File size: 39,501 Bytes
131da64 |
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 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 |
"""
A collection of assorted utilities for deep learning with no required dependencies aside from PyTorch, NumPy and the Python stdlib.
TODO: Validate required Python version.
"""
import contextlib
import builtins
import functools
import glob
import hashlib
import io
import os
import pickle
import subprocess
import sys
import time
import traceback
from collections import defaultdict
from datetime import datetime
from functools import lru_cache, partial, wraps
from importlib import import_module
from importlib.util import find_spec
from io import BytesIO
from pathlib import Path
from typing import Any, Optional
from urllib.parse import urlparse
import inspect
import numpy as np
import torch
import torch.distributed as dist
from torch import Tensor
# TPUs:
# wait_device_ops
# https://github.com/pytorch/xla/blob/2766f9f756e8c8150ea3b7df98762c2f82d66e39/examples/debug/train_resnet_benchmark.py#L18
# From diffusers
import importlib
import importlib.metadata
# TODO: This doesn't work for all packages (`bs4`, `faiss`, etc.) Talk to Sylvain to see how to do with it better.
def _is_package_available(pkg_name: str, return_version: bool = False):
# Check if the package spec exists and grab its version to avoid importing a local directory
package_exists = importlib.util.find_spec(pkg_name) is not None
package_version = "N/A"
if package_exists:
try:
# Primary method to get the package version
package_version = importlib.metadata.version(pkg_name)
except importlib.metadata.PackageNotFoundError:
# Fallback method: Only for "torch" and versions containing "dev"
if pkg_name == "torch":
try:
package = importlib.import_module(pkg_name)
temp_version = getattr(package, "__version__", "N/A")
# Check if the version contains "dev"
if "dev" in temp_version:
package_version = temp_version
package_exists = True
else:
package_exists = False
except ImportError:
# If the package can't be imported, it's not available
package_exists = False
else:
# For packages other than "torch", don't attempt the fallback and set as not available
package_exists = False
if return_version:
return package_exists, package_version
else:
return package_exists
# https://github.com/huggingface/transformers/blob/main/src/transformers/utils/import_utils.py#L281
@lru_cache
def is_torch_cuda_available():
return torch.cuda.is_available()
@lru_cache
def is_torch_xla_available(check_is_tpu=False, check_is_gpu=False):
"""
Check if `torch_xla` is available. To train a native pytorch job in an environment with torch xla installed, set
the USE_TORCH_XLA to false.
"""
assert not (check_is_tpu and check_is_gpu), "The check_is_tpu and check_is_gpu cannot both be true."
_torch_xla_available, _torch_xla_version = _is_package_available("torch_xla", return_version=True)
if not _torch_xla_available:
return False
import torch_xla
if check_is_gpu:
return torch_xla.runtime.device_type() in ["GPU", "CUDA"]
elif check_is_tpu:
return torch_xla.runtime.device_type() == "TPU"
return True
def get_available_backend():
if is_torch_cuda_available():
backend = "cuda"
elif is_torch_xla_available():
backend = "xla"
else:
backend = "cpu"
return backend
if is_torch_xla_available():
import torch_xla.core.xla_model as xm
builtins.HAS_XLA_SPAWNED = False
def use_dist():
return dist.is_available() and dist.is_initialized()
def get_device():
return torch.device(f"cuda:{get_rank()}")
def get_tpu_devices():
import torch_xla.core.xla_model as xm
return xm.get_xla_supported_devices()
def get_num_nodes():
if is_torch_cuda_available():
rprint(f"Warning: get_num_nodes() is not supported for CUDA. Returning world size // num devices.")
return get_world_size() // get_num_devices()
if is_torch_xla_available():
from torch_xla._internal import tpu
return tpu.num_tpu_workers()
def get_num_devices():
"""
Number of physical "devices" on a single node. In CUDA, this is the number of GPUs. In XLA, this is the number of TPUs *chips* so, e.g., a v4 slice will always return 4 (even if part of a larger slice).
"""
if is_torch_cuda_available():
return torch.cuda.device_count()
else:
return 1
def get_world_size():
if is_torch_xla_available():
import torch_xla.runtime as xr
return xr.world_size()
elif use_dist():
return dist.get_world_size()
elif 'WORLD_SIZE' in os.environ:
return int(os.environ['WORLD_SIZE'])
elif is_torch_cuda_available():
return torch.cuda.device_count()
else:
return 1
@lru_cache
def get_xla_rank():
# When using spmd, these return 0 regardless of node [e.g., essentially return the local rank]
# import torch_xla.core.xla_model as xm
# return xm.get_ordinal()
# from accelerate import PartialState
# return PartialState().local_process_index
from torch_xla._internal import tpu
task_id = tpu.task_id() # Num chips
worker_id = tpu.worker_id() # Num workers [e.g., which node]
return task_id if task_id is not None else worker_id
def get_rank(check_for_group: bool = False):
""" Helper function that returns torch.distributed.get_rank() if DDP has been initialized otherwise it returns 0.
"""
if os.environ.get("FORCE_MAIN_RANK", "0") == "1":
return 0
elif is_torch_xla_available():
if builtins.HAS_XLA_SPAWNED:
return get_xla_rank()
else:
return 0
elif use_dist():
return dist.get_rank()
elif (rank := os.environ.get("RANK", None)) is not None:
return int(rank) # RANK is set by torch.distributed.launch
elif (rank := os.environ.get("SLURM_PROCID", None)) is not None:
return int(rank) # SLURM_PROCID is set by SLURM
elif check_for_group:
# if neither pytorch, SLURM env vars are set
# check NODE_RANK/GROUP_RANK and LOCAL_RANK env vars
# assume global_rank is zero if undefined
node_rank = os.environ.get("NODE_RANK", os.environ.get("GROUP_RANK", 0))
local_rank = os.environ.get("LOCAL_RANK", 0)
return 0 if (int(node_rank) == 0 and int(local_rank) == 0) else 1
else:
return 0
def is_main_process():
return get_rank() == 0
def get_local_rank():
if is_torch_xla_available():
import torch_xla.core.xla_model as xm
return xm.get_local_ordinal()
else:
return int(os.environ.get("LOCAL_RANK", 0))
def is_local_main_process():
return get_local_rank() == 0
def get_num_gpus() -> int:
return get_world_size()
def rank_zero_fn(fn):
@wraps(fn)
def wrapped_fn(*args: Any, **kwargs: Any):
if is_main_process():
return fn(*args, **kwargs)
return None
return wrapped_fn
def barrier():
if use_dist() or getattr(builtins, "HAS_XLA_SPAWNED", False):
frame = inspect.currentframe().f_back
filename = frame.f_code.co_filename
lineno = frame.f_lineno
code_context = inspect.getframeinfo(frame).code_context[0].strip()
debug_log_func(f"Before barrier - {filename}:{lineno} - {code_context}")
if is_torch_xla_available() and getattr(builtins, "HAS_XLA_SPAWNED", False):
import torch_xla.core.xla_model as xm
xm.rendezvous('barrier')
else:
torch.distributed.barrier()
debug_log_func(f"After barrier - {filename}:{lineno} - {code_context}")
def get_hostname():
return __import__('socket').gethostname().removesuffix('.eth')
def get_slurm_job_id():
if os.environ.get("SLURM_ARRAY_JOB_ID", None) is not None and os.environ.get("SLURM_ARRAY_TASK_ID", None) is not None:
job_str = f"{os.environ.get('SLURM_ARRAY_JOB_ID')}_{os.environ.get('SLURM_ARRAY_TASK_ID')}"
elif os.environ.get("SLURM_JOB_ID", None) is not None:
job_str = os.environ.get("SLURM_JOB_ID")
else:
job_str = None
return job_str
def get_restart_str():
restart_str = f"r{os.environ.get('SLURM_RESTART_COUNT', '0')}_" if os.environ.get('SLURM_RESTART_COUNT', None) is not None else ""
if "TORCHELASTIC_RESTART_COUNT" in os.environ and os.environ.get("TORCHELASTIC_RESTART_COUNT", '0') != '0':
restart_str = f"r{os.environ.get('TORCHELASTIC_RESTART_COUNT', '0')}_"
return restart_str
def get_slurm_filename_info():
job_str = f"{get_slurm_job_id()}__" if get_slurm_job_id() is not None else ""
restart_str = get_restart_str()
if "SLURM_NODEID" in os.environ:
node_str = f"{os.environ.get('SLURM_NODEID', '')}_"
elif "SLURM_PROCID" in os.environ:
node_str = f"{os.environ.get('SLURM_PROCID', '')}_"
else:
node_str = ""
return f"{job_str}{restart_str}{node_str}{get_rank()}"
def get_slurm_log_prefix():
if "SLURM_NODEID" in os.environ:
slurm_nodestr = f", Node:{os.environ.get('SLURM_NODEID', 'N/A')}"
elif "SLURM_PROCID" in os.environ:
slurm_nodestr = f", Node:{os.environ.get('SLURM_PROCID', 'N/A')}"
else:
slurm_nodestr = ""
jobid = get_slurm_job_id()
jobid_str = f", JID:{jobid}" if jobid is not None else ""
restart_str = f", {get_restart_str()}" if get_restart_str() != "" else ""
return f"Rank:{get_rank()}{slurm_nodestr}{jobid_str}{restart_str}"
def slurm_prefix_func():
timestamp = datetime.now().strftime("[%Y-%m-%d %H:%M:%S]")
return f"{timestamp} [{get_slurm_log_prefix()}]"
try:
from unidisc.utils.logging_utils import log_info, log_debug, set_logger, log_memory
set_logger(__name__)
info_log_func = log_info
debug_log_func = log_debug
prefix_func = None
except Exception as e:
print(e)
def simple_print_func(*args, **kwargs):
print(*args)
info_log_func = simple_print_func
debug_log_func = simple_print_func
prefix_func = slurm_prefix_func
def gprint(*args, main_process_only=False, **kwargs):
"""
Prints to console + log as INFO on regardless of rank.
"""
if prefix_func is not None:
args = (prefix_func(), *args)
info_log_func(*args, main_process_only=main_process_only, **kwargs)
def dprint(*args, **kwargs):
"""
Prints to console + log as DEBUG on regardless of rank.
"""
if prefix_func is not None:
args = (prefix_func(), *args)
debug_log_func(*args, **kwargs)
def rprint(*args, **kwargs):
"""
Prints to console + log as INFO on main process only. All ranks also print to log file [as DEBUG] if called, but this is not required [e.g., no barrier used]
"""
gprint(*args, main_process_only=True, **kwargs)
def mprint(*args, **kwargs):
log_memory(*args, **kwargs)
# dprint(*args, **kwargs)
log_func = rprint
def process_file_prefix():
datetime_str = datetime.now().strftime("%Y_%m_%d-%H_%M_%S_%f")[:-5]
return f"{get_slurm_filename_info()}_{get_hostname()}_{datetime_str}"
def get_info():
return subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE).stdout.decode("utf-8")
def print_trainable_parameters(model):
"""
Prints the number of trainable parameters in the model.
"""
trainable_params = 0
all_param = 0
for _, param in model.named_parameters():
all_param += param.numel()
if param.requires_grad:
trainable_params += param.numel()
log_func(
f"trainable params: {trainable_params} || all params: {all_param} || trainable%: {100 * trainable_params / all_param:.2f}"
)
def print_params(model):
log_func(f"Total Parameters: {sum(p.numel() for p in model.parameters()):,}")
log_func(f"Unfrozen Parameters: {sum(p.numel() for p in model.parameters() if p.requires_grad):,}")
log_func(f"Frozen Parameters: {sum(p.numel() for p in model.parameters() if not p.requires_grad):,}")
def calculate_storage_size(obj, storage_view_sizes, count_views=False):
if isinstance(obj, torch.Tensor):
storage = obj.storage()
storage_id = id(storage)
element_size = storage.element_size()
storage_size = storage.size() * element_size
view_size = obj.numel() * element_size
# We count storage size only for the first time we encounter the storage
if storage_id not in storage_view_sizes:
storage_view_sizes[storage_id] = storage_size
print_size = storage_size
else:
print_size = 0 if not count_views or not obj._is_view() else view_size
if count_views or not obj._is_view():
log_func(f"{'View' if obj._is_view() else 'Storage'} Tensor: " f"shape {obj.size()}, size {print_size / (1024 ** 2):.2f} MB")
return print_size if count_views or not obj._is_view() else 0 # Count views only if requested
elif isinstance(obj, dict):
# Recurse for dictionaries
return sum(calculate_storage_size(v, storage_view_sizes, count_views) for v in obj.values())
elif isinstance(obj, (list, tuple)):
# Recurse for lists or tuples
return sum(calculate_storage_size(item, storage_view_sizes, count_views) for item in obj)
elif hasattr(obj, "__dataclass_fields__"):
# Recurse for dataclasses based on their fields
fields = getattr(obj, "__dataclass_fields__")
return sum(calculate_storage_size(getattr(obj, f), storage_view_sizes, count_views) for f in fields)
else:
# Non-Tensor, non-dict, non-list objects are not measured
return 0
def calculate_total_size(obj, count_views=False):
storage_view_sizes = defaultdict(int)
total_size = calculate_storage_size(obj, storage_view_sizes, count_views)
total_unique_storage_size = sum(storage_view_sizes.values())
log_func(f"Total unique storage size: {total_unique_storage_size / (1024 ** 2):.2f} MB")
if count_views: # Only add view sizes to total if requested
total_view_size = total_size - total_unique_storage_size
log_func(f"Total view size (if counted): {total_view_size / (1024 ** 2):.2f} MB")
else:
log_func(f"Total size (without counting views): {total_size / (1024 ** 2):.2f} MB")
return total_size
def save_tensor_dict(tensor_dict: dict, path):
output_dict = {}
for k, v in tensor_dict.items():
if isinstance(v, Tensor):
if v.dtype == torch.float16 or v.dtype == torch.bfloat16:
output_dict[k] = v.to(dtype=torch.float32).detach().cpu().numpy()
else:
output_dict[k] = v.detach().cpu().numpy()
elif isinstance(v, np.ndarray):
output_dict[f"np_{k}"] = v
else:
output_dict[k] = v
np.savez_compressed(path, **output_dict)
def load_tensor_dict(path: Path, object_keys=[]):
from jaxtyping import BFloat16 # TODO: Remove dependency
tensor_dict = {}
np_dict = np.load(path, allow_pickle=True)
for k, v in np_dict.items():
if k in object_keys:
tensor_dict[k] = v
elif v.dtype == BFloat16:
tensor_dict[k] = torch.from_numpy(v.astype(np.float32)).to(dtype=torch.bfloat16)
elif k.startswith("np_"):
tensor_dict[k.removeprefix("np_")] = v
else:
tensor_dict[k] = torch.from_numpy(v)
return tensor_dict
def tensor_hash(tensor):
"""Computes a SHA256 hash of a tensor. Useful for debugging to check equality in different places."""
tensor_bytes = tensor.detach().float().cpu().numpy().tobytes()
return hashlib.sha256(tensor_bytes).hexdigest()
def module_hash(module: Optional[dict] = None, state_dict: Optional[dict] = None):
assert module is not None or state_dict is not None
state_dict = module.state_dict() if module is not None else state_dict
sorted_state_dict = {k: state_dict[k] for k in sorted(state_dict)}
params_cat = torch.cat([v.flatten() for _, v in sorted_state_dict.items()])
return tensor_hash(params_cat)
def parameter_hash(params: list[torch.Tensor]):
return tensor_hash(torch.cat([p.cpu().flatten() for p in params]))
def find_diff_params(state_dict_1, state_dict_2):
diff_keys = set(state_dict_1.keys()) ^ set(state_dict_2.keys()) # Symmetric difference to find keys not in both
matched_keys = set(state_dict_1.keys()) & set(state_dict_2.keys()) # Intersection to find keys in both
# Check for differences in matched keys
for key in matched_keys:
if not torch.equal(state_dict_1[key], state_dict_2[key]):
diff_keys.add(key)
return diff_keys
def init_from_ckpt(module, path, ignore_keys=None, unfrozen_keys=None, strict=False, truncate=None, only_incl=None, verbose=True):
log_func(f"Loading {module.__class__.__name__} from checkpoint: {path}")
log_func(f"Strict Load: {strict}, Ignoring: {ignore_keys}, Unfreezing: {unfrozen_keys}, Truncating: {truncate}")
if ignore_keys is None:
ignore_keys = ()
if unfrozen_keys is None:
unfrozen_keys = ()
sd = torch.load(path, map_location="cpu")
if "state_dict" in sd.keys():
sd = sd["state_dict"]
elif "weight" in sd.keys():
sd = sd["weight"]
num_deleted = defaultdict(int)
for k in list(sd):
for ik in ignore_keys:
if k.startswith(ik):
num_deleted[ik] += 1
del sd[k]
for k, v in num_deleted.items():
log_func(f"Deleted {v} keys due to ignore_key: {k}")
if truncate is not None:
for k in list(sd):
if k.startswith(truncate):
sd[k.replace(truncate, "")] = sd[k]
del sd[k]
num_ignored = defaultdict(int)
for n in module.state_dict().keys():
if n not in sd.keys():
for ik in ignore_keys:
if ik in n:
num_ignored[ik] += 1
else:
log_func(f"Missing {n}")
if only_incl is not None:
for k in list(sd):
keep = False
for ik in only_incl:
if ik in k:
keep = True
if not keep:
del sd[k]
for k, v in num_ignored.items():
log_func(f"Missing {v} keys due to ignore_key: {k}")
for n in sd.keys():
if n not in module.state_dict().keys():
log_func(f"Unexpected {n}")
checkpoint_keys = set(sd.keys())
current_keys = set(module.state_dict().keys())
if verbose:
log_func(f"Loading: {checkpoint_keys.intersection(current_keys)}")
else:
log_func(f"Loading {len(checkpoint_keys.intersection(current_keys))} keys into the model: {str(module.__class__)}")
module.load_state_dict(sd, strict=strict)
if len(unfrozen_keys) > 0:
for n, p in module.named_parameters():
p.requires_grad_ = False
for unfrozen_name in unfrozen_keys:
if unfrozen_name in n:
p.requires_grad_ = True
log_func(f"Unfreezing: {n}")
log_func(f"Restored from {path}")
def check_gpu_memory_usage():
allocated = torch.cuda.memory_allocated()
reserved = torch.cuda.memory_reserved()
total_memory = torch.cuda.get_device_properties(int(get_local_rank())).total_memory
allocated_percent = (allocated / total_memory) * 100
reserved_percent = (reserved / total_memory) * 100
log_func(f"Allocated memory: {allocated_percent:.2f}%")
log_func(f"Reserved memory: {reserved_percent:.2f}%")
log_func(f'Available devices (CUDA_VISIBLE_DEVICES): {os.environ.get("CUDA_VISIBLE_DEVICES")}')
assert allocated_percent <= 5
assert reserved_percent <= 5
def load_checkpoint_from_url(url: str, file_path: Optional[str] = None) -> Path:
if file_path is None:
parts = urlparse(url)
filename = os.path.basename(parts.path)
if file_path is not None:
filename = file_path
file_path = Path.home() / ".cache" / "pretrained_weights" / filename
file_path.parent.mkdir(parents=True, exist_ok=True)
if not os.path.exists(file_path):
log_func(f'Downloading: "{url}" to {file_path}\n')
torch.hub.download_url_to_file(url, file_path, progress=True)
return file_path
# Copied from torch.profiler.profiler
def tensorboard_trace_handler(dir_name: str, record_memory: bool = False, worker_name: Optional[str] = None, use_gzip: bool = True):
"""
Outputs tracing files to directory of ``dir_name``, then that directory can be
directly delivered to tensorboard as logdir.
``worker_name`` should be unique for each worker in distributed scenario,
it will be set to '[hostname]_[pid]' by default.
"""
import os
import socket
import time
def handler_fn(prof: torch.profiler.profile) -> None:
nonlocal worker_name
if not os.path.isdir(dir_name):
try:
os.makedirs(dir_name, exist_ok=True)
except Exception as e:
raise RuntimeError("Can't create directory: " + dir_name) from e
if not worker_name:
worker_name = f"{socket.gethostname()}_{os.getpid()}"
# Use nanosecond here to avoid naming clash when exporting the trace
file_name = f"{worker_name}.{time.time_ns()}.pt.trace.json"
if use_gzip:
file_name = file_name + ".gz"
chrome_trace_path = os.path.join(dir_name, file_name)
memory_trace_path = os.path.join(dir_name, "memory_timeline.html")
if is_main_process() and record_memory:
try:
log_func(f"Exporting memory timeline: {memory_trace_path}")
prof.export_memory_timeline(memory_trace_path)
except Exception as e:
log_func(f"Failed to export memory timeline: {e}")
prof.key_averages().table(sort_by="self_cuda_time_total", row_limit=100)
log_func(f"Exporting chrome trace to {chrome_trace_path}")
prof.export_chrome_trace(chrome_trace_path)
return handler_fn
def get_date_time_str():
return datetime.now().strftime("%Y_%m_%d-%H_%M_%S.%f")[:-3]
def save_memory_profile(profile_dir):
import wandb
rank_postfix = f"_rank_{get_rank()}" if use_dist() else ""
log_func(f"Saving memory profile to {profile_dir}")
os.makedirs(profile_dir, exist_ok=True)
torch.cuda.memory._dump_snapshot(f"{profile_dir}/memory_snapshot{rank_postfix}.pickle")
os.system(
f"python -m torch.cuda._memory_viz trace_plot {profile_dir}/memory_snapshot{rank_postfix}.pickle -o {profile_dir}/memory_snapshot{rank_postfix}.html"
)
torch.cuda.memory._save_segment_usage(f"{profile_dir}/segment{rank_postfix}.svg")
torch.cuda.memory._save_memory_usage(f"{profile_dir}/memory{rank_postfix}.svg")
torch.cuda.memory._record_memory_history(enabled=None)
log_func(f"Saved memory snapshot at: {profile_dir}/memory_snapshot{rank_postfix}.pickle")
log_func(f"Run the following to view the snapshot:\npython -m http.server --directory {profile_dir.resolve()} 6008")
if is_main_process() and wandb.run is not None:
wandb.log({'profile': wandb.Html(f"{profile_dir}/memory_snapshot{rank_postfix}.html")})
wandb.log({'profile': wandb.Html(f"{profile_dir}/memory_timeline{rank_postfix}.html")})
def print_memory(verbose: bool = False, print_output: bool = True):
max_cur_reserved, max_peak_reserved, max_peak_allocated, max_cur_allocated = -1, -1, -1, -1
max_cur_reserved_device, max_peak_reserved_device, max_peak_allocated_device, max_cur_allocated_device = -1, -1, -1, -1
for device in range(torch.cuda.device_count()):
current_reserved_memory_MB = torch.cuda.memory_reserved(device=torch.device(f'cuda:{device}')) / (1024**2)
peak_reserved_memory_MB = torch.cuda.max_memory_reserved(device=torch.device(f'cuda:{device}')) / (1024**2)
peak_allocated_memory_MB = torch.cuda.max_memory_allocated(device=torch.device(f'cuda:{device}')) / (1024**2)
current_allocated_memory_MB = torch.cuda.memory_allocated(device=torch.device(f'cuda:{device}')) / (1024**2)
if current_reserved_memory_MB > max_cur_reserved:
max_cur_reserved = current_reserved_memory_MB
max_cur_reserved_device = device
if peak_reserved_memory_MB > max_peak_reserved:
max_peak_reserved = peak_reserved_memory_MB
max_peak_reserved_device = device
if peak_allocated_memory_MB > max_peak_allocated:
max_peak_allocated = peak_allocated_memory_MB
max_peak_allocated_device = device
if current_allocated_memory_MB > max_cur_allocated:
max_cur_allocated = current_allocated_memory_MB
max_cur_allocated_device = device
if verbose:
log_func(torch.cuda.memory_summary(abbreviated=False))
if print_output:
log_func(f"GPU Cur Reserved: {max_cur_reserved:.2f}MB on rank {max_cur_reserved_device}, Cur Allocated: {max_cur_allocated:.2f}MB on rank {max_cur_allocated_device}, Peak Reserved: {max_peak_reserved:.2f}MB on rank {max_peak_reserved_device}, Peak Allocated: {max_peak_allocated:.2f}MB on rank {max_peak_allocated_device}")
return max_cur_reserved
def print_memory_summary():
val = print_memory(verbose=False, print_output=False)
log_func(f"GPU Cur Reserved: {val:.2f}MB, {val / (torch.cuda.get_device_properties(0).total_memory / 1024**2) * 100:.2f}%")
@contextlib.contextmanager
def show_memory_usage(empty_cache: bool = True, verbose: bool = False, print_output=False, show_caller: bool = True):
synchronize_device()
if empty_cache: clear_cache()
if show_caller:
callers = [
f"{frame.function} (...{frame.filename[-10:]}:{frame.lineno})"
for frame in inspect.stack()
if "contextlib" not in frame.filename
][1:5]
decorated_func_name = ", ".join([inspect.currentframe().f_back.f_code.co_name, inspect.currentframe().f_back.f_back.f_code.co_name, inspect.currentframe().f_back.f_back.f_back.f_code.co_name])
caller_info = decorated_func_name + " " + " -> ".join(reversed(callers))
log_func(f"Before context (called by {caller_info}): {print_memory(verbose, print_output=False)}MB cur reserved")
caller_str = f", called by {caller_info} "
else:
caller_str = ""
yield
synchronize_device()
if empty_cache: clear_cache()
log_func(f"After context{caller_str}: {print_memory(verbose, print_output=print_output)}MB cur reserved")
@contextlib.contextmanager
def profile_memory(enable: bool = True, empty_cache: bool = True):
with contextlib.ExitStack() as stack:
stack.enter_context(show_memory_usage(empty_cache=empty_cache))
if enable and is_main_process(): torch.cuda.memory._record_memory_history()
yield
if enable: save_memory_profile(Path(f"output/profile/{get_date_time_str()}"))
def profile_memory_decorator(func):
@wraps(func)
def wrapper(*args, **kwargs):
with profile_memory():
return func(*args, **kwargs)
return wrapper
class Profiler:
def __init__(self, output_dir, warmup_steps: int = 5, active_steps: int = 3, record_memory: bool = False, record_memory_only: bool = False):
self.record_memory = record_memory
self.profile_dir = Path(output_dir) / "profile"
self.profile_dir.mkdir(parents=True, exist_ok=True)
wait, warmup, active, repeat = 0, warmup_steps, active_steps, 0
self.total_steps = (wait + warmup + active) * (1 + repeat)
schedule = torch.profiler.schedule(wait=wait, warmup=warmup, active=active, repeat=repeat)
profiler_kwargs = dict(record_shapes=True, with_stack=True)
if record_memory_only:
pass
else:
profiler_kwargs.update(
dict(
with_modules=True,
with_flops=True,
),
)
self.profiler = torch.profiler.profile(
schedule=schedule,
on_trace_ready=tensorboard_trace_handler(self.profile_dir, record_memory=record_memory),
profile_memory=record_memory,
**profiler_kwargs
)
self.profiler.start()
def step(self, global_step: int):
self.profiler.step()
return global_step >= (self.total_steps - 1)
def finish(self):
self.profiler.stop()
if use_dist():
torch.distributed.barrier()
if is_main_process():
import wandb
traces = glob.glob(f"{self.profile_dir}/*.pt.trace.json*")
for trace in traces:
log_func(f"Adding {trace}")
wandb.save(trace, base_path=self.profile_dir, policy="now")
if use_dist():
torch.distributed.barrier()
def get_pdb():
return import_module("pdb") if (any(["_pdbpp_path_hack" in str(p) for p in sys.path]) or find_spec("ipdb") is None) else import_module("ipdb")
def _breakpoint(rank: Optional[int] = None, traceback: Optional[Any] = None):
if get_num_gpus() > 1:
if (is_main_process() if rank is None else get_rank() == rank):
if is_torch_xla_available():
from fairseq import pdb
pdb.set_trace()
else:
old_stdin = None
if isinstance(sys.stdin, io.TextIOWrapper):
old_stdin = sys.stdin
sys.stdin = open(0)
try:
log_func('Breakpoint triggered. You may need to type "up" to get to the correct frame')
log_func(f'Traceback: {traceback}')
if traceback is not None:
get_pdb().post_mortem(traceback)
else:
get_pdb().set_trace()
finally:
if old_stdin is not None:
sys.stdin.close()
sys.stdin = old_stdin
barrier()
else:
if traceback is not None:
get_pdb().post_mortem(traceback)
else:
log_func("Breakpoint triggered. You may need to type \"up\" to get to the correct frame")
get_pdb().set_trace(sys._getframe(1))
def set_global_exists():
builtins.exists = lambda v: v is not None
def set_global_breakpoint():
import ipdb
builtins.breakpoint = _breakpoint
builtins.st = ipdb.set_trace # We import st everywhere
builtins.ug = lambda: globals().update(locals())
def set_timing_builtins(enable: bool = False, sync: bool = False):
builtins.start_timing = partial(start_timing, builtin=True, enable=False, sync=False)
builtins.end_timing = partial(end_timing, builtin=True, enable=False, sync=False)
builtins.ENABLE_TIMING = enable
builtins.ENABLE_TIMING_SYNC = sync
def synchronize_device():
if is_torch_cuda_available():
torch.cuda.synchronize()
elif is_torch_xla_available():
import torch_xla
torch_xla.sync()
xm.wait_device_ops()
def clear_cache():
if is_torch_cuda_available():
# This caused untold grief. Without calling `_cuda_clearCublasWorkspaces`,
# some model configurations would eventually cause a CUDA OOM during inference. See:
# https://github.com/pytorch/pytorch/issues/99835, https://github.com/pytorch/pytorch/issues/105181
torch._C._cuda_clearCublasWorkspaces()
torch._dynamo.reset()
import gc; gc.collect()
torch.cuda.empty_cache()
elif is_torch_xla_available():
rprint("Clearing cache not supported for XLA")
def start_timing(message: str, enable: bool = False, sync: bool = False, builtin: bool = False):
if (builtin and ENABLE_TIMING) or enable:
if (builtin and ENABLE_TIMING_SYNC) or sync:
synchronize_device()
torch.cuda.nvtx.range_push(f"[SYNC] {message}" if ((builtin and ENABLE_TIMING_SYNC) or sync) else message)
return time.time()
def end_timing(start_time: Optional[float] = None, enable: bool = False, sync: bool = False, builtin: bool = False):
if (builtin and ENABLE_TIMING) or enable:
if (builtin and ENABLE_TIMING_SYNC) or sync:
synchronize_device()
torch.cuda.nvtx.range_pop()
if start_time is not None:
return time.time() - start_time
@contextlib.contextmanager
def breakpoint_on_error():
set_global_breakpoint()
try:
yield
except Exception as e:
print("Exception...", e)
traceback.print_exc()
breakpoint(traceback=e.__traceback__)
raise e
@contextlib.contextmanager
def get_time_sync(enable: bool = True):
if enable and is_main_process():
synchronize_device()
start_time = time.time()
yield
if enable and is_main_process():
synchronize_device()
end_time = time.time()
print(f"Time taken: {end_time - start_time:.2f}s")
def write_to_file(path: Path, text: str):
try:
path.parent.mkdir(parents=True, exist_ok=True)
with open(path, "a") as file:
file.write(text + "\n")
except:
log_func(f"Could not write to {path}")
def to(obj, device):
if torch.is_tensor(obj):
return obj.to(device)
if isinstance(obj, dict):
return {k : to(v, device) for k, v in obj.items()}
if isinstance(obj, tuple):
return tuple(to(v, device) for v in obj)
if isinstance(obj, list):
return [to(v, device) for v in obj]
return obj
def all_gather(data):
"""
Run all_gather on arbitrary picklable data (not necessarily tensors)
Args:
data: any picklable object
Returns:
list[data]: list of data gathered from each rank
"""
world_size = get_num_gpus()
if world_size == 1:
return [data]
# serialized to a Tensor
buffer = pickle.dumps(data)
storage = torch.ByteStorage.from_buffer(buffer)
tensor = torch.ByteTensor(storage).to("cuda")
# obtain Tensor size of each rank
local_size = torch.tensor([tensor.numel()], device="cuda")
size_list = [torch.tensor([0], device="cuda") for _ in range(world_size)]
dist.all_gather(size_list, local_size)
size_list = [int(size.item()) for size in size_list]
max_size = max(size_list)
# receiving Tensor from all ranks
# we pad the tensor because torch all_gather does not support
# gathering tensors of different shapes
tensor_list = []
for _ in size_list:
tensor_list.append(torch.empty((max_size,), dtype=torch.uint8, device="cuda"))
if local_size != max_size:
padding = torch.empty(size=(max_size - local_size,), dtype=torch.uint8, device="cuda")
tensor = torch.cat((tensor, padding), dim=0)
dist.all_gather(tensor_list, tensor)
data_list = []
for size, tensor in zip(size_list, tensor_list):
buffer = tensor.cpu().numpy().tobytes()[:size]
data_list.append(to(pickle.loads(buffer), torch.device('cpu')))
return data_list
def get_modules(model: torch.nn.Module, cls: Any):
children = list(model.children())
if isinstance(model, cls):
return [model]
elif len(children) == 0:
return []
else:
return [ci for c in children for ci in get_modules(model=c, cls=cls)]
map_chars = {
"/" : "__",
" " : "_",
}
def sanitize_filename(filename: str) -> str:
return "".join(map_chars.get(c, c) for c in filename if c.isalnum() or map_chars.get(c, c) in (" ", ".", "_", "-", "__"))
def hash_str_as_int(s: str):
return int(hashlib.sha256(s.encode('utf-8')).hexdigest(), 16) % 10**8
def torch_to_numpy(arr: Tensor):
if arr.dtype == torch.bfloat16:
return arr.float().cpu().detach().numpy()
else:
return arr.cpu().detach().numpy()
def to_numpy(arr):
if isinstance(arr, Tensor):
return torch_to_numpy(arr)
else:
return arr
class try_except:
def __init__(self, raise_error: bool = False, write_error_to_file: bool = False, clear_cuda_cache: bool = False):
self.raise_error = raise_error
self.write_error_to_file = write_error_to_file
self.clear_cuda_cache = clear_cuda_cache
def __call__(self, f):
@functools.wraps(f)
def inner(*args, **kwargs):
with self:
return f(*args, **kwargs)
return inner
def __enter__(self):
if self.clear_cuda_cache:
clear_cache()
return self
def __exit__(self, exc_type, exc_value, tb):
if exc_type is not None:
try:
log_func(f"Exception caught: {exc_type}")
log_func(traceback.format_exc())
if self.write_error_to_file:
timestamp = int(time.time_ns())
with open(f"exception_{timestamp}_{process_file_prefix()}.out", "w") as file:
file.write(traceback.format_exc())
except Exception as e:
print(f"Error writing to file: {e}")
if self.clear_cuda_cache:
clear_cache()
if self.raise_error:
raise exc_value
return True # Suppress the exception if raise_error is False
def move_to(obj, device):
if torch.is_tensor(obj):
return obj.to(device=device)
elif isinstance(obj, dict):
res = {}
for k, v in obj.items():
res[k] = move_to(v, device)
return res
elif isinstance(obj, list):
res = []
for v in obj:
res.append(move_to(v, device))
return res
elif isinstance(obj, str):
return obj
else:
raise TypeError("Invalid type for move_to")
import types
def run_with_named_function(name, func, *args, **kwargs):
"""
Runs the given function inside a dynamically created function with the specified name.
This causes the specified name to appear in the stack trace.
E.g., x = run_with_named_function(f"iter_{i}", self._ddpm_update, x, t, dt, x0=x0, x0_unmask=x0_unmask, **kwargs)
Parameters:
- name (str): The desired name to appear in the stack trace.
- func (callable): The function to execute.
- *args: Positional arguments to pass to func.
- **kwargs: Keyword arguments to pass to func.
"""
def wrapper(*args, **kwargs):
return func(*args, **kwargs)
code = wrapper.__code__
new_code = types.CodeType(
code.co_argcount,
code.co_posonlyargcount if hasattr(code, "co_posonlyargcount") else 0,
code.co_kwonlyargcount,
code.co_nlocals,
code.co_stacksize,
code.co_flags,
code.co_code,
code.co_consts,
code.co_names,
code.co_varnames,
code.co_filename,
name, # Set the function name in the code object
code.co_firstlineno,
code.co_lnotab,
code.co_freevars,
code.co_cellvars
)
new_func = types.FunctionType(
new_code,
wrapper.__globals__,
name,
wrapper.__defaults__,
wrapper.__closure__,
)
return new_func(*args, **kwargs) |