File size: 16,059 Bytes
3a464db | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 | import torch
import numpy as np
import gradio as gr
import torch.nn.functional as F
from transformers import AutoModelForCausalLM, AutoTokenizer
import time
import random
import types
from generation_functions import setup_model_with_custom_generation
# Check available GPU
device_accelerated = 'cuda:0' if torch.cuda.is_available() else 'cpu'
print(f"Accelerated model using device: {device_accelerated}")
# Set random seed
def fix_seed(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
random.seed(seed)
fix_seed(42)
# Load model and tokenizer - using Fast_dLLM model
model_name = "Efficient-Large-Model/Fast_dLLM_v2_7B"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
# Load Fast_dLLM model instance
model_accelerated = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype="auto",
device_map=device_accelerated,
trust_remote_code=True
)
# Set up custom generation functions
model_accelerated = setup_model_with_custom_generation(model_accelerated)
# Constants
MASK_TOKEN = "[MASK]"
MASK_ID = 151665 # mask_id for Fast_dLLM model
question_ai = '''Write a piece of code to implement quick sort.'''
question_math = '''A deep-sea monster rises from the waters once every hundred years to feast on a ship and sate its hunger. Over three hundred years, it has consumed 847 people. Ships have been built larger over time, so each new ship has twice as many people as the last ship. How many people were on the ship the monster ate in the first hundred years?'''
question_gsm8k = '''Question: Skyler has 100 hats on his hand with the colors red, blue, and white. Half of the hats are red, 3/5 of the remaining hats are blue, and the rest are white. How many white hats does Skyler have?'''
# Removed parse_constraints function - no longer needed
def format_chat_history(history):
"""
Format chat history for the LLaDA model
Args:
history: List of [user_message, assistant_message] pairs
Returns:
Formatted conversation for the model
"""
messages = []
for user_msg, assistant_msg in history:
messages.append({"role": "user", "content": user_msg})
if assistant_msg: # Skip if None (for the latest user message)
messages.append({"role": "assistant", "content": assistant_msg})
return messages
@torch.no_grad()
def generate_response_with_visualization_fast_dllm(model, tokenizer, device, messages, max_new_tokens=1024,
temperature=0.0, block_length=32,
threshold=0.9, top_p=0.9):
"""
Generate text with Fast_dLLM model with visualization using custom generation function
Args:
messages: List of message dictionaries with 'role' and 'content'
max_new_tokens: Maximum number of tokens to generate
temperature: Sampling temperature
block_length: Block size for generation
threshold: Threshold for generation
top_p: Top-p sampling parameter
Yields:
Visualization states showing the progression and final text
"""
# Prepare the prompt using chat template
text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
)
model_inputs = tokenizer([text], return_tensors="pt").to(device)
# Use custom mdm_sample_with_visualization method
generator = model.mdm_sample_with_visualization(
model_inputs["input_ids"],
tokenizer=tokenizer,
max_new_tokens=max_new_tokens,
small_block_size=block_length,
temperature=temperature,
threshold=threshold,
top_p=top_p,
)
# Collect all states and final text
states = []
for item in generator:
if isinstance(item, list): # Visualization state
states.append(item)
yield item
else: # Final text
final_text = item
break
# Return final text
yield final_text
css = '''
.category-legend{display:none}
.message, .bubble, .chatbot .message, .chatbot .bubble {
max-width: 80% !important;
white-space: pre-wrap !important;
word-break: break-word !important;
box-sizing: border-box !important;
}
/* HighlightedText allows auto line wrapping and sets fixed height */
.highlighted-text-container {
white-space: pre-wrap !important;
word-break: break-word !important;
height: 200px !important;
overflow-y: auto !important;
}
.generating {
border: none;
}
#input-row {
align-items: center !important;
}
'''
def create_chatbot_demo():
with gr.Blocks(css=css) as demo:
gr.Markdown("# Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding")
gr.Markdown("[code](https://github.com/NVlabs/Fast-dLLM), [project page](https://nvlabs.github.io/Fast-dLLM/)")
# STATE MANAGEMENT
chat_history_cache = gr.State([])
# UI COMPONENTS
# Input area - moved above Fast-dLLM Accelerated section
with gr.Group():
with gr.Row(elem_id="input-row"):
user_input = gr.Textbox(
label="Your Message",
placeholder="Type your message here...",
show_label=False,
scale=8
)
send_btn = gr.Button("Send", scale=1)
clear_btn = gr.Button("Clear Conversation", scale=1)
# Fast-dLLM Accelerated conversation interface
gr.Markdown("## Fast-dLLM Model (7B Parameters)")
with gr.Row():
with gr.Column(scale=2):
chatbot_ui = gr.Chatbot(label="Conversation (Fast-dLLM Model)", height=520)
with gr.Column(scale=2):
with gr.Row():
generation_time = gr.Textbox(
label="Generation Time",
value="wait for generation",
interactive=False
)
throughput = gr.Textbox(
label="Generation Speed",
value="wait for generation",
interactive=False
)
output_vis = gr.HighlightedText(
label="Denoising Process Visualization (Real-time)",
combine_adjacent=False,
show_legend=True,
elem_classes=["highlighted-text-container"]
)
output_vis_slow = gr.HighlightedText(
label="Denoising Process Visualization (Slow Motion)",
combine_adjacent=False,
show_legend=True,
elem_classes=["highlighted-text-container"]
)
# Examples moved below the conversation interfaces
gr.Examples(
examples=[
[question_ai],
[question_gsm8k],
[question_math],
],
inputs=user_input,
label="Example Inputs"
)
# Advanced generation settings
with gr.Accordion("Generation Settings", open=True):
with gr.Row():
max_new_tokens = gr.Slider(
minimum=64, maximum=2048, value=1024, step=64,
label="Max New Tokens"
)
block_length = gr.Slider(
minimum=4, maximum=32, value=16, step=4,
label="Block Size"
)
with gr.Row():
temperature = gr.Slider(
minimum=0.0, maximum=2.0, value=0.0, step=0.1,
label="Temperature"
)
top_p = gr.Slider(
minimum=0.1, maximum=1.0, value=0.95, step=0.05,
label="Top-p"
)
with gr.Row():
threshold = gr.Slider(
minimum=0.5, maximum=1.0, value=0.95, step=0.05,
label="Threshold"
)
visualization_delay = gr.Slider(
minimum=0.0, maximum=1.0, value=0.1, step=0.1,
label="Visualization Delay (seconds)"
)
# Current response text box (hidden)
current_response = gr.Textbox(
label="Current Response",
placeholder="The assistant's response will appear here...",
lines=3,
visible=False
)
# HELPER FUNCTIONS
def add_message(history, message, response):
"""Add a message pair to the history and return the updated history"""
history = history.copy()
history.append([message, response])
return history
def user_message_submitted(message, history_cache, max_new_tokens):
"""Process a submitted user message"""
# Skip empty messages
if not message.strip():
# Return current state unchanged
history_cache_for_display = history_cache.copy()
return history_cache, history_cache_for_display, "", [], [], "wait for generation", "wait for generation"
# Add user message to history
history_cache = add_message(history_cache, message, None)
# Format for display - temporarily show user message with empty response
history_cache_for_display = history_cache.copy()
# Clear the input
message_out = ""
# Return immediately to update UI with user message
return history_cache, history_cache_for_display, message_out, [], [], "processing...", "processing..."
def accelerated_response(history_cache, max_new_tokens, temperature, top_p, block_length, threshold, visualization_delay):
"""Generate accelerated model response independently"""
if not history_cache:
return history_cache, [], [], "", "wait for generation", "wait for generation"
# Get the last user message
last_user_message = history_cache[-1][0]
try:
# Format all messages except the last one (which has no response yet)
messages = format_chat_history(history_cache[:-1])
# Add the last user message
messages.append({"role": "user", "content": last_user_message})
# Start timing
start_time = time.time()
# Generate with accelerated model and yield states in real-time
with torch.no_grad():
generator = generate_response_with_visualization_fast_dllm(
model_accelerated, tokenizer, device_accelerated,
messages, max_new_tokens, temperature, block_length, threshold, top_p
)
# Collect all states and get final text
states = []
for item in generator:
if isinstance(item, list): # Visualization state
states.append(item)
yield history_cache, item, [], "", "processing...", "processing..."
else: # Final text
cache_response_text = item
break
accelerated_complete_time = time.time() - start_time
cache_generation_time_str = f"{accelerated_complete_time:.2f}s"
# Calculate throughput
cache_response_tokens = tokenizer.encode(cache_response_text, add_special_tokens=False)
cache_num_tokens = len(cache_response_tokens)
cache_throughput = cache_num_tokens / accelerated_complete_time if accelerated_complete_time > 0 else 0
cache_throughput_str = f"{cache_throughput:.2f} tokens/s"
# Update history
history_cache[-1][1] = cache_response_text
# Final yield with complete information and start slow motion visualization
if states:
# First, yield the final real-time state
yield history_cache, states[-1], states[0], cache_response_text, cache_generation_time_str, cache_throughput_str
# Then animate through slow motion visualization
for state in states[1:]:
time.sleep(visualization_delay)
yield history_cache, states[-1], state, cache_response_text, cache_generation_time_str, cache_throughput_str
except Exception as e:
error_msg = f"Error: {str(e)}"
print(error_msg)
error_vis = [(error_msg, "red")]
yield history_cache, error_vis, error_vis, error_msg, "Error", "Error"
def clear_conversation():
"""Clear the conversation history"""
empty_history = []
empty_response = ""
empty_vis = []
time_str = "wait for generation"
throughput_str = "wait for generation"
return (
empty_history, # chat_history_cache
empty_history, # chatbot_ui
empty_response, # current_response
empty_vis, # output_vis
empty_vis, # output_vis_slow
time_str, # generation_time
throughput_str # throughput
)
# EVENT HANDLERS
# Clear button handler
clear_btn.click(
fn=clear_conversation,
inputs=[],
outputs=[chat_history_cache, chatbot_ui, current_response, output_vis, output_vis_slow, generation_time, throughput]
)
# User message submission flow (2-step process)
# Step 1: Add user message to history and update UI
msg_submit = user_input.submit(
fn=user_message_submitted,
inputs=[user_input, chat_history_cache, max_new_tokens],
outputs=[chat_history_cache, chatbot_ui, user_input, output_vis, output_vis_slow, generation_time, throughput]
)
# Also connect the send button
send_click = send_btn.click(
fn=user_message_submitted,
inputs=[user_input, chat_history_cache, max_new_tokens],
outputs=[chat_history_cache, chatbot_ui, user_input, output_vis, output_vis_slow, generation_time, throughput]
)
# Step 2: Generate accelerated model response
msg_submit.then(
fn=accelerated_response,
inputs=[
chat_history_cache, max_new_tokens,
temperature, top_p, block_length, threshold, visualization_delay
],
outputs=[chatbot_ui, output_vis, output_vis_slow, current_response, generation_time, throughput]
)
send_click.then(
fn=accelerated_response,
inputs=[
chat_history_cache, max_new_tokens,
temperature, top_p, block_length, threshold, visualization_delay
],
outputs=[chatbot_ui, output_vis, output_vis_slow, current_response, generation_time, throughput]
)
return demo
# Launch the demo
if __name__ == "__main__":
demo = create_chatbot_demo()
demo.queue().launch(server_port=10086, share=True) |