flagopen/starcoder-15b-taco

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: 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_dict = torch.load(f, map_location=device)147        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights148        # and also handle the case where it was already removed by another helper script149        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)150        state_dicts.append(state_dict)151 152    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:153        raise ValueError(f"{files[0]} is not a zero checkpoint")154    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]155    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]156 157    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert158    # parameters can be different from data parallelism for non-expert parameters. So we can just159    # use the max of the partition_count to get the dp world_size.160 161    if type(world_size) is list:162        world_size = max(world_size)163 164    if world_size != total_files:165        raise ValueError(166            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "167            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."168        )169 170    # the groups are named differently in each stage171    if zero_stage <= 2:172        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS173    elif zero_stage == 3:174        fp32_groups_key = FP32_FLAT_GROUPS175    else:176        raise ValueError(f"unknown zero stage {zero_stage}")177 178    if zero_stage <= 2:179        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]180    elif zero_stage == 3:181        # if there is more than one param group, there will be multiple flattened tensors - one182        # flattened tensor per group - for simplicity merge them into a single tensor183        #184        # XXX: could make the script more memory efficient for when there are multiple groups - it185        # will require matching the sub-lists of param_shapes for each param group flattened tensor186 187        fp32_flat_groups = [188            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))189        ]190 191    return zero_stage, world_size, fp32_flat_groups192 193 194def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):195    """196    Returns fp32 state_dict reconstructed from ds checkpoint197 198    Args:199        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)200 201    """202    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")203 204    optim_files = get_optim_files(ds_checkpoint_dir)205    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)206    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")207 208    model_files = get_model_state_files(ds_checkpoint_dir)209 210    zero_model_states = parse_model_states(model_files)211    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')212 213    if zero_stage <= 2:214        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)215    elif zero_stage == 3:216        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)217 218 219def _zero2_merge_frozen_params(state_dict, zero_model_states):220    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:221        return222 223    frozen_param_shapes = zero_model_states[0].frozen_param_shapes224    frozen_param_fragments = zero_model_states[0].frozen_param_fragments225 226    if debug:227        num_elem = sum(s.numel() for s in frozen_param_shapes.values())228        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')229 230        wanted_params = len(frozen_param_shapes)231        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())232        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])233        print(f'Frozen params: Have {avail_numel} numels to process.')234        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')235 236    total_params = 0237    total_numel = 0238    for name, shape in frozen_param_shapes.items():239        total_params += 1240        unpartitioned_numel = shape.numel()241        total_numel += unpartitioned_numel242 243        state_dict[name] = frozen_param_fragments[name]244 245        if debug:246            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")247 248    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")249 250 251def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):252    param_shapes = zero_model_states[0].param_shapes253 254    # Reconstruction protocol:255    #256    # XXX: document this257 258    if debug:259        for i in range(world_size):260            for j in range(len(fp32_flat_groups[0])):261                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")262 263    # XXX: memory usage doubles here (zero2)264    num_param_groups = len(fp32_flat_groups[0])265    merged_single_partition_of_fp32_groups = []266    for i in range(num_param_groups):267        merged_partitions = [sd[i] for sd in fp32_flat_groups]268        full_single_fp32_vector = torch.cat(merged_partitions, 0)269        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)270    avail_numel = sum(271        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])272 273    if debug:274        wanted_params = sum([len(shapes) for shapes in param_shapes])275        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])276        # not asserting if there is a mismatch due to possible padding277        print(f"Have {avail_numel} numels to process.")278        print(f"Need {wanted_numel} numels in {wanted_params} params.")279 280    # params281    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support282    # out-of-core computing solution283    total_numel = 0284    total_params = 0285    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):286        offset = 0287        avail_numel = full_single_fp32_vector.numel()288        for name, shape in shapes.items():289 290            unpartitioned_numel = shape.numel()291            total_numel += unpartitioned_numel292            total_params += 1293 294            if debug:295                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")296            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)297            offset += unpartitioned_numel298 299        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and300        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex301        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the302        # live optimizer object, so we are checking that the numbers are within the right range303        align_to = 2 * world_size304 305        def zero2_align(x):306            return align_to * math.ceil(x / align_to)307 308        if debug:309            print(f"original offset={offset}, avail_numel={avail_numel}")310 311        offset = zero2_align(offset)312        avail_numel = zero2_align(avail_numel)313 314        if debug:315            print(f"aligned  offset={offset}, avail_numel={avail_numel}")316 317        # Sanity check318        if offset != avail_numel:319            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")320 321    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")322 323 324def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):325    state_dict = OrderedDict()326 327    # buffers328    buffers = zero_model_states[0].buffers329    state_dict.update(buffers)330    if debug:331        print(f"added {len(buffers)} buffers")332 333    _zero2_merge_frozen_params(state_dict, zero_model_states)334 335    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)336 337    # recover shared parameters338    for pair in zero_model_states[0].shared_params:339        if pair[1] in state_dict:340            state_dict[pair[0]] = state_dict[pair[1]]341 342    return state_dict343 344 345def zero3_partitioned_param_info(unpartitioned_numel, world_size):346    remainder = unpartitioned_numel % world_size347    padding_numel = (world_size - remainder) if remainder else 0348    partitioned_numel = math.ceil(unpartitioned_numel / world_size)349    return partitioned_numel, padding_numel350 351 352def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):353    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:354        return355 356    if debug:357        for i in range(world_size):358            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())359            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')360 361        frozen_param_shapes = zero_model_states[0].frozen_param_shapes362        wanted_params = len(frozen_param_shapes)363        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())364        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size365        print(f'Frozen params: Have {avail_numel} numels to process.')366        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')367 368    total_params = 0369    total_numel = 0370    for name, shape in zero_model_states[0].frozen_param_shapes.items():371        total_params += 1372        unpartitioned_numel = shape.numel()373        total_numel += unpartitioned_numel374 375        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)376        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)377 378        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)379 380        if debug:381            print(382                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"383            )384 385    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")386 387 388def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):389    param_shapes = zero_model_states[0].param_shapes390    avail_numel = fp32_flat_groups[0].numel() * world_size391    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each392    # param, re-consolidating each param, while dealing with padding if any393 394    # merge list of dicts, preserving order395    param_shapes = {k: v for d in param_shapes for k, v in d.items()}396 397    if debug:398        for i in range(world_size):399            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")400 401        wanted_params = len(param_shapes)402        wanted_numel = sum(shape.numel() for shape in param_shapes.values())403        # not asserting if there is a mismatch due to possible padding404        avail_numel = fp32_flat_groups[0].numel() * world_size405        print(f"Trainable params: Have {avail_numel} numels to process.")406        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")407 408    # params409    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support410    # out-of-core computing solution411    offset = 0412    total_numel = 0413    total_params = 0414    for name, shape in param_shapes.items():415 416        unpartitioned_numel = shape.numel()417        total_numel += unpartitioned_numel418        total_params += 1419 420        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)421 422        if debug:423            print(424                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"425            )426 427        # XXX: memory usage doubles here428        state_dict[name] = torch.cat(429            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),430            0).narrow(0, 0, unpartitioned_numel).view(shape)431        offset += partitioned_numel432 433    offset *= world_size434 435    # Sanity check436    if offset != avail_numel:437        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")438 439    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")440 441 442def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):443    state_dict = OrderedDict()444 445    # buffers446    buffers = zero_model_states[0].buffers447    state_dict.update(buffers)448    if debug:449        print(f"added {len(buffers)} buffers")450 451    _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)452 453    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)454 455    # recover shared parameters456    for pair in zero_model_states[0].shared_params:457        if pair[1] in state_dict:458            state_dict[pair[0]] = state_dict[pair[1]]459 460    return state_dict461 462 463def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):464    """465    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with466    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example467    via a model hub.468 469    Args:470        - ``checkpoint_dir``: path to the desired checkpoint folder471        - ``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``472 473    Returns:474        - pytorch ``state_dict``475 476    Note: this approach may not work if your application doesn't have sufficient free CPU memory and477    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with478    the checkpoint.479 480    A typical usage might be ::481 482        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint483        # do the training and checkpoint saving484        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu485        model = model.cpu() # move to cpu486        model.load_state_dict(state_dict)487        # submit to model hub or save the model to share with others488 489    In this example the ``model`` will no longer be usable in the deepspeed context of the same490    application. i.e. you will need to re-initialize the deepspeed engine, since491    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.492 493    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.494 495    """496    if tag is None:497        latest_path = os.path.join(checkpoint_dir, 'latest')498        if os.path.isfile(latest_path):499            with open(latest_path, 'r') as fd:500                tag = fd.read().strip()501        else:502            raise ValueError(f"Unable to find 'latest' file at {latest_path}")503 504    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)505 506    if not os.path.isdir(ds_checkpoint_dir):507        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")508 509    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)510 511 512def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):513    """514    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be515    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.516 517    Args:518        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)519        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)520        - ``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``521    """522 523    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)524    print(f"Saving fp32 state dict to {output_file}")525    torch.save(state_dict, output_file)526 527 528def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):529    """530    1. Put the provided model to cpu531    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``532    3. Load it into the provided model533 534    Args:535        - ``model``: the model object to update536        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)537        - ``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``538 539    Returns:540        - ``model`: modified model541 542    Make sure you have plenty of CPU memory available before you call this function. If you don't543    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it544    conveniently placed for you in the checkpoint folder.545 546    A typical usage might be ::547 548        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint549        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)550        # submit to model hub or save the model to share with others551 552    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context553    of the same application. i.e. you will need to re-initialize the deepspeed engine, since554    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.555 556    """557    logger.info(f"Extracting fp32 weights")558    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)559 560    logger.info(f"Overwriting model with fp32 weights")561    model = model.cpu()562    model.load_state_dict(state_dict, strict=False)563 564    return model565 566 567if __name__ == "__main__":568 569    parser = argparse.ArgumentParser()570    parser.add_argument("checkpoint_dir",571                        type=str,572                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")573    parser.add_argument(574        "output_file",575        type=str,576        help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")577    parser.add_argument("-t",578                        "--tag",579                        type=str,580                        default=None,581                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")582    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")583    args = parser.parse_args()584 585    debug = args.debug586 587    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)588