vendor : update BoringSSL to 0.20260327.0 (#21211)

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

.devops/cpu.Dockerfile

@@ -36,7 +36,7 @@ RUN mkdir -p /app/full \
 FROM ubuntu:$UBUNTU_VERSION AS base
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 curl\
+    && apt-get install -y libgomp1 curl \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.devops/cuda-new.Dockerfile

@@ -1,6 +1,6 @@
 ARG UBUNTU_VERSION=24.04
 # This needs to generally match the container host's environment.
-ARG CUDA_VERSION=13.1.0
+ARG CUDA_VERSION=13.1.1
 # Target the CUDA build image
 ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
 
@@ -12,7 +12,9 @@ FROM ${BASE_CUDA_DEV_CONTAINER} AS build
 ARG CUDA_DOCKER_ARCH=default
 
 RUN apt-get update && \
-    apt-get install -y build-essential cmake python3 python3-pip git libssl-dev libgomp1
+    apt-get install -y gcc-14 g++-14 build-essential cmake python3 python3-pip git libssl-dev libgomp1
+
+ENV CC=gcc-14 CXX=g++-14 CUDAHOSTCXX=g++-14
 
 WORKDIR /app
 
@@ -39,7 +41,7 @@ RUN mkdir -p /app/full \
 FROM ${BASE_CUDA_RUN_CONTAINER} AS base
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 curl\
+    && apt-get install -y libgomp1 curl \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.devops/cuda.Dockerfile

@@ -1,6 +1,6 @@
-ARG UBUNTU_VERSION=22.04
+ARG UBUNTU_VERSION=24.04
 # This needs to generally match the container host's environment.
-ARG CUDA_VERSION=12.4.0
+ARG CUDA_VERSION=12.8.1
 # Target the CUDA build image
 ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
 
@@ -12,7 +12,9 @@ FROM ${BASE_CUDA_DEV_CONTAINER} AS build
 ARG CUDA_DOCKER_ARCH=default
 
 RUN apt-get update && \
-    apt-get install -y build-essential cmake python3 python3-pip git libssl-dev libgomp1
+    apt-get install -y gcc-14 g++-14 build-essential cmake python3 python3-pip git libssl-dev libgomp1
+
+ENV CC=gcc-14 CXX=g++-14 CUDAHOSTCXX=g++-14
 
 WORKDIR /app
 
@@ -39,7 +41,7 @@ RUN mkdir -p /app/full \
 FROM ${BASE_CUDA_RUN_CONTAINER} AS base
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 curl\
+    && apt-get install -y libgomp1 curl \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \
@@ -60,7 +62,8 @@ RUN apt-get update \
     git \
     python3 \
     python3-pip \
-    && pip install --upgrade pip setuptools wheel \
+    python3-wheel \
+    && pip install --break-system-packages --upgrade setuptools \
     && pip install --break-system-packages -r requirements.txt \
     && apt autoremove -y \
     && apt clean -y \

.devops/intel.Dockerfile

@@ -33,8 +33,25 @@ RUN mkdir -p /app/full \
 
 FROM intel/deep-learning-essentials:$ONEAPI_VERSION AS base
 
+ARG IGC_VERSION=v2.30.1
+ARG IGC_VERSION_FULL=2_2.30.1+20950
+ARG COMPUTE_RUNTIME_VERSION=26.09.37435.1
+ARG COMPUTE_RUNTIME_VERSION_FULL=26.09.37435.1-0
+ARG IGDGMM_VERSION=22.9.0
+RUN mkdir /tmp/neo/ && cd /tmp/neo/ \
+  && wget https://github.com/intel/intel-graphics-compiler/releases/download/$IGC_VERSION/intel-igc-core-${IGC_VERSION_FULL}_amd64.deb \
+  && wget https://github.com/intel/intel-graphics-compiler/releases/download/$IGC_VERSION/intel-igc-opencl-${IGC_VERSION_FULL}_amd64.deb \
+  && wget https://github.com/intel/compute-runtime/releases/download/$COMPUTE_RUNTIME_VERSION/intel-ocloc-dbgsym_${COMPUTE_RUNTIME_VERSION_FULL}_amd64.ddeb \
+  && wget https://github.com/intel/compute-runtime/releases/download/$COMPUTE_RUNTIME_VERSION/intel-ocloc_${COMPUTE_RUNTIME_VERSION_FULL}_amd64.deb \
+  && wget https://github.com/intel/compute-runtime/releases/download/$COMPUTE_RUNTIME_VERSION/intel-opencl-icd-dbgsym_${COMPUTE_RUNTIME_VERSION_FULL}_amd64.ddeb \
+  && wget https://github.com/intel/compute-runtime/releases/download/$COMPUTE_RUNTIME_VERSION/intel-opencl-icd_${COMPUTE_RUNTIME_VERSION_FULL}_amd64.deb \
+  && wget https://github.com/intel/compute-runtime/releases/download/$COMPUTE_RUNTIME_VERSION/libigdgmm12_${IGDGMM_VERSION}_amd64.deb \
+  && wget https://github.com/intel/compute-runtime/releases/download/$COMPUTE_RUNTIME_VERSION/libze-intel-gpu1-dbgsym_${COMPUTE_RUNTIME_VERSION_FULL}_amd64.ddeb \
+  && wget https://github.com/intel/compute-runtime/releases/download/$COMPUTE_RUNTIME_VERSION/libze-intel-gpu1_${COMPUTE_RUNTIME_VERSION_FULL}_amd64.deb \
+  && dpkg --install *.deb
+
 RUN apt-get update \
-    && apt-get install -y libgomp1 curl\
+    && apt-get install -y libgomp1 curl \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.devops/musa.Dockerfile

@@ -46,7 +46,7 @@ RUN mkdir -p /app/full \
 FROM ${BASE_MUSA_RUN_CONTAINER} AS base
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 curl\
+    && apt-get install -y libgomp1 curl \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.devops/openvino.Dockerfile

@@ -78,7 +78,7 @@ ARG http_proxy
 ARG https_proxy
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 libtbb12 curl\
+    && apt-get install -y libgomp1 libtbb12 curl \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.devops/rocm.Dockerfile

@@ -58,7 +58,7 @@ RUN mkdir -p /app/full \
 FROM ${BASE_ROCM_DEV_CONTAINER} AS base
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 curl\
+    && apt-get install -y libgomp1 curl \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \
@@ -79,7 +79,7 @@ RUN apt-get update \
     git \
     python3-pip \
     python3 \
-    python3-wheel\
+    python3-wheel \
     && pip install --break-system-packages --upgrade setuptools \
     && pip install --break-system-packages -r requirements.txt \
     && apt autoremove -y \

.devops/vulkan.Dockerfile

@@ -49,17 +49,20 @@ COPY --from=build /app/full /app
 
 WORKDIR /app
 
+ENV PATH="/root/.venv/bin:/root/.local/bin:${PATH}"
+
+# Flag for compatibility with pip
+ARG UV_INDEX_STRATEGY="unsafe-best-match"
 RUN apt-get update \
     && apt-get install -y \
     build-essential \
+    curl \
     git \
-    python3.13 \
-    python3.13-dev \
-    python3-pip \
-    python3-wheel \
-    && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.13 100 \
-    && pip install --break-system-packages --upgrade setuptools \
-    && pip install --break-system-packages -r requirements.txt \
+    ca-certificates \
+    && curl -LsSf https://astral.sh/uv/install.sh | sh \
+    && uv python install 3.13 \
+    && uv venv --python 3.13 /root/.venv \
+    && uv pip install --python /root/.venv/bin/python -r requirements.txt \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.github/workflows/docker.yml

@@ -25,184 +25,13 @@ permissions:
   packages: write
 
 jobs:
-  push_to_registry:
-    name: Push Docker image to Docker Hub
-
-    runs-on: ${{ matrix.config.runs_on }}
-    env:
-      COMMIT_SHA: ${{ github.sha }}
-    strategy:
-      fail-fast: false
-      matrix:
-        config:
-          # Multi-stage build
-          - { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/arm64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-24.04" }
-          - { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-24.04" }
-          - { tag: "cuda cuda12", dockerfile: ".devops/cuda.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-24.04", cuda_version: "12.4.0", ubuntu_version: "22.04" }
-          - { tag: "cuda13", dockerfile: ".devops/cuda-new.Dockerfile",  platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-24.04", cuda_version: "13.1.0", ubuntu_version: "24.04" }
-          - { tag: "musa",   dockerfile: ".devops/musa.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-24.04" }
-          - { tag: "intel",  dockerfile: ".devops/intel.Dockerfile",  platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-24.04" }
-          - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-24.04" }
-          - { tag: "s390x",  dockerfile: ".devops/s390x.Dockerfile",  platforms: "linux/s390x", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-24.04-s390x" }
-          - { tag: "rocm",   dockerfile: ".devops/rocm.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-24.04" }
-          - { tag: "openvino", dockerfile: ".devops/openvino.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-24.04" }
-    steps:
-      - name: Check out the repo
-        uses: actions/checkout@v6
-        with:
-          fetch-depth: 0 # preserve git history, so we can determine the build number
-
-      - name: Set up QEMU
-        if: ${{ matrix.config.tag != 's390x' }}
-        uses: docker/setup-qemu-action@c7c53464625b32c7a7e944ae62b3e17d2b600130 # v3
-        with:
-          image: tonistiigi/binfmt:qemu-v10.2.1
-
-      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@8d2750c68a42422c14e847fe6c8ac0403b4cbd6f # v3
-
-      - name: Log in to Docker Hub
-        uses: docker/login-action@c94ce9fb468520275223c153574b00df6fe4bcc9 # v3
-        with:
-          registry: ghcr.io
-          username: ${{ github.repository_owner }}
-          password: ${{ secrets.GITHUB_TOKEN }}
-
-      - name: Determine source tag name
-        id: srctag
-        uses: ./.github/actions/get-tag-name
-        env:
-          BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
-
-      - name: Determine image tag name
-        id: tag
-        shell: bash
-        run: |
-          REPO_OWNER="${GITHUB_REPOSITORY_OWNER@L}"  # to lower case
-          REPO_NAME="${{ github.event.repository.name }}"
-          PREFIX="ghcr.io/${REPO_OWNER}/${REPO_NAME}:"
-
-          # list all tags possible
-          tags="${{ matrix.config.tag }}"
-          for tag in $tags; do
-              if [[ "$tag" == "cpu" ]]; then
-                  TYPE=""
-              else
-                  TYPE="-$tag"
-              fi
-              CACHETAGS="${PREFIX}buildcache${TYPE}"
-              FULLTAGS="${FULLTAGS:+$FULLTAGS,}${PREFIX}full${TYPE},${PREFIX}full${TYPE}-${{ steps.srctag.outputs.name }}"
-              LIGHTTAGS="${LIGHTTAGS:+$LIGHTTAGS,}${PREFIX}light${TYPE},${PREFIX}light${TYPE}-${{ steps.srctag.outputs.name }}"
-              SERVERTAGS="${SERVERTAGS:+$SERVERTAGS,}${PREFIX}server${TYPE},${PREFIX}server${TYPE}-${{ steps.srctag.outputs.name }}"
-          done
-          echo "cache_output_tags=$CACHETAGS" >> $GITHUB_OUTPUT
-          echo "full_output_tags=$FULLTAGS" >> $GITHUB_OUTPUT
-          echo "light_output_tags=$LIGHTTAGS" >> $GITHUB_OUTPUT
-          echo "server_output_tags=$SERVERTAGS" >> $GITHUB_OUTPUT
-          echo "cache_output_tags=$CACHETAGS"  # print out for debugging
-          echo "full_output_tags=$FULLTAGS"  # print out for debugging
-          echo "light_output_tags=$LIGHTTAGS"  # print out for debugging
-          echo "server_output_tags=$SERVERTAGS"  # print out for debugging
-        env:
-          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
-
-      - name: Free Disk Space (Ubuntu)
-        if: ${{ matrix.config.free_disk_space == true }}
-        uses: ggml-org/free-disk-space@v1.3.1
-        with:
-          # this might remove tools that are actually needed,
-          # if set to "true" but frees about 6 GB
-          tool-cache: false
-
-          # all of these default to true, but feel free to set to
-          # "false" if necessary for your workflow
-          android: true
-          dotnet: true
-          haskell: true
-          large-packages: true
-          docker-images: true
-          swap-storage: true
-
-      - name: Build and push Full Docker image (tagged + versioned)
-        if: ${{ (github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && matrix.config.full == true }}
-        uses: docker/build-push-action@10e90e3645eae34f1e60eeb005ba3a3d33f178e8 # v6
-        with:
-          context: .
-          push: true
-          platforms: ${{ matrix.config.platforms }}
-          # tag list is generated from step above
-          tags: ${{ steps.tag.outputs.full_output_tags }}
-          file: ${{ matrix.config.dockerfile }}
-          target: full
-          provenance: false
-          build-args: |
-            ${{ matrix.config.ubuntu_version && format('UBUNTU_VERSION={0}', matrix.config.ubuntu_version) || '' }}
-            ${{ matrix.config.cuda_version && format('CUDA_VERSION={0}', matrix.config.cuda_version) || '' }}
-          # using github experimental cache
-          #cache-from: type=gha
-          #cache-to: type=gha,mode=max
-          # return to this if the experimental github cache is having issues
-          #cache-to: type=local,dest=/tmp/.buildx-cache
-          #cache-from: type=local,src=/tmp/.buildx-cache
-          # using registry cache (no storage limit)
-          cache-from: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }}
-          cache-to: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }},mode=max
-
-      - name: Build and push Light Docker image (tagged + versioned)
-        if: ${{ (github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && matrix.config.light == true }}
-        uses: docker/build-push-action@10e90e3645eae34f1e60eeb005ba3a3d33f178e8 # v6
-        with:
-          context: .
-          push: true
-          platforms: ${{ matrix.config.platforms }}
-          # tag list is generated from step above
-          tags: ${{ steps.tag.outputs.light_output_tags }}
-          file: ${{ matrix.config.dockerfile }}
-          target: light
-          provenance: false
-          build-args: |
-            ${{ matrix.config.ubuntu_version && format('UBUNTU_VERSION={0}', matrix.config.ubuntu_version) || '' }}
-            ${{ matrix.config.cuda_version && format('CUDA_VERSION={0}', matrix.config.cuda_version) || '' }}
-          # using github experimental cache
-          #cache-from: type=gha
-          #cache-to: type=gha,mode=max
-          # return to this if the experimental github cache is having issues
-          #cache-to: type=local,dest=/tmp/.buildx-cache
-          #cache-from: type=local,src=/tmp/.buildx-cache
-          # using registry cache (no storage limit)
-          cache-from: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }}
-          cache-to: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }},mode=max
-
-      - name: Build and push Server Docker image (tagged + versioned)
-        if: ${{ (github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && matrix.config.server == true }}
-        uses: docker/build-push-action@10e90e3645eae34f1e60eeb005ba3a3d33f178e8 # v6
-        with:
-          context: .
-          push: true
-          platforms: ${{ matrix.config.platforms }}
-          # tag list is generated from step above
-          tags: ${{ steps.tag.outputs.server_output_tags }}
-          file: ${{ matrix.config.dockerfile }}
-          target: server
-          provenance: false
-          build-args: |
-            ${{ matrix.config.ubuntu_version && format('UBUNTU_VERSION={0}', matrix.config.ubuntu_version) || '' }}
-            ${{ matrix.config.cuda_version && format('CUDA_VERSION={0}', matrix.config.cuda_version) || '' }}
-          # using github experimental cache
-          #cache-from: type=gha
-          #cache-to: type=gha,mode=max
-          # return to this if the experimental github cache is having issues
-          #cache-to: type=local,dest=/tmp/.buildx-cache
-          #cache-from: type=local,src=/tmp/.buildx-cache
-          # using registry cache (no storage limit)
-          cache-from: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }}
-          cache-to: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }},mode=max
-
   create_tag:
     name: Create and push git tag
-    runs-on: ubuntu-22.04
+    runs-on: ubuntu-slim
     permissions:
       contents: write
+    outputs:
+      source_tag: ${{ steps.srctag.outputs.name }}
 
     steps:
       - name: Clone
@@ -223,3 +52,391 @@ jobs:
         run: |
           git tag ${{ steps.srctag.outputs.name }} || exit 0
           git push origin ${{ steps.srctag.outputs.name }} || exit 0
+
+  prepare_matrices:
+    name: Prepare Docker matrices
+    runs-on: ubuntu-24.04
+    outputs:
+      build_matrix: ${{ steps.matrices.outputs.build_matrix }}
+      merge_matrix: ${{ steps.matrices.outputs.merge_matrix }}
+
+    steps:
+      - name: Generate build and merge matrices
+        id: matrices
+        shell: bash
+        run: |
+          set -euo pipefail
+
+          # Keep all build targets in one place and derive merge targets from it.
+          cat > build-matrix.json <<'JSON'
+          [
+            { "tag": "cpu", "dockerfile": ".devops/cpu.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": false, "runs_on": "ubuntu-24.04" },
+            { "tag": "cpu", "dockerfile": ".devops/cpu.Dockerfile", "platforms": "linux/arm64", "full": true, "light": true, "server": true, "free_disk_space": false, "runs_on": "ubuntu-24.04-arm" },
+            { "tag": "cpu", "dockerfile": ".devops/s390x.Dockerfile", "platforms": "linux/s390x", "full": true, "light": true, "server": true, "free_disk_space": false, "runs_on": "ubuntu-24.04-s390x" },
+            { "tag": "cuda cuda12", "dockerfile": ".devops/cuda.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
+            { "tag": "cuda cuda12", "dockerfile": ".devops/cuda.Dockerfile", "platforms": "linux/arm64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04-arm" },
+            { "tag": "cuda13", "dockerfile": ".devops/cuda-new.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
+            { "tag": "cuda13", "dockerfile": ".devops/cuda-new.Dockerfile", "platforms": "linux/arm64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04-arm" },
+            { "tag": "musa", "dockerfile": ".devops/musa.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
+            { "tag": "intel", "dockerfile": ".devops/intel.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
+            { "tag": "vulkan", "dockerfile": ".devops/vulkan.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": false, "runs_on": "ubuntu-24.04" },
+            { "tag": "vulkan", "dockerfile": ".devops/vulkan.Dockerfile", "platforms": "linux/arm64", "full": true, "light": true, "server": true, "free_disk_space": false, "runs_on": "ubuntu-24.04-arm" },
+            { "tag": "rocm", "dockerfile": ".devops/rocm.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
+            { "tag": "openvino", "dockerfile": ".devops/openvino.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": false, "runs_on": "ubuntu-24.04" }
+          ]
+          JSON
+
+          BUILD_MATRIX="$(jq -c . build-matrix.json)"
+          MERGE_MATRIX="$(jq -c '
+            reduce .[] as $entry ({}; .[$entry.tag] |= (
+              . // {
+                tag: $entry.tag,
+                arches: [],
+                full: false,
+                light: false,
+                server: false
+              }
+              | .full = (.full or ($entry.full // false))
+              | .light = (.light or ($entry.light // false))
+              | .server = (.server or ($entry.server // false))
+              | .arches += [($entry.platforms | sub("^linux/"; ""))]
+            ))
+            # Backward compatibility: s390x tags are aliases of cpu for the linux/s390x platform.
+            | if (has("cpu") and (((.cpu.arches // []) | index("s390x")) != null)) then
+                . + {
+                  s390x: {
+                    tag: "s390x",
+                    arches: ["s390x"],
+                    full: .cpu.full,
+                    light: .cpu.light,
+                    server: .cpu.server
+                  }
+                }
+              else
+                .
+              end
+            | [.[] | .arches = (.arches | unique | sort | join(" "))]
+          ' build-matrix.json)"
+
+          echo "build_matrix=$BUILD_MATRIX" >> "$GITHUB_OUTPUT"
+          echo "merge_matrix=$MERGE_MATRIX" >> "$GITHUB_OUTPUT"
+
+  push_to_registry:
+    name: Push Docker image to Docker Registry
+    needs: [prepare_matrices, create_tag]
+
+    runs-on: ${{ matrix.config.runs_on }}
+    strategy:
+      fail-fast: false
+      matrix:
+        config: ${{ fromJSON(needs.prepare_matrices.outputs.build_matrix) }}
+    steps:
+      - name: Check out the repo
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+          ref: ${{ needs.create_tag.outputs.source_tag }}
+
+      - name: Set up QEMU
+        if: ${{ contains(matrix.config.platforms, 'linux/amd64') }}
+        uses: docker/setup-qemu-action@ce360397dd3f832beb865e1373c09c0e9f86d70a # v4
+        with:
+          image: tonistiigi/binfmt:qemu-v10.2.1
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@4d04d5d9486b7bd6fa91e7baf45bbb4f8b9deedd # v4
+
+      - name: Log in to Docker Registry
+        uses: docker/login-action@b45d80f862d83dbcd57f89517bcf500b2ab88fb2 # v4
+        with:
+          registry: ghcr.io
+          username: ${{ github.repository_owner }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Determine image metadata
+        id: meta
+        shell: bash
+        run: |
+          set -euo pipefail
+
+          REPO_OWNER="${GITHUB_REPOSITORY_OWNER@L}"  # to lower case
+          REPO_NAME="${{ github.event.repository.name }}"
+          IMAGE_REPO="ghcr.io/${REPO_OWNER}/${REPO_NAME}"
+          PREFIX="${IMAGE_REPO}:"
+          PLATFORM="${{ matrix.config.platforms }}"
+          ARCH_SUFFIX="${PLATFORM#linux/}"
+
+          # list all tags possible
+          tags="${{ matrix.config.tag }}"
+          for tag in $tags; do
+              if [[ "$tag" == "cpu" ]]; then
+                  TYPE=""
+              else
+                  TYPE="-$tag"
+              fi
+              CACHETAG="${PREFIX}buildcache${TYPE}-${ARCH_SUFFIX}"
+          done
+
+          SAFE_TAGS="$(echo "$tags" | tr ' ' '_')"
+
+          echo "image_repo=$IMAGE_REPO" >> $GITHUB_OUTPUT
+          echo "arch_suffix=$ARCH_SUFFIX" >> $GITHUB_OUTPUT
+          echo "cache_output_tag=$CACHETAG" >> $GITHUB_OUTPUT
+          echo "digest_artifact_suffix=${SAFE_TAGS}-${ARCH_SUFFIX}" >> $GITHUB_OUTPUT
+          echo "cache_output_tag=$CACHETAG"  # print out for debugging
+        env:
+          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
+
+      - name: Free Disk Space (Ubuntu)
+        if: ${{ matrix.config.free_disk_space == true }}
+        uses: ggml-org/free-disk-space@v1.3.1
+        with:
+          # this might remove tools that are actually needed,
+          # if set to "true" but frees about 6 GB
+          tool-cache: false
+
+          # all of these default to true, but feel free to set to
+          # "false" if necessary for your workflow
+          android: true
+          dotnet: true
+          haskell: true
+          large-packages: true
+          docker-images: true
+          swap-storage: true
+
+      - name: Build and push Full Docker image by digest
+        id: build_full
+        if: ${{ (github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && matrix.config.full == true }}
+        uses: docker/build-push-action@d08e5c354a6adb9ed34480a06d141179aa583294 # v7
+        with:
+          context: .
+          platforms: ${{ matrix.config.platforms }}
+          outputs: type=image,name=${{ steps.meta.outputs.image_repo }},push-by-digest=true,name-canonical=true,push=true
+          file: ${{ matrix.config.dockerfile }}
+          target: full
+          provenance: false
+          build-args: |
+            ${{ matrix.config.ubuntu_version && format('UBUNTU_VERSION={0}', matrix.config.ubuntu_version) || '' }}
+            ${{ matrix.config.cuda_version && format('CUDA_VERSION={0}', matrix.config.cuda_version) || '' }}
+          # using github experimental cache
+          #cache-from: type=gha
+          #cache-to: type=gha,mode=max
+          # return to this if the experimental github cache is having issues
+          #cache-to: type=local,dest=/tmp/.buildx-cache
+          #cache-from: type=local,src=/tmp/.buildx-cache
+          # using registry cache (no storage limit)
+          cache-from: type=registry,ref=${{ steps.meta.outputs.cache_output_tag }}
+          cache-to: type=registry,ref=${{ steps.meta.outputs.cache_output_tag }},mode=max
+
+      - name: Build and push Light Docker image by digest
+        id: build_light
+        if: ${{ (github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && matrix.config.light == true }}
+        uses: docker/build-push-action@d08e5c354a6adb9ed34480a06d141179aa583294 # v7
+        with:
+          context: .
+          platforms: ${{ matrix.config.platforms }}
+          outputs: type=image,name=${{ steps.meta.outputs.image_repo }},push-by-digest=true,name-canonical=true,push=true
+          file: ${{ matrix.config.dockerfile }}
+          target: light
+          provenance: false
+          build-args: |
+            ${{ matrix.config.ubuntu_version && format('UBUNTU_VERSION={0}', matrix.config.ubuntu_version) || '' }}
+            ${{ matrix.config.cuda_version && format('CUDA_VERSION={0}', matrix.config.cuda_version) || '' }}
+          # using github experimental cache
+          #cache-from: type=gha
+          #cache-to: type=gha,mode=max
+          # return to this if the experimental github cache is having issues
+          #cache-to: type=local,dest=/tmp/.buildx-cache
+          #cache-from: type=local,src=/tmp/.buildx-cache
+          # using registry cache (no storage limit)
+          cache-from: type=registry,ref=${{ steps.meta.outputs.cache_output_tag }}
+          cache-to: type=registry,ref=${{ steps.meta.outputs.cache_output_tag }},mode=max
+
+      - name: Build and push Server Docker image by digest
+        id: build_server
+        if: ${{ (github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && matrix.config.server == true }}
+        uses: docker/build-push-action@d08e5c354a6adb9ed34480a06d141179aa583294 # v7
+        with:
+          context: .
+          platforms: ${{ matrix.config.platforms }}
+          outputs: type=image,name=${{ steps.meta.outputs.image_repo }},push-by-digest=true,name-canonical=true,push=true
+          file: ${{ matrix.config.dockerfile }}
+          target: server
+          provenance: false
+          build-args: |
+            ${{ matrix.config.ubuntu_version && format('UBUNTU_VERSION={0}', matrix.config.ubuntu_version) || '' }}
+            ${{ matrix.config.cuda_version && format('CUDA_VERSION={0}', matrix.config.cuda_version) || '' }}
+          # using github experimental cache
+          #cache-from: type=gha
+          #cache-to: type=gha,mode=max
+          # return to this if the experimental github cache is having issues
+          #cache-to: type=local,dest=/tmp/.buildx-cache
+          #cache-from: type=local,src=/tmp/.buildx-cache
+          # using registry cache (no storage limit)
+          cache-from: type=registry,ref=${{ steps.meta.outputs.cache_output_tag }}
+          cache-to: type=registry,ref=${{ steps.meta.outputs.cache_output_tag }},mode=max
+
+      - name: Export digest metadata
+        shell: bash
+        run: |
+            set -euo pipefail
+
+            TAGS="${{ matrix.config.tag }}"
+            ARCH_SUFFIX="${{ steps.meta.outputs.arch_suffix }}"
+            DIGEST_FILE="/tmp/digests/${{ steps.meta.outputs.digest_artifact_suffix }}.tsv"
+            mkdir -p /tmp/digests
+
+            add_digest_rows() {
+                local image_type="$1"
+                local digest="$2"
+
+                if [[ -z "$digest" ]]; then
+                  echo "Missing digest for image_type=${image_type}" >&2
+                  exit 1
+                fi
+
+                for tag in $TAGS; do
+                    printf '%s\t%s\t%s\t%s\n' "$tag" "$ARCH_SUFFIX" "$image_type" "$digest" >> "$DIGEST_FILE"
+                done
+            }
+
+            if [[ "${{ matrix.config.full }}" == "true" ]]; then
+                add_digest_rows "full" "${{ steps.build_full.outputs.digest }}"
+            fi
+
+            if [[ "${{ matrix.config.light }}" == "true" ]]; then
+                add_digest_rows "light" "${{ steps.build_light.outputs.digest }}"
+            fi
+
+            if [[ "${{ matrix.config.server }}" == "true" ]]; then
+                add_digest_rows "server" "${{ steps.build_server.outputs.digest }}"
+            fi
+
+      - name: Upload digest metadata
+        uses: actions/upload-artifact@bbbca2ddaa5d8feaa63e36b76fdaad77386f024f # v7
+        with:
+          name: digests-${{ steps.meta.outputs.digest_artifact_suffix }}
+          path: /tmp/digests/${{ steps.meta.outputs.digest_artifact_suffix }}.tsv
+          if-no-files-found: error
+
+  merge_arch_tags:
+    name: Create shared tags from digests
+    needs: [prepare_matrices, push_to_registry, create_tag]
+    runs-on: ubuntu-24.04
+    strategy:
+      fail-fast: false
+      matrix:
+        config: ${{ fromJSON(needs.prepare_matrices.outputs.merge_matrix) }}
+
+    steps:
+      - name: Check out the repo
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+
+      - name: Download digest metadata
+        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8
+        with:
+          pattern: digests-*
+          path: /tmp/digests
+          merge-multiple: true
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@4d04d5d9486b7bd6fa91e7baf45bbb4f8b9deedd # v4
+
+      - name: Log in to Docker Registry
+        uses: docker/login-action@b45d80f862d83dbcd57f89517bcf500b2ab88fb2 # v4
+        with:
+          registry: ghcr.io
+          username: ${{ github.repository_owner }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Create tags from digests
+        shell: bash
+        run: |
+          set -euo pipefail
+
+          REPO_OWNER="${GITHUB_REPOSITORY_OWNER@L}"  # to lower case
+          REPO_NAME="${{ github.event.repository.name }}"
+          IMAGE_REPO="ghcr.io/${REPO_OWNER}/${REPO_NAME}"
+          PREFIX="${IMAGE_REPO}:"
+          SRC_TAG="${{ needs.create_tag.outputs.source_tag }}"
+          TAGS="${{ matrix.config.tag }}"
+          ARCHES="${{ matrix.config.arches }}"
+          DIGEST_GLOB="/tmp/digests/*.tsv"
+
+          if ! ls ${DIGEST_GLOB} >/dev/null 2>&1; then
+              echo "No digest metadata found in /tmp/digests" >&2
+              exit 1
+          fi
+
+          if [[ -z "$SRC_TAG" ]]; then
+              echo "Missing source tag from create_tag" >&2
+              exit 1
+          fi
+
+          find_digest() {
+              local tag_name="$1"
+              local arch="$2"
+              local image_type="$3"
+              local digest
+
+              digest="$(awk -F '\t' -v t="$tag_name" -v a="$arch" -v i="$image_type" '$1 == t && $2 == a && $3 == i { print $4; exit }' ${DIGEST_GLOB})"
+
+              # Backward compatibility: s390x tags are aliases of cpu for the linux/s390x platform.
+              if [[ -z "$digest" && "$tag_name" == "s390x" && "$arch" == "s390x" ]]; then
+                digest="$(awk -F '\t' -v t="cpu" -v a="$arch" -v i="$image_type" '$1 == t && $2 == a && $3 == i { print $4; exit }' ${DIGEST_GLOB})"
+              fi
+
+              if [[ -z "$digest" ]]; then
+                echo "Missing digest for tag=${tag_name} arch=${arch} image_type=${image_type}" >&2
+                exit 1
+              fi
+
+              echo "$digest"
+          }
+
+          create_manifest_tags() {
+              local image_type="$1"
+              local tag_name="$2"
+              local suffix="$3"
+
+              local merged_tag="${PREFIX}${image_type}${suffix}"
+              local merged_versioned_tag="${merged_tag}-${SRC_TAG}"
+
+              local refs=()
+
+              for arch in $ARCHES; do
+                  local digest
+                  digest="$(find_digest "$tag_name" "$arch" "$image_type")"
+                  refs+=("${IMAGE_REPO}@${digest}")
+              done
+
+              echo "Creating ${merged_tag} from ${refs[*]}"
+              docker buildx imagetools create --tag "${merged_tag}" "${refs[@]}"
+
+              echo "Creating ${merged_versioned_tag} from ${refs[*]}"
+              docker buildx imagetools create --tag "${merged_versioned_tag}" "${refs[@]}"
+          }
+
+          for tag in $TAGS; do
+              if [[ "$tag" == "cpu" ]]; then
+                  TYPE=""
+              else
+                  TYPE="-$tag"
+              fi
+
+              if [[ "${{ matrix.config.full }}" == "true" ]]; then
+                  create_manifest_tags "full" "$tag" "$TYPE"
+              fi
+
+              if [[ "${{ matrix.config.light }}" == "true" ]]; then
+                  create_manifest_tags "light" "$tag" "$TYPE"
+              fi
+
+              if [[ "${{ matrix.config.server }}" == "true" ]]; then
+                  create_manifest_tags "server" "$tag" "$TYPE"
+              fi
+          done
+        env:
+          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'

.github/workflows/python-type-check.yml

@@ -31,7 +31,7 @@ jobs:
         uses: actions/setup-python@v6
         with:
           python-version: "3.11"
-          pip-install: -r requirements/requirements-all.txt ty==0.0.24
+          pip-install: -r requirements/requirements-all.txt ty==0.0.26
       # - name: Type-check with Pyright
       #   uses: jakebailey/pyright-action@v2
       #   with:

common/chat-auto-parser-generator.cpp

@@ -65,7 +65,7 @@ common_chat_params peg_generator::generate_parser(const common_chat_template &
         data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
             foreach_function(inputs.tools, [&](const json & tool) {
                 const auto & function = tool.at("function");
-                auto         schema   = function.at("parameters");
+                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
                 builder.resolve_refs(schema);
             });
             parser.build_grammar(builder, data.grammar_lazy);
@@ -221,7 +221,7 @@ common_peg_parser analyze_tools::build_tool_parser_tag_json(parser_build_context
     foreach_function(inputs.tools, [&](const json & tool) {
         const auto & func   = tool.at("function");
         std::string  name   = func.at("name");
-        const auto & schema = func.at("parameters");
+        const auto & schema = func.contains("parameters") ? func.at("parameters") : json::object();
 
         // Build call_id parser based on position (if supported)
         common_peg_parser call_id_section = p.eps();
@@ -282,19 +282,11 @@ common_peg_parser analyze_tools::build_tool_parser_tag_tagged(parser_build_conte
     common_peg_parser tool_choice = p.choice();
 
     foreach_function(inputs.tools, [&](const json & tool) {
-        const auto & func   = tool.at("function");
-        std::string  name   = func.at("name");
-        const auto & params = func.at("parameters");
-
-        if (!params.contains("properties") || !params.at("properties").is_object()) {
-            return;
-        }
-
-        const auto &          properties = params.at("properties");
+        const auto &          func       = tool.at("function");
+        std::string           name       = func.at("name");
+        const auto &          params     = func.contains("parameters") ? func.at("parameters") : json::object();
+        const auto &          properties = params.contains("properties") ? params.at("properties") : json::object();
         std::set<std::string> required;
-        if (params.contains("required") && params.at("required").is_array()) {
-            params.at("required").get_to(required);
-        }
 
         // Build parser for each argument, separating required and optional
         std::vector<common_peg_parser> required_parsers;
@@ -311,17 +303,18 @@ common_peg_parser analyze_tools::build_tool_parser_tag_tagged(parser_build_conte
                 }
             }
 
-            auto arg = p.tool_arg(
-                p.tool_arg_open(arguments.name_prefix + p.tool_arg_name(p.literal(param_name)) +
-                                arguments.name_suffix) +
-                arguments.value_prefix +
-                (type == "string" ? p.tool_arg_string_value(p.schema(p.until(arguments.value_suffix),
-                                                                     "tool-" + name + "-arg-" + param_name + "-schema",
-                                                                     param_schema, true)) :
-                                    p.tool_arg_json_value(p.schema(
-                                        p.json(), "tool-" + name + "-arg-" + param_name + "-schema", param_schema, false)) +
-                                        p.space()) +
-                p.tool_arg_close(p.literal(arguments.value_suffix)));
+            auto arg =
+                p.tool_arg(p.tool_arg_open(arguments.name_prefix + p.tool_arg_name(p.literal(param_name)) +
+                                           arguments.name_suffix) +
+                           arguments.value_prefix +
+                           (type == "string" ?
+                                p.tool_arg_string_value(p.schema(p.until(arguments.value_suffix),
+                                                                 "tool-" + name + "-arg-" + param_name + "-schema",
+                                                                 param_schema, true)) :
+                                p.tool_arg_json_value(p.schema(
+                                    p.json(), "tool-" + name + "-arg-" + param_name + "-schema", param_schema, false)) +
+                                    p.space()) +
+                           p.tool_arg_close(p.literal(arguments.value_suffix)));
 
             auto named_arg = p.rule("tool-" + name + "-arg-" + param_name, arg);
             if (is_required) {

common/chat.cpp

@@ -971,6 +971,7 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
     auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
     auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
     auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
+    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
 
     auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
         auto start           = p.rule("start", p.literal("<|start|>assistant"));
@@ -979,7 +980,13 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
         auto channel         = p.literal("<|channel|>") + (p.literal("commentary") | p.literal("analysis"));
         auto constrain_type  = p.chars("[A-Za-z0-9_-]", 1, -1);
 
-        auto analysis = p.rule("analysis", p.literal("<|channel|>analysis<|message|>") + p.reasoning(content) + end);
+        if (extract_reasoning) {
+            p.rule("analysis", p.literal("<|channel|>analysis<|message|>") + p.reasoning(content) + end);
+        } else {
+            p.rule("analysis", p.content(p.literal("<|channel|>analysis<|message|>") + content + end));
+        }
+
+        auto analysis = p.ref("analysis");
         auto preamble = p.rule("preamble", p.literal("<|channel|>commentary<|message|>") + p.content(content) + end);
         auto final_msg = p.rule("final", p.literal("<|channel|>final<|message|>") + p.content(content));
         auto any = p.rule("any", preamble | analysis);

common/common.cpp

@@ -656,14 +656,53 @@ bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_over
     return true;
 }
 
-// simple glob: * matches non-/ chars, ** matches anything including /
+static inline bool glob_class_match(const char c, const char * pattern, const char * class_end) {
+    const char * class_start = pattern;
+    bool negated = false;
+
+    if (*class_start == '!') {
+        negated = true;
+        class_start++;
+    }
+
+    // If first character after negation is ']' or '-', treat it as literal
+    if (*class_start == ']' || *class_start == '-') {
+        if (class_start < class_end && *class_start == c) {
+            return !negated;
+        }
+        class_start++;
+    }
+
+    bool matched = false;
+
+    while (class_start < class_end) {
+        if (class_start + 2 < class_end && class_start[1] == '-' && class_start[2] != ']') {
+            char start_char = *class_start;
+            char end_char = class_start[2];
+            if (c >= start_char && c <= end_char) {
+                matched = true;
+                break;
+            }
+            class_start += 3;
+        } else {
+            if (*class_start == c) {
+                matched = true;
+                break;
+            }
+            class_start++;
+        }
+    }
+
+    return negated ? !matched : matched;
+}
+
+// simple glob: * matches non-/ chars, ** matches anything including /, [] matches character class
 static inline bool glob_match(const char * pattern, const char * str) {
     if (*pattern == '\0') {
         return *str == '\0';
     }
     if (pattern[0] == '*' && pattern[1] == '*') {
         const char * p = pattern + 2;
-        if (*p == '/') p++;
         if (glob_match(p, str)) return true;
         if (*str != '\0') return glob_match(pattern, str + 1);
         return false;
@@ -678,6 +717,26 @@ static inline bool glob_match(const char * pattern, const char * str) {
     if (*pattern == '?' && *str != '\0' && *str != '/') {
         return glob_match(pattern + 1, str + 1);
     }
+    if (*pattern == '[') {
+        const char * class_end = pattern + 1;
+        // If first character after '[' is ']' or '-', treat it as literal
+        if (*class_end == ']' || *class_end == '-') {
+            class_end++;
+        }
+        while (*class_end != '\0' && *class_end != ']') {
+            class_end++;
+        }
+        if (*class_end == ']') {
+            if (*str == '\0') return false;
+            bool matched = glob_class_match(*str, pattern + 1, class_end);
+            return matched && glob_match(class_end + 1, str + 1);
+        } else {
+            if (*str == '[') {
+                return glob_match(pattern + 1, str + 1);
+            }
+            return false;
+        }
+    }
     if (*pattern == *str) {
         return glob_match(pattern + 1, str + 1);
     }

common/jinja/parser.cpp

@@ -539,6 +539,9 @@ private:
             statement_ptr step = slices.size() > 2 ? std::move(slices[2]) : nullptr;
             return mk_stmt<slice_expression>(start_pos, std::move(start), std::move(stop), std::move(step));
         }
+        if (slices.empty()) {
+            return mk_stmt<blank_expression>(start_pos);
+        }
         return std::move(slices[0]);
     }
 

common/jinja/runtime.cpp

@@ -771,10 +771,15 @@ value member_expression::execute_impl(context & ctx) {
     }
 
     JJ_DEBUG("Member expression on object type %s, property type %s", object->type().c_str(), property->type().c_str());
-    ensure_key_type_allowed(property);
-
     value val = mk_val<value_undefined>("object_property");
 
+    if (property->is_undefined()) {
+        JJ_DEBUG("%s", "Member expression property is undefined, returning undefined");
+        return val;
+    }
+
+    ensure_key_type_allowed(property);
+
     if (is_val<value_undefined>(object)) {
         JJ_DEBUG("%s", "Accessing property on undefined object, returning undefined");
         return val;

common/jinja/runtime.h

@@ -263,6 +263,14 @@ struct comment_statement : public statement {
 
 // Expressions
 
+// Represents an omitted expression in a computed member, e.g. `a[]`.
+struct blank_expression : public expression {
+    std::string type() const override { return "BlankExpression"; }
+    value execute_impl(context &) override {
+        return mk_val<value_undefined>();
+    }
+};
+
 struct member_expression : public expression {
     statement_ptr object;
     statement_ptr property;

common/json-schema-to-grammar.cpp

@@ -416,15 +416,30 @@ private:
                     i++;
                 } else if (c == '(') {
                     i++;
-                    if (i < length) {
-                        if (sub_pattern[i] == '?') {
+                    if (i < length && sub_pattern[i] == '?') {
+                        if (i + 1 < length && sub_pattern[i + 1] == ':') {
+                            i += 2; // skip "?:" for non-capturing group, treat as regular group
+                        } else {
+                            // lookahead/lookbehind (?=, ?!, ?<=, ?<!) - not supported
                             _warnings.push_back("Unsupported pattern syntax");
+                            // skip to matching ')' to avoid UB on empty seq
+                            int depth = 1;
+                            while (i < length && depth > 0) {
+                                if (sub_pattern[i] == '\\' && i + 1 < length) {
+                                    i += 2; // skip escaped character
+                                } else {
+                                    if (sub_pattern[i] == '(') depth++;
+                                    else if (sub_pattern[i] == ')') depth--;
+                                    i++;
+                                }
+                            }
+                            continue;
                         }
                     }
                     seq.emplace_back("(" + to_rule(transform()) + ")", false);
                 } else if (c == ')') {
                     i++;
-                    if (start > 0 && sub_pattern[start - 1] != '(') {
+                    if (start > 0 && sub_pattern[start - 1] != '(' && (start < 2 || sub_pattern[start - 2] != '?' || sub_pattern[start - 1] != ':')) {
                         _errors.push_back("Unbalanced parentheses");
                     }
                     return join_seq();

common/sampling.cpp

@@ -383,6 +383,12 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
         params.backend_sampling = false;
     }
 
+    if (rbudget && params.backend_sampling) {
+        LOG_WRN("%s: backend sampling is not compatible with reasoning budget, disabling\n", __func__);
+
+        params.backend_sampling = false;
+    }
+
     auto * result = new common_sampler {
         /* .params  = */ params,
         /* .grmr    = */ grmr,

convert_hf_to_gguf.py

@@ -31,10 +31,10 @@ import gguf
 from gguf.vocab import MistralTokenizerType, MistralVocab
 
 try:
-    from mistral_common.tokens.tokenizers.base import TokenizerVersion # type: ignore[import-not-found]
-    from mistral_common.tokens.tokenizers.multimodal import DATASET_MEAN as _MISTRAL_COMMON_DATASET_MEAN, DATASET_STD as _MISTRAL_COMMON_DATASET_STD # type: ignore[import-not-found]
-    from mistral_common.tokens.tokenizers.tekken import Tekkenizer # type: ignore[import-not-found]
-    from mistral_common.tokens.tokenizers.sentencepiece import ( # type: ignore[import-not-found]
+    from mistral_common.tokens.tokenizers.base import TokenizerVersion # type: ignore[import-not-found, ty:unresolved-import]
+    from mistral_common.tokens.tokenizers.multimodal import DATASET_MEAN as _MISTRAL_COMMON_DATASET_MEAN, DATASET_STD as _MISTRAL_COMMON_DATASET_STD # type: ignore[import-not-found, ty:unresolved-import]
+    from mistral_common.tokens.tokenizers.tekken import Tekkenizer # type: ignore[import-not-found, ty:unresolved-import]
+    from mistral_common.tokens.tokenizers.sentencepiece import ( # type: ignore[import-not-found, ty:unresolved-import]
         SentencePieceTokenizer,
     )
 

docs/docker.md

@@ -13,24 +13,30 @@ We have three Docker images available for this project:
 
 Additionally, there the following images, similar to the above:
 
-- `ghcr.io/ggml-org/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`)
-- `ghcr.io/ggml-org/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`)
-- `ghcr.io/ggml-org/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA support. (platforms: `linux/amd64`)
-- `ghcr.io/ggml-org/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
-- `ghcr.io/ggml-org/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
-- `ghcr.io/ggml-org/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggml-org/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA 12 support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggml-org/llama.cpp:full-cuda13`: Same as `full` but compiled with CUDA 13 support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggml-org/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA 12 support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggml-org/llama.cpp:light-cuda13`: Same as `light` but compiled with CUDA 13 support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggml-org/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA 12 support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggml-org/llama.cpp:server-cuda13`: Same as `server` but compiled with CUDA 13 support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggml-org/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:full-musa`: Same as `full` but compiled with MUSA support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:light-musa`: Same as `light` but compiled with MUSA support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:server-musa`: Same as `server` but compiled with MUSA support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:full-intel`: Same as `full` but compiled with SYCL support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:light-intel`: Same as `light` but compiled with SYCL support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:server-intel`: Same as `server` but compiled with SYCL support. (platforms: `linux/amd64`)
-- `ghcr.io/ggml-org/llama.cpp:full-vulkan`: Same as `full` but compiled with Vulkan support. (platforms: `linux/amd64`)
-- `ghcr.io/ggml-org/llama.cpp:light-vulkan`: Same as `light` but compiled with Vulkan support. (platforms: `linux/amd64`)
-- `ghcr.io/ggml-org/llama.cpp:server-vulkan`: Same as `server` but compiled with Vulkan support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:full-vulkan`: Same as `full` but compiled with Vulkan support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggml-org/llama.cpp:light-vulkan`: Same as `light` but compiled with Vulkan support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggml-org/llama.cpp:server-vulkan`: Same as `server` but compiled with Vulkan support. (platforms: `linux/amd64`, `linux/arm64`)
 - `ghcr.io/ggml-org/llama.cpp:full-openvino`: Same as `full` but compiled with OpenVino support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:light-openvino`: Same as `light` but compiled with OpenVino support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:server-openvino`: Same as `server` but compiled with OpenVino support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:full-s390x`: Identical to `full`, an alias for the `s390x` platform. (platforms: `linux/s390x`)
+- `ghcr.io/ggml-org/llama.cpp:light-s390x`: Identical to `light`, an alias for the `s390x` platform. (platforms: `linux/s390x`)
+- `ghcr.io/ggml-org/llama.cpp:server-s390x`: Identical to `server`, an alias for the `s390x` platform. (platforms: `linux/s390x`)
 
 The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](../.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](../.github/workflows/docker.yml). If you need different settings (for example, a different CUDA, ROCm or MUSA library, you'll need to build the images locally for now).
 
@@ -82,7 +88,7 @@ You may want to pass in some different `ARGS`, depending on the CUDA environment
 
 The defaults are:
 
-- `CUDA_VERSION` set to `12.4.0`
+- `CUDA_VERSION` set to `12.8.1`
 - `CUDA_DOCKER_ARCH` set to the cmake build default, which includes all the supported architectures
 
 The resulting images, are essentially the same as the non-CUDA images:

examples/model-conversion/scripts/causal/compare-logits.py

@@ -7,7 +7,7 @@ import os
 
 # Add utils directory to path for direct script execution
 sys.path.insert(0, str(Path(__file__).parent.parent / "utils"))
-from common import get_model_name_from_env_path, compare_tokens, exit_with_warning  # type: ignore[import-not-found]
+from common import get_model_name_from_env_path, compare_tokens, exit_with_warning  # type: ignore[import-not-found, ty:unresolved-import]
 
 def quick_logits_check(pytorch_file, llamacpp_file):
     """Lightweight sanity check before NMSE"""

examples/model-conversion/scripts/utils/check-nmse.py

@@ -5,7 +5,7 @@ import sys
 import os
 import argparse
 from pathlib import Path
-from common import get_model_name_from_env_path  # type: ignore[import-not-found]
+from common import get_model_name_from_env_path  # type: ignore[import-not-found, ty:unresolved-import]
 
 def calculate_nmse(reference, test):
     mse = np.mean((test - reference) ** 2)

examples/model-conversion/scripts/utils/compare_tokens.py

@@ -2,7 +2,7 @@
 
 import argparse
 import sys
-from common import compare_tokens  # type: ignore[import-not-found]
+from common import compare_tokens  # type: ignore[import-not-found, ty:unresolved-import]
 
 
 def parse_arguments():

examples/model-conversion/scripts/utils/semantic_check.py

@@ -7,7 +7,7 @@ import importlib
 from pathlib import Path
 
 from transformers import AutoTokenizer, AutoConfig, AutoModelForCausalLM, AutoModel
-from common import compare_tokens, exit_with_warning  # type: ignore[import-not-found]
+from common import compare_tokens, exit_with_warning  # type: ignore[import-not-found, ty:unresolved-import]
 
 unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
 

examples/sycl/build.sh

@@ -20,4 +20,4 @@ cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA
 #cmake --build . --config Release --target llama-bench
 
 #build all binary
-cmake --build . --config Release -j -v
+cmake --build . --config Release -j$((($(nproc)+1)/2)) -v

examples/sycl/run-llama2.sh

@@ -23,9 +23,9 @@ if [ $# -gt 0 ]; then
     GGML_SYCL_DEVICE=$1
     echo "use $GGML_SYCL_DEVICE as main GPU"
     #use signle GPU only
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-completion -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT} -mg $GGML_SYCL_DEVICE -sm none ${LOAD_MODE}
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-completion -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 200 -e -ngl ${NGL} -s 0 -c ${CONTEXT} -mg $GGML_SYCL_DEVICE -sm none ${LOAD_MODE}
 
 else
     #use multiple GPUs with same max compute units
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-completion -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT} ${LOAD_MODE}
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-completion -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 200 -e -ngl ${NGL} -s 0 -c ${CONTEXT} ${LOAD_MODE}
 fi

ggml/src/ggml-cuda/argsort.cu

@@ -47,9 +47,11 @@ void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
 #ifdef STRIDED_ITERATOR_AVAILABLE
     auto offset_iterator = cuda::make_strided_iterator(cuda::make_counting_iterator(0), ncols);
 #else
-    ggml_cuda_pool_alloc<int> offsets_alloc(pool, nrows + 1);
+    // offset_iterator needs to populate nrows + 1 elements, so we also have to ceildiv nrows + 1 by block_size
+    const int                 nrows_offset = nrows + 1;
+    ggml_cuda_pool_alloc<int> offsets_alloc(pool, nrows_offset);
     int *                     offset_iterator = offsets_alloc.get();
-    const dim3                offset_grid((nrows + block_size - 1) / block_size);
+    const dim3                offset_grid((nrows_offset + block_size - 1) / block_size);
     init_offsets<<<offset_grid, block_size, 0, stream>>>(offset_iterator, ncols, nrows);
 #endif
     CUDA_CHECK(cudaMemcpyAsync(temp_keys, x, ncols * nrows * sizeof(float), cudaMemcpyDeviceToDevice, stream));

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

@@ -2343,7 +2343,8 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
         static_assert(MMVQ_MAX_BATCH_SIZE == MMVF_MAX_BATCH_SIZE);
         if (ne2 <= MMVQ_MAX_BATCH_SIZE) {
             if (ggml_is_quantized(src0->type)) {
-                if (ne2 <= MMVQ_MMID_MAX_BATCH_SIZE) {
+                const int mmvq_mmid_max = get_mmvq_mmid_max_batch(src0->type, cc);
+                if (ne2 <= mmvq_mmid_max) {
                     ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst);
                     return;
                 }
@@ -2946,14 +2947,18 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
         }
 
         // [TAG_MUL_MAT_ID_CUDA_GRAPHS]
-        if (node->op == GGML_OP_MUL_MAT_ID && (!ggml_is_quantized(node->src[0]->type) || node->ne[2] > MMVQ_MMID_MAX_BATCH_SIZE)) {
-            // under these conditions, the mul_mat_id operation will need to synchronize the stream, so we cannot use CUDA graphs
-            // TODO: figure out a way to enable for larger batch sizes, without hurting performance
-            // ref: https://github.com/ggml-org/llama.cpp/pull/18958
-            use_cuda_graph = false;
+        if (node->op == GGML_OP_MUL_MAT_ID) {
+            const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+            const int mmvq_mmid_max = get_mmvq_mmid_max_batch(node->src[0]->type, cc);
+            if (!ggml_is_quantized(node->src[0]->type) || node->ne[2] > mmvq_mmid_max) {
+                // under these conditions, the mul_mat_id operation will need to synchronize the stream, so we cannot use CUDA graphs
+                // TODO: figure out a way to enable for larger batch sizes, without hurting performance
+                // ref: https://github.com/ggml-org/llama.cpp/pull/18958
+                use_cuda_graph = false;
 #ifndef NDEBUG
-            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported node type\n", __func__);
+                GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported node type\n", __func__);
 #endif
+            }
         }
 
         if (!use_cuda_graph) {

ggml/src/ggml-cuda/mmvq.cu

@@ -97,6 +97,194 @@ static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
     return MMVQ_PARAMETERS_GENERIC;
 }
 
+// Per-architecture maximum batch size for which MMVQ should be used for MUL_MAT_ID.
+// Returns a value <= MMVQ_MAX_BATCH_SIZE. Default is MMVQ_MAX_BATCH_SIZE.
+// Check https://github.com/ggml-org/llama.cpp/pull/20905#issuecomment-4145835627 for details
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_pascal_older(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ1_S:   return 6;
+        case GGML_TYPE_IQ1_M:   return 6;
+        case GGML_TYPE_IQ2_S:   return 4;
+        case GGML_TYPE_IQ2_XS:  return 5;
+        case GGML_TYPE_IQ2_XXS: return 5;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 4;
+        case GGML_TYPE_IQ4_NL:  return 6;
+        case GGML_TYPE_IQ4_XS:  return 5;
+        case GGML_TYPE_MXFP4:   return 4;
+        case GGML_TYPE_Q2_K:    return 4;
+        case GGML_TYPE_Q3_K:    return 4;
+        case GGML_TYPE_Q4_0:    return 6;
+        case GGML_TYPE_Q4_1:    return 6;
+        case GGML_TYPE_Q4_K:    return 5;
+        case GGML_TYPE_Q5_0:    return 6;
+        case GGML_TYPE_Q5_1:    return 6;
+        case GGML_TYPE_Q5_K:    return 5;
+        case GGML_TYPE_Q6_K:    return 4;
+        case GGML_TYPE_Q8_0:    return 4;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_turing_plus(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ2_S:   return 7;
+        case GGML_TYPE_IQ3_S:   return 6;
+        case GGML_TYPE_IQ3_XXS: return 7;
+        case GGML_TYPE_MXFP4:   return 7;
+        case GGML_TYPE_Q2_K:    return 7;
+        case GGML_TYPE_Q3_K:    return 5;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_gcn(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ1_S:   return 5;
+        case GGML_TYPE_IQ1_M:   return 5;
+        case GGML_TYPE_IQ2_S:   return 4;
+        case GGML_TYPE_IQ2_XS:  return 4;
+        case GGML_TYPE_IQ2_XXS: return 4;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 4;
+        case GGML_TYPE_IQ4_NL:  return 6;
+        case GGML_TYPE_IQ4_XS:  return 4;
+        case GGML_TYPE_Q2_K:    return 4;
+        case GGML_TYPE_Q3_K:    return 4;
+        case GGML_TYPE_Q4_0:    return 5;
+        case GGML_TYPE_Q4_1:    return 5;
+        case GGML_TYPE_Q4_K:    return 4;
+        case GGML_TYPE_Q5_K:    return 4;
+        case GGML_TYPE_Q6_K:    return 4;
+        case GGML_TYPE_Q8_0:    return 4;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_cdna(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ2_S:   return 5;
+        case GGML_TYPE_IQ2_XS:  return 5;
+        case GGML_TYPE_IQ2_XXS: return 5;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 5;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_rdna1_rdna2(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ2_S:   return 4;
+        case GGML_TYPE_IQ2_XS:  return 4;
+        case GGML_TYPE_IQ2_XXS: return 4;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 4;
+        case GGML_TYPE_Q2_K:    return 7;
+        case GGML_TYPE_Q3_K:    return 4;
+        case GGML_TYPE_Q4_K:    return 5;
+        case GGML_TYPE_Q5_K:    return 6;
+        case GGML_TYPE_Q6_K:    return 5;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_rdna3(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ1_S:   return 6;
+        case GGML_TYPE_IQ1_M:   return 6;
+        case GGML_TYPE_IQ2_S:   return 4;
+        case GGML_TYPE_IQ2_XS:  return 4;
+        case GGML_TYPE_IQ2_XXS: return 4;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 4;
+        case GGML_TYPE_IQ4_NL:  return 6;
+        case GGML_TYPE_IQ4_XS:  return 6;
+        case GGML_TYPE_Q4_K:    return 4;
+        case GGML_TYPE_Q5_K:    return 4;
+        case GGML_TYPE_Q6_K:    return 4;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_rdna4(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ1_S:   return 7;
+        case GGML_TYPE_IQ1_M:   return 7;
+        case GGML_TYPE_IQ2_S:   return 4;
+        case GGML_TYPE_IQ2_XS:  return 4;
+        case GGML_TYPE_IQ2_XXS: return 4;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 4;
+        case GGML_TYPE_IQ4_NL:  return 7;
+        case GGML_TYPE_IQ4_XS:  return 5;
+        case GGML_TYPE_MXFP4:   return 5;
+        case GGML_TYPE_Q3_K:    return 4;
+        case GGML_TYPE_Q4_0:    return 7;
+        case GGML_TYPE_Q4_1:    return 7;
+        case GGML_TYPE_Q4_K:    return 4;
+        case GGML_TYPE_Q5_0:    return 7;
+        case GGML_TYPE_Q5_1:    return 7;
+        case GGML_TYPE_Q5_K:    return 5;
+        case GGML_TYPE_Q6_K:    return 5;
+        case GGML_TYPE_Q8_0:    return 7;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+// Host function: returns the max batch size for the current arch+type at runtime.
+int get_mmvq_mmid_max_batch(ggml_type type, int cc) {
+    // NVIDIA: Volta, Ada Lovelace, and Blackwell always use MMVQ for MUL_MAT_ID.
+    if (cc == GGML_CUDA_CC_VOLTA || cc >= GGML_CUDA_CC_ADA_LOVELACE) {
+        return MMVQ_MAX_BATCH_SIZE;
+    }
+    if (cc >= GGML_CUDA_CC_TURING) {
+        return get_mmvq_mmid_max_batch_turing_plus(type);
+    }
+    if (GGML_CUDA_CC_IS_NVIDIA(cc)) {
+        return get_mmvq_mmid_max_batch_pascal_older(type);
+    }
+    // AMD
+    if (GGML_CUDA_CC_IS_RDNA4(cc)) {
+        return get_mmvq_mmid_max_batch_rdna4(type);
+    }
+    if (GGML_CUDA_CC_IS_RDNA3(cc)) {
+        return get_mmvq_mmid_max_batch_rdna3(type);
+    }
+    if (GGML_CUDA_CC_IS_RDNA1(cc) || GGML_CUDA_CC_IS_RDNA2(cc)) {
+        return get_mmvq_mmid_max_batch_rdna1_rdna2(type);
+    }
+    if (GGML_CUDA_CC_IS_CDNA(cc)) {
+        return get_mmvq_mmid_max_batch_cdna(type);
+    }
+    if (GGML_CUDA_CC_IS_GCN(cc)) {
+        return get_mmvq_mmid_max_batch_gcn(type);
+    }
+    return MMVQ_MAX_BATCH_SIZE;
+}
+
+// Device constexpr: returns the max batch size for the current arch+type at compile time.
+template <ggml_type type>
+static constexpr __device__ int get_mmvq_mmid_max_batch_for_device() {
+#if defined(RDNA4)
+    return get_mmvq_mmid_max_batch_rdna4(type);
+#elif defined(RDNA3)
+    return get_mmvq_mmid_max_batch_rdna3(type);
+#elif defined(RDNA2) || defined(RDNA1)
+    return get_mmvq_mmid_max_batch_rdna1_rdna2(type);
+#elif defined(CDNA)
+    return get_mmvq_mmid_max_batch_cdna(type);
+#elif defined(GCN)
+    return get_mmvq_mmid_max_batch_gcn(type);
+#elif defined(__CUDA_ARCH__) && (__CUDA_ARCH__ == GGML_CUDA_CC_VOLTA || __CUDA_ARCH__ >= GGML_CUDA_CC_ADA_LOVELACE)
+    return MMVQ_MAX_BATCH_SIZE;
+#elif defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
+    return get_mmvq_mmid_max_batch_turing_plus(type);
+#else
+    return get_mmvq_mmid_max_batch_pascal_older(type);
+#endif
+}
+
 static constexpr __host__ __device__ int calc_nwarps(ggml_type type, int ncols_dst, mmvq_parameter_table_id table_id) {
     if (table_id == MMVQ_PARAMETERS_GENERIC) {
         switch (ncols_dst) {
@@ -195,7 +383,7 @@ static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int
     return 1;
 }
 
-template <ggml_type type, int ncols_dst, bool has_fusion, bool is_multi_token_id = false, bool small_k = false>
+template <ggml_type type, int ncols_dst, bool has_fusion, bool small_k = false>
 __launch_bounds__(calc_nwarps(type, ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void mul_mat_vec_q(
         const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids, const ggml_cuda_mm_fusion_args_device fusion, float * __restrict__ dst,
@@ -222,22 +410,13 @@ static __global__ void mul_mat_vec_q(
 
     const uint32_t channel_dst = blockIdx.y;
 
-    uint32_t token_idx = 0;
     uint32_t channel_x;
     uint32_t channel_y;
     uint32_t sample_dst;
 
-    if constexpr (is_multi_token_id) {
-        // Multi-token MUL_MAT_ID path, adding these in the normal path causes a perf regression for n_tokens=1 case
-        token_idx  = blockIdx.z;
-        channel_x  = ids[channel_dst + token_idx * ids_stride];
-        channel_y  = fastmodulo(channel_dst, nchannels_y);
-        sample_dst = 0;
-    } else {
-        channel_x  = ncols_dst == 1 && ids ? ids[channel_dst]                     : fastdiv(channel_dst, channel_ratio);
-        channel_y  = ncols_dst == 1 && ids ? fastmodulo(channel_dst, nchannels_y) : channel_dst;
-        sample_dst = blockIdx.z;
-    }
+    channel_x  = ncols_dst == 1 && ids ? ids[channel_dst]                     : fastdiv(channel_dst, channel_ratio);
+    channel_y  = ncols_dst == 1 && ids ? fastmodulo(channel_dst, nchannels_y) : channel_dst;
+    sample_dst = blockIdx.z;
 
     const uint32_t sample_x    = fastdiv(sample_dst, sample_ratio);
     const uint32_t sample_y    = sample_dst;
@@ -294,9 +473,6 @@ static __global__ void mul_mat_vec_q(
     float tmp_gate[ncols_dst][rows_per_cuda_block] = {{0.0f}};
 
     const block_q8_1 * y = ((const block_q8_1 *) vy) + sample_y*stride_sample_y + channel_y*stride_channel_y;
-    if constexpr (is_multi_token_id) {
-        y += token_idx*stride_col_y;
-    }
     const int kbx_offset = sample_x*stride_sample_x + channel_x*stride_channel_x + row0*stride_row_x;
 
     for (int kbx = tid / (qi/vdr); kbx < blocks_per_row_x; kbx += blocks_per_iter) {
@@ -350,10 +526,6 @@ static __global__ void mul_mat_vec_q(
 
     dst += sample_dst*stride_sample_dst + channel_dst*stride_channel_dst + row0;
 
-    if constexpr (is_multi_token_id) {
-        dst += token_idx*stride_col_dst;
-    }
-
     // sum up partial sums and write back result
 #pragma unroll
     for (int j = 0; j < ncols_dst; ++j) {
@@ -413,6 +585,69 @@ static __global__ void mul_mat_vec_q(
     }
 }
 
+// Dedicated MoE multi-token kernel.
+// Grid: (ceil(nrows_x / c_rows_per_block), nchannels_dst)
+// Block: (warp_size, ncols_dst) - each warp handles one token independently.
+// No shared memory reduction needed since each warp works alone.
+template <ggml_type type, int c_rows_per_block>
+__launch_bounds__(get_mmvq_mmid_max_batch_for_device<type>()*ggml_cuda_get_physical_warp_size(), 1)
+static __global__ void mul_mat_vec_q_moe(
+        const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids,
+        float * __restrict__ dst,
+        const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t nrows_x,
+        const uint32_t stride_row_x, const uint32_t stride_col_y, const uint32_t stride_col_dst,
+        const uint32_t stride_channel_x, const uint32_t stride_channel_y, const uint32_t stride_channel_dst,
+        const uint32_t ncols_dst, const uint32_t ids_stride) {
+
+    constexpr int qk  = ggml_cuda_type_traits<type>::qk;
+    constexpr int qi  = ggml_cuda_type_traits<type>::qi;
+    constexpr int vdr = get_vdr_mmvq(type);
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+
+    constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
+
+    const uint32_t token_idx   = threadIdx.y;
+    const int      row0        = c_rows_per_block*blockIdx.x;
+    const int      blocks_per_row_x = ncols_x / qk;
+    constexpr int  blocks_per_iter  = vdr * warp_size / qi;
+
+    const uint32_t channel_dst = blockIdx.y;
+
+    if (token_idx >= ncols_dst) {
+        return;
+    }
+
+    const uint32_t channel_x = ids[channel_dst + token_idx * ids_stride];
+    const uint32_t channel_y = fastmodulo(channel_dst, nchannels_y);
+
+    const block_q8_1 * y = ((const block_q8_1 *) vy) + channel_y*stride_channel_y + token_idx*stride_col_y;
+    const int kbx_offset  = channel_x*stride_channel_x + row0*stride_row_x;
+
+    // partial sum for each thread
+    float tmp[c_rows_per_block] = {0.0f};
+
+    for (int kbx = threadIdx.x / (qi/vdr); kbx < blocks_per_row_x; kbx += blocks_per_iter) {
+        const int kby = kbx * (qk/QK8_1);
+        const int kqs = vdr * (threadIdx.x % (qi/vdr));
+
+#pragma unroll
+        for (int i = 0; i < c_rows_per_block; ++i) {
+            tmp[i] += vec_dot_q_cuda(vx, &y[kby], kbx_offset + i*stride_row_x + kbx, kqs);
+        }
+    }
+
+    // Warp-level reduction only - no shared memory needed
+#pragma unroll
+    for (int i = 0; i < c_rows_per_block; ++i) {
+        tmp[i] = warp_reduce_sum<warp_size>(tmp[i]);
+    }
+
+    // Write results
+    if (threadIdx.x < c_rows_per_block && (c_rows_per_block == 1 || uint32_t(row0 + threadIdx.x) < nrows_x)) {
+        dst[channel_dst*stride_channel_dst + token_idx*stride_col_dst + row0 + threadIdx.x] = tmp[threadIdx.x];
+    }
+}
+
 template<ggml_type type>
 static std::pair<dim3, dim3> calc_launch_params(
         const int ncols_dst, const int nrows_x, const int nchannels_dst, const int nsamples_or_ntokens,
@@ -425,7 +660,7 @@ static std::pair<dim3, dim3> calc_launch_params(
     return {block_nums, block_dims};
 }
 
-template<ggml_type type, int c_ncols_dst, bool is_multi_token_id = false, bool small_k = false>
+template<ggml_type type, int c_ncols_dst, bool small_k = false>
 static void mul_mat_vec_q_switch_fusion(
         const void * vx, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
@@ -438,7 +673,7 @@ static void mul_mat_vec_q_switch_fusion(
     const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
     if constexpr (c_ncols_dst == 1) {
         if (has_fusion) {
-            mul_mat_vec_q<type, c_ncols_dst, true, is_multi_token_id, small_k><<<block_nums, block_dims, nbytes_shared, stream>>>
+            mul_mat_vec_q<type, c_ncols_dst, true, small_k><<<block_nums, block_dims, nbytes_shared, stream>>>
                 (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
@@ -448,12 +683,33 @@ static void mul_mat_vec_q_switch_fusion(
 
     GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");
 
-    mul_mat_vec_q<type, c_ncols_dst, false, is_multi_token_id, small_k><<<block_nums, block_dims, nbytes_shared, stream>>>
+    mul_mat_vec_q<type, c_ncols_dst, false, small_k><<<block_nums, block_dims, nbytes_shared, stream>>>
         (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
         channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
         sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
 }
 
+template <ggml_type type>
+static void mul_mat_vec_q_moe_launch(
+        const void * vx, const void * vy, const int32_t * ids, float * dst,
+        const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t nrows_x,
+        const uint32_t stride_row_x, const uint32_t stride_col_y, const uint32_t stride_col_dst,
+        const uint32_t stride_channel_x, const uint32_t stride_channel_y, const uint32_t stride_channel_dst,
+        const uint32_t ncols_dst, const uint32_t ids_stride,
+        const int warp_size, const int nchannels_dst, cudaStream_t stream) {
+
+    constexpr int rows_per_block = 2; // 2 gives best perf based on tuning
+    const int64_t nblocks_rows = (nrows_x + rows_per_block - 1) / rows_per_block;
+    const dim3 block_nums(nblocks_rows, nchannels_dst);
+    const dim3 block_dims(warp_size, ncols_dst);
+
+    mul_mat_vec_q_moe<type, rows_per_block><<<block_nums, block_dims, 0, stream>>>(
+        vx, vy, ids, dst, ncols_x, nchannels_y, nrows_x,
+        stride_row_x, stride_col_y, stride_col_dst,
+        stride_channel_x, stride_channel_y, stride_channel_dst,
+        ncols_dst, ids_stride);
+}
+
 template <ggml_type type>
 static void mul_mat_vec_q_switch_ncols_dst(
         const void * vx, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
@@ -472,20 +728,62 @@ static void mul_mat_vec_q_switch_ncols_dst(
     const uint3 sample_ratio_fd  = init_fastdiv_values(nsamples_dst  / nsamples_x);
 
     const int device = ggml_cuda_get_device();
+    const int                     cc        = ggml_cuda_info().devices[device].cc;
     const int warp_size = ggml_cuda_info().devices[device].warp_size;
-    const mmvq_parameter_table_id table_id = get_device_table_id(ggml_cuda_info().devices[device].cc);
+    const mmvq_parameter_table_id table_id  = get_device_table_id(cc);
 
     const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
     const bool has_ids = ids != nullptr;
 
+    const auto should_use_small_k = [&](int c_ncols_dst) {
+        // When K is small, increase rows_per_block to match nwarps so each warp has more work to do
+        // Trigger when the full thread block covers all K blocks in a single loop iteration and few threads remain idle.
+        constexpr int qk                    = ggml_cuda_type_traits<type>::qk;
+        constexpr int qi                    = ggml_cuda_type_traits<type>::qi;
+        constexpr int vdr                   = get_vdr_mmvq(type);
+        const int     blocks_per_row_x      = ncols_x / qk;
+        const int     blocks_per_iter_1warp = vdr * warp_size / qi;
+        const int     nwarps                = calc_nwarps(type, c_ncols_dst, table_id);
+        bool          use                   = nwarps > 1 && blocks_per_row_x < nwarps * blocks_per_iter_1warp;
+
+        constexpr std::array<ggml_type, 2> iq_slow_turing = {
+            GGML_TYPE_IQ3_XXS,
+            GGML_TYPE_IQ3_S,
+        };
+        constexpr std::array<ggml_type, 8> iq_slow_other = {
+            GGML_TYPE_IQ1_S, GGML_TYPE_IQ1_M,   GGML_TYPE_IQ2_XXS, GGML_TYPE_IQ2_XS,
+            GGML_TYPE_IQ2_S, GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ3_S,   GGML_TYPE_IQ4_XS,
+        };
+        constexpr std::array<ggml_type, 3> slow_pascal = {
+            GGML_TYPE_IQ3_S,
+            GGML_TYPE_Q2_K,
+            GGML_TYPE_Q3_K,
+        };
+
+        const bool is_nvidia_turing_plus  = GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_TURING;
+        const bool is_nvidia_pascal_older = GGML_CUDA_CC_IS_NVIDIA(cc) && cc < GGML_CUDA_CC_VOLTA;
+
+        if (is_nvidia_turing_plus) {
+            if (ncols_dst == 1 &&
+                    std::find(iq_slow_turing.begin(), iq_slow_turing.end(), type) != iq_slow_turing.end()) {
+                use = false;
+            }
+        } else if ((ncols_dst == 1 && std::find(iq_slow_other.begin(), iq_slow_other.end(), type) != iq_slow_other.end()) ||
+                (is_nvidia_pascal_older && std::find(slow_pascal.begin(), slow_pascal.end(), type) != slow_pascal.end()) ||
+                GGML_CUDA_CC_IS_RDNA(cc)) {
+            use = false;
+        }
+
+        return use;
+    };
+
     if (has_ids && ncols_dst > 1) {
-        // Multi-token MUL_MAT_ID path only - single-token goes through regular path below
-        constexpr int c_ncols_dst = 1;
-        std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, ncols_dst, warp_size, table_id);
-        mul_mat_vec_q_switch_fusion<type, c_ncols_dst, true>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
-             channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-             sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-             dims.first, dims.second, 0, ids_stride, stream);
+        // Multi-token MUL_MAT_ID path - dedicated MoE kernel
+        mul_mat_vec_q_moe_launch<type>(
+            vx, vy, ids, dst, ncols_x, nchannels_y_fd, nrows_x,
+            stride_row_x, stride_col_y, stride_col_dst,
+            stride_channel_x, stride_channel_y, stride_channel_dst,
+            ncols_dst, ids_stride, warp_size, nchannels_dst, stream);
         return;
     }
 
@@ -493,31 +791,24 @@ static void mul_mat_vec_q_switch_ncols_dst(
         case 1: {
             constexpr int c_ncols_dst = 1;
 
-            // When K is small, increase rows_per_block to match nwarps so each warp has more work to do
-            // Trigger when the full thread block covers all K blocks in a single loop iteration and few threads remain idle.
-            constexpr int qk  = ggml_cuda_type_traits<type>::qk;
-            constexpr int qi  = ggml_cuda_type_traits<type>::qi;
-            constexpr int vdr = get_vdr_mmvq(type);
-            const int blocks_per_row_x = ncols_x / qk;
-            const int blocks_per_iter_1warp = vdr * warp_size / qi;
-            const int nwarps = calc_nwarps(type, c_ncols_dst, table_id);
-            const bool use_small_k = nwarps > 1 && blocks_per_row_x < nwarps * blocks_per_iter_1warp;
+            bool use_small_k = should_use_small_k(c_ncols_dst);
+
             if (use_small_k) {
-                std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst,
-                                                                    warp_size, table_id, true);
-                mul_mat_vec_q_switch_fusion<type, c_ncols_dst, false, true>(
+                std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst,
+                                                                        nsamples_dst, warp_size, table_id, true);
+                mul_mat_vec_q_switch_fusion<type, c_ncols_dst, true>(
                     vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
-                    channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                    sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                    dims.first, dims.second, 0, ids_stride, stream);
+                    channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio_fd,
+                    stride_sample_x, stride_sample_y, stride_sample_dst, dims.first, dims.second, 0, ids_stride,
+                    stream);
             } else {
-                std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst,
-                                                                    warp_size, table_id);
+                std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst,
+                                                                        nsamples_dst, warp_size, table_id);
                 mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(
                     vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
-                    channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                    sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                    dims.first, dims.second, 0, ids_stride, stream);
+                    channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio_fd,
+                    stride_sample_x, stride_sample_y, stride_sample_dst, dims.first, dims.second, 0, ids_stride,
+                    stream);
             }
         } break;
         case 2: {

ggml/src/ggml-cuda/mmvq.cuh

@@ -1,7 +1,10 @@
 #include "common.cuh"
 
 #define MMVQ_MAX_BATCH_SIZE 8 // Max. batch size for which to use MMVQ kernels.
-#define MMVQ_MMID_MAX_BATCH_SIZE 4 // Max. batch size for which to use MMVQ kernels for MUL_MAT_ID
+
+// Returns the maximum batch size for which MMVQ should be used for MUL_MAT_ID,
+// based on the quantization type and GPU architecture (compute capability).
+int get_mmvq_mmid_max_batch(ggml_type type, int cc);
 
 void ggml_cuda_mul_mat_vec_q(ggml_backend_cuda_context & ctx,
     const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst, const ggml_cuda_mm_fusion_args_host * fusion = nullptr);

ggml/src/ggml-hexagon/htp/flash-attn-ops.c

@@ -346,6 +346,9 @@ static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void *
 
     const HVX_Vector logit_cap = hvx_vec_splat_f32(factx->logit_softcap);
 
+    dma_cache m_cache;
+    dma_cache_init(&m_cache, spad_m, factx->size_m_block, DMA_CACHE_MAX_SIZE);
+
     for (uint32_t ir = ir0; ir < ir1; ++ir) {
         const uint32_t iq3 = fastdiv(ir, &factx->src0_div21);
         const uint32_t iq2 = fastdiv(ir - iq3*neq2*neq1, &factx->src0_div1);
@@ -389,9 +392,8 @@ static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void *
             // Mask
             if (mask) {
                 const uint8_t * m_src = (const uint8_t *) (mp_base + ic_start);
-                uint8_t * m_dst = spad_m + (ib % 2) * factx->size_m_block;
                 // Mask is 1D contiguous for this row
-                dma_queue_push(dma, dma_make_ptr(m_dst, m_src), current_block_size * 2, current_block_size * 2, current_block_size * 2, 1);
+                dma_cache_push(dma, &m_cache, m_src, current_block_size * 2, current_block_size * 2, current_block_size * 2, 1);
             }
 
             // FARF(HIGH, "fa %u: prefetch KVM: ir %u ib %u iq1 %u iq2 %u iq3 %u : size_k_row %u size_v_row %u bs %u: usec %u",
@@ -554,7 +556,7 @@ static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void *
                 // Mask
                 if (mask) {
                     const uint8_t * m_src = (const uint8_t *) (mp_base + next_ic_start);
-                    dma_queue_push(dma, dma_make_ptr(m_base, m_src), next_block_size * 2, next_block_size * 2, next_block_size * 2, 1);
+                    dma_cache_push(dma, &m_cache, m_src, next_block_size * 2, next_block_size * 2, next_block_size * 2, 1);
                 }
 
                 // FARF(HIGH, "fa %u: prefetch KVM: ir %u ib %u : iq1 %u iq2 %u iq3 %u : size_k_row %u size_v_row %u bs %u: usec %u",
@@ -684,7 +686,7 @@ int op_flash_attn_ext(struct htp_ops_context * octx) {
     octx->src0_spad.size_per_thread = size_q_block * 1;
     octx->src1_spad.size_per_thread = factx.size_k_block * 2;
     octx->src2_spad.size_per_thread = factx.size_v_block * 2;
-    octx->src3_spad.size_per_thread = mask ? factx.size_m_block * 2 : 0;
+    octx->src3_spad.size_per_thread = mask ? factx.size_m_block * DMA_CACHE_MAX_SIZE : 0;
     octx->dst_spad.size_per_thread  = size_vkq_acc;
 
     octx->src0_spad.size = octx->src0_spad.size_per_thread * octx->n_threads;
@@ -705,6 +707,8 @@ int op_flash_attn_ext(struct htp_ops_context * octx) {
     octx->src3_spad.data = octx->src2_spad.data + octx->src2_spad.size;
     octx->dst_spad.data  = octx->src3_spad.data + octx->src3_spad.size;
 
+    // FARF(ERROR, "fa: qrows-per-thread %u", factx.qrows_per_thread);
+
     if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
         worker_pool_run_func(octx->ctx->worker_pool, flash_attn_ext_f16_thread, &factx, octx->n_threads);
     }

ggml/src/ggml-hexagon/htp/hex-dma.h

@@ -143,7 +143,7 @@ static inline bool dma_queue_push_single_1d(dma_queue * q, dma_ptr dptr, size_t
     desc->desc_size  = 0; // 1D mode
     desc->src_bypass = dma_src_l2_bypass_on;
     desc->dst_bypass = dma_dst_l2_bypass_on;
-    desc->order      = 1;
+    desc->order      = 0;
     desc->done       = 0;
     desc->src        = (void *) dptr.src;
     desc->dst        = (void *) dptr.dst;
@@ -151,8 +151,12 @@ static inline bool dma_queue_push_single_1d(dma_queue * q, dma_ptr dptr, size_t
 
     q->dptr[q->push_idx] = dptr;
 
-    dmlink(q->tail, desc);
-    q->tail = (dma_descriptor_2d *) desc;
+    if (size) {
+        dmlink(q->tail, desc);
+        q->tail = (dma_descriptor_2d *) desc;
+    } else {
+        desc->done = 1;
+    }
 
     // FARF(ERROR, "dma-push: i %u row-size %u nrows %d dst %p src %p\n", q->push_idx, row_size, nrows, dptr.dst, dptr.src);
     q->push_idx = (q->push_idx + 1) & q->idx_mask;
@@ -175,7 +179,7 @@ static inline bool dma_queue_push_single_2d(dma_queue * q, dma_ptr dptr, size_t
     desc->dst_bypass     = dma_dst_l2_bypass_on;
     desc->src_comp       = 0;
     desc->dst_comp       = 0;
-    desc->order          = 1;
+    desc->order          = 0;
     desc->done           = 0;
     desc->src_stride     = src_stride;
     desc->dst_stride     = dst_stride;
@@ -197,8 +201,12 @@ static inline bool dma_queue_push_single_2d(dma_queue * q, dma_ptr dptr, size_t
 
     q->dptr[q->push_idx] = dptr;
 
-    dmlink(q->tail, desc);
-    q->tail = desc;
+    if (nrows) {
+        dmlink(q->tail, desc);
+        q->tail = desc;
+    } else {
+        desc->done = 1;
+    }
 
     // FARF(ERROR, "dma-push: i %u row-size %u nrows %d dst %p src %p\n", q->push_idx, row_size, nrows, dptr.dst, dptr.src);
     q->push_idx = (q->push_idx + 1) & q->idx_mask;
@@ -215,12 +223,9 @@ static inline dma_ptr dma_queue_pop(dma_queue * q) {
     dma_descriptor_2d * desc = &q->desc[q->pop_idx];
 
     // Wait for desc to complete
-    while (1) {
-        dmpoll();
-        if (desc->done) {
-            break;
-        }
+    while (!desc->done) {
         // FARF(ERROR, "dma-pop: waiting for DMA : %u\n", q->pop_idx);
+        dmpoll();
     }
 
     dptr = q->dptr[q->pop_idx];
@@ -312,6 +317,54 @@ static inline bool dma_queue_push_vtcm_to_ddr(dma_queue * q, dma_ptr dptr, size_
     return dma_queue_push(q, dptr, dst_row_size, src_row_size, dst_row_size, nrows);
 }
 
+#define DMA_CACHE_MAX_SIZE 64U
+
+typedef struct {
+    uint8_t *base;
+    uint32_t line_size;
+    uint32_t capacity;
+    uint32_t src[DMA_CACHE_MAX_SIZE];
+    uint16_t age[DMA_CACHE_MAX_SIZE];
+} dma_cache;
+
+static inline void dma_cache_init(dma_cache *c, uint8_t *base, uint32_t line_size, uint32_t capacity)
+{
+    c->capacity  = (capacity > DMA_CACHE_MAX_SIZE) ? DMA_CACHE_MAX_SIZE : capacity;
+    c->base      = base;
+    c->line_size = line_size;
+
+    for (unsigned i=0; i < c->capacity; i++) {
+        c->src[i] = 0;
+        c->age[i] = 0;
+    }
+}
+
+static inline bool dma_cache_push(dma_queue *q, dma_cache *c, const uint8_t * src, uint32_t dst_stride, uint32_t src_stride, uint32_t row_size, uint32_t nrows)
+{
+    uint32_t o_idx = 0;
+    uint16_t o_age = 0;
+    uint8_t *  dst = 0;
+
+    for (unsigned i=0; i < c->capacity; i++) {
+        if (c->src[i] == (uint32_t) src) {
+            c->age[i] = 0;
+            dst = c->base + (i * c->line_size); nrows = 0; // dummy dma
+            // FARF(ERROR, "dma-cache: found %p", src);
+        } else {
+            c->age[i]++;
+            if (c->age[i] > o_age) { o_age = c->age[i]; o_idx = i; }
+        }
+    }
+    if (!dst) {
+        // FARF(ERROR, "dma-cache: replacing #%u : age %u %p -> %p", o_idx, c->age[o_idx], (void *) c->src[o_idx], src);
+        c->age[o_idx] = 0;
+        c->src[o_idx] = (uint32_t) src;
+        dst = c->base + o_idx * c->line_size; // normal nrows dma
+    }
+
+    return dma_queue_push(q, dma_make_ptr(dst, src), dst_stride, src_stride, row_size, nrows);
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 #endif

ggml/src/ggml-hexagon/htp/rope-ops.c

@@ -333,8 +333,8 @@ static void rope_job_f32(unsigned int nth, unsigned int ith, void * data) {
                     //         (unsigned) HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - rctx->t_start));
                 }
 
-                // Skip DMA transactions from prev block (if any)
-                // No need to wait for these since the DMA is setup for in-order processing
+                // Skip output DMA transactions from prev block (if any)
+                // No need to wait for those here since we're explicitly waiting for the latest prefecthes below.
                 for (uint32_t d=0; d < dma_depth; d++) { dma_queue_pop_nowait(dma_queue); }
 
                 // Compute loop

ggml/src/ggml-opencl/CMakeLists.txt

@@ -114,6 +114,8 @@ set(GGML_OPENCL_KERNELS
     gemv_noshuffle_q4_1_f32
     gemm_noshuffle_q4_1_f32
     gemv_noshuffle_general_q8_0_f32
+    gemv_noshuffle_q4_k_f32
+    gemm_noshuffle_q4_k_f32
     gemv_noshuffle_q6_k_f32
     gemm_noshuffle_q6_k_f32
     mul

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -538,6 +538,8 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_restore_block_q4_0_noshuffle;
     cl_kernel kernel_convert_block_q4_1_noshuffle;
     cl_kernel kernel_restore_block_q4_1_noshuffle;
+    cl_kernel kernel_convert_block_q4_K_noshuffle;
+    cl_kernel kernel_restore_block_q4_K_noshuffle;
     cl_kernel kernel_convert_block_q4_K, kernel_restore_block_q4_K;
     cl_kernel kernel_convert_block_q6_K, kernel_restore_block_q6_K;
     cl_kernel kernel_mul_mat_q4_0_f32_1d_8x_flat, kernel_mul_mat_q4_0_f32_1d_16x_flat;
@@ -720,6 +722,8 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_gemm_noshuffle_q4_1_f32;
     cl_kernel kernel_mul_mm_q8_0_f32_8x4;
     cl_kernel CL_mul_mat_vec_q8_0_f32;
+    cl_kernel kernel_gemv_noshuffle_q4_k_f32;
+    cl_kernel kernel_gemm_noshuffle_q4_k_f32;
     cl_kernel kernel_gemv_noshuffle_q6_K_f32;
     cl_kernel kernel_gemm_noshuffle_q6_K_f32;
 #endif // GGML_OPENCL_USE_ADRENO_KERNELS
@@ -932,6 +936,8 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         CL_CHECK((backend_ctx->kernel_restore_block_q8_0_trans  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q8_0_trans", &err), err));
         CL_CHECK((backend_ctx->kernel_convert_block_q4_K  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q4_K", &err), err));
         CL_CHECK((backend_ctx->kernel_restore_block_q4_K  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q4_K", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_block_q4_K_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q4_K_noshuffle", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_q4_K_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q4_K_noshuffle", &err), err));
         CL_CHECK((backend_ctx->kernel_convert_block_q6_K  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q6_K", &err), err));
         CL_CHECK((backend_ctx->kernel_restore_block_q6_K  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q6_K", &err), err));
         CL_CHECK((backend_ctx->kernel_convert_block_q6_K_noshuffle  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q6_K_noshuffle", &err), err));
@@ -2619,6 +2625,45 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // gemm_noshuffle_q4_k_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "gemm_noshuffle_q4_k_f32.cl.h"
+       };
+#else
+        const std::string kernel_src = read_file("gemm_noshuffle_q4_k_f32.cl");
+#endif
+        cl_program prog = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+        CL_CHECK((backend_ctx->kernel_gemm_noshuffle_q4_k_f32 = clCreateKernel(prog, "kernel_gemm_noshuffle_q4_k_f32", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // gemv_noshuffle_q4_k_f32
+    {
+        std::string CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std +
+                                       " -cl-mad-enable ";
+        if (backend_ctx->has_vector_subgroup_broadcast) {
+            CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAST ";
+        }
+
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "gemv_noshuffle_q4_k_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("gemv_noshuffle_q4_k_f32.cl");
+#endif
+
+        cl_program prog = build_program_from_source(
+            backend_ctx->context, backend_ctx->device, kernel_src.c_str(), CL_gemv_compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_gemv_noshuffle_q4_k_f32 = clCreateKernel(prog, "kernel_gemv_noshuffle_q4_k_f32", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
     std::string CL_moe_compile_opts = std::string("-cl-std=") + opencl_c_std +
             " -cl-mad-enable "
             " -cl-fast-relaxed-math";
@@ -5060,12 +5105,25 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
             CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
         CL_CHECK(err);
 
+        #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
         cl_kernel kernel = backend_ctx->kernel_convert_block_q4_K;
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+            kernel = backend_ctx->kernel_convert_block_q4_K_noshuffle;
+        }
+        #else
+        cl_kernel kernel = backend_ctx->kernel_convert_block_q4_K;
+        #endif
+
+        cl_uchar mask_0F = 0x0F;
+        cl_uchar mask_F0 = 0xF0;
+
         CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
         CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->q));
         CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->s));
         CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra->d));
         CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extra->dm));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_uchar), &mask_0F));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_uchar), &mask_F0));
 
         size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
         size_t local_work_size[] = {64, 1, 1};
@@ -5076,6 +5134,20 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         CL_CHECK(clReleaseMemObject(data_device));
 
         tensor->extra  = extra;
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+
+            int M = tensor->ne[1];
+            int K = tensor->ne[0];
+
+            GGML_ASSERT(K % 32 == 0);
+
+            // Transpose q, d, dm as ushort
+            transpose_2d_as_16b(backend_ctx, extra->q, extra->q, size_q, K/4, M);
+            transpose_2d_as_16b(backend_ctx, extra->d, extra->d, size_d, K/256, M);
+            transpose_2d_as_16b(backend_ctx, extra->dm, extra->dm, size_dm, K/256, M);
+        }
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
         return;
     }
     if (tensor->type == GGML_TYPE_Q6_K) {
@@ -5516,12 +5588,60 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
             ggml_nbytes(tensor), NULL, &err);
         CL_CHECK(err);
 
+        cl_uchar mask_0F = 0x0F;
+        cl_uchar mask_F0 = 0xF0;
+
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+            int M = tensor->ne[1];
+            int K = tensor->ne[0];
+
+            size_t size_q  = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/2;
+            size_t size_d  = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(ggml_fp16_t);
+            size_t size_dm = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(ggml_fp16_t);
+
+            static ggml_cl_buffer buf_trans_q;
+            static ggml_cl_buffer buf_trans_d;
+            static ggml_cl_buffer buf_trans_dm;
+
+            buf_trans_q.allocate(backend_ctx->context, size_q);
+            buf_trans_d.allocate(backend_ctx->context, size_d);
+            buf_trans_dm.allocate(backend_ctx->context, size_dm);
+
+            // Transpose q, d, dm back
+            transpose_2d_as_16b(backend_ctx, extra->q,  buf_trans_q.buffer,  size_q,  M, K/4);
+            transpose_2d_as_16b(backend_ctx, extra->d,  buf_trans_d.buffer,  size_d,  M, K/256);
+            transpose_2d_as_16b(backend_ctx, extra->dm, buf_trans_dm.buffer, size_dm, M, K/256);
+
+            cl_kernel kernel = backend_ctx->kernel_restore_block_q4_K_noshuffle;
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &buf_trans_q.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->s));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &buf_trans_d.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &buf_trans_dm.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &data_device));
+            CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_uchar), &mask_0F));
+            CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_uchar), &mask_F0));
+
+            size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+            size_t local_work_size[] = {1, 1, 1};
+
+            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+                global_work_size, local_work_size, 0, NULL, NULL));
+            CL_CHECK(clEnqueueReadBuffer(queue, data_device, CL_TRUE, offset,
+                size, data, 0, NULL, NULL));
+            CL_CHECK(clReleaseMemObject(data_device));
+            return;
+        }
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
+
         cl_kernel kernel = backend_ctx->kernel_restore_block_q4_K;
         CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->q));
         CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->s));
         CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->d));
         CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra->dm));
         CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &data_device));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_uchar), &mask_0F));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_uchar), &mask_F0));
 
         size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
         size_t local_work_size[] = {1, 1, 1};
@@ -9688,6 +9808,192 @@ static void ggml_cl_mul_mat_q8_0_f32_adreno(ggml_backend_t backend, const ggml_t
 #endif
 }
 
+static void ggml_cl_mul_mat_q4_k_f32_adreno(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+    ggml_tensor_extra_cl_q4_K * extra0_q4_k = (ggml_tensor_extra_cl_q4_K *)src0->extra;
+
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int  ne00 = src0->ne[0];
+    const int  ne01 = src0->ne[1];
+
+    const int  ne1 = dst->ne[1];
+
+    GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0);
+
+    cl_context context = backend_ctx->context;
+    cl_kernel kernel;
+
+    cl_int              err;
+    cl_image_format     img_fmt;
+    cl_image_desc       img_desc;
+    cl_buffer_region    region;
+
+    int M = ne01;
+    int N = ne1;
+    int K = ne00;
+
+    cl_uchar mask_d6 = 0x3F;
+    cl_uchar mask_d4 = 0x0F;
+    cl_uchar mask_hi2 = 0xC0;
+
+    if (ne1 == 1) {
+        cl_mem q_img = nullptr;
+        cl_mem b_sub_buf = nullptr;
+        cl_mem b_img = nullptr;
+
+        // image for q
+        img_fmt = { CL_R, CL_UNSIGNED_INT32};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = M * K / 2 / 4;
+        img_desc.buffer = extra0_q4_k->q;
+        CL_CHECK((q_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        // subbuffer for activations
+        region.origin = offset1;
+        region.size = K * N * sizeof(float);
+        CL_CHECK((b_sub_buf = clCreateSubBuffer(extra1->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // image for activations
+        img_fmt = {CL_RGBA, CL_FLOAT};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = K * N / 4;
+        img_desc.buffer = b_sub_buf;
+        CL_CHECK((b_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        kernel = backend_ctx->kernel_gemv_noshuffle_q4_k_f32;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &q_img));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &extra0_q4_k->d));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra0_q4_k->dm));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &extra0_q4_k->s));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &b_img));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_int),   &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_int),   &ne01));
+        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_uchar), &mask_d6));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_uchar), &mask_d4));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_uchar), &mask_hi2));
+
+        size_t local_work_size[3] = {64, 4, 1};
+        size_t global_work_size[3] = {(size_t)CEIL_DIV(ne01/2, 64)*64, 4, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+
+        CL_CHECK(clReleaseMemObject(q_img));
+        CL_CHECK(clReleaseMemObject(b_sub_buf));
+        CL_CHECK(clReleaseMemObject(b_img));
+    } else {
+
+        cl_mem b_sub_buf = nullptr;
+        cl_mem b_sub_buf_trans = nullptr;
+        cl_mem b_img = nullptr;
+        cl_mem b_img_trans = nullptr;
+
+        // subbuffer for activations
+        region.origin = offset1;
+        region.size = K * N * sizeof(float);
+        CL_CHECK((b_sub_buf = clCreateSubBuffer(extra1->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // image for activations
+        img_fmt = {CL_RGBA, CL_FLOAT};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = K * N / 4;
+        img_desc.buffer = b_sub_buf;
+        CL_CHECK((b_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        // pad N to multiple of 8
+        int extra_elements = N % 8;
+        int padding = 0;
+        if (extra_elements > 0){
+            padding = 8 - extra_elements;
+        }
+
+        // subbuffer for transposed activations
+        region.origin = 0;
+        region.size = K * (N + padding) * sizeof(float)/2;
+        backend_ctx->prealloc_act_trans.allocate(context, region.size);
+        CL_CHECK((b_sub_buf_trans = clCreateSubBuffer(backend_ctx->prealloc_act_trans.buffer, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // image for transposed activations
+        img_fmt = {CL_RGBA, CL_HALF_FLOAT};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = K * (N + padding) / 4;
+        img_desc.buffer = b_sub_buf_trans;
+        CL_CHECK((b_img_trans = clCreateImage(context, 0, &img_fmt, &img_desc, NULL, &err), err));
+
+        // transpose activations
+        int height_B = N/4;
+        if (height_B == 0) {
+            height_B = 1;
+        }
+        int width_B = K/4;
+        int padded_height_B = (N + padding)/4;
+
+        kernel = backend_ctx->kernel_transpose_32_16;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &b_img));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &b_img_trans));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int),    &height_B));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int),    &width_B));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),    &padded_height_B));
+
+        size_t local_work_size_t[2] = { 1, 16 };
+        size_t global_work_size_t[2] = { (size_t)width_B, (size_t)padded_height_B };
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size_t, local_work_size_t, dst);
+
+        // gemm
+        kernel = backend_ctx->kernel_gemm_noshuffle_q4_k_f32;
+        int padded_N = N + padding;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0_q4_k->q));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &extra0_q4_k->s));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra0_q4_k->d));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &extra0_q4_k->dm));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &b_img_trans));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_int),   &ne01));
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_int),   &padded_N));
+        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_int),   &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_int),   &ne1));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_uchar), &mask_d6));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_uchar), &mask_d4));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_uchar), &mask_hi2));
+
+        size_t global_work_size[3] = {(size_t)CEIL_DIV(ne1, 8), (size_t)CEIL_DIV(ne01, 4), 1};
+        size_t local_work_size[3] = {1, 128, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+        CL_CHECK(clReleaseMemObject(b_sub_buf));
+        CL_CHECK(clReleaseMemObject(b_sub_buf_trans));
+        CL_CHECK(clReleaseMemObject(b_img));
+        CL_CHECK(clReleaseMemObject(b_img_trans));
+    }
+#else
+    GGML_UNUSED(backend);
+    GGML_UNUSED(src0);
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+#endif
+}
+
 static void ggml_cl_mul_mat_q6_K_f32_adreno(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
     GGML_ASSERT(src0);
@@ -10014,6 +10320,12 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
             return;
     }
 
+    // q4_k x fp32
+    if (src0t == GGML_TYPE_Q4_K && src1t == GGML_TYPE_F32) {
+            ggml_cl_mul_mat_q4_k_f32_adreno(backend, src0, src1, dst);
+            return;
+    }
+
     // q6_K x fp32
     if (src0t == GGML_TYPE_Q6_K && src1t == GGML_TYPE_F32) {
         ggml_cl_mul_mat_q6_K_f32_adreno(backend, src0, src1, dst);

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

@@ -424,13 +424,17 @@ kernel void kernel_restore_block_q8_0_trans(
 // Convert the block_q4_K format to 4 separate arrays (AOS -> SOA).
 // This kernel does not deshuffle the bits.
 // Each thread processes a super block.
+// Mask args are just to keep the signature consistent with the no-shuffle
+// version and they are not used in this kernel.
 //------------------------------------------------------------------------------
 kernel void kernel_convert_block_q4_K(
     global struct block_q4_K * src0,
     global uchar * dst_q,
     global uchar * dst_s,
     global half  * dst_d,
-    global half  * dst_dm
+    global half  * dst_dm,
+    uchar mask_0F,
+    uchar mask_F0
 ) {
     global struct block_q4_K * b = (global struct block_q4_K *) src0 + get_global_id(0);
     global uchar * q  = (global uchar *) dst_q  + QK_K/2*get_global_id(0);
@@ -451,12 +455,15 @@ kernel void kernel_convert_block_q4_K(
 
 // Restore block_q4_K from flattened arrays.
 // Each thread processes a super block.
+// Mask args are just to keep the signature consistent with the no-shuffle ones.
 kernel void kernel_restore_block_q4_K(
     global uchar * src_q,
     global uchar * src_s,
     global half  * src_d,
     global half  * src_dm,
-    global struct block_q4_K * dst
+    global struct block_q4_K * dst,
+    uchar mask_0F,
+    uchar mask_F0
 ) {
     global struct block_q4_K * b = (global struct block_q4_K *) dst + get_global_id(0);
     global uchar * q  = (global uchar *) src_q  + QK_K/2*get_global_id(0);
@@ -475,6 +482,70 @@ kernel void kernel_restore_block_q4_K(
     }
 }
 
+kernel void kernel_convert_block_q4_K_noshuffle(
+    global struct block_q4_K * src0,
+    global uchar * dst_q,
+    global uchar * dst_s,
+    global half  * dst_d,
+    global half  * dst_dm,
+    uchar mask_0F,
+    uchar mask_F0
+) {
+    global struct block_q4_K * b = (global struct block_q4_K *) src0 + get_global_id(0);
+    global uchar * q  = (global uchar *) dst_q  + QK_K/2 * get_global_id(0);
+    global uchar * s  = (global uchar *) dst_s  + K_SCALE_SIZE * get_global_id(0);
+    global half  * d  = (global half  *) dst_d  + get_global_id(0);
+    global half  * dm = (global half  *) dst_dm + get_global_id(0);
+
+    *d  = b->d;
+    *dm = b->dm;
+
+    for (int i = 0; i < QK_K / 64; ++i) {
+        for (int j = 0; j < 16; ++j) {
+            uchar x0 = b->q[i*32 + 2*j];
+            uchar x1 = b->q[i*32 + 2*j + 1];
+            q[i*32 + j]      = convert_uchar(x0 & mask_0F) | convert_uchar((x1 & mask_0F) << 4);
+            q[i*32 + j + 16] = convert_uchar((x0 & mask_F0) >> 4)   | convert_uchar(x1 & mask_F0);
+        }
+    }
+
+    for (int i = 0; i < K_SCALE_SIZE; ++i) {
+        s[i] = b->s[i];
+    }
+}
+
+kernel void kernel_restore_block_q4_K_noshuffle(
+    global uchar * src_q,
+    global uchar * src_s,
+    global half  * src_d,
+    global half  * src_dm,
+    global struct block_q4_K * dst,
+    uchar mask_0F,
+    uchar mask_F0
+) {
+    global struct block_q4_K * b = (global struct block_q4_K *) dst + get_global_id(0);
+    global uchar * q  = (global uchar *) src_q  + QK_K/2 * get_global_id(0);
+    global uchar * s  = (global uchar *) src_s  + K_SCALE_SIZE * get_global_id(0);
+    global half  * d  = (global half  *) src_d  + get_global_id(0);
+    global half  * dm = (global half  *) src_dm + get_global_id(0);
+
+    b->d  = *d;
+    b->dm = *dm;
+
+    for (int i = 0; i < QK_K / 64; ++i) {
+        for (int j = 0; j < 16; ++j) {
+            uchar lo = q[i*32 + j];
+            uchar hi = q[i*32 + j + 16];
+            b->q[i*32 + 2*j]     = convert_uchar((lo & mask_0F) | ((hi & mask_0F) << 4));
+            b->q[i*32 + 2*j + 1] = convert_uchar(((lo & mask_F0) >> 4) | (hi & mask_F0));
+        }
+    }
+
+    for (int i = 0; i < K_SCALE_SIZE; ++i) {
+        b->s[i] = s[i];
+    }
+}
+
 //------------------------------------------------------------------------------
 // kernel_convert_block_q6_K
 // Convert the block_q6_K format to 3 separate arrays (AOS -> SOA).

ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_k_f32.cl

@@ -0,0 +1,172 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+#define QK_K         256
+#define K_SCALE_SIZE 12
+
+inline void get_scale_min_k4(
+    int j,
+    global const uchar * q,
+    uchar * d,
+    uchar * m,
+    uchar mask_d6,
+    uchar mask_d4,
+    uchar mask_hi2
+) {
+    if (j < 4) {
+        *d = q[j]   & mask_d6;
+        *m = q[j+4] & mask_d6;
+    } else {
+        *d = (q[j+4] & mask_d4) | ((q[j-4] & mask_hi2) >> 2);
+        *m = ((q[j+4] >> 4) & mask_d4) | ((q[j]   & mask_hi2) >> 2);
+    }
+}
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_128
+#endif
+kernel void kernel_gemm_noshuffle_q4_k_f32(
+    global const ushort * src0_q,
+    global const uchar  * src0_s,
+    global const half   * src0_d,
+    global const half   * src0_dm,
+    read_only image1d_buffer_t src1,
+    global float * dst,
+    ulong offsetd,
+    int m,
+    int n,
+    int k,
+    int n_no_padding,
+    uchar mask_d6,
+    uchar mask_d4,
+    uchar mask_hi2
+) {
+    dst = (global float *)((global char *)dst + offsetd);
+    int n_4 = n >> 2;
+    int gy = get_global_id(0);
+    int gx = get_global_id(1);
+    int gx_2 = gx << 2;
+
+    half8 c0 = 0, c1 = 0, c2 = 0, c3 = 0;
+    half8 B;
+    half4 dequantized_weights;
+
+    int num_blocks_K = k / QK_K;
+
+    global const ushort * weight_ptr = src0_q + gx_2;
+    global const half   * d_ptr      = src0_d  + gx_2;
+    global const half   * dm_ptr     = src0_dm + gx_2;
+
+    for (int i = 0; i < k; i += 32) {
+        int sb_idx  = i / QK_K;
+        int sub_idx = (i / 32) % 8;
+
+        half4 d  = vload4(0, d_ptr  + sb_idx * m);
+        half4 dm = vload4(0, dm_ptr + sb_idx * m);
+
+        global const uchar * sc0 = src0_s + (gx_2+0) * num_blocks_K * K_SCALE_SIZE + sb_idx * K_SCALE_SIZE;
+        global const uchar * sc1 = src0_s + (gx_2+1) * num_blocks_K * K_SCALE_SIZE + sb_idx * K_SCALE_SIZE;
+        global const uchar * sc2 = src0_s + (gx_2+2) * num_blocks_K * K_SCALE_SIZE + sb_idx * K_SCALE_SIZE;
+        global const uchar * sc3 = src0_s + (gx_2+3) * num_blocks_K * K_SCALE_SIZE + sb_idx * K_SCALE_SIZE;
+
+        uchar sv0, mn0, sv1, mn1, sv2, mn2, sv3, mn3;
+        get_scale_min_k4(sub_idx, sc0, &sv0, &mn0, mask_d6, mask_d4, mask_hi2);
+        get_scale_min_k4(sub_idx, sc1, &sv1, &mn1, mask_d6, mask_d4, mask_hi2);
+        get_scale_min_k4(sub_idx, sc2, &sv2, &mn2, mask_d6, mask_d4, mask_hi2);
+        get_scale_min_k4(sub_idx, sc3, &sv3, &mn3, mask_d6, mask_d4, mask_hi2);
+
+        half4 scale = convert_half4(convert_float4(d)  * convert_float4((uchar4)(sv0, sv1, sv2, sv3)));
+        half4 mval  = convert_half4(convert_float4(dm) * convert_float4((uchar4)(mn0, mn1, mn2, mn3)));
+
+        for (int l = 0; l < 32; l += 4) {
+            int ki = i + l;
+            ushort4 bits4 = vload4(0, weight_ptr + (ki/4) * m);
+
+            // j=0
+            B.s0123 = read_imageh(src1, gy*2   + (ki+0) * n_4);
+            B.s4567 = read_imageh(src1, gy*2+1 + (ki+0) * n_4);
+            dequantized_weights.s0 = (bits4.s0 & 0x000F) * scale.s0 - mval.s0;
+            dequantized_weights.s1 = (bits4.s1 & 0x000F) * scale.s1 - mval.s1;
+            dequantized_weights.s2 = (bits4.s2 & 0x000F) * scale.s2 - mval.s2;
+            dequantized_weights.s3 = (bits4.s3 & 0x000F) * scale.s3 - mval.s3;
+            c0 += B * dequantized_weights.s0;
+            c1 += B * dequantized_weights.s1;
+            c2 += B * dequantized_weights.s2;
+            c3 += B * dequantized_weights.s3;
+
+            // j=1
+            B.s0123 = read_imageh(src1, gy*2   + (ki+1) * n_4);
+            B.s4567 = read_imageh(src1, gy*2+1 + (ki+1) * n_4);
+            dequantized_weights.s0 = ((bits4.s0 & 0x00F0) >> 4) * scale.s0 - mval.s0;
+            dequantized_weights.s1 = ((bits4.s1 & 0x00F0) >> 4) * scale.s1 - mval.s1;
+            dequantized_weights.s2 = ((bits4.s2 & 0x00F0) >> 4) * scale.s2 - mval.s2;
+            dequantized_weights.s3 = ((bits4.s3 & 0x00F0) >> 4) * scale.s3 - mval.s3;
+            c0 += B * dequantized_weights.s0;
+            c1 += B * dequantized_weights.s1;
+            c2 += B * dequantized_weights.s2;
+            c3 += B * dequantized_weights.s3;
+
+            // j=2
+            B.s0123 = read_imageh(src1, gy*2   + (ki+2) * n_4);
+            B.s4567 = read_imageh(src1, gy*2+1 + (ki+2) * n_4);
+            dequantized_weights.s0 = ((bits4.s0 & 0x0F00) >> 8) * scale.s0 - mval.s0;
+            dequantized_weights.s1 = ((bits4.s1 & 0x0F00) >> 8) * scale.s1 - mval.s1;
+            dequantized_weights.s2 = ((bits4.s2 & 0x0F00) >> 8) * scale.s2 - mval.s2;
+            dequantized_weights.s3 = ((bits4.s3 & 0x0F00) >> 8) * scale.s3 - mval.s3;
+            c0 += B * dequantized_weights.s0;
+            c1 += B * dequantized_weights.s1;
+            c2 += B * dequantized_weights.s2;
+            c3 += B * dequantized_weights.s3;
+
+            // j=3
+            B.s0123 = read_imageh(src1, gy*2   + (ki+3) * n_4);
+            B.s4567 = read_imageh(src1, gy*2+1 + (ki+3) * n_4);
+            dequantized_weights.s0 = ((bits4.s0 & 0xF000) >> 12) * scale.s0 - mval.s0;
+            dequantized_weights.s1 = ((bits4.s1 & 0xF000) >> 12) * scale.s1 - mval.s1;
+            dequantized_weights.s2 = ((bits4.s2 & 0xF000) >> 12) * scale.s2 - mval.s2;
+            dequantized_weights.s3 = ((bits4.s3 & 0xF000) >> 12) * scale.s3 - mval.s3;
+            c0 += B * dequantized_weights.s0;
+            c1 += B * dequantized_weights.s1;
+            c2 += B * dequantized_weights.s2;
+            c3 += B * dequantized_weights.s3;
+        }
+    }
+
+    int idx = (gy<<3)*m + (gx<<2);
+
+    if (idx+3 < m*n_no_padding) {
+        vstore4((float4)(c0.s0, c1.s0, c2.s0, c3.s0), 0, dst + idx);
+        idx += m;
+    }
+    if (idx+3 < m*n_no_padding) {
+        vstore4((float4)(c0.s1, c1.s1, c2.s1, c3.s1), 0, dst + idx);
+        idx += m;
+    }
+    if (idx+3 < m*n_no_padding) {
+        vstore4((float4)(c0.s2, c1.s2, c2.s2, c3.s2), 0, dst + idx);
+        idx += m;
+    }
+    if (idx+3 < m*n_no_padding) {
+        vstore4((float4)(c0.s3, c1.s3, c2.s3, c3.s3), 0, dst + idx);
+        idx += m;
+    }
+    if (idx+3 < m*n_no_padding) {
+        vstore4((float4)(c0.s4, c1.s4, c2.s4, c3.s4), 0, dst + idx);
+        idx += m;
+    }
+    if (idx+3 < m*n_no_padding) {
+        vstore4((float4)(c0.s5, c1.s5, c2.s5, c3.s5), 0, dst + idx);
+        idx += m;
+    }
+    if (idx+3 < m*n_no_padding) {
+        vstore4((float4)(c0.s6, c1.s6, c2.s6, c3.s6), 0, dst + idx);
+        idx += m;
+    }
+    if (idx+3 < m*n_no_padding) {
+        vstore4((float4)(c0.s7, c1.s7, c2.s7, c3.s7), 0, dst + idx);
+    }
+}

ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32.cl

@@ -0,0 +1,318 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
+#endif
+
+#define QK_K  256
+#define NSUBGROUPS 4
+#define SUBGROUP_SIZE 64
+
+inline void get_scale_min_k4(
+    int j,
+    global const uchar * q,
+    uchar * d,
+    uchar * m,
+    uchar mask_d6,
+    uchar mask_d4,
+    uchar mask_hi2
+) {
+    if (j < 4) {
+        *d = q[j]   & mask_d6;
+        *m = q[j+4] & mask_d6;
+    } else {
+        *d = (q[j+4] & mask_d4) | ((q[j-4] & mask_hi2) >> 2);
+        *m = ((q[j+4] >> 4) & mask_d4) | ((q[j]   & mask_hi2) >> 2);
+    }
+}
+
+#define dequantizeBlockAccum_ns_sgbroadcast_1_hi(total_sums, bits4, scale, minv, y) \
+    float shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 0); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 0); \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 0); \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 0); \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 0); \
+    total_sums.s0 += ((bits4.s2 & 0x000F) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 0); \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 0); \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 0); \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 1); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 1); \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 1); \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 1); \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 1); \
+    total_sums.s0 += ((bits4.s6 & 0x000F) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 1); \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 1); \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 1); \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_1_lo(total_sums, bits4, scale, minv, y) \
+    shared_y = sub_group_broadcast(y.s0, 2); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 2); \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 2); \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 2); \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 2); \
+    total_sums.s0 += ((bits4.s2 & 0x000F) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 2); \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 2); \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 2); \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 3); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 3); \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 3); \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 3); \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 3); \
+    total_sums.s0 += ((bits4.s6 & 0x000F) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 3); \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 3); \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 3); \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_8_hi(total_sums, bits4, scale, minv, y) \
+    float8 shared_y; \
+    shared_y = sub_group_broadcast(y, 0); \
+    total_sums.s0 += ((bits4.s0 & 0x000F)         * scale.s0 - minv.s0) * shared_y.s0; \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4)  * scale.s0 - minv.s0) * shared_y.s1; \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8)  * scale.s0 - minv.s0) * shared_y.s2; \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y.s3; \
+    total_sums.s0 += ((bits4.s2 & 0x000F)         * scale.s0 - minv.s0) * shared_y.s4; \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4)  * scale.s0 - minv.s0) * shared_y.s5; \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8)  * scale.s0 - minv.s0) * shared_y.s6; \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y.s7; \
+    total_sums.s1 += ((bits4.s1 & 0x000F)         * scale.s1 - minv.s1) * shared_y.s0; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4)  * scale.s1 - minv.s1) * shared_y.s1; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8)  * scale.s1 - minv.s1) * shared_y.s2; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y.s3; \
+    total_sums.s1 += ((bits4.s3 & 0x000F)         * scale.s1 - minv.s1) * shared_y.s4; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4)  * scale.s1 - minv.s1) * shared_y.s5; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8)  * scale.s1 - minv.s1) * shared_y.s6; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 1); \
+    total_sums.s0 += ((bits4.s4 & 0x000F)         * scale.s0 - minv.s0) * shared_y.s0; \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4)  * scale.s0 - minv.s0) * shared_y.s1; \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8)  * scale.s0 - minv.s0) * shared_y.s2; \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y.s3; \
+    total_sums.s0 += ((bits4.s6 & 0x000F)         * scale.s0 - minv.s0) * shared_y.s4; \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4)  * scale.s0 - minv.s0) * shared_y.s5; \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8)  * scale.s0 - minv.s0) * shared_y.s6; \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y.s7; \
+    total_sums.s1 += ((bits4.s5 & 0x000F)         * scale.s1 - minv.s1) * shared_y.s0; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4)  * scale.s1 - minv.s1) * shared_y.s1; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8)  * scale.s1 - minv.s1) * shared_y.s2; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y.s3; \
+    total_sums.s1 += ((bits4.s7 & 0x000F)         * scale.s1 - minv.s1) * shared_y.s4; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4)  * scale.s1 - minv.s1) * shared_y.s5; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8)  * scale.s1 - minv.s1) * shared_y.s6; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y.s7; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_8_lo(total_sums, bits4, scale, minv, y) \
+    shared_y = sub_group_broadcast(y, 2); \
+    total_sums.s0 += ((bits4.s0 & 0x000F)         * scale.s0 - minv.s0) * shared_y.s0; \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4)  * scale.s0 - minv.s0) * shared_y.s1; \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8)  * scale.s0 - minv.s0) * shared_y.s2; \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y.s3; \
+    total_sums.s0 += ((bits4.s2 & 0x000F)         * scale.s0 - minv.s0) * shared_y.s4; \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4)  * scale.s0 - minv.s0) * shared_y.s5; \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8)  * scale.s0 - minv.s0) * shared_y.s6; \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y.s7; \
+    total_sums.s1 += ((bits4.s1 & 0x000F)         * scale.s1 - minv.s1) * shared_y.s0; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4)  * scale.s1 - minv.s1) * shared_y.s1; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8)  * scale.s1 - minv.s1) * shared_y.s2; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y.s3; \
+    total_sums.s1 += ((bits4.s3 & 0x000F)         * scale.s1 - minv.s1) * shared_y.s4; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4)  * scale.s1 - minv.s1) * shared_y.s5; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8)  * scale.s1 - minv.s1) * shared_y.s6; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 3); \
+    total_sums.s0 += ((bits4.s4 & 0x000F)         * scale.s0 - minv.s0) * shared_y.s0; \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4)  * scale.s0 - minv.s0) * shared_y.s1; \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8)  * scale.s0 - minv.s0) * shared_y.s2; \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y.s3; \
+    total_sums.s0 += ((bits4.s6 & 0x000F)         * scale.s0 - minv.s0) * shared_y.s4; \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4)  * scale.s0 - minv.s0) * shared_y.s5; \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8)  * scale.s0 - minv.s0) * shared_y.s6; \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y.s7; \
+    total_sums.s1 += ((bits4.s5 & 0x000F)         * scale.s1 - minv.s1) * shared_y.s0; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4)  * scale.s1 - minv.s1) * shared_y.s1; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8)  * scale.s1 - minv.s1) * shared_y.s2; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y.s3; \
+    total_sums.s1 += ((bits4.s7 & 0x000F)         * scale.s1 - minv.s1) * shared_y.s4; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4)  * scale.s1 - minv.s1) * shared_y.s5; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8)  * scale.s1 - minv.s1) * shared_y.s6; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y.s7; \
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_gemv_noshuffle_q4_k_f32(
+        read_only  image1d_buffer_t src0_q,
+        global half2  * src0_d,
+        global half2  * src0_m,
+        global uchar  * src0_s,
+        read_only  image1d_buffer_t src1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        uchar mask_d6,
+        uchar mask_d4,
+        uchar mask_hi2)
+{
+    uint groupId = get_local_id(1);
+    uint gid     = get_global_id(0);
+    ushort slid  = get_sub_group_local_id();
+
+    uint K = ne00;
+    uint M = ne01;
+
+    uint LINE_STRIDE_A  = M / 2;
+    uint BLOCK_STRIDE_A = NSUBGROUPS * M;
+    uint scales_per_row = (K / QK_K) * 12;
+
+    private uint4     regA;
+    private half2     regS;
+    private half2     regM;
+    private float8    regB;
+
+    private float2 totalSum = (float2)(0.0f);
+
+    for (uint k = groupId; k < (K / 32); k += NSUBGROUPS) {
+        uint sb = k / 8;
+        uint j  = k % 8;
+
+        half2 d   = src0_d[gid + sb * LINE_STRIDE_A];
+        half2 dm  = src0_m[gid + sb * LINE_STRIDE_A];
+
+        global const uchar * sc0 = src0_s + 2 * gid * scales_per_row + sb * 12;
+        global const uchar * sc1 = src0_s + (2 * gid + 1) * scales_per_row + sb * 12;
+
+        uchar sv0, mn0, sv1, mn1;
+        get_scale_min_k4(j, sc0, &sv0, &mn0, mask_d6, mask_d4, mask_hi2);
+        get_scale_min_k4(j, sc1, &sv1, &mn1, mask_d6, mask_d4, mask_hi2);
+
+        regS = convert_half2(convert_float2(d)  * convert_float2((uchar2)(sv0, sv1)));
+        regM = convert_half2(convert_float2(dm) * convert_float2((uchar2)(mn0, mn1)));
+
+        if (slid < 4) {
+            regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));
+            regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));
+        }
+
+        // load half weights for two blocks in consecutive rows
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAST
+        dequantizeBlockAccum_ns_sgbroadcast_8_hi(totalSum, as_ushort8(regA), regS, regM, regB);
+#else
+        dequantizeBlockAccum_ns_sgbroadcast_1_hi(totalSum, as_ushort8(regA), regS, regM, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAST
+
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAST
+        dequantizeBlockAccum_ns_sgbroadcast_8_lo(totalSum, as_ushort8(regA), regS, regM, regB);
+#else
+        dequantizeBlockAccum_ns_sgbroadcast_1_lo(totalSum, as_ushort8(regA), regS, regM, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAST
+    }
+
+    // reduction in local memory, assumes #wave=4
+    local float2 reduceLM[SUBGROUP_SIZE * 3];
+    if (groupId == 1) {
+        reduceLM[SUBGROUP_SIZE * 0 + slid] = totalSum;
+    }
+    if (groupId == 2) {
+        reduceLM[SUBGROUP_SIZE * 1 + slid] = totalSum;
+    }
+    if (groupId == 3) {
+        reduceLM[SUBGROUP_SIZE * 2 + slid] = totalSum;
+    }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (groupId == 0) {
+        totalSum += reduceLM[SUBGROUP_SIZE * 0 + slid];
+    }
+    if (groupId == 0) {
+        totalSum += reduceLM[SUBGROUP_SIZE * 1 + slid];
+    }
+    if (groupId == 0) {
+        totalSum += reduceLM[SUBGROUP_SIZE * 2 + slid];
+    }
+
+    // 2 outputs per fiber in wave 0
+    if (groupId == 0) {
+        dst = (global float*)((global char*)dst + offsetd);
+        vstore2(totalSum, 0, &(dst[gid * 2]));
+    }
+
+}

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

@@ -1340,7 +1340,9 @@ bool rpc_server::init_tensor(const rpc_msg_init_tensor_req & request) {
     if (buffer && buffer->iface.init_tensor) {
         buffer->iface.init_tensor(buffer, tensor);
     } else {
-        GGML_LOG_ERROR("Null buffer for tensor passed to init_tensor function\n");
+        if (!buffer) {
+            GGML_LOG_ERROR("Tensor with null buffer passed to init_tensor function\n");
+        }
     }
 
     if (tensor->extra != nullptr) {

gguf-py/gguf/vocab.py

@@ -14,12 +14,12 @@ except ImportError:
     SentencePieceProcessor: Any = None
 
 try:
-    from mistral_common.tokens.tokenizers.mistral import MistralTokenizer # type: ignore[import-not-found]
-    from mistral_common.tokens.tokenizers.tekken import Tekkenizer # type: ignore[import-not-found]
-    from mistral_common.tokens.tokenizers.utils import ( # type: ignore[import-not-found]
+    from mistral_common.tokens.tokenizers.mistral import MistralTokenizer # type: ignore[import-not-found, ty:unresolved-import]
+    from mistral_common.tokens.tokenizers.tekken import Tekkenizer # type: ignore[import-not-found, ty:unresolved-import]
+    from mistral_common.tokens.tokenizers.utils import ( # type: ignore[import-not-found, ty:unresolved-import]
         _filter_valid_tokenizer_files,
     )
-    from mistral_common.tokens.tokenizers.sentencepiece import ( # type: ignore[import-not-found]
+    from mistral_common.tokens.tokenizers.sentencepiece import ( # type: ignore[import-not-found, ty:unresolved-import]
         SentencePieceTokenizer,
     )
 except ImportError:
@@ -32,7 +32,7 @@ else:
     _mistral_common_installed = True
 
 try:
-    from mistral_common.tokens.tokenizers.utils import ( # type: ignore[import-not-found]
+    from mistral_common.tokens.tokenizers.utils import ( # type: ignore[import-not-found, ty:unresolved-import]
         get_one_valid_tokenizer_file,
     )
 except ImportError:

models/templates/Qwen3.5-4B.jinja

@@ -0,0 +1,154 @@
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n' }}
+    {%- endif %}
+{%- endif %}

scripts/gen-unicode-data.py

@@ -147,7 +147,7 @@ ranges_nfd: list[tuple[int, int, int]] = [(0, 0, 0)]  # start, last, nfd
 for codepoint, norm in table_nfd:
     start = ranges_nfd[-1][0]
     if ranges_nfd[-1] != (start, codepoint - 1, norm):
-        ranges_nfd.append(None)  # type: ignore[arg-type]  # dummy, will be replaced below
+        ranges_nfd.append((0, 0, 0))  # dummy, will be replaced below
         start = codepoint
     ranges_nfd[-1] = (start, codepoint, norm)
 

src/llama-arch.cpp

@@ -557,6 +557,8 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_OUTPUT_NORM,
                 LLM_TENSOR_OUTPUT,
                 LLM_TENSOR_ROPE_FREQS,
+                LLM_TENSOR_ROPE_FACTORS_LONG,
+                LLM_TENSOR_ROPE_FACTORS_SHORT,
                 LLM_TENSOR_ATTN_NORM,
                 LLM_TENSOR_ATTN_Q,
                 LLM_TENSOR_ATTN_K,

src/llama-model-loader.cpp

@@ -1158,6 +1158,12 @@ struct ggml_tensor * llama_model_loader::create_tensor(
                     if (overrides->buft == ggml_backend_cpu_buffer_type()) {
                         // when overriding to a CPU buffer, consider the extra buffer types
                         buft = select_weight_buft(hparams, t_meta, op, buft_list_cpu);
+                        if (use_mmap) {
+                            static std::once_flag once;
+                            std::call_once(once, [] {
+                                LLAMA_LOG_WARN("llama_model_loader: tensor overrides to CPU are used with mmap enabled - consider using --no-mmap for better performance\n");
+                            });
+                        }
                     } else {
                         buft = overrides->buft;
                     }

tests/test-backend-ops.cpp

@@ -8424,6 +8424,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {1023, 2, 1, 3}, order));
         test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {1024, 2, 1, 3}, order));
         test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {1025, 2, 1, 3}, order));
+        test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {1025, 256, 1, 1}, order)); // test ceildiv in CUDA's CUB's DeviceSegmentedSort
         test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {2047, 2, 1, 3}, order));
         test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {2048, 2, 1, 3}, order));
         test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {2049, 2, 1, 3}, order));

tests/test-chat.cpp

@@ -425,6 +425,7 @@ static common_chat_tool special_function_tool_with_optional_param{
         "required": ["arg1"]
     })",
 };
+
 static common_chat_tool empty_args_tool{
     /* .name = */ "empty_args",
     /* .description = */ "A tool that takes no arguments",
@@ -433,6 +434,15 @@ static common_chat_tool empty_args_tool{
         "properties": {}
     })",
 };
+
+static common_chat_tool empty_args_tool_no_properties{
+    /* .name = */ "empty_args_no_props",
+    /* .description = */ "A tool that takes no arguments and has no properties",
+    /* .parameters = */ R"({
+        "type": "object"
+    })",
+};
+
 static common_chat_tool python_tool{
     /* .name = */ "python",
     /* .description = */ "an ipython interpreter",
@@ -1410,6 +1420,176 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
         }
     })";
 
+    {
+        // Qwen3.5 (basically same as Nemotron, but keeping separate tests just in case)
+        auto tst = peg_tester("models/templates/Qwen3.5-4B.jinja", detailed_debug);
+
+        tst.test("I'm\nthinking</think>Hello, world!\nWhat's up?")
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .enable_thinking(true)
+            .expect(message_assist_thoughts)
+            .run();
+
+                tst.test("I'm\nthinking\n</think>\nHello, world!\nWhat's up?")
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_NONE)
+            .expect_content("<think>\nI'm\nthinking\n</think>\nHello, world!\nWhat's up?")
+            .run();
+
+        tst.test("I'm\nthinking\n</think>\nHello, world!\nWhat's up?")
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .expect(message_assist_thoughts)
+            .run();
+
+        tst.test(
+               "<tool_call>\n"
+               "<function=special_function>\n"
+               "<parameter=arg1>\n1\n</parameter>\n"
+               "</function>\n"
+               "</tool_call>")
+            .enable_thinking(false)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({ special_function_tool })
+            .expect(message_assist_call)
+            .run();
+
+        tst.test(
+               "I'm\nthinking\n</think>\n"
+               "<tool_call>\n"
+               "<function=special_function>\n"
+               "<parameter=arg1>\n1\n</parameter>\n"
+               "</function>\n"
+               "</tool_call>")
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({ special_function_tool })
+            .expect(message_assist_call_thoughts)
+            .run();
+
+        tst.test(
+               "<tool_call>\n"
+               "<function=special_function>\n"
+               "<parameter=arg1>\n1\n</parameter>\n"
+               "</function>\n"
+               "</tool_call>\n"
+               "<tool_call>\n"
+               "<function=special_function_with_opt>\n"
+               "<parameter=arg1>\n1\n</parameter>\n"
+               "<parameter=arg2>\n2\n</parameter>\n"
+               "</function>\n"
+               "</tool_call>")
+            .enable_thinking(false)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .parallel_tool_calls(true)
+            .tools({
+                special_function_tool, special_function_tool_with_optional_param
+        })
+            .expect_tool_calls({
+                { "special_function", R"({"arg1": 1})", {} },
+                { "special_function_with_opt", R"({"arg1": 1, "arg2": 2})", {} },
+            })
+            .run();
+
+        tst.test(
+               "<tool_call>\n"
+               "<function=python>\n"
+               "<parameter=code>\n"
+               "def hello():\n"
+               "    print(\"Hello, world!\")\n"
+               "\n"
+               "hello()\n"
+               "</parameter>\n"
+               "</function>\n"
+               "</tool_call>")
+            .enable_thinking(false)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({
+                python_tool
+        })
+            .expect_tool_calls({
+                { "python", "{\"code\": \"def hello():\\n    print(\\\"Hello, world!\\\")\\n\\nhello()\"}", {} },
+            })
+            .run();
+
+        tst.test(
+               "I need to output the invoice details in JSON\n"
+               "</think>\n"
+               R"({"amount": 123.45, "date": "2025-12-03"})")
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .enable_thinking(true)
+            .json_schema(invoice_schema)
+            .expect_reasoning("I need to output the invoice details in JSON")
+            .expect_content(R"({"amount": 123.45, "date": "2025-12-03"})")
+            .run();
+
+        // tool call segment in reasoning
+        tst.test(
+               "Let's call a tool: <tool_call>\n"
+               "<function=python>\n"
+               "<parameter=code>\n"
+               "def hello():\n"
+               "    print(\"Not the real call!\")\n"
+               "\n"
+               "hello()\n"
+               "</parameter>\n"
+               "</function>\n"
+               "</tool_call></think>\n"
+               "<tool_call>\n"
+               "<function=python>\n"
+               "<parameter=code>\n"
+               "def hello():\n"
+               "    print(\"Hello, world!\")\n"
+               "\n"
+               "hello()\n"
+               "</parameter>\n"
+               "</function>\n"
+               "</tool_call>"
+            )
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({
+                python_tool
+        })
+            .expect_reasoning("Let's call a tool: <tool_call>\n"
+               "<function=python>\n"
+               "<parameter=code>\n"
+               "def hello():\n"
+               "    print(\"Not the real call!\")\n"
+               "\n"
+               "hello()\n"
+               "</parameter>\n"
+               "</function>\n"
+               "</tool_call>")
+            .expect_tool_calls({
+                { "python", "{\"code\": \"def hello():\\n    print(\\\"Hello, world!\\\")\\n\\nhello()\"}", {} },
+            })
+            .run();
+
+        // No args tool
+        tst.test(
+               "<tool_call>\n"
+               "<function=empty_args>\n"
+               "</function>\n"
+               "</tool_call>")
+            .enable_thinking(false)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({ empty_args_tool })
+            .expect(message_with_tool_calls("empty_args", "{}"))
+            .run();
+
+        // No args tool with no properties defined
+        tst.test(
+               "<tool_call>\n"
+               "<function=empty_args_no_props>\n"
+               "</function>\n"
+               "</tool_call>")
+            .enable_thinking(false)
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({ empty_args_tool_no_properties })
+            .expect(message_with_tool_calls("empty_args_no_props", "{}"))
+            .run();
+    }
+
     {
         // Ministral-3-14B-Reasoning-2512
         auto tst = peg_tester("models/templates/mistralai-Ministral-3-14B-Reasoning-2512.jinja", detailed_debug);
@@ -2796,6 +2976,14 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .expect(message_assist_thoughts)
             .run();
 
+        // Analysis channel (reasoning) with final channel (content) with reasoning_format = none
+        tst.test(
+               "<|channel|>analysis<|message|>I'm\nthinking<|end|><|start|>assistant<|channel|>final<|message|>Hello, world!\nWhat's "
+               "up?")
+            .reasoning_format(COMMON_REASONING_FORMAT_NONE)
+            .expect_content("<|channel|>analysis<|message|>I'm\nthinking<|end|>Hello, world!\nWhat's up?")
+            .run();
+
         // Analysis channel only (partial) - still works when reasoning format is set
         tst.test("<|channel|>analysis<|message|>I'm\nthinking")
             .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
@@ -2805,24 +2993,28 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
 
         // Tool call with recipient in role header: " to=functions.NAME<|channel|>analysis<|message|>JSON"
         tst.test(" to=functions.special_function<|channel|>analysis<|message|>{\"arg1\": 1}")
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
             .tools({ special_function_tool })
             .expect(message_assist_call)
             .run();
 
         // Tool call with recipient in channel header: "<|channel|>analysis to=functions.NAME<|message|>JSON"
         tst.test("<|channel|>analysis to=functions.special_function<|message|>{\"arg1\": 1}")
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
             .tools({ special_function_tool })
             .expect(message_assist_call)
             .run();
 
         // Tool call with constraint: " to=functions.NAME<|channel|>analysis <|constrain|>json<|message|>JSON"
         tst.test(" to=functions.special_function<|channel|>analysis <|constrain|>json<|message|>{\"arg1\": 1}")
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
             .tools({ special_function_tool })
             .expect(message_assist_call)
             .run();
 
         // Tool call in commentary channel (channel header variant)
         tst.test("<|channel|>commentary to=functions.special_function<|message|>{\"arg1\": 1}")
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
             .tools({ special_function_tool })
             .expect(message_assist_call)
             .run();

tests/test-jinja.cpp

@@ -387,6 +387,24 @@ static void test_expressions(testing & t) {
         "Bob"
     );
 
+    test_template(t, "empty computed member defaults to undefined",
+        "{{ a[]|default('fallback') }}",
+        {{"a", {{"name", "Bob"}}}},
+        "fallback"
+    );
+
+    test_template(t, "empty computed member is undefined",
+        "{{ a[] is undefined }}",
+        {{"a", {{"name", "Bob"}}}},
+        "True"
+    );
+
+    test_template(t, "undefined computed member is undefined",
+        "{{ a[undefined] is undefined }}",
+        {{"a", {{"name", "Bob"}}}},
+        "True"
+    );
+
     test_template(t, "array access",
         "{{ items[1] }}",
         {{"items", json::array({"a", "b", "c"})}},

tests/test-json-schema-to-grammar.cpp

@@ -1525,6 +1525,47 @@ int main() {
         }
     });
 
+    // C++ only tests (features not yet supported in JS/Python implementations)
+    {
+        fprintf(stderr, "#\n# Testing C++ only features\n#\n");
+        auto run = [](const TestCase & tc) {
+            fprintf(stderr, "- %s\n", tc.name.c_str());
+            try {
+                tc.verify(json_schema_to_grammar(nlohmann::ordered_json::parse(tc.schema), true));
+                tc.verify_status(SUCCESS);
+            } catch (const std::invalid_argument & ex) {
+                fprintf(stderr, "Error: %s\n", ex.what());
+                tc.verify_status(FAILURE);
+            }
+        };
+
+        run({
+            SUCCESS,
+            "regexp with non-capturing group",
+            R"""({
+                "type": "string",
+                "pattern": "^(?:foo|bar)baz$"
+            })""",
+            R"""(
+                root ::= "\"" (("foo" | "bar") "baz") "\"" space
+                space ::= | " " | "\n"{1,2} [ \t]{0,20}
+            )""",
+        });
+
+        run({
+            SUCCESS,
+            "regexp with nested non-capturing groups",
+            R"""({
+                "type": "string",
+                "pattern": "^(?:(?:ab)+c)?d$"
+            })""",
+            R"""(
+                root ::= "\"" ((("ab")+ "c")? "d") "\"" space
+                space ::= | " " | "\n"{1,2} [ \t]{0,20}
+            )""",
+        });
+    }
+
     if (getenv("LLAMA_SKIP_TESTS_SLOW_ON_EMULATOR")) {
         fprintf(stderr, "\033[33mWARNING: Skipping slow tests on emulator.\n\033[0m");
     } else {

tools/server/README.md

@@ -1775,6 +1775,16 @@ Apart from error types supported by OAI, we also have custom types that are spec
 }
 ```
 
+### Custom default Web UI preferences
+
+You can specify default preferences for the web UI using `--webui-config <JSON config>` or `--webui-config-file <path to JSON config>`. For example, you can disable pasting long text as attachments and enable rendering Markdown in user messages with this command:
+
+```bash
+./llama-server -m model.gguf --webui-config '{"pasteLongTextToFileLen": 0, "renderUserContentAsMarkdown": true}'
+```
+
+You may find available preferences in [settings-config.ts](webui/src/lib/constants/settings-config.ts).
+
 ### Legacy completion web UI
 
 A new chat-based UI has replaced the old completion-based since [this PR](https://github.com/ggml-org/llama.cpp/pull/10175). If you want to use the old completion, start the server with `--path ./tools/server/public_legacy`

tools/server/server-context.cpp

@@ -2493,7 +2493,7 @@ private:
                     bool has_mtmd = false;
 
                     // check if we should process the image
-                    if (slot.prompt.n_tokens() < slot.task->n_tokens() && input_tokens[slot.prompt.n_tokens()] == LLAMA_TOKEN_NULL) {
+                    while (slot.prompt.n_tokens() < slot.task->n_tokens() && input_tokens[slot.prompt.n_tokens()] == LLAMA_TOKEN_NULL) {
                         // process the image
                         size_t n_tokens_out = 0;
                         int32_t res = input_tokens.process_chunk(ctx, mctx, slot.prompt.n_tokens(), slot.prompt.tokens.pos_next(), slot.id, n_tokens_out);

tools/server/server-cors-proxy.h

@@ -32,13 +32,22 @@ static server_http_res_ptr proxy_request(const server_http_req & req, std::strin
 
     SRV_INF("proxying %s request to %s://%s:%i%s\n", method.c_str(), parsed_url.scheme.c_str(), parsed_url.host.c_str(), parsed_url.port, parsed_url.path.c_str());
 
+    std::map<std::string, std::string> headers;
+    for (auto [key, value] : req.headers) {
+        auto new_key = key;
+        if (string_starts_with(new_key, "X-Proxy-Header-")) {
+            string_replace_all(new_key, "X-Proxy-Header-", "");
+        }
+        headers[new_key] = value;
+    }
+
     auto proxy = std::make_unique<server_http_proxy>(
             method,
             parsed_url.scheme,
             parsed_url.host,
             parsed_url.port,
             parsed_url.path,
-            req.headers,
+            headers,
             req.body,
             req.should_stop,
             600, // timeout_read (default to 10 minutes)

tools/server/server-http.h

@@ -35,7 +35,7 @@ using server_http_res_ptr = std::unique_ptr<server_http_res>;
 
 struct server_http_req {
     std::map<std::string, std::string> params; // path_params + query_params
-    std::map<std::string, std::string> headers; // reserved for future use
+    std::map<std::string, std::string> headers; // used by MCP proxy
     std::string path;
     std::string query_string; // query parameters string (e.g. "action=save")
     std::string body;

tools/server/tests/utils.py

@@ -116,7 +116,7 @@ class ServerProcess:
             self.server_port = int(os.environ["PORT"])
         self.external_server = "DEBUG_EXTERNAL" in os.environ
 
-    def start(self, timeout_seconds: int | None = DEFAULT_HTTP_TIMEOUT) -> None:
+    def start(self, timeout_seconds: int = DEFAULT_HTTP_TIMEOUT) -> None:
         if self.external_server:
             print(f"[external_server]: Assuming external server running on {self.server_host}:{self.server_port}")
             return
@@ -288,7 +288,15 @@ class ServerProcess:
             server_instances.remove(self)
         if self.process:
             print(f"Stopping server with pid={self.process.pid}")
-            self.process.kill()
+            self.process.terminate()
+            try:
+                self.process.wait(timeout=5)
+            except subprocess.TimeoutExpired:
+                print(f"Server pid={self.process.pid} did not terminate in time, killing")
+                self.process.kill()
+                self.process.wait(timeout=5)
+            except Exception as e:
+                print(f"Error waiting for server: {e}")
             self.process = None
 
     def make_request(

tools/server/webui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActions.svelte

@@ -10,9 +10,9 @@
 		ModelsSelector,
 		ModelsSelectorSheet
 	} from '$lib/components/app';
-	import { DialogChatSettings } from '$lib/components/app/dialogs';
 	import { SETTINGS_SECTION_TITLES } from '$lib/constants';
 	import { mcpStore } from '$lib/stores/mcp.svelte';
+	import { getChatSettingsDialogContext } from '$lib/contexts';
 	import { FileTypeCategory } from '$lib/enums';
 	import { getFileTypeCategory } from '$lib/utils';
 	import { config } from '$lib/stores/settings.svelte';
@@ -169,7 +169,7 @@
 		selectorModelRef?.open();
 	}
 
-	let showChatSettingsDialogWithMcpSection = $state(false);
+	const chatSettingsDialog = getChatSettingsDialogContext();
 
 	let hasMcpPromptsSupport = $derived.by(() => {
 		const perChatOverrides = conversationsStore.getAllMcpServerOverrides();
@@ -197,7 +197,7 @@
 				{onSystemPromptClick}
 				{onMcpPromptClick}
 				{onMcpResourcesClick}
-				onMcpSettingsClick={() => (showChatSettingsDialogWithMcpSection = true)}
+				onMcpSettingsClick={() => chatSettingsDialog.open(SETTINGS_SECTION_TITLES.MCP)}
 			/>
 		{:else}
 			<ChatFormActionAttachmentsDropdown
@@ -210,13 +210,13 @@
 				{onSystemPromptClick}
 				{onMcpPromptClick}
 				{onMcpResourcesClick}
-				onMcpSettingsClick={() => (showChatSettingsDialogWithMcpSection = true)}
+				onMcpSettingsClick={() => chatSettingsDialog.open(SETTINGS_SECTION_TITLES.MCP)}
 			/>
 		{/if}
 
 		<McpServersSelector
 			{disabled}
-			onSettingsClick={() => (showChatSettingsDialogWithMcpSection = true)}
+			onSettingsClick={() => chatSettingsDialog.open(SETTINGS_SECTION_TITLES.MCP)}
 		/>
 	</div>
 
@@ -265,9 +265,3 @@
 		/>
 	{/if}
 </div>
-
-<DialogChatSettings
-	open={showChatSettingsDialogWithMcpSection}
-	onOpenChange={(open) => (showChatSettingsDialogWithMcpSection = open)}
-	initialSection={SETTINGS_SECTION_TITLES.MCP}
-/>

tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessage.svelte

@@ -180,6 +180,10 @@
 		chatActions.continueAssistantMessage(message);
 	}
 
+	function handleForkConversation(options: { name: string; includeAttachments: boolean }) {
+		chatActions.forkConversation(message, options);
+	}
+
 	function handleNavigateToSibling(siblingId: string) {
 		chatActions.navigateToSibling(siblingId);
 	}
@@ -285,6 +289,7 @@
 		onCopy={handleCopy}
 		onDelete={handleDelete}
 		onEdit={handleEdit}
+		onForkConversation={handleForkConversation}
 		onNavigateToSibling={handleNavigateToSibling}
 		onShowDeleteDialogChange={handleShowDeleteDialogChange}
 		{showDeleteDialog}
@@ -303,6 +308,7 @@
 		onCopy={handleCopy}
 		onDelete={handleDelete}
 		onEdit={handleEdit}
+		onForkConversation={handleForkConversation}
 		onNavigateToSibling={handleNavigateToSibling}
 		onRegenerate={handleRegenerate}
 		onShowDeleteDialogChange={handleShowDeleteDialogChange}

tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageActions.svelte

@@ -1,12 +1,16 @@
 <script lang="ts">
-	import { Edit, Copy, RefreshCw, Trash2, ArrowRight } from '@lucide/svelte';
+	import { Edit, Copy, RefreshCw, Trash2, ArrowRight, GitBranch } from '@lucide/svelte';
 	import {
 		ActionIcon,
 		ChatMessageBranchingControls,
 		DialogConfirmation
 	} from '$lib/components/app';
 	import { Switch } from '$lib/components/ui/switch';
+	import { Checkbox } from '$lib/components/ui/checkbox';
+	import Input from '$lib/components/ui/input/input.svelte';
+	import Label from '$lib/components/ui/label/label.svelte';
 	import { MessageRole } from '$lib/enums';
+	import { activeConversation } from '$lib/stores/conversations.svelte';
 
 	interface Props {
 		role: MessageRole.USER | MessageRole.ASSISTANT;
@@ -24,6 +28,7 @@
 		onEdit?: () => void;
 		onRegenerate?: () => void;
 		onContinue?: () => void;
+		onForkConversation?: (options: { name: string; includeAttachments: boolean }) => void;
 		onDelete: () => void;
 		onConfirmDelete: () => void;
 		onNavigateToSibling?: (siblingId: string) => void;
@@ -42,6 +47,7 @@
 		onConfirmDelete,
 		onContinue,
 		onDelete,
+		onForkConversation,
 		onNavigateToSibling,
 		onShowDeleteDialogChange,
 		onRegenerate,
@@ -53,10 +59,27 @@
 		onRawOutputToggle
 	}: Props = $props();
 
+	let showForkDialog = $state(false);
+	let forkName = $state('');
+	let forkIncludeAttachments = $state(true);
+
 	function handleConfirmDelete() {
 		onConfirmDelete();
 		onShowDeleteDialogChange(false);
 	}
+
+	function handleOpenForkDialog() {
+		const conv = activeConversation();
+
+		forkName = `Fork of ${conv?.name ?? 'Conversation'}`;
+		forkIncludeAttachments = true;
+		showForkDialog = true;
+	}
+
+	function handleConfirmFork() {
+		onForkConversation?.({ name: forkName.trim(), includeAttachments: forkIncludeAttachments });
+		showForkDialog = false;
+	}
 </script>
 
 <div class="relative {justify === 'start' ? 'mt-2' : ''} flex h-6 items-center justify-between">
@@ -86,6 +109,10 @@
 				<ActionIcon icon={ArrowRight} tooltip="Continue" onclick={onContinue} />
 			{/if}
 
+			{#if onForkConversation}
+				<ActionIcon icon={GitBranch} tooltip="Fork conversation" onclick={handleOpenForkDialog} />
+			{/if}
+
 			<ActionIcon icon={Trash2} tooltip="Delete" onclick={onDelete} />
 		</div>
 	</div>
@@ -116,3 +143,42 @@
 	onConfirm={handleConfirmDelete}
 	onCancel={() => onShowDeleteDialogChange(false)}
 />
+
+<DialogConfirmation
+	bind:open={showForkDialog}
+	title="Fork Conversation"
+	description="Create a new conversation branching from this message."
+	confirmText="Fork"
+	cancelText="Cancel"
+	icon={GitBranch}
+	onConfirm={handleConfirmFork}
+	onCancel={() => (showForkDialog = false)}
+>
+	<div class="flex flex-col gap-4 py-2">
+		<div class="flex flex-col gap-2">
+			<Label for="fork-name">Title</Label>
+
+			<Input
+				id="fork-name"
+				class="text-foreground"
+				placeholder="Enter fork name"
+				type="text"
+				bind:value={forkName}
+			/>
+		</div>
+
+		<div class="flex items-center gap-2">
+			<Checkbox
+				id="fork-attachments"
+				checked={forkIncludeAttachments}
+				onCheckedChange={(checked) => {
+					forkIncludeAttachments = checked === true;
+				}}
+			/>
+
+			<Label for="fork-attachments" class="cursor-pointer text-sm font-normal">
+				Include all attachments
+			</Label>
+		</div>
+	</div>
+</DialogConfirmation>

tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageAssistant.svelte

@@ -39,6 +39,7 @@
 		onContinue?: () => void;
 		onDelete: () => void;
 		onEdit?: () => void;
+		onForkConversation?: (options: { name: string; includeAttachments: boolean }) => void;
 		onNavigateToSibling?: (siblingId: string) => void;
 		onRegenerate: (modelOverride?: string) => void;
 		onShowDeleteDialogChange: (show: boolean) => void;
@@ -58,6 +59,7 @@
 		onCopy,
 		onDelete,
 		onEdit,
+		onForkConversation,
 		onNavigateToSibling,
 		onRegenerate,
 		onShowDeleteDialogChange,
@@ -345,6 +347,7 @@
 			onContinue={currentConfig.enableContinueGeneration && !hasReasoningMarkers
 				? onContinue
 				: undefined}
+			{onForkConversation}
 			{onDelete}
 			{onConfirmDelete}
 			{onNavigateToSibling}

tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageUser.svelte

@@ -21,6 +21,7 @@
 		onEdit: () => void;
 		onDelete: () => void;
 		onConfirmDelete: () => void;
+		onForkConversation?: (options: { name: string; includeAttachments: boolean }) => void;
 		onShowDeleteDialogChange: (show: boolean) => void;
 		onNavigateToSibling?: (siblingId: string) => void;
 		onCopy: () => void;
@@ -35,6 +36,7 @@
 		onEdit,
 		onDelete,
 		onConfirmDelete,
+		onForkConversation,
 		onShowDeleteDialogChange,
 		onNavigateToSibling,
 		onCopy
@@ -114,6 +116,7 @@
 					{onCopy}
 					{onDelete}
 					{onEdit}
+					{onForkConversation}
 					{onNavigateToSibling}
 					{onShowDeleteDialogChange}
 					{siblingInfo}

tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessages.svelte

@@ -79,6 +79,13 @@
 			onUserAction?.();
 			await chatStore.continueAssistantMessage(message.id);
 			refreshAllMessages();
+		},
+
+		forkConversation: async (
+			message: DatabaseMessage,
+			options: { name: string; includeAttachments: boolean }
+		) => {
+			await conversationsStore.forkConversation(message.id, options);
 		}
 	});
 

tools/server/webui/src/lib/components/app/chat/ChatScreen/ChatScreenHeader.svelte

@@ -1,16 +1,11 @@
 <script lang="ts">
 	import { Settings } from '@lucide/svelte';
-	import { DialogChatSettings } from '$lib/components/app';
 	import { Button } from '$lib/components/ui/button';
 	import { useSidebar } from '$lib/components/ui/sidebar';
+	import { getChatSettingsDialogContext } from '$lib/contexts';
 
 	const sidebar = useSidebar();
-
-	let settingsOpen = $state(false);
-
-	function toggleSettings() {
-		settingsOpen = true;
-	}
+	const chatSettingsDialog = getChatSettingsDialogContext();
 </script>
 
 <header
@@ -22,12 +17,10 @@
 		<Button
 			variant="ghost"
 			size="icon-lg"
-			onclick={toggleSettings}
+			onclick={() => chatSettingsDialog.open()}
 			class="rounded-full backdrop-blur-lg"
 		>
 			<Settings class="h-4 w-4" />
 		</Button>
 	</div>
 </header>
-
-<DialogChatSettings open={settingsOpen} onOpenChange={(open) => (settingsOpen = open)} />

tools/server/webui/src/lib/components/app/chat/ChatSidebar/ChatSidebar.svelte

@@ -1,13 +1,18 @@
 <script lang="ts">
 	import { goto } from '$app/navigation';
 	import { page } from '$app/state';
-	import { Trash2 } from '@lucide/svelte';
+	import { Trash2, Pencil } from '@lucide/svelte';
 	import { ChatSidebarConversationItem, DialogConfirmation } from '$lib/components/app';
+	import { Checkbox } from '$lib/components/ui/checkbox';
+	import Label from '$lib/components/ui/label/label.svelte';
 	import ScrollArea from '$lib/components/ui/scroll-area/scroll-area.svelte';
 	import * as Sidebar from '$lib/components/ui/sidebar';
-	import * as AlertDialog from '$lib/components/ui/alert-dialog';
 	import Input from '$lib/components/ui/input/input.svelte';
-	import { conversationsStore, conversations } from '$lib/stores/conversations.svelte';
+	import {
+		conversationsStore,
+		conversations,
+		buildConversationTree
+	} from '$lib/stores/conversations.svelte';
 	import { chatStore } from '$lib/stores/chat.svelte';
 	import { getPreviewText } from '$lib/utils';
 	import ChatSidebarActions from './ChatSidebarActions.svelte';
@@ -18,6 +23,7 @@
 	let isSearchModeActive = $state(false);
 	let searchQuery = $state('');
 	let showDeleteDialog = $state(false);
+	let deleteWithForks = $state(false);
 	let showEditDialog = $state(false);
 	let selectedConversation = $state<DatabaseConversation | null>(null);
 	let editedName = $state('');
@@ -35,10 +41,30 @@
 		return conversations();
 	});
 
+	let conversationTree = $derived(buildConversationTree(filteredConversations));
+
+	let selectedConversationHasDescendants = $derived.by(() => {
+		if (!selectedConversation) return false;
+
+		const allConvs = conversations();
+		const queue = [selectedConversation.id];
+
+		while (queue.length > 0) {
+			const parentId = queue.pop()!;
+
+			for (const c of allConvs) {
+				if (c.forkedFromConversationId === parentId) return true;
+			}
+		}
+
+		return false;
+	});
+
 	async function handleDeleteConversation(id: string) {
 		const conversation = conversations().find((conv) => conv.id === id);
 		if (conversation) {
 			selectedConversation = conversation;
+			deleteWithForks = false;
 			showDeleteDialog = true;
 		}
 	}
@@ -54,11 +80,14 @@
 
 	function handleConfirmDelete() {
 		if (selectedConversation) {
+			const convId = selectedConversation.id;
+			const withForks = deleteWithForks;
 			showDeleteDialog = false;
 
 			setTimeout(() => {
-				conversationsStore.deleteConversation(selectedConversation.id);
-				selectedConversation = null;
+				conversationsStore.deleteConversation(convId, {
+					deleteWithForks: withForks
+				});
 			}, 100); // Wait for animation to finish
 		}
 	}
@@ -110,7 +139,7 @@
 </script>
 
 <ScrollArea class="h-[100vh]">
-	<Sidebar.Header class=" top-0 z-10 gap-6 bg-sidebar/50 px-4 py-4 pb-2 backdrop-blur-lg md:sticky">
+	<Sidebar.Header class=" top-0 z-10 gap-4 bg-sidebar/50 p-4 pb-2 backdrop-blur-lg md:sticky">
 		<a href="#/" onclick={handleMobileSidebarItemClick}>
 			<h1 class="inline-flex items-center gap-1 px-2 text-xl font-semibold">llama.cpp</h1>
 		</a>
@@ -118,7 +147,7 @@
 		<ChatSidebarActions {handleMobileSidebarItemClick} bind:isSearchModeActive bind:searchQuery />
 	</Sidebar.Header>
 
-	<Sidebar.Group class="mt-4 space-y-2 p-0 px-4">
+	<Sidebar.Group class="mt-2 space-y-2 p-0 px-4">
 		{#if (filteredConversations.length > 0 && isSearchModeActive) || !isSearchModeActive}
 			<Sidebar.GroupLabel>
 				{isSearchModeActive ? 'Search results' : 'Conversations'}
@@ -127,15 +156,17 @@
 
 		<Sidebar.GroupContent>
 			<Sidebar.Menu>
-				{#each filteredConversations as conversation (conversation.id)}
-					<Sidebar.MenuItem class="mb-1">
+				{#each conversationTree as { conversation, depth } (conversation.id)}
+					<Sidebar.MenuItem class="mb-1 p-0">
 						<ChatSidebarConversationItem
 							conversation={{
 								id: conversation.id,
 								name: conversation.name,
 								lastModified: conversation.lastModified,
-								currNode: conversation.currNode
+								currNode: conversation.currNode,
+								forkedFromConversationId: conversation.forkedFromConversationId
 							}}
+							{depth}
 							{handleMobileSidebarItemClick}
 							isActive={currentChatId === conversation.id}
 							onSelect={selectConversation}
@@ -146,7 +177,7 @@
 					</Sidebar.MenuItem>
 				{/each}
 
-				{#if filteredConversations.length === 0}
+				{#if conversationTree.length === 0}
 					<div class="px-2 py-4 text-center">
 						<p class="mb-4 p-4 text-sm text-muted-foreground">
 							{searchQuery.length > 0
@@ -177,35 +208,40 @@
 		showDeleteDialog = false;
 		selectedConversation = null;
 	}}
-/>
+>
+	{#if selectedConversationHasDescendants}
+		<div class="flex items-center gap-2 py-2">
+			<Checkbox id="delete-with-forks" bind:checked={deleteWithForks} />
 
-<AlertDialog.Root bind:open={showEditDialog}>
-	<AlertDialog.Content>
-		<AlertDialog.Header>
-			<AlertDialog.Title>Edit Conversation Name</AlertDialog.Title>
-			<AlertDialog.Description>
-				<Input
-					class="mt-4 text-foreground"
-					onkeydown={(e) => {
-						if (e.key === 'Enter') {
-							e.preventDefault();
-							handleConfirmEdit();
-						}
-					}}
-					placeholder="Enter a new name"
-					type="text"
-					bind:value={editedName}
-				/>
-			</AlertDialog.Description>
-		</AlertDialog.Header>
-		<AlertDialog.Footer>
-			<AlertDialog.Cancel
-				onclick={() => {
-					showEditDialog = false;
-					selectedConversation = null;
-				}}>Cancel</AlertDialog.Cancel
-			>
-			<AlertDialog.Action onclick={handleConfirmEdit}>Save</AlertDialog.Action>
-		</AlertDialog.Footer>
-	</AlertDialog.Content>
-</AlertDialog.Root>
+			<Label for="delete-with-forks" class="text-sm">Also delete all forked conversations</Label>
+		</div>
+	{/if}
+</DialogConfirmation>
+
+<DialogConfirmation
+	bind:open={showEditDialog}
+	title="Edit Conversation Name"
+	description=""
+	confirmText="Save"
+	cancelText="Cancel"
+	icon={Pencil}
+	onConfirm={handleConfirmEdit}
+	onCancel={() => {
+		showEditDialog = false;
+		selectedConversation = null;
+	}}
+	onKeydown={(e) => {
+		if (e.key === 'Enter') {
+			e.preventDefault();
+			e.stopImmediatePropagation();
+			handleConfirmEdit();
+		}
+	}}
+>
+	<Input
+		class="text-foreground"
+		placeholder="Enter a new name"
+		type="text"
+		bind:value={editedName}
+	/>
+</DialogConfirmation>

tools/server/webui/src/lib/components/app/chat/ChatSidebar/ChatSidebarActions.svelte

@@ -3,6 +3,9 @@
 	import { KeyboardShortcutInfo } from '$lib/components/app';
 	import { Button } from '$lib/components/ui/button';
 	import { Input } from '$lib/components/ui/input';
+	import { McpLogo } from '$lib/components/app';
+	import { SETTINGS_SECTION_TITLES } from '$lib/constants';
+	import { getChatSettingsDialogContext } from '$lib/contexts';
 
 	interface Props {
 		handleMobileSidebarItemClick: () => void;
@@ -18,6 +21,8 @@
 
 	let searchInput: HTMLInputElement | null = $state(null);
 
+	const chatSettingsDialog = getChatSettingsDialogContext();
+
 	function handleSearchModeDeactivate() {
 		isSearchModeActive = false;
 		searchQuery = '';
@@ -30,7 +35,7 @@
 	});
 </script>
 
-<div class="space-y-0.5">
+<div class="my-1 space-y-1">
 	{#if isSearchModeActive}
 		<div class="relative">
 			<Search class="absolute top-2.5 left-2 h-4 w-4 text-muted-foreground" />
@@ -50,13 +55,14 @@
 		</div>
 	{:else}
 		<Button
-			class="w-full justify-between hover:[&>kbd]:opacity-100"
+			class="w-full justify-between backdrop-blur-none! hover:[&>kbd]:opacity-100"
 			href="?new_chat=true#/"
 			onclick={handleMobileSidebarItemClick}
 			variant="ghost"
 		>
 			<div class="flex items-center gap-2">
 				<SquarePen class="h-4 w-4" />
+
 				New chat
 			</div>
 
@@ -64,7 +70,7 @@
 		</Button>
 
 		<Button
-			class="w-full justify-between hover:[&>kbd]:opacity-100"
+			class="w-full justify-between backdrop-blur-none! hover:[&>kbd]:opacity-100"
 			onclick={() => {
 				isSearchModeActive = true;
 			}}
@@ -72,10 +78,25 @@
 		>
 			<div class="flex items-center gap-2">
 				<Search class="h-4 w-4" />
-				Search conversations
+
+				Search
 			</div>
 
 			<KeyboardShortcutInfo keys={['cmd', 'k']} />
 		</Button>
+
+		<Button
+			class="w-full justify-between backdrop-blur-none! hover:[&>kbd]:opacity-100"
+			onclick={() => {
+				chatSettingsDialog.open(SETTINGS_SECTION_TITLES.MCP);
+			}}
+			variant="ghost"
+		>
+			<div class="flex items-center gap-2">
+				<McpLogo class="h-4 w-4" />
+
+				MCP Servers
+			</div>
+		</Button>
 	{/if}
 </div>

tools/server/webui/src/lib/components/app/chat/ChatSidebar/ChatSidebarConversationItem.svelte

@@ -1,13 +1,23 @@
 <script lang="ts">
-	import { Trash2, Pencil, MoreHorizontal, Download, Loader2, Square } from '@lucide/svelte';
+	import {
+		Trash2,
+		Pencil,
+		MoreHorizontal,
+		Download,
+		Loader2,
+		Square,
+		GitBranch
+	} from '@lucide/svelte';
 	import { DropdownMenuActions } from '$lib/components/app';
 	import * as Tooltip from '$lib/components/ui/tooltip';
+	import { FORK_TREE_DEPTH_PADDING } from '$lib/constants';
 	import { getAllLoadingChats } from '$lib/stores/chat.svelte';
 	import { conversationsStore } from '$lib/stores/conversations.svelte';
 	import { onMount } from 'svelte';
 
 	interface Props {
 		isActive?: boolean;
+		depth?: number;
 		conversation: DatabaseConversation;
 		handleMobileSidebarItemClick?: () => void;
 		onDelete?: (id: string) => void;
@@ -23,7 +33,8 @@
 		onEdit,
 		onSelect,
 		onStop,
-		isActive = false
+		isActive = false,
+		depth = 0
 	}: Props = $props();
 
 	let renderActionsDropdown = $state(false);
@@ -88,14 +99,34 @@
 
 <!-- svelte-ignore a11y_mouse_events_have_key_events -->
 <button
-	class="group flex min-h-9 w-full cursor-pointer items-center justify-between space-x-3 rounded-lg px-3 py-1.5 text-left transition-colors hover:bg-foreground/10 {isActive
+	class="group flex min-h-9 w-full cursor-pointer items-center justify-between space-x-3 rounded-lg py-1.5 text-left transition-colors hover:bg-foreground/10 {isActive
 		? 'bg-foreground/5 text-accent-foreground'
-		: ''}"
+		: ''} px-3"
 	onclick={handleSelect}
 	onmouseover={handleMouseOver}
 	onmouseleave={handleMouseLeave}
 >
-	<div class="flex min-w-0 flex-1 items-center gap-2">
+	<div
+		class="flex min-w-0 flex-1 items-center gap-2"
+		style:padding-left="{depth * FORK_TREE_DEPTH_PADDING}px"
+	>
+		{#if depth > 0}
+			<Tooltip.Root>
+				<Tooltip.Trigger>
+					<a
+						href="#/chat/{conversation.forkedFromConversationId}"
+						class="flex shrink-0 items-center text-muted-foreground transition-colors hover:text-foreground"
+					>
+						<GitBranch class="h-3.5 w-3.5" />
+					</a>
+				</Tooltip.Trigger>
+
+				<Tooltip.Content>
+					<p>See parent conversation</p>
+				</Tooltip.Content>
+			</Tooltip.Root>
+		{/if}
+
 		{#if isLoading}
 			<Tooltip.Root>
 				<Tooltip.Trigger>

tools/server/webui/src/lib/components/app/dialogs/DialogConfirmation.svelte

@@ -1,6 +1,6 @@
 <script lang="ts">
 	import * as AlertDialog from '$lib/components/ui/alert-dialog';
-	import type { Component } from 'svelte';
+	import type { Component, Snippet } from 'svelte';
 	import { KeyboardKey } from '$lib/enums';
 
 	interface Props {
@@ -14,6 +14,7 @@
 		onConfirm: () => void;
 		onCancel: () => void;
 		onKeydown?: (event: KeyboardEvent) => void;
+		children?: Snippet;
 	}
 
 	let {
@@ -26,7 +27,8 @@
 		icon,
 		onConfirm,
 		onCancel,
-		onKeydown
+		onKeydown,
+		children
 	}: Props = $props();
 
 	function handleKeydown(event: KeyboardEvent) {
@@ -60,6 +62,10 @@
 			</AlertDialog.Description>
 		</AlertDialog.Header>
 
+		{#if children}
+			{@render children()}
+		{/if}
+
 		<AlertDialog.Footer>
 			<AlertDialog.Cancel onclick={onCancel}>{cancelText}</AlertDialog.Cancel>
 			<AlertDialog.Action

tools/server/webui/src/lib/components/app/models/ModelId.svelte

@@ -18,7 +18,8 @@
 		showRaw = undefined,
 		aliases,
 		tags,
-		class: className = ''
+		class: className = '',
+		...rest
 	}: Props = $props();
 
 	const badgeClass =
@@ -36,9 +37,9 @@
 </script>
 
 {#if resolvedShowRaw}
-	<TruncatedText class="font-medium {className}" showTooltip={false} text={modelId} />
+	<TruncatedText class="font-medium {className}" showTooltip={false} text={modelId} {...rest} />
 {:else}
-	<span class="flex min-w-0 flex-wrap items-center gap-1 {className}">
+	<span class="flex min-w-0 flex-wrap items-center gap-1 {className}" {...rest}>
 		<span class="min-w-0 truncate font-medium">
 			{#if showOrgName && parsed.orgName && !(aliases && aliases.length > 0)}{parsed.orgName}/{/if}{displayName}
 		</span>

tools/server/webui/src/lib/components/app/models/ModelsSelector.svelte

@@ -271,50 +271,49 @@
 		{#if isRouter}
 			<DropdownMenu.Root bind:open={isOpen} onOpenChange={handleOpenChange}>
 				<DropdownMenu.Trigger
-					disabled={disabled || updating}
-					onclick={(e) => {
-						e.preventDefault();
-						e.stopPropagation();
-					}}
-				>
-					<button
-						type="button"
-						class={cn(
-							`inline-grid cursor-pointer grid-cols-[1fr_auto_1fr] items-center gap-1.5 rounded-sm bg-muted-foreground/10 px-1.5 py-1 text-xs transition hover:text-foreground focus:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-2 disabled:cursor-not-allowed disabled:opacity-60`,
-							!isCurrentModelInCache
-								? 'bg-red-400/10 !text-red-400 hover:bg-red-400/20 hover:text-red-400'
-								: forceForegroundText
+					class={cn(
+						`inline-grid cursor-pointer grid-cols-[1fr_auto_1fr] items-center gap-1.5 rounded-sm bg-muted-foreground/10 px-1.5 py-1 text-xs transition hover:text-foreground focus:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-2 disabled:cursor-not-allowed disabled:opacity-60`,
+						!isCurrentModelInCache
+							? 'bg-red-400/10 !text-red-400 hover:bg-red-400/20 hover:text-red-400'
+							: forceForegroundText
+								? 'text-foreground'
+								: isHighlightedCurrentModelActive
 									? 'text-foreground'
-									: isHighlightedCurrentModelActive
-										? 'text-foreground'
-										: 'text-muted-foreground',
-							isOpen ? 'text-foreground' : ''
-						)}
-						style="max-width: min(calc(100cqw - 9rem), 20rem)"
-						disabled={disabled || updating}
-					>
-						<Package class="h-3.5 w-3.5" />
+									: 'text-muted-foreground',
+						isOpen ? 'text-foreground' : ''
+					)}
+					style="max-width: min(calc(100cqw - 9rem), 20rem)"
+					disabled={disabled || updating}
+				>
+					<Package class="h-3.5 w-3.5" />
 
-						{#if selectedOption}
-							<Tooltip.Root>
-								<Tooltip.Trigger class="min-w-0 overflow-hidden">
-									<ModelId modelId={selectedOption.model} class="min-w-0" showOrgName />
-								</Tooltip.Trigger>
+					{#if selectedOption}
+						<Tooltip.Root>
+							<Tooltip.Trigger>
+								<!-- prevent another nested button element -->
+								{#snippet child({ props })}
+									<ModelId
+										modelId={selectedOption.model}
+										class="min-w-0 overflow-hidden"
+										showOrgName
+										{...props}
+									/>
+								{/snippet}
+							</Tooltip.Trigger>
 
-								<Tooltip.Content>
-									<p class="font-mono">{selectedOption.model}</p>
-								</Tooltip.Content>
-							</Tooltip.Root>
-						{:else}
-							<span class="min-w-0 font-medium">Select model</span>
-						{/if}
+							<Tooltip.Content>
+								<p class="font-mono">{selectedOption.model}</p>
+							</Tooltip.Content>
+						</Tooltip.Root>
+					{:else}
+						<span class="min-w-0 font-medium">Select model</span>
+					{/if}
 
-						{#if updating || isLoadingModel}
-							<Loader2 class="h-3 w-3.5 animate-spin" />
-						{:else}
-							<ChevronDown class="h-3 w-3.5" />
-						{/if}
-					</button>
+					{#if updating || isLoadingModel}
+						<Loader2 class="h-3 w-3.5 animate-spin" />
+					{:else}
+						<ChevronDown class="h-3 w-3.5" />
+					{/if}
 				</DropdownMenu.Trigger>
 
 				<DropdownMenu.Content
@@ -407,8 +406,16 @@
 
 				{#if selectedOption}
 					<Tooltip.Root>
-						<Tooltip.Trigger class="min-w-0 overflow-hidden">
-							<ModelId modelId={selectedOption.model} class="min-w-0" showOrgName />
+						<Tooltip.Trigger>
+							<!-- prevent another nested button element -->
+							{#snippet child({ props })}
+								<ModelId
+									modelId={selectedOption.model}
+									class="min-w-0 overflow-hidden"
+									showOrgName
+									{...props}
+								/>
+							{/snippet}
 						</Tooltip.Trigger>
 
 						<Tooltip.Content>

tools/server/webui/src/lib/constants/context-keys.ts

@@ -0,0 +1,3 @@
+export const CONTEXT_KEY_MESSAGE_EDIT = 'chat-message-edit';
+export const CONTEXT_KEY_CHAT_ACTIONS = 'chat-actions';
+export const CONTEXT_KEY_CHAT_SETTINGS_DIALOG = 'chat-settings-dialog';

tools/server/webui/src/lib/constants/index.ts

@@ -10,6 +10,7 @@ export * from './cache';
 export * from './chat-form';
 export * from './code-blocks';
 export * from './code';
+export * from './context-keys';
 export * from './css-classes';
 export * from './favicon';
 export * from './floating-ui-constraints';

tools/server/webui/src/lib/constants/ui.ts

@@ -1 +1,2 @@
+export const FORK_TREE_DEPTH_PADDING = 8;
 export const SYSTEM_MESSAGE_PLACEHOLDER = 'System message';

tools/server/webui/src/lib/contexts/chat-actions.context.ts

@@ -1,4 +1,5 @@
 import { getContext, setContext } from 'svelte';
+import { CONTEXT_KEY_CHAT_ACTIONS } from '$lib/constants';
 
 export interface ChatActionsContext {
 	copy: (message: DatabaseMessage) => void;
@@ -21,9 +22,13 @@ export interface ChatActionsContext {
 	) => void;
 	regenerateWithBranching: (message: DatabaseMessage, modelOverride?: string) => void;
 	continueAssistantMessage: (message: DatabaseMessage) => void;
+	forkConversation: (
+		message: DatabaseMessage,
+		options: { name: string; includeAttachments: boolean }
+	) => void;
 }
 
-const CHAT_ACTIONS_KEY = Symbol.for('chat-actions');
+const CHAT_ACTIONS_KEY = Symbol.for(CONTEXT_KEY_CHAT_ACTIONS);
 
 export function setChatActionsContext(ctx: ChatActionsContext): ChatActionsContext {
 	return setContext(CHAT_ACTIONS_KEY, ctx);

tools/server/webui/src/lib/contexts/chat-settings-dialog.context.ts

@@ -0,0 +1,19 @@
+import { getContext, setContext } from 'svelte';
+import type { SettingsSectionTitle } from '$lib/constants';
+import { CONTEXT_KEY_CHAT_SETTINGS_DIALOG } from '$lib/constants';
+
+export interface ChatSettingsDialogContext {
+	open: (initialSection?: SettingsSectionTitle) => void;
+}
+
+const CHAT_SETTINGS_DIALOG_KEY = Symbol.for(CONTEXT_KEY_CHAT_SETTINGS_DIALOG);
+
+export function setChatSettingsDialogContext(
+	ctx: ChatSettingsDialogContext
+): ChatSettingsDialogContext {
+	return setContext(CHAT_SETTINGS_DIALOG_KEY, ctx);
+}
+
+export function getChatSettingsDialogContext(): ChatSettingsDialogContext {
+	return getContext(CHAT_SETTINGS_DIALOG_KEY);
+}

tools/server/webui/src/lib/contexts/index.ts

@@ -11,3 +11,9 @@ export {
 	setChatActionsContext,
 	type ChatActionsContext
 } from './chat-actions.context';
+
+export {
+	getChatSettingsDialogContext,
+	setChatSettingsDialogContext,
+	type ChatSettingsDialogContext
+} from './chat-settings-dialog.context';

tools/server/webui/src/lib/contexts/message-edit.context.ts

@@ -1,4 +1,5 @@
 import { getContext, setContext } from 'svelte';
+import { CONTEXT_KEY_MESSAGE_EDIT } from '$lib/constants';
 
 export interface MessageEditState {
 	readonly isEditing: boolean;
@@ -22,7 +23,7 @@ export interface MessageEditActions {
 
 export type MessageEditContext = MessageEditState & MessageEditActions;
 
-const MESSAGE_EDIT_KEY = Symbol.for('chat-message-edit');
+const MESSAGE_EDIT_KEY = Symbol.for(CONTEXT_KEY_MESSAGE_EDIT);
 
 /**
  * Sets the message edit context. Call this in the parent component (ChatMessage.svelte).

tools/server/webui/src/lib/services/database.service.ts

@@ -1,5 +1,6 @@
 import Dexie, { type EntityTable } from 'dexie';
-import { findDescendantMessages, uuid } from '$lib/utils';
+import { findDescendantMessages, uuid, filterByLeafNodeId } from '$lib/utils';
+import type { McpServerOverride } from '$lib/types/database';
 
 class LlamacppDatabase extends Dexie {
 	conversations!: EntityTable<DatabaseConversation, string>;
@@ -173,8 +174,47 @@ export class DatabaseService {
 	 *
 	 * @param id - Conversation ID
 	 */
-	static async deleteConversation(id: string): Promise<void> {
+	static async deleteConversation(
+		id: string,
+		options?: { deleteWithForks?: boolean }
+	): Promise<void> {
 		await db.transaction('rw', [db.conversations, db.messages], async () => {
+			if (options?.deleteWithForks) {
+				// Recursively collect all descendant IDs
+				const idsToDelete: string[] = [];
+				const queue = [id];
+
+				while (queue.length > 0) {
+					const parentId = queue.pop()!;
+					const children = await db.conversations
+						.filter((c) => c.forkedFromConversationId === parentId)
+						.toArray();
+
+					for (const child of children) {
+						idsToDelete.push(child.id);
+						queue.push(child.id);
+					}
+				}
+
+				for (const forkId of idsToDelete) {
+					await db.conversations.delete(forkId);
+					await db.messages.where('convId').equals(forkId).delete();
+				}
+			} else {
+				// Reparent direct children to deleted conv's parent
+				const conv = await db.conversations.get(id);
+				const newParent = conv?.forkedFromConversationId;
+				const directChildren = await db.conversations
+					.filter((c) => c.forkedFromConversationId === id)
+					.toArray();
+
+				for (const child of directChildren) {
+					await db.conversations.update(child.id, {
+						forkedFromConversationId: newParent ?? undefined
+					});
+				}
+			}
+
 			await db.conversations.delete(id);
 			await db.messages.where('convId').equals(id).delete();
 		});
@@ -364,4 +404,88 @@ export class DatabaseService {
 			return { imported: importedCount, skipped: skippedCount };
 		});
 	}
+
+	/**
+	 *
+	 *
+	 * Forking
+	 *
+	 *
+	 */
+
+	/**
+	 * Forks a conversation at a specific message, creating a new conversation
+	 * containing all messages from the root up to (and including) the target message.
+	 *
+	 * @param sourceConvId - The source conversation ID
+	 * @param atMessageId - The message ID to fork at (the new conversation ends here)
+	 * @param options - Fork options (name and whether to include attachments)
+	 * @returns The newly created conversation
+	 */
+	static async forkConversation(
+		sourceConvId: string,
+		atMessageId: string,
+		options: { name: string; includeAttachments: boolean }
+	): Promise<DatabaseConversation> {
+		return await db.transaction('rw', [db.conversations, db.messages], async () => {
+			const sourceConv = await db.conversations.get(sourceConvId);
+			if (!sourceConv) {
+				throw new Error(`Source conversation ${sourceConvId} not found`);
+			}
+
+			const allMessages = await db.messages.where('convId').equals(sourceConvId).toArray();
+
+			const pathMessages = filterByLeafNodeId(allMessages, atMessageId, true) as DatabaseMessage[];
+			if (pathMessages.length === 0) {
+				throw new Error(`Could not resolve message path to ${atMessageId}`);
+			}
+
+			const idMap = new Map<string, string>();
+
+			for (const msg of pathMessages) {
+				idMap.set(msg.id, uuid());
+			}
+
+			const newConvId = uuid();
+			const clonedMessages: DatabaseMessage[] = pathMessages.map((msg) => {
+				const newId = idMap.get(msg.id)!;
+				const newParent = msg.parent ? (idMap.get(msg.parent) ?? null) : null;
+				const newChildren = msg.children
+					.filter((childId: string) => idMap.has(childId))
+					.map((childId: string) => idMap.get(childId)!);
+
+				return {
+					...msg,
+					id: newId,
+					convId: newConvId,
+					parent: newParent,
+					children: newChildren,
+					extra: options.includeAttachments ? msg.extra : undefined
+				};
+			});
+
+			const lastClonedMessage = clonedMessages[clonedMessages.length - 1];
+			const newConv: DatabaseConversation = {
+				id: newConvId,
+				name: options.name,
+				lastModified: Date.now(),
+				currNode: lastClonedMessage.id,
+				forkedFromConversationId: sourceConvId,
+				mcpServerOverrides: sourceConv.mcpServerOverrides
+					? sourceConv.mcpServerOverrides.map((o: McpServerOverride) => ({
+							serverId: o.serverId,
+							enabled: o.enabled
+						}))
+					: undefined
+			};
+
+			await db.conversations.add(newConv);
+
+			for (const msg of clonedMessages) {
+				await db.messages.add(msg);
+			}
+
+			return newConv;
+		});
+	}
 }

tools/server/webui/src/lib/services/mcp.service.ts

@@ -39,7 +39,13 @@ import type {
 	MCPResourceContent,
 	MCPReadResourceResult
 } from '$lib/types';
-import { buildProxiedUrl, throwIfAborted, isAbortError, createBase64DataUrl } from '$lib/utils';
+import {
+	buildProxiedUrl,
+	buildProxiedHeaders,
+	throwIfAborted,
+	isAbortError,
+	createBase64DataUrl
+} from '$lib/utils';
 
 interface ToolResultContentItem {
 	type: string;
@@ -118,7 +124,7 @@ export class MCPService {
 		const requestInit: RequestInit = {};
 
 		if (config.headers) {
-			requestInit.headers = config.headers;
+			requestInit.headers = buildProxiedHeaders(config.headers);
 		}
 
 		if (config.credentials) {

tools/server/webui/src/lib/stores/chat.svelte.ts

@@ -28,6 +28,7 @@ import {
 	filterByLeafNodeId,
 	findDescendantMessages,
 	findLeafNode,
+	findMessageById,
 	isAbortError
 } from '$lib/utils';
 import {
@@ -416,7 +417,7 @@ class ChatStore {
 		if (!activeConv) return false;
 		try {
 			const allMessages = await conversationsStore.getConversationMessages(activeConv.id);
-			const systemMessage = allMessages.find((m) => m.id === messageId);
+			const systemMessage = findMessageById(allMessages, messageId);
 			if (!systemMessage || systemMessage.role !== MessageRole.SYSTEM) return false;
 			const rootMessage = allMessages.find((m) => m.type === 'root' && m.parent === null);
 			if (!rootMessage) return false;
@@ -878,7 +879,7 @@ class ChatStore {
 			const msg = conversationsStore.activeMessages[idx];
 			if (msg.role !== MessageRole.ASSISTANT) return;
 			const allMessages = await conversationsStore.getConversationMessages(activeConv.id);
-			const parentMessage = allMessages.find((m) => m.id === msg.parent);
+			const parentMessage = findMessageById(allMessages, msg.parent);
 			if (!parentMessage) return;
 			this.setChatLoading(activeConv.id, true);
 			this.clearChatStreaming(activeConv.id);
@@ -928,7 +929,7 @@ class ChatStore {
 		if (!activeConv)
 			return { totalCount: 0, userMessages: 0, assistantMessages: 0, messageTypes: [] };
 		const allMessages = await conversationsStore.getConversationMessages(activeConv.id);
-		const messageToDelete = allMessages.find((m) => m.id === messageId);
+		const messageToDelete = findMessageById(allMessages, messageId);
 
 		// For system messages, don't count descendants as they will be preserved (reparented to root)
 		if (messageToDelete?.role === MessageRole.SYSTEM) {
@@ -975,7 +976,7 @@ class ChatStore {
 		if (!activeConv) return;
 		try {
 			const allMessages = await conversationsStore.getConversationMessages(activeConv.id);
-			const messageToDelete = allMessages.find((m) => m.id === messageId);
+			const messageToDelete = findMessageById(allMessages, messageId);
 
 			if (!messageToDelete) return;
 
@@ -1024,7 +1025,7 @@ class ChatStore {
 			this.clearChatStreaming(activeConv.id);
 
 			const allMessages = await conversationsStore.getConversationMessages(activeConv.id);
-			const dbMessage = allMessages.find((m) => m.id === messageId);
+			const dbMessage = findMessageById(allMessages, messageId);
 
 			if (!dbMessage) {
 				this.setChatLoading(activeConv.id, false);
@@ -1265,35 +1266,56 @@ class ChatStore {
 		let result = this.getMessageByIdWithRole(messageId, MessageRole.USER);
 		if (!result) result = this.getMessageByIdWithRole(messageId, MessageRole.SYSTEM);
 		if (!result) return;
-		const { message: msg } = result;
+		const { message: msg, index: idx } = result;
 		try {
 			const allMessages = await conversationsStore.getConversationMessages(activeConv.id);
 			const rootMessage = allMessages.find((m) => m.type === 'root' && m.parent === null);
 			const isFirstUserMessage =
 				msg.role === MessageRole.USER && rootMessage && msg.parent === rootMessage.id;
-			const parentId = msg.parent || rootMessage?.id;
-			if (!parentId) return;
 			const extrasToUse =
 				newExtras !== undefined
 					? JSON.parse(JSON.stringify(newExtras))
 					: msg.extra
 						? JSON.parse(JSON.stringify(msg.extra))
 						: undefined;
-			const newMessage = await DatabaseService.createMessageBranch(
-				{
-					convId: msg.convId,
-					type: msg.type,
-					timestamp: Date.now(),
-					role: msg.role,
+
+			let messageIdForResponse: string;
+
+			const dbMsg = findMessageById(allMessages, msg.id);
+			const hasChildren = dbMsg ? dbMsg.children.length > 0 : msg.children.length > 0;
+
+			if (!hasChildren) {
+				// No responses after this message — update in place instead of branching
+				const updates: Partial<DatabaseMessage> = {
 					content: newContent,
-					toolCalls: msg.toolCalls || '',
-					children: [],
-					extra: extrasToUse,
-					model: msg.model
-				},
-				parentId
-			);
-			await conversationsStore.updateCurrentNode(newMessage.id);
+					timestamp: Date.now(),
+					extra: extrasToUse
+				};
+				await DatabaseService.updateMessage(msg.id, updates);
+				conversationsStore.updateMessageAtIndex(idx, updates);
+				messageIdForResponse = msg.id;
+			} else {
+				// Has children — create a new branch as sibling
+				const parentId = msg.parent || rootMessage?.id;
+				if (!parentId) return;
+				const newMessage = await DatabaseService.createMessageBranch(
+					{
+						convId: msg.convId,
+						type: msg.type,
+						timestamp: Date.now(),
+						role: msg.role,
+						content: newContent,
+						toolCalls: msg.toolCalls || '',
+						children: [],
+						extra: extrasToUse,
+						model: msg.model
+					},
+					parentId
+				);
+				await conversationsStore.updateCurrentNode(newMessage.id);
+				messageIdForResponse = newMessage.id;
+			}
+
 			conversationsStore.updateConversationTimestamp();
 			if (isFirstUserMessage && newContent.trim())
 				await conversationsStore.updateConversationTitleWithConfirmation(
@@ -1301,7 +1323,8 @@ class ChatStore {
 					newContent.trim()
 				);
 			await conversationsStore.refreshActiveMessages();
-			if (msg.role === MessageRole.USER) await this.generateResponseForMessage(newMessage.id);
+			if (msg.role === MessageRole.USER)
+				await this.generateResponseForMessage(messageIdForResponse);
 		} catch (error) {
 			console.error('Failed to edit message with branching:', error);
 		}

tools/server/webui/src/lib/stores/conversations.svelte.ts

@@ -39,6 +39,12 @@ import {
 	MULTIPLE_UNDERSCORE_REGEX,
 	MCP_DEFAULT_ENABLED_LOCALSTORAGE_KEY
 } from '$lib/constants';
+import { SvelteMap, SvelteSet } from 'svelte/reactivity';
+
+export interface ConversationTreeItem {
+	conversation: DatabaseConversation;
+	depth: number;
+}
 
 class ConversationsStore {
 	/**
@@ -300,15 +306,45 @@ class ConversationsStore {
 	 * Deletes a conversation and all its messages
 	 * @param convId - The conversation ID to delete
 	 */
-	async deleteConversation(convId: string): Promise<void> {
+	async deleteConversation(convId: string, options?: { deleteWithForks?: boolean }): Promise<void> {
 		try {
-			await DatabaseService.deleteConversation(convId);
+			await DatabaseService.deleteConversation(convId, options);
 
-			this.conversations = this.conversations.filter((c) => c.id !== convId);
+			if (options?.deleteWithForks) {
+				// Collect all descendants recursively
+				const idsToRemove = new SvelteSet([convId]);
+				const queue = [convId];
+				while (queue.length > 0) {
+					const parentId = queue.pop()!;
+					for (const c of this.conversations) {
+						if (c.forkedFromConversationId === parentId && !idsToRemove.has(c.id)) {
+							idsToRemove.add(c.id);
+							queue.push(c.id);
+						}
+					}
+				}
+				this.conversations = this.conversations.filter((c) => !idsToRemove.has(c.id));
 
-			if (this.activeConversation?.id === convId) {
-				this.clearActiveConversation();
-				await goto(`?new_chat=true#/`);
+				if (this.activeConversation && idsToRemove.has(this.activeConversation.id)) {
+					this.clearActiveConversation();
+					await goto(`?new_chat=true#/`);
+				}
+			} else {
+				// Reparent direct children to deleted conv's parent (or promote to top-level)
+				const deletedConv = this.conversations.find((c) => c.id === convId);
+				const newParent = deletedConv?.forkedFromConversationId;
+				this.conversations = this.conversations
+					.filter((c) => c.id !== convId)
+					.map((c) =>
+						c.forkedFromConversationId === convId
+							? { ...c, forkedFromConversationId: newParent }
+							: c
+					);
+
+				if (this.activeConversation?.id === convId) {
+					this.clearActiveConversation();
+					await goto(`?new_chat=true#/`);
+				}
 			}
 		} catch (error) {
 			console.error('Failed to delete conversation:', error);
@@ -658,6 +694,42 @@ class ConversationsStore {
 		this.saveMcpDefaults();
 	}
 
+	/**
+	 * Forks a conversation at a specific message, creating a new conversation
+	 * containing messages from root up to the target message, then navigates to it.
+	 *
+	 * @param messageId - The message ID to fork at
+	 * @param options - Fork options (name and whether to include attachments)
+	 * @returns The new conversation ID, or null if fork failed
+	 */
+	async forkConversation(
+		messageId: string,
+		options: { name: string; includeAttachments: boolean }
+	): Promise<string | null> {
+		if (!this.activeConversation) return null;
+
+		try {
+			const newConv = await DatabaseService.forkConversation(
+				this.activeConversation.id,
+				messageId,
+				options
+			);
+
+			this.conversations = [newConv, ...this.conversations];
+
+			await goto(`#/chat/${newConv.id}`);
+
+			toast.success('Conversation forked');
+
+			return newConv.id;
+		} catch (error) {
+			console.error('Failed to fork conversation:', error);
+			toast.error('Failed to fork conversation');
+
+			return null;
+		}
+	}
+
 	/**
 	 *
 	 *
@@ -830,3 +902,53 @@ export const conversations = () => conversationsStore.conversations;
 export const activeConversation = () => conversationsStore.activeConversation;
 export const activeMessages = () => conversationsStore.activeMessages;
 export const isConversationsInitialized = () => conversationsStore.isInitialized;
+
+/**
+ * Builds a flat tree of conversations with depth levels for nested forks.
+ * Accepts a pre-filtered list so search filtering stays in the component.
+ */
+export function buildConversationTree(convs: DatabaseConversation[]): ConversationTreeItem[] {
+	const childrenByParent = new SvelteMap<string, DatabaseConversation[]>();
+	const forkIds = new SvelteSet<string>();
+
+	for (const conv of convs) {
+		if (conv.forkedFromConversationId) {
+			forkIds.add(conv.id);
+
+			const siblings = childrenByParent.get(conv.forkedFromConversationId) || [];
+
+			siblings.push(conv);
+			childrenByParent.set(conv.forkedFromConversationId, siblings);
+		}
+	}
+
+	const result: ConversationTreeItem[] = [];
+	const visited = new SvelteSet<string>();
+
+	function walk(conv: DatabaseConversation, depth: number) {
+		visited.add(conv.id);
+		result.push({ conversation: conv, depth });
+
+		const children = childrenByParent.get(conv.id);
+		if (children) {
+			children.sort((a, b) => b.lastModified - a.lastModified);
+
+			for (const child of children) {
+				walk(child, depth + 1);
+			}
+		}
+	}
+
+	const roots = convs.filter((c) => !forkIds.has(c.id));
+	for (const root of roots) {
+		walk(root, 0);
+	}
+
+	for (const conv of convs) {
+		if (!visited.has(conv.id)) {
+			walk(conv, 1);
+		}
+	}
+
+	return result;
+}

tools/server/webui/src/lib/types/database.d.ts

@@ -12,6 +12,7 @@ export interface DatabaseConversation {
 	lastModified: number;
 	name: string;
 	mcpServerOverrides?: McpServerOverride[];
+	forkedFromConversationId?: string;
 }
 
 export interface DatabaseMessageExtraAudioFile {

tools/server/webui/src/lib/utils/branching.ts

@@ -17,6 +17,17 @@
 
 import { MessageRole } from '$lib/enums';
 
+/**
+ * Finds a message by its ID in the given messages array.
+ */
+export function findMessageById(
+	messages: readonly DatabaseMessage[],
+	id: string | null | undefined
+): DatabaseMessage | undefined {
+	if (!id) return undefined;
+	return messages.find((m) => m.id === id);
+}
+
 /**
  * Filters messages to get the conversation path from root to a specific leaf node.
  * If the leafNodeId doesn't exist, returns the path with the latest timestamp.

tools/server/webui/src/lib/utils/cors-proxy.ts

@@ -19,6 +19,21 @@ export function buildProxiedUrl(targetUrl: string): URL {
 	return proxyUrl;
 }
 
+/**
+ * Wrap original headers for proxying through the CORS proxy. This avoids issues with duplicated llama.cpp-specific and target headers when using the CORS proxy.
+ * @param headers - The original headers to be proxied to target
+ * @returns List of "wrapped" headers to be sent to the CORS proxy
+ */
+export function buildProxiedHeaders(headers: Record<string, string>): Record<string, string> {
+	const proxiedHeaders: Record<string, string> = {};
+
+	for (const [key, value] of Object.entries(headers)) {
+		proxiedHeaders[`X-Proxy-Header-${key}`] = value;
+	}
+
+	return proxiedHeaders;
+}
+
 /**
  * Get a proxied URL string for use in fetch requests.
  * @param targetUrl - The original URL to proxy

tools/server/webui/src/lib/utils/index.ts

@@ -22,6 +22,7 @@ export { default as autoResizeTextarea } from './autoresize-textarea';
 // Branching utilities
 export {
 	filterByLeafNodeId,
+	findMessageById,
 	findLeafNode,
 	findDescendantMessages,
 	getMessageSiblings,
@@ -38,7 +39,7 @@ export { highlightCode, detectIncompleteCodeBlock, type IncompleteCodeBlock } fr
 export { setConfigValue, getConfigValue, configToParameterRecord } from './config-helpers';
 
 // CORS Proxy
-export { buildProxiedUrl, getProxiedUrlString } from './cors-proxy';
+export { buildProxiedUrl, getProxiedUrlString, buildProxiedHeaders } from './cors-proxy';
 
 // Conversation utilities
 export { createMessageCountMap, getMessageCount } from './conversation-utils';

tools/server/webui/src/routes/+layout.svelte

@@ -4,7 +4,11 @@
 	import { browser } from '$app/environment';
 	import { page } from '$app/state';
 	import { untrack } from 'svelte';
-	import { ChatSidebar, DialogConversationTitleUpdate } from '$lib/components/app';
+	import {
+		ChatSidebar,
+		DialogConversationTitleUpdate,
+		DialogChatSettings
+	} from '$lib/components/app';
 	import { isLoading } from '$lib/stores/chat.svelte';
 	import { conversationsStore, activeMessages } from '$lib/stores/conversations.svelte';
 	import * as Sidebar from '$lib/components/ui/sidebar/index.js';
@@ -17,8 +21,10 @@
 	import { modelsStore } from '$lib/stores/models.svelte';
 	import { mcpStore } from '$lib/stores/mcp.svelte';
 	import { TOOLTIP_DELAY_DURATION } from '$lib/constants';
+	import type { SettingsSectionTitle } from '$lib/constants';
 	import { KeyboardKey } from '$lib/enums';
 	import { IsMobile } from '$lib/hooks/is-mobile.svelte';
+	import { setChatSettingsDialogContext } from '$lib/contexts';
 
 	let { children } = $props();
 
@@ -42,6 +48,16 @@
 	let titleUpdateNewTitle = $state('');
 	let titleUpdateResolve: ((value: boolean) => void) | null = null;
 
+	let chatSettingsDialogOpen = $state(false);
+	let chatSettingsDialogInitialSection = $state<SettingsSectionTitle | undefined>(undefined);
+
+	setChatSettingsDialogContext({
+		open: (initialSection?: SettingsSectionTitle) => {
+			chatSettingsDialogInitialSection = initialSection;
+			chatSettingsDialogOpen = true;
+		}
+	});
+
 	// Global keyboard shortcuts
 	function handleKeydown(event: KeyboardEvent) {
 		const isCtrlOrCmd = event.ctrlKey || event.metaKey;
@@ -213,6 +229,12 @@
 
 	<Toaster richColors />
 
+	<DialogChatSettings
+		open={chatSettingsDialogOpen}
+		onOpenChange={(open) => (chatSettingsDialogOpen = open)}
+		initialSection={chatSettingsDialogInitialSection}
+	/>
+
 	<DialogConversationTitleUpdate
 		bind:open={titleUpdateDialogOpen}
 		currentTitle={titleUpdateCurrentTitle}

tools/server/webui/tests/stories/ChatSidebar.stories.svelte

@@ -73,7 +73,7 @@
 			conversationsStore.conversations = mockConversations;
 		}, 0));
 		
-		const searchTrigger = screen.getByText('Search conversations');
+		const searchTrigger = screen.getByText('Search');
 		userEvent.click(searchTrigger);
 	}}
 >

vendor/cpp-httplib/CMakeLists.txt

@@ -39,7 +39,7 @@ if (LLAMA_BUILD_BORINGSSL)
     set(FIPS OFF CACHE BOOL "Enable FIPS (BoringSSL)")
 
     set(BORINGSSL_GIT "https://boringssl.googlesource.com/boringssl" CACHE STRING "BoringSSL git repository")
-    set(BORINGSSL_VERSION "0.20260211.0" CACHE STRING "BoringSSL version")
+    set(BORINGSSL_VERSION "0.20260327.0" CACHE STRING "BoringSSL version")
 
     message(STATUS "Fetching BoringSSL version ${BORINGSSL_VERSION}")