colossalai.py

import functools
import warnings
from typing import Optional

import torch
import torch.distributed as dist
import torch.nn as nn
from transformers.tokenization_utils_base import PreTrainedTokenizerBase

import colossalai
from colossalai.booster.plugin import GeminiPlugin, LowLevelZeroPlugin
from colossalai.booster.plugin.gemini_plugin import GeminiModel
from colossalai.booster.plugin.low_level_zero_plugin import LowLevelZeroModel
from colossalai.tensor import ProcessGroup, ShardSpec
from colossalai.utils import get_current_device
from colossalai.zero import ColoInitContext
from colossalai.zero.gemini.gemini_ddp import GeminiDDP

from .ddp import DDPStrategy


class ColossalAIStrategy(DDPStrategy):
    """
        The strategy for training with ColossalAI.

    Args:
        stage(int): The stage to use in ZeRO. Choose in (1, 2, 3)
        precision(str): The precision to use. Choose in ('fp32', 'fp16'). Stage 3 only supports fp16.
        seed(int): The seed for the random number generator.
        shard_init(bool): Whether to shard the model parameters during initialization. Only for ZeRO-3.
            This is not compatible with `from_pretrained()`. We temporarily disable this and will support it in the future.
        placement_policy(str): The placement policy for gemini. Choose in ('cpu', 'cuda')
                          If it is “cpu”, parameters, gradients and optimizer states will be offloaded to CPU,
                          If it is “cuda”, they will not be offloaded, which means max CUDA memory will be used. It is the fastest.
        pin_memory(bool): Whether to pin the memory for the data loader. Only for ZeRO-3.
        force_outputs_fp32(bool): Whether to force the outputs to be fp32. Only for ZeRO-3.
        search_range_mb(int): The search range in MB for the chunk size. Only for ZeRO-3.
        hidden_dim(optional, int): The hidden dimension for the gemini. Only for ZeRO-3.
        min_chunk_size_mb(float): The minimum chunk size in MB. Only for ZeRO-3.
        gpu_margin_mem_ratio(float): The margin memory ratio for the GPU. Only for ZeRO-3.
        reduce_bucket_size(int): The reduce bucket size in bytes. Only for ZeRO-1 and ZeRO-2.
        overlap_communication(bool): Whether to overlap communication and computation. Only for ZeRO-1 and ZeRO-2.
        initial_scale(float): The initial scale for the optimizer.
        growth_factor(float): The growth factor for the optimizer.
        backoff_factor(float): The backoff factor for the optimizer.
        growth_interval(int): The growth interval for the optimizer.
        hysteresis(int): The hysteresis for the optimizer.
        min_scale(float): The minimum scale for the optimizer.
        max_scale(float): The maximum scale for the optimizer.
        max_norm(float): The maximum norm for the optimizer.
        norm_type(float): The norm type for the optimizer.

    """

    def __init__(
            self,
            stage: int = 3,
            precision: str = 'fp16',
            seed: int = 42,
            shard_init: bool = False,    # only for stage 3
            placement_policy: str = 'cuda',
            pin_memory: bool = True,    # only for stage 3
            force_outputs_fp32: bool = False,    # only for stage 3
            search_range_mb: int = 32,    # only for stage 3
            hidden_dim: Optional[int] = None,    # only for stage 3
            min_chunk_size_mb: float = 32,    # only for stage 3
            gpu_margin_mem_ratio: float = 0.0,    # only for stage 3
            reduce_bucket_size: int = 12 * 1024**2,    # only for stage 1&2
            overlap_communication: bool = True,    # only for stage 1&2
            initial_scale: float = 2**16,
            growth_factor: float = 2,
            backoff_factor: float = 0.5,
            growth_interval: int = 1000,
            hysteresis: int = 2,
            min_scale: float = 1,
            max_scale: float = 2**32,
            max_norm: float = 0.0,
            norm_type: float = 2.0) -> None:

        assert stage in (1, 2, 3), f'Unsupported stage "{stage}"'
        assert placement_policy in ('cpu', 'cuda'), f'Unsupported placement policy "{placement_policy}"'
        assert precision in ('fp32', 'fp16'), f'Unsupported precision "{precision}"'

        # TODO(ver217): support shard_init when using from_pretrained()
        if shard_init:
            warnings.warn(
                f'Shard init is not supported model.from_pretrained() yet. '
                'Please load weights after strategy.prepare()'
            )
        if stage == 3 and precision == 'fp32':
            warnings.warn(f'Stage 3 only supports fp16. Precision is set to fp16.')
            precision = 'fp16'
        self.precision = precision
        self.shard_init = shard_init

        optim_kwargs = dict(
            initial_scale=initial_scale,
            growth_factor=growth_factor,
            backoff_factor=backoff_factor,
            growth_interval=growth_interval,
            hysteresis=hysteresis,
            min_scale=min_scale,
            max_scale=max_scale,
            max_norm=max_norm,
            norm_type=norm_type
        )
        # NOTE: dist should be initialized before calling get_current_device()
        if stage == 3:
            plugin_initializer = lambda: GeminiPlugin(
                # gemini_config
                device=get_current_device(),
                placement_policy=placement_policy,
                precision=precision,
                pin_memory=pin_memory,
                force_outputs_fp32=force_outputs_fp32,
                strict_ddp_mode=shard_init,
                search_range_mb=search_range_mb,
                hidden_dim=hidden_dim,
                min_chunk_size_mb=min_chunk_size_mb,
                # zero_optim_config
                gpu_margin_mem_ratio=gpu_margin_mem_ratio,
                # optim_config
                **optim_kwargs
            )
        else:
            plugin_initializer = lambda: LowLevelZeroPlugin(
                # zero_config
                stage=stage,
                precision=precision,
                # zero_optim_config
                reduce_bucket_size_in_m=reduce_bucket_size,
                overlap_communication=overlap_communication,
                cpu_offload=(placement_policy == 'cpu'),
                # optim_config
                **optim_kwargs
            )

        super().__init__(seed, plugin_initializer)

    def _post_init(self) -> None:
        assert isinstance(self.plugin, (LowLevelZeroPlugin, GeminiPlugin)), \
            f'{type(self).__name__}\'s plugin is not initialized properly.'

    def setup_distributed(self) -> None:
        colossalai.launch_from_torch({}, seed=self.seed)

    def model_init_context(self):
        if isinstance(self.plugin, GeminiPlugin):
            world_size = dist.get_world_size()
            shard_pg = ProcessGroup(tp_degree=world_size) if self.shard_init else None
            default_dist_spec = ShardSpec([-1], [world_size]) if self.shard_init else None
            return ColoInitContext(device=get_current_device(),
                                   dtype=torch.half,
                                   default_pg=shard_pg,
                                   default_dist_spec=default_dist_spec)
        return super().model_init_context()

    def unwrap_model(self, model: nn.Module) -> nn.Module:
        if isinstance(self.plugin, GeminiPlugin):
            assert isinstance(model, GeminiModel)
            ddp_model = model.unwrap()
            assert isinstance(ddp_model, GeminiDDP)
            return ddp_model.module
        elif isinstance(self.plugin, LowLevelZeroPlugin):
            assert isinstance(model, LowLevelZeroModel)
            return model.module
        else:
            raise RuntimeError(f'Unsupported plugin {type(self.plugin)}')

    def save_pretrained(self,
                        model: nn.Module,
                        path: str,
                        only_rank0: bool = True,
                        tokenizer: Optional[PreTrainedTokenizerBase] = None) -> None:
        if isinstance(self.plugin, GeminiPlugin):
            raise RuntimeError('ColossalAI strategy with stage-3 does not support save_pretrained() now')
        super().save_pretrained(model, path, only_rank0, tokenizer)

    def get_model_state_dict_shard(self, model: nn.Module, **config):
        if not isinstance(self.plugin, GeminiPlugin):
            yield from super().get_model_state_dict_shard(model, **config)
        else:
            # unwrapped_model = self._unwrap_model(model)
            # for module in unwrapped_model.modules():
            #     if isinstance(module, LoraLinear):
            #         module.merge_weights = True
            #         module.eval()
            assert isinstance(model, LowLevelZeroModel)
            yield from model.state_dict_shard(max_shard_size=1024, only_rank_0=False)