Spaces:
Paused
Paused
File size: 39,444 Bytes
01115c6 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 |
# Updated generator.py with proper function order
from dataclasses import dataclass
from typing import List, Tuple
import torch
import torchaudio
import logging
import os
from huggingface_hub import hf_hub_download
from transformers import AutoTokenizer
from tokenizers.processors import TemplateProcessing
from app.models import Segment
from app.text_normalizer import clean_text_for_tts
from app.text_normalizer import TextNormalizer
# Set up logging
logger = logging.getLogger(__name__)
# Import the CSM watermarking code
try:
from app.watermarking import CSM_1B_GH_WATERMARK, load_watermarker, watermark
except ImportError:
# Define stubs for watermarking if the module is not available
CSM_1B_GH_WATERMARK = "CSM1B"
def load_watermarker(device="cpu"):
return None
def watermark(watermarker, audio, sample_rate, key):
return audio, sample_rate
def load_llama3_tokenizer():
"""
Load tokenizer for Llama 3.2, using unsloth's open version
instead of the gated meta-llama version.
"""
try:
# Use the unsloth version which is not gated
tokenizer_name = "unsloth/Llama-3.2-1B"
tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
bos = tokenizer.bos_token
eos = tokenizer.eos_token
tokenizer._tokenizer.post_processor = TemplateProcessing(
single=f"{bos}:0 $A:0 {eos}:0",
pair=f"{bos}:0 $A:0 {eos}:0 {bos}:1 $B:1 {eos}:1",
special_tokens=[(f"{bos}", tokenizer.bos_token_id), (f"{eos}", tokenizer.eos_token_id)],
)
logger.info("Successfully loaded tokenizer from unsloth/Llama-3.2-1B")
return tokenizer
except Exception as e:
logger.error(f"Error loading tokenizer from unsloth: {e}")
# Fallback to a simpler tokenizer if needed
try:
from transformers import GPT2Tokenizer
logger.warning("Falling back to GPT2Tokenizer")
tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
tokenizer.pad_token = tokenizer.eos_token
return tokenizer
except Exception as fallback_e:
logger.error(f"Fallback tokenizer also failed: {fallback_e}")
raise RuntimeError("Could not load any suitable tokenizer")
class Generator:
"""Generator class for CSM-1B model."""
def __init__(self, model):
"""Initialize generator with model."""
self._model = model
self._model.setup_caches(1)
self._text_tokenizer = load_llama3_tokenizer()
device = next(model.parameters()).device
# Load Mimi codec for audio tokenization
try:
logger.info("Loading Mimi audio codec...")
from huggingface_hub import hf_hub_download
# First try to import from moshi
try:
from moshi.models import loaders
DEFAULT_REPO = loaders.DEFAULT_REPO
MIMI_NAME = loaders.MIMI_NAME
get_mimi = loaders.get_mimi
except ImportError:
logger.warning("moshi.models.loaders not found, using fallback")
# Fallback values if moshi.models.loaders is not available
DEFAULT_REPO = "kyutai/mimi"
MIMI_NAME = "mimi-december.pt"
# Fallback function to load mimi
def get_mimi(checkpoint_path, device):
from moshi.models.vqvae_model import MiMiModule
checkpoint = torch.load(checkpoint_path, map_location=device)
model = MiMiModule.init_from_checkpoint(checkpoint, device=device)
return model
mimi_weight = hf_hub_download(DEFAULT_REPO, MIMI_NAME)
mimi = get_mimi(mimi_weight, device=device)
mimi.set_num_codebooks(32)
self._audio_tokenizer = mimi
self.sample_rate = mimi.sample_rate
logger.info(f"Mimi codec loaded successfully with sample rate {self.sample_rate}")
except Exception as e:
logger.error(f"Error loading Mimi codec: {e}")
self._audio_tokenizer = None
self.sample_rate = 24000 # Default sample rate
logger.warning(f"Using fallback sample rate: {self.sample_rate}")
raise RuntimeError(f"Failed to load Mimi codec: {e}")
try:
self._watermarker = load_watermarker(device=device)
logger.info("Watermarker loaded successfully")
except Exception as e:
logger.warning(f"Error loading watermarker: {e}. Watermarking will be disabled.")
self._watermarker = None
self.device = device
# Optimize for CUDA throughput
if torch.cuda.is_available():
torch.backends.cudnn.benchmark = True
torch.cuda.empty_cache()
logger.info("CUDA optimizations enabled")
def _tokenize_text_segment(self, text: str, speaker: int) -> Tuple[torch.Tensor, torch.Tensor]:
"""Tokenize a text segment."""
frame_tokens = []
frame_masks = []
# Strip any voice instructions in square brackets to avoid them being read out
text = self._clean_text_input(text)
text_tokens = self._text_tokenizer.encode(f"[{speaker}]{text}")
text_frame = torch.zeros(len(text_tokens), 33).long()
text_frame_mask = torch.zeros(len(text_tokens), 33).bool()
text_frame[:, -1] = torch.tensor(text_tokens)
text_frame_mask[:, -1] = True
frame_tokens.append(text_frame.to(self.device))
frame_masks.append(text_frame_mask.to(self.device))
return torch.cat(frame_tokens, dim=0), torch.cat(frame_masks, dim=0)
def _clean_text_input(self, text: str) -> str:
"""Clean and normalize text for TTS."""
return clean_text_for_tts(text)
def _tokenize_audio(self, audio: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
"""Tokenize audio."""
if self._audio_tokenizer is None:
raise RuntimeError("Audio tokenizer not initialized")
frame_tokens = []
frame_masks = []
# (K, T)
audio = audio.to(self.device)
audio_tokens = self._audio_tokenizer.encode(audio.unsqueeze(0).unsqueeze(0))[0]
# add EOS frame
eos_frame = torch.zeros(audio_tokens.size(0), 1).to(self.device)
audio_tokens = torch.cat([audio_tokens, eos_frame], dim=1)
audio_frame = torch.zeros(audio_tokens.size(1), 33).long().to(self.device)
audio_frame_mask = torch.zeros(audio_tokens.size(1), 33).bool().to(self.device)
audio_frame[:, :-1] = audio_tokens.transpose(0, 1)
audio_frame_mask[:, :-1] = True
frame_tokens.append(audio_frame)
frame_masks.append(audio_frame_mask)
return torch.cat(frame_tokens, dim=0), torch.cat(frame_masks, dim=0)
def _tokenize_segment(self, segment: Segment) -> Tuple[torch.Tensor, torch.Tensor]:
"""Tokenize a segment of text and audio."""
text_tokens, text_masks = self._tokenize_text_segment(segment.text, segment.speaker)
audio_tokens, audio_masks = self._tokenize_audio(segment.audio)
return torch.cat([text_tokens, audio_tokens], dim=0), torch.cat([text_masks, audio_masks], dim=0)
def generate_quick(
self,
text: str,
speaker: int,
context: List[Segment],
max_audio_length_ms: float = 2000, # Short for quick generation
temperature: float = 0.7, # Lower for more predictable output
topk: int = 20, # Lower for faster beam selection
) -> torch.Tensor:
"""Generate audio quickly for real-time streaming."""
# Similar to generate() but optimized for speed
self._model.reset_caches()
# Convert max_audio_length_ms to frames - limit for faster generation
max_audio_frames = min(int(max_audio_length_ms / 80), 128) # Smaller limit
# Process text
cleaned_text = clean_text_for_tts(text)
# Prepare tokens
tokens, tokens_mask = [], []
# Add context segments (limited to 1 for speed)
if context:
segment_tokens, segment_tokens_mask = self._tokenize_segment(context[0])
tokens.append(segment_tokens)
tokens_mask.append(segment_tokens_mask)
# Add text tokens
gen_segment_tokens, gen_segment_tokens_mask = self._tokenize_text_segment(cleaned_text, speaker)
tokens.append(gen_segment_tokens)
tokens_mask.append(gen_segment_tokens_mask)
prompt_tokens = torch.cat(tokens, dim=0).long().to(self.device)
prompt_tokens_mask = torch.cat(tokens_mask, dim=0).bool().to(self.device)
# Generate with larger batch size for fewer iterations
curr_tokens = prompt_tokens.unsqueeze(0)
curr_tokens_mask = prompt_tokens_mask.unsqueeze(0)
curr_pos = torch.arange(0, prompt_tokens.size(0)).unsqueeze(0).long().to(self.device)
# Use larger batch size
batch_size = 64 # Generate more frames at once
all_samples = []
for start_idx in range(0, max_audio_frames, batch_size):
end_idx = min(start_idx + batch_size, max_audio_frames)
batch_frames = end_idx - start_idx
samples_batch = []
for i in range(batch_frames):
sample = self._model.generate_frame(curr_tokens, curr_tokens_mask, curr_pos, temperature, topk)
samples_batch.append(sample)
if torch.all(sample == 0):
break
curr_tokens = torch.cat([sample, torch.zeros(1, 1).long().to(self.device)], dim=1).unsqueeze(1)
curr_tokens_mask = torch.cat(
[torch.ones_like(sample).bool(), torch.zeros(1, 1).bool().to(self.device)], dim=1
).unsqueeze(1)
curr_pos = curr_pos[:, -1:] + 1
all_samples.extend(samples_batch)
if len(samples_batch) < batch_frames:
break
if not all_samples:
return torch.zeros(10, device=self.device) # Return short empty audio
# Decode audio
audio = self._audio_tokenizer.decode(torch.stack(all_samples).permute(1, 2, 0)).squeeze(0).squeeze(0)
return audio
@torch.inference_mode()
def generate(
self,
text: str,
speaker: int,
context: List[Segment],
max_audio_length_ms: float = 90_000,
temperature: float = 0.9,
topk: int = 50,
) -> torch.Tensor:
"""Generate audio from text."""
if self._audio_tokenizer is None:
raise RuntimeError("Audio tokenizer not initialized")
# Start timing
start_time = torch.cuda.Event(enable_timing=True)
end_time = torch.cuda.Event(enable_timing=True)
start_time.record()
self._model.reset_caches()
# Convert max_audio_length_ms to frames - this controls the maximum generation length
max_audio_frames = min(int(max_audio_length_ms / 80), 1024) # Limit to reasonable size
max_seq_len = 2048 - max_audio_frames
# Check if text is long and should be split
if len(text) > 200:
logger.info(f"Long text detected ({len(text)} chars), processing in segments")
sentences = TextNormalizer.split_into_sentences(text)
logger.info(f"Split into {len(sentences)} segments")
# Process sentences individually and concatenate the results
all_audio_segments = []
# Use the first sentence to establish voice
first_sentence = sentences[0]
cleaned_text = clean_text_for_tts(first_sentence)
# Generate the first segment
tokens, tokens_mask = [], []
# Add context segments for the first sentence
for segment in context:
segment_tokens, segment_tokens_mask = self._tokenize_segment(segment)
tokens.append(segment_tokens)
tokens_mask.append(segment_tokens_mask)
# Add first sentence tokens
gen_segment_tokens, gen_segment_tokens_mask = self._tokenize_text_segment(cleaned_text, speaker)
tokens.append(gen_segment_tokens)
tokens_mask.append(gen_segment_tokens_mask)
prompt_tokens = torch.cat(tokens, dim=0).long().to(self.device)
prompt_tokens_mask = torch.cat(tokens_mask, dim=0).bool().to(self.device)
# Check context size and truncate if needed
if prompt_tokens.size(0) >= max_seq_len:
logger.warning(f"Inputs too long ({prompt_tokens.size(0)} tokens), truncating to {max_seq_len - 50}")
prompt_tokens = prompt_tokens[-max_seq_len+50:]
prompt_tokens_mask = prompt_tokens_mask[-max_seq_len+50:]
# Generate first sentence audio
curr_tokens = prompt_tokens.unsqueeze(0)
curr_tokens_mask = prompt_tokens_mask.unsqueeze(0)
curr_pos = torch.arange(0, prompt_tokens.size(0)).unsqueeze(0).long().to(self.device)
# Generate first segment
first_segment_samples = []
for start_idx in range(0, max_audio_frames, 32):
end_idx = min(start_idx + 32, max_audio_frames)
batch_frames = end_idx - start_idx
samples_batch = []
for i in range(batch_frames):
sample = self._model.generate_frame(curr_tokens, curr_tokens_mask, curr_pos, temperature, topk)
samples_batch.append(sample)
if torch.all(sample == 0):
break
curr_tokens = torch.cat([sample, torch.zeros(1, 1).long().to(self.device)], dim=1).unsqueeze(1)
curr_tokens_mask = torch.cat(
[torch.ones_like(sample).bool(), torch.zeros(1, 1).bool().to(self.device)], dim=1
).unsqueeze(1)
curr_pos = curr_pos[:, -1:] + 1
first_segment_samples.extend(samples_batch)
if len(samples_batch) < batch_frames:
break
if not first_segment_samples:
raise RuntimeError("No audio generated for first segment")
# Decode first segment
first_segment_audio = self._audio_tokenizer.decode(
torch.stack(first_segment_samples).permute(1, 2, 0)
).squeeze(0).squeeze(0)
all_audio_segments.append(first_segment_audio)
# Now process remaining sentences using the first as context
for i, sentence in enumerate(sentences[1:], 1):
logger.info(f"Generating segment {i+1}/{len(sentences)}")
cleaned_text = clean_text_for_tts(sentence)
# Create a context segment from the previous generation
prev_segment = Segment(
speaker=speaker,
text=sentences[i-1],
audio=all_audio_segments[-1]
)
# Generate with this segment as context
segment_tokens, segment_tokens_mask = [], []
segment_tokens.append(self._tokenize_segment(prev_segment)[0])
segment_tokens_mask.append(self._tokenize_segment(prev_segment)[1])
# Add current segment tokens
current_tokens, current_tokens_mask = self._tokenize_text_segment(cleaned_text, speaker)
segment_tokens.append(current_tokens)
segment_tokens_mask.append(current_tokens_mask)
segment_prompt_tokens = torch.cat(segment_tokens, dim=0).long().to(self.device)
segment_prompt_tokens_mask = torch.cat(segment_tokens_mask, dim=0).bool().to(self.device)
# Check length and truncate if needed
if segment_prompt_tokens.size(0) >= max_seq_len:
segment_prompt_tokens = segment_prompt_tokens[-max_seq_len+50:]
segment_prompt_tokens_mask = segment_prompt_tokens_mask[-max_seq_len+50:]
# Generate audio for this segment
curr_tokens = segment_prompt_tokens.unsqueeze(0)
curr_tokens_mask = segment_prompt_tokens_mask.unsqueeze(0)
curr_pos = torch.arange(0, segment_prompt_tokens.size(0)).unsqueeze(0).long().to(self.device)
# Generate segment
segment_samples = []
for start_idx in range(0, max_audio_frames, 32):
end_idx = min(start_idx + 32, max_audio_frames)
batch_frames = end_idx - start_idx
samples_batch = []
for i in range(batch_frames):
sample = self._model.generate_frame(curr_tokens, curr_tokens_mask, curr_pos, temperature, topk)
samples_batch.append(sample)
if torch.all(sample == 0):
break
curr_tokens = torch.cat([sample, torch.zeros(1, 1).long().to(self.device)], dim=1).unsqueeze(1)
curr_tokens_mask = torch.cat(
[torch.ones_like(sample).bool(), torch.zeros(1, 1).bool().to(self.device)], dim=1
).unsqueeze(1)
curr_pos = curr_pos[:, -1:] + 1
segment_samples.extend(samples_batch)
if len(samples_batch) < batch_frames:
break
if not segment_samples:
logger.warning(f"No audio generated for segment {i+1}")
continue
# Decode segment
segment_audio = self._audio_tokenizer.decode(
torch.stack(segment_samples).permute(1, 2, 0)
).squeeze(0).squeeze(0)
all_audio_segments.append(segment_audio)
# Combine all segments with small pauses
pause_samples = int(0.3 * self.sample_rate) # 300ms pause
pause = torch.zeros(pause_samples, device=self.device)
audio_parts = []
for i, segment_audio in enumerate(all_audio_segments):
audio_parts.append(segment_audio)
if i < len(all_audio_segments) - 1:
audio_parts.append(pause)
audio = torch.cat(audio_parts)
logger.info(f"Combined {len(all_audio_segments)} segments into final audio")
else:
# For shorter text, standard processing
tokens, tokens_mask = [], []
# Add context segments
for segment in context:
segment_tokens, segment_tokens_mask = self._tokenize_segment(segment)
tokens.append(segment_tokens)
tokens_mask.append(segment_tokens_mask)
# Process text
cleaned_text = clean_text_for_tts(text)
gen_segment_tokens, gen_segment_tokens_mask = self._tokenize_text_segment(cleaned_text, speaker)
tokens.append(gen_segment_tokens)
tokens_mask.append(gen_segment_tokens_mask)
prompt_tokens = torch.cat(tokens, dim=0).long().to(self.device)
prompt_tokens_mask = torch.cat(tokens_mask, dim=0).bool().to(self.device)
# Check context size
if prompt_tokens.size(0) >= max_seq_len:
logger.warning(f"Inputs too long ({prompt_tokens.size(0)} tokens), truncating to {max_seq_len - 50}")
prompt_tokens = prompt_tokens[-max_seq_len+50:]
prompt_tokens_mask = prompt_tokens_mask[-max_seq_len+50:]
# Generate audio - optimized batch generation
curr_tokens = prompt_tokens.unsqueeze(0)
curr_tokens_mask = prompt_tokens_mask.unsqueeze(0)
curr_pos = torch.arange(0, prompt_tokens.size(0)).unsqueeze(0).long().to(self.device)
# Using optimized batch generation
batch_size = 32 # Generate this many frames at once
all_samples = []
for start_idx in range(0, max_audio_frames, batch_size):
end_idx = min(start_idx + batch_size, max_audio_frames)
batch_frames = end_idx - start_idx
samples_batch = []
for i in range(batch_frames):
sample = self._model.generate_frame(curr_tokens, curr_tokens_mask, curr_pos, temperature, topk)
samples_batch.append(sample)
if torch.all(sample == 0):
break
curr_tokens = torch.cat([sample, torch.zeros(1, 1).long().to(self.device)], dim=1).unsqueeze(1)
curr_tokens_mask = torch.cat(
[torch.ones_like(sample).bool(), torch.zeros(1, 1).bool().to(self.device)], dim=1
).unsqueeze(1)
curr_pos = curr_pos[:, -1:] + 1
all_samples.extend(samples_batch)
if len(samples_batch) < batch_frames:
logger.info(f"Early stopping at frame {start_idx + len(samples_batch)}/{max_audio_frames}")
break
if not all_samples:
raise RuntimeError("No audio generated - model produced empty output")
# Decode audio
audio = self._audio_tokenizer.decode(torch.stack(all_samples).permute(1, 2, 0)).squeeze(0).squeeze(0)
# Apply watermark
if self._watermarker is not None:
try:
audio, wm_sample_rate = watermark(self._watermarker, audio, self.sample_rate, CSM_1B_GH_WATERMARK)
audio = torchaudio.functional.resample(audio, orig_freq=wm_sample_rate, new_freq=self.sample_rate)
except Exception as e:
logger.warning(f"Error applying watermark: {e}. Continuing without watermark.")
# Record execution time
end_time.record()
torch.cuda.synchronize()
execution_ms = start_time.elapsed_time(end_time)
audio_length_ms = (audio.shape[0] / self.sample_rate) * 1000
# Calculate real-time factor (RTF)
rtf = execution_ms / audio_length_ms
logger.info(f"Audio generated in {execution_ms:.2f}ms, length: {audio_length_ms:.2f}ms, RTF: {rtf:.2f}x")
return audio
# Define helper functions for multi-GPU support
def _manual_device_map(model, state_dict, strategy="balanced"):
"""Apply manual device mapping for multi-GPU setups.
Args:
model: The model to map
state_dict: Model state dict
strategy: Mapping strategy ('balanced', 'sequential')
Returns:
Model with weights distributed across GPUs
"""
num_gpus = torch.cuda.device_count()
if num_gpus <= 1:
# No need for mapping with single GPU
model.load_state_dict(state_dict)
model = model.to("cuda")
return model
logger.info(f"Applying manual {strategy} device mapping across {num_gpus} GPUs")
# Get all layer names from state dict
layer_names = [name for name in state_dict.keys() if "layers" in name]
backbone_layers = [name for name in layer_names if "backbone.layers" in name]
decoder_layers = [name for name in layer_names if "decoder.layers" in name]
# Count number of backbone and decoder layers
backbone_layer_indices = set()
for name in backbone_layers:
parts = name.split('.')
if len(parts) > 2:
try:
backbone_layer_indices.add(int(parts[2]))
except ValueError:
pass
decoder_layer_indices = set()
for name in decoder_layers:
parts = name.split('.')
if len(parts) > 2:
try:
decoder_layer_indices.add(int(parts[2]))
except ValueError:
pass
num_backbone_layers = len(backbone_layer_indices)
num_decoder_layers = len(decoder_layer_indices)
# Create device map
device_map = {}
if strategy == "balanced":
# Distribute layers evenly across GPUs
layers_per_gpu = (num_backbone_layers + num_decoder_layers) // num_gpus
remainder = (num_backbone_layers + num_decoder_layers) % num_gpus
# Assign backbone layers
for i in backbone_layer_indices:
gpu_idx = min(i // layers_per_gpu, num_gpus - 1)
device_map[f"backbone.layers.{i}"] = f"cuda:{gpu_idx}"
# Assign decoder layers
for i in decoder_layer_indices:
gpu_idx = min((i + num_backbone_layers) // layers_per_gpu, num_gpus - 1)
device_map[f"decoder.layers.{i}"] = f"cuda:{gpu_idx}"
elif strategy == "sequential":
# Fill each GPU sequentially
# Backbone layers on first GPU(s)
backbone_per_gpu = max(1, num_backbone_layers // ((num_gpus + 1) // 2))
for i in backbone_layer_indices:
gpu_idx = min(i // backbone_per_gpu, (num_gpus + 1) // 2 - 1)
device_map[f"backbone.layers.{i}"] = f"cuda:{gpu_idx}"
# Decoder layers on remaining GPU(s)
decoder_per_gpu = max(1, num_decoder_layers // (num_gpus - (num_gpus + 1) // 2 + 1))
for i in decoder_layer_indices:
gpu_idx = min(i // decoder_per_gpu + (num_gpus + 1) // 2 - 1, num_gpus - 1)
device_map[f"decoder.layers.{i}"] = f"cuda:{gpu_idx}"
# Assign embeddings and other components
device_map["text_embeddings"] = "cuda:0"
device_map["audio_embeddings"] = "cuda:0"
device_map["projection"] = "cuda:0"
device_map["codebook0_head"] = "cuda:0"
device_map["audio_head"] = "cuda:0"
# Load state dict with device mapping
model.load_state_dict(state_dict)
# Move model parts to assigned devices
for name, device in device_map.items():
if "backbone.layers" in name:
layer_idx = int(name.split('.')[-1])
if hasattr(model.backbone, 'layers') and layer_idx < len(model.backbone.layers):
model.backbone.layers[layer_idx] = model.backbone.layers[layer_idx].to(device)
elif "decoder.layers" in name:
layer_idx = int(name.split('.')[-1])
if hasattr(model.decoder, 'layers') and layer_idx < len(model.decoder.layers):
model.decoder.layers[layer_idx] = model.decoder.layers[layer_idx].to(device)
elif hasattr(model, name):
setattr(model, name, getattr(model, name).to(device))
logger.info(f"Model distributed across GPUs with {strategy} strategy")
return model
def load_csm_1b(ckpt_path: str = "ckpt.pt", device: str = "cuda", device_map: str = None) -> Generator:
"""Load CSM-1B model and create generator with performance optimizations.
Args:
ckpt_path: Path to model checkpoint
device: Device to load model on ('cuda', 'cpu', or specific CUDA device)
device_map: Optional device mapping strategy ('auto', 'balanced', 'sequential', or None)
Returns:
Generator instance with optimized settings
"""
try:
# Import models module for CSM
from app.torchtune_models import Model, ModelArgs
# Create model
model_args = ModelArgs(
backbone_flavor="llama-1B",
decoder_flavor="llama-100M",
text_vocab_size=128256,
audio_vocab_size=2051,
audio_num_codebooks=32,
)
# Load model
logger.info(f"Loading CSM-1B model from {ckpt_path} with device={device}, device_map={device_map}")
# Check for CUDA availability
cuda_available = device == "cuda" and torch.cuda.is_available()
# Set up torch for optimized inference
if cuda_available:
# Check if we should enable TF32 (faster but slightly less precise)
enable_tf32 = os.environ.get("ENABLE_TF32", "true").lower() == "true"
if enable_tf32:
logger.info("Enabling TF32 for faster matrix multiplications")
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
# Check for available precision modes
use_bfloat16 = torch.cuda.is_bf16_supported()
use_float16 = not use_bfloat16 and torch.cuda.is_available() # Fallback to float16
if use_bfloat16:
dtype = torch.bfloat16
logger.info("Using bfloat16 precision for faster inference")
elif use_float16:
dtype = torch.float16
logger.info("Using float16 precision for faster inference")
else:
dtype = torch.float32
logger.info("Using float32 precision (mixed precision not available)")
# Enable Flash Attention if available
try:
import flash_attn
if os.environ.get("ENABLE_FLASH_ATTN", "true").lower() == "true":
logger.info("Flash Attention detected - enabling for faster attention")
os.environ["PYTORCH_FLASH_ATTENTION_ENABLED"] = "1"
except ImportError:
logger.info("Flash Attention not available (install flash-attn for faster inference)")
else:
# CPU-only mode
dtype = torch.float32
logger.info("Using CPU mode with float32 precision")
# Check for quantization
enable_quantization = os.environ.get("ENABLE_QUANTIZATION", "false").lower() == "true"
is_quantized = False
# Check for multi-GPU setup
if device_map and torch.cuda.device_count() > 1:
logger.info(f"Using device_map={device_map} across {torch.cuda.device_count()} GPUs")
# Create model with device map
model = Model(model_args)
# Load state dict
state_dict = torch.load(ckpt_path, map_location='cpu')
# Try quantization before device mapping if enabled
if enable_quantization and cuda_available:
try:
from bitsandbytes.nn import Linear8bitLt
def replace_with_8bit(model):
"""Replace linear layers with 8-bit quantized versions"""
for name, module in model.named_modules():
if isinstance(module, torch.nn.Linear) and module.out_features > 256:
parent_name = name.rsplit('.', 1)[0] if '.' in name else ''
parent = model
if parent_name:
for attr in parent_name.split('.'):
parent = getattr(parent, attr)
child_name = name.rsplit('.', 1)[1] if '.' in name else name
setattr(parent, child_name, Linear8bitLt.from_float(module))
return model
logger.info("Applying 8-bit quantization to linear layers")
model = replace_with_8bit(model)
is_quantized = True
except ImportError:
logger.warning("bitsandbytes not available, skipping quantization")
# Apply device mapping
if device_map == "auto":
# Use accelerate for automatic device mapping
try:
from accelerate import init_empty_weights, load_checkpoint_and_dispatch
# Initialize empty model
with init_empty_weights():
empty_model = Model(model_args)
# Load and dispatch model across GPUs
model = load_checkpoint_and_dispatch(
empty_model,
ckpt_path,
device_map="auto",
no_split_module_classes=["TransformerLayer"],
# Offload CPU if very large model
offload_folder="offload" if os.environ.get("OFFLOAD_TO_CPU", "false").lower() == "true" else None
)
logger.info("Model loaded with automatic device mapping")
except ImportError:
logger.warning("accelerate package not found, falling back to manual device mapping")
model = _manual_device_map(model, state_dict, "balanced")
except Exception as mapping_error:
logger.error(f"Auto device mapping failed: {mapping_error}, falling back to manual")
model = _manual_device_map(model, state_dict, "balanced")
else:
# Manual device mapping
model = _manual_device_map(model, state_dict, device_map or "balanced")
else:
# Single GPU or CPU setup
# Try quantization before loading if enabled (GPU only)
if enable_quantization and cuda_available and not is_quantized:
try:
# First load to CPU for quantization
model = Model(model_args).to("cpu")
state_dict = torch.load(ckpt_path, map_location="cpu")
model.load_state_dict(state_dict)
from bitsandbytes.nn import Linear8bitLt
def replace_with_8bit(model):
"""Replace linear layers with 8-bit quantized versions"""
for name, module in model.named_modules():
if isinstance(module, torch.nn.Linear) and module.out_features > 256:
parent_name = name.rsplit('.', 1)[0] if '.' in name else ''
parent = model
if parent_name:
for attr in parent_name.split('.'):
parent = getattr(parent, attr)
child_name = name.rsplit('.', 1)[1] if '.' in name else name
setattr(parent, child_name, Linear8bitLt.from_float(module))
return model
logger.info("Applying 8-bit quantization to linear layers")
model = replace_with_8bit(model)
model = model.to(device=device)
is_quantized = True
except ImportError:
logger.warning("bitsandbytes not available, loading without quantization")
# Load the standard way
model = Model(model_args).to(device=device, dtype=dtype)
state_dict = torch.load(ckpt_path, map_location=device)
model.load_state_dict(state_dict)
except Exception as quant_error:
logger.error(f"Quantization failed: {quant_error}, loading without quantization")
# Load the standard way
model = Model(model_args).to(device=device, dtype=dtype)
state_dict = torch.load(ckpt_path, map_location=device)
model.load_state_dict(state_dict)
else:
# Standard load without quantization
model = Model(model_args).to(device=device, dtype=dtype)
state_dict = torch.load(ckpt_path, map_location=device)
model.load_state_dict(state_dict)
# Apply torch.compile if available (PyTorch 2.0+)
compile_mode = os.environ.get("TORCH_COMPILE_MODE", "none")
if hasattr(torch, 'compile') and compile_mode != "none" and cuda_available:
try:
logger.info(f"Using torch.compile with mode '{compile_mode}' for faster inference")
if compile_mode == "default":
model = torch.compile(model)
else:
model = torch.compile(model, mode=compile_mode)
except Exception as compile_error:
logger.warning(f"Torch compile failed (requires PyTorch 2.0+): {compile_error}")
# Try to optimize CUDA graphs for faster inference (advanced)
use_cuda_graphs = os.environ.get("USE_CUDA_GRAPHS", "false").lower() == "true"
if use_cuda_graphs and cuda_available and hasattr(torch.cuda, 'CUDAGraph'):
try:
logger.info("Setting up CUDA graphs for repeated inference patterns")
# This requires custom integration inside the model's forward method
# Just flagging that CUDA graphs should be used
model.use_cuda_graphs = True
except Exception as cuda_graph_error:
logger.warning(f"CUDA graphs setup failed: {cuda_graph_error}")
model.use_cuda_graphs = False
# Set optimal settings for CUDA context
if cuda_available:
# Set benchmark mode for hardware-specific optimizations
torch.backends.cudnn.benchmark = True
# Clean up CUDA cache before creating generator
torch.cuda.empty_cache()
# Ensure all CUDA work is completed to avoid launch delays
torch.cuda.synchronize()
# Create generator
logger.info("Creating generator with optimized settings")
generator = Generator(model)
# Log memory usage if on CUDA
if cuda_available:
memory_allocated = torch.cuda.memory_allocated() / (1024**3)
memory_reserved = torch.cuda.memory_reserved() / (1024**3)
logger.info(f"Model loaded, CUDA memory: {memory_allocated:.2f}GB allocated, {memory_reserved:.2f}GB reserved")
logger.info(f"Generator created successfully: precision={dtype}, quantized={is_quantized}")
return generator
except Exception as e:
logger.error(f"Failed to load CSM-1B model: {e}")
import traceback
logger.error(traceback.format_exc())
raise |