Merge branch 'main' into fix/api-reload

.gitignore

@@ -2,6 +2,7 @@ cache/*
 characters/*
 extensions/silero_tts/outputs/*
 extensions/elevenlabs_tts/outputs/*
+extensions/sd_api_pictures/outputs/*
 logs/*
 loras/*
 models/*

README.md

@@ -84,10 +84,6 @@ pip install -r requirements.txt
 > 
 > For bitsandbytes and `--load-in-8bit` to work on Linux/WSL, this dirty fix is currently necessary: https://github.com/oobabooga/text-generation-webui/issues/400#issuecomment-1474876859
 
-### Alternative: native Windows installation
-
-As an alternative to the recommended WSL method, you can install the web UI natively on Windows using this guide. It will be a lot harder and the performance may be slower: [Installation instructions for human beings](https://github.com/oobabooga/text-generation-webui/wiki/Installation-instructions-for-human-beings).
-
 ### Alternative: one-click installers
 
 [oobabooga-windows.zip](https://github.com/oobabooga/one-click-installers/archive/refs/heads/oobabooga-windows.zip)
@@ -101,7 +97,13 @@ Just download the zip above, extract it, and double click on "install". The web
 
 Source codes: https://github.com/oobabooga/one-click-installers
 
-This method lags behind the newest developments and does not support 8-bit mode on Windows without additional set up: https://github.com/oobabooga/text-generation-webui/issues/147#issuecomment-1456040134, https://github.com/oobabooga/text-generation-webui/issues/20#issuecomment-1411650652
+> **Note**
+> 
+> To get 8-bit and 4-bit models working in your 1-click Windows installation, you can use the [one-click-bandaid](https://github.com/ClayShoaf/oobabooga-one-click-bandaid).
+
+### Alternative: native Windows installation
+
+As an alternative to the recommended WSL method, you can install the web UI natively on Windows using this guide. It will be a lot harder and the performance may be slower: [Installation instructions for human beings](https://github.com/oobabooga/text-generation-webui/wiki/Installation-instructions-for-human-beings).
 
 ### Alternative: Docker
 
@@ -175,15 +177,17 @@ Optionally, you can use the following command-line flags:
 | `--cpu`          | Use the CPU to generate text.|
 | `--load-in-8bit` | Load the model with 8-bit precision.|
 | `--load-in-4bit` | DEPRECATED: use `--gptq-bits 4` instead. |
-| `--gptq-bits GPTQ_BITS` |  Load a pre-quantized model with specified precision. 2, 3, 4 and 8 (bit) are supported. Currently only works with LLaMA and OPT. |
-| `--gptq-model-type MODEL_TYPE` |  Model type of pre-quantized model. Currently only LLaMa and OPT are supported. |
+| `--gptq-bits GPTQ_BITS` |  GPTQ: Load a pre-quantized model with specified precision. 2, 3, 4 and 8 (bit) are supported. Currently only works with LLaMA and OPT. |
+| `--gptq-model-type MODEL_TYPE` |  GPTQ: Model type of pre-quantized model. Currently only LLaMa and OPT are supported. |
+| `--gptq-pre-layer GPTQ_PRE_LAYER` |  GPTQ: The number of layers to preload. |
 | `--bf16`         | Load the model with bfloat16 precision. Requires NVIDIA Ampere GPU. |
 | `--auto-devices` | Automatically split the model across the available GPU(s) and CPU.|
 | `--disk`         | If the model is too large for your GPU(s) and CPU combined, send the remaining layers to the disk. |
 | `--disk-cache-dir DISK_CACHE_DIR` | Directory to save the disk cache to. Defaults to `cache/`. |
-|  `--gpu-memory GPU_MEMORY [GPU_MEMORY ...]` |  Maxmimum GPU memory in GiB to be allocated per GPU. Example: `--gpu-memory 10` for a single GPU, `--gpu-memory 10 5` for two GPUs. |
+|  `--gpu-memory GPU_MEMORY [GPU_MEMORY ...]` |  Maxmimum GPU memory in GiB to be allocated per GPU. Example: `--gpu-memory 10` for a single GPU, `--gpu-memory 10 5` for two GPUs. You can also set values in MiB like `--gpu-memory 3500MiB`. |
 | `--cpu-memory CPU_MEMORY` | Maximum CPU memory in GiB to allocate for offloaded weights. Must be an integer number. Defaults to 99.|
-| `--flexgen`      |         Enable the use of FlexGen offloading. |
+| `--no-cache`     | Set `use_cache` to False while generating text. This reduces the VRAM usage a bit with a performance cost. |
+| `--flexgen`      | Enable the use of FlexGen offloading. |
 |  `--percent PERCENT [PERCENT ...]` |  FlexGen: allocation percentages. Must be 6 numbers separated by spaces (default: 0, 100, 100, 0, 100, 0). |
 |  `--compress-weight` |  FlexGen: Whether to compress weight (default: False).|
 |  `--pin-weight [PIN_WEIGHT]` |       FlexGen: whether to pin weights (setting this to False reduces CPU memory by 20%). |
@@ -201,7 +205,7 @@ Optionally, you can use the following command-line flags:
 | `--auto-launch`  | Open the web UI in the default browser upon launch. |
 | `--verbose`      | Print the prompts to the terminal. |
 
-Out of memory errors? [Check this guide](https://github.com/oobabooga/text-generation-webui/wiki/Low-VRAM-guide).
+Out of memory errors? [Check the low VRAM guide](https://github.com/oobabooga/text-generation-webui/wiki/Low-VRAM-guide).
 
 ## Presets
 

api-example-stream.py

@@ -34,6 +34,7 @@ async def run(context):
         'penalty_alpha': 0,
         'length_penalty': 1,
         'early_stopping': False,
+        'seed': -1,
     }
     session = random_hash()
 
@@ -44,14 +45,14 @@ async def run(context):
                 case "send_hash":
                     await websocket.send(json.dumps({
                         "session_hash": session,
-                        "fn_index": 7
+                        "fn_index": 12
                     }))
                 case "estimation":
                     pass
                 case "send_data":
                     await websocket.send(json.dumps({
                         "session_hash": session,
-                        "fn_index": 7,
+                        "fn_index": 12,
                         "data": [
                             context,
                             params['max_new_tokens'],
@@ -68,6 +69,7 @@ async def run(context):
                             params['penalty_alpha'],
                             params['length_penalty'],
                             params['early_stopping'],
+                            params['seed'],
                         ]
                     }))
                 case "process_starts":

api-example.py

@@ -32,6 +32,7 @@ params = {
     'penalty_alpha': 0,
     'length_penalty': 1,
     'early_stopping': False,
+    'seed': -1,
 }
 
 # Input prompt
@@ -54,6 +55,7 @@ response = requests.post(f"http://{server}:7860/run/textgen", json={
         params['penalty_alpha'],
         params['length_penalty'],
         params['early_stopping'],
+        params['seed'],
     ]
 }).json()
 

css/main.css

@@ -50,3 +50,7 @@ ol li p, ul li p {
 #main, #parameters, #chat-settings, #interface-mode, #lora {
   border: 0;
 }
+
+.gradio-container-3-18-0 .prose * h1, h2, h3, h4 {
+  color: white;
+}

download-model.py

@@ -116,10 +116,11 @@ def get_download_links_from_huggingface(model, branch):
 
             is_pytorch = re.match("(pytorch|adapter)_model.*\.bin", fname)
             is_safetensors = re.match("model.*\.safetensors", fname)
+            is_pt = re.match(".*\.pt", fname)
             is_tokenizer = re.match("tokenizer.*\.model", fname)
-            is_text = re.match(".*\.(txt|json)", fname) or is_tokenizer
+            is_text = re.match(".*\.(txt|json|py)", fname) or is_tokenizer
 
-            if any((is_pytorch, is_safetensors, is_text, is_tokenizer)):
+            if any((is_pytorch, is_safetensors, is_pt, is_tokenizer, is_text)):
                 if is_text:
                     links.append(f"https://huggingface.co/{model}/resolve/{branch}/{fname}")
                     classifications.append('text')
@@ -132,7 +133,8 @@ def get_download_links_from_huggingface(model, branch):
                     elif is_pytorch:
                         has_pytorch = True
                         classifications.append('pytorch')
-
+                    elif is_pt:
+                        classifications.append('pt')
 
         cursor = base64.b64encode(f'{{"file_name":"{dict[-1]["path"]}"}}'.encode()) + b':50'
         cursor = base64.b64encode(cursor)

extensions/api/script.py

@@ -57,6 +57,7 @@ class Handler(BaseHTTPRequestHandler):
                 penalty_alpha=0, 
                 length_penalty=1,
                 early_stopping=False,
+                seed=-1,
             )
 
             answer = ''

extensions/elevenlabs_tts/script.py

@@ -1,6 +1,8 @@
+import re
 from pathlib import Path
 
 import gradio as gr
+import modules.shared as shared
 from elevenlabslib import ElevenLabsUser
 from elevenlabslib.helpers import save_bytes_to_path
 
@@ -15,7 +17,10 @@ wav_idx = 0
 user = ElevenLabsUser(params['api_key'])
 user_info = None
 
-
+if not shared.args.no_stream:
+    print("Please add --no-stream. This extension is not meant to be used with streaming.")
+    raise ValueError
+    
 # Check if the API is valid and refresh the UI accordingly.
 def check_valid_api():
     
@@ -47,14 +52,9 @@ def refresh_voices():
         return
 
 def remove_surrounded_chars(string):
-    new_string = ""
-    in_star = False
-    for char in string:
-        if char == '*':
-            in_star = not in_star
-        elif not in_star:
-            new_string += char
-    return new_string
+    # this expression matches to 'as few symbols as possible (0 upwards) between any asterisks' OR
+    # 'as few symbols as possible (0 upwards) between an asterisk and the end of the string'
+    return re.sub('\*[^\*]*?(\*|$)','',string)
 
 def input_modifier(string):
     """
@@ -110,4 +110,4 @@ def ui():
     voice.change(lambda x: params.update({'selected_voice': x}), voice, None)
     api_key.change(lambda x: params.update({'api_key': x}), api_key, None)
     connect.click(check_valid_api, [], connection_status)
-    connect.click(refresh_voices, [], voice)
+    connect.click(refresh_voices, [], voice)

extensions/sd_api_pictures/script.py

@@ -0,0 +1,179 @@
+import base64
+import io
+import re
+from pathlib import Path
+
+import gradio as gr
+import modules.chat as chat
+import modules.shared as shared
+import requests
+import torch
+from PIL import Image
+
+torch._C._jit_set_profiling_mode(False)
+
+# parameters which can be customized in settings.json of webui  
+params = {
+    'enable_SD_api': False,
+    'address': 'http://127.0.0.1:7860',
+    'save_img': False,
+    'SD_model': 'NeverEndingDream', # not really used right now
+    'prompt_prefix': '(Masterpiece:1.1), (solo:1.3), detailed, intricate, colorful',
+    'negative_prompt': '(worst quality, low quality:1.3)',
+    'side_length': 512,
+    'restore_faces': False
+}
+
+SD_models = ['NeverEndingDream'] # TODO: get with http://{address}}/sdapi/v1/sd-models and allow user to select
+
+streaming_state = shared.args.no_stream # remember if chat streaming was enabled
+picture_response = False # specifies if the next model response should appear as a picture
+pic_id = 0
+
+def remove_surrounded_chars(string):
+    # this expression matches to 'as few symbols as possible (0 upwards) between any asterisks' OR
+    # 'as few symbols as possible (0 upwards) between an asterisk and the end of the string'
+    return re.sub('\*[^\*]*?(\*|$)','',string)
+
+# I don't even need input_hijack for this as visible text will be commited to history as the unmodified string
+def input_modifier(string):
+    """
+    This function is applied to your text inputs before
+    they are fed into the model.
+    """
+    global params, picture_response
+    if not params['enable_SD_api']:
+        return string
+
+    commands = ['send', 'mail', 'me']
+    mediums = ['image', 'pic', 'picture', 'photo']
+    subjects = ['yourself', 'own']
+    lowstr = string.lower()
+
+    # TODO: refactor out to separate handler and also replace detection with a regexp
+    if any(command in lowstr for command in commands) and any(case in lowstr for case in mediums): # trigger the generation if a command signature and a medium signature is found
+        picture_response = True
+        shared.args.no_stream = True                                                               # Disable streaming cause otherwise the SD-generated picture would return as a dud
+        shared.processing_message = "*Is sending a picture...*"
+        string = "Please provide a detailed description of your surroundings, how you look and the situation you're in and what you are doing right now"
+        if any(target in lowstr for target in subjects):                                           # the focus of the image should be on the sending character
+            string = "Please provide a detailed and vivid description of how you look and what you are wearing"
+
+    return string
+
+# Get and save the Stable Diffusion-generated picture
+def get_SD_pictures(description):
+
+    global params, pic_id
+
+    payload = {
+        "prompt": params['prompt_prefix'] + description,
+        "seed": -1,
+        "sampler_name": "DPM++ 2M Karras",
+        "steps": 32,
+        "cfg_scale": 7,
+        "width": params['side_length'],
+        "height": params['side_length'],
+        "restore_faces": params['restore_faces'],
+        "negative_prompt": params['negative_prompt']
+    }
+    
+    response = requests.post(url=f'{params["address"]}/sdapi/v1/txt2img', json=payload)
+    r = response.json()
+
+    visible_result = ""
+    for img_str in r['images']:
+        image = Image.open(io.BytesIO(base64.b64decode(img_str.split(",",1)[0])))
+        if params['save_img']:
+            output_file = Path(f'extensions/sd_api_pictures/outputs/{pic_id:06d}.png')
+            image.save(output_file.as_posix())
+            pic_id += 1
+        # lower the resolution of received images for the chat, otherwise the log size gets out of control quickly with all the base64 values in visible history
+        image.thumbnail((300, 300))
+        buffered = io.BytesIO()
+        image.save(buffered, format="JPEG")
+        buffered.seek(0)
+        image_bytes = buffered.getvalue()
+        img_str = "data:image/jpeg;base64," + base64.b64encode(image_bytes).decode()
+        visible_result = visible_result + f'<img src="{img_str}" alt="{description}">\n'
+    
+    return visible_result
+
+# TODO: how do I make the UI history ignore the resulting pictures (I don't want HTML to appear in history)
+# and replace it with 'text' for the purposes of logging?
+def output_modifier(string):
+    """
+    This function is applied to the model outputs.
+    """
+    global pic_id, picture_response, streaming_state
+
+    if not picture_response:
+        return string
+
+    string = remove_surrounded_chars(string)
+    string = string.replace('"', '')
+    string = string.replace('“', '')
+    string = string.replace('\n', ' ')
+    string = string.strip()
+
+    if string == '':
+        string = 'no viable description in reply, try regenerating'
+
+    # I can't for the love of all that's holy get the name from shared.gradio['name1'], so for now it will be like this
+    text = f'*Description: "{string}"*'
+
+    image = get_SD_pictures(string)
+
+    picture_response = False
+
+    shared.processing_message = "*Is typing...*"
+    shared.args.no_stream = streaming_state
+    return image + "\n" + text
+
+def bot_prefix_modifier(string):
+    """
+    This function is only applied in chat mode. It modifies
+    the prefix text for the Bot and can be used to bias its
+    behavior.
+    """
+
+    return string
+
+def force_pic():
+    global picture_response
+    picture_response = True
+
+def ui():
+
+    # Gradio elements
+    with gr.Accordion("Stable Diffusion api integration", open=True):
+        with gr.Row():
+            with gr.Column():
+                enable = gr.Checkbox(value=params['enable_SD_api'], label='Activate SD Api integration')
+                save_img = gr.Checkbox(value=params['save_img'], label='Keep original received images in the outputs subdir')
+            with gr.Column():
+                address = gr.Textbox(placeholder=params['address'], value=params['address'], label='Stable Diffusion host address')
+        
+        with gr.Row():
+            force_btn = gr.Button("Force the next response to be a picture")
+            generate_now_btn = gr.Button("Generate an image response to the input")
+
+        with gr.Accordion("Generation parameters", open=False):
+            prompt_prefix = gr.Textbox(placeholder=params['prompt_prefix'], value=params['prompt_prefix'], label='Prompt Prefix (best used to describe the look of the character)')
+            with gr.Row():
+                negative_prompt = gr.Textbox(placeholder=params['negative_prompt'], value=params['negative_prompt'], label='Negative Prompt')
+                dimensions = gr.Slider(256,702,value=params['side_length'],step=64,label='Image dimensions')
+                # model = gr.Dropdown(value=SD_models[0], choices=SD_models, label='Model')
+    
+    # Event functions to update the parameters in the backend
+    enable.change(lambda x: params.update({"enable_SD_api": x}), enable, None)
+    save_img.change(lambda x: params.update({"save_img": x}), save_img, None)
+    address.change(lambda x: params.update({"address": x}), address, None)
+    prompt_prefix.change(lambda x: params.update({"prompt_prefix": x}), prompt_prefix, None)
+    negative_prompt.change(lambda x: params.update({"negative_prompt": x}), negative_prompt, None)
+    dimensions.change(lambda x: params.update({"side_length": x}), dimensions, None)
+    # model.change(lambda x: params.update({"SD_model": x}), model, None)
+
+    force_btn.click(force_pic)
+    generate_now_btn.click(force_pic)
+    generate_now_btn.click(eval('chat.cai_chatbot_wrapper'), shared.input_params, shared.gradio['display'], show_progress=shared.args.no_stream)

extensions/send_pictures/script.py

@@ -2,11 +2,11 @@ import base64
 from io import BytesIO
 
 import gradio as gr
-import torch
-from transformers import BlipForConditionalGeneration, BlipProcessor
-
 import modules.chat as chat
 import modules.shared as shared
+import torch
+from PIL import Image
+from transformers import BlipForConditionalGeneration, BlipProcessor
 
 # If 'state' is True, will hijack the next chat generation with
 # custom input text given by 'value' in the format [text, visible_text]
@@ -25,10 +25,12 @@ def caption_image(raw_image):
 
 def generate_chat_picture(picture, name1, name2):
     text = f'*{name1} sends {name2} a picture that contains the following: "{caption_image(picture)}"*'
+    # lower the resolution of sent images for the chat, otherwise the log size gets out of control quickly with all the base64 values in visible history
+    picture.thumbnail((300, 300))
     buffer = BytesIO()
     picture.save(buffer, format="JPEG")
     img_str = base64.b64encode(buffer.getvalue()).decode('utf-8')
-    visible_text = f'<img src="data:image/jpeg;base64,{img_str}">'
+    visible_text = f'<img src="data:image/jpeg;base64,{img_str}" alt="{text}">'
     return text, visible_text
 
 def ui():

extensions/silero_tts/script.py

@@ -1,11 +1,11 @@
+import re
 import time
 from pathlib import Path
 
 import gradio as gr
-import torch
-
 import modules.chat as chat
 import modules.shared as shared
+import torch
 
 torch._C._jit_set_profiling_mode(False)
 
@@ -46,14 +46,9 @@ def load_model():
 model = load_model()
 
 def remove_surrounded_chars(string):
-    new_string = ""
-    in_star = False
-    for char in string:
-        if char == '*':
-            in_star = not in_star
-        elif not in_star:
-            new_string += char
-    return new_string
+    # this expression matches to 'as few symbols as possible (0 upwards) between any asterisks' OR
+    # 'as few symbols as possible (0 upwards) between an asterisk and the end of the string'
+    return re.sub('\*[^\*]*?(\*|$)','',string)
 
 def remove_tts_from_history(name1, name2):
     for i, entry in enumerate(shared.history['internal']):
@@ -166,4 +161,4 @@ def ui():
     autoplay.change(lambda x: params.update({"autoplay": x}), autoplay, None)
     voice.change(lambda x: params.update({"speaker": x}), voice, None)
     v_pitch.change(lambda x: params.update({"voice_pitch": x}), v_pitch, None)
-    v_speed.change(lambda x: params.update({"voice_speed": x}), v_speed, None)
+    v_speed.change(lambda x: params.update({"voice_speed": x}), v_speed, None)

modules/GPTQ_loader.py

@@ -1,3 +1,4 @@
+import re
 import sys
 from pathlib import Path
 
@@ -8,6 +9,7 @@ import modules.shared as shared
 
 sys.path.insert(0, str(Path("repositories/GPTQ-for-LLaMa")))
 import llama
+import llama_inference_offload
 import opt
 
 
@@ -23,7 +25,10 @@ def load_quantized(model_name):
         model_type = shared.args.gptq_model_type.lower()
 
     if model_type == 'llama':
-        load_quant = llama.load_quant
+        if not shared.args.gptq_pre_layer:
+            load_quant = llama.load_quant
+        else:
+            load_quant = llama_inference_offload.load_quant
     elif model_type == 'opt':
         load_quant = opt.load_quant
     else:
@@ -52,20 +57,28 @@ def load_quantized(model_name):
         print(f"Could not find {pt_model}, exiting...")
         exit()
 
-    model = load_quant(str(path_to_model), str(pt_path), shared.args.gptq_bits)
-
-    # Multiple GPUs or GPU+CPU
-    if shared.args.gpu_memory:
-        max_memory = {}
-        for i in range(len(shared.args.gpu_memory)):
-            max_memory[i] = f"{shared.args.gpu_memory[i]}GiB"
-        max_memory['cpu'] = f"{shared.args.cpu_memory or '99'}GiB"
-
-        device_map = accelerate.infer_auto_device_map(model, max_memory=max_memory, no_split_module_classes=["LlamaDecoderLayer"])
-        model = accelerate.dispatch_model(model, device_map=device_map)
-
-    # Single GPU
+    # qwopqwop200's offload
+    if shared.args.gptq_pre_layer:
+        model = load_quant(str(path_to_model), str(pt_path), shared.args.gptq_bits, shared.args.gptq_pre_layer)
     else:
-        model = model.to(torch.device('cuda:0'))
+        model = load_quant(str(path_to_model), str(pt_path), shared.args.gptq_bits)
+
+        # accelerate offload (doesn't work properly)
+        if shared.args.gpu_memory:
+            memory_map = list(map(lambda x : x.strip(), shared.args.gpu_memory))
+            max_cpu_memory = shared.args.cpu_memory.strip() if shared.args.cpu_memory is not None else '99GiB'
+            max_memory = {}
+            for i in range(len(memory_map)):
+                max_memory[i] = f'{memory_map[i]}GiB' if not re.match('.*ib$', memory_map[i].lower()) else memory_map[i]
+            max_memory['cpu'] = max_cpu_memory
+
+            device_map = accelerate.infer_auto_device_map(model, max_memory=max_memory, no_split_module_classes=["LlamaDecoderLayer"])
+            print("Using the following device map for the 4-bit model:", device_map)
+            # https://huggingface.co/docs/accelerate/package_reference/big_modeling#accelerate.dispatch_model
+            model = accelerate.dispatch_model(model, device_map=device_map, offload_buffers=True)
+
+        # No offload
+        elif not shared.args.cpu:
+            model = model.to(torch.device('cuda:0'))
 
     return model

modules/LoRA.py

@@ -2,21 +2,36 @@ from pathlib import Path
 
 import modules.shared as shared
 from modules.models import load_model
+from modules.text_generation import clear_torch_cache
 
 
+def reload_model():
+    shared.model = shared.tokenizer = None
+    clear_torch_cache()
+    shared.model, shared.tokenizer = load_model(shared.model_name)
+
 def add_lora_to_model(lora_name):
 
     from peft import PeftModel
 
-    # Is there a more efficient way of returning to the base model?
-    if lora_name == "None":
-        print("Reloading the model to remove the LoRA...")
-        shared.model, shared.tokenizer = load_model(shared.model_name)
-    else:
-        # Why doesn't this work in 16-bit mode?
-        print(f"Adding the LoRA {lora_name} to the model...")
+    # If a LoRA had been previously loaded, or if we want
+    # to unload a LoRA, reload the model
+    if shared.lora_name != "None" or lora_name == "None":
+        reload_model()
+    shared.lora_name = lora_name
 
+    if lora_name != "None":
+        print(f"Adding the LoRA {lora_name} to the model...")
         params = {}
-        #params['device_map'] = {'': 0}
-        #params['dtype'] = shared.model.dtype
+        if not shared.args.cpu:
+            params['dtype'] = shared.model.dtype
+            if hasattr(shared.model, "hf_device_map"):
+                params['device_map'] = {"base_model.model."+k: v for k, v in shared.model.hf_device_map.items()}
+            elif shared.args.load_in_8bit:
+                params['device_map'] = {'': 0}
+            
         shared.model = PeftModel.from_pretrained(shared.model, Path(f"loras/{lora_name}"), **params)
+        if not shared.args.load_in_8bit and not shared.args.cpu:
+            shared.model.half()
+            if not hasattr(shared.model, "hf_device_map"):
+                shared.model.cuda()

modules/RWKV.py

@@ -45,11 +45,11 @@ class RWKVModel:
             token_stop = token_stop
         )
 
-        return context+self.pipeline.generate(context, token_count=token_count, args=args, callback=callback)
+        return self.pipeline.generate(context, token_count=token_count, args=args, callback=callback)
 
     def generate_with_streaming(self, **kwargs):
         with Iteratorize(self.generate, kwargs, callback=None) as generator:
-            reply = kwargs['context']
+            reply = ''
             for token in generator:
                 reply += token
                 yield reply

modules/callbacks.py

@@ -11,24 +11,22 @@ import modules.shared as shared
 # Copied from https://github.com/PygmalionAI/gradio-ui/
 class _SentinelTokenStoppingCriteria(transformers.StoppingCriteria):
 
-    def __init__(self, sentinel_token_ids: torch.LongTensor,
-                 starting_idx: int):
+    def __init__(self, sentinel_token_ids: list[torch.LongTensor], starting_idx: int):
         transformers.StoppingCriteria.__init__(self)
         self.sentinel_token_ids = sentinel_token_ids
         self.starting_idx = starting_idx
 
-    def __call__(self, input_ids: torch.LongTensor,
-                 _scores: torch.FloatTensor) -> bool:
+    def __call__(self, input_ids: torch.LongTensor, _scores: torch.FloatTensor) -> bool:
         for sample in input_ids:
             trimmed_sample = sample[self.starting_idx:]
-            # Can't unfold, output is still too tiny. Skip.
-            if trimmed_sample.shape[-1] < self.sentinel_token_ids.shape[-1]:
-                continue
 
-            for window in trimmed_sample.unfold(
-                    0, self.sentinel_token_ids.shape[-1], 1):
-                if torch.all(torch.eq(self.sentinel_token_ids, window)):
-                    return True
+            for i in range(len(self.sentinel_token_ids)):
+                # Can't unfold, output is still too tiny. Skip.
+                if trimmed_sample.shape[-1] < self.sentinel_token_ids[i].shape[-1]:
+                    continue
+                for window in trimmed_sample.unfold(0, self.sentinel_token_ids[i].shape[-1], 1):
+                    if torch.all(torch.eq(self.sentinel_token_ids[i], window)):
+                        return True
         return False
 
 class Stream(transformers.StoppingCriteria):

modules/chat.py

@@ -51,47 +51,37 @@ def generate_chat_prompt(user_input, max_new_tokens, name1, name2, context, chat
     prompt = ''.join(rows)
     return prompt
 
-def extract_message_from_reply(question, reply, name1, name2, check, impersonate=False):
+def extract_message_from_reply(reply, name1, name2, check):
     next_character_found = False
 
-    asker = name1 if not impersonate else name2
-    replier = name2 if not impersonate else name1
-
-    previous_idx = [m.start() for m in re.finditer(f"(^|\n){re.escape(replier)}:", question)]
-    idx = [m.start() for m in re.finditer(f"(^|\n){re.escape(replier)}:", reply)]
-    idx = idx[max(len(previous_idx)-1, 0)]
-
-    if not impersonate:
-        reply = reply[idx + 1 + len(apply_extensions(f"{replier}:", "bot_prefix")):]
-    else:
-        reply = reply[idx + 1 + len(f"{replier}:"):]
-
     if check:
         lines = reply.split('\n')
         reply = lines[0].strip()
         if len(lines) > 1:
             next_character_found = True
     else:
-        idx = reply.find(f"\n{asker}:")
-        if idx != -1:
-            reply = reply[:idx]
-            next_character_found = True
-        reply = fix_newlines(reply)
+        for string in [f"\n{name1}:", f"\n{name2}:"]:
+            idx = reply.find(string)
+            if idx != -1:
+                reply = reply[:idx]
+                next_character_found = True
 
         # If something like "\nYo" is generated just before "\nYou:"
         # is completed, trim it
-        next_turn = f"\n{asker}:"
-        for j in range(len(next_turn)-1, 0, -1):
-            if reply[-j:] == next_turn[:j]:
-                reply = reply[:-j]
-                break
+        if not next_character_found:
+            for string in [f"\n{name1}:", f"\n{name2}:"]:
+                for j in range(len(string)-1, 0, -1):
+                    if reply[-j:] == string[:j]:
+                        reply = reply[:-j]
+                        break
 
+    reply = fix_newlines(reply)
     return reply, next_character_found
 
 def stop_everything_event():
     shared.stop_everything = True
 
-def chatbot_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, chat_generation_attempts=1, regenerate=False):
+def chatbot_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, seed, name1, name2, context, check, chat_prompt_size, chat_generation_attempts=1, regenerate=False):
     shared.stop_everything = False
     just_started = True
     eos_token = '\n' if check else None
@@ -125,12 +115,13 @@ def chatbot_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical
         yield shared.history['visible']+[[visible_text, shared.processing_message]]
 
     # Generate
-    reply = ''
+    cumulative_reply = ''
     for i in range(chat_generation_attempts):
-        for reply in generate_reply(f"{prompt}{' ' if len(reply) > 0 else ''}{reply}", max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, eos_token=eos_token, stopping_string=f"\n{name1}:"):
+        for reply in generate_reply(f"{prompt}{' ' if len(cumulative_reply) > 0 else ''}{cumulative_reply}", max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, seed, eos_token=eos_token, stopping_strings=[f"\n{name1}:", f"\n{name2}:"]):
+            reply = cumulative_reply + reply
 
             # Extracting the reply
-            reply, next_character_found = extract_message_from_reply(prompt, reply, name1, name2, check)
+            reply, next_character_found = extract_message_from_reply(reply, name1, name2, check)
             visible_reply = re.sub("(<USER>|<user>|{{user}})", name1_original, reply)
             visible_reply = apply_extensions(visible_reply, "output")
             if shared.args.chat:
@@ -152,9 +143,11 @@ def chatbot_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical
             if next_character_found:
                 break
 
+        cumulative_reply = reply
+
     yield shared.history['visible']
 
-def impersonate_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, chat_generation_attempts=1):
+def impersonate_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, seed, name1, name2, context, check, chat_prompt_size, chat_generation_attempts=1):
     eos_token = '\n' if check else None
 
     if 'pygmalion' in shared.model_name.lower():
@@ -162,22 +155,27 @@ def impersonate_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typ
 
     prompt = generate_chat_prompt(text, max_new_tokens, name1, name2, context, chat_prompt_size, impersonate=True)
 
-    reply = ''
     # Yield *Is typing...*
     yield shared.processing_message
+
+    cumulative_reply = ''
     for i in range(chat_generation_attempts):
-        for reply in generate_reply(prompt+reply, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, eos_token=eos_token, stopping_string=f"\n{name2}:"):
-            reply, next_character_found = extract_message_from_reply(prompt, reply, name1, name2, check, impersonate=True)
+        for reply in generate_reply(f"{prompt}{' ' if len(cumulative_reply) > 0 else ''}{cumulative_reply}", max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, seed, eos_token=eos_token, stopping_strings=[f"\n{name1}:", f"\n{name2}:"]):
+            reply = cumulative_reply + reply
+            reply, next_character_found = extract_message_from_reply(reply, name1, name2, check)
             yield reply
             if next_character_found:
                 break
-        yield reply
 
-def cai_chatbot_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, chat_generation_attempts=1):
-    for _history in chatbot_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, chat_generation_attempts):
+        cumulative_reply = reply
+
+    yield reply
+
+def cai_chatbot_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, seed, name1, name2, context, check, chat_prompt_size, chat_generation_attempts=1):
+    for _history in chatbot_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, seed, name1, name2, context, check, chat_prompt_size, chat_generation_attempts):
         yield generate_chat_html(_history, name1, name2, shared.character)
 
-def regenerate_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, chat_generation_attempts=1):
+def regenerate_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, seed, name1, name2, context, check, chat_prompt_size, chat_generation_attempts=1):
     if (shared.character != 'None' and len(shared.history['visible']) == 1) or len(shared.history['internal']) == 0:
         yield generate_chat_output(shared.history['visible'], name1, name2, shared.character)
     else:
@@ -185,7 +183,7 @@ def regenerate_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typi
         last_internal = shared.history['internal'].pop()
         # Yield '*Is typing...*'
         yield generate_chat_output(shared.history['visible']+[[last_visible[0], shared.processing_message]], name1, name2, shared.character)
-        for _history in chatbot_wrapper(last_internal[0], max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, chat_generation_attempts, regenerate=True):
+        for _history in chatbot_wrapper(last_internal[0], max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, seed, name1, name2, context, check, chat_prompt_size, chat_generation_attempts, regenerate=True):
             if shared.args.cai_chat:
                 shared.history['visible'][-1] = [last_visible[0], _history[-1][1]]
             else:

modules/models.py

@@ -1,5 +1,6 @@
 import json
 import os
+import re
 import time
 import zipfile
 from pathlib import Path
@@ -120,11 +121,12 @@ def load_model(model_name):
                 params["torch_dtype"] = torch.float16
 
             if shared.args.gpu_memory:
-                memory_map = shared.args.gpu_memory
+                memory_map = list(map(lambda x : x.strip(), shared.args.gpu_memory))
+                max_cpu_memory = shared.args.cpu_memory.strip() if shared.args.cpu_memory is not None else '99GiB'
                 max_memory = {}
                 for i in range(len(memory_map)):
-                    max_memory[i] = f'{memory_map[i]}GiB'
-                max_memory['cpu'] = f'{shared.args.cpu_memory or 99}GiB'
+                    max_memory[i] = f'{memory_map[i]}GiB' if not re.match('.*ib$', memory_map[i].lower()) else memory_map[i]
+                max_memory['cpu'] = max_cpu_memory
                 params['max_memory'] = max_memory
             elif shared.args.auto_devices:
                 total_mem = (torch.cuda.get_device_properties(0).total_memory / (1024*1024))

modules/shared.py

@@ -27,9 +27,9 @@ settings = {
     'max_new_tokens': 200,
     'max_new_tokens_min': 1,
     'max_new_tokens_max': 2000,
-    'name1': 'Person 1',
-    'name2': 'Person 2',
-    'context': 'This is a conversation between two people.',
+    'name1': 'You',
+    'name2': 'Assistant',
+    'context': 'This is a conversation with your Assistant. The Assistant is very helpful and is eager to chat with you and answer your questions.',
     'stop_at_newline': False,
     'chat_prompt_size': 2048,
     'chat_prompt_size_min': 0,
@@ -56,7 +56,7 @@ settings = {
     },
     'lora_prompts': {
         'default': 'Common sense questions and answers\n\nQuestion: \nFactual answer:',
-        'alpaca-lora-7b': "Below is an instruction that describes a task. Write a response that appropriately completes the request.\n### Instruction:\nWrite a poem about the transformers Python library. \nMention the word \"large language models\" in that poem.\n### Response:\n"
+        '(alpaca-lora-7b|alpaca-lora-13b|alpaca-lora-30b)': "Below is an instruction that describes a task. Write a response that appropriately completes the request.\n### Instruction:\nWrite a poem about the transformers Python library. \nMention the word \"large language models\" in that poem.\n### Response:\n"
     }
 }
 
@@ -79,14 +79,16 @@ parser.add_argument('--cai-chat', action='store_true', help='Launch the web UI i
 parser.add_argument('--cpu', action='store_true', help='Use the CPU to generate text.')
 parser.add_argument('--load-in-8bit', action='store_true', help='Load the model with 8-bit precision.')
 parser.add_argument('--load-in-4bit', action='store_true', help='DEPRECATED: use --gptq-bits 4 instead.')
-parser.add_argument('--gptq-bits', type=int, default=0, help='Load a pre-quantized model with specified precision. 2, 3, 4 and 8bit are supported. Currently only works with LLaMA and OPT.')
-parser.add_argument('--gptq-model-type', type=str, help='Model type of pre-quantized model. Currently only LLaMa and OPT are supported.')
+parser.add_argument('--gptq-bits', type=int, default=0, help='GPTQ: Load a pre-quantized model with specified precision. 2, 3, 4 and 8bit are supported. Currently only works with LLaMA and OPT.')
+parser.add_argument('--gptq-model-type', type=str, help='GPTQ: Model type of pre-quantized model. Currently only LLaMa and OPT are supported.')
+parser.add_argument('--gptq-pre-layer', type=int, default=0, help='GPTQ: The number of layers to preload.')
 parser.add_argument('--bf16', action='store_true', help='Load the model with bfloat16 precision. Requires NVIDIA Ampere GPU.')
 parser.add_argument('--auto-devices', action='store_true', help='Automatically split the model across the available GPU(s) and CPU.')
 parser.add_argument('--disk', action='store_true', help='If the model is too large for your GPU(s) and CPU combined, send the remaining layers to the disk.')
 parser.add_argument('--disk-cache-dir', type=str, default="cache", help='Directory to save the disk cache to. Defaults to "cache".')
-parser.add_argument('--gpu-memory', type=int, nargs="+", help='Maxmimum GPU memory in GiB to be allocated per GPU. Example: --gpu-memory 10 for a single GPU, --gpu-memory 10 5 for two GPUs.')
-parser.add_argument('--cpu-memory', type=int, help='Maximum CPU memory in GiB to allocate for offloaded weights. Must be an integer number. Defaults to 99.')
+parser.add_argument('--gpu-memory', type=str, nargs="+", help='Maxmimum GPU memory in GiB to be allocated per GPU. Example: --gpu-memory 10 for a single GPU, --gpu-memory 10 5 for two GPUs.')
+parser.add_argument('--cpu-memory', type=str, help='Maximum CPU memory in GiB to allocate for offloaded weights. Must be an integer number. Defaults to 99.')
+parser.add_argument('--no-cache', action='store_true', help='Set use_cache to False while generating text. This reduces the VRAM usage a bit at a performance cost.')
 parser.add_argument('--flexgen', action='store_true', help='Enable the use of FlexGen offloading.')
 parser.add_argument('--percent', type=int, nargs="+", default=[0, 100, 100, 0, 100, 0], help='FlexGen: allocation percentages. Must be 6 numbers separated by spaces (default: 0, 100, 100, 0, 100, 0).')
 parser.add_argument("--compress-weight", action="store_true", help="FlexGen: activate weight compression.")

modules/text_generation.py

@@ -1,6 +1,7 @@
 import gc
 import re
 import time
+import traceback
 
 import numpy as np
 import torch
@@ -92,30 +93,16 @@ def clear_torch_cache():
     if not shared.args.cpu:
         torch.cuda.empty_cache()
 
-def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, eos_token=None, stopping_string=None):
-    clear_torch_cache()
-    t0 = time.time()
+def set_manual_seed(seed):
+    if seed != -1:
+        torch.manual_seed(seed)
+        if torch.cuda.is_available():
+            torch.cuda.manual_seed_all(seed)
 
-    # These models are not part of Hugging Face, so we handle them
-    # separately and terminate the function call earlier
-    if shared.is_RWKV:
-        try:
-            if shared.args.no_stream:
-                reply = shared.model.generate(context=question, token_count=max_new_tokens, temperature=temperature, top_p=top_p, top_k=top_k)
-                yield formatted_outputs(reply, shared.model_name)
-            else:
-                if not (shared.args.chat or shared.args.cai_chat):
-                    yield formatted_outputs(question, shared.model_name)
-                # RWKV has proper streaming, which is very nice.
-                # No need to generate 8 tokens at a time.
-                for reply in shared.model.generate_with_streaming(context=question, token_count=max_new_tokens, temperature=temperature, top_p=top_p, top_k=top_k):
-                    yield formatted_outputs(reply, shared.model_name)
-        finally:
-            t1 = time.time()
-            output = encode(reply)[0]
-            input_ids = encode(question)
-            print(f"Output generated in {(t1-t0):.2f} seconds ({(len(output)-len(input_ids[0]))/(t1-t0):.2f} tokens/s, {len(output)-len(input_ids[0])} tokens)")
-            return
+def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, seed, eos_token=None, stopping_strings=[]):
+    clear_torch_cache()
+    set_manual_seed(seed)
+    t0 = time.time()
 
     original_question = question
     if not (shared.args.chat or shared.args.cai_chat):
@@ -123,17 +110,46 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
     if shared.args.verbose:
         print(f"\n\n{question}\n--------------------\n")
 
+    # These models are not part of Hugging Face, so we handle them
+    # separately and terminate the function call earlier
+    if shared.is_RWKV:
+        try:
+            if shared.args.no_stream:
+                reply = shared.model.generate(context=question, token_count=max_new_tokens, temperature=temperature, top_p=top_p, top_k=top_k)
+                if not (shared.args.chat or shared.args.cai_chat):
+                    reply = original_question + apply_extensions(reply, "output")
+                yield formatted_outputs(reply, shared.model_name)
+            else:
+                if not (shared.args.chat or shared.args.cai_chat):
+                    yield formatted_outputs(question, shared.model_name)
+
+                # RWKV has proper streaming, which is very nice.
+                # No need to generate 8 tokens at a time.
+                for reply in shared.model.generate_with_streaming(context=question, token_count=max_new_tokens, temperature=temperature, top_p=top_p, top_k=top_k):
+                    if not (shared.args.chat or shared.args.cai_chat):
+                        reply = original_question + apply_extensions(reply, "output")
+                    yield formatted_outputs(reply, shared.model_name)
+
+        except Exception:
+            traceback.print_exc()
+        finally:
+            t1 = time.time()
+            output = encode(reply)[0]
+            input_ids = encode(question)
+            print(f"Output generated in {(t1-t0):.2f} seconds ({(len(output)-len(input_ids[0]))/(t1-t0):.2f} tokens/s, {len(output)-len(input_ids[0])} tokens)")
+            return
+
     input_ids = encode(question, max_new_tokens)
     original_input_ids = input_ids
     output = input_ids[0]
+
     cuda = not any((shared.args.cpu, shared.args.deepspeed, shared.args.flexgen))
     eos_token_ids = [shared.tokenizer.eos_token_id] if shared.tokenizer.eos_token_id is not None else []
     if eos_token is not None:
         eos_token_ids.append(int(encode(eos_token)[0][-1]))
     stopping_criteria_list = transformers.StoppingCriteriaList()
-    if stopping_string is not None:
-        # Copied from https://github.com/PygmalionAI/gradio-ui/blob/master/src/model.py
-        t = encode(stopping_string, 0, add_special_tokens=False)
+    if type(stopping_strings) is list and len(stopping_strings) > 0:
+        t = [encode(string, 0, add_special_tokens=False) for string in stopping_strings]
         stopping_criteria_list.append(_SentinelTokenStoppingCriteria(sentinel_token_ids=t, starting_idx=len(input_ids[0])))
 
     generate_params = {}
@@ -163,6 +179,8 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
             "temperature": temperature,
             "stop": eos_token_ids[-1],
         })
+    if shared.args.no_cache:
+        generate_params.update({"use_cache": False})
     if shared.args.deepspeed:
         generate_params.update({"synced_gpus": True})
     if shared.soft_prompt:
@@ -182,9 +200,10 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
             if shared.soft_prompt:
                 output = torch.cat((input_ids[0], output[filler_input_ids.shape[1]:]))
 
-            reply = decode(output)
+            new_tokens = len(output) - len(input_ids[0])
+            reply = decode(output[-new_tokens:])
             if not (shared.args.chat or shared.args.cai_chat):
-                reply = original_question + apply_extensions(reply[len(question):], "output")
+                reply = original_question + apply_extensions(reply, "output")
 
             yield formatted_outputs(reply, shared.model_name)
 
@@ -207,10 +226,11 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
                 for output in generator:
                     if shared.soft_prompt:
                         output = torch.cat((input_ids[0], output[filler_input_ids.shape[1]:]))
-                    reply = decode(output)
 
+                    new_tokens = len(output) - len(input_ids[0])
+                    reply = decode(output[-new_tokens:])
                     if not (shared.args.chat or shared.args.cai_chat):
-                        reply = original_question + apply_extensions(reply[len(question):], "output")
+                        reply = original_question + apply_extensions(reply, "output")
 
                     if output[-1] in eos_token_ids:
                         break
@@ -226,10 +246,11 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
                     output = shared.model.generate(**generate_params)[0]
                 if shared.soft_prompt:
                     output = torch.cat((input_ids[0], output[filler_input_ids.shape[1]:]))
-                reply = decode(output)
 
+                new_tokens = len(output) - len(original_input_ids[0])
+                reply = decode(output[-new_tokens:])
                 if not (shared.args.chat or shared.args.cai_chat):
-                    reply = original_question + apply_extensions(reply[len(question):], "output")
+                    reply = original_question + apply_extensions(reply, "output")
 
                 if np.count_nonzero(np.isin(input_ids[0], eos_token_ids)) < np.count_nonzero(np.isin(output, eos_token_ids)):
                     break
@@ -238,9 +259,15 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
                 input_ids = np.reshape(output, (1, output.shape[0]))
                 if shared.soft_prompt:
                     inputs_embeds, filler_input_ids = generate_softprompt_input_tensors(input_ids)
+                    generate_params.update({"inputs_embeds": inputs_embeds})
+                    generate_params.update({"inputs": filler_input_ids})
+                else:
+                    generate_params.update({"inputs": input_ids})
 
             yield formatted_outputs(reply, shared.model_name)
 
+    except Exception:
+        traceback.print_exc()
     finally:
         t1 = time.time()
         print(f"Output generated in {(t1-t0):.2f} seconds ({(len(output)-len(original_input_ids[0]))/(t1-t0):.2f} tokens/s, {len(output)-len(original_input_ids[0])} tokens)")

presets/Default.txt

@@ -1,12 +1,7 @@
 do_sample=True
-temperature=1
-top_p=1
-typical_p=1
-repetition_penalty=1
-top_k=50
-num_beams=1
-penalty_alpha=0
-min_length=0
-length_penalty=1
-no_repeat_ngram_size=0
+top_p=0.5
+top_k=40
+temperature=0.7
+repetition_penalty=1.2
+typical_p=1.0
 early_stopping=False

presets/Individual Today.txt

@@ -1,6 +0,0 @@
-do_sample=True
-top_p=0.9
-top_k=50
-temperature=1.39
-repetition_penalty=1.08
-typical_p=0.2

requirements.txt

@@ -6,6 +6,7 @@ markdown
 numpy
 peft==0.2.0
 requests
+rwkv==0.7.0
 safetensors==0.3.0
 sentencepiece
 tqdm

server.py

@@ -1,4 +1,3 @@
-import gc
 import io
 import json
 import re
@@ -8,7 +7,6 @@ import zipfile
 from pathlib import Path
 
 import gradio as gr
-import torch
 
 import modules.chat as chat
 import modules.extensions as extensions_module
@@ -17,7 +15,7 @@ import modules.ui as ui
 from modules.html_generator import generate_chat_html
 from modules.LoRA import add_lora_to_model
 from modules.models import load_model, load_soft_prompt
-from modules.text_generation import generate_reply
+from modules.text_generation import clear_torch_cache, generate_reply
 
 # Loading custom settings
 settings_file = None
@@ -56,21 +54,14 @@ def load_model_wrapper(selected_model):
     if selected_model != shared.model_name:
         shared.model_name = selected_model
         shared.model = shared.tokenizer = None
-        if not shared.args.cpu:
-            gc.collect()
-            torch.cuda.empty_cache()
+        clear_torch_cache()
         shared.model, shared.tokenizer = load_model(shared.model_name)
 
     return selected_model
 
 def load_lora_wrapper(selected_lora):
-    shared.lora_name = selected_lora
-    default_text = shared.settings['lora_prompts'][next((k for k in shared.settings['lora_prompts'] if re.match(k.lower(), shared.lora_name.lower())), 'default')]
-
-    if not shared.args.cpu:
-        gc.collect()
-        torch.cuda.empty_cache()
     add_lora_to_model(selected_lora)
+    default_text = shared.settings['lora_prompts'][next((k for k in shared.settings['lora_prompts'] if re.match(k.lower(), shared.lora_name.lower())), 'default')]
 
     return selected_lora, default_text
 
@@ -102,7 +93,7 @@ def load_preset_values(preset_menu, return_dict=False):
     if return_dict:
         return generate_params
     else:
-        return generate_params['do_sample'], generate_params['temperature'], generate_params['top_p'], generate_params['typical_p'], generate_params['repetition_penalty'], generate_params['encoder_repetition_penalty'], generate_params['top_k'], generate_params['min_length'], generate_params['no_repeat_ngram_size'], generate_params['num_beams'], generate_params['penalty_alpha'], generate_params['length_penalty'], generate_params['early_stopping']
+        return preset_menu, generate_params['do_sample'], generate_params['temperature'], generate_params['top_p'], generate_params['typical_p'], generate_params['repetition_penalty'], generate_params['encoder_repetition_penalty'], generate_params['top_k'], generate_params['min_length'], generate_params['no_repeat_ngram_size'], generate_params['num_beams'], generate_params['penalty_alpha'], generate_params['length_penalty'], generate_params['early_stopping']
 
 def upload_soft_prompt(file):
     with zipfile.ZipFile(io.BytesIO(file)) as zf:
@@ -160,6 +151,12 @@ def create_settings_menus(default_preset):
                         shared.gradio['length_penalty'] = gr.Slider(-5, 5, value=generate_params['length_penalty'], label='length_penalty')
                 shared.gradio['early_stopping'] = gr.Checkbox(value=generate_params['early_stopping'], label='early_stopping')
 
+            shared.gradio['seed'] = gr.Number(value=-1, label='Seed (-1 for random)')
+
+    with gr.Row():
+        shared.gradio['preset_menu_mirror'] = gr.Dropdown(choices=available_presets, value=default_preset if not shared.args.flexgen else 'Naive', label='Generation parameters preset')
+        ui.create_refresh_button(shared.gradio['preset_menu_mirror'], lambda : None, lambda : {'choices': get_available_presets()}, 'refresh-button')
+
     with gr.Row():
         shared.gradio['lora_menu'] = gr.Dropdown(choices=available_loras, value=shared.lora_name, label='LoRA')
         ui.create_refresh_button(shared.gradio['lora_menu'], lambda : None, lambda : {'choices': get_available_loras()}, 'refresh-button')
@@ -174,7 +171,8 @@ def create_settings_menus(default_preset):
             shared.gradio['upload_softprompt'] = gr.File(type='binary', file_types=['.zip'])
 
     shared.gradio['model_menu'].change(load_model_wrapper, [shared.gradio['model_menu']], [shared.gradio['model_menu']], show_progress=True)
-    shared.gradio['preset_menu'].change(load_preset_values, [shared.gradio['preset_menu']], [shared.gradio['do_sample'], shared.gradio['temperature'], shared.gradio['top_p'], shared.gradio['typical_p'], shared.gradio['repetition_penalty'], shared.gradio['encoder_repetition_penalty'], shared.gradio['top_k'], shared.gradio['min_length'], shared.gradio['no_repeat_ngram_size'], shared.gradio['num_beams'], shared.gradio['penalty_alpha'], shared.gradio['length_penalty'], shared.gradio['early_stopping']])
+    shared.gradio['preset_menu'].change(load_preset_values, [shared.gradio['preset_menu']], [shared.gradio[k] for k in ['preset_menu_mirror', 'do_sample', 'temperature', 'top_p', 'typical_p', 'repetition_penalty', 'encoder_repetition_penalty', 'top_k', 'min_length', 'no_repeat_ngram_size', 'num_beams', 'penalty_alpha', 'length_penalty', 'early_stopping']])
+    shared.gradio['preset_menu_mirror'].change(load_preset_values, [shared.gradio['preset_menu_mirror']], [shared.gradio[k] for k in ['preset_menu', 'do_sample', 'temperature', 'top_p', 'typical_p', 'repetition_penalty', 'encoder_repetition_penalty', 'top_k', 'min_length', 'no_repeat_ngram_size', 'num_beams', 'penalty_alpha', 'length_penalty', 'early_stopping']])
     shared.gradio['lora_menu'].change(load_lora_wrapper, [shared.gradio['lora_menu']], [shared.gradio['lora_menu'], shared.gradio['textbox']], show_progress=True)
     shared.gradio['softprompts_menu'].change(load_soft_prompt, [shared.gradio['softprompts_menu']], [shared.gradio['softprompts_menu']], show_progress=True)
     shared.gradio['upload_softprompt'].upload(upload_soft_prompt, [shared.gradio['upload_softprompt']], [shared.gradio['softprompts_menu']])
@@ -235,9 +233,7 @@ else:
     shared.model_name = available_models[i]
 shared.model, shared.tokenizer = load_model(shared.model_name)
 if shared.args.lora:
-    print(shared.args.lora)
-    shared.lora_name = shared.args.lora
-    add_lora_to_model(shared.lora_name)
+    add_lora_to_model(shared.args.lora)
 
 # Default UI settings
 default_preset = shared.settings['presets'][next((k for k in shared.settings['presets'] if re.match(k.lower(), shared.model_name.lower())), 'default')]
@@ -325,13 +321,13 @@ def create_interface():
                 create_settings_menus(default_preset)
 
             function_call = 'chat.cai_chatbot_wrapper' if shared.args.cai_chat else 'chat.chatbot_wrapper'
-            shared.input_params = [shared.gradio[k] for k in ['textbox', 'max_new_tokens', 'do_sample', 'temperature', 'top_p', 'typical_p', 'repetition_penalty', 'encoder_repetition_penalty', 'top_k', 'min_length', 'no_repeat_ngram_size', 'num_beams', 'penalty_alpha', 'length_penalty', 'early_stopping', 'name1', 'name2', 'context', 'check', 'chat_prompt_size_slider', 'chat_generation_attempts']]
+            shared.input_params = [shared.gradio[k] for k in ['textbox', 'max_new_tokens', 'do_sample', 'temperature', 'top_p', 'typical_p', 'repetition_penalty', 'encoder_repetition_penalty', 'top_k', 'min_length', 'no_repeat_ngram_size', 'num_beams', 'penalty_alpha', 'length_penalty', 'early_stopping', 'seed', 'name1', 'name2', 'context', 'check', 'chat_prompt_size_slider', 'chat_generation_attempts']]
 
             gen_events.append(shared.gradio['Generate'].click(eval(function_call), shared.input_params, shared.gradio['display'], show_progress=shared.args.no_stream))
             gen_events.append(shared.gradio['textbox'].submit(eval(function_call), shared.input_params, shared.gradio['display'], show_progress=shared.args.no_stream))
             gen_events.append(shared.gradio['Regenerate'].click(chat.regenerate_wrapper, shared.input_params, shared.gradio['display'], show_progress=shared.args.no_stream))
             gen_events.append(shared.gradio['Impersonate'].click(chat.impersonate_wrapper, shared.input_params, shared.gradio['textbox'], show_progress=shared.args.no_stream))
-            shared.gradio['Stop'].click(chat.stop_everything_event, [], [], cancels=gen_events)
+            shared.gradio['Stop'].click(chat.stop_everything_event, [], [], cancels=gen_events, queue=False)
 
             shared.gradio['Copy last reply'].click(chat.send_last_reply_to_input, [], shared.gradio['textbox'], show_progress=shared.args.no_stream)
             shared.gradio['Replace last reply'].click(chat.replace_last_reply, [shared.gradio['textbox'], shared.gradio['name1'], shared.gradio['name2']], shared.gradio['display'], show_progress=shared.args.no_stream)
@@ -388,7 +384,7 @@ def create_interface():
             with gr.Tab("Parameters", elem_id="parameters"):
                 create_settings_menus(default_preset)
 
-            shared.input_params = [shared.gradio[k] for k in ['textbox', 'max_new_tokens', 'do_sample', 'temperature', 'top_p', 'typical_p', 'repetition_penalty', 'encoder_repetition_penalty', 'top_k', 'min_length', 'no_repeat_ngram_size', 'num_beams', 'penalty_alpha', 'length_penalty', 'early_stopping']]
+            shared.input_params = [shared.gradio[k] for k in ['textbox', 'max_new_tokens', 'do_sample', 'temperature', 'top_p', 'typical_p', 'repetition_penalty', 'encoder_repetition_penalty', 'top_k', 'min_length', 'no_repeat_ngram_size', 'num_beams', 'penalty_alpha', 'length_penalty', 'early_stopping', 'seed']]
             output_params = [shared.gradio[k] for k in ['textbox', 'markdown', 'html']]
             gen_events.append(shared.gradio['Generate'].click(generate_reply, shared.input_params, output_params, show_progress=shared.args.no_stream, api_name='textgen'))
             gen_events.append(shared.gradio['textbox'].submit(generate_reply, shared.input_params, output_params, show_progress=shared.args.no_stream))
@@ -420,7 +416,7 @@ def create_interface():
             with gr.Tab("Parameters", elem_id="parameters"):
                 create_settings_menus(default_preset)
 
-            shared.input_params = [shared.gradio[k] for k in ['textbox', 'max_new_tokens', 'do_sample', 'temperature', 'top_p', 'typical_p', 'repetition_penalty', 'encoder_repetition_penalty', 'top_k', 'min_length', 'no_repeat_ngram_size', 'num_beams', 'penalty_alpha', 'length_penalty', 'early_stopping']]
+            shared.input_params = [shared.gradio[k] for k in ['textbox', 'max_new_tokens', 'do_sample', 'temperature', 'top_p', 'typical_p', 'repetition_penalty', 'encoder_repetition_penalty', 'top_k', 'min_length', 'no_repeat_ngram_size', 'num_beams', 'penalty_alpha', 'length_penalty', 'early_stopping', 'seed']]
             output_params = [shared.gradio[k] for k in ['output_textbox', 'markdown', 'html']]
             gen_events.append(shared.gradio['Generate'].click(generate_reply, shared.input_params, output_params, show_progress=shared.args.no_stream, api_name='textgen'))
             gen_events.append(shared.gradio['textbox'].submit(generate_reply, shared.input_params, output_params, show_progress=shared.args.no_stream))

settings-template.json

@@ -2,9 +2,9 @@
     "max_new_tokens": 200,
     "max_new_tokens_min": 1,
     "max_new_tokens_max": 2000,
-    "name1": "Person 1",
-    "name2": "Person 2",
-    "context": "This is a conversation between two people.",
+    "name1": "You",
+    "name2": "Assistant",
+    "context": "This is a conversation with your Assistant. The Assistant is very helpful and is eager to chat with you and answer your questions.",
     "stop_at_newline": false,
     "chat_prompt_size": 2048,
     "chat_prompt_size_min": 0,