2026-04-08 16:13:41 +00:00
70 changed files with 888 additions and 1831 deletions
--- a/.github/workflows/build-everything-tgw.yml
+++ b/.github/workflows/build-everything-tgw.yml
@ -41,13 +41,6 @@ jobs:
      version: ${{ inputs.version }}
      config: 'os:ubuntu-22.04'
  build_release_rocm_windows:
    name: ROCm Windows
    uses: ./.github/workflows/build-portable-release-rocm.yml
    with:
      version: ${{ inputs.version }}
      config: 'os:windows-2022'
  build_release_rocm_linux:
    name: ROCm Linux
    uses: ./.github/workflows/build-portable-release-rocm.yml
@ -75,31 +68,3 @@ jobs:
    with:
      version: ${{ inputs.version }}
      config: 'os:macos-15-intel,macos-14'
  build_release_ik_cuda_windows:
    name: ik CUDA Windows
    uses: ./.github/workflows/build-portable-release-ik-cuda.yml
    with:
      version: ${{ inputs.version }}
      config: 'os:windows-2022'
  build_release_ik_cuda_linux:
    name: ik CUDA Linux
    uses: ./.github/workflows/build-portable-release-ik-cuda.yml
    with:
      version: ${{ inputs.version }}
      config: 'os:ubuntu-22.04'
  build_release_ik_cpu_windows:
    name: ik CPU Windows
    uses: ./.github/workflows/build-portable-release-ik.yml
    with:
      version: ${{ inputs.version }}
      config: 'os:windows-2022'
  build_release_ik_cpu_linux:
    name: ik CPU Linux
    uses: ./.github/workflows/build-portable-release-ik.yml
    with:
      version: ${{ inputs.version }}
      config: 'os:ubuntu-22.04'
--- a/.github/workflows/build-portable-release-ik-cuda.yml
+++ b/.github/workflows/build-portable-release-ik-cuda.yml
@ -1,178 +0,0 @@
 name: Build ik CUDA
 on:
  workflow_dispatch:
    inputs:
      version:
        description: 'Version tag of text-generation-webui to build: v3.0'
        default: 'v3.0'
        required: true
        type: string
      config:
        description: 'Override configurations to build: key1:item1-1,item1-2;key2:item2-1,item2-2'
        default: 'Default'
        required: false
        type: string
      exclude:
        description: 'Exclude build configurations: key1-1:item1-1,key1-2:item1-2;key2-1:item2-1,key2-2:item2-2'
        default: 'None'
        required: false
        type: string
  workflow_call:
    inputs:
      version:
        description: 'Version tag of text-generation-webui to build: v3.0'
        default: 'v3.0'
        required: true
        type: string
      config:
        description: 'Configurations to build: key1:item1-1,item1-2;key2:item2-1,item2-2'
        default: 'Default'
        required: false
        type: string
      exclude:
        description: 'Exclude build configurations: key1-1:item1-1,key1-2:item1-2;key2-1:item2-1,key2-2:item2-2'
        default: 'None'
        required: false
        type: string
 permissions:
  contents: write
 jobs:
  define_matrix:
    name: Define Build Matrix
    runs-on: ubuntu-latest
    outputs:
      matrix: ${{ steps.set-matrix.outputs.matrix }}
    defaults:
      run:
        shell: pwsh
    env:
      CONFIGIN: ${{ inputs.config }}
      EXCLUDEIN: ${{ inputs.exclude }}
    steps:
      - name: Define Job Output
        id: set-matrix
        run: |
          $matrix = @{
              'os' = @('ubuntu-22.04', 'windows-2022')
              'pyver' = @("3.13")
              'cuda' = @("12.4", "13.1")
          }
          if ($env:CONFIGIN -ne 'Default') {$env:CONFIGIN.split(';').foreach({$matrix[$_.split(':')[0]] = $_.split(':')[1].split(',')})}
          if ($env:EXCLUDEIN -ne 'None') {
              $exclusions = @()
              $exclusions += $env:EXCLUDEIN.split(';').replace(':','=').replace(',',"`n") | ConvertFrom-StringData
              $matrix['exclude'] = $exclusions
          }
          $matrixOut = ConvertTo-Json $matrix -Compress
          Write-Output ('matrix=' + $matrixOut) >> $env:GITHUB_OUTPUT
  build_wheels:
    name: ${{ matrix.os }} ${{ matrix.pyver }} CUDA ${{ matrix.cuda }}
    needs: define_matrix
    runs-on: ${{ matrix.os }}
    strategy:
      matrix: ${{ fromJSON(needs.define_matrix.outputs.matrix) }}
    defaults:
      run:
        shell: pwsh
    env:
      PCKGVER: ${{ inputs.version }}
    steps:
      - uses: actions/checkout@v6
        with:
          repository: 'oobabooga/text-generation-webui'
          ref: ${{ inputs.version }}
          submodules: 'recursive'
      - uses: actions/setup-python@v6
        with:
          python-version: ${{ matrix.pyver }}
      - name: Build Package
        shell: bash
        run: |
            VERSION_CLEAN="${{ inputs.version }}"
            VERSION_CLEAN="${VERSION_CLEAN#v}"
            cd ..
            cp -r text-generation-webui "text-generation-webui-ik-${VERSION_CLEAN}"
            cd "text-generation-webui-ik-${VERSION_CLEAN}"
            # Remove extensions that need additional requirements
            allowed=("character_bias" "gallery" "sd_api_pictures")
            find extensions/ -mindepth 1 -maxdepth 1 -type d | grep -v -E "$(printf '%s|' "${allowed[@]}" | sed 's/|$//')" | xargs rm -rf
            # Define common variables
            CUDA_VERSION="${{ matrix.cuda }}"
            VERSION="${{ inputs.version }}"
            # 1. Set platform-specific variables
            if [[ "$RUNNER_OS" == "Windows" ]]; then
                PLATFORM="windows"
                PYTHON_URL="https://github.com/astral-sh/python-build-standalone/releases/download/20260303/cpython-3.13.12+20260303-x86_64-pc-windows-msvc-install_only_stripped.tar.gz"
                PIP_PATH="portable_env/python.exe -m pip"
                PACKAGES_PATH="portable_env/Lib/site-packages"
                rm start_linux.sh start_macos.sh
            else
                PLATFORM="linux"
                PYTHON_URL="https://github.com/astral-sh/python-build-standalone/releases/download/20260303/cpython-3.13.12+20260303-x86_64-unknown-linux-gnu-install_only_stripped.tar.gz"
                PIP_PATH="portable_env/bin/python -m pip"
                PACKAGES_PATH="portable_env/lib/python3.13/site-packages"
                rm start_macos.sh start_windows.bat
            fi
            # 2. Download and extract Python
            cd ..
            echo "Downloading Python for $PLATFORM..."
            curl -L -o python-build.tar.gz "$PYTHON_URL"
            tar -xzf python-build.tar.gz
            mv python "text-generation-webui-ik-${VERSION_CLEAN}/portable_env"
            # 3. Prepare requirements file based on CUDA version
            cd "text-generation-webui-ik-${VERSION_CLEAN}"
            if [[ "$CUDA_VERSION" == "13.1" ]]; then
                REQ_FILE="requirements/portable/requirements_ik_cuda131.txt"
            else
                REQ_FILE="requirements/portable/requirements_ik.txt"
            fi
            # 4. Inject --ik into start scripts
            sed -i 's/--portable/--portable --ik/g' start_linux.sh start_windows.bat 2>/dev/null || true
            # 5. Install packages
            echo "Installing Python packages from $REQ_FILE..."
            $PIP_PATH install --target="./$PACKAGES_PATH" -r "$REQ_FILE"
            # 6. Clean up
            rm -rf .git cmd* update_wizard* Colab-TextGen-GPU.ipynb docker setup.cfg .github .gitignore requirements/ one_click.py
            # 7. Create archive
            cd ..
            if [[ "$RUNNER_OS" == "Windows" ]]; then
                ARCHIVE_NAME="textgen-portable-ik-${VERSION_CLEAN}-${PLATFORM}-cuda${CUDA_VERSION}.zip"
                echo "Creating archive: $ARCHIVE_NAME"
                powershell -Command "Compress-Archive -Path text-generation-webui-ik-${VERSION_CLEAN} -DestinationPath $ARCHIVE_NAME"
            else
                ARCHIVE_NAME="textgen-portable-ik-${VERSION_CLEAN}-${PLATFORM}-cuda${CUDA_VERSION}.tar.gz"
                echo "Creating archive: $ARCHIVE_NAME"
                tar czf "$ARCHIVE_NAME" "text-generation-webui-ik-${VERSION_CLEAN}"
            fi
      - name: Upload files to a GitHub release
        id: upload-release
        uses: svenstaro/upload-release-action@2.7.0
        continue-on-error: true
        with:
          repo_token: ${{ secrets.GITHUB_TOKEN }}
          file: ../textgen-portable-ik-*
          tag: ${{ inputs.version }}
          file_glob: true
          make_latest: false
          overwrite: true
--- a/.github/workflows/build-portable-release-ik.yml
+++ b/.github/workflows/build-portable-release-ik.yml
@ -1,173 +0,0 @@
 name: Build ik CPU
 on:
  workflow_dispatch:
    inputs:
      version:
        description: 'Version tag of text-generation-webui to build: v3.0'
        default: 'v3.0'
        required: true
        type: string
      config:
        description: 'Override configurations to build: key1:item1-1,item1-2;key2:item2-1,item2-2'
        default: 'Default'
        required: false
        type: string
      exclude:
        description: 'Exclude build configurations: key1-1:item1-1,key1-2:item1-2;key2-1:item2-1,key2-2:item2-2'
        default: 'None'
        required: false
        type: string
  workflow_call:
    inputs:
      version:
        description: 'Version tag of text-generation-webui to build: v3.0'
        default: 'v3.0'
        required: true
        type: string
      config:
        description: 'Configurations to build: key1:item1-1,item1-2;key2:item2-1,item2-2'
        default: 'Default'
        required: false
        type: string
      exclude:
        description: 'Exclude build configurations: key1-1:item1-1,key1-2:item1-2;key2-1:item2-1,key2-2:item2-2'
        default: 'None'
        required: false
        type: string
 permissions:
  contents: write
 jobs:
  define_matrix:
    name: Define Build Matrix
    runs-on: ubuntu-latest
    outputs:
      matrix: ${{ steps.set-matrix.outputs.matrix }}
    defaults:
      run:
        shell: pwsh
    env:
      CONFIGIN: ${{ inputs.config }}
      EXCLUDEIN: ${{ inputs.exclude }}
    steps:
      - name: Define Job Output
        id: set-matrix
        run: |
          $matrix = @{
              'os' = @('ubuntu-22.04', 'windows-2022')
              'pyver' = @("3.13")
          }
          if ($env:CONFIGIN -ne 'Default') {$env:CONFIGIN.split(';').foreach({$matrix[$_.split(':')[0]] = $_.split(':')[1].split(',')})}
          if ($env:EXCLUDEIN -ne 'None') {
              $exclusions = @()
              $exclusions += $env:EXCLUDEIN.split(';').replace(':','=').replace(',',"`n") | ConvertFrom-StringData
              $matrix['exclude'] = $exclusions
          }
          $matrixOut = ConvertTo-Json $matrix -Compress
          Write-Output ('matrix=' + $matrixOut) >> $env:GITHUB_OUTPUT
  build_wheels:
    name: ${{ matrix.os }} ${{ matrix.pyver }}
    needs: define_matrix
    runs-on: ${{ matrix.os }}
    strategy:
      matrix: ${{ fromJSON(needs.define_matrix.outputs.matrix) }}
    defaults:
      run:
        shell: pwsh
    env:
      PCKGVER: ${{ inputs.version }}
    steps:
      - uses: actions/checkout@v6
        with:
          repository: 'oobabooga/text-generation-webui'
          ref: ${{ inputs.version }}
          submodules: 'recursive'
      - uses: actions/setup-python@v6
        with:
          python-version: ${{ matrix.pyver }}
      - name: Build Package
        shell: bash
        run: |
            VERSION_CLEAN="${{ inputs.version }}"
            VERSION_CLEAN="${VERSION_CLEAN#v}"
            cd ..
            cp -r text-generation-webui "text-generation-webui-ik-${VERSION_CLEAN}"
            cd "text-generation-webui-ik-${VERSION_CLEAN}"
            # Remove extensions that need additional requirements
            allowed=("character_bias" "gallery" "sd_api_pictures")
            find extensions/ -mindepth 1 -maxdepth 1 -type d | grep -v -E "$(printf '%s|' "${allowed[@]}" | sed 's/|$//')" | xargs rm -rf
            # Define common variables
            VERSION="${{ inputs.version }}"
            # 1. Set platform-specific variables
            if [[ "$RUNNER_OS" == "Windows" ]]; then
                PLATFORM="windows-cpu"
                PYTHON_URL="https://github.com/astral-sh/python-build-standalone/releases/download/20260303/cpython-3.13.12+20260303-x86_64-pc-windows-msvc-install_only_stripped.tar.gz"
                PIP_PATH="portable_env/python.exe -m pip"
                PACKAGES_PATH="portable_env/Lib/site-packages"
                rm start_linux.sh start_macos.sh
            else
                PLATFORM="linux-cpu"
                PYTHON_URL="https://github.com/astral-sh/python-build-standalone/releases/download/20260303/cpython-3.13.12+20260303-x86_64-unknown-linux-gnu-install_only_stripped.tar.gz"
                PIP_PATH="portable_env/bin/python -m pip"
                PACKAGES_PATH="portable_env/lib/python3.13/site-packages"
                rm start_macos.sh start_windows.bat
            fi
            # 2. Download and extract Python
            echo "Downloading Python for $PLATFORM..."
            cd ..
            curl -L -o python-build.tar.gz "$PYTHON_URL"
            tar -xzf python-build.tar.gz
            mv python "text-generation-webui-ik-${VERSION_CLEAN}/portable_env"
            # 3. Prepare requirements file
            cd "text-generation-webui-ik-${VERSION_CLEAN}"
            REQ_FILE="requirements/portable/requirements_ik_cpu_only.txt"
            echo "Using requirements file: $REQ_FILE"
            # 4. Inject --ik into start scripts
            sed -i 's/--portable/--portable --ik/g' start_linux.sh start_windows.bat 2>/dev/null || true
            # 5. Install packages
            echo "Installing Python packages from $REQ_FILE..."
            $PIP_PATH install --target="./$PACKAGES_PATH" -r "$REQ_FILE"
            # 6. Clean up
            rm -rf .git cmd* update_wizard* Colab-TextGen-GPU.ipynb docker setup.cfg .github .gitignore requirements/ one_click.py
            # 7. Create archive
            cd ..
            if [[ "$RUNNER_OS" == "Windows" ]]; then
                ARCHIVE_NAME="textgen-portable-ik-${VERSION_CLEAN}-${PLATFORM}.zip"
                echo "Creating archive: $ARCHIVE_NAME"
                powershell -Command "Compress-Archive -Path text-generation-webui-ik-${VERSION_CLEAN} -DestinationPath $ARCHIVE_NAME"
            else
                ARCHIVE_NAME="textgen-portable-ik-${VERSION_CLEAN}-${PLATFORM}.tar.gz"
                echo "Creating archive: $ARCHIVE_NAME"
                tar czf "$ARCHIVE_NAME" "text-generation-webui-ik-${VERSION_CLEAN}"
            fi
      - name: Upload files to a GitHub release
        id: upload-release
        uses: svenstaro/upload-release-action@2.7.0
        continue-on-error: true
        with:
          repo_token: ${{ secrets.GITHUB_TOKEN }}
          file: ../textgen-portable-ik-*
          tag: ${{ inputs.version }}
          file_glob: true
          make_latest: false
          overwrite: true
--- a/README.md
+++ b/README.md
@ -24,9 +24,9 @@ A Gradio web UI for running Large Language Models locally. 100% private and offl
 ## Features
 - **Easy setup**: [Portable builds](https://github.com/oobabooga/text-generation-webui/releases) (zero setup, just unzip and run) for GGUF models on Windows/Linux/macOS, or a one-click installer for the full feature set.
- **Multiple backends**: [llama.cpp](https://github.com/ggerganov/llama.cpp), [ik_llama.cpp](https://github.com/ikawrakow/ik_llama.cpp), [Transformers](https://github.com/huggingface/transformers), [ExLlamaV3](https://github.com/turboderp-org/exllamav3), and [TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM). Switch between backends and models without restarting.
+- **Multiple backends**: [llama.cpp](https://github.com/ggerganov/llama.cpp), [Transformers](https://github.com/huggingface/transformers), [ExLlamaV3](https://github.com/turboderp-org/exllamav3), and [TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM). Switch between backends and models without restarting.
 - **OpenAI/Anthropic-compatible API**: Chat, Completions, and Messages endpoints with tool-calling support. Use as a local drop-in replacement for the OpenAI/Anthropic APIs ([examples](https://github.com/oobabooga/text-generation-webui/wiki/12-%E2%80%90-OpenAI-API#examples)).
- **Tool-calling**: Models can call custom functions during chat — web search, page fetching, math, and more. Each tool is a single `.py` file. MCP servers are also supported ([tutorial](https://github.com/oobabooga/text-generation-webui/wiki/Tool-Calling-Tutorial)).
+- **Tool-calling**: Models can call custom functions during chat — web search, page fetching, math, and more. Each tool is a single `.py` file, easy to create and extend ([tutorial](https://github.com/oobabooga/text-generation-webui/wiki/Tool-Calling-Tutorial)).
 - **Vision (multimodal)**: Attach images to messages for visual understanding ([tutorial](https://github.com/oobabooga/text-generation-webui/wiki/Multimodal-Tutorial)).
 - **File attachments**: Upload text files, PDF documents, and .docx documents to talk about their contents.
 - **Training**: Fine-tune LoRAs on multi-turn chat or raw text datasets. Supports resuming interrupted runs ([tutorial](https://github.com/oobabooga/text-generation-webui/wiki/05-%E2%80%90-Training-Tab)).
--- a/css/html_instruct_style.css
+++ b/css/html_instruct_style.css
@ -13,12 +13,21 @@
    line-height: 28px !important;
 }
-.dark .chat .message-body :is(p,li),
+.dark .chat .message-body :is(p,li,h1,h2,h3,h4,h5,h6),
 .dark .chat .message-body em:not(:is(h1,h2,h3,h4,h5,h6,b,strong) em),
 .dark .chat .message-body q:not(:is(h1,h2,h3,h4,h5,h6,b,strong) q) {
    color: #d1d5db !important;
 }
 .chat .message-body :is(th, td),
 .prose hr {
    border-color: #40404096 !important;
 }
 .dark .chat .message-body :is(th, td),
 .dark .prose hr {
    border-color: rgb(255 255 255 / 30%) !important;
 }
 .chat .message-body :is(p, ul, ol) {
    margin: 1.25em 0 !important;
--- a/css/main.css
+++ b/css/main.css
@ -22,17 +22,6 @@
    font-style: italic;
 }
 /* Hide spin buttons on number inputs (look bad on Windows) */
 input[type="number"]::-webkit-outer-spin-button,
 input[type="number"]::-webkit-inner-spin-button {
    -webkit-appearance: none;
    margin: 0;
 }
 input[type="number"] {
    -moz-appearance: textfield;
 }
 .padded.svelte-12cmxck {
    padding: 3px 0;
 }
@ -257,8 +246,8 @@ button {
 .pretty_scrollbar::-webkit-scrollbar,
 #image-history-gallery > :nth-child(2)::-webkit-scrollbar {
-    width: 7px;
+    width: 8px;
-    height: 7px;
+    height: 8px;
 }
 .pretty_scrollbar::-webkit-scrollbar-track,
@ -271,7 +260,7 @@ button {
 #image-history-gallery > :nth-child(2)::-webkit-scrollbar-thumb,
 #image-history-gallery > :nth-child(2)::-webkit-scrollbar-thumb:hover {
    background: var(--neutral-300);
-    border-radius: 9999px;
+    border-radius: 30px;
 }
 .dark .pretty_scrollbar::-webkit-scrollbar-thumb,
@ -279,17 +268,18 @@ button {
 .dark #image-history-gallery > :nth-child(2)::-webkit-scrollbar-thumb,
 .dark #image-history-gallery > :nth-child(2)::-webkit-scrollbar-thumb:hover {
    background: rgb(255 255 255 / 6.25%);
-    border-radius: 9999px;
+    border-radius: 30px;
 }
 .pretty_scrollbar::-webkit-resizer,
 #image-history-gallery > :nth-child(2)::-webkit-resizer {
-    background: transparent;
+    background: #d2d2d8;
 }
 .dark .pretty_scrollbar::-webkit-resizer,
 .dark #image-history-gallery > :nth-child(2)::-webkit-resizer {
-    background: transparent;
+    background: rgb(255 255 255 / 10%);
    border-radius: 10px;
 }
 .pretty_scrollbar::-webkit-scrollbar-corner,
@ -446,25 +436,15 @@ audio {
 .dark .message-body h4,
 .dark .message-body h5,
 .dark .message-body h6 {
-    color: #e8e8e8 !important;
+    color: white !important;
 }
-.message-body blockquote {
+.dark .message-body blockquote {
-    border-left-width: 4px;
+    border-left-color: rgb(255 255 255 / 30%);
    border-left-color: var(--border-color-primary);
 }
 .message-body h1,
 .message-body h2,
 .message-body h3,
 .message-body h4,
 .message-body h5,
 .message-body h6 {
    color: #1a1a1a;
 }
 .message-body h1 {
-    font-weight: 700;
+    font-weight: 800;
    font-size: 2.25em;
    margin-top: 0;
    margin-bottom: 0.8888889em;
@ -496,13 +476,13 @@ audio {
 }
 .message-body h5 {
-    font-weight: 600;
+    font-weight: normal;
    font-size: 1em;
    margin: 0;
 }
 .message-body h6 {
-    font-weight: 600;
+    font-weight: normal;
    font-size: 1em;
    margin: 0;
 }
@ -610,7 +590,7 @@ audio {
 }
 #chat-input textarea::-webkit-scrollbar {
-    width: 7px;
+    width: 8px;
 }
 #chat-input textarea::-webkit-scrollbar-track {
@ -619,7 +599,7 @@ audio {
 #chat-input textarea::-webkit-scrollbar-thumb {
    background: var(--neutral-300);
-    border-radius: 9999px;
+    border-radius: 30px;
 }
 .dark #chat-input textarea::-webkit-scrollbar-thumb {
@ -653,10 +633,6 @@ audio {
    background: transparent;
 }
 #chat-input .thumbnails {
    padding-top: 3px;
 }
 .chat-input-positioned {
    max-width: 54rem;
    left: 50%;
@ -759,30 +735,7 @@ audio {
 .hover-element {
    position: relative;
-    padding-top: 4px;
+    font-size: 24px;
 }
 #hover-element-button {
    display: flex;
    align-items: center;
    justify-content: center;
    width: 32px;
    height: 32px;
    border-radius: 0.5rem;
    cursor: pointer;
    color: gray;
 }
 #hover-element-button:hover {
    background-color: var(--background-fill-secondary);
 }
 #hover-element-button svg {
    color: inherit;
 }
 .dark #hover-element-button:hover {
    background-color: var(--selected-item-color-dark);
 }
 .hover-menu {
@ -790,40 +743,27 @@ audio {
    position: absolute;
    bottom: 100%;
    left: 0;
-    background: white;
+    box-shadow: 0 2px 12px rgb(0 0 0 / 15%);
-    border: 1px solid rgba(0, 0, 0, 0.1);
+    border-radius: 0.5rem;
    box-shadow: 0 4px 16px rgb(0 0 0 / 12%), 0 1px 3px rgb(0 0 0 / 8%);
    border-radius: 0.75rem;
    z-index: 10000;
    min-width: 330px;
    flex-direction: column;
-    padding: 4px;
+    overflow: hidden;
 }
 .hover-menu::before {
    content: '';
    position: absolute;
    top: 100%;
    left: 0;
    width: 100%;
    height: 8px;
 }
 .hover-menu > * {
    border: none !important;
    box-shadow: none !important;
 }
 .hover-menu button {
    width: 100%;
-    background: transparent !important;
+    background: white !important;
-    border: none !important;
+    border-radius: 0 !important;
    border-radius: 0.5rem !important;
    justify-content: space-between;
    margin: 0 !important;
    height: 36px;
-    font-weight: 500;
+    border-color: transparent !important;
-    box-shadow: none !important;
+    transition: background-color 0.15s ease;
 }
 .hover-menu button:not(#clear-history-confirm) {
    border-bottom: 0 !important;
 }
 .hover-menu button:hover {
@ -835,26 +775,19 @@ audio {
 }
 #show-controls {
-    background-color: transparent;
+    background-color: white;
-    border: none !important;
+    border-color: transparent !important;
    height: 36px;
-    border-radius: 0.5rem;
+    border-radius: 0;
    border-bottom: 0 !important;
    padding-top: 3px;
    padding-left: 4px;
    display: flex;
    font-weight: normal;
 }
 #show-controls:hover {
    background-color: #dbeafe;
 }
 .dark #show-controls {
-    background-color: transparent;
+    background-color: var(--darker-gray);
 }
 .dark #show-controls:hover {
    background-color: var(--selected-item-color-dark);
 }
 #show-controls label {
@ -864,12 +797,12 @@ audio {
    width: 100%;
    padding-right: 12px;
    gap: 10px;
-    font-weight: 500;
+    font-weight: 600;
    color: var(--button-secondary-text-color);
 }
 #show-controls label input {
-    margin-top: 5px;
+    margin-top: 4px;
 }
 .transparent-substring {
@ -909,7 +842,7 @@ audio {
 }
 #chat-input-row {
-    padding: 0.5rem 1rem 1rem;
+    padding: 1rem;
 }
 #chat-col {
@ -1275,14 +1208,9 @@ audio {
    color: #9ca3af;
 }
 .dark .hover-menu {
    background: var(--darker-gray);
    border-color: transparent;
    box-shadow: 0 4px 16px rgb(0 0 0 / 40%);
 }
 .dark .hover-menu button {
-    background-color: transparent !important;
+    border-color: var(--border-color-primary);
    background-color: var(--darker-gray) !important;
 }
 .dark #chat-controls,
@ -1444,7 +1372,8 @@ audio {
 }
 .footer-button svg {
-    stroke: rgb(140 140 148);
+    stroke: rgb(156 163 175);
    transition: stroke 0.2s;
 }
 .footer-button:hover svg {
@ -1459,12 +1388,12 @@ audio {
    stroke: rgb(209 213 219);
 }
-.block:has(> .label-wrap) {
+.tgw-accordion {
    padding: 10px 12px !important;
    border: 1px solid #d2d2d8;
 }
-.dark .block:has(> .label-wrap) {
+.dark .tgw-accordion {
    border: 1px solid var(--border-color-dark);
 }
@ -1974,24 +1903,14 @@ table, tr, td, th, thead {
    border: 0;
 }
 .prose hr {
    border-color: var(--border-color-primary);
 }
 td + td,
-th + th {
+th + th { border-left: 1px solid; }
    border-left: 1px solid var(--border-color-primary) !important;
 }
 tr + tr td,
-tr + tr th {
+tr + tr th { border-top: 1px solid; }
    border-top: 1px solid var(--border-color-primary) !important;
 }
 thead + tbody tr:first-child td,
-thead + tbody tr:first-child th {
+thead + tbody tr:first-child th { border-top: 1px solid; }
    border-top: 1px solid var(--border-color-primary) !important;
 }
 /* ------------------------------------------------
   Tools CheckboxGroup - vertical DragDrop-like style
@ -2023,8 +1942,8 @@ thead + tbody tr:first-child th {
 /* Pretty scrollbar for the tools list */
 #tools-group .wrap::-webkit-scrollbar {
-    width: 7px;
+    width: 8px;
-    height: 7px;
+    height: 8px;
 }
 #tools-group .wrap::-webkit-scrollbar-track {
@ -2034,13 +1953,13 @@ thead + tbody tr:first-child th {
 #tools-group .wrap::-webkit-scrollbar-thumb,
 #tools-group .wrap::-webkit-scrollbar-thumb:hover {
    background: var(--neutral-300);
-    border-radius: 9999px;
+    border-radius: 30px;
 }
 .dark #tools-group .wrap::-webkit-scrollbar-thumb,
 .dark #tools-group .wrap::-webkit-scrollbar-thumb:hover {
    background: rgb(255 255 255 / 6.25%);
-    border-radius: 9999px;
+    border-radius: 30px;
 }
 #tools-group .wrap::-webkit-scrollbar-corner {
--- a/docs/01
+++ b/docs/01
@ -112,7 +112,7 @@ Used for talking to an instruction-following model using the prompt format defin
 The prompt format is defined by the **Instruction template** parameter in "Parameters" > "Instruction template", which represents a Jinja2 template.
-Note that when you load a model in the "Model" tab, the web UI will try to automatically detect its instruction template (if any) from the model metadata (e.g. `tokenizer_config.json` or GGUF metadata), and will update the values under "Parameters" > "Instruction template" accordingly. You should check the model card on Hugging Face to see if you are using the correct prompt format.
+Note that when you load a model in the "Model" tab, the web UI will try to automatically detect its instruction template (if any), and will update the values under "Parameters" > "Instruction template" accordingly. This is done using a set of regular expressions defined in `user_data/models/config.yaml`. This detection is not guaranteed to be accurate. You should check the model card on Hugging Face to see if you are using the correct prompt format.
 ### Chat-instruct
--- a/docs/12
+++ b/docs/12
@ -39,7 +39,7 @@ curl http://127.0.0.1:5000/v1/completions \
 #### Chat completions
-Works best with instruction-following models. If the "instruction_template" variable is not provided, it will be detected automatically from the model metadata.
+Works best with instruction-following models. If the "instruction_template" variable is not provided, it will be guessed automatically based on the model name using the regex patterns in `user_data/models/config.yaml`.
 ```shell
 curl http://127.0.0.1:5000/v1/chat/completions \
@ -232,17 +232,6 @@ curl -k http://127.0.0.1:5000/v1/internal/model/load \
  }'
 ```
 You can also set a default instruction template for all subsequent API requests by passing `instruction_template` (a template name from `user_data/instruction-templates/`) or `instruction_template_str` (a raw Jinja2 string):
 ```shell
 curl -k http://127.0.0.1:5000/v1/internal/model/load \
  -H "Content-Type: application/json" \
  -d '{
    "model_name": "Qwen_Qwen3-0.6B-Q4_K_M.gguf",
    "instruction_template": "Alpaca"
  }'
 ```
 #### Python chat example
 ```python
--- a/Tutorial.md
+++ b/Tutorial.md
@ -80,19 +80,6 @@ def execute(arguments):
 You can open the built-in tools in `user_data/tools/` for more examples.
 ## MCP servers
 You can connect to remote [MCP (Model Context Protocol)](https://modelcontextprotocol.io/) servers to use their tools alongside local ones.
 In the chat sidebar, open the **MCP servers** accordion and enter one server URL per line. For servers that require authentication, append headers after the URL separated by commas:
 ```
 https://example.com/mcp
 https://other.com/mcp,Authorization: Bearer sk-xxx
 ```
 All tools from the configured servers are automatically discovered and made available to the model during generation. If an MCP tool has the same name as a selected local tool, the local tool takes priority.
 ## Tool calling over the API
 Tool calling over the API follows the [OpenAI API](https://platform.openai.com/docs/guides/function-calling) convention. Define your tools, send them with your messages, and handle tool calls in a loop until the model gives a final answer.
--- a/download-model.py
+++ b/download-model.py
@ -158,21 +158,28 @@ class ModelDownloader:
        # Also if GGUF and safetensors are available, download only safetensors
        if (has_pytorch or has_pt or has_gguf) and has_safetensors:
            has_gguf = False
-            keep = [i for i, c in enumerate(classifications) if c not in ['pytorch', 'pt', 'gguf']]
+            for i in range(len(classifications) - 1, -1, -1):
-            links = [links[i] for i in keep]
+                if classifications[i] in ['pytorch', 'pt', 'gguf']:
-            file_sizes = [file_sizes[i] for i in keep]
+                    links.pop(i)
                    file_sizes.pop(i)
        # For GGUF, try to download only the Q4_K_M if no specific file is specified.
        if has_gguf and specific_file is None:
-            has_q4km = any('q4_k_m' in link.lower() for link in links)
+            has_q4km = False
            for i in range(len(classifications) - 1, -1, -1):
                if 'q4_k_m' in links[i].lower():
                    has_q4km = True
            if has_q4km:
-                keep = [i for i, link in enumerate(links) if 'q4_k_m' in link.lower()]
+                for i in range(len(classifications) - 1, -1, -1):
                    if 'q4_k_m' not in links[i].lower():
                        links.pop(i)
                        file_sizes.pop(i)
            else:
-                keep = [i for i, link in enumerate(links) if not link.lower().endswith('.gguf')]
+                for i in range(len(classifications) - 1, -1, -1):
-
+                    if links[i].lower().endswith('.gguf'):
-            links = [links[i] for i in keep]
+                        links.pop(i)
-            file_sizes = [file_sizes[i] for i in keep]
+                        file_sizes.pop(i)
        is_llamacpp = has_gguf and specific_file is not None
        return links, sha256, is_lora, is_llamacpp, file_sizes
--- a/extensions/superbooga/download_urls.py
+++ b/extensions/superbooga/download_urls.py
@ -2,11 +2,8 @@ import concurrent.futures
 import requests
 from modules.web_search import _validate_url
 def download_single(url):
    _validate_url(url)
    headers = {
        'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/58.0.3029.110 Safari/537.3'
    }
--- a/extensions/superboogav2/download_urls.py
+++ b/extensions/superboogav2/download_urls.py
@ -5,14 +5,12 @@ import requests
 from bs4 import BeautifulSoup
 import extensions.superboogav2.parameters as parameters
 from modules.web_search import _validate_url
 from .data_processor import process_and_add_to_collector
 from .utils import create_metadata_source
 def _download_single(url):
    _validate_url(url)
    response = requests.get(url, timeout=5)
    if response.status_code == 200:
        return response.content
--- a/js/dark_theme.js
+++ b/js/dark_theme.js
@ -1,6 +1,6 @@
 function toggleDarkMode() {
  document.body.classList.toggle("dark");
-  const currentCSS = document.getElementById("highlight-css");
+  var currentCSS = document.getElementById("highlight-css");
  if (currentCSS.getAttribute("href") === "file/css/highlightjs/github-dark.min.css") {
    currentCSS.setAttribute("href", "file/css/highlightjs/github.min.css");
  } else {
@ -9,10 +9,12 @@ function toggleDarkMode() {
  // Re-highlight all code blocks once stylesheet loads
  currentCSS.onload = function() {
-    // Clear data-highlighted so hljs will re-process with the new theme
+    const messageBodies = document.getElementById("chat").querySelectorAll(".message-body");
-    document.querySelectorAll("#chat .message-body pre code[data-highlighted]").forEach((codeBlock) => {
+    messageBodies.forEach((messageBody) => {
-      delete codeBlock.dataset.highlighted;
+      const codeBlocks = messageBody.querySelectorAll("pre code");
      codeBlocks.forEach((codeBlock) => {
        hljs.highlightElement(codeBlock);
      });
    });
    doSyntaxHighlighting();
  };
 }
--- a/js/global_scope_js.js
+++ b/js/global_scope_js.js
@ -1,35 +1,11 @@
 // -------------------------------------------------
 // Shared helpers
 // -------------------------------------------------
 function getProfilePictureUrl() {
  return "/file/user_data/cache/pfp_character.png?time=" + Date.now();
 }
 const MESSAGE_SELECTOR = ".message, .user-message, .assistant-message";
 function getMessageElement(element) {
  if (!element) return null;
  return element.closest(MESSAGE_SELECTOR);
 }
 function isUserRole(messageElement) {
  return messageElement.classList.contains("user-message") ||
         messageElement.querySelector(".text-you") !== null ||
         messageElement.querySelector(".circle-you") !== null;
 }
 // Trigger a synthetic 'input' event so Gradio picks up programmatic value changes
 function dispatchGradioInput(element) {
  element.dispatchEvent(new Event("input", { bubbles: true }));
 }
 // -------------------------------------------------
 // Event handlers
 // -------------------------------------------------
 function copyToClipboard(element) {
-  const messageElement = getMessageElement(element);
+  if (!element) return;
  const messageElement = element.closest(".message, .user-message, .assistant-message");
  if (!messageElement) return;
  const rawText = messageElement.getAttribute("data-raw");
@ -72,7 +48,9 @@ function fallbackCopyToClipboard(text) {
 }
 function branchHere(element) {
-  const messageElement = getMessageElement(element);
+  if (!element) return;
  const messageElement = element.closest(".message, .user-message, .assistant-message");
  if (!messageElement) return;
  const index = messageElement.getAttribute("data-index");
@ -91,7 +69,11 @@ function branchHere(element) {
  }
  branchIndexInput.value = index;
-  dispatchGradioInput(branchIndexInput);
+
  // Trigger any 'change' or 'input' events Gradio might be listening for
  const event = new Event("input", { bubbles: true });
  branchIndexInput.dispatchEvent(event);
  branchButton.click();
 }
@ -100,7 +82,9 @@ function branchHere(element) {
 // -------------------------------------------------
 function editHere(buttonElement) {
-  const messageElement = getMessageElement(buttonElement);
+  if (!buttonElement) return;
  const messageElement = buttonElement.closest(".message, .user-message, .assistant-message");
  if (!messageElement) return;
  const messageBody = messageElement.querySelector(".message-body");
@ -113,7 +97,12 @@ function editHere(buttonElement) {
    return;
  }
-  startEditing(messageElement, messageBody, isUserRole(messageElement));
+  // Determine role based on message element - handle different chat modes
  const isUserMessage = messageElement.classList.contains("user-message") ||
                       messageElement.querySelector(".text-you") !== null ||
                       messageElement.querySelector(".circle-you") !== null;
  startEditing(messageElement, messageBody, isUserMessage);
 }
 function startEditing(messageElement, messageBody, isUserMessage) {
@ -220,22 +209,30 @@ function submitMessageEdit(index, newText, isUserMessage) {
  editTextInput.value = newText;
  editRoleInput.value = isUserMessage ? "user" : "assistant";
-  dispatchGradioInput(editIndexInput);
+  editIndexInput.dispatchEvent(new Event("input", { bubbles: true }));
-  dispatchGradioInput(editTextInput);
+  editTextInput.dispatchEvent(new Event("input", { bubbles: true }));
-  dispatchGradioInput(editRoleInput);
+  editRoleInput.dispatchEvent(new Event("input", { bubbles: true }));
  editButton.click();
  return true;
 }
 function navigateVersion(element, direction) {
-  const messageElement = getMessageElement(element);
+  if (!element) return;
  const messageElement = element.closest(".message, .user-message, .assistant-message");
  if (!messageElement) return;
  const index = messageElement.getAttribute("data-index");
  if (!index) return;
-  const role = isUserRole(messageElement) ? "user" : "assistant";
+  // Determine role based on message element classes
  let role = "assistant"; // Default role
  if (messageElement.classList.contains("user-message") ||
      messageElement.querySelector(".text-you") ||
      messageElement.querySelector(".circle-you")) {
    role = "user";
  }
  const indexInput = document.getElementById("Navigate-message-index")?.querySelector("input");
  const directionInput = document.getElementById("Navigate-direction")?.querySelector("textarea");
@ -251,9 +248,11 @@ function navigateVersion(element, direction) {
  directionInput.value = direction;
  roleInput.value = role;
-  dispatchGradioInput(indexInput);
+  // Trigger 'input' events for Gradio to pick up changes
-  dispatchGradioInput(directionInput);
+  const event = new Event("input", { bubbles: true });
-  dispatchGradioInput(roleInput);
+  indexInput.dispatchEvent(event);
  directionInput.dispatchEvent(event);
  roleInput.dispatchEvent(event);
  navigateButton.click();
 }
@ -314,7 +313,7 @@ function handleMorphdomUpdate(data) {
 function applyMorphdomUpdate(data) {
  // Determine target element and use it as query scope
-  let target_element, target_html;
+  var target_element, target_html;
  if (data.last_message_only) {
    const childNodes = document.getElementsByClassName("messages")[0].childNodes;
    target_element = childNodes[childNodes.length - 1];
--- a/js/main.js
+++ b/js/main.js
@ -4,9 +4,8 @@
 // Sync highlight.js theme with the actual Gradio theme
 var defined_hljs_css = document.body.classList.contains("dark") ? "file/css/highlightjs/github-dark.min.css" : "file/css/highlightjs/github.min.css";
-var hljsCssElement = document.getElementById("highlight-css");
+if (document.getElementById("highlight-css").getAttribute("href") !== defined_hljs_css) {
-if (hljsCssElement.getAttribute("href") !== defined_hljs_css) {
+  document.getElementById("highlight-css").setAttribute("href", defined_hljs_css);
  hljsCssElement.setAttribute("href", defined_hljs_css);
 }
 let main_parent = document.getElementById("chat-tab").parentNode;
@ -50,18 +49,21 @@ document.querySelector(".header_bar").addEventListener("click", function(event)
 //------------------------------------------------
 // --- Helper functions --- //
-function isModifiedKeyboardEvent(event) {
+function isModifiedKeyboardEvent() {
-  return event instanceof KeyboardEvent &&
+  return (event instanceof KeyboardEvent &&
-    (event.shiftKey || event.ctrlKey || event.altKey || event.metaKey);
+    event.shiftKey ||
    event.ctrlKey ||
    event.altKey ||
    event.metaKey);
 }
-function isFocusedOnEditableTextbox(event) {
+function isFocusedOnEditableTextbox() {
  if (event.target.tagName === "INPUT" || event.target.tagName === "TEXTAREA") {
    return !!event.target.value;
  }
  return false;
 }
 let previousTabId = "chat-tab-button";
 document.addEventListener("keydown", function(event) {
  // Stop generation on Esc pressed
  if (event.key === "Escape") {
@ -115,14 +117,14 @@ document.addEventListener("keydown", function(event) {
  }
  // --- Simple version navigation --- //
-  if (!isFocusedOnEditableTextbox(event)) {
+  if (!isFocusedOnEditableTextbox()) {
    // Version navigation on Arrow keys (horizontal)
-    if (!isModifiedKeyboardEvent(event) && event.key === "ArrowLeft") {
+    if (!isModifiedKeyboardEvent() && event.key === "ArrowLeft") {
      event.preventDefault();
      navigateLastAssistantMessage("left");
    }
-    else if (!isModifiedKeyboardEvent(event) && event.key === "ArrowRight") {
+    else if (!isModifiedKeyboardEvent() && event.key === "ArrowRight") {
      event.preventDefault();
      if (!navigateLastAssistantMessage("right")) {
        // If can't navigate right (last version), regenerate
@ -157,8 +159,9 @@ targetElement.addEventListener("scroll", function() {
  let diff = targetElement.scrollHeight - targetElement.clientHeight;
  let isAtBottomNow = Math.abs(targetElement.scrollTop - diff) <= 10 || diff <= 0;
  // Add scrolling class to disable hover effects
  if (window.isScrolled || !isAtBottomNow) {
-    targetElement.classList.add("scrolling"); // Disables hover effects during scroll
+    targetElement.classList.add("scrolling");
  }
  if(isAtBottomNow) {
@ -199,8 +202,12 @@ const observer = new MutationObserver(function() {
 });
 // Only watch for attribute changes on targetElement (e.g. _generating class)
 const config = {
  attributes: true
 };
 // Start observing the target element
-observer.observe(targetElement, { attributes: true });
+observer.observe(targetElement, config);
 //------------------------------------------------
 // Handle syntax highlighting / LaTeX
@ -221,7 +228,7 @@ window.doSyntaxHighlighting = function() {
  if (messageBodies.length > 0) {
    let hasSeenVisible = false;
-    // Go from last message to first so we can early-exit once past visible area
+    // Go from last message to first
    for (let i = messageBodies.length - 1; i >= 0; i--) {
      const messageBody = messageBodies[i];
@ -236,8 +243,8 @@ window.doSyntaxHighlighting = function() {
          codeBlock.classList.add("pretty_scrollbar");
        });
        // Only render math in visible elements
        const mathContainers = messageBody.querySelectorAll("p, span, li, td, th, h1, h2, h3, h4, h5, h6, blockquote, figcaption, caption, dd, dt");
        // Only render math in individually visible containers (the outer check is on the message body)
        mathContainers.forEach(container => {
          if (isElementVisibleOnScreen(container)) {
            renderMathInElement(container, {
@ -264,7 +271,7 @@ const doSyntaxHighlighting = window.doSyntaxHighlighting;
 // Add some scrollbars
 //------------------------------------------------
 const scrollbarElements = document.querySelectorAll(".add_scrollbar textarea, .add_scrollbar .drag-drop-list");
-for(let i = 0; i < scrollbarElements.length; i++) {
+for(i = 0; i < scrollbarElements.length; i++) {
  scrollbarElements[i].classList.remove("scroll-hide");
  scrollbarElements[i].classList.add("pretty_scrollbar");
  scrollbarElements[i].style.resize = "none";
@ -291,13 +298,13 @@ if (toolsInfo) {
 // Remove some backgrounds
 //------------------------------------------------
 const noBackgroundelements = document.querySelectorAll(".no-background");
-for(let i = 0; i < noBackgroundelements.length; i++) {
+for(i = 0; i < noBackgroundelements.length; i++) {
  noBackgroundelements[i].parentNode.style.border = "none";
  noBackgroundelements[i].parentNode.parentNode.parentNode.style.alignItems = "center";
 }
 const slimDropdownElements = document.querySelectorAll(".slim-dropdown");
-for (let i = 0; i < slimDropdownElements.length; i++) {
+for (i = 0; i < slimDropdownElements.length; i++) {
  const parentNode = slimDropdownElements[i].parentNode;
  parentNode.style.background = "transparent";
  parentNode.style.border = "0";
@ -309,19 +316,18 @@ for (let i = 0; i < slimDropdownElements.length; i++) {
 // https://github.com/SillyTavern/SillyTavern/blob/6c8bd06308c69d51e2eb174541792a870a83d2d6/public/script.js
 //------------------------------------------------
 var buttonsInChat = document.querySelectorAll("#chat-tab #chat-buttons button, #chat-tab #chat-buttons #show-controls");
 var hoverContainer = document.getElementById("gr-hover-container");
 var button = document.getElementById("hover-element-button");
 var menu = document.getElementById("hover-menu");
 var istouchscreen = (navigator.maxTouchPoints > 0) || "ontouchstart" in document.documentElement;
 function showMenu() {
-  menu.style.display = "flex";
+  menu.style.display = "flex"; // Show the menu
 }
 function hideMenu() {
-  menu.style.display = "none";
+  menu.style.display = "none"; // Hide the menu
  if (!istouchscreen) {
-    document.querySelector("#chat-input textarea").focus();
+    document.querySelector("#chat-input textarea").focus(); // Focus on the chat input
  }
 }
@ -330,6 +336,7 @@ if (buttonsInChat.length > 0) {
    const thisButton = buttonsInChat[i];
    menu.appendChild(thisButton);
    // Only apply transformations to button elements
    if (thisButton.tagName.toLowerCase() === "button") {
      thisButton.addEventListener("click", () => {
        hideMenu();
@ -339,6 +346,7 @@ if (buttonsInChat.length > 0) {
      const matches = buttonText.match(/(\(.*?\))/);
      if (matches && matches.length > 1) {
        // Apply the transparent-substring class to the matched substring
        const substring = matches[1];
        const newText = buttonText.replace(substring, `&nbsp;<span class="transparent-substring">${substring.slice(1, -1)}</span>`);
        thisButton.innerHTML = newText;
@ -347,19 +355,16 @@ if (buttonsInChat.length > 0) {
  }
 }
-var menuInteracting = false;
+function isMouseOverButtonOrMenu() {
  return menu.matches(":hover") || button.matches(":hover");
 }
-hoverContainer.addEventListener("mouseenter", function () {
+button.addEventListener("mouseenter", function () {
  if (!istouchscreen) {
    showMenu();
  }
 });
 hoverContainer.addEventListener("mousedown", function () {
  menuInteracting = true;
  setTimeout(function () { menuInteracting = false; }, 300);
 });
 button.addEventListener("click", function () {
  if (menu.style.display === "flex") {
    hideMenu();
@ -369,39 +374,49 @@ button.addEventListener("click", function () {
  }
 });
-hoverContainer.addEventListener("mouseleave", function () {
+// Add event listener for mouseleave on the button
-  if (!istouchscreen) {
+button.addEventListener("mouseleave", function () {
-    setTimeout(function () {
+  // Delay to prevent menu hiding when the mouse leaves the button into the menu
-      if (!hoverContainer.matches(":hover") && !menu.matches(":hover")) {
+  setTimeout(function () {
-        hideMenu();
+    if (!isMouseOverButtonOrMenu()) {
-      }
+      hideMenu();
-    }, 50);
+    }
-  }
+  }, 100);
 });
 // Add event listener for mouseleave on the menu
 menu.addEventListener("mouseleave", function () {
  // Delay to prevent menu hide when the mouse leaves the menu into the button
  setTimeout(function () {
    if (!isMouseOverButtonOrMenu()) {
      hideMenu();
    }
  }, 100);
 });
 // Add event listener for click anywhere in the document
 document.addEventListener("click", function (event) {
  const target = event.target;
  // Check if the click is outside the button/menu and the menu is visible
-  if (!menuInteracting && !event.target.closest("#gr-hover-container") && menu.style.display === "flex") {
+  if (!isMouseOverButtonOrMenu() && menu.style.display === "flex") {
    hideMenu();
  }
-  const target = event.target;
+  if (event.target.classList.contains("pfp_character")) {
  if (target.classList.contains("pfp_character")) {
    toggleBigPicture();
  }
  // Handle sidebar clicks on mobile
  if (isMobile()) {
-    // Check if the click did NOT originate from any of the specified toggle buttons or elements
+  // Check if the click did NOT originate from any of the specified toggle buttons or elements
    if (
      target.closest("#navigation-toggle") !== navigationToggle &&
-      target.closest("#past-chats-toggle") !== pastChatsToggle &&
+    target.closest("#past-chats-toggle") !== pastChatsToggle &&
-      target.closest("#chat-controls-toggle") !== chatControlsToggle &&
+    target.closest("#chat-controls-toggle") !== chatControlsToggle &&
-      target.closest(".header_bar") !== headerBar &&
+    target.closest(".header_bar") !== headerBar &&
-      target.closest("#past-chats-row") !== pastChatsRow &&
+    target.closest("#past-chats-row") !== pastChatsRow &&
-      target.closest("#chat-controls") !== chatControlsRow
+    target.closest("#chat-controls") !== chatControlsRow
    ) {
      handleIndividualSidebarClose(event);
    }
@ -418,19 +433,27 @@ document.getElementById("chat-input-row").classList.add("chat-input-positioned")
 //------------------------------------------------
 const chatTextArea = document.getElementById("chat-input").querySelector("textarea");
-function focusOnVisible(element) {
+function respondToChatInputVisibility(element, callback) {
-  var observer = new IntersectionObserver((entries) => {
+  var options = {
    root: document.documentElement,
  };
  var observer = new IntersectionObserver((entries, observer) => {
    entries.forEach(entry => {
-      if (entry.intersectionRatio > 0) {
+      callback(entry.intersectionRatio > 0);
        element.focus();
      }
    });
-  }, { root: document.documentElement });
+  }, options);
  observer.observe(element);
 }
-focusOnVisible(chatTextArea);
+function handleChatInputVisibilityChange(isVisible) {
  if (isVisible) {
    chatTextArea.focus();
  }
 }
 respondToChatInputVisibility(chatTextArea, handleChatInputVisibilityChange);
 //------------------------------------------------
 // Show enlarged character picture when the profile
@ -440,7 +463,8 @@ let bigPictureVisible = false;
 function addBigPicture() {
  var imgElement = document.createElement("img");
-  imgElement.src = getProfilePictureUrl();
+  var timestamp = new Date().getTime();
  imgElement.src = "/file/user_data/cache/pfp_character.png?time=" + timestamp;
  imgElement.classList.add("bigProfilePicture");
  imgElement.addEventListener("load", function () {
    this.style.visibility = "visible";
@ -454,8 +478,9 @@ function addBigPicture() {
 }
 function deleteBigPicture() {
-  document.querySelectorAll(".bigProfilePicture").forEach(function (element) {
+  var bigProfilePictures = document.querySelectorAll(".bigProfilePicture");
-    element.remove();
+  bigProfilePictures.forEach(function (element) {
    element.parentNode.removeChild(element);
  });
 }
@ -469,11 +494,44 @@ function toggleBigPicture() {
  }
 }
 //------------------------------------------------
 // Handle the chat input box growth
 //------------------------------------------------
 // Cache DOM elements
 const chatContainer = document.getElementById("chat").parentNode.parentNode.parentNode;
 const chatInput = document.querySelector("#chat-input textarea");
 // Variables to store current dimensions
 let currentChatInputHeight = chatInput.clientHeight;
 //------------------------------------------------
 // Focus on the rename text area when it becomes visible
 //------------------------------------------------
 const renameTextArea = document.getElementById("rename-row").querySelector("textarea");
-focusOnVisible(renameTextArea);
+
 function respondToRenameVisibility(element, callback) {
  var options = {
    root: document.documentElement,
  };
  var observer = new IntersectionObserver((entries, observer) => {
    entries.forEach(entry => {
      callback(entry.intersectionRatio > 0);
    });
  }, options);
  observer.observe(element);
 }
 function handleVisibilityChange(isVisible) {
  if (isVisible) {
    renameTextArea.focus();
  }
 }
 respondToRenameVisibility(renameTextArea, handleVisibilityChange);
 //------------------------------------------------
 // Adjust the chat tab margin if no extension UI
@ -679,21 +737,21 @@ function handleIndividualSidebarClose(event) {
  // Close navigation bar if click is outside and it is open
  if (!headerBar.contains(target) && !headerBar.classList.contains("sidebar-hidden")) {
-    toggleSidebar(headerBar, navigationToggle);
+    toggleSidebar(headerBar, navigationToggle, true);
  }
  // Close past chats row if click is outside and it is open
  if (!pastChatsRow.contains(target) && !pastChatsRow.classList.contains("sidebar-hidden")) {
-    toggleSidebar(pastChatsRow, pastChatsToggle);
+    toggleSidebar(pastChatsRow, pastChatsToggle, true);
  }
  // Close chat controls row if click is outside and it is open
  if (!chatControlsRow.contains(target) && !chatControlsRow.classList.contains("sidebar-hidden")) {
-    toggleSidebar(chatControlsRow, chatControlsToggle);
+    toggleSidebar(chatControlsRow, chatControlsToggle, true);
  }
 }
-function toggleSidebar(sidebar, toggle) {
+function toggleSidebar(sidebar, toggle, forceClose = false) {
  const isCurrentlyHidden = sidebar.classList.contains("sidebar-hidden");
  const shouldClose = !isCurrentlyHidden;
@ -718,6 +776,11 @@ function toggleSidebar(sidebar, toggle) {
    toggle.classList.toggle("chat-controls-open", !shouldClose);
    toggle.innerHTML = shouldClose ? leftArrowSVG : rightArrowSVG;
  }
  // Mobile handling
  if (isMobile()) {
    sidebar.classList.toggle("sidebar-shown", !shouldClose);
  }
 }
 // Function to check if the device is mobile
@ -777,17 +840,17 @@ pastChatsToggle.addEventListener("click", () => {
  const isCurrentlyOpen = !pastChatsRow.classList.contains("sidebar-hidden");
  toggleSidebar(pastChatsRow, pastChatsToggle);
-  // On desktop, sync both sidebars together
+  // On desktop, open/close both sidebars at the same time
  if (!isMobile()) {
    if (isCurrentlyOpen) {
      // If we just closed the left sidebar, also close the right sidebar
      if (!chatControlsRow.classList.contains("sidebar-hidden")) {
-        toggleSidebar(chatControlsRow, chatControlsToggle);
+        toggleSidebar(chatControlsRow, chatControlsToggle, true);
      }
    } else {
      // If we just opened the left sidebar, also open the right sidebar
      if (chatControlsRow.classList.contains("sidebar-hidden")) {
-        toggleSidebar(chatControlsRow, chatControlsToggle);
+        toggleSidebar(chatControlsRow, chatControlsToggle, false);
      }
    }
  }
@ -797,17 +860,17 @@ chatControlsToggle.addEventListener("click", () => {
  const isCurrentlyOpen = !chatControlsRow.classList.contains("sidebar-hidden");
  toggleSidebar(chatControlsRow, chatControlsToggle);
-  // On desktop, sync both sidebars together
+  // On desktop, open/close both sidebars at the same time
  if (!isMobile()) {
    if (isCurrentlyOpen) {
      // If we just closed the right sidebar, also close the left sidebar
      if (!pastChatsRow.classList.contains("sidebar-hidden")) {
-        toggleSidebar(pastChatsRow, pastChatsToggle);
+        toggleSidebar(pastChatsRow, pastChatsToggle, true);
      }
    } else {
      // If we just opened the right sidebar, also open the left sidebar
      if (pastChatsRow.classList.contains("sidebar-hidden")) {
-        toggleSidebar(pastChatsRow, pastChatsToggle);
+        toggleSidebar(pastChatsRow, pastChatsToggle, false);
      }
    }
  }
@ -827,7 +890,7 @@ if (isMobile()) {
  const textarea = document.querySelector("#chat-input textarea");
  if (textarea) {
-    // Force textarea height recalculation by simulating content change
+    // Simulate adding and removing a newline
    textarea.value += "\n";
    textarea.dispatchEvent(new Event("input", { bubbles: true }));
    textarea.value = textarea.value.slice(0, -1);
--- a/js/save_files.js
+++ b/js/save_files.js
@ -1,9 +1,10 @@
 // Functions for downloading JSON files
 function getCurrentTimestamp() {
  const now = new Date();
-  const timezoneOffset = now.getTimezoneOffset() * 60000; // Convert minutes to milliseconds
+  const timezoneOffset = now.getTimezoneOffset() * 60000; // Convert to milliseconds
  const localTime = new Date(now.getTime() - timezoneOffset);
-  return localTime.toISOString().replace(/[-:]/g, "").slice(0, 15);
+  const formattedTimestamp = localTime.toISOString().replace(/[-:]/g, "").slice(0, 15);
  return formattedTimestamp;
 }
 function saveFile(contents, filename) {
@ -17,18 +18,23 @@ function saveFile(contents, filename) {
 }
 function saveHistory(history, character, mode) {
-  let path;
+  let path = null;
  if (["chat", "chat-instruct"].includes(mode) && character && character.trim() !== "") {
    path = `history_${character}_${getCurrentTimestamp()}.json`;
  } else {
-    path = `history_${mode || "unknown"}_${getCurrentTimestamp()}.json`;
+    try {
      path = `history_${mode}_${getCurrentTimestamp()}.json`;
    } catch (error) {
      path = `history_${getCurrentTimestamp()}.json`;
    }
  }
  saveFile(history, path);
 }
 function saveSession(session) {
-  const path = `session_${getCurrentTimestamp()}.json`;
+  let path = null;
  path = `session_${getCurrentTimestamp()}.json`;
  saveFile(session, path);
 }
--- a/js/show_controls.js
+++ b/js/show_controls.js
@ -1,11 +1,13 @@
 const chatParent = document.querySelector(".chat-parent");
 function toggle_controls(value) {
  const navToggle = document.getElementById("navigation-toggle");
  const pastChatsToggle = document.getElementById("past-chats-toggle");
  const extensions = document.querySelector("#extensions");
  const galleryExtension = document.getElementById("gallery-extension");
  if (value) {
    // SHOW MODE: Click toggles to show hidden sidebars
    const navToggle = document.getElementById("navigation-toggle");
    const pastChatsToggle = document.getElementById("past-chats-toggle");
    if (navToggle && document.querySelector(".header_bar")?.classList.contains("sidebar-hidden")) {
      navToggle.click();
    }
@ -17,11 +19,17 @@ function toggle_controls(value) {
    if (extensions) {
      extensions.style.display = "inherit";
    }
-    if (galleryExtension) {
+
-      galleryExtension.style.display = "block";
+    let gallery_element = document.getElementById("gallery-extension");
    if (gallery_element) {
      gallery_element.style.display = "block";
    }
  } else {
    // HIDE MODE: Click toggles to hide visible sidebars
    const navToggle = document.getElementById("navigation-toggle");
    const pastChatsToggle = document.getElementById("past-chats-toggle");
    if (navToggle && !document.querySelector(".header_bar")?.classList.contains("sidebar-hidden")) {
      navToggle.click();
    }
@ -33,8 +41,5 @@ function toggle_controls(value) {
    if (extensions) {
      extensions.style.display = "none";
    }
    if (galleryExtension) {
      galleryExtension.style.display = "none";
    }
  }
 }
--- a/js/switch_tabs.js
+++ b/js/switch_tabs.js
@ -2,9 +2,17 @@ function scrollToTop() {
  window.scrollTo({ top: 0 });
 }
-function findButtonsByText(buttonText, container = document) {
+function findButtonsByText(buttonText) {
-  return Array.from(container.getElementsByTagName("button"))
+  const buttons = document.getElementsByTagName("button");
-    .filter(btn => btn.textContent.trim() === buttonText);
+  const matchingButtons = [];
  for (let i = 0; i < buttons.length; i++) {
    if (buttons[i].textContent.trim() === buttonText) {
      matchingButtons.push(buttons[i]);
    }
  }
  return matchingButtons;
 }
 function switch_to_chat() {
@ -31,9 +39,13 @@ function switch_to_character() {
 function switch_to_image_ai_generate() {
  const container = document.querySelector("#image-ai-tab");
-  const generateBtn = findButtonsByText("Generate", container)[0];
+  const buttons = container.getElementsByTagName("button");
-  if (generateBtn) {
+
-    generateBtn.click();
+  for (let i = 0; i < buttons.length; i++) {
    if (buttons[i].textContent.trim() === "Generate") {
      buttons[i].click();
      break;
    }
  }
  scrollToTop();
--- a/js/update_big_picture.js
+++ b/js/update_big_picture.js
@ -1,6 +1,7 @@
 function updateBigPicture() {
  var existingElement = document.querySelector(".bigProfilePicture");
  if (existingElement) {
-    existingElement.src = getProfilePictureUrl();
+    var timestamp = new Date().getTime();
    existingElement.src = "/file/user_data/cache/pfp_character.png?time=" + timestamp;
  }
 }
--- a/modules/api/completions.py
+++ b/modules/api/completions.py
@ -39,146 +39,6 @@ def load_chat_template_file(filepath):
    return text
 def _first_token_display_str(token_id, prompt, tokenizer):
    """Return the display string for the first prompt token.
    Returns empty string for BOS or tokens that don't appear at the start
    of the prompt text, so they don't shift text_offset for subsequent tokens.
    """
    token_id = int(token_id)
    bos_id = getattr(tokenizer, 'bos_token_id', None)
    if bos_id is not None and token_id == bos_id:
        return ""
    import torch
    tok = tokenizer.decode(torch.tensor([token_id]))
    if not prompt.startswith(tok):
        return ""
    return tok
 def _compute_prompt_logprob_entries(prompt, logprobs_count, input_ids=None):
    """Compute logprob entries for prompt tokens via a forward pass.
    Returns a list of logprob entries in the standard format.
    The first token gets a null entry (no conditioning context).
    Supported for HF-compatible loaders (Transformers, ExLlamav3_HF, etc.)
    via a single forward pass, and for llama.cpp via the server's
    prompt_logprobs parameter. Returns [] for unsupported loaders.
    """
    if input_ids is None:
        input_ids = encode(prompt)  # (1, seq_len) tensor or array
    token_ids = input_ids[0]
    n_tokens = len(token_ids)
    if n_tokens == 0:
        return []
    loader = shared.args.loader
    model = shared.model
    if loader == 'llama.cpp':
        return model.get_prompt_logprob_entries(token_ids, max(logprobs_count, 1), prompt=prompt)
    first_token_str = _first_token_display_str(token_ids[0], prompt, shared.tokenizer)
    if n_tokens <= 1:
        return [{"token": first_token_str, "null_logprob": True}]
    import torch
    from modules.torch_utils import clear_torch_cache
    if hasattr(model, 'get_prompt_logits'):
        logits = model.get_prompt_logits(input_ids)
    elif hasattr(model, 'forward'):
        # HF-compatible loaders (Transformers, ExLlamav3_HF, etc.)
        input_ids_tensor = input_ids if isinstance(input_ids, torch.Tensor) else torch.tensor(input_ids, dtype=torch.long)
        if hasattr(model, 'device'):
            input_ids_tensor = input_ids_tensor.to(model.device)
        with torch.no_grad():
            # Pass labels to ensure logits are returned for ALL positions,
            # not just the last token (some HF wrappers like ExLlamav3_HF
            # only compute the last-token logits when labels are absent).
            outputs = model(input_ids=input_ids_tensor, labels=input_ids_tensor)
            logits = outputs.logits  # keep on GPU, (1, seq_len, vocab) in model dtype
            del outputs
    else:
        return []
    entries = [{"token": first_token_str, "null_logprob": True}]
    logprobs_count = max(logprobs_count, 1)
    k = min(logprobs_count, logits.shape[-1])
    chunk_size = 2048
    unique_ids = set(int(tid) for tid in token_ids[1:])
    # Process logits in chunks on GPU, only move top-K results to CPU
    all_top_log_probs_list = []
    all_top_indices_list = []
    all_actual_lps = []
    for start in range(0, n_tokens - 1, chunk_size):
        end = min(start + chunk_size, n_tokens - 1)
        chunk_logits = logits[0, start:end].float()  # (chunk, vocab) on GPU
        chunk_lse = torch.logsumexp(chunk_logits, dim=-1)
        chunk_top_values, chunk_top_indices = torch.topk(chunk_logits, k=k, dim=-1)
        chunk_top_log_probs = chunk_top_values - chunk_lse.unsqueeze(-1)
        # Compute logprob for actual next tokens in this chunk
        chunk_top_sets = [set(chunk_top_indices[j].tolist()) for j in range(end - start)]
        for j in range(end - start):
            actual_tid = int(token_ids[start + j + 1])
            if actual_tid not in chunk_top_sets[j]:
                all_actual_lps.append((chunk_logits[j, actual_tid] - chunk_lse[j]).item())
            else:
                all_actual_lps.append(None)  # will use top_log_probs
        all_top_log_probs_list.append(chunk_top_log_probs.cpu())
        all_top_indices_list.append(chunk_top_indices.cpu())
        unique_ids.update(int(tid) for tid in chunk_top_indices.flatten().tolist())
        del chunk_logits, chunk_lse, chunk_top_values
    del logits
    clear_torch_cache()
    all_top_log_probs = torch.cat(all_top_log_probs_list, dim=0)
    all_top_indices = torch.cat(all_top_indices_list, dim=0)
    unique_ids_list = sorted(unique_ids)
    decoded_list = shared.tokenizer.batch_decode([[tid] for tid in unique_ids_list]) if hasattr(shared.tokenizer, 'batch_decode') else [shared.tokenizer.decode(torch.tensor([tid])) for tid in unique_ids_list]
    decoded_strs = dict(zip(unique_ids_list, decoded_list))
    for i in range(1, n_tokens):
        token_id = int(token_ids[i])
        idx = i - 1
        top_log_probs = all_top_log_probs[idx]
        top_ids = all_top_indices[idx].tolist()
        actual_token_str = decoded_strs[token_id]
        if token_id in top_ids:
            actual_lp = top_log_probs[top_ids.index(token_id)].item()
            alternatives = [
                {"token": decoded_strs[top_ids[j]], "token_id": top_ids[j], "logprob": top_log_probs[j].item()}
                for j in range(k) if top_ids[j] != token_id
            ]
        else:
            actual_lp = all_actual_lps[idx]
            alternatives = [
                {"token": decoded_strs[top_ids[j]], "token_id": top_ids[j], "logprob": top_log_probs[j].item()}
                for j in range(k - 1)
            ]
        entry = {"top_logprobs": [{"token": actual_token_str, "token_id": token_id, "logprob": actual_lp}] + alternatives}
        entries.append(entry)
    return entries
 def _get_raw_logprob_entries(offset=0):
    """Get raw logprob entries from llama.cpp/ExLlamav3 backend, starting from offset.
@ -205,21 +65,6 @@ def _parse_entry_top(entry):
    return entry.get('top_logprobs', entry.get('top_probs', []))
 def _extract_sampled_token(entry, top):
    """Get the actually sampled token and its logprob from a logprob entry.
    Uses the entry-level token/logprob when available (the actually sampled
    token), falling back to top[0] (highest-probability alternative) which
    may differ with non-greedy sampling.
    """
    if 'token' in entry:
        return entry['token'], entry.get('logprob', entry.get('prob', 0))
    token_str = top[0].get('token', '')
    token_logprob = top[0].get('logprob', top[0].get('prob', 0))
    return token_str, token_logprob
 def format_chat_logprobs(entries):
    """Format logprob entries into OpenAI chat completions logprobs format.
@ -234,7 +79,9 @@ def format_chat_logprobs(entries):
        if not top:
            continue
-        token_str, token_logprob = _extract_sampled_token(entry, top)
+        chosen = top[0]
        token_str = chosen.get('token', '')
        token_logprob = chosen.get('logprob', chosen.get('prob', 0))
        top_list = []
        for item in top:
@ -259,7 +106,7 @@ def format_chat_logprobs(entries):
 def format_completion_logprobs(entries):
    """Format logprob entries into OpenAI completions logprobs format.
-    Output: {"tokens", "token_logprobs", "top_logprobs": [{token: prob}], "top_logprobs_ids": [{token_id: prob}], "text_offset"}
+    Output: {"tokens", "token_logprobs", "top_logprobs": [{token: prob}], "text_offset"}
    """
    if not entries:
        return None
@ -267,27 +114,17 @@ def format_completion_logprobs(entries):
    tokens = []
    token_logprobs = []
    top_logprobs = []
    top_logprobs_ids = []
    text_offset = []
    offset = 0
    for entry in entries:
        # Handle null logprob entries (first prompt token with echo)
        if entry.get("null_logprob"):
            token_str = entry.get("token", "")
            tokens.append(token_str)
            token_logprobs.append(None)
            top_logprobs.append(None)
            top_logprobs_ids.append(None)
            text_offset.append(offset)
            offset += len(token_str)
            continue
        top = _parse_entry_top(entry)
        if not top:
            continue
-        token_str, token_logprob = _extract_sampled_token(entry, top)
+        chosen = top[0]
        token_str = chosen.get('token', '')
        token_logprob = chosen.get('logprob', chosen.get('prob', 0))
        tokens.append(token_str)
        token_logprobs.append(token_logprob)
@ -295,29 +132,21 @@ def format_completion_logprobs(entries):
        offset += len(token_str)
        top_dict = {}
        top_dict_ids = {}
        for item in top:
            t = item.get('token', '')
            lp = item.get('logprob', item.get('prob', 0))
            top_dict[t] = lp
            tid = item.get('token_id', item.get('id'))
            if tid is not None:
                top_dict_ids[tid] = lp
        top_logprobs.append(top_dict)
        top_logprobs_ids.append(top_dict_ids if top_dict_ids else None)
    if not tokens:
        return None
-    result = {
+    return {
        "tokens": tokens,
        "token_logprobs": token_logprobs,
        "top_logprobs": top_logprobs,
        "text_offset": text_offset
    }
    if any(x is not None for x in top_logprobs_ids):
        result["top_logprobs_ids"] = top_logprobs_ids
    return result
 def process_parameters(body, is_legacy=False):
@ -578,10 +407,7 @@ def chat_completions_common(body: dict, is_legacy: bool = False, stream=False, p
    })
    max_tokens = generate_params['max_new_tokens']
-    if max_tokens is not None and max_tokens <= 0:
+    if max_tokens in [None, 0]:
        raise InvalidRequestError(message="max_tokens must be greater than 0.", param="max_tokens")
    if max_tokens is None:
        generate_params['max_new_tokens'] = 512
        generate_params['auto_max_new_tokens'] = True
@ -826,15 +652,6 @@ def completions_common(body: dict, is_legacy: bool = False, stream=False, stop_e
    # common params
    generate_params = process_parameters(body, is_legacy=is_legacy)
    max_tokens = generate_params['max_new_tokens']
    if max_tokens is None:
        generate_params['max_new_tokens'] = 512
        generate_params['auto_max_new_tokens'] = True
        max_tokens = 512
    elif max_tokens < 0:
        raise InvalidRequestError(message="max_tokens must be greater than or equal to 0.", param="max_tokens")
    elif max_tokens == 0 and body.get('logprobs') is None:
        raise InvalidRequestError(message="max_tokens is 0 but no logprobs parameter was specified.", param="max_tokens")
    generate_params['stream'] = stream
    if stop_event is not None:
        generate_params['stop_event'] = stop_event
@ -883,17 +700,9 @@ def completions_common(body: dict, is_legacy: bool = False, stream=False, stop_e
                        prompt = decode(prompt)[0]
            prefix = prompt if echo else ''
-            prompt_input_ids = encode(prompt)
+            token_count = len(encode(prompt)[0])
            token_count = len(prompt_input_ids[0])
            total_prompt_token_count += token_count
            # Compute prompt logprobs once per prompt (shared across n_completions)
            logprobs_val = body.get('logprobs', None)
            if echo and logprobs_val is not None and logprobs_val >= 0:
                prompt_entries = _compute_prompt_logprob_entries(prompt, logprobs_val, input_ids=prompt_input_ids)
            else:
                prompt_entries = None
            original_seed = generate_params.get('seed', -1)
            for _n in range(n_completions):
                # Increment seed for each completion to ensure diversity (matches llama.cpp native behavior)
@ -904,41 +713,29 @@ def completions_common(body: dict, is_legacy: bool = False, stream=False, stop_e
                    logprob_proc.token_alternatives_history.clear()
                # generate reply #######################################
-                if max_tokens == 0:
+                debug_msg({'prompt': prompt, 'generate_params': generate_params})
-                    answer = ''
+                generator = generate_reply(prompt, generate_params, is_chat=False)
-                    completion_token_count = 0
+                answer = ''
                    stop_reason = "stop"
                else:
                    debug_msg({'prompt': prompt, 'generate_params': generate_params})
                    generator = generate_reply(prompt, generate_params, is_chat=False)
                    answer = ''
-                    for a in generator:
+                for a in generator:
-                        answer = a
+                    answer = a
                    completion_token_count = len(encode(answer)[0])
                    stop_reason = "stop"
                    if token_count + completion_token_count >= generate_params['truncation_length'] or completion_token_count >= max_tokens:
                        stop_reason = "length"
                completion_token_count = len(encode(answer)[0])
                total_completion_token_count += completion_token_count
                stop_reason = "stop"
                if token_count + completion_token_count >= generate_params['truncation_length'] or completion_token_count >= max_tokens:
                    stop_reason = "length"
-                if max_tokens == 0:
+                if logprob_proc:
                    all_entries = []
                    for alt in logprob_proc.token_alternatives_history:
                        all_entries.extend(_dict_to_logprob_entries(alt))
                    completion_logprobs = format_completion_logprobs(all_entries)
                elif shared.args.loader in ('llama.cpp', 'ExLlamav3'):
                    raw = getattr(shared.model, 'last_completion_probabilities', None)
                    completion_logprobs = format_completion_logprobs(raw)
                else:
-                    if logprob_proc:
+                    completion_logprobs = None
                        all_entries = []
                        for alt in logprob_proc.token_alternatives_history:
                            all_entries.extend(_dict_to_logprob_entries(alt))
                    elif shared.args.loader in ('llama.cpp', 'ExLlamav3'):
                        all_entries = getattr(shared.model, 'last_completion_probabilities', None) or []
                    else:
                        all_entries = []
                if prompt_entries:
                    all_entries = prompt_entries + all_entries
                completion_logprobs = format_completion_logprobs(all_entries) if all_entries else None
                respi = {
                    "index": choice_index,
@ -978,8 +775,7 @@ def completions_common(body: dict, is_legacy: bool = False, stream=False, stop_e
                raise InvalidRequestError(message="API Batched generation not yet supported.", param=prompt_str)
        prefix = prompt if echo else ''
-        prompt_input_ids = encode(prompt)
+        token_count = len(encode(prompt)[0])
        token_count = len(prompt_input_ids[0])
        # Check if usage should be included in streaming chunks per OpenAI spec
        stream_options = body.get('stream_options')
@ -1012,57 +808,37 @@ def completions_common(body: dict, is_legacy: bool = False, stream=False, stop_e
            return chunk
        logprobs_val = body.get('logprobs', None)
        if echo and logprobs_val is not None and logprobs_val >= 0:
            prompt_entries = _compute_prompt_logprob_entries(prompt, logprobs_val, input_ids=prompt_input_ids)
            prompt_logprobs_formatted = format_completion_logprobs(prompt_entries) if prompt_entries else None
        else:
            prompt_logprobs_formatted = None
        # Clear stale logprobs from any previous request before building the
        # first chunk, so text_streaming_chunk doesn't pick up old data.
        if hasattr(shared.model, 'last_completion_probabilities'):
            shared.model.last_completion_probabilities = []
        cmpl_logprobs_offset[0] = 0
        chunk = text_streaming_chunk(prefix)
        if prompt_logprobs_formatted is not None:
            chunk[resp_list][0]["logprobs"] = prompt_logprobs_formatted
        if include_usage:
            chunk['usage'] = None
        yield chunk
        # generate reply #######################################
-        if max_tokens == 0:
+        debug_msg({'prompt': prompt, 'generate_params': generate_params})
-            answer = ''
+        generator = generate_reply(prompt, generate_params, is_chat=False)
-            completion_token_count = 0
+        answer = ''
-            stop_reason = "stop"
+        seen_content = ''
-        else:
+        completion_token_count = 0
            debug_msg({'prompt': prompt, 'generate_params': generate_params})
            generator = generate_reply(prompt, generate_params, is_chat=False)
            answer = ''
            seen_content = ''
            completion_token_count = 0
-            for a in generator:
+        for a in generator:
-                answer = a
+            answer = a
-                len_seen = len(seen_content)
+            len_seen = len(seen_content)
-                new_content = answer[len_seen:]
+            new_content = answer[len_seen:]
-                if not new_content or chr(0xfffd) in new_content:  # partial unicode character, don't send it yet.
+            if not new_content or chr(0xfffd) in new_content:  # partial unicode character, don't send it yet.
-                    continue
+                continue
-                seen_content = answer
+            seen_content = answer
-                chunk = text_streaming_chunk(new_content)
+            chunk = text_streaming_chunk(new_content)
-                if include_usage:
+            if include_usage:
-                    chunk['usage'] = None
+                chunk['usage'] = None
-                yield chunk
+            yield chunk
-            completion_token_count = len(encode(answer)[0])
+        completion_token_count = len(encode(answer)[0])
-            stop_reason = "stop"
+        stop_reason = "stop"
-            if token_count + completion_token_count >= generate_params['truncation_length'] or completion_token_count >= max_tokens:
+        if token_count + completion_token_count >= generate_params['truncation_length'] or completion_token_count >= max_tokens:
-                stop_reason = "length"
+            stop_reason = "length"
        chunk = text_streaming_chunk(suffix)
        chunk[resp_list][0]["finish_reason"] = stop_reason
--- a/modules/api/embeddings.py
+++ b/modules/api/embeddings.py
@ -6,7 +6,6 @@ from transformers import AutoModel
 from .errors import ServiceUnavailableError
 from .utils import debug_msg, float_list_to_base64
 from modules.logging_colors import logger
 from modules import shared
 embeddings_params_initialized = False
@ -42,7 +41,7 @@ def load_embedding_model(model: str):
    try:
        logger.info(f"Try embedding model: {model} on {embeddings_device}")
        if 'jina-embeddings' in model:
-            embeddings_model = AutoModel.from_pretrained(model, trust_remote_code=shared.args.trust_remote_code)
+            embeddings_model = AutoModel.from_pretrained(model, trust_remote_code=True)  # trust_remote_code is needed to use the encode method
            embeddings_model = embeddings_model.to(embeddings_device)
        else:
            embeddings_model = SentenceTransformer(model, device=embeddings_device)
--- a/modules/api/images.py
+++ b/modules/api/images.py
@ -4,11 +4,8 @@ OpenAI-compatible image generation using local diffusion models.
 import base64
 import io
 import json
 import time
 from PIL.PngImagePlugin import PngInfo
 from .errors import ServiceUnavailableError
 from modules import shared
@ -18,7 +15,7 @@ def generations(request):
    Generate images using the loaded diffusion model.
    Returns dict with 'created' timestamp and 'data' list of images.
    """
-    from modules.ui_image_generation import build_generation_metadata, generate
+    from modules.ui_image_generation import generate
    if shared.image_model is None:
        raise ServiceUnavailableError("No image model loaded. Load a model via the UI first.")
@ -49,18 +46,10 @@ def generations(request):
    if not images:
        raise ServiceUnavailableError("Image generation failed or produced no images.")
-    # Build response with per-batch metadata (seed increments per batch)
+    # Build response
    base_seed = state.get('image_seed_resolved', state['image_seed'])
    batch_size = int(state['image_batch_size'])
    resp = {'created': int(time.time()), 'data': []}
-    for idx, img in enumerate(images):
+    for img in images:
-        batch_seed = base_seed + idx // batch_size
+        b64 = _image_to_base64(img)
        metadata = build_generation_metadata(state, batch_seed)
        metadata_json = json.dumps(metadata, ensure_ascii=False)
        png_info = PngInfo()
        png_info.add_text("image_gen_settings", metadata_json)
        b64 = _image_to_base64(img, png_info)
        image_obj = {'revised_prompt': request.prompt}
@ -74,7 +63,7 @@ def generations(request):
    return resp
-def _image_to_base64(image, png_info=None) -> str:
+def _image_to_base64(image) -> str:
    buffered = io.BytesIO()
-    image.save(buffered, format="PNG", pnginfo=png_info)
+    image.save(buffered, format="PNG")
    return base64.b64encode(buffered.getvalue()).decode('utf-8')
--- a/modules/api/models.py
+++ b/modules/api/models.py
@ -2,7 +2,7 @@ from modules import loaders, shared
 from modules.logging_colors import logger
 from modules.LoRA import add_lora_to_model
 from modules.models import load_model, unload_model
-from modules.models_settings import get_model_metadata, load_instruction_template, update_model_parameters
+from modules.models_settings import get_model_metadata, update_model_parameters
 from modules.utils import get_available_loras, get_available_models
@ -42,10 +42,12 @@ def model_info_dict(model_name: str) -> dict:
 def _load_model(data):
    model_name = data["model_name"]
-    args = data.get("args")
+    args = data["args"]
    settings = data["settings"]
    unload_model()
    model_settings = get_model_metadata(model_name)
    update_model_parameters(model_settings)
    # Update shared.args with custom model loading settings
    # Security: only allow keys that correspond to model loading
@ -53,16 +55,6 @@ def _load_model(data):
    # flags like trust_remote_code or extra_flags to be set via the API.
    blocked_keys = {'extra_flags'}
    allowed_keys = set(loaders.list_model_elements()) - blocked_keys
    # Reset all loader args to their startup values before applying new ones,
    # so settings from a previous API load don't leak into this one.
    # Include blocked keys in the reset (safe: restores startup value, not API-controlled).
    for k in allowed_keys | blocked_keys:
        if hasattr(shared.args, k) and hasattr(shared.original_args, k):
            setattr(shared.args, k, getattr(shared.original_args, k))
    update_model_parameters(model_settings)
    if args:
        for k in args:
            if k in allowed_keys and hasattr(shared.args, k):
@ -70,12 +62,15 @@ def _load_model(data):
    shared.model, shared.tokenizer = load_model(model_name)
-    if data.get("instruction_template_str") is not None:
+    # Update shared.settings with custom generation defaults
-        shared.settings['instruction_template_str'] = data["instruction_template_str"]
+    if settings:
-        logger.info("INSTRUCTION TEMPLATE: set to custom Jinja2 string")
+        for k in settings:
-    elif data.get("instruction_template") is not None:
+            if k in shared.settings:
-        shared.settings['instruction_template_str'] = load_instruction_template(data["instruction_template"])
+                shared.settings[k] = settings[k]
-        logger.info(f"INSTRUCTION TEMPLATE: {data['instruction_template']}")
+                if k == 'truncation_length':
                    logger.info(f"TRUNCATION LENGTH (UPDATED): {shared.settings['truncation_length']}")
                elif k == 'instruction_template':
                    logger.info(f"INSTRUCTION TEMPLATE (UPDATED): {shared.settings['instruction_template']}")
 def list_loras():
--- a/modules/api/script.py
+++ b/modules/api/script.py
@ -475,8 +475,10 @@ async def handle_list_models():
@app.post("/v1/internal/model/load", dependencies=check_admin_key)
 async def handle_load_model(request_data: LoadModelRequest):
    '''
-    The "args" parameter can be used to modify loader flags before loading
+    This endpoint is experimental and may change in the future.
-    a model. Example:
+
    The "args" parameter can be used to modify flags like "--load-in-4bit"
    or "--n-gpu-layers" before loading a model. Example:
    ```
    "args": {
@ -485,13 +487,18 @@ async def handle_load_model(request_data: LoadModelRequest):
    }
    ```
-    Loader args are reset to their startup defaults between loads, so
+    Note that those settings will remain after loading the model. So you
-    settings from a previous load do not leak into the next one.
+    may need to change them back to load a second model.
-    The "instruction_template" parameter sets the default instruction
+    The "settings" parameter is also a dict but with keys for the
-    template by name (from user_data/instruction-templates/). The
+    shared.settings object. It can be used to modify the default instruction
-    "instruction_template_str" parameter sets it as a raw Jinja2 string
+    template like this:
-    and takes precedence over "instruction_template".
+
    ```
    "settings": {
      "instruction_template": "Alpaca"
    }
    ```
    '''
    try:
@ -537,8 +544,8 @@ async def handle_unload_loras():
 def find_available_port(starting_port):
    """Try the starting port, then find an available one if it's taken."""
    try:
        # Try to create a socket with the starting port
        with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
            s.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
            s.bind(('', starting_port))
            return starting_port
    except OSError:
@ -563,7 +570,7 @@ def run_server():
        server_addrs.append(shared.args.listen_host)
    else:
        if os.environ.get('OPENEDAI_ENABLE_IPV6', shared.args.api_enable_ipv6):
-            server_addrs.append('::' if shared.args.listen else '::1')
+            server_addrs.append('[::]' if shared.args.listen else '[::1]')
        if not os.environ.get('OPENEDAI_DISABLE_IPV4', shared.args.api_disable_ipv4):
            server_addrs.append('0.0.0.0' if shared.args.listen else '127.0.0.1')
@ -580,7 +587,7 @@ def run_server():
        )
    else:
        url_proto = 'https://' if (ssl_certfile and ssl_keyfile) else 'http://'
-        urls = [f'{url_proto}[{addr}]:{port}/v1' if ':' in addr else f'{url_proto}{addr}:{port}/v1' for addr in server_addrs]
+        urls = [f'{url_proto}{addr}:{port}/v1' for addr in server_addrs]
        if len(urls) > 1:
            logger.info('OpenAI/Anthropic-compatible API URLs:\n\n' + '\n'.join(urls) + '\n')
        else:
--- a/modules/api/typing.py
+++ b/modules/api/typing.py
@ -271,8 +271,7 @@ class ModelListResponse(BaseModel):
 class LoadModelRequest(BaseModel):
    model_name: str
    args: dict | None = None
-    instruction_template: str | None = Field(default=None, description="An instruction template defined under text-generation-webui/user_data/instruction-templates. Sets the default template for all subsequent API requests.")
+    settings: dict | None = None
    instruction_template_str: str | None = Field(default=None, description="A Jinja2 instruction template string. If set, takes precedence over instruction_template.")
 class LoraListResponse(BaseModel):
--- a/modules/chat.py
+++ b/modules/chat.py
@ -210,57 +210,6 @@ def _expand_tool_sequence(tool_seq):
    return messages
 def _convert_to_tool_responses(messages):
    """Convert role:'tool' messages to tool_responses format.
    Templates like Gemma 4 expect tool results as a ``tool_responses``
    attribute on a message rather than separate ``role: 'tool'`` messages.
    This function groups consecutive tool messages and rewrites them.
    """
    result = []
    tc_id_to_name = {}
    i = 0
    while i < len(messages):
        msg = messages[i]
        if msg.get('tool_calls'):
            for tc in msg['tool_calls']:
                tc_id = tc.get('id', '')
                func_name = tc.get('function', {}).get('name', 'unknown')
                if tc_id:
                    tc_id_to_name[tc_id] = func_name
        if msg.get('role') == 'tool':
            tool_responses = []
            while i < len(messages) and messages[i].get('role') == 'tool':
                tool_msg = messages[i]
                tc_id = tool_msg.get('tool_call_id', '')
                func_name = tc_id_to_name.get(tc_id, 'unknown')
                content = tool_msg.get('content', '')
                try:
                    response = json.loads(content)
                except (json.JSONDecodeError, ValueError, TypeError):
                    response = content
                tool_responses.append({
                    'name': func_name,
                    'response': response,
                })
                i += 1
            result.append({
                'role': 'tool',
                'tool_responses': tool_responses,
            })
        else:
            result.append(msg)
            i += 1
    return result
 def _format_attachments(attachments, include_text=True):
    """Build image ref and text attachment strings from a list of attachments."""
    attachments_text = ""
@ -318,9 +267,6 @@ def generate_chat_prompt(user_input, state, **kwargs):
        tools=state['tools'] if 'tools' in state else None,
    )
    active_template_str = state['instruction_template_str'] if state['mode'] == 'instruct' else chat_template_str
    uses_tool_responses = 'tool_responses' in active_template_str
    messages = []
    if state['mode'] == 'instruct':
@ -557,9 +503,6 @@ def generate_chat_prompt(user_input, state, **kwargs):
        return prompt
    if uses_tool_responses:
        messages = _convert_to_tool_responses(messages)
    prompt = make_prompt(messages)
    # Handle truncation
@ -568,24 +511,13 @@ def generate_chat_prompt(user_input, state, **kwargs):
        encoded_length = get_encoded_length(prompt)
        while len(messages) > 0 and encoded_length > max_length:
            # Remove old message, save system message
            if len(messages) > 2 and messages[0]['role'] == 'system':
-                pop_idx = 1
+                messages.pop(1)
            # Remove old message when no system message is present
            elif len(messages) > 1 and messages[0]['role'] != 'system':
-                pop_idx = 0
+                messages.pop(0)
            else:
                pop_idx = None
            if pop_idx is not None:
                messages.pop(pop_idx)
                # Remove orphaned tool-call/tool-result messages that
                # would be invalid without their partner.
                while pop_idx < len(messages):
                    msg = messages[pop_idx]
                    if msg.get('role') == 'tool' or (msg.get('role') == 'assistant' and msg.get('tool_calls')):
                        messages.pop(pop_idx)
                    else:
                        break
            # Resort to truncating the user input
            else:
@ -705,7 +637,7 @@ def get_stopping_strings(state):
        # Find positions of each message content
        first_user_end = prompt.find("first user message") + len("first user message")
        first_assistant_start = prompt.find("first assistant message")
-        first_assistant_end = first_assistant_start + len("first assistant message")
+        first_assistant_end = prompt.find("first assistant message") + len("first assistant message")
        second_user_start = prompt.find("second user message")
        second_assistant_end = prompt.find("second assistant message") + len("second assistant message")
@ -739,10 +671,7 @@ def get_stopping_strings(state):
    # Handle GPT-OSS as a special case
    if '<|channel|>final<|message|>' in state['instruction_template_str'] and "<|end|>" in result:
        result.remove("<|end|>")
-        if '<|result|>' in state['instruction_template_str']:
+        result.append("<|result|>")
            result.append("<|result|>")
        elif '<|return|>' in state['instruction_template_str']:
            result.append("<|return|>")
        result = list(set(result))
    if shared.args.verbose:
@ -1194,7 +1123,7 @@ def chatbot_wrapper(text, state, regenerate=False, _continue=False, loading_mess
    # visible text from before buffering started so raw markup doesn't flash
    # in the UI.  The internal text is left intact so the caller can still
    # parse tool calls from it.
-    if is_stream and _check_tool_markers and streaming_tool_buffer_check(output['internal'][-1][1], markers=_streaming_markers, tool_names=_tool_names, check_bare_names=_check_bare_names, partial_match=False):
+    if is_stream and _check_tool_markers and streaming_tool_buffer_check(output['internal'][-1][1], markers=_streaming_markers, tool_names=_tool_names, check_bare_names=_check_bare_names):
        output['visible'][-1][1] = _last_visible_before_tool_buffer or ''
    yield output
@ -1275,23 +1204,14 @@ def generate_chat_reply_wrapper(text, state, regenerate=False, _continue=False):
    # Load tools if any are selected
    selected = state.get('selected_tools', [])
    mcp_servers = state.get('mcp_servers', '')
    parse_tool_call = None
    _tool_parsers = None
-    if selected or mcp_servers:
+    if selected:
-        from modules.tool_use import load_tools, load_mcp_tools, execute_tool
+        from modules.tool_use import load_tools, execute_tool
        from modules.tool_parsing import parse_tool_call, get_tool_call_id, detect_tool_call_format
    if selected:
        tool_defs, tool_executors = load_tools(selected)
        if mcp_servers:
            mcp_defs, mcp_executors = load_mcp_tools(mcp_servers)
            for td in mcp_defs:
                fn = td['function']['name']
                if fn in tool_executors:
                    logger.warning(f'MCP tool "{fn}" conflicts with a local tool. Skipping.')
                    continue
                tool_defs.append(td)
                tool_executors[fn] = mcp_executors[fn]
        state['tools'] = tool_defs
        tool_func_names = [t['function']['name'] for t in tool_defs]
        _template_str = state.get('instruction_template_str', '') if state.get('mode') == 'instruct' else state.get('chat_template_str', '')
@ -1839,8 +1759,7 @@ def load_history(unique_id, character, mode):
    if not p.exists():
        return {'internal': [], 'visible': [], 'metadata': {}}
-    with open(p, 'rb') as fh:
+    f = json.loads(open(p, 'rb').read())
        f = json.loads(fh.read())
    if 'internal' in f and 'visible' in f:
        history = f
    else:
@ -1904,17 +1823,19 @@ def generate_pfp_cache(character):
    if not cache_folder.exists():
        cache_folder.mkdir()
-    for extension in ['png', 'jpg', 'jpeg']:
+    for path in [shared.user_data_dir / 'characters' / f"{character}.{extension}" for extension in ['png', 'jpg', 'jpeg']]:
        path = shared.user_data_dir / 'characters' / f"{character}.{extension}"
        if path.exists():
            original_img = Image.open(path)
-            pfp_path = cache_folder / 'pfp_character.png'
+            # Define file paths
-            thumb_path = cache_folder / 'pfp_character_thumb.png'
+            pfp_path = Path(f'{cache_folder}/pfp_character.png')
            thumb_path = Path(f'{cache_folder}/pfp_character_thumb.png')
            # Save main picture and thumbnail
            original_img.save(pfp_path, format='PNG')
            thumb = make_thumbnail(original_img)
            thumb.save(thumb_path, format='PNG')
            # Return the path to the thumbnail, not the in-memory PIL Image object.
            return str(thumb_path)
    return None
@ -1935,13 +1856,13 @@ def load_character(character, name1, name2):
        logger.error(f"Could not find the character \"{character}\" inside {shared.user_data_dir}/characters. No character has been loaded.")
        raise ValueError
-    with open(filepath, 'r', encoding='utf-8') as fh:
+    file_contents = open(filepath, 'r', encoding='utf-8').read()
        file_contents = fh.read()
    data = json.loads(file_contents) if extension == "json" else yaml.safe_load(file_contents)
    cache_folder = Path(shared.args.disk_cache_dir)
-    for path in [cache_folder / "pfp_character.png", cache_folder / "pfp_character_thumb.png"]:
+    for path in [Path(f"{cache_folder}/pfp_character.png"), Path(f"{cache_folder}/pfp_character_thumb.png")]:
-        path.unlink(missing_ok=True)
+        if path.exists():
            path.unlink()
    picture = generate_pfp_cache(character)
@ -1997,7 +1918,9 @@ def clear_character_for_ui(state):
    # Clear the cache files
    cache_folder = Path(shared.args.disk_cache_dir)
    for cache_file in ['pfp_character.png', 'pfp_character_thumb.png']:
-        (cache_folder / cache_file).unlink(missing_ok=True)
+        cache_path = Path(f'{cache_folder}/{cache_file}')
        if cache_path.exists():
            cache_path.unlink()
    return state, state['name2'], state['context'], state['greeting'], None
@ -2092,10 +2015,11 @@ def upload_your_profile_picture(img_path):
        cache_folder.mkdir()
    if img is None:
-        (cache_folder / "pfp_me.png").unlink(missing_ok=True)
+        if Path(f"{cache_folder}/pfp_me.png").exists():
            Path(f"{cache_folder}/pfp_me.png").unlink()
    else:
        img = make_thumbnail(img)
-        img.save(cache_folder / 'pfp_me.png')
+        img.save(Path(f'{cache_folder}/pfp_me.png'))
        logger.info(f'Profile picture saved to "{cache_folder}/pfp_me.png"')
@ -2151,12 +2075,13 @@ def generate_user_pfp_cache(user):
    if not cache_folder.exists():
        cache_folder.mkdir()
-    for extension in ['png', 'jpg', 'jpeg']:
+    for path in [shared.user_data_dir / 'users' / f"{user}.{extension}" for extension in ['png', 'jpg', 'jpeg']]:
        path = shared.user_data_dir / 'users' / f"{user}.{extension}"
        if path.exists():
            original_img = Image.open(path)
-            pfp_path = cache_folder / 'pfp_me.png'
+            # Define file paths
            pfp_path = Path(f'{cache_folder}/pfp_me.png')
            # Save thumbnail
            thumb = make_thumbnail(original_img)
            thumb.save(pfp_path, format='PNG')
            logger.info(f'User profile picture cached to "{pfp_path}"')
@ -2188,7 +2113,9 @@ def load_user(user_name, name1, user_bio):
    # Clear existing user picture cache
    cache_folder = Path(shared.args.disk_cache_dir)
-    (cache_folder / "pfp_me.png").unlink(missing_ok=True)
+    pfp_path = Path(f"{cache_folder}/pfp_me.png")
    if pfp_path.exists():
        pfp_path.unlink()
    # Generate new picture cache
    picture = generate_user_pfp_cache(user_name)
@ -2612,13 +2539,15 @@ def handle_character_picture_change(picture_path):
    if picture is not None:
        # Save to cache
-        picture.save(cache_folder / 'pfp_character.png', format='PNG')
+        picture.save(Path(f'{cache_folder}/pfp_character.png'), format='PNG')
        thumb = make_thumbnail(picture)
-        thumb.save(cache_folder / 'pfp_character_thumb.png', format='PNG')
+        thumb.save(Path(f'{cache_folder}/pfp_character_thumb.png'), format='PNG')
    else:
        # Remove cache files when picture is cleared
        for cache_file in ['pfp_character.png', 'pfp_character_thumb.png']:
-            (cache_folder / cache_file).unlink(missing_ok=True)
+            cache_path = Path(f'{cache_folder}/{cache_file}')
            if cache_path.exists():
                cache_path.unlink()
 def handle_mode_change(state):
--- a/modules/exllamav3.py
+++ b/modules/exllamav3.py
@ -423,15 +423,6 @@ class Exllamav3Model:
        if logit_bias:
            filters.append(LogitBiasFilter(self.tokenizer, logit_bias))
        # Suppress EOS tokens via logit bias so they are never sampled
        if state['ban_eos_token']:
            eos_bias = {}
            for eos_id in self.config.eos_token_id_list:
                if eos_id is not None:
                    eos_bias[str(eos_id)] = float('-inf')
            if eos_bias:
                filters.append(LogitBiasFilter(self.tokenizer, eos_bias))
        # Logprobs support (OpenAI API)
        logprobs = state.get('logprobs', 0) or 0
        return_top_tokens = logprobs if logprobs > 0 else 0
@ -489,35 +480,15 @@ class Exllamav3Model:
            return
        id_to_piece = self.tokenizer.get_id_to_piece_list(True)
        sampled_ids = result.get("token_ids")    # (batch, seq_len) - actually sampled tokens
        sampled_probs = result.get("token_probs")  # (batch, seq_len) - their probabilities
        def _piece(tid):
            s = id_to_piece[tid] if tid < len(id_to_piece) else f"<{tid}>"
            return s.replace('\u2581', ' ')
        def _logprob(prob):
            return math.log(prob) if prob > 0 else float("-inf")
        # top_k_tokens shape: (batch, seq_len, k), top_k_probs same
        for seq_idx in range(top_k_tokens.shape[1]):
            entry = {"top_logprobs": []}
            for k_idx in range(top_k_tokens.shape[2]):
                token_id = top_k_tokens[0, seq_idx, k_idx].item()
                prob = top_k_probs[0, seq_idx, k_idx].item()
-                entry["top_logprobs"].append({"token": _piece(token_id), "logprob": _logprob(prob)})
+                token_str = id_to_piece[token_id] if token_id < len(id_to_piece) else f"<{token_id}>"
-
+                logprob = math.log(prob) if prob > 0 else float("-inf")
-            # Record the actually sampled token at the entry level so
+                entry["top_logprobs"].append({"token": token_str, "logprob": logprob})
            # format_completion_logprobs uses it instead of top_logprobs[0]
            # (they differ with non-greedy sampling).
            if sampled_ids is not None:
                sid = sampled_ids[0, seq_idx].item()
                entry["token"] = _piece(sid)
                if sampled_probs is not None:
                    entry["logprob"] = _logprob(sampled_probs[0, seq_idx].item())
                else:
                    entry["logprob"] = None
            self.last_completion_probabilities.append(entry)
    def generate(self, prompt, state):
@ -527,31 +498,42 @@ class Exllamav3Model:
        return output
    def get_prompt_logits(self, input_ids):
        """Return logits for all positions via a single no-cache forward pass.
        Used by prompt logprobs computation. Returns (1, seq_len, vocab) on CPU in float32.
        """
        import torch
        input_ids_tensor = input_ids if isinstance(input_ids, torch.Tensor) else torch.tensor(input_ids, dtype=torch.long)
        input_ids_tensor = input_ids_tensor.view(1, -1).cpu()
        with torch.no_grad():
            return self.model.forward(
                input_ids=input_ids_tensor,
                params={"attn_mode": "flash_attn_nc"}
            ).cpu().float()
    def get_logits(self, token_ids, **kwargs):
        """
        Process a batch of token_ids and return the logits for the last token.
-        Uses flash_attn_nc (no cache) for correct results with recurrent models.
+        This will reset and overwrite the model's cache.
        """
        # Initialize a single params dictionary that will be updated in-place
        params = {
            "cache": self.cache,
            "reconstruct": False,
            "attn_mode": "flash_attn",
            "batch_shape": (1, self.max_tokens),
            "past_len": 0
        }
        params.update(kwargs)
        # Process prefix tokens to fill the cache and generate recurrent state
        if token_ids.shape[-1] > 1:
            prefix_ids = token_ids[:, :-1]
            # This forward call updates the 'params' dict with the recurrent state
            self.model.forward(
                input_ids=prefix_ids,
                params=params
            )
            # Update past_len for the next call
            params["past_len"] = prefix_ids.shape[-1]
        # Process the last token, now using the state-filled 'params' dict
        last_token_ids = token_ids[:, -1:]
        logits = self.model.forward(
-            input_ids=token_ids,
+            input_ids=last_token_ids,
-            params={"attn_mode": "flash_attn_nc"}
+            params=params
        )
-        return logits[:, -1:, :].float().cpu()
+        return logits.float().cpu()
    def encode(self, string, **kwargs):
        add_bos = kwargs.pop('add_bos', True)
--- a/modules/exllamav3_hf.py
+++ b/modules/exllamav3_hf.py
@ -26,9 +26,6 @@ except Exception:
 class Exllamav3HF(PreTrainedModel, GenerationMixin):
    def __init__(self, model_dir):
        hf_config = PretrainedConfig.from_pretrained(model_dir)
        # Ensure text_config is a proper object, not a dict (fixes qwen3_5_moe + transformers compat)
        if isinstance(getattr(hf_config, 'text_config', None), dict):
            hf_config.text_config = PretrainedConfig(**hf_config.text_config)
        super().__init__(hf_config)
        exl3_config = Config.from_directory(model_dir)
@ -202,11 +199,30 @@ class Exllamav3HF(PreTrainedModel, GenerationMixin):
                }
            ).to(input_ids.device).float()
        else:
-            # Labels path: single pass without cache for correct logits
+            # Labels path: use cache for cross-chunk attention.
-            logits = self.ex_model.forward(
+            tokens_to_process = seq_tensor
-                input_ids=seq_tensor.view(1, -1),
+            all_logits = None
-                params={"attn_mode": "flash_attn_nc"}
+            current_len = 0
-            ).float().cpu()
+
            for i in range(0, tokens_to_process.shape[0], max_chunk_size):
                chunk = tokens_to_process[i:i + max_chunk_size]
                chunk_logits = self.ex_model.forward(
                    input_ids=chunk.view(1, -1),
                    params={
                        "attn_mode": "flash_attn",
                        "cache": ex_cache,
                        "past_len": current_len,
                        "batch_shape": (1, self.max_tokens),
                    }
                ).float()
                current_len += chunk.shape[0]
                if all_logits is None:
                    all_logits = chunk_logits
                else:
                    all_logits = torch.cat([all_logits, chunk_logits], dim=1)
            logits = all_logits
        if is_negative:
            self.past_seq_negative = seq_tensor
--- a/modules/extensions.py
+++ b/modules/extensions.py
@ -191,19 +191,21 @@ def _apply_custom_generate_reply():
 def _apply_custom_css():
-    return ''.join(
+    all_css = ''
-        getattr(extension, 'custom_css')()
+    for extension, _ in iterator():
-        for extension, _ in iterator()
+        if hasattr(extension, 'custom_css'):
-        if hasattr(extension, 'custom_css')
+            all_css += getattr(extension, 'custom_css')()
-    )
+
    return all_css
 def _apply_custom_js():
-    return ''.join(
+    all_js = ''
-        getattr(extension, 'custom_js')()
+    for extension, _ in iterator():
-        for extension, _ in iterator()
+        if hasattr(extension, 'custom_js'):
-        if hasattr(extension, 'custom_js')
+            all_js += getattr(extension, 'custom_js')()
-    )
+
    return all_js
 def create_extensions_block():
--- a/modules/html_generator.py
+++ b/modules/html_generator.py
@ -14,13 +14,6 @@ from modules.reasoning import extract_reasoning
 from modules.sane_markdown_lists import SaneListExtension
 from modules.utils import get_available_chat_styles
 # Pre-compiled regex for protecting markdown-sensitive characters inside LaTeX.
 # Covers $$...$$, \[...\], \(...\), and inline $...$ (when content contains \\).
 _LATEX_PATTERN = re.compile(
    r'((?:^|[\r\n\s])\$\$[^`]*?\$\$)|\\\[(.*?)\\\]|\\\((.*?)\\\)|(?<!\$)\$(?!\$)([^\$\n]*\\\\[^\$\n]*?)\$(?!\$)',
    re.DOTALL
 )
 # This is to store the paths to the thumbnails of the profile pictures
 image_cache = {}
@ -192,29 +185,28 @@ def process_markdown_content(string):
    if not string:
        return ""
-    # Define unique placeholders for LaTeX characters that conflict with markdown
+    # Define unique placeholders for LaTeX asterisks and underscores
    LATEX_ASTERISK_PLACEHOLDER = "LATEXASTERISKPLACEHOLDER"
    LATEX_UNDERSCORE_PLACEHOLDER = "LATEXUNDERSCOREPLACEHOLDER"
    LATEX_PIPE_PLACEHOLDER = "LATEXPIPEPLACEHOLDER"
    def protect_latex_content(content):
        """Protect markdown-sensitive characters inside LaTeX."""
        content = content.replace('*', LATEX_ASTERISK_PLACEHOLDER)
        content = content.replace('_', LATEX_UNDERSCORE_PLACEHOLDER)
        content = content.replace('|', LATEX_PIPE_PLACEHOLDER)
        return content
    def protect_asterisks_underscores_in_latex(match):
-        """A replacer function for re.sub to protect markdown-sensitive characters in multiple LaTeX formats."""
+        """A replacer function for re.sub to protect asterisks and underscores in multiple LaTeX formats."""
        # Check which delimiter group was captured
        if match.group(1) is not None:  # Content from $$...$$
-            return protect_latex_content(match.group(1))
+            content = match.group(1)
            modified_content = content.replace('*', LATEX_ASTERISK_PLACEHOLDER)
            modified_content = modified_content.replace('_', LATEX_UNDERSCORE_PLACEHOLDER)
            return f'{modified_content}'
        elif match.group(2) is not None:  # Content from \[...\]
-            return f'\\[{protect_latex_content(match.group(2))}\\]'
+            content = match.group(2)
            modified_content = content.replace('*', LATEX_ASTERISK_PLACEHOLDER)
            modified_content = modified_content.replace('_', LATEX_UNDERSCORE_PLACEHOLDER)
            return f'\\[{modified_content}\\]'
        elif match.group(3) is not None:  # Content from \(...\)
-            return f'\\({protect_latex_content(match.group(3))}\\)'
+            content = match.group(3)
-        elif match.group(4) is not None:  # Content from $...$
+            modified_content = content.replace('*', LATEX_ASTERISK_PLACEHOLDER)
-            return f'${protect_latex_content(match.group(4).strip())}$'
+            modified_content = modified_content.replace('_', LATEX_UNDERSCORE_PLACEHOLDER)
            return f'\\({modified_content}\\)'
        return match.group(0)  # Fallback
@ -248,7 +240,9 @@ def process_markdown_content(string):
    string = re.sub(r"(.)```", r"\1\n```", string)
    # Protect asterisks and underscores within all LaTeX blocks before markdown conversion
-    string = _LATEX_PATTERN.sub(protect_asterisks_underscores_in_latex, string)
+    latex_pattern = re.compile(r'((?:^|[\r\n\s])\$\$[^`]*?\$\$)|\\\[(.*?)\\\]|\\\((.*?)\\\)',
                               re.DOTALL)
    string = latex_pattern.sub(protect_asterisks_underscores_in_latex, string)
    result = ''
    is_code = False
@ -312,7 +306,6 @@ def process_markdown_content(string):
    # Restore the LaTeX asterisks and underscores after markdown conversion
    html_output = html_output.replace(LATEX_ASTERISK_PLACEHOLDER, '*')
    html_output = html_output.replace(LATEX_UNDERSCORE_PLACEHOLDER, '_')
    html_output = html_output.replace(LATEX_PIPE_PLACEHOLDER, '|')
    # Remove extra newlines before </code>
    html_output = re.sub(r'\s*</code>', '</code>', html_output)
--- a/modules/image_models.py
+++ b/modules/image_models.py
@ -10,49 +10,72 @@ def get_quantization_config(quant_method):
    Get the appropriate quantization config based on the selected method.
    Applies quantization to both the transformer and the text_encoder.
    """
    if quant_method == 'none' or not quant_method:
        return None
    import torch
    # Import BitsAndBytesConfig from BOTH libraries to be safe
    from diffusers import BitsAndBytesConfig as DiffusersBnBConfig
    from diffusers import TorchAoConfig
    from diffusers.quantizers import PipelineQuantizationConfig
    from transformers import BitsAndBytesConfig as TransformersBnBConfig
-    torchao_methods = {
+    if quant_method == 'none' or not quant_method:
-        'torchao-int8wo': 'int8wo',
+        return None
        'torchao-fp4': 'fp4_e2m1',
        'torchao-float8wo': 'float8wo',
    }
-    if quant_method == 'bnb-8bit':
+    # Bitsandbytes 8-bit quantization
    elif quant_method == 'bnb-8bit':
        return PipelineQuantizationConfig(
            quant_mapping={
-                "transformer": DiffusersBnBConfig(load_in_8bit=True),
+                "transformer": DiffusersBnBConfig(
-                "text_encoder": TransformersBnBConfig(load_in_8bit=True)
+                    load_in_8bit=True
                ),
                "text_encoder": TransformersBnBConfig(
                    load_in_8bit=True
                )
            }
        )
    # Bitsandbytes 4-bit quantization
    elif quant_method == 'bnb-4bit':
        bnb_4bit_kwargs = dict(
            load_in_4bit=True,
            bnb_4bit_quant_type="nf4",
            bnb_4bit_compute_dtype=torch.bfloat16,
            bnb_4bit_use_double_quant=True
        )
        return PipelineQuantizationConfig(
            quant_mapping={
-                "transformer": DiffusersBnBConfig(**bnb_4bit_kwargs),
+                "transformer": DiffusersBnBConfig(
-                "text_encoder": TransformersBnBConfig(**bnb_4bit_kwargs)
+                    load_in_4bit=True,
                    bnb_4bit_quant_type="nf4",
                    bnb_4bit_compute_dtype=torch.bfloat16,
                    bnb_4bit_use_double_quant=True
                ),
                "text_encoder": TransformersBnBConfig(
                    load_in_4bit=True,
                    bnb_4bit_quant_type="nf4",
                    bnb_4bit_compute_dtype=torch.bfloat16,
                    bnb_4bit_use_double_quant=True
                )
            }
        )
-    elif quant_method in torchao_methods:
+    # torchao int8 weight-only
-        ao_type = torchao_methods[quant_method]
+    elif quant_method == 'torchao-int8wo':
        return PipelineQuantizationConfig(
            quant_mapping={
-                "transformer": TorchAoConfig(ao_type),
+                "transformer": TorchAoConfig("int8wo"),
-                "text_encoder": TorchAoConfig(ao_type)
+                "text_encoder": TorchAoConfig("int8wo")
            }
        )
    # torchao fp4 (e2m1)
    elif quant_method == 'torchao-fp4':
        return PipelineQuantizationConfig(
            quant_mapping={
                "transformer": TorchAoConfig("fp4_e2m1"),
                "text_encoder": TorchAoConfig("fp4_e2m1")
            }
        )
    # torchao float8 weight-only
    elif quant_method == 'torchao-float8wo':
        return PipelineQuantizationConfig(
            quant_mapping={
                "transformer": TorchAoConfig("float8wo"),
                "text_encoder": TorchAoConfig("float8wo")
            }
        )
@ -129,7 +152,7 @@ def load_image_model(model_name, dtype='bfloat16', attn_backend='sdpa', cpu_offl
        modules = ["transformer", "unet"]
-        # Set attention backend (diffusers defaults to native/SDPA)
+        # Set attention backend
        if attn_backend == 'flash_attention_2':
            for name in modules:
                mod = getattr(pipe, name, None)
--- a/modules/llama_cpp_server.py
+++ b/modules/llama_cpp_server.py
@ -11,6 +11,7 @@ import time
 from pathlib import Path
 from typing import Any, List
 import llama_cpp_binaries
 import requests
 from modules import shared
@ -310,45 +311,8 @@ class LlamaServer:
        else:
            raise Exception(f"Unexpected response format: 'completion_probabilities' not found in {result}")
    def get_prompt_logprob_entries(self, token_ids, n_probs=5, prompt=""):
        """Get logprob entries for prompt tokens via a single n_predict=0 request.
        Requires llama.cpp server with prompt_logprobs support.
        Returns entries in the standard format for format_completion_logprobs().
        """
        token_ids_list = token_ids.tolist() if hasattr(token_ids, 'tolist') else list(token_ids)
        url = f"http://127.0.0.1:{self.port}/completion"
        payload = {
            "prompt": token_ids_list,
            "n_predict": 0,
            "n_probs": n_probs,
            "prompt_logprobs": True,
            "stream": False,
            "cache_prompt": False,
        }
        response = self.session.post(url, json=payload)
        result = response.json()
        prompt_probs = result.get("prompt_probabilities", [])
        if not prompt_probs:
            return []
        # Null first token (no conditioning context); use empty string for BOS
        # or tokens that don't appear at the start of the prompt text.
        first_token_str = self.decode([token_ids_list[0]])
        if self.bos_token and first_token_str == self.bos_token:
            first_token_str = ""
        elif not prompt.startswith(first_token_str):
            first_token_str = ""
        entries = [{"token": first_token_str, "null_logprob": True}]
        entries.extend(prompt_probs)
        return entries
    def _get_vocabulary_size(self):
-        """Get and store the model's vocabulary size."""
+        """Get and store the model's maximum context length."""
        url = f"http://127.0.0.1:{self.port}/v1/models"
        response = self.session.get(url).json()
@ -373,7 +337,6 @@ class LlamaServer:
        """Check if a port is available for use."""
        with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
            try:
                s.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
                s.bind(('', port))
                return True
            except OSError:
@ -394,16 +357,7 @@ class LlamaServer:
        """Start the llama.cpp server and wait until it's ready."""
        # Determine the server path
        if self.server_path is None:
-            if shared.args.ik:
+            self.server_path = llama_cpp_binaries.get_binary_path()
                try:
                    import ik_llama_cpp_binaries
                except ImportError:
                    raise ImportError("--ik requires the ik_llama_cpp_binaries package. Install it with: pip install <ik_llama_cpp_binaries wheel URL>")
                self.server_path = ik_llama_cpp_binaries.get_binary_path()
            else:
                import llama_cpp_binaries
                self.server_path = llama_cpp_binaries.get_binary_path()
        # Build the command
        cmd = [
@ -516,10 +470,6 @@ class LlamaServer:
                        else:
                            cmd.append(f"--{flag_item}")
        # Patch flags for ik_llama.cpp compatibility
        if shared.args.ik:
            cmd = _patch_cmd_for_ik(cmd)
        env = os.environ.copy()
        if os.name == 'posix':
            current_path = env.get('LD_LIBRARY_PATH', '')
@ -657,49 +607,3 @@ def filter_stderr_with_progress(process_stderr):
            process_stderr.close()
        except Exception:
            pass
 def _patch_cmd_for_ik(cmd):
    """
    Rewrite upstream llama.cpp flags to ik_llama.cpp equivalents:
      --no-webui           → --webui none
      --fit off            → (removed)
      --fit on / --fit-ctx → --fit (bare flag)
      --fit-target         → --fit-margin
      --cache-reuse        → (removed, unsupported)
      --swa-full           → (removed, unsupported)
    """
    # Add Hadamard KV cache rotation when using quantized cache types.
    # This significantly improves quantized cache quality (especially q4_0)
    # and is a no-op for MLA models like DeepSeek.
    if shared.args.cache_type in ("q8_0", "q4_0"):
        cmd += ["-khad", "-vhad"]
    patched = []
    i = 0
    while i < len(cmd):
        arg = cmd[i]
        if arg == "--no-webui":
            patched += ["--webui", "none"]
        elif arg == "--fit" and i + 1 < len(cmd) and cmd[i + 1] in ("on", "off"):
            val = cmd[i + 1]
            i += 1
            if val == "on":
                patched.append("--fit")
            # "off" → drop entirely
        elif arg == "--fit-ctx":
            patched.append("--fit")
            i += 1  # skip the value
        elif arg == "--fit-target":
            patched.append("--fit-margin")
        elif arg == "--cache-reuse":
            i += 1  # skip the value
        elif arg == "--swa-full":
            pass  # bare flag, just drop it
        else:
            patched.append(arg)
        i += 1
    return patched
--- a/modules/loaders.py
+++ b/modules/loaders.py
@ -20,7 +20,6 @@ loaders_and_params = OrderedDict({
        'no_mmap',
        'mlock',
        'numa',
        'ik',
        'parallel',
        'model_draft',
        'draft_max',
@ -291,21 +290,16 @@ def blacklist_samplers(loader, dynamic_temperature):
@functools.cache
 def get_all_params():
    from modules import shared
    all_params = set()
    for k in loaders_and_params:
        for el in loaders_and_params[k]:
            all_params.add(el)
    if shared.args.portable:
        all_params.discard('ik')
    return sorted(all_params)
@functools.cache
 def list_model_elements():
-    elements = [
+    return [
        'filter_by_loader',
        'loader',
        'cpu_memory',
@ -353,12 +347,6 @@ def list_model_elements():
        'mmproj',
    ]
    from modules import shared
    if not shared.args.portable:
        elements.append('ik')
    return elements
 def make_loader_params_visible(loader):
    import gradio as gr
--- a/modules/logits.py
+++ b/modules/logits.py
@ -4,6 +4,7 @@ import numpy as np
 from modules import models, shared
 from modules.logging_colors import logger
 from modules.models import load_model
 from modules.text_generation import generate_reply
 from modules.utils import check_model_loaded
@ -11,7 +12,8 @@ global_scores = None
 def get_next_logits(*args, **kwargs):
-    models.load_model_if_idle_unloaded()
+    if shared.args.idle_timeout > 0 and shared.model is None and shared.model_name not in [None, 'None']:
        shared.model, shared.tokenizer = load_model(shared.model_name)
    needs_lock = not args[2]  # use_samplers
    if needs_lock:
--- a/modules/models.py
+++ b/modules/models.py
@ -1,5 +1,4 @@
 import sys
 import threading
 import time
 import modules.shared as shared
@ -8,15 +7,6 @@ from modules.models_settings import get_model_metadata
 from modules.utils import resolve_model_path
 last_generation_time = time.time()
 active_generation_count = 0
 _generation_count_lock = threading.Lock()
 def load_model_if_idle_unloaded():
    global last_generation_time
    if shared.args.idle_timeout > 0 and shared.model is None and shared.model_name not in [None, 'None']:
        shared.model, shared.tokenizer = load_model(shared.model_name)
        last_generation_time = time.time()
 def load_model(model_name, loader=None):
@ -76,7 +66,8 @@ def load_model(model_name, loader=None):
    logger.info(f"Loaded \"{model_name}\" in {(time.time()-t0):.2f} seconds.")
    logger.info(f"LOADER: \"{loader}\"")
-    logger.info(f"CONTEXT LENGTH: {shared.settings['truncation_length']}")
+    logger.info(f"TRUNCATION LENGTH: {shared.settings['truncation_length']}")
    logger.info(f"INSTRUCTION TEMPLATE: \"{metadata['instruction_template']}\"")
    return model, tokenizer
@ -168,10 +159,7 @@ def unload_model_if_idle():
    while True:
        shared.generation_lock.acquire()
        try:
-            with _generation_count_lock:
+            if time.time() - last_generation_time > shared.args.idle_timeout * 60:
                is_active = active_generation_count > 0
            if not is_active and time.time() - last_generation_time > shared.args.idle_timeout * 60:
                if shared.model is not None:
                    logger.info("Unloading the model for inactivity.")
                    unload_model(keep_model_name=True)
--- a/modules/models_settings.py
+++ b/modules/models_settings.py
@ -23,9 +23,14 @@ def get_fallback_settings():
 def get_model_metadata(model):
    model_path = resolve_model_path(model)
    model_settings = {}
-    # Fallback settings
+    # Get settings from user_data/models/config.yaml and user_data/models/config-user.yaml
-    model_settings = get_fallback_settings()
+    settings = shared.model_config
    for pat in settings:
        if re.match(pat.lower(), Path(model).name.lower()):
            for k in settings[pat]:
                model_settings[k] = settings[pat][k]
    path = model_path / 'config.json'
    if path.exists():
@ -400,19 +405,14 @@ def load_instruction_template(template):
    if template == 'None':
        return ''
-    for name in (template, 'Alpaca'):
+    for filepath in [shared.user_data_dir / 'instruction-templates' / f'{template}.yaml', shared.user_data_dir / 'instruction-templates' / 'Alpaca.yaml']:
-        path = shared.user_data_dir / 'instruction-templates' / f'{name}.yaml'
+        if filepath.exists():
-        try:
+            break
            with open(path, 'r', encoding='utf-8') as f:
                file_contents = f.read()
        except FileNotFoundError:
            if name == template:
                logger.warning(f"Instruction template '{template}' not found, falling back to Alpaca")
            continue
        break
    else:
        return ''
    with open(filepath, 'r', encoding='utf-8') as f:
        file_contents = f.read()
    data = yaml.safe_load(file_contents)
    if 'instruction_template' in data:
        return data['instruction_template']
--- a/modules/prompts.py
+++ b/modules/prompts.py
@ -1,7 +1,6 @@
 from pathlib import Path
 from modules import shared, utils
 from modules.utils import sanitize_filename
 from modules.text_generation import get_encoded_length
@ -19,7 +18,6 @@ def load_prompt(fname):
        return initial_content
    fname = sanitize_filename(fname)
    file_path = shared.user_data_dir / 'logs' / 'notebook' / f'{fname}.txt'
    if file_path.exists():
        with open(file_path, 'r', encoding='utf-8') as f:
--- a/modules/reasoning.py
+++ b/modules/reasoning.py
@ -7,7 +7,6 @@ THINKING_FORMATS = [
    ('<|channel|>analysis<|message|>', '<|end|>', '<|channel|>final<|message|>'),
    ('<|channel|>commentary<|message|>', '<|end|>', '<|channel|>final<|message|>'),
    ('<seed:think>', '</seed:think>', None),
    ('<|channel>thought', '<channel|>', None),  # Gemma 4
    ('<|think|>', '<|end|>', '<|content|>'),  # Solar Open
    # ('Thinking Process:', '</think>', None),  # Qwen3.5 verbose thinking outside tags -- removed: too prone to false positives in streaming
    (None, '</think>', None),  # End-only variant (e.g., Qwen3-next)
@ -73,16 +72,9 @@ def extract_reasoning(text, html_escaped=False):
                if content_pos != -1:
                    content_start = content_pos + len(content_esc)
                else:
-                    # Content tag not present yet.  In GPT-OSS the region
+                    # Content tag not present — fall back to content after
-                    # between <|end|> and the content tag contains internal
+                    # end_tag (e.g. GPT-OSS tool calls skip the final channel).
-                    # markup (<|start|>assistant…) that must not be shown.
+                    content_start = end_pos + len(end_esc)
                    # Suppress it to prevent tag leaks during streaming.
                    remainder = text[end_pos + len(end_esc):].lstrip()
                    framing_token = esc('<|start|>')
                    if not remainder or remainder.startswith(framing_token) or framing_token.startswith(remainder):
                        content_start = len(text)
                    else:
                        content_start = end_pos + len(end_esc)
            else:
                content_start = end_pos + len(end_esc)
--- a/modules/shared.py
+++ b/modules/shared.py
@ -110,7 +110,6 @@ group.add_argument('--numa', action='store_true', help='Activate NUMA task alloc
 group.add_argument('--parallel', type=int, default=1, help='Number of parallel request slots. The context size is divided equally among slots. For example, to have 4 slots with 8192 context each, set ctx_size to 32768.')
 group.add_argument('--fit-target', type=str, default='512', help='Target VRAM margin per device for auto GPU layers, comma-separated list of values in MiB. A single value is broadcast across all devices.')
 group.add_argument('--extra-flags', type=str, default=None, help='Extra flags to pass to llama-server. Example: "--jinja --rpc 192.168.1.100:50052"')
 group.add_argument('--ik', action='store_true', help='Use ik_llama.cpp instead of upstream llama.cpp. Requires the ik_llama_cpp_binaries package to be installed.')
 # Transformers/Accelerate
 group = parser.add_argument_group('Transformers/Accelerate')
@ -259,7 +258,6 @@ settings = {
    'enable_web_search': False,
    'web_search_pages': 3,
    'selected_tools': [],
    'mcp_servers': '',
    'prompt-notebook': '',
    'preset': 'Top-P' if (user_data_dir / 'presets/Top-P.yaml').exists() else None,
    'max_new_tokens': 512,
@ -364,7 +362,7 @@ settings = {
    'image_llm_variations_prompt': 'Write a variation of the image generation prompt above. Consider the intent of the user with that prompt and write something that will likely please them, with added details. Output only the new prompt. Do not add any explanations, prefixes, or additional text.',
    'image_model_menu': 'None',
    'image_dtype': 'bfloat16',
-    'image_attn_backend': 'sdpa',
+    'image_attn_backend': 'flash_attention_2',
    'image_cpu_offload': False,
    'image_compile': False,
    'image_quant': 'none',
@ -456,7 +454,17 @@ def load_user_config():
 args.loader = fix_loader_name(args.loader)
 # Load model-specific settings
 p = Path(f'{args.model_dir}/config.yaml')
 if p.exists():
    model_config = yaml.safe_load(open(p, 'r').read())
 else:
    model_config = {}
 del p
 # Load custom model-specific settings
 user_config = load_user_config()
 model_config = OrderedDict(model_config)
 user_config = OrderedDict(user_config)
--- a/modules/text_generation.py
+++ b/modules/text_generation.py
@ -17,7 +17,9 @@ from modules.utils import check_model_loaded
 def generate_reply(*args, **kwargs):
-    models.load_model_if_idle_unloaded()
+    if shared.args.idle_timeout > 0 and shared.model is None and shared.model_name not in [None, 'None']:
        from modules.models import load_model
        shared.model, shared.tokenizer = load_model(shared.model_name)
    state = args[1] if len(args) > 1 else kwargs.get('state', {})
    use_parallel = (
@ -29,16 +31,10 @@ def generate_reply(*args, **kwargs):
    if not use_parallel:
        shared.generation_lock.acquire()
    with models._generation_count_lock:
        models.active_generation_count += 1
    try:
        for result in _generate_reply(*args, **kwargs):
            yield result
    finally:
        with models._generation_count_lock:
            models.active_generation_count -= 1
        models.last_generation_time = time.time()
        if not use_parallel:
            shared.generation_lock.release()
@ -130,9 +126,7 @@ def _generate_reply(question, state, stopping_strings=None, is_chat=False, escap
 def encode(prompt, add_special_tokens=True, add_bos_token=True, truncation_length=None):
    if shared.tokenizer is None:
-        models.load_model_if_idle_unloaded()
+        raise ValueError('No tokenizer is loaded')
        if shared.tokenizer is None:
            raise ValueError('No tokenizer is loaded')
    # llama.cpp case
    if shared.model.__class__.__name__ == 'LlamaServer':
@ -182,9 +176,7 @@ def encode(prompt, add_special_tokens=True, add_bos_token=True, truncation_lengt
 def decode(output_ids, skip_special_tokens=True):
    if shared.tokenizer is None:
-        models.load_model_if_idle_unloaded()
+        raise ValueError('No tokenizer is loaded')
        if shared.tokenizer is None:
            raise ValueError('No tokenizer is loaded')
    return shared.tokenizer.decode(output_ids, skip_special_tokens=skip_special_tokens)
--- a/modules/tool_parsing.py
+++ b/modules/tool_parsing.py
@ -27,11 +27,10 @@ TOOL_CALL_OPENING_MARKERS = [
    '[TOOL_CALLS]',
    'to=functions.',
    '<|channel|>commentary',
    '<|tool_call>call:',
 ]
-def streaming_tool_buffer_check(text, markers=None, tool_names=None, check_bare_names=False, partial_match=True):
+def streaming_tool_buffer_check(text, markers=None, tool_names=None, check_bare_names=False):
    '''
    Check whether streaming output should be withheld because it may
    contain tool-call markup.
@ -43,10 +42,6 @@ def streaming_tool_buffer_check(text, markers=None, tool_names=None, check_bare_
        tool_names: List of tool function names.
        check_bare_names: Whether to do partial-prefix matching on tool
                          names (for models with unknown template format).
        partial_match: Whether to check partial prefixes of markers/names.
                       Set to False for end-of-generation checks where a
                       partial prefix is just normal text, not an incomplete
                       tool call.
    '''
    # Strip thinking blocks so tool-call syntax inside <think> doesn't
    # trigger false positives.
@ -64,9 +59,6 @@ def streaming_tool_buffer_check(text, markers=None, tool_names=None, check_bare_
            if name + '{' in text or name + ' {' in text:
                return True
    if not partial_match:
        return False
    # Partial-prefix matching: only for template-specific markers.
    for marker in (markers if markers is not None else TOOL_CALL_OPENING_MARKERS):
        for prefix_len in range(min(len(marker) - 1, len(text)), 0, -1):
@ -408,78 +400,6 @@ def _parse_glm_tool_calls(answer: str, tool_names: list[str]):
    return matches, start_pos
 def _extract_gemma4_balanced(text, start):
    """Extract balanced braces from Gemma 4 format, using <|"|> as string delimiters."""
    if start >= len(text) or text[start] != '{':
        return None
    depth = 0
    in_string = False
    quote_token = '<|"|>'
    quote_len = len(quote_token)
    i = start
    while i < len(text):
        if text[i:i + quote_len] == quote_token:
            in_string = not in_string
            i += quote_len
            continue
        if in_string:
            i += 1
            continue
        c = text[i]
        if c == '{':
            depth += 1
        elif c == '}':
            depth -= 1
            if depth == 0:
                return text[start:i + 1]
        i += 1
    return None
 def _parse_gemma4_tool_calls(answer: str, tool_names: list[str]):
    """Parse Gemma 4-style tool calls.
    Format:
        <|tool_call>call:func_name{key:<|"|>value<|"|>,...}<tool_call|>
    Values use <|"|> tokens instead of standard JSON quotes, and keys are
    bare identifiers.
    """
    matches = []
    start_pos = None
    for m in re.finditer(r'<\|tool_call>call:([^\s{]+)\s*', answer):
        func_name = m.group(1).strip()
        if func_name not in tool_names:
            continue
        brace_start = m.end()
        if brace_start >= len(answer) or answer[brace_start] != '{':
            continue
        content = _extract_gemma4_balanced(answer, brace_start)
        if content is None:
            continue
        # Convert to JSON: split on <|"|> tokens so that key quoting
        # only applies outside string values (even-indexed parts),
        # then rejoin with real quotes.
        parts = content.split('<|"|>')
        for idx in range(0, len(parts), 2):
            parts[idx] = re.sub(r'(^|[{,\[])\s*(\w+)\s*:', r'\1"\2":', parts[idx])
        json_str = '"'.join(parts)
        try:
            arguments = json.loads(json_str)
            if start_pos is None:
                start_pos = m.start()
            matches.append(_make_tool_call(func_name, arguments))
        except (json.JSONDecodeError, ValueError):
            pass
    return matches, start_pos
 def _parse_pythonic_tool_calls(answer: str, tool_names: list[str]):
    """Parse pythonic-style tool calls used by Llama 4 and similar models.
@ -552,11 +472,6 @@ TOOL_CALL_FORMATS = [
        'parser': _parse_channel_tool_calls,
        'markers': ['to=functions.', '<|channel|>commentary'],
    },
    {
        'template_hints': ['<|tool_call>call:'],
        'parser': _parse_gemma4_tool_calls,
        'markers': ['<|tool_call>call:'],
    },
    {
        'template_hints': ['minimax:tool_call'],
        'parser': _parse_minimax_tool_calls,
@ -589,7 +504,6 @@ ALL_PARSERS = [
    _parse_deep_seek_tool_calls,
    _parse_kimi_tool_calls,
    _parse_channel_tool_calls,
    _parse_gemma4_tool_calls,
    _parse_minimax_tool_calls,
    _parse_glm_tool_calls,
    _parse_xml_param_tool_calls,
@ -638,15 +552,9 @@ def parse_tool_call(answer: str, tool_names: list[str], return_prefix: bool = Fa
    # Strip thinking blocks so tool-call syntax inside <think> is ignored.
    original_answer = answer
    _, answer = extract_reasoning(answer)
-    # Reasoning extraction returns empty content when GPT-OSS internal
+    # Offset between original and stripped text, used to map start_pos
-    # markup (<|start|>assistant…) follows the thinking block without a
+    # back to the original string when returning a prefix.
-    # content tag.  Fall back to the full text so tool-call markers can
+    reasoning_offset = len(original_answer) - len(answer)
    # be found.
    if not answer.strip():
        answer = original_answer
        reasoning_offset = 0
    else:
        reasoning_offset = len(original_answer) - len(answer)
    matches = []
    start_pos = None
@ -712,8 +620,6 @@ def parse_tool_call(answer: str, tool_names: list[str], return_prefix: bool = Fa
                if not isinstance(candidates, list):
                    candidates = [candidates]
                for candidate_dict in candidates:
                    if not isinstance(candidate_dict, dict):
                        continue
                    checked_candidate = check_and_sanitize_tool_call_candidate(candidate_dict, tool_names)
                    if checked_candidate is not None:
                        matches.append(checked_candidate)
--- a/modules/tool_use.py
+++ b/modules/tool_use.py
@ -1,4 +1,3 @@
 import asyncio
 import importlib.util
 import json
@ -56,119 +55,6 @@ def load_tools(selected_names):
    return tool_defs, executors
 def _parse_mcp_servers(servers_str):
    """Parse MCP servers textbox: one server per line, format 'url' or 'url,Header: value,Header2: value2'."""
    servers = []
    for line in servers_str.strip().splitlines():
        line = line.strip()
        if not line:
            continue
        parts = line.split(',')
        url = parts[0].strip()
        headers = {}
        for part in parts[1:]:
            part = part.strip()
            if ':' in part:
                key, val = part.split(':', 1)
                headers[key.strip()] = val.strip()
        servers.append((url, headers))
    return servers
 def _mcp_tool_to_openai(tool):
    """Convert an MCP Tool object to OpenAI-format tool dict."""
    return {
        "type": "function",
        "function": {
            "name": tool.name,
            "description": tool.description or "",
            "parameters": tool.inputSchema or {"type": "object", "properties": {}}
        }
    }
 async def _mcp_session(url, headers, callback):
    """Open an MCP session and pass it to the callback."""
    from mcp.client.streamable_http import streamablehttp_client
    from mcp import ClientSession
    async with streamablehttp_client(url, headers=headers or None) as (read_stream, write_stream, _):
        async with ClientSession(read_stream, write_stream) as session:
            await session.initialize()
            return await callback(session)
 def _make_mcp_executor(name, url, headers):
    def executor(arguments):
        return asyncio.run(_call_mcp_tool(name, arguments, url, headers))
    return executor
 async def _connect_mcp_server(url, headers):
    """Connect to one MCP server and return (tool_defs, executors)."""
    async def _discover(session):
        result = await session.list_tools()
        tool_defs = []
        executors = {}
        for tool in result.tools:
            tool_defs.append(_mcp_tool_to_openai(tool))
            executors[tool.name] = _make_mcp_executor(tool.name, url, headers)
        return tool_defs, executors
    return await _mcp_session(url, headers, _discover)
 async def _call_mcp_tool(name, arguments, url, headers):
    """Connect to an MCP server and call a single tool."""
    async def _invoke(session):
        result = await session.call_tool(name, arguments)
        parts = []
        for content in result.content:
            if hasattr(content, 'text'):
                parts.append(content.text)
            else:
                parts.append(str(content))
        return '\n'.join(parts) if parts else ''
    return await _mcp_session(url, headers, _invoke)
 async def _connect_all_mcp_servers(servers):
    """Connect to all MCP servers concurrently."""
    results = await asyncio.gather(
        *(_connect_mcp_server(url, headers) for url, headers in servers),
        return_exceptions=True
    )
    all_defs = []
    all_executors = {}
    for (url, _), result in zip(servers, results):
        if isinstance(result, Exception):
            logger.exception(f'Failed to connect to MCP server "{url}"', exc_info=result)
            continue
        defs, execs = result
        for td, (fn, ex) in zip(defs, execs.items()):
            if fn in all_executors:
                logger.warning(f'MCP tool "{fn}" from {url} conflicts with an already loaded tool. Skipping.')
                continue
            all_defs.append(td)
            all_executors[fn] = ex
    return all_defs, all_executors
 def load_mcp_tools(servers_str):
    """
    Parse MCP servers string and discover tools from each server.
    Returns (tool_defs, executors) in the same format as load_tools.
    """
    servers = _parse_mcp_servers(servers_str)
    if not servers:
        return [], {}
    return asyncio.run(_connect_all_mcp_servers(servers))
 def execute_tool(func_name, arguments, executors):
    """Execute a tool by function name. Returns result as a JSON string."""
    fn = executors.get(func_name)
--- a/modules/training.py
+++ b/modules/training.py
@ -52,7 +52,7 @@ def create_ui():
                        with gr.Column():
                            always_override = gr.Checkbox(label='Override Existing Files', value=False, info='If the name is the same, checking will replace the existing file, and unchecking will load and continue from it (the rank must be the same).', elem_classes=['no-background'])
-                    with gr.Accordion(label='Target Modules', open=False):
+                    with gr.Accordion(label='Target Modules', open=False, elem_classes='tgw-accordion'):
                        gr.Markdown("Selects which modules to target in training. Targeting more modules is closer to a full fine-tune at the cost of increased VRAM and adapter size.")
                        all_linear = gr.Checkbox(label='Target all linear layers', value=True, info='Targets every nn.Linear layer except lm_head. Works for any model architecture. When checked, the individual module checkboxes below are ignored.', elem_classes=['no-background'])
                        with gr.Row():
@ -87,7 +87,7 @@ def create_ui():
                            with gr.Row():
                                lr_scheduler_type = gr.Dropdown(label='LR Scheduler', value='cosine', choices=['linear', 'constant', 'constant_with_warmup', 'cosine', 'cosine_with_restarts', 'polynomial', 'inverse_sqrt'], info='Learning rate scheduler - defines how the learning rate changes over time. "Constant" means never change, "linear" means to go in a straight line from the learning rate down to 0, cosine follows a curve, etc.', elem_classes=['slim-dropdown'])
-                    with gr.Accordion(label='Advanced Options', open=False):
+                    with gr.Accordion(label='Advanced Options', open=False, elem_classes='tgw-accordion'):
                        with gr.Row():
                            with gr.Column():
                                optimizer = gr.Dropdown(label='Optimizer', value='adamw_torch', choices=['adamw_hf', 'adamw_torch', 'adamw_torch_fused', 'adamw_torch_xla', 'adamw_apex_fused', 'adafactor', 'adamw_bnb_8bit', 'adamw_anyprecision', 'sgd', 'adagrad'], info='Optimizer algorithm. adamw_torch is the standard choice. adamw_bnb_8bit uses less VRAM. adafactor is memory-efficient for large models.', elem_classes=['slim-dropdown'])
--- a/modules/transformers_loader.py
+++ b/modules/transformers_loader.py
@ -109,6 +109,7 @@ def load_model_HF(model_name):
    params = {
        'low_cpu_mem_usage': True,
        'attn_implementation': shared.args.attn_implementation,
        'torch_dtype': torch.bfloat16 if shared.args.bf16 else torch.float16,
    }
    if shared.original_args.trust_remote_code:
@ -119,17 +120,6 @@ def load_model_HF(model_name):
    config = AutoConfig.from_pretrained(path_to_model, trust_remote_code=shared.original_args.trust_remote_code)
    # Determine torch_dtype: respect --bf16 flag, otherwise autodetect
    # from model config, but never allow float32.
    if shared.args.bf16:
        params['torch_dtype'] = torch.bfloat16
    else:
        dtype = getattr(config, 'torch_dtype', None) or getattr(getattr(config, 'text_config', None), 'torch_dtype', None)
        if dtype in (torch.float16, torch.bfloat16):
            params['torch_dtype'] = dtype
        else:
            params['torch_dtype'] = torch.float16
    if 'chatglm' in model_name.lower():
        LoaderClass = AutoModel
    else:
--- a/modules/ui.py
+++ b/modules/ui.py
@ -75,7 +75,7 @@ if not shared.args.old_colors:
        background_fill_primary_dark='var(--darker-gray, #1C1C1D)',
        body_background_fill="white",
        block_background_fill="transparent",
-        body_text_color='#1a1a1a',
+        body_text_color='rgb(64, 64, 64)',
        button_secondary_background_fill="white",
        button_secondary_border_color="var(--border-color-primary)",
        block_title_text_color='*body_text_color',
@ -209,7 +209,6 @@ def list_interface_input_elements():
        'textbox',
        'start_with',
        'selected_tools',
        'mcp_servers',
        'mode',
        'chat_style',
        'chat-instruct_command',
@ -435,7 +434,6 @@ def setup_auto_save():
        'custom_system_message',
        'chat_template_str',
        'selected_tools',
        'mcp_servers',
        # Parameters tab (ui_parameters.py) - Generation parameters
        'preset_menu',
--- a/modules/ui_chat.py
+++ b/modules/ui_chat.py
@ -52,7 +52,7 @@ def create_ui():
                shared.gradio['html_display'] = gr.HTML(value=chat_html_wrapper({'internal': [], 'visible': [], 'metadata': {}}, '', '', 'chat', 'cai-chat', '')['html'], visible=True)
                with gr.Row(elem_id="chat-input-row"):
                    with gr.Column(scale=1, elem_id='gr-hover-container'):
-                        gr.HTML(value='<div class="hover-element" onclick="void(0)"><span id="hover-element-button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><line x1="4" y1="6" x2="20" y2="6"></line><line x1="4" y1="12" x2="20" y2="12"></line><line x1="4" y1="18" x2="20" y2="18"></line></svg></span><div class="hover-menu" id="hover-menu"></div></div>', elem_id='gr-hover')
+                        gr.HTML(value='<div class="hover-element" onclick="void(0)"><span style="width: 100px; display: block" id="hover-element-button">&#9776;</span><div class="hover-menu" id="hover-menu"></div>', elem_id='gr-hover')
                    with gr.Column(scale=10, elem_id='chat-input-container'):
                        shared.gradio['textbox'] = gr.MultimodalTextbox(label='', placeholder='Send a message', file_types=['text', '.pdf', 'image'], file_count="multiple", elem_id='chat-input', elem_classes=['add_scrollbar'])
@ -82,7 +82,7 @@ def create_ui():
                gr.HTML("<div class='sidebar-vertical-separator'></div>")
                shared.gradio['reasoning_effort'] = gr.Dropdown(value=shared.settings['reasoning_effort'], choices=['low', 'medium', 'high'], label='Reasoning effort', info='Used by GPT-OSS.')
-                shared.gradio['enable_thinking'] = gr.Checkbox(value=shared.settings['enable_thinking'], label='Enable thinking', info='For models with thinking support.')
+                shared.gradio['enable_thinking'] = gr.Checkbox(value=shared.settings['enable_thinking'], label='Enable thinking', info='Used by Seed-OSS and pre-2507 Qwen3.')
                gr.HTML("<div class='sidebar-vertical-separator'></div>")
@ -105,9 +105,6 @@ def create_ui():
                shared.gradio['selected_tools'].change(fn=sync_web_tools, inputs=[shared.gradio['selected_tools']], outputs=[shared.gradio['selected_tools']], show_progress=False)
                with gr.Accordion('MCP servers', open=False):
                    shared.gradio['mcp_servers'] = gr.Textbox(value=shared.settings.get('mcp_servers', ''), lines=3, max_lines=3, label='', info='One url per line. For headers, write url,Header: value,Header2: value2', elem_classes=['add_scrollbar'])
                gr.HTML("<div class='sidebar-vertical-separator'></div>")
                with gr.Row():
--- a/modules/ui_default.py
+++ b/modules/ui_default.py
@ -10,7 +10,7 @@ from modules.text_generation import (
    stop_everything_event
 )
 from modules.ui_notebook import store_notebook_state_and_debounce
-from modules.utils import gradio, sanitize_filename
+from modules.utils import gradio
 inputs = ('textbox-default', 'interface_state')
 outputs = ('output_textbox', 'html-default')
@ -167,7 +167,6 @@ def handle_new_prompt():
 def handle_delete_prompt_confirm_default(prompt_name):
    prompt_name = sanitize_filename(prompt_name)
    available_prompts = utils.get_available_prompts()
    current_index = available_prompts.index(prompt_name) if prompt_name in available_prompts else 0
@ -200,8 +199,6 @@ def handle_rename_prompt_click_default(current_name):
 def handle_rename_prompt_confirm_default(new_name, current_name):
    new_name = sanitize_filename(new_name)
    current_name = sanitize_filename(current_name)
    old_path = shared.user_data_dir / "logs" / "notebook" / f"{current_name}.txt"
    new_path = shared.user_data_dir / "logs" / "notebook" / f"{new_name}.txt"
--- a/modules/ui_image_generation.py
+++ b/modules/ui_image_generation.py
@ -798,9 +798,6 @@ def generate(state, save_images=True):
        if seed == -1:
            seed = random.randint(0, 2**32 - 1)
        # Store resolved seed back so callers (e.g. API) can access it
        state['image_seed_resolved'] = seed
        device = get_device()
        if device is None:
            device = "cpu"
--- a/modules/ui_model_menu.py
+++ b/modules/ui_model_menu.py
@ -51,9 +51,7 @@ def create_ui():
                        with gr.Column():
                            shared.gradio['vram_info'] = gr.HTML(value=get_initial_vram_info())
-                            if not shared.args.portable:
+                            shared.gradio['cpu_moe'] = gr.Checkbox(label="cpu-moe", value=shared.args.cpu_moe, info='Move the experts to the CPU. Saves VRAM on MoE models.')
                                shared.gradio['ik'] = gr.Checkbox(label="ik", value=shared.args.ik, info='Use ik_llama.cpp instead of upstream llama.cpp.')
                            shared.gradio['streaming_llm'] = gr.Checkbox(label="streaming-llm", value=shared.args.streaming_llm, info='Activate StreamingLLM to avoid re-evaluating the entire prompt when old messages are removed.')
                            shared.gradio['load_in_8bit'] = gr.Checkbox(label="load-in-8bit", value=shared.args.load_in_8bit)
                            shared.gradio['load_in_4bit'] = gr.Checkbox(label="load-in-4bit", value=shared.args.load_in_4bit)
@ -66,13 +64,13 @@ def create_ui():
                            )
                            # Multimodal
-                            with gr.Accordion("Multimodal (vision)", open=False) as shared.gradio['mmproj_accordion']:
+                            with gr.Accordion("Multimodal (vision)", open=False, elem_classes='tgw-accordion') as shared.gradio['mmproj_accordion']:
                                with gr.Row():
                                    shared.gradio['mmproj'] = gr.Dropdown(label="mmproj file", choices=utils.get_available_mmproj(), value=lambda: shared.args.mmproj or 'None', elem_classes='slim-dropdown', info=f'Select a file that matches your model. Must be placed in {shared.user_data_dir}/mmproj/', interactive=not mu)
                                    ui.create_refresh_button(shared.gradio['mmproj'], lambda: None, lambda: {'choices': utils.get_available_mmproj()}, 'refresh-button', interactive=not mu)
                            # Speculative decoding
-                            with gr.Accordion("Speculative decoding", open=False) as shared.gradio['speculative_decoding_accordion']:
+                            with gr.Accordion("Speculative decoding", open=False, elem_classes='tgw-accordion') as shared.gradio['speculative_decoding_accordion']:
                                shared.gradio['draft_max'] = gr.Number(label="draft-max", precision=0, step=1, value=shared.args.draft_max, info='Maximum number of tokens to draft for speculative decoding. Recommended: 4 for draft model, 64 for n-gram.')
                                gr.Markdown('#### Draft model')
@ -91,7 +89,7 @@ def create_ui():
                                shared.gradio['spec_ngram_min_hits'] = gr.Number(label="spec-ngram-min-hits", precision=0, step=1, value=shared.args.spec_ngram_min_hits, info='Minimum n-gram hits for ngram-map speculative decoding.', visible=shared.args.spec_type != 'none')
                    gr.Markdown("## Other options")
-                    with gr.Accordion("See more options", open=False):
+                    with gr.Accordion("See more options", open=False, elem_classes='tgw-accordion'):
                        with gr.Row():
                            with gr.Column():
                                shared.gradio['parallel'] = gr.Slider(label="parallel", minimum=1, step=1, maximum=64, value=shared.args.parallel, info='Number of parallel request slots for the API. The context size is divided equally among slots. For example, to have 4 slots with 8192 context each, set ctx_size to 32768.')
@ -108,7 +106,6 @@ def create_ui():
                            with gr.Column():
                                shared.gradio['cpu'] = gr.Checkbox(label="cpu", value=shared.args.cpu, info='Use PyTorch in CPU mode.')
                                shared.gradio['disk'] = gr.Checkbox(label="disk", value=shared.args.disk)
                                shared.gradio['cpu_moe'] = gr.Checkbox(label="cpu-moe", value=shared.args.cpu_moe, info='Move the experts to the CPU. Saves VRAM on MoE models.')
                                shared.gradio['row_split'] = gr.Checkbox(label="row_split", value=shared.args.row_split, info='Split the model by rows across GPUs. This may improve multi-gpu performance.')
                                shared.gradio['no_kv_offload'] = gr.Checkbox(label="no_kv_offload", value=shared.args.no_kv_offload, info='Do not offload the K, Q, V to the GPU. This saves VRAM but reduces performance.')
                                shared.gradio['no_mmap'] = gr.Checkbox(label="no-mmap", value=shared.args.no_mmap)
--- a/modules/ui_notebook.py
+++ b/modules/ui_notebook.py
@ -11,7 +11,7 @@ from modules.text_generation import (
    get_token_ids,
    stop_everything_event
 )
-from modules.utils import gradio, sanitize_filename
+from modules.utils import gradio
 _notebook_file_lock = threading.Lock()
 _notebook_auto_save_timer = None
@ -202,7 +202,6 @@ def handle_new_prompt():
 def handle_delete_prompt_confirm_notebook(prompt_name):
    prompt_name = sanitize_filename(prompt_name)
    available_prompts = utils.get_available_prompts()
    current_index = available_prompts.index(prompt_name) if prompt_name in available_prompts else 0
@ -234,8 +233,6 @@ def handle_rename_prompt_click_notebook(current_name):
 def handle_rename_prompt_confirm_notebook(new_name, current_name):
    new_name = sanitize_filename(new_name)
    current_name = sanitize_filename(current_name)
    old_path = shared.user_data_dir / "logs" / "notebook" / f"{current_name}.txt"
    new_path = shared.user_data_dir / "logs" / "notebook" / f"{new_name}.txt"
@ -252,7 +249,6 @@ def handle_rename_prompt_confirm_notebook(new_name, current_name):
 def autosave_prompt(text, prompt_name):
    """Automatically save the text to the selected prompt file"""
    prompt_name = sanitize_filename(prompt_name)
    if prompt_name and text.strip():
        prompt_path = shared.user_data_dir / "logs" / "notebook" / f"{prompt_name}.txt"
        prompt_path.parent.mkdir(parents=True, exist_ok=True)
--- a/modules/utils.py
+++ b/modules/utils.py
@ -105,9 +105,6 @@ def resolve_model_path(model_name_or_path, image_model=False):
    before the default models directory.
    """
    if model_name_or_path is None:
        raise FileNotFoundError("No model specified.")
    path_candidate = Path(model_name_or_path)
    if path_candidate.exists():
        return path_candidate
--- a/requirements/full/requirements.txt
+++ b/requirements/full/requirements.txt
@ -1,4 +1,4 @@
-accelerate==1.13.*
+accelerate==1.12.*
 audioop-lts<1.0; python_version >= "3.13"
 bitsandbytes==0.49.*
 datasets
@ -9,7 +9,6 @@ flash-linear-attention==0.4.*
 huggingface-hub==1.5.*
 jinja2==3.1.6
 markdown
 mcp==1.27.0
 numpy==2.2.*
 pandas
 peft==0.18.*
@ -26,14 +25,14 @@ sentencepiece
 tensorboard
 torchao==0.15.*
 trafilatura==2.0.0
-transformers==5.5.*
+transformers==5.3.*
 triton-windows==3.5.1.post24; platform_system == "Windows"
 tqdm
 wandb
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio-4.37.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio-4.37.2+custom.12-py3-none-any.whl
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio_client-1.0.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio_client-1.0.2+custom.12-py3-none-any.whl
 # API
 flask_cloudflared==0.0.15
@ -41,11 +40,9 @@ sse-starlette==1.6.5
 tiktoken
 # CUDA wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/llama_cpp_binaries-0.110.0+cu124-py3-none-win_amd64.whl; platform_system == "Windows"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.98.0/llama_cpp_binaries-0.98.0+cu124-py3-none-win_amd64.whl; platform_system == "Windows"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/llama_cpp_binaries-0.110.0+cu124-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.98.0/llama_cpp_binaries-0.98.0+cu124-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/ik_llama_cpp_binaries-0.110.0+cu124-py3-none-win_amd64.whl; platform_system == "Windows"
+https://github.com/turboderp-org/exllamav3/releases/download/v0.0.26/exllamav3-0.0.26+cu128.torch2.9.0-cp313-cp313-win_amd64.whl; platform_system == "Windows" and python_version == "3.13"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/ik_llama_cpp_binaries-0.110.0+cu124-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/turboderp-org/exllamav3/releases/download/v0.0.26/exllamav3-0.0.26+cu128.torch2.9.0-cp313-cp313-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.13"
 https://github.com/turboderp-org/exllamav3/releases/download/v0.0.28/exllamav3-0.0.28+cu128.torch2.9.0-cp313-cp313-win_amd64.whl; platform_system == "Windows" and python_version == "3.13"
 https://github.com/turboderp-org/exllamav3/releases/download/v0.0.28/exllamav3-0.0.28+cu128.torch2.9.0-cp313-cp313-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.13"
 https://github.com/kingbri1/flash-attention/releases/download/v2.8.3/flash_attn-2.8.3+cu128torch2.9.0cxx11abiFALSE-cp313-cp313-win_amd64.whl; platform_system == "Windows" and python_version == "3.13"
 https://github.com/kingbri1/flash-attention/releases/download/v2.8.3/flash_attn-2.8.3+cu128torch2.9.0cxx11abiFALSE-cp313-cp313-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.13"
--- a/requirements/full/requirements_amd.txt
+++ b/requirements/full/requirements_amd.txt
@ -1,4 +1,4 @@
-accelerate==1.13.*
+accelerate==1.12.*
 audioop-lts<1.0; python_version >= "3.13"
 datasets
 diffusers==0.37.*
@ -7,7 +7,6 @@ fastapi==0.112.4
 huggingface-hub==1.5.*
 jinja2==3.1.6
 markdown
 mcp==1.27.0
 numpy==2.2.*
 pandas
 peft==0.18.*
@ -23,14 +22,14 @@ scipy
 sentencepiece
 tensorboard
 torchao==0.15.*
-transformers==5.5.*
+transformers==5.3.*
 tqdm
 trafilatura==2.0.0
 wandb
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio-4.37.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio-4.37.2+custom.12-py3-none-any.whl
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio_client-1.0.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio_client-1.0.2+custom.12-py3-none-any.whl
 # API
 flask_cloudflared==0.0.15
@ -38,5 +37,5 @@ sse-starlette==1.6.5
 tiktoken
 # AMD wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/llama_cpp_binaries-0.110.0+rocm7.2-py3-none-win_amd64.whl; platform_system == "Windows"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.98.0/llama_cpp_binaries-0.98.0+rocm7.2-py3-none-win_amd64.whl; platform_system == "Windows"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/llama_cpp_binaries-0.110.0+rocm7.2-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.98.0/llama_cpp_binaries-0.98.0+rocm7.2-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
--- a/requirements/full/requirements_apple_intel.txt
+++ b/requirements/full/requirements_apple_intel.txt
@ -1,4 +1,4 @@
-accelerate==1.13.*
+accelerate==1.12.*
 audioop-lts<1.0; python_version >= "3.13"
 datasets
 diffusers==0.37.*
@ -7,7 +7,6 @@ fastapi==0.112.4
 huggingface-hub==1.5.*
 jinja2==3.1.6
 markdown
 mcp==1.27.0
 numpy==2.2.*
 pandas
 peft==0.18.*
@ -23,14 +22,14 @@ scipy
 sentencepiece
 tensorboard
 torchao==0.15.*
-transformers==5.5.*
+transformers==5.3.*
 tqdm
 trafilatura==2.0.0
 wandb
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio-4.37.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio-4.37.2+custom.12-py3-none-any.whl
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio_client-1.0.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio_client-1.0.2+custom.12-py3-none-any.whl
 # API
 flask_cloudflared==0.0.15
@ -38,4 +37,4 @@ sse-starlette==1.6.5
 tiktoken
 # Mac wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/llama_cpp_binaries-0.110.0-py3-none-macosx_13_0_x86_64.whl; platform_system == "Darwin"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.98.0/llama_cpp_binaries-0.98.0-py3-none-macosx_13_0_x86_64.whl; platform_system == "Darwin"
--- a/requirements/full/requirements_apple_silicon.txt
+++ b/requirements/full/requirements_apple_silicon.txt
@ -1,4 +1,4 @@
-accelerate==1.13.*
+accelerate==1.12.*
 audioop-lts<1.0; python_version >= "3.13"
 datasets
 diffusers==0.37.*
@ -7,7 +7,6 @@ fastapi==0.112.4
 huggingface-hub==1.5.*
 jinja2==3.1.6
 markdown
 mcp==1.27.0
 numpy==2.2.*
 pandas
 peft==0.18.*
@ -23,14 +22,14 @@ scipy
 sentencepiece
 tensorboard
 torchao==0.15.*
-transformers==5.5.*
+transformers==5.3.*
 tqdm
 trafilatura==2.0.0
 wandb
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio-4.37.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio-4.37.2+custom.12-py3-none-any.whl
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio_client-1.0.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio_client-1.0.2+custom.12-py3-none-any.whl
 # API
 flask_cloudflared==0.0.15
@ -38,4 +37,4 @@ sse-starlette==1.6.5
 tiktoken
 # Mac wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/llama_cpp_binaries-0.110.0-py3-none-macosx_13_0_arm64.whl; platform_system == "Darwin"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.98.0/llama_cpp_binaries-0.98.0-py3-none-macosx_13_0_arm64.whl; platform_system == "Darwin"
--- a/requirements/full/requirements_cpu_only.txt
+++ b/requirements/full/requirements_cpu_only.txt
@ -1,4 +1,4 @@
-accelerate==1.13.*
+accelerate==1.12.*
 audioop-lts<1.0; python_version >= "3.13"
 datasets
 diffusers==0.37.*
@ -7,7 +7,6 @@ fastapi==0.112.4
 huggingface-hub==1.5.*
 jinja2==3.1.6
 markdown
 mcp==1.27.0
 numpy==2.2.*
 pandas
 peft==0.18.*
@ -23,14 +22,14 @@ scipy
 sentencepiece
 tensorboard
 torchao==0.15.*
-transformers==5.5.*
+transformers==5.3.*
 tqdm
 trafilatura==2.0.0
 wandb
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio-4.37.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio-4.37.2+custom.12-py3-none-any.whl
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio_client-1.0.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio_client-1.0.2+custom.12-py3-none-any.whl
 # API
 flask_cloudflared==0.0.15
@ -38,7 +37,5 @@ sse-starlette==1.6.5
 tiktoken
 # llama.cpp (CPU only)
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/llama_cpp_binaries-0.110.0+cpu-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.98.0/llama_cpp_binaries-0.98.0+cpu-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/llama_cpp_binaries-0.110.0+cpu-py3-none-win_amd64.whl; platform_system == "Windows"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.98.0/llama_cpp_binaries-0.98.0+cpu-py3-none-win_amd64.whl; platform_system == "Windows"
 https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/ik_llama_cpp_binaries-0.110.0+cpu-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
 https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/ik_llama_cpp_binaries-0.110.0+cpu-py3-none-win_amd64.whl; platform_system == "Windows"
--- a/requirements/full/requirements_nowheels.txt
+++ b/requirements/full/requirements_nowheels.txt
@ -1,4 +1,4 @@
-accelerate==1.13.*
+accelerate==1.12.*
 audioop-lts<1.0; python_version >= "3.13"
 datasets
 diffusers==0.37.*
@ -7,7 +7,6 @@ fastapi==0.112.4
 huggingface-hub==1.5.*
 jinja2==3.1.6
 markdown
 mcp==1.27.0
 numpy==2.2.*
 pandas
 peft==0.18.*
@ -23,14 +22,14 @@ scipy
 sentencepiece
 tensorboard
 torchao==0.15.*
-transformers==5.5.*
+transformers==5.3.*
 tqdm
 trafilatura==2.0.0
 wandb
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio-4.37.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio-4.37.2+custom.12-py3-none-any.whl
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio_client-1.0.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio_client-1.0.2+custom.12-py3-none-any.whl
 # API
 flask_cloudflared==0.0.15
--- a/requirements/portable/requirements.txt
+++ b/requirements/portable/requirements.txt
@ -3,7 +3,6 @@ fastapi==0.112.4
 huggingface-hub==1.5.*
 jinja2==3.1.6
 markdown
 mcp==1.27.0
 numpy==2.2.*
 pydantic==2.11.0
 pymupdf==1.27.*
@ -15,8 +14,8 @@ trafilatura==2.0.0
 tqdm
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio-4.37.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio-4.37.2+custom.12-py3-none-any.whl
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio_client-1.0.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio_client-1.0.2+custom.12-py3-none-any.whl
 # API
 flask_cloudflared==0.0.15
@ -24,5 +23,5 @@ sse-starlette==1.6.5
 tiktoken
 # CUDA wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/llama_cpp_binaries-0.110.0+cu124-py3-none-win_amd64.whl; platform_system == "Windows"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.98.0/llama_cpp_binaries-0.98.0+cu124-py3-none-win_amd64.whl; platform_system == "Windows"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/llama_cpp_binaries-0.110.0+cu124-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.98.0/llama_cpp_binaries-0.98.0+cu124-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
--- a/requirements/portable/requirements_amd.txt
+++ b/requirements/portable/requirements_amd.txt
@ -3,7 +3,6 @@ fastapi==0.112.4
 huggingface-hub==1.5.*
 jinja2==3.1.6
 markdown
 mcp==1.27.0
 numpy==2.2.*
 pydantic==2.11.0
 pymupdf==1.27.*
@ -15,8 +14,8 @@ trafilatura==2.0.0
 tqdm
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio-4.37.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio-4.37.2+custom.12-py3-none-any.whl
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio_client-1.0.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio_client-1.0.2+custom.12-py3-none-any.whl
 # API
 flask_cloudflared==0.0.15
@ -24,5 +23,5 @@ sse-starlette==1.6.5
 tiktoken
 # AMD wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/llama_cpp_binaries-0.110.0+rocm7.2-py3-none-win_amd64.whl; platform_system == "Windows"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.98.0/llama_cpp_binaries-0.98.0+rocm7.2-py3-none-win_amd64.whl; platform_system == "Windows"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/llama_cpp_binaries-0.110.0+rocm7.2-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.98.0/llama_cpp_binaries-0.98.0+rocm7.2-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
--- a/requirements/portable/requirements_apple_intel.txt
+++ b/requirements/portable/requirements_apple_intel.txt
@ -3,7 +3,6 @@ fastapi==0.112.4
 huggingface-hub==1.5.*
 jinja2==3.1.6
 markdown
 mcp==1.27.0
 numpy==2.2.*
 pydantic==2.11.0
 pymupdf==1.27.*
@ -15,8 +14,8 @@ trafilatura==2.0.0
 tqdm
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio-4.37.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio-4.37.2+custom.12-py3-none-any.whl
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio_client-1.0.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio_client-1.0.2+custom.12-py3-none-any.whl
 # API
 flask_cloudflared==0.0.15
@ -24,4 +23,4 @@ sse-starlette==1.6.5
 tiktoken
 # Mac wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/llama_cpp_binaries-0.110.0-py3-none-macosx_13_0_x86_64.whl; platform_system == "Darwin"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.98.0/llama_cpp_binaries-0.98.0-py3-none-macosx_13_0_x86_64.whl; platform_system == "Darwin"
--- a/requirements/portable/requirements_apple_silicon.txt
+++ b/requirements/portable/requirements_apple_silicon.txt
@ -3,7 +3,6 @@ fastapi==0.112.4
 huggingface-hub==1.5.*
 jinja2==3.1.6
 markdown
 mcp==1.27.0
 numpy==2.2.*
 pydantic==2.11.0
 pymupdf==1.27.*
@ -15,8 +14,8 @@ trafilatura==2.0.0
 tqdm
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio-4.37.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio-4.37.2+custom.12-py3-none-any.whl
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio_client-1.0.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio_client-1.0.2+custom.12-py3-none-any.whl
 # API
 flask_cloudflared==0.0.15
@ -24,4 +23,4 @@ sse-starlette==1.6.5
 tiktoken
 # Mac wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/llama_cpp_binaries-0.110.0-py3-none-macosx_13_0_arm64.whl; platform_system == "Darwin"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.98.0/llama_cpp_binaries-0.98.0-py3-none-macosx_13_0_arm64.whl; platform_system == "Darwin"
--- a/requirements/portable/requirements_cpu_only.txt
+++ b/requirements/portable/requirements_cpu_only.txt
@ -3,7 +3,6 @@ fastapi==0.112.4
 huggingface-hub==1.5.*
 jinja2==3.1.6
 markdown
 mcp==1.27.0
 numpy==2.2.*
 pydantic==2.11.0
 pymupdf==1.27.*
@ -15,8 +14,8 @@ trafilatura==2.0.0
 tqdm
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio-4.37.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio-4.37.2+custom.12-py3-none-any.whl
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio_client-1.0.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio_client-1.0.2+custom.12-py3-none-any.whl
 # API
 flask_cloudflared==0.0.15
@ -24,5 +23,5 @@ sse-starlette==1.6.5
 tiktoken
 # llama.cpp (CPU only)
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/llama_cpp_binaries-0.110.0+cpu-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.98.0/llama_cpp_binaries-0.98.0+cpu-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/llama_cpp_binaries-0.110.0+cpu-py3-none-win_amd64.whl; platform_system == "Windows"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.98.0/llama_cpp_binaries-0.98.0+cpu-py3-none-win_amd64.whl; platform_system == "Windows"
--- a/requirements/portable/requirements_cuda131.txt
+++ b/requirements/portable/requirements_cuda131.txt
@ -3,7 +3,6 @@ fastapi==0.112.4
 huggingface-hub==1.5.*
 jinja2==3.1.6
 markdown
 mcp==1.27.0
 numpy==2.2.*
 pydantic==2.11.0
 pymupdf==1.27.*
@ -15,8 +14,8 @@ trafilatura==2.0.0
 tqdm
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio-4.37.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio-4.37.2+custom.12-py3-none-any.whl
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio_client-1.0.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio_client-1.0.2+custom.12-py3-none-any.whl
 # API
 flask_cloudflared==0.0.15
@ -24,5 +23,5 @@ sse-starlette==1.6.5
 tiktoken
 # CUDA wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/llama_cpp_binaries-0.110.0+cu131-py3-none-win_amd64.whl; platform_system == "Windows"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.98.0/llama_cpp_binaries-0.98.0+cu131-py3-none-win_amd64.whl; platform_system == "Windows"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/llama_cpp_binaries-0.110.0+cu131-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.98.0/llama_cpp_binaries-0.98.0+cu131-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
--- a/requirements/portable/requirements_ik.txt
+++ b/requirements/portable/requirements_ik.txt
@ -1,28 +0,0 @@
 audioop-lts<1.0; python_version >= "3.13"
 fastapi==0.112.4
 huggingface-hub==1.5.*
 jinja2==3.1.6
 markdown
 mcp==1.27.0
 numpy==2.2.*
 pydantic==2.11.0
 pymupdf==1.27.*
 python-docx==1.1.2
 pyyaml
 requests
 rich
 trafilatura==2.0.0
 tqdm
 # Gradio
 https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio-4.37.2+custom.19-py3-none-any.whl
 https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio_client-1.0.2+custom.19-py3-none-any.whl
 # API
 flask_cloudflared==0.0.15
 sse-starlette==1.6.5
 tiktoken
 # CUDA wheels
 https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/ik_llama_cpp_binaries-0.110.0+cu124-py3-none-win_amd64.whl; platform_system == "Windows"
 https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/ik_llama_cpp_binaries-0.110.0+cu124-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
--- a/requirements/portable/requirements_ik_cpu_only.txt
+++ b/requirements/portable/requirements_ik_cpu_only.txt
@ -1,28 +0,0 @@
 audioop-lts<1.0; python_version >= "3.13"
 fastapi==0.112.4
 huggingface-hub==1.5.*
 jinja2==3.1.6
 markdown
 mcp==1.27.0
 numpy==2.2.*
 pydantic==2.11.0
 pymupdf==1.27.*
 python-docx==1.1.2
 pyyaml
 requests
 rich
 trafilatura==2.0.0
 tqdm
 # Gradio
 https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio-4.37.2+custom.19-py3-none-any.whl
 https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio_client-1.0.2+custom.19-py3-none-any.whl
 # API
 flask_cloudflared==0.0.15
 sse-starlette==1.6.5
 tiktoken
 # ik_llama.cpp (CPU only)
 https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/ik_llama_cpp_binaries-0.110.0+cpu-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
 https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/ik_llama_cpp_binaries-0.110.0+cpu-py3-none-win_amd64.whl; platform_system == "Windows"
--- a/requirements/portable/requirements_ik_cuda131.txt
+++ b/requirements/portable/requirements_ik_cuda131.txt
@ -1,28 +0,0 @@
 audioop-lts<1.0; python_version >= "3.13"
 fastapi==0.112.4
 huggingface-hub==1.5.*
 jinja2==3.1.6
 markdown
 mcp==1.27.0
 numpy==2.2.*
 pydantic==2.11.0
 pymupdf==1.27.*
 python-docx==1.1.2
 pyyaml
 requests
 rich
 trafilatura==2.0.0
 tqdm
 # Gradio
 https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio-4.37.2+custom.19-py3-none-any.whl
 https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio_client-1.0.2+custom.19-py3-none-any.whl
 # API
 flask_cloudflared==0.0.15
 sse-starlette==1.6.5
 tiktoken
 # CUDA wheels
 https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/ik_llama_cpp_binaries-0.110.0+cu131-py3-none-win_amd64.whl; platform_system == "Windows"
 https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/ik_llama_cpp_binaries-0.110.0+cu131-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
--- a/requirements/portable/requirements_nowheels.txt
+++ b/requirements/portable/requirements_nowheels.txt
@ -3,7 +3,6 @@ fastapi==0.112.4
 huggingface-hub==1.5.*
 jinja2==3.1.6
 markdown
 mcp==1.27.0
 numpy==2.2.*
 pydantic==2.11.0
 pymupdf==1.27.*
@ -15,8 +14,8 @@ trafilatura==2.0.0
 tqdm
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio-4.37.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio-4.37.2+custom.12-py3-none-any.whl
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio_client-1.0.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio_client-1.0.2+custom.12-py3-none-any.whl
 # API
 flask_cloudflared==0.0.15
--- a/requirements/portable/requirements_vulkan.txt
+++ b/requirements/portable/requirements_vulkan.txt
@ -3,7 +3,6 @@ fastapi==0.112.4
 huggingface-hub==1.5.*
 jinja2==3.1.6
 markdown
 mcp==1.27.0
 numpy==2.2.*
 pydantic==2.11.0
 pymupdf==1.27.*
@ -15,8 +14,8 @@ trafilatura==2.0.0
 tqdm
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio-4.37.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio-4.37.2+custom.12-py3-none-any.whl
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.19/gradio_client-1.0.2+custom.19-py3-none-any.whl
+https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.12/gradio_client-1.0.2+custom.12-py3-none-any.whl
 # API
 flask_cloudflared==0.0.15
@ -24,5 +23,5 @@ sse-starlette==1.6.5
 tiktoken
 # Vulkan wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/llama_cpp_binaries-0.110.0+vulkan-py3-none-win_amd64.whl; platform_system == "Windows"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.98.0/llama_cpp_binaries-0.98.0+vulkan-py3-none-win_amd64.whl; platform_system == "Windows"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/llama_cpp_binaries-0.110.0+vulkan-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.98.0/llama_cpp_binaries-0.98.0+vulkan-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
--- a/server.py
+++ b/server.py
@ -18,6 +18,7 @@ import modules.extensions as extensions_module
 from modules.LoRA import add_lora_to_model
 from modules.models import load_model, unload_model_if_idle
 from modules.models_settings import (
    get_fallback_settings,
    get_model_metadata,
    update_model_parameters
 )
@ -270,6 +271,10 @@ if __name__ == "__main__":
    # Apply CLI overrides for image model settings (CLI flags take precedence over saved settings)
    shared.apply_image_model_cli_overrides()
    # Fallback settings for models
    shared.model_config['.*'] = get_fallback_settings()
    shared.model_config.move_to_end('.*', last=False)  # Move to the beginning
    # Activate the extensions listed on settings.yaml
    extensions_module.available_extensions = utils.get_available_extensions()
    for extension in shared.settings['default_extensions']:
--- a/user_data/models/config.yaml
+++ b/user_data/models/config.yaml
@ -0,0 +1,203 @@
 .*(llama|alpac|vicuna|guanaco|koala|llava|wizardlm|metharme|pygmalion-7b|pygmalion-2|mythalion|wizard-mega|openbuddy|vigogne|h2ogpt-research|manticore):
  model_type: 'llama'
 .*(opt-|opt_|opt1|opt3|optfor|galactica|galpaca|pygmalion-350m):
  model_type: 'opt'
 .*(gpt-j|gptj|gpt4all-j|malion-6b|pygway|pygmalion-6b|dolly-v1):
  model_type: 'gptj'
 .*(gpt-neox|koalpaca-polyglot|polyglot.*koalpaca|polyglot-ko|polyglot_ko|pythia|stablelm|incite|dolly-v2|polycoder|h2ogpt-oig|h2ogpt-oasst1|h2ogpt-gm):
  model_type: 'gptneox'
 .*bloom:
  model_type: 'bloom'
 .*gpt2:
  model_type: 'gpt2'
 .*falcon:
  model_type: 'falcon'
 .*mpt:
  model_type: 'mpt'
 .*(starcoder|starchat):
  model_type: 'starcoder'
 .*dolly-v2:
  model_type: 'dollyv2'
 .*replit:
  model_type: 'replit'
 .*(oasst|openassistant-|stablelm-7b-sft-v7-epoch-3):
  instruction_template: 'Open Assistant'
  skip_special_tokens: false
 (?!.*galactica)(?!.*reward).*openassistant:
  instruction_template: 'Open Assistant'
  skip_special_tokens: false
 .*galactica:
  skip_special_tokens: false
 .*dolly-v[0-9]-[0-9]*b:
  instruction_template: 'Alpaca'
  skip_special_tokens: false
 .*alpaca-native-4bit:
  instruction_template: 'Alpaca'
 .*llava:
  instruction_template: 'LLaVA'
 .*llava.*1.5:
  instruction_template: 'Vicuna-v1.1'
 .*wizard.*mega:
  instruction_template: 'Wizard-Mega'
 .*starchat-beta:
  instruction_template: 'Starchat-Beta'
 (?!.*v0)(?!.*1.1)(?!.*1_1)(?!.*stable)(?!.*chinese).*vicuna:
  instruction_template: 'Vicuna-v0'
 .*vicuna.*v0:
  instruction_template: 'Vicuna-v0'
 .*vicuna.*(1.1|1_1|1.3|1_3):
  instruction_template: 'Vicuna-v1.1'
 .*vicuna.*(1.5|1_5):
  instruction_template: 'Vicuna-v1.1'
 .*stable.*vicuna:
  instruction_template: 'StableVicuna'
 (?!.*chat).*chinese-vicuna:
  instruction_template: 'Alpaca'
 .*chinese-vicuna.*chat:
  instruction_template: 'Chinese-Vicuna-Chat'
 .*alpaca:
  instruction_template: 'Alpaca'
 .*koala:
  instruction_template: 'Koala'
 .*chatglm:
  instruction_template: 'ChatGLM'
 .*(metharme|pygmalion|mythalion):
  instruction_template: 'Metharme'
 .*raven:
  instruction_template: 'RWKV-Raven'
 .*moss-moon.*sft:
  instruction_template: 'MOSS'
 .*stablelm-tuned:
  instruction_template: 'StableLM'
 .*galactica.*finetuned:
  instruction_template: 'Galactica Finetuned'
 .*galactica.*-v2:
  instruction_template: 'Galactica v2'
 (?!.*finetuned)(?!.*-v2).*galactica:
  instruction_template: 'Galactica'
 .*guanaco:
  instruction_template: 'Guanaco non-chat'
 .*baize:
  instruction_template: 'Baize'
 .*mpt-.*instruct:
  instruction_template: 'Alpaca'
 .*mpt-.*chat:
  instruction_template: 'ChatML'
 (?!.*-flan-)(?!.*-t5-).*lamini-:
  instruction_template: 'Alpaca'
 .*incite.*chat:
  instruction_template: 'INCITE-Chat'
 .*incite.*instruct:
  instruction_template: 'INCITE-Instruct'
 .*ziya-:
  instruction_template: 'Ziya'
 .*koalpaca:
  instruction_template: 'KoAlpaca'
 .*openbuddy:
  instruction_template: 'OpenBuddy'
 (?!.*chat).*vigogne:
  instruction_template: 'Vigogne-Instruct'
 .*vigogne.*chat:
  instruction_template: 'Vigogne-Chat'
 .*(llama-deus|supercot|llama-natural-instructions|open-llama-0.3t-7b-instruct-dolly-hhrlhf|open-llama-0.3t-7b-open-instruct):
  instruction_template: 'Alpaca'
 .*bactrian:
  instruction_template: 'Bactrian'
 .*(h2ogpt-oig-|h2ogpt-oasst1-|h2ogpt-research-oasst1-):
  instruction_template: 'INCITE-Chat'
 .*h2ogpt-gm-:
  instruction_template: 'H2O-prompt_answer'
 .*manticore:
  instruction_template: 'Manticore Chat'
 .*bluemoonrp-(30|13)b:
  instruction_template: 'Bluemoon'
 .*Nous-Hermes-13b:
  instruction_template: 'Alpaca'
 .*airoboros:
  instruction_template: 'Vicuna-v1.1'
 .*airoboros.*1.2:
  instruction_template: 'Airoboros-v1.2'
 .*alpa(cino|sta):
  instruction_template: 'Alpaca'
 .*hippogriff:
  instruction_template: 'Hippogriff'
 .*lazarus:
  instruction_template: 'Alpaca'
 .*guanaco-.*(7|13|33|65)b:
  instruction_template: 'Vicuna-v0'
 .*hypermantis:
  instruction_template: 'Alpaca'
 .*open-llama-.*-open-instruct:
  instruction_template: 'Alpaca'
 .*starcoder-gpteacher-code-instruct:
  instruction_template: 'Alpaca'
 .*tulu:
  instruction_template: 'Tulu'
 .*chronos:
  instruction_template: 'Alpaca'
 .*samantha:
  instruction_template: 'Samantha'
 .*wizardcoder:
  instruction_template: 'Alpaca'
 .*minotaur:
  instruction_template: 'Manticore Chat'
 .*orca_mini:
  instruction_template: 'Orca Mini'
 .*(platypus|gplatty|superplatty):
  instruction_template: 'Alpaca'
 .*(openorca-platypus2):
  instruction_template: 'OpenOrca-Platypus2'
 .*longchat:
  instruction_template: 'Vicuna-v1.1'
 .*vicuna-33b:
  instruction_template: 'Vicuna-v1.1'
 .*redmond-hermes-coder:
  instruction_template: 'Alpaca'
 .*wizardcoder-15b:
  instruction_template: 'Alpaca'
 .*wizardlm:
  instruction_template: 'Vicuna-v1.1'
 .*godzilla:
  instruction_template: 'Alpaca'
 .*llama(-?)(2|v2).*chat:
  instruction_template: 'Llama-v2'
 .*newhope:
  instruction_template: 'NewHope'
 .*stablebeluga2:
  instruction_template: 'StableBeluga2'
 .*openchat:
  instruction_template: 'OpenChat'
 .*codellama.*instruct:
  instruction_template: 'Llama-v2'
 .*(mistral|mixtral).*instruct:
  instruction_template: 'Mistral'
 .*mistral.*openorca:
  instruction_template: 'ChatML'
 .*(WizardCoder-Python-34B-V1.0|Phind-CodeLlama-34B-v2|CodeBooga-34B-v0.1):
  instruction_template: 'Alpaca'
 .*orca-2-(13|7)b:
  instruction_template: 'ChatML'
 .*openhermes.*mistral:
  instruction_template: 'ChatML'
 .*Yi-34B-Chat:
  instruction_template: 'ChatML'
 (dolphin).*:
  instruction_template: 'ChatML'
 .*synthia:
  instruction_template: 'Synthia'
 .*(hercules|hyperion):
  instruction_template: 'ChatML'
 .*command-r:
  instruction_template: 'Command-R'
 .*xwin-lm-70b-v0.1:
  instruction_template: 'Vicuna-v1.1'
 .*platypus-yi-34b:
  instruction_template: 'Vicuna-v1.1'
 .*CausalLM-RP-34B:
  instruction_template: 'ChatML'
 34b-beta:
  instruction_template: 'ChatML'
 .*airoboros-3_1-yi-34b-200k:
  instruction_template: 'Llama-v2'
 .*chatqa:
  instruction_template: 'NVIDIA-ChatQA'