Merge pull request oobabooga#6421 from oobabooga/dev

Merge dev branch

Author: oobabooga

README.md

@@ -10,33 +10,29 @@ Its goal is to become the [AUTOMATIC1111/stable-diffusion-webui](https://github.
 
 ## Features
 
-* Multiple backends for text generation in a single UI and API, including [Transformers](https://github.com/huggingface/transformers), [llama.cpp](https://github.com/ggerganov/llama.cpp) (through [llama-cpp-python](https://github.com/abetlen/llama-cpp-python)), [ExLlamaV2](https://github.com/turboderp/exllamav2), [AutoGPTQ](https://github.com/PanQiWei/AutoGPTQ), and [TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM). [AutoAWQ](https://github.com/casper-hansen/AutoAWQ), [HQQ](https://github.com/mobiusml/hqq), and [AQLM](https://github.com/Vahe1994/AQLM) are also supported through the Transformers loader.
-* OpenAI-compatible API server with Chat and Completions endpoints – see the [examples](https://github.com/oobabooga/text-generation-webui/wiki/12-%E2%80%90-OpenAI-API#examples).
-* Automatic prompt formatting for each model using the Jinja2 template in its metadata.
-* Three chat modes: `instruct`, `chat-instruct`, and `chat`, allowing for both instruction-following and casual conversations with characters. `chat-instruct` mode automatically applies the model's template to the chat prompt, ensuring high-quality outputs without manual setup.
-* "Past chats" menu to quickly switch between conversations and start new ones.
-* Free-form generation in the Default/Notebook tabs without being limited to chat turns. Send formatted chat conversations from the Chat tab to these tabs.
-* Multiple sampling parameters and generation options for sophisticated text generation control.
-* Easy switching between different models through the UI without restarting, using the "Model" tab.
-* Simple LoRA fine-tuning tool to customize models with your data.
-* All in one folder. The requirements are installed in a self-contained `installer_files` folder that doesn't interfere with the system's environment.
-* Extensions support, including numerous built-in and user-contributed extensions. See [the wiki](https://github.com/oobabooga/text-generation-webui/wiki/07-%E2%80%90-Extensions) and [the extensions directory](https://github.com/oobabooga/text-generation-webui-extensions) for details.
+- Supports multiple text generation backends in one UI/API, including [Transformers](https://github.com/huggingface/transformers), [llama.cpp](https://github.com/ggerganov/llama.cpp), and [ExLlamaV2](https://github.com/turboderp/exllamav2). [TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM), [AutoGPTQ](https://github.com/PanQiWei/AutoGPTQ), [AutoAWQ](https://github.com/casper-hansen/AutoAWQ), [HQQ](https://github.com/mobiusml/hqq), and [AQLM](https://github.com/Vahe1994/AQLM) are also supported but you need to install them manually.
+- 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).
+- Automatic prompt formatting using Jinja2 templates.
+- Three chat modes: `instruct`, `chat-instruct`, and `chat`, with automatic prompt templates in `chat-instruct`.
+- "Past chats" menu to quickly switch between conversations.
+- 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.
+- Simple LoRA fine-tuning tool.
+- Requirements installed in a self-contained `installer_files` directory that doesn't interfere with the system environment.
+- 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.
 
 ## How to install
 
-1) Clone or [download](https://github.com/oobabooga/text-generation-webui/archive/refs/heads/main.zip) the repository.
-2) Run the `start_linux.sh`, `start_windows.bat`, `start_macos.sh`, or `start_wsl.bat` script depending on your OS.
+1) Clone or [download the repository](https://github.com/oobabooga/text-generation-webui/archive/refs/heads/main.zip).
+2) Run the script that matches your OS: `start_linux.sh`, `start_windows.bat`, `start_macos.sh`, or `start_wsl.bat`.
 3) Select your GPU vendor when asked.
 4) Once the installation ends, browse to `http://localhost:7860`.
 5) Have fun!
 
-To restart the web UI in the future, run the `start_` script again.
+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.
 
-This script creates an `installer_files` folder where it sets up the project's requirements. If you need to reinstall the requirements, just delete that folder and start the web UI again.
-
-The script accepts command-line flags, such as `./start_linux.sh --help`. Alternatively, you can edit the `CMD_FLAGS.txt` file with a text editor and add your flags there, such as `--api` in case you need to use the API.
-
-To get updates in the future, 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 `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`.
 
 <details>
 <summary>
@@ -80,12 +76,12 @@ conda activate textgen
 
 | System | GPU | Command |
 |--------|---------|---------|
-| Linux/WSL | NVIDIA | `pip3 install torch==2.2.2 torchvision==0.17.2 torchaudio==2.2.2 --index-url https://download.pytorch.org/whl/cu121` |
-| Linux/WSL | CPU only | `pip3 install torch==2.2.2 torchvision==0.17.2 torchaudio==2.2.2 --index-url https://download.pytorch.org/whl/cpu` |
-| Linux | AMD | `pip3 install torch==2.2.2 torchvision==0.17.2 torchaudio==2.2.2 --index-url https://download.pytorch.org/whl/rocm5.6` |
-| MacOS + MPS | Any | `pip3 install torch==2.2.2 torchvision==0.17.2 torchaudio==2.2.2` |
-| Windows | NVIDIA | `pip3 install torch==2.2.2 torchvision==0.17.2 torchaudio==2.2.2 --index-url https://download.pytorch.org/whl/cu121` |
-| Windows | CPU only | `pip3 install torch==2.2.2 torchvision==0.17.2 torchaudio==2.2.2` |
+| Linux/WSL | NVIDIA | `pip3 install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu121` |
+| Linux/WSL | CPU only | `pip3 install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cpu` |
+| Linux | AMD | `pip3 install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/rocm6.1` |
+| MacOS + MPS | Any | `pip3 install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1` |
+| Windows | NVIDIA | `pip3 install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu121` |
+| Windows | CPU only | `pip3 install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1` |
 
 The up-to-date commands can be found here: https://pytorch.org/get-started/locally/.
 
@@ -150,7 +146,7 @@ Then browse to
 1) For Kepler GPUs and older, you will need to install CUDA 11.8 instead of 12:
 
 ```
-pip3 install torch==2.2.2 torchvision==0.17.2 torchaudio==2.2.2 --index-url https://download.pytorch.org/whl/cu118
+pip3 install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu118
 conda install -y -c "nvidia/label/cuda-11.8.0" cuda-runtime
 ```
 

docs/12 - OpenAI API.md

@@ -19,7 +19,7 @@ Add `--api` to your command-line flags.
 
 ### Examples
 
-For the documentation with all the parameters and their types, consult `http://127.0.0.1:5000/docs` or the [typing.py](https://github.com/oobabooga/text-generation-webui/blob/main/extensions/openai/typing.py) file.
+For the documentation with all the endpoints, parameters and their types, consult `http://127.0.0.1:5000/docs` or the [typing.py](https://github.com/oobabooga/text-generation-webui/blob/main/extensions/openai/typing.py) file.
 
 The official examples in the [OpenAI documentation](https://platform.openai.com/docs/api-reference) should also work, and the same parameters apply (although the API here has more optional parameters).
 
@@ -114,6 +114,30 @@ curl -k http://127.0.0.1:5000/v1/internal/logits \
   }'
 ```
 
+#### List models
+
+```shell
+curl -k http://127.0.0.1:5000/v1/internal/model/list \
+  -H "Content-Type: application/json"
+```
+
+#### Load model
+
+```shell
+curl -k http://127.0.0.1:5000/v1/internal/model/load \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model_name": "model_name",
+    "args": {
+      "load_in_4bit": true,
+      "n_gpu_layers": 12
+    },
+    "settings": {
+      "instruction_template": "Alpaca"
+    }
+  }'
+```
+
 #### Python chat example
 
 ```python

extensions/Training_PRO/script.py

@@ -241,7 +241,7 @@ def ui():
                         stride_length = gr.Slider(label='Stride', minimum=1, maximum=2048, value=512, step=1, info='Used to make the evaluation faster at the cost of accuracy. 1 = slowest but most accurate. 512 is a common value.')
 
                     with gr.Column():
-                        max_length = gr.Slider(label='max_length', minimum=0, maximum=shared.settings['truncation_length_max'], value=0, step=1, info='The context for each evaluation. If set to 0, the maximum context length for the model will be used.')
+                        max_length = gr.Number(label='max_length', precision=0, step=256, value=0, info='The context for each evaluation. If set to 0, the maximum context length for the model will be used.')
 
                 with gr.Row():
                     start_current_evaluation = gr.Button("Evaluate loaded model")

extensions/openai/completions.py

@@ -154,8 +154,9 @@ def convert_history(history):
                     elif item['type'] == 'text' and isinstance(item['text'], str):
                         content = item['text']
 
-                if image_url and content:
+                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)
@@ -234,7 +235,7 @@ def chat_completions_common(body: dict, is_legacy: bool = False, stream=False, p
             raise InvalidRequestError(message="messages: missing content", param='messages')
 
     # Chat Completions
-    object_type = 'chat.completions' if not stream else 'chat.completions.chunk'
+    object_type = 'chat.completion' if not stream else 'chat.completion.chunk'
     created_time = int(time.time())
     cmpl_id = "chatcmpl-%d" % (int(time.time() * 1000000000))
     resp_list = 'data' if is_legacy else 'choices'

extensions/openai/typing.py

@@ -37,6 +37,8 @@ class GenerationOptions(BaseModel):
     dry_base: float = 1.75
     dry_allowed_length: int = 2
     dry_sequence_breakers: str = '"\\n", ":", "\\"", "*"'
+    xtc_threshold: float = 0.1
+    xtc_probability: float = 0
     truncation_length: int = 0
     max_tokens_second: int = 0
     prompt_lookup_num_tokens: int = 0

js/main.js

@@ -600,4 +600,15 @@ headerBar.addEventListener("click", (e) => {
   }
 });
 
+//------------------------------------------------
+// Add a confirmation dialog when leaving the page
+// Useful to avoid data loss
+//------------------------------------------------
+window.addEventListener('beforeunload', function (event) {
+  // Cancel the event
+  event.preventDefault();
+  // Chrome requires returnValue to be set
+  event.returnValue = '';
+});
+
 moveToChatTab();

modules/LoRA.py

@@ -1,7 +1,6 @@
 from pathlib import Path
 
 import torch
-from peft import PeftModel
 from transformers import is_torch_xpu_available
 
 import modules.shared as shared
@@ -85,6 +84,9 @@ def add_lora_autogptq(lora_names):
 
 
 def add_lora_transformers(lora_names):
+
+    from peft import PeftModel
+
     prior_set = set(shared.lora_names)
     added_set = set(lora_names) - prior_set
     removed_set = prior_set - set(lora_names)

modules/chat.py

@@ -1059,7 +1059,12 @@ def handle_start_new_chat_click(state):
 
     convert_to_markdown.cache_clear()
 
-    return [history, html, gr.update(choices=histories, value=histories[0][1])]
+    if len(histories) > 0:
+        past_chats_update = gr.update(choices=histories, value=histories[0][1])
+    else:
+        past_chats_update = gr.update(choices=histories)
+
+    return [history, html, past_chats_update]
 
 
 def handle_delete_chat_confirm_click(state):
@@ -1110,10 +1115,15 @@ def handle_upload_chat_history(load_chat_history, state):
 
     convert_to_markdown.cache_clear()
 
+    if len(histories) > 0:
+        past_chats_update = gr.update(choices=histories, value=histories[0][1])
+    else:
+        past_chats_update = gr.update(choices=histories)
+
     return [
         history,
         html,
-        gr.update(choices=histories, value=histories[0][1])
+        past_chats_update
     ]
 
 
@@ -1132,6 +1142,11 @@ def handle_character_menu_change(state):
 
     convert_to_markdown.cache_clear()
 
+    if len(histories) > 0:
+        past_chats_update = gr.update(choices=histories, value=histories[0][1])
+    else:
+        past_chats_update = gr.update(choices=histories)
+
     return [
         history,
         html,
@@ -1140,7 +1155,7 @@ def handle_character_menu_change(state):
         picture,
         greeting,
         context,
-        gr.update(choices=histories, value=histories[0][1]),
+        past_chats_update,
     ]
 
 
@@ -1151,12 +1166,17 @@ def handle_mode_change(state):
 
     convert_to_markdown.cache_clear()
 
+    if len(histories) > 0:
+        past_chats_update = gr.update(choices=histories, value=histories[0][1])
+    else:
+        past_chats_update = gr.update(choices=histories)
+
     return [
         history,
         html,
         gr.update(visible=state['mode'] != 'instruct'),
         gr.update(visible=state['mode'] == 'chat-instruct'),
-        gr.update(choices=histories, value=histories[0][1])
+        past_chats_update
     ]
 
 

modules/exllamav2.py

@@ -7,6 +7,7 @@
     ExLlamaV2Cache,
     ExLlamaV2Cache_8bit,
     ExLlamaV2Cache_Q4,
+    ExLlamaV2Cache_TP,
     ExLlamaV2Config,
     ExLlamaV2Tokenizer
 )
@@ -18,14 +19,6 @@
 
 try:
     import flash_attn
-except ModuleNotFoundError:
-    logger.warning(
-        'You are running ExLlamaV2 without flash-attention. This will cause the VRAM usage '
-        'to be a lot higher than it could be.\n'
-        'Try installing flash-attention following the instructions here: '
-        'https://github.com/Dao-AILab/flash-attention#installation-and-features'
-    )
-    pass
 except Exception:
     logger.warning('Failed to load flash-attention due to the following error:\n')
     traceback.print_exc()
@@ -54,21 +47,30 @@ def from_pretrained(self, path_to_model):
 
         model = ExLlamaV2(config)
 
-        if not shared.args.autosplit:
-            split = None
-            if shared.args.gpu_split:
-                split = [float(alloc) for alloc in shared.args.gpu_split.split(",")]
+        split = None
+        if shared.args.gpu_split:
+            split = [float(alloc) for alloc in shared.args.gpu_split.split(",")]
 
+        if shared.args.enable_tp:
+            model.load_tp(split)
+        elif not shared.args.autosplit:
             model.load(split)
 
+        # Determine the correct cache type
         if shared.args.cache_8bit:
-            cache = ExLlamaV2Cache_8bit(model, lazy=shared.args.autosplit)
+            cache_type = ExLlamaV2Cache_8bit
         elif shared.args.cache_4bit:
-            cache = ExLlamaV2Cache_Q4(model, lazy=shared.args.autosplit)
+            cache_type = ExLlamaV2Cache_Q4
         else:
-            cache = ExLlamaV2Cache(model, lazy=shared.args.autosplit)
+            cache_type = ExLlamaV2Cache
 
-        if shared.args.autosplit:
+        # Use TP if specified
+        if shared.args.enable_tp:
+            cache = ExLlamaV2Cache_TP(model, base=cache_type)
+        else:
+            cache = cache_type(model, lazy=shared.args.autosplit)
+
+        if shared.args.autosplit and not shared.args.enable_tp:
             model.load_autosplit(cache)
 
         tokenizer = ExLlamaV2Tokenizer(config)

modules/exllamav2_hf.py

@@ -9,6 +9,7 @@
     ExLlamaV2Cache,
     ExLlamaV2Cache_8bit,
     ExLlamaV2Cache_Q4,
+    ExLlamaV2Cache_TP,
     ExLlamaV2Config
 )
 from torch.nn import CrossEntropyLoss
@@ -20,14 +21,6 @@
 
 try:
     import flash_attn
-except ModuleNotFoundError:
-    logger.warning(
-        'You are running ExLlamaV2 without flash-attention. This will cause the VRAM usage '
-        'to be a lot higher than it could be.\n'
-        'Try installing flash-attention following the instructions here: '
-        'https://github.com/Dao-AILab/flash-attention#installation-and-features'
-    )
-    pass
 except Exception:
     logger.warning('Failed to load flash-attention due to the following error:\n')
     traceback.print_exc()
@@ -42,21 +35,30 @@ def __init__(self, config: ExLlamaV2Config):
 
         self.ex_model = ExLlamaV2(config)
 
-        if not shared.args.autosplit:
-            split = None
-            if shared.args.gpu_split:
-                split = [float(alloc) for alloc in shared.args.gpu_split.split(",")]
+        split = None
+        if shared.args.gpu_split:
+            split = [float(alloc) for alloc in shared.args.gpu_split.split(",")]
 
+        if shared.args.enable_tp:
+            self.ex_model.load_tp(split)
+        elif not shared.args.autosplit:
             self.ex_model.load(split)
 
+        # Determine the correct cache type
         if shared.args.cache_8bit:
-            self.ex_cache = ExLlamaV2Cache_8bit(self.ex_model, lazy=shared.args.autosplit)
+            cache_type = ExLlamaV2Cache_8bit
         elif shared.args.cache_4bit:
-            self.ex_cache = ExLlamaV2Cache_Q4(self.ex_model, lazy=shared.args.autosplit)
+            cache_type = ExLlamaV2Cache_Q4
         else:
-            self.ex_cache = ExLlamaV2Cache(self.ex_model, lazy=shared.args.autosplit)
+            cache_type = ExLlamaV2Cache
 
-        if shared.args.autosplit:
+        # Use TP if specified
+        if shared.args.enable_tp:
+            self.ex_cache = ExLlamaV2Cache_TP(self.ex_model, base=cache_type)
+        else:
+            self.ex_cache = cache_type(self.ex_model, lazy=shared.args.autosplit)
+
+        if shared.args.autosplit and not shared.args.enable_tp:
             self.ex_model.load_autosplit(self.ex_cache)
 
         self.past_seq = None