Merge pull request #7471 from oobabooga/dev

Merge dev branch
Remove hardcoded trust_remote_code=True in embedding loader
2026-04-07 07:33:47 +00:00 · 2026-04-06 21:54:02 -03:00 · 2026-04-06 17:47:50 -07:00 · 2026-04-06 17:04:49 -07:00 · 2026-04-06 14:13:01 -07:00 · 2026-04-06 12:06:28 -07:00
49 changed files with 749 additions and 325 deletions
--- a/.github/workflows/build-everything-tgw.yml
+++ b/.github/workflows/build-everything-tgw.yml
@ -41,6 +41,13 @@ 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: ${{ inputs.version }}
      config: 'os:windows-2022'
  build_release_rocm_linux:
    name: ROCm Linux
    uses: ./.github/workflows/build-portable-release-rocm.yml
--- a/.github/workflows/build-portable-release-ik-cuda.yml
+++ b/.github/workflows/build-portable-release-ik-cuda.yml
@ -102,8 +102,8 @@ jobs:
            VERSION_CLEAN="${{ inputs.version }}"
            VERSION_CLEAN="${VERSION_CLEAN#v}"
            cd ..
-            cp -r text-generation-webui "text-generation-webui-${VERSION_CLEAN}"
+            cp -r text-generation-webui "text-generation-webui-ik-${VERSION_CLEAN}"
-            cd "text-generation-webui-${VERSION_CLEAN}"
+            cd "text-generation-webui-ik-${VERSION_CLEAN}"
            # Remove extensions that need additional requirements
            allowed=("character_bias" "gallery" "sd_api_pictures")
@ -133,10 +133,10 @@ jobs:
            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-${VERSION_CLEAN}/portable_env"
+            mv python "text-generation-webui-ik-${VERSION_CLEAN}/portable_env"
            # 3. Prepare requirements file based on CUDA version
-            cd "text-generation-webui-${VERSION_CLEAN}"
+            cd "text-generation-webui-ik-${VERSION_CLEAN}"
            if [[ "$CUDA_VERSION" == "13.1" ]]; then
                REQ_FILE="requirements/portable/requirements_ik_cuda131.txt"
            else
@ -158,11 +158,11 @@ jobs:
            if [[ "$RUNNER_OS" == "Windows" ]]; then
                ARCHIVE_NAME="textgen-portable-ik-${VERSION_CLEAN}-${PLATFORM}-cuda${CUDA_VERSION}.zip"
                echo "Creating archive: $ARCHIVE_NAME"
-                powershell -Command "Compress-Archive -Path text-generation-webui-${VERSION_CLEAN} -DestinationPath $ARCHIVE_NAME"
+                powershell -Command "Compress-Archive -Path text-generation-webui-ik-${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-${VERSION_CLEAN}"
+                tar czf "$ARCHIVE_NAME" "text-generation-webui-ik-${VERSION_CLEAN}"
            fi
      - name: Upload files to a GitHub release
--- a/.github/workflows/build-portable-release-ik.yml
+++ b/.github/workflows/build-portable-release-ik.yml
@ -101,8 +101,8 @@ jobs:
            VERSION_CLEAN="${{ inputs.version }}"
            VERSION_CLEAN="${VERSION_CLEAN#v}"
            cd ..
-            cp -r text-generation-webui "text-generation-webui-${VERSION_CLEAN}"
+            cp -r text-generation-webui "text-generation-webui-ik-${VERSION_CLEAN}"
-            cd "text-generation-webui-${VERSION_CLEAN}"
+            cd "text-generation-webui-ik-${VERSION_CLEAN}"
            # Remove extensions that need additional requirements
            allowed=("character_bias" "gallery" "sd_api_pictures")
@ -131,10 +131,10 @@ jobs:
            cd ..
            curl -L -o python-build.tar.gz "$PYTHON_URL"
            tar -xzf python-build.tar.gz
-            mv python "text-generation-webui-${VERSION_CLEAN}/portable_env"
+            mv python "text-generation-webui-ik-${VERSION_CLEAN}/portable_env"
            # 3. Prepare requirements file
-            cd "text-generation-webui-${VERSION_CLEAN}"
+            cd "text-generation-webui-ik-${VERSION_CLEAN}"
            REQ_FILE="requirements/portable/requirements_ik_cpu_only.txt"
            echo "Using requirements file: $REQ_FILE"
@ -153,11 +153,11 @@ jobs:
            if [[ "$RUNNER_OS" == "Windows" ]]; then
                ARCHIVE_NAME="textgen-portable-ik-${VERSION_CLEAN}-${PLATFORM}.zip"
                echo "Creating archive: $ARCHIVE_NAME"
-                powershell -Command "Compress-Archive -Path text-generation-webui-${VERSION_CLEAN} -DestinationPath $ARCHIVE_NAME"
+                powershell -Command "Compress-Archive -Path text-generation-webui-ik-${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-${VERSION_CLEAN}"
+                tar czf "$ARCHIVE_NAME" "text-generation-webui-ik-${VERSION_CLEAN}"
            fi
      - name: Upload files to a GitHub release
--- 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), [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), [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.
 - **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, easy to create and extend ([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. MCP servers are also supported ([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
@ -13,21 +13,12 @@
    line-height: 28px !important;
 }
-.dark .chat .message-body :is(p,li,h1,h2,h3,h4,h5,h6),
+.dark .chat .message-body :is(p,li),
 .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
@ -22,6 +22,17 @@
    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;
 }
@ -246,8 +257,8 @@ button {
 .pretty_scrollbar::-webkit-scrollbar,
 #image-history-gallery > :nth-child(2)::-webkit-scrollbar {
-    width: 8px;
+    width: 7px;
-    height: 8px;
+    height: 7px;
 }
 .pretty_scrollbar::-webkit-scrollbar-track,
@ -260,7 +271,7 @@ 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: 30px;
+    border-radius: 9999px;
 }
 .dark .pretty_scrollbar::-webkit-scrollbar-thumb,
@ -268,18 +279,17 @@ 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: 30px;
+    border-radius: 9999px;
 }
 .pretty_scrollbar::-webkit-resizer,
 #image-history-gallery > :nth-child(2)::-webkit-resizer {
-    background: #d2d2d8;
+    background: transparent;
 }
 .dark .pretty_scrollbar::-webkit-resizer,
 .dark #image-history-gallery > :nth-child(2)::-webkit-resizer {
-    background: rgb(255 255 255 / 10%);
+    background: transparent;
    border-radius: 10px;
 }
 .pretty_scrollbar::-webkit-scrollbar-corner,
@ -436,15 +446,25 @@ audio {
 .dark .message-body h4,
 .dark .message-body h5,
 .dark .message-body h6 {
-    color: white !important;
+    color: #e8e8e8 !important;
 }
-.dark .message-body blockquote {
+.message-body blockquote {
-    border-left-color: rgb(255 255 255 / 30%);
+    border-left-width: 4px;
    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: 800;
+    font-weight: 700;
    font-size: 2.25em;
    margin-top: 0;
    margin-bottom: 0.8888889em;
@ -476,13 +496,13 @@ audio {
 }
 .message-body h5 {
-    font-weight: normal;
+    font-weight: 600;
    font-size: 1em;
    margin: 0;
 }
 .message-body h6 {
-    font-weight: normal;
+    font-weight: 600;
    font-size: 1em;
    margin: 0;
 }
@ -590,7 +610,7 @@ audio {
 }
 #chat-input textarea::-webkit-scrollbar {
-    width: 8px;
+    width: 7px;
 }
 #chat-input textarea::-webkit-scrollbar-track {
@ -599,7 +619,7 @@ audio {
 #chat-input textarea::-webkit-scrollbar-thumb {
    background: var(--neutral-300);
-    border-radius: 30px;
+    border-radius: 9999px;
 }
 .dark #chat-input textarea::-webkit-scrollbar-thumb {
@ -633,6 +653,10 @@ audio {
    background: transparent;
 }
 #chat-input .thumbnails {
    padding-top: 3px;
 }
 .chat-input-positioned {
    max-width: 54rem;
    left: 50%;
@ -735,7 +759,30 @@ audio {
 .hover-element {
    position: relative;
-    font-size: 24px;
+    padding-top: 4px;
 }
 #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 {
@ -743,27 +790,40 @@ audio {
    position: absolute;
    bottom: 100%;
    left: 0;
-    box-shadow: 0 2px 12px rgb(0 0 0 / 15%);
+    background: white;
-    border-radius: 0.5rem;
+    border: 1px solid rgba(0, 0, 0, 0.1);
    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;
-    overflow: hidden;
+    padding: 4px;
 }
 .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: white !important;
+    background: transparent !important;
-    border-radius: 0 !important;
+    border: none !important;
    border-radius: 0.5rem !important;
    justify-content: space-between;
    margin: 0 !important;
    height: 36px;
-    border-color: transparent !important;
+    font-weight: 500;
-    transition: background-color 0.15s ease;
+    box-shadow: none !important;
 }
 .hover-menu button:not(#clear-history-confirm) {
    border-bottom: 0 !important;
 }
 .hover-menu button:hover {
@ -775,19 +835,26 @@ audio {
 }
 #show-controls {
-    background-color: white;
+    background-color: transparent;
-    border-color: transparent !important;
+    border: none !important;
    height: 36px;
-    border-radius: 0;
+    border-radius: 0.5rem;
    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: var(--darker-gray);
+    background-color: transparent;
 }
 .dark #show-controls:hover {
    background-color: var(--selected-item-color-dark);
 }
 #show-controls label {
@ -797,12 +864,12 @@ audio {
    width: 100%;
    padding-right: 12px;
    gap: 10px;
-    font-weight: 600;
+    font-weight: 500;
    color: var(--button-secondary-text-color);
 }
 #show-controls label input {
-    margin-top: 4px;
+    margin-top: 5px;
 }
 .transparent-substring {
@ -842,7 +909,7 @@ audio {
 }
 #chat-input-row {
-    padding: 1rem;
+    padding: 0.5rem 1rem 1rem;
 }
 #chat-col {
@ -1208,9 +1275,14 @@ 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 {
-    border-color: var(--border-color-primary);
+    background-color: transparent !important;
    background-color: var(--darker-gray) !important;
 }
 .dark #chat-controls,
@ -1372,8 +1444,7 @@ audio {
 }
 .footer-button svg {
-    stroke: rgb(156 163 175);
+    stroke: rgb(140 140 148);
    transition: stroke 0.2s;
 }
 .footer-button:hover svg {
@ -1388,12 +1459,12 @@ audio {
    stroke: rgb(209 213 219);
 }
-.tgw-accordion {
+.block:has(> .label-wrap) {
    padding: 10px 12px !important;
    border: 1px solid #d2d2d8;
 }
-.dark .tgw-accordion {
+.dark .block:has(> .label-wrap) {
    border: 1px solid var(--border-color-dark);
 }
@ -1903,14 +1974,24 @@ table, tr, td, th, thead {
    border: 0;
 }
 .prose hr {
    border-color: var(--border-color-primary);
 }
 td + td,
-th + th { border-left: 1px solid; }
+th + th {
    border-left: 1px solid var(--border-color-primary) !important;
 }
 tr + tr td,
-tr + tr th { border-top: 1px solid; }
+tr + tr th {
    border-top: 1px solid var(--border-color-primary) !important;
 }
 thead + tbody tr:first-child td,
-thead + tbody tr:first-child th { border-top: 1px solid; }
+thead + tbody tr:first-child th {
    border-top: 1px solid var(--border-color-primary) !important;
 }
 /* ------------------------------------------------
   Tools CheckboxGroup - vertical DragDrop-like style
@ -1942,8 +2023,8 @@ thead + tbody tr:first-child th { border-top: 1px solid; }
 /* Pretty scrollbar for the tools list */
 #tools-group .wrap::-webkit-scrollbar {
-    width: 8px;
+    width: 7px;
-    height: 8px;
+    height: 7px;
 }
 #tools-group .wrap::-webkit-scrollbar-track {
@ -1953,13 +2034,13 @@ thead + tbody tr:first-child th { border-top: 1px solid; }
 #tools-group .wrap::-webkit-scrollbar-thumb,
 #tools-group .wrap::-webkit-scrollbar-thumb:hover {
    background: var(--neutral-300);
-    border-radius: 30px;
+    border-radius: 9999px;
 }
 .dark #tools-group .wrap::-webkit-scrollbar-thumb,
 .dark #tools-group .wrap::-webkit-scrollbar-thumb:hover {
    background: rgb(255 255 255 / 6.25%);
-    border-radius: 30px;
+    border-radius: 9999px;
 }
 #tools-group .wrap::-webkit-scrollbar-corner {
--- a/docs/12
+++ b/docs/12
@ -232,6 +232,17 @@ 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
@ -80,6 +80,19 @@ 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/js/main.js
+++ b/js/main.js
@ -309,18 +309,19 @@ 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"; // Show the menu
+  menu.style.display = "flex";
 }
 function hideMenu() {
-  menu.style.display = "none"; // Hide the menu
+  menu.style.display = "none";
  if (!istouchscreen) {
-    document.querySelector("#chat-input textarea").focus(); // Focus on the chat input
+    document.querySelector("#chat-input textarea").focus();
  }
 }
@ -329,7 +330,6 @@ 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();
@ -339,7 +339,6 @@ 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;
@ -348,16 +347,19 @@ if (buttonsInChat.length > 0) {
  }
 }
-function isMouseOverButtonOrMenu() {
+var menuInteracting = false;
  return menu.matches(":hover") || button.matches(":hover");
 }
-button.addEventListener("mouseenter", function () {
+hoverContainer.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();
@ -367,24 +369,20 @@ button.addEventListener("click", function () {
  }
 });
-// Delay to prevent menu hiding when the mouse leaves the button or menu
+hoverContainer.addEventListener("mouseleave", function () {
-function delayedHideMenu() {
+  if (!istouchscreen) {
-  setTimeout(function () {
+    setTimeout(function () {
-    if (!isMouseOverButtonOrMenu()) {
+      if (!hoverContainer.matches(":hover") && !menu.matches(":hover")) {
-      hideMenu();
+        hideMenu();
-    }
+      }
-  }, 100);
+    }, 50);
-}
+  }
-
+});
 // Add event listener for mouseleave on the button
 button.addEventListener("mouseleave", delayedHideMenu);
 // Add event listener for mouseleave on the menu
 menu.addEventListener("mouseleave", delayedHideMenu);
 // Add event listener for click anywhere in the document
 document.addEventListener("click", function (event) {
  // Check if the click is outside the button/menu and the menu is visible
-  if (!isMouseOverButtonOrMenu() && menu.style.display === "flex") {
+  if (!menuInteracting && !event.target.closest("#gr-hover-container") && menu.style.display === "flex") {
    hideMenu();
  }
--- a/modules/api/embeddings.py
+++ b/modules/api/embeddings.py
@ -6,6 +6,7 @@ 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
@ -41,7 +42,7 @@ 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=True)  # trust_remote_code is needed to use the encode method
+            embeddings_model = AutoModel.from_pretrained(model, trust_remote_code=shared.args.trust_remote_code)
            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
@ -4,8 +4,11 @@ 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
@ -15,7 +18,7 @@ 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 generate
+    from modules.ui_image_generation import build_generation_metadata, generate
    if shared.image_model is None:
        raise ServiceUnavailableError("No image model loaded. Load a model via the UI first.")
@ -46,10 +49,18 @@ def generations(request):
    if not images:
        raise ServiceUnavailableError("Image generation failed or produced no images.")
-    # Build response
+    # Build response with per-batch metadata (seed increments per batch)
    base_seed = state.get('image_seed_resolved', state['image_seed'])
    batch_size = int(state['image_batch_size'])
    resp = {'created': int(time.time()), 'data': []}
-    for img in images:
+    for idx, img in enumerate(images):
-        b64 = _image_to_base64(img)
+        batch_seed = base_seed + idx // batch_size
        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}
@ -63,7 +74,7 @@ def generations(request):
    return resp
-def _image_to_base64(image) -> str:
+def _image_to_base64(image, png_info=None) -> str:
    buffered = io.BytesIO()
-    image.save(buffered, format="PNG")
+    image.save(buffered, format="PNG", pnginfo=png_info)
    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, update_model_parameters
+from modules.models_settings import get_model_metadata, load_instruction_template, update_model_parameters
 from modules.utils import get_available_loras, get_available_models
@ -42,12 +42,10 @@ def model_info_dict(model_name: str) -> dict:
 def _load_model(data):
    model_name = data["model_name"]
-    args = data["args"]
+    args = data.get("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
@ -55,6 +53,16 @@ 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):
@ -62,15 +70,12 @@ def _load_model(data):
    shared.model, shared.tokenizer = load_model(model_name)
-    # Update shared.settings with custom generation defaults
+    if data.get("instruction_template_str") is not None:
-    if settings:
+        shared.settings['instruction_template_str'] = data["instruction_template_str"]
-        for k in settings:
+        logger.info("INSTRUCTION TEMPLATE: set to custom Jinja2 string")
-            if k in shared.settings:
+    elif data.get("instruction_template") is not None:
-                shared.settings[k] = settings[k]
+        shared.settings['instruction_template_str'] = load_instruction_template(data["instruction_template"])
-                if k == 'truncation_length':
+        logger.info(f"INSTRUCTION TEMPLATE: {data['instruction_template']}")
                    logger.info(f"CONTEXT 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
@ -475,10 +475,8 @@ async def handle_list_models():
@app.post("/v1/internal/model/load", dependencies=check_admin_key)
 async def handle_load_model(request_data: LoadModelRequest):
    '''
-    This endpoint is experimental and may change in the future.
+    The "args" parameter can be used to modify loader flags before loading
-
+    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": {
@ -487,18 +485,13 @@ async def handle_load_model(request_data: LoadModelRequest):
    }
    ```
-    Note that those settings will remain after loading the model. So you
+    Loader args are reset to their startup defaults between loads, so
-    may need to change them back to load a second model.
+    settings from a previous load do not leak into the next one.
-    The "settings" parameter is also a dict but with keys for the
+    The "instruction_template" parameter sets the default instruction
-    shared.settings object. It can be used to modify the default instruction
+    template by name (from user_data/instruction-templates/). The
-    template like this:
+    "instruction_template_str" parameter sets it as a raw Jinja2 string
-
+    and takes precedence over "instruction_template".
    ```
    "settings": {
      "instruction_template": "Alpaca"
    }
    ```
    '''
    try:
@ -544,8 +537,8 @@ 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:
@ -570,7 +563,7 @@ 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')
@ -587,7 +580,7 @@ def run_server():
        )
    else:
        url_proto = 'https://' if (ssl_certfile and ssl_keyfile) else 'http://'
-        urls = [f'{url_proto}{addr}:{port}/v1' for addr in server_addrs]
+        urls = [f'{url_proto}[{addr}]:{port}/v1' if ':' in addr else 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
@ -271,7 +271,8 @@ class ModelListResponse(BaseModel):
 class LoadModelRequest(BaseModel):
    model_name: str
    args: dict | None = None
-    settings: 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.")
    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
@ -210,6 +210,57 @@ 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 = ""
@ -267,6 +318,9 @@ 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':
@ -503,6 +557,9 @@ 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
@ -511,13 +568,24 @@ 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':
-                messages.pop(1)
+                pop_idx = 1
            # Remove old message when no system message is present
            elif len(messages) > 1 and messages[0]['role'] != 'system':
-                messages.pop(0)
+                pop_idx = 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:
@ -637,7 +705,7 @@ 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 = prompt.find("first assistant message") + len("first assistant message")
+        first_assistant_end = first_assistant_start + len("first assistant message")
        second_user_start = prompt.find("second user message")
        second_assistant_end = prompt.find("second assistant message") + len("second assistant message")
@ -1126,7 +1194,7 @@ 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):
+    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):
        output['visible'][-1][1] = _last_visible_before_tool_buffer or ''
    yield output
@ -1207,14 +1275,23 @@ 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:
+    if selected or mcp_servers:
-        from modules.tool_use import load_tools, execute_tool
+        from modules.tool_use import load_tools, load_mcp_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', '')
@ -1762,7 +1839,8 @@ def load_history(unique_id, character, mode):
    if not p.exists():
        return {'internal': [], 'visible': [], 'metadata': {}}
-    f = json.loads(open(p, 'rb').read())
+    with open(p, 'rb') as fh:
        f = json.loads(fh.read())
    if 'internal' in f and 'visible' in f:
        history = f
    else:
@ -1826,19 +1904,17 @@ def generate_pfp_cache(character):
    if not cache_folder.exists():
        cache_folder.mkdir()
-    for path in [shared.user_data_dir / 'characters' / f"{character}.{extension}" for extension in ['png', 'jpg', 'jpeg']]:
+    for extension in ['png', 'jpg', 'jpeg']:
        path = shared.user_data_dir / 'characters' / f"{character}.{extension}"
        if path.exists():
            original_img = Image.open(path)
-            # Define file paths
+            pfp_path = cache_folder / 'pfp_character.png'
-            pfp_path = Path(f'{cache_folder}/pfp_character.png')
+            thumb_path = cache_folder / 'pfp_character_thumb.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
@ -1859,13 +1935,13 @@ 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
-    file_contents = open(filepath, 'r', encoding='utf-8').read()
+    with open(filepath, 'r', encoding='utf-8') as fh:
        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 [Path(f"{cache_folder}/pfp_character.png"), Path(f"{cache_folder}/pfp_character_thumb.png")]:
+    for path in [cache_folder / "pfp_character.png", cache_folder / "pfp_character_thumb.png"]:
-        if path.exists():
+        path.unlink(missing_ok=True)
            path.unlink()
    picture = generate_pfp_cache(character)
@ -1921,9 +1997,7 @@ 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_path = Path(f'{cache_folder}/{cache_file}')
+        (cache_folder / cache_file).unlink(missing_ok=True)
        if cache_path.exists():
            cache_path.unlink()
    return state, state['name2'], state['context'], state['greeting'], None
@ -2018,11 +2092,10 @@ def upload_your_profile_picture(img_path):
        cache_folder.mkdir()
    if img is None:
-        if Path(f"{cache_folder}/pfp_me.png").exists():
+        (cache_folder / "pfp_me.png").unlink(missing_ok=True)
            Path(f"{cache_folder}/pfp_me.png").unlink()
    else:
        img = make_thumbnail(img)
-        img.save(Path(f'{cache_folder}/pfp_me.png'))
+        img.save(cache_folder / 'pfp_me.png')
        logger.info(f'Profile picture saved to "{cache_folder}/pfp_me.png"')
@ -2078,13 +2151,12 @@ def generate_user_pfp_cache(user):
    if not cache_folder.exists():
        cache_folder.mkdir()
-    for path in [shared.user_data_dir / 'users' / f"{user}.{extension}" for extension in ['png', 'jpg', 'jpeg']]:
+    for extension in ['png', 'jpg', 'jpeg']:
        path = shared.user_data_dir / 'users' / f"{user}.{extension}"
        if path.exists():
            original_img = Image.open(path)
-            # Define file paths
+            pfp_path = cache_folder / 'pfp_me.png'
            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}"')
@ -2116,9 +2188,7 @@ def load_user(user_name, name1, user_bio):
    # Clear existing user picture cache
    cache_folder = Path(shared.args.disk_cache_dir)
-    pfp_path = Path(f"{cache_folder}/pfp_me.png")
+    (cache_folder / "pfp_me.png").unlink(missing_ok=True)
    if pfp_path.exists():
        pfp_path.unlink()
    # Generate new picture cache
    picture = generate_user_pfp_cache(user_name)
@ -2542,15 +2612,13 @@ def handle_character_picture_change(picture_path):
    if picture is not None:
        # Save to cache
-        picture.save(Path(f'{cache_folder}/pfp_character.png'), format='PNG')
+        picture.save(cache_folder / 'pfp_character.png', format='PNG')
        thumb = make_thumbnail(picture)
-        thumb.save(Path(f'{cache_folder}/pfp_character_thumb.png'), format='PNG')
+        thumb.save(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_path = Path(f'{cache_folder}/{cache_file}')
+            (cache_folder / cache_file).unlink(missing_ok=True)
            if cache_path.exists():
                cache_path.unlink()
 def handle_mode_change(state):
--- a/modules/html_generator.py
+++ b/modules/html_generator.py
@ -14,6 +14,13 @@ 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 = {}
@ -185,28 +192,29 @@ def process_markdown_content(string):
    if not string:
        return ""
-    # Define unique placeholders for LaTeX asterisks and underscores
+    # Define unique placeholders for LaTeX characters that conflict with markdown
    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 asterisks and underscores in multiple LaTeX formats."""
+        """A replacer function for re.sub to protect markdown-sensitive characters in multiple LaTeX formats."""
        # Check which delimiter group was captured
        if match.group(1) is not None:  # Content from $$...$$
-            content = match.group(1)
+            return protect_latex_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 \[...\]
-            content = match.group(2)
+            return f'\\[{protect_latex_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 \(...\)
-            content = match.group(3)
+            return f'\\({protect_latex_content(match.group(3))}\\)'
-            modified_content = content.replace('*', LATEX_ASTERISK_PLACEHOLDER)
+        elif match.group(4) is not None:  # Content from $...$
-            modified_content = modified_content.replace('_', LATEX_UNDERSCORE_PLACEHOLDER)
+            return f'${protect_latex_content(match.group(4).strip())}$'
            return f'\\({modified_content}\\)'
        return match.group(0)  # Fallback
@ -240,9 +248,7 @@ def process_markdown_content(string):
    string = re.sub(r"(.)```", r"\1\n```", string)
    # Protect asterisks and underscores within all LaTeX blocks before markdown conversion
-    latex_pattern = re.compile(r'((?:^|[\r\n\s])\$\$[^`]*?\$\$)|\\\[(.*?)\\\]|\\\((.*?)\\\)',
+    string = _LATEX_PATTERN.sub(protect_asterisks_underscores_in_latex, string)
                               re.DOTALL)
    string = latex_pattern.sub(protect_asterisks_underscores_in_latex, string)
    result = ''
    is_code = False
@ -306,6 +312,7 @@ 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
@ -10,72 +10,49 @@ 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
-    if quant_method == 'none' or not quant_method:
+    torchao_methods = {
-        return None
+        'torchao-int8wo': 'int8wo',
        'torchao-fp4': 'fp4_e2m1',
        'torchao-float8wo': 'float8wo',
    }
-    # Bitsandbytes 8-bit quantization
+    if quant_method == 'bnb-8bit':
    elif quant_method == 'bnb-8bit':
        return PipelineQuantizationConfig(
            quant_mapping={
-                "transformer": DiffusersBnBConfig(
+                "transformer": DiffusersBnBConfig(load_in_8bit=True),
-                    load_in_8bit=True
+                "text_encoder": TransformersBnBConfig(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(
+                "transformer": DiffusersBnBConfig(**bnb_4bit_kwargs),
-                    load_in_4bit=True,
+                "text_encoder": TransformersBnBConfig(**bnb_4bit_kwargs)
                    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
                )
            }
        )
-    # torchao int8 weight-only
+    elif quant_method in torchao_methods:
-    elif quant_method == 'torchao-int8wo':
+        ao_type = torchao_methods[quant_method]
        return PipelineQuantizationConfig(
            quant_mapping={
-                "transformer": TorchAoConfig("int8wo"),
+                "transformer": TorchAoConfig(ao_type),
-                "text_encoder": TorchAoConfig("int8wo")
+                "text_encoder": TorchAoConfig(ao_type)
            }
        )
    # 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")
            }
        )
@ -152,7 +129,7 @@ def load_image_model(model_name, dtype='bfloat16', attn_backend='sdpa', cpu_offl
        modules = ["transformer", "unet"]
-        # Set attention backend
+        # Set attention backend (diffusers defaults to native/SDPA)
        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
@ -373,6 +373,7 @@ 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:
--- a/modules/models_settings.py
+++ b/modules/models_settings.py
@ -400,14 +400,19 @@ def load_instruction_template(template):
    if template == 'None':
        return ''
-    for filepath in [shared.user_data_dir / 'instruction-templates' / f'{template}.yaml', shared.user_data_dir / 'instruction-templates' / 'Alpaca.yaml']:
+    for name in (template, 'Alpaca'):
-        if filepath.exists():
+        path = shared.user_data_dir / 'instruction-templates' / f'{name}.yaml'
-            break
+        try:
            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
@ -1,6 +1,7 @@
 from pathlib import Path
 from modules import shared, utils
 from modules.utils import sanitize_filename
 from modules.text_generation import get_encoded_length
@ -18,6 +19,7 @@ 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
@ -7,6 +7,7 @@ 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)
@ -72,9 +73,16 @@ def extract_reasoning(text, html_escaped=False):
                if content_pos != -1:
                    content_start = content_pos + len(content_esc)
                else:
-                    # Content tag not present — fall back to content after
+                    # Content tag not present yet.  In GPT-OSS the region
-                    # end_tag (e.g. GPT-OSS tool calls skip the final channel).
+                    # between <|end|> and the content tag contains internal
-                    content_start = end_pos + len(end_esc)
+                    # markup (<|start|>assistant…) that must not be shown.
                    # 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
@ -259,6 +259,7 @@ 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,
@ -363,7 +364,7 @@ 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': 'flash_attention_2',
+    'image_attn_backend': 'sdpa',
    'image_cpu_offload': False,
    'image_compile': False,
    'image_quant': 'none',
--- a/modules/tool_parsing.py
+++ b/modules/tool_parsing.py
@ -27,10 +27,11 @@ 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):
+def streaming_tool_buffer_check(text, markers=None, tool_names=None, check_bare_names=False, partial_match=True):
    '''
    Check whether streaming output should be withheld because it may
    contain tool-call markup.
@ -42,6 +43,10 @@ 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.
@ -59,6 +64,9 @@ 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):
@ -400,6 +408,78 @@ 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.
@ -472,6 +552,11 @@ 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,
@ -504,6 +589,7 @@ 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,
@ -552,9 +638,15 @@ 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)
-    # Offset between original and stripped text, used to map start_pos
+    # Reasoning extraction returns empty content when GPT-OSS internal
-    # back to the original string when returning a prefix.
+    # markup (<|start|>assistant…) follows the thinking block without a
-    reasoning_offset = len(original_answer) - len(answer)
+    # content tag.  Fall back to the full text so tool-call markers can
    # be found.
    if not answer.strip():
        answer = original_answer
        reasoning_offset = 0
    else:
        reasoning_offset = len(original_answer) - len(answer)
    matches = []
    start_pos = None
@ -620,6 +712,8 @@ 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
@ -1,3 +1,4 @@
 import asyncio
 import importlib.util
 import json
@ -55,6 +56,119 @@ 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, elem_classes='tgw-accordion'):
+                    with gr.Accordion(label='Target Modules', open=False):
                        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, elem_classes='tgw-accordion'):
+                    with gr.Accordion(label='Advanced Options', open=False):
                        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/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='rgb(64, 64, 64)',
+        body_text_color='#1a1a1a',
        button_secondary_background_fill="white",
        button_secondary_border_color="var(--border-color-primary)",
        block_title_text_color='*body_text_color',
@ -209,6 +209,7 @@ def list_interface_input_elements():
        'textbox',
        'start_with',
        'selected_tools',
        'mcp_servers',
        'mode',
        'chat_style',
        'chat-instruct_command',
@ -434,6 +435,7 @@ 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 style="width: 100px; display: block" id="hover-element-button">&#9776;</span><div class="hover-menu" id="hover-menu"></div>', elem_id='gr-hover')
+                        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')
                    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'])
@ -105,6 +105,9 @@ 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
+from modules.utils import gradio, sanitize_filename
 inputs = ('textbox-default', 'interface_state')
 outputs = ('output_textbox', 'html-default')
@ -167,6 +167,7 @@ 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
@ -199,6 +200,8 @@ 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
@ -798,6 +798,9 @@ 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
@ -54,7 +54,6 @@ def create_ui():
                            if not shared.args.portable:
                                shared.gradio['ik'] = gr.Checkbox(label="ik", value=shared.args.ik, info='Use ik_llama.cpp instead of upstream llama.cpp.')
                            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['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)
@ -67,13 +66,13 @@ def create_ui():
                            )
                            # Multimodal
-                            with gr.Accordion("Multimodal (vision)", open=False, elem_classes='tgw-accordion') as shared.gradio['mmproj_accordion']:
+                            with gr.Accordion("Multimodal (vision)", open=False) 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, elem_classes='tgw-accordion') as shared.gradio['speculative_decoding_accordion']:
+                            with gr.Accordion("Speculative decoding", open=False) 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')
@ -92,7 +91,7 @@ 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, elem_classes='tgw-accordion'):
+                    with gr.Accordion("See more options", open=False):
                        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.')
@ -109,6 +108,7 @@ 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
+from modules.utils import gradio, sanitize_filename
 _notebook_file_lock = threading.Lock()
 _notebook_auto_save_timer = None
@ -202,6 +202,7 @@ 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
@ -233,6 +234,8 @@ 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"
@ -249,6 +252,7 @@ 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
@ -105,6 +105,9 @@ 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
@ -9,6 +9,7 @@ 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.*
@ -31,8 +32,8 @@ tqdm
 wandb
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.14/gradio-4.37.2+custom.14-py3-none-any.whl
+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.14/gradio_client-1.0.2+custom.14-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
@ -40,10 +41,10 @@ sse-starlette==1.6.5
 tiktoken
 # CUDA wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/llama_cpp_binaries-0.102.0+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-win_amd64.whl; platform_system == "Windows"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/llama_cpp_binaries-0.102.0+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/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/v0.102.0/ik_llama_cpp_binaries-0.102.0+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-win_amd64.whl; platform_system == "Windows"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/ik_llama_cpp_binaries-0.102.0+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-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
 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"
--- a/requirements/full/requirements_amd.txt
+++ b/requirements/full/requirements_amd.txt
@ -7,6 +7,7 @@ fastapi==0.112.4
 huggingface-hub==1.5.*
 jinja2==3.1.6
 markdown
 mcp==1.27.0
 numpy==2.2.*
 pandas
 peft==0.18.*
@ -28,8 +29,8 @@ trafilatura==2.0.0
 wandb
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.14/gradio-4.37.2+custom.14-py3-none-any.whl
+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.14/gradio_client-1.0.2+custom.14-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
@ -37,5 +38,5 @@ sse-starlette==1.6.5
 tiktoken
 # AMD wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/llama_cpp_binaries-0.102.0+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-win_amd64.whl; platform_system == "Windows"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/llama_cpp_binaries-0.102.0+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/v0.110.0/llama_cpp_binaries-0.110.0+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
@ -7,6 +7,7 @@ fastapi==0.112.4
 huggingface-hub==1.5.*
 jinja2==3.1.6
 markdown
 mcp==1.27.0
 numpy==2.2.*
 pandas
 peft==0.18.*
@ -28,8 +29,8 @@ trafilatura==2.0.0
 wandb
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.14/gradio-4.37.2+custom.14-py3-none-any.whl
+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.14/gradio_client-1.0.2+custom.14-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
@ -37,4 +38,4 @@ sse-starlette==1.6.5
 tiktoken
 # Mac wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/llama_cpp_binaries-0.102.0-py3-none-macosx_13_0_x86_64.whl; platform_system == "Darwin"
+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"
--- a/requirements/full/requirements_apple_silicon.txt
+++ b/requirements/full/requirements_apple_silicon.txt
@ -7,6 +7,7 @@ fastapi==0.112.4
 huggingface-hub==1.5.*
 jinja2==3.1.6
 markdown
 mcp==1.27.0
 numpy==2.2.*
 pandas
 peft==0.18.*
@ -28,8 +29,8 @@ trafilatura==2.0.0
 wandb
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.14/gradio-4.37.2+custom.14-py3-none-any.whl
+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.14/gradio_client-1.0.2+custom.14-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
@ -37,4 +38,4 @@ sse-starlette==1.6.5
 tiktoken
 # Mac wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/llama_cpp_binaries-0.102.0-py3-none-macosx_13_0_arm64.whl; platform_system == "Darwin"
+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"
--- a/requirements/full/requirements_cpu_only.txt
+++ b/requirements/full/requirements_cpu_only.txt
@ -7,6 +7,7 @@ fastapi==0.112.4
 huggingface-hub==1.5.*
 jinja2==3.1.6
 markdown
 mcp==1.27.0
 numpy==2.2.*
 pandas
 peft==0.18.*
@ -28,8 +29,8 @@ trafilatura==2.0.0
 wandb
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.14/gradio-4.37.2+custom.14-py3-none-any.whl
+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.14/gradio_client-1.0.2+custom.14-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
@ -37,7 +38,7 @@ sse-starlette==1.6.5
 tiktoken
 # llama.cpp (CPU only)
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/llama_cpp_binaries-0.102.0+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-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/llama_cpp_binaries-0.102.0+cpu-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+cpu-py3-none-win_amd64.whl; platform_system == "Windows"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/ik_llama_cpp_binaries-0.102.0+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-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/ik_llama_cpp_binaries-0.102.0+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-win_amd64.whl; platform_system == "Windows"
--- a/requirements/full/requirements_nowheels.txt
+++ b/requirements/full/requirements_nowheels.txt
@ -7,6 +7,7 @@ fastapi==0.112.4
 huggingface-hub==1.5.*
 jinja2==3.1.6
 markdown
 mcp==1.27.0
 numpy==2.2.*
 pandas
 peft==0.18.*
@ -28,8 +29,8 @@ trafilatura==2.0.0
 wandb
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.14/gradio-4.37.2+custom.14-py3-none-any.whl
+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.14/gradio_client-1.0.2+custom.14-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
--- a/requirements/portable/requirements.txt
+++ b/requirements/portable/requirements.txt
@ -3,6 +3,7 @@ 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.*
@ -14,8 +15,8 @@ trafilatura==2.0.0
 tqdm
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.14/gradio-4.37.2+custom.14-py3-none-any.whl
+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.14/gradio_client-1.0.2+custom.14-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
@ -23,5 +24,5 @@ sse-starlette==1.6.5
 tiktoken
 # CUDA wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/llama_cpp_binaries-0.102.0+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-win_amd64.whl; platform_system == "Windows"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/llama_cpp_binaries-0.102.0+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/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_amd.txt
+++ b/requirements/portable/requirements_amd.txt
@ -3,6 +3,7 @@ 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.*
@ -14,8 +15,8 @@ trafilatura==2.0.0
 tqdm
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.14/gradio-4.37.2+custom.14-py3-none-any.whl
+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.14/gradio_client-1.0.2+custom.14-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
@ -23,5 +24,5 @@ sse-starlette==1.6.5
 tiktoken
 # AMD wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/llama_cpp_binaries-0.102.0+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-win_amd64.whl; platform_system == "Windows"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/llama_cpp_binaries-0.102.0+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/v0.110.0/llama_cpp_binaries-0.110.0+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
@ -3,6 +3,7 @@ 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.*
@ -14,8 +15,8 @@ trafilatura==2.0.0
 tqdm
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.14/gradio-4.37.2+custom.14-py3-none-any.whl
+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.14/gradio_client-1.0.2+custom.14-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
@ -23,4 +24,4 @@ sse-starlette==1.6.5
 tiktoken
 # Mac wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/llama_cpp_binaries-0.102.0-py3-none-macosx_13_0_x86_64.whl; platform_system == "Darwin"
+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"
--- a/requirements/portable/requirements_apple_silicon.txt
+++ b/requirements/portable/requirements_apple_silicon.txt
@ -3,6 +3,7 @@ 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.*
@ -14,8 +15,8 @@ trafilatura==2.0.0
 tqdm
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.14/gradio-4.37.2+custom.14-py3-none-any.whl
+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.14/gradio_client-1.0.2+custom.14-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
@ -23,4 +24,4 @@ sse-starlette==1.6.5
 tiktoken
 # Mac wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/llama_cpp_binaries-0.102.0-py3-none-macosx_13_0_arm64.whl; platform_system == "Darwin"
+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"
--- a/requirements/portable/requirements_cpu_only.txt
+++ b/requirements/portable/requirements_cpu_only.txt
@ -3,6 +3,7 @@ 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.*
@ -14,8 +15,8 @@ trafilatura==2.0.0
 tqdm
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.14/gradio-4.37.2+custom.14-py3-none-any.whl
+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.14/gradio_client-1.0.2+custom.14-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
@ -23,5 +24,5 @@ sse-starlette==1.6.5
 tiktoken
 # llama.cpp (CPU only)
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/llama_cpp_binaries-0.102.0+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-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/llama_cpp_binaries-0.102.0+cpu-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+cpu-py3-none-win_amd64.whl; platform_system == "Windows"
--- a/requirements/portable/requirements_cuda131.txt
+++ b/requirements/portable/requirements_cuda131.txt
@ -3,6 +3,7 @@ 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.*
@ -14,8 +15,8 @@ trafilatura==2.0.0
 tqdm
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.14/gradio-4.37.2+custom.14-py3-none-any.whl
+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.14/gradio_client-1.0.2+custom.14-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
@ -23,5 +24,5 @@ sse-starlette==1.6.5
 tiktoken
 # CUDA wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/llama_cpp_binaries-0.102.0+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-win_amd64.whl; platform_system == "Windows"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/llama_cpp_binaries-0.102.0+cu131-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/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_ik.txt
+++ b/requirements/portable/requirements_ik.txt
@ -3,6 +3,7 @@ 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.*
@ -14,8 +15,8 @@ trafilatura==2.0.0
 tqdm
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.14/gradio-4.37.2+custom.14-py3-none-any.whl
+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.14/gradio_client-1.0.2+custom.14-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
@ -23,5 +24,5 @@ sse-starlette==1.6.5
 tiktoken
 # CUDA wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/ik_llama_cpp_binaries-0.102.0+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-win_amd64.whl; platform_system == "Windows"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/ik_llama_cpp_binaries-0.102.0+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-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
@ -3,6 +3,7 @@ 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.*
@ -14,8 +15,8 @@ trafilatura==2.0.0
 tqdm
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.14/gradio-4.37.2+custom.14-py3-none-any.whl
+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.14/gradio_client-1.0.2+custom.14-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
@ -23,5 +24,5 @@ sse-starlette==1.6.5
 tiktoken
 # ik_llama.cpp (CPU only)
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/ik_llama_cpp_binaries-0.102.0+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-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/ik_llama_cpp_binaries-0.102.0+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-win_amd64.whl; platform_system == "Windows"
--- a/requirements/portable/requirements_ik_cuda131.txt
+++ b/requirements/portable/requirements_ik_cuda131.txt
@ -3,6 +3,7 @@ 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.*
@ -14,8 +15,8 @@ trafilatura==2.0.0
 tqdm
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.14/gradio-4.37.2+custom.14-py3-none-any.whl
+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.14/gradio_client-1.0.2+custom.14-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
@ -23,5 +24,5 @@ sse-starlette==1.6.5
 tiktoken
 # CUDA wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/ik_llama_cpp_binaries-0.102.0+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-win_amd64.whl; platform_system == "Windows"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/ik_llama_cpp_binaries-0.102.0+cu131-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/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
@ -3,6 +3,7 @@ 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.*
@ -14,8 +15,8 @@ trafilatura==2.0.0
 tqdm
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.14/gradio-4.37.2+custom.14-py3-none-any.whl
+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.14/gradio_client-1.0.2+custom.14-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
--- a/requirements/portable/requirements_vulkan.txt
+++ b/requirements/portable/requirements_vulkan.txt
@ -3,6 +3,7 @@ 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.*
@ -14,8 +15,8 @@ trafilatura==2.0.0
 tqdm
 # Gradio
-https://github.com/oobabooga/gradio/releases/download/4.37.2-custom.14/gradio-4.37.2+custom.14-py3-none-any.whl
+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.14/gradio_client-1.0.2+custom.14-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
@ -23,5 +24,5 @@ sse-starlette==1.6.5
 tiktoken
 # Vulkan wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/llama_cpp_binaries-0.102.0+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-win_amd64.whl; platform_system == "Windows"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.102.0/llama_cpp_binaries-0.102.0+vulkan-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.110.0/llama_cpp_binaries-0.110.0+vulkan-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
Author	SHA1	Message	Date
oobabooga	9dcf574160	Merge pull request #7471 from oobabooga/dev Merge dev branch	2026-04-06 21:54:02 -03:00
oobabooga	e18f32cba7	Remove hardcoded trust_remote_code=True in embedding loader	2026-04-06 17:47:50 -07:00
oobabooga	778e1c4d52	Update llama.cpp/ik_llama.cpp	2026-04-06 17:04:49 -07:00
oobabooga	775c913de2	Fix crash when truncating prompts with tool call messages	2026-04-06 14:13:01 -07:00
oobabooga	cb511928e2	Fix GPT-OSS tag leak during streaming between thinking and tool calls	2026-04-06 12:06:28 -07:00
oobabooga	193424cc93	API: Fix IPv6 address formatting	2026-04-06 10:07:52 -07:00
oobabooga	c26ffdd24c	API: add instruction_template support to the model load endpoint	2026-04-06 07:02:53 -07:00
oobabooga	4d6230a944	Follow-up to `d78fc46114`	2026-04-06 06:48:48 -07:00
oobabooga	7b2f15e34a	Minor change after `b1d06dcf96`	2026-04-05 21:16:32 -07:00
oobabooga	05e4842033	Fix image generation: default to SDPA attention backend	2026-04-05 20:03:06 -07:00
oobabooga	b1d06dcf96	UI: Add MCP server support	2026-04-05 19:46:01 -07:00
oobabooga	abc3487f4d	UI: Move cpu-moe checkbox to extra flags (no longer useful now that --fit exists)	2026-04-05 18:24:26 -07:00
oobabooga	223dd4b801	UI: Hide spin buttons on number inputs	2026-04-05 18:22:50 -07:00
oobabooga	f8db23b362	Call ik portable build folders text-generation-webui-ik-version	2026-04-05 17:12:28 -07:00
oobabooga	d78fc46114	Fix "address already in use" on server restart (Linux/macOS)	2026-04-05 16:42:27 -07:00
oobabooga	422f42ca7f	Pre-compile LaTeX regex in html_generator.py	2026-04-04 23:51:15 -07:00
oobabooga	544fcb0b7f	Simplify modules/image_models.py	2026-04-04 23:29:57 -07:00
oobabooga	c63a79ee48	Image generation: Embed generation metadata in API image responses	2026-04-04 23:15:14 -07:00
oobabooga	9805ddcde9	Update the custom gradio wheels	2026-04-04 21:34:09 -07:00
oobabooga	91f9b01516	UI: Minor change	2026-04-04 21:13:20 -07:00
oobabooga	1f49a64e1a	UI: Improve blockquote border width and color	2026-04-04 20:44:37 -07:00
oobabooga	e8b31c063a	UI: Soften message action icons in light mode	2026-04-04 20:38:31 -07:00
oobabooga	ee917cd5ed	UI: Make table and hr borders more subtle	2026-04-04 20:35:27 -07:00
oobabooga	dfd8ec9c49	UI: Make accordion outline styling global	2026-04-04 20:13:20 -07:00
oobabooga	0c033caf0e	UI: Reduce spacing above chat input	2026-04-04 20:09:28 -07:00
oobabooga	1b403a4ffa	UI: Fix inline LaTeX rendering by protecting $...$ from markdown (closes #7423 )	2026-04-04 19:33:05 -07:00
oobabooga	8cb7fe9c47	UI: Improve message action icon visibility in light mode	2026-04-04 19:14:17 -07:00
oobabooga	41bce3f4de	UI: Improve scrollbars style	2026-04-04 19:07:36 -07:00
oobabooga	ffea8f282e	UI: Improve message text contrast	2026-04-04 18:53:13 -07:00
oobabooga	7fed60f90a	UI: Improve the hover menu looks	2026-04-04 18:29:36 -07:00
oobabooga	2eef90a323	API: Remove deprecated "settings" parameter from model load endpoint	2026-04-04 11:00:14 -07:00
oobabooga	9183dc444e	API: Fix loader args leaking between sequential model loads	2026-04-04 10:48:53 -07:00
oobabooga	e0ad4e60df	UI: Fix tool buffer check truncating visible text at end of generation	2026-04-04 09:57:07 -07:00
oobabooga	16af11f868	Update README	2026-04-04 04:22:37 -07:00
oobabooga	54b2f39c78	Cleanup modules/chat.py	2026-04-03 22:07:21 -07:00
oobabooga	b5afecc63b	Merge pull request #7464 from oobabooga/dev Merge dev branch	2026-04-04 00:56:02 -03:00
oobabooga	2fbaee58cd	Add Windows + ROCm portable builds	2026-04-03 20:54:28 -07:00
oobabooga	62e67adb55	Merge pull request #7463 from oobabooga/dev Merge dev branch	2026-04-03 20:58:32 -03:00
oobabooga	fc35acab9b	API: Fix tool call parser crash on non-dict JSON output	2026-04-03 16:56:15 -07:00
oobabooga	8ecdb41078	fix(security): sanitize filenames in all prompt file operations (CWE-22) (#7462 ) --------- Co-authored-by: Alex Chen <ffulbtech@gmail.com>	2026-04-03 19:36:50 -03:00
oobabooga	5fb8c4fbd6	Update the custom gradio wheels	2026-04-03 11:02:00 -07:00
oobabooga	0050a33f37	Merge pull request #7461 from oobabooga/dev Merge dev branch	2026-04-03 14:07:42 -03:00
oobabooga	6b66da84d2	Update the custom gradio wheels	2026-04-03 10:01:51 -07:00
oobabooga	8e8e1ba898	Update the custom gradio wheels	2026-04-03 09:50:15 -07:00
oobabooga	131a9a0140	Update llama.cpp	2026-04-03 09:15:03 -07:00
oobabooga	95d6c53e13	Revert "API: Add warning about vanilla llama-server not supporting prompt logprobs + instructions" This reverts commit `42dfcdfc5b`.	2026-04-03 07:30:48 -07:00
oobabooga	8bba9ecc3f	Update the custom gradio wheels	2026-04-03 05:58:05 -07:00
oobabooga	66d1a22c73	Fix crash when no model is selected (None passed to resolve_model_path)	2026-04-03 05:56:36 -07:00
oobabooga	000d776967	Revert "llama.cpp: Disable jinja by default (we use Python jinja, not cpp jinja)" This reverts commit `a1cb5b5dc0`.	2026-04-03 05:49:03 -07:00
oobabooga	a1cb5b5dc0	llama.cpp: Disable jinja by default (we use Python jinja, not cpp jinja) This was causing template compilation issues with qwen models.	2026-04-02 21:56:40 -07:00
oobabooga	b11379f328	Merge pull request #7455 from oobabooga/dev Merge dev branch	2026-04-03 00:50:06 -03:00
oobabooga	42dfcdfc5b	API: Add warning about vanilla llama-server not supporting prompt logprobs + instructions	2026-04-02 20:46:27 -07:00
oobabooga	6e2b70bde6	Add Gemma 4 tool-calling support	2026-04-02 20:26:27 -07:00