Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -11,8 +11,7 @@ from typing import List, Dict, Tuple, Optional
|
|
11 |
import torch.distributions as dists # Added import
|
12 |
|
13 |
# --- START: Copied Helper functions from generation_utils.py ---
|
14 |
-
#
|
15 |
-
|
16 |
def top_p_logits(logits, top_p=None):
|
17 |
""" Applies top-p filtering to logits. """
|
18 |
if top_p is None or top_p >= 1.0:
|
@@ -42,39 +41,38 @@ def top_k_logits(logits, top_k=None):
|
|
42 |
def sample_tokens(logits, temperature=0.0, top_p=None, top_k=None, margin_confidence=False, neg_entropy=False):
|
43 |
""" Samples tokens based on logits and calculates confidence. """
|
44 |
if temperature > 0:
|
45 |
-
|
|
|
|
|
46 |
if top_p is not None and top_p < 1.0: # Apply top_p if valid
|
47 |
logits = top_p_logits(logits, top_p)
|
48 |
if top_k is not None and top_k > 0: # Apply top_k if valid
|
49 |
logits = top_k_logits(logits, top_k)
|
50 |
|
51 |
# Ensure logits are not all -inf after filtering, if so, sample uniformly? Or handle error.
|
52 |
-
#
|
53 |
-
|
54 |
-
|
55 |
-
|
|
|
|
|
56 |
|
57 |
probs = torch.softmax(logits, dim=-1)
|
58 |
|
|
|
|
|
|
|
|
|
|
|
|
|
59 |
if temperature > 0:
|
60 |
try:
|
61 |
-
# Check for NaNs or Infs in probs before sampling
|
62 |
-
if torch.isnan(probs).any() or torch.isinf(probs).any():
|
63 |
-
print("Warning: NaN or Inf detected in probabilities before sampling. Attempting to recover.")
|
64 |
-
# Simple recovery: Sample from uniform distribution or highest prob token
|
65 |
-
probs = torch.nan_to_num(probs, nan=0.0, posinf=0.0, neginf=0.0)
|
66 |
-
if probs.sum() == 0: # If all probabilities became zero
|
67 |
-
print("Warning: All probabilities became zero. Sampling uniformly.")
|
68 |
-
probs = torch.ones_like(probs) / probs.shape[-1]
|
69 |
-
else:
|
70 |
-
probs = probs / probs.sum(dim=-1, keepdim=True) # Re-normalize
|
71 |
-
|
72 |
x0 = dists.Categorical(probs=probs).sample()
|
73 |
confidence = torch.gather(probs, -1, x0.unsqueeze(-1)).squeeze(-1)
|
74 |
except Exception as e: # Catch broader exceptions during sampling
|
75 |
print(f"Warning: Error during Categorical sampling: {e}. Falling back to argmax.")
|
76 |
confidence, x0 = probs.max(dim=-1)
|
77 |
-
else:
|
78 |
confidence, x0 = probs.max(dim=-1)
|
79 |
|
80 |
if margin_confidence:
|
@@ -86,14 +84,18 @@ def sample_tokens(logits, temperature=0.0, top_p=None, top_k=None, margin_confid
|
|
86 |
|
87 |
if neg_entropy:
|
88 |
epsilon = 1e-10
|
|
|
89 |
log_probs = torch.log(probs + epsilon)
|
90 |
confidence = torch.sum(probs * log_probs, dim=-1) # Should be negative entropy
|
91 |
|
92 |
-
|
|
|
93 |
|
|
|
94 |
# --- END: Copied Helper functions ---
|
95 |
|
96 |
|
|
|
97 |
# Load model configuration to get special token IDs
|
98 |
config = AutoConfig.from_pretrained("Dream-org/Dream-v0-Instruct-7B", trust_remote_code=True)
|
99 |
# Use AutoModel for the base model loading, relying on trust_remote_code=True
|
@@ -113,7 +115,7 @@ model = AutoModel.from_pretrained(
|
|
113 |
model_path,
|
114 |
torch_dtype=torch.bfloat16 if device == 'cuda' else torch.float32, # Use bfloat16 only on CUDA
|
115 |
trust_remote_code=True,
|
116 |
-
|
117 |
)
|
118 |
model = model.to(device).eval()
|
119 |
print("Model loaded.")
|
@@ -123,27 +125,17 @@ MASK_TOKEN = tokenizer.mask_token
|
|
123 |
MASK_ID = tokenizer.mask_token_id # Use tokenizer's mask_token_id directly
|
124 |
PAD_ID = tokenizer.pad_token_id # Use tokenizer's pad_token_id
|
125 |
EOS_ID = tokenizer.eos_token_id # Use tokenizer's eos_token_id
|
126 |
-
# Use attributes from loaded config/tokenizer objects
|
127 |
-
# MASK_ID = config.mask_token_id # Can use this too, should be consistent
|
128 |
-
# PAD_ID = config.pad_token_id
|
129 |
-
# EOS_ID = config.eos_token_id
|
130 |
|
131 |
-
# Ensure mask_token_id is correctly identified
|
132 |
if MASK_ID is None:
|
133 |
print("Warning: Mask token ID not found in config/tokenizer. Trying to fetch from tokenizer...")
|
134 |
-
# Try getting from tokenizer directly if config doesn't have it or it's None
|
135 |
mask_token_special = tokenizer.mask_token
|
136 |
if mask_token_special:
|
137 |
MASK_ID = tokenizer.convert_tokens_to_ids(mask_token_special)
|
138 |
print(f"Found MASK_ID from tokenizer: {MASK_ID}")
|
139 |
else:
|
140 |
-
# Fallback or raise error if still not found
|
141 |
raise ValueError("Cannot determine MASK_ID. Check model's tokenizer configuration.")
|
142 |
|
143 |
-
# Make sure EOS_ID and PAD_ID are handled correctly; Dream uses the same ID for both
|
144 |
SPECIAL_TOKEN_IDS = {PAD_ID, EOS_ID, MASK_ID}
|
145 |
-
# Add other special tokens defined in tokenizer_config.json if needed for hiding
|
146 |
-
# Get IDs for im_start, im_end etc. if they should also be hidden/handled specially
|
147 |
try:
|
148 |
IM_START_ID = tokenizer.convert_tokens_to_ids("<|im_start|>")
|
149 |
IM_END_ID = tokenizer.convert_tokens_to_ids("<|im_end|>")
|
@@ -156,7 +148,6 @@ except KeyError:
|
|
156 |
|
157 |
|
158 |
# --- Helper Functions ---
|
159 |
-
|
160 |
def parse_constraints(constraints_text: str) -> Dict[int, List[int]]:
|
161 |
"""
|
162 |
Parse constraints in format: 'position:word, position:word, ...'
|
@@ -174,43 +165,17 @@ def parse_constraints(constraints_text: str) -> Dict[int, List[int]]:
|
|
174 |
continue
|
175 |
pos_str, word = part.split(':', 1)
|
176 |
try:
|
177 |
-
# Position relative to the start of the *generation*
|
178 |
pos = int(pos_str.strip())
|
179 |
word = word.strip() # Strip whitespace from word
|
180 |
-
|
181 |
-
|
182 |
-
|
183 |
-
|
184 |
-
|
185 |
-
# Assume we want the word as it would appear mid-sentence unless pos is 0.
|
186 |
-
token_ids = tokenizer.encode(word, add_special_tokens=False)
|
187 |
-
# Add space prefix if needed based on position? This is tricky.
|
188 |
-
# Let's assume the user provides the word how they want it tokenized,
|
189 |
-
# potentially including a leading space if necessary.
|
190 |
-
# Example: " 5: word" might be tokenized differently than "5:word".
|
191 |
-
# Simplest approach: Tokenize exactly what the user provided.
|
192 |
-
# Let's refine: add space prefix automatically if pos > 0, unless word already starts with space?
|
193 |
-
# This seems more robust for typical usage.
|
194 |
-
if pos > 0 and not word.startswith(" "):
|
195 |
-
token_ids_with_space = tokenizer.encode(" " + word, add_special_tokens=False)
|
196 |
-
# Check if adding space actually changes tokenization significantly
|
197 |
-
# Heuristic: if the first token ID changes, use the space-prefixed version.
|
198 |
-
first_token_no_space = tokenizer.encode(word, add_special_tokens=False)[0] if token_ids else None
|
199 |
-
first_token_with_space = tokenizer.encode(" " + word, add_special_tokens=False)[0] if token_ids_with_space else None
|
200 |
-
|
201 |
-
if first_token_no_space != first_token_with_space:
|
202 |
-
token_ids = token_ids_with_space
|
203 |
-
# If tokenization doesn't change much, maybe stick to original? Less surprising.
|
204 |
-
# Let's stick to adding the space if pos > 0 for consistency, like original code.
|
205 |
-
token_ids = tokenizer.encode(" " + word, add_special_tokens=False)
|
206 |
-
|
207 |
-
elif pos == 0:
|
208 |
-
token_ids = tokenizer.encode(word, add_special_tokens=False)
|
209 |
-
|
210 |
|
211 |
if token_ids and pos >= 0:
|
212 |
constraints[pos] = token_ids
|
213 |
-
elif not token_ids:
|
214 |
print(f"Warning: Could not tokenize constraint word '{word}'")
|
215 |
except ValueError:
|
216 |
print(f"Warning: Invalid position '{pos_str}' in constraint part '{part}'")
|
@@ -219,26 +184,16 @@ def parse_constraints(constraints_text: str) -> Dict[int, List[int]]:
|
|
219 |
print(f"Warning: Error processing constraint '{part}': {e}")
|
220 |
continue
|
221 |
|
222 |
-
print(f"Parsed constraints: {constraints}") # Debugging
|
223 |
return constraints
|
224 |
|
225 |
|
226 |
def format_chat_history(history: List[List[Optional[str]]]) -> List[Dict[str, str]]:
|
227 |
-
"""
|
228 |
-
Format chat history for the Dream model's chat template.
|
229 |
-
|
230 |
-
Args:
|
231 |
-
history: List of [user_message, assistant_message] pairs.
|
232 |
-
The last assistant_message might be None.
|
233 |
-
|
234 |
-
Returns:
|
235 |
-
Formatted list of message dictionaries for tokenizer.apply_chat_template.
|
236 |
-
"""
|
237 |
messages = []
|
238 |
for user_msg, assistant_msg in history:
|
239 |
-
if user_msg:
|
240 |
messages.append({"role": "user", "content": user_msg})
|
241 |
-
# Add assistant message only if it exists (it won't for the last turn before generation)
|
242 |
if assistant_msg:
|
243 |
messages.append({"role": "assistant", "content": assistant_msg})
|
244 |
return messages
|
@@ -250,19 +205,17 @@ def apply_constraints_to_state(
|
|
250 |
parsed_constraints: Dict[int, List[int]],
|
251 |
current_step: Optional[int] = None # For logging/debugging
|
252 |
) -> torch.Tensor:
|
253 |
-
"""Applies constraints directly to the state tensor `x`."""
|
254 |
-
modified_x = x
|
|
|
255 |
for rel_pos, word_token_ids in parsed_constraints.items():
|
256 |
abs_start_pos = prompt_length + rel_pos
|
257 |
abs_end_pos = abs_start_pos + len(word_token_ids)
|
258 |
|
259 |
-
# Ensure the constraint fits within the generation length
|
260 |
if abs_start_pos < total_length and abs_end_pos <= total_length:
|
261 |
try:
|
262 |
constraint_tensor = torch.tensor(word_token_ids, dtype=torch.long, device=modified_x.device)
|
263 |
-
# Force the constraint tokens onto the sequence
|
264 |
modified_x[0, abs_start_pos:abs_end_pos] = constraint_tensor
|
265 |
-
# print(f"Debug (Step {current_step}): Applied constraint {tokenizer.decode(word_token_ids)} at pos {rel_pos}") # Debug
|
266 |
except IndexError:
|
267 |
print(f"Warning (Step {current_step}): Constraint at {rel_pos} ('{tokenizer.decode(word_token_ids)}') goes out of bounds.")
|
268 |
except Exception as e:
|
@@ -286,27 +239,7 @@ def generate_dream_response(
|
|
286 |
alg_temp: Optional[float],
|
287 |
visualization_delay: float
|
288 |
) -> List[Tuple[str, str]]:
|
289 |
-
"""
|
290 |
-
Generates text using the Dream model step-by-step and yields visualization states live.
|
291 |
-
|
292 |
-
Args:
|
293 |
-
history: Chat history.
|
294 |
-
gen_length: Max new tokens to generate.
|
295 |
-
steps: Number of diffusion steps.
|
296 |
-
constraints_text: User-provided constraints string.
|
297 |
-
temperature: Sampling temperature.
|
298 |
-
top_p: Top-p sampling nucleus. Clamp to < 1.0 or None.
|
299 |
-
top_k: Top-k sampling. Clamp to > 0 or None.
|
300 |
-
alg: Remasking algorithm ('origin', 'maskgit_plus', 'topk_margin', 'entropy').
|
301 |
-
alg_temp: Temperature for confidence-based algorithms.
|
302 |
-
visualization_delay: Delay between visualization steps.
|
303 |
-
|
304 |
-
Yields:
|
305 |
-
Tuple[List[List[Optional[str]]], List[Tuple[str, Optional[str]]], str]:
|
306 |
-
- Updated history (may be intermediate until final response)
|
307 |
-
- Visualization data for HighlightedText for the current step
|
308 |
-
- Intermediate or Final response text (yielded repeatedly)
|
309 |
-
"""
|
310 |
|
311 |
if not history or not history[-1][0]:
|
312 |
yield history, [("No input message found.", "red")], ""
|
@@ -315,74 +248,62 @@ def generate_dream_response(
|
|
315 |
# --- 1. Preparation ---
|
316 |
last_user_message = history[-1][0]
|
317 |
messages_for_template = format_chat_history(history) # Includes the latest user message
|
|
|
318 |
|
319 |
-
# Parse constraints relative to the *generated* sequence
|
320 |
-
parsed_constraints = parse_constraints(constraints_text) # Dict[rel_pos, List[token_id]]
|
321 |
-
|
322 |
-
# Prepare inputs using the chat template
|
323 |
try:
|
324 |
inputs = tokenizer.apply_chat_template(
|
325 |
messages_for_template,
|
326 |
return_tensors="pt",
|
327 |
return_dict=True,
|
328 |
-
add_generation_prompt=True
|
329 |
)
|
330 |
input_ids = inputs.input_ids.to(device)
|
331 |
-
|
|
|
332 |
prompt_length = input_ids.shape[1]
|
333 |
except Exception as e:
|
334 |
print(f"Error applying chat template: {e}")
|
335 |
yield history, [("Error preparing input.", "red")], ""
|
336 |
return
|
337 |
|
338 |
-
|
339 |
-
|
340 |
-
# Ensure top_p and top_k have valid values for filtering functions
|
341 |
-
top_p_val = top_p if top_p is not None and top_p < 1.0 else None
|
342 |
top_k_val = top_k if top_k is not None and top_k > 0 else None
|
343 |
-
alg_temp_val = alg_temp if alg in ['maskgit_plus', 'topk_margin', 'entropy'] else None
|
344 |
|
345 |
# --- 2. Initialize Generation State ---
|
346 |
total_length = prompt_length + gen_length
|
347 |
-
|
348 |
-
# Initial state: prompt + MASK tokens
|
349 |
initial_generation_part = torch.full((1, gen_length), MASK_ID, dtype=torch.long, device=device)
|
350 |
x = torch.cat((input_ids, initial_generation_part), dim=1)
|
351 |
|
352 |
-
# Prepare
|
353 |
generation_attention_mask = torch.ones((1, gen_length), dtype=torch.long, device=device)
|
354 |
-
|
355 |
-
|
356 |
-
#
|
357 |
-
#
|
358 |
-
|
359 |
-
#
|
360 |
-
#
|
361 |
-
|
362 |
-
|
363 |
-
#
|
364 |
-
|
365 |
-
|
366 |
-
|
367 |
-
|
368 |
-
|
369 |
-
|
370 |
-
|
371 |
-
) # Shape [B, 1, N, N]
|
372 |
-
else:
|
373 |
-
tok_idx = None
|
374 |
-
attention_mask_for_model = None # Let the model handle full attention if mask is None
|
375 |
-
|
376 |
-
# Timesteps for diffusion
|
377 |
timesteps = torch.linspace(1, eps, steps + 1, device=device)
|
378 |
|
379 |
-
# Apply initial constraints
|
380 |
-
x = apply_constraints_to_state(x, prompt_length, total_length, parsed_constraints, current_step=-1)
|
381 |
|
382 |
# --- 3. Visualization Setup ---
|
383 |
-
previous_tokens_vis = None
|
384 |
-
final_response_text = ""
|
385 |
-
history_copy = [list(item) for item in history] #
|
386 |
|
387 |
# --- 4. Initial Yield (Masked State) ---
|
388 |
initial_generated_tokens = x[0, prompt_length:].cpu()
|
@@ -393,84 +314,53 @@ def generate_dream_response(
|
|
393 |
vis_data_initial.append((display_token, color))
|
394 |
|
395 |
previous_tokens_vis = initial_generated_tokens
|
396 |
-
yield history_copy, vis_data_initial, ""
|
397 |
time.sleep(visualization_delay)
|
398 |
|
399 |
# --- 5. Step-by-Step Diffusion Loop ---
|
400 |
try:
|
401 |
start_time = time.time()
|
402 |
for i in range(steps):
|
403 |
-
|
404 |
-
|
405 |
-
if not mask_index.any(): # Stop if no masks left
|
406 |
print(f"No mask tokens left at step {i}. Stopping early.")
|
407 |
break
|
408 |
|
409 |
-
#
|
410 |
-
#
|
411 |
-
|
412 |
-
# Call the model - ensure attention mask format is correct
|
413 |
-
# The model forward expects `attention_mask` usually of shape [B, N] or broadcastable.
|
414 |
-
# If we use `attention_mask_for_model = None`, it implies full attention.
|
415 |
-
# If we computed `attention_mask_for_model` as [B, 1, N, N], pass that.
|
416 |
-
# Let's try passing the [B, N] mask and let the model handle broadcasting/causality internally.
|
417 |
outputs = model(
|
418 |
input_ids=x,
|
419 |
-
attention_mask=
|
420 |
-
position_ids=None,
|
421 |
-
use_cache=False,
|
422 |
return_dict=True
|
423 |
)
|
424 |
logits = outputs.logits
|
|
|
425 |
|
426 |
-
|
427 |
-
|
428 |
-
# This seems to align logits with the *previous* token's prediction. Is this correct for diffusion?
|
429 |
-
# Let's assume the original code did this for a reason, perhaps related to how the model was trained or expects inputs.
|
430 |
-
# Update: Looking at standard LM forward pass, logits[t] predicts token[t+1].
|
431 |
-
# The shift aligns logits[t] with token[t]. Let's keep it.
|
432 |
-
logits = torch.cat([logits[:,:1], logits[:, :-1]], dim=1)
|
433 |
-
|
434 |
-
|
435 |
-
# Select logits for masked positions
|
436 |
-
# Ensure mask_index has the same batch dimension size as logits
|
437 |
-
# mask_index shape is [B, N], logits shape is [B, N, V]
|
438 |
-
# We need to select elements from the last dim of logits where mask is True
|
439 |
-
mask_logits = logits[mask_index] # This correctly selects [num_masked_tokens, V]
|
440 |
-
|
441 |
-
if mask_logits.numel() == 0: # If no masks, logits selection is empty
|
442 |
print(f"No masked tokens found for logit selection at step {i}. Stopping.")
|
443 |
break
|
444 |
|
445 |
-
# print(f"Step {i}: mask_logits shape: {mask_logits.shape}") # Debug
|
446 |
-
|
447 |
# --- Sampling / Remasking Logic ---
|
448 |
t = timesteps[i]
|
449 |
s = timesteps[i + 1]
|
450 |
-
|
451 |
x_new_masked_part = torch.full_like(x[mask_index], MASK_ID, device=device, dtype=torch.long)
|
452 |
|
453 |
if alg == 'origin':
|
454 |
-
|
455 |
-
p_transfer = (1.0 - s / t) if i < steps - 1 else 1.0 # Ensure float division
|
456 |
-
# Sample only for the tokens to be revealed in this step
|
457 |
num_masked = mask_logits.shape[0]
|
458 |
transfer_indices_relative = torch.rand(num_masked, device=device) < p_transfer
|
459 |
logits_to_sample = mask_logits[transfer_indices_relative]
|
460 |
|
461 |
if logits_to_sample.numel() > 0:
|
462 |
-
# print(f"Step {i} (origin): Sampling {logits_to_sample.shape[0]} tokens.") # Debug
|
463 |
_, sampled_tokens = sample_tokens(logits_to_sample, temperature=temperature, top_p=top_p_val, top_k=top_k_val)
|
464 |
-
# Place sampled tokens into the correct positions within the masked part
|
465 |
x_new_masked_part[transfer_indices_relative] = sampled_tokens
|
466 |
-
# else:
|
467 |
-
# print(f"Step {i} (origin): No tokens to sample (p_transfer={p_transfer}).") # Debug
|
468 |
|
469 |
-
else:
|
470 |
-
# Confidence-based algorithms (maskgit_plus, topk_margin, entropy)
|
471 |
use_margin = (alg == 'topk_margin')
|
472 |
use_entropy = (alg == 'entropy')
|
473 |
-
# print(f"Step {i} ({alg}): Sampling all {mask_logits.shape[0]} masked tokens for confidence.") # Debug
|
474 |
confidence, x0_candidates = sample_tokens(
|
475 |
mask_logits,
|
476 |
temperature=temperature,
|
@@ -479,67 +369,92 @@ def generate_dream_response(
|
|
479 |
margin_confidence=use_margin,
|
480 |
neg_entropy=use_entropy
|
481 |
)
|
482 |
-
# print(f"Step {i} ({alg}): Confidence range: [{confidence.min():.2f}, {confidence.max():.2f}]") # Debug
|
483 |
-
|
484 |
|
485 |
num_mask_token = mask_logits.shape[0]
|
486 |
-
# Calculate number to reveal based on time steps, ensure it's an int
|
487 |
target_num_revealed_float = num_mask_token * (1.0 - s / t)
|
488 |
number_transfer_tokens = int(target_num_revealed_float) if i < steps - 1 else num_mask_token
|
489 |
|
490 |
-
|
491 |
if number_transfer_tokens > 0:
|
492 |
-
|
493 |
-
if
|
494 |
-
|
495 |
-
|
496 |
-
|
497 |
-
|
498 |
-
|
499 |
-
|
500 |
-
|
501 |
-
|
502 |
-
|
503 |
-
#
|
504 |
-
|
505 |
-
|
506 |
-
|
507 |
-
|
508 |
-
|
509 |
-
|
510 |
-
|
511 |
-
|
512 |
-
|
513 |
-
|
514 |
-
|
515 |
-
|
516 |
-
|
517 |
-
|
518 |
-
|
519 |
-
|
520 |
-
|
521 |
-
|
522 |
-
|
523 |
-
|
524 |
-
|
525 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
526 |
|
527 |
# Update the global state `x` only at the masked positions
|
528 |
x[mask_index] = x_new_masked_part
|
529 |
|
530 |
# --- Apply Constraints ---
|
531 |
-
# Constraints should be applied *after* sampling/revealing tokens for the step
|
532 |
x = apply_constraints_to_state(x, prompt_length, total_length, parsed_constraints, current_step=i)
|
533 |
|
534 |
# --- Yield Visualization ---
|
535 |
-
current_generated_tokens = x[0, prompt_length:].cpu()
|
536 |
vis_data = []
|
|
|
537 |
for j in range(gen_length):
|
538 |
current_tok_id = current_generated_tokens[j].item()
|
539 |
-
previous_tok_id = previous_tokens_vis[j].item() if previous_tokens_vis is not None else MASK_ID
|
540 |
|
541 |
try:
|
542 |
-
|
|
|
543 |
display_token = MASK_TOKEN if current_tok_id == MASK_ID else decoded_token
|
544 |
except Exception:
|
545 |
display_token = f"[ID:{current_tok_id}]" # Fallback
|
@@ -554,31 +469,17 @@ def generate_dream_response(
|
|
554 |
else: # Token was already revealed
|
555 |
color = "#6699CC" # Light Blue
|
556 |
|
557 |
-
# Hide special tokens (PAD/EOS) if they were already revealed (LLaDA effect)
|
558 |
-
# Ensure PAD_ID and EOS_ID are not None before checking
|
559 |
should_hide = (PAD_ID is not None and current_tok_id == PAD_ID) or \
|
560 |
(EOS_ID is not None and current_tok_id == EOS_ID)
|
561 |
if should_hide and previous_tok_id == current_tok_id:
|
562 |
token_to_display = "" # Hide by making empty
|
563 |
color = None # No color for hidden
|
564 |
|
565 |
-
|
566 |
if token_to_display:
|
567 |
vis_data.append((token_to_display, color))
|
568 |
-
|
569 |
-
|
570 |
-
|
571 |
-
elif len(vis_data) > 0 and not isinstance(vis_data[-1], tuple) and vis_data[-1] == " ":
|
572 |
-
pass # Already added a space
|
573 |
-
elif len(vis_data) > 0 :
|
574 |
-
# Add a single space if hiding follows a visible token, improves readability slightly
|
575 |
-
# Let's simplify: just omit hidden tokens. Adding spaces might be complex.
|
576 |
-
pass
|
577 |
-
|
578 |
-
# Update previous state for the next iteration
|
579 |
-
previous_tokens_vis = current_generated_tokens
|
580 |
-
|
581 |
-
# Decode intermediate response (might be partial) - skip specials for readability
|
582 |
intermediate_response_tokens = x[0, prompt_length:]
|
583 |
intermediate_response_text = tokenizer.decode(
|
584 |
intermediate_response_tokens,
|
@@ -586,9 +487,6 @@ def generate_dream_response(
|
|
586 |
clean_up_tokenization_spaces=True
|
587 |
).strip()
|
588 |
|
589 |
-
# Yield current state
|
590 |
-
# We yield the *current* history, the vis data for this step, and intermediate text
|
591 |
-
# The final text will overwrite the intermediate text in the UI eventually
|
592 |
yield history_copy, vis_data, intermediate_response_text
|
593 |
time.sleep(visualization_delay)
|
594 |
|
@@ -598,65 +496,47 @@ def generate_dream_response(
|
|
598 |
# --- 6. Final Processing & Yield ---
|
599 |
final_sequence = x[0]
|
600 |
response_tokens = final_sequence[prompt_length:]
|
601 |
-
|
602 |
-
# Decode the final response text
|
603 |
final_response_text = tokenizer.decode(
|
604 |
response_tokens,
|
605 |
-
skip_special_tokens=True,
|
606 |
clean_up_tokenization_spaces=True
|
607 |
).strip()
|
608 |
-
|
609 |
-
# Update history with the final response *before* the last yield
|
610 |
history_copy[-1][1] = final_response_text
|
611 |
|
612 |
-
# Yield the final state (which might be the same as the last yielded state if loop finished)
|
613 |
-
# Need to format vis_data one last time based on the final `x`
|
614 |
final_generated_tokens = x[0, prompt_length:].cpu()
|
615 |
vis_data_final = []
|
|
|
616 |
for j in range(gen_length):
|
617 |
current_tok_id = final_generated_tokens[j].item()
|
618 |
-
previous_tok_id = previous_tokens_vis[j].item() if previous_tokens_vis is not None else MASK_ID
|
619 |
-
|
620 |
try:
|
621 |
-
decoded_token = tokenizer.decode([current_tok_id], skip_special_tokens=False)
|
622 |
display_token = MASK_TOKEN if current_tok_id == MASK_ID else decoded_token
|
623 |
except Exception:
|
624 |
display_token = f"[ID:{current_tok_id}]" # Fallback
|
625 |
-
|
626 |
color = None
|
627 |
token_to_display = display_token
|
628 |
-
|
629 |
-
|
630 |
-
|
631 |
-
elif previous_tok_id == MASK_ID:
|
632 |
-
color = "#66CC66"
|
633 |
-
else:
|
634 |
-
color = "#6699CC"
|
635 |
-
|
636 |
should_hide = (PAD_ID is not None and current_tok_id == PAD_ID) or \
|
637 |
(EOS_ID is not None and current_tok_id == EOS_ID)
|
638 |
if should_hide and previous_tok_id == current_tok_id:
|
639 |
-
token_to_display = ""
|
640 |
-
|
641 |
-
|
642 |
-
if token_to_display:
|
643 |
-
vis_data_final.append((token_to_display, color))
|
644 |
|
645 |
-
# Yield the final history, final visualization, and final text
|
646 |
yield history_copy, vis_data_final, final_response_text
|
647 |
print("Visualization streaming complete.")
|
648 |
|
649 |
-
|
650 |
except Exception as e:
|
651 |
print(f"Error during generation or processing: {e}")
|
652 |
import traceback
|
653 |
traceback.print_exc()
|
654 |
-
# Update history with error message? Or leave as None? Let's leave as None.
|
655 |
yield history_copy, [("Error during generation.", "red")], ""
|
656 |
return
|
657 |
|
658 |
|
659 |
-
# --- Gradio UI (
|
660 |
css = '''
|
661 |
.category-legend{display:none}
|
662 |
button{min-height: 60px}
|
@@ -669,12 +549,8 @@ def create_chatbot_demo():
|
|
669 |
"[[Blog](https://hkunlp.github.io/blog/2025/dream/)]" # Note: Link might be hypothetical
|
670 |
)
|
671 |
|
672 |
-
# STATE MANAGEMENT
|
673 |
-
# chat_history = gr.State([]) # Use gr.Chatbot's internal state implicitly if possible, or manage manually
|
674 |
-
# Let's manage manually with a list for clarity with yielding updates
|
675 |
_chat_history_store = gr.State([]) # Hidden state to store actual history list
|
676 |
|
677 |
-
# UI COMPONENTS
|
678 |
with gr.Row():
|
679 |
with gr.Column(scale=3):
|
680 |
chatbot_ui = gr.Chatbot(
|
@@ -682,10 +558,7 @@ def create_chatbot_demo():
|
|
682 |
height=500,
|
683 |
show_copy_button=True,
|
684 |
bubble_full_width=False,
|
685 |
-
# value=[] # Initialize chatbot UI empty
|
686 |
)
|
687 |
-
|
688 |
-
# Message input
|
689 |
with gr.Group():
|
690 |
with gr.Row():
|
691 |
user_input = gr.Textbox(
|
@@ -694,11 +567,9 @@ def create_chatbot_demo():
|
|
694 |
scale=7,
|
695 |
autofocus=True,
|
696 |
show_label=False,
|
697 |
-
container=False
|
698 |
)
|
699 |
send_btn = gr.Button("Send", scale=1, variant="primary")
|
700 |
-
|
701 |
-
|
702 |
constraints_input = gr.Textbox(
|
703 |
label="Word Constraints (Optional)",
|
704 |
info="Place words at specific positions (0-indexed from start of generation). Format: 'pos:word, pos:word,...'. Example: '0:Once, 5:upon, 10:time'",
|
@@ -708,125 +579,72 @@ def create_chatbot_demo():
|
|
708 |
with gr.Column(scale=2):
|
709 |
output_vis = gr.HighlightedText(
|
710 |
label="Denoising Process Visualization",
|
711 |
-
combine_adjacent=
|
712 |
-
show_legend=
|
713 |
-
interactive=False
|
714 |
)
|
715 |
-
# Add a text box to display the final/intermediate response clearly
|
716 |
response_text_display = gr.Textbox(
|
717 |
label="Generated Response",
|
718 |
interactive=False,
|
719 |
-
lines=5
|
720 |
)
|
721 |
|
722 |
-
|
723 |
-
# Advanced generation settings
|
724 |
with gr.Accordion("Generation Settings", open=False):
|
725 |
-
|
726 |
-
gen_length = gr.Slider(
|
727 |
-
|
728 |
-
|
729 |
-
)
|
730 |
-
|
731 |
-
|
732 |
-
|
733 |
-
)
|
734 |
-
|
735 |
-
|
736 |
-
|
737 |
-
|
738 |
-
)
|
739 |
-
alg_temp = gr.Slider(
|
740 |
-
minimum=0.0, maximum=1.0, value=0.1, step=0.05,
|
741 |
-
label="Remasking Temp (Confidence Algs)"
|
742 |
-
)
|
743 |
-
|
744 |
-
with gr.Row():
|
745 |
-
top_p = gr.Slider(
|
746 |
-
minimum=0.0, maximum=1.0, value=0.95, step=0.05,
|
747 |
-
label="Top-P (<=0 or >=1 disables)" # Clarify disabling condition
|
748 |
-
)
|
749 |
-
top_k = gr.Slider(
|
750 |
-
minimum=0, maximum=200, value=0, step=5,
|
751 |
-
label="Top-K (0 disables)"
|
752 |
-
)
|
753 |
-
|
754 |
-
with gr.Row():
|
755 |
-
remasking_strategy = gr.Radio(
|
756 |
-
choices=['origin', 'maskgit_plus', 'topk_margin', 'entropy'],
|
757 |
-
value='entropy', # Default to entropy as in example
|
758 |
-
label="Remasking Strategy (Algorithm)"
|
759 |
-
)
|
760 |
-
|
761 |
-
with gr.Row():
|
762 |
-
visualization_delay = gr.Slider(
|
763 |
-
minimum=0.0, maximum=0.5, value=0.03, step=0.01, # Slightly faster default
|
764 |
-
label="Visualization Delay (seconds)"
|
765 |
-
)
|
766 |
|
767 |
-
# Clear button
|
768 |
clear_btn = gr.Button("Clear Conversation")
|
769 |
|
770 |
-
# --- Event Handlers ---
|
771 |
-
|
772 |
def add_user_message_to_history(message: str, history_store: List[List[Optional[str]]]):
|
773 |
-
"""Adds user message, clears input, prepares for bot response."""
|
774 |
if not message.strip():
|
775 |
gr.Warning("Please enter a message.")
|
776 |
-
# Return unchanged history, empty vis, empty response text
|
777 |
return history_store, history_store, "", [], ""
|
778 |
-
|
779 |
-
# Add user message with placeholder for bot response
|
780 |
history_store.append([message, None])
|
781 |
-
# Return updated history store, history for chatbot UI, empty input, empty vis, empty response
|
782 |
return history_store, history_store, "", [], ""
|
783 |
|
784 |
def clear_conversation():
|
785 |
-
|
786 |
-
return [], [], "", [], "" # History store, chatbot UI, input, vis, response text
|
787 |
|
788 |
-
# --- Connect UI elements ---
|
789 |
-
|
790 |
-
# Define the inputs for the generation function once
|
791 |
generation_inputs = [
|
792 |
_chat_history_store, gen_length, steps, constraints_input,
|
793 |
temperature, top_p, top_k, remasking_strategy, alg_temp,
|
794 |
visualization_delay
|
795 |
]
|
796 |
-
# Define the outputs for the generation function
|
797 |
-
# Now yields: history_copy, vis_data, intermediate_response_text
|
798 |
-
# Map these to: chatbot_ui, output_vis, response_text_display
|
799 |
generation_outputs = [chatbot_ui, output_vis, response_text_display]
|
800 |
|
801 |
-
# Handle Textbox Submission (Enter key)
|
802 |
submit_listener = user_input.submit(
|
803 |
fn=add_user_message_to_history,
|
804 |
inputs=[user_input, _chat_history_store],
|
805 |
-
outputs=[_chat_history_store, chatbot_ui, user_input, output_vis, response_text_display]
|
806 |
-
)
|
807 |
-
# Chain the bot response generation after the user message is added
|
808 |
-
submit_listener.then(
|
809 |
fn=generate_dream_response,
|
810 |
inputs=generation_inputs,
|
811 |
-
outputs=generation_outputs,
|
812 |
-
show_progress="hidden"
|
813 |
)
|
814 |
|
815 |
-
# Handle Send Button Click
|
816 |
click_listener = send_btn.click(
|
817 |
fn=add_user_message_to_history,
|
818 |
inputs=[user_input, _chat_history_store],
|
819 |
-
outputs=[_chat_history_store, chatbot_ui, user_input, output_vis, response_text_display]
|
820 |
-
)
|
821 |
-
# Chain the bot response generation after the user message is added
|
822 |
-
click_listener.then(
|
823 |
fn=generate_dream_response,
|
824 |
inputs=generation_inputs,
|
825 |
-
outputs=generation_outputs,
|
826 |
show_progress="hidden"
|
827 |
)
|
828 |
|
829 |
-
# Clear Button Action
|
830 |
clear_btn.click(
|
831 |
clear_conversation,
|
832 |
inputs=[],
|
@@ -838,5 +656,4 @@ def create_chatbot_demo():
|
|
838 |
# --- Launch ---
|
839 |
if __name__ == "__main__":
|
840 |
demo = create_chatbot_demo()
|
841 |
-
|
842 |
-
demo.queue().launch(debug=True, share=False) # Set share=True for public link
|
|
|
11 |
import torch.distributions as dists # Added import
|
12 |
|
13 |
# --- START: Copied Helper functions from generation_utils.py ---
|
14 |
+
# [Keep the copied functions: top_p_logits, top_k_logits, sample_tokens]
|
|
|
15 |
def top_p_logits(logits, top_p=None):
|
16 |
""" Applies top-p filtering to logits. """
|
17 |
if top_p is None or top_p >= 1.0:
|
|
|
41 |
def sample_tokens(logits, temperature=0.0, top_p=None, top_k=None, margin_confidence=False, neg_entropy=False):
|
42 |
""" Samples tokens based on logits and calculates confidence. """
|
43 |
if temperature > 0:
|
44 |
+
# Prevent division by zero or negative temperatures
|
45 |
+
safe_temp = max(temperature, 1e-6)
|
46 |
+
logits = logits / safe_temp
|
47 |
if top_p is not None and top_p < 1.0: # Apply top_p if valid
|
48 |
logits = top_p_logits(logits, top_p)
|
49 |
if top_k is not None and top_k > 0: # Apply top_k if valid
|
50 |
logits = top_k_logits(logits, top_k)
|
51 |
|
52 |
# Ensure logits are not all -inf after filtering, if so, sample uniformly? Or handle error.
|
53 |
+
# Add a check here: if all logits are -inf, assign uniform probability.
|
54 |
+
is_all_neg_inf = torch.all(logits == torch.finfo(logits.dtype).min, dim=-1, keepdim=True)
|
55 |
+
if torch.any(is_all_neg_inf):
|
56 |
+
# print("Warning: All logits became -inf after filtering. Assigning uniform probabilities.")
|
57 |
+
uniform_logits = torch.zeros_like(logits)
|
58 |
+
logits = torch.where(is_all_neg_inf, uniform_logits, logits)
|
59 |
|
60 |
probs = torch.softmax(logits, dim=-1)
|
61 |
|
62 |
+
# Clamp probabilities to avoid NaNs in sampling, ensure they sum to 1
|
63 |
+
probs = torch.clamp(probs, min=0.0) # Ensure non-negative
|
64 |
+
probs = probs / probs.sum(dim=-1, keepdim=True) # Re-normalize
|
65 |
+
probs = torch.nan_to_num(probs, nan=0.0) # Handle any remaining NaNs
|
66 |
+
|
67 |
+
|
68 |
if temperature > 0:
|
69 |
try:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
70 |
x0 = dists.Categorical(probs=probs).sample()
|
71 |
confidence = torch.gather(probs, -1, x0.unsqueeze(-1)).squeeze(-1)
|
72 |
except Exception as e: # Catch broader exceptions during sampling
|
73 |
print(f"Warning: Error during Categorical sampling: {e}. Falling back to argmax.")
|
74 |
confidence, x0 = probs.max(dim=-1)
|
75 |
+
else: # Greedy decoding (temperature == 0)
|
76 |
confidence, x0 = probs.max(dim=-1)
|
77 |
|
78 |
if margin_confidence:
|
|
|
84 |
|
85 |
if neg_entropy:
|
86 |
epsilon = 1e-10
|
87 |
+
# Ensure probs are > 0 for log
|
88 |
log_probs = torch.log(probs + epsilon)
|
89 |
confidence = torch.sum(probs * log_probs, dim=-1) # Should be negative entropy
|
90 |
|
91 |
+
# Ensure confidence is not NaN
|
92 |
+
confidence = torch.nan_to_num(confidence, nan=0.0)
|
93 |
|
94 |
+
return confidence, x0
|
95 |
# --- END: Copied Helper functions ---
|
96 |
|
97 |
|
98 |
+
# [Keep model loading, constants, helper functions: parse_constraints, format_chat_history, apply_constraints_to_state]
|
99 |
# Load model configuration to get special token IDs
|
100 |
config = AutoConfig.from_pretrained("Dream-org/Dream-v0-Instruct-7B", trust_remote_code=True)
|
101 |
# Use AutoModel for the base model loading, relying on trust_remote_code=True
|
|
|
115 |
model_path,
|
116 |
torch_dtype=torch.bfloat16 if device == 'cuda' else torch.float32, # Use bfloat16 only on CUDA
|
117 |
trust_remote_code=True,
|
118 |
+
attn_implementation="sdpa" # Explicitly request SDPA if available/desired
|
119 |
)
|
120 |
model = model.to(device).eval()
|
121 |
print("Model loaded.")
|
|
|
125 |
MASK_ID = tokenizer.mask_token_id # Use tokenizer's mask_token_id directly
|
126 |
PAD_ID = tokenizer.pad_token_id # Use tokenizer's pad_token_id
|
127 |
EOS_ID = tokenizer.eos_token_id # Use tokenizer's eos_token_id
|
|
|
|
|
|
|
|
|
128 |
|
|
|
129 |
if MASK_ID is None:
|
130 |
print("Warning: Mask token ID not found in config/tokenizer. Trying to fetch from tokenizer...")
|
|
|
131 |
mask_token_special = tokenizer.mask_token
|
132 |
if mask_token_special:
|
133 |
MASK_ID = tokenizer.convert_tokens_to_ids(mask_token_special)
|
134 |
print(f"Found MASK_ID from tokenizer: {MASK_ID}")
|
135 |
else:
|
|
|
136 |
raise ValueError("Cannot determine MASK_ID. Check model's tokenizer configuration.")
|
137 |
|
|
|
138 |
SPECIAL_TOKEN_IDS = {PAD_ID, EOS_ID, MASK_ID}
|
|
|
|
|
139 |
try:
|
140 |
IM_START_ID = tokenizer.convert_tokens_to_ids("<|im_start|>")
|
141 |
IM_END_ID = tokenizer.convert_tokens_to_ids("<|im_end|>")
|
|
|
148 |
|
149 |
|
150 |
# --- Helper Functions ---
|
|
|
151 |
def parse_constraints(constraints_text: str) -> Dict[int, List[int]]:
|
152 |
"""
|
153 |
Parse constraints in format: 'position:word, position:word, ...'
|
|
|
165 |
continue
|
166 |
pos_str, word = part.split(':', 1)
|
167 |
try:
|
|
|
168 |
pos = int(pos_str.strip())
|
169 |
word = word.strip() # Strip whitespace from word
|
170 |
+
token_ids = []
|
171 |
+
if word: # Only encode if word is not empty
|
172 |
+
# Add space prefix automatically if pos > 0 and word doesn't start with space
|
173 |
+
text_to_encode = (" " + word) if (pos > 0 and not word.startswith(" ")) else word
|
174 |
+
token_ids = tokenizer.encode(text_to_encode, add_special_tokens=False)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
175 |
|
176 |
if token_ids and pos >= 0:
|
177 |
constraints[pos] = token_ids
|
178 |
+
elif not token_ids and word: # Don't warn for empty words after split
|
179 |
print(f"Warning: Could not tokenize constraint word '{word}'")
|
180 |
except ValueError:
|
181 |
print(f"Warning: Invalid position '{pos_str}' in constraint part '{part}'")
|
|
|
184 |
print(f"Warning: Error processing constraint '{part}': {e}")
|
185 |
continue
|
186 |
|
187 |
+
# print(f"Parsed constraints: {constraints}") # Debugging
|
188 |
return constraints
|
189 |
|
190 |
|
191 |
def format_chat_history(history: List[List[Optional[str]]]) -> List[Dict[str, str]]:
|
192 |
+
""" Formats chat history for the template. """
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
193 |
messages = []
|
194 |
for user_msg, assistant_msg in history:
|
195 |
+
if user_msg:
|
196 |
messages.append({"role": "user", "content": user_msg})
|
|
|
197 |
if assistant_msg:
|
198 |
messages.append({"role": "assistant", "content": assistant_msg})
|
199 |
return messages
|
|
|
205 |
parsed_constraints: Dict[int, List[int]],
|
206 |
current_step: Optional[int] = None # For logging/debugging
|
207 |
) -> torch.Tensor:
|
208 |
+
""" Applies constraints directly to the state tensor `x`. """
|
209 |
+
modified_x = x # Modify in place maybe okay? Let's stick with clone for safety.
|
210 |
+
modified_x = x.clone()
|
211 |
for rel_pos, word_token_ids in parsed_constraints.items():
|
212 |
abs_start_pos = prompt_length + rel_pos
|
213 |
abs_end_pos = abs_start_pos + len(word_token_ids)
|
214 |
|
|
|
215 |
if abs_start_pos < total_length and abs_end_pos <= total_length:
|
216 |
try:
|
217 |
constraint_tensor = torch.tensor(word_token_ids, dtype=torch.long, device=modified_x.device)
|
|
|
218 |
modified_x[0, abs_start_pos:abs_end_pos] = constraint_tensor
|
|
|
219 |
except IndexError:
|
220 |
print(f"Warning (Step {current_step}): Constraint at {rel_pos} ('{tokenizer.decode(word_token_ids)}') goes out of bounds.")
|
221 |
except Exception as e:
|
|
|
239 |
alg_temp: Optional[float],
|
240 |
visualization_delay: float
|
241 |
) -> List[Tuple[str, str]]:
|
242 |
+
""" Generates text step-by-step and yields visualization states live. """
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
243 |
|
244 |
if not history or not history[-1][0]:
|
245 |
yield history, [("No input message found.", "red")], ""
|
|
|
248 |
# --- 1. Preparation ---
|
249 |
last_user_message = history[-1][0]
|
250 |
messages_for_template = format_chat_history(history) # Includes the latest user message
|
251 |
+
parsed_constraints = parse_constraints(constraints_text)
|
252 |
|
|
|
|
|
|
|
|
|
253 |
try:
|
254 |
inputs = tokenizer.apply_chat_template(
|
255 |
messages_for_template,
|
256 |
return_tensors="pt",
|
257 |
return_dict=True,
|
258 |
+
add_generation_prompt=True
|
259 |
)
|
260 |
input_ids = inputs.input_ids.to(device)
|
261 |
+
# Ensure prompt_attention_mask is also on the correct device
|
262 |
+
prompt_attention_mask = inputs.attention_mask.to(device) if 'attention_mask' in inputs else torch.ones_like(input_ids)
|
263 |
prompt_length = input_ids.shape[1]
|
264 |
except Exception as e:
|
265 |
print(f"Error applying chat template: {e}")
|
266 |
yield history, [("Error preparing input.", "red")], ""
|
267 |
return
|
268 |
|
269 |
+
eps = 1e-3
|
270 |
+
top_p_val = top_p if top_p is not None and 0.0 < top_p < 1.0 else None # Make sure top_p is > 0
|
|
|
|
|
271 |
top_k_val = top_k if top_k is not None and top_k > 0 else None
|
272 |
+
alg_temp_val = alg_temp if alg in ['maskgit_plus', 'topk_margin', 'entropy'] and alg_temp is not None and alg_temp > 0 else None # Ensure > 0
|
273 |
|
274 |
# --- 2. Initialize Generation State ---
|
275 |
total_length = prompt_length + gen_length
|
|
|
|
|
276 |
initial_generation_part = torch.full((1, gen_length), MASK_ID, dtype=torch.long, device=device)
|
277 |
x = torch.cat((input_ids, initial_generation_part), dim=1)
|
278 |
|
279 |
+
# --- Prepare Attention Mask for SDPA ---
|
280 |
generation_attention_mask = torch.ones((1, gen_length), dtype=torch.long, device=device)
|
281 |
+
full_attention_mask_long = torch.cat((prompt_attention_mask, generation_attention_mask), dim=1) # Shape [B, N], dtype torch.long
|
282 |
+
|
283 |
+
# Convert attention mask for SDPA: Needs float matching query dtype.
|
284 |
+
# Where mask is 1 (attend), value should be 0.0. Where mask is 0 (don't attend), value should be -inf.
|
285 |
+
attention_mask_for_model = full_attention_mask_long.to(model.dtype) # Convert to model's dtype (e.g., bfloat16)
|
286 |
+
# Invert the mask logic: (1.0 - mask) gives 0s for attend, 1s for mask
|
287 |
+
# Multiply by large negative number (min value for dtype) for masked positions
|
288 |
+
large_neg_val = torch.finfo(model.dtype).min
|
289 |
+
attention_mask_for_model = (1.0 - attention_mask_for_model) * large_neg_val
|
290 |
+
# Ensure the shape is broadcastable, SDPA usually handles [B, N] -> [B, H, N, N] if needed.
|
291 |
+
# However, explicitly making it [B, 1, 1, N] or [B, 1, N, N] can be safer.
|
292 |
+
# Let's try passing [B, N] first, if it fails, reshape.
|
293 |
+
# Reshape to [B, 1, 1, N] which is commonly expected for additive masks by HF models
|
294 |
+
attention_mask_for_model = attention_mask_for_model.unsqueeze(1).unsqueeze(2)
|
295 |
+
# Now shape is [B, 1, 1, N]
|
296 |
+
|
297 |
+
# --- Timesteps ---
|
|
|
|
|
|
|
|
|
|
|
|
|
298 |
timesteps = torch.linspace(1, eps, steps + 1, device=device)
|
299 |
|
300 |
+
# Apply initial constraints
|
301 |
+
x = apply_constraints_to_state(x, prompt_length, total_length, parsed_constraints, current_step=-1)
|
302 |
|
303 |
# --- 3. Visualization Setup ---
|
304 |
+
previous_tokens_vis = None
|
305 |
+
final_response_text = ""
|
306 |
+
history_copy = [list(item) for item in history] # Mutable copy
|
307 |
|
308 |
# --- 4. Initial Yield (Masked State) ---
|
309 |
initial_generated_tokens = x[0, prompt_length:].cpu()
|
|
|
314 |
vis_data_initial.append((display_token, color))
|
315 |
|
316 |
previous_tokens_vis = initial_generated_tokens
|
317 |
+
yield history_copy, vis_data_initial, ""
|
318 |
time.sleep(visualization_delay)
|
319 |
|
320 |
# --- 5. Step-by-Step Diffusion Loop ---
|
321 |
try:
|
322 |
start_time = time.time()
|
323 |
for i in range(steps):
|
324 |
+
mask_index = (x == MASK_ID)
|
325 |
+
if not mask_index.any():
|
|
|
326 |
print(f"No mask tokens left at step {i}. Stopping early.")
|
327 |
break
|
328 |
|
329 |
+
# --- Model Forward Pass ---
|
330 |
+
# Pass the correctly formatted float mask
|
|
|
|
|
|
|
|
|
|
|
|
|
331 |
outputs = model(
|
332 |
input_ids=x,
|
333 |
+
attention_mask=attention_mask_for_model, # Pass the [B, 1, 1, N] float mask
|
334 |
+
position_ids=None,
|
335 |
+
use_cache=False,
|
336 |
return_dict=True
|
337 |
)
|
338 |
logits = outputs.logits
|
339 |
+
logits = torch.cat([logits[:,:1], logits[:, :-1]], dim=1) # Align logits
|
340 |
|
341 |
+
mask_logits = logits[mask_index]
|
342 |
+
if mask_logits.numel() == 0:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
343 |
print(f"No masked tokens found for logit selection at step {i}. Stopping.")
|
344 |
break
|
345 |
|
|
|
|
|
346 |
# --- Sampling / Remasking Logic ---
|
347 |
t = timesteps[i]
|
348 |
s = timesteps[i + 1]
|
|
|
349 |
x_new_masked_part = torch.full_like(x[mask_index], MASK_ID, device=device, dtype=torch.long)
|
350 |
|
351 |
if alg == 'origin':
|
352 |
+
p_transfer = (1.0 - s / t) if i < steps - 1 else 1.0
|
|
|
|
|
353 |
num_masked = mask_logits.shape[0]
|
354 |
transfer_indices_relative = torch.rand(num_masked, device=device) < p_transfer
|
355 |
logits_to_sample = mask_logits[transfer_indices_relative]
|
356 |
|
357 |
if logits_to_sample.numel() > 0:
|
|
|
358 |
_, sampled_tokens = sample_tokens(logits_to_sample, temperature=temperature, top_p=top_p_val, top_k=top_k_val)
|
|
|
359 |
x_new_masked_part[transfer_indices_relative] = sampled_tokens
|
|
|
|
|
360 |
|
361 |
+
else: # Confidence-based algorithms
|
|
|
362 |
use_margin = (alg == 'topk_margin')
|
363 |
use_entropy = (alg == 'entropy')
|
|
|
364 |
confidence, x0_candidates = sample_tokens(
|
365 |
mask_logits,
|
366 |
temperature=temperature,
|
|
|
369 |
margin_confidence=use_margin,
|
370 |
neg_entropy=use_entropy
|
371 |
)
|
|
|
|
|
372 |
|
373 |
num_mask_token = mask_logits.shape[0]
|
|
|
374 |
target_num_revealed_float = num_mask_token * (1.0 - s / t)
|
375 |
number_transfer_tokens = int(target_num_revealed_float) if i < steps - 1 else num_mask_token
|
376 |
|
|
|
377 |
if number_transfer_tokens > 0:
|
378 |
+
num_samples = min(number_transfer_tokens, num_mask_token) # Ensure k <= num_mask_token
|
379 |
+
if num_samples > 0: # Proceed only if we need to sample > 0 tokens
|
380 |
+
if alg_temp_val is None or alg_temp_val <= 0: # Top-k confidence
|
381 |
+
sort_metric = confidence if alg != 'entropy' else -confidence # Lower entropy = higher confidence
|
382 |
+
# Ensure k is not greater than the number of elements
|
383 |
+
k_topk = min(num_samples, sort_metric.numel())
|
384 |
+
if k_topk > 0:
|
385 |
+
_, transfer_indices_relative = torch.topk(sort_metric, k=k_topk)
|
386 |
+
else:
|
387 |
+
transfer_indices_relative = torch.tensor([], dtype=torch.long, device=device)
|
388 |
+
|
389 |
+
else: # Sample based on confidence temperature
|
390 |
+
# Ensure confidence has elements before processing
|
391 |
+
if confidence.numel() > 0:
|
392 |
+
conf_probs = confidence / alg_temp_val
|
393 |
+
# Handle potential inf/-inf before softmax, ensure non-negative probabilities
|
394 |
+
conf_probs = torch.nan_to_num(conf_probs, nan=0.0, posinf=1e9, neginf=-1e9) # Use large numbers instead of inf
|
395 |
+
conf_probs = torch.clamp(conf_probs - conf_probs.max(), min=-30) # Prevent large positive values leading to inf in exp
|
396 |
+
conf_probs = F.softmax(conf_probs, dim=-1)
|
397 |
+
conf_probs = torch.clamp(conf_probs, min=0.0) # Ensure non-negative
|
398 |
+
conf_probs = torch.nan_to_num(conf_probs, nan=0.0) # Handle NaNs
|
399 |
+
|
400 |
+
# Normalize probabilities if they don't sum to 1
|
401 |
+
prob_sum = conf_probs.sum()
|
402 |
+
if not torch.isclose(prob_sum, torch.tensor(1.0, device=device), atol=1e-4) and prob_sum > 0:
|
403 |
+
# print(f"Warning step {i}: Confidence probabilities sum {prob_sum:.4f} != 1. Re-normalizing.")
|
404 |
+
conf_probs = conf_probs / prob_sum
|
405 |
+
|
406 |
+
if conf_probs.numel() > 0 and num_samples > 0 and torch.all(conf_probs >= 0) and torch.isclose(conf_probs.sum(), torch.tensor(1.0, device=device)):
|
407 |
+
try:
|
408 |
+
transfer_indices_relative = torch.multinomial(conf_probs, num_samples=num_samples, replacement=False)
|
409 |
+
except RuntimeError as e:
|
410 |
+
print(f"Warning step {i}: Multinomial sampling failed ('{e}'). Falling back to top-k.")
|
411 |
+
sort_metric = confidence if alg != 'entropy' else -confidence
|
412 |
+
k_multinomial_fallback = min(num_samples, sort_metric.numel())
|
413 |
+
if k_multinomial_fallback > 0:
|
414 |
+
_, transfer_indices_relative = torch.topk(sort_metric, k=k_multinomial_fallback)
|
415 |
+
else:
|
416 |
+
transfer_indices_relative = torch.tensor([], dtype=torch.long, device=device)
|
417 |
+
else: # Handle cases where multinomial is not possible
|
418 |
+
# print(f"Warning step {i}: Invalid probabilities for multinomial sampling. Falling back to top-k.")
|
419 |
+
sort_metric = confidence if alg != 'entropy' else -confidence
|
420 |
+
k_multinomial_fallback = min(num_samples, sort_metric.numel())
|
421 |
+
if k_multinomial_fallback > 0:
|
422 |
+
_, transfer_indices_relative = torch.topk(sort_metric, k=k_multinomial_fallback)
|
423 |
+
else:
|
424 |
+
transfer_indices_relative = torch.tensor([], dtype=torch.long, device=device)
|
425 |
+
else: # No confidence values to sample from
|
426 |
+
transfer_indices_relative = torch.tensor([], dtype=torch.long, device=device)
|
427 |
+
|
428 |
+
# Apply the transfer
|
429 |
+
if transfer_indices_relative.numel() > 0:
|
430 |
+
# Ensure indices are within bounds of x0_candidates
|
431 |
+
valid_indices = transfer_indices_relative < x0_candidates.shape[0]
|
432 |
+
valid_transfer_indices = transfer_indices_relative[valid_indices]
|
433 |
+
|
434 |
+
if valid_transfer_indices.numel() > 0:
|
435 |
+
# Ensure indices are also within bounds of x_new_masked_part
|
436 |
+
if valid_transfer_indices.max() < x_new_masked_part.shape[0]:
|
437 |
+
x_new_masked_part[valid_transfer_indices] = x0_candidates[valid_transfer_indices].clone()
|
438 |
+
else:
|
439 |
+
print(f"Warning step {i}: transfer_indices out of bounds for x_new_masked_part.")
|
440 |
|
441 |
# Update the global state `x` only at the masked positions
|
442 |
x[mask_index] = x_new_masked_part
|
443 |
|
444 |
# --- Apply Constraints ---
|
|
|
445 |
x = apply_constraints_to_state(x, prompt_length, total_length, parsed_constraints, current_step=i)
|
446 |
|
447 |
# --- Yield Visualization ---
|
448 |
+
current_generated_tokens = x[0, prompt_length:].cpu()
|
449 |
vis_data = []
|
450 |
+
# [Keep visualization formatting logic the same]
|
451 |
for j in range(gen_length):
|
452 |
current_tok_id = current_generated_tokens[j].item()
|
453 |
+
previous_tok_id = previous_tokens_vis[j].item() if previous_tokens_vis is not None and j < len(previous_tokens_vis) else MASK_ID
|
454 |
|
455 |
try:
|
456 |
+
# Use replace to handle potential bytes rendering issues
|
457 |
+
decoded_token = tokenizer.decode([current_tok_id], skip_special_tokens=False, clean_up_tokenization_spaces=False)
|
458 |
display_token = MASK_TOKEN if current_tok_id == MASK_ID else decoded_token
|
459 |
except Exception:
|
460 |
display_token = f"[ID:{current_tok_id}]" # Fallback
|
|
|
469 |
else: # Token was already revealed
|
470 |
color = "#6699CC" # Light Blue
|
471 |
|
|
|
|
|
472 |
should_hide = (PAD_ID is not None and current_tok_id == PAD_ID) or \
|
473 |
(EOS_ID is not None and current_tok_id == EOS_ID)
|
474 |
if should_hide and previous_tok_id == current_tok_id:
|
475 |
token_to_display = "" # Hide by making empty
|
476 |
color = None # No color for hidden
|
477 |
|
|
|
478 |
if token_to_display:
|
479 |
vis_data.append((token_to_display, color))
|
480 |
+
|
481 |
+
previous_tokens_vis = current_generated_tokens # Update for next step
|
482 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
483 |
intermediate_response_tokens = x[0, prompt_length:]
|
484 |
intermediate_response_text = tokenizer.decode(
|
485 |
intermediate_response_tokens,
|
|
|
487 |
clean_up_tokenization_spaces=True
|
488 |
).strip()
|
489 |
|
|
|
|
|
|
|
490 |
yield history_copy, vis_data, intermediate_response_text
|
491 |
time.sleep(visualization_delay)
|
492 |
|
|
|
496 |
# --- 6. Final Processing & Yield ---
|
497 |
final_sequence = x[0]
|
498 |
response_tokens = final_sequence[prompt_length:]
|
|
|
|
|
499 |
final_response_text = tokenizer.decode(
|
500 |
response_tokens,
|
501 |
+
skip_special_tokens=True,
|
502 |
clean_up_tokenization_spaces=True
|
503 |
).strip()
|
|
|
|
|
504 |
history_copy[-1][1] = final_response_text
|
505 |
|
|
|
|
|
506 |
final_generated_tokens = x[0, prompt_length:].cpu()
|
507 |
vis_data_final = []
|
508 |
+
# [Keep final visualization formatting logic the same]
|
509 |
for j in range(gen_length):
|
510 |
current_tok_id = final_generated_tokens[j].item()
|
511 |
+
previous_tok_id = previous_tokens_vis[j].item() if previous_tokens_vis is not None and j < len(previous_tokens_vis) else MASK_ID
|
|
|
512 |
try:
|
513 |
+
decoded_token = tokenizer.decode([current_tok_id], skip_special_tokens=False, clean_up_tokenization_spaces=False)
|
514 |
display_token = MASK_TOKEN if current_tok_id == MASK_ID else decoded_token
|
515 |
except Exception:
|
516 |
display_token = f"[ID:{current_tok_id}]" # Fallback
|
|
|
517 |
color = None
|
518 |
token_to_display = display_token
|
519 |
+
if current_tok_id == MASK_ID: color = "#444444"
|
520 |
+
elif previous_tok_id == MASK_ID: color = "#66CC66"
|
521 |
+
else: color = "#6699CC"
|
|
|
|
|
|
|
|
|
|
|
522 |
should_hide = (PAD_ID is not None and current_tok_id == PAD_ID) or \
|
523 |
(EOS_ID is not None and current_tok_id == EOS_ID)
|
524 |
if should_hide and previous_tok_id == current_tok_id:
|
525 |
+
token_to_display = ""; color = None
|
526 |
+
if token_to_display: vis_data_final.append((token_to_display, color))
|
|
|
|
|
|
|
527 |
|
|
|
528 |
yield history_copy, vis_data_final, final_response_text
|
529 |
print("Visualization streaming complete.")
|
530 |
|
|
|
531 |
except Exception as e:
|
532 |
print(f"Error during generation or processing: {e}")
|
533 |
import traceback
|
534 |
traceback.print_exc()
|
|
|
535 |
yield history_copy, [("Error during generation.", "red")], ""
|
536 |
return
|
537 |
|
538 |
|
539 |
+
# --- Gradio UI (No changes needed here) ---
|
540 |
css = '''
|
541 |
.category-legend{display:none}
|
542 |
button{min-height: 60px}
|
|
|
549 |
"[[Blog](https://hkunlp.github.io/blog/2025/dream/)]" # Note: Link might be hypothetical
|
550 |
)
|
551 |
|
|
|
|
|
|
|
552 |
_chat_history_store = gr.State([]) # Hidden state to store actual history list
|
553 |
|
|
|
554 |
with gr.Row():
|
555 |
with gr.Column(scale=3):
|
556 |
chatbot_ui = gr.Chatbot(
|
|
|
558 |
height=500,
|
559 |
show_copy_button=True,
|
560 |
bubble_full_width=False,
|
|
|
561 |
)
|
|
|
|
|
562 |
with gr.Group():
|
563 |
with gr.Row():
|
564 |
user_input = gr.Textbox(
|
|
|
567 |
scale=7,
|
568 |
autofocus=True,
|
569 |
show_label=False,
|
570 |
+
container=False
|
571 |
)
|
572 |
send_btn = gr.Button("Send", scale=1, variant="primary")
|
|
|
|
|
573 |
constraints_input = gr.Textbox(
|
574 |
label="Word Constraints (Optional)",
|
575 |
info="Place words at specific positions (0-indexed from start of generation). Format: 'pos:word, pos:word,...'. Example: '0:Once, 5:upon, 10:time'",
|
|
|
579 |
with gr.Column(scale=2):
|
580 |
output_vis = gr.HighlightedText(
|
581 |
label="Denoising Process Visualization",
|
582 |
+
combine_adjacent=True,
|
583 |
+
show_legend=False,
|
584 |
+
interactive=False
|
585 |
)
|
|
|
586 |
response_text_display = gr.Textbox(
|
587 |
label="Generated Response",
|
588 |
interactive=False,
|
589 |
+
lines=5
|
590 |
)
|
591 |
|
|
|
|
|
592 |
with gr.Accordion("Generation Settings", open=False):
|
593 |
+
with gr.Row():
|
594 |
+
gen_length = gr.Slider(minimum=16, maximum=512, value=128, step=8, label="Max New Tokens")
|
595 |
+
steps = gr.Slider(minimum=8, maximum=512, value=128, step=8, label="Diffusion Steps")
|
596 |
+
with gr.Row():
|
597 |
+
temperature = gr.Slider(minimum=0.0, maximum=1.0, value=0.4, step=0.05, label="Temperature (0 = greedy)")
|
598 |
+
alg_temp = gr.Slider(minimum=0.0, maximum=1.0, value=0.1, step=0.05, label="Remasking Temp (Confidence Algs)")
|
599 |
+
with gr.Row():
|
600 |
+
top_p = gr.Slider(minimum=0.0, maximum=1.0, value=0.95, step=0.05, label="Top-P (0 disables)")
|
601 |
+
top_k = gr.Slider(minimum=0, maximum=200, value=0, step=5, label="Top-K (0 disables)")
|
602 |
+
with gr.Row():
|
603 |
+
remasking_strategy = gr.Radio(choices=['origin', 'maskgit_plus', 'topk_margin', 'entropy'], value='entropy', label="Remasking Strategy (Algorithm)")
|
604 |
+
with gr.Row():
|
605 |
+
visualization_delay = gr.Slider(minimum=0.0, maximum=0.5, value=0.03, step=0.01, label="Visualization Delay (seconds)")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
606 |
|
|
|
607 |
clear_btn = gr.Button("Clear Conversation")
|
608 |
|
|
|
|
|
609 |
def add_user_message_to_history(message: str, history_store: List[List[Optional[str]]]):
|
|
|
610 |
if not message.strip():
|
611 |
gr.Warning("Please enter a message.")
|
|
|
612 |
return history_store, history_store, "", [], ""
|
|
|
|
|
613 |
history_store.append([message, None])
|
|
|
614 |
return history_store, history_store, "", [], ""
|
615 |
|
616 |
def clear_conversation():
|
617 |
+
return [], [], "", [], ""
|
|
|
618 |
|
|
|
|
|
|
|
619 |
generation_inputs = [
|
620 |
_chat_history_store, gen_length, steps, constraints_input,
|
621 |
temperature, top_p, top_k, remasking_strategy, alg_temp,
|
622 |
visualization_delay
|
623 |
]
|
|
|
|
|
|
|
624 |
generation_outputs = [chatbot_ui, output_vis, response_text_display]
|
625 |
|
|
|
626 |
submit_listener = user_input.submit(
|
627 |
fn=add_user_message_to_history,
|
628 |
inputs=[user_input, _chat_history_store],
|
629 |
+
outputs=[_chat_history_store, chatbot_ui, user_input, output_vis, response_text_display]
|
630 |
+
).then(
|
|
|
|
|
631 |
fn=generate_dream_response,
|
632 |
inputs=generation_inputs,
|
633 |
+
outputs=generation_outputs,
|
634 |
+
show_progress="hidden"
|
635 |
)
|
636 |
|
|
|
637 |
click_listener = send_btn.click(
|
638 |
fn=add_user_message_to_history,
|
639 |
inputs=[user_input, _chat_history_store],
|
640 |
+
outputs=[_chat_history_store, chatbot_ui, user_input, output_vis, response_text_display]
|
641 |
+
).then(
|
|
|
|
|
642 |
fn=generate_dream_response,
|
643 |
inputs=generation_inputs,
|
644 |
+
outputs=generation_outputs,
|
645 |
show_progress="hidden"
|
646 |
)
|
647 |
|
|
|
648 |
clear_btn.click(
|
649 |
clear_conversation,
|
650 |
inputs=[],
|
|
|
656 |
# --- Launch ---
|
657 |
if __name__ == "__main__":
|
658 |
demo = create_chatbot_demo()
|
659 |
+
demo.queue().launch(debug=True, share=False)
|
|