m-a-p/OpenCodeInterpreter-DS-1.3B

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 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: python zero_to_fp32.py . pytorch_model.bin14 15import argparse16import torch17import glob18import math19import os20import re21from collections import OrderedDict22from dataclasses import dataclass23 24# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with25# DeepSpeed data structures it has to be available in the current python environment.26from deepspeed.utils import logger27from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,28                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,29                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)30 31 32@dataclass33class zero_model_state:34    buffers: dict()35    param_shapes: dict()36    shared_params: list37    ds_version: int38    frozen_param_shapes: dict()39    frozen_param_fragments: dict()40 41 42debug = 043 44# load to cpu45device = torch.device('cpu')46 47 48def atoi(text):49    return int(text) if text.isdigit() else text50 51 52def natural_keys(text):53    '''54    alist.sort(key=natural_keys) sorts in human order55    http://nedbatchelder.com/blog/200712/human_sorting.html56    (See Toothy's implementation in the comments)57    '''58    return [atoi(c) for c in re.split(r'(\d+)', text)]59 60 61def get_model_state_file(checkpoint_dir, zero_stage):62    if not os.path.isdir(checkpoint_dir):63        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")64 65    # there should be only one file66    if zero_stage == 2:67        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")68    elif zero_stage == 3:69        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")70 71    if not os.path.exists(file):72        raise FileNotFoundError(f"can't find model states file at '{file}'")73 74    return file75 76 77def get_checkpoint_files(checkpoint_dir, glob_pattern):78    # XXX: need to test that this simple glob rule works for multi-node setup too79    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)80 81    if len(ckpt_files) == 0:82        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")83 84    return ckpt_files85 86 87def get_optim_files(checkpoint_dir):88    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")89 90 91def get_model_state_files(checkpoint_dir):92    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")93 94 95def parse_model_states(files):96    zero_model_states = []97    for file in files:98        state_dict = torch.load(file, map_location=device)99 100        if BUFFER_NAMES not in state_dict:101            raise ValueError(f"{file} is not a model state checkpoint")102        buffer_names = state_dict[BUFFER_NAMES]103        if debug:104            print("Found buffers:", buffer_names)105 106        # recover just the buffers while restoring them to fp32 if they were saved in fp16107        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}108        param_shapes = state_dict[PARAM_SHAPES]109 110        # collect parameters that are included in param_shapes111        param_names = []112        for s in param_shapes:113            for name in s.keys():114                param_names.append(name)115 116        # update with frozen parameters117        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)118        if frozen_param_shapes is not None:119            if debug:120                print(f"Found frozen_param_shapes: {frozen_param_shapes}")121            param_names += list(frozen_param_shapes.keys())122 123        # handle shared params124        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]125 126        ds_version = state_dict.get(DS_VERSION, None)127 128        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)129 130        z_model_state = zero_model_state(buffers=buffers,131                                         param_shapes=param_shapes,132                                         shared_params=shared_params,133                                         ds_version=ds_version,134                                         frozen_param_shapes=frozen_param_shapes,135                                         frozen_param_fragments=frozen_param_fragments)136        zero_model_states.append(z_model_state)137 138    return zero_model_states139 140 141def parse_optim_states(files, ds_checkpoint_dir):142 143    total_files = len(files)144    state_dicts = []145    for f in files:146        state_dicts.append(torch.load(f, map_location=device))147 148    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:149        raise ValueError(f"{files[0]} is not a zero checkpoint")150    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]151    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]152 153    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert154    # parameters can be different from data parallelism for non-expert parameters. So we can just155    # use the max of the partition_count to get the dp world_size.156 157    if type(world_size) is list:158        world_size = max(world_size)159 160    if world_size != total_files:161        raise ValueError(162            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "163            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."164        )165 166    # the groups are named differently in each stage167    if zero_stage == 2:168        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS169    elif zero_stage == 3:170        fp32_groups_key = FP32_FLAT_GROUPS171    else:172        raise ValueError(f"unknown zero stage {zero_stage}")173 174    if zero_stage == 2:175        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]176    elif zero_stage == 3:177        # if there is more than one param group, there will be multiple flattened tensors - one178        # flattened tensor per group - for simplicity merge them into a single tensor179        #180        # XXX: could make the script more memory efficient for when there are multiple groups - it181        # will require matching the sub-lists of param_shapes for each param group flattened tensor182 183        fp32_flat_groups = [184            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))185        ]186 187    return zero_stage, world_size, fp32_flat_groups188 189 190def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):191    """192    Returns fp32 state_dict reconstructed from ds checkpoint193 194    Args:195        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)196 197    """198    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")199 200    optim_files = get_optim_files(ds_checkpoint_dir)201    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)202    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")203 204    model_files = get_model_state_files(ds_checkpoint_dir)205 206    zero_model_states = parse_model_states(model_files)207    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')208 209    if zero_stage == 2:210        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)211    elif zero_stage == 3:212        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)213 214 215def _zero2_merge_frozen_params(state_dict, zero_model_states):216    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:217        return218 219    frozen_param_shapes = zero_model_states[0].frozen_param_shapes220    frozen_param_fragments = zero_model_states[0].frozen_param_fragments221 222    if debug:223        num_elem = sum(s.numel() for s in frozen_param_shapes.values())224        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')225 226        wanted_params = len(frozen_param_shapes)227        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())228        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])229        print(f'Frozen params: Have {avail_numel} numels to process.')230        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')231 232    total_params = 0233    total_numel = 0234    for name, shape in frozen_param_shapes.items():235        total_params += 1236        unpartitioned_numel = shape.numel()237        total_numel += unpartitioned_numel238 239        state_dict[name] = frozen_param_fragments[name]240 241        if debug:242            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")243 244    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")245 246 247def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):248    param_shapes = zero_model_states[0].param_shapes249 250    # Reconstruction protocol:251    #252    # XXX: document this253 254    if debug:255        for i in range(world_size):256            for j in range(len(fp32_flat_groups[0])):257                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")258 259    # XXX: memory usage doubles here (zero2)260    num_param_groups = len(fp32_flat_groups[0])261    merged_single_partition_of_fp32_groups = []262    for i in range(num_param_groups):263        merged_partitions = [sd[i] for sd in fp32_flat_groups]264        full_single_fp32_vector = torch.cat(merged_partitions, 0)265        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)266    avail_numel = sum(267        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])268 269    if debug:270        wanted_params = sum([len(shapes) for shapes in param_shapes])271        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])272        # not asserting if there is a mismatch due to possible padding273        print(f"Have {avail_numel} numels to process.")274        print(f"Need {wanted_numel} numels in {wanted_params} params.")275 276    # params277    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support278    # out-of-core computing solution279    total_numel = 0280    total_params = 0281    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):282        offset = 0283        avail_numel = full_single_fp32_vector.numel()284        for name, shape in shapes.items():285 286            unpartitioned_numel = shape.numel()287            total_numel += unpartitioned_numel288            total_params += 1289 290            if debug:291                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")292            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)293            offset += unpartitioned_numel294 295        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and296        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex297        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the298        # live optimizer object, so we are checking that the numbers are within the right range299        align_to = 2 * world_size300 301        def zero2_align(x):302            return align_to * math.ceil(x / align_to)303 304        if debug:305            print(f"original offset={offset}, avail_numel={avail_numel}")306 307        offset = zero2_align(offset)308        avail_numel = zero2_align(avail_numel)309 310        if debug:311            print(f"aligned  offset={offset}, avail_numel={avail_numel}")312 313        # Sanity check314        if offset != avail_numel:315            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")316 317    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")318 319 320def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):321    state_dict = OrderedDict()322 323    # buffers324    buffers = zero_model_states[0].buffers325    state_dict.update(buffers)326    if debug:327        print(f"added {len(buffers)} buffers")328 329    _zero2_merge_frozen_params(state_dict, zero_model_states)330 331    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)332 333    # recover shared parameters334    for pair in zero_model_states[0].shared_params:335        if pair[1] in state_dict:336            state_dict[pair[0]] = state_dict[pair[1]]337 338    return state_dict339 340 341def zero3_partitioned_param_info(unpartitioned_numel, world_size):342    remainder = unpartitioned_numel % world_size343    padding_numel = (world_size - remainder) if remainder else 0344    partitioned_numel = math.ceil(unpartitioned_numel / world_size)345    return partitioned_numel, padding_numel346 347 348def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):349    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:350        return351 352    if debug:353        for i in range(world_size):354            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())355            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')356 357        frozen_param_shapes = zero_model_states[0].frozen_param_shapes358        wanted_params = len(frozen_param_shapes)359        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())360        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size361        print(f'Frozen params: Have {avail_numel} numels to process.')362        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')363 364    total_params = 0365    total_numel = 0366    for name, shape in zero_model_states[0].frozen_param_shapes.items():367        total_params += 1368        unpartitioned_numel = shape.numel()369        total_numel += unpartitioned_numel370 371        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)372        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)373 374        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)375 376        if debug:377            print(378                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"379            )380 381    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")382 383 384def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):385    param_shapes = zero_model_states[0].param_shapes386    avail_numel = fp32_flat_groups[0].numel() * world_size387    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each388    # param, re-consolidating each param, while dealing with padding if any389 390    # merge list of dicts, preserving order391    param_shapes = {k: v for d in param_shapes for k, v in d.items()}392 393    if debug:394        for i in range(world_size):395            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")396 397        wanted_params = len(param_shapes)398        wanted_numel = sum(shape.numel() for shape in param_shapes.values())399        # not asserting if there is a mismatch due to possible padding400        avail_numel = fp32_flat_groups[0].numel() * world_size401        print(f"Trainable params: Have {avail_numel} numels to process.")402        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")403 404    # params405    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support406    # out-of-core computing solution407    offset = 0408    total_numel = 0409    total_params = 0410    for name, shape in param_shapes.items():411 412        unpartitioned_numel = shape.numel()413        total_numel += unpartitioned_numel414        total_params += 1415 416        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)417 418        if debug:419            print(420                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"421            )422 423        # XXX: memory usage doubles here424        state_dict[name] = torch.cat(425            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),426            0).narrow(0, 0, unpartitioned_numel).view(shape)427        offset += partitioned_numel428 429    offset *= world_size430 431    # Sanity check432    if offset != avail_numel:433        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")434 435    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")436 437 438def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):439    state_dict = OrderedDict()440 441    # buffers442    buffers = zero_model_states[0].buffers443    state_dict.update(buffers)444    if debug:445        print(f"added {len(buffers)} buffers")446 447    _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)448 449    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)450 451    # recover shared parameters452    for pair in zero_model_states[0].shared_params:453        if pair[1] in state_dict:454            state_dict[pair[0]] = state_dict[pair[1]]455 456    return state_dict457 458 459def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):460    """461    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with462    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example463    via a model hub.464 465    Args:466        - ``checkpoint_dir``: path to the desired checkpoint folder467        - ``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``468 469    Returns:470        - pytorch ``state_dict``471 472    Note: this approach may not work if your application doesn't have sufficient free CPU memory and473    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with474    the checkpoint.475 476    A typical usage might be ::477 478        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint479        # do the training and checkpoint saving480        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu481        model = model.cpu() # move to cpu482        model.load_state_dict(state_dict)483        # submit to model hub or save the model to share with others484 485    In this example the ``model`` will no longer be usable in the deepspeed context of the same486    application. i.e. you will need to re-initialize the deepspeed engine, since487    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.488 489    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.490 491    """492    if tag is None:493        latest_path = os.path.join(checkpoint_dir, 'latest')494        if os.path.isfile(latest_path):495            with open(latest_path, 'r') as fd:496                tag = fd.read().strip()497        else:498            raise ValueError(f"Unable to find 'latest' file at {latest_path}")499 500    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)501 502    if not os.path.isdir(ds_checkpoint_dir):503        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")504 505    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)506 507 508def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):509    """510    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be511    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.512 513    Args:514        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)515        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)516        - ``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``517    """518 519    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)520    print(f"Saving fp32 state dict to {output_file}")521    torch.save(state_dict, output_file)522 523 524def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):525    """526    1. Put the provided model to cpu527    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``528    3. Load it into the provided model529 530    Args:531        - ``model``: the model object to update532        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)533        - ``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``534 535    Returns:536        - ``model`: modified model537 538    Make sure you have plenty of CPU memory available before you call this function. If you don't539    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it540    conveniently placed for you in the checkpoint folder.541 542    A typical usage might be ::543 544        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint545        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)546        # submit to model hub or save the model to share with others547 548    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context549    of the same application. i.e. you will need to re-initialize the deepspeed engine, since550    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.551 552    """553    logger.info(f"Extracting fp32 weights")554    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)555 556    logger.info(f"Overwriting model with fp32 weights")557    model = model.cpu()558    model.load_state_dict(state_dict, strict=False)559 560    return model561 562 563if __name__ == "__main__":564 565    parser = argparse.ArgumentParser()566    parser.add_argument("checkpoint_dir",567                        type=str,568                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")569    parser.add_argument(570        "output_file",571        type=str,572        help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")573    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")574    args = parser.parse_args()575 576    debug = args.debug577 578    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file)579