File size: 19,224 Bytes
0ebbe7f 609fa8e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 | # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: LicenseRef-NvidiaProprietary
#
# NVIDIA CORPORATION, its affiliates and licensors retain all intellectual
# property and proprietary rights in and to this material, related
# documentation and any modifications thereto. Any use, reproduction,
# disclosure or distribution of this material and related documentation
# without an express license agreement from NVIDIA CORPORATION or
# its affiliates is strictly prohibited.
import numpy as np
import torch
import torch.nn.functional as F
from transformers.utils import ModelOutput
from dataclasses import dataclass
from transformers.cache_utils import Cache, DynamicCache
@dataclass
class SimpleOutputWithPast(ModelOutput):
loss: torch.FloatTensor | None = None
logits: torch.FloatTensor | None = None
causal_logits: torch.FloatTensor | None = None
past_key_values: Cache | None = None
hidden_states: tuple[torch.FloatTensor, ...] | None = None
attentions: tuple[torch.FloatTensor, ...] | None = None
from .nemotron_diffusion_image_utils import maybe_truncate_last_dim, pad_along_last_dim
def wte(model,x,t2i_inference=False,gen_shape=None,x_gen=None,inputs_embeds_curr=None,new_token_mask=None):
if t2i_inference:
assert x_gen is not None
if new_token_mask is None:
new_token_mask = x >= INT_MAX
# if x_gen is None:
# x_gen = x[new_token_mask] - OFFSET
# else:
# x_gen = x_gen - OFFSET
gen_latents_comp_embeds = model.call_gen_embedding(x_gen,gen_shape)
if inputs_embeds_curr is None:
x_txt_only = x.clone()
# replace consequtent [1] * 4096 to [1] * 1024
x_txt_only[new_token_mask] = 0
inputs_embeds_curr = model.embed_tokens(x_txt_only)
inputs_embeds_curr[new_token_mask] = pad_along_last_dim(gen_latents_comp_embeds,inputs_embeds_curr.shape[-1]).view(-1,inputs_embeds_curr.shape[-1])
else:
inputs_embeds_curr = model.embed_tokens(x)
new_token_mask = None
return inputs_embeds_curr,new_token_mask
INT_MAX = 1_000_000
def get_logits(model,input_emnbeddings,modality_indices=None,t2i_inference=False,past_key_values=None,gen_shape=None,timesteps=None,input_modality_indices=None):
if t2i_inference:
if input_modality_indices is None:
input_modality_indices =modality_indices
output = model(None,input_embeddings=input_emnbeddings,modality_indices=input_modality_indices,output_hidden_states=True,past_key_values=past_key_values,
is_training=False,
overwrite_attn_impl='flash_attn'
)
hidden_states = output.hidden_states[-1]
gen_hidden_states = hidden_states[modality_indices]
gen_hidden_states = maybe_truncate_last_dim(gen_hidden_states,model.config.d_model_gen)
gen_logits = model.call_gen_predictor(gen_hidden_states,gen_shape,timesteps=timesteps) # * 8 D
seq_len_per_img = np.prod(gen_shape)
if len(gen_logits.shape) == 2:
gen_logits = gen_logits.view(-1,seq_len_per_img,gen_logits.shape[-1])
else:
gen_logits = gen_logits.view(-1,seq_len_per_img,*gen_logits.shape[-2:])
# N L 8 D
return gen_logits
final_logits = torch.zeros(*gen_logits.shape[:-1],OFFSET+gen_logits.shape[-1],dtype=output.logits.dtype,device=output.logits.device)
final_logits[:] = float('-inf')
final_logits[...,OFFSET:] = gen_logits
# breakpoint()
# inal_logits = torch.zeros(*hidden_states.shape[:-1],OFFSET+gen_logits.shape[-1],dtype=output.logits.dtype,device=output.logits.device)
# final_logits = final_logits + float('-inf')
# final_logits[...,:output.logits.shape[-1]] = output.logits
# final_logits[modality_indices] = float('-inf')
# local = final_logits[modality_indices]
# local[...,OFFSET:] = gen_logits
# final_logits[modality_indices] = local
logits = final_logits
return logits
else:
modality_indices = torch.zeros(input_emnbeddings.shape[:-1],device=input_emnbeddings.device,dtype=torch.bool)
logits = model(None,input_embeddings=input_emnbeddings,modality_indices=modality_indices,past_key_values=past_key_values).logits
return logits
def add_gumbel_noise(logits, temperature):
'''
The Gumbel max is a method for sampling categorical distributions.
According to arXiv:2409.02908, for MDM, low-precision Gumbel Max improves perplexity score but reduces generation quality.
Thus, we use float64.
'''
if temperature == 0:
return logits
logits = logits.to(torch.float64)
noise = torch.rand_like(logits, dtype=torch.float64)
gumbel_noise = (- torch.log(noise)) ** temperature
return logits.exp() / gumbel_noise
def get_transfer_index(logits, temperature, remasking, mask_index, x, num_transfer_tokens, threshold=None, neg_entropy=False):
logits_with_noise = add_gumbel_noise(logits, temperature=temperature)
x0 = torch.argmax(logits_with_noise, dim=-1)
if remasking == 'low_confidence':
# p = F.softmax(logits.to(torch.float64), dim=-1)
p = F.softmax(logits, dim=-1)
x0_p = torch.squeeze(
torch.gather(p, dim=-1, index=torch.unsqueeze(x0, -1)), -1) # b, l
elif remasking == 'top_p_margin':
# Compute probabilities
p = F.softmax(logits, dim=-1) # (B, L, V)
# Top-2 per position
top2 = torch.topk(p, k=2, dim=-1).values # (B, L, 2)
margin = top2[..., 0] - top2[..., 1] # (B, L)
# Normalize margin to [0,1] over MASKED positions per row
plus_inf = torch.full_like(margin, float('inf'))
minus_inf = torch.full_like(margin, float('-inf'))
masked_for_min = torch.where(mask_index, margin, plus_inf)
masked_for_max = torch.where(mask_index, margin, minus_inf)
row_min = masked_for_min.amin(dim=1, keepdim=True) # (B, 1)
row_max = masked_for_max.amax(dim=1, keepdim=True) # (B, 1)
denom = (row_max - row_min)
# If denom==0 (all equal), set normalized=1 on masked; 0 elsewhere by default
normalized = torch.zeros_like(margin)
nonzero = denom > 0
normalized = torch.where(
mask_index & nonzero,
(margin - row_min) / (denom + 1e-12),
normalized
)
normalized = torch.where(
mask_index & (~nonzero),
torch.ones_like(normalized),
normalized
)
x0_p = normalized # ∈ [0,1] on masked positions
elif remasking == 'random':
x0_p = torch.rand((x0.shape[0], x0.shape[1]), device=x0.device)
else:
raise NotImplementedError(remasking)
# Calculate negative entropy if requested
if neg_entropy:
# p = F.softmax(logits.to(torch.float64), dim=-1)
p = F.softmax(logits, dim=-1)
epsilon = 1e-10
log_probs = torch.log(p + epsilon)
confidence_scores = torch.sum(p * log_probs, dim=-1) # negative entropy per position
else:
confidence_scores = x0_p
x0 = torch.where(mask_index, x0, x)
confidence = torch.where(mask_index, confidence_scores, -np.inf)
transfer_index = torch.zeros_like(x0, dtype=torch.bool, device=x0.device)
if threshold is not None:
num_transfer_tokens = mask_index.sum(dim=1, keepdim=True)
# print(f'confidence: {confidence}')
for j in range(confidence.shape[0]):
_, select_index = torch.topk(confidence[j], k=num_transfer_tokens[j])
transfer_index[j, select_index] = True
if threshold is not None:
for k in range(1, num_transfer_tokens[j]):
if confidence[j, select_index[k]] < threshold:
transfer_index[j, select_index[k]] = False
return x0, transfer_index
def get_num_transfer_tokens(mask_index, steps: int):
mask_num = mask_index.sum(dim=1, keepdim=True)
base = mask_num // steps
remainder = mask_num % steps
num_transfer_tokens = torch.zeros(mask_num.size(0), steps, device=mask_index.device, dtype=torch.int64) + base
for i in range(mask_num.size(0)):
num_transfer_tokens[i, : int(remainder[i])] += 1
return num_transfer_tokens
def simple_fwd(model,input_ids=None,inputs_embeds=None,attention_mask=None,position_ids=None,past_key_values=None,**kwargs):
enc_out = model.encoder(
past_key_values=past_key_values,
input_ids=input_ids,
inputs_embeds=inputs_embeds,
attention_mask=attention_mask,
position_ids=position_ids,
is_training=False,
overwrite_attn_impl='flash_attn',
# overwrite_attn_impl='flash_attn',
# overwrite_block_mask='full',
**kwargs,
)
logits = model.diffusion_head(enc_out.last_hidden_state)
return SimpleOutputWithPast(
loss=logits,
logits=logits,
causal_logits=None,
past_key_values=enc_out.past_key_values,
hidden_states=None,
attentions=None,
)
@torch.no_grad()
def generate_with_prefix_cache_block_diff(
model,
prompt=None,
prompt_embeds=None,
steps=128,
gen_length=128,
block_length=128,
temperature=0.,
remasking='low_confidence',
mask_id=126336,
threshold=None,
factor=None,
shift_logits=False,
neg_entropy=False,
causal_context=False,
eos_token_id=None,
max_thinking_tokens=None,
end_think_token_id=None,
):
dream_style=shift_logits
if (prompt is None) == (prompt_embeds is None):
raise ValueError("Exactly one of `prompt` or `prompt_embeds` must be provided.")
if prompt is not None:
prompt_ids = prompt
prompt_len = prompt_ids.shape[1]
x_accum = prompt_ids.clone()
B = prompt_ids.shape[0]
token_device = prompt_ids.device
token_dtype = prompt_ids.dtype
else:
prompt_ids = None
prompt_len = prompt_embeds.shape[1]
B = prompt_embeds.shape[0]
token_device = prompt_embeds.device
token_dtype = torch.long
# Keep prefix slots so block slicing by prompt_len stays identical.
x_accum = torch.full((B, prompt_len), mask_id, dtype=token_dtype, device=token_device)
assert gen_length % block_length == 0
num_blocks = gen_length // block_length
assert steps % num_blocks == 0
steps_per_block = steps // num_blocks
nfe = 0
model_module = model.module if hasattr(model, "module") else model
for layer in model_module.encoder.layers:
layer.self_attn.mode = 'bidirectional'
if causal_context:
for layer in model_module.encoder.layers:
if hasattr(layer.self_attn, 'diffusion_lm'):
layer.self_attn.diffusion_lm=False
# Compute KV cache for the prompt initially
output = simple_fwd(model,
input_ids=prompt_ids,
inputs_embeds=prompt_embeds,
use_cache=True,
use_causal_mask=causal_context,
)
past_key_values = output.past_key_values
if causal_context:
for layer in model_module.encoder.layers:
if hasattr(layer.self_attn, 'diffusion_lm'):
layer.self_attn.diffusion_lm=True
# Causal prefill: next token from last position (same as linear_spec_generate).
next_token = None
if causal_context:
last_logit = output.logits[:, -1, :]
if temperature > 0:
probs = torch.softmax(last_logit / temperature, dim=-1)
next_token = torch.multinomial(probs, num_samples=1)
else:
next_token = torch.argmax(last_logit, dim=-1, keepdim=True)
# For dream_style: store the "next token logit" of the context
next_logits_context = None
if dream_style:
next_logits_context = output.logits[:, -1:, :] # (B, 1, V)
for num_block in range(num_blocks):
# Create a new block with mask tokens; under causal context, seed position 0
# with the next-token prediction from the previous causal forward (prefill or
# post-block encode), matching linear_spec_generate.
mask_block = torch.ones(
(B, block_length),
dtype=token_dtype,
device=token_device,
) * mask_id
if causal_context:
mask_block[:, 0] = next_token[:, 0]
# Append the block of masks
x_accum = torch.cat([x_accum, mask_block], dim=1)
current_block_start = prompt_len + num_block * block_length
block_slice = slice(current_block_start, current_block_start + block_length)
# ---- thinking budget enforcement ----
# If we've generated >= max_thinking_tokens without a </think>, inject one.
if end_think_token_id is not None and max_thinking_tokens is not None:
tokens_before_block = num_block * block_length
tokens_after_block = tokens_before_block + block_length
if tokens_after_block > max_thinking_tokens:
gen_so_far = x_accum[:, prompt_len:current_block_start]
has_end_think = (
(gen_so_far == end_think_token_id).any(dim=1)
if gen_so_far.size(1) > 0
else torch.zeros(B, dtype=torch.bool, device=token_device)
)
if not has_end_think.all():
if tokens_before_block < max_thinking_tokens:
offset = max_thinking_tokens - tokens_before_block
else:
offset = 0
inject_pos = current_block_start + offset
for b in range(B):
if not has_end_think[b]:
x_accum[b, inject_pos] = end_think_token_id
# Build the initial mask for this block
mask_block_idx0 = (x_accum[:, block_slice] == mask_id) # (B, Lb)
# Precompute the transfer schedule for this block
if dream_style:
# masked positions only (position 0 may be causal-seeded, not mask_id)
schedule_mask = mask_block_idx0
else:
schedule_mask = mask_block_idx0
num_transfer_tokens = get_num_transfer_tokens(schedule_mask, steps_per_block) # (B, steps)
# Denoise the current block
for i in range(steps_per_block):
mask_block_idx = (x_accum[:, block_slice] == mask_id) # (B, Lb)
if mask_block_idx.sum() == 0:
break
nfe += 1
# Forward only the current noisy block using cached context
logits_block = simple_fwd(model,
x_accum[:, block_slice],
past_key_values=past_key_values,
use_cache=False
).logits
if dream_style:
# Align logits so that each masked position has a predictor:
# prepend context-next logit, then use logits_block[:-1]
if block_length == 1:
logits_use = next_logits_context # (B, 1, V)
else:
logits_use = torch.cat(
[next_logits_context, logits_block[:, :-1, :]],
dim=1
) # (B, Lb, V)
mask_use = mask_block_idx # (B, Lb)
x_use = x_accum[:, block_slice] # (B, Lb)
x0, transfer_idx = get_transfer_index(
logits_use, temperature, remasking, mask_use, x_use,
num_transfer_tokens=num_transfer_tokens[:, i],
threshold=threshold, neg_entropy=neg_entropy
)
cur = x_accum[:, block_slice].clone()
cur[transfer_idx] = x0[transfer_idx]
x_accum[:, block_slice] = cur
else:
# non-AR (same-position) case
x0, transfer_idx = get_transfer_index(
logits_block, temperature, remasking, mask_block_idx,
x_accum[:, block_slice],
num_transfer_tokens=num_transfer_tokens[:, i],
threshold=threshold, neg_entropy=neg_entropy
)
cur = x_accum[:, block_slice].clone()
cur[transfer_idx] = x0[transfer_idx]
x_accum[:, block_slice] = cur
if eos_token_id is not None:
block_tokens = x_accum[:, block_slice] # (B, Lb)
eos_mask = (block_tokens == eos_token_id) # (B, Lb)
any_eos = eos_mask.any(dim=1) # (B,)
if any_eos.any():
after_eos = eos_mask.cumsum(dim=1).bool() # (B, Lb)
mask_before = (block_tokens == mask_id) & ~after_eos
if (any_eos & ~mask_before.any(dim=1)).any():
break
if causal_context:
for layer in model_module.encoder.layers:
if hasattr(layer.self_attn, 'diffusion_lm'):
layer.self_attn.diffusion_lm=False
# after block is fully denoised, update KV cache
output = simple_fwd(model,
x_accum[:, block_slice],
past_key_values=past_key_values,
use_cache=True,
use_causal_mask=causal_context
)
past_key_values = output.past_key_values
nfe += 1
if causal_context:
for layer in model_module.encoder.layers:
if hasattr(layer.self_attn, 'diffusion_lm'):
layer.self_attn.diffusion_lm=True
# Next block's first position = greedy/sampled next token from this causal encode
last_logit = output.logits[:, -1, :]
if temperature > 0:
probs = torch.softmax(last_logit / temperature, dim=-1)
next_token = torch.multinomial(probs, num_samples=1)
else:
next_token = torch.argmax(last_logit, dim=-1, keepdim=True)
if dream_style and num_block < num_blocks - 1:
# refresh context-next logit for the next block
next_logits_context = output.logits[:, -1:, :] # (B, 1, V)
if eos_token_id is not None:
gen_so_far = x_accum[:, prompt_len:] # (B, gen_len_so_far)
is_eos = (gen_so_far == eos_token_id) # (B, gen_len_so_far)
has_eos = is_eos.any(dim=1) # (B,)
if has_eos.all():
first_eos_pos = is_eos.to(torch.int64).argmax(dim=1) # (B,)
max_eos = first_eos_pos.max().item()
if prompt_ids is None:
return x_accum[:, prompt_len : prompt_len + max_eos + 1], nfe
return x_accum[:, : prompt_len + max_eos + 1], nfe
if prompt_ids is None:
return x_accum[:, prompt_len:], nfe
return x_accum, nfe
|