Merge pull request #6984 from oobabooga/dev

Merge dev branch
2025-06-07 06:06:20 -04:00 · 2025-05-16 17:13:26 -03:00 · 2025-05-16 17:13:26 -03:00 · dc3094549e
commit dc3094549e
parent ace8afb825 e4d3f4449d
65 changed files with 955 additions and 406 deletions
--- a/.github/workflows/build-portable-release-cuda.yml
+++ b/.github/workflows/build-portable-release-cuda.yml
@ -102,6 +102,8 @@ jobs:
        shell: bash
        run: |
            rm -rf .git cmd* update_wizard* Colab-TextGen-GPU.ipynb docker
+            allowed=("character_bias" "gallery" "openai" "sd_api_pictures")
+            find extensions/ -mindepth 1 -maxdepth 1 -type d | grep -v -E "$(printf '%s|' "${allowed[@]}" | sed 's/|$//')" | xargs rm -rf

            # Define common variables
            CUDA_VERSION="${{ matrix.cuda }}"
--- a/.github/workflows/build-portable-release-vulkan.yml
+++ b/.github/workflows/build-portable-release-vulkan.yml
@ -101,6 +101,8 @@ jobs:
        shell: bash
        run: |
            rm -rf .git cmd* update_wizard* Colab-TextGen-GPU.ipynb docker
+            allowed=("character_bias" "gallery" "openai" "sd_api_pictures")
+            find extensions/ -mindepth 1 -maxdepth 1 -type d | grep -v -E "$(printf '%s|' "${allowed[@]}" | sed 's/|$//')" | xargs rm -rf

            # Define common variables
            AVX_SUPPORT="${{ matrix.avx }}"
--- a/.github/workflows/build-portable-release.yml
+++ b/.github/workflows/build-portable-release.yml
@ -101,6 +101,8 @@ jobs:
        shell: bash
        run: |
            rm -rf .git cmd* update_wizard* Colab-TextGen-GPU.ipynb docker
+            allowed=("character_bias" "gallery" "openai" "sd_api_pictures")
+            find extensions/ -mindepth 1 -maxdepth 1 -type d | grep -v -E "$(printf '%s|' "${allowed[@]}" | sed 's/|$//')" | xargs rm -rf

            # Define common variables
            AVX_SUPPORT="${{ matrix.avx }}"
--- a/README.md
+++ b/README.md
@ -22,7 +22,7 @@ Its goal is to become the [AUTOMATIC1111/stable-diffusion-webui](https://github.
 - Free-form text generation in the Default/Notebook tabs without being limited to chat turns. You can send formatted conversations from the Chat tab to these.
 - Multiple sampling parameters and generation options for sophisticated text generation control.
 - Switch between different models easily in the UI without restarting, with fine control over settings.
- OpenAI-compatible API with Chat and Completions endpoints – see [examples](https://github.com/oobabooga/text-generation-webui/wiki/12-%E2%80%90-OpenAI-API#examples).
+- OpenAI-compatible API with Chat and Completions endpoints, including tool-calling support – see [examples](https://github.com/oobabooga/text-generation-webui/wiki/12-%E2%80%90-OpenAI-API#examples).
 - 100% offline and private, with zero telemetry, external resources, or remote update requests.
 - Extension support, with numerous built-in and user-contributed extensions available. See the [wiki](https://github.com/oobabooga/text-generation-webui/wiki/07-%E2%80%90-Extensions) and [extensions directory](https://github.com/oobabooga/text-generation-webui-extensions) for details.

@ -44,7 +44,7 @@ Download from: https://github.com/oobabooga/text-generation-webui/releases

 To restart the web UI later, just run the same `start_` script. If you need to reinstall, delete the `installer_files` folder created during setup and run the script again.

-You can use command-line flags, like `./start_linux.sh --help`, or add them to `user_data/CMD_FLAGS.txt` (such as `--api` to enable API use). To update the project, run `update_wizard_linux.sh`, `update_wizard_windows.bat`, `update_wizard_macos.sh`, or `update_wizard_wsl.bat`.
+You can use command-line flags, like `./start_linux.sh --help`, or add them to `user_data/CMD_FLAGS.txt` (such as `--api` to enable API use). To update the project, run `update_wizard_linux.sh`, `update_wizard_windows.bat`, or `update_wizard_macos.sh`.

 <details>
 <summary>
@ -55,12 +55,12 @@ Setup details and information about installing manually

 The script uses Miniconda to set up a Conda environment in the `installer_files` folder.

-If you ever need to install something manually in the `installer_files` environment, you can launch an interactive shell using the cmd script: `cmd_linux.sh`, `cmd_windows.bat`, `cmd_macos.sh`, or `cmd_wsl.bat`.
+If you ever need to install something manually in the `installer_files` environment, you can launch an interactive shell using the cmd script: `cmd_linux.sh`, `cmd_windows.bat`, or `cmd_macos.sh`.

 * There is no need to run any of those scripts (`start_`, `update_wizard_`, or `cmd_`) as admin/root.
 * To install the requirements for extensions, you can use the `extensions_reqs` script for your OS. At the end, this script will install the main requirements for the project to make sure that they take precedence in case of version conflicts.
 * For additional instructions about AMD and WSL setup, consult [the documentation](https://github.com/oobabooga/text-generation-webui/wiki).
-* For automated installation, you can use the `GPU_CHOICE`, `USE_CUDA118`, `LAUNCH_AFTER_INSTALL`, and `INSTALL_EXTENSIONS` environment variables. For instance: `GPU_CHOICE=A USE_CUDA118=FALSE LAUNCH_AFTER_INSTALL=FALSE INSTALL_EXTENSIONS=TRUE ./start_linux.sh`.
+* For automated installation, you can use the `GPU_CHOICE`, `LAUNCH_AFTER_INSTALL`, and `INSTALL_EXTENSIONS` environment variables. For instance: `GPU_CHOICE=A LAUNCH_AFTER_INSTALL=FALSE INSTALL_EXTENSIONS=TRUE ./start_linux.sh`.

 ### Manual installation using Conda

@ -90,7 +90,7 @@ conda activate textgen
 |--------|---------|---------|
 | Linux/WSL | NVIDIA | `pip3 install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu124` |
 | Linux/WSL | CPU only | `pip3 install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cpu` |
-| Linux | AMD | `pip3 install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/rocm6.1` |
+| Linux | AMD | `pip3 install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/rocm6.2.4` |
 | MacOS + MPS | Any | `pip3 install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0` |
 | Windows | NVIDIA | `pip3 install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu124` |
 | Windows | CPU only | `pip3 install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0` |
--- a/css/chat_style-Dark.css
+++ b/css/chat_style-Dark.css
@ -1,7 +1,9 @@
 .message {
    display: grid;
+    align-items: start;
    grid-template-columns: 60px minmax(0, 1fr);
-    padding-bottom: 28px;
+    padding-bottom: 14px;
+    padding-top: 14px;
    font-size: 18px;
    font-family: Roboto, Arial, sans-serif; /* Modern font */
    line-height: 1.5;
@ -102,6 +104,7 @@
@media screen and (width <= 688px) {
    .message {
        display: grid;
+        align-items: start;
        grid-template-columns: 60px minmax(0, 1fr);
        padding-bottom: 25px;
        font-size: 15px;
--- a/css/chat_style-TheEncrypted777.css
+++ b/css/chat_style-TheEncrypted777.css
@ -2,8 +2,10 @@

 .message {
    display: grid;
+    align-items: start;
    grid-template-columns: 60px minmax(0, 1fr);
-    padding-bottom: 28px;
+    padding-bottom: 14px;
+    padding-top: 14px;
    font-size: 18px;
    font-family: 'Noto Sans', Arial, sans-serif;
    line-height: 1.428571429;
@ -100,6 +102,7 @@
@media screen and (width <= 688px) {
    .message {
        display: grid;
+        align-items: start;
        grid-template-columns: 60px minmax(0, 1fr);
        padding-bottom: 25px;
        font-size: 15px;
--- a/css/chat_style-cai-chat-square.css
+++ b/css/chat_style-cai-chat-square.css
@ -16,6 +16,7 @@
 }

 .message {
-    padding-bottom: 2em;
+    padding-bottom: 1em;
+    padding-top: 1em;
    grid-template-columns: 70px minmax(0, 1fr);
 }
--- a/css/chat_style-cai-chat.css
+++ b/css/chat_style-cai-chat.css
@ -1,7 +1,9 @@
 .message {
    display: grid;
+    align-items: start;
    grid-template-columns: 60px minmax(0, 1fr);
-    padding-bottom: 2em;
+    padding-bottom: 1em;
+    padding-top: 1em;
    font-size: 15px;
    font-family: 'Noto Sans', Helvetica, Arial, sans-serif;
    line-height: 22.5px !important;
--- a/css/chat_style-messenger.css
+++ b/css/chat_style-messenger.css
@ -1,5 +1,6 @@
 .message {
-    padding-bottom: 25px;
+    padding-bottom: 12.5px;
+    padding-top: 12.5px;
    font-size: 15px;
    font-family: 'Noto Sans', Helvetica, Arial, sans-serif;
    line-height: 1.428571429;
--- a/css/chat_style-wpp.css
+++ b/css/chat_style-wpp.css
@ -1,5 +1,6 @@
 .message {
-    padding-bottom: 25px;
+    padding-bottom: 12.5px;
+    padding-top: 12.5px;
    font-size: 15px;
    font-family: 'Noto Sans', Helvetica, Arial, sans-serif;
    line-height: 1.428571429;
--- a/css/html_instruct_style.css
+++ b/css/html_instruct_style.css
@ -8,10 +8,6 @@
    padding-top: 0 !important;
 }

-.chat > .messages > :last-child {
-    margin-bottom: 1.7rem !important;
-}
-
 .chat .message-body p, .chat .message-body li {
    font-size: 1rem !important;
    line-height: 28px !important;
@ -46,7 +42,7 @@
 }

 .chat .user-message {
-    background: #f5f5f5;
+    background: #f3f4f6;
    padding: 1.5rem 1rem;
    padding-bottom: 2rem;
    border-radius: 0;
@ -61,16 +57,16 @@
 }

 .dark .chat .user-message {
-    background: transparent;
+    background: var(--light-gray);
 }

 .dark .chat .assistant-message {
-    background: var(--light-gray);
+    background: transparent;
 }

 .chat .user-message .text,
 .chat .assistant-message .text {
-    max-width: 645px;
+    max-width: 700px;
    margin-left: auto;
    margin-right: auto;
 }
--- a/css/main.css
+++ b/css/main.css
@ -2,7 +2,7 @@
    --darker-gray: #202123;
    --dark-gray: #343541;
    --light-gray: #444654;
-    --light-theme-gray: #f5f5f5;
+    --light-theme-gray: #f9fbff;
    --border-color-dark: #525252;
    --header-width: 112px;
    --selected-item-color-dark: #32333e;
@ -389,7 +389,7 @@ div.svelte-362y77>*, div.svelte-362y77>.form>* {
 .chat {
    margin-left: auto;
    margin-right: auto;
-    min-height: var(--chat-height);
+    flex: 1;
    overflow-y: auto;
    display: flex;
    flex-direction: column;
@ -401,10 +401,9 @@ div.svelte-362y77>*, div.svelte-362y77>.form>* {
 }

 .chat-parent {
-    height: calc(100dvh - 98px - var(--input-delta));
+    flex: 1;
    overflow: auto !important;
    border-radius: 0 !important;
-    margin-bottom: var(--input-delta) !important;
 }

 .chat-parent .prose {
@ -421,13 +420,13 @@ div.svelte-362y77>*, div.svelte-362y77>.form>* {
 }

 .chat-parent.bigchat {
-    height: calc(100dvh - 98px - var(--input-delta)) !important;
-    margin-bottom: var(--input-delta) !important;
+    flex: 1;
 }

 .chat > .messages {
    display: flex;
    flex-direction: column;
+    min-height: calc(100vh - 102px);
 }

 .chat > .messages > :first-child {
@ -546,7 +545,7 @@ div.svelte-362y77>*, div.svelte-362y77>.form>* {
    border-radius: 5px;
    font-size: 82%;
    padding: 1px 3px;
-    background: white !important;
+    background: #f3f4f6 !important;
    color: #1f2328;
 }

@ -560,18 +559,17 @@ div.svelte-362y77>*, div.svelte-362y77>.form>* {
    padding: 15px;
 }

-.message-body :not(pre) > code::before {
-    content: "`";
-}
-
-.message-body :not(pre) > code::after {
-    content: "`";
-}
-
 .message-body :not(pre) > code {
    white-space: normal !important;
    font-weight: bold;
-    font-family: unset;
+    font-size: 0.95em;
+    font-family: Menlo,"Roboto Mono","Courier New",Courier,monospace,Inter,sans-serif;
+    padding: .15rem .3rem;
+    background-color: #ececec;
+}
+
+.dark .message-body :not(pre) > code {
+    background-color: rgb(255 255 255 / 10%);
 }

 #chat-input {
@ -582,7 +580,10 @@ div.svelte-362y77>*, div.svelte-362y77>.form>* {
 }

 #chat-input textarea {
+    background: #f3f4f6;
    padding: 0.65rem 2.5rem;
+    border: 0;
+    box-shadow: 0;
 }

 #chat-input textarea::placeholder {
@ -603,8 +604,6 @@ div.svelte-362y77>*, div.svelte-362y77>.form>* {
 }

 .chat-input-positioned {
-    position: absolute;
-    bottom: 0;
    max-width: 54rem;
    left: 50%;
    transform: translateX(-50%);
@ -744,7 +743,7 @@ div.svelte-362y77>*, div.svelte-362y77>.form>* {

 .hover-menu button {
    width: 100%;
-    background: transparent !important;
+    background: white !important;
    border-radius: 0 !important;
    justify-content: space-between;
    margin: 0 !important;
@ -760,7 +759,11 @@ div.svelte-362y77>*, div.svelte-362y77>.form>* {
 }

 .hover-menu button:hover {
-    background: var(--button-secondary-background-fill-hover) !important;
+    background: #dbeafe !important;
+}
+
+.dark .hover-menu button:hover {
+    background: var(--selected-item-color-dark) !important;
 }

 .transparent-substring {
@ -789,6 +792,8 @@ div.svelte-362y77>*, div.svelte-362y77>.form>* {
 }

 #chat-input-container {
+    display: flex;
+    flex-direction: column;
    min-width: 0 !important;
 }

@ -798,9 +803,8 @@ div.svelte-362y77>*, div.svelte-362y77>.form>* {
 }

 #chat-input-row {
-    padding-bottom: 1.5em;
-    padding-left: 1rem;
-    padding-right: 1rem;
+    padding: 1rem;
+    padding-top: 0;
 }

 #chat-input-row.bigchat {
@ -808,22 +812,17 @@ div.svelte-362y77>*, div.svelte-362y77>.form>* {
 }

 #chat-col {
-    padding-bottom: 100px;
+    height: 100dvh;
+    display: flex;
+    flex-direction: column;
+    padding-bottom: 0;
+    gap: 0;
 }

@media screen and (width <= 924px) {
    #chat-col {
-        padding-bottom: 100px;
        margin-top: 32px;
-        position: relative; /* Ensure positioning for the pseudo-element */
-    }
-
-    .chat-parent {
-        height: calc(100dvh - 98px - var(--input-delta) - 32px);
-    }
-
-    .chat-parent.bigchat {
-        height: calc(100dvh - 98px - var(--input-delta) - 32px) !important;
+        height: calc(100dvh - 32px);
    }
 }

@ -985,6 +984,11 @@ div.svelte-362y77>*, div.svelte-362y77>.form>* {
    cursor: pointer;
 }

+#past-chats .selected,
+#past-chats label:hover {
+    background-color: #dbeafe !important;
+}
+
 #past-chats-buttons,
 #delete-chat-row,
 #rename-row {
@ -993,7 +997,6 @@ div.svelte-362y77>*, div.svelte-362y77>.form>* {
    gap: 9px;
 }

-
 #past-chats-row,
 #chat-controls {
    width: 260px;
@ -1111,12 +1114,9 @@ div.svelte-362y77>*, div.svelte-362y77>.form>* {
    color: #9ca3af;
 }

-.dark .hover-menu {
-    background-color: var(--darker-gray);
-}
-
 .dark .hover-menu button {
    border-color: var(--border-color-primary);
+    background-color: var(--darker-gray) !important;
 }

 .dark #chat-controls,
@ -1125,8 +1125,8 @@ div.svelte-362y77>*, div.svelte-362y77>.form>* {
    border: 0 !important;
 }

-.dark #past-chats .selected,
-.dark #past-chats label:hover {
+.dark gradio-app .gradio-container.gradio-container-4-37-2 .contain #past-chats .selected,
+.dark gradio-app .gradio-container.gradio-container-4-37-2 .contain #past-chats label:hover {
    background-color: var(--selected-item-color-dark) !important;
 }

@ -1163,7 +1163,7 @@ div.svelte-362y77>*, div.svelte-362y77>.form>* {
 }

 .header_bar button.selected {
-    background: #E0E0E0;
+    background: #dbeafe;
 }

 #chat-controls,
@ -1382,3 +1382,19 @@ div.svelte-362y77>*, div.svelte-362y77>.form>* {
    50% { opacity: 1; }
    100% { opacity: 0.6; }
 }
+
+strong {
+    font-weight: bold;
+}
+
+.min.svelte-1ybaih5 {
+    min-height: 0;
+}
+
+#vram-info .value {
+    color: #008d00;
+}
+
+.dark #vram-info .value {
+    color: #07ff07;
+}
--- a/docker/amd/docker-compose.yml
+++ b/docker/amd/docker-compose.yml
@ -22,7 +22,7 @@ services:
        TORCH_CUDA_ARCH_LIST: ${TORCH_CUDA_ARCH_LIST:-7.5} 
        BUILD_EXTENSIONS: ${BUILD_EXTENSIONS:-}
        APP_GID: ${APP_GID:-6972} 
-        APP_UID: ${APP_UID-6972} 
+        APP_UID: ${APP_UID:-6972} 
    env_file: .env
    user: "${APP_RUNTIME_UID:-6972}:${APP_RUNTIME_GID:-6972}"
    ports:
--- a/docker/cpu/docker-compose.yml
+++ b/docker/cpu/docker-compose.yml
@ -22,7 +22,7 @@ services:
        TORCH_CUDA_ARCH_LIST: ${TORCH_CUDA_ARCH_LIST:-7.5} 
        BUILD_EXTENSIONS: ${BUILD_EXTENSIONS:-}
        APP_GID: ${APP_GID:-6972} 
-        APP_UID: ${APP_UID-6972} 
+        APP_UID: ${APP_UID:-6972} 
    env_file: .env
    user: "${APP_RUNTIME_UID:-6972}:${APP_RUNTIME_GID:-6972}"
    ports:
--- a/docker/intel/docker-compose.yml
+++ b/docker/intel/docker-compose.yml
@ -22,7 +22,7 @@ services:
        TORCH_CUDA_ARCH_LIST: ${TORCH_CUDA_ARCH_LIST:-7.5} 
        BUILD_EXTENSIONS: ${BUILD_EXTENSIONS:-}
        APP_GID: ${APP_GID:-6972} 
-        APP_UID: ${APP_UID-6972} 
+        APP_UID: ${APP_UID:-6972} 
    env_file: .env
    user: "${APP_RUNTIME_UID:-6972}:${APP_RUNTIME_GID:-6972}"
    ports:
--- a/docker/nvidia/Dockerfile
+++ b/docker/nvidia/Dockerfile
@ -14,7 +14,7 @@ WORKDIR /home/app/
 RUN git clone https://github.com/oobabooga/text-generation-webui.git 
 WORKDIR /home/app/text-generation-webui
 RUN GPU_CHOICE=A LAUNCH_AFTER_INSTALL=FALSE INSTALL_EXTENSIONS=TRUE ./start_linux.sh --verbose
-COPY CMD_FLAGS.txt /home/app/text-generation-webui/
+COPY /user_data/CMD_FLAGS.txt /home/app/text-generation-webui/user_data
 EXPOSE ${CONTAINER_PORT:-7860} ${CONTAINER_API_PORT:-5000} ${CONTAINER_API_STREAM_PORT:-5005}
 WORKDIR /home/app/text-generation-webui
 # set umask to ensure group read / write at runtime
--- a/docker/nvidia/docker-compose.yml
+++ b/docker/nvidia/docker-compose.yml
@ -22,7 +22,7 @@ services:
        TORCH_CUDA_ARCH_LIST: ${TORCH_CUDA_ARCH_LIST:-7.5} 
        BUILD_EXTENSIONS: ${BUILD_EXTENSIONS:-}
        APP_GID: ${APP_GID:-6972} 
-        APP_UID: ${APP_UID-6972} 
+        APP_UID: ${APP_UID:-6972} 
    env_file: .env
    user: "${APP_RUNTIME_UID:-6972}:${APP_RUNTIME_GID:-6972}"
    ports:
@ -31,17 +31,7 @@ services:
    stdin_open: true
    tty: true
    volumes:
-      - ./cache:/home/app/text-generation-webui/cache
-      - ./characters:/home/app/text-generation-webui/characters
-      - ./extensions:/home/app/text-generation-webui/extensions
-      - ./loras:/home/app/text-generation-webui/loras
-      - ./logs:/home/app/text-generation-webui/logs
-      - ./models:/home/app/text-generation-webui/models
-      - ./presets:/home/app/text-generation-webui/presets
-      - ./prompts:/home/app/text-generation-webui/prompts
-      - ./softprompts:/home/app/text-generation-webui/softprompts
-      - ./training:/home/app/text-generation-webui/training
-      - ./cloudflared:/etc/cloudflared
+      - ./user_data:/home/app/text-generation-webui/user_data
    deploy:
      resources:
        reservations:
--- a/docs/12
+++ b/docs/12
@ -257,6 +257,85 @@ headers = {

 in any of the examples above.

+#### Tool/Function Calling Example
+
+You need to use a model with tools support. The prompt will be automatically formatted using the model's Jinja2 template.
+
+Request:
+
+```
+curl http://127.0.0.1:5000/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "messages": [
+      {
+        "role": "system",
+        "content": "You are a helpful assistant."
+      },
+      {
+        "role": "user",
+        "content": "What time is it currently in New York City?"
+      }
+    ],
+    "tools": [
+      {
+        "type": "function",
+        "function": {
+          "name": "get_current_time",
+          "description": "Get current time in a specific timezones",
+          "parameters": {
+            "type": "object",
+            "required": ["timezone"],
+            "properties": {
+              "timezone": {
+                "type": "string",
+                "description": "IANA timezone name (e.g., America/New_York, Europe/London). Use Europe/Berlin as local timezone if no timezone provided by the user."
+              }
+            }
+          }
+        }
+      }
+    ]
+  }'
+```
+
+Sample response:
+
+```
+{
+    "id": "chatcmpl-1746532051477984256",
+    "object": "chat.completion",
+    "created": 1746532051,
+    "model": "qwen2.5-coder-14b-instruct-q4_k_m.gguf",
+    "choices": [
+        {
+            "index": 0,
+            "finish_reason": "tool_calls",
+            "message": {
+                "role": "assistant",
+                "content": "```xml\n<function>\n{\n  \"name\": \"get_current_time\",\n  \"arguments\": {\n    \"timezone\": \"America/New_York\"\n  }\n}\n</function>\n```"
+            },
+            "tool_calls": [
+                {
+                    "type": "function",
+                    "function": {
+                        "name": "get_current_time",
+                        "arguments": "{\"timezone\": \"America/New_York\"}"
+                    },
+                    "id": "call_52ij07mh",
+                    "index": "0"
+                }
+            ]
+        }
+    ],
+    "usage": {
+        "prompt_tokens": 224,
+        "completion_tokens": 38,
+        "total_tokens": 262
+    }
+}
+```
+
 ### Environment variables

 The following environment variables can be used (they take precedence over everything else):
--- a/extensions/openai/completions.py
+++ b/extensions/openai/completions.py
@ -1,16 +1,14 @@
-import base64
 import copy
-import re
+import json
 import time
 from collections import deque
-from io import BytesIO

-import requests
 import tiktoken
-from PIL import Image
+from pydantic import ValidationError

 from extensions.openai.errors import InvalidRequestError
-from extensions.openai.utils import debug_msg
+from extensions.openai.typing import ToolDefinition
+from extensions.openai.utils import debug_msg, getToolCallId, parseToolCall
 from modules import shared
 from modules.chat import (
    generate_chat_prompt,
@ -96,72 +94,32 @@ def convert_history(history):
    user_input_last = True
    system_message = ""

-    # Multimodal: convert OpenAI format to multimodal extension format
-    if any('content' in entry and isinstance(entry['content'], list) for entry in history):
-        new_history = []
-        for entry in history:
-            if isinstance(entry['content'], list):
-                for item in entry['content']:
-                    if not isinstance(item, dict):
-                        continue
-
-                    image_url = None
-                    content = None
-                    if item['type'] == 'image_url' and isinstance(item['image_url'], dict):
-                        image_url = item['image_url']['url']
-                    elif item['type'] == 'text' and isinstance(item['text'], str):
-                        content = item['text']
-                    if image_url:
-                        new_history.append({"image_url": image_url, "role": "user"})
-                    if content:
-                        new_history.append({"content": content, "role": "user"})
-            else:
-                new_history.append(entry)
-
-        history = new_history
-
    for entry in history:
-        if "image_url" in entry:
-            image_url = entry['image_url']
-            if "base64" in image_url:
-                image_url = re.sub('^data:image/.+;base64,', '', image_url)
-                img = Image.open(BytesIO(base64.b64decode(image_url)))
-            else:
-                try:
-                    my_res = requests.get(image_url)
-                    img = Image.open(BytesIO(my_res.content))
-                except Exception:
-                    raise 'Image cannot be loaded from the URL!'
-
-            buffered = BytesIO()
-            if img.mode in ("RGBA", "P"):
-                img = img.convert("RGB")
-
-            img.save(buffered, format="JPEG")
-            img_str = base64.b64encode(buffered.getvalue()).decode('utf-8')
-            content = f'<img src="data:image/jpeg;base64,{img_str}">'
-        else:
-            content = entry["content"]
-
+        content = entry["content"]
        role = entry["role"]

        if role == "user":
            user_input = content
            user_input_last = True
            if current_message:
-                chat_dialogue.append([current_message, ''])
+                chat_dialogue.append([current_message, '', ''])
                current_message = ""

            current_message = content
        elif role == "assistant":
+            if "tool_calls" in entry and isinstance(entry["tool_calls"], list) and len(entry["tool_calls"]) > 0 and content.strip() == "":
+                continue  # skip tool calls
            current_reply = content
            user_input_last = False
            if current_message:
-                chat_dialogue.append([current_message, current_reply])
+                chat_dialogue.append([current_message, current_reply, ''])
                current_message = ""
                current_reply = ""
            else:
-                chat_dialogue.append(['', current_reply])
+                chat_dialogue.append(['', current_reply, ''])
+        elif role == "tool":
+            user_input_last = False
+            chat_dialogue.append(['', '', content])
        elif role == "system":
            system_message += f"\n{content}" if system_message else content

@ -181,6 +139,10 @@ def chat_completions_common(body: dict, is_legacy: bool = False, stream=False, p
    if 'messages' not in body:
        raise InvalidRequestError(message="messages is required", param='messages')

+    tools = None
+    if 'tools' in body and body['tools'] is not None and isinstance(body['tools'], list) and len(body['tools']) > 0:
+        tools = validateTools(body['tools'])  # raises InvalidRequestError if validation fails
+
    messages = body['messages']
    for m in messages:
        if 'role' not in m:
@ -238,6 +200,7 @@ def chat_completions_common(body: dict, is_legacy: bool = False, stream=False, p
        'custom_system_message': custom_system_message,
        'chat_template_str': chat_template_str,
        'chat-instruct_command': chat_instruct_command,
+        'tools': tools,
        'history': history,
        'stream': stream
    })
@ -250,7 +213,7 @@ def chat_completions_common(body: dict, is_legacy: bool = False, stream=False, p
    requested_model = generate_params.pop('model')
    logprob_proc = generate_params.pop('logprob_proc', None)

-    def chat_streaming_chunk(content):
+    def chat_streaming_chunk(content, chunk_tool_calls=None):
        # begin streaming
        chunk = {
            "id": cmpl_id,
@ -260,7 +223,7 @@ def chat_completions_common(body: dict, is_legacy: bool = False, stream=False, p
            resp_list: [{
                "index": 0,
                "finish_reason": None,
-                "delta": {'role': 'assistant', 'content': content},
+                "delta": {'role': 'assistant', 'content': content, 'tool_calls': chunk_tool_calls},
            }],
        }

@ -269,6 +232,7 @@ def chat_completions_common(body: dict, is_legacy: bool = False, stream=False, p
            chunk[resp_list][0]["logprobs"] = {'top_logprobs': [top_logprobs]}
        # else:
        #    chunk[resp_list][0]["logprobs"] = None
+
        return chunk

    # generate reply #######################################
@ -277,8 +241,6 @@ def chat_completions_common(body: dict, is_legacy: bool = False, stream=False, p
        yield {'prompt': prompt}
        return

-    debug_msg({'prompt': prompt, 'generate_params': generate_params})
-
    if stream:
        yield chat_streaming_chunk('')

@ -288,8 +250,23 @@ def chat_completions_common(body: dict, is_legacy: bool = False, stream=False, p
    answer = ''
    seen_content = ''

+    tool_calls = []
+    end_last_tool_call = 0
+    supported_tools = [x["function"]["name"] for x in tools] if tools is not None else None
+
    for a in generator:
        answer = a['internal'][-1][1]
+
+        if supported_tools is not None:
+            tool_call = parseToolCall(answer[end_last_tool_call:], supported_tools) if len(answer) > 0 else []
+            if len(tool_call) > 0:
+                for tc in tool_call:
+                    tc["id"] = getToolCallId()
+                    tc["index"] = str(len(tool_calls))
+                    tc["function"]["arguments"] = json.dumps(tc["function"]["arguments"])
+                    tool_calls.append(tc)
+                end_last_tool_call = len(answer)
+
        if stream:
            len_seen = len(seen_content)
            new_content = answer[len_seen:]
@ -297,18 +274,25 @@ def chat_completions_common(body: dict, is_legacy: bool = False, stream=False, p
            if not new_content or chr(0xfffd) in new_content:  # partial unicode character, don't send it yet.
                continue

-            seen_content = answer
            chunk = chat_streaming_chunk(new_content)
+
+            seen_content = answer
            yield chunk

+        # stop generation if tool_calls were generated previously
+        if len(tool_calls) > 0:
+            break
+
    token_count = len(encode(prompt)[0])
    completion_token_count = len(encode(answer)[0])
    stop_reason = "stop"
+    if len(tool_calls) > 0:
+        stop_reason = "tool_calls"
    if token_count + completion_token_count >= generate_params['truncation_length'] or completion_token_count >= generate_params['max_new_tokens']:
        stop_reason = "length"

    if stream:
-        chunk = chat_streaming_chunk('')
+        chunk = chat_streaming_chunk('', tool_calls)
        chunk[resp_list][0]['finish_reason'] = stop_reason
        chunk['usage'] = {
            "prompt_tokens": token_count,
@ -326,7 +310,8 @@ def chat_completions_common(body: dict, is_legacy: bool = False, stream=False, p
            resp_list: [{
                "index": 0,
                "finish_reason": stop_reason,
-                "message": {"role": "assistant", "content": answer}
+                "message": {"role": "assistant", "content": answer},
+                "tool_calls": tool_calls
            }],
            "usage": {
                "prompt_tokens": token_count,
@ -515,3 +500,19 @@ def completions(body: dict, is_legacy: bool = False) -> dict:
 def stream_completions(body: dict, is_legacy: bool = False):
    for resp in completions_common(body, is_legacy, stream=True):
        yield resp
+
+
+def validateTools(tools: list[dict]):
+    # Validate each tool definition in the JSON array
+    valid_tools = None
+    for idx in range(len(tools)):
+        tool = tools[idx]
+        try:
+            tool_definition = ToolDefinition(**tool)
+            if valid_tools is None:
+                valid_tools = []
+            valid_tools.append(tool)
+        except ValidationError:
+            raise InvalidRequestError(message=f"Invalid tool specification at index {idx}.", param='tools')
+
+    return valid_tools
--- a/extensions/openai/script.py
+++ b/extensions/openai/script.py
@ -14,6 +14,7 @@ from fastapi.requests import Request
 from fastapi.responses import JSONResponse
 from pydub import AudioSegment
 from sse_starlette import EventSourceResponse
+from starlette.concurrency import iterate_in_threadpool

 import extensions.openai.completions as OAIcompletions
 import extensions.openai.images as OAIimages
@ -115,7 +116,7 @@ async def openai_completions(request: Request, request_data: CompletionRequest):
        async def generator():
            async with streaming_semaphore:
                response = OAIcompletions.stream_completions(to_dict(request_data), is_legacy=is_legacy)
-                for resp in response:
+                async for resp in iterate_in_threadpool(response):
                    disconnected = await request.is_disconnected()
                    if disconnected:
                        break
@ -125,7 +126,12 @@ async def openai_completions(request: Request, request_data: CompletionRequest):
        return EventSourceResponse(generator())  # SSE streaming

    else:
-        response = OAIcompletions.completions(to_dict(request_data), is_legacy=is_legacy)
+        response = await asyncio.to_thread(
+            OAIcompletions.completions,
+            to_dict(request_data),
+            is_legacy=is_legacy
+        )
+
        return JSONResponse(response)


@ -138,7 +144,7 @@ async def openai_chat_completions(request: Request, request_data: ChatCompletion
        async def generator():
            async with streaming_semaphore:
                response = OAIcompletions.stream_chat_completions(to_dict(request_data), is_legacy=is_legacy)
-                for resp in response:
+                async for resp in iterate_in_threadpool(response):
                    disconnected = await request.is_disconnected()
                    if disconnected:
                        break
@ -148,7 +154,12 @@ async def openai_chat_completions(request: Request, request_data: ChatCompletion
        return EventSourceResponse(generator())  # SSE streaming

    else:
-        response = OAIcompletions.chat_completions(to_dict(request_data), is_legacy=is_legacy)
+        response = await asyncio.to_thread(
+            OAIcompletions.chat_completions,
+            to_dict(request_data),
+            is_legacy=is_legacy
+        )
+
        return JSONResponse(response)


@ -436,7 +447,7 @@ def run_server():

    # Start server
    logging.getLogger("uvicorn.error").propagate = False
-    uvicorn.run(app, host=server_addrs, port=port, ssl_certfile=ssl_certfile, ssl_keyfile=ssl_keyfile)
+    uvicorn.run(app, host=server_addrs, port=port, ssl_certfile=ssl_certfile, ssl_keyfile=ssl_keyfile, access_log=False)


 def setup():
--- a/extensions/openai/typing.py
+++ b/extensions/openai/typing.py
@ -1,8 +1,8 @@
 import json
 import time
-from typing import Dict, List
+from typing import Dict, List, Optional

-from pydantic import BaseModel, Field
+from pydantic import BaseModel, Field, validator


 class GenerationOptions(BaseModel):
@ -54,6 +54,48 @@ class GenerationOptions(BaseModel):
    grammar_string: str = ""


+class ToolDefinition(BaseModel):
+    function: 'ToolFunction'
+    type: str
+
+
+class ToolFunction(BaseModel):
+    description: str
+    name: str
+    parameters: 'ToolParameters'
+
+
+class ToolParameters(BaseModel):
+    properties: Optional[Dict[str, 'ToolProperty']] = None
+    required: Optional[list[str]] = None
+    type: str
+    description: Optional[str] = None
+
+
+class ToolProperty(BaseModel):
+    description: Optional[str] = None
+    type: Optional[str] = None  # we are faced with definitions like anyOf, e.g. {'type': 'function', 'function': {'name': 'git_create_branch', 'description': 'Creates a new branch from an optional base branch', 'parameters': {'type': 'object', 'properties': {'repo_path': {'title': 'Repo Path', 'type': 'string'}, 'branch_name': {'title': 'Branch Name', 'type': 'string'}, 'base_branch': {'anyOf': [{'type': 'string'}, {'type': 'null'}], 'default': None, 'title': 'Base Branch'}}, 'required': ['repo_path', 'branch_name'], 'title': 'GitCreateBranch'}}}
+
+
+class FunctionCall(BaseModel):
+    name: str
+    arguments: Optional[str] = None
+    parameters: Optional[str] = None
+
+    @validator('arguments', allow_reuse=True)
+    def checkPropertyArgsOrParams(cls, v, values, **kwargs):
+        if not v and not values.get('parameters'):
+            raise ValueError("At least one of 'arguments' or 'parameters' must be provided as property in FunctionCall type")
+        return v
+
+
+class ToolCall(BaseModel):
+    id: str
+    index: int
+    type: str
+    function: FunctionCall
+
+
 class CompletionRequestParams(BaseModel):
    model: str | None = Field(default=None, description="Unused parameter. To change the model, use the /v1/internal/model/load endpoint.")
    prompt: str | List[str]
@ -92,6 +134,7 @@ class ChatCompletionRequestParams(BaseModel):
    frequency_penalty: float | None = 0
    function_call: str | dict | None = Field(default=None, description="Unused parameter.")
    functions: List[dict] | None = Field(default=None, description="Unused parameter.")
+    tools: List[dict] | None = Field(default=None, description="Tools signatures passed via MCP.")
    logit_bias: dict | None = None
    max_tokens: int | None = None
    n: int | None = Field(default=1, description="Unused parameter.")
--- a/extensions/openai/utils.py
+++ b/extensions/openai/utils.py
@ -1,5 +1,8 @@
 import base64
+import json
 import os
+import random
+import re
 import time
 import traceback
 from typing import Callable, Optional
@ -52,3 +55,94 @@ def _start_cloudflared(port: int, tunnel_id: str, max_attempts: int = 3, on_star
            time.sleep(3)

        raise Exception('Could not start cloudflared.')
+
+
+def getToolCallId() -> str:
+    letter_bytes = "abcdefghijklmnopqrstuvwxyz0123456789"
+    b = [random.choice(letter_bytes) for _ in range(8)]
+    return "call_" + "".join(b).lower()
+
+
+def checkAndSanitizeToolCallCandidate(candidate_dict: dict, tool_names: list[str]):
+    # check if property 'function' exists and is a dictionary, otherwise adapt dict
+    if 'function' not in candidate_dict and 'name' in candidate_dict and isinstance(candidate_dict['name'], str):
+        candidate_dict = {"type": "function", "function": candidate_dict}
+    if 'function' in candidate_dict and isinstance(candidate_dict['function'], str):
+        candidate_dict['name'] = candidate_dict['function']
+        del candidate_dict['function']
+        candidate_dict = {"type": "function", "function": candidate_dict}
+    if 'function' in candidate_dict and isinstance(candidate_dict['function'], dict):
+        # check if 'name' exists within 'function' and is part of known tools
+        if 'name' in candidate_dict['function'] and candidate_dict['function']['name'] in tool_names:
+            candidate_dict["type"] = "function"  # ensure required property 'type' exists and has the right value
+            # map property 'parameters' used by some older models to 'arguments'
+            if "arguments" not in candidate_dict["function"] and "parameters" in candidate_dict["function"]:
+                candidate_dict["function"]["arguments"] = candidate_dict["function"]["parameters"]
+                del candidate_dict["function"]["parameters"]
+            return candidate_dict
+    return None
+
+
+def parseToolCall(answer: str, tool_names: list[str]):
+    matches = []
+
+    # abort on very short answers to save computation cycles
+    if len(answer) < 10:
+        return matches
+
+    # Define the regex pattern to find the JSON content wrapped in <function>, <tools>, <tool_call>, and other tags observed from various models
+    patterns = [r"(```[^\n]*)\n(.*?)```", r"<([^>]+)>(.*?)</\1>"]
+
+    for pattern in patterns:
+        for match in re.finditer(pattern, answer, re.DOTALL):
+            # print(match.group(2))
+            if match.group(2) is None:
+                continue
+            # remove backtick wraps if present
+            candidate = re.sub(r"^```(json|xml|python[^\n]*)\n", "", match.group(2).strip())
+            candidate = re.sub(r"```$", "", candidate.strip())
+            # unwrap inner tags
+            candidate = re.sub(pattern, r"\2", candidate.strip(), flags=re.DOTALL)
+            # llm might have generated multiple json objects separated by linebreaks, check for this pattern and try parsing each object individually
+            if re.search(r"\}\s*\n\s*\{", candidate) is not None:
+                candidate = re.sub(r"\}\s*\n\s*\{", "},\n{", candidate)
+            if not candidate.strip().startswith("["):
+                candidate = "[" + candidate + "]"
+
+            candidates = []
+            try:
+                # parse the candidate JSON into a dictionary
+                candidates = json.loads(candidate)
+                if not isinstance(candidates, list):
+                    candidates = [candidates]
+            except json.JSONDecodeError:
+                # Ignore invalid JSON silently
+                continue
+
+            for candidate_dict in candidates:
+                checked_candidate = checkAndSanitizeToolCallCandidate(candidate_dict, tool_names)
+                if checked_candidate is not None:
+                    matches.append(checked_candidate)
+
+        # last resort if nothing has been mapped: LLM might have produced plain json tool call without xml-like tags
+        if len(matches) == 0:
+            try:
+                candidate = answer
+                # llm might have generated multiple json objects separated by linebreaks, check for this pattern and try parsing each object individually
+                if re.search(r"\}\s*\n\s*\{", candidate) is not None:
+                    candidate = re.sub(r"\}\s*\n\s*\{", "},\n{", candidate)
+                if not candidate.strip().startswith("["):
+                    candidate = "[" + candidate + "]"
+                # parse the candidate JSON into a dictionary
+                candidates = json.loads(candidate)
+                if not isinstance(candidates, list):
+                    candidates = [candidates]
+                for candidate_dict in candidates:
+                    checked_candidate = checkAndSanitizeToolCallCandidate(candidate_dict, tool_names)
+                    if checked_candidate is not None:
+                        matches.append(checked_candidate)
+            except json.JSONDecodeError:
+                # Ignore invalid JSON silently
+                pass
+
+    return matches
--- a/extensions/superboogav2/chromadb.py
+++ b/extensions/superboogav2/chromadb.py
@ -1,10 +1,11 @@
 import math
 import random
 import threading
-import torch
+
 import chromadb
 import numpy as np
 import posthog
+import torch
 from chromadb.config import Settings
 from chromadb.utils import embedding_functions

@ -292,6 +293,8 @@ class ChromaCollector():

        for doc in documents:
            doc_tokens = encode(doc)[0]
+            if isinstance(doc_tokens, np.ndarray):
+                doc_tokens = doc_tokens.tolist()
            doc_token_count = len(doc_tokens)
            if current_token_count + doc_token_count > max_token_count:
                # If adding this document would exceed the max token count,
--- a/js/main.js
+++ b/js/main.js
@ -150,6 +150,16 @@ const observer = new MutationObserver(function(mutations) {
  if (!isScrolled && targetElement.scrollTop !== targetElement.scrollHeight) {
    targetElement.scrollTop = targetElement.scrollHeight;
  }
+
+  const chatElement = document.getElementById("chat");
+  if (chatElement) {
+    const messagesContainer = chatElement.querySelector(".messages");
+    const lastChild = messagesContainer?.lastElementChild;
+    const prevSibling = lastChild?.previousElementSibling;
+    if (lastChild && prevSibling) {
+      lastChild.style.minHeight = `calc(max(70vh, 100vh - ${prevSibling.offsetHeight}px - 102px))`;
+    }
+  }
 });

 // Configure the observer to watch for changes in the subtree and attributes
@ -442,12 +452,6 @@ function updateCssProperties() {

  // Check if the chat container is visible
  if (chatContainer.clientHeight > 0) {
-    const chatContainerParentHeight = chatContainer.parentNode.clientHeight;
-    const newChatHeight = `${chatContainerParentHeight - chatInputHeight - 80}px`;
-
-    document.documentElement.style.setProperty("--chat-height", newChatHeight);
-    document.documentElement.style.setProperty("--input-delta", `${chatInputHeight - 40}px`);
-
    // Adjust scrollTop based on input height change
    if (chatInputHeight !== currentChatInputHeight) {
      const deltaHeight = chatInputHeight - currentChatInputHeight;
@ -720,7 +724,7 @@ function isMobile() {
 // Function to initialize sidebars
 function initializeSidebars() {
  const isOnMobile = isMobile();
-  
+
  if (isOnMobile) {
    // Mobile state: Hide sidebars and set closed states
    [pastChatsRow, chatControlsRow, headerBar].forEach(el => {
--- a/modules/chat.py
+++ b/modules/chat.py
@ -5,6 +5,7 @@ import html
 import json
 import pprint
 import re
+import time
 from datetime import datetime
 from functools import partial
 from pathlib import Path
@ -145,7 +146,7 @@ def generate_chat_prompt(user_input, state, **kwargs):
    instruct_renderer = partial(
        instruction_template.render,
        builtin_tools=None,
-        tools=None,
+        tools=state['tools'] if 'tools' in state else None,
        tools_in_user_message=False,
        add_generation_prompt=False
    )
@ -171,9 +172,13 @@ def generate_chat_prompt(user_input, state, **kwargs):
            messages.append({"role": "system", "content": context})

    insert_pos = len(messages)
-    for user_msg, assistant_msg in reversed(history):
-        user_msg = user_msg.strip()
-        assistant_msg = assistant_msg.strip()
+    for entry in reversed(history):
+        user_msg = entry[0].strip()
+        assistant_msg = entry[1].strip()
+        tool_msg = entry[2].strip() if len(entry) > 2 else ''
+
+        if tool_msg:
+            messages.insert(insert_pos, {"role": "tool", "content": tool_msg})

        if assistant_msg:
            messages.insert(insert_pos, {"role": "assistant", "content": assistant_msg})
@ -394,16 +399,13 @@ def chatbot_wrapper(text, state, regenerate=False, _continue=False, loading_mess

        # Extract the reply
        if state['mode'] in ['chat', 'chat-instruct']:
-            visible_reply = re.sub("(<USER>|<user>|{{user}})", state['name1'], reply + '▍')
+            visible_reply = re.sub("(<USER>|<user>|{{user}})", state['name1'], reply)
        else:
-            visible_reply = reply + '▍'
+            visible_reply = reply

        visible_reply = html.escape(visible_reply)

        if shared.stop_everything:
-            if output['visible'][-1][1].endswith('▍'):
-                output['visible'][-1][1] = output['visible'][-1][1][:-1]
-
            output['visible'][-1][1] = apply_extensions('output', output['visible'][-1][1], state, is_chat=True)
            yield output
            return
@ -419,9 +421,6 @@ def chatbot_wrapper(text, state, regenerate=False, _continue=False, loading_mess
            if is_stream:
                yield output

-    if output['visible'][-1][1].endswith('▍'):
-        output['visible'][-1][1] = output['visible'][-1][1][:-1]
-
    output['visible'][-1][1] = apply_extensions('output', output['visible'][-1][1], state, is_chat=True)
    yield output

@ -481,9 +480,17 @@ def generate_chat_reply_wrapper(text, state, regenerate=False, _continue=False):
        send_dummy_reply(state['start_with'], state)

    history = state['history']
+    last_save_time = time.monotonic()
+    save_interval = 8
    for i, history in enumerate(generate_chat_reply(text, state, regenerate, _continue, loading_message=True, for_ui=True)):
        yield chat_html_wrapper(history, state['name1'], state['name2'], state['mode'], state['chat_style'], state['character_menu']), history

+        current_time = time.monotonic()
+        # Save on first iteration or if save_interval seconds have passed
+        if i == 0 or (current_time - last_save_time) >= save_interval:
+            save_history(history, state['unique_id'], state['character_menu'], state['mode'])
+            last_save_time = current_time
+
    save_history(history, state['unique_id'], state['character_menu'], state['mode'])


--- a/modules/exllamav3_hf.py
+++ b/modules/exllamav3_hf.py
@ -119,7 +119,7 @@ class Exllamav3HF(PreTrainedModel, GenerationMixin):
        reset = True

        # Maximum number of tokens to process in a single forward pass
-        max_chunk_size = 2048
+        max_chunk_size = 256

        # Make the forward call
        if labels is None:
--- a/modules/llama_cpp_server.py
+++ b/modules/llama_cpp_server.py
@ -66,7 +66,7 @@ class LlamaServer:
            "top_k": state["top_k"],
            "top_p": state["top_p"],
            "min_p": state["min_p"],
-            "tfs_z": state["tfs"],
+            "top_n_sigma": state["top_n_sigma"] if state["top_n_sigma"] > 0 else -1,
            "typical_p": state["typical_p"],
            "repeat_penalty": state["repetition_penalty"],
            "repeat_last_n": state["repetition_penalty_range"],
@ -102,8 +102,10 @@ class LlamaServer:

            penalty_found = False
            for s in samplers:
-                if s.strip() in ["dry", "top_k", "typ_p", "top_p", "min_p", "xtc", "temperature"]:
+                if s.strip() in ["dry", "top_k", "top_p", "top_n_sigma", "min_p", "temperature", "xtc"]:
                    filtered_samplers.append(s.strip())
+                elif s.strip() == "typical_p":
+                    filtered_samplers.append("typ_p")
                elif not penalty_found and s.strip() == "repetition_penalty":
                    filtered_samplers.append("penalties")
                    penalty_found = True
@ -144,8 +146,9 @@ class LlamaServer:
            pprint.PrettyPrinter(indent=4, sort_dicts=False).pprint(printable_payload)
            print()

-        # Make a direct request with streaming enabled using a context manager
-        with self.session.post(url, json=payload, stream=True) as response:
+        # Make the generation request
+        response = self.session.post(url, json=payload, stream=True)
+        try:
            response.raise_for_status()  # Raise an exception for HTTP errors

            full_text = ""
@ -182,6 +185,8 @@ class LlamaServer:
                    print(f"JSON decode error: {e}")
                    print(f"Problematic line: {line}")
                    continue
+        finally:
+            response.close()

    def generate(self, prompt, state):
        output = ""
@ -210,14 +215,15 @@ class LlamaServer:
            pprint.PrettyPrinter(indent=4, sort_dicts=False).pprint(printable_payload)
            print()

-        response = self.session.post(url, json=payload)
-        result = response.json()
+        for retry in range(5):
+            response = self.session.post(url, json=payload)
+            result = response.json()

-        if "completion_probabilities" in result:
-            if use_samplers:
-                return result["completion_probabilities"][0]["top_probs"]
-            else:
-                return result["completion_probabilities"][0]["top_logprobs"]
+            if "completion_probabilities" in result:
+                if use_samplers:
+                    return result["completion_probabilities"][0]["top_probs"]
+                else:
+                    return result["completion_probabilities"][0]["top_logprobs"]
        else:
            raise Exception(f"Unexpected response format: 'completion_probabilities' not found in {result}")

@ -255,9 +261,10 @@ class LlamaServer:
            self.server_path,
            "--model", self.model_path,
            "--ctx-size", str(shared.args.ctx_size),
-            "--n-gpu-layers", str(shared.args.n_gpu_layers),
+            "--gpu-layers", str(shared.args.gpu_layers),
            "--batch-size", str(shared.args.batch_size),
            "--port", str(self.port),
+            "--no-webui",
        ]

        if shared.args.flash_attn:
@ -278,8 +285,10 @@ class LlamaServer:
            cmd.append("--no-kv-offload")
        if shared.args.row_split:
            cmd += ["--split-mode", "row"]
+        cache_type = "fp16"
        if shared.args.cache_type != "fp16" and shared.args.cache_type in llamacpp_valid_cache_types:
            cmd += ["--cache-type-k", shared.args.cache_type, "--cache-type-v", shared.args.cache_type]
+            cache_type = shared.args.cache_type
        if shared.args.compress_pos_emb != 1:
            cmd += ["--rope-freq-scale", str(1.0 / shared.args.compress_pos_emb)]
        if shared.args.rope_freq_base > 0:
@ -316,9 +325,15 @@ class LlamaServer:
            for flag_item in extra_flags.split(','):
                if '=' in flag_item:
                    flag, value = flag_item.split('=', 1)
-                    cmd += [f"--{flag}", value]
+                    if len(flag) <= 3:
+                        cmd += [f"-{flag}", value]
+                    else:
+                        cmd += [f"--{flag}", value]
                else:
-                    cmd.append(f"--{flag_item}")
+                    if len(flag_item) <= 3:
+                        cmd.append(f"-{flag_item}")
+                    else:
+                        cmd.append(f"--{flag_item}")

        env = os.environ.copy()
        if os.name == 'posix':
@ -333,6 +348,7 @@ class LlamaServer:
            print(' '.join(str(item) for item in cmd[1:]))
            print()

+        logger.info(f"Using gpu_layers={shared.args.gpu_layers} | ctx_size={shared.args.ctx_size} | cache_type={cache_type}")
        # Start the server with pipes for output
        self.process = subprocess.Popen(
            cmd,
--- a/modules/loaders.py
+++ b/modules/loaders.py
@ -5,7 +5,7 @@ import gradio as gr

 loaders_and_params = OrderedDict({
    'llama.cpp': [
-        'n_gpu_layers',
+        'gpu_layers',
        'threads',
        'threads_batch',
        'batch_size',
@ -28,6 +28,7 @@ loaders_and_params = OrderedDict({
        'device_draft',
        'ctx_size_draft',
        'speculative_decoding_accordion',
+        'vram_info',
    ],
    'Transformers': [
        'gpu_split',
@ -84,7 +85,6 @@ loaders_and_params = OrderedDict({
        'no_flash_attn',
        'no_xformers',
        'no_sdpa',
-        'exllamav2_info',
        'model_draft',
        'draft_max',
        'ctx_size_draft',
@ -299,7 +299,7 @@ loaders_samplers = {
        'typical_p',
        'xtc_threshold',
        'xtc_probability',
-        'tfs',
+        'top_n_sigma',
        'dry_multiplier',
        'dry_allowed_length',
        'dry_base',
--- a/modules/logits.py
+++ b/modules/logits.py
@ -7,6 +7,7 @@ from modules import models, shared
 from modules.logging_colors import logger
 from modules.models import load_model
 from modules.text_generation import generate_reply
+from modules.utils import check_model_loaded

 global_scores = None

@ -33,9 +34,9 @@ def get_next_logits(*args, **kwargs):


 def _get_next_logits(prompt, state, use_samplers, previous, top_logits=25, return_dict=False):
-    if shared.model is None:
-        logger.error("No model is loaded! Select one in the Model tab.")
-        return 'Error: No model is loaded1 Select one in the Model tab.', previous
+    model_is_loaded, error_message = check_model_loaded()
+    if not model_is_loaded:
+        return error_message, previous

    # llama.cpp case
    if shared.model.__class__.__name__ == 'LlamaServer':
--- a/modules/models.py
+++ b/modules/models.py
@ -71,7 +71,6 @@ def llama_cpp_server_loader(model_name):
    else:
        model_file = sorted(Path(f'{shared.args.model_dir}/{model_name}').glob('*.gguf'))[0]

-    logger.info(f"llama.cpp weights detected: \"{model_file}\"")
    try:
        model = LlamaServer(model_file)
        return model, model
--- a/modules/models_settings.py
+++ b/modules/models_settings.py
@ -1,7 +1,11 @@
+import functools
 import json
 import re
+import subprocess
+from math import exp
 from pathlib import Path

+import gradio as gr
 import yaml

 from modules import chat, loaders, metadata_gguf, shared, ui
@ -54,7 +58,7 @@ def get_model_metadata(model):
        else:
            model_file = list(path.glob('*.gguf'))[0]

-        metadata = metadata_gguf.load_metadata(model_file)
+        metadata = load_gguf_metadata_with_cache(model_file)

        for k in metadata:
            if k.endswith('context_length'):
@ -67,7 +71,7 @@ def get_model_metadata(model):
            elif k.endswith('rope.scaling.factor'):
                model_settings['compress_pos_emb'] = metadata[k]
            elif k.endswith('block_count'):
-                model_settings['n_gpu_layers'] = metadata[k] + 1
+                model_settings['gpu_layers'] = metadata[k] + 1

        if 'tokenizer.chat_template' in metadata:
            template = metadata['tokenizer.chat_template']
@ -209,15 +213,27 @@ def apply_model_settings_to_state(model, state):
    model_settings = get_model_metadata(model)
    if 'loader' in model_settings:
        loader = model_settings.pop('loader')
-
-        # If the user is using an alternative loader for the same model type, let them keep using it
        if not (loader == 'ExLlamav2_HF' and state['loader'] in ['ExLlamav2']):
            state['loader'] = loader

    for k in model_settings:
-        if k in state:
+        if k in state and k != 'gpu_layers':  # Skip gpu_layers, handle separately
            state[k] = model_settings[k]

+    # Handle GPU layers and VRAM update for llama.cpp
+    if state['loader'] == 'llama.cpp' and 'gpu_layers' in model_settings:
+        vram_info, gpu_layers_update = update_gpu_layers_and_vram(
+            state['loader'],
+            model,
+            model_settings['gpu_layers'],
+            state['ctx_size'],
+            state['cache_type'],
+            auto_adjust=True
+        )
+
+        state['gpu_layers'] = gpu_layers_update
+        state['vram_info'] = vram_info
+
    return state


@ -277,3 +293,186 @@ def save_instruction_template(model, template):
        yield (f"Instruction template for `{model}` unset in `{p}`, as the value for template was `{template}`.")
    else:
        yield (f"Instruction template for `{model}` saved to `{p}` as `{template}`.")
+
+
+@functools.lru_cache(maxsize=1)
+def load_gguf_metadata_with_cache(model_file):
+    return metadata_gguf.load_metadata(model_file)
+
+
+def get_model_size_mb(model_file: Path) -> float:
+    filename = model_file.name
+
+    # Check for multipart pattern
+    match = re.match(r'(.+)-\d+-of-\d+\.gguf$', filename)
+
+    if match:
+        # It's a multipart file, find all matching parts
+        base_pattern = match.group(1)
+        part_files = sorted(model_file.parent.glob(f'{base_pattern}-*-of-*.gguf'))
+        total_size = sum(p.stat().st_size for p in part_files)
+    else:
+        # Single part
+        total_size = model_file.stat().st_size
+
+    return total_size / (1024 ** 2)  # Return size in MB
+
+
+def estimate_vram(gguf_file, gpu_layers, ctx_size, cache_type):
+    model_file = Path(f'{shared.args.model_dir}/{gguf_file}')
+    metadata = load_gguf_metadata_with_cache(model_file)
+    size_in_mb = get_model_size_mb(model_file)
+
+    # Extract values from metadata
+    n_layers = None
+    n_kv_heads = None
+    embedding_dim = None
+    context_length = None
+    feed_forward_dim = None
+
+    for key, value in metadata.items():
+        if key.endswith('.block_count'):
+            n_layers = value
+        elif key.endswith('.attention.head_count_kv'):
+            n_kv_heads = value
+        elif key.endswith('.embedding_length'):
+            embedding_dim = value
+        elif key.endswith('.context_length'):
+            context_length = value
+        elif key.endswith('.feed_forward_length'):
+            feed_forward_dim = value
+
+    if gpu_layers > n_layers:
+        gpu_layers = n_layers
+
+    # Convert cache_type to numeric
+    if cache_type == 'q4_0':
+        cache_type = 4
+    elif cache_type == 'q8_0':
+        cache_type = 8
+    else:
+        cache_type = 16
+
+    # Derived features
+    size_per_layer = size_in_mb / max(n_layers, 1e-6)
+    context_per_layer = context_length / max(n_layers, 1e-6)
+    ffn_per_embedding = feed_forward_dim / max(embedding_dim, 1e-6)
+    kv_cache_factor = n_kv_heads * cache_type * ctx_size
+
+    # Helper function for smaller
+    def smaller(x, y):
+        return 1 if x < y else 0
+
+    # Calculate VRAM using the model
+    # Details: https://oobabooga.github.io/blog/posts/gguf-vram-formula/
+    vram = (
+        (size_per_layer - 21.19195204848197)
+        * exp(0.0001047328491557063 * size_in_mb * smaller(ffn_per_embedding, 2.671096993407845))
+        + 0.0006621544775632052 * context_per_layer
+        + 3.34664386576376e-05 * kv_cache_factor
+    ) * (1.363306170123392 + gpu_layers) + 1255.163594536052
+
+    return vram
+
+
+def get_nvidia_free_vram():
+    """
+    Calculates the total free VRAM across all NVIDIA GPUs by parsing nvidia-smi output.
+
+    Returns:
+        int: The total free VRAM in MiB summed across all detected NVIDIA GPUs.
+             Returns -1 if nvidia-smi command fails (not found, error, etc.).
+             Returns 0 if nvidia-smi succeeds but no GPU memory info found.
+    """
+    try:
+        # Execute nvidia-smi command
+        result = subprocess.run(
+            ['nvidia-smi'],
+            capture_output=True,
+            text=True,
+            check=False
+        )
+
+        # Check if nvidia-smi returned an error
+        if result.returncode != 0:
+            return -1
+
+        # Parse the output for memory usage patterns
+        output = result.stdout
+
+        # Find memory usage like "XXXXMiB / YYYYMiB"
+        # Captures used and total memory for each GPU
+        matches = re.findall(r"(\d+)\s*MiB\s*/\s*(\d+)\s*MiB", output)
+
+        if not matches:
+            # No GPUs found in expected format
+            return 0
+
+        total_free_vram_mib = 0
+        for used_mem_str, total_mem_str in matches:
+            try:
+                used_mib = int(used_mem_str)
+                total_mib = int(total_mem_str)
+                total_free_vram_mib += (total_mib - used_mib)
+            except ValueError:
+                # Skip malformed entries
+                pass
+
+        return total_free_vram_mib
+
+    except FileNotFoundError:
+        # nvidia-smi not found (likely no NVIDIA drivers installed)
+        return -1
+    except Exception:
+        # Handle any other unexpected exceptions
+        return -1
+
+
+def update_gpu_layers_and_vram(loader, model, gpu_layers, ctx_size, cache_type, auto_adjust=False, for_ui=True):
+    """
+    Unified function to handle GPU layers and VRAM updates.
+
+    Args:
+        for_ui: If True, returns Gradio updates. If False, returns raw values.
+
+    Returns:
+        - If for_ui=True: (vram_info_update, gpu_layers_update) or just vram_info_update
+        - If for_ui=False: (vram_usage, adjusted_layers) or just vram_usage
+    """
+    if loader != 'llama.cpp' or model in ["None", None] or not model.endswith(".gguf"):
+        vram_info = "<div id=\"vram-info\"'>Estimated VRAM to load the model:</span>"
+        if for_ui:
+            return (vram_info, gr.update()) if auto_adjust else vram_info
+        else:
+            return (0, gpu_layers) if auto_adjust else 0
+
+    current_layers = gpu_layers
+    max_layers = gpu_layers
+
+    if auto_adjust:
+        # Get max layers from model metadata
+        model_settings = get_model_metadata(model)
+        max_layers = model_settings.get('gpu_layers', gpu_layers)
+
+        # Auto-adjust based on available VRAM
+        available_vram = get_nvidia_free_vram()
+        if available_vram > 0:
+            tolerance = 906
+            current_layers = max_layers
+            while current_layers > 0 and estimate_vram(model, current_layers, ctx_size, cache_type) > available_vram - tolerance:
+                current_layers -= 1
+
+    # Calculate VRAM with current layers
+    vram_usage = estimate_vram(model, current_layers, ctx_size, cache_type)
+
+    if for_ui:
+        vram_info = f"<div id=\"vram-info\"'>Estimated VRAM to load the model: <span class=\"value\">{vram_usage:.0f} MiB</span>"
+        if auto_adjust:
+            return vram_info, gr.update(value=current_layers, maximum=max_layers)
+        else:
+            return vram_info
+    else:
+        if auto_adjust:
+            return vram_usage, current_layers
+        else:
+            return vram_usage
--- a/modules/presets.py
+++ b/modules/presets.py
@ -11,7 +11,7 @@ from modules.logging_colors import logger


 def default_preset():
-    return {
+    result = {
        'temperature': 1,
        'dynatemp_low': 1,
        'dynatemp_high': 1,
@ -46,10 +46,17 @@ def default_preset():
        'do_sample': True,
        'dynamic_temperature': False,
        'temperature_last': False,
-        'sampler_priority': 'repetition_penalty\npresence_penalty\nfrequency_penalty\ndry\ntemperature\ndynamic_temperature\nquadratic_sampling\ntop_n_sigma\ntop_k\ntop_p\ntypical_p\nepsilon_cutoff\neta_cutoff\ntfs\ntop_a\nmin_p\nmirostat\nxtc\nencoder_repetition_penalty\nno_repeat_ngram',
+        'sampler_priority': 'repetition_penalty\npresence_penalty\nfrequency_penalty\ndry\ntop_n_sigma\ntemperature\ndynamic_temperature\nquadratic_sampling\ntop_k\ntop_p\ntypical_p\nepsilon_cutoff\neta_cutoff\ntfs\ntop_a\nmin_p\nmirostat\nxtc\nencoder_repetition_penalty\nno_repeat_ngram',
        'dry_sequence_breakers': '"\\n", ":", "\\"", "*"',
    }

+    if shared.args.portable:
+        samplers = result['sampler_priority'].split('\n')
+        samplers = [sampler for sampler in samplers if sampler in ["dry", "top_k", "top_p", "top_n_sigma", "min_p", "temperature", "xtc", "typical_p", "repetition_penalty"]]
+        result['sampler_priority'] = '\n'.join(samplers)
+
+    return result
+

 def presets_params():
    return [k for k in default_preset()]
--- a/modules/shared.py
+++ b/modules/shared.py
@ -47,7 +47,6 @@ settings = {
    'max_new_tokens_max': 4096,
    'prompt_lookup_num_tokens': 0,
    'max_tokens_second': 0,
-    'max_updates_second': 12,
    'auto_max_new_tokens': True,
    'ban_eos_token': False,
    'add_bos_token': True,
@ -60,7 +59,6 @@ settings = {
    'custom_stopping_strings': '',
    'custom_token_bans': '',
    'negative_prompt': '',
-    'autoload_model': False,
    'dark_theme': True,
    'default_extensions': [],
    'instruction_template_str': "{%- set ns = namespace(found=false) -%}\n{%- for message in messages -%}\n    {%- if message['role'] == 'system' -%}\n        {%- set ns.found = true -%}\n    {%- endif -%}\n{%- endfor -%}\n{%- if not ns.found -%}\n    {{- '' + 'Below is an instruction that describes a task. Write a response that appropriately completes the request.' + '\\n\\n' -}}\n{%- endif %}\n{%- for message in messages %}\n    {%- if message['role'] == 'system' -%}\n        {{- '' + message['content'] + '\\n\\n' -}}\n    {%- else -%}\n        {%- if message['role'] == 'user' -%}\n            {{-'### Instruction:\\n' + message['content'] + '\\n\\n'-}}\n        {%- else -%}\n            {{-'### Response:\\n' + message['content'] + '\\n\\n' -}}\n        {%- endif -%}\n    {%- endif -%}\n{%- endfor -%}\n{%- if add_generation_prompt -%}\n    {{-'### Response:\\n'-}}\n{%- endif -%}",
@ -121,7 +119,7 @@ group.add_argument('--threads-batch', type=int, default=0, help='Number of threa
 group.add_argument('--batch-size', type=int, default=256, help='Maximum number of prompt tokens to batch together when calling llama_eval.')
 group.add_argument('--no-mmap', action='store_true', help='Prevent mmap from being used.')
 group.add_argument('--mlock', action='store_true', help='Force the system to keep the model in RAM.')
-group.add_argument('--n-gpu-layers', type=int, default=0, help='Number of layers to offload to the GPU.')
+group.add_argument('--gpu-layers', '--n-gpu-layers', type=int, default=256, metavar='N', help='Number of layers to offload to the GPU.')
 group.add_argument('--tensor-split', type=str, default=None, help='Split the model across multiple GPUs. Comma-separated list of proportions. Example: 60,40.')
 group.add_argument('--numa', action='store_true', help='Activate NUMA task allocation for llama.cpp.')
 group.add_argument('--no-kv-offload', action='store_true', help='Do not offload the  K, Q, V to the GPU. This saves VRAM but reduces the performance.')
@ -130,9 +128,9 @@ group.add_argument('--extra-flags', type=str, default=None, help='Extra flags to
 group.add_argument('--streaming-llm', action='store_true', help='Activate StreamingLLM to avoid re-evaluating the entire prompt when old messages are removed.')

 # Cache
-group = parser.add_argument_group('Context and cache management')
+group = parser.add_argument_group('Context and cache')
 group.add_argument('--ctx-size', '--n_ctx', '--max_seq_len', type=int, default=8192, metavar='N', help='Context size in tokens.')
-group.add_argument('--cache_type', type=str, default='fp16', help='KV cache type; valid options: llama.cpp - fp16, q8_0, q4_0; ExLlamaV2 - fp16, fp8, q8, q6, q4; ExLlamaV3 - fp16, q2 to q8 (can specify k_bits and v_bits separately, e.g. q4_q8).')
+group.add_argument('--cache-type', '--cache_type', type=str, default='fp16', metavar='N', help='KV cache type; valid options: llama.cpp - fp16, q8_0, q4_0; ExLlamaV2 - fp16, fp8, q8, q6, q4; ExLlamaV3 - fp16, q2 to q8 (can specify k_bits and v_bits separately, e.g. q4_q8).')

 # Speculative decoding
 group = parser.add_argument_group('Speculative decoding')
@ -161,10 +159,6 @@ group.add_argument('--hqq-backend', type=str, default='PYTORCH_COMPILE', help='B
 group = parser.add_argument_group('TensorRT-LLM')
 group.add_argument('--cpp-runner', action='store_true', help='Use the ModelRunnerCpp runner, which is faster than the default ModelRunner but doesn\'t support streaming yet.')

-# Cache
-group = parser.add_argument_group('Cache')
-group.add_argument('--cache_type', type=str, default='fp16', help='KV cache type; valid options: llama.cpp - fp16, q8_0, q4_0; ExLlamaV2 - fp16, fp8, q8, q6, q4.')
-
 # DeepSpeed
 group = parser.add_argument_group('DeepSpeed')
 group.add_argument('--deepspeed', action='store_true', help='Enable the use of DeepSpeed ZeRO-3 for inference via the Transformers integration.')
@ -190,6 +184,7 @@ group.add_argument('--ssl-keyfile', type=str, help='The path to the SSL certific
 group.add_argument('--ssl-certfile', type=str, help='The path to the SSL certificate cert file.', default=None)
 group.add_argument('--subpath', type=str, help='Customize the subpath for gradio, use with reverse proxy')
 group.add_argument('--old-colors', action='store_true', help='Use the legacy Gradio colors, before the December/2024 update.')
+group.add_argument('--portable', action='store_true', help='Hide features not available in portable mode like training.')

 # API
 group = parser.add_argument_group('API')
@ -311,11 +306,13 @@ if args.api or args.public_api:
    add_extension('openai', last=True)

 # Load model-specific settings
-with Path(f'{args.model_dir}/config.yaml') as p:
-    if p.exists():
-        model_config = yaml.safe_load(open(p, 'r').read())
-    else:
-        model_config = {}
+p = Path(f'{args.model_dir}/config.yaml')
+if p.exists():
+    model_config = yaml.safe_load(open(p, 'r').read())
+else:
+    model_config = {}
+del p
+

 # Load custom model-specific settings
 user_config = load_user_config()
--- a/modules/tensorrt_llm.py
+++ b/modules/tensorrt_llm.py
@ -1,15 +1,15 @@
 from pathlib import Path

-import torch
-
 import tensorrt_llm
+import torch
+from tensorrt_llm.runtime import ModelRunner, ModelRunnerCpp
+
 from modules import shared
 from modules.logging_colors import logger
 from modules.text_generation import (
    get_max_prompt_length,
    get_reply_from_output_ids
 )
-from tensorrt_llm.runtime import ModelRunner, ModelRunnerCpp


 class TensorRTLLMModel:
--- a/modules/text_generation.py
+++ b/modules/text_generation.py
@ -14,6 +14,7 @@ from modules.callbacks import Iteratorize
 from modules.extensions import apply_extensions
 from modules.html_generator import generate_basic_html
 from modules.logging_colors import logger
+from modules.utils import check_model_loaded


 def generate_reply(*args, **kwargs):
@ -34,8 +35,8 @@ def _generate_reply(question, state, stopping_strings=None, is_chat=False, escap
    # Find the appropriate generation function
    generate_func = apply_extensions('custom_generate_reply')
    if generate_func is None:
-        if shared.model_name == 'None' or shared.model is None:
-            logger.error("No model is loaded! Select one in the Model tab.")
+        model_is_loaded, error_message = check_model_loaded()
+        if not model_is_loaded:
            yield ''
            return

@ -64,41 +65,39 @@ def _generate_reply(question, state, stopping_strings=None, is_chat=False, escap
            all_stop_strings += st

    shared.stop_everything = False
-    last_update = -1
    reply = ''
    is_stream = state['stream']
    if len(all_stop_strings) > 0 and not state['stream']:
        state = copy.deepcopy(state)
        state['stream'] = True

-    min_update_interval = 0
-    if state.get('max_updates_second', 0) > 0:
-        min_update_interval = 1 / state['max_updates_second']
-
    # Generate
+    last_update = -1
+    latency_threshold = 1 / 1000
    for reply in generate_func(question, original_question, state, stopping_strings, is_chat=is_chat):
+        cur_time = time.monotonic()
        reply, stop_found = apply_stopping_strings(reply, all_stop_strings)
        if escape_html:
            reply = html.escape(reply)

        if is_stream:
-            cur_time = time.time()
-
            # Limit number of tokens/second to make text readable in real time
            if state['max_tokens_second'] > 0:
                diff = 1 / state['max_tokens_second'] - (cur_time - last_update)
                if diff > 0:
                    time.sleep(diff)

-                last_update = time.time()
+                last_update = time.monotonic()
                yield reply

            # Limit updates to avoid lag in the Gradio UI
            # API updates are not limited
            else:
-                if cur_time - last_update > min_update_interval:
-                    last_update = cur_time
+                # If 'generate_func' takes less than 0.001 seconds to yield the next token
+                # (equivalent to more than 1000 tok/s), assume that the UI is lagging behind and skip yielding
+                if (cur_time - last_update) > latency_threshold:
                    yield reply
+                last_update = time.monotonic()

        if stop_found or (state['max_tokens_second'] > 0 and shared.stop_everything):
            break
@ -471,7 +470,7 @@ def generate_reply_HF(question, original_question, state, stopping_strings=None,
        t1 = time.time()
        original_tokens = len(original_input_ids[0])
        new_tokens = len(output) - (original_tokens if not shared.is_seq2seq else 0)
-        print(f'Output generated in {(t1-t0):.2f} seconds ({new_tokens/(t1-t0):.2f} tokens/s, {new_tokens} tokens, context {original_tokens}, seed {seed})')
+        logger.info(f'Output generated in {(t1-t0):.2f} seconds ({new_tokens/(t1-t0):.2f} tokens/s, {new_tokens} tokens, context {original_tokens}, seed {seed})')
        return


@ -480,7 +479,7 @@ def generate_reply_custom(question, original_question, state, stopping_strings=N
    For models that do not use the transformers library for sampling
    """

-    seed = set_manual_seed(state['seed'])
+    state['seed'] = set_manual_seed(state['seed'])
    t0 = time.time()
    reply = ''
    try:
@ -500,7 +499,7 @@ def generate_reply_custom(question, original_question, state, stopping_strings=N
        t1 = time.time()
        original_tokens = len(encode(original_question)[0])
        new_tokens = len(encode(original_question + reply)[0]) - original_tokens
-        print(f'Output generated in {(t1-t0):.2f} seconds ({new_tokens/(t1-t0):.2f} tokens/s, {new_tokens} tokens, context {original_tokens}, seed {seed})')
+        logger.info(f'Output generated in {(t1-t0):.2f} seconds ({new_tokens/(t1-t0):.2f} tokens/s, {new_tokens} tokens, context {original_tokens}, seed {state["seed"]})')
        return


--- a/modules/ui.py
+++ b/modules/ui.py
@ -61,7 +61,7 @@ if not shared.args.old_colors:
        background_fill_primary_dark='var(--darker-gray)',
        body_background_fill="white",
        block_background_fill="transparent",
-        body_text_color="#333",
+        body_text_color='rgb(64, 64, 64)',
        button_secondary_background_fill="#f4f4f4",
        button_secondary_border_color="var(--border-color-primary)",

@ -105,7 +105,7 @@ def list_model_elements():
        'filter_by_loader',
        'loader',
        'cpu_memory',
-        'n_gpu_layers',
+        'gpu_layers',
        'threads',
        'threads_batch',
        'batch_size',
@ -192,7 +192,6 @@ def list_interface_input_elements():
        'max_new_tokens',
        'prompt_lookup_num_tokens',
        'max_tokens_second',
-        'max_updates_second',
        'do_sample',
        'dynamic_temperature',
        'temperature_last',
--- a/modules/ui_chat.py
+++ b/modules/ui_chat.py
@ -46,8 +46,8 @@ def create_ui():

        with gr.Row():
            with gr.Column(elem_id='chat-col'):
-                shared.gradio['html_display'] = gr.HTML(value=chat_html_wrapper({'internal': [], 'visible': []}, '', '', 'chat', 'cai-chat', '')['html'], visible=True)
                shared.gradio['display'] = gr.JSON(value={}, visible=False)  # Hidden buffer
+                shared.gradio['html_display'] = gr.HTML(value=chat_html_wrapper({'internal': [], 'visible': []}, '', '', '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')
--- a/modules/ui_model_menu.py
+++ b/modules/ui_model_menu.py
@ -14,6 +14,7 @@ from modules.models_settings import (
    get_model_metadata,
    save_instruction_template,
    save_model_settings,
+    update_gpu_layers_and_vram,
    update_model_parameters
 )
 from modules.utils import gradio
@ -26,71 +27,36 @@ def create_ui():
        with gr.Row():
            with gr.Column():
                with gr.Row():
-                    with gr.Column():
-                        with gr.Row():
-                            shared.gradio['model_menu'] = gr.Dropdown(choices=utils.get_available_models(), value=lambda: shared.model_name, label='Model', elem_classes='slim-dropdown', interactive=not mu)
-                            ui.create_refresh_button(shared.gradio['model_menu'], lambda: None, lambda: {'choices': utils.get_available_models()}, 'refresh-button', interactive=not mu)
-                            shared.gradio['load_model'] = gr.Button("Load", visible=not shared.settings['autoload_model'], elem_classes='refresh-button', interactive=not mu)
-                            shared.gradio['unload_model'] = gr.Button("Unload", elem_classes='refresh-button', interactive=not mu)
-                            shared.gradio['save_model_settings'] = gr.Button("Save settings", elem_classes='refresh-button', interactive=not mu)
+                    shared.gradio['model_menu'] = gr.Dropdown(choices=utils.get_available_models(), value=lambda: shared.model_name, label='Model', elem_classes='slim-dropdown', interactive=not mu)
+                    ui.create_refresh_button(shared.gradio['model_menu'], lambda: None, lambda: {'choices': utils.get_available_models()}, 'refresh-button', interactive=not mu)
+                    shared.gradio['load_model'] = gr.Button("Load", elem_classes='refresh-button', interactive=not mu)
+                    shared.gradio['unload_model'] = gr.Button("Unload", elem_classes='refresh-button', interactive=not mu)
+                    shared.gradio['save_model_settings'] = gr.Button("Save settings", elem_classes='refresh-button', interactive=not mu)

-                    with gr.Column():
-                        with gr.Row():
-                            shared.gradio['lora_menu'] = gr.Dropdown(multiselect=True, choices=utils.get_available_loras(), value=shared.lora_names, label='LoRA(s)', elem_classes='slim-dropdown', interactive=not mu)
-                            ui.create_refresh_button(shared.gradio['lora_menu'], lambda: None, lambda: {'choices': utils.get_available_loras(), 'value': shared.lora_names}, 'refresh-button', interactive=not mu)
-                            shared.gradio['lora_menu_apply'] = gr.Button(value='Apply LoRAs', elem_classes='refresh-button', interactive=not mu)
-
-        with gr.Row():
-            with gr.Column():
-                shared.gradio['loader'] = gr.Dropdown(label="Model loader", choices=loaders.loaders_and_params.keys(), value=None)
+                shared.gradio['loader'] = gr.Dropdown(label="Model loader", choices=loaders.loaders_and_params.keys() if not shared.args.portable else ['llama.cpp'], value=None)
                with gr.Blocks():
+                    gr.Markdown("## Main options")
                    with gr.Row():
                        with gr.Column():
-                            shared.gradio['n_gpu_layers'] = gr.Slider(label="n-gpu-layers", minimum=0, maximum=256, value=shared.args.n_gpu_layers, info='Must be greater than 0 for the GPU to be used. ⚠️ Lower this value if you can\'t load the model.')
-                            shared.gradio['threads'] = gr.Slider(label="threads", minimum=0, step=1, maximum=256, value=shared.args.threads)
-                            shared.gradio['threads_batch'] = gr.Slider(label="threads_batch", minimum=0, step=1, maximum=256, value=shared.args.threads_batch)
-                            shared.gradio['batch_size'] = gr.Slider(label="batch_size", minimum=1, maximum=4096, step=1, value=shared.args.batch_size)
-                            shared.gradio['hqq_backend'] = gr.Dropdown(label="hqq_backend", choices=["PYTORCH", "PYTORCH_COMPILE", "ATEN"], value=shared.args.hqq_backend)
-                            shared.gradio['ctx_size'] = gr.Number(label='ctx-size', precision=0, step=256, value=shared.args.ctx_size, info='Context length. ⚠️ Lower this value if you can\'t load the model. Common values: 2048, 4096, 8192, 16384, 32768, 65536.')
-                            shared.gradio['cache_type'] = gr.Dropdown(label="cache_type", choices=['fp16', 'q8_0', 'q4_0', 'fp8', 'q8', 'q7', 'q6', 'q5', 'q4', 'q3', 'q2'], value=shared.args.cache_type, allow_custom_value=True, info='Valid options: llama.cpp - fp16, q8_0, q4_0; ExLlamaV2 - fp16, fp8, q8, q6, q4; ExLlamaV3 - fp16, q2 to q8. For ExLlamaV3, you can type custom combinations for separate k/v bits (e.g. q4_q8).')
-                            shared.gradio['tensor_split'] = gr.Textbox(label='tensor_split', info='List of proportions to split the model across multiple GPUs. Example: 60,40')
+                            shared.gradio['gpu_layers'] = gr.Slider(label="gpu-layers", minimum=0, maximum=get_initial_gpu_layers_max(), step=1, value=shared.args.gpu_layers, info='Must be greater than 0 for the GPU to be used. ⚠️ Lower this value if you can\'t load the model.')
+                            shared.gradio['ctx_size'] = gr.Slider(label='ctx-size', minimum=256, maximum=131072, step=256, value=shared.args.ctx_size, info='Context length. ⚠️ Lower this value if you can\'t load the model.')
                            shared.gradio['gpu_split'] = gr.Textbox(label='gpu-split', info='Comma-separated list of VRAM (in GB) to use per GPU. Example: 20,7,7')
-                            shared.gradio['extra_flags'] = gr.Textbox(label='extra-flags', info='Additional flags to pass to llama-server. Format: "flag1=value1,flag2,flag3=value3". Example: "override-tensor=exps=CPU"', value=shared.args.extra_flags)
-                            shared.gradio['cpu_memory'] = gr.Number(label="Maximum CPU memory in GiB. Use this for CPU offloading.", value=shared.args.cpu_memory)
-                            shared.gradio['alpha_value'] = gr.Number(label='alpha_value', value=shared.args.alpha_value, precision=2, info='Positional embeddings alpha factor for NTK RoPE scaling. Recommended values (NTKv1): 1.75 for 1.5x context, 2.5 for 2x context. Use either this or compress_pos_emb, not both.')
-                            shared.gradio['rope_freq_base'] = gr.Number(label='rope_freq_base', value=shared.args.rope_freq_base, precision=0, info='Positional embeddings frequency base for NTK RoPE scaling. Related to alpha_value by rope_freq_base = 10000 * alpha_value ^ (64 / 63). 0 = from model.')
-                            shared.gradio['compress_pos_emb'] = gr.Number(label='compress_pos_emb', value=shared.args.compress_pos_emb, precision=2, info='Positional embeddings compression factor. Should be set to (context length) / (model\'s original context length). Equal to 1/rope_freq_scale.')
-                            shared.gradio['compute_dtype'] = gr.Dropdown(label="compute_dtype", choices=["bfloat16", "float16", "float32"], value=shared.args.compute_dtype, info='Used by load-in-4bit.')
-                            shared.gradio['quant_type'] = gr.Dropdown(label="quant_type", choices=["nf4", "fp4"], value=shared.args.quant_type, info='Used by load-in-4bit.')
-                            shared.gradio['num_experts_per_token'] = gr.Number(label="Number of experts per token", value=shared.args.num_experts_per_token, info='Only applies to MoE models like Mixtral.')
+                            shared.gradio['cache_type'] = gr.Dropdown(label="cache-type", choices=['fp16', 'q8_0', 'q4_0', 'fp8', 'q8', 'q7', 'q6', 'q5', 'q4', 'q3', 'q2'], value=shared.args.cache_type, allow_custom_value=True, info='Valid options: llama.cpp - fp16, q8_0, q4_0; ExLlamaV2 - fp16, fp8, q8, q6, q4; ExLlamaV3 - fp16, q2 to q8. For ExLlamaV3, you can type custom combinations for separate k/v bits (e.g. q4_q8).')
+                            shared.gradio['hqq_backend'] = gr.Dropdown(label="hqq_backend", choices=["PYTORCH", "PYTORCH_COMPILE", "ATEN"], value=shared.args.hqq_backend)

                        with gr.Column():
+                            shared.gradio['vram_info'] = gr.HTML(value=get_initial_vram_info())
+                            shared.gradio['flash_attn'] = gr.Checkbox(label="flash-attn", value=shared.args.flash_attn, info='Use flash-attention.')
+                            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)
                            shared.gradio['torch_compile'] = gr.Checkbox(label="torch-compile", value=shared.args.torch_compile, info='Compile the model with torch.compile for improved performance.')
-                            shared.gradio['flash_attn'] = gr.Checkbox(label="flash_attn", value=shared.args.flash_attn, info='Use flash-attention.')
                            shared.gradio['use_flash_attention_2'] = gr.Checkbox(label="use_flash_attention_2", value=shared.args.use_flash_attention_2, info='Set use_flash_attention_2=True while loading the model.')
-                            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['cpu'] = gr.Checkbox(label="cpu", value=shared.args.cpu, info='llama.cpp: Use llama-cpp-python compiled without GPU acceleration. Transformers: use PyTorch in CPU mode.')
-                            shared.gradio['disk'] = gr.Checkbox(label="disk", value=shared.args.disk)
-                            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 the performance.')
-                            shared.gradio['no_mmap'] = gr.Checkbox(label="no-mmap", value=shared.args.no_mmap)
-                            shared.gradio['mlock'] = gr.Checkbox(label="mlock", value=shared.args.mlock)
-                            shared.gradio['numa'] = gr.Checkbox(label="numa", value=shared.args.numa, info='NUMA support can help on some systems with non-uniform memory access.')
                            shared.gradio['use_double_quant'] = gr.Checkbox(label="use_double_quant", value=shared.args.use_double_quant, info='Used by load-in-4bit.')
-                            shared.gradio['use_eager_attention'] = gr.Checkbox(label="use_eager_attention", value=shared.args.use_eager_attention, info='Set attn_implementation= eager while loading the model.')
-                            shared.gradio['bf16'] = gr.Checkbox(label="bf16", value=shared.args.bf16)
                            shared.gradio['autosplit'] = gr.Checkbox(label="autosplit", value=shared.args.autosplit, info='Automatically split the model tensors across the available GPUs.')
                            shared.gradio['enable_tp'] = gr.Checkbox(label="enable_tp", value=shared.args.enable_tp, info='Enable Tensor Parallelism (TP).')
-                            shared.gradio['no_flash_attn'] = gr.Checkbox(label="no_flash_attn", value=shared.args.no_flash_attn)
-                            shared.gradio['no_xformers'] = gr.Checkbox(label="no_xformers", value=shared.args.no_xformers)
-                            shared.gradio['no_sdpa'] = gr.Checkbox(label="no_sdpa", value=shared.args.no_sdpa)
-                            shared.gradio['cfg_cache'] = gr.Checkbox(label="cfg-cache", value=shared.args.cfg_cache, info='Necessary to use CFG with this loader.')
                            shared.gradio['cpp_runner'] = gr.Checkbox(label="cpp-runner", value=shared.args.cpp_runner, info='Enable inference with ModelRunnerCpp, which is faster than the default ModelRunner.')
                            shared.gradio['trust_remote_code'] = gr.Checkbox(label="trust-remote-code", value=shared.args.trust_remote_code, info='Set trust_remote_code=True while loading the tokenizer/model. To enable this option, start the web UI with the --trust-remote-code flag.', interactive=shared.args.trust_remote_code)
-                            shared.gradio['no_use_fast'] = gr.Checkbox(label="no_use_fast", value=shared.args.no_use_fast, info='Set use_fast=False while loading the tokenizer.')
-                            shared.gradio['exllamav2_info'] = gr.Markdown("ExLlamav2_HF is recommended over ExLlamav2 for better integration with extensions and more consistent sampling behavior across loaders.")
                            shared.gradio['tensorrt_llm_info'] = gr.Markdown('* TensorRT-LLM has to be installed manually in a separate Python 3.10 environment at the moment. For a guide, consult the description of [this PR](https://github.com/oobabooga/text-generation-webui/pull/5715). \n\n* `ctx_size` is only used when `cpp-runner` is checked.\n\n* `cpp_runner` does not support streaming at the moment.')

                            # Speculative decoding
@ -99,15 +65,50 @@ def create_ui():
                                    shared.gradio['model_draft'] = gr.Dropdown(label="model-draft", choices=utils.get_available_models(), value=lambda: shared.args.model_draft, elem_classes='slim-dropdown', info='Draft model. Speculative decoding only works with models sharing the same vocabulary (e.g., same model family).', interactive=not mu)
                                    ui.create_refresh_button(shared.gradio['model_draft'], lambda: None, lambda: {'choices': utils.get_available_models()}, 'refresh-button', interactive=not mu)

-                                shared.gradio['draft_max'] = gr.Number(label="draft-max", precision=0, step=1, value=shared.args.draft_max, info='Number of tokens to draft for speculative decoding.')
                                shared.gradio['gpu_layers_draft'] = gr.Slider(label="gpu-layers-draft", minimum=0, maximum=256, value=shared.args.gpu_layers_draft, info='Number of layers to offload to the GPU for the draft model.')
+                                shared.gradio['draft_max'] = gr.Number(label="draft-max", precision=0, step=1, value=shared.args.draft_max, info='Number of tokens to draft for speculative decoding. Recommended value: 4.')
                                shared.gradio['device_draft'] = gr.Textbox(label="device-draft", value=shared.args.device_draft, info='Comma-separated list of devices to use for offloading the draft model. Example: CUDA0,CUDA1')
                                shared.gradio['ctx_size_draft'] = gr.Number(label="ctx-size-draft", precision=0, step=256, value=shared.args.ctx_size_draft, info='Size of the prompt context for the draft model. If 0, uses the same as the main model.')

-            with gr.Column():
-                with gr.Row():
-                    shared.gradio['autoload_model'] = gr.Checkbox(value=shared.settings['autoload_model'], label='Autoload the model', info='Whether to load the model as soon as it is selected in the Model dropdown.', interactive=not mu)
+                    gr.Markdown("## Other options")
+                    with gr.Accordion("See more options", open=False, elem_classes='tgw-accordion'):
+                        with gr.Row():
+                            with gr.Column():
+                                shared.gradio['threads'] = gr.Slider(label="threads", minimum=0, step=1, maximum=256, value=shared.args.threads)
+                                shared.gradio['threads_batch'] = gr.Slider(label="threads_batch", minimum=0, step=1, maximum=256, value=shared.args.threads_batch)
+                                shared.gradio['batch_size'] = gr.Slider(label="batch_size", minimum=1, maximum=4096, step=1, value=shared.args.batch_size)
+                                shared.gradio['tensor_split'] = gr.Textbox(label='tensor_split', info='List of proportions to split the model across multiple GPUs. Example: 60,40')
+                                shared.gradio['extra_flags'] = gr.Textbox(label='extra-flags', info='Additional flags to pass to llama-server. Format: "flag1=value1,flag2,flag3=value3". Example: "override-tensor=exps=CPU"', value=shared.args.extra_flags)
+                                shared.gradio['cpu_memory'] = gr.Number(label="Maximum CPU memory in GiB. Use this for CPU offloading.", value=shared.args.cpu_memory)
+                                shared.gradio['alpha_value'] = gr.Number(label='alpha_value', value=shared.args.alpha_value, precision=2, info='Positional embeddings alpha factor for NTK RoPE scaling. Recommended values (NTKv1): 1.75 for 1.5x context, 2.5 for 2x context. Use either this or compress_pos_emb, not both.')
+                                shared.gradio['rope_freq_base'] = gr.Number(label='rope_freq_base', value=shared.args.rope_freq_base, precision=0, info='Positional embeddings frequency base for NTK RoPE scaling. Related to alpha_value by rope_freq_base = 10000 * alpha_value ^ (64 / 63). 0 = from model.')
+                                shared.gradio['compress_pos_emb'] = gr.Number(label='compress_pos_emb', value=shared.args.compress_pos_emb, precision=2, info='Positional embeddings compression factor. Should be set to (context length) / (model\'s original context length). Equal to 1/rope_freq_scale.')
+                                shared.gradio['compute_dtype'] = gr.Dropdown(label="compute_dtype", choices=["bfloat16", "float16", "float32"], value=shared.args.compute_dtype, info='Used by load-in-4bit.')
+                                shared.gradio['quant_type'] = gr.Dropdown(label="quant_type", choices=["nf4", "fp4"], value=shared.args.quant_type, info='Used by load-in-4bit.')
+                                shared.gradio['num_experts_per_token'] = gr.Number(label="Number of experts per token", value=shared.args.num_experts_per_token, info='Only applies to MoE models like Mixtral.')

+                            with gr.Column():
+                                shared.gradio['cpu'] = gr.Checkbox(label="cpu", value=shared.args.cpu, info='llama.cpp: Use llama-cpp-python compiled without GPU acceleration. Transformers: use PyTorch in CPU mode.')
+                                shared.gradio['disk'] = gr.Checkbox(label="disk", value=shared.args.disk)
+                                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 the performance.')
+                                shared.gradio['no_mmap'] = gr.Checkbox(label="no-mmap", value=shared.args.no_mmap)
+                                shared.gradio['mlock'] = gr.Checkbox(label="mlock", value=shared.args.mlock)
+                                shared.gradio['numa'] = gr.Checkbox(label="numa", value=shared.args.numa, info='NUMA support can help on some systems with non-uniform memory access.')
+                                shared.gradio['use_eager_attention'] = gr.Checkbox(label="use_eager_attention", value=shared.args.use_eager_attention, info='Set attn_implementation= eager while loading the model.')
+                                shared.gradio['bf16'] = gr.Checkbox(label="bf16", value=shared.args.bf16)
+                                shared.gradio['no_flash_attn'] = gr.Checkbox(label="no_flash_attn", value=shared.args.no_flash_attn)
+                                shared.gradio['no_xformers'] = gr.Checkbox(label="no_xformers", value=shared.args.no_xformers)
+                                shared.gradio['no_sdpa'] = gr.Checkbox(label="no_sdpa", value=shared.args.no_sdpa)
+                                shared.gradio['cfg_cache'] = gr.Checkbox(label="cfg-cache", value=shared.args.cfg_cache, info='Necessary to use CFG with this loader.')
+                                shared.gradio['no_use_fast'] = gr.Checkbox(label="no_use_fast", value=shared.args.no_use_fast, info='Set use_fast=False while loading the tokenizer.')
+                                if not shared.args.portable:
+                                    with gr.Row():
+                                        shared.gradio['lora_menu'] = gr.Dropdown(multiselect=True, choices=utils.get_available_loras(), value=shared.lora_names, label='LoRA(s)', elem_classes='slim-dropdown', interactive=not mu)
+                                        ui.create_refresh_button(shared.gradio['lora_menu'], lambda: None, lambda: {'choices': utils.get_available_loras(), 'value': shared.lora_names}, 'refresh-button', interactive=not mu)
+                                        shared.gradio['lora_menu_apply'] = gr.Button(value='Apply LoRAs', elem_classes='refresh-button', interactive=not mu)
+
+            with gr.Column():
                with gr.Tab("Download"):
                    shared.gradio['custom_model_menu'] = gr.Textbox(label="Download model or LoRA", info="Enter the Hugging Face username/model path, for instance: facebook/galactica-125m. To specify a branch, add it at the end after a \":\" character like this: facebook/galactica-125m:main. To download a single file, enter its name in the second box.", interactive=not mu)
                    shared.gradio['download_specific_file'] = gr.Textbox(placeholder="File name (for GGUF models)", show_label=False, max_lines=1, interactive=not mu)
@ -132,11 +133,10 @@ def create_event_handlers():

    # In this event handler, the interface state is read and updated
    # with the model defaults (if any), and then the model is loaded
-    # unless "autoload_model" is unchecked
    shared.gradio['model_menu'].change(
        ui.gather_interface_values, gradio(shared.input_elements), gradio('interface_state')).then(
        handle_load_model_event_initial, gradio('model_menu', 'interface_state'), gradio(ui.list_interface_input_elements()) + gradio('interface_state'), show_progress=False).then(
-        load_model_wrapper, gradio('model_menu', 'loader', 'autoload_model'), gradio('model_status'), show_progress=True).success(
+        partial(load_model_wrapper, autoload=False), gradio('model_menu', 'loader'), gradio('model_status'), show_progress=True).success(
        handle_load_model_event_final, gradio('truncation_length', 'loader', 'interface_state'), gradio('truncation_length', 'filter_by_loader'), show_progress=False)

    shared.gradio['load_model'].click(
@ -145,15 +145,31 @@ def create_event_handlers():
        partial(load_model_wrapper, autoload=True), gradio('model_menu', 'loader'), gradio('model_status'), show_progress=True).success(
        handle_load_model_event_final, gradio('truncation_length', 'loader', 'interface_state'), gradio('truncation_length', 'filter_by_loader'), show_progress=False)

-    shared.gradio['unload_model'].click(handle_unload_model_click, None, gradio('model_status'), show_progress=False)
+    shared.gradio['unload_model'].click(handle_unload_model_click, None, gradio('model_status'), show_progress=False).then(
+        partial(update_gpu_layers_and_vram, auto_adjust=True), gradio('loader', 'model_menu', 'gpu_layers', 'ctx_size', 'cache_type'), gradio('vram_info', 'gpu_layers'), show_progress=False)
+
    shared.gradio['save_model_settings'].click(
        ui.gather_interface_values, gradio(shared.input_elements), gradio('interface_state')).then(
        save_model_settings, gradio('model_menu', 'interface_state'), gradio('model_status'), show_progress=False)

-    shared.gradio['lora_menu_apply'].click(load_lora_wrapper, gradio('lora_menu'), gradio('model_status'), show_progress=False)
+    # For ctx_size and cache_type - auto-adjust GPU layers
+    for param in ['ctx_size', 'cache_type']:
+        shared.gradio[param].change(
+            partial(update_gpu_layers_and_vram, auto_adjust=True),
+            gradio('loader', 'model_menu', 'gpu_layers', 'ctx_size', 'cache_type'),
+            gradio('vram_info', 'gpu_layers'), show_progress=False)
+
+    # For manual gpu_layers changes - only update VRAM
+    shared.gradio['gpu_layers'].change(
+        partial(update_gpu_layers_and_vram, auto_adjust=False),
+        gradio('loader', 'model_menu', 'gpu_layers', 'ctx_size', 'cache_type'),
+        gradio('vram_info'), show_progress=False)
+
+    if not shared.args.portable:
+        shared.gradio['lora_menu_apply'].click(load_lora_wrapper, gradio('lora_menu'), gradio('model_status'), show_progress=False)
+
    shared.gradio['download_model_button'].click(download_model_wrapper, gradio('custom_model_menu', 'download_specific_file'), gradio('model_status'), show_progress=True)
    shared.gradio['get_file_list'].click(partial(download_model_wrapper, return_links=True), gradio('custom_model_menu', 'download_specific_file'), gradio('model_status'), show_progress=True)
-    shared.gradio['autoload_model'].change(lambda x: gr.update(visible=not x), gradio('autoload_model'), gradio('load_model'))
    shared.gradio['customized_template_submit'].click(save_instruction_template, gradio('model_menu', 'customized_template'), gradio('model_status'), show_progress=True)


@ -192,6 +208,26 @@ def load_lora_wrapper(selected_loras):

 def download_model_wrapper(repo_id, specific_file, progress=gr.Progress(), return_links=False, check=False):
    try:
+        # Handle direct GGUF URLs
+        if repo_id.startswith("https://") and ("huggingface.co" in repo_id) and (repo_id.endswith(".gguf") or repo_id.endswith(".gguf?download=true")):
+            try:
+                path = repo_id.split("huggingface.co/")[1]
+
+                # Extract the repository ID (first two parts of the path)
+                parts = path.split("/")
+                if len(parts) >= 2:
+                    extracted_repo_id = f"{parts[0]}/{parts[1]}"
+
+                    # Extract the filename (last part of the path)
+                    filename = repo_id.split("/")[-1]
+                    if "?download=true" in filename:
+                        filename = filename.replace("?download=true", "")
+
+                    repo_id = extracted_repo_id
+                    specific_file = filename
+            except:
+                pass
+
        if repo_id == "":
            yield ("Please enter a model path")
            return
@ -205,6 +241,18 @@ def download_model_wrapper(repo_id, specific_file, progress=gr.Progress(), retur

        yield ("Getting the download links from Hugging Face")
        links, sha256, is_lora, is_llamacpp = downloader.get_download_links_from_huggingface(model, branch, text_only=False, specific_file=specific_file)
+
+        # Check for multiple GGUF files
+        gguf_files = [link for link in links if link.lower().endswith('.gguf')]
+        if len(gguf_files) > 1 and not specific_file:
+            output = "Multiple GGUF files found. Please copy one of the following filenames to the 'File name' field:\n\n```\n"
+            for link in gguf_files:
+                output += f"{Path(link).name}\n"
+
+            output += "```"
+            yield output
+            return
+
        if return_links:
            output = "```\n"
            for link in links:
@ -252,10 +300,34 @@ def update_truncation_length(current_length, state):
    return current_length


+def get_initial_vram_info():
+    if shared.model_name != 'None' and shared.args.loader == 'llama.cpp':
+        return update_gpu_layers_and_vram(
+            shared.args.loader,
+            shared.model_name,
+            shared.args.gpu_layers,
+            shared.args.ctx_size,
+            shared.args.cache_type,
+            auto_adjust=False,
+            for_ui=True
+        )
+
+    return "<div id=\"vram-info\"'>Estimated VRAM to load the model:</span>"
+
+
+def get_initial_gpu_layers_max():
+    if shared.model_name != 'None' and shared.args.loader == 'llama.cpp':
+        model_settings = get_model_metadata(shared.model_name)
+        return model_settings.get('gpu_layers', 256)
+
+    return 256
+
+
 def handle_load_model_event_initial(model, state):
    state = apply_model_settings_to_state(model, state)
    output = ui.apply_interface_values(state)
-    update_model_parameters(state)
+    update_model_parameters(state)  # This updates the command-line flags
+
    return output + [state]


--- a/modules/ui_parameters.py
+++ b/modules/ui_parameters.py
@ -21,7 +21,7 @@ def create_ui(default_preset):
                        shared.gradio['random_preset'] = gr.Button('🎲', elem_classes='refresh-button')

                with gr.Column():
-                    shared.gradio['filter_by_loader'] = gr.Dropdown(label="Filter by loader", choices=["All"] + list(loaders.loaders_and_params.keys()), value="All", elem_classes='slim-dropdown')
+                    shared.gradio['filter_by_loader'] = gr.Dropdown(label="Filter by loader", choices=["All"] + list(loaders.loaders_and_params.keys()) if not shared.args.portable else ['llama.cpp'], value="All", elem_classes='slim-dropdown')

            with gr.Row():
                with gr.Column():
@ -71,8 +71,6 @@ def create_ui(default_preset):
                            shared.gradio['max_new_tokens'] = gr.Slider(minimum=shared.settings['max_new_tokens_min'], maximum=shared.settings['max_new_tokens_max'], value=shared.settings['max_new_tokens'], step=1, label='max_new_tokens', info='⚠️ Setting this too high can cause prompt truncation.')
                            shared.gradio['prompt_lookup_num_tokens'] = gr.Slider(value=shared.settings['prompt_lookup_num_tokens'], minimum=0, maximum=10, step=1, label='prompt_lookup_num_tokens', info='Activates Prompt Lookup Decoding.')
                            shared.gradio['max_tokens_second'] = gr.Slider(value=shared.settings['max_tokens_second'], minimum=0, maximum=20, step=1, label='Maximum tokens/second', info='To make text readable in real time.')
-                            shared.gradio['max_updates_second'] = gr.Slider(value=shared.settings['max_updates_second'], minimum=0, maximum=24, step=1, label='Maximum UI updates/second', info='Set this if you experience lag in the UI during streaming.')
-
                with gr.Column():
                    with gr.Row():
                        with gr.Column():
@ -82,7 +80,7 @@ def create_ui(default_preset):
                            shared.gradio['auto_max_new_tokens'] = gr.Checkbox(value=shared.settings['auto_max_new_tokens'], label='auto_max_new_tokens', info='Expand max_new_tokens to the available context length.')
                            shared.gradio['ban_eos_token'] = gr.Checkbox(value=shared.settings['ban_eos_token'], label='Ban the eos_token', info='Forces the model to never end the generation prematurely.')
                            shared.gradio['add_bos_token'] = gr.Checkbox(value=shared.settings['add_bos_token'], label='Add the bos_token to the beginning of prompts', info='Disabling this can make the replies more creative.')
-                            shared.gradio['enable_thinking'] = gr.Checkbox(value=shared.settings['enable_thinking'], label='enable_thinking', info='Used by Qwen3 to toggle <think> mode.')
+                            shared.gradio['enable_thinking'] = gr.Checkbox(value=shared.settings['enable_thinking'], label='Enable thinking', info='Used by Qwen3 to toggle <think> mode.')
                            shared.gradio['skip_special_tokens'] = gr.Checkbox(value=shared.settings['skip_special_tokens'], label='Skip special tokens', info='Some specific models need this unset.')
                            shared.gradio['stream'] = gr.Checkbox(value=shared.settings['stream'], label='Activate text streaming')
                            shared.gradio['static_cache'] = gr.Checkbox(value=shared.settings['static_cache'], label='Static KV cache', info='Use a static cache for improved performance.')
--- a/modules/ui_session.py
+++ b/modules/ui_session.py
@ -23,11 +23,15 @@ def create_ui():
                        shared.gradio['bool_menu'] = gr.CheckboxGroup(choices=get_boolean_arguments(), value=get_boolean_arguments(active=True), label="Boolean command-line flags", elem_classes='checkboxgroup-table')

            with gr.Column():
-                extension_name = gr.Textbox(lines=1, label='Install or update an extension', info='Enter the GitHub URL below and press Enter. For a list of extensions, see: https://github.com/oobabooga/text-generation-webui-extensions ⚠️  WARNING ⚠️ : extensions can execute arbitrary code. Make sure to inspect their source code before activating them.', interactive=not mu)
-                extension_status = gr.Markdown()
+                if not shared.args.portable:
+                    extension_name = gr.Textbox(lines=1, label='Install or update an extension', info='Enter the GitHub URL below and press Enter. For a list of extensions, see: https://github.com/oobabooga/text-generation-webui-extensions ⚠️  WARNING ⚠️ : extensions can execute arbitrary code. Make sure to inspect their source code before activating them.', interactive=not mu)
+                    extension_status = gr.Markdown()
+                else:
+                    pass

        shared.gradio['theme_state'] = gr.Textbox(visible=False, value='dark' if shared.settings['dark_theme'] else 'light')
-        extension_name.submit(clone_or_pull_repository, extension_name, extension_status, show_progress=False)
+        if not shared.args.portable:
+            extension_name.submit(clone_or_pull_repository, extension_name, extension_status, show_progress=False)

        # Reset interface event
        shared.gradio['reset_interface'].click(
--- a/modules/utils.py
+++ b/modules/utils.py
@ -72,6 +72,20 @@ def natural_keys(text):
    return [atoi(c) for c in re.split(r'(\d+)', text)]


+def check_model_loaded():
+    if shared.model_name == 'None' or shared.model is None:
+        if len(get_available_models()) <= 1:
+            error_msg = "No model is loaded.\n\nTo get started:\n1) Place a GGUF file in your user_data/models folder\n2) Go to the Model tab and select it"
+            logger.error(error_msg)
+            return False, error_msg
+        else:
+            error_msg = "No model is loaded. Please select one in the Model tab."
+            logger.error(error_msg)
+            return False, error_msg
+
+    return True, None
+
+
 def get_available_models():
    # Get all GGUF files
    gguf_files = get_available_ggufs()
--- a/one_click.py
+++ b/one_click.py
@ -126,7 +126,7 @@ def check_env():
        sys.exit(1)

    # Ensure this is a new environment and not the base environment
-    if os.environ["CONDA_DEFAULT_ENV"] == "base":
+    if os.environ.get("CONDA_DEFAULT_ENV", "") == "base":
        print("Create an environment for this project and activate it. Exiting...")
        sys.exit(1)

@ -222,7 +222,7 @@ def update_pytorch_and_python():
    if "+cu" in torver:
        install_cmd = f"{base_cmd} --index-url https://download.pytorch.org/whl/cu124"
    elif "+rocm" in torver:
-        install_cmd = f"{base_cmd} --index-url https://download.pytorch.org/whl/rocm6.1"
+        install_cmd = f"{base_cmd} --index-url https://download.pytorch.org/whl/rocm6.2.4"
    elif "+cpu" in torver:
        install_cmd = f"{base_cmd} --index-url https://download.pytorch.org/whl/cpu"
    elif "+cxx11" in torver:
@ -273,7 +273,7 @@ def install_webui():
            "What is your GPU?",
            {
                'A': 'NVIDIA - CUDA 12.4',
-                'B': 'AMD - Linux/macOS only, requires ROCm 6.1',
+                'B': 'AMD - Linux/macOS only, requires ROCm 6.2.4',
                'C': 'Apple M Series',
                'D': 'Intel Arc (beta)',
                'N': 'CPU mode'
@ -314,7 +314,7 @@ def install_webui():
    if selected_gpu == "NVIDIA":
        install_pytorch += "--index-url https://download.pytorch.org/whl/cu124"
    elif selected_gpu == "AMD":
-        install_pytorch += "--index-url https://download.pytorch.org/whl/rocm6.1"
+        install_pytorch += "--index-url https://download.pytorch.org/whl/rocm6.2.4"
    elif selected_gpu in ["APPLE", "NONE"]:
        install_pytorch += "--index-url https://download.pytorch.org/whl/cpu"
    elif selected_gpu == "INTEL":
--- a/requirements/full/requirements.txt
+++ b/requirements/full/requirements.txt
@ -30,12 +30,12 @@ sse-starlette==1.6.5
 tiktoken

 # CUDA wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+cu124-py3-none-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+cu124-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
-https://github.com/oobabooga/exllamav3/releases/download/v0.0.1a6/exllamav3-0.0.1a6+cu124.torch2.6.0-cp311-cp311-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
-https://github.com/oobabooga/exllamav3/releases/download/v0.0.1a6/exllamav3-0.0.1a6+cu124.torch2.6.0-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
-https://github.com/oobabooga/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9+cu124.torch2.6.0-cp311-cp311-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
-https://github.com/oobabooga/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9+cu124.torch2.6.0-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
-https://github.com/oobabooga/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9-py3-none-any.whl; platform_system == "Linux" and platform_machine != "x86_64"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+cu124-py3-none-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+cu124-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
+https://github.com/oobabooga/exllamav3/releases/download/v0.0.1a9/exllamav3-0.0.1a9+cu124.torch2.6.0-cp311-cp311-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
+https://github.com/oobabooga/exllamav3/releases/download/v0.0.1a9/exllamav3-0.0.1a9+cu124.torch2.6.0-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
+https://github.com/turboderp-org/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9+cu124.torch2.6.0-cp311-cp311-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
+https://github.com/turboderp-org/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9+cu124.torch2.6.0-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
+https://github.com/turboderp-org/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9-py3-none-any.whl; platform_system == "Linux" and platform_machine != "x86_64"
 https://github.com/oobabooga/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu124torch2.6.0cxx11abiFALSE-cp311-cp311-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
 https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.6cxx11abiFALSE-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
--- a/requirements/full/requirements_amd.txt
+++ b/requirements/full/requirements_amd.txt
@ -29,6 +29,7 @@ sse-starlette==1.6.5
 tiktoken

 # AMD wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+rocm6.1.2-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
-https://github.com/oobabooga/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9+rocm6.1.torch2.6.0-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
-https://github.com/oobabooga/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9-py3-none-any.whl; platform_system != "Darwin" and platform_machine != "x86_64"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+vulkan-py3-none-win_amd64.whl; platform_system == "Windows"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+vulkan-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/turboderp-org/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9+rocm6.2.4.torch2.6.0-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
+https://github.com/turboderp-org/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9-py3-none-any.whl; platform_system != "Darwin" and platform_machine != "x86_64"
--- a/requirements/full/requirements_amd_noavx2.txt
+++ b/requirements/full/requirements_amd_noavx2.txt
@ -29,6 +29,7 @@ sse-starlette==1.6.5
 tiktoken

 # AMD wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+rocm6.1.2avx-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
-https://github.com/oobabooga/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9+rocm6.1.torch2.6.0-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
-https://github.com/oobabooga/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9-py3-none-any.whl; platform_system != "Darwin" and platform_machine != "x86_64"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+vulkanavx-py3-none-win_amd64.whl; platform_system == "Windows"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+vulkanavx-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/turboderp-org/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9+rocm6.2.4.torch2.6.0-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
+https://github.com/turboderp-org/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9-py3-none-any.whl; platform_system != "Darwin" and platform_machine != "x86_64"
--- a/requirements/full/requirements_apple_intel.txt
+++ b/requirements/full/requirements_apple_intel.txt
@ -29,7 +29,7 @@ sse-starlette==1.6.5
 tiktoken

 # Mac wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0-py3-none-macosx_15_0_x86_64.whl; platform_system == "Darwin" and platform_release >= "24.0.0" and platform_release < "25.0.0" and python_version == "3.11"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0-py3-none-macosx_14_0_x86_64.whl; platform_system == "Darwin" and platform_release >= "23.0.0" and platform_release < "24.0.0" and python_version == "3.11"
-https://github.com/oobabooga/exllamav3/releases/download/v0.0.1a6/exllamav3-0.0.1a6-py3-none-any.whl
-https://github.com/oobabooga/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9-py3-none-any.whl
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0-py3-none-macosx_15_0_x86_64.whl; platform_system == "Darwin" and platform_release >= "24.0.0" and platform_release < "25.0.0" and python_version == "3.11"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0-py3-none-macosx_14_0_x86_64.whl; platform_system == "Darwin" and platform_release >= "23.0.0" and platform_release < "24.0.0" and python_version == "3.11"
+https://github.com/oobabooga/exllamav3/releases/download/v0.0.1a9/exllamav3-0.0.1a9-py3-none-any.whl
+https://github.com/turboderp-org/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9-py3-none-any.whl
--- a/requirements/full/requirements_apple_silicon.txt
+++ b/requirements/full/requirements_apple_silicon.txt
@ -29,8 +29,8 @@ sse-starlette==1.6.5
 tiktoken

 # Mac wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0-py3-none-macosx_15_0_arm64.whl; platform_system == "Darwin" and platform_release >= "24.0.0" and platform_release < "25.0.0" and python_version == "3.11"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0-py3-none-macosx_14_0_arm64.whl; platform_system == "Darwin" and platform_release >= "23.0.0" and platform_release < "24.0.0" and python_version == "3.11"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0-py3-none-macosx_13_0_arm64.whl; platform_system == "Darwin" and platform_release >= "22.0.0" and platform_release < "23.0.0" and python_version == "3.11"
-https://github.com/oobabooga/exllamav3/releases/download/v0.0.1a6/exllamav3-0.0.1a6-py3-none-any.whl
-https://github.com/oobabooga/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9-py3-none-any.whl
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0-py3-none-macosx_15_0_arm64.whl; platform_system == "Darwin" and platform_release >= "24.0.0" and platform_release < "25.0.0" and python_version == "3.11"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0-py3-none-macosx_14_0_arm64.whl; platform_system == "Darwin" and platform_release >= "23.0.0" and platform_release < "24.0.0" and python_version == "3.11"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0-py3-none-macosx_13_0_arm64.whl; platform_system == "Darwin" and platform_release >= "22.0.0" and platform_release < "23.0.0" and python_version == "3.11"
+https://github.com/oobabooga/exllamav3/releases/download/v0.0.1a9/exllamav3-0.0.1a9-py3-none-any.whl
+https://github.com/turboderp-org/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9-py3-none-any.whl
--- a/requirements/full/requirements_cpu_only.txt
+++ b/requirements/full/requirements_cpu_only.txt
@ -29,5 +29,5 @@ sse-starlette==1.6.5
 tiktoken

 # llama.cpp (CPU only, AVX2)
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+cpuavx2-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+cpuavx2-py3-none-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+cpuavx2-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+cpuavx2-py3-none-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
--- a/requirements/full/requirements_cpu_only_noavx2.txt
+++ b/requirements/full/requirements_cpu_only_noavx2.txt
@ -29,5 +29,5 @@ sse-starlette==1.6.5
 tiktoken

 # llama.cpp (CPU only, no AVX2)
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+cpuavx-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+cpuavx-py3-none-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+cpuavx-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+cpuavx-py3-none-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
--- a/requirements/full/requirements_noavx2.txt
+++ b/requirements/full/requirements_noavx2.txt
@ -30,12 +30,12 @@ sse-starlette==1.6.5
 tiktoken

 # CUDA wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+cu124avx-py3-none-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+cu124avx-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
-https://github.com/oobabooga/exllamav3/releases/download/v0.0.1a6/exllamav3-0.0.1a6+cu124.torch2.6.0-cp311-cp311-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
-https://github.com/oobabooga/exllamav3/releases/download/v0.0.1a6/exllamav3-0.0.1a6+cu124.torch2.6.0-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
-https://github.com/oobabooga/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9+cu124.torch2.6.0-cp311-cp311-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
-https://github.com/oobabooga/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9+cu124.torch2.6.0-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
-https://github.com/oobabooga/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9-py3-none-any.whl; platform_system == "Linux" and platform_machine != "x86_64"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+cu124avx-py3-none-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+cu124avx-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
+https://github.com/oobabooga/exllamav3/releases/download/v0.0.1a9/exllamav3-0.0.1a9+cu124.torch2.6.0-cp311-cp311-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
+https://github.com/oobabooga/exllamav3/releases/download/v0.0.1a9/exllamav3-0.0.1a9+cu124.torch2.6.0-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
+https://github.com/turboderp-org/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9+cu124.torch2.6.0-cp311-cp311-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
+https://github.com/turboderp-org/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9+cu124.torch2.6.0-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
+https://github.com/turboderp-org/exllamav2/releases/download/v0.2.9/exllamav2-0.2.9-py3-none-any.whl; platform_system == "Linux" and platform_machine != "x86_64"
 https://github.com/oobabooga/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu124torch2.6.0cxx11abiFALSE-cp311-cp311-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
 https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.6cxx11abiFALSE-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
--- a/requirements/portable/requirements.txt
+++ b/requirements/portable/requirements.txt
@ -15,5 +15,5 @@ sse-starlette==1.6.5
 tiktoken

 # CUDA wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+cu124-py3-none-win_amd64.whl; platform_system == "Windows"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+cu124-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+cu124-py3-none-win_amd64.whl; platform_system == "Windows"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+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
@ -1,18 +0,0 @@
-fastapi==0.112.4
-gradio==4.37.*
-jinja2==3.1.6
-markdown
-numpy==1.26.*
-pydantic==2.8.2
-pyyaml
-requests
-rich
-tqdm
-
-# API
-flask_cloudflared==0.0.14
-sse-starlette==1.6.5
-tiktoken
-
-# AMD wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+rocm6.1.2-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
--- a/requirements/portable/requirements_amd_noavx2.txt
+++ b/requirements/portable/requirements_amd_noavx2.txt
@ -1,18 +0,0 @@
-fastapi==0.112.4
-gradio==4.37.*
-jinja2==3.1.6
-markdown
-numpy==1.26.*
-pydantic==2.8.2
-pyyaml
-requests
-rich
-tqdm
-
-# API
-flask_cloudflared==0.0.14
-sse-starlette==1.6.5
-tiktoken
-
-# AMD wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+rocm6.1.2avx-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
@ -15,5 +15,5 @@ sse-starlette==1.6.5
 tiktoken

 # Mac wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0-py3-none-macosx_15_0_x86_64.whl; platform_system == "Darwin" and platform_release >= "24.0.0" and platform_release < "25.0.0"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0-py3-none-macosx_14_0_x86_64.whl; platform_system == "Darwin" and platform_release >= "23.0.0" and platform_release < "24.0.0"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0-py3-none-macosx_15_0_x86_64.whl; platform_system == "Darwin" and platform_release >= "24.0.0" and platform_release < "25.0.0"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0-py3-none-macosx_14_0_x86_64.whl; platform_system == "Darwin" and platform_release >= "23.0.0" and platform_release < "24.0.0"
--- a/requirements/portable/requirements_apple_silicon.txt
+++ b/requirements/portable/requirements_apple_silicon.txt
@ -15,6 +15,6 @@ sse-starlette==1.6.5
 tiktoken

 # Mac wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0-py3-none-macosx_15_0_arm64.whl; platform_system == "Darwin" and platform_release >= "24.0.0" and platform_release < "25.0.0"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0-py3-none-macosx_14_0_arm64.whl; platform_system == "Darwin" and platform_release >= "23.0.0" and platform_release < "24.0.0"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0-py3-none-macosx_13_0_arm64.whl; platform_system == "Darwin" and platform_release >= "22.0.0" and platform_release < "23.0.0"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0-py3-none-macosx_15_0_arm64.whl; platform_system == "Darwin" and platform_release >= "24.0.0" and platform_release < "25.0.0"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0-py3-none-macosx_14_0_arm64.whl; platform_system == "Darwin" and platform_release >= "23.0.0" and platform_release < "24.0.0"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0-py3-none-macosx_13_0_arm64.whl; platform_system == "Darwin" and platform_release >= "22.0.0" and platform_release < "23.0.0"
--- a/requirements/portable/requirements_cpu_only.txt
+++ b/requirements/portable/requirements_cpu_only.txt
@ -15,5 +15,5 @@ sse-starlette==1.6.5
 tiktoken

 # llama.cpp (CPU only, AVX2)
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+cpuavx2-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+cpuavx2-py3-none-win_amd64.whl; platform_system == "Windows"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+cpuavx2-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+cpuavx2-py3-none-win_amd64.whl; platform_system == "Windows"
--- a/requirements/portable/requirements_cpu_only_noavx2.txt
+++ b/requirements/portable/requirements_cpu_only_noavx2.txt
@ -15,5 +15,5 @@ sse-starlette==1.6.5
 tiktoken

 # llama.cpp (CPU only, no AVX2)
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+cpuavx-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+cpuavx-py3-none-win_amd64.whl; platform_system == "Windows"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+cpuavx-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+cpuavx-py3-none-win_amd64.whl; platform_system == "Windows"
--- a/requirements/portable/requirements_noavx2.txt
+++ b/requirements/portable/requirements_noavx2.txt
@ -15,5 +15,5 @@ sse-starlette==1.6.5
 tiktoken

 # CUDA wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+cu124avx-py3-none-win_amd64.whl; platform_system == "Windows"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+cu124avx-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+cu124avx-py3-none-win_amd64.whl; platform_system == "Windows"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+cu124avx-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
--- a/requirements/portable/requirements_vulkan.txt
+++ b/requirements/portable/requirements_vulkan.txt
@ -15,5 +15,5 @@ sse-starlette==1.6.5
 tiktoken

 # CUDA wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+vulkan-py3-none-win_amd64.whl; platform_system == "Windows"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+vulkan-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+vulkan-py3-none-win_amd64.whl; platform_system == "Windows"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+vulkan-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
--- a/requirements/portable/requirements_vulkan_noavx2.txt
+++ b/requirements/portable/requirements_vulkan_noavx2.txt
@ -15,5 +15,5 @@ sse-starlette==1.6.5
 tiktoken

 # CUDA wheels
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+vulkanavx-py3-none-win_amd64.whl; platform_system == "Windows"
-https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.8.0/llama_cpp_binaries-0.8.0+vulkanavx-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+vulkanavx-py3-none-win_amd64.whl; platform_system == "Windows"
+https://github.com/oobabooga/llama-cpp-binaries/releases/download/v0.12.0/llama_cpp_binaries-0.12.0+vulkanavx-py3-none-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
--- a/server.py
+++ b/server.py
@ -51,6 +51,7 @@ from modules.models import load_model, unload_model_if_idle
 from modules.models_settings import (
    get_fallback_settings,
    get_model_metadata,
+    update_gpu_layers_and_vram,
    update_model_parameters
 )
 from modules.shared import do_cmd_flags_warnings
@ -90,7 +91,7 @@ def create_interface():
        'instruction_template_str': shared.settings['instruction_template_str'],
        'prompt_menu-default': shared.settings['prompt-default'],
        'prompt_menu-notebook': shared.settings['prompt-notebook'],
-        'filter_by_loader': shared.args.loader or 'All'
+        'filter_by_loader': (shared.args.loader or 'All') if not shared.args.portable else 'llama.cpp'
    })

    if Path("user_data/cache/pfp_character.png").exists():
@ -127,7 +128,8 @@ def create_interface():

        ui_parameters.create_ui(shared.settings['preset'])  # Parameters tab
        ui_model_menu.create_ui()  # Model tab
-        training.create_ui()  # Training tab
+        if not shared.args.portable:
+            training.create_ui()  # Training tab
        ui_session.create_ui()  # Session tab

        # Generation events
@ -247,6 +249,20 @@ if __name__ == "__main__":
        model_settings = get_model_metadata(model_name)
        update_model_parameters(model_settings, initial=True)  # hijack the command-line arguments

+        # Auto-adjust GPU layers if not provided by user and it's a llama.cpp model
+        if 'gpu_layers' not in shared.provided_arguments and shared.args.loader == 'llama.cpp' and 'gpu_layers' in model_settings:
+            vram_usage, adjusted_layers = update_gpu_layers_and_vram(
+                shared.args.loader,
+                model_name,
+                model_settings['gpu_layers'],
+                shared.args.ctx_size,
+                shared.args.cache_type,
+                auto_adjust=True,
+                for_ui=False
+            )
+
+            shared.args.gpu_layers = adjusted_layers
+
        # Load the model
        shared.model, shared.tokenizer = load_model(model_name)
        if shared.args.lora:
--- a/start_linux.sh
+++ b/start_linux.sh
@ -1,10 +1,15 @@
 #!/usr/bin/env bash

+# environment isolation
+export PYTHONNOUSERSITE=1
+unset PYTHONPATH
+unset PYTHONHOME
+
 cd "$(dirname "${BASH_SOURCE[0]}")"

 # Portable install case
 if [ -d "portable_env" ]; then
-    ./portable_env/bin/python3 server.py --api --auto-launch "$@"
+    ./portable_env/bin/python3 server.py --portable --api --auto-launch "$@"
    exit $?
 fi

@ -61,10 +66,6 @@ if [ ! -e "$INSTALL_ENV_DIR/bin/python" ]; then
    exit
 fi

-# environment isolation
-export PYTHONNOUSERSITE=1
-unset PYTHONPATH
-unset PYTHONHOME
 export CUDA_PATH="$INSTALL_ENV_DIR"
 export CUDA_HOME="$CUDA_PATH"

--- a/start_macos.sh
+++ b/start_macos.sh
@ -1,10 +1,15 @@
 #!/bin/bash

+# environment isolation
+export PYTHONNOUSERSITE=1
+unset PYTHONPATH
+unset PYTHONHOME
+
 cd "$(dirname "${BASH_SOURCE[0]}")"

 # Portable install case
 if [ -d "portable_env" ]; then
-    ./portable_env/bin/python3 server.py --api --auto-launch --api-port 5005 "$@"
+    ./portable_env/bin/python3 server.py --portable --api --auto-launch --api-port 5005 "$@"
    exit $?
 fi

@ -61,10 +66,6 @@ if [ ! -e "$INSTALL_ENV_DIR/bin/python" ]; then
    exit
 fi

-# environment isolation
-export PYTHONNOUSERSITE=1
-unset PYTHONPATH
-unset PYTHONHOME
 export CUDA_PATH="$INSTALL_ENV_DIR"
 export CUDA_HOME="$CUDA_PATH"

--- a/start_windows.bat
+++ b/start_windows.bat
@ -1,11 +1,16 @@
@echo off
 setlocal enabledelayedexpansion

+@rem environment isolation
+set PYTHONNOUSERSITE=1
+set PYTHONPATH=
+set PYTHONHOME=
+
 cd /D "%~dp0"

@rem Portable install case
 if exist "portable_env" (
-    .\portable_env\python.exe server.py --api --auto-launch %*
+    .\portable_env\python.exe server.py --portable --api --auto-launch %*
    exit /b %errorlevel%
 )

@ -87,10 +92,6 @@ if not exist "%INSTALL_ENV_DIR%" (
@rem check if conda environment was actually created
 if not exist "%INSTALL_ENV_DIR%\python.exe" ( echo. && echo Conda environment is empty. && goto end )

-@rem environment isolation
-set PYTHONNOUSERSITE=1
-set PYTHONPATH=
-set PYTHONHOME=
 set "CUDA_PATH=%INSTALL_ENV_DIR%"
 set "CUDA_HOME=%CUDA_PATH%"

--- a/user_data/settings-template.yaml
+++ b/user_data/settings-template.yaml
@ -18,7 +18,6 @@ max_new_tokens_min: 1
 max_new_tokens_max: 4096
 prompt_lookup_num_tokens: 0
 max_tokens_second: 0
-max_updates_second: 12
 auto_max_new_tokens: true
 ban_eos_token: false
 add_bos_token: true
@ -31,7 +30,6 @@ seed: -1
 custom_stopping_strings: ''
 custom_token_bans: ''
 negative_prompt: ''
-autoload_model: false
 dark_theme: true
 default_extensions: []
 instruction_template_str: |-