Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -4,519 +4,335 @@ import numpy as np
|
|
4 |
import gradio as gr
|
5 |
import spaces
|
6 |
import torch.nn.functional as F
|
7 |
-
from transformers import AutoTokenizer, AutoModel
|
8 |
-
from transformers.generation.configuration_utils import GenerationConfig
|
9 |
import time
|
10 |
-
import
|
11 |
-
import torch.distributions as dists # Import dists for sampling logic
|
12 |
|
13 |
-
#
|
14 |
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
15 |
print(f"Using device: {device}")
|
16 |
|
17 |
-
#
|
18 |
model_path = "Dream-org/Dream-v0-Instruct-7B"
|
19 |
-
|
20 |
-
|
|
|
|
|
21 |
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
|
22 |
-
model = AutoModel.from_pretrained(model_path, torch_dtype=torch.bfloat16, trust_remote_code=True)
|
23 |
-
model = model.to(device).eval()
|
24 |
-
print("Model and Tokenizer loaded.")
|
25 |
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
|
32 |
# --- Helper Functions ---
|
33 |
|
34 |
-
def parse_constraints(constraints_text):
|
35 |
"""Parse constraints in format: 'position:word, position:word, ...'"""
|
36 |
constraints = {}
|
|
|
37 |
if not constraints_text:
|
38 |
-
return constraints
|
39 |
|
40 |
parts = constraints_text.split(',')
|
41 |
for part in parts:
|
42 |
if ':' not in part:
|
43 |
continue
|
|
|
44 |
try:
|
45 |
-
pos_str, word = part.split(':', 1)
|
46 |
pos = int(pos_str.strip())
|
47 |
-
# Use strip() and lower() for robustness if needed, but preserve case for now
|
48 |
word = word.strip()
|
49 |
if word and pos >= 0:
|
50 |
-
#
|
51 |
-
|
|
|
|
|
52 |
prefix = " " if pos > 0 else ""
|
53 |
tokens = tokenizer.encode(prefix + word, add_special_tokens=False)
|
54 |
for i, token_id in enumerate(tokens):
|
55 |
-
#
|
56 |
-
|
57 |
-
|
58 |
-
constraints[pos + i] = token_id
|
59 |
except ValueError:
|
60 |
continue
|
61 |
except Exception as e:
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
return constraints
|
66 |
|
|
|
|
|
|
|
|
|
|
|
67 |
|
68 |
def format_chat_history(history):
|
69 |
"""
|
70 |
-
Format chat history for the Dream model
|
71 |
|
72 |
Args:
|
73 |
history: List of [user_message, assistant_message] pairs
|
74 |
|
75 |
Returns:
|
76 |
-
Formatted
|
77 |
"""
|
78 |
messages = []
|
79 |
-
#
|
80 |
-
if history
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
messages.append({"role": "user", "content": user_msg})
|
88 |
-
if assistant_msg is not None: # Skip if None (for the latest user message)
|
89 |
messages.append({"role": "assistant", "content": assistant_msg})
|
90 |
|
91 |
return messages
|
92 |
|
93 |
-
# --- Core Generation Logic
|
94 |
|
95 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
96 |
"""
|
97 |
-
|
98 |
-
Returns confidence and chosen token ID.
|
99 |
"""
|
100 |
-
|
101 |
-
|
102 |
-
|
103 |
-
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
# Calculate probabilities
|
121 |
-
probs = torch.softmax(logits, dim=-1)
|
122 |
-
|
123 |
-
# Sample or Argmax
|
124 |
-
if temperature > 0:
|
125 |
-
# Use torch distributions for robust sampling
|
126 |
-
dist = dists.Categorical(probs=probs)
|
127 |
-
x0 = dist.sample()
|
128 |
-
# Gather confidence for the sampled token
|
129 |
-
confidence = torch.gather(probs, -1, x0.unsqueeze(-1)).squeeze(-1)
|
130 |
-
else:
|
131 |
-
# Argmax for deterministic generation
|
132 |
-
confidence, x0 = torch.max(probs, dim=-1)
|
133 |
-
|
134 |
-
# --- Calculate specific confidence metrics if requested ---
|
135 |
-
# Note: These modify the 'confidence' variable *after* sampling x0
|
136 |
-
if margin_confidence:
|
137 |
-
if probs.shape[-1] >= 2:
|
138 |
-
# Ensure logits weren't completely masked, handle edge cases
|
139 |
-
if not torch.isinf(logits).all(dim=-1).any():
|
140 |
-
# Sort probabilities to get top1 and top2
|
141 |
-
sorted_probs, _ = torch.sort(probs, dim=-1, descending=True)
|
142 |
-
top1_probs = sorted_probs[..., 0]
|
143 |
-
top2_probs = sorted_probs[..., 1]
|
144 |
-
confidence = top1_probs - top2_probs
|
145 |
-
else:
|
146 |
-
# Fallback if all logits are -inf (shouldn't normally happen)
|
147 |
-
confidence.fill_(0.0) # Or some other indicator
|
148 |
-
else:
|
149 |
-
# Only one possible token, margin is undefined or 1? Set to top1 prob.
|
150 |
-
confidence, _ = torch.max(probs, dim=-1)
|
151 |
-
|
152 |
-
elif neg_entropy:
|
153 |
-
epsilon = 1e-9 # Slightly smaller epsilon
|
154 |
-
log_probs = torch.log(probs + epsilon)
|
155 |
-
# Negative entropy is sum(p * log(p))
|
156 |
-
confidence = torch.sum(probs * log_probs, dim=-1) # Lower value (more negative) is higher confidence
|
157 |
-
|
158 |
-
return confidence, x0
|
159 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
160 |
|
161 |
-
|
162 |
-
|
163 |
-
|
164 |
-
|
165 |
-
|
166 |
-
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
172 |
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
steps: Number of diffusion steps
|
177 |
-
constraints: Dictionary mapping positions to *token IDs*
|
178 |
-
temperature: Sampling temperature
|
179 |
-
top_p: Nucleus sampling probability
|
180 |
-
top_k: Top-k sampling
|
181 |
-
alg: Remasking strategy ('origin', 'maskgit_plus', 'topk_margin', 'entropy')
|
182 |
-
alg_temp: Temperature for confidence-based remasking randomness
|
183 |
-
yield_intermediate: Whether to yield intermediate states for visualization
|
184 |
|
185 |
-
|
186 |
-
|
187 |
-
|
188 |
-
|
189 |
-
constraints = {} # keys are positions relative to start of response
|
190 |
-
|
191 |
-
# --- Prepare Input ---
|
192 |
-
chat_input_text = tokenizer.apply_chat_template(
|
193 |
-
messages, add_generation_prompt=True, tokenize=False
|
194 |
-
)
|
195 |
-
input_ids = tokenizer(chat_input_text, return_tensors="pt")['input_ids'].to(device)
|
196 |
-
prompt_length = input_ids.shape[1]
|
197 |
-
max_length = prompt_length + gen_length
|
198 |
-
|
199 |
-
# Clamp max_length if it exceeds model capacity (use config value if available)
|
200 |
-
model_max_len = getattr(config, 'max_position_embeddings', 2048) # Default fallback
|
201 |
-
if max_length > model_max_len:
|
202 |
-
print(f"Warning: Requested length ({max_length}) exceeds model max ({model_max_len}). Clamping.")
|
203 |
-
max_length = model_max_len
|
204 |
-
gen_length = max_length - prompt_length
|
205 |
-
if gen_length <= 0:
|
206 |
-
print("Warning: Prompt is already at or exceeding model max length. Cannot generate.")
|
207 |
-
if yield_intermediate:
|
208 |
-
yield [], "Error: Prompt too long."
|
209 |
-
return
|
210 |
-
else:
|
211 |
-
return [], "Error: Prompt too long."
|
212 |
|
|
|
|
|
|
|
|
|
|
|
213 |
|
214 |
-
|
215 |
-
|
216 |
-
|
217 |
|
218 |
-
|
219 |
-
|
220 |
-
|
221 |
-
|
222 |
-
|
223 |
-
|
224 |
-
|
225 |
-
|
226 |
-
print(f"Warning: Skipped constraint for special token ID {token_id} at pos {rel_pos}")
|
227 |
|
|
|
|
|
|
|
228 |
|
229 |
-
# --- Visualization Setup ---
|
230 |
-
visualization_states = []
|
231 |
-
revealed_eos_pad = set() # Track positions where EOS/PAD was shown once
|
232 |
-
|
233 |
-
def get_vis_state(current_x, old_x, step_confidences=None):
|
234 |
-
nonlocal revealed_eos_pad
|
235 |
-
state = []
|
236 |
-
newly_revealed_in_step = False # Flag if any token changed from MASK
|
237 |
-
current_revealed_eos_pad = set() # Track EOS/PAD revealed *in this step*
|
238 |
-
|
239 |
-
for i in range(gen_length):
|
240 |
-
abs_pos = prompt_length + i
|
241 |
-
current_token_id = current_x[0, abs_pos].item()
|
242 |
-
old_token_id = old_x[0, abs_pos].item()
|
243 |
-
|
244 |
-
is_eos_or_pad = (current_token_id == EOS_ID or current_token_id == PAD_ID)
|
245 |
-
|
246 |
-
# Handle EOS/PAD hiding: Show once, then hide
|
247 |
-
if is_eos_or_pad and abs_pos in revealed_eos_pad:
|
248 |
-
state.append(("", "#FFFFFF")) # Make it invisible (white on white/transparent)
|
249 |
-
continue # Skip rest of logic for this pos
|
250 |
-
|
251 |
-
token_str = tokenizer.decode([current_token_id], skip_special_tokens=False) # Decode even specials initially
|
252 |
-
|
253 |
-
if current_token_id == MASK_ID:
|
254 |
-
color = "#444444" # Dark Gray for Mask
|
255 |
-
token_str = MASK_TOKEN # Display mask token string
|
256 |
-
elif old_token_id == MASK_ID: # Newly revealed in this step
|
257 |
-
newly_revealed_in_step = True
|
258 |
-
confidence = step_confidences.get(abs_pos, 0.5) # Get confidence if available, default 0.5
|
259 |
-
|
260 |
-
# Color based on confidence (adjust thresholds as needed)
|
261 |
-
# Note: Entropy confidence is negative, more negative = higher confidence
|
262 |
-
if alg == 'entropy':
|
263 |
-
# Example thresholds for negative entropy (adjust based on observation)
|
264 |
-
if confidence > -1.0: # Low confidence (high entropy)
|
265 |
-
color = "#FF6666" # Light Red
|
266 |
-
elif confidence > -3.0: # Medium confidence
|
267 |
-
color = "#FFAA33" # Orange
|
268 |
-
else: # High confidence (low entropy)
|
269 |
-
color = "#66CC66" # Light Green
|
270 |
-
else: # Standard confidence (probability or margin)
|
271 |
-
if confidence < 0.3:
|
272 |
-
color = "#FF6666" # Light Red
|
273 |
-
elif confidence < 0.7:
|
274 |
-
color = "#FFAA33" # Orange
|
275 |
-
else:
|
276 |
-
color = "#66CC66" # Light Green
|
277 |
-
|
278 |
-
# If it's EOS/PAD revealed now, mark for future hiding
|
279 |
-
if is_eos_or_pad:
|
280 |
-
current_revealed_eos_pad.add(abs_pos)
|
281 |
-
|
282 |
-
else: # Previously revealed
|
283 |
-
color = "#6699CC" # Light Blue
|
284 |
-
|
285 |
-
# Clean up token string for display (optional)
|
286 |
-
# token_str = token_str.replace(" ", " ") # Keep spaces visible
|
287 |
-
|
288 |
-
state.append((token_str, color))
|
289 |
-
|
290 |
-
# Update the global set of revealed EOS/PAD positions
|
291 |
-
revealed_eos_pad.update(current_revealed_eos_pad)
|
292 |
-
|
293 |
-
return state, newly_revealed_in_step
|
294 |
-
|
295 |
-
# Add initial state (all masked, constraints applied)
|
296 |
-
initial_vis_state, _ = get_vis_state(x, x) # Pass x as old_x initially
|
297 |
-
visualization_states.append(initial_vis_state)
|
298 |
-
if yield_intermediate:
|
299 |
-
yield initial_vis_state # Yield the starting state
|
300 |
-
|
301 |
-
# --- Diffusion Loop ---
|
302 |
-
timesteps = torch.linspace(1.0, 1e-3, steps + 1, device=device) # Use epsilon from Dream's defaults if needed
|
303 |
-
|
304 |
-
# Store the state before the loop starts
|
305 |
-
old_x = x.clone()
|
306 |
-
|
307 |
-
for i in range(steps):
|
308 |
-
# --- Core Dream Step ---
|
309 |
-
mask_index = (x == MASK_ID)
|
310 |
-
if not mask_index.any(): # Stop if no masks left
|
311 |
-
print(f"No masks left at step {i}. Stopping generation.")
|
312 |
-
break
|
313 |
-
|
314 |
-
# Prepare attention mask (full attention for Dream unless specified otherwise)
|
315 |
-
# Dream's modeling code handles standard causal masking internally based on position_ids
|
316 |
-
# For diffusion, we typically allow attending to everything (masked or not)
|
317 |
-
# The `model` forward pass expects a standard causal mask or None
|
318 |
-
# Let's use None, assuming the model handles positions correctly
|
319 |
-
attention_mask = None # Or potentially create a full mask: torch.ones_like(x)
|
320 |
-
|
321 |
-
# Create position_ids (simple range for now)
|
322 |
-
position_ids = torch.arange(0, x.shape[1], device=device).unsqueeze(0)
|
323 |
-
|
324 |
-
# Model forward pass
|
325 |
-
outputs = model(input_ids=x, attention_mask=attention_mask, position_ids=position_ids)
|
326 |
-
logits = outputs.logits
|
327 |
-
# logits = torch.cat([logits[:,:1], logits[:, :-1]], dim=1) # Dream applies shift in utils, replicate if needed
|
328 |
-
|
329 |
-
# Select logits for masked positions ONLY
|
330 |
-
# Need to handle batch dimension (which is 1 here)
|
331 |
-
current_mask_indices_flat = torch.where(mask_index.flatten())[0]
|
332 |
-
if len(current_mask_indices_flat) == 0:
|
333 |
-
print(f"No mask indices found flat at step {i}. Stopping generation.")
|
334 |
-
break
|
335 |
-
|
336 |
-
# Use advanced indexing to get logits for masked positions
|
337 |
-
# Logits shape: [batch_size, seq_len, vocab_size]
|
338 |
-
# Mask_index shape: [batch_size, seq_len]
|
339 |
-
# We need logits corresponding to True values in mask_index
|
340 |
-
# Example: batch_idx = torch.where(mask_index)[0], seq_idx = torch.where(mask_index)[1]
|
341 |
-
# mask_logits = logits[batch_idx, seq_idx]
|
342 |
-
batch_indices, seq_indices = torch.where(mask_index)
|
343 |
-
mask_logits = logits[batch_indices, seq_indices] # Shape: [num_masked_tokens, vocab_size]
|
344 |
-
|
345 |
-
if mask_logits.numel() == 0: # Double check after indexing
|
346 |
-
print(f"No mask logits selected at step {i}. Stopping generation.")
|
347 |
-
break
|
348 |
-
|
349 |
-
t = timesteps[i]
|
350 |
-
s = timesteps[i + 1]
|
351 |
-
|
352 |
-
# --- Remasking Logic (Simplified from Dream's _sample) ---
|
353 |
-
step_confidences = {} # Store confidences for revealed tokens in this step {abs_pos: confidence}
|
354 |
-
|
355 |
-
if alg == 'origin':
|
356 |
-
p_transfer = (1.0 - s / t) if i < steps - 1 else 1.0
|
357 |
-
# Sample for all masked positions
|
358 |
-
confidence, x0_masked = sample_tokens_for_vis(mask_logits, temperature=temperature, top_p=top_p, top_k=top_k)
|
359 |
-
# Decide which ones to transfer based on random probability
|
360 |
-
transfer_mask = torch.rand(x0_masked.shape, device=device) < p_transfer
|
361 |
-
# Create a tensor of MASK_IDs, and fill in the transferred tokens
|
362 |
-
updates_for_masked_pos = torch.full_like(x0_masked, MASK_ID)
|
363 |
-
updates_for_masked_pos[transfer_mask] = x0_masked[transfer_mask]
|
364 |
-
# Update x at the masked positions
|
365 |
-
x[mask_index] = updates_for_masked_pos
|
366 |
-
|
367 |
-
# Store confidences for the *transferred* tokens for visualization
|
368 |
-
transferred_indices_flat = current_mask_indices_flat[transfer_mask]
|
369 |
-
transferred_confidences = confidence[transfer_mask]
|
370 |
-
for flat_idx, conf in zip(transferred_indices_flat, transferred_confidences):
|
371 |
-
abs_pos = flat_idx.item() # Convert flat index back to seq position (assuming batch=1)
|
372 |
-
step_confidences[abs_pos] = conf.item()
|
373 |
-
|
374 |
-
|
375 |
-
else: # Confidence-based algorithms ('maskgit_plus', 'topk_margin', 'entropy')
|
376 |
-
use_margin = (alg == 'topk_margin')
|
377 |
-
use_entropy = (alg == 'entropy')
|
378 |
-
# Sample potential replacements for ALL masked positions first
|
379 |
-
confidence, x0_masked = sample_tokens_for_vis(
|
380 |
-
mask_logits,
|
381 |
-
temperature=temperature,
|
382 |
-
top_p=top_p,
|
383 |
-
top_k=top_k,
|
384 |
-
margin_confidence=use_margin,
|
385 |
-
neg_entropy=use_entropy
|
386 |
-
)
|
387 |
|
388 |
-
|
389 |
-
# Calculate how many tokens to unmask/transfer in this step
|
390 |
-
num_transfer_tokens = int(num_mask_tokens * (1.0 - s / t)) if i < steps - 1 else num_mask_tokens
|
391 |
-
|
392 |
-
if num_transfer_tokens > 0 and confidence.numel() > 0:
|
393 |
-
transfer_indices_relative = None # Indices relative to the masked tokens
|
394 |
-
if alg_temp is None or alg_temp <= 0:
|
395 |
-
# Deterministic: Select top-k confidence scores among masked tokens
|
396 |
-
# Ensure k is not larger than the number of masked tokens
|
397 |
-
k = min(num_transfer_tokens, confidence.shape[0])
|
398 |
-
if k > 0:
|
399 |
-
_, transfer_indices_relative = torch.topk(confidence, k)
|
400 |
-
else:
|
401 |
-
# Stochastic: Sample based on confidence scores
|
402 |
-
# Ensure probabilities are valid
|
403 |
-
conf_probs = F.softmax(confidence / alg_temp, dim=-1)
|
404 |
-
if not torch.isnan(conf_probs).any() and not torch.isinf(conf_probs).any() and conf_probs.sum() > 1e-6:
|
405 |
-
# Ensure k is not larger than the number of masked tokens
|
406 |
-
k = min(num_transfer_tokens, confidence.shape[0])
|
407 |
-
if k > 0:
|
408 |
-
transfer_indices_relative = torch.multinomial(conf_probs, num_samples=k, replacement=False)
|
409 |
-
else:
|
410 |
-
print(f"Warning: Invalid confidence probabilities at step {i}. Falling back to top-k.")
|
411 |
-
# Fallback to deterministic if sampling fails
|
412 |
-
k = min(num_transfer_tokens, confidence.shape[0])
|
413 |
-
if k > 0:
|
414 |
-
_, transfer_indices_relative = torch.topk(confidence, k)
|
415 |
-
|
416 |
-
|
417 |
-
if transfer_indices_relative is not None and transfer_indices_relative.numel() > 0:
|
418 |
-
# Create updates, initially all MASK_ID
|
419 |
-
updates_for_masked_pos = torch.full_like(x0_masked, MASK_ID)
|
420 |
-
# Place the selected sampled tokens into the updates tensor
|
421 |
-
updates_for_masked_pos[transfer_indices_relative] = x0_masked[transfer_indices_relative]
|
422 |
-
# Update x at the original masked positions
|
423 |
-
x[mask_index] = updates_for_masked_pos
|
424 |
-
|
425 |
-
# Store confidences for the *transferred* tokens for visualization
|
426 |
-
selected_confidences = confidence[transfer_indices_relative]
|
427 |
-
# Get the absolute positions corresponding to these relative indices
|
428 |
-
original_indices_flat = current_mask_indices_flat[transfer_indices_relative]
|
429 |
-
for flat_idx, conf in zip(original_indices_flat, selected_confidences):
|
430 |
-
abs_pos = flat_idx.item()
|
431 |
-
step_confidences[abs_pos] = conf.item()
|
432 |
|
433 |
-
|
434 |
-
# No tokens were selected to transfer, x remains unchanged for masked parts
|
435 |
-
pass # x[mask_index] remains MASK_ID essentially
|
436 |
|
437 |
-
else:
|
438 |
-
# If num_transfer_tokens is 0, x remains unchanged for masked parts
|
439 |
-
pass
|
440 |
-
|
441 |
-
# --- Apply Constraints and Finalize Step ---
|
442 |
-
# Ensure constraints are always maintained AFTER updates
|
443 |
-
for rel_pos, token_id in constraints.items():
|
444 |
-
abs_pos = prompt_length + rel_pos
|
445 |
-
if abs_pos < max_length:
|
446 |
-
# Check if the position was masked before applying constraint
|
447 |
-
# if mask_index[0, abs_pos]: # Only apply if it *was* a mask, maybe? Optional.
|
448 |
-
x[:, abs_pos] = token_id
|
449 |
-
|
450 |
-
# --- Visualization Update ---
|
451 |
-
current_vis_state, newly_revealed = get_vis_state(x, old_x, step_confidences)
|
452 |
-
|
453 |
-
# Only add/yield if something actually changed or if it's the last step
|
454 |
-
if newly_revealed or i == steps - 1:
|
455 |
-
visualization_states.append(current_vis_state)
|
456 |
-
if yield_intermediate:
|
457 |
-
yield current_vis_state
|
458 |
-
|
459 |
-
# Update old_x for the next iteration
|
460 |
-
old_x = x.clone()
|
461 |
-
|
462 |
-
|
463 |
-
# --- Final Output ---
|
464 |
-
response_tokens = x[0, prompt_length:]
|
465 |
-
# Decode, cleaning up potential special tokens unless they are intended
|
466 |
-
final_text = tokenizer.decode(response_tokens,
|
467 |
-
skip_special_tokens=True, # Skip things like <|mask|> in final output
|
468 |
-
clean_up_tokenization_spaces=True)
|
469 |
-
|
470 |
-
# If not yielding intermediates, return the full list now
|
471 |
-
if not yield_intermediate:
|
472 |
-
return visualization_states, final_text
|
473 |
-
else:
|
474 |
-
# If yielding intermediates, we still need a way to signal completion
|
475 |
-
# and return the final text. Gradio's yield typically handles this if
|
476 |
-
# the last yielded value is the final one. We'll return the final text
|
477 |
-
# separately after the loop finishes in the calling function.
|
478 |
-
# The loop yields states, the calling function returns the final text.
|
479 |
-
pass # Final text is handled outside the generator function
|
480 |
-
|
481 |
-
|
482 |
-
# --- Gradio UI ---
|
483 |
css = '''
|
484 |
.category-legend{display:none}
|
485 |
button{height: 60px}
|
486 |
-
.token-
|
487 |
-
|
488 |
-
.token-new-high { background-color: #66CC66; color: black; padding: 1px 2px; margin: 1px; border-radius: 3px; display: inline-block; }
|
489 |
-
.token-new-mid { background-color: #FFAA33; color: black; padding: 1px 2px; margin: 1px; border-radius: 3px; display: inline-block; }
|
490 |
-
.token-new-low { background-color: #FF6666; color: black; padding: 1px 2px; margin: 1px; border-radius: 3px; display: inline-block; }
|
491 |
-
.token-old { background-color: #6699CC; color: white; padding: 1px 2px; margin: 1px; border-radius: 3px; display: inline-block; }
|
492 |
-
.token-hidden { display: none; } /* Hide EOS/PAD after first reveal */
|
493 |
'''
|
494 |
-
|
495 |
def create_chatbot_demo():
|
496 |
with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
|
497 |
gr.Markdown("# Dream 7B - Diffusion Language Model Demo")
|
498 |
gr.Markdown(
|
499 |
"[[Model Card](https://huggingface.co/Dream-org/Dream-v0-Instruct-7B)] "
|
500 |
-
"[[Blog](https://hkunlp.github.io/blog/2025/dream/)] "
|
501 |
-
"
|
502 |
-
)
|
503 |
-
gr.Markdown(
|
504 |
-
"**Note:** This demo visualizes the diffusion process in real-time. "
|
505 |
-
"Tokens start masked (<font color='#444444'>[MASK]</font>) and are revealed step-by-step. "
|
506 |
-
"Colors indicate confidence: <font color='#66CC66'>High</font>, "
|
507 |
-
"<font color='#FFAA33'>Medium</font>, <font color='#FF6666'>Low</font>. "
|
508 |
-
"Previously revealed tokens are <font color='#6699CC'>blue</font>. "
|
509 |
-
f"EOS/PAD tokens ({tokenizer.decode([EOS_ID])}) are hidden after appearing once."
|
510 |
)
|
511 |
|
512 |
# STATE MANAGEMENT
|
513 |
chat_history = gr.State([])
|
514 |
-
|
|
|
515 |
|
516 |
# UI COMPONENTS
|
517 |
with gr.Row():
|
518 |
with gr.Column(scale=3):
|
519 |
-
chatbot_ui = gr.Chatbot(
|
|
|
|
|
|
|
|
|
520 |
|
521 |
# Message input
|
522 |
with gr.Group():
|
@@ -524,229 +340,219 @@ def create_chatbot_demo():
|
|
524 |
user_input = gr.Textbox(
|
525 |
label="Your Message",
|
526 |
placeholder="Type your message here...",
|
527 |
-
|
528 |
-
|
529 |
)
|
530 |
send_btn = gr.Button("Send", scale=1)
|
531 |
|
532 |
constraints_input = gr.Textbox(
|
533 |
-
label="Word Constraints (
|
534 |
-
info="Place words at
|
535 |
-
placeholder="0:
|
536 |
value=""
|
537 |
)
|
538 |
with gr.Column(scale=2):
|
539 |
-
# Use HighlightedText with specific classes for better styling control
|
540 |
output_vis = gr.HighlightedText(
|
541 |
label="Denoising Process Visualization",
|
542 |
-
|
543 |
-
#
|
544 |
-
|
545 |
-
#
|
546 |
-
# color_map={ # This might not work directly with dynamic classes, CSS is better
|
547 |
-
# "MASK": "#444444", "NEW_H": "#66CC66", "NEW_M": "#FFAA33",
|
548 |
-
# "NEW_L": "#FF6666", "OLD": "#6699CC", "HIDDEN": "#FFFFFF"
|
549 |
-
# }
|
550 |
-
combine_adjacent=False, # Keep tokens separate
|
551 |
-
height=550, # Adjust height as needed
|
552 |
)
|
553 |
|
554 |
-
|
555 |
# Advanced generation settings
|
556 |
with gr.Accordion("Generation Settings", open=False):
|
557 |
with gr.Row():
|
558 |
gen_length = gr.Slider(
|
559 |
-
minimum=16, maximum=512, value=
|
560 |
label="Max New Tokens"
|
561 |
)
|
562 |
steps = gr.Slider(
|
563 |
-
minimum=8, maximum=512, value=
|
564 |
-
label="
|
565 |
)
|
566 |
with gr.Row():
|
567 |
temperature = gr.Slider(
|
568 |
-
minimum=0.0, maximum=1.
|
569 |
label="Temperature"
|
570 |
)
|
571 |
top_p = gr.Slider(
|
572 |
-
minimum=0.
|
573 |
-
label="Top-P
|
574 |
-
)
|
575 |
-
# top_k = gr.Slider(
|
576 |
-
# minimum=0, maximum=200, value=0, step=5, # Allow Top-K=0 (disabled)
|
577 |
-
# label="Top-K (0 to disable)"
|
578 |
-
# )
|
579 |
-
with gr.Row():
|
580 |
-
# Dream specific algorithm choice
|
581 |
-
alg_strategy = gr.Radio(
|
582 |
-
choices=["entropy", "maskgit_plus", "topk_margin", "origin"],
|
583 |
-
value="entropy",
|
584 |
-
label="Algorithm (`alg`)"
|
585 |
)
|
586 |
-
|
587 |
-
minimum=0
|
588 |
-
label="
|
589 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
590 |
with gr.Row():
|
591 |
visualization_delay = gr.Slider(
|
592 |
-
minimum=0.0, maximum=0.5, value=0.
|
593 |
label="Visualization Delay (seconds)"
|
594 |
)
|
595 |
|
596 |
# Clear button
|
597 |
clear_btn = gr.Button("Clear Conversation")
|
598 |
|
599 |
-
# ---
|
600 |
-
def
|
601 |
-
"""Add a message pair to the history
|
|
|
|
|
|
|
602 |
history.append([message, response])
|
603 |
return history
|
604 |
|
605 |
-
def
|
606 |
-
"""
|
607 |
-
if not message
|
608 |
-
return history, history, "", []
|
609 |
-
|
610 |
-
# Add user message
|
611 |
-
history =
|
612 |
-
|
613 |
-
|
614 |
-
|
615 |
-
|
616 |
-
|
617 |
-
|
618 |
-
|
619 |
-
|
620 |
-
|
621 |
-
|
622 |
-
|
623 |
-
|
624 |
-
|
625 |
-
|
626 |
-
|
627 |
-
|
628 |
-
|
629 |
-
|
630 |
-
# Format history for the model (excluding the last None response)
|
631 |
-
messages = format_chat_history(history[:-1])
|
632 |
-
# Add the current user message
|
633 |
-
messages.append({"role": "user", "content": last_user_message})
|
634 |
-
|
635 |
-
# Parse constraints into token IDs
|
636 |
-
parsed_constraints = parse_constraints(constraints_str)
|
637 |
-
print(f"Parsed constraints: {parsed_constraints}")
|
638 |
-
|
639 |
-
|
640 |
-
final_text = "" # Initialize final_text
|
641 |
-
|
642 |
-
# Use the generator function
|
643 |
-
response_generator = generate_response_with_visualization_dream(
|
644 |
-
messages,
|
645 |
-
gen_length=gen_length,
|
646 |
-
steps=steps,
|
647 |
-
constraints=parsed_constraints,
|
648 |
-
temperature=temperature,
|
649 |
-
top_p=top_p if top_p > 0 else None, # Pass None if 0
|
650 |
-
top_k=None, # Pass None if 0 top_k if top_k > 0 else None,
|
651 |
-
alg=alg,
|
652 |
-
alg_temp=alg_temp if alg_temp > 0 else None, # Pass None if 0
|
653 |
-
yield_intermediate=True
|
654 |
-
)
|
655 |
|
656 |
-
|
657 |
-
|
658 |
-
|
659 |
-
|
660 |
-
|
661 |
-
|
662 |
-
|
663 |
-
|
664 |
-
|
665 |
-
|
666 |
-
|
667 |
-
|
668 |
-
|
669 |
-
|
670 |
-
|
671 |
-
|
672 |
-
|
673 |
-
|
674 |
-
|
675 |
-
|
676 |
-
|
677 |
-
|
678 |
-
|
679 |
-
|
680 |
|
|
|
|
|
681 |
|
682 |
-
# Update the history with the actual final response
|
683 |
-
history[-1][1] = final_text.strip() if final_text else "[No response]"
|
684 |
|
685 |
-
|
686 |
-
|
|
|
|
|
|
|
|
|
687 |
|
688 |
-
except Exception as e:
|
689 |
-
import traceback
|
690 |
-
print(f"Error during generation: {e}")
|
691 |
-
traceback.print_exc()
|
692 |
-
error_msg = f"Error: {str(e)}"
|
693 |
-
history[-1][1] = error_msg # Show error in chat
|
694 |
-
# Show error in visualization (red text)
|
695 |
-
error_vis = [(error_msg, "#FF0000")]
|
696 |
-
yield history, error_vis, error_msg
|
697 |
|
|
|
|
|
|
|
|
|
|
|
|
|
698 |
|
699 |
-
|
700 |
-
|
701 |
-
return [], [], "", [] # History, Chatbot UI, Response Text, Visualization
|
702 |
|
|
|
|
|
|
|
703 |
|
704 |
-
|
|
|
705 |
|
706 |
-
# 1. User Submits Message (Textbox Enter or Button Click)
|
707 |
-
submit_triggers = [user_input.submit, send_btn.click]
|
708 |
-
for trigger in submit_triggers:
|
709 |
-
trigger.then(
|
710 |
-
fn=user_message_action,
|
711 |
-
inputs=[user_input, chat_history],
|
712 |
-
outputs=[chat_history, chatbot_ui, user_input, output_vis, current_response_text], # Update history state, chatbot UI, clear input, clear vis, clear response state
|
713 |
-
queue=True # Enable queue for handling multiple users
|
714 |
-
).then(
|
715 |
-
# 2. Trigger Bot Response Generation (Generator Function)
|
716 |
-
fn=bot_response_generator,
|
717 |
-
inputs=[
|
718 |
-
chat_history, gen_length, steps, constraints_input, visualization_delay,
|
719 |
-
temperature, top_p, # top_k,
|
720 |
-
alg_strategy, alg_temp
|
721 |
-
],
|
722 |
-
outputs=[chatbot_ui, output_vis, current_response_text] # Stream updates to chatbot, visualization, and store final text
|
723 |
-
)
|
724 |
|
725 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
726 |
clear_btn.click(
|
727 |
-
fn=
|
728 |
inputs=[],
|
729 |
-
outputs=[chat_history, chatbot_ui,
|
730 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
731 |
)
|
732 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
733 |
return demo
|
734 |
|
735 |
-
# --- Launch ---
|
736 |
if __name__ == "__main__":
|
737 |
-
# Make sure the necessary Dream model files (modeling_dream.py, configuration_dream.py etc.)
|
738 |
-
# are in the same directory or accessible in the Python path.
|
739 |
-
# Also ensure 'generation_utils.py' is available if needed by the model loading/config.
|
740 |
-
# Check if 'modeling_dream.py' exists before launching
|
741 |
-
import os
|
742 |
-
if not os.path.exists("modeling_dream.py") or not os.path.exists("configuration_dream.py"):
|
743 |
-
print("\nERROR: Could not find 'modeling_dream.py' and/or 'configuration_dream.py'.")
|
744 |
-
print("Please make sure these files (from the 'dream_model.txt' source) are in the same directory as this script.")
|
745 |
-
print("You might need to extract them from the provided text file.")
|
746 |
-
# exit() # Optional: stop execution if files are missing
|
747 |
-
|
748 |
-
print("Creating Gradio Demo...")
|
749 |
demo = create_chatbot_demo()
|
750 |
-
|
751 |
-
|
752 |
-
demo.queue().launch(share=True, debug=True) # Enable debug for more detailed logs
|
|
|
4 |
import gradio as gr
|
5 |
import spaces
|
6 |
import torch.nn.functional as F
|
7 |
+
from transformers import AutoTokenizer, AutoModel, AutoConfig
|
|
|
8 |
import time
|
9 |
+
import copy
|
|
|
10 |
|
11 |
+
# Determine device
|
12 |
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
13 |
print(f"Using device: {device}")
|
14 |
|
15 |
+
# --- Model and Tokenizer Loading ---
|
16 |
model_path = "Dream-org/Dream-v0-Instruct-7B"
|
17 |
+
|
18 |
+
print(f"Loading tokenizer from {model_path}...")
|
19 |
+
# Load configuration first to get special token IDs
|
20 |
+
config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
|
21 |
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
|
|
|
|
|
|
|
22 |
|
23 |
+
print(f"Loading model from {model_path}...")
|
24 |
+
model = AutoModel.from_pretrained(
|
25 |
+
model_path,
|
26 |
+
torch_dtype=torch.bfloat16,
|
27 |
+
trust_remote_code=True
|
28 |
+
).to(device).eval()
|
29 |
+
print("Model loaded successfully.")
|
30 |
+
|
31 |
+
# --- Constants from Dream Model ---
|
32 |
+
# Get IDs directly from config or tokenizer if available
|
33 |
+
MASK_TOKEN = tokenizer.mask_token
|
34 |
+
MASK_ID = config.mask_token_id if hasattr(config, 'mask_token_id') else tokenizer.mask_token_id
|
35 |
+
EOS_ID = config.eos_token_id if hasattr(config, 'eos_token_id') else tokenizer.eos_token_id
|
36 |
+
PAD_ID = config.pad_token_id if hasattr(config, 'pad_token_id') else tokenizer.pad_token_id # Often same as EOS
|
37 |
+
|
38 |
+
print(f"MASK_TOKEN: '{MASK_TOKEN}', MASK_ID: {MASK_ID}")
|
39 |
+
print(f"EOS_ID: {EOS_ID}, PAD_ID: {PAD_ID}")
|
40 |
+
if MASK_ID is None:
|
41 |
+
raise ValueError("Could not determine MASK_ID from model config or tokenizer.")
|
42 |
+
if EOS_ID is None:
|
43 |
+
raise ValueError("Could not determine EOS_ID from model config or tokenizer.")
|
44 |
+
if PAD_ID is None:
|
45 |
+
raise ValueError("Could not determine PAD_ID from model config or tokenizer.")
|
46 |
+
|
47 |
|
48 |
# --- Helper Functions ---
|
49 |
|
50 |
+
def parse_constraints(constraints_text, tokenizer):
|
51 |
"""Parse constraints in format: 'position:word, position:word, ...'"""
|
52 |
constraints = {}
|
53 |
+
processed_constraints_tokens = {}
|
54 |
if not constraints_text:
|
55 |
+
return constraints, processed_constraints_tokens
|
56 |
|
57 |
parts = constraints_text.split(',')
|
58 |
for part in parts:
|
59 |
if ':' not in part:
|
60 |
continue
|
61 |
+
pos_str, word = part.split(':', 1)
|
62 |
try:
|
|
|
63 |
pos = int(pos_str.strip())
|
|
|
64 |
word = word.strip()
|
65 |
if word and pos >= 0:
|
66 |
+
# Store original word constraint for display/debugging if needed
|
67 |
+
constraints[pos] = word
|
68 |
+
# Tokenize the word (add space for consistency if not BOS)
|
69 |
+
# Note: Dream tokenizer might handle spaces differently, adjust if needed
|
70 |
prefix = " " if pos > 0 else ""
|
71 |
tokens = tokenizer.encode(prefix + word, add_special_tokens=False)
|
72 |
for i, token_id in enumerate(tokens):
|
73 |
+
# Prevent overwriting multi-token constraints partially
|
74 |
+
if pos + i not in processed_constraints_tokens:
|
75 |
+
processed_constraints_tokens[pos + i] = token_id
|
|
|
76 |
except ValueError:
|
77 |
continue
|
78 |
except Exception as e:
|
79 |
+
print(f"Error tokenizing constraint word '{word}': {e}")
|
80 |
+
continue
|
|
|
|
|
81 |
|
82 |
+
# Sort by position for consistent application
|
83 |
+
processed_constraints_tokens = dict(sorted(processed_constraints_tokens.items()))
|
84 |
+
print(f"Parsed Constraints (Word): {constraints}")
|
85 |
+
print(f"Parsed Constraints (Tokens): {processed_constraints_tokens}")
|
86 |
+
return constraints, processed_constraints_tokens
|
87 |
|
88 |
def format_chat_history(history):
|
89 |
"""
|
90 |
+
Format chat history for the Dream model using its chat template convention.
|
91 |
|
92 |
Args:
|
93 |
history: List of [user_message, assistant_message] pairs
|
94 |
|
95 |
Returns:
|
96 |
+
Formatted list of message dictionaries for the model
|
97 |
"""
|
98 |
messages = []
|
99 |
+
# Add system prompt if not present (standard practice)
|
100 |
+
if not history or history[0][0] is None or history[0][0].lower() != "system":
|
101 |
+
messages.append({"role": "system", "content": "You are a helpful assistant."})
|
102 |
+
|
103 |
+
for user_msg, assistant_msg in history:
|
104 |
+
if user_msg is not None: # Handle potential system message case
|
105 |
+
messages.append({"role": "user", "content": user_msg})
|
106 |
+
if assistant_msg: # Skip if None (for the latest user message)
|
|
|
|
|
107 |
messages.append({"role": "assistant", "content": assistant_msg})
|
108 |
|
109 |
return messages
|
110 |
|
111 |
+
# --- Core Generation Logic with Visualization Hook ---
|
112 |
|
113 |
+
@spaces.GPU
|
114 |
+
def generate_response_with_visualization(
|
115 |
+
messages, # List of message dictionaries
|
116 |
+
gen_length=64,
|
117 |
+
steps=64,
|
118 |
+
constraints_text="", # Raw constraint text
|
119 |
+
temperature=0.2,
|
120 |
+
top_p=0.95,
|
121 |
+
top_k=None, # Added for Dream
|
122 |
+
alg="entropy", # Changed from remasking
|
123 |
+
alg_temp=0.0, # Added for Dream
|
124 |
+
visualization_delay=0.05,
|
125 |
+
tokenizer=tokenizer,
|
126 |
+
model=model,
|
127 |
+
device=device,
|
128 |
+
MASK_ID=MASK_ID,
|
129 |
+
EOS_ID=EOS_ID,
|
130 |
+
PAD_ID=PAD_ID
|
131 |
+
):
|
132 |
"""
|
133 |
+
Generate text with Dream model with real-time visualization using a hook.
|
|
|
134 |
"""
|
135 |
+
visualization_states = []
|
136 |
+
final_text = ""
|
137 |
+
# Use a list to hold previous_x, allowing nonlocal modification
|
138 |
+
# Initialize with None, it will be set after the first hook call
|
139 |
+
shared_state = {'previous_x': None}
|
140 |
+
|
141 |
+
|
142 |
+
try:
|
143 |
+
# --- 1. Prepare Inputs ---
|
144 |
+
_, parsed_constraints_tokens = parse_constraints(constraints_text, tokenizer)
|
145 |
+
|
146 |
+
# Apply chat template
|
147 |
+
# Important: Keep tokenize=False initially to get prompt length correctly
|
148 |
+
# The template adds roles and special tokens like <|im_start|> etc.
|
149 |
+
chat_input_text = tokenizer.apply_chat_template(
|
150 |
+
messages,
|
151 |
+
add_generation_prompt=True, # Adds the prompt for the assistant's turn
|
152 |
+
tokenize=False
|
153 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
154 |
|
155 |
+
# Tokenize the full templated chat string
|
156 |
+
inputs = tokenizer(chat_input_text, return_tensors="pt", return_dict=True)
|
157 |
+
input_ids = inputs.input_ids.to(device)
|
158 |
+
attention_mask = inputs.attention_mask.to(device) # Use mask from tokenizer
|
159 |
+
|
160 |
+
prompt_length = input_ids.shape[1]
|
161 |
+
total_length = prompt_length + gen_length
|
162 |
+
|
163 |
+
# --- 2. Initialize Generation Sequence ---
|
164 |
+
# Start with the prompt, pad the rest with MASK_ID
|
165 |
+
x = torch.full((1, total_length), MASK_ID, dtype=torch.long, device=device)
|
166 |
+
x[:, :prompt_length] = input_ids.clone()
|
167 |
+
attention_mask = F.pad(attention_mask, (0, gen_length), value=1) # Extend attention mask
|
168 |
+
|
169 |
+
# Apply initial constraints to the masked sequence `x`
|
170 |
+
for pos, token_id in parsed_constraints_tokens.items():
|
171 |
+
absolute_pos = prompt_length + pos
|
172 |
+
if absolute_pos < total_length:
|
173 |
+
print(f"Applying initial constraint at pos {absolute_pos}: token {token_id}")
|
174 |
+
x[:, absolute_pos] = token_id
|
175 |
+
|
176 |
+
# Store initial state (prompt + all masked) for visualization
|
177 |
+
initial_state_vis = []
|
178 |
+
# Add prompt tokens (optional visualization, could be grayed out or skipped)
|
179 |
+
# for i in range(prompt_length):
|
180 |
+
# token_str = tokenizer.decode([x[0, i].item()], skip_special_tokens=True)
|
181 |
+
# initial_state_vis.append((token_str if token_str else " ", "#AAAAAA")) # Gray for prompt
|
182 |
+
|
183 |
+
# Add masked tokens for the generation part
|
184 |
+
for _ in range(gen_length):
|
185 |
+
initial_state_vis.append((MASK_TOKEN, "#444444")) # Dark gray for masks
|
186 |
+
visualization_states.append(initial_state_vis)
|
187 |
+
shared_state['previous_x'] = x.clone() # Initialize previous_x
|
188 |
+
|
189 |
+
|
190 |
+
# --- 3. Define the Visualization Hook ---
|
191 |
+
def generation_tokens_hook_func(step, current_x_hook, logits):
|
192 |
+
# nonlocal previous_x # Allow modification of the outer scope variable
|
193 |
+
current_x_hook = current_x_hook.clone() # Work on a copy
|
194 |
+
|
195 |
+
# --- Apply constraints within the hook ---
|
196 |
+
# This ensures constraints are respected even if the model tries to overwrite them
|
197 |
+
for pos, token_id in parsed_constraints_tokens.items():
|
198 |
+
absolute_pos = prompt_length + pos
|
199 |
+
if absolute_pos < total_length:
|
200 |
+
current_x_hook[:, absolute_pos] = token_id
|
201 |
+
# --- End Constraint Application ---
|
202 |
+
|
203 |
+
if shared_state['previous_x'] is None: # First call
|
204 |
+
shared_state['previous_x'] = current_x_hook.clone()
|
205 |
+
return current_x_hook # Must return the (potentially modified) sequence
|
206 |
+
|
207 |
+
# Generate visualization state for this step
|
208 |
+
current_state_vis = []
|
209 |
+
prev_x_step = shared_state['previous_x']
|
210 |
+
|
211 |
+
for i in range(gen_length):
|
212 |
+
pos = prompt_length + i # Absolute position in the sequence
|
213 |
+
current_token_id = current_x_hook[0, pos].item()
|
214 |
+
prev_token_id = prev_x_step[0, pos].item()
|
215 |
+
|
216 |
+
# Decode token, handling special tokens we want to hide
|
217 |
+
token_str = ""
|
218 |
+
color = "#444444" # Default: Dark Gray (Mask)
|
219 |
+
token_str_raw = tokenizer.decode([current_token_id], skip_special_tokens=False) # Keep special tokens for ID check
|
220 |
+
|
221 |
+
if current_token_id == MASK_ID:
|
222 |
+
token_str = MASK_TOKEN
|
223 |
+
color = "#444444" # Dark gray
|
224 |
+
elif current_token_id == EOS_ID or current_token_id == PAD_ID:
|
225 |
+
token_str = "" # Hide EOS/PAD visually
|
226 |
+
color = "#DDDDDD" # Use a light gray or make transparent if possible
|
227 |
+
else:
|
228 |
+
# Decode without special tokens for display if it's not MASK/EOS/PAD
|
229 |
+
token_str = tokenizer.decode([current_token_id], skip_special_tokens=True).strip()
|
230 |
+
if not token_str: token_str = token_str_raw # Fallback if strip removes everything (e.g., space)
|
231 |
|
232 |
+
if prev_token_id == MASK_ID:
|
233 |
+
# Newly revealed in this step
|
234 |
+
color = "#66CC66" # Light green (Simplified from confidence levels)
|
235 |
+
else:
|
236 |
+
# Previously revealed
|
237 |
+
color = "#6699CC" # Light blue
|
238 |
+
|
239 |
+
current_state_vis.append((token_str if token_str else " ", color)) # Ensure non-empty tuple element
|
240 |
+
|
241 |
+
visualization_states.append(current_state_vis)
|
242 |
+
shared_state['previous_x'] = current_x_hook.clone() # Update previous_x for the next step
|
243 |
+
|
244 |
+
return current_x_hook # Return the sequence (constraints applied)
|
245 |
+
|
246 |
+
# --- 4. Run Diffusion Generation ---
|
247 |
+
print("Starting diffusion generation...")
|
248 |
+
start_time = time.time()
|
249 |
+
output = model.diffusion_generate(
|
250 |
+
input_ids=x[:, :prompt_length], # Pass only the initial prompt to diffusion_generate
|
251 |
+
# as it handles the masking internally based on MASK_ID
|
252 |
+
attention_mask=attention_mask, # Provide the full attention mask
|
253 |
+
max_new_tokens=gen_length,
|
254 |
+
output_history=False, # We capture history via the hook
|
255 |
+
return_dict_in_generate=True,
|
256 |
+
steps=steps,
|
257 |
+
temperature=temperature,
|
258 |
+
top_p=top_p,
|
259 |
+
top_k=top_k,
|
260 |
+
alg=alg,
|
261 |
+
alg_temp=alg_temp if alg != 'origin' else None, # alg_temp only for confidence-based
|
262 |
+
# Pass the hook function
|
263 |
+
generation_tokens_hook_func=generation_tokens_hook_func,
|
264 |
+
# Ensure the initial masked sequence `x` is used correctly if needed by internal logic
|
265 |
+
# Depending on the exact implementation of diffusion_generate, passing x directly might be needed
|
266 |
+
# Check Dream's generation_utils if issues arise. For now, assume it uses input_ids + max_new_tokens
|
267 |
+
)
|
268 |
+
end_time = time.time()
|
269 |
+
print(f"Diffusion generation finished in {end_time - start_time:.2f} seconds.")
|
270 |
|
271 |
+
# --- 5. Process Final Output ---
|
272 |
+
# The hook has already built visualization_states
|
273 |
+
final_sequence = output.sequences[0]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
274 |
|
275 |
+
# Decode the generated part, skipping special tokens for the final text output
|
276 |
+
response_tokens = final_sequence[prompt_length:]
|
277 |
+
# Filter out PAD tokens before final decode, keep EOS if needed conceptually, but skip for clean text
|
278 |
+
response_tokens_cleaned = [tok for tok in response_tokens if tok != PAD_ID] # Keep EOS initially if needed
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
279 |
|
280 |
+
final_text = tokenizer.decode(
|
281 |
+
response_tokens_cleaned,
|
282 |
+
skip_special_tokens=True, # Skip EOS, BOS, etc.
|
283 |
+
clean_up_tokenization_spaces=True # Recommended for cleaner output
|
284 |
+
).strip()
|
285 |
|
286 |
+
# Ensure the last state in visualization matches the final text (debug check)
|
287 |
+
# print(f"Last Vis State Tokens: {''.join([t[0] for t in visualization_states[-1]]).strip()}")
|
288 |
+
# print(f"Final Decoded Text: {final_text}")
|
289 |
|
290 |
+
except Exception as e:
|
291 |
+
print(f"Error during generation: {e}")
|
292 |
+
import traceback
|
293 |
+
traceback.print_exc()
|
294 |
+
# Add error message to visualization
|
295 |
+
error_msg = f"Error: {str(e)}"
|
296 |
+
visualization_states.append([(error_msg, "red")])
|
297 |
+
final_text = error_msg # Display error in the chatbot too
|
|
|
298 |
|
299 |
+
# Make sure at least the initial state is present
|
300 |
+
if not visualization_states:
|
301 |
+
visualization_states.append([("Error: No states generated.", "red")])
|
302 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
303 |
|
304 |
+
return visualization_states, final_text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
305 |
|
306 |
+
# --- Gradio UI Definition ---
|
|
|
|
|
307 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
308 |
css = '''
|
309 |
.category-legend{display:none}
|
310 |
button{height: 60px}
|
311 |
+
.token-text { white-space: pre; } /* Preserve spaces in tokens */
|
312 |
+
footer { display: none !important; visibility: hidden !important; }
|
|
|
|
|
|
|
|
|
|
|
313 |
'''
|
|
|
314 |
def create_chatbot_demo():
|
315 |
with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
|
316 |
gr.Markdown("# Dream 7B - Diffusion Language Model Demo")
|
317 |
gr.Markdown(
|
318 |
"[[Model Card](https://huggingface.co/Dream-org/Dream-v0-Instruct-7B)] "
|
319 |
+
"[[Blog Post](https://hkunlp.github.io/blog/2025/dream/)] "
|
320 |
+
"(Note: Visualization shows token reveal steps, colors indicate status: Gray=Masked, Green=Newly Revealed, Blue=Previously Revealed)"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
321 |
)
|
322 |
|
323 |
# STATE MANAGEMENT
|
324 |
chat_history = gr.State([])
|
325 |
+
# Store constraints parsed into token IDs
|
326 |
+
parsed_constraints_state = gr.State({})
|
327 |
|
328 |
# UI COMPONENTS
|
329 |
with gr.Row():
|
330 |
with gr.Column(scale=3):
|
331 |
+
chatbot_ui = gr.Chatbot(
|
332 |
+
label="Conversation",
|
333 |
+
height=500,
|
334 |
+
bubble_full_width=False # Makes bubbles wrap content
|
335 |
+
)
|
336 |
|
337 |
# Message input
|
338 |
with gr.Group():
|
|
|
340 |
user_input = gr.Textbox(
|
341 |
label="Your Message",
|
342 |
placeholder="Type your message here...",
|
343 |
+
show_label=False,
|
344 |
+
scale=7
|
345 |
)
|
346 |
send_btn = gr.Button("Send", scale=1)
|
347 |
|
348 |
constraints_input = gr.Textbox(
|
349 |
+
label="Word Constraints (Experimental)",
|
350 |
+
info="Place specific words at positions (0-indexed). Format: 'pos:word, pos:word'. Example: '0:Once, 5:upon, 10:time'. Multi-token words supported.",
|
351 |
+
placeholder="0:The, 10:story",
|
352 |
value=""
|
353 |
)
|
354 |
with gr.Column(scale=2):
|
|
|
355 |
output_vis = gr.HighlightedText(
|
356 |
label="Denoising Process Visualization",
|
357 |
+
combine_adjacent=False,
|
358 |
+
show_legend=False, # Legend not very informative here
|
359 |
+
height=560, # Match chatbot height + input box approx
|
360 |
+
elem_classes=["token-text"] # Apply custom class for styling if needed
|
|
|
|
|
|
|
|
|
|
|
|
|
361 |
)
|
362 |
|
|
|
363 |
# Advanced generation settings
|
364 |
with gr.Accordion("Generation Settings", open=False):
|
365 |
with gr.Row():
|
366 |
gen_length = gr.Slider(
|
367 |
+
minimum=16, maximum=512, value=128, step=8,
|
368 |
label="Max New Tokens"
|
369 |
)
|
370 |
steps = gr.Slider(
|
371 |
+
minimum=8, maximum=512, value=128, step=4,
|
372 |
+
label="Denoising Steps"
|
373 |
)
|
374 |
with gr.Row():
|
375 |
temperature = gr.Slider(
|
376 |
+
minimum=0.0, maximum=1.0, value=0.2, step=0.05,
|
377 |
label="Temperature"
|
378 |
)
|
379 |
top_p = gr.Slider(
|
380 |
+
minimum=0.1, maximum=1.0, value=0.95, step=0.05,
|
381 |
+
label="Top-P"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
382 |
)
|
383 |
+
top_k = gr.Slider(
|
384 |
+
minimum=0, maximum=200, value=0, step=5,
|
385 |
+
label="Top-K (0=disabled)"
|
386 |
)
|
387 |
+
with gr.Row():
|
388 |
+
alg = gr.Radio(
|
389 |
+
choices=['origin', 'maskgit_plus', 'topk_margin', 'entropy'],
|
390 |
+
value='entropy',
|
391 |
+
label="Sampling Algorithm (`alg`)"
|
392 |
+
)
|
393 |
+
alg_temp = gr.Slider(
|
394 |
+
minimum=0.0, maximum=1.0, value=0.0, step=0.05,
|
395 |
+
label="Algorithm Temp (`alg_temp`, adds randomness to confidence-based `alg`)"
|
396 |
+
)
|
397 |
+
|
398 |
with gr.Row():
|
399 |
visualization_delay = gr.Slider(
|
400 |
+
minimum=0.0, maximum=0.5, value=0.02, step=0.01,
|
401 |
label="Visualization Delay (seconds)"
|
402 |
)
|
403 |
|
404 |
# Clear button
|
405 |
clear_btn = gr.Button("Clear Conversation")
|
406 |
|
407 |
+
# --- Event Handlers ---
|
408 |
+
def add_message(history, message, response):
|
409 |
+
"""Add a message pair to the history and return the updated history"""
|
410 |
+
# Ensure history is a list
|
411 |
+
if not isinstance(history, list):
|
412 |
+
history = []
|
413 |
history.append([message, response])
|
414 |
return history
|
415 |
|
416 |
+
def user_message_submitted(message, history):
|
417 |
+
"""Process a submitted user message"""
|
418 |
+
if not message.strip():
|
419 |
+
return history, history, "", [] # No change if empty
|
420 |
+
|
421 |
+
# Add user message (response is None for now)
|
422 |
+
history = add_message(history, message, None)
|
423 |
+
|
424 |
+
# Return updated history for display, clear input box
|
425 |
+
return history, history, "", [] # history, chatbot_ui, user_input, output_vis
|
426 |
+
|
427 |
+
|
428 |
+
def bot_response_stream(
|
429 |
+
history, # Current chat history (list of lists)
|
430 |
+
gen_length, steps, constraints, # Generation settings
|
431 |
+
temperature, top_p, top_k, alg, alg_temp, # Sampling settings
|
432 |
+
delay # Visualization delay
|
433 |
+
):
|
434 |
+
"""Generate bot response and stream visualization states"""
|
435 |
+
if not history or history[-1][1] is not None: # Check if history is present and last response isn't already set
|
436 |
+
print("Skipping bot response generation: No new user message.")
|
437 |
+
# Yield empty state if needed to prevent errors downstream
|
438 |
+
# Ensure history is returned correctly if nothing happens
|
439 |
+
yield history, [], "Internal Error: No user message found."
|
440 |
+
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
441 |
|
442 |
+
# Format messages for the model
|
443 |
+
# Exclude the last entry as it only contains the user message
|
444 |
+
messages_for_model = format_chat_history(history) # Already includes system prompt
|
445 |
+
|
446 |
+
print("\n--- Generating Bot Response ---")
|
447 |
+
print(f"History: {history}")
|
448 |
+
print(f"Messages for model: {messages_for_model}")
|
449 |
+
print(f"Constraints text: '{constraints}'")
|
450 |
+
print(f"Gen length: {gen_length}, Steps: {steps}, Temp: {temperature}, Top-P: {top_p}, Top-K: {top_k}, Alg: {alg}, Alg Temp: {alg_temp}")
|
451 |
+
|
452 |
+
# Call the generation function
|
453 |
+
vis_states, response_text = generate_response_with_visualization(
|
454 |
+
messages_for_model,
|
455 |
+
gen_length=gen_length,
|
456 |
+
steps=steps,
|
457 |
+
constraints_text=constraints,
|
458 |
+
temperature=temperature,
|
459 |
+
top_p=top_p if top_p < 1.0 else None, # None disables top-p
|
460 |
+
top_k=top_k if top_k > 0 else None, # None disables top-k
|
461 |
+
alg=alg,
|
462 |
+
alg_temp=alg_temp,
|
463 |
+
visualization_delay=delay,
|
464 |
+
# Pass other necessary args like tokenizer, model if not global
|
465 |
+
)
|
466 |
|
467 |
+
print(f"Generated response text: '{response_text}'")
|
468 |
+
print(f"Number of visualization states: {len(vis_states)}")
|
469 |
|
|
|
|
|
470 |
|
471 |
+
# Update the history with the final response
|
472 |
+
# Make sure history is mutable if needed or reassign
|
473 |
+
if history:
|
474 |
+
history[-1][1] = response_text
|
475 |
+
else:
|
476 |
+
print("Warning: History was empty when trying to update response.")
|
477 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
478 |
|
479 |
+
# Stream the visualization states
|
480 |
+
if not vis_states:
|
481 |
+
print("Warning: No visualization states were generated.")
|
482 |
+
# Yield something to prevent downstream errors
|
483 |
+
yield history, [("Error: No visualization.", "red")], response_text
|
484 |
+
return
|
485 |
|
486 |
+
# Yield initial state immediately if desired, or just start loop
|
487 |
+
# yield history, vis_states[0], response_text
|
|
|
488 |
|
489 |
+
for state in vis_states:
|
490 |
+
yield history, state, response_text # Yield updated history, current vis state, final text
|
491 |
+
time.sleep(delay) # Pause between steps
|
492 |
|
493 |
+
# Final yield to ensure the last state is displayed
|
494 |
+
yield history, vis_states[-1], response_text
|
495 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
496 |
|
497 |
+
def clear_conversation():
|
498 |
+
"""Clear the conversation history and visualization"""
|
499 |
+
return [], [], "", [] # history, chatbot, user_input, output_vis
|
500 |
+
|
501 |
+
# --- Event Wiring ---
|
502 |
+
|
503 |
+
# Clear button
|
504 |
clear_btn.click(
|
505 |
+
fn=clear_conversation,
|
506 |
inputs=[],
|
507 |
+
outputs=[chat_history, chatbot_ui, user_input, output_vis]
|
508 |
+
)
|
509 |
+
|
510 |
+
# User message submission flow (2-step using .then)
|
511 |
+
# 1. User submits message -> Update history and chatbot UI immediately
|
512 |
+
submit_action = user_input.submit(
|
513 |
+
fn=user_message_submitted,
|
514 |
+
inputs=[user_input, chat_history],
|
515 |
+
outputs=[chat_history, chatbot_ui, user_input, output_vis] # Update chatbot, clear input
|
516 |
+
)
|
517 |
+
|
518 |
+
# Connect send button to the same function
|
519 |
+
send_action = send_btn.click(
|
520 |
+
fn=user_message_submitted,
|
521 |
+
inputs=[user_input, chat_history],
|
522 |
+
outputs=[chat_history, chatbot_ui, user_input, output_vis]
|
523 |
+
)
|
524 |
+
|
525 |
+
# 2. After UI update -> Trigger bot response generation and streaming
|
526 |
+
# Use the updated chat_history from the first step
|
527 |
+
submit_action.then(
|
528 |
+
fn=bot_response_stream,
|
529 |
+
inputs=[
|
530 |
+
chat_history, gen_length, steps, constraints_input,
|
531 |
+
temperature, top_p, top_k, alg, alg_temp,
|
532 |
+
visualization_delay
|
533 |
+
],
|
534 |
+
outputs=[chatbot_ui, output_vis, user_input] # Update chatbot, visualization. Keep user_input as output to potentially display final text/error? (Check Gradio docs for Textbox output binding on yield) Let's remove user_input from outputs here.
|
535 |
)
|
536 |
|
537 |
+
send_action.then(
|
538 |
+
fn=bot_response_stream,
|
539 |
+
inputs=[
|
540 |
+
chat_history, gen_length, steps, constraints_input,
|
541 |
+
temperature, top_p, top_k, alg, alg_temp,
|
542 |
+
visualization_delay
|
543 |
+
],
|
544 |
+
outputs=[chatbot_ui, output_vis] # Update chatbot and visualization
|
545 |
+
)
|
546 |
+
|
547 |
+
# Clear input after send/submit (already handled in user_message_submitted)
|
548 |
+
# submit_action.then(lambda: "", outputs=user_input)
|
549 |
+
# send_action.then(lambda: "", outputs=user_input)
|
550 |
+
|
551 |
+
|
552 |
return demo
|
553 |
|
554 |
+
# --- Launch the Gradio App ---
|
555 |
if __name__ == "__main__":
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
556 |
demo = create_chatbot_demo()
|
557 |
+
# Using queue for streaming and handling multiple users
|
558 |
+
demo.queue().launch(debug=True, share=True)
|
|