llama: add token ID tracking to KV cell (#27762 )

* kv: track token id * rm get_prev_tokens, move it to the main pr * nits * add get_prev_tokens
ui: Move Settings and MCP Servers routes to dialog-based views (#27744 )
2026-08-27 02:27:34 +02:00 · 2026-08-26 23:34:28 +02:00 · 2026-08-26 21:07:24 +02:00 · 2026-08-26 20:19:19 +02:00 · 2026-08-26 19:02:06 +03:00 · 2026-08-26 16:49:32 +02:00
58 changed files with 1349 additions and 900 deletions
@@ -64,7 +64,7 @@ jobs:
    needs: create_tag
    uses: ./.github/workflows/ui-build.yml
    with:
-      hf_ui_version: ${{ needs.create_tag.outputs.source_tag }}
+      ui_version: ${{ needs.create_tag.outputs.source_tag }}

  prepare_matrices:
    name: Prepare Docker matrices
@@ -162,7 +162,7 @@ jobs:
        if: ${{ matrix.config.prebuilt_ui == true }}
        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8
        with:
-          name: ui-build
+          name: llama-ui.zip
          path: tools/ui/dist

      - name: Set up QEMU
@@ -61,31 +61,8 @@ jobs:
            echo "should_release=false" >> $GITHUB_OUTPUT
          fi

-  get-version:
-    runs-on: ubuntu-slim
-    outputs:
-      ui_version: ${{ steps.version.outputs.ui_version }}
-    steps:
-      - uses: actions/checkout@v6
-        with:
-          fetch-depth: 0
-      - id: version
-        run: |
-          # Resolve UI version: BUILD_NUMBER from cmake/build-info.cmake > git hash + epoch > fallback
-          version=""
-          if grep -q "BUILD_NUMBER" cmake/build-info.cmake; then
-            build_number=$(grep "set(BUILD_NUMBER" cmake/build-info.cmake | grep -oP '\d+')
-            if [ -n "$build_number" ] && [ "$build_number" -gt 0 ]; then
-              version="b${build_number}"
-            fi
-          fi
-          if [ -z "$version" ]; then
-            version=$(git rev-parse --short HEAD)-$(date +%s)
-          fi
-          echo "ui_version=${version}" >> $GITHUB_OUTPUT
-
  macos-cpu:
-    needs: [check-release, get-version]
+    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
    strategy:
      matrix:
@@ -119,12 +96,11 @@ jobs:
        with:
          fetch-depth: 0

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
+      - name: Download UI build
+        uses: actions/download-artifact@v7
        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          name: llama-ui.zip
+          path: tools/ui/dist

      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
@@ -141,7 +117,6 @@ jobs:
            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
            -DLLAMA_FATAL_WARNINGS=ON \
            -DLLAMA_BUILD_BORINGSSL=ON \
-            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
            ${{ env.CMAKE_ARGS }}
          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)

@@ -167,7 +142,7 @@ jobs:
          key: release-${{ matrix.os }}-${{ matrix.arch }}

  ubuntu-cpu:
-    needs: [check-release, get-version]
+    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
    strategy:
      matrix:
@@ -191,12 +166,11 @@ jobs:
        with:
          fetch-depth: 0

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
+      - name: Download UI build
+        uses: actions/download-artifact@v7
        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          name: llama-ui.zip
+          path: tools/ui/dist

      - name: Dependencies
        id: depends
@@ -227,7 +201,6 @@ jobs:
            -DGGML_NATIVE=OFF \
            -DGGML_CPU_ALL_VARIANTS=ON \
            -DLLAMA_FATAL_WARNINGS=ON \
-            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
            ${{ env.CMAKE_ARGS }}
          cmake --build build --config Release -j $(nproc)

@@ -254,7 +227,7 @@ jobs:
          key: release-${{ matrix.os }}-cpu

  ubuntu-vulkan:
-    needs: [check-release, get-version]
+    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

    strategy:
@@ -277,12 +250,11 @@ jobs:
        with:
          fetch-depth: 0

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
+      - name: Download UI build
+        uses: actions/download-artifact@v7
        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          name: llama-ui.zip
+          path: tools/ui/dist

      - name: Dependencies
        id: depends
@@ -314,7 +286,6 @@ jobs:
            -DGGML_NATIVE=OFF \
            -DGGML_CPU_ALL_VARIANTS=ON \
            -DGGML_VULKAN=ON \
-            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
            ${{ env.CMAKE_ARGS }}
          cmake --build build --config Release -j $(nproc)

@@ -340,7 +311,7 @@ jobs:
          key: release-${{ matrix.os }}-vulkan

  android-arm64:
-    needs: [check-release, get-version]
+    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

    runs-on: ubuntu-latest
@@ -358,12 +329,11 @@ jobs:
        with:
          fetch-depth: 0

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
+      - name: Download UI build
+        uses: actions/download-artifact@v7
        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          name: llama-ui.zip
+          path: tools/ui/dist

      - name: Set up JDK
        uses: actions/setup-java@v5
@@ -407,7 +377,6 @@ jobs:
            -DLLAMA_FATAL_WARNINGS=ON \
            -DGGML_OPENMP=OFF \
            -DLLAMA_BUILD_BORINGSSL=ON \
-            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
            ${{ env.CMAKE_ARGS }}
          cmake --build build --config Release -j $(nproc)

@@ -433,7 +402,7 @@ jobs:
          name: llama-bin-android-arm64.tar.gz

  ubuntu-24-openvino:
-    needs: [check-release, get-version]
+    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

    runs-on: ubuntu-24.04
@@ -460,12 +429,11 @@ jobs:
        with:
          fetch-depth: 0

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
+      - name: Download UI build
+        uses: actions/download-artifact@v7
        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          name: llama-ui.zip
+          path: tools/ui/dist

      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
@@ -508,7 +476,6 @@ jobs:
            -DGGML_OPENVINO=ON \
            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
-            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
            ${{ env.CMAKE_ARGS }}
          cmake --build build/ReleaseOV --config Release --parallel

@@ -552,7 +519,7 @@ jobs:
          key: release-ubuntu-24.04-openvino-release-no-preset-v1

  windows-openvino:
-    needs: [check-release]
+    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

    runs-on: windows-2022
@@ -577,12 +544,11 @@ jobs:
        with:
            fetch-depth: 0

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
+      - name: Download UI build
+        uses: actions/download-artifact@v7
        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          name: llama-ui.zip
+          path: tools/ui/dist

      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
@@ -682,7 +648,7 @@ jobs:

  windows-cpu:
    name: windows-cpu / ${{ matrix.arch }}
-    needs: [check-release]
+    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

    runs-on: windows-2025-vs2026
@@ -702,12 +668,11 @@ jobs:
        with:
          fetch-depth: 0

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
+      - name: Download UI build
+        uses: actions/download-artifact@v7
        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          name: llama-ui.zip
+          path: tools/ui/dist

      - name: Install Ninja
        run: |
@@ -749,8 +714,10 @@ jobs:
        with:
          key: release-windows-2025-vs2026-${{ matrix.arch }}-cpu

+  # TODO: build only the ggml-hip backend like the other windows backend jobs
+  #       (windows-cuda, windows-sycl), then drop the ui-build dependency
  windows-rocm:
-    needs: [check-release]
+    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

    runs-on: windows-2022
@@ -769,6 +736,12 @@ jobs:
        with:
          fetch-depth: 0

+      - name: Download UI build
+        uses: actions/download-artifact@v7
+        with:
+          name: llama-ui.zip
+          path: tools/ui/dist
+
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
        with:
@@ -879,6 +852,8 @@ jobs:
        with:
          key: windows-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}

+  # note: builds only the backend library - llama-server (with the embedded UI)
+  #       is injected from the windows-cpu zip during the release "Merge artifacts" step
  windows:
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -909,13 +884,6 @@ jobs:
        id: checkout
        uses: actions/checkout@v6

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
-        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
-
      - name: Install Vulkan SDK
        id: get_vulkan
        if: ${{ matrix.backend == 'vulkan' }}
@@ -978,6 +946,8 @@ jobs:
          path: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
          name: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip

+  # note: builds only the ggml-cuda backend - llama-server is injected from the
+  #       windows-cpu zip during the release "Merge artifacts" step
  windows-cuda:
    name: windows-cuda (${{ matrix.cuda }}, ${{ matrix.arch }})
    needs: [check-release]
@@ -1006,13 +976,6 @@ jobs:
        id: checkout
        uses: actions/checkout@v6

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
-        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
-
      - name: Install Cuda Toolkit
        uses: ./.github/actions/windows-setup-cuda
        with:
@@ -1084,6 +1047,8 @@ jobs:
        with:
          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}

+  # note: builds only the ggml-sycl backend - llama-server is injected from the
+  #       windows-cpu zip during the release "Merge artifacts" step
  windows-sycl:
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1118,13 +1083,6 @@ jobs:
          Expand-Archive -Path "level-zero-win-sdk.zip" -DestinationPath "C:/level-zero-sdk" -Force
          "LEVEL_ZERO_V1_SDK_PATH=C:/level-zero-sdk" | Out-File -FilePath $env:GITHUB_ENV -Append

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
-        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
-
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
        with:
@@ -1195,7 +1153,7 @@ jobs:
          key: release-windows-2022-x64-sycl

  ubuntu-24-sycl:
-    needs: [check-release]
+    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

    strategy:
@@ -1237,12 +1195,11 @@ jobs:
          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero-devel_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero-devel.deb
          sudo apt-get install -y ./level-zero.deb ./level-zero-devel.deb

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
+      - name: Download UI build
+        uses: actions/download-artifact@v7
        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          name: llama-ui.zip
+          path: tools/ui/dist

      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
@@ -1288,7 +1245,7 @@ jobs:
          key: release-ubuntu-24.04-sycl-${{ matrix.build }}

  ubuntu-24-rocm:
-    needs: [check-release, get-version]
+    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

    runs-on: ubuntu-24.04
@@ -1310,12 +1267,11 @@ jobs:
        with:
          fetch-depth: 0

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
+      - name: Download UI build
+        uses: actions/download-artifact@v7
        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          name: llama-ui.zip
+          path: tools/ui/dist

      - name: Free up disk space
        uses: ggml-org/free-disk-space@v1.3.1
@@ -1388,7 +1344,6 @@ jobs:
            -DGPU_TARGETS="${{ matrix.gpu_targets }}" \
            -DGGML_HIP=ON \
            -DHIP_PLATFORM=amd \
-            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
            ${{ env.CMAKE_ARGS }}
          cmake --build build --config Release -j $(nproc)

@@ -1417,7 +1372,7 @@ jobs:
          key: release-ubuntu-24.04-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}

  ios-xcode:
-    needs: [check-release, get-version]
+    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
    runs-on: macos-26

@@ -1445,8 +1400,7 @@ jobs:
            -DLLAMA_BUILD_SERVER=OFF \
            -DCMAKE_SYSTEM_NAME=iOS \
            -DCMAKE_OSX_DEPLOYMENT_TARGET=16.0 \
-            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml \
-            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }}
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO

      - name: xcodebuild for swift package
@@ -1569,11 +1523,9 @@ jobs:
 #          name: llama-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}${{ matrix.use_acl_graph == 'on' && '-aclgraph' || '' }}.tar.gz

  ui-build:
-    needs: [check-release, get-version]
+    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
    uses: ./.github/workflows/ui-build.yml
-    with:
-      hf_ui_version: ${{ needs.get-version.outputs.ui_version }}

  release:
    if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
@@ -1588,7 +1540,6 @@ jobs:
    runs-on: ubuntu-slim

    needs:
-      - get-version
      - windows
      - windows-cpu
      - windows-cuda
@@ -1628,24 +1579,27 @@ jobs:
          path: ./artifact
          merge-multiple: true

-      - name: Move artifacts
+      - name: Merge artifacts
        id: move_artifacts
        run: |
          mkdir -p release

-          echo "Adding CPU backend files to existing zips..."
+          # the windows-cpu zip contains the full toolset (llama-server with the embedded
+          # UI, ggml-cpu) - inject it into the other windows zips so that every archive
+          # ships the same binaries, only with a different backend library on top
+          echo "Injecting windows-cpu binaries (llama-server + CPU backend) into the backend zips..."
          for arch in x64 arm64; do
            cpu_zip="artifact/llama-bin-win-cpu-${arch}.zip"
            temp_dir=$(mktemp -d)
-            echo "Extracting CPU backend for $arch..."
+            echo "Extracting windows-cpu-${arch} package..."
            unzip "$cpu_zip" -d "$temp_dir"

-            echo "Adding CPU files to $arch zips..."
+            echo "Merging into $arch zips..."
            for target_zip in artifact/llama-bin-win-*-${arch}.zip; do
              if [[ "$target_zip" == "$cpu_zip" ]]; then
                continue
              fi
-              echo "Adding CPU backend to $(basename "$target_zip")"
+              echo "Injecting into $(basename "$target_zip")"
              realpath_target_zip=$(realpath "$target_zip")
              (cd "$temp_dir" && zip -r "$realpath_target_zip" .)
            done
@@ -1669,7 +1623,7 @@ jobs:
        id: download_ui
        uses: actions/download-artifact@v7
        with:
-          name: ui-build
+          name: llama-ui.zip
          path: ./ui-dist

      - name: Package UI
@@ -73,13 +73,6 @@ jobs:
          fetch-depth: 0
          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
-        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
-
      - name: Build
        id: cmake_build
        run: |
@@ -31,6 +31,6 @@ jobs:
      - name: Upload built UI
        uses: actions/upload-artifact@v6
        with:
-          name: ui-build
+          name: llama-ui.zip
          path: tools/ui/dist/
          retention-days: 1
@@ -3,8 +3,8 @@ name: UI Build
 on:
  workflow_call:
    inputs:
-      hf_ui_version:
-        description: 'Version string for version.json (e.g. 12345)'
+      ui_version:
+        description: 'Version string embedded in build.json (e.g. b1234); defaults to b<commit-count>'
        required: false
        type: string

@@ -17,6 +17,17 @@ jobs:
    steps:
      - name: Checkout code
        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+
+      - name: Resolve UI version
+        id: version
+        run: |
+          version="${{ inputs.ui_version }}"
+          if [ -z "$version" ]; then
+            version="b$(git rev-list --count HEAD)"
+          fi
+          echo "ui_version=${version}" >> $GITHUB_OUTPUT

      - name: Setup Node.js
        uses: actions/setup-node@v6
@@ -31,8 +42,7 @@ jobs:

      - name: Build application
        env:
-          HF_UI_VERSION: ${{ inputs.hf_ui_version || '' }}
-          LLAMA_BUILD_NUMBER: ${{ inputs.hf_ui_version || 'b0000' }}
+          LLAMA_BUILD_NUMBER: ${{ steps.version.outputs.ui_version }}
        run: npm run build
        working-directory: tools/ui

@@ -43,6 +53,6 @@ jobs:
      - name: Upload built UI
        uses: actions/upload-artifact@v6
        with:
-          name: ui-build
+          name: llama-ui.zip
          path: tools/ui/dist/
          retention-days: 1
@@ -37,7 +37,7 @@ jobs:
      - name: Download UI build artifact
        uses: actions/download-artifact@v7
        with:
-          name: ui-build
+          name: llama-ui.zip
          path: tools/ui/dist/

      - name: Create distribution archive
@@ -64,7 +64,7 @@ jobs:
      - name: Download built UI artifacts
        uses: actions/download-artifact@v6
        with:
-          name: ui-build
+          name: llama-ui.zip
          path: tools/ui/dist/

      - name: Run type checking
@@ -106,7 +106,7 @@ jobs:
      - name: Download built UI artifacts
        uses: actions/download-artifact@v6
        with:
-          name: ui-build
+          name: llama-ui.zip
          path: tools/ui/dist/

      - name: Build Storybook
@@ -63,7 +63,7 @@ jobs:
      - name: Download built UI artifacts
        uses: actions/download-artifact@v6
        with:
-          name: ui-build
+          name: llama-ui.zip
          path: tools/ui/dist/

      - name: Install dependencies
@@ -126,7 +126,7 @@ jobs:
      - name: Download built UI artifacts (reuses ui-build)
        uses: actions/download-artifact@v6
        with:
-          name: ui-build
+          name: llama-ui.zip
          path: tools/ui/dist/

      - name: Install Playwright browsers
@@ -134,8 +134,8 @@ option(LLAMA_BUILD_TOOLS     "llama: build tools"
 option(LLAMA_BUILD_EXAMPLES  "llama: build examples"                                                             ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER    "llama: build server example"                                                       ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_APP       "llama: build the unified binary"                                                   ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_UI        "llama: build the embedded Web UI for server"                                       ON)
-option(LLAMA_USE_PREBUILT_UI "llama: use prebuilt UI from HF Bucket when available (requires LLAMA_BUILD_UI=ON)" ON)
+option(LLAMA_BUILD_UI        "llama: build the embedded Web UI for server"                                       OFF)
+option(LLAMA_USE_PREBUILT_UI "llama: use prebuilt UI from HF Bucket when available" ON)

 option(LLAMA_TOOLS_INSTALL "llama: install tools" ${LLAMA_TOOLS_INSTALL_DEFAULT})
 option(LLAMA_TESTS_INSTALL "llama: install tests" ON)
@@ -35,8 +35,8 @@ Legend:
 |                              COS | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
-|               CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|               CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                           CUMSUM | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             DIAG | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
@@ -19292,10 +19292,10 @@
 "Vulkan0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=8,nr23=[4,1],kv=4096,nb=512,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_K=f16,type_V=f16,permute=[0,1,2,3]","support","1","yes","Vulkan"
 "Vulkan0","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[6,1],kv=16384,nb=512,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_K=f16,type_V=f16,permute=[0,1,2,3]","support","1","yes","Vulkan"
 "Vulkan0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=8,nr23=[4,1],kv=16384,nb=512,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_K=f16,type_V=f16,permute=[0,1,2,3]","support","1","yes","Vulkan"
-"Vulkan0","CROSS_ENTROPY_LOSS","type=f32,ne=[10,5,4,3]","support","0","no","Vulkan"
-"Vulkan0","CROSS_ENTROPY_LOSS","type=f32,ne=[30000,1,1,1]","support","0","no","Vulkan"
-"Vulkan0","CROSS_ENTROPY_LOSS_BACK","type=f32,ne=[10,5,4,3]","support","0","no","Vulkan"
-"Vulkan0","CROSS_ENTROPY_LOSS_BACK","type=f32,ne=[30000,1,1,1]","support","0","no","Vulkan"
+"Vulkan0","CROSS_ENTROPY_LOSS","type=f32,ne=[10,5,4,3]","support","1","yes","Vulkan"
+"Vulkan0","CROSS_ENTROPY_LOSS","type=f32,ne=[30000,1,1,1]","support","1","yes","Vulkan"
+"Vulkan0","CROSS_ENTROPY_LOSS_BACK","type=f32,ne=[10,5,4,3]","support","1","yes","Vulkan"
+"Vulkan0","CROSS_ENTROPY_LOSS_BACK","type=f32,ne=[30000,1,1,1]","support","1","yes","Vulkan"
 "Vulkan0","OPT_STEP_ADAMW","type=f32,ne=[10,5,4,3]","support","1","yes","Vulkan"
 "Vulkan0","OPT_STEP_SGD","type=f32,ne=[10,5,4,3]","support","1","yes","Vulkan"
 "Vulkan0","GATED_DELTA_NET","type=f32,head_count=32,head_size=128,n_seq_tokens=1,n_seqs=1,v_repeat=1,permuted=0,kda=0,K=1","support","1","yes","Vulkan"
@@ -6,8 +6,8 @@
 extern "C" {
 #endif

-#define RPC_PROTO_MAJOR_VERSION    5
-#define RPC_PROTO_MINOR_VERSION    1
+#define RPC_PROTO_MAJOR_VERSION    6
+#define RPC_PROTO_MINOR_VERSION    0
 #define RPC_PROTO_PATCH_VERSION    0

 #ifdef  __cplusplus
@@ -9,6 +9,9 @@
 #include <optional>
 #include <string>
 #include <vector>
+#include <queue>
+#include <condition_variable>
+#include <future>
 #include <memory>
 #include <mutex>
 #include <unordered_map>
@@ -17,6 +20,8 @@
 #include <fstream>
 #include <filesystem>
 #include <algorithm>
+#include <atomic>
+#include <thread>

 static const char * RPC_DEBUG = std::getenv("GGML_RPC_DEBUG");

@@ -72,6 +77,7 @@ enum rpc_cmd {
    RPC_CMD_DEVICE_COUNT,
    RPC_CMD_GRAPH_RECOMPUTE,
    RPC_CMD_MEMSET_TENSOR,
+    RPC_CMD_NONE,
    RPC_CMD_COUNT,
 };

@@ -223,24 +229,24 @@ struct ggml_backend_rpc_buffer_type_context {
    size_t      max_size;
 };

+class rpc_dispatcher;
 struct ggml_backend_rpc_context {
-    std::string endpoint;
-    uint32_t    device;
-    std::string name;
+    std::shared_ptr<rpc_dispatcher> dispatcher;
+    uint32_t                        device;
+    std::string                     name;
 };

 struct ggml_backend_rpc_buffer_context {
-    std::shared_ptr<socket_t> sock;
-    void * base_ptr;
-    uint64_t remote_ptr;
+    std::shared_ptr<rpc_dispatcher>   dispatcher;
+    void                            * base_ptr;
+    uint64_t                          remote_ptr;
 };

 // RPC helper functions

 // Computes FNV-1a hash of the data
-static uint64_t fnv_hash(const uint8_t * data, size_t len) {
+static uint64_t fnv_hash(const uint8_t * data, size_t len, uint64_t hash = 0xcbf29ce484222325ULL) {
    const uint64_t fnv_prime = 0x100000001b3ULL;
-    uint64_t hash = 0xcbf29ce484222325ULL;

    for (size_t i = 0; i < len; ++i) {
        hash ^= data[i];
@@ -357,44 +363,248 @@ static bool negotiate_hello(const std::shared_ptr<socket_t> & sock) {
    return true;
 }

-static std::shared_ptr<socket_t> get_socket(const std::string & endpoint) {
-    static std::mutex mutex;
-    std::lock_guard<std::mutex> lock(mutex);
-    static std::unordered_map<std::string, std::weak_ptr<socket_t>> sockets;
+template <typename T>
+class message_queue {
+public:
+    message_queue() {}

-    auto it = sockets.find(endpoint);
-    if (it != sockets.end()) {
-        if (auto sock = it->second.lock()) {
-            return sock;
+    bool push(const T &value) {
+        std::unique_lock<std::mutex> lock(mutex);
+        if (interrupted) {
+            return false;
        }
+        queue.push(value);
+        cvar.notify_all();
+        return true;
    }
+
+    bool pop(T* out) {
+        std::unique_lock<std::mutex> lock(mutex);
+        cvar.wait(lock, [this] { return !queue.empty() || interrupted; });
+        if (interrupted) {
+            return false;
+        }
+        *out = queue.front();
+        queue.pop();
+        return true;
+    }
+
+    void interrupt() {
+        std::unique_lock<std::mutex> lock(mutex);
+        interrupted = true;
+        lock.unlock();
+        cvar.notify_all();
+    }
+
+private:
+    bool interrupted = false;
+    std::queue<T> queue;
+    std::mutex mutex;
+    std::condition_variable cvar;
+};
+
+class rpc_dispatcher {
+public:
+    rpc_dispatcher() {
+    }
+
+    void send(enum rpc_cmd cmd, std::shared_ptr<const void> input, size_t input_size);
+    void send(enum rpc_cmd cmd, std::shared_ptr<const void> input, size_t input_size, void * output, size_t output_size);
+    void send_async(enum rpc_cmd cmd, std::shared_ptr<const void> input, size_t input_size);
+    void send_async(enum rpc_cmd cmd, std::shared_ptr<const void> input, size_t input_size, void * output, size_t output_size);
+
+    ggml_backend_event_t event_new(ggml_backend_dev_t dev);
+    void event_free(ggml_backend_event_t event);
+    void event_synchronize(ggml_backend_event_t event);
+    void event_record(ggml_backend_event_t event);
+    void synchronize();
+
+    void start(const std::string & endpoint);
+    void work();
+
+    ~rpc_dispatcher();
+
+private:
+    struct rpc_msg {
+        rpc_cmd                       cmd;
+        std::shared_ptr<const void>   input;
+        size_t                        input_size;
+        void                        * output;
+        size_t                        output_size;
+        std::promise<void>            completion;
+    };
+    using rpc_msg_ptr   = std::shared_ptr<rpc_msg>;
+    using rpc_msg_queue = message_queue<rpc_msg_ptr>;
+    struct rpc_event {
+        rpc_msg_ptr              msg;
+        std::shared_future<void> sf;
+    };
+    rpc_msg_queue    queue;
+    socket_ptr       sock;
+    std::atomic_bool running;
+    std::thread      thread;
+};
+
+static void rpc_dispatcher_trampoline(rpc_dispatcher * dispatcher)
+{
+    dispatcher->work();
+}
+
+void rpc_dispatcher::send(enum rpc_cmd cmd, std::shared_ptr<const void> input, size_t input_size) {
+    auto msg = std::make_shared<rpc_msg>();
+    msg->cmd = cmd;
+    msg->input = input;
+    msg->input_size = input_size;
+    msg->output = nullptr;
+    msg->output_size = 0;
+    GGML_ASSERT(queue.push(msg));
+    auto future = msg->completion.get_future();
+    future.wait();
+}
+
+void rpc_dispatcher::send_async(enum rpc_cmd cmd, std::shared_ptr<const void> input, size_t input_size) {
+    auto msg = std::make_shared<rpc_msg>();
+    msg->cmd = cmd;
+    msg->input = input;
+    msg->input_size = input_size;
+    msg->output = nullptr;
+    msg->output_size = 0;
+    GGML_ASSERT(queue.push(msg));
+}
+
+void rpc_dispatcher::send(enum rpc_cmd cmd, std::shared_ptr<const void> input, size_t input_size, void * output, size_t output_size) {
+    auto msg = std::make_shared<rpc_msg>();
+    msg->cmd = cmd;
+    msg->input = input;
+    msg->input_size = input_size;
+    msg->output = output;
+    msg->output_size = output_size;
+    GGML_ASSERT(queue.push(msg));
+    auto future = msg->completion.get_future();
+    future.wait();
+}
+
+void rpc_dispatcher::send_async(enum rpc_cmd cmd, std::shared_ptr<const void> input, size_t input_size, void * output, size_t output_size) {
+    auto msg = std::make_shared<rpc_msg>();
+    msg->cmd = cmd;
+    msg->input = input;
+    msg->input_size = input_size;
+    msg->output = output;
+    msg->output_size = output_size;
+    GGML_ASSERT(queue.push(msg));
+}
+
+ggml_backend_event_t rpc_dispatcher::event_new(ggml_backend_dev_t dev) {
+    rpc_event * ev = new rpc_event;
+    ev->msg = std::make_shared<rpc_msg>();
+    ev->msg->cmd = RPC_CMD_NONE;
+    ev->sf = ev->msg->completion.get_future().share();
+    GGML_ASSERT(queue.push(ev->msg));
+    return new ggml_backend_event {
+        /* .device  = */ dev,
+        /* .context = */ ev,
+    };
+}
+
+void rpc_dispatcher::event_free(ggml_backend_event_t event) {
+    rpc_event * ev = (rpc_event *)event->context;
+    delete ev;
+}
+
+void rpc_dispatcher::event_synchronize(ggml_backend_event_t event) {
+    rpc_event * ev = (rpc_event *)event->context;
+    ev->sf.wait();
+}
+
+void rpc_dispatcher::event_record(ggml_backend_event_t event) {
+    rpc_event * ev = (rpc_event *)event->context;
+    ev->msg = std::make_shared<rpc_msg>();
+    ev->msg->cmd = RPC_CMD_NONE;
+    ev->sf = ev->msg->completion.get_future().share();
+    GGML_ASSERT(queue.push(ev->msg));
+}
+
+void rpc_dispatcher::synchronize() {
+    // to ensure all messages are processed, submit dummy message and wait for it to complete
+    auto msg = std::make_shared<rpc_msg>();
+    msg->cmd = RPC_CMD_NONE;
+    GGML_ASSERT(queue.push(msg));
+    msg->completion.get_future().wait();
+}
+
+void rpc_dispatcher::start(const std::string & endpoint) {
    std::string host;
    int port;
    if (!parse_endpoint(endpoint, host, port)) {
-        GGML_LOG_ERROR("Failed to parse endpoint: %s\n", endpoint.c_str());
-        return nullptr;
+        GGML_ABORT("Failed to parse endpoint: %s\n", endpoint.c_str());
+    }
+    if (!rpc_transport_init()) {
+        GGML_ABORT("RPC transport initialization failed\n");
    }

-    if (!rpc_transport_init()) {
-        return nullptr;
-    }
-    auto sock = socket_t::connect(host.c_str(), port);
+    sock = socket_t::connect(host.c_str(), port);
    if (sock == nullptr) {
-        return nullptr;
+        GGML_ABORT("Failed to connect to %s\n", endpoint.c_str());
    }
    if (!negotiate_hello(sock)) {
-        return nullptr;
+        GGML_ABORT("RPC handshake failed for %s\n", endpoint.c_str());
    }
    LOG_DBG("[%s] connected to %s\n", __func__, endpoint.c_str());
-    sockets[endpoint] = sock;
-    return sock;
+    running = true;
+    thread = std::thread(rpc_dispatcher_trampoline, this);
+}
+
+void rpc_dispatcher::work() {
+    while (running) {
+        rpc_msg_ptr msg_ptr;
+        if (!queue.pop(&msg_ptr)) {
+            break;
+        }
+        if (msg_ptr->cmd != RPC_CMD_NONE) {
+            if (msg_ptr->output) {
+                bool status = send_rpc_cmd(sock, msg_ptr->cmd, msg_ptr->input.get(), msg_ptr->input_size, msg_ptr->output, msg_ptr->output_size);
+                RPC_STATUS_ASSERT(status);
+            } else {
+                bool status = send_rpc_cmd(sock, msg_ptr->cmd, msg_ptr->input.get(), msg_ptr->input_size);
+                RPC_STATUS_ASSERT(status);
+            }
+        }
+        msg_ptr->completion.set_value();
+    }
+}
+
+rpc_dispatcher::~rpc_dispatcher() {
+    running = false;
+    queue.interrupt();
+    sock = nullptr;
+    if (thread.joinable()) {
+        thread.join();
+    }
+}
+
+static std::shared_ptr<rpc_dispatcher> get_dispatcher(const std::string & endpoint) {
+    static std::mutex mutex;
+    std::lock_guard<std::mutex> lock(mutex);
+    static std::unordered_map<std::string, std::weak_ptr<rpc_dispatcher>> dispatchers;
+
+    auto it = dispatchers.find(endpoint);
+    if (it != dispatchers.end()) {
+        if (auto dispatcher = it->second.lock()) {
+            return dispatcher;
+        }
+    }
+
+    auto dispatcher = std::make_shared<rpc_dispatcher>();
+    dispatcher->start(endpoint);
+    dispatchers[endpoint] = dispatcher;
+    return dispatcher;
 }

 static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
    ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
-    rpc_msg_free_buffer_req request = {ctx->remote_ptr};
-    bool status = send_rpc_cmd(ctx->sock, RPC_CMD_FREE_BUFFER, &request, sizeof(request), nullptr, 0);
-    RPC_STATUS_ASSERT(status);
+    auto request = std::make_shared<rpc_msg_free_buffer_req>();
+    request->remote_ptr = ctx->remote_ptr;
+    ctx->dispatcher->send(RPC_CMD_FREE_BUFFER, request, sizeof(*request));
    delete ctx;
 }

@@ -403,10 +613,10 @@ static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) {
    if (ctx->base_ptr != nullptr) {
        return ctx->base_ptr;
    }
-    rpc_msg_buffer_get_base_req request = {ctx->remote_ptr};
+    auto request = std::make_shared<rpc_msg_buffer_get_base_req>();
+    request->remote_ptr = ctx->remote_ptr;
    rpc_msg_buffer_get_base_rsp response;
-    bool status = send_rpc_cmd(ctx->sock, RPC_CMD_BUFFER_GET_BASE, &request, sizeof(request), &response, sizeof(response));
-    RPC_STATUS_ASSERT(status);
+    ctx->dispatcher->send(RPC_CMD_BUFFER_GET_BASE, request, sizeof(*request), &response, sizeof(response));
    ctx->base_ptr = reinterpret_cast<void *>(response.base_ptr);
    return ctx->base_ptr;
 }
@@ -463,12 +673,9 @@ static enum ggml_status ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_
    // Due to bandwidth constraints, we only call the server init tensor functions if necessary.
    // In particular, only quantized tensors need padding
    if (ggml_is_quantized(tensor->type) && (tensor->ne[0] % 512 != 0) && (tensor->view_src == nullptr)) {
-        rpc_msg_init_tensor_req request;
-
-        request.tensor = serialize_tensor(tensor);
-
-        bool status = send_rpc_cmd(ctx->sock, RPC_CMD_INIT_TENSOR, &request, sizeof(request), nullptr, 0);
-        RPC_STATUS_ASSERT(status);
+        auto request = std::make_shared<rpc_msg_init_tensor_req>();
+        request->tensor = serialize_tensor(tensor);
+        ctx->dispatcher->send(RPC_CMD_INIT_TENSOR, request, sizeof(*request));
    }
    return GGML_STATUS_SUCCESS;
 }
@@ -476,27 +683,24 @@ static enum ggml_status ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_
 static void ggml_backend_rpc_buffer_memset_tensor(
        ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
    ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
-    rpc_msg_memset_tensor_req request = {
-        /* .tensor = */ serialize_tensor(tensor),
-        /* .offset = */ offset,
-        /* .size   = */ size,
-        /* .value  = */ value,
-    };
-    bool status = send_rpc_cmd(ctx->sock, RPC_CMD_MEMSET_TENSOR, &request, sizeof(request), nullptr, 0);
-    RPC_STATUS_ASSERT(status);
+    auto request = std::make_shared<rpc_msg_memset_tensor_req>();
+    request->tensor = serialize_tensor(tensor);
+    request->offset = offset;
+    request->size   = size;
+    request->value  = value;
+    ctx->dispatcher->send(RPC_CMD_MEMSET_TENSOR, request, sizeof(*request));
 }

 static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
    ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
    rpc_tensor rpc_tensor = serialize_tensor(tensor);
    if (size > HASH_THRESHOLD) {
-        rpc_msg_set_tensor_hash_req request;
-        request.tensor = rpc_tensor;
-        request.offset = offset;
-        request.hash = fnv_hash((const uint8_t*)data, size);
+        auto request = std::make_shared<rpc_msg_set_tensor_hash_req>();
+        request->tensor = rpc_tensor;
+        request->offset = offset;
+        request->hash = fnv_hash((const uint8_t*)data, size);
        rpc_msg_set_tensor_hash_rsp response;
-        bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR_HASH, &request, sizeof(request), &response, sizeof(response));
-        RPC_STATUS_ASSERT(status);
+        ctx->dispatcher->send(RPC_CMD_SET_TENSOR_HASH, request, sizeof(*request), &response, sizeof(response));
        if (response.result) {
            // the server has the same data, no need to send it
            return;
@@ -504,22 +708,21 @@ static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggm
    }
    // input serialization format: | rpc_tensor | offset (8 bytes) | data (size bytes)
    size_t input_size = sizeof(rpc_tensor) + sizeof(uint64_t) + size;
-    std::vector<uint8_t> input(input_size, 0);
-    memcpy(input.data(), &rpc_tensor, sizeof(rpc_tensor));
-    memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset));
-    memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), data, size);
-    bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR, input.data(), input.size());
-    RPC_STATUS_ASSERT(status);
+    uint8_t * input = new uint8_t[input_size]();
+    memcpy(input, &rpc_tensor, sizeof(rpc_tensor));
+    memcpy(input + sizeof(rpc_tensor), &offset, sizeof(offset));
+    memcpy(input + sizeof(rpc_tensor) + sizeof(offset), data, size);
+    std::shared_ptr<uint8_t> input_ptr(input, std::default_delete<uint8_t[]>());
+    ctx->dispatcher->send(RPC_CMD_SET_TENSOR, input_ptr, input_size);
 }

 static void ggml_backend_rpc_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
    ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
-    rpc_msg_get_tensor_req request;
-    request.tensor = serialize_tensor(tensor);
-    request.offset = offset;
-    request.size = size;
-    bool status = send_rpc_cmd(ctx->sock, RPC_CMD_GET_TENSOR, &request, sizeof(request), data, size);
-    RPC_STATUS_ASSERT(status);
+    auto request = std::make_shared<rpc_msg_get_tensor_req>();
+    request->tensor = serialize_tensor(tensor);
+    request->offset = offset;
+    request->size = size;
+    ctx->dispatcher->send(RPC_CMD_GET_TENSOR, request, sizeof(*request), data, size);
 }

 static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
@@ -529,16 +732,15 @@ static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, con
        ggml_backend_rpc_buffer_context * src_ctx = (ggml_backend_rpc_buffer_context *)src_buffer->context;
        ggml_backend_buffer_t dst_buffer = dst->buffer;
        ggml_backend_rpc_buffer_context * dst_ctx = (ggml_backend_rpc_buffer_context *)dst_buffer->context;
-        if (src_ctx->sock != dst_ctx->sock) {
+        if (src_ctx->dispatcher != dst_ctx->dispatcher) {
            return false;
        }
        ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
-        rpc_msg_copy_tensor_req request;
-        request.src = serialize_tensor(src);
-        request.dst = serialize_tensor(dst);
+        auto request = std::make_shared<rpc_msg_copy_tensor_req>();
+        request->src = serialize_tensor(src);
+        request->dst = serialize_tensor(dst);
        rpc_msg_copy_tensor_rsp response;
-        bool status = send_rpc_cmd(ctx->sock, RPC_CMD_COPY_TENSOR, &request, sizeof(request), &response, sizeof(response));
-        RPC_STATUS_ASSERT(status);
+        ctx->dispatcher->send(RPC_CMD_COPY_TENSOR, request, sizeof(*request), &response, sizeof(response));
        return response.result;
    }
    return false;
@@ -546,9 +748,10 @@ static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, con

 static void ggml_backend_rpc_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
    ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
-    rpc_msg_buffer_clear_req request = {ctx->remote_ptr, value};
-    bool status = send_rpc_cmd(ctx->sock, RPC_CMD_BUFFER_CLEAR, &request, sizeof(request), nullptr, 0);
-    RPC_STATUS_ASSERT(status);
+    auto request = std::make_shared<rpc_msg_buffer_clear_req>();
+    request->remote_ptr = ctx->remote_ptr;
+    request->value = value;
+    ctx->dispatcher->send(RPC_CMD_BUFFER_CLEAR, request, sizeof(*request));
 }

 static ggml_backend_buffer_i ggml_backend_rpc_buffer_interface = {
@@ -572,15 +775,17 @@ static const char * ggml_backend_rpc_buffer_type_name(ggml_backend_buffer_type_t

 static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
    ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
-    rpc_msg_alloc_buffer_req request = {buft_ctx->device, size};
+    auto request = std::make_shared<rpc_msg_alloc_buffer_req>();
+    request->device = buft_ctx->device;
+    request->size = size;
    rpc_msg_alloc_buffer_rsp response;
-    auto sock = get_socket(buft_ctx->endpoint);
-    bool status = send_rpc_cmd(sock, RPC_CMD_ALLOC_BUFFER, &request, sizeof(request), &response, sizeof(response));
-    RPC_STATUS_ASSERT(status);
+
+    auto dispatcher = get_dispatcher(buft_ctx->endpoint);
+    dispatcher->send(RPC_CMD_ALLOC_BUFFER, request, sizeof(*request), &response, sizeof(response));
    if (response.remote_ptr != 0) {
        ggml_backend_buffer_t buffer = ggml_backend_buffer_init(buft,
            ggml_backend_rpc_buffer_interface,
-            new ggml_backend_rpc_buffer_context{sock, nullptr, response.remote_ptr},
+            new ggml_backend_rpc_buffer_context{dispatcher, nullptr, response.remote_ptr},
            response.remote_size);
        return buffer;
    } else {
@@ -588,11 +793,11 @@ static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer(ggml_back
    }
 }

-static size_t get_alignment(const std::shared_ptr<socket_t> & sock, uint32_t device) {
-    rpc_msg_get_alignment_req request = {device};
+static size_t get_alignment(const std::shared_ptr<rpc_dispatcher> & dispatcher, uint32_t device) {
+    auto request = std::make_shared<rpc_msg_get_alignment_req>();
+    request->device = device;
    rpc_msg_get_alignment_rsp response;
-    bool status = send_rpc_cmd(sock, RPC_CMD_GET_ALIGNMENT, &request, sizeof(request), &response, sizeof(response));
-    RPC_STATUS_ASSERT(status);
+    dispatcher->send(RPC_CMD_GET_ALIGNMENT, request, sizeof(*request), &response, sizeof(response));
    return response.alignment;
 }

@@ -601,11 +806,11 @@ static size_t ggml_backend_rpc_buffer_type_get_alignment(ggml_backend_buffer_typ
    return buft_ctx->alignment;
 }

-static size_t get_max_size(const std::shared_ptr<socket_t> & sock, uint32_t device) {
-    rpc_msg_get_max_size_req request = {device};
+static size_t get_max_size(const std::shared_ptr<rpc_dispatcher> & dispatcher, uint32_t device) {
+    auto request = std::make_shared<rpc_msg_get_max_size_req>();
+    request->device = device;
    rpc_msg_get_max_size_rsp response;
-    bool status = send_rpc_cmd(sock, RPC_CMD_GET_MAX_SIZE, &request, sizeof(request), &response, sizeof(response));
-    RPC_STATUS_ASSERT(status);
+    dispatcher->send(RPC_CMD_GET_MAX_SIZE, request, sizeof(*request), &response, sizeof(response));
    return response.max_size;
 }

@@ -628,23 +833,63 @@ static size_t ggml_backend_rpc_buffer_type_get_alloc_size(ggml_backend_buffer_ty

    if (rpc_get) {
        ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
-        auto sock = get_socket(buft_ctx->endpoint);

-        rpc_msg_get_alloc_size_req request = {
-            /*.device =*/ buft_ctx->device,
-            /*.tensor =*/ serialize_tensor(tensor),
-            /*.srcs   =*/ {},
+        // Cache key for calls to read the alloc_size.
+        // We deliberately exclude src tensor dimensions from the key because:
+        // 1. For CPU backends, alloc_size = ggml_nbytes(output) regardless of src shapes
+        // 2. For GPU backends, the reservation graph uses max dimensions, so the
+        //    cached value from reservation is always >= any subsequent request
+        // 3. Including src dims causes cache misses per-ubatch (e.g. growing KV cache)
+        //    which blocks the main thread behind in-flight GRAPH_COMPUTE commands
+        struct alloc_size_cache_key {
+            uint32_t device;
+            uint32_t type;
+            uint32_t op;
+            int32_t  op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
+            uint32_t ne[GGML_MAX_DIMS];
        };

+        alloc_size_cache_key key = {};
+        key.device = buft_ctx->device;
+        key.type = tensor->type;
+        key.op = tensor->op;
+        memcpy(key.op_params, tensor->op_params, sizeof(key.op_params));
+        for (int i = 0; i < GGML_MAX_DIMS; i++) {
+            key.ne[i] = (uint32_t)tensor->ne[i];
+        }
+
+        uint64_t cache_hash = fnv_hash((const uint8_t *)&key, sizeof(key));
+        cache_hash = fnv_hash((const uint8_t *)buft_ctx->endpoint.data(), buft_ctx->endpoint.size(), cache_hash);
+
+        // alloc sizes are immutable for a given tensor configuration
+        static std::mutex cache_mutex;
+        static std::unordered_map<uint64_t, size_t> cache;
+
+        {
+            std::lock_guard<std::mutex> lock(cache_mutex);
+            auto it = cache.find(cache_hash);
+            if (it != cache.end()) {
+                return it->second;
+            }
+        }
+
+        auto request = std::make_shared<rpc_msg_get_alloc_size_req>();
+        request->device = buft_ctx->device;
+        request->tensor = serialize_tensor(tensor);
+
        // .get_alloc_size could be a function of the tensor's srcs, so we must serialize them as well
        for (int i = 0; i < GGML_MAX_SRC; i++) {
-            request.srcs[i] = serialize_tensor(tensor->src[i]);
+            request->srcs[i] = serialize_tensor(tensor->src[i]);
        }

-        // TODO: cache the alloc responses to avoid extra RPC calls?
        rpc_msg_get_alloc_size_rsp response;
-        bool status = send_rpc_cmd(sock, RPC_CMD_GET_ALLOC_SIZE, &request, sizeof(request), &response, sizeof(response));
-        RPC_STATUS_ASSERT(status);
+        auto dispatcher = get_dispatcher(buft_ctx->endpoint);
+        dispatcher->send(RPC_CMD_GET_ALLOC_SIZE, request, sizeof(*request), &response, sizeof(response));
+
+        {
+            std::lock_guard<std::mutex> lock(cache_mutex);
+            cache[cache_hash] = response.alloc_size;
+        }

        return response.alloc_size;
    }
@@ -673,9 +918,44 @@ static void ggml_backend_rpc_free(ggml_backend_t backend) {
    delete backend;
 }

+static void ggml_backend_rpc_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    ggml_backend_rpc_context * ctx = (ggml_backend_rpc_context *)backend->context;
+    rpc_tensor rpc_tensor = serialize_tensor(tensor);
+    if (size > HASH_THRESHOLD) {
+        auto request = std::make_shared<rpc_msg_set_tensor_hash_req>();
+        request->tensor = rpc_tensor;
+        request->offset = offset;
+        request->hash = fnv_hash((const uint8_t*)data, size);
+        rpc_msg_set_tensor_hash_rsp response;
+        // TODO: make this async
+        ctx->dispatcher->send(RPC_CMD_SET_TENSOR_HASH, request, sizeof(*request), &response, sizeof(response));
+        if (response.result) {
+            // the server has the same data, no need to send it
+            return;
+        }
+    }
+    // input serialization format: | rpc_tensor | offset (8 bytes) | data (size bytes)
+    size_t input_size = sizeof(rpc_tensor) + sizeof(uint64_t) + size;
+    uint8_t * input = new uint8_t[input_size]();
+    memcpy(input, &rpc_tensor, sizeof(rpc_tensor));
+    memcpy(input + sizeof(rpc_tensor), &offset, sizeof(offset));
+    memcpy(input + sizeof(rpc_tensor) + sizeof(offset), data, size);
+    std::shared_ptr<uint8_t> input_ptr(input, std::default_delete<uint8_t[]>());
+    ctx->dispatcher->send_async(RPC_CMD_SET_TENSOR, input_ptr, input_size);
+}
+
+static void ggml_backend_rpc_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    ggml_backend_rpc_context * ctx = (ggml_backend_rpc_context *)backend->context;
+    auto request = std::make_shared<rpc_msg_get_tensor_req>();
+    request->tensor = serialize_tensor(tensor);
+    request->offset = offset;
+    request->size = size;
+    ctx->dispatcher->send_async(RPC_CMD_GET_TENSOR, request, sizeof(*request), data, size);
+}
+
 static void ggml_backend_rpc_synchronize(ggml_backend_t backend) {
-    GGML_UNUSED(backend);
-    // this is no-op because we don't have any async operations
+    ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
+    rpc_ctx->dispatcher->synchronize();
 }

 static void add_tensor(ggml_tensor * tensor, const ggml_cgraph * cgraph, std::vector<rpc_tensor> & tensors, std::unordered_set<ggml_tensor*> & visited) {
@@ -698,7 +978,7 @@ static void add_tensor(ggml_tensor * tensor, const ggml_cgraph * cgraph, std::ve
    tensors.push_back(result);
 }

-static void serialize_graph(uint32_t device, const ggml_cgraph * cgraph, std::vector<uint8_t> & output) {
+static uint8_t * serialize_graph(uint32_t device, const ggml_cgraph * cgraph, size_t * output_size) {
    uint32_t n_nodes = cgraph->n_nodes;
    std::vector<rpc_tensor> tensors;
    std::unordered_set<ggml_tensor*> visited;
@@ -708,9 +988,9 @@ static void serialize_graph(uint32_t device, const ggml_cgraph * cgraph, std::ve
    // serialization format:
    // | device (4 bytes) | n_nodes (4 bytes) | nodes (n_nodes * sizeof(uint64_t) | n_tensors (4 bytes) | tensors (n_tensors * sizeof(rpc_tensor)) |
    uint32_t n_tensors = tensors.size();
-    int output_size = 2*sizeof(uint32_t) + n_nodes * sizeof(uint64_t) + sizeof(uint32_t) + n_tensors * sizeof(rpc_tensor);
-    output.resize(output_size, 0);
-    uint8_t * dest = output.data();
+    *output_size = 2*sizeof(uint32_t) + n_nodes * sizeof(uint64_t) + sizeof(uint32_t) + n_tensors * sizeof(rpc_tensor);
+    uint8_t * output = new uint8_t[*output_size]();
+    uint8_t * dest = output;
    memcpy(dest, &device, sizeof(device));
    dest += sizeof(device);
    memcpy(dest, &n_nodes, sizeof(n_nodes));
@@ -723,6 +1003,7 @@ static void serialize_graph(uint32_t device, const ggml_cgraph * cgraph, std::ve
    dest += sizeof(n_tensors);
    rpc_tensor * out_tensors = (rpc_tensor *)dest;
    memcpy(out_tensors, tensors.data(), n_tensors * sizeof(rpc_tensor));
+    return output;
 }

 static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
@@ -733,27 +1014,35 @@ static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t backend, g
    GGML_ASSERT(cgraph->n_nodes > 0);
    bool reuse = cgraph->uid != 0 && rpc_dev_ctx->last_graph_uid == cgraph->uid;
    if (reuse) {
-        rpc_msg_graph_recompute_req request;
-        request.device = rpc_ctx->device;
-        auto sock = get_socket(rpc_ctx->endpoint);
-        bool status = send_rpc_cmd(sock, RPC_CMD_GRAPH_RECOMPUTE, &request, sizeof(request));
-        RPC_STATUS_ASSERT(status);
+        auto request = std::make_shared<rpc_msg_graph_recompute_req>();
+        request->device = rpc_ctx->device;
+        rpc_ctx->dispatcher->send_async(RPC_CMD_GRAPH_RECOMPUTE, request, sizeof(*request));
    } else {
        rpc_dev_ctx->last_graph_uid = cgraph->uid;
-        std::vector<uint8_t> input;
-        serialize_graph(rpc_ctx->device, cgraph, input);
-        auto sock = get_socket(rpc_ctx->endpoint);
-        bool status = send_rpc_cmd(sock, RPC_CMD_GRAPH_COMPUTE, input.data(), input.size());
-        RPC_STATUS_ASSERT(status);
+        size_t input_size = 0;
+        uint8_t * input = serialize_graph(rpc_ctx->device, cgraph, &input_size);
+        std::shared_ptr<uint8_t> input_ptr(input, std::default_delete<uint8_t[]>());
+        rpc_ctx->dispatcher->send_async(RPC_CMD_GRAPH_COMPUTE, input_ptr, input_size);
    }
    return GGML_STATUS_SUCCESS;
 }

+static void ggml_backend_rpc_event_record(ggml_backend_t backend, ggml_backend_event_t event) {
+    ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
+    rpc_ctx->dispatcher->event_record(event);
+}
+
+static void ggml_backend_rpc_event_wait(ggml_backend_t backend, ggml_backend_event_t event) {
+    // this is noop for RPC as we have a single stream
+    GGML_UNUSED(backend);
+    GGML_UNUSED(event);
+}
+
 static ggml_backend_i ggml_backend_rpc_interface = {
    /* .get_name                = */ ggml_backend_rpc_name,
    /* .free                    = */ ggml_backend_rpc_free,
-    /* .set_tensor_async        = */ NULL,
-    /* .get_tensor_async        = */ NULL,
+    /* .set_tensor_async        = */ ggml_backend_rpc_set_tensor_async,
+    /* .get_tensor_async        = */ ggml_backend_rpc_get_tensor_async,
    /* .set_tensor_2d_async     = */ NULL,
    /* .get_tensor_2d_async     = */ NULL,
    /* .cpy_tensor_async        = */ NULL,
@@ -763,8 +1052,8 @@ static ggml_backend_i ggml_backend_rpc_interface = {
    /* .graph_plan_update       = */ NULL,
    /* .graph_plan_compute      = */ NULL,
    /* .graph_compute           = */ ggml_backend_rpc_graph_compute,
-    /* .event_record            = */ NULL,
-    /* .event_wait              = */ NULL,
+    /* .event_record            = */ ggml_backend_rpc_event_record,
+    /* .event_wait              = */ ggml_backend_rpc_event_wait,
    /* .graph_optimize          = */ NULL,
 };

@@ -778,13 +1067,9 @@ ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint, u
    if (it != buft_map.end()) {
        return it->second;
    }
-    auto sock = get_socket(endpoint);
-    if (sock == nullptr) {
-        GGML_LOG_ERROR("Failed to connect to %s\n", endpoint);
-        return nullptr;
-    }
-    size_t alignment = get_alignment(sock, device);
-    size_t max_size = get_max_size(sock, device);
+    auto dispatcher = get_dispatcher(endpoint);
+    size_t alignment = get_alignment(dispatcher, device);
+    size_t max_size = get_max_size(dispatcher, device);
    ggml_backend_rpc_buffer_type_context * buft_ctx = new ggml_backend_rpc_buffer_type_context {
        /* .endpoint  = */ endpoint,
        /* .device    = */ device,
@@ -804,10 +1089,11 @@ ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint, u

 ggml_backend_t ggml_backend_rpc_init(const char * endpoint, uint32_t device) {
    std::string dev_name = "RPC" + std::to_string(device) + "[" + std::string(endpoint) + "]";
+    auto dispatcher = get_dispatcher(endpoint);
    ggml_backend_rpc_context * ctx = new ggml_backend_rpc_context {
-        /* .endpoint       = */ endpoint,
-        /* .device         = */ device,
-        /* .name           = */ dev_name,
+        /* .dispatcher = */ dispatcher,
+        /* .device     = */ device,
+        /* .name       = */ dev_name,
    };
    auto reg = ggml_backend_rpc_add_server(endpoint);
    ggml_backend_t backend = new ggml_backend {
@@ -823,26 +1109,16 @@ bool ggml_backend_is_rpc(ggml_backend_t backend) {
    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_rpc_guid());
 }

-static void get_device_memory(const std::shared_ptr<socket_t> & sock, uint32_t device, size_t * free, size_t * total) {
-    rpc_msg_get_device_memory_req request;
-    request.device = device;
+void ggml_backend_rpc_get_device_memory(const char * endpoint, uint32_t device, size_t * free, size_t * total) {
+    auto dispatcher = get_dispatcher(endpoint);
+    auto request = std::make_shared<rpc_msg_get_device_memory_req>();
+    request->device = device;
    rpc_msg_get_device_memory_rsp response;
-    bool status = send_rpc_cmd(sock, RPC_CMD_GET_DEVICE_MEMORY, &request, sizeof(request), &response, sizeof(response));
-    RPC_STATUS_ASSERT(status);
+    dispatcher->send(RPC_CMD_GET_DEVICE_MEMORY, request, sizeof(*request), &response, sizeof(response));
    *free = response.free_mem;
    *total = response.total_mem;
 }

-void ggml_backend_rpc_get_device_memory(const char * endpoint, uint32_t device, size_t * free, size_t * total) {
-    auto sock = get_socket(endpoint);
-    if (sock == nullptr) {
-        *free = 0;
-        *total = 0;
-        return;
-    }
-    get_device_memory(sock, device, free, total);
-}
-
 // RPC server-side implementation

 class rpc_server {
@@ -1647,9 +1923,6 @@ static void rpc_serve_client(const std::vector<ggml_backend_t> & backends, const
                if (!server.free_buffer(request)) {
                    return;
                }
-                if (!send_msg(sock, nullptr, 0)) {
-                    return;
-                }
                break;
            }
            case RPC_CMD_BUFFER_CLEAR: {
@@ -1660,9 +1933,6 @@ static void rpc_serve_client(const std::vector<ggml_backend_t> & backends, const
                if (!server.buffer_clear(request)) {
                    return;
                }
-                if (!send_msg(sock, nullptr, 0)) {
-                    return;
-                }
                break;
            }
            case RPC_CMD_MEMSET_TENSOR: {
@@ -1673,9 +1943,6 @@ static void rpc_serve_client(const std::vector<ggml_backend_t> & backends, const
                if (!server.memset_tensor(request)) {
                    return;
                }
-                if (!send_msg(sock, nullptr, 0)) {
-                    return;
-                }
                break;
            }
            case RPC_CMD_SET_TENSOR: {
@@ -1710,9 +1977,6 @@ static void rpc_serve_client(const std::vector<ggml_backend_t> & backends, const
                if (!server.init_tensor(request)) {
                    return;
                }
-                if (!send_msg(sock, nullptr, 0)) {
-                    return;
-                }
                break;
            }
            case RPC_CMD_GET_TENSOR: {
@@ -1889,10 +2153,10 @@ static void ggml_backend_rpc_device_get_props(ggml_backend_dev_t dev, struct ggm
    props->type        = ggml_backend_rpc_device_get_type(dev);
    ggml_backend_rpc_device_get_memory(dev, &props->memory_free, &props->memory_total);
    props->caps = {
-        /* .async                 = */ false,
+        /* .async                 = */ true,
        /* .host_buffer           = */ false,
        /* .buffer_from_host_ptr  = */ false,
-        /* .events                = */ false,
+        /* .events                = */ true,
        /* .mmap_support          = */ true,
    };
 }
@@ -1929,6 +2193,24 @@ static bool ggml_backend_rpc_device_supports_buft(ggml_backend_dev_t dev, ggml_b
    return buft_ctx->endpoint == dev_ctx->endpoint && buft_ctx->device == dev_ctx->device;
 }

+static ggml_backend_event_t ggml_backend_rpc_device_event_new(ggml_backend_dev_t dev) {
+    ggml_backend_rpc_device_context * ctx = (ggml_backend_rpc_device_context *)dev->context;
+    auto dispatcher = get_dispatcher(ctx->endpoint);
+    return dispatcher->event_new(dev);
+}
+
+static void ggml_backend_rpc_device_event_free(ggml_backend_dev_t dev, ggml_backend_event_t event) {
+    ggml_backend_rpc_device_context * ctx = (ggml_backend_rpc_device_context *)dev->context;
+    auto dispatcher = get_dispatcher(ctx->endpoint);
+    dispatcher->event_free(event);
+}
+
+static void ggml_backend_rpc_device_event_synchronize(ggml_backend_dev_t dev, ggml_backend_event_t event) {
+    ggml_backend_rpc_device_context * ctx = (ggml_backend_rpc_device_context *)dev->context;
+    auto dispatcher = get_dispatcher(ctx->endpoint);
+    dispatcher->event_synchronize(event);
+}
+
 static const struct ggml_backend_device_i ggml_backend_rpc_device_i = {
    /* .get_name             = */ ggml_backend_rpc_device_get_name,
    /* .get_description      = */ ggml_backend_rpc_device_get_description,
@@ -1942,9 +2224,9 @@ static const struct ggml_backend_device_i ggml_backend_rpc_device_i = {
    /* .supports_op          = */ ggml_backend_rpc_device_supports_op,
    /* .supports_buft        = */ ggml_backend_rpc_device_supports_buft,
    /* .offload_op           = */ NULL,
-    /* .event_new            = */ NULL,
-    /* .event_free           = */ NULL,
-    /* .event_synchronize    = */ NULL,
+    /* .event_new            = */ ggml_backend_rpc_device_event_new,
+    /* .event_free           = */ ggml_backend_rpc_device_event_free,
+    /* .event_synchronize    = */ ggml_backend_rpc_device_event_synchronize,
 };

 // backend reg interface
@@ -2004,14 +2286,9 @@ ggml_backend_reg_t ggml_backend_rpc_reg(void) {
 }

 static uint32_t ggml_backend_rpc_get_device_count(const char * endpoint) {
-    auto sock = get_socket(endpoint);
-    if (sock == nullptr) {
-        GGML_LOG_ERROR("Failed to connect to %s\n", endpoint);
-        return 0;
-    }
+    auto dispatcher = get_dispatcher(endpoint);
    rpc_msg_device_count_rsp response;
-    bool status = send_rpc_cmd(sock, RPC_CMD_DEVICE_COUNT, nullptr, 0, &response, sizeof(response));
-    RPC_STATUS_ASSERT(status);
+    dispatcher->send(RPC_CMD_DEVICE_COUNT, nullptr, 0, &response, sizeof(response));
    return response.device_count;
 }

@@ -1042,6 +1042,8 @@ struct vk_device_struct {
    vk_pipeline pipeline_argsort_large_f32[num_argsort_pipelines];
    vk_pipeline pipeline_topk_f32[num_topk_pipelines];
    vk_pipeline pipeline_sum_rows_f32;
+    vk_pipeline pipeline_cross_entropy_loss_f32, pipeline_cross_entropy_loss_f32_wg512;
+    vk_pipeline pipeline_cross_entropy_loss_back_f32, pipeline_cross_entropy_loss_back_f32_wg512;
    vk_pipeline pipeline_fwht_f32[4];
    vk_pipeline pipeline_cumsum_f32;
    vk_pipeline pipeline_cumsum_small_f32;
@@ -4169,10 +4171,16 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
    const uint32_t subgroup_size_16 = std::max(device->subgroup_size, 16u);
    const uint32_t subgroup_size_32 = std::max(device->subgroup_size, 32u);

+    // clamp WARP for l_/m_ warptiles so WM <= BM (breaks on subgroupSize > 64)
+    const uint32_t mm_warp_8  = std::min(subgroup_size_8,  64u);
+    const uint32_t mm_warp_16 = std::min(subgroup_size_16, 64u);
+
    const uint32_t mul_mat_subgroup_size = (device->vendor_id == VK_VENDOR_ID_INTEL && device->subgroup_size_control) ? device->subgroup_min_size : device->subgroup_size;
    const uint32_t mul_mat_subgroup_size_8 = std::max(mul_mat_subgroup_size, 8u);
    const uint32_t mul_mat_subgroup_size_16 = std::max(mul_mat_subgroup_size, 16u);
    const uint32_t mul_mat_subgroup_size_32 = std::max(mul_mat_subgroup_size, 32u);
+    const uint32_t mul_mat_mm_warp_8  = std::min(mul_mat_subgroup_size_8,  64u);
+    const uint32_t mul_mat_mm_warp_16 = std::min(mul_mat_subgroup_size_16, 64u);

    const bool subgroup_min_size_16 = (!device->subgroup_size_control && device->subgroup_size >= 16) ||
                                      (device->subgroup_size_control && device->subgroup_max_size >= 16);
@@ -4253,39 +4261,39 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {

        const uint32_t s_warptile_wm = device->subgroup_size == 8 ? 8 : 32;

-        l_warptile = { 128,             128, 128, 16, subgroup_size_8 * 2, 64, 2, tm_l, tn_l, tk_l, subgroup_size_8 };
-        m_warptile = { 128,              64,  64, 16, subgroup_size_8,     32, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
-        s_warptile = { subgroup_size_32, 32,  32, 16, s_warptile_wm,       32, 2, tm_s, tn_s, tk_s, subgroup_size_8 };
+        l_warptile = { 128,             128, 128, 16, mm_warp_8 * 2, 64, 2, tm_l, tn_l, tk_l, mm_warp_8 };
+        m_warptile = { 128,              64,  64, 16, mm_warp_8,     32, 2, tm_m, tn_m, tk_m, mm_warp_8 };
+        s_warptile = { subgroup_size_32, 32,  32, 16, s_warptile_wm, 32, 2, tm_s, tn_s, tk_s, subgroup_size_8 };

-        l_warptile_mmq = { 128,             128, 128, 32, subgroup_size_8 * 2, 64, 2, tm_l, tn_l, tk_l, subgroup_size_8 };
-        m_warptile_mmq = { 128,              64,  64, 32, subgroup_size_8,     32, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
-        s_warptile_mmq = { subgroup_size_32, 32,  32, 32, s_warptile_wm,       32, 2, tm_s, tn_s, tk_s, subgroup_size_8 };
+        l_warptile_mmq = { 128,             128, 128, 32, mm_warp_8 * 2, 64, 2, tm_l, tn_l, tk_l, mm_warp_8 };
+        m_warptile_mmq = { 128,              64,  64, 32, mm_warp_8,     32, 2, tm_m, tn_m, tk_m, mm_warp_8 };
+        s_warptile_mmq = { subgroup_size_32, 32,  32, 32, s_warptile_wm, 32, 2, tm_s, tn_s, tk_s, subgroup_size_8 };

        // Integer MMQ has a smaller shared memory profile, but heavier register use
-        l_warptile_mmq_int = { 128,             128, 128, 32, subgroup_size_8 * 2, 64, 2, 4, 4, 1, subgroup_size_8 };
-        m_warptile_mmq_int = { 128,              64,  64, 32, subgroup_size_8,     32, 2, 2, 2, 1, subgroup_size_8 };
-        s_warptile_mmq_int = { subgroup_size_32, 32,  32, 32, s_warptile_wm,       32, 2, 2, 1, 1, subgroup_size_8 };
+        l_warptile_mmq_int = { 128,             128, 128, 32, mm_warp_8 * 2, 64, 2, 4, 4, 1, mm_warp_8 };
+        m_warptile_mmq_int = { 128,              64,  64, 32, mm_warp_8,     32, 2, 2, 2, 1, mm_warp_8 };
+        s_warptile_mmq_int = { subgroup_size_32, 32,  32, 32, s_warptile_wm, 32, 2, 2, 1, 1, subgroup_size_8 };

        // K-quants use even more registers, mitigate by setting WMITER to 1
-        l_warptile_mmq_int_k = { 128,               128, 128, 32, subgroup_size_8 * 2, 64, 1, 4, 4, 1, subgroup_size_8 };
-        m_warptile_mmq_int_k = { 128,                64,  64, 32, subgroup_size_8,     32, 1, 2, 2, 1, subgroup_size_8 };
-        s_warptile_mmq_int_k = { subgroup_size_32,   32,  32, 32, s_warptile_wm,       32, 1, 2, 1, 1, subgroup_size_8 };
+        l_warptile_mmq_int_k = { 128,               128, 128, 32, mm_warp_8 * 2, 64, 1, 4, 4, 1, mm_warp_8 };
+        m_warptile_mmq_int_k = { 128,                64,  64, 32, mm_warp_8,     32, 1, 2, 2, 1, mm_warp_8 };
+        s_warptile_mmq_int_k = { subgroup_size_32,   32,  32, 32, s_warptile_wm, 32, 1, 2, 1, 1, subgroup_size_8 };

-        l_warptile_id = { 128,                      128, 128, 16, mul_mat_subgroup_size_16 * 2, 64, 2, tm_l, tn_l, tk_l, mul_mat_subgroup_size_16 };
-        m_warptile_id = { 128,                       64,  64, 16, mul_mat_subgroup_size_16,     32, 2, tm_m, tn_m, tk_m, mul_mat_subgroup_size_16 };
-        s_warptile_id = { mul_mat_subgroup_size_16,  32,  32, 16, s_warptile_wm,                32, 2, tm_s, tn_s, tk_s, mul_mat_subgroup_size_16 };
+        l_warptile_id = { 128,                      128, 128, 16, mul_mat_mm_warp_16 * 2, 64, 2, tm_l, tn_l, tk_l, mul_mat_mm_warp_16 };
+        m_warptile_id = { 128,                       64,  64, 16, mul_mat_mm_warp_16,     32, 2, tm_m, tn_m, tk_m, mul_mat_mm_warp_16 };
+        s_warptile_id = { mul_mat_subgroup_size_16,  32,  32, 16, s_warptile_wm,          32, 2, tm_s, tn_s, tk_s, mul_mat_subgroup_size_16 };

-        l_warptile_mmqid = { 128,                       128, 128, 32, mul_mat_subgroup_size_8 * 2, 64, 2, tm_l, tn_l, tk_l, mul_mat_subgroup_size_8 };
-        m_warptile_mmqid = { 128,                        64,  64, 32, mul_mat_subgroup_size_8,     32, 2, tm_m, tn_m, tk_m, mul_mat_subgroup_size_8 };
-        s_warptile_mmqid = { mul_mat_subgroup_size_32,   32,  32, 32, s_warptile_wm,               32, 2, tm_s, tn_s, tk_s, mul_mat_subgroup_size_8 };
+        l_warptile_mmqid = { 128,                       128, 128, 32, mul_mat_mm_warp_8 * 2, 64, 2, tm_l, tn_l, tk_l, mul_mat_mm_warp_8 };
+        m_warptile_mmqid = { 128,                        64,  64, 32, mul_mat_mm_warp_8,     32, 2, tm_m, tn_m, tk_m, mul_mat_mm_warp_8 };
+        s_warptile_mmqid = { mul_mat_subgroup_size_32,   32,  32, 32, s_warptile_wm,         32, 2, tm_s, tn_s, tk_s, mul_mat_subgroup_size_8 };

-        l_warptile_mmqid_int = { 128,                       128, 128, 32, mul_mat_subgroup_size_8 * 2, 64, 2, 4, 4, 1, mul_mat_subgroup_size_8 };
-        m_warptile_mmqid_int = { 128,                        64,  64, 32, mul_mat_subgroup_size_8,     32, 2, 2, 2, 1, mul_mat_subgroup_size_8 };
-        s_warptile_mmqid_int = { mul_mat_subgroup_size_32,   32,  32, 32, s_warptile_wm,               32, 2, 2, 1, 1, mul_mat_subgroup_size_8 };
+        l_warptile_mmqid_int = { 128,                       128, 128, 32, mul_mat_mm_warp_8 * 2, 64, 2, 4, 4, 1, mul_mat_mm_warp_8 };
+        m_warptile_mmqid_int = { 128,                        64,  64, 32, mul_mat_mm_warp_8,     32, 2, 2, 2, 1, mul_mat_mm_warp_8 };
+        s_warptile_mmqid_int = { mul_mat_subgroup_size_32,   32,  32, 32, s_warptile_wm,         32, 2, 2, 1, 1, mul_mat_subgroup_size_8 };

-        l_warptile_mmqid_int_k = { 128,                     128, 128, 32, mul_mat_subgroup_size_16 * 2, 64, 1, 4, 4, 1, mul_mat_subgroup_size_16 };
-        m_warptile_mmqid_int_k = { 128,                      64,  64, 32, mul_mat_subgroup_size_16,     32, 1, 2, 2, 1, mul_mat_subgroup_size_16 };
-        s_warptile_mmqid_int_k = { mul_mat_subgroup_size_32, 32,  32, 32, s_warptile_wm,                32, 1, 2, 1, 1, mul_mat_subgroup_size_16 };
+        l_warptile_mmqid_int_k = { 128,                     128, 128, 32, mul_mat_mm_warp_16 * 2, 64, 1, 4, 4, 1, mul_mat_mm_warp_16 };
+        m_warptile_mmqid_int_k = { 128,                      64,  64, 32, mul_mat_mm_warp_16,     32, 1, 2, 2, 1, mul_mat_mm_warp_16 };
+        s_warptile_mmqid_int_k = { mul_mat_subgroup_size_32, 32,  32, 32, s_warptile_wm,          32, 1, 2, 1, 1, mul_mat_subgroup_size_16 };

        // chip specific tuning
        if ((device->architecture == AMD_GCN) && (device->driver_id != vk::DriverId::eAmdProprietary)) {
@@ -4293,13 +4301,13 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
            m_warptile_mmqid = m_warptile_mmqid_int = { 256, 64, 64, 32, 16, 16, 2, 2, 2, 1, 16 };
        } else if (device->vendor_id == VK_VENDOR_ID_AMD && device->coopmat_support && device->driver_id != vk::DriverId::eAmdProprietary) {
            // This is intentionally using tx_m values, slight performance increase
-            l_warptile = { 256, 128, 128, 16, subgroup_size_8, 64, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
-            l_warptile_mmq = l_warptile_mmq_int = { 256, 128, 128, 32, subgroup_size_8, 64, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
-            l_warptile_mmq_int_k = { 256, 128, 128, 32, subgroup_size_16, 64, 1, 4, 2, 1, subgroup_size_16 };
+            l_warptile = { 256, 128, 128, 16, mm_warp_8, 64, 2, tm_m, tn_m, tk_m, mm_warp_8 };
+            l_warptile_mmq = l_warptile_mmq_int = { 256, 128, 128, 32, mm_warp_8, 64, 2, tm_m, tn_m, tk_m, mm_warp_8 };
+            l_warptile_mmq_int_k = { 256, 128, 128, 32, mm_warp_16, 64, 1, 4, 2, 1, mm_warp_16 };
        } else if (device->vendor_id == VK_VENDOR_ID_INTEL && device->coopmat_support) {
            // Xe2/Xe3 with coopmat enabled - warptile performance tuning
-            l_warptile = { 512, 128, 128, 16, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
-            l_warptile_mmq = { 512, 128, 128, 32, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
+            l_warptile = { 512, 128, 128, 16, mm_warp_8, 32, 2, tm_m, tn_m, tk_m, mm_warp_8 };
+            l_warptile_mmq = { 512, 128, 128, 32, mm_warp_8, 32, 2, tm_m, tn_m, tk_m, mm_warp_8 };
        }

        l_mmq_wg_denoms = l_wg_denoms = {128, 128, 1 };
@@ -5172,8 +5180,8 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
        const uint32_t s_warptile_wm = device->subgroup_size == 8 ? 8 : 32;

        // use scalar tile sizes
-        l_warptile = { 128, 128, 128, 16, subgroup_size_8 * 2, 64, 2, 4, 4, 1, subgroup_size_8 };
-        m_warptile = { 128,  64,  64, 16, subgroup_size_8, 32, 2, 4, 2, 1, subgroup_size_8 };
+        l_warptile = { 128, 128, 128, 16, mm_warp_8 * 2, 64, 2, 4, 4, 1, mm_warp_8 };
+        m_warptile = { 128,  64,  64, 16, mm_warp_8, 32, 2, 4, 2, 1, mm_warp_8 };
        s_warptile = { subgroup_size_32, 32, 32, 16, s_warptile_wm, 32, 2, 2, 2, 1, subgroup_size_8 };

        l_wg_denoms = {128, 128, 1 };
@@ -5758,6 +5766,10 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
    ggml_vk_create_pipeline(device, device->pipeline_argmax_f32, "argmax_f32", argmax_f32_len, argmax_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);

    ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cross_entropy_loss_f32, "cross_entropy_loss_f32", cross_entropy_loss_f32_len, cross_entropy_loss_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cross_entropy_loss_f32_wg512, "cross_entropy_loss_f32_wg512", cross_entropy_loss_f32_len, cross_entropy_loss_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, { 512 }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cross_entropy_loss_back_f32, "cross_entropy_loss_back_f32", cross_entropy_loss_back_f32_len, cross_entropy_loss_back_f32_data, "main", 4, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cross_entropy_loss_back_f32_wg512, "cross_entropy_loss_back_f32_wg512", cross_entropy_loss_back_f32_len, cross_entropy_loss_back_f32_data, "main", 4, sizeof(vk_op_push_constants), {1, 1, 1}, { 512 }, 1);
    // Intel Windows driver in range [32.0.101.8509, 32.0.101.8860) will crash when using fwht kernels so we gate that here
    const bool can_use_fwht = device->driver_id != vk::DriverId::eIntelProprietaryWindows ||
        !ggml_vk_intel_windows_driver_in_range(device->properties.driverVersion, 101, 8509, 101, 8860);
@@ -11577,6 +11589,17 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
            return ctx->device->pipeline_sum_rows_f32;
        }
        return nullptr;
+    case GGML_OP_CROSS_ENTROPY_LOSS:
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return src0->ne[0] > 1024 ? ctx->device->pipeline_cross_entropy_loss_f32_wg512 : ctx->device->pipeline_cross_entropy_loss_f32;
+        }
+        return nullptr;
+    case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
+        // src0 is the scalar grad; src1 is logits
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && src2 && src2->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return src1->ne[0] > 1024 ? ctx->device->pipeline_cross_entropy_loss_back_f32_wg512 : ctx->device->pipeline_cross_entropy_loss_back_f32;
+        }
+        return nullptr;
    case GGML_OP_CUMSUM:
        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
            if (src0->ne[0] <= 512) {
@@ -13942,6 +13965,103 @@ static void ggml_vk_cumsum(ggml_backend_vk_context * ctx, vk_context& subctx, co
    ctx->prealloc_split_k_need_sync = true;
 }

+static std::array<uint32_t, 3> ggml_vk_nrows_elements(uint32_t nr) {
+    if (nr > 262144) {
+        return { 512, 512, CEIL_DIV(nr, 262144) };
+    }
+    if (nr > 512) {
+        return { 512, CEIL_DIV(nr, 512), 1 };
+    }
+    return { nr, 1, 1 };
+}
+
+static void ggml_vk_cross_entropy_loss(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src1));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+    GGML_ASSERT(ggml_are_same_shape(src0, src1));
+    GGML_ASSERT(ggml_is_scalar(dst));
+
+    const uint32_t nclasses = (uint32_t)src0->ne[0];
+    const uint32_t nrows    = (uint32_t)ggml_nrows(src0);
+
+    vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, nullptr, dst, GGML_OP_CROSS_ENTROPY_LOSS);
+    GGML_ASSERT(pipeline != nullptr);
+
+    ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
+    ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_sum_rows_f32, 1);
+
+    vk_subbuffer src0_buf = ggml_vk_tensor_subbuffer(ctx, src0);
+    vk_subbuffer src1_buf = ggml_vk_tensor_subbuffer(ctx, src1);
+    vk_subbuffer dst_buf  = ggml_vk_tensor_subbuffer(ctx, dst, true);
+
+    const vk_op_push_constants pc = { nclasses, nrows, 0.0f, 0.0f, 0.0f, 0.0f };
+
+    const size_t tmp_size = (size_t)nrows * sizeof(float);
+    if (ctx->prealloc_size_x < tmp_size) {
+        ctx->prealloc_size_x = tmp_size;
+        ggml_vk_preallocate_buffers(ctx, subctx);
+    }
+    if (ctx->prealloc_x_need_sync) {
+        ggml_vk_sync_buffers(ctx, subctx);
+    }
+
+    vk_subbuffer tmp_buf = { ctx->prealloc_x, 0, tmp_size };
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src0_buf, src1_buf, tmp_buf }, pc, ggml_vk_nrows_elements(nrows));
+    ggml_vk_sync_buffers(ctx, subctx);
+
+    vk_op_sum_rows_push_constants sp = {};
+    sp.n_cols = nrows;
+    sp.ne01 = 1;
+    sp.ne02 = 1;
+    sp.weight = 1.0f;
+    init_pushconst_fastdiv(sp);
+    sp.misalign_offsets = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
+
+    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_sum_rows_f32, { tmp_buf, dst_buf }, sp, { 1, 1, 1 });
+    ctx->prealloc_x_need_sync = true;
+}
+
+static void ggml_vk_cross_entropy_loss_back(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst) {
+    const ggml_tensor * grad   = dst->src[0];
+    const ggml_tensor * logits = dst->src[1];
+    const ggml_tensor * labels = dst->src[2];
+
+    GGML_ASSERT(grad->type   == GGML_TYPE_F32);
+    GGML_ASSERT(logits->type == GGML_TYPE_F32);
+    GGML_ASSERT(labels->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type    == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_scalar(grad));
+    GGML_ASSERT(ggml_is_contiguous(grad));
+    GGML_ASSERT(ggml_is_contiguous(logits));
+    GGML_ASSERT(ggml_is_contiguous(labels));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+    GGML_ASSERT(ggml_are_same_shape(logits, labels));
+    GGML_ASSERT(ggml_are_same_shape(logits, dst));
+
+    const uint32_t nclasses = (uint32_t)logits->ne[0];
+    const uint32_t nrows    = (uint32_t)ggml_nrows(logits);
+
+    vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, grad, logits, labels, dst, GGML_OP_CROSS_ENTROPY_LOSS_BACK);
+    GGML_ASSERT(pipeline != nullptr);
+
+    ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
+
+    vk_subbuffer grad_buf   = ggml_vk_tensor_subbuffer(ctx, grad);
+    vk_subbuffer logits_buf = ggml_vk_tensor_subbuffer(ctx, logits);
+    vk_subbuffer labels_buf = ggml_vk_tensor_subbuffer(ctx, labels);
+    vk_subbuffer dst_buf    = ggml_vk_tensor_subbuffer(ctx, dst);
+
+    const vk_op_push_constants pc = { nclasses, nrows, 0.0f, 0.0f, 0.0f, 0.0f };
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { grad_buf, logits_buf, labels_buf, dst_buf }, pc, ggml_vk_nrows_elements(nrows));
+}
+
 static void ggml_vk_argmax(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_ARGMAX, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], 0.0f, 0.0f, 0.0f, 0.0f });
 }
@@ -15687,6 +15807,14 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
    case GGML_OP_ARGMAX:
        ggml_vk_argmax(ctx, compute_ctx, src0, node);

+        break;
+    case GGML_OP_CROSS_ENTROPY_LOSS:
+        ggml_vk_cross_entropy_loss(ctx, compute_ctx, node);
+
+        break;
+    case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
+        ggml_vk_cross_entropy_loss_back(ctx, compute_ctx, node);
+
        break;
    case GGML_OP_COUNT_EQUAL:
        ggml_vk_count_equal(ctx, compute_ctx, src0, src1, node);
@@ -18511,6 +18639,18 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
            }
        case GGML_OP_ARGMAX:
            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_CROSS_ENTROPY_LOSS:
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32
+                && ggml_is_contiguous(op->src[1]) && op->src[1]->type == GGML_TYPE_F32
+                && ggml_are_same_shape(op->src[0], op->src[1])
+                && ggml_is_contiguous(op) && ggml_is_scalar(op) && op->type == GGML_TYPE_F32;
+        case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32 && ggml_is_scalar(op->src[0])
+                && ggml_is_contiguous(op->src[1]) && op->src[1]->type == GGML_TYPE_F32
+                && ggml_is_contiguous(op->src[2]) && op->src[2]->type == GGML_TYPE_F32
+                && ggml_are_same_shape(op->src[1], op->src[2])
+                && ggml_are_same_shape(op->src[1], op)
+                && ggml_is_contiguous(op) && op->type == GGML_TYPE_F32;
        case GGML_OP_COUNT_EQUAL:
            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_I32
                && ggml_is_contiguous(op->src[1]) && op->src[1]->type == GGML_TYPE_I32;
@@ -19437,6 +19577,10 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
            tensor_clone = ggml_mean(ggml_ctx, src_clone[0]);
        } else if (tensor->op == GGML_OP_ARGMAX) {
            tensor_clone = ggml_argmax(ggml_ctx, src_clone[0]);
+        } else if (tensor->op == GGML_OP_CROSS_ENTROPY_LOSS) {
+            tensor_clone = ggml_cross_entropy_loss(ggml_ctx, src_clone[0], src_clone[1]);
+        } else if (tensor->op == GGML_OP_CROSS_ENTROPY_LOSS_BACK) {
+            tensor_clone = ggml_cross_entropy_loss_back(ggml_ctx, src_clone[0], src_clone[1], src_clone[2]);
        } else if (tensor->op == GGML_OP_COUNT_EQUAL) {
            tensor_clone = ggml_count_equal(ggml_ctx, src_clone[0], src_clone[1]);
        } else if (tensor->op == GGML_OP_SOLVE_TRI) {
@@ -0,0 +1,78 @@
+#version 450
+
+#include "generic_head.glsl"
+#include "types.glsl"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(constant_id = 0) const uint BLOCK_SIZE = 32;
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+
+shared FLOAT_TYPE tmp[BLOCK_SIZE];
+
+FLOAT_TYPE wg_reduce_max(FLOAT_TYPE v) {
+    const uint tid = gl_LocalInvocationID.x;
+    tmp[tid] = v;
+    barrier();
+    [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] = max(tmp[tid], tmp[tid + s]);
+        }
+        barrier();
+    }
+    v = tmp[0];
+    barrier();
+    return v;
+}
+
+FLOAT_TYPE wg_reduce_sum(FLOAT_TYPE v) {
+    const uint tid = gl_LocalInvocationID.x;
+    tmp[tid] = v;
+    barrier();
+    [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+    v = tmp[0];
+    barrier();
+    return v;
+}
+
+void main() {
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
+    const uint tid = gl_LocalInvocationID.x;
+
+    if (row >= p.KY) {
+        return;
+    }
+
+    const uint off = row * p.KX;
+
+    FLOAT_TYPE max_logit = FLOAT_TYPE(uintBitsToFloat(0xFF800000));
+    for (uint i = tid; i < p.KX; i += BLOCK_SIZE) {
+        max_logit = max(max_logit, FLOAT_TYPE(data_a[off + i]));
+    }
+    max_logit = wg_reduce_max(max_logit);
+
+    FLOAT_TYPE sum_exp = FLOAT_TYPE(0.0f);
+    for (uint i = tid; i < p.KX; i += BLOCK_SIZE) {
+        sum_exp += exp(FLOAT_TYPE(data_a[off + i]) - max_logit);
+    }
+    const FLOAT_TYPE log_sum = log(wg_reduce_sum(sum_exp));
+
+    FLOAT_TYPE loss = FLOAT_TYPE(0.0f);
+    for (uint i = tid; i < p.KX; i += BLOCK_SIZE) {
+        loss += (FLOAT_TYPE(data_a[off + i]) - max_logit - log_sum) * FLOAT_TYPE(data_b[off + i]);
+    }
+    loss = -wg_reduce_sum(loss) / FLOAT_TYPE(p.KY);
+
+    if (tid == 0) {
+        data_d[row] = D_TYPE(loss);
+    }
+}
@@ -0,0 +1,75 @@
+#version 450
+
+#include "generic_head.glsl"
+#include "types.glsl"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(constant_id = 0) const uint BLOCK_SIZE = 32;
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer G {A_TYPE data_g[];};
+layout (binding = 1) readonly buffer X {B_TYPE data_x[];};
+layout (binding = 2) readonly buffer Y {B_TYPE data_y[];};
+layout (binding = 3) writeonly buffer D {D_TYPE data_d[];};
+
+shared FLOAT_TYPE tmp[BLOCK_SIZE];
+
+FLOAT_TYPE wg_reduce_max(FLOAT_TYPE v) {
+    const uint tid = gl_LocalInvocationID.x;
+    tmp[tid] = v;
+    barrier();
+    [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] = max(tmp[tid], tmp[tid + s]);
+        }
+        barrier();
+    }
+    v = tmp[0];
+    barrier();
+    return v;
+}
+
+FLOAT_TYPE wg_reduce_sum(FLOAT_TYPE v) {
+    const uint tid = gl_LocalInvocationID.x;
+    tmp[tid] = v;
+    barrier();
+    [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+    v = tmp[0];
+    barrier();
+    return v;
+}
+
+void main() {
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
+    const uint tid = gl_LocalInvocationID.x;
+
+    if (row >= p.KY) {
+        return;
+    }
+
+    const uint off = row * p.KX;
+    const FLOAT_TYPE d_by_nrows = FLOAT_TYPE(data_g[0]) / FLOAT_TYPE(p.KY);
+
+    FLOAT_TYPE max_logit = FLOAT_TYPE(uintBitsToFloat(0xFF800000));
+    for (uint i = tid; i < p.KX; i += BLOCK_SIZE) {
+        max_logit = max(max_logit, FLOAT_TYPE(data_x[off + i]));
+    }
+    max_logit = wg_reduce_max(max_logit);
+
+    FLOAT_TYPE sum_exp = FLOAT_TYPE(0.0f);
+    for (uint i = tid; i < p.KX; i += BLOCK_SIZE) {
+        sum_exp += exp(FLOAT_TYPE(data_x[off + i]) - max_logit);
+    }
+    const FLOAT_TYPE inv_sum = FLOAT_TYPE(1.0f) / wg_reduce_sum(sum_exp);
+
+    for (uint i = tid; i < p.KX; i += BLOCK_SIZE) {
+        const FLOAT_TYPE sm = exp(FLOAT_TYPE(data_x[off + i]) - max_logit) * inv_sum;
+        data_d[off + i] = D_TYPE((sm - FLOAT_TYPE(data_y[off + i])) * d_by_nrows);
+    }
+}
@@ -1029,6 +1029,8 @@ void process_shaders() {

    string_to_spv("argmax_f32", "argmax.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "int"}}));
    string_to_spv("sum_rows_f32", "sum_rows.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+    string_to_spv("cross_entropy_loss_f32", "cross_entropy_loss.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
+    string_to_spv("cross_entropy_loss_back_f32", "cross_entropy_loss_back.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
    string_to_spv("fwht_f32", "fwht.comp", {});
    string_to_spv("fwht_shmem_f32", "fwht.comp", {{"FWHT_SHMEM", "1"}});
    string_to_spv("count_equal_i32", "count_equal.comp", merge_maps(base_dict, {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}}));
@@ -43,10 +43,10 @@
 #define LLAMA_FILE_MAGIC_GGSQ 0x67677371u // 'ggsq'

 #define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
-#define LLAMA_SESSION_VERSION 9
+#define LLAMA_SESSION_VERSION 10

 #define LLAMA_STATE_SEQ_MAGIC   LLAMA_FILE_MAGIC_GGSQ
-#define LLAMA_STATE_SEQ_VERSION 2
+#define LLAMA_STATE_SEQ_VERSION 3

 #ifdef __cplusplus
 extern "C" {
@@ -12,6 +12,7 @@
 #include <limits>
 #include <map>
 #include <stdexcept>
+#include <unordered_map>

 static bool ggml_is_power_of_2(int n) {
    return (n & (n - 1)) == 0;
@@ -1128,11 +1129,18 @@ void llama_kv_cache::apply_ubatch(const slot_info & sinfo, const llama_ubatch &

            cells.pos_set(idx, ubatch.pos[i]);

-            if (ubatch.is_pos_2d()) {
-                llama_kv_cell_ext ext {
-                    /*.x =*/ ubatch.pos[i + ubatch.n_tokens*2],
-                    /*.y =*/ ubatch.pos[i + ubatch.n_tokens],
-                };
+            if (ubatch.is_pos_2d() || ubatch.token) {
+                llama_kv_cell_ext ext;
+
+                if (ubatch.is_pos_2d()) {
+                    ext.x = ubatch.pos[i + ubatch.n_tokens*2];
+                    ext.y = ubatch.pos[i + ubatch.n_tokens];
+                }
+
+                if (ubatch.token) {
+                    ext.tok = ubatch.token[i];
+                }
+
                cells.ext_set(idx, ext);
            }

@@ -1805,6 +1813,69 @@ void llama_kv_cache::set_input_v_rot(ggml_tensor * dst) const {
    memcpy(dst->data, attn_rot_hadamard.at(n_rot).data(), ggml_nbytes(dst));
 }

+bool llama_kv_cache::has_cell_ext() const {
+    return hparams.n_pos_per_embd() > 1;
+}
+
+void llama_kv_cache::get_prev_tokens(const llama_ubatch & ubatch, uint32_t n, std::vector<llama_token> & res) const {
+    const uint32_t n_tokens = ubatch.n_tokens;
+
+    res.clear();
+    res.resize(n_tokens*n, LLAMA_TOKEN_NULL);
+
+    if (n == 0) {
+        return;
+    }
+
+    // note: apply_ubatch() has already stored the current ubatch
+    //       the window below thus covers tokens of this very ubatch as well, which is what we want
+    llama_pos p_min = std::numeric_limits<llama_pos>::max();
+    llama_pos p_max = std::numeric_limits<llama_pos>::min();
+
+    std::bitset<LLAMA_MAX_SEQ> seqs;
+
+    for (uint32_t i = 0; i < n_tokens; ++i) {
+        p_min = std::min(p_min, ubatch.pos[i]);
+        p_max = std::max(p_max, ubatch.pos[i]);
+    }
+
+    for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
+        seqs.set(ubatch.seq_id_unq[s]);
+    }
+
+    // (seq_id, pos) -> token, for every cell that could be a predecessor of a ubatch token
+    std::unordered_map<uint64_t, llama_token> hist;
+
+    const auto key = [](llama_seq_id seq_id, llama_pos pos) {
+        return ((uint64_t) seq_id << 32) | (uint32_t) pos;
+    };
+
+    for (uint32_t s = 0; s < n_stream; ++s) {
+        v_cells[s].for_each_token_in(seqs, p_min - (llama_pos) n, p_max,
+            [&](llama_seq_id seq_id, llama_pos pos, llama_token tok) {
+                hist[key(seq_id, pos)] = tok;
+            });
+    }
+
+    for (uint32_t i = 0; i < n_tokens; ++i) {
+        // TODO: a token that belongs to more than one sequence has an ambiguous history.
+        //       the n-gram architectures have to reject such batches
+        const llama_seq_id seq_id = ubatch.seq_id[i][0];
+
+        for (uint32_t j = 0; j < n; ++j) {
+            const llama_pos p = ubatch.pos[i] - (llama_pos) (n - j);
+            if (p < 0) {
+                continue;
+            }
+
+            const auto it = hist.find(key(seq_id, p));
+            if (it != hist.end()) {
+                res[i*n + j] = it->second;
+            }
+        }
+    }
+}
+
 size_t llama_kv_cache::total_size() const {
    size_t size = 0;

@@ -2106,7 +2177,7 @@ void llama_kv_cache::state_write_meta(llama_io_write_i & io, const cell_ranges_t
            io.write(&pos,      sizeof(pos));
            io.write(&n_seq_id, sizeof(n_seq_id));

-            if (hparams.n_pos_per_embd() > 1) {
+            if (has_cell_ext()) {
                const llama_kv_cell_ext ext = cells.ext_get(i);
                io.write(&ext, sizeof(ext));
            }
@@ -2243,12 +2314,17 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
                return false;
            }

-            if (hparams.n_pos_per_embd() > 1) {
+            if (has_cell_ext()) {
                llama_kv_cell_ext ext;
                io.read(&ext, sizeof(ext));

-                ubatch.pos[i + ubatch.n_tokens]   = ext.y;
-                ubatch.pos[i + ubatch.n_tokens*2] = ext.x;
+                if (hparams.n_pos_per_embd() > 1) {
+                    ubatch.pos[i + ubatch.n_tokens]   = ext.y;
+                    ubatch.pos[i + ubatch.n_tokens*2] = ext.x;
+                }
+
+                // apply_ubatch() below restores ext.tok from the ubatch tokens
+                ubatch.token[i] = ext.tok;
            }

            // read the sequence id, but directly discard it - we will use dest_seq_id instead
@@ -2268,7 +2344,8 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
            return false;
        }

-        // TODO: we cannot yet restore llama_kv_cell_ext as the apply_ubatch() does not support it yet
+        // note: apply_ubatch() rebuilds llama_kv_cell_ext from the ubatch
+        //       only ext.tok and the M-RoPE 2D position round-trip through it
        //       see: https://github.com/ggml-org/llama.cpp/pull/16825#issuecomment-3460868350
        apply_ubatch(sinfo, ubatch);

@@ -2301,7 +2378,7 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32

            cells.pos_set(i, pos);

-            if (hparams.n_pos_per_embd() > 1) {
+            if (has_cell_ext()) {
                llama_kv_cell_ext ext;
                io.read(&ext, sizeof(ext));
                cells.ext_set(i, ext);
@@ -2652,3 +2729,7 @@ void llama_kv_cache_context::set_input_k_rot(ggml_tensor * dst) const {
 void llama_kv_cache_context::set_input_v_rot(ggml_tensor * dst) const {
    kv->set_input_v_rot(dst);
 }
+
+void llama_kv_cache_context::get_prev_tokens(const llama_ubatch & ubatch, uint32_t n, std::vector<llama_token> & res) const {
+    kv->get_prev_tokens(ubatch, n, res);
+}
@@ -219,6 +219,14 @@ public:
    void set_input_k_rot(ggml_tensor * dst) const;
    void set_input_v_rot(ggml_tensor * dst) const;

+    // true if llama_kv_cell_ext holds information that has to survive a state save/restore
+    bool has_cell_ext() const;
+
+    // for every token of the ubatch, the ids of the n tokens that precede it in its sequence
+    // entries with no matching cell are set to LLAMA_TOKEN_NULL
+    // note: used by n-gram input embeddings
+    void get_prev_tokens(const llama_ubatch & ubatch, uint32_t n, std::vector<llama_token> & res) const;
+
 private:
    const llama_model & model;
    const llama_hparams & hparams;
@@ -401,6 +409,9 @@ public:
    void set_input_k_rot(ggml_tensor * dst) const;
    void set_input_v_rot(ggml_tensor * dst) const;

+    // see llama_kv_cache::get_prev_tokens()
+    void get_prev_tokens(const llama_ubatch & ubatch, uint32_t n, std::vector<llama_token> & res) const;
+
 private:
    llama_memory_status status;

@@ -15,6 +15,10 @@ struct llama_kv_cell_ext {
    llama_pos x = 0;
    llama_pos y = 0;

+    // when tok = LLAMA_TOKEN_NULL when the cell is produced by embedding input (i.e. multimodal)
+    // use case: n-gram embeddings hash
+    llama_token tok = LLAMA_TOKEN_NULL;
+
    // return true if the current 2D spatial position is greater than other
    bool is_2d_gt(llama_pos ox, llama_pos oy) const {
        return (y > oy) || (y == oy && x > ox);
@@ -23,7 +27,7 @@ struct llama_kv_cell_ext {
    void reset() {
        static_assert(std::is_trivially_copyable_v<llama_kv_cell_ext>);

-        memset(this, 0, sizeof(*this));
+        *this = llama_kv_cell_ext{};
    }
 };

@@ -305,6 +309,29 @@ public:
        return seq[i].test(seq_id);
    }

+    // gather the token ids of the cells in `seqs` with position in [p0, p1)
+    // the callback receives (seq_id, pos, token) for every such (cell, seq) pair
+    // note: used by n-gram input embeddings to recover the tokens preceding a ubatch
+    template<typename F>
+    void for_each_token_in(const std::bitset<LLAMA_MAX_SEQ> & seqs, llama_pos p0, llama_pos p1, F && f) const {
+        for (const auto & i : used) {
+            if (pos[i] < p0 || pos[i] >= p1) {
+                continue;
+            }
+
+            const auto m = seq[i] & seqs;
+            if (m.none()) {
+                continue;
+            }
+
+            for (llama_seq_id s = 0; s < LLAMA_MAX_SEQ; ++s) {
+                if (m.test(s)) {
+                    f(s, pos[i], ext[i].tok);
+                }
+            }
+        }
+    }
+
    // note: call only if the cell is not empty and the seq_id is not in the cell
    void seq_add(uint32_t i, llama_seq_id seq_id) {
        assert(i < pos.size());
@@ -8,7 +8,8 @@
 		ChatFormInputFileInputInvisible,
 		ChatFormMcpResourcesList,
 		ChatFormPickers,
-		DialogMcpResourcesBrowser
+		DialogMcpResourcesBrowser,
+		DialogMcpServers
 	} from '$lib/components/app';
 	import {
 		CLIPBOARD_CONTENT_QUOTE_PREFIX,
@@ -183,6 +184,9 @@
 	let isResourceDialogOpen = $state(false);
 	let preSelectedResourceUri = $state<string | undefined>(undefined);

+	// MCP Servers Dialog State
+	let isMcpServersDialogOpen = $state(false);
+
 	let currentConfig = $derived(settingsStore.config);

 	let pasteLongTextToFileLength = $derived.by(() => {
@@ -618,6 +622,7 @@
 				onFileUpload={handleFileUpload}
 				onMcpPromptClick={showMcpPromptButton ? () => pickers.openPromptPicker() : undefined}
 				onMcpResourcesClick={() => (isResourceDialogOpen = true)}
+				onMcpSettingsClick={() => (isMcpServersDialogOpen = true)}
 				onMicClick={handleMicClick}
 				{onStop}
 				onSystemPromptClick={() => onSystemPromptClick?.({ files: uploadedFiles, message: value })}
@@ -656,3 +661,5 @@
 	}}
 	preSelectedUri={preSelectedResourceUri}
 />
+
+<DialogMcpServers bind:open={isMcpServersDialogOpen} />
@@ -1,10 +1,6 @@
 <script lang="ts">
-	import { File, FolderOpen, MessageSquare, Plus, Zap } from '@lucide/svelte';
-	import {
-		ChatFormActionAddMcpServersSubmenu,
-		ChatFormActionAddReasoningSubmenu,
-		ChatFormActionAddToolsSubmenu
-	} from '$lib/components/app';
+	import { File, MessageSquare, Plus } from '@lucide/svelte';
+	import { ChatFormActionAddToolsSubmenu, McpLogo } from '$lib/components/app';
 	import { buttonVariants } from '$lib/components/ui/button';
 	import * as DropdownMenu from '$lib/components/ui/dropdown-menu';
 	import * as Tooltip from '$lib/components/ui/tooltip';
@@ -31,11 +27,6 @@
 	// must not restore focus to the trigger on close
 	let suppressCloseAutoFocus = false;

-	function handleMcpSettingsClick() {
-		dropdownOpen = false;
-		chatFormActions.onMcpSettingsClick?.();
-	}
-
 	const attachmentMenu = useAttachmentMenu(
 		() => ({
 			hasAudioModality: chatFormActions.hasAudioModality,
@@ -93,10 +84,6 @@
 				}
 			}}
 		>
-			<ChatFormActionAddReasoningSubmenu />
-
-			<DropdownMenu.Separator />
-
 			<DropdownMenu.Sub>
 				<DropdownMenu.SubTrigger class="flex cursor-pointer items-center gap-2">
 					<File class={ICON_CLASS_DEFAULT} />
@@ -156,31 +143,14 @@

 			<ChatFormActionAddToolsSubmenu />

-			<ChatFormActionAddMcpServersSubmenu onMcpSettingsClick={handleMcpSettingsClick} />
+			<DropdownMenu.Item
+				class="flex cursor-pointer items-center gap-2"
+				onclick={chatFormActions.onMcpSettingsClick}
+			>
+				<McpLogo class={ICON_CLASS_DEFAULT} />

-			{#if chatFormActions.hasMcpPromptsSupport}
-				<DropdownMenu.Separator />
-
-				<DropdownMenu.Item
-					class="flex cursor-pointer items-center gap-2"
-					onclick={chatFormActions.onMcpPromptClick}
-				>
-					<Zap class={ICON_CLASS_DEFAULT} />
-
-					<span>MCP Prompt</span>
-				</DropdownMenu.Item>
-			{/if}
-
-			{#if chatFormActions.hasMcpResourcesSupport}
-				<DropdownMenu.Item
-					class="flex cursor-pointer items-center gap-2"
-					onclick={chatFormActions.onMcpResourcesClick}
-				>
-					<FolderOpen class={ICON_CLASS_DEFAULT} />
-
-					<span>MCP Resources</span>
-				</DropdownMenu.Item>
-			{/if}
+				<span>MCP Servers</span>
+			</DropdownMenu.Item>
 		</DropdownMenu.Content>
 	</DropdownMenu.Root>
 </div>
@@ -1,152 +0,0 @@
-<script lang="ts">
-	import { Plus, Settings } from '@lucide/svelte';
-	import { goto } from '$app/navigation';
-	import { DropdownMenuSearchable, McpLogo, McpServerIdentity } from '$lib/components/app';
-	import * as DropdownMenu from '$lib/components/ui/dropdown-menu';
-	import { Switch } from '$lib/components/ui/switch';
-	import { ICON_CLASS_DEFAULT, ROUTES } from '$lib/constants';
-	import { HealthCheckStatus } from '$lib/enums';
-	import { conversationsStore, mcpStore } from '$lib/stores';
-	import type { MCPServerSettingsEntry } from '$lib/types';
-
-	interface Props {
-		onMcpSettingsClick?: () => void;
-	}
-
-	let { onMcpSettingsClick }: Props = $props();
-
-	let mcpSearchQuery = $state('');
-	// Every configured server is listed; `enabled` is an on/off state,
-	// not a visibility filter, so a disabled server stays toggleable.
-	let mcpServers = $derived(mcpStore.getServers());
-	let hasMcpServers = $derived(mcpServers.length > 0);
-	let filteredMcpServers = $derived.by(() => {
-		const query = mcpSearchQuery.toLowerCase().trim();
-
-		if (!query) return mcpServers;
-
-		return mcpServers.filter((s) => {
-			const name = getServerLabel(s).toLowerCase();
-			const url = s.url.toLowerCase();
-
-			return name.includes(query) || url.includes(query);
-		});
-	});
-
-	function getServerLabel(server: MCPServerSettingsEntry): string {
-		return mcpStore.getServerLabel(server);
-	}
-
-	function isServerEnabledForChat(serverId: string): boolean {
-		return conversationsStore.preferences.isMcpServerEnabledForChat(serverId);
-	}
-
-	async function toggleServerForChat(serverId: string) {
-		await conversationsStore.preferences.toggleMcpServerForChat(serverId);
-	}
-
-	function handleMcpSubMenuOpen(open: boolean) {
-		if (open) {
-			mcpSearchQuery = '';
-			mcpStore.runHealthChecksForServers(mcpServers);
-		}
-	}
-
-	function handleMcpSettingsClick() {
-		onMcpSettingsClick?.();
-
-		goto(`${hasMcpServers ? '' : '?add'}${ROUTES.MCP_SERVERS}`);
-	}
-</script>
-
-<DropdownMenu.Root>
-	<DropdownMenu.Sub onOpenChange={handleMcpSubMenuOpen}>
-		<DropdownMenu.SubTrigger class="flex cursor-pointer items-center gap-2">
-			<McpLogo class={ICON_CLASS_DEFAULT} />
-
-			<span>MCP Servers</span>
-		</DropdownMenu.SubTrigger>
-
-		<DropdownMenu.SubContent class="w-72 pt-0">
-			{#if hasMcpServers}
-				<DropdownMenuSearchable
-					bind:searchValue={mcpSearchQuery}
-					emptyMessage="No servers found"
-					isEmpty={filteredMcpServers.length === 0}
-					placeholder="Search servers..."
-				>
-					<div class="max-h-64 overflow-y-auto">
-						{#each filteredMcpServers as server (server.id)}
-							{@const healthState = mcpStore.getHealthCheckState(server.id)}
-							{@const hasError = healthState.status === HealthCheckStatus.ERROR}
-							{@const isEnabledForChat = isServerEnabledForChat(server.id)}
-							{@const displayName = getServerLabel(server)}
-							{@const faviconUrl = mcpStore.getServerFavicon(server.id)}
-
-							<button
-								class="flex w-full items-center justify-between gap-2 rounded-sm px-2 py-2 text-left transition-colors hover:bg-accent disabled:cursor-not-allowed disabled:opacity-50"
-								disabled={hasError}
-								onclick={() => !hasError && toggleServerForChat(server.id)}
-								type="button"
-							>
-								<div class="flex min-w-0 flex-1 items-center gap-2">
-									<div class="min-w-0 flex-1">
-										<McpServerIdentity
-											{displayName}
-											{faviconUrl}
-											iconClass={ICON_CLASS_DEFAULT}
-											iconRounded="rounded-sm"
-											nameClass="text-sm"
-											showVersion={false}
-										/>
-									</div>
-
-									{#if hasError}
-										<span
-											class="shrink-0 rounded bg-destructive/15 px-1.5 py-0.5 text-xs text-destructive"
-										>
-											Error
-										</span>
-									{/if}
-								</div>
-
-								<Switch
-									checked={isEnabledForChat}
-									disabled={hasError}
-									onCheckedChange={() => toggleServerForChat(server.id)}
-									onclick={(e) => e.stopPropagation()}
-								/>
-							</button>
-						{/each}
-					</div>
-
-					{#snippet footer()}
-						<DropdownMenu.Item
-							class="flex cursor-pointer items-center gap-2"
-							onclick={handleMcpSettingsClick}
-						>
-							<Settings class={ICON_CLASS_DEFAULT} />
-
-							<span>Manage MCP Servers</span>
-						</DropdownMenu.Item>
-					{/snippet}
-				</DropdownMenuSearchable>
-			{:else}
-				<div class="px-2 py-3 text-center text-sm text-muted-foreground">
-					No MCP servers configured
-				</div>
-
-				<DropdownMenu.Separator />
-
-				<DropdownMenu.Item
-					class="flex cursor-pointer items-center gap-2"
-					onclick={handleMcpSettingsClick}
-				>
-					<Plus class={ICON_CLASS_DEFAULT} />
-
-					<span>Add MCP Servers</span>
-				</DropdownMenu.Item>
-			{/if}
-		</DropdownMenu.SubContent>
-	</DropdownMenu.Sub>
-</DropdownMenu.Root>
@@ -0,0 +1,51 @@
+<script lang="ts">
+	import { FolderOpen, Server, Zap } from '@lucide/svelte';
+	import { McpLogo } from '$lib/components/app';
+	import * as DropdownMenu from '$lib/components/ui/dropdown-menu';
+	import { ICON_CLASS_DEFAULT } from '$lib/constants';
+	import { getChatFormActionsContext } from '$lib/contexts';
+
+	const chatFormActions = getChatFormActionsContext();
+
+	function handleServersClick() {
+		chatFormActions.onMcpSettingsClick?.();
+	}
+</script>
+
+<DropdownMenu.Sub>
+	<DropdownMenu.SubTrigger class="flex cursor-pointer items-center gap-2">
+		<McpLogo class={ICON_CLASS_DEFAULT} />
+
+		<span>MCP</span>
+	</DropdownMenu.SubTrigger>
+
+	<DropdownMenu.SubContent class="w-48">
+		<DropdownMenu.Item class="flex cursor-pointer items-center gap-2" onclick={handleServersClick}>
+			<Server class={ICON_CLASS_DEFAULT} />
+
+			<span>Servers</span>
+		</DropdownMenu.Item>
+
+		{#if chatFormActions.hasMcpPromptsSupport}
+			<DropdownMenu.Item
+				class="flex cursor-pointer items-center gap-2"
+				onclick={chatFormActions.onMcpPromptClick}
+			>
+				<Zap class={ICON_CLASS_DEFAULT} />
+
+				<span>Prompts</span>
+			</DropdownMenu.Item>
+		{/if}
+
+		{#if chatFormActions.hasMcpResourcesSupport}
+			<DropdownMenu.Item
+				class="flex cursor-pointer items-center gap-2"
+				onclick={chatFormActions.onMcpResourcesClick}
+			>
+				<FolderOpen class={ICON_CLASS_DEFAULT} />
+
+				<span>Resources</span>
+			</DropdownMenu.Item>
+		{/if}
+	</DropdownMenu.SubContent>
+</DropdownMenu.Sub>
@@ -1,6 +1,5 @@
 <script lang="ts">
 	import { SkipForward, Square } from '@lucide/svelte';
-	import { goto } from '$app/navigation';
 	import { page } from '$app/state';
 	import {
 		ChatFormActionModels,
@@ -10,7 +9,7 @@
 		ChatFormContextGauge
 	} from '$lib/components/app';
 	import { Button } from '$lib/components/ui/button';
-	import { ICON_CLASS_DEFAULT, ROUTES } from '$lib/constants';
+	import { ICON_CLASS_DEFAULT } from '$lib/constants';
 	import { setChatFormActionsContext } from '$lib/contexts';
 	import { FileTypeCategory, MessageRole } from '$lib/enums';
 	import { ChatService } from '$lib/services';
@@ -34,6 +33,7 @@
 		onSystemPromptClick?: () => void;
 		onMcpPromptClick?: () => void;
 		onMcpResourcesClick?: () => void;
+		onMcpSettingsClick?: () => void;
 	}

 	let {
@@ -47,6 +47,7 @@
 		onFileUpload,
 		onMcpPromptClick,
 		onMcpResourcesClick,
+		onMcpSettingsClick,
 		onMicClick,
 		onStop,
 		onSystemPromptClick,
@@ -163,7 +164,7 @@
 			return onMcpResourcesClick;
 		},
 		get onMcpSettingsClick() {
-			return () => goto(ROUTES.MCP_SERVERS);
+			return onMcpSettingsClick;
 		},
 		get onSystemPromptClick() {
 			return onSystemPromptClick;
@@ -221,25 +221,17 @@ export { default as ChatFormActionModels } from './ChatForm/ChatFormActions/Chat
 export { default as ChatFormActionAddToolsSubmenu } from './ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddToolsSubmenu.svelte';

 /**
- * Dropdown submenu for managing MCP servers in the chat form.
+ * Dropdown submenu for MCP prompts and resources in the chat form.
 *
- * Displays a searchable list of enabled MCP servers with toggle switches
- * to enable/disable each server for chat. Shows server favicon, health status,
- * and a "Manage MCP Servers" settings link.
- *
- * Features:
- * - Search/filter servers by name or URL
- * - Per-server toggle to enable/disable for chat
- * - Health check indicator (shows "Error" badge for failed servers)
- * - Server favicon display
- * - Settings link to manage MCP server configuration
+ * Shows an "MCP" sub-menu item with entries for MCP Prompts and MCP
+ * Resources. Only visible when the server supports them.
 *
 * @example
 * ```svelte
- * <ChatFormActionAddMcpServersSubmenu onMcpSettingsClick={handleMcpSettingsClick} />
+ * <ChatFormActionAddMcpSubmenu />
 * ```
 */
-export { default as ChatFormActionAddMcpServersSubmenu } from './ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddMcpServersSubmenu.svelte';
+export { default as ChatFormActionAddMcpSubmenu } from './ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddMcpSubmenu.svelte';

 /**
 * Dropdown submenu for selecting reasoning effort level.
@@ -253,7 +253,7 @@
 </script>

 <Dialog.Root onOpenChange={handleOpenChange} {open}>
-	<Dialog.Content class="max-h-[80vh] !max-w-4xl overflow-hidden p-0">
+	<Dialog.Content class="max-h-[80vh] md:max-w-4xl! w-full! overflow-hidden p-0">
 		<Dialog.Header class="border-b border-border/30 px-6 py-4">
 			<Dialog.Title class="flex items-center gap-2">
 				<FolderOpen class="h-5 w-5" />
@@ -246,7 +246,7 @@
 </script>

 <Dialog.Root onOpenChange={handleOpenChange} {open}>
-	<Dialog.Content class="sm:max-w-2xl">
+	<Dialog.Content class="max-w-2xl!">
 		<Dialog.Header>
 			<Dialog.Title class="select-none">Add New MCP Server</Dialog.Title>
 		</Dialog.Header>
@@ -0,0 +1,33 @@
+<script lang="ts">
+	import { McpLogo } from '$lib/components/app';
+	import { SettingsMcpServers } from '$lib/components/app/settings';
+	import * as Dialog from '$lib/components/ui/dialog';
+
+	interface Props {
+		open?: boolean;
+		onOpenChange?: (open: boolean) => void;
+	}
+
+	let { onOpenChange, open = $bindable(false) }: Props = $props();
+
+	function handleOpenChange(value: boolean) {
+		open = value;
+		onOpenChange?.(value);
+	}
+</script>
+
+<Dialog.Root onOpenChange={handleOpenChange} {open}>
+	<Dialog.Content
+		class="md:h-[calc(100vh-4rem)]! md:max-h-240! md:w-[calc(100vw-4rem)]! md:max-w-360! flex flex-col"
+	>
+		<Dialog.Header>
+			<Dialog.Title class="flex items-center gap-2">
+				<McpLogo class="h-5 w-5" />
+
+				<span>MCP Servers</span>
+			</Dialog.Title>
+		</Dialog.Header>
+
+		<SettingsMcpServers class="mt-4" />
+	</Dialog.Content>
+</Dialog.Root>
@@ -14,6 +14,7 @@
 <Dialog.Root bind:open {onOpenChange}>
 	<Dialog.Content
 		class="z-999999 grid max-h-full max-w-full! grid-rows-[1fr_auto] overflow-hidden p-0 md:h-[90vh] md:max-w-[90vw]!"
+		showCloseButton
 	>
 		<MermaidPreview {svgHtml} />
 	</Dialog.Content>
@@ -0,0 +1,34 @@
+<script lang="ts">
+	import { Settings } from '@lucide/svelte';
+	import { SettingsChat } from '$lib/components/app/settings';
+	import * as Dialog from '$lib/components/ui/dialog';
+
+	interface Props {
+		open?: boolean;
+		onOpenChange?: (open: boolean) => void;
+		initialSection?: string;
+	}
+
+	let { initialSection, onOpenChange, open = $bindable(false) }: Props = $props();
+
+	function handleOpenChange(value: boolean) {
+		open = value;
+		onOpenChange?.(value);
+	}
+</script>
+
+<Dialog.Root onOpenChange={handleOpenChange} {open}>
+	<Dialog.Content
+		class="md:h-[calc(100vh-4rem)]! md:max-h-240! md:w-[calc(100vw-4rem)]! md:max-w-6xl! flex flex-col p-0 md:p-6 gap-0"
+	>
+		<Dialog.Header class="md:p-0 p-4">
+			<Dialog.Title class="flex items-center gap-2">
+				<Settings class="h-5 w-5" />
+
+				<span>Settings</span>
+			</Dialog.Title>
+		</Dialog.Header>
+
+		<SettingsChat {initialSection} onClose={() => (open = false)} onSectionChange={() => {}} />
+	</Dialog.Content>
+</Dialog.Root>
@@ -18,6 +18,23 @@
 */
 export { default as DialogMcpServerAddNew } from './DialogMcpServerAddNew.svelte';

+/**
+ * **DialogMcpServers** - MCP servers dialog shown from the chat form
+ *
+ * Shows the same MCP server list as the `/mcp-servers` route inside a modal
+ * dialog.
+ */
+export { default as DialogMcpServers } from './DialogMcpServers.svelte';
+
+/**
+ * **DialogSettingsChat** - Chat settings shown in a modal dialog
+ *
+ * Wraps the full SettingsChat layout (sidebar, mobile header, fields, footer)
+ * inside a ShadCN Dialog instead of a dedicated route. Section switching is
+ * handled in-app via `onSectionChange` rather than URL navigation.
+ */
+export { default as DialogSettingsChat } from './DialogSettingsChat.svelte';
+
 /**
 * **DialogExportSettings** - Settings export dialog with sensitive data warning
 *
@@ -1,13 +1,22 @@
 <script lang="ts">
-	import { Database, FileText, ListChecks, MessageSquare, Sparkles, Wrench } from '@lucide/svelte';
+	import {
+		Database,
+		ExternalLink,
+		FileText,
+		ListChecks,
+		MessageSquare,
+		Sparkles,
+		Wrench
+	} from '@lucide/svelte';
 	import { Badge } from '$lib/components/ui/badge';
 	import type { MCPCapabilitiesInfo } from '$lib/types';

 	interface Props {
 		capabilities?: MCPCapabilitiesInfo;
+		onBrowseResources?: () => void;
 	}

-	let { capabilities }: Props = $props();
+	let { capabilities, onBrowseResources }: Props = $props();
 </script>

 {#if capabilities}
@@ -20,10 +29,24 @@
 	{/if}

 	{#if capabilities.server.resources}
-		<Badge class="h-5 gap-1 bg-blue-50 px-1.5 text-[10px] dark:bg-blue-950" variant="outline">
+		<Badge
+			class="h-5 cursor-pointer gap-1 bg-blue-50 px-1.5 text-[10px] transition-colors hover:bg-blue-100 dark:bg-blue-950 dark:hover:bg-blue-900"
+			onclick={onBrowseResources}
+			onkeydown={(e) => {
+				if (e.key === 'Enter' || e.key === ' ') {
+					e.preventDefault();
+					onBrowseResources?.();
+				}
+			}}
+			role="button"
+			tabindex={0}
+			variant="outline"
+		>
 			<Database class="h-3 w-3 text-blue-600 dark:text-blue-400" />

 			Resources
+
+			<ExternalLink class="h-3 w-3 text-blue-600 dark:text-blue-400" />
 		</Badge>
 	{/if}

@@ -22,9 +22,10 @@
 		onToggle: (enabled: boolean) => void;
 		onUpdate: (updates: Partial<MCPServerSettingsEntry>) => void;
 		onDelete: () => void;
+		onBrowseResources?: () => void;
 	}

-	let { enabled, onDelete, onToggle, onUpdate, server }: Props = $props();
+	let { enabled, onBrowseResources, onDelete, onToggle, onUpdate, server }: Props = $props();

 	let healthState = $derived<HealthCheckState>(mcpStore.getHealthCheckState(server.id));
 	let displayName = $derived(mcpStore.getServerLabel(server));
@@ -125,6 +126,7 @@
 			{displayName}
 			enabled={enabled ?? server.enabled}
 			{faviconUrl}
+			{onBrowseResources}
 			{onToggle}
 			{serverInfo}
 			{transportType}
@@ -191,12 +193,14 @@
 				</div>
 			{/if}

-			<McpServerCardActions
-				{isHealthChecking}
-				onDelete={handleDeleteClick}
-				onEdit={startEditing}
-				onRefresh={handleHealthCheck}
-			/>
+			<div class="flex items-center gap-2">
+				<McpServerCardActions
+					{isHealthChecking}
+					onDelete={handleDeleteClick}
+					onEdit={startEditing}
+					onRefresh={handleHealthCheck}
+				/>
+			</div>
 		</div>
 	{/if}
 </Card.Root>
@@ -12,6 +12,7 @@
 		enabled: boolean;
 		disabled?: boolean;
 		onToggle: (enabled: boolean) => void;
+		onBrowseResources?: () => void;
 		serverInfo?: MCPServerInfo;
 		capabilities?: MCPCapabilitiesInfo;
 		transportType?: MCPTransportType;
@@ -23,6 +24,7 @@
 		displayName,
 		enabled,
 		faviconUrl,
+		onBrowseResources,
 		onToggle,
 		serverInfo,
 		transportType
@@ -57,7 +59,7 @@
 					{/if}

 					{#if capabilities}
-						<McpCapabilitiesBadges {capabilities} />
+						<McpCapabilitiesBadges {capabilities} {onBrowseResources} />
 					{/if}
 				</div>
 			{/if}
@@ -5,6 +5,7 @@
 	import {
 		ActionIcon,
 		DialogConversationRename,
+		DialogSettingsChat,
 		Logo,
 		SidebarNavigationActions,
 		SidebarNavigationConversationList
@@ -91,6 +92,7 @@
 	let selectedIds = new SvelteSet<string>();

 	let renameDialogOpen = $state(false);
+	let settingsDialogOpen = $state(false);
 	let renameTargetConversationId = $state<string | null>(null);
 	let renameDraft = $state('');
 	let renameOriginalTitle = $state('');
@@ -308,7 +310,7 @@

 <svelte:window bind:innerWidth onkeydown={handleKeydown} />

-{#if innerWidth > 768 || (!page.url.hash.includes(ROUTES.SETTINGS) && !page.url.hash.includes(ROUTES.MCP_SERVERS) && !page.url.hash.includes(ROUTES.SEARCH))}
+{#if innerWidth > 768 || !page.url.hash.includes(ROUTES.SEARCH)}
 	<aside
 		class={[
 			'fixed md:sticky top-2 left-2 md:left-0 md:ml-2 md:mt-2 pt-2 z-10 w-[calc(100dvw-1rem)]',
@@ -400,6 +402,7 @@
 					isSearchModeActive = false;
 					searchQuery = '';
 				}}
+				onSettingsClick={() => (settingsDialogOpen = true)}
 			/>

 			{#if uiStore.isSidebarExpanded || isOnMobile}
@@ -447,6 +450,8 @@
 	onConfirm={handleRenameConfirm}
 />

+<DialogSettingsChat bind:open={settingsDialogOpen} />
+
 <style>
 	aside {
 		@media (max-width: 768px) {
@@ -26,6 +26,7 @@
 		onSearchDeactivated?: () => void;
 		onSearchClick?: () => void;
 		onNewChat?: () => void;
+		onSettingsClick?: () => void;
 	}

 	let {
@@ -35,6 +36,7 @@
 		onNewChat,
 		onSearchClick,
 		onSearchDeactivated,
+		onSettingsClick,
 		searchQuery = $bindable('')
 	}: Props = $props();

@@ -115,14 +117,16 @@
 							onNewChat?.();
 							void conversationsStore.openNewChat();
 						}
-					: item.route
-						? () => {
-								onNewChat?.();
-								goto(item.route!);
-							}
-						: isSearchOnMobile
-							? undefined
-							: onSearchClick}
+					: item.action === SidebarAction.SETTINGS
+						? () => onSettingsClick?.()
+						: item.route
+							? () => {
+									onNewChat?.();
+									goto(item.route!);
+								}
+							: isSearchOnMobile
+								? undefined
+								: onSearchClick}
 			{@const itemTransition = {
 				delay: !initialized ? i * ICON_STRIP_TRANSITION_DELAY_MULTIPLIER : 0,
 				duration: ICON_STRIP_TRANSITION_DURATION,
@@ -169,14 +173,16 @@
 							onNewChat?.();
 							void conversationsStore.openNewChat();
 						}
-					: item.route
-						? () => {
-								onNewChat?.();
-								goto(item.route!);
-							}
-						: isSearchOnMobile
-							? undefined
-							: onSearchClick}
+					: item.action === SidebarAction.SETTINGS
+						? () => onSettingsClick?.()
+						: item.route
+							? () => {
+									onNewChat?.();
+									goto(item.route!);
+								}
+							: isSearchOnMobile
+								? undefined
+								: onSearchClick}
 			{@const itemTransition = {
 				delay: !initialized ? i * ICON_STRIP_TRANSITION_DELAY_MULTIPLIER : 0,
 				duration: ICON_STRIP_TRANSITION_DURATION,
@@ -1,7 +1,5 @@
 <script lang="ts">
 	import { RefreshCw } from '@lucide/svelte';
-	import { goto } from '$app/navigation';
-	import { page } from '$app/state';
 	import {
 		SettingsChatDesktopSidebar,
 		SettingsChatFields,
@@ -18,21 +16,29 @@
 		SETTINGS_SECTION_SLUGS
 	} from '$lib/constants';
 	import { ColorMode } from '$lib/enums/ui.enums';
-	import { RouterService } from '$lib/services/router.service';
-	import { modelsStore, serverStore, settingsReferrer, settingsStore } from '$lib/stores';
-	import type { SettingsSection } from '$lib/types';
+	import { modelsStore, serverStore, settingsStore } from '$lib/stores';
+	import type { SettingsSection, SettingsSectionTitle } from '$lib/types';
 	import { setMode } from 'mode-watcher';
 	import { fade } from 'svelte/transition';
 	interface Props {
 		initialSection?: string;
-		getSectionHref?: (section: SettingsSection) => string;
+		onSectionChange?: (section: SettingsSectionTitle) => void;
+		onClose?: () => void;
 	}

-	let { getSectionHref, initialSection }: Props = $props();
+	let { initialSection, onClose, onSectionChange }: Props = $props();

-	let activeSlug = $derived(
-		initialSection ?? (page.params as Record<string, string | undefined>).section ?? 'general'
-	);
+	let activeSlug = $derived(initialSection ?? 'general');
+
+	function handleSectionChange(section: SettingsSectionTitle) {
+		const found = SETTINGS_CHAT_SECTIONS.find((s) => s.title === section);
+
+		if (found) {
+			activeSlug = found.slug;
+		}
+
+		onSectionChange?.(section);
+	}

 	let currentSection = $derived(
 		SETTINGS_CHAT_SECTIONS.find((section) => section.slug === activeSlug) ||
@@ -115,7 +121,7 @@
 		}

 		settingsStore.updateMultipleConfig(processedConfig);
-		goto(settingsReferrer.url);
+		onClose?.();
 	}

 	export function reset() {
@@ -123,32 +129,24 @@
 	}
 </script>

-<div in:fade={{ duration: 150 }} class="mx-auto flex h-full w-full flex-col md:pl-8">
-	<div class="flex flex-1 flex-col gap-4 md:flex-row">
+<div in:fade={{ duration: 150 }} class="mx-auto flex h-full w-full flex-col">
+	<div class="flex flex-1 flex-col md:flex-row md:gap-4">
 		<SettingsChatDesktopSidebar
-			getHref={getSectionHref ??
-				((section: SettingsSection) => RouterService.settings(section.slug))}
 			isActive={(section: SettingsSection) => section.slug === activeSlug}
+			onSectionChange={handleSectionChange}
 			sections={SETTINGS_CHAT_SECTIONS}
 		/>

 		<SettingsChatMobileHeader
 			bind:this={mobileHeader}
-			getHref={getSectionHref ??
-				((section: SettingsSection) => RouterService.settings(section.slug))}
 			isActive={(section: SettingsSection) => section.slug === activeSlug}
+			onSectionChange={handleSectionChange}
 			sections={SETTINGS_CHAT_SECTIONS}
 		/>

-		<div class="mx-auto max-w-3xl flex-1">
-			<div class="space-y-6 p-4 md:p-6 md:pt-28">
+		<div class="mx-auto max-w-2xl px-4 flex-1 md:mt-4">
+			<div class="space-y-6 pt-3">
 				<div class="grid">
-					<div class="mb-6 flex items-center gap-2 border-b border-border/30 pb-6 md:flex">
-						<currentSection.icon class="h-5 w-5" />
-
-						<h3 class="text-lg font-semibold">{currentSection.title}</h3>
-					</div>
-
 					{#if currentSection.slug === SETTINGS_SECTION_SLUGS.TOOLS}
 						<SettingsChatToolsTab />
 					{:else if currentSection.slug === SETTINGS_SECTION_SLUGS.IMPORT_EXPORT}
@@ -1,54 +1,31 @@
 <script lang="ts">
-	import { Settings } from '@lucide/svelte';
 	import { ICON_CLASS_DEFAULT } from '$lib/constants';
 	import type { SettingsSection, SettingsSectionTitle } from '$lib/types';

 	interface Props {
 		sections: SettingsSection[];
 		isActive: (section: SettingsSection) => boolean;
-		getHref?: (section: SettingsSection) => string;
 		onSectionChange?: (section: SettingsSectionTitle) => void;
 	}

-	let { getHref, isActive, onSectionChange, sections }: Props = $props();
+	let { isActive, onSectionChange, sections }: Props = $props();
 </script>

-<div class="sticky top-2 hidden w-64 flex-col self-start bg-background py-4 md:flex gap-6">
-	<div class="flex items-center gap-2 py-2">
-		<Settings class="h-5 w-5 md:h-6 md:w-6" />
-
-		<h1 class="text-xl font-semibold md:text-2xl">Settings</h1>
-	</div>
-
+<div class="sticky top-12 hidden w-64 flex-col self-start bg-background md:flex gap-6">
 	<nav class="space-y-1">
 		{#each sections as section (section.title)}
-			{#if getHref}
-				<a
-					class="flex w-full cursor-pointer items-center gap-3 rounded-lg px-3 py-2 text-left text-sm no-underline transition-colors hover:bg-accent {isActive(
-						section
-					)
-						? 'bg-accent text-accent-foreground'
-						: 'text-muted-foreground'}"
-					href={getHref(section)}
-				>
-					<section.icon class={ICON_CLASS_DEFAULT} />
+			<button
+				class="flex w-full cursor-pointer items-center gap-3 rounded-lg px-3 py-2 text-left text-sm transition-colors hover:bg-accent {isActive(
+					section
+				)
+					? 'bg-accent text-accent-foreground'
+					: 'text-muted-foreground'}"
+				onclick={() => onSectionChange?.(section.title)}
+			>
+				<section.icon class={ICON_CLASS_DEFAULT} />

-					<span class="ml-2">{section.title}</span>
-				</a>
-			{:else}
-				<button
-					class="flex w-full cursor-pointer items-center gap-3 rounded-lg px-3 py-2 text-left text-sm transition-colors hover:bg-accent {isActive(
-						section
-					)
-						? 'bg-accent text-accent-foreground'
-						: 'text-muted-foreground'}"
-					onclick={() => onSectionChange?.(section.title)}
-				>
-					<section.icon class={ICON_CLASS_DEFAULT} />
-
-					<span class="ml-2">{section.title}</span>
-				</button>
-			{/if}
+				<span class="ml-2">{section.title}</span>
+			</button>
 		{/each}
 	</nav>
 </div>
@@ -1,5 +1,4 @@
 <script lang="ts">
-	import { Settings } from '@lucide/svelte';
 	import { ScrollCarousel } from '$lib/components/app';
 	import { ICON_CLASS_DEFAULT, UI_DATA_ATTRS } from '$lib/constants';
 	import { BooleanString } from '$lib/enums';
@@ -10,11 +9,10 @@
 	interface Props {
 		sections: SettingsSection[];
 		isActive: (section: SettingsSection) => boolean;
-		getHref?: (section: SettingsSection) => string;
 		onSectionChange?: (section: SettingsSectionTitle) => void;
 	}

-	let { getHref, isActive, onSectionChange, sections }: Props = $props();
+	let { isActive, onSectionChange, sections }: Props = $props();

 	const carousel = useScrollCarousel();

@@ -37,51 +35,26 @@
 	}
 </script>

-<div class="sticky top-0 z-10 flex flex-col bg-background md:hidden">
-	<div class="flex items-center gap-2 px-4 pt-4 pb-2 md:pt-6">
-		<Settings class="h-5 w-5 md:h-6 md:w-6" />
-
-		<h1 class="text-xl font-semibold md:text-2xl">Settings</h1>
-	</div>
-
-	<div class="border-b border-border/30 py-2">
+<div class="flex flex-col bg-background md:hidden sticky top-13 z-50">
+	<div class="border-b border-border/30">
 		<ScrollCarousel alwaysShowArrows {carousel} containerClass="py-2" innerClass="gap-2">
 			{#each sections as section (section.title)}
-				{#if getHref}
-					<a
-						class="flex cursor-pointer items-center gap-2 rounded-lg px-3 py-2 text-sm whitespace-nowrap no-underline transition-colors first:ml-4 last:mr-4 hover:bg-accent {isActive(
-							section
-						)
-							? 'bg-accent text-accent-foreground'
-							: 'text-muted-foreground'}"
-						{...{ [UI_DATA_ATTRS.ACTIVE]: isActive(section) }}
-						href={getHref(section)}
-						onclick={(e: MouseEvent) => {
-							carousel.scrollToCenter(e.currentTarget as HTMLElement);
-						}}
-					>
-						<section.icon class="{ICON_CLASS_DEFAULT} flex-shrink-0" />
+				<button
+					class="flex cursor-pointer items-center gap-2 rounded-lg px-3 py-2 text-sm whitespace-nowrap transition-colors first:ml-4 last:mr-4 hover:bg-accent {isActive(
+						section
+					)
+						? 'bg-accent text-accent-foreground'
+						: 'text-muted-foreground'}"
+					{...{ [UI_DATA_ATTRS.ACTIVE]: isActive(section) }}
+					onclick={(e: MouseEvent) => {
+						onSectionChange?.(section.title);
+						carousel.scrollToCenter(e.currentTarget as HTMLElement);
+					}}
+				>
+					<section.icon class="{ICON_CLASS_DEFAULT} flex-shrink-0" />

-						<span>{section.title}</span>
-					</a>
-				{:else}
-					<button
-						class="flex cursor-pointer items-center gap-2 rounded-lg px-3 py-2 text-sm whitespace-nowrap transition-colors first:ml-4 last:mr-4 hover:bg-accent {isActive(
-							section
-						)
-							? 'bg-accent text-accent-foreground'
-							: 'text-muted-foreground'}"
-						{...{ [UI_DATA_ATTRS.ACTIVE]: isActive(section) }}
-						onclick={(e: MouseEvent) => {
-							onSectionChange?.(section.title);
-							carousel.scrollToCenter(e.currentTarget as HTMLElement);
-						}}
-					>
-						<section.icon class="{ICON_CLASS_DEFAULT} flex-shrink-0" />
-
-						<span>{section.title}</span>
-					</button>
-				{/if}
+					<span>{section.title}</span>
+				</button>
 			{/each}
 		</ScrollCarousel>
 	</div>
@@ -29,7 +29,7 @@
 	}
 </script>

-<div class="sticky bottom-0 mx-auto mt-4 flex w-full justify-between p-6">
+<div class="sticky bottom-0 mx-auto mt-4 flex w-full justify-between pb-4 md:pb-0">
 	<div class="flex gap-2">
 		<Button onclick={handleResetClick} variant="outline">
 			<RotateCcw class="h-3 w-3" />
@@ -1,14 +1,11 @@
 <script lang="ts">
-	import McpLogo from '../mcp/McpLogo.svelte';
-	import { Plus, X } from '@lucide/svelte';
-	import { browser } from '$app/environment';
-	import { goto, replaceState } from '$app/navigation';
+	import { Plus } from '@lucide/svelte';
+	import { replaceState } from '$app/navigation';
 	import { page } from '$app/state';
-	import { ActionIcon, McpServerCard, McpServerCardSkeleton } from '$lib/components/app';
-	import { DialogMcpServerAddNew } from '$lib/components/app/dialogs';
+	import { McpServerCard, McpServerCardSkeleton } from '$lib/components/app';
+	import { DialogMcpResourcesBrowser, DialogMcpServerAddNew } from '$lib/components/app/dialogs';
 	import { Button } from '$lib/components/ui/button';
 	import * as Empty from '$lib/components/ui/empty';
-	import { ROUTES } from '$lib/constants';
 	import { HealthCheckStatus } from '$lib/enums';
 	import { conversationsStore, mcpStore, toolsStore } from '$lib/stores';
 	import { onMount } from 'svelte';
@@ -23,26 +20,7 @@
 	let servers = $derived(mcpStore.getServers());

 	let isAddingServer = $state(false);
-
-	let previousRouteId = $state<string | null>(null);
-
-	$effect(() => {
-		const currentId = page.route.id;
-
-		return () => {
-			previousRouteId = currentId;
-		};
-	});
-
-	function handleClose() {
-		const prevIsMcpServers = previousRouteId === '/mcp-servers';
-
-		if (browser && window.history.length > 1 && !prevIsMcpServers) {
-			history.back();
-		} else {
-			goto(ROUTES.START);
-		}
-	}
+	let isResourcesDialogOpen = $state(false);

 	onMount(() => {
 		if (page.url.searchParams.has('add')) {
@@ -71,25 +49,13 @@
 	}
 </script>

-<div in:fade={{ duration: 150 }} class="flex min-h-[calc(100dvh-4rem)] flex-col">
-	<div class="fixed top-4.5 right-4 z-50 md:hidden">
-		<ActionIcon icon={X} onclick={handleClose} tooltip="Close" />
-	</div>
-
-	<div
-		class="sticky top-0 z-10 mt-4 mb-2 flex items-start gap-4 md:p-4 p-0 px-4 md:justify-between md:px-8"
-	>
-		<div class="flex items-center gap-2">
-			<McpLogo class="h-5 w-5 md:h-6 md:w-6" />
-
-			<h1 class="text-lg font-semibold md:text-2xl">MCP Servers</h1>
-		</div>
-	</div>
-
+<div in:fade={{ duration: 150 }} class="flex flex-col h-full">
 	<DialogMcpServerAddNew bind:open={isAddingServer} />

+	<DialogMcpResourcesBrowser bind:open={isResourcesDialogOpen} />
+
 	{#if servers.length === 0}
-		<div class="flex flex-1 items-center justify-center py-16">
+		<div class="flex flex-1 items-center justify-center pb-20 pt-10 my-auto">
 			<Empty.Root class="max-w-md">
 				<Empty.Header>
 					<Empty.Media variant="icon">
@@ -112,8 +78,8 @@
 		</div>
 	{:else}
 		<div
-			class="grid gap-3 {className}"
-			style="grid-template-columns: repeat(auto-fill, minmax(min(32rem, calc(100dvw - 2rem)), 1fr));"
+			class="grid gap-4 {className}"
+			style="grid-template-columns: repeat(auto-fill, minmax(min(25rem, calc(100dvw - 4rem)), 1fr));"
 		>
 			{#each servers as server (server.id)}
 				{#if isServerPending(server.id, server.enabled)}
@@ -121,6 +87,7 @@
 				{:else}
 					<McpServerCard
 						enabled={conversationsStore.preferences.isMcpServerEnabledForChat(server.id)}
+						onBrowseResources={() => (isResourcesDialogOpen = true)}
 						onDelete={() => mcpStore.removeServer(server.id)}
 						onToggle={async () => {
 							const wasEnabled = conversationsStore.preferences.isMcpServerEnabledForChat(
@@ -1,21 +1,21 @@
 /**
 * Full chat settings page layout with sidebar, mobile header, and content area.
 * Manages local configuration state, section navigation, and context setup.
- * Accepts an optional `initialSection` prop to override the URL-based section resolution.
+ * Accepts an optional `initialSection` prop to set the initial active section.
 */
 export { default as SettingsChat } from './SettingsChat/SettingsChat.svelte';

 /**
 * Desktop sidebar navigation for chat settings.
 * Displays a list of settings sections with icons and titles.
- * Supports both hash-link navigation (via `getHref`) and in-app section switching (via `onSectionChange`).
+ * Switches sections in-app via `onSectionChange`.
 */
 export { default as SettingsChatDesktopSidebar } from './SettingsChatDesktopSidebar.svelte';

 /**
 * Mobile header with a horizontally scrollable section picker for chat settings.
 * Shows chevron buttons for scroll navigation and highlights the active section.
- * Supports both hash-link navigation (via `getHref`) and in-app section switching (via `onSectionChange`).
+ * Switches sections in-app via `onSectionChange`.
 */
 export { default as SettingsChatMobileHeader } from './SettingsChatMobileHeader.svelte';

@@ -5,3 +5,9 @@
 </script>

 <DialogPrimitive.Close bind:ref data-slot="dialog-close" {...restProps} />
+
+<style>
+	:global([data-dialog-close]) {
+		z-index: 999;
+	}
+</style>
@@ -10,7 +10,7 @@
 		class: className,
 		portalProps,
 		ref = $bindable(null),
-		showCloseButton = true,
+		showCloseButton = false,
 		...restProps
 	}: WithoutChildrenOrChild<DialogPrimitive.ContentProps> & {
 		portalProps?: DialogPrimitive.PortalProps;
@@ -25,7 +25,8 @@
 	<DialogPrimitive.Content
 		bind:ref
 		class={cn(
-			`fixed top-[50%] left-[50%] z-50 grid max-h-[100dvh] w-full max-w-[calc(100%-2rem)] translate-x-[-50%] translate-y-[-50%] gap-4 overflow-y-auto rounded-lg border border-border/30 bg-background p-6 shadow-lg duration-200 data-[state=closed]:animate-out data-[state=closed]:fade-out-0 data-[state=closed]:fill-mode-forwards data-[state=closed]:zoom-out-95 data-[state=open]:animate-in data-[state=open]:fade-in-0 data-[state=open]:zoom-in-95 sm:max-w-lg md:max-h-[100vh]`,
+			`fixed top-[50%] left-[50%] z-50 grid translate-x-[-50%] translate-y-[-50%] gap-4 overflow-y-auto rounded-lg border border-border/30 bg-background p-6 shadow-lg duration-200 data-[state=closed]:animate-out data-[state=closed]:fade-out-0 data-[state=closed]:fill-mode-forwards data-[state=closed]:zoom-out-95 data-[state=open]:animate-in data-[state=open]:fade-in-0 data-[state=open]:zoom-in-95`,
+			'max-h-[100dvh] max-w-full size-auto sm:max-w-lg md:w-auto md:size-auto md:max-w-[calc(100%-2rem)]',
 			className
 		)}
 		data-slot="dialog-content"
@@ -1,20 +1,43 @@
 <script lang="ts">
+	import XIcon from '@lucide/svelte/icons/x';
 	import { cn, type WithElementRef } from '$lib/components/ui/utils';
+	import { Dialog as DialogPrimitive } from 'bits-ui';
 	import type { HTMLAttributes } from 'svelte/elements';

 	let {
 		children,
 		class: className,
 		ref = $bindable(null),
+		showCloseButton = true,
 		...restProps
-	}: WithElementRef<HTMLAttributes<HTMLDivElement>> = $props();
+	}: WithElementRef<HTMLAttributes<HTMLDivElement>> & {
+		showCloseButton?: boolean;
+	} = $props();
 </script>

+<!--
+	Header is `sticky`, so it stays at the top while the dialog body scrolls. The close
+	button lives here (not in the body) so it sticks together with the title. `sticky`
+	makes it the containing block, so the close can be absolutely placed at its corner.
+-->
 <div
 	bind:this={ref}
-	class={cn('flex flex-col gap-2 text-center sm:text-left', className)}
+	class={cn(
+		'flex flex-col gap-2 text-center sm:text-left sticky top-0 z-50 bg-background md:bg-transparent',
+		className
+	)}
 	data-slot="dialog-header"
 	{...restProps}
 >
 	{@render children?.()}
+
+	{#if showCloseButton}
+		<DialogPrimitive.Close
+			class="absolute top-0 right-0 rounded-xs opacity-70 ring-offset-background transition-opacity hover:opacity-100 focus:ring-2 focus:ring-ring focus:ring-offset-2 focus:outline-hidden disabled:pointer-events-none [&_svg]:pointer-events-none [&_svg]:shrink-0 [&_svg:not([class*='size-'])]:size-4"
+		>
+			<XIcon />
+
+			<span class="sr-only">Close</span>
+		</DialogPrimitive.Close>
+	{/if}
 </div>
@@ -15,10 +15,6 @@ export const ROUTES = {
 	MCP_SERVERS: '#/mcp-servers',
 	/** Search — mobile-only full-page conversation search. */
 	SEARCH: '#/search',
-	/** Settings base — for dynamic settings URLs use RouterService. */
-	SETTINGS: '#/settings',
-	/** Exit destination for the settings view (fallback when no referrer). */
-	SETTINGS_EXIT: '#/',
 	/** Root — start of the app. */
 	START: '#/'
 } as const;
@@ -1,6 +1,4 @@
-import { ROUTES } from './routes.constants';
 import { Package, Search, Settings, SquarePen } from '@lucide/svelte';
-import McpLogo from '$lib/components/app/mcp/McpLogo.svelte';
 import { SidebarAction, ToolSource } from '$lib/enums';
 import type { DesktopIconStripItem } from '$lib/types';

@@ -64,15 +62,8 @@ export const SIDEBAR_ACTIONS_ITEMS: DesktopIconStripItem[] = [
 	},
 	{ icon: Search, keys: ['cmd', 'k'], tooltip: 'Search' },
 	{
-		activeRouteId: '/mcp-servers',
-		icon: McpLogo,
-		route: ROUTES.MCP_SERVERS,
-		tooltip: 'MCP Servers'
-	},
-	{
-		activeUrlIncludes: '#/settings',
+		action: SidebarAction.SETTINGS,
 		icon: Settings,
-		route: `${ROUTES.SETTINGS}/general`,
 		tooltip: 'Settings'
 	}
 ];
@@ -23,7 +23,8 @@ export enum ScrollCarouselVariant {
 * Sidebar icon strip actions handled directly by the sidebar.
 */
 export enum SidebarAction {
-	NEW_CHAT = 'new-chat'
+	NEW_CHAT = 'new-chat',
+	SETTINGS = 'settings'
 }

 /**
@@ -1,45 +0,0 @@
-import { beforeNavigate } from '$app/navigation';
-import { page } from '$app/state';
-import { ROUTES } from '$lib/constants';
-import { settingsReferrer } from '$lib/stores';
-
-export interface ChatSettings {
-	reset: () => void;
-}
-
-export function useSettingsNavigation() {
-	const subroute = $state({
-		activePanel: 'chat' as 'chat' | 'settings' | 'mcp',
-		chatSettingsRef: undefined as ChatSettings | undefined
-	});
-	const isSettingsRoute = $derived(!!page.route.id?.startsWith('/settings'));
-
-	beforeNavigate(({ from, to }) => {
-		if (to?.route?.id?.startsWith('/settings') && !from?.route?.id?.startsWith('/settings')) {
-			settingsReferrer.url = window.location.hash || ROUTES.START;
-		}
-	});
-
-	$effect(() => {
-		if (subroute.activePanel === 'settings' && subroute.chatSettingsRef) {
-			subroute.chatSettingsRef.reset();
-		}
-	});
-
-	// Return to chat when navigating to a new route
-	$effect(() => {
-		void page.url;
-
-		subroute.activePanel = 'chat';
-	});
-
-	return {
-		get isSettingsRoute() {
-			return isSettingsRoute;
-		},
-
-		get panel() {
-			return subroute;
-		}
-	};
-}
@@ -307,7 +307,6 @@ export { SandboxService } from './sandbox.service';
 *
 * **Key Responsibilities:**
 * - Build chat URLs for specific conversations: `RouterService.chat(id)` → `#/chat/:id`
- * - Build settings URLs for sections: `RouterService.settings(section)` → `#/settings/:section`
 *
 * @see ROUTES in constants/routes.ts — static route base paths
 */
@@ -1,8 +1,7 @@
 /**
 * RouterService - Builds app route paths
 *
- * Returns chat and settings route strings from a single source of truth
- * (ROUTES). No state.
+ * Returns chat route strings from a single source of truth (ROUTES). No state.
 */

 import { ROUTES } from '$lib/constants';
@@ -11,8 +10,4 @@ export class RouterService {
 	static chat(id: string): string {
 		return `${ROUTES.CHAT}/${id}`;
 	}
-
-	static settings(section: string): string {
-		return `${ROUTES.SETTINGS}/${section}`;
-	}
 }
@@ -49,8 +49,6 @@ export { uiStore } from './ui.svelte';
 // SETTINGS / UI PREFERENCES
 export { settingsStore } from './settings/index.svelte';

-export { settingsReferrer } from './settings/referrer.svelte';
-
 export { permissionsStore } from './permissions.svelte';

 // TOOLS
@@ -1,19 +0,0 @@
-/**
- * settingsReferrer - Remembers the settings route to return to after exit
- *
- * Tracks the last settings section the user was on so the app can return
- * there after a fallback exit. Standalone reactive value, no host.
- */
-
-import { ROUTES } from '$lib/constants';
-
-let _url = $state<string>(ROUTES.SETTINGS_EXIT);
-
-export const settingsReferrer = {
-	get url() {
-		return _url;
-	},
-	set url(value: string) {
-		_url = value;
-	}
-};
@@ -1,5 +0,0 @@
-<script lang="ts">
-	import { SettingsMcpServers } from '$lib/components/app/settings';
-</script>
-
-<SettingsMcpServers class="mx-auto w-full p-4 md:p-8 md:py-8" />
@@ -1,38 +0,0 @@
-<script lang="ts">
-	import { X } from '@lucide/svelte';
-	import { browser } from '$app/environment';
-	import { goto } from '$app/navigation';
-	import { page } from '$app/state';
-	import { ActionIcon } from '$lib/components/app';
-	import { ROUTES } from '$lib/constants';
-
-	let { children } = $props();
-
-	let previousRouteId = $state<string | null>(null);
-
-	$effect(() => {
-		const currentId = page.route.id;
-
-		return () => {
-			previousRouteId = currentId;
-		};
-	});
-
-	function handleClose() {
-		const prevIsSettings = previousRouteId?.startsWith('/settings');
-
-		if (browser && window.history.length > 1 && !prevIsSettings) {
-			history.back();
-		} else {
-			goto(ROUTES.SETTINGS_EXIT);
-		}
-	}
-</script>
-
-<div class="fixed top-4.5 right-4 z-50 md:hidden">
-	<ActionIcon icon={X} onclick={handleClose} tooltip="Close" />
-</div>
-
-<div class="min-h-full">
-	{@render children?.()}
-</div>
@@ -1,15 +0,0 @@
-<script lang="ts">
-	import { afterNavigate, replaceState } from '$app/navigation';
-	import { page } from '$app/state';
-	import { SettingsChat } from '$lib/components/app/settings';
-	import { SETTINGS_SECTION_SLUGS } from '$lib/constants';
-	import { RouterService } from '$lib/services';
-
-	afterNavigate(() => {
-		if (!page.params.section) {
-			replaceState(RouterService.settings(SETTINGS_SECTION_SLUGS.GENERAL), {});
-		}
-	});
-</script>
-
-<SettingsChat initialSection={(page.params as Record<string, string | undefined>).section} />
Author	SHA1	Message	Date
Xuan-Son NguyenandGitHub	925e117994	llama: add token ID tracking to KV cell (#27762 ) * kv: track token id * rm get_prev_tokens, move it to the main pr * nits * add get_prev_tokens	2026-08-26 23:34:28 +02:00
Aleksander GrygierandGitHub	539f24529b	ui: Move Settings and MCP Servers routes to dialog-based views (#27744 ) * ui : open MCP servers in a dialog from the chat form Replace the MCP servers submenu with a single "MCP Servers" item that opens a new DialogMcpServers dialog instead of navigating to the /mcp-servers route. Assisted-by: pi * ui : browse MCP resources from the server card Make the Resources capability badge clickable so it opens the MCP resources browser dialog, and drop the page-only chrome from SettingsMcpServers. Assisted-by: pi * ui : remove mcp-servers route and sidebar entry MCP servers are now managed in a dialog, so drop the dedicated route and the sidebar icon that navigated to it. Assisted-by: pi * ui : remove unused MCP servers submenu component The submenu was replaced by the MCP servers dialog, so delete the component and its export. Assisted-by: pi * feat(ui): add DialogSettingsChat dialog * refactor(ui): switch SettingsChat to in-app section navigation * feat(ui): open settings as dialog from sidebar * refactor(ui): remove settings route and URL-based settings navigation * fix(ui): adjust MCP dialogs for new base sizing * chore: Formatting & linting	2026-08-26 21:07:24 +02:00
Aleksander GrygierandGitHub	0379a19f09	ui: Update Dialog component styling (#27743 ) * feat(ui): make base dialog responsive and support sticky headers * ui: move dialog close button to the sticky header Assisted-by: pi * chore: Formatting & linting	2026-08-26 20:19:19 +02:00
Ruben OrtlamandGitHub	5e6a37cb11	vulkan: warptiles currently assume warp sizes <= 64, clamp to work around larger warps (#27726 )	2026-08-26 19:02:06 +03:00
Pranav UttarkarandGitHub	bf94216469	Implemented vulkan cross_entropy_loss and cross_entropy_loss_back (#27216 )	2026-08-26 16:49:32 +02:00
Radoslav GerganovandGitHub	d0132a680a	rpc : implement event and async backend APIs (#18626 ) * rpc : implement event and async backend APIs * cache responses from RPC_CMD_GET_ALLOC_SIZE	2026-08-26 17:34:46 +03:00
Aleksander GrygierandGitHub	4d19b28769	ci: Clean up UI builds from releases (#27706 ) * ci : inline UI version resolution into ui-build.yml * ci : build UI once and reuse the artifact in release jobs Server jobs now extract the ui-build artifact into tools/ui/dist instead of npm-building the UI. Also removes the get-version job and the no-op -DHF_UI_VERSION flags. Assisted-by: pi:Kimi-K3 * ui : disable the npm UI build by default (LLAMA_BUILD_UI=OFF) The flag now only controls building the UI from source via npm. The UI is still embedded by default from local tools/ui/dist or the prebuilt download (LLAMA_USE_PREBUILT_UI=ON). CI jobs no longer npm-build the UI; server-sanitize does not need node anymore. Assisted-by: pi:Kimi-K3 * ci : rename the ui-build artifact to llama-ui.zip Consistent with the other artifact names in the Actions summary. Assisted-by: pi:Kimi-K3 * ci : clarify the windows artifact merge in release.yml The windows-cuda/vulkan/sycl jobs build only the backend library; llama-server (with the embedded UI) is injected into their zips from the windows-cpu package during the release. State this in the job comments and use accurate wording in the merge step. Assisted-by: pi:Kimi-K3	2026-08-26 14:12:09 +02:00