Arc-Intelligence/ATLAS-8B-Instruct
220
1#!/usr/bin/env python2 3# Copyright (c) Microsoft Corporation.4# SPDX-License-Identifier: Apache-2.05 6# DeepSpeed Team7 8# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets9# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in10# the future. Once extracted, the weights don't require DeepSpeed and can be used in any11# application.12#13# example:14# python zero_to_fp32.py . output_dir/15# or16# python zero_to_fp32.py . output_dir/ --safe_serialization17 18import argparse19import torch20import glob21import math22import os23import re24import json25from tqdm import tqdm26from collections import OrderedDict27from dataclasses import dataclass28 29# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with30# DeepSpeed data structures it has to be available in the current python environment.31from deepspeed.utils import logger32from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,33 FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,34 FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)35 36 37@dataclass38class zero_model_state:39 buffers: dict()40 param_shapes: dict()41 shared_params: list42 ds_version: int43 frozen_param_shapes: dict()44 frozen_param_fragments: dict()45 46 47debug = 048 49# load to cpu50device = torch.device('cpu')51 52 53def atoi(text):54 return int(text) if text.isdigit() else text55 56 57def natural_keys(text):58 '''59 alist.sort(key=natural_keys) sorts in human order60 http://nedbatchelder.com/blog/200712/human_sorting.html61 (See Toothy's implementation in the comments)62 '''63 return [atoi(c) for c in re.split(r'(\d+)', text)]64 65 66def get_model_state_file(checkpoint_dir, zero_stage):67 if not os.path.isdir(checkpoint_dir):68 raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")69 70 # there should be only one file71 if zero_stage <= 2:72 file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")73 elif zero_stage == 3:74 file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")75 76 if not os.path.exists(file):77 raise FileNotFoundError(f"can't find model states file at '{file}'")78 79 return file80 81 82def get_checkpoint_files(checkpoint_dir, glob_pattern):83 # XXX: need to test that this simple glob rule works for multi-node setup too84 ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)85 86 if len(ckpt_files) == 0:87 raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")88 89 return ckpt_files90 91 92def get_optim_files(checkpoint_dir):93 return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")94 95 96def get_model_state_files(checkpoint_dir):97 return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")98 99 100def parse_model_states(files):101 zero_model_states = []102 for file in files:103 state_dict = torch.load(file, map_location=device)104 105 if BUFFER_NAMES not in state_dict:106 raise ValueError(f"{file} is not a model state checkpoint")107 buffer_names = state_dict[BUFFER_NAMES]108 if debug:109 print("Found buffers:", buffer_names)110 111 # recover just the buffers while restoring them to fp32 if they were saved in fp16112 buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}113 param_shapes = state_dict[PARAM_SHAPES]114 115 # collect parameters that are included in param_shapes116 param_names = []117 for s in param_shapes:118 for name in s.keys():119 param_names.append(name)120 121 # update with frozen parameters122 frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)123 if frozen_param_shapes is not None:124 if debug:125 print(f"Found frozen_param_shapes: {frozen_param_shapes}")126 param_names += list(frozen_param_shapes.keys())127 128 # handle shared params129 shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]130 131 ds_version = state_dict.get(DS_VERSION, None)132 133 frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)134 135 z_model_state = zero_model_state(buffers=buffers,136 param_shapes=param_shapes,137 shared_params=shared_params,138 ds_version=ds_version,139 frozen_param_shapes=frozen_param_shapes,140 frozen_param_fragments=frozen_param_fragments)141 zero_model_states.append(z_model_state)142 143 return zero_model_states144 145 146def parse_optim_states(files, ds_checkpoint_dir):147 total_files = len(files)148 state_dicts = []149 for f in files:150 state_dict = torch.load(f, map_location=device)151 # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights152 # and also handle the case where it was already removed by another helper script153 state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)154 state_dicts.append(state_dict)155 156 if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:157 raise ValueError(f"{files[0]} is not a zero checkpoint")158 zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]159 world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]160 161 # For ZeRO-2 each param group can have different partition_count as data parallelism for expert162 # parameters can be different from data parallelism for non-expert parameters. So we can just163 # use the max of the partition_count to get the dp world_size.164 165 if type(world_size) is list:166 world_size = max(world_size)167 168 if world_size != total_files:169 raise ValueError(170 f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "171 "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."172 )173 174 # the groups are named differently in each stage175 if zero_stage <= 2:176 fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS177 elif zero_stage == 3:178 fp32_groups_key = FP32_FLAT_GROUPS179 else:180 raise ValueError(f"unknown zero stage {zero_stage}")181 182 if zero_stage <= 2:183 fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]184 elif zero_stage == 3:185 # if there is more than one param group, there will be multiple flattened tensors - one186 # flattened tensor per group - for simplicity merge them into a single tensor187 #188 # XXX: could make the script more memory efficient for when there are multiple groups - it189 # will require matching the sub-lists of param_shapes for each param group flattened tensor190 191 fp32_flat_groups = [192 torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))193 ]194 195 return zero_stage, world_size, fp32_flat_groups196 197 198def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):199 """200 Returns fp32 state_dict reconstructed from ds checkpoint201 202 Args:203 - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)204 205 """206 print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")207 208 optim_files = get_optim_files(ds_checkpoint_dir)209 zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)210 print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")211 212 model_files = get_model_state_files(ds_checkpoint_dir)213 214 zero_model_states = parse_model_states(model_files)215 print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')216 217 if zero_stage <= 2:218 return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,219 exclude_frozen_parameters)220 elif zero_stage == 3:221 return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,222 exclude_frozen_parameters)223 224 225def _zero2_merge_frozen_params(state_dict, zero_model_states):226 if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:227 return228 229 frozen_param_shapes = zero_model_states[0].frozen_param_shapes230 frozen_param_fragments = zero_model_states[0].frozen_param_fragments231 232 if debug:233 num_elem = sum(s.numel() for s in frozen_param_shapes.values())234 print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')235 236 wanted_params = len(frozen_param_shapes)237 wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())238 avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])239 print(f'Frozen params: Have {avail_numel} numels to process.')240 print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')241 242 total_params = 0243 total_numel = 0244 for name, shape in frozen_param_shapes.items():245 total_params += 1246 unpartitioned_numel = shape.numel()247 total_numel += unpartitioned_numel248 249 state_dict[name] = frozen_param_fragments[name]250 251 if debug:252 print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")253 254 print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")255 256 257def _has_callable(obj, fn):258 attr = getattr(obj, fn, None)259 return callable(attr)260 261 262def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):263 param_shapes = zero_model_states[0].param_shapes264 265 # Reconstruction protocol:266 #267 # XXX: document this268 269 if debug:270 for i in range(world_size):271 for j in range(len(fp32_flat_groups[0])):272 print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")273 274 # XXX: memory usage doubles here (zero2)275 num_param_groups = len(fp32_flat_groups[0])276 merged_single_partition_of_fp32_groups = []277 for i in range(num_param_groups):278 merged_partitions = [sd[i] for sd in fp32_flat_groups]279 full_single_fp32_vector = torch.cat(merged_partitions, 0)280 merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)281 avail_numel = sum(282 [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])283 284 if debug:285 wanted_params = sum([len(shapes) for shapes in param_shapes])286 wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])287 # not asserting if there is a mismatch due to possible padding288 print(f"Have {avail_numel} numels to process.")289 print(f"Need {wanted_numel} numels in {wanted_params} params.")290 291 # params292 # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support293 # out-of-core computing solution294 total_numel = 0295 total_params = 0296 for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):297 offset = 0298 avail_numel = full_single_fp32_vector.numel()299 for name, shape in shapes.items():300 301 unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)302 total_numel += unpartitioned_numel303 total_params += 1304 305 if debug:306 print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")307 state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)308 offset += unpartitioned_numel309 310 # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and311 # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex312 # paddings performed in the code it's almost impossible to predict the exact numbers w/o the313 # live optimizer object, so we are checking that the numbers are within the right range314 align_to = 2 * world_size315 316 def zero2_align(x):317 return align_to * math.ceil(x / align_to)318 319 if debug:320 print(f"original offset={offset}, avail_numel={avail_numel}")321 322 offset = zero2_align(offset)323 avail_numel = zero2_align(avail_numel)324 325 if debug:326 print(f"aligned offset={offset}, avail_numel={avail_numel}")327 328 # Sanity check329 if offset != avail_numel:330 raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")331 332 print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")333 334 335def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,336 exclude_frozen_parameters):337 state_dict = OrderedDict()338 339 # buffers340 buffers = zero_model_states[0].buffers341 state_dict.update(buffers)342 if debug:343 print(f"added {len(buffers)} buffers")344 345 if not exclude_frozen_parameters:346 _zero2_merge_frozen_params(state_dict, zero_model_states)347 348 _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)349 350 # recover shared parameters351 for pair in zero_model_states[0].shared_params:352 if pair[1] in state_dict:353 state_dict[pair[0]] = state_dict[pair[1]]354 355 return state_dict356 357 358def zero3_partitioned_param_info(unpartitioned_numel, world_size):359 remainder = unpartitioned_numel % world_size360 padding_numel = (world_size - remainder) if remainder else 0361 partitioned_numel = math.ceil(unpartitioned_numel / world_size)362 return partitioned_numel, padding_numel363 364 365def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):366 if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:367 return368 369 if debug:370 for i in range(world_size):371 num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())372 print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')373 374 frozen_param_shapes = zero_model_states[0].frozen_param_shapes375 wanted_params = len(frozen_param_shapes)376 wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())377 avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size378 print(f'Frozen params: Have {avail_numel} numels to process.')379 print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')380 381 total_params = 0382 total_numel = 0383 for name, shape in zero_model_states[0].frozen_param_shapes.items():384 total_params += 1385 unpartitioned_numel = shape.numel()386 total_numel += unpartitioned_numel387 388 param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)389 state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)390 391 partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)392 393 if debug:394 print(395 f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"396 )397 398 print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")399 400 401def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):402 param_shapes = zero_model_states[0].param_shapes403 avail_numel = fp32_flat_groups[0].numel() * world_size404 # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each405 # param, re-consolidating each param, while dealing with padding if any406 407 # merge list of dicts, preserving order408 param_shapes = {k: v for d in param_shapes for k, v in d.items()}409 410 if debug:411 for i in range(world_size):412 print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")413 414 wanted_params = len(param_shapes)415 wanted_numel = sum(shape.numel() for shape in param_shapes.values())416 # not asserting if there is a mismatch due to possible padding417 avail_numel = fp32_flat_groups[0].numel() * world_size418 print(f"Trainable params: Have {avail_numel} numels to process.")419 print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")420 421 # params422 # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support423 # out-of-core computing solution424 offset = 0425 total_numel = 0426 total_params = 0427 for name, shape in tqdm(param_shapes.items(), desc='Gathering Sharded Weights'):428 unpartitioned_numel = shape.numel()429 total_numel += unpartitioned_numel430 total_params += 1431 partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)432 433 if debug:434 print(435 f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"436 )437 438 # XXX: memory usage doubles here439 state_dict[name] = torch.cat(440 tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),441 0).narrow(0, 0, unpartitioned_numel).view(shape)442 offset += partitioned_numel443 444 offset *= world_size445 446 # Sanity check447 if offset != avail_numel:448 raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")449 450 print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")451 452 453def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,454 exclude_frozen_parameters):455 state_dict = OrderedDict()456 457 # buffers458 buffers = zero_model_states[0].buffers459 state_dict.update(buffers)460 if debug:461 print(f"added {len(buffers)} buffers")462 463 if not exclude_frozen_parameters:464 _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)465 466 _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)467 468 # recover shared parameters469 for pair in zero_model_states[0].shared_params:470 if pair[1] in state_dict:471 state_dict[pair[0]] = state_dict[pair[1]]472 473 return state_dict474 475 476def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None, exclude_frozen_parameters=False):477 """478 Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with479 ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example480 via a model hub.481 482 Args:483 - ``checkpoint_dir``: path to the desired checkpoint folder484 - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``485 - ``exclude_frozen_parameters``: exclude frozen parameters486 487 Returns:488 - pytorch ``state_dict``489 490 Note: this approach may not work if your application doesn't have sufficient free CPU memory and491 you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with492 the checkpoint.493 494 A typical usage might be ::495 496 from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint497 # do the training and checkpoint saving498 state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu499 model = model.cpu() # move to cpu500 model.load_state_dict(state_dict)501 # submit to model hub or save the model to share with others502 503 In this example the ``model`` will no longer be usable in the deepspeed context of the same504 application. i.e. you will need to re-initialize the deepspeed engine, since505 ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.506 507 If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.508 509 """510 if tag is None:511 latest_path = os.path.join(checkpoint_dir, 'latest')512 if os.path.isfile(latest_path):513 with open(latest_path, 'r') as fd:514 tag = fd.read().strip()515 else:516 raise ValueError(f"Unable to find 'latest' file at {latest_path}")517 518 ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)519 520 if not os.path.isdir(ds_checkpoint_dir):521 raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")522 523 return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)524 525 526def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,527 output_dir,528 max_shard_size="5GB",529 safe_serialization=False,530 tag=None,531 exclude_frozen_parameters=False):532 """533 Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be534 loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.535 536 Args:537 - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)538 - ``output_dir``: directory to the pytorch fp32 state_dict output files539 - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB540 - ``safe_serialization``: whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).541 - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``542 - ``exclude_frozen_parameters``: exclude frozen parameters543 """544 # Dependency pre-check545 if safe_serialization:546 try:547 from safetensors.torch import save_file548 except ImportError:549 print('If you want to use `safe_serialization`, please `pip install safetensors`')550 raise551 if max_shard_size is not None:552 try:553 from huggingface_hub import split_torch_state_dict_into_shards554 except ImportError:555 print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')556 raise557 558 # Convert zero checkpoint to state_dict559 state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag, exclude_frozen_parameters)560 561 # Shard the model if it is too big.562 weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"563 if max_shard_size is not None:564 filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")565 state_dict_split = split_torch_state_dict_into_shards(state_dict,566 filename_pattern=filename_pattern,567 max_shard_size=max_shard_size)568 else:569 from collections import namedtuple570 StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])571 state_dict_split = StateDictSplit(is_sharded=False,572 filename_to_tensors={weights_name: list(state_dict.keys())})573 574 # Save the model575 filename_to_tensors = state_dict_split.filename_to_tensors.items()576 for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):577 shard = {tensor: state_dict[tensor].contiguous() for tensor in tensors}578 output_path = os.path.join(output_dir, shard_file)579 if safe_serialization:580 save_file(shard, output_path, metadata={"format": "pt"})581 else:582 torch.save(shard, output_path)583 584 # Save index if sharded585 if state_dict_split.is_sharded:586 index = {587 "metadata": state_dict_split.metadata,588 "weight_map": state_dict_split.tensor_to_filename,589 }590 save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"591 save_index_file = os.path.join(output_dir, save_index_file)592 with open(save_index_file, "w", encoding="utf-8") as f:593 content = json.dumps(index, indent=2, sort_keys=True) + "\n"594 f.write(content)595 596 597def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):598 """599 1. Put the provided model to cpu600 2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``601 3. Load it into the provided model602 603 Args:604 - ``model``: the model object to update605 - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)606 - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``607 608 Returns:609 - ``model`: modified model610 611 Make sure you have plenty of CPU memory available before you call this function. If you don't612 have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it613 conveniently placed for you in the checkpoint folder.614 615 A typical usage might be ::616 617 from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint618 model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)619 # submit to model hub or save the model to share with others620 621 Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context622 of the same application. i.e. you will need to re-initialize the deepspeed engine, since623 ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.624 625 """626 logger.info(f"Extracting fp32 weights")627 state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)628 629 logger.info(f"Overwriting model with fp32 weights")630 model = model.cpu()631 model.load_state_dict(state_dict, strict=False)632 633 return model634 635 636if __name__ == "__main__":637 parser = argparse.ArgumentParser()638 parser.add_argument("checkpoint_dir",639 type=str,640 help="path to the desired checkpoint folder, e.g., path/checkpoint-12")641 parser.add_argument("output_dir",642 type=str,643 help="directory to the pytorch fp32 state_dict output files"644 "(e.g. path/checkpoint-12-output/)")645 parser.add_argument(646 "--max_shard_size",647 type=str,648 default="5GB",649 help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"650 "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"651 "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"652 "without CPU OOM issues.")653 parser.add_argument(654 "--safe_serialization",655 default=False,656 action='store_true',657 help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")658 parser.add_argument("-t",659 "--tag",660 type=str,661 default=None,662 help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")663 parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")664 parser.add_argument("-d", "--debug", action='store_true', help="enable debug")665 args = parser.parse_args()666 667 debug = args.debug668 669 convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,670 args.output_dir,671 max_shard_size=args.max_shard_size,672 safe_serialization=args.safe_serialization,673 tag=args.tag,674 exclude_frozen_parameters=args.exclude_frozen_parameters)675