generic();
     }
 }
 
@@ -248,7 +254,7 @@ static void ggml_cpy_f32_q8_0_cuda(
 
     GGML_ASSERT(ne % QK8_0 == 0);
     const int64_t num_blocks = ne / QK8_0;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
     cpy_f32_q<cpy_blck_f32_q8_0, QK8_0><<<num_blocks, 1, 0, stream>>>
         (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
@@ -259,7 +265,7 @@ static void ggml_cpy_q8_0_f32_cuda(
     const int64_t nb03, const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13, cudaStream_t stream) {
 
     const int64_t num_blocks = ne;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
     cpy_q_f32<cpy_blck_q8_0_f32, QK8_0><<<num_blocks, 1, 0, stream>>>
         (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
@@ -271,7 +277,7 @@ static void ggml_cpy_f32_q4_0_cuda(
 
     GGML_ASSERT(ne % QK4_0 == 0);
     const int64_t num_blocks = ne / QK4_0;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
     cpy_f32_q<cpy_blck_f32_q4_0, QK4_0><<<num_blocks, 1, 0, stream>>>
         (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
@@ -284,7 +290,7 @@ static void ggml_cpy_q4_0_f32_cuda(
     const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
     cudaStream_t stream) {
     const int64_t num_blocks = ne;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
     cpy_q_f32<cpy_blck_q_f32<dequantize_q4_0, QK4_0>, QK4_0><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
          ne10, ne11, ne12, nb10, nb11, nb12, nb13);
@@ -297,7 +303,7 @@ static void ggml_cpy_f32_q4_1_cuda(
 
     GGML_ASSERT(ne % QK4_1 == 0);
     const int64_t num_blocks = ne / QK4_1;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
     cpy_f32_q<cpy_blck_f32_q4_1, QK4_1><<<num_blocks, 1, 0, stream>>>
         (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
@@ -310,7 +316,7 @@ static void ggml_cpy_q4_1_f32_cuda(
     const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
     cudaStream_t stream) {
     const int64_t num_blocks = ne;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
     cpy_q_f32<cpy_blck_q_f32<dequantize_q4_1, QK4_1>, QK4_1><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
          ne10, ne11, ne12, nb10, nb11, nb12, nb13);
@@ -323,7 +329,7 @@ static void ggml_cpy_f32_q5_0_cuda(
 
     GGML_ASSERT(ne % QK5_0 == 0);
     const int64_t num_blocks = ne / QK5_0;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
     cpy_f32_q<cpy_blck_f32_q5_0, QK5_0><<<num_blocks, 1, 0, stream>>>
         (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
@@ -336,7 +342,7 @@ static void ggml_cpy_q5_0_f32_cuda(
     const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
     cudaStream_t stream) {
     const int64_t num_blocks = ne;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
     cpy_q_f32<cpy_blck_q_f32<dequantize_q5_0, QK5_0>, QK5_0><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
         ne10, ne11, ne12, nb10, nb11, nb12, nb13);
@@ -349,7 +355,7 @@ static void ggml_cpy_f32_q5_1_cuda(
 
     GGML_ASSERT(ne % QK5_1 == 0);
     const int64_t num_blocks = ne / QK5_1;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
     cpy_f32_q<cpy_blck_f32_q5_1, QK5_1><<<num_blocks, 1, 0, stream>>>
         (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
@@ -362,7 +368,7 @@ static void ggml_cpy_q5_1_f32_cuda(
     const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
     cudaStream_t stream) {
     const int64_t num_blocks = ne;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
     cpy_q_f32<cpy_blck_q_f32<dequantize_q5_1, QK5_1>, QK5_1><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
         ne10, ne11, ne12, nb10, nb11, nb12, nb13);
@@ -375,11 +381,51 @@ static void ggml_cpy_f32_iq4_nl_cuda(
 
     GGML_ASSERT(ne % QK4_NL == 0);
     const int64_t num_blocks = ne / QK4_NL;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
     cpy_f32_q<cpy_blck_f32_iq4_nl, QK4_NL><<<num_blocks, 1, 0, stream>>>
         (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
+// check if a same-type copy reduces to a 2D strided copy (height rows of width
+// contiguous bytes), so it can use cudaMemcpy2DAsync instead of the scalar kernel
+static bool ggml_cuda_cpy_as_memcpy_2d(const ggml_tensor * src0, const ggml_tensor * src1,
+        size_t & width, size_t & height, size_t & spitch, size_t & dpitch) {
+    // require matching shape: a reshaped copy maps elements by flat order, which the
+    // prefix walk below does not handle
+    if (src0->type != src1->type || !ggml_are_same_shape(src0, src1)) {
+        return false;
+    }
+
+    // grow the contiguous prefix block shared by both tensors
+    size_t block_nb = ggml_element_size(src0);
+    int d = 0;
+    for (; d < GGML_MAX_DIMS; ++d) {
+        if (src0->nb[d] != block_nb || src1->nb[d] != block_nb) {
+            break;
+        }
+        block_nb *= src0->ne[d];
+    }
+
+    // d == 0: nothing contiguous; d == GGML_MAX_DIMS: fully contiguous (handled by memcpy)
+    if (d == 0 || d == GGML_MAX_DIMS) {
+        return false;
+    }
+
+    // dim d carries the rows; everything above it must be a single element
+    for (int i = d + 1; i < GGML_MAX_DIMS; ++i) {
+        if (src0->ne[i] != 1) {
+            return false;
+        }
+    }
+
+    width  = block_nb;
+    height = src0->ne[d];
+    spitch = src0->nb[d];
+    dpitch = src1->nb[d];
+
+    return spitch >= width && dpitch >= width;
+}
+
 void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1) {
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
@@ -415,6 +461,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
     const bool can_be_transposed = nb01 == (int64_t)ggml_element_size(src0) &&
         src0->ne[3] == 1 && nb02 == ne00 * ne01 * (int64_t)ggml_element_size(src0);
 
+    size_t mc_width = 0, mc_height = 0, mc_spitch = 0, mc_dpitch = 0;
+
     if (src0->type == src1->type && contiguous_srcs) {
         GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1));
 #if defined(GGML_USE_MUSA) && defined(GGML_MUSA_MUDNN_COPY)
@@ -425,6 +473,9 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
         {
             CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
         }
+    } else if (ggml_cuda_cpy_as_memcpy_2d(src0, src1, mc_width, mc_height, mc_spitch, mc_dpitch)) {
+        CUDA_CHECK(cudaMemcpy2DAsync(src1_ddc, mc_dpitch, src0_ddc, mc_spitch,
+                                     mc_width, mc_height, cudaMemcpyDeviceToDevice, main_stream));
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
         if (can_be_transposed) {
             ggml_cpy_scalar_cuda<float, float, true>

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

@@ -3192,11 +3192,24 @@ static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_
     ggml_backend_buffer_t buf_src = src->view_src ? src->view_src->buffer : src->buffer;
     ggml_backend_buffer_t buf_dst = dst->view_src ? dst->view_src->buffer : dst->buffer;
 
-    if (!ggml_backend_is_cuda(backend_src) || !ggml_backend_is_cuda(backend_dst)) {
+    // Enables async copies from CPU to CUDA, instead of only CUDA-to-CUDA
+    // Excluding this path for HIP and MUSA as a precaution.
+    // According to the summary in https://github.com/ggml-org/llama.cpp/pull/20793#issuecomment-4275794315, this change is not beneficial for hip anyways.
+    // Additionally, there is a lot of anectodal evidence that hip/musa stream behavior might not always 1:1 match CUDA behavior.
+    // e.g. https://github.com/ROCm/rocm-systems/issues/5109
+    // It thus makes sense to exclude this path for HIP and MUSA. This PR was not aimed these backends, the majority of testing happened on CUDA.
+    // This can be revisited in the future if enabling copy_from_host benefits hip/MUSA, and if the PR author can extensively test on these backends.
+#if defined(GGML_USE_HIP) || defined(GGML_USE_MUSA)
+    const bool copy_from_host = false;
+#else
+    const bool copy_from_host = ggml_backend_buffer_is_host(buf_src) && ggml_backend_dev_type(backend_src->device) == GGML_BACKEND_DEVICE_TYPE_CPU;
+#endif
+
+    if (!(copy_from_host || ggml_backend_is_cuda(backend_src)) || !ggml_backend_is_cuda(backend_dst)) {
         return false;
     }
 
-    if (!ggml_backend_buffer_is_cuda(buf_src) || !ggml_backend_buffer_is_cuda(buf_dst)) {
+    if (!(copy_from_host || ggml_backend_buffer_is_cuda(buf_src)) || !ggml_backend_buffer_is_cuda(buf_dst)) {
         return false;
     }
 
@@ -3207,14 +3220,17 @@ static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_
     ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *) buf_src->context;
     ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *) buf_dst->context;
 
-    if (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device) {
+    if ((copy_from_host && cuda_ctx_dst->device != buf_ctx_dst->device) ||
+        !copy_from_host && (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device)) {
 #ifndef NDEBUG
         GGML_LOG_DEBUG("%s: backend and buffer devices do not match\n", __func__);
 #endif // NDEBUG
         return false;
     }
 
-    if (backend_src != backend_dst) {
+    if (copy_from_host) {
+        CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyHostToDevice, cuda_ctx_dst->stream()));
+    } else if (backend_src != backend_dst) {
         // copy on src stream
         if (cuda_ctx_src->device == cuda_ctx_dst->device) {
             CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream()));

ggml/src/ggml-cuda/out-prod.cu

@@ -2,6 +2,28 @@
 
 #include <cstdint>
 
+static __global__ void k_compute_out_prod_ptrs(
+        const float * src0_d, const float * src1_d, float * dst_d,
+        const float ** ptrs_a, const float ** ptrs_b, float ** ptrs_c,
+        const int64_t ne2, const int64_t ne3,
+        const int64_t dps2, const int64_t dps3,
+        const size_t s02, const size_t s03,
+        const size_t s12, const size_t s13,
+        const size_t s2,  const size_t s3) {
+    const int64_t i2 = blockIdx.x*blockDim.x + threadIdx.x;
+    const int64_t i3 = blockIdx.y*blockDim.y + threadIdx.y;
+
+    if (i2 >= ne2 || i3 >= ne3) {
+        return;
+    }
+
+    const int64_t idx = i3*ne2 + i2;
+
+    ptrs_a[idx] = src0_d + (i3/dps3)*s03 + (i2/dps2)*s02;
+    ptrs_b[idx] = src1_d +  i3      *s13 +  i2      *s12;
+    ptrs_c[idx] = dst_d  +  i3      *s3  +  i2      *s2;
+}
+
 void ggml_cuda_out_prod(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
@@ -67,18 +89,39 @@ void ggml_cuda_out_prod(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
                         &beta,  dst_d  +  i3     *s3,  ldc, s2,
                         batch_count));
         }
+    } else if (ne2 > 1 || ne3 > 1) {
+        // dps2 > 1 (src0 broadcast along dim 2 with non-uniform stride) or multiple GEMMs
+        // along dim 3: compute per-GEMM pointers on the device and use a single batched GEMM.
+        GGML_ASSERT(ne3 > 0);
+        GGML_ASSERT(ne2 <= (int64_t) std::numeric_limits<int>::max() / ne3);
+        const int batch_count = (int) (ne2 * ne3);
+
+        ggml_cuda_pool_alloc<const float *> ptrs_a(ctx.pool(), batch_count);
+        ggml_cuda_pool_alloc<const float *> ptrs_b(ctx.pool(), batch_count);
+        ggml_cuda_pool_alloc<      float *> ptrs_c(ctx.pool(), batch_count);
+
+        const dim3 block_dims(16, 16);
+        const dim3 grid_dims((ne2 + block_dims.x - 1)/block_dims.x, (ne3 + block_dims.y - 1)/block_dims.y);
+        k_compute_out_prod_ptrs<<<grid_dims, block_dims, 0, stream>>>(
+            src0_d, src1_d, dst_d,
+            ptrs_a.get(), ptrs_b.get(), ptrs_c.get(),
+            ne2, ne3, dps2, dps3, s02, s03, s12, s13, s2, s3);
+        CUDA_CHECK(cudaGetLastError());
+
+        CUBLAS_CHECK(
+            cublasSgemmBatched(handle, CUBLAS_OP_N, src1_cublas_op,
+                    ne0, ne1, ne01,
+                    &alpha, ptrs_a.get(), lda,
+                            ptrs_b.get(), ldb,
+                    &beta,  ptrs_c.get(), ldc,
+                    batch_count));
     } else {
-        // Fallback: ne2 == 1 (no batching benefit) or dps2 > 1 (src0 broadcast along dim 2
-        // with non-uniform stride; would need cublasSgemmBatched with pointer arrays).
-        for (int64_t i3 = 0; i3 < ne3; ++i3) {
-            for (int64_t i2 = 0; i2 < ne2; ++i2) {
-                CUBLAS_CHECK(
-                    cublasSgemm(handle, CUBLAS_OP_N, src1_cublas_op,
-                            ne0, ne1, ne01,
-                            &alpha, src0_d + (i3/dps3)*s03 + (i2/dps2)*s02, lda,
-                                    src1_d +  i3      *s13 +  i2      *s12, ldb,
-                            &beta,  dst_d  +  i3      *s3  +  i2      *s2,  ldc));
-            }
-        }
+        // ne2 == 1 && ne3 == 1: single GEMM
+        CUBLAS_CHECK(
+            cublasSgemm(handle, CUBLAS_OP_N, src1_cublas_op,
+                    ne0, ne1, ne01,
+                    &alpha, src0_d, lda,
+                            src1_d, ldb,
+                    &beta,  dst_d,  ldc));
     }
 }

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

@@ -48,6 +48,7 @@
 #define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS
 #define cublasSetStream hipblasSetStream
 #define cublasSgemm hipblasSgemm
+#define cublasSgemmBatched hipblasSgemmBatched
 #define cublasSgemmStridedBatched hipblasSgemmStridedBatched
 #define cublasStatus_t hipblasStatus_t
 #define cublasOperation_t hipblasOperation_t

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

@@ -32,6 +32,7 @@
 #define cublasSetMathMode mublasSetMathMode
 #define cublasSetStream mublasSetStream
 #define cublasSgemm mublasSgemm
+#define cublasSgemmBatched mublasSgemmBatched
 #define cublasSgemmStridedBatched mublasSgemmStridedBatched
 #define cublasStatus_t mublasStatus_t
 #define cublasOperation_t mublasOperation_t

ggml/src/ggml-opencl/CMakeLists.txt

@@ -192,7 +192,10 @@ set(GGML_OPENCL_KERNELS
     mul_mm_f16_f32_kq_kqv
     conv2d
     conv2d_f16_f32
+    flash_attn_pre_f16
     flash_attn_f32_f16
+    flash_attn_f32_q8_0
+    flash_attn_f32_q4_0
     flash_attn_f16
     flash_attn_f32
 )

ggml/src/ggml-opencl/fa_tune.h

@@ -0,0 +1,91 @@
+#pragma once
+
+// Flash-attention per-(dk,dv) tile tuning for the Adreno OpenCL backend.
+// Isolated from ggml-opencl.cpp so the tuning numbers are easy to find and
+// edit; the FA dispatch and kernel-compile logic stay in the main file.
+// This header is a file section — it is #included exactly once, at the point
+// in ggml-opencl.cpp where the ggml logging macros are already in scope.
+
+// Per-(dk, dv) FA config; shared by dispatch and supports_op.
+struct ggml_opencl_fa_dim {
+    int dk; int dv; int bm; int bn; int n_split; int nkv_split_threshold;
+};
+
+// Split variant fires when n_kv >= threshold (threshold=0 -> always split).
+// Default tuning covers Adreno 7xx/8xx mobile and X1-series laptop GPUs.
+static const ggml_opencl_fa_dim g_fa_dims_adreno_default[] = {
+    { 40,  40, 64, 32, 1, 0}, { 64,  64, 64, 32, 2, 64},
+    { 80,  80, 64, 32, 2, 64}, { 96,  96, 64, 32, 2, 64},
+    {112, 112, 64, 32, 2, 64}, {128, 128, 64, 32, 2, 64},
+    {192, 128, 16, 16, 1, 0},
+    {192, 192, 16, 16, 1, 0},
+    {256, 256, 16, 16, 16, 0},
+};
+
+struct ggml_opencl_fa_dim_table {
+    const ggml_opencl_fa_dim * data;
+    size_t                     count;
+
+    const ggml_opencl_fa_dim * begin() const { return data; }
+    const ggml_opencl_fa_dim * end()   const { return data + count; }
+};
+
+// Mutable copy of the active table; GGML_OPENCL_FA_TUNE patches entries here
+// at backend init without touching the const source table.
+static ggml_opencl_fa_dim g_fa_dims_runtime[
+    sizeof(g_fa_dims_adreno_default) / sizeof(g_fa_dims_adreno_default[0])];
+
+static ggml_opencl_fa_dim_table g_opencl_fa_dims = {
+    g_fa_dims_adreno_default,
+    sizeof(g_fa_dims_adreno_default) / sizeof(g_fa_dims_adreno_default[0]),
+};
+
+// GGML_OPENCL_FA_TUNE=dk:dv:bm:bn:nsplit:thr[,…] — patches matching entries
+// in the active table at backend init, before the first FA kernel compiles.
+// Unmatched (dk,dv) pairs are warned and ignored.
+static void ggml_opencl_fa_apply_env_overrides() {
+    const char * e = std::getenv("GGML_OPENCL_FA_TUNE");
+    if (!e || !e[0]) {
+        return;
+    }
+
+    std::string s = e;
+    size_t pos = 0;
+    while (pos < s.size()) {
+        size_t comma = s.find(',', pos);
+        std::string entry = s.substr(pos, comma == std::string::npos ? std::string::npos : comma - pos);
+        int dk, dv, bm, bn, nsplit, thr;
+        if (std::sscanf(entry.c_str(), "%d:%d:%d:%d:%d:%d", &dk, &dv, &bm, &bn, &nsplit, &thr) == 6) {
+            bool patched = false;
+            for (size_t i = 0; i < g_opencl_fa_dims.count; ++i) {
+                ggml_opencl_fa_dim & d = g_fa_dims_runtime[i];
+                if (d.dk == dk && d.dv == dv) {
+                    d.bm = bm; d.bn = bn; d.n_split = nsplit; d.nkv_split_threshold = thr;
+                    GGML_LOG_INFO("ggml_opencl: FA tune override DK=%d DV=%d -> bm=%d bn=%d n_split=%d thr=%d\n",
+                                  dk, dv, bm, bn, nsplit, thr);
+                    patched = true;
+                    break;
+                }
+            }
+            if (!patched) {
+                GGML_LOG_WARN("ggml_opencl: FA tune override DK=%d DV=%d ignored (no matching dim)\n", dk, dv);
+            }
+        } else {
+            GGML_LOG_WARN("ggml_opencl: FA tune override entry malformed: '%s'\n", entry.c_str());
+        }
+        if (comma == std::string::npos) break;
+        pos = comma + 1;
+    }
+}
+
+// Copy the default table into the mutable runtime buffer and apply any
+// GGML_OPENCL_FA_TUNE overrides. A per-generation table can be added here
+// once it has been tuned on hardware.
+static void ggml_cl_init_fa_dims_table() {
+    const size_t count = sizeof(g_fa_dims_adreno_default) / sizeof(g_fa_dims_adreno_default[0]);
+    for (size_t i = 0; i < count; ++i) {
+        g_fa_dims_runtime[i] = g_fa_dims_adreno_default[i];
+    }
+    g_opencl_fa_dims = { g_fa_dims_runtime, count };
+    ggml_opencl_fa_apply_env_overrides();
+}

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

@@ -1582,6 +1582,158 @@ kernel void kernel_restore_block_q8_0(
     }
 }
 
+// View-aware AoS q8_0 -> f32 dequant (f32/f32 FA path).
+kernel void kernel_dequant_q8_0_f32_view_aos(
+    global char * src,
+    ulong         src_offset,
+    ulong         src_nb1,
+    ulong         src_nb2,
+    ulong         src_nb3,
+    int           nblk0,
+    int           ne1,
+    int           ne2,
+    int           ne3,
+    global float * dst
+) {
+    int blk_i0 = get_global_id(0);
+    int i1     = get_global_id(1);
+    int batch  = get_global_id(2);
+
+    if (blk_i0 >= nblk0) return;
+    if (i1     >= ne1)   return;
+
+    int i2 = batch % ne2;
+    int i3 = batch / ne2;
+    if (i3 >= ne3) return;
+
+    global char * block = src + src_offset + (ulong)i3*src_nb3 + (ulong)i2*src_nb2 + (ulong)i1*src_nb1 + (ulong)blk_i0 * (2 + QK8_0);
+    float d = vload_half(0, (global half *)block);
+    global char * qs = block + 2;
+
+    ulong dst_row_base = ((ulong)i3 * ne2 * ne1 + (ulong)i2 * ne1 + (ulong)i1) * nblk0;
+    global float * out = dst + (dst_row_base + blk_i0) * QK8_0;
+
+    for (int i = 0; i < QK8_0; ++i) {
+        out[i] = d * (float)qs[i];
+    }
+}
+
+// View-aware AoS q8_0 -> f16 dequant. Rows tight, batch strides may be gapped.
+kernel void kernel_dequant_q8_0_f16_view_aos(
+    global char * src,
+    ulong         src_offset,
+    ulong         src_nb1,
+    ulong         src_nb2,
+    ulong         src_nb3,
+    int           nblk0,
+    int           ne1,
+    int           ne2,
+    int           ne3,
+    global half * dst
+) {
+    int blk_i0 = get_global_id(0);
+    int i1     = get_global_id(1);
+    int batch  = get_global_id(2);
+
+    if (blk_i0 >= nblk0) return;
+    if (i1     >= ne1)   return;
+
+    int i2 = batch % ne2;
+    int i3 = batch / ne2;
+    if (i3 >= ne3) return;
+
+    global char * block = src + src_offset + (ulong)i3*src_nb3 + (ulong)i2*src_nb2 + (ulong)i1*src_nb1 + (ulong)blk_i0 * (2 + QK8_0);
+    float d = vload_half(0, (global half *)block);
+    global char * qs = block + 2;
+
+    ulong dst_row_base = ((ulong)i3 * ne2 * ne1 + (ulong)i2 * ne1 + (ulong)i1) * nblk0;
+    global half * out = dst + (dst_row_base + blk_i0) * QK8_0;
+
+    for (int i = 0; i < QK8_0; ++i) {
+        out[i] = (half)(d * (float)qs[i]);
+    }
+}
+
+// View-aware AoS q4_0 -> f32 dequant (mirrors the q8_0 view variant).
+kernel void kernel_dequant_q4_0_f32_view_aos(
+    global char * src,
+    ulong         src_offset,
+    ulong         src_nb1,
+    ulong         src_nb2,
+    ulong         src_nb3,
+    int           nblk0,
+    int           ne1,
+    int           ne2,
+    int           ne3,
+    global float * dst
+) {
+    int blk_i0 = get_global_id(0);
+    int i1     = get_global_id(1);
+    int batch  = get_global_id(2);
+
+    if (blk_i0 >= nblk0) return;
+    if (i1     >= ne1)   return;
+
+    int i2 = batch % ne2;
+    int i3 = batch / ne2;
+    if (i3 >= ne3) return;
+
+    global char * block = src + src_offset + (ulong)i3*src_nb3 + (ulong)i2*src_nb2 + (ulong)i1*src_nb1 + (ulong)blk_i0 * (2 + QK4_0/2);
+    float d = vload_half(0, (global half *)block);
+    global uchar * qs = (global uchar *)(block + 2);
+
+    ulong dst_row_base = ((ulong)i3 * ne2 * ne1 + (ulong)i2 * ne1 + (ulong)i1) * nblk0;
+    global float * out = dst + (dst_row_base + blk_i0) * QK4_0;
+
+    for (int i = 0; i < QK4_0/2; ++i) {
+        uchar byte = qs[i];
+        int q0 = (int)(byte & 0x0F) - 8;
+        int q1 = (int)(byte >> 4)   - 8;
+        out[i]            = d * (float)q0;
+        out[i + QK4_0/2]  = d * (float)q1;
+    }
+}
+
+// View-aware AoS q4_0 -> f16 dequant (mirrors the q8_0 view variant).
+kernel void kernel_dequant_q4_0_f16_view_aos(
+    global char * src,
+    ulong         src_offset,
+    ulong         src_nb1,
+    ulong         src_nb2,
+    ulong         src_nb3,
+    int           nblk0,
+    int           ne1,
+    int           ne2,
+    int           ne3,
+    global half * dst
+) {
+    int blk_i0 = get_global_id(0);
+    int i1     = get_global_id(1);
+    int batch  = get_global_id(2);
+
+    if (blk_i0 >= nblk0) return;
+    if (i1     >= ne1)   return;
+
+    int i2 = batch % ne2;
+    int i3 = batch / ne2;
+    if (i3 >= ne3) return;
+
+    global char * block = src + src_offset + (ulong)i3*src_nb3 + (ulong)i2*src_nb2 + (ulong)i1*src_nb1 + (ulong)blk_i0 * (2 + QK4_0/2);
+    float d = vload_half(0, (global half *)block);
+    global uchar * qs = (global uchar *)(block + 2);
+
+    ulong dst_row_base = ((ulong)i3 * ne2 * ne1 + (ulong)i2 * ne1 + (ulong)i1) * nblk0;
+    global half * out = dst + (dst_row_base + blk_i0) * QK4_0;
+
+    for (int i = 0; i < QK4_0/2; ++i) {
+        uchar byte = qs[i];
+        int q0 = (int)(byte & 0x0F) - 8;
+        int q1 = (int)(byte >> 4)   - 8;
+        out[i]          = (half)(d * (float)q0);
+        out[i + QK4_0/2] = (half)(d * (float)q1);
+    }
+}
+
 kernel void kernel_restore_block_q8_0_trans(
     global uchar * src_q,
     global half  * src_d,

ggml/src/ggml-opencl/kernels/flash_attn_f16.cl

@@ -4,14 +4,26 @@
 #define ACC_TYPE4 float4
 #define DATA_TYPE half
 #define DATA_TYPE4 half4
-#define CONVERT_ACC4(x) convert_float4(x)
-#define CONVERT_DATA4(x) convert_half4(x)
+#define CONVERT_ACC4(x) ((float4)((float)(x).s0, (float)(x).s1, (float)(x).s2, (float)(x).s3))
+#define CONVERT_DATA4(x) ((half4)((half)(x).s0, (half)(x).s1, (half)(x).s2, (half)(x).s3))
 
 #define DK_VEC (DK/4)
 #define DV_VEC (DV/4)
 #define WG_SIZE (BLOCK_M)
 #define Q1_WG_SIZE 64
 
+// The kernels are built with -cl-finite-math-only. On some older Adreno GPUs,
+// infinite operand can cause undefined behavior and miscompilation for exp.
+// Therefore, a large negative value is used instead.
+#define FA_M_INIT (-3.0e38f)
+
+// Drop full unroll at DK>=192 — Adreno compiler host-memory budget.
+#if DK >= 192
+#define FA_UNROLL
+#else
+#define FA_UNROLL _Pragma("unroll")
+#endif
+
 inline float get_alibi_slope(
     const float max_bias, const uint h, const uint n_head_log2, const float m0, const float m1
 ) {
@@ -81,18 +93,18 @@ __kernel void flash_attn_f16(
     if (my_query_row < n_q) {
         const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2 + my_query_row * q_nb1;
         const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset);
-        #pragma unroll
+        FA_UNROLL
         for (int i = 0; i < DK_VEC; ++i) {
             q_priv[i] = CONVERT_ACC4(q_ptr[i]);
         }
     }
 
     ACC_TYPE4 o_acc[DV_VEC];
-    #pragma unroll
+    FA_UNROLL
     for (int i = 0; i < DV_VEC; ++i) {
         o_acc[i] = (ACC_TYPE4)(0.0f);
     }
-    ACC_TYPE m_i = -INFINITY;
+    ACC_TYPE m_i = FA_M_INIT;
     ACC_TYPE l_i = 0.0f;
 
     float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
@@ -125,49 +137,72 @@ __kernel void flash_attn_f16(
             continue;
         }
 
-        for (int j = 0; j < BLOCK_N; j += 2) {
+        for (int j = 0; j < BLOCK_N; j += 4) {
             const int k_row0 = k_start + j;
             const int k_row1 = k_start + j + 1;
+            const int k_row2 = k_start + j + 2;
+            const int k_row3 = k_start + j + 3;
 
             ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f);
             ACC_TYPE4 dot_acc1 = (ACC_TYPE4)(0.0f);
-            #pragma unroll
+            ACC_TYPE4 dot_acc2 = (ACC_TYPE4)(0.0f);
+            ACC_TYPE4 dot_acc3 = (ACC_TYPE4)(0.0f);
+            FA_UNROLL
             for (int k = 0; k < DK_VEC; k++) {
-                dot_acc0 = mad(q_priv[k], CONVERT_ACC4(l_k[j][k]), dot_acc0);
-                dot_acc1 = mad(q_priv[k], CONVERT_ACC4(l_k[j+1][k]), dot_acc1);
+                const ACC_TYPE4 qk = q_priv[k];
+                dot_acc0 = mad(qk, CONVERT_ACC4(l_k[j][k]),   dot_acc0);
+                dot_acc1 = mad(qk, CONVERT_ACC4(l_k[j+1][k]), dot_acc1);
+                dot_acc2 = mad(qk, CONVERT_ACC4(l_k[j+2][k]), dot_acc2);
+                dot_acc3 = mad(qk, CONVERT_ACC4(l_k[j+3][k]), dot_acc3);
             }
-            ACC_TYPE score0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale;
-            ACC_TYPE score1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale;
+            ACC_TYPE s0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale;
+            ACC_TYPE s1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale;
+            ACC_TYPE s2 = (dot_acc2.s0 + dot_acc2.s1 + dot_acc2.s2 + dot_acc2.s3) * scale;
+            ACC_TYPE s3 = (dot_acc3.s0 + dot_acc3.s1 + dot_acc3.s2 + dot_acc3.s3) * scale;
 
             if (is_causal) {
-                if (k_row0 > (n_kv - n_q + my_query_row)) score0 = -INFINITY;
-                if (k_row1 > (n_kv - n_q + my_query_row)) score1 = -INFINITY;
+                const int causal_limit = n_kv - n_q + my_query_row;
+                if (k_row0 > causal_limit) s0 = FA_M_INIT;
+                if (k_row1 > causal_limit) s1 = FA_M_INIT;
+                if (k_row2 > causal_limit) s2 = FA_M_INIT;
+                if (k_row3 > causal_limit) s3 = FA_M_INIT;
             }
-
-            if (k_row0 >= n_kv) score0 = -INFINITY;
-            if (k_row1 >= n_kv) score1 = -INFINITY;
+            if (k_row0 >= n_kv) s0 = FA_M_INIT;
+            if (k_row1 >= n_kv) s1 = FA_M_INIT;
+            if (k_row2 >= n_kv) s2 = FA_M_INIT;
+            if (k_row3 >= n_kv) s3 = FA_M_INIT;
 
             if (mask_base != NULL) {
                 const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base + my_query_row * mask_nb1);
-                if (k_row0 < n_kv) score0 += slope * (ACC_TYPE)mask_ptr[k_row0];
-                if (k_row1 < n_kv) score1 += slope * (ACC_TYPE)mask_ptr[k_row1];
+                if (k_row0 < n_kv) s0 += slope * (ACC_TYPE)mask_ptr[k_row0];
+                if (k_row1 < n_kv) s1 += slope * (ACC_TYPE)mask_ptr[k_row1];
+                if (k_row2 < n_kv) s2 += slope * (ACC_TYPE)mask_ptr[k_row2];
+                if (k_row3 < n_kv) s3 += slope * (ACC_TYPE)mask_ptr[k_row3];
             }
 
             if (logit_softcap > 0.0f) {
-                score0 = logit_softcap * tanh(score0 / logit_softcap);
-                score1 = logit_softcap * tanh(score1 / logit_softcap);
+                s0 = logit_softcap * tanh(s0 / logit_softcap);
+                s1 = logit_softcap * tanh(s1 / logit_softcap);
+                s2 = logit_softcap * tanh(s2 / logit_softcap);
+                s3 = logit_softcap * tanh(s3 / logit_softcap);
             }
 
-            const ACC_TYPE m_new = max(m_i, max(score0, score1));
-            const ACC_TYPE p0 = exp(score0 - m_new);
-            const ACC_TYPE p1 = exp(score1 - m_new);
-            const ACC_TYPE scale_prev = exp(m_i - m_new);
+            const ACC_TYPE m_new      = max(m_i, max(max(s0, s1), max(s2, s3)));
+            const ACC_TYPE scale_prev = native_exp(m_i - m_new);
+            const ACC_TYPE p0         = native_exp(s0 - m_new);
+            const ACC_TYPE p1         = native_exp(s1 - m_new);
+            const ACC_TYPE p2         = native_exp(s2 - m_new);
+            const ACC_TYPE p3         = native_exp(s3 - m_new);
 
-            #pragma unroll
+            FA_UNROLL
             for (int i = 0; i < DV_VEC; ++i) {
-                o_acc[i] = o_acc[i] * scale_prev + p0 * CONVERT_ACC4(l_v[j][i]) + p1 * CONVERT_ACC4(l_v[j+1][i]);
+                o_acc[i] = mad(p3, CONVERT_ACC4(l_v[j+3][i]),
+                           mad(p2, CONVERT_ACC4(l_v[j+2][i]),
+                           mad(p1, CONVERT_ACC4(l_v[j+1][i]),
+                           mad(p0, CONVERT_ACC4(l_v[j][i]),
+                           o_acc[i] * scale_prev))));
             }
-            l_i = l_i * scale_prev + p0 + p1;
+            l_i = l_i * scale_prev + p0 + p1 + p2 + p3;
             m_i = m_new;
         }
     }
@@ -179,7 +214,7 @@ __kernel void flash_attn_f16(
             const ACC_TYPE m_final = max(m_i, m_sink);
 
             const ACC_TYPE scale_o = exp(m_i - m_final);
-            #pragma unroll
+            FA_UNROLL
             for (int i = 0; i < DV_VEC; ++i) {
                 o_acc[i] *= scale_o;
             }
@@ -191,12 +226,12 @@ __kernel void flash_attn_f16(
         global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
         if (l_i > 0.0f) {
             const ACC_TYPE l_inv = 1.0f / l_i;
-            #pragma unroll
+            FA_UNROLL
             for (int i = 0; i < DV_VEC; ++i) {
                 o_row[i] = CONVERT_DATA4(o_acc[i] * l_inv);
             }
         } else {
-            #pragma unroll
+            FA_UNROLL
             for (int i = 0; i < DV_VEC; ++i) {
                 o_row[i] = (DATA_TYPE4)(0.0f);
             }
@@ -258,7 +293,7 @@ __kernel void flash_attn_f16_q1(
     ACC_TYPE4 q_priv[DK_VEC];
     const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2;
     const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset);
-    #pragma unroll
+    FA_UNROLL
     for (int i = 0; i < DK_VEC; ++i) {
         q_priv[i] = CONVERT_ACC4(q_ptr[i]);
     }
@@ -270,12 +305,12 @@ __kernel void flash_attn_f16_q1(
         sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset);
     }
 
-    ACC_TYPE m_i = (sinks_ptr != NULL) ? sinks_ptr[head_idx] : -INFINITY;
+    ACC_TYPE m_i = (sinks_ptr != NULL) ? sinks_ptr[head_idx] : FA_M_INIT;
     for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
         const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
         const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
         ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
-        #pragma unroll
+        FA_UNROLL
         for (int k = 0; k < DK_VEC; k++) {
             dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc);
         }
@@ -293,7 +328,7 @@ __kernel void flash_attn_f16_q1(
     __local ACC_TYPE local_m[Q1_WG_SIZE];
     local_m[tid] = m_i;
     barrier(CLK_LOCAL_MEM_FENCE);
-    #pragma unroll
+    FA_UNROLL
     for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
         if (tid < s) local_m[tid] = max(local_m[tid], local_m[tid + s]);
         barrier(CLK_LOCAL_MEM_FENCE);
@@ -301,7 +336,7 @@ __kernel void flash_attn_f16_q1(
     const ACC_TYPE m_final = local_m[0];
 
     ACC_TYPE4 o_acc[DV_VEC];
-    #pragma unroll
+    FA_UNROLL
     for (int i = 0; i < DV_VEC; ++i) o_acc[i] = (ACC_TYPE4)(0.0f);
     ACC_TYPE l_i = 0.0f;
 
@@ -311,7 +346,7 @@ __kernel void flash_attn_f16_q1(
         const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
         const global DATA_TYPE4* v_ptr = (const global DATA_TYPE4*)(v_base + v_row_offset);
         ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
-        #pragma unroll
+        FA_UNROLL
         for (int k = 0; k < DK_VEC; k++) {
             dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc);
         }
@@ -325,7 +360,7 @@ __kernel void flash_attn_f16_q1(
         }
         const ACC_TYPE p = exp(score - m_final);
         l_i += p;
-        #pragma unroll
+        FA_UNROLL
         for (int i = 0; i < DV_VEC; i++) {
             o_acc[i] = mad(p, CONVERT_ACC4(v_ptr[i]), o_acc[i]);
         }
@@ -335,7 +370,7 @@ __kernel void flash_attn_f16_q1(
     __local ACC_TYPE4 local_o_comp[Q1_WG_SIZE];
     local_l[tid] = l_i;
     barrier(CLK_LOCAL_MEM_FENCE);
-    #pragma unroll
+    FA_UNROLL
     for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
         if (tid < s) local_l[tid] += local_l[tid + s];
         barrier(CLK_LOCAL_MEM_FENCE);
@@ -354,7 +389,7 @@ __kernel void flash_attn_f16_q1(
         for (int i = 0; i < DV_VEC; i++) {
             local_o_comp[tid] = o_acc[i];
             barrier(CLK_LOCAL_MEM_FENCE);
-            #pragma unroll
+            FA_UNROLL
             for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
                 if (tid < s) local_o_comp[tid] += local_o_comp[tid + s];
                 barrier(CLK_LOCAL_MEM_FENCE);
@@ -364,7 +399,7 @@ __kernel void flash_attn_f16_q1(
             }
         }
     } else if (tid == 0) {
-        #pragma unroll
+        FA_UNROLL
         for (int i = 0; i < DV_VEC; ++i) o_row[i] = (DATA_TYPE4)(0.0f);
     }
 }

ggml/src/ggml-opencl/kernels/flash_attn_f32.cl

@@ -13,6 +13,18 @@
 #define WG_SIZE (BLOCK_M)
 #define Q1_WG_SIZE 64
 
+// The kernels are built with -cl-finite-math-only. On some older Adreno GPUs,
+// infinite operand can cause undefined behavior and miscompilation for exp.
+// Therefore, a large negative value is used instead.
+#define FA_M_INIT (-3.0e38f)
+
+// Drop full unroll at DK>=192 — Adreno compiler host-memory budget.
+#if DK >= 192
+#define FA_UNROLL
+#else
+#define FA_UNROLL _Pragma("unroll")
+#endif
+
 inline float get_alibi_slope(
     const float max_bias, const uint h, const uint n_head_log2, const float m0, const float m1
 ) {
@@ -82,18 +94,18 @@ __kernel void flash_attn_f32(
     if (my_query_row < n_q) {
         const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2 + my_query_row * q_nb1;
         const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset);
-        #pragma unroll
+        FA_UNROLL
         for (int i = 0; i < DK_VEC; ++i) {
             q_priv[i] = CONVERT_ACC4(q_ptr[i]);
         }
     }
 
     ACC_TYPE4 o_acc[DV_VEC];
-    #pragma unroll
+    FA_UNROLL
     for (int i = 0; i < DV_VEC; ++i) {
         o_acc[i] = (ACC_TYPE4)(0.0f);
     }
-    ACC_TYPE m_i = -INFINITY;
+    ACC_TYPE m_i = FA_M_INIT;
     ACC_TYPE l_i = 0.0f;
 
     float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
@@ -126,49 +138,72 @@ __kernel void flash_attn_f32(
             continue;
         }
 
-        for (int j = 0; j < BLOCK_N; j += 2) {
+        for (int j = 0; j < BLOCK_N; j += 4) {
             const int k_row0 = k_start + j;
             const int k_row1 = k_start + j + 1;
+            const int k_row2 = k_start + j + 2;
+            const int k_row3 = k_start + j + 3;
 
             ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f);
             ACC_TYPE4 dot_acc1 = (ACC_TYPE4)(0.0f);
-            #pragma unroll
+            ACC_TYPE4 dot_acc2 = (ACC_TYPE4)(0.0f);
+            ACC_TYPE4 dot_acc3 = (ACC_TYPE4)(0.0f);
+            FA_UNROLL
             for (int k = 0; k < DK_VEC; k++) {
-                dot_acc0 = mad(q_priv[k], CONVERT_ACC4(l_k[j][k]), dot_acc0);
-                dot_acc1 = mad(q_priv[k], CONVERT_ACC4(l_k[j+1][k]), dot_acc1);
+                const ACC_TYPE4 qk = q_priv[k];
+                dot_acc0 = mad(qk, CONVERT_ACC4(l_k[j][k]),   dot_acc0);
+                dot_acc1 = mad(qk, CONVERT_ACC4(l_k[j+1][k]), dot_acc1);
+                dot_acc2 = mad(qk, CONVERT_ACC4(l_k[j+2][k]), dot_acc2);
+                dot_acc3 = mad(qk, CONVERT_ACC4(l_k[j+3][k]), dot_acc3);
             }
-            ACC_TYPE score0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale;
-            ACC_TYPE score1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale;
+            ACC_TYPE s0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale;
+            ACC_TYPE s1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale;
+            ACC_TYPE s2 = (dot_acc2.s0 + dot_acc2.s1 + dot_acc2.s2 + dot_acc2.s3) * scale;
+            ACC_TYPE s3 = (dot_acc3.s0 + dot_acc3.s1 + dot_acc3.s2 + dot_acc3.s3) * scale;
 
             if (is_causal) {
-                if (k_row0 > (n_kv - n_q + my_query_row)) score0 = -INFINITY;
-                if (k_row1 > (n_kv - n_q + my_query_row)) score1 = -INFINITY;
+                const int causal_limit = n_kv - n_q + my_query_row;
+                if (k_row0 > causal_limit) s0 = FA_M_INIT;
+                if (k_row1 > causal_limit) s1 = FA_M_INIT;
+                if (k_row2 > causal_limit) s2 = FA_M_INIT;
+                if (k_row3 > causal_limit) s3 = FA_M_INIT;
             }
-
-            if (k_row0 >= n_kv) score0 = -INFINITY;
-            if (k_row1 >= n_kv) score1 = -INFINITY;
+            if (k_row0 >= n_kv) s0 = FA_M_INIT;
+            if (k_row1 >= n_kv) s1 = FA_M_INIT;
+            if (k_row2 >= n_kv) s2 = FA_M_INIT;
+            if (k_row3 >= n_kv) s3 = FA_M_INIT;
 
             if (mask_base != NULL) {
                 const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base + my_query_row * mask_nb1);
-                if (k_row0 < n_kv) score0 += slope * (ACC_TYPE)mask_ptr[k_row0];
-                if (k_row1 < n_kv) score1 += slope * (ACC_TYPE)mask_ptr[k_row1];
+                if (k_row0 < n_kv) s0 += slope * (ACC_TYPE)mask_ptr[k_row0];
+                if (k_row1 < n_kv) s1 += slope * (ACC_TYPE)mask_ptr[k_row1];
+                if (k_row2 < n_kv) s2 += slope * (ACC_TYPE)mask_ptr[k_row2];
+                if (k_row3 < n_kv) s3 += slope * (ACC_TYPE)mask_ptr[k_row3];
             }
 
             if (logit_softcap > 0.0f) {
-                score0 = logit_softcap * tanh(score0 / logit_softcap);
-                score1 = logit_softcap * tanh(score1 / logit_softcap);
+                s0 = logit_softcap * tanh(s0 / logit_softcap);
+                s1 = logit_softcap * tanh(s1 / logit_softcap);
+                s2 = logit_softcap * tanh(s2 / logit_softcap);
+                s3 = logit_softcap * tanh(s3 / logit_softcap);
             }
 
-            const ACC_TYPE m_new = max(m_i, max(score0, score1));
-            const ACC_TYPE p0 = exp(score0 - m_new);
-            const ACC_TYPE p1 = exp(score1 - m_new);
-            const ACC_TYPE scale_prev = exp(m_i - m_new);
+            const ACC_TYPE m_new      = max(m_i, max(max(s0, s1), max(s2, s3)));
+            const ACC_TYPE scale_prev = native_exp(m_i - m_new);
+            const ACC_TYPE p0         = native_exp(s0 - m_new);
+            const ACC_TYPE p1         = native_exp(s1 - m_new);
+            const ACC_TYPE p2         = native_exp(s2 - m_new);
+            const ACC_TYPE p3         = native_exp(s3 - m_new);
 
-            #pragma unroll
+            FA_UNROLL
             for (int i = 0; i < DV_VEC; ++i) {
-                o_acc[i] = o_acc[i] * scale_prev + p0 * CONVERT_ACC4(l_v[j][i]) + p1 * CONVERT_ACC4(l_v[j+1][i]);
+                o_acc[i] = mad(p3, CONVERT_ACC4(l_v[j+3][i]),
+                           mad(p2, CONVERT_ACC4(l_v[j+2][i]),
+                           mad(p1, CONVERT_ACC4(l_v[j+1][i]),
+                           mad(p0, CONVERT_ACC4(l_v[j][i]),
+                           o_acc[i] * scale_prev))));
             }
-            l_i = l_i * scale_prev + p0 + p1;
+            l_i = l_i * scale_prev + p0 + p1 + p2 + p3;
             m_i = m_new;
         }
     }
@@ -180,7 +215,7 @@ __kernel void flash_attn_f32(
             const ACC_TYPE m_final = max(m_i, m_sink);
 
             const ACC_TYPE scale_o = exp(m_i - m_final);
-            #pragma unroll
+            FA_UNROLL
             for (int i = 0; i < DV_VEC; ++i) {
                 o_acc[i] *= scale_o;
             }
@@ -192,12 +227,12 @@ __kernel void flash_attn_f32(
         global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
         if (l_i > 0.0f) {
             const ACC_TYPE l_inv = 1.0f / l_i;
-            #pragma unroll
+            FA_UNROLL
             for (int i = 0; i < DV_VEC; ++i) {
                 o_row[i] = CONVERT_DATA4(o_acc[i] * l_inv);
             }
         } else {
-            #pragma unroll
+            FA_UNROLL
             for (int i = 0; i < DV_VEC; ++i) {
                 o_row[i] = (DATA_TYPE4)(0.0f);
             }
@@ -259,7 +294,7 @@ __kernel void flash_attn_f32_q1(
     ACC_TYPE4 q_priv[DK_VEC];
     const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2;
     const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset);
-    #pragma unroll
+    FA_UNROLL
     for (int i = 0; i < DK_VEC; ++i) {
         q_priv[i] = CONVERT_ACC4(q_ptr[i]);
     }
@@ -271,12 +306,12 @@ __kernel void flash_attn_f32_q1(
         sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset);
     }
 
-    ACC_TYPE m_i = (sinks_ptr != NULL) ? sinks_ptr[head_idx] : -INFINITY;
+    ACC_TYPE m_i = (sinks_ptr != NULL) ? sinks_ptr[head_idx] : FA_M_INIT;
     for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
         const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
         const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
         ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
-        #pragma unroll
+        FA_UNROLL
         for (int k = 0; k < DK_VEC; k++) {
             dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc);
         }
@@ -294,7 +329,7 @@ __kernel void flash_attn_f32_q1(
     __local ACC_TYPE local_m[Q1_WG_SIZE];
     local_m[tid] = m_i;
     barrier(CLK_LOCAL_MEM_FENCE);
-    #pragma unroll
+    FA_UNROLL
     for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
         if (tid < s) local_m[tid] = max(local_m[tid], local_m[tid + s]);
         barrier(CLK_LOCAL_MEM_FENCE);
@@ -302,7 +337,7 @@ __kernel void flash_attn_f32_q1(
     const ACC_TYPE m_final = local_m[0];
 
     ACC_TYPE4 o_acc[DV_VEC];
-    #pragma unroll
+    FA_UNROLL
     for (int i = 0; i < DV_VEC; ++i) o_acc[i] = (ACC_TYPE4)(0.0f);
     ACC_TYPE l_i = 0.0f;
 
@@ -312,7 +347,7 @@ __kernel void flash_attn_f32_q1(
         const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
         const global DATA_TYPE4* v_ptr = (const global DATA_TYPE4*)(v_base + v_row_offset);
         ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
-        #pragma unroll
+        FA_UNROLL
         for (int k = 0; k < DK_VEC; k++) {
             dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc);
         }
@@ -326,7 +361,7 @@ __kernel void flash_attn_f32_q1(
         }
         const ACC_TYPE p = exp(score - m_final);
         l_i += p;
-        #pragma unroll
+        FA_UNROLL
         for (int i = 0; i < DV_VEC; i++) {
             o_acc[i] = mad(p, CONVERT_ACC4(v_ptr[i]), o_acc[i]);
         }
@@ -336,7 +371,7 @@ __kernel void flash_attn_f32_q1(
     __local ACC_TYPE4 local_o_comp[Q1_WG_SIZE];
     local_l[tid] = l_i;
     barrier(CLK_LOCAL_MEM_FENCE);
-    #pragma unroll
+    FA_UNROLL
     for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
         if (tid < s) local_l[tid] += local_l[tid + s];
         barrier(CLK_LOCAL_MEM_FENCE);
@@ -355,7 +390,7 @@ __kernel void flash_attn_f32_q1(
         for (int i = 0; i < DV_VEC; i++) {
             local_o_comp[tid] = o_acc[i];
             barrier(CLK_LOCAL_MEM_FENCE);
-            #pragma unroll
+            FA_UNROLL
             for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
                 if (tid < s) local_o_comp[tid] += local_o_comp[tid + s];
                 barrier(CLK_LOCAL_MEM_FENCE);
@@ -365,7 +400,7 @@ __kernel void flash_attn_f32_q1(
             }
         }
     } else if (tid == 0) {
-        #pragma unroll
+        FA_UNROLL
         for (int i = 0; i < DV_VEC; ++i) o_row[i] = (DATA_TYPE4)(0.0f);
     }
 }

ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl

@@ -1,5 +1,13 @@
 #pragma OPENCL EXTENSION cl_khr_fp16 : enable
 
+#ifdef cl_khr_subgroup_shuffle
+#pragma OPENCL EXTENSION cl_khr_subgroup_shuffle : enable
+#define HAS_SUBGROUP_SHUFFLE 1
+#elif defined(cl_qcom_subgroup_shuffle)
+#pragma OPENCL EXTENSION cl_qcom_subgroup_shuffle : enable
+#define HAS_SUBGROUP_SHUFFLE 1
+#endif
+
 #define ACC_TYPE float
 #define ACC_TYPE4 float4
 #define Q_DATA_TYPE4 float4
@@ -12,9 +20,34 @@
 
 #define DK_VEC (DK/4)
 #define DV_VEC (DV/4)
-#define WG_SIZE (BLOCK_M)
 #define Q1_WG_SIZE 64
 
+// The kernels are built with -cl-finite-math-only. On some older Adreno GPUs,
+// infinite operand can cause undefined behavior and miscompilation for exp.
+// Therefore, a large negative value is used instead.
+#define FA_M_INIT (-3.0e38f)
+
+// Drop full unroll at DK>=192 — Adreno compiler host-memory budget.
+#if DK >= 192
+#define FA_UNROLL
+#else
+#define FA_UNROLL _Pragma("unroll")
+#endif
+
+// N_SPLIT>1 splits DK/DV across threads to cut per-thread register use.
+#ifndef N_SPLIT
+#define N_SPLIT 1
+#endif
+
+#define SPLIT_DK_VEC (DK_VEC / N_SPLIT)
+#define SPLIT_DV_VEC (DV_VEC / N_SPLIT)
+
+#if N_SPLIT > 1
+#define WG_SIZE (BLOCK_M * N_SPLIT)
+#else
+#define WG_SIZE (BLOCK_M)
+#endif
+
 inline float get_alibi_slope(
     const float max_bias, const uint h, const uint n_head_log2, const float m0, const float m1
 ) {
@@ -54,19 +87,38 @@ __kernel void flash_attn_f32_f16(
     const int mask_ne2,
     const int mask_ne3,
     const global void* sinks_void,
-    const ulong sinks_offset
+    const ulong sinks_offset,
+    const global void * k_pad_void,
+    const global void * v_pad_void,
+    const global void * mask_pad_void,
+    const global char * blk,
+    const int n_kv_blocks,
+    const ulong mask_pad_nb1,
+    const ulong mask_pad_nb2,
+    const ulong mask_pad_nb3
 ) {
     const int tid = get_local_id(0);
     const int block_q_idx = get_group_id(0);
     const int head_batch_idx = get_global_id(1);
 
-    const int my_query_row = block_q_idx * BLOCK_M + tid;
+#if N_SPLIT > 1
+    const int q_lane    = tid / N_SPLIT;
+    const int split_idx = tid % N_SPLIT;
+#else
+    const int q_lane    = tid;
+    const int split_idx = 0;
+#endif
+
+    const int my_query_row = block_q_idx * BLOCK_M + q_lane;
+    const int query_valid = my_query_row < n_q;
 
     const int batch_idx = head_batch_idx / n_head;
     const int head_idx = head_batch_idx % n_head;
 
     const int gqa_ratio = n_head / n_head_kv;
     const int head_kv_idx = head_idx / gqa_ratio;
+    const int mask_head_idx = mask_void != NULL ? head_idx % mask_ne2 : 0;
+    const int mask_batch_idx = mask_void != NULL ? batch_idx % mask_ne3 : 0;
 
     const global char* q_base = (const global char*)q_void + q_offset;
     const global char* k_base = (const global char*)k_void + k_offset;
@@ -75,27 +127,41 @@ __kernel void flash_attn_f32_f16(
 
     const global char* mask_base = NULL;
     if (mask_void != NULL) {
-        const int mask_head_idx = head_idx % mask_ne2;
-        const int mask_batch_idx = batch_idx % mask_ne3;
         mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
     }
+    const global char* mask_pad_base = NULL;
+    if (mask_pad_void != NULL) {
+        mask_pad_base = (const global char*)mask_pad_void + mask_batch_idx * mask_pad_nb3 + mask_head_idx * mask_pad_nb2;
+    }
+    const global char* blk_base = NULL;
+    if (blk != NULL) {
+        const int n_q_blocks = (n_q + BLOCK_M - 1) / BLOCK_M;
+        blk_base = blk + (((mask_batch_idx * mask_ne2) + mask_head_idx) * n_q_blocks + block_q_idx) * n_kv_blocks;
+    }
 
-    ACC_TYPE4 q_priv[DK_VEC];
-    if (my_query_row < n_q) {
+    ACC_TYPE4 q_priv[SPLIT_DK_VEC];
+    const int dk_off = split_idx * SPLIT_DK_VEC;
+    if (query_valid) {
         const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2 + my_query_row * q_nb1;
         const global Q_DATA_TYPE4* q_ptr = (const global Q_DATA_TYPE4*)(q_base + q_row_offset);
-        #pragma unroll
-        for (int i = 0; i < DK_VEC; ++i) {
-            q_priv[i] = CONVERT_Q_ACC4(q_ptr[i]);
+        FA_UNROLL
+        for (int i = 0; i < SPLIT_DK_VEC; ++i) {
+            q_priv[i] = CONVERT_Q_ACC4(q_ptr[dk_off + i]);
+        }
+    } else {
+        FA_UNROLL
+        for (int i = 0; i < SPLIT_DK_VEC; ++i) {
+            q_priv[i] = (ACC_TYPE4)(0.0f);
         }
     }
 
-    ACC_TYPE4 o_acc[DV_VEC];
-    #pragma unroll
-    for (int i = 0; i < DV_VEC; ++i) {
+    ACC_TYPE4 o_acc[SPLIT_DV_VEC];
+    FA_UNROLL
+    for (int i = 0; i < SPLIT_DV_VEC; ++i) {
         o_acc[i] = (ACC_TYPE4)(0.0f);
     }
-    ACC_TYPE m_i = -INFINITY;
+
+    ACC_TYPE m_i = FA_M_INIT;
     ACC_TYPE l_i = 0.0f;
 
     float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
@@ -103,86 +169,369 @@ __kernel void flash_attn_f32_f16(
     __local KV_DATA_TYPE4 l_k[BLOCK_N][DK_VEC];
     __local KV_DATA_TYPE4 l_v[BLOCK_N][DV_VEC];
 
+#if N_SPLIT > 1 && !defined(HAS_SUBGROUP_SHUFFLE)
+    __local ACC_TYPE local_partial[BLOCK_N][WG_SIZE];
+    __local ACC_TYPE local_p[BLOCK_M][BLOCK_N];
+    __local ACC_TYPE local_softmax_scale[BLOCK_M];
+    __local ACC_TYPE local_l_inv[BLOCK_M];
+#endif
+
     for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) {
+        char blk_cur = 1;
+        if (blk_base != NULL) {
+            blk_cur = blk_base[k_start / BLOCK_N];
+            if (blk_cur == 0) continue;
+        }
+
+        const int use_kv_pad = k_pad_void != NULL && k_start + BLOCK_N > n_kv;
+        const int k_tile_start = use_kv_pad ? 0 : k_start;
+        const ulong k_tile_nb2 = use_kv_pad ? (ulong) BLOCK_N * k_nb1 : k_nb2;
+        const ulong k_tile_nb3 = use_kv_pad ? (ulong) n_head_kv * k_tile_nb2 : k_nb3;
+        const ulong v_tile_nb2 = use_kv_pad ? (ulong) BLOCK_N * v_nb1 : v_nb2;
+        const ulong v_tile_nb3 = use_kv_pad ? (ulong) n_head_kv * v_tile_nb2 : v_nb3;
+        const global char* k_tile_base = use_kv_pad ? (const global char*) k_pad_void : k_base;
+        const global char* v_tile_base = use_kv_pad ? (const global char*) v_pad_void : v_base;
+
         for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) {
             const int row = i / DK_VEC;
             const int col = i % DK_VEC;
-            const int k_row_idx = k_start + row;
-            if (k_row_idx < n_kv) {
-                const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_row_idx * k_nb1;
-                l_k[row][col] = ((__global KV_DATA_TYPE4*)(k_base + k_row_offset))[col];
+            const int k_row_idx = k_tile_start + row;
+            if (use_kv_pad || k_row_idx < n_kv) {
+                const ulong k_row_offset = batch_idx * k_tile_nb3 + head_kv_idx * k_tile_nb2 + k_row_idx * k_nb1;
+                l_k[row][col] = ((__global KV_DATA_TYPE4*)(k_tile_base + k_row_offset))[col];
+            } else {
+                l_k[row][col] = (KV_DATA_TYPE4)(0.0h);
             }
         }
         for (int i = tid; i < BLOCK_N * DV_VEC; i += WG_SIZE) {
             const int row = i / DV_VEC;
             const int col = i % DV_VEC;
-            const int v_row_idx = k_start + row;
-            if (v_row_idx < n_kv) {
-                const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + v_row_idx * v_nb1;
-                l_v[row][col] = ((__global KV_DATA_TYPE4*)(v_base + v_row_offset))[col];
+            const int v_row_idx = k_tile_start + row;
+            if (use_kv_pad || v_row_idx < n_kv) {
+                const ulong v_row_offset = batch_idx * v_tile_nb3 + head_kv_idx * v_tile_nb2 + v_row_idx * v_nb1;
+                l_v[row][col] = ((__global KV_DATA_TYPE4*)(v_tile_base + v_row_offset))[col];
+            } else {
+                l_v[row][col] = (KV_DATA_TYPE4)(0.0h);
             }
         }
         barrier(CLK_LOCAL_MEM_FENCE);
 
-        if (my_query_row >= n_q) {
-            continue;
+#if N_SPLIT > 1 && defined(HAS_SUBGROUP_SHUFFLE)
+        {
+            const int dv_off = split_idx * SPLIT_DV_VEC;
+            for (int j = 0; j < BLOCK_N; j += 2) {
+                const int k_row0 = k_start + j;
+                const int k_row1 = k_start + j + 1;
+
+                ACC_TYPE partial0 = 0.0f;
+                ACC_TYPE partial1 = 0.0f;
+                FA_UNROLL
+                for (int k = 0; k < SPLIT_DK_VEC; k++) {
+                    const ACC_TYPE4 qk = q_priv[k];
+                    ACC_TYPE4 dot0 = qk * CONVERT_KV_ACC4(l_k[j  ][dk_off + k]);
+                    ACC_TYPE4 dot1 = qk * CONVERT_KV_ACC4(l_k[j+1][dk_off + k]);
+                    partial0 += dot0.s0 + dot0.s1 + dot0.s2 + dot0.s3;
+                    partial1 += dot1.s0 + dot1.s1 + dot1.s2 + dot1.s3;
+                }
+
+                FA_UNROLL
+                for (int step = 1; step < N_SPLIT; step <<= 1) {
+                    partial0 += sub_group_shuffle_xor(partial0, step);
+                    partial1 += sub_group_shuffle_xor(partial1, step);
+                }
+
+                ACC_TYPE score0 = partial0 * scale;
+                ACC_TYPE score1 = partial1 * scale;
+
+                if (!query_valid) { score0 = FA_M_INIT; score1 = FA_M_INIT; }
+                if (is_causal) {
+                    if (k_row0 > (n_kv - n_q + my_query_row)) score0 = FA_M_INIT;
+                    if (k_row1 > (n_kv - n_q + my_query_row)) score1 = FA_M_INIT;
+                }
+                if (k_row0 >= n_kv) score0 = FA_M_INIT;
+                if (k_row1 >= n_kv) score1 = FA_M_INIT;
+
+                if (query_valid && mask_base != NULL && blk_cur != 2) {
+                    if (use_kv_pad && mask_pad_base != NULL) {
+                        const global MASK_DATA_TYPE* mask_ptr =
+                            (const global MASK_DATA_TYPE*)(mask_pad_base + my_query_row * mask_pad_nb1);
+                        score0 += slope * (ACC_TYPE)mask_ptr[j];
+                        score1 += slope * (ACC_TYPE)mask_ptr[j + 1];
+                    } else {
+                        const global MASK_DATA_TYPE* mask_ptr =
+                            (const global MASK_DATA_TYPE*)(mask_base + my_query_row * mask_nb1);
+                        if (k_row0 < n_kv) score0 += slope * (ACC_TYPE)mask_ptr[k_row0];
+                        if (k_row1 < n_kv) score1 += slope * (ACC_TYPE)mask_ptr[k_row1];
+                    }
+                }
+
+                if (logit_softcap > 0.0f) {
+                    score0 = logit_softcap * tanh(score0 / logit_softcap);
+                    score1 = logit_softcap * tanh(score1 / logit_softcap);
+                }
+
+                const ACC_TYPE m_new = max(m_i, max(score0, score1));
+                // Whole tile masked (m_new == FA_M_INIT): force the exp() args
+                // far negative so the tile contributes 0, not exp(0)=1.
+                const ACC_TYPE m_exp = (m_new == FA_M_INIT) ? 0.0f : m_new;
+                const ACC_TYPE sp    = native_exp(m_i - m_exp);
+                const ACC_TYPE p0    = native_exp(score0 - m_exp);
+                const ACC_TYPE p1    = native_exp(score1 - m_exp);
+
+                FA_UNROLL
+                for (int i = 0; i < SPLIT_DV_VEC; ++i) {
+                    o_acc[i] = o_acc[i] * sp
+                             + p0 * CONVERT_KV_ACC4(l_v[j  ][dv_off + i])
+                             + p1 * CONVERT_KV_ACC4(l_v[j+1][dv_off + i]);
+                }
+                l_i = l_i * sp + p0 + p1;
+                m_i = m_new;
+            }
         }
-
-        for (int j = 0; j < BLOCK_N; j += 2) {
-            const int k_row0 = k_start + j;
-            const int k_row1 = k_start + j + 1;
-
-            ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f);
-            ACC_TYPE4 dot_acc1 = (ACC_TYPE4)(0.0f);
-            #pragma unroll
-            for (int k = 0; k < DK_VEC; k++) {
-                dot_acc0 = mad(q_priv[k], CONVERT_KV_ACC4(l_k[j][k]), dot_acc0);
-                dot_acc1 = mad(q_priv[k], CONVERT_KV_ACC4(l_k[j+1][k]), dot_acc1);
+#elif N_SPLIT > 1
+        // N_SPLIT>1 fallback (no shuffle): 3-phase local-memory reduction.
+        // Phase 1 — partial dots for all BLOCK_N tokens.
+        for (int j = 0; j < BLOCK_N; ++j) {
+            ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+            FA_UNROLL
+            for (int k = 0; k < SPLIT_DK_VEC; k++) {
+                dot_acc = mad(q_priv[k], CONVERT_KV_ACC4(l_k[j][dk_off + k]), dot_acc);
             }
-            ACC_TYPE score0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale;
-            ACC_TYPE score1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale;
-
-            if (is_causal) {
-                if (k_row0 > (n_kv - n_q + my_query_row)) score0 = -INFINITY;
-                if (k_row1 > (n_kv - n_q + my_query_row)) score1 = -INFINITY;
-            }
-
-            if (k_row0 >= n_kv) score0 = -INFINITY;
-            if (k_row1 >= n_kv) score1 = -INFINITY;
-
-            if (mask_base != NULL) {
-                const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base + my_query_row * mask_nb1);
-                if (k_row0 < n_kv) score0 += slope * (ACC_TYPE)mask_ptr[k_row0];
-                if (k_row1 < n_kv) score1 += slope * (ACC_TYPE)mask_ptr[k_row1];
-            }
-
-            if (logit_softcap > 0.0f) {
-                score0 = logit_softcap * tanh(score0 / logit_softcap);
-                score1 = logit_softcap * tanh(score1 / logit_softcap);
-            }
-
-            const ACC_TYPE m_new = max(m_i, max(score0, score1));
-            const ACC_TYPE p0 = exp(score0 - m_new);
-            const ACC_TYPE p1 = exp(score1 - m_new);
-            const ACC_TYPE scale_prev = exp(m_i - m_new);
-
-            #pragma unroll
-            for (int i = 0; i < DV_VEC; ++i) {
-                o_acc[i] = o_acc[i] * scale_prev + p0 * CONVERT_KV_ACC4(l_v[j][i]) + p1 * CONVERT_KV_ACC4(l_v[j+1][i]);
-            }
-            l_i = l_i * scale_prev + p0 + p1;
-            m_i = m_new;
+            local_partial[j][tid] =
+                dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3;
         }
+        barrier(CLK_LOCAL_MEM_FENCE);  // 1 barrier: partial dots visible
+
+        // Phase 2 — split_idx==0 reduces partial sums and computes block softmax.
+        if (split_idx == 0) {
+            if (query_valid) {
+                ACC_TYPE m_new = m_i;
+                for (int j = 0; j < BLOCK_N; ++j) {
+                    const int k_row = k_start + j;
+                    ACC_TYPE score = 0.0f;
+                    FA_UNROLL
+                    for (int s = 0; s < N_SPLIT; s++) {
+                        score += local_partial[j][q_lane * N_SPLIT + s];
+                    }
+                    score *= scale;
+
+                    if (is_causal && k_row > (n_kv - n_q + my_query_row)) score = FA_M_INIT;
+                    if (k_row >= n_kv) score = FA_M_INIT;
+
+                    if (mask_base != NULL && blk_cur != 2) {
+                        if (use_kv_pad && mask_pad_base != NULL) {
+                            const global MASK_DATA_TYPE* mask_ptr =
+                                (const global MASK_DATA_TYPE*)(mask_pad_base + my_query_row * mask_pad_nb1);
+                            score += slope * (ACC_TYPE)mask_ptr[j];
+                        } else {
+                            const global MASK_DATA_TYPE* mask_ptr =
+                                (const global MASK_DATA_TYPE*)(mask_base + my_query_row * mask_nb1);
+                            if (k_row < n_kv) score += slope * (ACC_TYPE)mask_ptr[k_row];
+                        }
+                    }
+
+                    if (logit_softcap > 0.0f) {
+                        score = logit_softcap * tanh(score / logit_softcap);
+                    }
+
+                    m_new = max(m_new, score);
+                    local_p[q_lane][j] = score;
+                }
+
+                const ACC_TYPE m_exp = (m_new == FA_M_INIT) ? 0.0f : m_new;
+                const ACC_TYPE sp = native_exp(m_i - m_exp);
+                ACC_TYPE l_new = l_i * sp;
+                for (int j = 0; j < BLOCK_N; ++j) {
+                    const ACC_TYPE p = native_exp(local_p[q_lane][j] - m_exp);
+                    local_p[q_lane][j] = p;
+                    l_new += p;
+                }
+                local_softmax_scale[q_lane] = sp;
+                l_i = l_new;
+                m_i = m_new;
+            } else {
+                local_softmax_scale[q_lane] = 1.0f;
+                for (int j = 0; j < BLOCK_N; ++j) local_p[q_lane][j] = 0.0f;
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        // Phase 3 — V accumulate using broadcast probabilities.
+        {
+            const ACC_TYPE sp_block = local_softmax_scale[q_lane];
+            const int dv_off = split_idx * SPLIT_DV_VEC;
+            FA_UNROLL
+            for (int i = 0; i < SPLIT_DV_VEC; ++i) {
+                o_acc[i] *= sp_block;
+            }
+            for (int j = 0; j < BLOCK_N; ++j) {
+                const ACC_TYPE p = local_p[q_lane][j];
+                FA_UNROLL
+                for (int i = 0; i < SPLIT_DV_VEC; ++i) {
+                    o_acc[i] = mad(p, CONVERT_KV_ACC4(l_v[j][dv_off + i]), o_acc[i]);
+                }
+            }
+        }
+#else
+        // N_SPLIT==1: j+=4 unroll. Requires BLOCK_N % 4 == 0.
+        if (query_valid) {
+            for (int j = 0; j < BLOCK_N; j += 4) {
+                const int k_row0 = k_start + j;
+                const int k_row1 = k_start + j + 1;
+                const int k_row2 = k_start + j + 2;
+                const int k_row3 = k_start + j + 3;
+
+                ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f);
+                ACC_TYPE4 dot_acc1 = (ACC_TYPE4)(0.0f);
+                ACC_TYPE4 dot_acc2 = (ACC_TYPE4)(0.0f);
+                ACC_TYPE4 dot_acc3 = (ACC_TYPE4)(0.0f);
+                FA_UNROLL
+                for (int k = 0; k < DK_VEC; k++) {
+                    const ACC_TYPE4 qk = q_priv[k];
+                    dot_acc0 = mad(qk, CONVERT_KV_ACC4(l_k[j][k]),   dot_acc0);
+                    dot_acc1 = mad(qk, CONVERT_KV_ACC4(l_k[j+1][k]), dot_acc1);
+                    dot_acc2 = mad(qk, CONVERT_KV_ACC4(l_k[j+2][k]), dot_acc2);
+                    dot_acc3 = mad(qk, CONVERT_KV_ACC4(l_k[j+3][k]), dot_acc3);
+                }
+                ACC_TYPE s0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale;
+                ACC_TYPE s1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale;
+                ACC_TYPE s2 = (dot_acc2.s0 + dot_acc2.s1 + dot_acc2.s2 + dot_acc2.s3) * scale;
+                ACC_TYPE s3 = (dot_acc3.s0 + dot_acc3.s1 + dot_acc3.s2 + dot_acc3.s3) * scale;
+
+                if (is_causal) {
+                    const int causal_limit = n_kv - n_q + my_query_row;
+                    if (k_row0 > causal_limit) s0 = FA_M_INIT;
+                    if (k_row1 > causal_limit) s1 = FA_M_INIT;
+                    if (k_row2 > causal_limit) s2 = FA_M_INIT;
+                    if (k_row3 > causal_limit) s3 = FA_M_INIT;
+                }
+                if (k_row0 >= n_kv) s0 = FA_M_INIT;
+                if (k_row1 >= n_kv) s1 = FA_M_INIT;
+                if (k_row2 >= n_kv) s2 = FA_M_INIT;
+                if (k_row3 >= n_kv) s3 = FA_M_INIT;
+
+                if (mask_base != NULL && blk_cur != 2) {
+                    if (use_kv_pad && mask_pad_base != NULL) {
+                        const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_pad_base + my_query_row * mask_pad_nb1);
+                        s0 += slope * (ACC_TYPE)mask_ptr[j];
+                        s1 += slope * (ACC_TYPE)mask_ptr[j + 1];
+                        s2 += slope * (ACC_TYPE)mask_ptr[j + 2];
+                        s3 += slope * (ACC_TYPE)mask_ptr[j + 3];
+                    } else {
+                        const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base + my_query_row * mask_nb1);
+                        if (k_row0 < n_kv) s0 += slope * (ACC_TYPE)mask_ptr[k_row0];
+                        if (k_row1 < n_kv) s1 += slope * (ACC_TYPE)mask_ptr[k_row1];
+                        if (k_row2 < n_kv) s2 += slope * (ACC_TYPE)mask_ptr[k_row2];
+                        if (k_row3 < n_kv) s3 += slope * (ACC_TYPE)mask_ptr[k_row3];
+                    }
+                }
+
+                if (logit_softcap > 0.0f) {
+                    s0 = logit_softcap * tanh(s0 / logit_softcap);
+                    s1 = logit_softcap * tanh(s1 / logit_softcap);
+                    s2 = logit_softcap * tanh(s2 / logit_softcap);
+                    s3 = logit_softcap * tanh(s3 / logit_softcap);
+                }
+
+                const ACC_TYPE m_new      = max(m_i, max(max(s0, s1), max(s2, s3)));
+                // Whole tile masked (m_new == FA_M_INIT): force the exp() args
+                // far negative so the tile contributes 0, not exp(0)=1.
+                const ACC_TYPE m_exp      = (m_new == FA_M_INIT) ? 0.0f : m_new;
+                const ACC_TYPE scale_prev = native_exp(m_i - m_exp);
+                const ACC_TYPE p0         = native_exp(s0 - m_exp);
+                const ACC_TYPE p1         = native_exp(s1 - m_exp);
+                const ACC_TYPE p2         = native_exp(s2 - m_exp);
+                const ACC_TYPE p3         = native_exp(s3 - m_exp);
+
+                FA_UNROLL
+                for (int i = 0; i < DV_VEC; ++i) {
+                    o_acc[i] = mad(p3, CONVERT_KV_ACC4(l_v[j+3][i]),
+                               mad(p2, CONVERT_KV_ACC4(l_v[j+2][i]),
+                               mad(p1, CONVERT_KV_ACC4(l_v[j+1][i]),
+                               mad(p0, CONVERT_KV_ACC4(l_v[j][i]),
+                               o_acc[i] * scale_prev))));
+                }
+                l_i = l_i * scale_prev + p0 + p1 + p2 + p3;
+                m_i = m_new;
+            }
+        }
+#endif
+        // End of tile: every thread must finish reading l_k/l_v before the
+        // next iteration's load overwrites them (WAR hazard on local memory).
+        barrier(CLK_LOCAL_MEM_FENCE);
     }
 
-    if (my_query_row < n_q) {
+    // Write output.
+#if N_SPLIT > 1 && defined(HAS_SUBGROUP_SHUFFLE)
+    if (query_valid) {
+        ACC_TYPE sinks_sp = 1.0f;
+        if (sinks_void != NULL) {
+            const global ACC_TYPE* sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset);
+            const ACC_TYPE m_sink  = sinks_ptr[head_idx];
+            const ACC_TYPE m_final = max(m_i, m_sink);
+            sinks_sp = exp(m_i - m_final);
+            l_i = l_i * sinks_sp + exp(m_sink - m_final);
+            m_i = m_final;
+        }
+        const ACC_TYPE l_inv = (l_i > 0.0f) ? (1.0f / l_i) : 0.0f;
+        const int dv_off = split_idx * SPLIT_DV_VEC;
+        const ulong o_row_offset = batch_idx * o_nb3 + my_query_row * o_nb2 + head_idx * o_nb1;
+        global O_DATA_TYPE4 *o_row = (global O_DATA_TYPE4 *)(o_base + o_row_offset);
+        if (l_inv > 0.0f) {
+            FA_UNROLL
+            for (int i = 0; i < SPLIT_DV_VEC; ++i) {
+                o_row[dv_off + i] = CONVERT_O_DATA4(o_acc[i] * sinks_sp * l_inv);
+            }
+        } else {
+            FA_UNROLL
+            for (int i = 0; i < SPLIT_DV_VEC; ++i) {
+                o_row[dv_off + i] = (O_DATA_TYPE4)(0.0f);
+            }
+        }
+    }
+#elif N_SPLIT > 1
+    if (split_idx == 0) {
+        ACC_TYPE sinks_sp = 1.0f;
+        if (query_valid && sinks_void != NULL) {
+            const global ACC_TYPE* sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset);
+            const ACC_TYPE m_sink = sinks_ptr[head_idx];
+            const ACC_TYPE m_final = max(m_i, m_sink);
+            sinks_sp = exp(m_i - m_final);
+            l_i = l_i * sinks_sp + exp(m_sink - m_final);
+            m_i = m_final;
+        }
+        local_softmax_scale[q_lane] = sinks_sp;
+        local_l_inv[q_lane] = (query_valid && l_i > 0.0f) ? (1.0f / l_i) : 0.0f;
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (query_valid) {
+        const ACC_TYPE sinks_sp = local_softmax_scale[q_lane];
+        const ACC_TYPE l_inv    = local_l_inv[q_lane];
+        const int dv_off = split_idx * SPLIT_DV_VEC;
+        const ulong o_row_offset = batch_idx * o_nb3 + my_query_row * o_nb2 + head_idx * o_nb1;
+        global O_DATA_TYPE4 *o_row = (global O_DATA_TYPE4 *)(o_base + o_row_offset);
+        if (l_inv > 0.0f) {
+            FA_UNROLL
+            for (int i = 0; i < SPLIT_DV_VEC; ++i) {
+                o_row[dv_off + i] = CONVERT_O_DATA4(o_acc[i] * sinks_sp * l_inv);
+            }
+        } else {
+            FA_UNROLL
+            for (int i = 0; i < SPLIT_DV_VEC; ++i) {
+                o_row[dv_off + i] = (O_DATA_TYPE4)(0.0f);
+            }
+        }
+    }
+#else
+    if (query_valid) {
         if (sinks_void != NULL) {
             const global ACC_TYPE* sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset);
             const ACC_TYPE m_sink = sinks_ptr[head_idx];
             const ACC_TYPE m_final = max(m_i, m_sink);
 
             const ACC_TYPE scale_o = exp(m_i - m_final);
-            #pragma unroll
+            FA_UNROLL
             for (int i = 0; i < DV_VEC; ++i) {
                 o_acc[i] *= scale_o;
             }
@@ -194,17 +543,18 @@ __kernel void flash_attn_f32_f16(
         global O_DATA_TYPE4 *o_row = (global O_DATA_TYPE4 *)(o_base + o_row_offset);
         if (l_i > 0.0f) {
             const ACC_TYPE l_inv = 1.0f / l_i;
-            #pragma unroll
+            FA_UNROLL
             for (int i = 0; i < DV_VEC; ++i) {
                 o_row[i] = CONVERT_O_DATA4(o_acc[i] * l_inv);
             }
         } else {
-            #pragma unroll
+            FA_UNROLL
             for (int i = 0; i < DV_VEC; ++i) {
                 o_row[i] = (O_DATA_TYPE4)(0.0f);
             }
         }
     }
+#endif
 }
 
 __kernel void flash_attn_f32_f16_q1(
@@ -258,13 +608,16 @@ __kernel void flash_attn_f32_f16_q1(
         mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
     }
 
-    ACC_TYPE4 q_priv[DK_VEC];
+    // Q is uniform across WG threads (n_q=1). Share via local memory to
+    // avoid per-thread q_priv[DK_VEC] dynamic-indexed private array that
+    // spills to DDR on Adreno.
+    __local ACC_TYPE4 q_shared[DK_VEC];
     const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2;
     const global Q_DATA_TYPE4* q_ptr = (const global Q_DATA_TYPE4*)(q_base + q_row_offset);
-    #pragma unroll
-    for (int i = 0; i < DK_VEC; ++i) {
-        q_priv[i] = CONVERT_Q_ACC4(q_ptr[i]);
+    for (int i = tid; i < DK_VEC; i += Q1_WG_SIZE) {
+        q_shared[i] = CONVERT_Q_ACC4(q_ptr[i]);
     }
+    barrier(CLK_LOCAL_MEM_FENCE);
 
     float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
 
@@ -273,14 +626,14 @@ __kernel void flash_attn_f32_f16_q1(
         sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset);
     }
 
-    ACC_TYPE m_i = (sinks_ptr != NULL) ? sinks_ptr[head_idx] : -INFINITY;
+    ACC_TYPE m_i = (sinks_ptr != NULL) ? sinks_ptr[head_idx] : FA_M_INIT;
     for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
         const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
         const global KV_DATA_TYPE4* k_ptr = (const global KV_DATA_TYPE4*)(k_base + k_row_offset);
         ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
-        #pragma unroll
+        FA_UNROLL
         for (int k = 0; k < DK_VEC; k++) {
-            dot_acc = mad(q_priv[k], CONVERT_KV_ACC4(k_ptr[k]), dot_acc);
+            dot_acc = mad(q_shared[k], CONVERT_KV_ACC4(k_ptr[k]), dot_acc);
         }
         ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
         if (mask_base != NULL) {
@@ -296,7 +649,7 @@ __kernel void flash_attn_f32_f16_q1(
     __local ACC_TYPE local_m[Q1_WG_SIZE];
     local_m[tid] = m_i;
     barrier(CLK_LOCAL_MEM_FENCE);
-    #pragma unroll
+    FA_UNROLL
     for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
         if (tid < s) local_m[tid] = max(local_m[tid], local_m[tid + s]);
         barrier(CLK_LOCAL_MEM_FENCE);
@@ -304,7 +657,7 @@ __kernel void flash_attn_f32_f16_q1(
     const ACC_TYPE m_final = local_m[0];
 
     ACC_TYPE4 o_acc[DV_VEC];
-    #pragma unroll
+    FA_UNROLL
     for (int i = 0; i < DV_VEC; ++i) o_acc[i] = (ACC_TYPE4)(0.0f);
     ACC_TYPE l_i = 0.0f;
 
@@ -314,9 +667,9 @@ __kernel void flash_attn_f32_f16_q1(
         const global KV_DATA_TYPE4* k_ptr = (const global KV_DATA_TYPE4*)(k_base + k_row_offset);
         const global KV_DATA_TYPE4* v_ptr = (const global KV_DATA_TYPE4*)(v_base + v_row_offset);
         ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
-        #pragma unroll
+        FA_UNROLL
         for (int k = 0; k < DK_VEC; k++) {
-            dot_acc = mad(q_priv[k], CONVERT_KV_ACC4(k_ptr[k]), dot_acc);
+            dot_acc = mad(q_shared[k], CONVERT_KV_ACC4(k_ptr[k]), dot_acc);
         }
         ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
         if (mask_base != NULL) {
@@ -328,7 +681,7 @@ __kernel void flash_attn_f32_f16_q1(
         }
         const ACC_TYPE p = exp(score - m_final);
         l_i += p;
-        #pragma unroll
+        FA_UNROLL
         for (int i = 0; i < DV_VEC; i++) {
             o_acc[i] = mad(p, CONVERT_KV_ACC4(v_ptr[i]), o_acc[i]);
         }
@@ -338,7 +691,7 @@ __kernel void flash_attn_f32_f16_q1(
     __local ACC_TYPE4 local_o_comp[Q1_WG_SIZE];
     local_l[tid] = l_i;
     barrier(CLK_LOCAL_MEM_FENCE);
-    #pragma unroll
+    FA_UNROLL
     for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
         if (tid < s) local_l[tid] += local_l[tid + s];
         barrier(CLK_LOCAL_MEM_FENCE);
@@ -357,7 +710,7 @@ __kernel void flash_attn_f32_f16_q1(
         for (int i = 0; i < DV_VEC; i++) {
             local_o_comp[tid] = o_acc[i];
             barrier(CLK_LOCAL_MEM_FENCE);
-            #pragma unroll
+            FA_UNROLL
             for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
                 if (tid < s) local_o_comp[tid] += local_o_comp[tid + s];
                 barrier(CLK_LOCAL_MEM_FENCE);
@@ -367,7 +720,257 @@ __kernel void flash_attn_f32_f16_q1(
             }
         }
     } else if (tid == 0) {
-        #pragma unroll
+        FA_UNROLL
         for (int i = 0; i < DV_VEC; ++i) o_row[i] = (O_DATA_TYPE4)(0.0f);
     }
 }
+
+// Flash-decoding split pass. gid(2) = q_idx * n_splits + split_idx.
+// Partial record per split: [m, l, O[DV]]. Merge kernel applies sink + norm.
+#define FA_PARTIAL_FLOATS (2 + DV)
+
+__kernel void flash_attn_f32_f16_q1_split(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void * mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3,
+    global float * partial_void,
+    const int n_splits,
+    const int kv_per_split
+) {
+    const int tid              = get_local_id(0);
+    const int head_batch_idx   = get_global_id(1);
+    const int split_q_idx      = get_global_id(2);
+    const int split_idx        = split_q_idx % n_splits;
+    const int q_idx            = split_q_idx / n_splits;
+    const int batch_idx        = head_batch_idx / n_head;
+    const int head_idx         = head_batch_idx % n_head;
+    const int gqa_ratio        = n_head / n_head_kv;
+    const int head_kv_idx      = head_idx / gqa_ratio;
+
+    const int kv_start = split_idx * kv_per_split;
+    const int kv_end   = min(kv_start + kv_per_split, n_kv);
+
+    const ulong record_stride = (ulong) FA_PARTIAL_FLOATS;
+    const ulong record_idx    = ((((ulong) batch_idx * n_head + head_idx) * n_q + q_idx)
+                                 * n_splits + split_idx);
+    global float  * rec       = partial_void + record_idx * record_stride;
+    global float4 * rec_o     = (global float4 *) (rec + 2);
+
+    if (kv_start >= kv_end) {
+        // Empty split: leave sentinel partial for merge.
+        if (tid == 0) {
+            rec[0] = FA_M_INIT;
+            rec[1] = 0.0f;
+        }
+        return;
+    }
+
+    const global char * q_base = (const global char *) q_void + q_offset;
+    const global char * k_base = (const global char *) k_void + k_offset;
+    const global char * v_base = (const global char *) v_void + v_offset;
+
+    const global char * mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx  = head_idx  % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char *) mask_void + mask_offset +
+                    mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2 +
+                    (ulong) q_idx * mask_nb1;
+    }
+
+    // Share Q via local memory (n_q=1 per split -> uniform across WG).
+    __local ACC_TYPE4 q_shared[DK_VEC];
+    const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2 + (ulong) q_idx * q_nb1;
+    const global Q_DATA_TYPE4 * q_ptr = (const global Q_DATA_TYPE4 *) (q_base + q_row_offset);
+    for (int i = tid; i < DK_VEC; i += Q1_WG_SIZE) {
+        q_shared[i] = CONVERT_Q_ACC4(q_ptr[i]);
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    const float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    // Pass 1a — split-local max.
+    ACC_TYPE m_i = FA_M_INIT;
+    for (int k_idx = kv_start + tid; k_idx < kv_end; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const global KV_DATA_TYPE4 * k_ptr = (const global KV_DATA_TYPE4 *) (k_base + k_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; ++k) {
+            dot_acc = mad(q_shared[k], CONVERT_KV_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global MASK_DATA_TYPE * mask_ptr = (const global MASK_DATA_TYPE *) (mask_base);
+            score += slope * (ACC_TYPE) mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        m_i = max(m_i, score);
+    }
+
+    __local ACC_TYPE local_m[Q1_WG_SIZE];
+    local_m[tid] = m_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_m[tid] = max(local_m[tid], local_m[tid + s]);
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    const ACC_TYPE m_c = local_m[0];
+
+    // Pass 1b — softmax-weighted V accumulate.
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) o_acc[i] = (ACC_TYPE4)(0.0f);
+    ACC_TYPE l_i = 0.0f;
+
+    for (int k_idx = kv_start + tid; k_idx < kv_end; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + k_idx * v_nb1;
+        const global KV_DATA_TYPE4 * k_ptr = (const global KV_DATA_TYPE4 *) (k_base + k_row_offset);
+        const global KV_DATA_TYPE4 * v_ptr = (const global KV_DATA_TYPE4 *) (v_base + v_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; ++k) {
+            dot_acc = mad(q_shared[k], CONVERT_KV_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global MASK_DATA_TYPE * mask_ptr = (const global MASK_DATA_TYPE *) (mask_base);
+            score += slope * (ACC_TYPE) mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        const ACC_TYPE p = exp(score - m_c);
+        l_i += p;
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; ++i) {
+            o_acc[i] = mad(p, CONVERT_KV_ACC4(v_ptr[i]), o_acc[i]);
+        }
+    }
+
+    __local ACC_TYPE  local_l[Q1_WG_SIZE];
+    __local ACC_TYPE4 local_o[Q1_WG_SIZE];
+    local_l[tid] = l_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_l[tid] += local_l[tid + s];
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    const ACC_TYPE l_c = local_l[0];
+
+    if (tid == 0) {
+        rec[0] = (float) m_c;
+        rec[1] = (float) l_c;
+    }
+    for (int i = 0; i < DV_VEC; ++i) {
+        local_o[tid] = o_acc[i];
+        barrier(CLK_LOCAL_MEM_FENCE);
+        #pragma unroll
+        for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+            if (tid < s) local_o[tid] += local_o[tid + s];
+            barrier(CLK_LOCAL_MEM_FENCE);
+        }
+        if (tid == 0) {
+            rec_o[i] = local_o[0];
+        }
+    }
+}
+
+// FD Pass 2: merge per-split partials into final O. Empty splits drop via exp(-INF)=0.
+__kernel void flash_attn_f32_merge(
+    const global float * partial_void,
+    global void * o_void,
+    const ulong o_offset,
+    const int n_head,
+    const int n_splits,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const global void * sinks_void,
+    const ulong sinks_offset,
+    const int n_q
+) {
+    const int lane           = get_local_id(0);  // 0..DV_VEC-1
+    const int head_batch_idx = get_global_id(1);
+    const int q_idx          = get_global_id(2);
+    const int batch_idx      = head_batch_idx / n_head;
+    const int head_idx       = head_batch_idx % n_head;
+
+    const ulong record_stride = (ulong) FA_PARTIAL_FLOATS;
+    const ulong record_idx_0  = (((ulong) batch_idx * n_head + head_idx) * n_q + q_idx) * n_splits;
+    const global float * rec0 = partial_void + record_idx_0 * record_stride;
+
+    __local ACC_TYPE m_final_shared;
+    __local ACC_TYPE l_final_shared;
+    if (lane == 0) {
+        ACC_TYPE m = FA_M_INIT;
+        for (int c = 0; c < n_splits; ++c) {
+            const ACC_TYPE m_c = rec0[c * record_stride + 0];
+            m = max(m, m_c);
+        }
+        ACC_TYPE m_sink = 0.0f;
+        bool has_sink = false;
+        if (sinks_void != NULL) {
+            const global ACC_TYPE * sinks_ptr =
+                (const global ACC_TYPE *) ((const global char *) sinks_void + sinks_offset);
+            m_sink = sinks_ptr[head_idx];
+            has_sink = true;
+            m = max(m, m_sink);
+        }
+        ACC_TYPE l = 0.0f;
+        for (int c = 0; c < n_splits; ++c) {
+            const ACC_TYPE m_c = rec0[c * record_stride + 0];
+            const ACC_TYPE l_c = rec0[c * record_stride + 1];
+            if (m_c > FA_M_INIT) {
+                l += l_c * exp(m_c - m);
+            }
+        }
+        if (has_sink) {
+            l += exp(m_sink - m);
+        }
+        m_final_shared = m;
+        l_final_shared = l;
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+    const ACC_TYPE m_final = m_final_shared;
+    const ACC_TYPE l_final = l_final_shared;
+    const ACC_TYPE l_inv   = (l_final > 0.0f) ? (1.0f / l_final) : 0.0f;
+
+    ACC_TYPE4 o = (ACC_TYPE4)(0.0f);
+    for (int c = 0; c < n_splits; ++c) {
+        const global float * rec_c   = rec0 + c * record_stride;
+        const ACC_TYPE       m_c     = rec_c[0];
+        if (m_c <= FA_M_INIT) continue;
+        const global float4 * rec_oc = (const global float4 *) (rec_c + 2);
+        const ACC_TYPE scale_c = exp(m_c - m_final);
+        o = mad((ACC_TYPE4)(scale_c), rec_oc[lane], o);
+    }
+    o = o * l_inv;
+
+    const ulong o_row_offset = (ulong) batch_idx * o_nb3 + (ulong) q_idx * o_nb2 + (ulong) head_idx * o_nb1;
+    global O_DATA_TYPE4 * o_row = (global O_DATA_TYPE4 *) ((global char *) o_void + o_offset + o_row_offset);
+    o_row[lane] = CONVERT_O_DATA4(o);
+}

ggml/src/ggml-opencl/kernels/flash_attn_pre_f16.cl

@@ -0,0 +1,156 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+__kernel void flash_attn_kv_pad_f16(
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * k_pad_void,
+    global void * v_pad_void,
+    const int n_kv,
+    const int n_head_kv,
+    const int n_batch,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3
+) {
+    const int row_idx = get_global_id(0);
+    const int head_kv_idx = get_global_id(1);
+    const int batch_idx = get_global_id(2);
+
+    if (row_idx >= BLOCK_N || head_kv_idx >= n_head_kv || batch_idx >= n_batch) {
+        return;
+    }
+
+    const int tail_start = n_kv - (n_kv % BLOCK_N);
+    const int src_row_idx = tail_start + row_idx;
+
+    const global char * k_src = (const global char *) k_void + k_offset;
+    const global char * v_src = (const global char *) v_void + v_offset;
+    global char * k_pad = (global char *) k_pad_void;
+    global char * v_pad = (global char *) v_pad_void;
+
+    const ulong k_dst_offset = ((ulong) batch_idx * (ulong) n_head_kv + (ulong) head_kv_idx) * ((ulong) BLOCK_N * k_nb1) + (ulong) row_idx * k_nb1;
+    const ulong v_dst_offset = ((ulong) batch_idx * (ulong) n_head_kv + (ulong) head_kv_idx) * ((ulong) BLOCK_N * v_nb1) + (ulong) row_idx * v_nb1;
+
+    if (src_row_idx < n_kv) {
+        const ulong k_src_offset = (ulong) batch_idx * k_nb3 + (ulong) head_kv_idx * k_nb2 + (ulong) src_row_idx * k_nb1;
+        const ulong v_src_offset = (ulong) batch_idx * v_nb3 + (ulong) head_kv_idx * v_nb2 + (ulong) src_row_idx * v_nb1;
+
+        for (ulong i = 0; i < k_nb1; ++i) {
+            k_pad[k_dst_offset + i] = k_src[k_src_offset + i];
+        }
+        for (ulong i = 0; i < v_nb1; ++i) {
+            v_pad[v_dst_offset + i] = v_src[v_src_offset + i];
+        }
+    } else {
+        for (ulong i = 0; i < k_nb1; ++i) {
+            k_pad[k_dst_offset + i] = 0;
+        }
+        for (ulong i = 0; i < v_nb1; ++i) {
+            v_pad[v_dst_offset + i] = 0;
+        }
+    }
+}
+
+__kernel void flash_attn_mask_pad_f16(
+    const global void * mask_void, ulong mask_offset,
+    global void * mask_pad_void,
+    const int n_q,
+    const int n_kv,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3
+) {
+    const int col_idx = get_global_id(0);
+    const int q_row = get_global_id(1);
+    const int mask_slice = get_global_id(2);
+
+    if (col_idx >= BLOCK_N || q_row >= n_q || mask_slice >= mask_ne2 * mask_ne3) {
+        return;
+    }
+
+    const int tail_start = n_kv - (n_kv % BLOCK_N);
+    const int src_col_idx = tail_start + col_idx;
+    const int mask_head_idx = mask_slice % mask_ne2;
+    const int mask_batch_idx = mask_slice / mask_ne2;
+
+    const global char * mask_src_base = (const global char *) mask_void + mask_offset +
+        (ulong) mask_batch_idx * mask_nb3 +
+        (ulong) mask_head_idx * mask_nb2 +
+        (ulong) q_row * mask_nb1;
+    const global half * mask_src = (const global half *) mask_src_base;
+
+    global half * mask_pad = (global half *) mask_pad_void;
+    const ulong dst_idx =
+        (((ulong) mask_batch_idx * (ulong) mask_ne2 + (ulong) mask_head_idx) * (ulong) n_q + (ulong) q_row) * (ulong) BLOCK_N +
+        (ulong) col_idx;
+
+    mask_pad[dst_idx] = src_col_idx < n_kv ? mask_src[src_col_idx] : (half) (-INFINITY);
+}
+
+// Per-KV-tile mask class. 0=all -inf (skip tile), 1=mixed (apply mask),
+// 2=all zero, no -inf (skip mask lookup). Causal diagonal tiles are class 1.
+__kernel void flash_attn_blk_f16(
+    const global void * mask_void, ulong mask_offset,
+    global char * blk,
+    const int n_q,
+    const int n_kv,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3
+) {
+    const int kv_block_idx = get_global_id(0);
+    const int q_block_idx = get_global_id(1);
+    const int mask_slice = get_global_id(2);
+
+    const int n_q_blocks = (n_q + BLOCK_M - 1) / BLOCK_M;
+    const int n_kv_blocks = (n_kv + BLOCK_N - 1) / BLOCK_N;
+    if (kv_block_idx >= n_kv_blocks || q_block_idx >= n_q_blocks || mask_slice >= mask_ne2 * mask_ne3) {
+        return;
+    }
+
+    const int mask_head_idx = mask_slice % mask_ne2;
+    const int mask_batch_idx = mask_slice / mask_ne2;
+    const int q_start = q_block_idx * BLOCK_M;
+    const int k_start = kv_block_idx * BLOCK_N;
+    const int q_count = min(BLOCK_M, n_q - q_start);
+    const int k_count = min(BLOCK_N, n_kv - k_start);
+
+    const half neg_max_half = (half) (-65504.0f);
+    char has_unmasked = 0;
+    char has_masked = 0;
+    char has_nonzero = 0;
+
+    const global char * mask_base = (const global char *) mask_void + mask_offset +
+        (ulong) mask_batch_idx * mask_nb3 +
+        (ulong) mask_head_idx * mask_nb2;
+
+    for (int qi = 0; qi < q_count; ++qi) {
+        const global half * mask_row = (const global half *) (mask_base + (ulong) (q_start + qi) * mask_nb1) + k_start;
+        for (int ki = 0; ki < k_count; ++ki) {
+            const half v = mask_row[ki];
+            if (v <= neg_max_half) {
+                has_masked = 1;
+            } else {
+                has_unmasked = 1;
+                if (v != (half) 0.0f) {
+                    has_nonzero = 1;
+                }
+            }
+        }
+        if (has_masked && has_unmasked) break;  // mixed tile — short-circuit.
+    }
+
+    char res;
+    if (has_unmasked == 0) {
+        res = 0;
+    } else if (has_masked || has_nonzero) {
+        res = 1;
+    } else {
+        res = 2;
+    }
+
+    blk[((ulong) mask_slice * (ulong) n_q_blocks + (ulong) q_block_idx) * (ulong) n_kv_blocks + (ulong) kv_block_idx] = res;
+}

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

@@ -24,6 +24,7 @@ kernel void kernel_norm(
         int ne01,
         int ne02,
         int ne03,
+        ulong nb00,
         ulong nb01,
         ulong nb02,
         ulong nb03,
@@ -43,7 +44,8 @@ kernel void kernel_norm(
     // parallel sum
     sum[get_local_id(0)] = 0.0f;
     for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
-        sum[get_local_id(0)] += x[i00];
+        // this kernel handles float, nb00/4 translates byte offset to element offset
+        sum[get_local_id(0)] += x[i00*nb00/4];
     }
     // reduce
     barrier(CLK_LOCAL_MEM_FENCE);
@@ -60,7 +62,8 @@ kernel void kernel_norm(
     global float * y = dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
     sum[get_local_id(0)] = 0.0f;
     for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
-        y[i00] = x[i00] - mean;
+        // this kernel handles float, nb00/4 translates byte offset to element offset
+        y[i00] = x[i00*nb00/4] - mean;
         sum[get_local_id(0)] += y[i00] * y[i00];
     }
 

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

@@ -158,6 +158,239 @@ kernel void kernel_set_rows_f32_i32(
     }
 }
 
+// f32 -> q8_0 quantize set_rows. Block = half d + char qs[32].
+#define QK8_0 32
+
+inline void quantize_q8_0_block(global float * x, global char * qs, global half * d_out) {
+    float amax = 0.0f;
+    for (int j = 0; j < QK8_0; j++) {
+        amax = fmax(amax, fabs(x[j]));
+    }
+
+    float d  = amax / 127.0f;
+    float id = (d != 0.0f) ? 127.0f / amax : 0.0f;
+
+    vstore_half(d, 0, d_out);
+
+    for (int j = 0; j < QK8_0; j++) {
+        qs[j] = (char)((int)round(x[j] * id));
+    }
+}
+
+kernel void kernel_set_rows_q8_0_i64(
+        global char * src0,
+        ulong         offset0,
+        global char * src1,
+        ulong         offset1,
+        global char * dst,
+        ulong         offsetd,
+        int           ne01,
+        ulong         nb01,
+        ulong         nb02,
+        ulong         nb03,
+        uint4         ne11,
+        uint4         ne12,
+        ulong         nb10,
+        ulong         nb11,
+        ulong         nb12,
+        int           nblk0,
+        ulong         nb1,
+        ulong         nb2,
+        ulong         nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
+
+    if (i01 >= ne01) {
+        return;
+    }
+
+    int i12 = fastmod(i03, ne12);
+    int i11 = fastmod(i02, ne11);
+
+    int i10 = i01;
+    long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
+
+    global char  * dst_row = (global char  *) (dst  +  i1*nb1  + i02*nb2  + i03*nb3);
+    global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
+
+    for (int blk = get_local_id(0); blk < nblk0; blk += get_local_size(0)) {
+        global float * x = src_row + blk * QK8_0;
+        global char  * y = dst_row + blk * (2 + QK8_0);
+
+        quantize_q8_0_block(x, y + 2, (global half *)y);
+    }
+}
+
+kernel void kernel_set_rows_q8_0_i32(
+        global char * src0,
+        ulong         offset0,
+        global char * src1,
+        ulong         offset1,
+        global char * dst,
+        ulong         offsetd,
+        int           ne01,
+        ulong         nb01,
+        ulong         nb02,
+        ulong         nb03,
+        uint4         ne11,
+        uint4         ne12,
+        ulong         nb10,
+        ulong         nb11,
+        ulong         nb12,
+        int           nblk0,
+        ulong         nb1,
+        ulong         nb2,
+        ulong         nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
+
+    if (i01 >= ne01) {
+        return;
+    }
+
+    int i12 = fastmod(i03, ne12);
+    int i11 = fastmod(i02, ne11);
+
+    int i10 = i01;
+    int i1  = ((global int *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
+
+    global char  * dst_row = (global char  *) (dst  +  i1*nb1  + i02*nb2  + i03*nb3);
+    global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
+
+    for (int blk = get_local_id(0); blk < nblk0; blk += get_local_size(0)) {
+        global float * x = src_row + blk * QK8_0;
+        global char  * y = dst_row + blk * (2 + QK8_0);
+
+        quantize_q8_0_block(x, y + 2, (global half *)y);
+    }
+}
+
+// SoA q8_0 variants. dst_q: int8[QK8_0] per block; dst_d: fp16 scale per block.
+// Layout matches kernel_convert_block_q8_0; block index follows dst element order.
+kernel void kernel_set_rows_q8_0_soa_i64(
+        global char * src0,
+        ulong         offset0,
+        global char * src1,
+        ulong         offset1,
+        global char * dst_q,
+        ulong         offset_q,
+        global char * dst_d,
+        ulong         offset_d,
+        int           ne01,
+        ulong         nb01,
+        ulong         nb02,
+        ulong         nb03,
+        uint4         ne11,
+        uint4         ne12,
+        ulong         nb10,
+        ulong         nb11,
+        ulong         nb12,
+        int           nblk0,
+        int           ne1_dst,
+        int           ne2_dst,
+        int           ne3_dst
+) {
+    src0  = src0  + offset0;
+    src1  = src1  + offset1;
+    dst_q = dst_q + offset_q;
+    dst_d = dst_d + offset_d;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
+
+    if (i01 >= ne01) {
+        return;
+    }
+
+    int i12 = fastmod(i03, ne12);
+    int i11 = fastmod(i02, ne11);
+
+    int i10 = i01;
+    long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
+
+    long row_blk_base = ((long)i03 * ne2_dst * ne1_dst + (long)i02 * ne1_dst + i1) * nblk0;
+
+    global half  * d_row = (global half  *)(dst_d) + row_blk_base;
+    global char  * q_row = (global char  *)(dst_q) + row_blk_base * QK8_0;
+    global float * src_row = (global float *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
+
+    for (int blk = get_local_id(0); blk < nblk0; blk += get_local_size(0)) {
+        global float * x = src_row + blk * QK8_0;
+        global char  * q = q_row + blk * QK8_0;
+
+        quantize_q8_0_block(x, q, d_row + blk);
+    }
+}
+
+kernel void kernel_set_rows_q8_0_soa_i32(
+        global char * src0,
+        ulong         offset0,
+        global char * src1,
+        ulong         offset1,
+        global char * dst_q,
+        ulong         offset_q,
+        global char * dst_d,
+        ulong         offset_d,
+        int           ne01,
+        ulong         nb01,
+        ulong         nb02,
+        ulong         nb03,
+        uint4         ne11,
+        uint4         ne12,
+        ulong         nb10,
+        ulong         nb11,
+        ulong         nb12,
+        int           nblk0,
+        int           ne1_dst,
+        int           ne2_dst,
+        int           ne3_dst
+) {
+    src0  = src0  + offset0;
+    src1  = src1  + offset1;
+    dst_q = dst_q + offset_q;
+    dst_d = dst_d + offset_d;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
+
+    if (i01 >= ne01) {
+        return;
+    }
+
+    int i12 = fastmod(i03, ne12);
+    int i11 = fastmod(i02, ne11);
+
+    int i10 = i01;
+    int i1  = ((global int *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
+
+    long row_blk_base = ((long)i03 * ne2_dst * ne1_dst + (long)i02 * ne1_dst + i1) * nblk0;
+
+    global half  * d_row = (global half  *)(dst_d) + row_blk_base;
+    global char  * q_row = (global char  *)(dst_q) + row_blk_base * QK8_0;
+    global float * src_row = (global float *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
+
+    for (int blk = get_local_id(0); blk < nblk0; blk += get_local_size(0)) {
+        global float * x = src_row + blk * QK8_0;
+        global char  * q = q_row + blk * QK8_0;
+
+        quantize_q8_0_block(x, q, d_row + blk);
+    }
+}
+
 kernel void kernel_set_rows_f16_i32(
         global char * src0,
         ulong         offset0,
@@ -206,3 +439,270 @@ kernel void kernel_set_rows_f16_i32(
         dst_row[ind] = src_row[ind];
     }
 }
+
+// f32 -> q4_0 quantize set_rows. Block = half d + uchar qs[16] (shuffled
+// nibbles: qs[j] low/high = elem j / j+16).
+// Dequant: val[i] = d * (nibble_i - 8)
+// nblk0 = number of q4_0 blocks per row = ne00 / 32.
+#define QK4_0 32
+#define Q4_0_BLOCK_SIZE 18
+
+inline void quantize_q4_0_block(global float * x, global uchar * qs, global half * d_out) {
+    // Find the signed value with the largest absolute magnitude (matches ggml ref).
+    float max  = 0.0f;
+    float amax = 0.0f;
+    for (int j = 0; j < QK4_0; j++) {
+        float v = x[j];
+        float a = fabs(v);
+        if (a > amax) {
+            amax = a;
+            max  = v;
+        }
+    }
+
+    float d  = max / -8.0f;
+    float id = (d != 0.0f) ? 1.0f / d : 0.0f;
+
+    vstore_half(d, 0, d_out);
+
+    for (int j = 0; j < QK4_0/2; j++) {
+        float x0 = x[j]           * id;
+        float x1 = x[j + QK4_0/2] * id;
+
+        int i0 = (int)(x0 + 8.5f);
+        int i1 = (int)(x1 + 8.5f);
+        if (i0 < 0)  i0 = 0;
+        if (i0 > 15) i0 = 15;
+        if (i1 < 0)  i1 = 0;
+        if (i1 > 15) i1 = 15;
+
+        qs[j] = (uchar)i0 | ((uchar)i1 << 4);
+    }
+}
+
+kernel void kernel_set_rows_q4_0_i64(
+        global char * src0,
+        ulong         offset0,
+        global char * src1,
+        ulong         offset1,
+        global char * dst,
+        ulong         offsetd,
+        int           ne01,
+        ulong         nb01,
+        ulong         nb02,
+        ulong         nb03,
+        uint4         ne11,
+        uint4         ne12,
+        ulong         nb10,
+        ulong         nb11,
+        ulong         nb12,
+        int           nblk0,
+        ulong         nb1,
+        ulong         nb2,
+        ulong         nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
+
+    if (i01 >= ne01) {
+        return;
+    }
+
+    int i12 = fastmod(i03, ne12);
+    int i11 = fastmod(i02, ne11);
+
+    int i10 = i01;
+    long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
+
+    global char  * dst_row = (global char  *) (dst  +  i1*nb1  + i02*nb2  + i03*nb3);
+    global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
+
+    for (int blk = get_local_id(0); blk < nblk0; blk += get_local_size(0)) {
+        global float * x    = src_row + blk * QK4_0;
+        global char  * y    = dst_row + blk * Q4_0_BLOCK_SIZE;
+        global half  * yd   = (global half  *)(y);
+        global uchar * yqs  = (global uchar *)(y + 2);
+
+        quantize_q4_0_block(x, yqs, yd);
+    }
+}
+
+kernel void kernel_set_rows_q4_0_i32(
+        global char * src0,
+        ulong         offset0,
+        global char * src1,
+        ulong         offset1,
+        global char * dst,
+        ulong         offsetd,
+        int           ne01,
+        ulong         nb01,
+        ulong         nb02,
+        ulong         nb03,
+        uint4         ne11,
+        uint4         ne12,
+        ulong         nb10,
+        ulong         nb11,
+        ulong         nb12,
+        int           nblk0,
+        ulong         nb1,
+        ulong         nb2,
+        ulong         nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
+
+    if (i01 >= ne01) {
+        return;
+    }
+
+    int i12 = fastmod(i03, ne12);
+    int i11 = fastmod(i02, ne11);
+
+    int i10 = i01;
+    int i1  = ((global int *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
+
+    global char  * dst_row = (global char  *) (dst  +  i1*nb1  + i02*nb2  + i03*nb3);
+    global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
+
+    for (int blk = get_local_id(0); blk < nblk0; blk += get_local_size(0)) {
+        global float * x    = src_row + blk * QK4_0;
+        global char  * y    = dst_row + blk * Q4_0_BLOCK_SIZE;
+        global half  * yd   = (global half  *)(y);
+        global uchar * yqs  = (global uchar *)(y + 2);
+
+        quantize_q4_0_block(x, yqs, yd);
+    }
+}
+
+// SoA variants for q4_0 dst. Used when the backend has split block_q4_0 records
+// into separate quant (dst_q) and scale (dst_d) sub-buffers — same pattern as
+// the q8_0 SoA variants above.
+//
+// Layout (matches kernel_convert_block_q4_0, the "shuffled" variant):
+//   dst_q: contiguous 16 packed nibbles per block, block i at offset i * 16 bytes.
+//   dst_d: contiguous fp16 scales, block i at offset i * 2 bytes.
+// Nibble layout inside each byte is unchanged from AoS: qs[j] low nibble = element j,
+// qs[j] high nibble = element j+16. kernel_restore_block_q4_0 copies bytes as-is.
+kernel void kernel_set_rows_q4_0_soa_i64(
+        global char * src0,
+        ulong         offset0,
+        global char * src1,
+        ulong         offset1,
+        global char * dst_q,
+        ulong         offset_q,
+        global char * dst_d,
+        ulong         offset_d,
+        int           ne01,
+        ulong         nb01,
+        ulong         nb02,
+        ulong         nb03,
+        uint4         ne11,
+        uint4         ne12,
+        ulong         nb10,
+        ulong         nb11,
+        ulong         nb12,
+        int           nblk0,
+        int           ne1_dst,
+        int           ne2_dst,
+        int           ne3_dst
+) {
+    src0  = src0  + offset0;
+    src1  = src1  + offset1;
+    dst_q = dst_q + offset_q;
+    dst_d = dst_d + offset_d;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
+
+    if (i01 >= ne01) {
+        return;
+    }
+
+    int i12 = fastmod(i03, ne12);
+    int i11 = fastmod(i02, ne11);
+
+    int i10 = i01;
+    long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
+
+    long row_blk_base = ((long)i03 * ne2_dst * ne1_dst + (long)i02 * ne1_dst + i1) * nblk0;
+
+    global half  * d_row   = (global half  *)(dst_d) + row_blk_base;
+    global uchar * q_row   = (global uchar *)(dst_q) + row_blk_base * (QK4_0/2);
+    global float * src_row = (global float *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
+
+    for (int blk = get_local_id(0); blk < nblk0; blk += get_local_size(0)) {
+        global float * x    = src_row + blk * QK4_0;
+        global uchar * qs   = q_row   + blk * (QK4_0/2);
+        global half  * d_bk = d_row   + blk;
+
+        quantize_q4_0_block(x, qs, d_bk);
+    }
+}
+
+kernel void kernel_set_rows_q4_0_soa_i32(
+        global char * src0,
+        ulong         offset0,
+        global char * src1,
+        ulong         offset1,
+        global char * dst_q,
+        ulong         offset_q,
+        global char * dst_d,
+        ulong         offset_d,
+        int           ne01,
+        ulong         nb01,
+        ulong         nb02,
+        ulong         nb03,
+        uint4         ne11,
+        uint4         ne12,
+        ulong         nb10,
+        ulong         nb11,
+        ulong         nb12,
+        int           nblk0,
+        int           ne1_dst,
+        int           ne2_dst,
+        int           ne3_dst
+) {
+    src0  = src0  + offset0;
+    src1  = src1  + offset1;
+    dst_q = dst_q + offset_q;
+    dst_d = dst_d + offset_d;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
+
+    if (i01 >= ne01) {
+        return;
+    }
+
+    int i12 = fastmod(i03, ne12);
+    int i11 = fastmod(i02, ne11);
+
+    int i10 = i01;
+    int i1  = ((global int *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
+
+    long row_blk_base = ((long)i03 * ne2_dst * ne1_dst + (long)i02 * ne1_dst + i1) * nblk0;
+
+    global half  * d_row   = (global half  *)(dst_d) + row_blk_base;
+    global uchar * q_row   = (global uchar *)(dst_q) + row_blk_base * (QK4_0/2);
+    global float * src_row = (global float *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
+
+    for (int blk = get_local_id(0); blk < nblk0; blk += get_local_size(0)) {
+        global float * x    = src_row + blk * QK4_0;
+        global uchar * qs   = q_row   + blk * (QK4_0/2);
+        global half  * d_bk = d_row   + blk;
+
+        quantize_q4_0_block(x, qs, d_bk);
+    }
+}

ggml/src/ggml-openvino/ggml-decoder.cpp

@@ -1270,77 +1270,14 @@ void GgmlOvDecoder::visit_subgraph(std::function<void(std::shared_ptr<GgmlDecode
 }
 
 std::string GgmlOvDecoder::compute_op_type(const ggml_tensor * node) {
-    static const std::map<ggml_op, std::string> ops = {
-        {GGML_OP_NONE,            "GGML_OP_NONE"           },
-        {GGML_OP_ACC,             "GGML_OP_ACC"            },
-        {GGML_OP_ADD,             "GGML_OP_ADD"            },
-        {GGML_OP_ADD1,            "GGML_OP_ADD1"           },
-        {GGML_OP_ADD_ID,          "GGML_OP_ADD_ID"         },
-        {GGML_OP_CONCAT,          "GGML_OP_CONCAT"         },
-        {GGML_OP_CONT,            "GGML_OP_CONT"           },
-        {GGML_OP_DIV,             "GGML_OP_DIV"            },
-        {GGML_OP_DUP,             "GGML_OP_DUP"            },
-        {GGML_OP_GET_ROWS,        "GGML_OP_GET_ROWS"       },
-        {GGML_OP_MUL,             "GGML_OP_MUL"            },
-        {GGML_OP_MUL_MAT,         "GGML_OP_MUL_MAT"        },
-        {GGML_OP_MUL_MAT_ID,      "GGML_OP_MUL_MAT_ID"     },
-        {GGML_OP_PERMUTE,         "GGML_OP_PERMUTE"        },
-        {GGML_OP_RESHAPE,         "GGML_OP_RESHAPE"        },
-        {GGML_OP_RMS_NORM,        "GGML_OP_RMS_NORM"       },
-        {GGML_OP_NORM,            "GGML_OP_NORM"           },
-        {GGML_OP_ROPE,            "GGML_OP_ROPE"           },
-        {GGML_OP_SCALE,           "GGML_OP_SCALE"          },
-        {GGML_OP_SOFT_MAX,        "GGML_OP_SOFT_MAX"       },
-        {GGML_OP_SUM_ROWS,        "GGML_OP_SUM_ROWS"       },
-        {GGML_OP_SUB,             "GGML_OP_SUB"            },
-        {GGML_OP_TRANSPOSE,       "GGML_OP_TRANSPOSE"      },
-        {GGML_OP_VIEW,            "GGML_OP_VIEW"           },
-        {GGML_OP_SET_ROWS,        "GGML_OP_SET_ROWS"       },
-        {GGML_OP_CPY,             "GGML_OP_CPY"            },
-        {GGML_OP_FLASH_ATTN_EXT,  "GGML_OP_FLASH_ATTN_EXT" },
-        {GGML_OP_L2_NORM,         "GGML_OP_L2_NORM"        },
-        {GGML_OP_CLAMP,           "GGML_OP_CLAMP"          },
-        {GGML_OP_PAD,             "GGML_OP_PAD"            },
-        {GGML_OP_SSM_CONV,        "GGML_OP_SSM_CONV"       },
-        {GGML_OP_GATED_DELTA_NET, "GGML_OP_GATED_DELTA_NET"},
-        {GGML_OP_ARGSORT,         "GGML_OP_ARGSORT"        },
-        {GGML_OP_REPEAT,          "GGML_OP_REPEAT"         },
-        {GGML_OP_IM2COL,          "GGML_OP_IM2COL"         }
-    };
-    static const std::map<ggml_unary_op, std::string> unary_ops = {
-        {GGML_UNARY_OP_ABS,         "GGML_UNARY_OP_ABS"        },
-        {GGML_UNARY_OP_SGN,         "GGML_UNARY_OP_SGN"        },
-        {GGML_UNARY_OP_NEG,         "GGML_UNARY_OP_NEG"        },
-        {GGML_UNARY_OP_STEP,        "GGML_UNARY_OP_STEP"       },
-        {GGML_UNARY_OP_TANH,        "GGML_UNARY_OP_TANH"       },
-        {GGML_UNARY_OP_ELU,         "GGML_UNARY_OP_ELU"        },
-        {GGML_UNARY_OP_RELU,        "GGML_UNARY_OP_RELU"       },
-        {GGML_UNARY_OP_SIGMOID,     "GGML_UNARY_OP_SIGMOID"    },
-        {GGML_UNARY_OP_GELU,        "GGML_UNARY_OP_GELU"       },
-        {GGML_UNARY_OP_GELU_QUICK,  "GGML_UNARY_OP_GELU_QUICK" },
-        {GGML_UNARY_OP_SILU,        "GGML_UNARY_OP_SILU"       },
-        {GGML_UNARY_OP_SOFTPLUS,    "GGML_UNARY_OP_SOFTPLUS"   },
-        {GGML_UNARY_OP_HARDSWISH,   "GGML_UNARY_OP_HARDSWISH"  },
-        {GGML_UNARY_OP_HARDSIGMOID, "GGML_UNARY_OP_HARDSIGMOID"},
-        {GGML_UNARY_OP_EXP,         "GGML_UNARY_OP_EXP"        },
-        {GGML_UNARY_OP_COUNT,       "GGML_UNARY_OP_COUNT"      }
-    };
-    static const std::map<ggml_glu_op, std::string> glu_ops = {
-        {GGML_GLU_OP_SWIGLU, "GGML_GLU_OP_SWIGLU"},
-        {GGML_GLU_OP_GEGLU,  "GGML_GLU_OP_GEGLU" },
-        {GGML_GLU_OP_REGLU,  "GGML_GLU_OP_REGLU" }
-    };
-
     switch (node->op) {
     case GGML_OP_UNARY:
-        return unary_ops.at(ggml_get_unary_op(node));
+        return std::string("GGML_UNARY_OP_") + ggml_unary_op_name(ggml_get_unary_op(node));
     case GGML_OP_GLU:
-        return glu_ops.at(ggml_get_glu_op(node));
+        return std::string("GGML_GLU_OP_") + ggml_glu_op_name(ggml_get_glu_op(node));
     default:
-        return ops.at(node->op);
+        return std::string("GGML_OP_") + ggml_op_name(node->op);
     }
-    static const std::string unknown_op = "UNKNOWN_GGML_OP";
-    return unknown_op;
 }
 
 const std::string & GgmlOvDecoder::get_op_type(int node_idx) const {

ggml/src/ggml-openvino/ggml-openvino.cpp

@@ -1053,6 +1053,10 @@ static bool is_op_unsupported_case(const ggml_tensor * op) {
             (op->ne[0] == 2 && op->ne[1] == 4 && op->ne[2] == 3 && op->ne[3] == 2)) {
             return true;
         }
+        // CPY into a strided view of a larger buffer (recurrent-state snapshots) not supported
+        if (op->view_src && ggml_nbytes(op) != ggml_nbytes(op->view_src)) {
+            return true;
+        }
         break;
     }
     case GGML_OP_MUL_MAT: {

ggml/src/ggml-openvino/openvino/op/add_id.cpp

@@ -17,6 +17,22 @@ namespace frontend {
 namespace ggml {
 namespace op {
 
+static ov::Output<ov::Node> reshape_add_id_input_to_2d(const ov::Output<ov::Node> & input,
+                                                       const ov::PartialShape & input_shape,
+                                                       const std::vector<int> & dims) {
+    const auto actual_shape = input.get_partial_shape();
+    if (actual_shape.rank().is_static() && actual_shape.rank().get_length() == 2) {
+        return input;
+    }
+
+    if (input_shape.rank().is_static() && input_shape.rank().get_length() == 2) {
+        return input;
+    }
+
+    auto shape = std::make_shared<ov::op::v3::ShapeOf>(input, ov::element::i64);
+    return std::make_shared<ov::op::v1::Reshape>(input, get_dimensions(shape, dims), false);
+}
+
 OutputVector translate_add_id(const NodeContext & context) {
     num_inputs_check(context, 3, 3);
 
@@ -28,11 +44,9 @@ OutputVector translate_add_id(const NodeContext & context) {
     //   input: [1, n_token, n_used, n_embd]
     //   bias:  [1, 1, n_expert, n_embd]
     //   ids:   [1, 1, n_token, n_used]
-    auto bias_shape_4d = std::make_shared<ov::op::v3::ShapeOf>(bias, ov::element::i64);
-    auto ids_shape_4d = std::make_shared<ov::op::v3::ShapeOf>(ids, ov::element::i64);
-
-    bias = std::make_shared<ov::op::v1::Reshape>(bias, get_dimensions(bias_shape_4d, {2, 3}), false);
-    ids = std::make_shared<ov::op::v1::Reshape>(ids, get_dimensions(ids_shape_4d, {2, 3}), false);
+    // Model bias constants may already be stored as [n_expert, n_embd].
+    bias = reshape_add_id_input_to_2d(bias, context.get_input_shape(1), {2, 3});
+    ids = reshape_add_id_input_to_2d(ids, context.get_input_shape(2), {2, 3});
 
     if (ids.get_element_type() != ov::element::i32 && ids.get_element_type() != ov::element::i64) {
         ids = std::make_shared<ov::op::v0::Convert>(ids, ov::element::i32);

ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp

@@ -3,8 +3,11 @@
 #include "../utils.h"
 
 #include <cstdint>
+#include <limits>
 #include <memory>
 #include <openvino/core/node_output.hpp>
+#include <openvino/op/add.hpp>
+#include <openvino/op/clamp.hpp>
 #include <openvino/op/constant.hpp>
 #include <openvino/op/multiply.hpp>
 #include <openvino/op/sigmoid.hpp>
@@ -15,7 +18,7 @@ namespace frontend {
 namespace ggml {
 namespace op {
 
-OutputVector translate_glu_swiglu(const NodeContext & context) {
+static std::pair<ov::Output<ov::Node>, ov::Output<ov::Node>> get_glu_inputs(const NodeContext & context) {
     num_inputs_check(context, 1, 2);
 
     ov::Output<ov::Node> src0;
@@ -52,6 +55,12 @@ OutputVector translate_glu_swiglu(const NodeContext & context) {
         std::swap(src0, src1);
     }
 
+    return {src0, src1};
+}
+
+OutputVector translate_glu_swiglu(const NodeContext & context) {
+    auto [src0, src1] = get_glu_inputs(context);
+
     auto sigmoid = std::make_shared<ov::op::v0::Sigmoid>(src0);
     auto silu = std::make_shared<ov::op::v1::Multiply>(src0, sigmoid);
     auto res = std::make_shared<ov::op::v1::Multiply>(silu, src1);
@@ -59,6 +68,27 @@ OutputVector translate_glu_swiglu(const NodeContext & context) {
     return rename_outputs_with_suffix({res}, context.get_name());
 }
 
+OutputVector translate_glu_swiglu_oai(const NodeContext & context) {
+    auto [src0, src1] = get_glu_inputs(context);
+
+    const int32_t * params = context.get_output_op_params();
+    const float alpha = reinterpret_cast<const float *>(params)[2];
+    const float limit = reinterpret_cast<const float *>(params)[3];
+
+    auto gate = std::make_shared<ov::op::v0::Clamp>(src0, -std::numeric_limits<float>::infinity(), limit);
+    auto alpha_const = ov::op::v0::Constant::create(ov::element::f32, {}, {alpha});
+    auto scaled_gate = std::make_shared<ov::op::v1::Multiply>(gate, alpha_const);
+    auto sigmoid = std::make_shared<ov::op::v0::Sigmoid>(scaled_gate);
+    auto out_glu = std::make_shared<ov::op::v1::Multiply>(gate, sigmoid);
+
+    auto up = std::make_shared<ov::op::v0::Clamp>(src1, -limit, limit);
+    auto one = ov::op::v0::Constant::create(ov::element::f32, {}, {1.0f});
+    auto up_plus_one = std::make_shared<ov::op::v1::Add>(up, one);
+    auto res = std::make_shared<ov::op::v1::Multiply>(out_glu, up_plus_one);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
 }  // namespace op
 }  // namespace ggml
 }  // namespace frontend

ggml/src/ggml-openvino/openvino/op/mul_mat_id.cpp

@@ -2,23 +2,135 @@
 #include "../op_table.h"
 #include "../utils.h"
 
+#include <cstdint>
+#include <cstring>
+#include <limits>
 #include <memory>
+#include <openvino/op/bitwise_and.hpp>
+#include <openvino/op/bitwise_right_shift.hpp>
 #include <openvino/op/broadcast.hpp>
 #include <openvino/op/concat.hpp>
 #include <openvino/op/constant.hpp>
 #include <openvino/op/convert.hpp>
 #include <openvino/op/gather.hpp>
 #include <openvino/op/matmul.hpp>
+#include <openvino/op/multiply.hpp>
 #include <openvino/op/reshape.hpp>
 #include <openvino/op/shape_of.hpp>
-#include <openvino/op/squeeze.hpp>
+#include <openvino/op/slice.hpp>
 #include <openvino/op/unsqueeze.hpp>
+#include <vector>
 
 namespace ov {
 namespace frontend {
 namespace ggml {
 namespace op {
 
+namespace {
+
+std::shared_ptr<ov::op::v0::Constant> const_i64(const std::vector<int64_t> & values) {
+    return ov::op::v0::Constant::create(ov::element::i64, ov::Shape{values.size()}, values);
+}
+
+ov::Output<ov::Node> slice_axis(const ov::Output<ov::Node> & input, int64_t axis, int64_t begin, int64_t end) {
+    return std::make_shared<ov::op::v8::Slice>(input, const_i64({begin}), const_i64({end}), const_i64({1}),
+                                              const_i64({axis}));
+}
+
+ov::Output<ov::Node> translate_mul_mat_id_mxfp4_packed(const NodeContext & context,
+                                                       ov::Output<ov::Node> expert_weights,
+                                                       ov::Output<ov::Node> activations,
+                                                       ov::Output<ov::Node> ids) {
+    auto packed_shape = expert_weights.get_partial_shape().to_shape();
+    FRONT_END_OP_CONVERSION_CHECK(packed_shape.size() == 5 && packed_shape[4] == 17,
+                                  "Expected packed MXFP4 expert weights with shape [1, n_expert, m, k_blocks, 17]");
+
+    const int64_t n_expert = static_cast<int64_t>(packed_shape[1]);
+    const int64_t rows = static_cast<int64_t>(packed_shape[2]);
+    const int64_t k_blocks = static_cast<int64_t>(packed_shape[3]);
+    const int64_t qk = 32;
+    const int64_t cols = k_blocks * qk;
+
+    auto packed_shape_4d = const_i64({n_expert, rows, k_blocks, 17});
+    expert_weights = std::make_shared<ov::op::v1::Reshape>(expert_weights, packed_shape_4d, false);
+
+    auto activations_shape_4d = std::make_shared<ov::op::v3::ShapeOf>(activations, ov::element::i64);
+    auto ids_shape_4d = std::make_shared<ov::op::v3::ShapeOf>(ids, ov::element::i64);
+    auto activations_shape_3d = get_dimensions(activations_shape_4d, {1, 2, 3});
+    auto ids_shape_2d = get_dimensions(ids_shape_4d, {2, 3});
+
+    activations = std::make_shared<ov::op::v1::Reshape>(activations, activations_shape_3d, false);
+    ids = std::make_shared<ov::op::v1::Reshape>(ids, ids_shape_2d, false);
+    if (ids.get_element_type() != ov::element::i32 && ids.get_element_type() != ov::element::i64) {
+        ids = std::make_shared<ov::op::v0::Convert>(ids, ov::element::i32);
+    }
+
+    auto gather_axis = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{}, {0});
+
+    static const std::vector<float> f4e2m1_lut = {0.0f,  0.5f,  1.0f,  1.5f,  2.0f,  3.0f,  4.0f,  6.0f,
+                                                  -0.0f, -0.5f, -1.0f, -1.5f, -2.0f, -3.0f, -4.0f, -6.0f};
+    std::vector<float> e8m0_lut(256);
+    for (size_t i = 0; i < e8m0_lut.size(); ++i) {
+        uint32_t bits = static_cast<uint32_t>(i) << 23;
+        memcpy(&e8m0_lut[i], &bits, sizeof(float));
+    }
+    e8m0_lut[0] = std::numeric_limits<float>::min() / 2.0f;
+    e8m0_lut[255] = std::numeric_limits<float>::quiet_NaN();
+
+    auto f4_lut = ov::op::v0::Constant::create(ov::element::f32, ov::Shape{f4e2m1_lut.size()}, f4e2m1_lut);
+    auto scale_lut = ov::op::v0::Constant::create(ov::element::f32, ov::Shape{e8m0_lut.size()}, e8m0_lut);
+
+    auto selected_packed_weights = std::make_shared<ov::op::v8::Gather>(expert_weights, ids, gather_axis);
+    auto scale_byte = slice_axis(selected_packed_weights, 4, 0, 1);
+    auto qs = slice_axis(selected_packed_weights, 4, 1, 17);
+    auto low = std::make_shared<ov::op::v13::BitwiseAnd>(
+        qs, ov::op::v0::Constant::create(ov::element::u8, ov::Shape{}, {0x0F}), ov::op::AutoBroadcastType::NUMPY);
+    auto high_shift = std::make_shared<ov::op::v15::BitwiseRightShift>(
+        qs, ov::op::v0::Constant::create(ov::element::u8, ov::Shape{}, {4}), ov::op::AutoBroadcastType::NUMPY);
+    auto nibbles = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{low, high_shift}, 4);
+    auto nibble_indices = std::make_shared<ov::op::v0::Convert>(nibbles, ov::element::i32);
+    auto weights_f32 = std::make_shared<ov::op::v8::Gather>(f4_lut, nibble_indices, gather_axis);
+
+    auto scale_indices = std::make_shared<ov::op::v0::Convert>(scale_byte, ov::element::i32);
+    auto scales_f32 = std::make_shared<ov::op::v8::Gather>(scale_lut, scale_indices, gather_axis);
+    ov::Output<ov::Node> selected_weights = std::make_shared<ov::op::v1::Multiply>(weights_f32, scales_f32,
+                                                                                  ov::op::AutoBroadcastType::NUMPY);
+
+    auto ids_shape = std::make_shared<ov::op::v3::ShapeOf>(ids, ov::element::i64);
+    auto selected_weights_target_dims = std::make_shared<ov::op::v0::Concat>(
+        ov::OutputVector{get_dimensions(ids_shape, {0, 1}), const_i64({rows, cols})}, 0);
+    selected_weights = std::make_shared<ov::op::v1::Reshape>(selected_weights, selected_weights_target_dims, false);
+
+    auto activations_shape = std::make_shared<ov::op::v3::ShapeOf>(activations, ov::element::i64);
+    ov::Output<ov::Node> acts_target_dims = std::make_shared<ov::op::v0::Concat>(
+        ov::OutputVector{
+            get_dimensions(activations_shape, {0}),
+            get_dimensions(ids_shape, {1}),
+            get_dimensions(activations_shape, {2}),
+        },
+        0);
+    ov::Output<ov::Node> acts_broadcasted =
+        std::make_shared<ov::op::v3::Broadcast>(activations, acts_target_dims, ov::op::BroadcastType::BIDIRECTIONAL);
+
+    auto activations_expanded = std::make_shared<ov::op::v0::Unsqueeze>(acts_broadcasted, const_i64({2}));
+    ov::Output<ov::Node> result =
+        std::make_shared<ov::op::v0::MatMul>(activations_expanded, selected_weights, false, true);
+
+    auto batch_dim = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+    auto row_dim = ov::op::v0::Constant::create(ov::element::i64, {1}, {rows});
+    auto result_target_dims = std::make_shared<ov::op::v0::Concat>(
+        ov::OutputVector{batch_dim, get_dimensions(ids_shape, {0, 1}), row_dim}, 0);
+    result = std::make_shared<ov::op::v1::Reshape>(result, result_target_dims, false);
+
+    const auto output_type = context.get_output_type();
+    if (result.get_element_type() != output_type) {
+        result = std::make_shared<ov::op::v0::Convert>(result, output_type);
+    }
+    return result;
+}
+
+}  // namespace
+
 OutputVector translate_mul_mat_id(const NodeContext & context) {
     num_inputs_check(context, 3, 3);
 
@@ -26,6 +138,12 @@ OutputVector translate_mul_mat_id(const NodeContext & context) {
     auto activations = process_view_input_new(context, 1);
     auto ids = process_view_input_new(context, 2);
 
+    if (expert_weights.get_element_type() == ov::element::u8 && expert_weights.get_partial_shape().rank().is_static() &&
+        expert_weights.get_partial_shape().rank().get_length() == 5) {
+        return rename_outputs_with_suffix({translate_mul_mat_id_mxfp4_packed(context, expert_weights, activations, ids)},
+                                          context.get_name());
+    }
+
     // OpenVINO sees GGML tensors in reversed dimension order:
     //   weights: [1, n_expert, m, k]
     //   activations: [1, n_tokens, n_used_or_1, k]

ggml/src/ggml-openvino/openvino/op/softmax.cpp

@@ -6,12 +6,16 @@
 #include <cstdint>
 #include <cstring>
 #include <memory>
+#include <openvino/op/broadcast.hpp>
 #include <openvino/frontend/exception.hpp>
 #include <openvino/op/add.hpp>
+#include <openvino/op/concat.hpp>
 #include <openvino/op/constant.hpp>
 #include <openvino/op/convert.hpp>
 #include <openvino/op/multiply.hpp>
 #include <openvino/op/reshape.hpp>
+#include <openvino/op/shape_of.hpp>
+#include <openvino/op/slice.hpp>
 #include <openvino/op/softmax.hpp>
 #include <vector>
 
@@ -20,12 +24,31 @@ namespace frontend {
 namespace ggml {
 namespace op {
 
+static bool is_static_one(const ov::Dimension & dim) {
+    return dim.is_static() && dim.get_length() == 1;
+}
+
+static bool same_static_dim(const ov::Dimension & lhs, const ov::Dimension & rhs) {
+    return lhs.is_static() && rhs.is_static() && lhs.get_length() == rhs.get_length();
+}
+
+static bool is_attention_sinks_input_shape(const ov::PartialShape & candidate, const ov::PartialShape & logits_shape) {
+    if (candidate.rank().is_dynamic() || logits_shape.rank().is_dynamic() || candidate.rank().get_length() != 4 ||
+        logits_shape.rank().get_length() != 4) {
+        return false;
+    }
+
+    return is_static_one(candidate[0]) && is_static_one(candidate[1]) && is_static_one(candidate[2]) &&
+           same_static_dim(candidate[3], logits_shape[1]);
+}
+
 // Reimplementation of GGML_OP_SOFT_MAX semantics for OpenVINO backend:
 // 1) logits = src0 * scale
 // 2) logits += mask (if provided)
-// 3) softmax over the last dimension
+// 3) append attention sinks as hidden logits (if provided)
+// 4) softmax over the last dimension and remove the hidden sink column
 OutputVector translate_soft_max(const NodeContext & context) {
-    num_inputs_check(context, 1, 2);
+    num_inputs_check(context, 1, 3);
 
     float scale = 1.0f;
     float max_bias = 0.0f;
@@ -33,6 +56,11 @@ OutputVector translate_soft_max(const NodeContext & context) {
     memcpy(&max_bias, (float *) context.get_output_op_params() + 1, sizeof(float));
 
     ov::Output<ov::Node> logits = context.get_input(0);
+    const bool second_input_is_sinks =
+        context.get_input_size() == 2 && is_attention_sinks_input_shape(context.get_input_shape(1), context.get_output_shape());
+    const bool has_mask = context.get_input_size() > 1 && !second_input_is_sinks;
+    const bool has_sinks = second_input_is_sinks || context.get_input_size() > 2;
+    const size_t sinks_input_idx = second_input_is_sinks ? 1 : 2;
 
     // Apply scale first: logits = src0 * scale
     if (scale != 1.0f) {
@@ -41,12 +69,12 @@ OutputVector translate_soft_max(const NodeContext & context) {
         logits = std::make_shared<ov::op::v1::Multiply>(logits, scale_const);
     }
 
-    FRONT_END_CHECK_IMPLEMENTED(!(max_bias > 0.0f && context.get_input_size() < 2),
+    FRONT_END_CHECK_IMPLEMENTED(!(max_bias > 0.0f && !has_mask),
                                 "OpenVINO softmax ALiBi path requires mask input");
 
     // Optional mask add: logits += mask
     // For max_bias > 0 (ALiBi), apply per-head slope to mask before adding.
-    if (context.get_input_size() > 1) {
+    if (has_mask) {
         ov::Output<ov::Node> mask = context.get_input(1);
 
         // For stateful
@@ -94,8 +122,40 @@ OutputVector translate_soft_max(const NodeContext & context) {
         logits = std::make_shared<ov::op::v1::Add>(logits, mask);
     }
 
+    ov::Output<ov::Node> softmax_input = logits;
+    if (has_sinks) {
+        ov::Output<ov::Node> sinks = context.get_input(sinks_input_idx);
+        if (sinks.get_element_type() != logits.get_element_type()) {
+            sinks = std::make_shared<ov::op::v0::Convert>(sinks, logits.get_element_type());
+        }
+
+        auto sink_shape = ov::op::v0::Constant::create(ov::element::i64, {4}, {1, -1, 1, 1});
+        auto sinks_4d = std::make_shared<ov::op::v1::Reshape>(sinks, sink_shape, false);
+
+        auto logits_shape = std::make_shared<ov::op::v3::ShapeOf>(logits, ov::element::i64);
+        auto zero = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+        auto one = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+        auto three = ov::op::v0::Constant::create(ov::element::i64, {1}, {3});
+        auto four = ov::op::v0::Constant::create(ov::element::i64, {1}, {4});
+        auto shape_axis = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+
+        auto sink_prefix_shape = std::make_shared<ov::op::v8::Slice>(logits_shape, zero, three, one, shape_axis);
+        auto sink_last_dim = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+        auto sink_broadcast_shape = std::make_shared<ov::op::v0::Concat>(
+            ov::OutputVector{sink_prefix_shape, sink_last_dim}, 0);
+        auto sink_column = std::make_shared<ov::op::v3::Broadcast>(sinks_4d, sink_broadcast_shape,
+                                                                   ov::op::BroadcastType::BIDIRECTIONAL);
+        softmax_input = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{logits, sink_column}, 3);
+
+        auto softmax_with_sink = std::make_shared<ov::op::v8::Softmax>(softmax_input, -1);
+        auto original_last_dim = std::make_shared<ov::op::v8::Slice>(logits_shape, three, four, one, shape_axis);
+        auto res = std::make_shared<ov::op::v8::Slice>(softmax_with_sink, zero, original_last_dim, one, three);
+
+        return rename_outputs_with_suffix({res}, context.get_name());
+    }
+
     // Softmax along last dimension (equivalent to ggml softmax over ne[0]).
-    auto res = std::make_shared<ov::op::v8::Softmax>(logits, -1);
+    auto res = std::make_shared<ov::op::v8::Softmax>(softmax_input, -1);
 
     return rename_outputs_with_suffix({res}, context.get_name());
 }

ggml/src/ggml-openvino/openvino/op_table.cpp

@@ -47,6 +47,7 @@ std::unordered_map<std::string, CreatorFunction> get_supported_ops() {
         {"GGML_UNARY_OP_TANH",      op::translate_1to1_match_1_input<v0::Tanh>     },
         {"GGML_OP_VIEW",            op::translate_view                             },
         {"GGML_GLU_OP_SWIGLU",      op::translate_glu_swiglu                       },
+        {"GGML_GLU_OP_SWIGLU_OAI",  op::translate_glu_swiglu_oai                   },
         {"GGML_GLU_OP_GEGLU",       op::translate_glu_geglu                        },
         {"GGML_OP_SET_ROWS",        op::translate_set_rows                         },
         {"GGML_OP_CPY",             op::translate_cpy                              },

ggml/src/ggml-openvino/openvino/op_table.h

@@ -32,6 +32,7 @@ GGML_OP_CONVERTER(translate_soft_max);
 GGML_OP_CONVERTER(translate_transpose);
 GGML_OP_CONVERTER(translate_view);
 GGML_OP_CONVERTER(translate_glu_swiglu);
+GGML_OP_CONVERTER(translate_glu_swiglu_oai);
 GGML_OP_CONVERTER(translate_glu_geglu);
 GGML_OP_CONVERTER(translate_set_rows);
 GGML_OP_CONVERTER(translate_cpy);

ggml/src/ggml-sycl/conv3d.cpp

@@ -103,8 +103,8 @@ void ggml_sycl_op_conv_3d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     // allocate packed arrays: A_packed (k x m), B_packed (k x n)
     ggml_sycl_pool_alloc<float> A_packed_alloc(ctx.pool());
     ggml_sycl_pool_alloc<float> B_packed_alloc(ctx.pool());
-    A_packed_alloc.alloc((size_t) knl_n_total * patch_total * sizeof(float));
-    B_packed_alloc.alloc((size_t) knl_n_total * oc * sizeof(float));
+    A_packed_alloc.alloc((size_t) knl_n_total * patch_total);
+    B_packed_alloc.alloc((size_t) knl_n_total * oc);
 
     float * A_packed = A_packed_alloc.get();
     float * B_packed = B_packed_alloc.get();
@@ -115,10 +115,16 @@ void ggml_sycl_op_conv_3d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
 
     // Combined kernel: im2col -> pack A, and pack B simultaneously
     const char * src1_base = (const char *) src1->data;
+    const char * src0_base = (const char *) src0->data;
     const int64_t src1_nb0 = src1->nb[0];
     const int64_t src1_nb1 = src1->nb[1];
     const int64_t src1_nb2 = src1->nb[2];
     const int64_t src1_nb3 = src1->nb[3];
+    const int64_t src1_w = src1->ne[0];
+    const int64_t src1_h = src1->ne[1];
+    const int64_t src1_d = src1->ne[2];
+
+    const bool src0_is_f32 = (src0->type == GGML_TYPE_F32);
 
     // Compute correct strides for src0 as (knl_n_total, oc) matrix
     const int64_t src0_packed_nb0 = kernel_type_size;
@@ -165,7 +171,7 @@ void ggml_sycl_op_conv_3d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
         const int64_t sz = dst_z * s2 + kz * d2 - p2;
 
         float val = 0.0f;
-        if (sx >= 0 && sx < src1->ne[0] && sy >= 0 && sy < src1->ne[1] && sz >= 0 && sz < src1->ne[2]) {
+        if (sx >= 0 && sx < src1_w && sy >= 0 && sy < src1_h && sz >= 0 && sz < src1_d) {
             const int64_t channel_idx = batch_idx * c + ic;
             const char * ptr = src1_base + sx * src1_nb0 + sy * src1_nb1 + sz * src1_nb2 + channel_idx * src1_nb3;
             val = *(const float *) ptr;
@@ -184,9 +190,9 @@ void ggml_sycl_op_conv_3d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
 
         const int64_t row = t % k;
         const int64_t col = t / k;
-        const char * src_ptr = (const char *) src0->data + row * src0_packed_nb0 + col * src0_packed_nb1;
+        const char * src_ptr = src0_base + row * src0_packed_nb0 + col * src0_packed_nb1;
         float v;
-        if (src0->type == GGML_TYPE_F32) {
+        if (src0_is_f32) {
             v = *(const float *) src_ptr;
         } else {
             v = sycl::vec<sycl::half, 1>(*(const sycl::half *) src_ptr).convert<float, sycl::rounding_mode::automatic>()[0];

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

@@ -5859,6 +5859,250 @@ static ggml_backend_dev_t ggml_backend_sycl_reg_get_device(ggml_backend_reg_t re
     return ctx->devices[index];
 }
 
+// ==========================================================================
+// Tensor parallelism (--split-mode tensor) for the SYCL backend.
+//
+// The meta-backend invokes these three entry points via get_proc_address:
+//   * ggml_backend_sycl_comm_init             - one-time per-graph setup
+//   * ggml_backend_sycl_comm_allreduce_tensor - per-allreduce step
+//   * ggml_backend_sycl_comm_free             - tear-down
+//
+// For N=2 (dual-GPU), this is a degenerate ring allreduce with dual paths
+// chosen by tensor size:
+//
+//   * Small (nelem < 32K): FP32 direct memcpy + per-device ADD
+//     kernel. The kernel depends_on() its corresponding memcpy event
+//     so it doesn't read partial data. Both devices run in parallel.
+//
+//   * Large (nelem >= 32K): BF16-compressed. Each device compresses
+//     its FP32 partial to BF16 locally, cross-device memcpys
+//     to the peer (half the PCI bandwidth), where it is decompressed
+//     and added into the local FP32 partial. 6 SYCL submissions per
+//     allreduce (2 compress + 2 memcpy + 2 decompress-add) vs the
+//     4 for the small path, but the bandwidth saving > 6 GB/s PCIe x 2
+//     dominates for larger tensors.
+//
+// Storage: A persistent uint8_t buffer per device, sized to
+// 4 * nelem bytes. Both paths reinterpret the same bytes (small path
+// as nelem floats; large path as outbox + inbox = 2*nelem uint16_t
+// each, using the full 4*nelem byte budget either way). Single
+// alloc+free per device keeps the SYCL pool's strict-LIFO invariant
+// trivial.
+//
+// For non-(N=2 FP32 contiguous) cases, comm_init or comm_allreduce_tensor
+// returns null/false, causing the meta-backend to use its generic
+// butterfly all-reduce fallback.
+// ==========================================================================
+
+struct ggml_backend_sycl_comm_context {
+    std::vector<ggml_backend_t> backends;
+    // ONE persistent per-device byte buffer, 4*nelem bytes.  Both the
+    // FP32 small-tensor path and the BF16 large-tensor path share it
+    // by reinterpreting.
+    std::unique_ptr<ggml_sycl_pool_alloc<uint8_t>> buf0;
+    std::unique_ptr<ggml_sycl_pool_alloc<uint8_t>> buf1;
+    int64_t buf_nelem = 0;
+};
+
+void * ggml_backend_sycl_comm_init(ggml_backend_t * backends, size_t n_backends) try {
+    for (size_t i = 0; i < n_backends; ++i) {
+        if (!ggml_backend_is_sycl(backends[i])) {
+            return nullptr;
+        }
+    }
+
+    // Initial version: N=2 only. For N!=2, returning null makes the
+    // meta-backend skip this backend-specific allreduce entirely.
+    if (n_backends != 2) {
+        return nullptr;
+    }
+
+    auto * ctx = new ggml_backend_sycl_comm_context;
+    ctx->backends.assign(backends, backends + n_backends);
+    auto * sctx0 = (ggml_backend_sycl_context *) backends[0]->context;
+    auto * sctx1 = (ggml_backend_sycl_context *) backends[1]->context;
+    ctx->buf0 = std::make_unique<ggml_sycl_pool_alloc<uint8_t>>(sctx0->pool());
+    ctx->buf1 = std::make_unique<ggml_sycl_pool_alloc<uint8_t>>(sctx1->pool());
+    return ctx;
+}
+catch (const sycl::exception &) { return nullptr; }
+catch (...)                     { return nullptr; }
+
+void ggml_backend_sycl_comm_free(void * comm_ctx_v) {
+    auto * comm_ctx = static_cast<ggml_backend_sycl_comm_context *>(comm_ctx_v);
+    if (comm_ctx == nullptr) {
+        return;
+    }
+
+    // Sync both per-device queues so the pool_alloc destructors don't
+    // return memory still in use by the last kernel.
+    if (comm_ctx->backends.size() == 2) {
+        auto * sctx0 = (ggml_backend_sycl_context *) comm_ctx->backends[0]->context;
+        auto * sctx1 = (ggml_backend_sycl_context *) comm_ctx->backends[1]->context;
+        try {
+            sctx0->stream()->wait();
+            sctx1->stream()->wait();
+        } catch (...) { /* best effort during shutdown */ }
+    }
+
+    delete comm_ctx;
+}
+
+bool ggml_backend_sycl_comm_allreduce_tensor(void * comm_ctx_v, struct ggml_tensor ** tensors) try {
+    if (comm_ctx_v == nullptr) {
+        return false;
+    }
+
+    auto * comm_ctx = static_cast<ggml_backend_sycl_comm_context *>(comm_ctx_v);
+    const size_t n_backends = comm_ctx->backends.size();
+
+    // Fast path: N=2, F32/F16, contiguous, matching shapes.
+    if (n_backends != 2) {
+        return false;
+    }
+    // Accept F32 or F16 inputs natively (types must match). F16 takes the
+    // direct 2-byte memcpy + add path below; other types return false so the
+    // meta-backend uses its generic all-reduce.
+    if (tensors[0]->type != tensors[1]->type) {
+        return false;
+    }
+    if (tensors[0]->type != GGML_TYPE_F32 && tensors[0]->type != GGML_TYPE_F16) {
+        return false;
+    }
+    if (!ggml_is_contiguous(tensors[0]) || !ggml_is_contiguous(tensors[1])) {
+        return false;
+    }
+    if (ggml_nelements(tensors[0]) != ggml_nelements(tensors[1])) {
+        return false;
+    }
+
+    const int64_t nelem  = ggml_nelements(tensors[0]);
+    const size_t  nbytes = ggml_nbytes(tensors[0]);
+    if (nelem == 0) {
+        return true;
+    }
+
+    auto * ctx0 = (ggml_backend_sycl_context *) comm_ctx->backends[0]->context;
+    auto * ctx1 = (ggml_backend_sycl_context *) comm_ctx->backends[1]->context;
+    queue_ptr q0 = ctx0->stream();
+    queue_ptr q1 = ctx1->stream();
+
+    // Grow per-device byte buffers if needed (4 * nelem bytes each).
+    if (comm_ctx->buf_nelem < nelem) {
+        comm_ctx->buf0->realloc(nelem * 4);
+        comm_ctx->buf1->realloc(nelem * 4);
+        comm_ctx->buf_nelem = nelem;
+    }
+    uint8_t * buf0 = comm_ctx->buf0->get();
+    uint8_t * buf1 = comm_ctx->buf1->get();
+
+    // F16 native path: direct 2-byte cross-device copy + add, skipping the
+    // F32 round-trip the meta-backend fallback would force. Cross-device copies
+    // go through dev2dev_memcpy because the two devices are in separate SYCL
+    // contexts (a raw peer-USM q->memcpy would be a silent no-op).
+    if (tensors[0]->type == GGML_TYPE_F16) {
+        sycl::half * f16_out0 = (sycl::half *) tensors[0]->data;
+        sycl::half * f16_out1 = (sycl::half *) tensors[1]->data;
+        sycl::half * f16_tmp0 = (sycl::half *) buf0;
+        sycl::half * f16_tmp1 = (sycl::half *) buf1;
+
+        q0->wait();
+        q1->wait();
+        dev2dev_memcpy(ctx0->device, *q0, ctx1->device, *q1, f16_tmp0, tensors[1]->data, nbytes);
+        dev2dev_memcpy(ctx1->device, *q1, ctx0->device, *q0, f16_tmp1, tensors[0]->data, nbytes);
+
+        q0->submit([&](sycl::handler & h) {
+            h.parallel_for(sycl::range<1>(nelem), [=](sycl::id<1> i) {
+                f16_out0[i] = (sycl::half) ((float) f16_out0[i] + (float) f16_tmp0[i]);
+            });
+        });
+        q1->submit([&](sycl::handler & h) {
+            h.parallel_for(sycl::range<1>(nelem), [=](sycl::id<1> i) {
+                f16_out1[i] = (sycl::half) ((float) f16_out1[i] + (float) f16_tmp1[i]);
+            });
+        });
+        return true;
+    }
+
+    float * out0 = (float *) tensors[0]->data;
+    float * out1 = (float *) tensors[1]->data;
+
+    // BF16 threshold: above this, the PCIe savings from halving the
+    // cross-device bytes outweigh the 2 extra compress kernels.
+    // Below: stay on the FP32 fast path.  Threshold mirrors the CUDA
+    // NCCL allreduce pattern for n_backends=2.
+    static constexpr int64_t BF16_THRESHOLD = 32768;
+
+    if (nelem < BF16_THRESHOLD) {
+        // FP32 small path: 4 SYCL submissions per allreduce.
+        float * tmp0 = (float *) buf0;
+        float * tmp1 = (float *) buf1;
+
+        // COMM-D2D-FIX: the two devices are in SEPARATE SYCL contexts, so a raw
+        // q->memcpy of a peer USM pointer is a silent no-op. Route cross-device
+        // copies through dev2dev_memcpy (L0 direct copy / host staging). It is
+        // synchronous, so wait for the local partials to be produced first.
+        q0->wait();
+        q1->wait();
+        dev2dev_memcpy(ctx0->device, *q0, ctx1->device, *q1, tmp0, tensors[1]->data, nbytes);
+        dev2dev_memcpy(ctx1->device, *q1, ctx0->device, *q0, tmp1, tensors[0]->data, nbytes);
+
+        q0->submit([&](sycl::handler & h) {
+            h.parallel_for(sycl::range<1>(nelem), [=](sycl::id<1> i) {
+                out0[i] += tmp0[i];
+            });
+        });
+        q1->submit([&](sycl::handler & h) {
+            h.parallel_for(sycl::range<1>(nelem), [=](sycl::id<1> i) {
+                out1[i] += tmp1[i];
+            });
+        });
+        return true;
+    }
+
+    // BF16 large path: 6 SYCL submissions per allreduce, but the
+    // cross-device memcpy is HALF the bytes. Pure bit-shift
+    // conversion (no rounding) — matches ggml's truncating fp32->bf16.
+    uint16_t * outbox0 = (uint16_t *) buf0;
+    uint16_t * inbox0  = outbox0 + nelem;
+    uint16_t * outbox1 = (uint16_t *) buf1;
+    uint16_t * inbox1  = outbox1 + nelem;
+
+    // Phase A: compress each device's local partial in parallel.
+    sycl::event c0 = q0->parallel_for(sycl::range<1>(nelem), [=](sycl::id<1> i) {
+        outbox0[i] = (uint16_t) (sycl::bit_cast<uint32_t>(out0[i]) >> 16);
+    });
+
+    sycl::event c1 = q1->parallel_for(sycl::range<1>(nelem), [=](sycl::id<1> i) {
+        outbox1[i] = (uint16_t) (sycl::bit_cast<uint32_t>(out1[i]) >> 16);
+    });
+
+    // Phase B: COMM-D2D-FIX-BF16 cross-device copy of compressed bytes via
+    // dev2dev_memcpy (separate SYCL contexts; sync copy after compress).
+    const size_t bf16_bytes = nelem * sizeof(uint16_t);
+    c0.wait();
+    c1.wait();
+    dev2dev_memcpy(ctx0->device, *q0, ctx1->device, *q1, inbox0, outbox1, bf16_bytes);
+    dev2dev_memcpy(ctx1->device, *q1, ctx0->device, *q0, inbox1, outbox0, bf16_bytes);
+
+    // Phase C: decompress + add into local FP32 partial.
+    q0->submit([&](sycl::handler & h) {
+        h.parallel_for(sycl::range<1>(nelem), [=](sycl::id<1> i) {
+            out0[i] += sycl::bit_cast<float>(((uint32_t) inbox0[i]) << 16);
+        });
+    });
+
+    q1->submit([&](sycl::handler & h) {
+        h.parallel_for(sycl::range<1>(nelem), [=](sycl::id<1> i) {
+            out1[i] += sycl::bit_cast<float>(((uint32_t) inbox1[i]) << 16);
+        });
+    });
+
+    return true;
+}
+catch (const sycl::exception &) { return false; }
+catch (...)                     { return false; }
+
 static void *ggml_backend_sycl_reg_get_proc_address(ggml_backend_reg_t reg, const char *name) {
     GGML_UNUSED(reg);
 
@@ -5866,6 +6110,17 @@ static void *ggml_backend_sycl_reg_get_proc_address(ggml_backend_reg_t reg, cons
         return (void *)ggml_backend_sycl_split_buffer_type;
     }
 
+    // Tensor parallelism (--split-mode tensor) entry points.
+    if (strcmp(name, "ggml_backend_comm_init") == 0) {
+        return (void *)ggml_backend_sycl_comm_init;
+    }
+    if (strcmp(name, "ggml_backend_comm_free") == 0) {
+        return (void *)ggml_backend_sycl_comm_free;
+    }
+    if (strcmp(name, "ggml_backend_comm_allreduce_tensor") == 0) {
+        return (void *)ggml_backend_sycl_comm_allreduce_tensor;
+    }
+
     // SYCL doesn't support registering host memory, left here for reference
     // "ggml_backend_register_host_buffer"
     // "ggml_backend_unregister_host_buffer"

ggml/src/ggml-sycl/norm.cpp

@@ -2,8 +2,10 @@
 #include "ggml-sycl/common.hpp"
 #include "ggml-sycl/presets.hpp"
 
-static void norm_f32(const float* x, float* dst, const int ncols, const int64_t stride_row, const int64_t stride_channel,
-        const int64_t stride_sample, const float eps, const sycl::nd_item<3>& item_ct1, sycl::float2* s_sum, int block_size) {
+static void norm_f32(const float* x, float* dst, const int ncols,
+    const int64_t src_stride_col, const int64_t src_stride_row, const int64_t src_stride_channel, const int64_t src_stride_sample,
+    const int64_t dst_stride_col, const int64_t dst_stride_row, const int64_t dst_stride_channel, const int64_t dst_stride_sample,
+    const float eps, const sycl::nd_item<3>& item_ct1, sycl::float2* s_sum, int block_size) {
 
     const int nrows = item_ct1.get_group_range(2);
     const int nchannels = item_ct1.get_group_range(1);
@@ -16,16 +18,16 @@ static void norm_f32(const float* x, float* dst, const int ncols, const int64_t
     const int tid = item_ct1.get_local_id(2);
     const int nwarps = nthreads / WARP_SIZE;
 
-    const auto strided_offset = calculate_offset<3>({stride_sample, stride_channel, stride_row}, {sample, channel, row});
-    const auto packed_offset = calculate_offset<3>({nchannels * nrows * ncols, nrows * ncols, ncols}, {sample, channel, row});
+    const auto src_offset = calculate_offset<3>({src_stride_sample, src_stride_channel, src_stride_row}, {sample, channel, row});
+    const auto dst_offset = calculate_offset<3>({dst_stride_sample, dst_stride_channel, dst_stride_row}, {sample, channel, row});
 
-    x += strided_offset;
-    dst += packed_offset;
+    x += src_offset;
+    dst += dst_offset;
 
     sycl::float2 mean_var = sycl::float2(0.f, 0.f);
 
     for (int col = tid; col < ncols; col += block_size) {
-        const float xi = x[col];
+        const float xi = x[col * src_stride_col];
         mean_var.x() += xi;
         mean_var.y() += xi * xi;
     }
@@ -54,7 +56,7 @@ static void norm_f32(const float* x, float* dst, const int ncols, const int64_t
     const float inv_std = sycl::rsqrt(var + eps);
 
     for (int col = tid; col < ncols; col += block_size) {
-        dst[col] = (x[col] - mean) * inv_std;
+        dst[col * dst_stride_col] = (x[col * src_stride_col] - mean) * inv_std;
     }
 }
 
@@ -145,8 +147,10 @@ static void group_norm_f32(const float* x, float* dst, const int group_size, con
     }
 }
 
-static void rms_norm_f32(const float* x, float* dst, const int ncols, const int64_t stride_row, const int64_t stride_channel,
-        const int64_t stride_sample, const float eps, const sycl::nd_item<3>& item_ct1, float* s_sum, int block_size) {
+static void rms_norm_f32(const float* x, float* dst, const int ncols,
+    const int64_t src_stride_col, const int64_t src_stride_row, const int64_t src_stride_channel, const int64_t src_stride_sample,
+    const int64_t dst_stride_col, const int64_t dst_stride_row, const int64_t dst_stride_channel, const int64_t dst_stride_sample,
+    const float eps, const sycl::nd_item<3>& item_ct1, float* s_sum, int block_size) {
 
     const int nrows = item_ct1.get_group_range(2);
     const int nchannels = item_ct1.get_group_range(1);
@@ -160,17 +164,17 @@ static void rms_norm_f32(const float* x, float* dst, const int ncols, const int6
     const int tid = item_ct1.get_local_id(2);
     const int nwarps = nthreads / WARP_SIZE;
 
-    const auto strided_offset = calculate_offset<3>({stride_sample, stride_channel, stride_row}, {sample, channel, row});
-    const auto packed_offset = calculate_offset<3>({nchannels * nrows * ncols, nrows * ncols, ncols}, {sample, channel, row});
+    const auto src_offset = calculate_offset<3>({src_stride_sample, src_stride_channel, src_stride_row}, {sample, channel, row});
+    const auto dst_offset = calculate_offset<3>({dst_stride_sample, dst_stride_channel, dst_stride_row}, {sample, channel, row});
 
-    x   += strided_offset;
-    dst += packed_offset;
+    x   += src_offset;
+    dst += dst_offset;
 
 
     float tmp = 0.0f; // partial sum for thread in warp
 
     for (int col = tid; col < ncols; col += block_size) {
-        const float xi = x[col];
+        const float xi = x[col * src_stride_col];
         tmp += xi * xi;
     }
 
@@ -198,14 +202,15 @@ static void rms_norm_f32(const float* x, float* dst, const int ncols, const int6
     const float scale = sycl::rsqrt(mean + eps);
 
     for (int col = tid; col < ncols; col += block_size) {
-        dst[col] = scale * x[col];
+        dst[col * dst_stride_col] = scale * x[col * src_stride_col];
     }
 }
 
 template<int warp_size>
 static void l2_norm_f32(const float * x, float * dst, const int ncols,
-    const int64_t stride_row, const int64_t stride_channel,
-    const int64_t stride_sample, const float eps,
+    const int64_t src_stride_col, const int64_t src_stride_row, const int64_t src_stride_channel,
+    const int64_t src_stride_sample, const int64_t dst_stride_col, const int64_t dst_stride_row,
+    const int64_t dst_stride_channel, const int64_t dst_stride_sample, const float eps,
     const sycl::nd_item<3>& item_ct1, float* s_sum, const int block_size) {
     const int nrows     = item_ct1.get_group_range(2);
     const int nchannels = item_ct1.get_group_range(1);
@@ -215,13 +220,13 @@ static void l2_norm_f32(const float * x, float * dst, const int ncols,
     const int sample  = item_ct1.get_group(0);
     const int tid     = item_ct1.get_local_id(2);
 
-    x   += sample*stride_sample + channel*stride_channel + row*stride_row;
-    dst += ((sample*nchannels + channel)*nrows + row)*ncols;
+    x   += sample*src_stride_sample + channel*src_stride_channel + row*src_stride_row;
+    dst += sample*dst_stride_sample + channel*dst_stride_channel + row*dst_stride_row;
 
     float tmp = 0.0f; // partial sum for thread in warp
 
     for (int col = tid; col < ncols; col += block_size) {
-        const float xi = x[col];
+        const float xi = x[col * src_stride_col];
         tmp += xi * xi;
     }
 
@@ -229,12 +234,13 @@ static void l2_norm_f32(const float * x, float * dst, const int ncols,
     const float scale = sycl::rsqrt(sycl::fmax(tmp, eps * eps));
 
     for (int col = tid; col < ncols; col += block_size) {
-        dst[col] = scale * x[col];
+        dst[col * dst_stride_col] = scale * x[col * src_stride_col];
     }
 }
 
 static void norm_f32_sycl(const float * x, float * dst, const int ncols, const int nrows, const int nchannels, const int nsamples,
-        const int64_t stride_row, const int64_t stride_channel, const int64_t stride_sample,
+    const int64_t src_stride_col, const int64_t src_stride_row, const int64_t src_stride_channel, const int64_t src_stride_sample,
+    const int64_t dst_stride_col, const int64_t dst_stride_row, const int64_t dst_stride_channel, const int64_t dst_stride_sample,
         const float eps, queue_ptr stream, int device) {
 
     const sycl::range<3> global_dims(nsamples, nchannels, nrows);
@@ -245,7 +251,10 @@ static void norm_f32_sycl(const float * x, float * dst, const int ncols, const i
                 sycl::nd_range<3>(global_dims * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
                 [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, nullptr, WARP_SIZE);
+                    norm_f32(x, dst, ncols,
+                        src_stride_col, src_stride_row, src_stride_channel, src_stride_sample,
+                        dst_stride_col, dst_stride_row, dst_stride_channel, dst_stride_sample,
+                        eps, item_ct1, nullptr, WARP_SIZE);
                 });
             });
     }
@@ -265,7 +274,10 @@ static void norm_f32_sycl(const float * x, float * dst, const int ncols, const i
                 sycl::nd_range<3>(global_dims * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
                 [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size);
+                    norm_f32(x, dst, ncols,
+                        src_stride_col, src_stride_row, src_stride_channel, src_stride_sample,
+                        dst_stride_col, dst_stride_row, dst_stride_channel, dst_stride_sample,
+                        eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size);
                 });
             });
     }
@@ -319,7 +331,9 @@ static void group_norm_f32_sycl(const float* x, float* dst,
 }
 
 static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const int nrows, const int nchannels, const int nsamples,
-        const int64_t stride_row, const int64_t stride_channel, const int64_t stride_sample, const float eps, queue_ptr stream, int device) {
+    const int64_t src_stride_col, const int64_t src_stride_row, const int64_t src_stride_channel, const int64_t src_stride_sample,
+    const int64_t dst_stride_col, const int64_t dst_stride_row, const int64_t dst_stride_channel, const int64_t dst_stride_sample,
+    const float eps, queue_ptr stream, int device) {
     // printf("%s ncols=%d, nrows=%d, WARP_SIZE=%d\n", __func__, ncols, nrows, WARP_SIZE);
 
     const sycl::range<3> global_dims(nsamples, nchannels, nrows);
@@ -330,7 +344,10 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const
                 sycl::nd_range<3>(global_dims * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
                 [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, nullptr, WARP_SIZE);
+                    rms_norm_f32(x, dst, ncols,
+                        src_stride_col, src_stride_row, src_stride_channel, src_stride_sample,
+                        dst_stride_col, dst_stride_row, dst_stride_channel, dst_stride_sample,
+                        eps, item_ct1, nullptr, WARP_SIZE);
                 });
             });
     }
@@ -350,7 +367,10 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const
                 sycl::nd_range<3>(global_dims * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
                 [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size);
+                    rms_norm_f32(x, dst, ncols,
+                        src_stride_col, src_stride_row, src_stride_channel, src_stride_sample,
+                        dst_stride_col, dst_stride_row, dst_stride_channel, dst_stride_sample,
+                        eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size);
                 });
             });
     }
@@ -363,9 +383,14 @@ static void l2_norm_f32_sycl(const float *   x,
                              const int       nrows,
                              const int       nchannels,
                              const int       nsamples,
-                             const int64_t   stride_row,
-                             const int64_t   stride_channel,
-                             const int64_t   stride_sample,
+                             const int64_t   src_stride_col,
+                             const int64_t   src_stride_row,
+                             const int64_t   src_stride_channel,
+                             const int64_t   src_stride_sample,
+                             const int64_t   dst_stride_col,
+                             const int64_t   dst_stride_row,
+                             const int64_t   dst_stride_channel,
+                             const int64_t   dst_stride_sample,
                              const float     eps,
                              queue_ptr       stream,
                              int             device) {
@@ -379,7 +404,10 @@ static void l2_norm_f32_sycl(const float *   x,
                     block_dims),
                 [=](sycl::nd_item<3> item_ct1)
                 [[sycl::reqd_sub_group_size(warp_size)]] {
-                    l2_norm_f32<warp_size>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1,
+                    l2_norm_f32<warp_size>(x, dst, ncols,
+                        src_stride_col, src_stride_row, src_stride_channel, src_stride_sample,
+                        dst_stride_col, dst_stride_row, dst_stride_channel, dst_stride_sample,
+                        eps, item_ct1,
                         nullptr, warp_size);
                 });
             });
@@ -398,7 +426,9 @@ static void l2_norm_f32_sycl(const float *   x,
                     block_dims),
                 [=](sycl::nd_item<3> item_ct1)
                 [[sycl::reqd_sub_group_size(warp_size)]] {
-                    l2_norm_f32<warp_size>(x, dst, ncols, stride_row, stride_channel, stride_sample,
+                    l2_norm_f32<warp_size>(x, dst, ncols,
+                        src_stride_col, src_stride_row, src_stride_channel, src_stride_sample,
+                        dst_stride_col, dst_stride_row, dst_stride_channel, dst_stride_sample,
                         eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size);
                 });
             });
@@ -421,12 +451,20 @@ void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
     memcpy(&eps, dst->op_params, sizeof(float));
     GGML_ASSERT(eps >= 0.0f);
     const size_t ts0 = ggml_type_size(src0->type);
-    GGML_ASSERT(nb00 == ts0);
-    const int64_t s01 = nb01 / ts0;
-    const int64_t s02 = nb02 / ts0;
-    const int64_t s03 = nb03 / ts0;
+    const size_t tdst = ggml_type_size(dst->type);
+    GGML_ASSERT(nb00 % ts0 == 0 && nb01 % ts0 == 0 && nb02 % ts0 == 0 && nb03 % ts0 == 0);
+    GGML_ASSERT(nb0 % tdst == 0 && nb1 % tdst == 0 && nb2 % tdst == 0 && nb3 % tdst == 0);
+    const int64_t ss0 = nb00 / ts0;
+    const int64_t ss1 = nb01 / ts0;
+    const int64_t ss2 = nb02 / ts0;
+    const int64_t ss3 = nb03 / ts0;
+    const int64_t ds0 = nb0 / tdst;
+    const int64_t ds1 = nb1 / tdst;
+    const int64_t ds2 = nb2 / tdst;
+    const int64_t ds3 = nb3 / tdst;
 
-    norm_f32_sycl(src0_dd, dst_dd, ne00, ne01, ne02, ne03, s01, s02, s03, eps, main_stream, ctx.device);
+    norm_f32_sycl(src0_dd, dst_dd, ne00, ne01, ne02, ne03,
+        ss0, ss1, ss2, ss3, ds0, ds1, ds2, ds3, eps, main_stream, ctx.device);
 }
 
 void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
@@ -465,11 +503,19 @@ void ggml_sycl_op_rms_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
 
     GGML_TENSOR_UNARY_OP_LOCALS
     const size_t ts0 = ggml_type_size(src0->type);
-    GGML_ASSERT(nb00 == ts0);
-    const int64_t s01 = nb01 / ts0;
-    const int64_t s02 = nb02 / ts0;
-    const int64_t s03 = nb03 / ts0;
-    rms_norm_f32_sycl(src0_dd, dst_dd, ne00, ne01, ne02, ne03, s01, s02, s03, eps, main_stream, ctx.device);
+    const size_t tdst = ggml_type_size(dst->type);
+    GGML_ASSERT(nb00 % ts0 == 0 && nb01 % ts0 == 0 && nb02 % ts0 == 0 && nb03 % ts0 == 0);
+    GGML_ASSERT(nb0 % tdst == 0 && nb1 % tdst == 0 && nb2 % tdst == 0 && nb3 % tdst == 0);
+    const int64_t ss0 = nb00 / ts0;
+    const int64_t ss1 = nb01 / ts0;
+    const int64_t ss2 = nb02 / ts0;
+    const int64_t ss3 = nb03 / ts0;
+    const int64_t ds0 = nb0 / tdst;
+    const int64_t ds1 = nb1 / tdst;
+    const int64_t ds2 = nb2 / tdst;
+    const int64_t ds3 = nb3 / tdst;
+    rms_norm_f32_sycl(src0_dd, dst_dd, ne00, ne01, ne02, ne03,
+        ss0, ss1, ss2, ss3, ds0, ds1, ds2, ds3, eps, main_stream, ctx.device);
 }
 
 void ggml_sycl_op_rms_norm_back(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
@@ -644,13 +690,21 @@ void ggml_sycl_op_l2_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
     GGML_ASSERT(eps >= 0.0f);
 
     const size_t ts0 = ggml_type_size(src0->type);
-    GGML_ASSERT(nb00 == ts0);
-    const int64_t s01 = nb01 / ts0;
-    const int64_t s02 = nb02 / ts0;
-    const int64_t s03 = nb03 / ts0;
+    const size_t tdst = ggml_type_size(dst->type);
+    GGML_ASSERT(nb00 % ts0 == 0 && nb01 % ts0 == 0 && nb02 % ts0 == 0 && nb03 % ts0 == 0);
+    GGML_ASSERT(nb0 % tdst == 0 && nb1 % tdst == 0 && nb2 % tdst == 0 && nb3 % tdst == 0);
+    const int64_t ss0 = nb00 / ts0;
+    const int64_t ss1 = nb01 / ts0;
+    const int64_t ss2 = nb02 / ts0;
+    const int64_t ss3 = nb03 / ts0;
+    const int64_t ds0 = nb0 / tdst;
+    const int64_t ds1 = nb1 / tdst;
+    const int64_t ds2 = nb2 / tdst;
+    const int64_t ds3 = nb3 / tdst;
 
     /*support both WARP_SIZE or WARP_32_SIZE in code
       choose by hardware for better performance
     */
-    l2_norm_f32_sycl<WARP_SIZE>(src0_d, dst_d, ne00, ne01, ne02, ne03, s01, s02, s03, eps, stream, ctx.device);
+    l2_norm_f32_sycl<WARP_SIZE>(src0_d, dst_d, ne00, ne01, ne02, ne03,
+            ss0, ss1, ss2, ss3, ds0, ds1, ds2, ds3, eps, stream, ctx.device);
 }

ggml/src/ggml-sycl/softmax.cpp

@@ -126,7 +126,7 @@ static void soft_max_f32(const float *         x,
             break;
         }
 
-        const float val = sycl::native::exp(vals[col] - max_val);
+        const float val = sycl::native::exp(sycl::max(vals[col] - max_val, -80.0f));
         tmp += val;
         vals[col] = val;
     }
@@ -154,7 +154,7 @@ static void soft_max_f32(const float *         x,
         tmp = warp_reduce_sum<WARP_SIZE>(tmp);
     }
     if (sinks) {
-        tmp += sycl::native::exp(sinks[i02] - max_val);
+        tmp += sycl::native::exp(sycl::max(sinks[i02] - max_val, -80.0f));
     }
     const float inv_sum = 1.0f / tmp;
 

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

@@ -308,6 +308,7 @@ enum vk_device_architecture {
     AMD_RDNA1,
     AMD_RDNA2,
     AMD_RDNA3,
+    INTEL_XE1,
     INTEL_XE2,
     NVIDIA_PRE_TURING,
     NVIDIA_TURING,
@@ -365,21 +366,26 @@ static vk_device_architecture get_device_architecture(const vk::PhysicalDevice&
         const std::vector<vk::ExtensionProperties> ext_props = device.enumerateDeviceExtensionProperties();
 
         bool subgroup_size_control = false;
+        bool integer_dot_product = false;
 
         for (const auto& properties : ext_props) {
             if (strcmp("VK_EXT_subgroup_size_control", properties.extensionName) == 0) {
                 subgroup_size_control = true;
+            } else if (strcmp("VK_KHR_shader_integer_dot_product", properties.extensionName) == 0) {
+                integer_dot_product = true;
             }
         }
 
-        if (!subgroup_size_control) {
+        if (!subgroup_size_control || !integer_dot_product) {
             return vk_device_architecture::OTHER;
         }
 
         vk::PhysicalDeviceProperties2 props2;
         vk::PhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_size_control_props;
+        vk::PhysicalDeviceShaderIntegerDotProductPropertiesKHR integer_dot_props;
 
         props2.pNext = &subgroup_size_control_props;
+        subgroup_size_control_props.pNext = &integer_dot_props;
         device.getProperties2(&props2);
 
         if (subgroup_size_control_props.minSubgroupSize == 16) {
@@ -388,6 +394,9 @@ static vk_device_architecture get_device_architecture(const vk::PhysicalDevice&
             // https://www.intel.com/content/www/us/en/content-details/824434/2024-intel-tech-tour-xe2-and-lunar-lake-s-gpu.html
             // https://www.intel.com/content/www/us/en/docs/oneapi/optimization-guide-gpu/2025-0/intel-xe-gpu-architecture.html
             return vk_device_architecture::INTEL_XE2;
+        } else if (subgroup_size_control_props.minSubgroupSize == 8 &&
+                 integer_dot_product && integer_dot_props.integerDotProduct4x8BitPackedSignedAccelerated) {
+            return vk_device_architecture::INTEL_XE1;
         }
     } else if (props.vendorID == VK_VENDOR_ID_NVIDIA) {
         const std::vector<vk::ExtensionProperties> ext_props = device.enumerateDeviceExtensionProperties();
@@ -3837,7 +3846,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
             l_warptile = { 256, 128, 128, 16, subgroup_size_8, 64, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
             l_warptile_mmq = l_warptile_mmq_int = { 256, 128, 128, 32, subgroup_size_8, 64, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
             l_warptile_mmq_int_k = { 256, 128, 128, 32, subgroup_size_16, 64, 1, 4, 2, 1, subgroup_size_16 };
-        } else if (device->vendor_id == VK_VENDOR_ID_INTEL && device->coopmat_support && device->architecture == INTEL_XE2) {
+        } else if (device->vendor_id == VK_VENDOR_ID_INTEL && device->coopmat_support) {
             // Xe2/Xe3 with coopmat enabled - warptile performance tuning
             l_warptile = { 512, 128, 128, 16, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
             l_warptile_mmq = { 512, 128, 128, 32, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
@@ -4710,7 +4719,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     }
     uint32_t rm_iq = 2 * rm_kq;
 
-    const bool use_subgroups = device->subgroup_arithmetic && device->architecture != vk_device_architecture::AMD_GCN;
+    const bool use_subgroups = device->subgroup_arithmetic;
     // Ensure a subgroup size >= 16 is available
     const bool use_subgroups16 = use_subgroups && subgroup_min_size_16;
 
@@ -6361,9 +6370,8 @@ static vk_device ggml_vk_get_device(size_t idx) {
                 break;
             case VK_VENDOR_ID_INTEL: {
                 // Current Windows driver does not expose BF16 support.
-                // We only want to use l_warptile if coopmat is available and is Xe2+
-                const bool xe2_with_coopmat = device->coopmat_support && device->architecture == INTEL_XE2;
-                const bool use_l_warptile = (i == GGML_TYPE_BF16) ? (device->coopmat_bf16_support && xe2_with_coopmat) : xe2_with_coopmat;
+                // We only want to use l_warptile if coopmat is available
+                const bool use_l_warptile = (i == GGML_TYPE_BF16) ? (device->coopmat_bf16_support && device->coopmat_support) : device->coopmat_support;
                 device->mul_mat_l[i] = use_l_warptile;
                 device->mul_mat_id_l[i] = use_l_warptile;
                 device->mul_mat_m[i] = true;
@@ -17890,9 +17898,9 @@ static bool ggml_vk_device_is_supported(const vk::PhysicalDevice & vkdev) {
 static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props, vk_device_architecture arch) {
     switch (props.vendorID) {
     case VK_VENDOR_ID_INTEL:
-        // Only allowing Xe2 GPU at the moment since Xe2 GPU can gain significant performance boost,
-        // while some older hardware (ex. Arc A770) has performance regressions
-        return arch == vk_device_architecture::INTEL_XE2;
+        // Only allowing Xe2/Xe3 GPU and integrated Xe GPUs at the moment since older hardware (ex. Arc A770) has performance regressions.
+        return (arch == vk_device_architecture::INTEL_XE2) ||
+            (arch == vk_device_architecture::INTEL_XE1 && props.deviceType == vk::PhysicalDeviceType::eIntegratedGpu && driver_props.driverID == vk::DriverId::eIntelProprietaryWindows);
     case VK_VENDOR_ID_AMD:
         if (driver_props.driverID == vk::DriverId::eAmdProprietary || driver_props.driverID == vk::DriverId::eAmdOpenSource) {
             // Workaround for AMD proprietary driver reporting support on all GPUs
@@ -17940,6 +17948,8 @@ static uint32_t ggml_vk_intel_shader_core_count(const vk::PhysicalDevice& vkdev)
     case 0xE20B:  // B580
     case 0xE211:  // Pro B60
         return 20;
+    case 0xB080:  // PTL Xe3 LPG 2x6 (12 subslices)
+        return 12;
     default:
         return 0;
     }

ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp

@@ -158,7 +158,7 @@ const uint32_t Csh_stride = BS_NPQ;
 #ifdef COOPMAT
 const uint32_t Csh_len    = BS_K * Csh_stride;
 #else
-const uint32_t Csh_len    = csh_store != 0 ? BS_K * Csh_stride : 1;
+const uint32_t Csh_len    = csh_store != 0 ? BS_K * Csh_stride : 8; // 8 to workaround compiler bug
 #endif
 shared SHMEM_TYPE Csh[Csh_len];  // K x NPQ
 #endif

ggml/src/ggml-vulkan/vulkan-shaders/conv3d_mm.comp

@@ -144,7 +144,7 @@ const uint32_t Csh_stride = BS_NPQ;
 #ifdef COOPMAT
 const uint32_t Csh_len    = BS_K * Csh_stride;
 #else
-const uint32_t Csh_len    = csh_store != 0 ? BS_K * Csh_stride : 1;
+const uint32_t Csh_len    = csh_store != 0 ? BS_K * Csh_stride : 8; // 8 to workaround compiler bug
 #endif
 shared SHMEM_TYPE Csh[Csh_len];  // K x NPQ
 #endif

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

@@ -28,13 +28,10 @@ vec2 cache_b_ds;
 
 #include "mul_mat_vecq_funcs.glsl"
 
-void iter(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i) {
+void iter(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const uint first_row, const uint num_rows, const uint col, const uint b_qs_idx) {
     [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
-        const uint col = i*BLOCK_SIZE + tid*K_PER_ITER;
-
         // Preload data_b block
         const uint b_block_idx = (j*p.batch_stride_b + col) / QUANT_K_Q8_1 + b_offset;
-        const uint b_qs_idx = tid % (32 / K_PER_ITER);
         const uint b_block_idx_outer = b_block_idx / 4;
         const uint b_block_idx_inner = b_block_idx % 4;
         cache_b_ds = vec2(data_b[b_block_idx_outer].ds[b_block_idx_inner]);
@@ -91,35 +88,35 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
         }
     }
 
-    uint num_iters = p.ncols / (K_PER_ITER * BLOCK_SIZE);
-    if (num_iters * K_PER_ITER * BLOCK_SIZE + K_PER_ITER*tid < p.ncols) {
+    const uint col_stride = K_PER_ITER * BLOCK_SIZE;
+    uint num_iters = p.ncols / col_stride;
+    if (num_iters * col_stride + K_PER_ITER * tid < p.ncols) {
         num_iters++;
     }
-    int unroll_count = 4;
-    uint unrolled_iters = num_iters & ~(unroll_count - 1);
 
-    uint i = 0;
-    while (i < unrolled_iters) {
+    const uint b_qs_idx = tid % (32 / K_PER_ITER);
+    uint col = tid * K_PER_ITER;
+    while (num_iters >= 4) {
         // Manually partially unroll the loop
-        [[unroll]] for (uint k = 0; k < unroll_count; ++k) {
-            iter(temp, first_row, num_rows, tid, i*K_PER_ITER);
-            i++;
+        [[unroll]] for (uint k = 0; k < 4; ++k) {
+            iter(temp, first_row, num_rows, col, b_qs_idx);
+            col += col_stride;
         }
+
+        num_iters -= 4;
     }
 
-    unroll_count = 2;
-    unrolled_iters = num_iters & ~(unroll_count - 1);
-
-    while (i < unrolled_iters) {
+    if (num_iters >= 2) {
         // Manually partially unroll the loop
-        [[unroll]] for (uint k = 0; k < unroll_count; ++k) {
-            iter(temp, first_row, num_rows, tid, i*K_PER_ITER);
-            i++;
-        }
+        iter(temp, first_row, num_rows, col, b_qs_idx);
+        col += col_stride;
+        iter(temp, first_row, num_rows, col, b_qs_idx);
+        col += col_stride;
+        num_iters -= 2;
     }
-    while (i < num_iters) {
-        iter(temp, first_row, num_rows, tid, i*K_PER_ITER);
-        i++;
+
+    if (num_iters > 0) {
+        iter(temp, first_row, num_rows, col, b_qs_idx);
     }
 
     reduce_result(temp, d_offset, first_row, num_rows, tid);

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

@@ -42,7 +42,7 @@ float op_leaky_relu(float x) {
 }
 
 float op_step(float x) {
-    return x >= 0.0f ? 1.0f : 0.0f;
+    return x > 0.0f ? 1.0f : 0.0f;
 }
 
 float op_tanh(float x) {

scripts/sync-ggml.last

@@ -1 +1 @@
-707321c4cf6d21cb4bc831aa8b687dbf01a521ce
+eced84c86f8b012c752c016f7fe789adea168e1e

src/llama-model.cpp

@@ -700,6 +700,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_160M:          return "160M";
         case LLM_TYPE_190M:          return "190M";
         case LLM_TYPE_220M:          return "220M";
+        case LLM_TYPE_230M:          return "230M";
         case LLM_TYPE_250M:          return "250M";
         case LLM_TYPE_256M:          return "256M";
         case LLM_TYPE_270M:          return "270M";

src/llama-model.h

@@ -36,6 +36,7 @@ enum llm_type {
     LLM_TYPE_160M,
     LLM_TYPE_190M,
     LLM_TYPE_220M,
+    LLM_TYPE_230M,
     LLM_TYPE_250M,
     LLM_TYPE_256M,
     LLM_TYPE_270M,

src/llama-quant.cpp

@@ -847,7 +847,7 @@ static void init_quantize_state_counters(quantize_state_impl & qs, std::vector<t
             qs.has_tied_embeddings = false;
         }
     }
-    qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)qs.model.hparams.n_layer();
+    qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)qs.model.hparams.n_layer_all;
 }
 
 //

src/models/lfm2.cpp

@@ -13,6 +13,7 @@ void llama_model_lfm2::load_arch_hparams(llama_model_loader & ml) {
     hparams.n_layer_dense_lead = hparams.n_layer();
 
     switch (hparams.n_ff()) {
+        case  2560: type = LLM_TYPE_230M; break;
         case  4608: type = LLM_TYPE_350M; break;
         case  6912: type = LLM_TYPE_700M; break;
         case  8192: type = LLM_TYPE_1_2B; break;

src/models/mamba-base.cpp

@@ -169,7 +169,6 @@ ggml_tensor * llm_build_mamba_base::build_mamba2_layer(llm_graph_input_rs * inp,
     GGML_ASSERT(ubatch.equal_seqs());
     GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
     GGML_ASSERT(d_inner % n_head  == 0);
-    GGML_ASSERT(d_inner % d_state == 0);
     GGML_ASSERT(d_inner % n_group == 0);
 
     ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);

src/models/mamba2.cpp

@@ -39,10 +39,11 @@ void llama_model_mamba2::load_arch_tensors(llama_model_loader &) {
     const int64_t d_inner = hparams.ssm_d_inner;
     const int64_t d_state = hparams.ssm_d_state;
     const int64_t n_group = hparams.ssm_n_group;
-    const int64_t d_in_proj = 2*d_inner + 2*n_group*d_state + n_head;
+    const int64_t dt_rank  = hparams.ssm_dt_rank;
+
+    const int64_t conv_dim = d_inner + 2 * n_group * d_state;
+    const int64_t d_in_proj = d_inner + conv_dim + dt_rank;
 
-    // only an expansion factor of 2 is supported for now
-    GGML_ASSERT(2 * n_embd == d_inner);
 
     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -68,11 +69,11 @@ void llama_model_mamba2::load_arch_tensors(llama_model_loader &) {
         layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner + 2*n_group*d_state}, 0);
         layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner + 2*n_group*d_state}, 0);
 
-        layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {n_head}, 0);
+        layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {dt_rank}, 0);
 
         // no "weight" suffix for these
-        layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_head}, 0);
-        layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {1, n_head}, 0);
+        layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, dt_rank}, 0);
+        layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {1, dt_rank}, 0);
 
         layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {d_inner / n_group, n_group}, 0);
 

tests/CMakeLists.txt

@@ -302,9 +302,9 @@ target_link_libraries(${TEST_TARGET} PRIVATE llama)
 llama_build_and_test(test-alloc.cpp)
 target_include_directories(test-alloc PRIVATE ${PROJECT_SOURCE_DIR}/ggml/src)
 
-llama_build(export-graph-ops.cpp)
-target_include_directories(export-graph-ops PRIVATE ${PROJECT_SOURCE_DIR}/ggml/src)
+llama_build(test-export-graph-ops.cpp)
+target_include_directories(test-export-graph-ops PRIVATE ${PROJECT_SOURCE_DIR}/ggml/src)
 if (TARGET gguf-model-data)
-    target_link_libraries(export-graph-ops PRIVATE gguf-model-data)
-    target_compile_definitions(export-graph-ops PRIVATE LLAMA_HF_FETCH)
+    target_link_libraries(test-export-graph-ops PRIVATE gguf-model-data)
+    target_compile_definitions(test-export-graph-ops PRIVATE LLAMA_HF_FETCH)
 endif()

tests/test-backend-ops.cpp

@@ -2890,12 +2890,17 @@ struct test_cpy : public test_case {
     const std::array<int64_t, 4> ne_dst;
     const std::array<int64_t, 4> permute_src;
     const std::array<int64_t, 4> permute_dst;
+    const std::array<int64_t, 4> dst_alloc; // if set, dst is a view into a larger buffer (strided)
     bool _src_use_permute;
     bool _dst_use_permute;
     bool _src_transpose;
     bool _use_dst_shape;
+    bool _use_dst_alloc;
 
     std::string vars() override {
+        if (_use_dst_alloc) {
+            return VARS_TO_STR8(type_src, type_dst, ne_src, ne_dst, permute_src, permute_dst, _src_transpose, dst_alloc);
+        }
         if (_use_dst_shape) {
             return VARS_TO_STR7(type_src, type_dst, ne_src, ne_dst, permute_src, permute_dst, _src_transpose);
         }
@@ -2943,12 +2948,15 @@ struct test_cpy : public test_case {
             std::array<int64_t, 4> ne_dst = {-1, -1, -1, -1},
             std::array<int64_t, 4> permute_src = {0, 0, 0, 0},
             std::array<int64_t, 4> permute_dst = {0, 0, 0, 0},
-            bool transpose_src = false)
+            bool transpose_src = false,
+            std::array<int64_t, 4> dst_alloc = {0, 0, 0, 0})
         : type_src(type_src), type_dst(type_dst), ne_src(ne_src), ne_dst(ne_dst), permute_src(permute_src), permute_dst(permute_dst),
+          dst_alloc(dst_alloc),
           _src_use_permute(permute_src[0] + permute_src[1] + permute_src[2] + permute_src[3] > 0),
           _dst_use_permute(permute_dst[0] + permute_dst[1] + permute_dst[2] + permute_dst[3] > 0),
           _src_transpose(transpose_src),
-          _use_dst_shape(ne_dst[0] >= 0 && ne_dst[1] >= 0 && ne_dst[2] >= 0 && ne_dst[3] >= 0){}
+          _use_dst_shape(ne_dst[0] >= 0 && ne_dst[1] >= 0 && ne_dst[2] >= 0 && ne_dst[3] >= 0),
+          _use_dst_alloc(dst_alloc[0] > 0){}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * src = ggml_new_tensor(ctx, type_src, 4, ne_src.data());
@@ -2966,12 +2974,23 @@ struct test_cpy : public test_case {
         }
 
         std::array<int64_t, 4> dst_ne = _use_dst_shape ? ne_dst : std::array<int64_t, 4>{src->ne[0], src->ne[1], src->ne[2], src->ne[3]};
-        ggml_tensor * dst = ggml_new_tensor(ctx, type_dst, 4, dst_ne.data());
-        ggml_set_name(dst, "dst");
+        ggml_tensor * dst;
 
-        if (_dst_use_permute) {
-            dst = ggml_permute(ctx, dst, permute_dst[0], permute_dst[1], permute_dst[2], permute_dst[3]);
-            ggml_set_name(dst, "dst_permuted");
+        if (_use_dst_alloc) {
+            // view a sub-block of a larger buffer -> strided dst
+            ggml_tensor * dst_buf = ggml_new_tensor(ctx, type_dst, 4, dst_alloc.data());
+            ggml_set_name(dst_buf, "dst_buf");
+            dst = ggml_view_4d(ctx, dst_buf, dst_ne[0], dst_ne[1], dst_ne[2], dst_ne[3],
+                dst_buf->nb[1], dst_buf->nb[2], dst_buf->nb[3], 0);
+            ggml_set_name(dst, "dst_view");
+        } else {
+            dst = ggml_new_tensor(ctx, type_dst, 4, dst_ne.data());
+            ggml_set_name(dst, "dst");
+
+            if (_dst_use_permute) {
+                dst = ggml_permute(ctx, dst, permute_dst[0], permute_dst[1], permute_dst[2], permute_dst[3]);
+                ggml_set_name(dst, "dst_permuted");
+            }
         }
 
         ggml_tensor * out = ggml_cpy(ctx, src, dst);
@@ -7973,6 +7992,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
             }
         }
     }
+    for (auto kernel_type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
+        test_cases.emplace_back(new test_conv_2d({ 256, 256, 192, 1 }, { 3, 3, 192, 96 }, kernel_type, 1, 1, 1, 1, 1, 1, false));
+    }
 
     // sycl backend will limit task global_range < MAX_INT
     // test cases for 2D im2col with large input W and H (occurs in stable-diffusion)
@@ -8176,6 +8198,10 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_cpy(GGML_TYPE_I32, GGML_TYPE_I32, {256, 4, 1, 1}, {-1,-1,-1,-1}, {0, 0, 0, 0}, {0, 0, 0, 0}, true));
     test_cases.emplace_back(new test_cpy(GGML_TYPE_I32, GGML_TYPE_I32, {256, 1, 4, 1}, {-1,-1,-1,-1}, {1, 2, 0, 3}, {0, 0, 0, 0}));
     test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {256, 1, 4, 1}, {-1,-1,-1,-1}, {1, 2, 0, 3}, {0, 0, 0, 0}));
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {2, 2097121, 1, 1}, {-1,-1,-1,-1}, {1, 0, 2, 3}));
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {2, 2, 524281, 1}, {-1,-1,-1,-1}, {1, 0, 2, 3}));
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {128, 2, 3, 1}, {128, 2, 3, 1}, {0, 0, 0, 0}, {0, 0, 0, 0}, false, {128, 4, 3, 1})); // strided dst
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_F16, GGML_TYPE_F16, {128, 2, 3, 1}, {128, 2, 3, 1}, {0, 0, 0, 0}, {0, 0, 0, 0}, false, {128, 4, 3, 1})); // strided dst
 
     // CPY - different src/dst shapes (reshaping via CPY)
     // Use permutations of {3, 5, 7, 32}. Total elements: 3*5*7*32 = 3360.
@@ -8670,6 +8696,12 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                                                   256, 16, 16, {ne2, 1}, {1, 1}));
     }
 
+    // nr2 sweep to cover the cublasSgemmBatched pointer-array path (dps2 > 1)
+    for (int64_t nr2 : {8, 16, 32}) {
+        test_cases.emplace_back(new test_out_prod(GGML_TYPE_F32, GGML_TYPE_F32,
+                                                  256, 16, 16, {1, 1}, {nr2, 1}));
+    }
+
     // add_id
     for (ggml_type type_a : {GGML_TYPE_F32}) {
         for (ggml_type type_b : {GGML_TYPE_F32}) {
@@ -9932,7 +9964,7 @@ static void usage(char ** argv) {
     printf("    --output specifies output format (default: console, options: console, sql, csv)\n");
     printf("    --list-ops lists all available GGML operations\n");
     printf("    --show-coverage shows test coverage\n");
-    printf("    --test-file reads test operators from a test file generated by llama-export-graph-ops\n");
+    printf("    --test-file reads test operators from a test file generated by test-export-graph-ops\n");
     printf("    -j <n> runs tests using <n> parallel worker threads (default: 1, test mode only)\n");
 }
 

tests/test-chat-template.cpp

@@ -135,7 +135,7 @@ int main(int argc, char ** argv) {
             output_path = args[i + 1];
             i++;
         } else if (args[i] == "--no-common") {
-            use_common = true;
+            use_common = false;
         } else if (tmpl_path.empty()) {
             tmpl_path = args[i];
         } else {

tests/test-export-graph-ops.cpp

@@ -185,7 +185,7 @@ int main(int argc, char ** argv) {
             return 1;
         }
 #else
-        LOG_ERR("export-graph-ops compiled without HF fetch support\n");
+        LOG_ERR("test-export-graph-ops compiled without HF fetch support\n");
         return 1;
 #endif
     }

tests/test-quantize-fns.cpp

@@ -102,21 +102,34 @@ static float dot_product_error(const ggml_type_traits * qfns, const ggml_type_tr
     return fabsf(result - dot_ref) / test_size;
 }
 
-int main(int argc, char * argv[]) {
-    bool verbose = false;
-    const size_t test_size = 32 * 128;
+static int test_vec_dot_f32(bool verbose) {
+    const auto * f32 = ggml_get_type_traits_cpu(GGML_TYPE_F32);
+    int num_failed = 0;
+    for (int n : {1, 2, 3, 5, 7, 8, 15, 16, 17, 31, 33, 63, 67, 127, 129, 193, 255, 1023}) {
+        std::vector<float> a(n);
+        std::vector<float> b(n);
+        generate_data(0.0, n, a.data());
+        generate_data(1.0, n, b.data());
 
-    std::string arg;
-    for (int i = 1; i < argc; i++) {
-        arg = argv[i];
+        float result = 0.0f;
+        f32->vec_dot(n, &result, 0, a.data(), 0, b.data(), 0, 1);
+        const float ref = dot_product(a.data(), b.data(), n);
+        const float error = fabsf(result - ref) / n;
 
-        if (arg == "-v") {
-            verbose = true;
-        } else {
-            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
-            return 1;
+        const bool failed = !(error < MAX_QUANTIZATION_REFERENCE_ERROR);
+        num_failed += failed;
+        if (failed || verbose) {
+            printf(" f32 vec_dot n=%4d:                 %s (ref=%f got=%f err=%f)\n",
+                   n, RESULT_STR[failed], ref, result, error);
         }
     }
+    return num_failed;
+}
+
+static int test_vec_dot_q(bool verbose) {
+    int num_failed = 0;
+
+    const size_t test_size = 32 * 128;
 
     std::vector<float> test_data(test_size);
     std::vector<float> test_data2(test_size);
@@ -124,11 +137,6 @@ int main(int argc, char * argv[]) {
     generate_data(0.0, test_data.size(), test_data.data());
     generate_data(1.0, test_data2.size(), test_data2.data());
 
-    ggml_cpu_init();
-
-    int num_failed = 0;
-    bool failed = false;
-
     for (int i = 0; i < GGML_TYPE_COUNT; i++) {
         ggml_type type = (ggml_type) i;
         const auto * qfns = ggml_get_type_traits(type);
@@ -156,7 +164,7 @@ int main(int argc, char * argv[]) {
                 type == GGML_TYPE_IQ3_S   ? MAX_QUANTIZATION_TOTAL_ERROR_3BITS :
                 type == GGML_TYPE_IQ3_XXS ? MAX_QUANTIZATION_TOTAL_ERROR_3BITS_XXS :
                 type == GGML_TYPE_NVFP4   ? MAX_QUANTIZATION_TOTAL_ERROR_FP4 : MAX_QUANTIZATION_TOTAL_ERROR;
-            failed = !(total_error < max_quantization_error);
+            bool failed = !(total_error < max_quantization_error);
             num_failed += failed;
             if (failed || verbose) {
                 printf("%5s absolute quantization error:    %s (%f)\n", ggml_type_name(type), RESULT_STR[failed], total_error);
@@ -171,15 +179,15 @@ int main(int argc, char * argv[]) {
 
             const float vec_dot_error = dot_product_error(qfns, qfns_cpu, test_size, test_data.data(), test_data2.data());
             const float max_allowed_error = type == GGML_TYPE_Q2_K || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ2_XXS ||
-                                            type == GGML_TYPE_IQ3_XXS || type == GGML_TYPE_IQ3_S || type == GGML_TYPE_IQ2_S
-                                          ? MAX_DOT_PRODUCT_ERROR_LOWBIT
-                                          : type == GGML_TYPE_Q1_0
-                                          ? MAX_DOT_PRODUCT_ERROR_BINARY
-                                          : type == GGML_TYPE_TQ1_0 || type == GGML_TYPE_TQ2_0
-                                          ? MAX_DOT_PRODUCT_ERROR_TERNARY
-                                          : type == GGML_TYPE_NVFP4
-                                          ? MAX_DOT_PRODUCT_ERROR_FP4
-                                          : MAX_DOT_PRODUCT_ERROR;
+                type == GGML_TYPE_IQ3_XXS || type == GGML_TYPE_IQ3_S || type == GGML_TYPE_IQ2_S
+                ? MAX_DOT_PRODUCT_ERROR_LOWBIT
+                : type == GGML_TYPE_Q1_0
+                ? MAX_DOT_PRODUCT_ERROR_BINARY
+                : type == GGML_TYPE_TQ1_0 || type == GGML_TYPE_TQ2_0
+                ? MAX_DOT_PRODUCT_ERROR_TERNARY
+                : type == GGML_TYPE_NVFP4
+                ? MAX_DOT_PRODUCT_ERROR_FP4
+                : MAX_DOT_PRODUCT_ERROR;
             failed = !(vec_dot_error < max_allowed_error);
             num_failed += failed;
             if (failed || verbose) {
@@ -188,6 +196,31 @@ int main(int argc, char * argv[]) {
         }
     }
 
+    return num_failed;
+}
+
+int main(int argc, char * argv[]) {
+    bool verbose = false;
+
+    std::string arg;
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+
+        if (arg == "-v") {
+            verbose = true;
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            return 1;
+        }
+    }
+
+    ggml_cpu_init();
+
+    int num_failed = 0;
+
+    num_failed += test_vec_dot_f32(verbose);
+    num_failed += test_vec_dot_q(verbose);
+
     if (num_failed || verbose) {
         printf("%d tests failed\n", num_failed);
     }

tests/test-thread-safety.cpp

@@ -146,6 +146,8 @@ int main(int argc, char ** argv) {
                 }
 
                 LOG_INF("Model %d/%d, Context %d/%d: %s\n\n", m + 1, num_models, c + 1, num_contexts, result.c_str());
+
+                llama_synchronize(ctx.get());
             });
         }
     }

tools/llama-bench/llama-bench.cpp

@@ -1035,25 +1035,23 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
 
     if (!params.hf_repo.empty()) {
         for (size_t i = 0; i < params.hf_repo.size(); i++) {
-            common_params_model model;
-
-            if (params.hf_file.empty() || params.hf_file[i].empty()) {
-                model.hf_repo = params.hf_repo[i];
-            } else {
-                model.hf_repo = params.hf_repo[i];
-                model.hf_file = params.hf_file[i];
+            common_params p;
+            p.hf_token      = params.hf_token;
+            p.offline       = params.offline;
+            p.model.hf_repo = params.hf_repo[i];
+            if (!params.hf_file.empty() && !params.hf_file[i].empty()) {
+                p.model.hf_file = params.hf_file[i];
             }
 
-            common_download_opts opts;
-            opts.bearer_token = params.hf_token;
-            opts.offline         = params.offline;
-            auto download_result = common_download_model(model, opts);
-            if (download_result.model_path.empty()) {
+            // only the text model file is needed
+            common_models_handler models_handler = common_models_handler_init(p, LLAMA_EXAMPLE_BENCH);
+            common_models_handler_apply(models_handler, p);
+            if (p.model.path.empty()) {
                 fprintf(stderr, "error: failed to download model from HuggingFace\n");
                 exit(1);
             }
 
-            params.model.push_back(download_result.model_path);
+            params.model.push_back(p.model.path);
         }
     }
 

tools/mtmd/CMakeLists.txt

@@ -115,22 +115,28 @@ if (TARGET mtmd)
     endif()
 endif()
 
-add_executable(llama-llava-cli    deprecation-warning.cpp)
-add_executable(llama-gemma3-cli   deprecation-warning.cpp)
-add_executable(llama-minicpmv-cli deprecation-warning.cpp)
-add_executable(llama-qwen2vl-cli  deprecation-warning.cpp)
+# Gate CLI binaries on LLAMA_BUILD_TOOLS so that standalone library-only
+# builds (LLAMA_BUILD_MTMD=ON with LLAMA_BUILD_TOOLS=OFF — e.g. Apple
+# XCFramework packaging) skip the executables entirely. LLAMA_BUILD_COMMON
+# defaults to ON in standalone builds, so we cannot rely on it for gating.
+if (LLAMA_BUILD_TOOLS)
+    add_executable(llama-llava-cli    deprecation-warning.cpp)
+    add_executable(llama-gemma3-cli   deprecation-warning.cpp)
+    add_executable(llama-minicpmv-cli deprecation-warning.cpp)
+    add_executable(llama-qwen2vl-cli  deprecation-warning.cpp)
 
-set(TARGET llama-mtmd-cli)
-add_executable         (${TARGET} mtmd-cli.cpp)
-set_target_properties  (${TARGET} PROPERTIES OUTPUT_NAME llama-mtmd-cli)
-if(LLAMA_TOOLS_INSTALL)
-    install(TARGETS ${TARGET} RUNTIME)
+    set(TARGET llama-mtmd-cli)
+    add_executable         (${TARGET} mtmd-cli.cpp)
+    set_target_properties  (${TARGET} PROPERTIES OUTPUT_NAME llama-mtmd-cli)
+    if(LLAMA_TOOLS_INSTALL)
+        install(TARGETS ${TARGET} RUNTIME)
+    endif()
+    target_link_libraries  (${TARGET} PRIVATE llama-common mtmd Threads::Threads)
+    target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+    # mtmd-debug tool
+    add_executable(llama-mtmd-debug debug/mtmd-debug.cpp)
+    set_target_properties(llama-mtmd-debug PROPERTIES OUTPUT_NAME llama-mtmd-debug)
+    target_link_libraries(llama-mtmd-debug PRIVATE llama-common mtmd Threads::Threads)
+    target_compile_features(llama-mtmd-debug PRIVATE cxx_std_17)
 endif()
-target_link_libraries  (${TARGET} PRIVATE llama-common mtmd Threads::Threads)
-target_compile_features(${TARGET} PRIVATE cxx_std_17)
-
-# mtmd-debug tool
-add_executable(llama-mtmd-debug debug/mtmd-debug.cpp)
-set_target_properties(llama-mtmd-debug PROPERTIES OUTPUT_NAME llama-mtmd-debug)
-target_link_libraries(llama-mtmd-debug PRIVATE llama-common mtmd Threads::Threads)
-target_compile_features(llama-mtmd-debug PRIVATE cxx_std_17)

tools/mtmd/clip-model.h

@@ -55,8 +55,7 @@ struct clip_hparams {
     int32_t n_head = 0;
     int32_t n_head_kv = 0;
     int32_t n_layer = 0;
-    // idefics3
-    int32_t n_merge = 0; // number of patch merges **per-side**
+    int32_t n_merge = 1; // number of patch merges **per-side**
 
     // for preprocessor
     int32_t image_longest_edge = 0;
@@ -135,8 +134,7 @@ struct clip_hparams {
     int32_t custom_image_max_tokens = -1;
 
     void set_limit_image_tokens(int n_tokens_min, int n_tokens_max) {
-        const int cur_merge = n_merge == 0 ? 1 : n_merge;
-        const int patch_area = patch_size * patch_size * cur_merge * cur_merge;
+        const int patch_area = patch_size * patch_size * n_merge * n_merge;
         image_min_pixels = (custom_image_min_tokens > 0 ? custom_image_min_tokens : n_tokens_min) * patch_area;
         image_max_pixels = (custom_image_max_tokens > 0 ? custom_image_max_tokens : n_tokens_max) * patch_area;
         warmup_image_size = static_cast<int>(std::sqrt(image_max_pixels));
@@ -145,8 +143,7 @@ struct clip_hparams {
     void set_warmup_n_tokens(int n_tokens) {
         int n_tok_per_side = static_cast<int>(std::sqrt(n_tokens));
         GGML_ASSERT(n_tok_per_side * n_tok_per_side == n_tokens && "n_tokens must be n*n");
-        const int cur_merge = n_merge == 0 ? 1 : n_merge;
-        warmup_image_size = n_tok_per_side * patch_size * cur_merge;
+        warmup_image_size = n_tok_per_side * patch_size * n_merge;
         // TODO: support warmup size for custom token numbers
     }
     // sam vit deepseek-ocr

tools/mtmd/clip.cpp

@@ -1210,6 +1210,9 @@ struct clip_model_loader {
             {
                 std::vector<int> pinpoints;
                 get_arr_int(KEY_IMAGE_GRID_PINPOINTS, pinpoints, false);
+                if (pinpoints.size() % 2 != 0) {
+                    throw std::runtime_error(string_format("%s: image_grid_pinpoints must have an even number of elements, got %zu\n", __func__, pinpoints.size()));
+                }
                 if (!pinpoints.empty()) {
                     for (size_t i = 0; i < pinpoints.size(); i += 2) {
                         hparams.image_res_candidates.push_back({
@@ -1252,15 +1255,16 @@ struct clip_model_loader {
             }
 
             if (is_vision) {
-                int idx_mean = gguf_find_key(ctx_gguf.get(), KEY_IMAGE_MEAN);
-                int idx_std  = gguf_find_key(ctx_gguf.get(), KEY_IMAGE_STD);
-                GGML_ASSERT(idx_mean >= 0 && "image_mean not found");
-                GGML_ASSERT(idx_std >= 0  && "image_std not found");
-                const float * mean_data = (const float *) gguf_get_arr_data(ctx_gguf.get(), idx_mean);
-                const float * std_data  = (const float *) gguf_get_arr_data(ctx_gguf.get(), idx_std);
+                std::vector<float> image_mean;
+                std::vector<float> image_std;
+                get_arr_f32(KEY_IMAGE_MEAN, image_mean, false);
+                get_arr_f32(KEY_IMAGE_STD , image_std, false);
+                if (image_mean.size() < 3 || image_std.size() < 3) {
+                    throw std::runtime_error(string_format("%s: image_mean/image_std arrays must have at least 3 elements, got %zu and %zu\n", __func__, image_mean.size(), image_std.size()));
+                }
                 for (int i = 0; i < 3; ++i) {
-                    hparams.image_mean[i] = mean_data[i];
-                    hparams.image_std[i]  = std_data[i];
+                    hparams.image_mean[i] = image_mean[i];
+                    hparams.image_std[i]  = image_std[i];
                 }
             }
 
@@ -1686,8 +1690,8 @@ struct clip_model_loader {
                 if (hparams.image_size > 65536) {
                     throw std::runtime_error(string_format("%s: image_size (%d) is too large (max 65536)\n", __func__, hparams.image_size));
                 }
-                if (hparams.patch_size <= 0) {
-                    throw std::runtime_error(string_format("%s: patch_size (%d) must be greater than 0\n", __func__, hparams.patch_size));
+                if (hparams.patch_size <= 0 || hparams.patch_size >= 65536) {
+                    throw std::runtime_error(string_format("%s: patch_size (%d) must be positive and less than 65536\n", __func__, hparams.patch_size));
                 }
                 if (hparams.n_embd <= 0) {
                     throw std::runtime_error(string_format("%s: n_embd (%d) must be greater than 0\n", __func__, hparams.n_embd));
@@ -1695,6 +1699,9 @@ struct clip_model_loader {
                 if (hparams.image_max_pixels < hparams.image_min_pixels) {
                     throw std::runtime_error(string_format("%s: image_max_pixels (%d) is less than image_min_pixels (%d)\n", __func__, hparams.image_max_pixels, hparams.image_min_pixels));
                 }
+                if (hparams.n_merge < 0 || hparams.n_merge >= 65536) {
+                    throw std::runtime_error(string_format("%s: n_merge (%d) must be greater than 0 and less than 65536\n", __func__, hparams.n_merge));
+                }
             }
 
             LOG_INF("%s: projector:          %s\n", __func__, proj_type.c_str());
@@ -3067,6 +3074,29 @@ struct clip_model_loader {
         output = gguf_get_val_f32(ctx_gguf.get(), i);
     }
 
+    void get_arr_f32(const std::string & key, std::vector<float> & output, bool required = true) const {
+        const int i = gguf_find_key(ctx_gguf.get(), key.c_str());
+        if (i < 0) {
+            if (required) {
+                throw std::runtime_error("Key not found: " + key);
+            }
+            return;
+        }
+        const auto type = gguf_get_arr_type(ctx_gguf.get(), i);
+        if (type != GGUF_TYPE_FLOAT32) {
+            throw std::runtime_error(string_format("%s: array '%s' has type %d, expected %d (GGUF_TYPE_FLOAT32)\n", __func__, key.c_str(), type, GGUF_TYPE_FLOAT32));
+        }
+        const size_t n = gguf_get_arr_n(ctx_gguf.get(), i);
+        if (n > (size_t) std::numeric_limits<int>::max()) {
+            throw std::runtime_error(string_format("%s: array '%s' is too large (%zu elements)\n", __func__, key.c_str(), n));
+        }
+        output.resize(n);
+        const float * values = (const float *)gguf_get_arr_data(ctx_gguf.get(), i);
+        for (size_t j = 0; j < n; ++j) {
+            output[j] = values[j];
+        }
+    }
+
     void get_string(const std::string & key, std::string & output, bool required = true) const {
         const int i = gguf_find_key(ctx_gguf.get(), key.c_str());
         if (i < 0) {
@@ -3086,11 +3116,18 @@ struct clip_model_loader {
             }
             return;
         }
-        int n = gguf_get_arr_n(ctx_gguf.get(), i);
+        const auto type = gguf_get_arr_type(ctx_gguf.get(), i);
+        if (type != GGUF_TYPE_INT32) {
+            throw std::runtime_error(string_format("%s: array '%s' has type %d, expected %d (GGUF_TYPE_INT32)\n", __func__, key.c_str(), type, GGUF_TYPE_INT32));
+        }
+        const size_t n = gguf_get_arr_n(ctx_gguf.get(), i);
+        if (n > (size_t) std::numeric_limits<int>::max()) {
+            throw std::runtime_error(string_format("%s: array '%s' is too large (%zu elements)\n", __func__, key.c_str(), n));
+        }
         output.resize(n);
         const int32_t * values = (const int32_t *)gguf_get_arr_data(ctx_gguf.get(), i);
-        for (int i = 0; i < n; ++i) {
-            output[i] = values[i];
+        for (size_t j = 0; j < n; ++j) {
+            output[j] = values[j];
         }
     }
 
@@ -3364,8 +3401,8 @@ int clip_n_output_tokens(const clip_ctx * ctx, const clip_image_f32 * img) {
             {
                 // dynamic size
                 int n_merge = ctx->model.hparams.n_merge;
-                int n_patches_x = img->nx() / patch_size / (n_merge > 0 ? n_merge : 1);
-                int n_patches_y = img->ny() / patch_size / (n_merge > 0 ? n_merge : 1);
+                int n_patches_x = img->nx() / patch_size / n_merge;
+                int n_patches_y = img->ny() / patch_size / n_merge;
                 if (ctx->model.token_embd_img_break) {
                     n_patches = n_patches_y * n_patches_x + n_patches_y - 1; // + one [IMG_BREAK] per row, except the last row
                 } else {

tools/mtmd/models/pixtral.cpp

@@ -63,8 +63,8 @@ ggml_cgraph * clip_graph_pixtral::build() {
         // and then concatenate the [IMG_BREAK] token to the end of each row, aka n_patches_per_row dimension
         // after the concatenation, we have a tensor with shape [n_embd, n_patches_per_row + 1, n_rows]
 
-        const int p_y             = n_merge > 0 ? n_patches_y / n_merge : n_patches_y;
-        const int p_x             = n_merge > 0 ? n_patches_x / n_merge : n_patches_x;
+        const int p_y             = n_patches_y / n_merge;
+        const int p_x             = n_patches_x / n_merge;
         const int p_total         = p_x * p_y;
         const int n_embd_text     = cur->ne[0];
         const int n_tokens_output = p_total + p_y - 1; // one [IMG_BREAK] per row, except the last row

tools/mtmd/mtmd-image.cpp

@@ -628,7 +628,7 @@ mtmd_image_preproc_out mtmd_image_preprocessor_llava_uhd::preprocess(const clip_
 mtmd_image_preprocessor_llava_uhd::slice_instructions mtmd_image_preprocessor_llava_uhd::get_slice_instructions(const clip_image_size & original_size) {
     mtmd_image_preprocessor_llava_uhd::slice_instructions res;
     // align slices by patch_size * n_merge so an integer number of merger output tokens fits per slice
-    const int n_merge         = hparams.n_merge > 0 ? hparams.n_merge : 1;
+    const int n_merge         = hparams.n_merge;
     const int patch_size      = hparams.patch_size * n_merge;
     const int slice_size      = hparams.image_size;
     const int original_width  = original_size.width;
@@ -894,7 +894,7 @@ mtmd_image_preproc_out mtmd_image_preprocessor_dyn_size::preprocess(const clip_i
     clip_image_u8 resized_image;
     const clip_image_size original_size = img.get_size();
     // the original pixtral model doesn't have n_merge
-    const int cur_merge = hparams.n_merge == 0 ? 1 : hparams.n_merge;
+    const int cur_merge = hparams.n_merge;
     const clip_image_size target_size = img_tool::calc_size_preserved_ratio(
         original_size,
         hparams.patch_size * cur_merge,

tools/parser/debug-template-parser.cpp

@@ -40,6 +40,7 @@ struct debug_options {
     bool               enable_reasoning  = true;
     bool               debug_jinja       = false;
     bool               force_tool_call   = false;
+    bool               parallel_tool_calls = true;
     output_mode        mode              = output_mode::BOTH;
     input_message_type input_message     = input_message_type::NONE;
 };
@@ -87,6 +88,7 @@ static void print_usage(const char * program_name) {
     LOG_ERR("\nOptions:\n");
     LOG_ERR("  --no-tools              Disable tool definitions\n");
     LOG_ERR("  --force-tool-call       Set tool calls to forced\n");
+    LOG_ERR("  --parallel-tool-calls=0|1 Set parallel_tool_calls (default: 1)\n");
     LOG_ERR("  --generation-prompt=0|1 Set add_generation_prompt (default: 1)\n");
     LOG_ERR("  --enable-reasoning=0|1  Enable reasoning parsing (default: 1)\n");
     LOG_ERR("  --output=MODE           Output mode: analysis, template, both (default: both)\n");
@@ -121,6 +123,8 @@ static bool parse_options(int argc, char ** argv, debug_options & opts) {
             opts.debug_jinja = true;
         } else if (arg == "--no-tools") {
             opts.with_tools = false;
+        } else if (arg.rfind("--parallel-tool-calls=", 0) == 0) {
+            opts.parallel_tool_calls = parse_bool_option(arg.substr(22));
         } else if (arg.rfind("--generation-prompt=", 0) == 0) {
             opts.generation_prompt = parse_bool_option(arg.substr(20));
         } else if (arg.rfind("--enable-reasoning=", 0) == 0) {
@@ -349,7 +353,7 @@ static autoparser::generation_params prepare_params(const debug_options & opts,
         params.tools       = json();
         params.tool_choice = COMMON_CHAT_TOOL_CHOICE_NONE;
     }
-    params.parallel_tool_calls = false;
+    params.parallel_tool_calls = opts.parallel_tool_calls;
     return params;
 }
 

tools/rpc/CMakeLists.txt

@@ -1,4 +1,4 @@
-set(TARGET rpc-server)
+set(TARGET ggml-rpc-server)
 add_executable(${TARGET} rpc-server.cpp)
 target_link_libraries(${TARGET} PRIVATE ggml)
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

tools/rpc/README.md

@@ -4,8 +4,8 @@
 > This example and the RPC backend are currently in a proof-of-concept development stage. As such, the functionality is fragile and
 > insecure. **Never run the RPC server on an open network or in a sensitive environment!**
 
-The `rpc-server` allows exposing `ggml` devices on a remote host.
-The RPC backend communicates with one or several instances of `rpc-server` and offloads computations to them.
+The `ggml-rpc-server` allows exposing `ggml` devices on a remote host.
+The RPC backend communicates with one or several instances of `ggml-rpc-server` and offloads computations to them.
 This can be used for distributed LLM inference with `llama.cpp` in the following way:
 
 ```mermaid
@@ -14,15 +14,15 @@ flowchart TD
     rpcb<-->|TCP|srvb
     rpcb<-.->|TCP|srvn
     subgraph hostn[Host N]
-    srvn[rpc-server]<-.->dev4["CUDA0"]
-    srvn[rpc-server]<-.->dev5["CPU"]
+    srvn[ggml-rpc-server]<-.->dev4["CUDA0"]
+    srvn[ggml-rpc-server]<-.->dev5["CPU"]
     end
     subgraph hostb[Host B]
-    srvb[rpc-server]<-->dev3["Metal"]
+    srvb[ggml-rpc-server]<-->dev3["Metal"]
     end
     subgraph hosta[Host A]
-    srva[rpc-server]<-->dev["CUDA0"]
-    srva[rpc-server]<-->dev2["CUDA1"]
+    srva[ggml-rpc-server]<-->dev["CUDA0"]
+    srva[ggml-rpc-server]<-->dev2["CUDA1"]
     end
     subgraph host[Main Host]
     local["Local devices"]<-->ggml[llama-cli]
@@ -33,7 +33,7 @@ flowchart TD
     class local,dev,dev2,dev3,dev4,dev5 devcls
 ```
 
-By default, `rpc-server` exposes all available accelerator devices on the host.
+By default, `ggml-rpc-server` exposes all available accelerator devices on the host.
 If there are no accelerators, it exposes a single `CPU` device.
 
 ## Usage
@@ -41,7 +41,7 @@ If there are no accelerators, it exposes a single `CPU` device.
 ### Remote hosts
 
 On each remote host, build the backends for each accelerator by adding `-DGGML_RPC=ON` to the build options.
-For example, to build the `rpc-server` with support for CUDA accelerators:
+For example, to build the `ggml-rpc-server` with support for CUDA accelerators:
 
 ```bash
 mkdir build-rpc-cuda
@@ -50,10 +50,10 @@ cmake .. -DGGML_CUDA=ON -DGGML_RPC=ON
 cmake --build . --config Release
 ```
 
-When started, the `rpc-server` will detect and expose all available `CUDA` devices:
+When started, the `ggml-rpc-server` will detect and expose all available `CUDA` devices:
 
 ```bash
-$ bin/rpc-server
+$ bin/ggml-rpc-server
 ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
 ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
 ggml_cuda_init: found 1 CUDA devices:
@@ -67,14 +67,14 @@ Devices:
 
 You can control the set of exposed CUDA devices with the `CUDA_VISIBLE_DEVICES` environment variable or the `--device` command line option. The following two commands have the same effect:
 ```bash
-$ CUDA_VISIBLE_DEVICES=0 bin/rpc-server -p 50052
-$ bin/rpc-server --device CUDA0 -p 50052
+$ CUDA_VISIBLE_DEVICES=0 bin/ggml-rpc-server -p 50052
+$ bin/ggml-rpc-server --device CUDA0 -p 50052
 ```
 
 ### Main host
 
 On the main host build `llama.cpp` with the backends for the local devices and add `-DGGML_RPC=ON` to the build options.
-Finally, when running `llama-cli` or `llama-server`, use the `--rpc` option to specify the host and port of each `rpc-server`:
+Finally, when running `llama-cli` or `llama-server`, use the `--rpc` option to specify the host and port of each `ggml-rpc-server`:
 
 ```bash
 $ llama-cli -hf ggml-org/gemma-3-1b-it-GGUF -ngl 99 --rpc 192.168.88.10:50052,192.168.88.11:50052
@@ -90,7 +90,7 @@ This can speed up model loading significantly, especially when using large model
 To enable the cache, use the `-c` option:
 
 ```bash
-$ bin/rpc-server -c
+$ bin/ggml-rpc-server -c
 ```
 
 By default, the cache is stored in the `$HOME/.cache/llama.cpp/rpc` directory and can be controlled via the `LLAMA_CACHE` environment variable.
@@ -103,8 +103,8 @@ RDMA is enabled by default when `libibverbs` is found at build time.
 
 ### Troubleshooting
 
-Use the `GGML_RPC_DEBUG` environment variable to enable debug messages from `rpc-server`:
+Use the `GGML_RPC_DEBUG` environment variable to enable debug messages from `ggml-rpc-server`:
 ```bash
-$ GGML_RPC_DEBUG=1 bin/rpc-server
+$ GGML_RPC_DEBUG=1 bin/ggml-rpc-server
 ```
 

tools/server/CMakeLists.txt

@@ -15,6 +15,8 @@ add_library(${TARGET} STATIC
     server-common.h
     server-context.cpp
     server-context.h
+    server-stream.cpp
+    server-stream.h
     server-tools.cpp
     server-tools.h
     server-schema.cpp

tools/server/README-dev.md

@@ -57,6 +57,7 @@ The core architecture consists of the following components:
 - `server_tokens`: Unified representation of token sequences (supports both text and multimodal tokens); used by `server_task` and `server_slot`.
 - `server_prompt_checkpoint`: For recurrent (e.g., RWKV) and SWA models, stores snapshots of KV cache state. Enables reuse when subsequent requests share the same prompt prefix, saving redundant computation.
 - `server_models`: Standalone component for managing multiple backend instances (used in router mode). It is completely independent of `server_context`.
+- `stream_session_manager`: Process wide owner of resumable SSE stream sessions (`g_stream_sessions`), keyed by conversation id. Backs the replay buffer that lets a client reattach to a generation after an HTTP disconnect. See the "Resumable streaming" section below.
 
 ```mermaid
 graph TD
@@ -117,6 +118,58 @@ Here is an example trace of an API request for text completion:
 - As the response is stateless, `server_res_generator` calls `response->update()` to update the response with the current state.
 - `server_res_generator` then calls `response->to_json()` and passes the response to the HTTP layer.
 
+### Resumable streaming (SSE replay buffer)
+
+By default a streaming generation is bound to its HTTP socket: when the socket drops (refresh, tab close, mobile background, transient network) the generation aborts and the live stream is lost. This feature keeps the generation running server side and lets a client reattach.
+
+It is opt in via the `X-Conversation-Id` header on `POST /v1/chat/completions`. Without the header the OAI strict path is unchanged. The conversation id is the only identity end to end (server map key, client localStorage key, route path), with an optional `::model` suffix for direct routing in router mode.
+
+The feature lives entirely in `server-stream.{h,cpp}` and rests on three types:
+
+- `stream_session`: a bounded ring buffer (4 MiB cap, oldest bytes drop first) plus a condvar. `append` pushes raw SSE bytes, `read_from` drains from any offset and blocks for live bytes or finalize, `finalize` wakes readers, `cancel` stops the producer. One conv maps to at most one live session.
+- `stream_session_manager` (`g_stream_sessions`): owns all sessions keyed by conv id, enforces the one conv one session invariant via `create_or_replace`, and runs a GC thread that drops completed sessions past their TTL.
+- `stream_pipe_producer` / `stream_pipe_consumer`: the write and read ends. The producer owns the session lifetime and finalizes it on destruction; the consumer is read only and never finalizes, so a reader detaching cannot kill a running generation.
+
+Producer side: `server_res_generator` attaches a producer pipe when the header is present. The HTTP content provider mirrors every chunk into the ring before writing it to the socket. While a pipe is attached, `stream_aware_should_stop` ignores peer disconnect, so a dropped socket does not stop generation: only an explicit `DELETE` does. When the peer leaves early, `on_complete` calls `close()`, which drains the rest of the generation into the ring on the http worker.
+
+Lifetime safety: the producer pipe holds a shared `alive` flag also captured by the session cancel hook. `~server_res_generator` calls `cleanup()` to clear that hook while the reader is still alive, so a `cancel` arriving during teardown can never call `stop()` on a freed response. This ordering is the most fragile part of the feature: finalizing or destroying the producer before `cleanup()` runs reintroduces a use after free.
+
+Consumer side: `GET /v1/stream/<conv_id>?from=N` opens a `text/event-stream` that replays buffered bytes from offset `N` and blocks for live bytes, so the browser reattaches like a fresh EventSource. An offset below the dropped prefix returns 400.
+
+Routes:
+
+- `GET /v1/stream/:conv_id?from=N`: replay or live reattach.
+- `POST /v1/streams/lookup` with `{"conversation_ids": [...]}`: returns session status only for ids the caller already owns. There is no listing route, so live sessions cannot be enumerated (an earlier `GET /v1/streams` was removed for exactly this reason).
+- `DELETE /v1/stream/:conv_id`: explicit Stop, idempotent (`evict_and_cancel`).
+
+Router mode binds the same paths to proxy handlers. A `conv_id -> child` map (`conv_models`), populated when a POST is routed, resolves the owning child in one lookup with no polling. The lookup groups ids per child; GET and DELETE proxy straight to the owner. This loopback REST hop is expected to move to a websocket IPC later, swapping only the transport.
+
+Lifecycle: `g_stream_sessions.start_gc()` runs in main after common init, `stop_gc()` runs first in `clean_up()` and finalizes every live session so no reader hangs. Reader blocking and the post drop drain both run on httplib worker threads, which block on a condvar rather than spin.
+
+| Constant | Value | Role |
+| --- | --- | --- |
+| `STREAM_SESSION_TTL_SECONDS` | 300 | retention of a completed session before GC |
+| `STREAM_SESSION_MAX_BYTES` | 4 MiB | ring cap per session |
+| `STREAM_SESSION_GC_INTERVAL_SECONDS` | 60 | GC tick |
+| `STREAM_READ_WAKE_INTERVAL_MS` | 200 | read_from wake to recheck should_stop |
+| `STREAM_LOOKUP_TIMEOUT_MS` | 250 | router to child loopback budget |
+
+```mermaid
+graph TD
+    Client -- "POST + X-Conversation-Id" --> RG[server_res_generator]
+    RG -- attach --> Prod[stream_pipe_producer]
+    Prod -- "write, drain on peer drop" --> Sess
+    subgraph g_stream_sessions
+        Sess[stream_session: ring buffer, 4 MiB]
+        GC[GC thread] -- drop after TTL --> Sess
+    end
+    Sess -- read_from offset --> Cons[stream_pipe_consumer]
+    Cons -- "GET /v1/stream/:id?from=N" --> Client
+    DEL[DELETE /v1/stream/:id] -- evict_and_cancel --> Sess
+```
+
+The diagram shows the buffer touch points. The live wire (chunks streamed to the original client during a normal generation) is the producer's default output, described under "Producer side" above.
+
 ### Testing
 
 `llama-server` includes an automated test suite based on `pytest`.
@@ -223,6 +276,7 @@ The flow for downloading a new model:
 - Speculative decoding: https://github.com/ggml-org/llama.cpp/pull/17808 and rework in https://github.com/ggml-org/llama.cpp/pull/17808
 - INI presets: https://github.com/ggml-org/llama.cpp/pull/17859 (+ refactoring: https://github.com/ggml-org/llama.cpp/pull/18169)
 - Sleeping mode: https://github.com/ggml-org/llama.cpp/pull/18228
+- Resumable streaming (SSE replay buffer): https://github.com/ggml-org/llama.cpp/pull/23226
 
 
 

tools/server/server-context.cpp

@@ -5,6 +5,7 @@
 #include "server-task.h"
 #include "server-queue.h"
 #include "server-schema.h"
+#include "server-stream.h"
 
 #include "build-info.h"
 #include "common.h"
@@ -4022,6 +4023,15 @@ struct server_res_generator : server_http_res {
             queue_tasks.wait_until_no_sleep();
         }
     }
+    ~server_res_generator() override {
+        // cleanup() must run while rd is still alive (rd is destroyed after this body returns)
+        if (spipe) {
+            spipe->cleanup();
+        }
+    }
+    void stop() override {
+        rd.stop();
+    }
     void ok(const json & response_data) {
         status = 200;
         data = safe_json_to_str(response_data);
@@ -4210,8 +4220,10 @@ std::unique_ptr<server_res_generator> server_routes::handle_completions_impl(
                 }
             };
 
+            auto effective_should_stop = stream_aware_should_stop(res_this, req.should_stop);
+
             try {
-                if (req.should_stop()) {
+                if (effective_should_stop()) {
                     SRV_DBG("%s", "stopping streaming due to should_stop condition\n");
                     return false; // should_stop condition met
                 }
@@ -4245,8 +4257,8 @@ std::unique_ptr<server_res_generator> server_routes::handle_completions_impl(
                 // receive subsequent results
                 bool timeout = false;
                 int64_t start_time = ggml_time_ms();
-                auto result = rd.next([&timeout, &req, &start_time, &params]() {
-                    if (req.should_stop()) {
+                auto result = rd.next([&timeout, &start_time, &params, &effective_should_stop]() {
+                    if (effective_should_stop()) {
                         return true; // should_stop condition met
                     } else if (params.sse_ping_interval > 0 && ggml_time_ms() - start_time > (int64_t)params.sse_ping_interval * 1000) {
                         timeout = true;
@@ -4264,7 +4276,7 @@ std::unique_ptr<server_res_generator> server_routes::handle_completions_impl(
 
                 if (result == nullptr) {
                     SRV_DBG("%s", "stopping streaming due to should_stop condition\n");
-                    GGML_ASSERT(req.should_stop());
+                    GGML_ASSERT(effective_should_stop());
                     return false; // should_stop condition met
                 }
 
@@ -4302,6 +4314,10 @@ std::unique_ptr<server_res_generator> server_routes::handle_completions_impl(
         };
     }
 
+    // attach a producer pipe to the response when X-Conversation-Id is present.
+    // the pipe mirrors SSE chunks into the ring buffer and wires up the cancel hook.
+    stream_session_attach_pipe(*res, req.headers);
+
     return res;
 }
 

tools/server/server-http.cpp

@@ -1,5 +1,6 @@
 #include "common.h"
 #include "server-http.h"
+#include "server-stream.h"
 #include "server-common.h"
 #include "ui.h"
 
@@ -456,13 +457,40 @@ static void set_headers(httplib::Response & res, const std::map<std::string, std
     }
 }
 
+// percent-decode a path component (%XX). path params arrive raw from httplib, unlike query
+// params, so a conv id like "conv::model" sent as "conv%3A%3Amodel" must be decoded here to
+// match the value the client put in the X-Conversation-Id header
+static std::string decode_path_component(const std::string & in) {
+    std::string out;
+    out.reserve(in.size());
+    for (size_t i = 0; i < in.size(); i++) {
+        if (in[i] == '%' && i + 2 < in.size()) {
+            auto hex = [](char c) -> int {
+                if (c >= '0' && c <= '9') return c - '0';
+                if (c >= 'a' && c <= 'f') return c - 'a' + 10;
+                if (c >= 'A' && c <= 'F') return c - 'A' + 10;
+                return -1;
+            };
+            int hi = hex(in[i + 1]);
+            int lo = hex(in[i + 2]);
+            if (hi >= 0 && lo >= 0) {
+                out.push_back(char((hi << 4) | lo));
+                i += 2;
+                continue;
+            }
+        }
+        out.push_back(in[i]);
+    }
+    return out;
+}
+
 static std::map<std::string, std::string> get_params(const httplib::Request & req) {
     std::map<std::string, std::string> params;
     for (const auto & [key, value] : req.params) {
         params[key] = value;
     }
     for (const auto & [key, value] : req.path_params) {
-        params[key] = value;
+        params[key] = decode_path_component(value);
     }
     return params;
 }
@@ -497,26 +525,41 @@ static void process_handler_response(server_http_req_ptr && request, server_http
         set_headers(res, response->headers);
         const std::string content_type = response->content_type;
         // convert to shared_ptr as both chunked_content_provider() and on_complete() need to use it
-        std::shared_ptr q_ptr = std::move(request);
-        std::shared_ptr r_ptr = std::move(response);
-        const auto chunked_content_provider = [response = r_ptr](size_t, const httplib::DataSink & sink) -> bool {
+        std::shared_ptr<server_http_req> q_ptr = std::move(request);
+        std::shared_ptr<server_http_res> r_ptr = std::move(response);
+
+        const auto chunked_content_provider = [response = r_ptr](size_t, httplib::DataSink & sink) -> bool {
             std::string chunk;
             const bool has_next = response->next(chunk);
             if (!chunk.empty()) {
+                // mirror into the ring buffer first, the session must reflect every SSE chunk
+                // whether or not the wire write below succeeds
+                if (response->spipe) {
+                    response->spipe->write(chunk.data(), chunk.size());
+                }
                 if (!sink.write(chunk.data(), chunk.size())) {
+                    // peer is gone, stop the wire path here
                     return false;
                 }
                 SRV_DBG("http: streamed chunk: %s\n", chunk.c_str());
             }
             if (!has_next) {
+                // producer reached its natural end on the wire, a later close() skips the drain
+                if (response->spipe) {
+                    response->spipe->done();
+                }
                 sink.done();
                 SRV_DBG("%s", "http: stream ended\n");
             }
             return has_next;
         };
         const auto on_complete = [request = q_ptr, response = r_ptr](bool) mutable {
-            response.reset(); // trigger the destruction of the response object
-            request.reset();  // trigger the destruction of the request object
+            // on a dropped peer, close() drains the rest of the generation into the ring buffer
+            if (response->spipe) {
+                response->spipe->close();
+            }
+            response.reset(); // spipe destructor finalizes the session if attached
+            request.reset();
         };
         res.set_chunked_content_provider(content_type, chunked_content_provider, on_complete);
     } else {

tools/server/server-http.h

@@ -3,6 +3,7 @@
 #include <atomic>
 #include <functional>
 #include <map>
+#include <memory>
 #include <string>
 #include <thread>
 #include <vector>
@@ -10,6 +11,7 @@
 #include <unordered_map>
 
 struct common_params;
+struct stream_pipe_producer; // defined in server-stream.h
 
 // generator-like API for HTTP response generation
 // this object response with one of the 2 modes:
@@ -23,12 +25,20 @@ struct server_http_res {
     std::string data;
     std::map<std::string, std::string> headers;
 
-    // TODO: move this to a virtual function once we have proper polymorphism support
+    // if set, the stream survives a client disconnect: the producer pipe keeps draining into the
+    // ring buffer and finalizes the session on destruction, so no explicit on_stream_end is needed.
+    // shared_ptr (not unique_ptr) so the forward-declared type is safe to delete here.
+    std::shared_ptr<stream_pipe_producer> spipe;
+
     std::function<bool(std::string &)> next = nullptr;
     bool is_stream() const {
         return next != nullptr;
     }
 
+    // called when the session is cancelled (e.g. DELETE /v1/stream/<conv_id>).
+    // server_res_generator overrides this to stop its reader; the default is a no-op.
+    virtual void stop() {}
+
     virtual ~server_http_res() = default;
 };
 

tools/server/server-models.cpp

@@ -1,12 +1,14 @@
 #include "server-common.h"
 #include "server-models.h"
 #include "server-context.h"
+#include "server-stream.h"
 
 #include "build-info.h"
 #include "preset.h"
 #include "download.h"
 
 #include <cpp-httplib/httplib.h> // TODO: remove this once we use HTTP client from download.h
+#include <optional>
 #include <sheredom/subprocess.h>
 
 #include <functional>
@@ -92,6 +94,9 @@ struct server_subproc {
     }
 };
 
+// short loopback budget for the resumable stream router to child JSON calls (probe, lookup,
+// delete). distinct from params.timeout_read/write which only applies to the generation proxy
+static constexpr int STREAM_LOOKUP_TIMEOUT_MS = 250;
 
 static std::filesystem::path get_server_exec_path() {
 #if defined(_WIN32)
@@ -223,8 +228,8 @@ void server_model_meta::update_caps() {
             "LLAMA_ARG_HF_REPO_FILE",
         });
         params.offline = true;
-        // params.skip_download = true; // TODO: ideally, we should validate the model here, but it takes too much time
-        common_params_handle_models(params, LLAMA_EXAMPLE_SERVER, {});
+        common_models_handler handler = common_models_handler_init(params, LLAMA_EXAMPLE_SERVER);
+        common_models_handler_apply(handler, params); // note: this won't download the model because offline=true
         if (params.mmproj.path.empty()) {
             multimodal = { false, false };
         } else {
@@ -1393,9 +1398,8 @@ struct server_download_state : public common_download_callback {
 
     bool run(common_params & params) {
         try {
-            common_params_handle_models_params p;
-            p.callback = this;
-            common_params_handle_models(params, LLAMA_EXAMPLE_SERVER, p);
+            common_models_handler handler = common_models_handler_init(params, LLAMA_EXAMPLE_SERVER);
+            common_models_handler_apply(handler, params, this);
             is_ok = true;
         } catch (const std::exception & e) {
             auto model_name = params.model.get_name();
@@ -1581,6 +1585,45 @@ static bool is_autoload(const common_params & params, const server_http_req & re
     }
 }
 
+// percent encode one query or path component, covers reserved chars without pulling in
+// httplib::detail. used by the stream routes to forward conversation_id to children safely
+static std::string encode_qs(const std::string & in) {
+    std::string out;
+    out.reserve(in.size() * 3);
+    for (unsigned char c : in) {
+        bool safe = (c >= 'A' && c <= 'Z') || (c >= 'a' && c <= 'z') || (c >= '0' && c <= '9')
+                 || c == '-' || c == '_' || c == '.' || c == '~';
+        if (safe) {
+            out.push_back(char(c));
+        } else {
+            char buf[4];
+            std::snprintf(buf, sizeof(buf), "%%%02X", c);
+            out.append(buf, 3);
+        }
+    }
+    return out;
+}
+
+// resolve the child that owns a conversation's stream session via the conv_id -> model map
+// populated when the POST was routed. single map lookup then a meta lookup, no polling, no
+// parsing of the conv id. returns nullopt when nothing maps, the caller answers not found and
+// the client recovers
+static std::optional<server_model_meta> resolve_child_for_conv(
+        server_models & models, const std::string & conversation_id) {
+    if (conversation_id.empty()) {
+        return std::nullopt;
+    }
+    auto tracked = models.conv_models.lookup(conversation_id);
+    if (!tracked.has_value()) {
+        return std::nullopt;
+    }
+    auto meta = models.get_meta(*tracked);
+    if (meta.has_value() && meta->is_ready()) {
+        return meta;
+    }
+    return std::nullopt;
+}
+
 void server_models_routes::init_routes() {
     this->get_router_props = [this](const server_http_req & req) {
         std::string name = req.get_param("model");
@@ -1629,6 +1672,12 @@ void server_models_routes::init_routes() {
         if (!router_validate_model(name, models, autoload, error_res)) {
             return error_res;
         }
+        // remember which child serves this conversation so the stream routes can route straight
+        // to it without polling, keyed on the exact conv id from the header
+        std::string conv_id = stream_conv_id_from_headers(req.headers);
+        if (!conv_id.empty()) {
+            models.conv_models.remember(conv_id, name);
+        }
         return models.proxy_request(req, method, name, true); // update last usage for POST request only
     };
 
@@ -1768,23 +1817,14 @@ void server_models_routes::init_routes() {
             throw std::invalid_argument("model must be a non-empty string");
         }
 
-        common_params_model model;
-        common_download_opts opts;
+        common_params p;
+        p.model.hf_repo  = name;
+        p.hf_token       = params.hf_token;
 
-        model.hf_repo        = name;
-        opts.bearer_token    = params.hf_token;
-        // note: we only check main model, no need sidecar here
-        opts.download_mmproj = false;
-        opts.download_mtp    = false;
-
-        // first, only check if the model is valid and can be downloaded
-        opts.skip_download = true;
+        // validate by fetching metadata
         bool ok = false;
         try {
-            auto validation = common_download_model(model, opts);
-            ok = !validation.model_path.empty();
-        } catch (const common_skip_download_exception &) {
-            // model is valid and will be downloaded
+            common_models_handler_init(p, LLAMA_EXAMPLE_SERVER);
             ok = true;
         } catch (...) {
             SRV_ERR("unknown error while validating model '%s'\n", name.c_str());
@@ -1829,6 +1869,128 @@ void server_models_routes::init_routes() {
         res_ok(res, {{"success", true}});
         return res;
     };
+
+    this->router_stream_get = [this](const server_http_req & req) {
+        // GET /v1/stream/<conv_id>?from=N. resolve the owning child from the conv_id -> model
+        // map, 404 when nothing maps
+        auto res = std::make_unique<server_http_res>();
+        std::string conv_id = req.get_param("conv_id");
+        if (conv_id.empty()) {
+            res_err(res, format_error_response("Missing conversation id in path", ERROR_TYPE_INVALID_REQUEST));
+            return res;
+        }
+        std::optional<server_model_meta> owner = resolve_child_for_conv(models, conv_id);
+        if (!owner.has_value()) {
+            res_err(res, format_error_response("Stream not found or expired", ERROR_TYPE_NOT_FOUND));
+            return res;
+        }
+        std::string from = req.get_param("from");
+        std::string child_path = "/v1/stream/" + encode_qs(conv_id);
+        if (!from.empty()) {
+            child_path += "?from=" + from;
+        }
+        SRV_INF("proxying stream resume to model %s on port %d, path=%s\n",
+                owner->name.c_str(), owner->port, child_path.c_str());
+        auto proxy = std::make_unique<server_http_proxy>(
+                "GET",
+                "http",
+                CHILD_ADDR,
+                owner->port,
+                child_path,
+                req.headers,
+                req.body,
+                req.files,
+                req.should_stop,
+                params.timeout_read,
+                params.timeout_write);
+        return std::unique_ptr<server_http_res>(std::move(proxy));
+    };
+
+    this->router_streams_lookup = [this](const server_http_req & req) {
+        // POST /v1/streams/lookup. resolve each requested conv id to its owning child via the
+        // map, group the ids per child, and query only the children that actually own some of
+        // them instead of fanning out to every ready child. a child only answers for the ids
+        // it owns, never lists anything else
+        auto res = std::make_unique<server_http_res>();
+        std::vector<std::string> requested;
+        try {
+            json body = json::parse(req.body);
+            if (body.contains("conversation_ids") && body["conversation_ids"].is_array()) {
+                for (const auto & v : body["conversation_ids"]) {
+                    if (v.is_string() && !v.get<std::string>().empty()) {
+                        requested.push_back(v.get<std::string>());
+                    }
+                }
+            }
+        } catch (const std::exception &) {
+            res_ok(res, json::array());
+            return res;
+        }
+
+        // group requested ids by the child port that owns them, drop ids that map to nothing
+        std::unordered_map<int, json> per_child;
+        for (const auto & cid : requested) {
+            auto owner = resolve_child_for_conv(models, cid);
+            if (!owner.has_value()) {
+                continue;
+            }
+            per_child[owner->port].push_back(cid);
+        }
+
+        json aggregated = json::array();
+        for (auto & [port, ids] : per_child) {
+            json child_body = {{"conversation_ids", ids}};
+            httplib::Client cli(CHILD_ADDR, port);
+            cli.set_connection_timeout(0, STREAM_LOOKUP_TIMEOUT_MS * 1000);
+            cli.set_read_timeout(0, STREAM_LOOKUP_TIMEOUT_MS * 1000);
+            cli.set_write_timeout(0, STREAM_LOOKUP_TIMEOUT_MS * 1000);
+            auto resp = cli.Post("/v1/streams/lookup", child_body.dump(), "application/json");
+            if (!resp || resp->status != 200) {
+                continue;
+            }
+            try {
+                json child_arr = json::parse(resp->body);
+                if (!child_arr.is_array()) {
+                    continue;
+                }
+                for (auto & entry : child_arr) {
+                    if (entry.is_object()) {
+                        aggregated.push_back(entry);
+                    }
+                }
+            } catch (const std::exception &) {
+                continue;
+            }
+        }
+        res_ok(res, aggregated);
+        return res;
+    };
+
+    this->router_stream_delete = [this](const server_http_req & req) {
+        // DELETE /v1/stream/<conv_id>. resolve the owning child via the map and forward only to
+        // it, evict_and_cancel is idempotent on the child
+        auto res = std::make_unique<server_http_res>();
+        std::string conv_id = req.get_param("conv_id");
+        if (conv_id.empty()) {
+            res_err(res, format_error_response("Missing conversation id in path", ERROR_TYPE_INVALID_REQUEST));
+            return res;
+        }
+        std::string child_path = "/v1/stream/" + encode_qs(conv_id);
+        auto owner = resolve_child_for_conv(models, conv_id);
+        if (owner.has_value()) {
+            httplib::Client cli(CHILD_ADDR, owner->port);
+            cli.set_connection_timeout(0, STREAM_LOOKUP_TIMEOUT_MS * 1000);
+            cli.set_read_timeout(0, STREAM_LOOKUP_TIMEOUT_MS * 1000);
+            cli.set_write_timeout(0, STREAM_LOOKUP_TIMEOUT_MS * 1000);
+            auto resp = cli.Delete(child_path.c_str());
+            (void) resp; // best effort, 404 and network errors are equivalent to no op
+        }
+        // drop the tracking entry, the session is being torn down
+        models.conv_models.forget(conv_id);
+        res->status = 204;
+        res->content_type = "application/json";
+        return res;
+    };
 }