Spaces:
Running
Running
File size: 43,659 Bytes
8c93973 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 |
"""
BlockSwap Module for SeedVR2
This module implements dynamic block swapping between GPU and CPU memory
to enable running large models on limited VRAM systems.
Key Features:
- Dynamic transformer block offloading during inference
- Non-blocking GPU transfers for optimal performance
- RoPE computation fallback to CPU on OOM
- Minimal performance overhead with intelligent caching
- I/O component offloading for maximum memory savings
"""
import time
import types
import torch
import weakref
from typing import Dict, Any, List, Optional
from .memory_manager import clear_memory
from .compatibility import call_rope_with_stability
from ..common.distributed import get_device
def is_blockswap_enabled(config: Optional[Dict[str, Any]]) -> bool:
"""
Check if BlockSwap configuration indicates BlockSwap should be enabled.
BlockSwap is enabled if either blocks_to_swap > 0 OR swap_io_components is True.
This is the authoritative function for determining BlockSwap status from configuration.
Args:
config: BlockSwap configuration dictionary with optional keys:
- blocks_to_swap: Number of blocks to offload (0 = disabled)
- swap_io_components: Whether to offload I/O components
Returns:
True if BlockSwap should be active, False otherwise
"""
if not config:
return False
blocks_to_swap = config.get("blocks_to_swap", 0)
swap_io_components = config.get("swap_io_components", False)
return blocks_to_swap > 0 or swap_io_components
def validate_blockswap_config(
block_swap_config: Optional[Dict[str, Any]],
dit_device: 'torch.device',
dit_offload_device: Optional['torch.device'],
debug: 'Debug'
) -> Optional[Dict[str, Any]]:
"""
Validate and potentially modify BlockSwap configuration.
Performs platform-specific validation and configuration adjustment:
- On macOS (MPS): Auto-disables BlockSwap since unified memory makes it meaningless
- On other platforms: Validates that offload_device is properly configured
This is the single authoritative validation point for BlockSwap configuration,
called early in configure_runner() before any model loading.
Args:
block_swap_config: BlockSwap configuration dictionary (may be None)
dit_device: Target device for DiT model inference
dit_offload_device: Device for offloading DiT blocks (may be None)
debug: Debug instance for logging warnings/errors
Returns:
Validated/modified block_swap_config (may be None or modified copy)
Raises:
ValueError: If BlockSwap is enabled but offload_device is invalid (non-MPS only)
"""
if not is_blockswap_enabled(block_swap_config):
return block_swap_config
blocks_to_swap = block_swap_config.get("blocks_to_swap", 0)
swap_io_components = block_swap_config.get("swap_io_components", False)
# Check for macOS unified memory - BlockSwap is meaningless there
if dit_device.type == "mps":
debug.log(
f"BlockSwap disabled: macOS uses unified memory (no separate VRAM/RAM). "
f"Ignoring blocks_to_swap={blocks_to_swap}, swap_io_components={swap_io_components}",
level="WARNING", category="blockswap", force=True
)
# Return disabled config
return {
**block_swap_config,
"blocks_to_swap": 0,
"swap_io_components": False
}
# Validate offload_device is set and different from dit_device
offload_device_valid = (
dit_offload_device is not None and
str(dit_offload_device) != str(dit_device)
)
if not offload_device_valid:
config_details = []
if blocks_to_swap > 0:
config_details.append(f"blocks_to_swap={blocks_to_swap}")
if swap_io_components:
config_details.append("swap_io_components=True")
offload_str = str(dit_offload_device) if dit_offload_device else "none"
raise ValueError(
f"BlockSwap enabled ({', '.join(config_details)}) but dit_offload_device is invalid. "
f"Current: device='{dit_device}', dit_offload_device='{offload_str}'. "
f"BlockSwap requires offload_device on the DiT Model to be set and different from device. "
f"Set --dit_offload_device cpu or disable BlockSwap."
)
return block_swap_config
# Timing helpers marked to skip torch.compile tracing
# These functions are excluded from Dynamo's graph tracing to avoid warnings
# about non-traceable builtins like time.time(), but they still execute normally
@torch._dynamo.disable
def _get_swap_start_time(debug, enabled: bool) -> Optional[float]:
"""Get start time for swap operation if debug is enabled."""
return time.time() if debug and enabled else None
@torch._dynamo.disable
def _log_swap_timing(debug, t_start: Optional[float], component_id, component_type: str) -> None:
"""Log swap timing if start time was captured."""
if debug and t_start is not None:
debug.log_swap_time(
component_id=component_id,
duration=time.time() - t_start,
component_type=component_type
)
def get_module_memory_mb(module: torch.nn.Module) -> float:
"""
Calculate memory usage of a module in MB.
Args:
module: PyTorch module to measure
Returns:
Memory usage in megabytes
"""
total_bytes = sum(
param.nelement() * param.element_size()
for param in module.parameters()
if param.data is not None
)
return total_bytes / (1024 * 1024)
def apply_block_swap_to_dit(
runner: 'VideoDiffusionInfer',
block_swap_config: Dict[str, Any],
debug: 'Debug'
) -> None:
"""
Apply block swapping configuration to a DiT model with OOM protection.
This is the main entry point for configuring block swapping on a model.
Handles block selection, I/O component offloading, device placement, and
forward method wrapping for dynamic memory management.
Args:
runner: VideoDiffusionInfer instance containing the model
block_swap_config: Configuration dictionary with keys:
- blocks_to_swap: Number of blocks to swap (from the start)
- swap_io_components: Whether to offload I/O components
- enable_debug: Whether to enable debug logging
- offload_device: Device to offload to (default: 'cpu')
debug: Debug instance for logging (required)
"""
# Early return if BlockSwap not enabled
if not is_blockswap_enabled(block_swap_config):
return
blocks_to_swap = block_swap_config.get("blocks_to_swap", 0)
swap_io_components = block_swap_config.get("swap_io_components", False)
# Early return only if both block swap and I/O swap are disabled
if blocks_to_swap <= 0 and not swap_io_components:
return
if debug is None:
if hasattr(runner, 'debug') and runner.debug is not None:
debug = runner.debug
else:
raise ValueError("Debug instance must be provided to apply_block_swap_to_dit")
debug.start_timer("apply_blockswap")
# Get the actual model (handle CompatibleDiT wrapper)
model = runner.dit
if hasattr(model, "dit_model"):
model = model.dit_model
# Determine devices
if hasattr(runner, '_dit_device'):
device = runner._dit_device
else:
device = get_device()
offload_device = block_swap_config.get("offload_device", torch.device('cpu'))
# Validate model structure
if not hasattr(model, "blocks"):
debug.log("Model doesn't have 'blocks' attribute for BlockSwap", level="ERROR", category="blockswap", force=True)
return
total_blocks = len(model.blocks)
# Clamp blocks_to_swap to available blocks BEFORE logging
effective_blocks = min(blocks_to_swap, total_blocks) if blocks_to_swap > 0 else 0
# Log configuration clearly based on what's enabled
block_text = "block" if effective_blocks <= 1 else "blocks"
if effective_blocks > 0 and swap_io_components:
debug.log(f"BlockSwap: {effective_blocks}/{total_blocks} transformer {block_text} + I/O components offloaded to {str(offload_device).upper()}", category="blockswap", force=True)
elif effective_blocks > 0:
debug.log(f"BlockSwap: {effective_blocks}/{total_blocks} transformer {block_text} offloaded to {str(offload_device).upper()}", category="blockswap", force=True)
elif swap_io_components:
debug.log(f"BlockSwap: I/O components offloaded to {str(offload_device).upper()} (0/{total_blocks} blocks swapped)", category="blockswap", force=True)
# Configure model with blockswap attributes
if blocks_to_swap > 0:
model.blocks_to_swap = effective_blocks - 1 # Convert to 0-indexed
else:
# No block swapping, set to -1 so no blocks match the swap condition
model.blocks_to_swap = -1
model.main_device = device
model.offload_device = offload_device
# Configure I/O components
io_config = _configure_io_components(model, device, offload_device,
swap_io_components, debug)
memory_stats = _configure_blocks(model, device, offload_device, debug)
memory_stats['io_components'] = io_config['components']
memory_stats['io_memory_mb'] = io_config['memory_mb']
memory_stats['gpu_components'] = io_config['gpu_components']
memory_stats['io_gpu_memory_mb'] = io_config['gpu_memory_mb']
# Log memory summary
_log_memory_summary(memory_stats, offload_device, device, swap_io_components,
debug)
# Initialize Nunchaku-style async management object
if blocks_to_swap > 0:
# normalize device objects
if isinstance(device, str):
device = torch.device(device)
model._swap_stream = torch.cuda.Stream(device=device)
model._block_ready_events = {}
# Preload first swapped block to seed pipeline (non-blocking on swap_stream)
try:
first_idx = 0
if first_idx <= model.blocks_to_swap:
with torch.cuda.stream(model._swap_stream):
model.blocks[first_idx].to(device, non_blocking=True)
ev = torch.cuda.Event(blocking=False)
ev.record(model._swap_stream) # record on swap_stream -> event gets device-bound here
model._block_ready_events[first_idx] = ev
except Exception as e:
debug.log(f"Failed to initialize swap-stream prefetch: {e}", level="WARNING", category="blockswap", force=True)
# Wrap block forward methods for dynamic swapping (only if blocks_to_swap > 0)
if blocks_to_swap > 0:
for b, block in enumerate(model.blocks):
if b <= model.blocks_to_swap:
_wrap_block_forward(block, b, model, debug)
# Patch RoPE modules for robust error handling
_patch_rope_for_blockswap(model, debug)
# Mark BlockSwap as active
runner._blockswap_active = True
# Store configuration for debugging and cleanup
model._block_swap_config = {
"blocks_swapped": blocks_to_swap,
"swap_io_components": swap_io_components,
"total_blocks": total_blocks,
"offload_device": offload_device,
"main_device": device,
"offload_memory": memory_stats['offload_memory'],
"main_memory": memory_stats['main_memory']
}
# Protect model from being moved entirely
_protect_model_from_move(model, runner, debug)
debug.log("BlockSwap configuration complete", category="success")
debug.end_timer("apply_blockswap", "BlockSwap configuration application")
def _configure_io_components(
model: torch.nn.Module,
device: torch.device,
offload_device: torch.device,
swap_io_components: bool,
debug: 'Debug'
) -> Dict[str, Any]:
"""
Configure I/O component placement and wrapping with memory tracking.
Handles all non-block modules (embeddings, normalization layers, etc.) by
either keeping them on GPU or offloading them with dynamic swapping wrappers.
Args:
model: DiT model containing named children to configure
device: Main computation device (typically GPU)
offload_device: Device for offloaded components (typically CPU)
swap_io_components: If True, offload I/O components with dynamic swapping
debug: Debug instance for logging (required)
Returns:
Dictionary containing:
- components: List of offloaded component names
- memory_mb: Total memory of offloaded components in MB
- gpu_components: List of components remaining on GPU
- gpu_memory_mb: Total memory of GPU components in MB
"""
io_components_offloaded = []
io_components_on_gpu = []
io_memory_mb = 0.0
io_gpu_memory_mb = 0.0
# Check for pin memory condition
use_pin_memory = (offload_device == "cpu") if isinstance(offload_device, str) else (offload_device.type == "cpu")
# Handle I/O modules with dynamic swapping
for name, module in model.named_children():
if name != "blocks":
module_memory = get_module_memory_mb(module)
if swap_io_components:
module.to(offload_device)
# Enable Pin Memory for I/O components
if use_pin_memory:
for p in module.parameters():
if not p.is_pinned():
p.data = p.data.pin_memory()
for buf in module.buffers():
if not buf.is_pinned():
buf.data = buf.data.pin_memory()
_wrap_io_forward(module, name, model, debug)
io_components_offloaded.append(name)
io_memory_mb += module_memory
debug.log(f"{name} → {str(offload_device).upper()} ({module_memory:.2f}MB, dynamic swapping)", category="blockswap", indent_level=1)
else:
module.to(device)
io_components_on_gpu.append(name)
io_gpu_memory_mb += module_memory
debug.log(f"{name} → {str(device).upper()} ({module_memory:.2f}MB)", category="blockswap", indent_level=1)
return {
'components': io_components_offloaded,
'memory_mb': io_memory_mb,
'gpu_components': io_components_on_gpu,
'gpu_memory_mb': io_gpu_memory_mb
}
def _configure_blocks(
model: torch.nn.Module,
device: torch.device,
offload_device: torch.device,
debug: 'Debug'
) -> Dict[str, float]:
"""
Configure transformer block placement and calculate memory statistics.
Moves blocks to their designated devices based on model.blocks_to_swap
attribute. Blocks with index <= blocks_to_swap go to offload device,
others stay on main device.
Args:
model: DiT model with blocks attribute and blocks_to_swap configured
device: Main computation device for non-swapped blocks
offload_device: Device for swapped blocks
debug: Debug instance for logging (required)
Returns:
Dictionary containing:
- offload_memory: Total memory of offloaded blocks in MB
- main_memory: Total memory of blocks on main device in MB
- io_components: Empty list (populated by caller)
"""
total_offload_memory = 0.0
total_main_memory = 0.0
# Check if we should pin memory (if offloading to CPU)
# Nunchaku uses pinned memory for faster async transfers
use_pin_memory = (offload_device == "cpu") if isinstance(offload_device, str) else (offload_device.type == "cpu")
# Move blocks based on swap configuration
for b, block in enumerate(model.blocks):
block_memory = get_module_memory_mb(block)
if b > model.blocks_to_swap:
block.to(device)
total_main_memory += block_memory
else:
block.to(offload_device, non_blocking=False)
total_offload_memory += block_memory
# Enable Pin Memory optimization for CPU Offload transfer speed
if use_pin_memory:
for p in block.parameters():
if not p.is_pinned():
p.data = p.data.pin_memory()
for buf in block.buffers():
if not buf.is_pinned():
buf.data = buf.data.pin_memory()
# Ensure all buffers match their containing module's device
for b, block in enumerate(model.blocks):
target_device = device if b > model.blocks_to_swap else offload_device
for name, buffer in block.named_buffers():
if buffer.device != torch.device(target_device):
# Apply pinning if needed
if use_pin_memory and target_device.type == "cpu" and not buffer.is_pinned():
buffer.data = buffer.data.pin_memory()
buffer.data = buffer.data.to(target_device, non_blocking=False)
return {
"offload_memory": total_offload_memory,
"main_memory": total_main_memory,
"io_components": [] # Will be populated by caller
}
def _log_memory_summary(
memory_stats: Dict[str, float],
offload_device: torch.device,
device: torch.device,
swap_io_components: bool,
debug: 'Debug'
) -> None:
"""
Log comprehensive memory usage summary for BlockSwap configuration.
Displays detailed breakdown of memory distribution across devices,
including transformer blocks and I/O components.
Args:
memory_stats: Dictionary containing:
- offload_memory: Memory offloaded from blocks (MB)
- main_memory: Memory remaining on main device (MB)
- io_memory_mb: Memory from offloaded I/O components (MB)
- io_gpu_memory_mb: Memory from I/O components on GPU (MB)
offload_device: Device used for offloading
device: Main computation device
swap_io_components: Whether I/O components are being swapped
debug: Debug instance for logging (required)
"""
debug.log("BlockSwap memory configuration:", category="blockswap")
# Log transformer blocks memory
blocks_offloaded = memory_stats['offload_memory']
blocks_on_gpu = memory_stats['main_memory']
offload_str = str(offload_device)
device_str = str(device)
if blocks_on_gpu == 0:
debug.log(f"Transformer blocks: {blocks_offloaded:.2f}MB on {offload_str} (dynamic swapping)", category="blockswap", indent_level=1)
else:
debug.log(f"Transformer blocks: {blocks_on_gpu:.2f}MB on {device_str}, {blocks_offloaded:.2f}MB on {offload_str}", category="blockswap", indent_level=1)
# Always log I/O components (whether swapping or not)
io_memory = memory_stats.get('io_memory_mb', 0.0)
io_gpu_memory = memory_stats.get('io_gpu_memory_mb', 0.0)
if swap_io_components and io_memory > 0:
io_components = memory_stats.get('io_components', [])
debug.log(f"I/O components: {io_memory:.2f}MB on {offload_str} (dynamic swapping)", category="blockswap", indent_level=1)
debug.log(f"{', '.join(io_components)}", category="blockswap", indent_level=2)
elif io_gpu_memory > 0:
io_gpu_components = memory_stats.get('gpu_components', [])
debug.log(f"I/O components: {io_gpu_memory:.2f}MB on {device_str}", category="blockswap", indent_level=1)
debug.log(f"{', '.join(io_gpu_components)}", category="blockswap", indent_level=2)
# Log total VRAM savings
total_offloaded = blocks_offloaded + (io_memory if swap_io_components else 0)
if total_offloaded > 0:
debug.log(f"Total VRAM saved: {total_offloaded:.2f}MB (~{total_offloaded/1024:.2f}GB)", category="blockswap", indent_level=1)
def _wrap_block_forward(
block: torch.nn.Module,
block_idx: int,
model: torch.nn.Module,
debug: 'Debug'
) -> None:
"""
Wrap individual transformer block forward for dynamic device swapping.
Implements Nunchaku-style pipelining: Prefetch Next -> Compute Current -> Offload Current.
https://github.com/nunchaku-tech/nunchaku/blob/main/nunchaku/models/utils.py
Creates a wrapped forward method that automatically:
1. Moves block to GPU before computation
2. Executes original forward pass
3. Moves block back to offload device after computation
4. Logs timing and manages memory pressure
Uses weak references to prevent memory leaks from closure retention.
Args:
block: Individual transformer block to wrap
block_idx: Index of this block in model.blocks
model: Parent DiT model (used for device references)
debug: Debug instance for logging (required)
"""
if hasattr(block, '_original_forward'):
return # Already wrapped
# Store original forward method
original_forward = block.forward
# Create weak references
model_ref = weakref.ref(model)
debug_ref = weakref.ref(debug) if debug is not None else (lambda: None)
# Store block_idx on the block itself to avoid closure issues
block._block_idx = block_idx
def wrapped_forward(self, *args, **kwargs):
# Retrieve weak references
model = model_ref()
debug = debug_ref()
if not model:
# Model has been garbage collected, fall back to original
return original_forward(*args, **kwargs)
# Check if block swap is active for this block
if hasattr(model, 'blocks_to_swap') and self._block_idx <= model.blocks_to_swap:
# Use dynamo-disabled helper to get start time (avoids compilation warnings)
t_start = _get_swap_start_time(debug, debug.enabled if debug else False)
# Only move to GPU if necessary
current_device = next(self.parameters()).device
target_device = torch.device(model.main_device)
# 1. Ensure CURRENT block is ready on GPU
# Check if we have a prefetch event waiting
if hasattr(model, '_block_ready_events') and self._block_idx in model._block_ready_events:
# Wait for the swap stream to finish moving this block
torch.cuda.current_stream().wait_event(model._block_ready_events[self._block_idx])
# Cleanup event
del model._block_ready_events[self._block_idx]
elif current_device != target_device:
# Fallback: First block or missed prefetch, move synchronously (but non-blocking)
debug.log(f"[blockswap] Block {self._block_idx} missing prefetch event, moving synchronously", level="WARNING", category="blockswap", force=True)
self.to(model.main_device, non_blocking=True)
# 2. Trigger Prefetch for NEXT block (Pipelining)
# Nunchaku logic: Start moving i+1 while i is computing
next_idx = self._block_idx + 1
if next_idx <= model.blocks_to_swap:
next_block = model.blocks[next_idx]
# Use the dedicated swap stream
with torch.cuda.stream(model._swap_stream):
next_block.to(model.main_device, non_blocking=True)
# Record event so next iteration knows when to wait
event = torch.cuda.Event(blocking=False)
event.record(model._swap_stream)
model._block_ready_events[next_idx] = event
# 3. Execute forward pass (Compute)
# This runs on the default stream, overlapping with the prefetch above
output = original_forward(*args, **kwargs)
# 4. Offload CURRENT block (Async)
# We record an event on compute stream to ensure we don't move data while it's being used
compute_done_event = torch.cuda.Event(blocking=False)
compute_done_event.record(torch.cuda.current_stream())
with torch.cuda.stream(model._swap_stream):
# Wait for compute to finish before moving memory out
model._swap_stream.wait_event(compute_done_event)
# Move back to offload device
self.to(model.offload_device, non_blocking=True)
# Use dynamo-disabled helper to log timing (avoids compilation warnings)
_log_swap_timing(debug, t_start, self._block_idx, "block (pipelined)")
# Only clear cache under memory pressure
clear_memory(debug=debug, deep=False, force=False, timer_name="wrap_block_forward")
else:
output = original_forward(*args, **kwargs)
return output
# Bind the wrapped function as a method to the block
block.forward = types.MethodType(wrapped_forward, block)
# Store reference to original forward for cleanup
block._original_forward = original_forward
def _wrap_io_forward(
module: torch.nn.Module,
module_name: str,
model: torch.nn.Module,
debug: 'Debug'
) -> None:
"""
Wrap I/O component forward for dynamic device swapping.
Similar to _wrap_block_forward but for I/O components (embeddings,
normalization layers, etc.). Handles swapping between GPU and CPU
during forward passes.
Uses weak references to prevent circular dependencies and memory leaks.
Args:
module: I/O component module to wrap
module_name: Name identifier for logging (e.g., 'x_embedder')
model: Parent DiT model (used for device references)
debug: Debug instance for logging (required)
"""
if hasattr(module, '_is_io_wrapped') and module._is_io_wrapped:
debug.log(f"Reusing existing I/O wrapper for {module_name}", category="reuse")
return # Already wrapped
# Store original forward method
original_forward = module.forward
# Create weak references
model_ref = weakref.ref(model)
debug_ref = weakref.ref(debug) if debug else lambda: None
# Store module name on the module itself
module._module_name = module_name
module._original_forward = original_forward
def wrapped_io_forward(self, *args, **kwargs):
# Retrieve weak references
model = model_ref()
debug = debug_ref()
if not model:
# Model has been garbage collected, fall back to original
return self._original_forward(*args, **kwargs)
# Use dynamo-disabled helper to get start time (avoids compilation warnings)
t_start = _get_swap_start_time(debug, debug.enabled if debug else False)
# Check current device to avoid unnecessary moves
current_device = next(self.parameters()).device
target_device = torch.device(model.main_device)
# Move to GPU for computation if needed
if current_device != target_device:
self.to(model.main_device, non_blocking=False)
# Execute forward pass
output = self._original_forward(*args, **kwargs)
# Move back to offload device
self.to(model.offload_device, non_blocking=False)
# Use dynamo-disabled helper to log timing (avoids compilation warnings)
_log_swap_timing(debug, t_start, self._module_name, "I/O")
# Only clear cache under memory pressure
clear_memory(debug=debug, deep=False, force=False, timer_name="wrap_block_forward")
return output
# Bind as a method
module.forward = types.MethodType(wrapped_io_forward, module)
module._is_io_wrapped = True
# Store module reference for restoration
if not hasattr(model, '_io_swappers'):
model._io_swappers = []
model._io_swappers.append((module, module_name))
def _patch_rope_for_blockswap(
model: torch.nn.Module,
debug: 'Debug'
) -> None:
"""
Patch RoPE (Rotary Position Embedding) modules for device-aware fallback.
Adds CPU fallback logic to RoPE modules to handle device mismatch errors
that can occur during BlockSwap operations. Complements the stability
wrapper from compatibility.py with device-specific error handling.
Args:
model: DiT model containing RoPE modules to patch
debug: Debug instance for logging (required)
"""
rope_patches = []
for name, module in model.named_modules():
if "rope" in name.lower() and hasattr(module, "get_axial_freqs"):
# Skip if already wrapped by blockswap
if hasattr(module, '_blockswap_wrapped') and module._blockswap_wrapped:
continue
# Get current method (might be stability-wrapped)
current_method = module.get_axial_freqs
# Create device-aware wrapper with proper closure handling
def make_device_aware_wrapper(module_name, current_fn):
def device_aware_rope_wrapper(self, *args, **kwargs):
try:
# Try current method (original or stability-wrapped)
return current_fn(*args, **kwargs)
except (RuntimeError, torch.cuda.OutOfMemoryError) as e:
error_msg = str(e).lower()
# Only handle device/memory specific errors
if any(x in error_msg for x in ["device", "memory", "allocation"]):
debug.log(f"RoPE OOM for {module_name}", level="WARNING", category="rope", force=True)
debug.log(f"Clearing RoPE cache and retrying", category="info", force=True)
# Get current device from parameters
try:
current_device = next(self.parameters()).device
except StopIteration:
# Fallback: use model's main_device if BlockSwap has set it, else use offload_device
if hasattr(model, 'main_device'):
current_device = torch.device(model.main_device)
elif hasattr(model, 'offload_device'):
current_device = torch.device(model.offload_device)
# Try clearing cache first (non-invasive fix)
if hasattr(current_fn, 'cache_clear'):
current_fn.cache_clear()
try:
# Retry on same device after clearing cache
return current_fn(*args, **kwargs)
except Exception as retry_error:
# Cache clear wasn't enough, need more drastic measures
debug.log(f"Cache clear insufficient for {module_name}, falling back to CPU", level="WARNING", category="rope", force=True)
# Fallback to CPU computation with stability
self.cpu()
try:
# Use call_rope_with_stability for CPU computation
# This ensures cache is cleared and autocast disabled
original_fn = getattr(self, '_original_get_axial_freqs', current_fn)
result = call_rope_with_stability(original_fn, *args, **kwargs)
# Move module back to original device
self.to(current_device)
# Move result to appropriate device if it's a tensor
if hasattr(result, 'to'):
target_device = args[0].device if len(args) > 0 and hasattr(args[0], 'device') else current_device
return result.to(target_device)
return result
except Exception as cpu_error:
# Always restore device even on error
self.to(current_device)
raise cpu_error
else:
# Not a device error, let it bubble up
raise
return device_aware_rope_wrapper
# Apply wrapper
module.get_axial_freqs = types.MethodType(
make_device_aware_wrapper(name, current_method),
module
)
module._blockswap_wrapped = True
# Store for cleanup (use original or previously stored)
original_method = getattr(module, '_original_get_axial_freqs', current_method)
rope_patches.append((module, original_method))
if rope_patches:
model._rope_patches = rope_patches
debug.log(f"Patched {len(rope_patches)} RoPE modules with device handling", category="success")
def _protect_model_from_move(
model: torch.nn.Module,
runner: 'VideoDiffusionInfer',
debug: 'Debug'
) -> None:
"""
Protect model from unintended full device movement during BlockSwap.
Wraps model.to() method to prevent other code from accidentally moving
the entire model to GPU, which would defeat BlockSwap's memory savings.
Allows movement only when explicitly bypassed via model flag.
Args:
model: DiT model to protect
runner: VideoDiffusionInfer instance (for active status check)
debug: Debug instance for logging (required)
"""
if not hasattr(model, '_original_to'):
# Store runner reference as weak reference to avoid circular refs
model._blockswap_runner_ref = weakref.ref(runner)
model._original_to = model.to
# Define the protected method without closures
def protected_model_to(self, device, *args, **kwargs):
# Check if protection is temporarily bypassed for offloading
# Flag is stored on model itself (not runner) to survive runner recreation
if getattr(self, "_blockswap_bypass_protection", False):
# Protection bypassed, allow movement
if hasattr(self, '_original_to'):
return self._original_to(device, *args, **kwargs)
# Get configured offload device directly from model
blockswap_offload_device = "cpu" # default
if hasattr(self, "_block_swap_config"):
blockswap_offload_device = self._block_swap_config.get("offload_device", "cpu")
# Check if BlockSwap is currently active via runner weak reference
runner_ref = getattr(self, '_blockswap_runner_ref', None)
blockswap_is_active = False
if runner_ref:
runner_obj = runner_ref()
if runner_obj and hasattr(runner_obj, "_blockswap_active"):
blockswap_is_active = runner_obj._blockswap_active
# Block attempts to move model away from configured offload device when active
if blockswap_is_active and str(device) != str(blockswap_offload_device):
# Get debug instance from runner if available
debug_instance = None
if runner_ref:
runner_obj = runner_ref()
if runner_obj and hasattr(runner_obj, 'debug'):
debug_instance = runner_obj.debug
if debug_instance:
debug_instance.log(
f"Blocked attempt to move BlockSwap model from {blockswap_offload_device} to {device}",
level="WARNING", category="blockswap", force=True
)
return self
# Allow movement (either bypass is enabled or target is offload device)
if hasattr(self, '_original_to'):
return self._original_to(device, *args, **kwargs)
else:
# Fallback - shouldn't happen
return super(type(self), self).to(device, *args, **kwargs)
# Bind as a method to the model instance
model.to = types.MethodType(protected_model_to, model)
def set_blockswap_bypass(runner, bypass: bool, debug):
"""
Set or unset bypass flag for BlockSwap protection.
Used for offloading to temporarily allow model movement.
Args:
runner: Runner instance with BlockSwap
bypass: True to bypass protection, False to enforce it
debug: Debug instance for logging
"""
if not hasattr(runner, "_blockswap_active") or not runner._blockswap_active:
return
# Get the actual model (handle CompatibleDiT wrapper)
model = runner.dit
if hasattr(model, "dit_model"):
model = model.dit_model
# Store on model so it survives runner recreation during caching
model._blockswap_bypass_protection = bypass
if bypass:
debug.log("BlockSwap protection disabled to allow model DiT offloading", category="success")
else:
debug.log("BlockSwap protection renabled to avoid accidentally offloading the entire DiT model", category="success")
def cleanup_blockswap(runner, keep_state_for_cache=False):
"""
Clean up BlockSwap configuration based on caching mode.
When caching (keep_state_for_cache=True):
- Keep all BlockSwap configuration intact
- Only mark as inactive for safety during non-inference operations
When not caching (keep_state_for_cache=False):
- Full cleanup of all BlockSwap state
Args:
runner: VideoDiffusionInfer instance to clean up
keep_state_for_cache: If True, preserve BlockSwap state for reuse
"""
# Get debug instance from runner
if not hasattr(runner, 'debug') or runner.debug is None:
raise ValueError("Debug instance must be available on runner for cleanup_blockswap")
debug = runner.debug
# Get the actual model (handle CompatibleDiT wrapper)
model = runner.dit
if hasattr(model, "dit_model"):
model = model.dit_model
# Check if there's any BlockSwap state to clean up (check both runner and model)
has_blockswap_state = (
hasattr(runner, "_blockswap_active") or
hasattr(model, "_block_swap_config") or
hasattr(model, "_blockswap_bypass_protection")
)
if not has_blockswap_state:
return
debug.log("Starting BlockSwap cleanup", category="cleanup")
if keep_state_for_cache:
# Minimal cleanup for caching - just mark as inactive and allow offloading
# Everything else stays intact for fast reactivation
if hasattr(runner, "_blockswap_active") and runner._blockswap_active:
if not getattr(model, "_blockswap_bypass_protection", False):
set_blockswap_bypass(runner=runner, bypass=True, debug=debug)
runner._blockswap_active = False
debug.log("BlockSwap deactivated for caching (configuration preserved)", category="success")
return
# Full cleanup when not caching
# Get the actual model (handle CompatibleDiT wrapper)
model = runner.dit
if hasattr(model, "dit_model"):
model = model.dit_model
# 1. Restore block forward methods
if hasattr(model, 'blocks'):
restored_count = 0
for block in model.blocks:
if hasattr(block, '_original_forward'):
block.forward = block._original_forward
delattr(block, '_original_forward')
restored_count += 1
# Clean up wrapper attributes
for attr in ['_block_idx', '_model_ref', '_debug_ref', '_blockswap_wrapped']:
if hasattr(block, attr):
delattr(block, attr)
if restored_count > 0:
debug.log(f"Restored {restored_count} block forward methods", category="success")
# 2. Restore RoPE patches
if hasattr(model, '_rope_patches'):
for module, original_method in model._rope_patches:
module.get_axial_freqs = original_method
# Clean up wrapper attributes
for attr in ['_rope_wrapped', '_original_get_axial_freqs']:
if hasattr(module, attr):
delattr(module, attr)
debug.log(f"Restored {len(model._rope_patches)} RoPE methods", category="success")
delattr(model, '_rope_patches')
# 3. Restore I/O component forward methods and move to offload device
if hasattr(model, '_io_swappers'):
for module, module_name in model._io_swappers:
if hasattr(module, '_original_forward'):
module.forward = module._original_forward
# Clean up wrapper attributes
for attr in ['_original_forward', '_model_ref', '_debug_ref',
'_module_name', '_is_io_wrapped']:
if hasattr(module, attr):
delattr(module, attr)
debug.log(f"Restored {len(model._io_swappers)} I/O components", category="success")
delattr(model, '_io_swappers')
# Move all IO components to offload device during full cleanup
if hasattr(model, 'offload_device'):
offload_device = model.offload_device
moved_count = 0
for name, module in model.named_children():
if name != "blocks":
module.to(offload_device)
moved_count += 1
if moved_count > 0:
debug.log(f"Moved {moved_count} IO components to offload device", category="success")
# 4. Restore original .to() method
if hasattr(model, '_original_to'):
model.to = model._original_to
delattr(model, '_original_to')
debug.log("Restored original .to() method", category="success")
# 5. Clean up BlockSwap-specific attributes
for attr in ['_blockswap_runner_ref', 'blocks_to_swap', 'main_device',
'offload_device']:
if hasattr(model, attr):
delattr(model, attr)
# 6. Clean up runner attributes
runner._blockswap_active = False
# Clean up pipelining resources on model (synchronize first)
if hasattr(model, '_swap_stream'):
try:
model._swap_stream.synchronize()
except Exception:
pass
for attr in ['_swap_stream', '_block_ready_events']:
if hasattr(model, attr):
delattr(model, attr)
# Remove all config attributes
for attr in ['_cached_blockswap_config', '_block_swap_config', '_blockswap_debug']:
if hasattr(runner, attr):
delattr(runner, attr)
debug.log("BlockSwap cleanup complete", category="success")
|