File size: 33,976 Bytes
7cd774e |
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 |
import copy
import gc
import os
import subprocess
import pytz
import sys
import shutil
import tempfile
import threading
import time
from functools import partial
from pathlib import Path
from typing import Optional
import bittensor as bt
import hivemind
import psutil
import torch
from memory_profiler import profile
from datetime import datetime
from hivemind.compression import deserialize_torch_tensor
from hivemind.proto import averaging_pb2
from hivemind.utils import get_logger
from hivemind.utils.asyncio import aiter_with_timeout
from hivemind.utils.streaming import combine_from_streaming
from huggingface_hub import (
create_tag,
hf_hub_download,
list_repo_refs,
list_repo_files,
scan_cache_dir,
upload_folder,
)
from huggingface_hub.utils import (
HfHubHTTPError,
RepositoryNotFoundError,
EntryNotFoundError,
)
from huggingface_hub.constants import HF_HUB_CACHE
from transformers import (
AutoModelForCausalLM,
AutoConfig,
get_cosine_schedule_with_warmup,
)
from distributed_training import __run__
from distributed_training.averaging.averagers import DTGradAverager, DTStateAverager
from distributed_training.utils.progress_tracker import (
get_global_epoch,
get_local_inner_step,
get_min_local_inner_Step,
)
from distributed_training.averaging.avg_handler import AveragingHandler
from huggingface_hub import list_repo_commits
hivemind_logger = get_logger(__name__)
class ModelLoadingManager:
def __init__(self):
self.loading_lock = threading.Lock()
self._is_loading = False
self._last_loaded_epoch = None
@property
def is_loading(self):
with self.loading_lock:
return self._is_loading
@property
def last_loaded_epoch(self):
with self.loading_lock:
return self._last_loaded_epoch
def set_loading_state(self, is_loading, epoch=None):
with self.loading_lock:
self._is_loading = is_loading
if not is_loading and epoch is not None:
self._last_loaded_epoch = epoch
class FastModelLoader:
def __init__(self, model_name: str, cache_dir: str = None):
"""
Initialize the fast model loader with HF downloader integration.
Args:
model_name (str): The HuggingFace model name (e.g., 'organization/model-name')
cache_dir (str, optional): Directory to store downloaded files. Defaults to HF cache.
"""
self.model_name = model_name
self.cache_dir = cache_dir or os.path.expanduser("~/.cache/huggingface/hub")
self._downloaded_files = {} # Cache of downloaded files
def download_files(self, revision: str = None, files: list = None):
"""
Download files using hfdownloader.
Args:
revision (str, optional): Git revision/epoch number
files (list, optional): List of specific files to download with patterns
Returns:
str: Path to downloaded files
"""
# Generate cache key
cache_key = f"{revision}_{','.join(files) if files else 'default'}"
# Check if we already downloaded these files
if cache_key in self._downloaded_files:
return self._downloaded_files[cache_key]
model_path = os.path.join(self.cache_dir, self.model_name.replace("/", "_"))
os.makedirs(model_path, exist_ok=True)
cmd = [
"hfdownloader",
"-r",
self.model_name,
"download",
"-c",
"10",
"-y",
]
if revision:
cmd.extend(["-b", revision])
# Add file patterns if specified, otherwise default to both model and optimizer
if files:
for file_pattern in files:
cmd.extend(["-f", f"{file_pattern}"])
else:
cmd.extend(
[
"-f",
"*.safetensors",
"-f",
"optimizer.pt",
"--skip-verify",
]
)
bt.logging.debug(f"Executing hfdownloader command: {' '.join(cmd)}")
try:
process = subprocess.Popen(
cmd,
cwd=model_path,
stdout=subprocess.PIPE,
stderr=subprocess.PIPE,
text=True,
bufsize=1,
universal_newlines=True,
env={
**os.environ,
"PYTHONUNBUFFERED": "1",
}, # Force Python unbuffered output
)
# Use select to handle both stdout and stderr
import select
outputs = [process.stdout, process.stderr]
while True:
# Wait for output on either stdout or stderr
readable, _, _ = select.select(outputs, [], [])
for output in readable:
line = output.readline()
if line:
# Don't buffer the print
print(line.rstrip(), flush=True)
# Check if process has finished
if process.poll() is not None:
break
# Get any remaining output
remaining_stdout, remaining_stderr = process.communicate()
if remaining_stdout:
print(remaining_stdout.rstrip(), flush=True)
if remaining_stderr:
print(
f"Error: {remaining_stderr.rstrip()}", file=sys.stderr, flush=True
)
if process.returncode != 0:
raise RuntimeError(
f"hfdownloader failed with return code {process.returncode}"
)
except Exception as e:
bt.logging.error(f"Download failed! Error: {str(e)}")
raise RuntimeError(f"hfdownloader failed: {str(e)}")
return model_path
def load_model_and_optimizer(self, epoch: int = None):
"""
Load both model and optimizer states in a single download operation.
Args:
epoch (int, optional): Epoch number for specific revision
Returns:
tuple: (model_state_dict, optimizer_state_dict)
"""
revision = str(epoch) if epoch is not None else None
# Download both model and optimizer files in one go
model_path = self.download_files(revision=revision)
# Load model state
model_files = list(Path(model_path).rglob("*.safetensors"))
if not model_files:
raise FileNotFoundError(f"No model files found in {model_path}")
bt.logging.info(f"Loading model state from: {[f.name for f in model_files]}")
state_dict = {}
for model_file in model_files:
from safetensors.torch import load_file
state = load_file(model_file)
state_dict.update(state)
# Load optimizer state
optimizer_file = Path(model_path) / "optimizer.pt"
if not optimizer_file.exists():
raise FileNotFoundError(f"Optimizer state not found at {optimizer_file}")
bt.logging.info(f"Loading optimizer state from: {optimizer_file}")
optimizer_state = torch.load(str(optimizer_file), map_location="cpu")
return state_dict, optimizer_state
def check_model_exists(repo_id: str, revision: Optional[str] = None) -> bool:
try:
if revision and revision != "None":
list_repo_files(repo_id, revision=revision)
else:
list_repo_files(repo_id)
return True
except Exception as e:
bt.logging.info(f"Model or revision check failed with error: {e}")
return False
# @profile
def load_model_optimizer_gradient_averager(
self,
local_model_name,
epoch,
reload_inner_optimizer=True,
reload_outer_optimizer=True,
revision=None,
use_fallback_model=True,
reset_block_list=True,
):
"""
Pytorch currently have an ongoing issue with memory leaks:
https://github.com/pytorch/pytorch/issues/64043. To mitigate
against this for now gc.collect() is run after each component
with optimizers and state averagers are deleted.
"""
bt.logging.debug(
f"CPU Memory Before Loading State {psutil.virtual_memory().available / 10**9} GB"
)
global_model_name = self.config.neuron.global_model_name
self.global_model_config = AutoConfig.from_pretrained(
global_model_name, trust_remote_code=True
)
if use_fallback_model:
model_name_list = [local_model_name, global_model_name]
else:
model_name_list = [local_model_name]
if (revision is None) and (local_model_name != global_model_name):
revision = f"{__run__}.{epoch}.{self.local_progress.inner_step}"
elif (revision is None) and (local_model_name == global_model_name):
revision = f"{__run__}.{epoch}.0"
# Delete Gradient and State Averagers
if hasattr(self, "state_averager"):
self.grad_averager.shutdown()
while self.grad_averager.is_alive():
time.sleep(1)
del self.grad_averager.main_parameters
del self.grad_averager.offloaded_optimizer
del self.grad_averager._averaged_tensors
del self.grad_averager
gc.collect()
torch.cuda.empty_cache()
self.state_averager.shutdown()
while self.state_averager.is_alive():
time.sleep(1)
del self.state_averager.optimizer.param_groups
del self.state_averager.optimizer
del self.state_averager.main_parameters
del self.state_averager._averaged_tensors
del self.state_averager
gc.collect()
torch.cuda.empty_cache()
bt.logging.info("Deleted State Averager and Gradient Averager")
# Delete existing averag handler
if hasattr(self, "avg_handler"):
del self.avg_handler.model
del self.avg_handler.inner_optimizer
del self.avg_handler.grad_averager
del self.avg_handler.state_averager
del self.avg_handler
gc.collect()
torch.cuda.empty_cache()
bt.logging.info("Deleted Average Handler")
for model_name in model_name_list:
optimizer_state = None
# Load Model & Inner Optimizer
try:
if model_name == global_model_name:
revision = ".".join(revision.split(".")[:-1] + ["0"])
if not check_model_exists(
model_name,
revision=revision,
):
continue
# Delete existing model
if hasattr(self, "model"):
transformer = self.model.model.transformer
for component in ["wte", "wpe"]:
if hasattr(transformer, component):
comp = getattr(transformer, component)
if hasattr(comp, "weight"):
del comp.weight
gc.collect()
torch.cuda.empty_cache()
if hasattr(comp, "norm"):
del comp.norm
gc.collect()
torch.cuda.empty_cache()
delattr(transformer, component)
del self.model
gc.collect()
torch.cuda.empty_cache()
bt.logging.info("Deleted Model")
self.model = AutoModelForCausalLM.from_pretrained(
model_name,
revision=revision,
trust_remote_code=True,
)
bt.logging.info(
f"Successfully Loaded Model From {model_name} With Revision {revision}"
)
# Move model to device
self.model = self.model.to(self.device)
self.model.config.block_list = []
self.local_progress.inner_step = (
self.model.config.inner_step
if "inner_step" in self.model.config.__dict__
else 0
)
if (model_name == global_model_name) and (
epoch == self.global_progress.epoch
):
self.allreduce_status_dict = (
self.model.config.all_reduce_scores
if "all_reduce_scores" in self.model.config.__dict__
else {}
)
if reload_inner_optimizer:
# Delete existing inner optimizer
if hasattr(self, "inner_optimizer"):
for i in self.inner_optimizer.param_groups[0]["params"]:
del i
gc.collect()
torch.cuda.empty_cache()
del self.inner_optimizer
gc.collect()
torch.cuda.empty_cache()
bt.logging.info("Deleted Inner Optimizer")
self.inner_optimizer = torch.optim.AdamW(
self.model.parameters(),
lr=self.learning_rate_maximum,
betas=(0.9, 0.95),
weight_decay=0.1,
)
bt.logging.info(f"Loaded Inner Optimizer")
self.scheduler = get_cosine_schedule_with_warmup(
self.inner_optimizer,
num_warmup_steps=1000,
num_training_steps=88000,
)
try:
optimizer_state = torch.load(
os.path.join(
model_name.split("/")[-1],
"inner_optimizer.pt",
),
weights_only=True,
map_location="cpu",
)
except:
optimizer_state = torch.load(
hf_hub_download(
repo_id=model_name,
filename="inner_optimizer.pt",
revision=revision,
),
weights_only=True,
map_location="cpu",
)
# Load optimizer state if available
if "optimizer_state_dict" in optimizer_state:
self.inner_optimizer.load_state_dict(
optimizer_state["optimizer_state_dict"]
)
if "learning_rate" in optimizer_state:
for group in self.inner_optimizer.param_groups:
group["lr"] = optimizer_state["learning_rate"]
if "scheduler_state" in optimizer_state:
self.scheduler.load_state_dict(optimizer_state["scheduler_state"])
bt.logging.info(
f"Successfully Loaded Inner Optimizer State From {model_name} For Revision {revision}"
)
break
except Exception as e:
if model_name == model_name_list[-1]:
raise Exception(f"Failed to load model despite repo existing: {str(e)}")
else:
bt.logging.info(f"Failed to load model despite repo existing: {str(e)}")
finally:
if isinstance(optimizer_state, dict):
keys = list(optimizer_state.keys())
for k in keys:
del optimizer_state[k]
gc.collect()
del optimizer_state
gc.collect()
torch.cuda.empty_cache()
# Set outer optimizer
self.outer_optimizer = partial(torch.optim.SGD, lr=0.7, momentum=0.9, nesterov=True)
# Load a new state averager
self.state_averager = DTStateAverager(
dht=self.dht,
prefix=f"{self.config.neuron.run_id}_state_averager",
optimizer=self.outer_optimizer,
params=self.model.parameters(),
initialize_optimizer=True,
offload_optimizer=self.offload_optimizer,
custom_gradients=self.offload_optimizer,
min_group_size=self.config.neuron.min_group_size,
min_matchmaking_time=30.0,
request_timeout=10.0,
next_chunk_timeout=45.0,
allreduce_timeout=self.allreduce_timeout - 30.0 - 15.0,
start=True,
)
bt.logging.info("Successfully Loaded Gradient Averager")
# Load a new gradient averager
self.grad_averager = DTGradAverager(
dht=self.dht,
main_parameters=self.state_averager.main_parameters,
offloaded_optimizer=self.state_averager.optimizer,
prefix=f"{self.config.neuron.run_id}_grad_averager",
compression=hivemind.Uniform8BitQuantization(),
state_compression=hivemind.Uniform8BitQuantization(),
min_group_size=self.config.neuron.min_group_size,
min_matchmaking_time=30.0,
request_timeout=10.0,
next_chunk_timeout=45.0,
allreduce_timeout=self.allreduce_timeout - 30.0 - 15.0,
start=True,
)
bt.logging.info("Successfully Loaded State Averager")
if reload_outer_optimizer:
optimizer_state = None
try:
optimizer_state = torch.load(
hf_hub_download(
repo_id=global_model_name,
filename="outer_optimizer.pt",
revision=".".join(revision.split(".")[:-1] + ["0"]),
),
weights_only=True,
map_location="cpu",
)
# Load optimizer state if available
if "optimizer_state_dict" in optimizer_state:
self.state_averager.optimizer.load_state_dict(
optimizer_state["optimizer_state_dict"]
)
bt.logging.info(
f"Successfully Loaded Outer Optimizer State From {global_model_name} For Revision {'.'.join(revision.split('.')[:-1] + ['0'])}"
)
except Exception as e:
bt.logging.warning(
f"No optimizer state found or failed to load: {str(e)}. Initializing fresh optimizer."
)
finally:
if isinstance(optimizer_state, dict):
keys = list(optimizer_state.keys())
for k in keys:
del optimizer_state[k]
gc.collect()
del optimizer_state
gc.collect()
torch.cuda.empty_cache()
self.avg_handler = AveragingHandler(
self.model,
self.inner_optimizer,
self.grad_averager,
self.state_averager,
self.retry_limit,
self.retry_delay,
self.uid,
self.config.neuron.local_batch_size_train,
self.config.neuron.local_batch_size_train_effective,
self.tokenizer,
self.device,
)
self.scaler = torch.amp.GradScaler(enabled=True)
if (self.local_progress.inner_step != 0) and ("." in revision):
self.state_averager.reset_main_parameters(
model_name,
revision=".".join(
revision.split(".")[:-1]
+ [str(get_min_local_inner_Step(self, model_name, epoch=epoch))]
),
)
bt.logging.debug(
f"CPU Memory After Loading State {psutil.virtual_memory().available / 10**9} GB"
)
def load_state_from_peer(
self,
repo_id=None,
epoch=None,
reload_inner_optimizer=True,
reload_outer_optimizer=True,
revision=None,
use_fallback_model=True,
):
try:
state_loaded = False
if epoch is None:
self.global_progress.epoch = get_global_epoch(self)
epoch = self.global_progress.epoch
if repo_id is None:
repo_id = self.config.neuron.global_model_name
self.local_progress.inner_step = get_local_inner_step(
self, repo_id, epoch=self.global_progress.epoch
)
bt.logging.debug("Model Weights Before Loading State")
current_model_weights_sample = copy.copy(
[layer for layer in self.model.parameters()][-2][-10:].tolist()
)
bt.logging.debug(current_model_weights_sample)
bt.logging.debug(f"Old Model Tag: {self.local_progress.epoch}")
if self.global_progress.epoch is not None:
bt.logging.debug(
f"Latest Model State Found On The HF Hub With The Tag: {self.global_progress.epoch}. Loading That Model State."
)
# Load model state with max retries
MAX_ATTEMPTS = 3
attempt = 0
while attempt < MAX_ATTEMPTS:
try:
load_model_optimizer_gradient_averager(
self,
local_model_name=repo_id,
epoch=epoch,
reload_inner_optimizer=reload_inner_optimizer,
reload_outer_optimizer=reload_outer_optimizer,
revision=revision,
use_fallback_model=use_fallback_model,
)
break
except Exception as e:
attempt += 1
if attempt == MAX_ATTEMPTS:
raise Exception(
f"Failed to load model after {MAX_ATTEMPTS} attempts: {str(e)}"
)
bt.logging.warning(
f"Failed to load model, retrying. Attempt {attempt}/{MAX_ATTEMPTS}. Error {str(e)}"
)
state_loaded = True
bt.logging.debug("Model Weights After Loading State")
new_model_weights_sample = copy.copy(
[layer for layer in self.model.parameters()][-2][-10:].tolist()
)
bt.logging.debug(new_model_weights_sample)
self.local_progress.epoch = epoch
self.local_progress.samples_accumulated = 0
bt.logging.debug(f"New Model Tag: {self.global_progress.epoch}")
# Clean up old cache
try:
cleanup_old_cache(self, repo_id, revision)
except Exception as e:
bt.logging.warning(f"Failed to cleanup cache: {str(e)}")
else:
bt.logging.debug(f"Model With Tag: {epoch} Does Not Exist")
return state_loaded
except Exception as e:
bt.logging.error(f"Error loading state: {str(e)}")
return False
# TODO Remove this if score_bandwidth is deprecated
async def load_state_from_miner(self, peer, timeout: Optional[float] = None):
metadata = None
hivemind_logger.info(f"Downloading parameters from peer {peer}")
try:
stub = self.grad_averager.get_stub(
self._p2p,
peer,
namespace=self.grad_averager.matchmaking_kwargs["prefix"],
)
stream = await stub.rpc_download_state_partial(averaging_pb2.DownloadRequest())
current_tensor_parts, tensors = [], []
# TODO merge this with hivemind.compression.deserialize_tensor_stream
async for message in aiter_with_timeout(stream, timeout=timeout):
if message.metadata:
metadata = self.grad_averager.serializer.loads(message.metadata)
if message.tensor_part.dtype and current_tensor_parts:
# tensor_part.dtype indicates the start of the new tensor, so we should wrap up this one
tensors.append(
deserialize_torch_tensor(
combine_from_streaming(current_tensor_parts)
)
)
current_tensor_parts = []
current_tensor_parts.append(message.tensor_part)
if current_tensor_parts:
tensors.append(
deserialize_torch_tensor(combine_from_streaming(current_tensor_parts))
)
if not metadata:
hivemind_logger.exception(f"Peer {peer} did not send its state")
return
hivemind_logger.info(f"Finished downloading state from {peer}")
return metadata, tensors
except Exception as e:
hivemind_logger.exception(f"Failed to download state from {peer} - {repr(e)}")
return None, None
def cleanup_old_cache(self, repo_id=None, current_revision=None):
"""Helper method to clean up old cache files"""
if repo_id is None:
repo_id = self.config.neuron.global_model_name
current_revision = self.model.config._commit_hash
cache_info = scan_cache_dir()
broken_cache_list = [str(warning) for warning in cache_info.warnings]
cache_dir = HF_HUB_CACHE
cache_dir = Path(cache_dir).expanduser().resolve()
bt.logging.info("Cache clearing warnings:")
bt.logging.info(f"{cache_info.warnings}")
# Delete cache using preferred huggingface cache clearing method
if current_revision is None:
for cache in cache_dir.iterdir():
if repo_id.replace("/", "--") in str(cache):
bt.logging.info(f"Deleting the entire cache folder for repo {repo_id}.")
try:
shutil.rmtree(str(cache))
except OSError as e:
bt.logging.info(
"Error: %s - %s deleting the entire cache folder for the repo: %s"
% (e.filename, e.strerror, repo_id)
)
else:
for repo in cache_info.repos:
if repo.repo_id == repo_id:
revisions = sorted(
repo.revisions, key=lambda r: r.last_modified, reverse=True
)
bt.logging.info(
f"Found {len(revisions)} model revisions in .cache folder. Proceeding to delete all non-current revision."
)
for revision in revisions:
if (current_revision is not None) and (
revision.commit_hash == current_revision
):
bt.logging.info(
f"Skipping cache for current revision {revision.commit_hash}"
)
continue
else:
bt.logging.info(
f"Deleting cache for revision {revision.commit_hash}"
)
cache_info.delete_revisions(revision.commit_hash).execute()
break
# Forcefully remove the entire cache folder for a model if it's corrupted
if len(broken_cache_list) > 1:
for cache in cache_dir.iterdir():
if str(cache) in str(broken_cache_list):
bt.logging.info(
f"Found repo {repo_id} in HF cache warning message. Proceeding to delete the entire cache folder."
)
try:
shutil.rmtree(str(cache))
except OSError as e:
bt.logging.info(
"Error: %s - %s deleting the entire cache folder for the repo: %s"
% (e.filename, e.strerror, repo_id)
)
def upload_new_state(self, epoch: int, results: dict, block: int = None):
attempt = 0
while attempt < self.model_upload_retry_limit:
try:
bt.logging.info(
f"Pushing new model and optimizer state to HF Hub with tag {epoch}"
)
# Save and upload both model and optimizer state
upload_success = save_and_upload_state(
self, epoch=epoch, results=results, block=block
)
if upload_success:
# Verify the upload
updated_refs = list_repo_refs(
self.config.neuron.global_model_name,
repo_type="model",
)
new_tag = (
max(
[
int(tag.name.split(".")[1])
for tag in updated_refs.tags
if (
(len(tag.name.split(".")) == 3)
and (tag.name.split(".")[0] == __run__)
)
]
)
if updated_refs.tags
else 0
)
bt.logging.info(f"Successfully pushed new model with tag {new_tag}")
# Wait to allow out of sync miners to download new model state
time.sleep(self.load_state_timeout)
break
except HfHubHTTPError as e:
attempt += 1
bt.logging.info(f"{e}. Loading State from Peer.")
state_loaded = load_state_from_peer(self, epoch=self.global_progress.epoch)
if state_loaded:
break
except Exception:
attempt += 1
bt.logging.warning(
f"Failed To Upload Model To HF hub, Retrying. Attempt {attempt}/{self.model_upload_retry_limit}."
)
if attempt < self.model_upload_retry_limit:
time.sleep(self.model_upload_retry_delay)
else:
bt.logging.error(
"Maximum Retry Limit Reached. Unable To Upload Model To HF Hub."
)
raise
return upload_success
def save_and_upload_state(self, epoch: int, results: dict, block: int = None):
"""Unified function to save and upload both model and optimizer state"""
batch_size = sum(
[result for result in results["gathered"].values() if result is not None]
)
participating_peers = results["participating_peers"]
failed_peers = results["failed_peers"]
attempt = 0
while attempt < self.model_upload_retry_limit:
try:
with tempfile.TemporaryDirectory() as tmp_folder:
bt.logging.info(
f"Preparing model and optimizer state for epoch {epoch}"
)
if block is not None:
self.model.config.last_allreduce_block = block
self.model.config.inner_step = 0
self.model.save_pretrained(tmp_folder)
# Save outer optimizer state
outer_optimizer_state = {
"optimizer_state_dict": self.state_averager.optimizer.state_dict(),
"learning_rate": self.state_averager.optimizer.param_groups[0][
"lr"
],
"epoch": epoch,
}
torch.save(
outer_optimizer_state,
os.path.join(tmp_folder, "outer_optimizer.pt"),
)
# Save outer optimizer state
inner_optimizer_state = {
"optimizer_state_dict": self.inner_optimizer.state_dict(),
"learning_rate": self.inner_optimizer.param_groups[0]["lr"],
"scheduler_state": self.scheduler.state_dict(),
"epoch": epoch,
}
torch.save(
inner_optimizer_state,
os.path.join(tmp_folder, "inner_optimizer.pt"),
)
bt.logging.info(
f"Uploading model and optimizer states to repo: {self.config.neuron.global_model_name}"
)
# Upload everything in one go
commit_message = f"Run {__run__}. Outer Step {epoch}. Inner Step {0}. Peers {len(participating_peers) - len(failed_peers)}."
upload_folder(
folder_path=tmp_folder,
repo_id=self.config.neuron.global_model_name,
repo_type="model",
commit_message=commit_message,
)
# Create a tag for this version
create_tag(
self.config.neuron.global_model_name,
repo_type="model",
tag=f"{__run__}.{epoch}.{0}",
tag_message=commit_message,
)
bt.logging.info(
f"Successfully pushed new model and optimizer state with tag {epoch} to repo: {self.config.neuron.global_model_name}"
)
return True
except Exception as e:
attempt += 1
bt.logging.warning(
f"Failed to upload state to HF hub, Retrying. Attempt {attempt}/{self.model_upload_retry_limit}. Error: {str(e)}"
)
if attempt < self.model_upload_retry_limit:
time.sleep(self.model_upload_retry_delay)
else:
bt.logging.error(
"Maximum retry limit reached. Unable to upload state to HF Hub."
)
raise
return False
def get_top_uid(self):
all_reduce_scores_uids = [
k
for k, v in self.allreduce_status_dict.items()
if (v == "SUCCESS")
and (self.uid_tracker[int(k)]["model_huggingface_id"] is not None)
and (
(
datetime.now(pytz.utc)
- list_repo_commits(
self.uid_tracker[int(k)]["model_huggingface_id"], repo_type="model"
)[0].created_at
).seconds
< (60 * 60)
)
]
top_uid_list = [
k
for k, v in sorted(
{
u: self.metagraph.incentive[int(u)].item()
for u in all_reduce_scores_uids
}.items(),
key=lambda item: item[1],
)
]
if top_uid_list != []:
top_uid = top_uid_list[-1]
bt.logging.info(f"Top UID Identified As {top_uid}")
return top_uid
|