ColossalAI/colossalai/zero/utils/zero_hook_v2.py

import torch
from colossalai.tensor.param_op_hook import ParamOpHook
from colossalai.tensor.chunk import ChunkManager, TensorState
from enum import Enum
from typing import List
from contextlib import contextmanager
from functools import partial
from colossalai.gemini.gemini_mgr import GeminiManager


class TrainingPhase(Enum):
    FORWARD = 0
    BACKWARD = 1


class ZeROHookV2(ParamOpHook):

    def __init__(self, gemini_manager: GeminiManager) -> None:
        super().__init__()
        self._gemini_manager = gemini_manager
        self._chunk_manager = gemini_manager.chunk_manager
        self._training_phase = TrainingPhase.FORWARD

    def pre_op(self, params):
        chunks = self._chunk_manager.get_chunks(params)
        for p in params:
            self._chunk_manager.trans_tensor_state(p, TensorState.COMPUTE)
        self._chunk_manager.exec_lazy_release()
        self._gemini_manager.sample_overall_data()
        self._gemini_manager.adjust_layout(chunks, 'fp16_param')
        for chunk in chunks:
            self._chunk_manager.access_chunk(chunk)
        self._gemini_manager.sample_model_data()

    def post_op(self, params):
        for p in params:
            tensor_state = TensorState.HOLD if self._training_phase == TrainingPhase.FORWARD or not p.requires_grad else TensorState.HOLD_AFTER_BWD
            self._chunk_manager.trans_tensor_state(p, tensor_state)
        self._chunk_manager.add_lazy_release_tensors(params)

    def pre_forward(self, params: List[torch.Tensor]) -> None:
        self.pre_op(params)

    def post_forward(self, params: List[torch.Tensor]) -> None:
        self.post_op(params)

    def pre_backward(self, params: List[torch.Tensor]) -> None:
        self.pre_op(params)

    def post_backward(self, params: List[torch.Tensor]) -> None:
        self.post_op(params)

    @contextmanager
    def switch_training_phase(self, training_phase: TrainingPhase = TrainingPhase.BACKWARD):
        try:
            old_training_phase = self._training_phase
            self._training_phase = training_phase
            yield
        finally:
            self._training_phase = old_training_phase

    switch_to_backward = switch_training_phase
    switch_to_forward = partial(switch_to_backward, training_phase=TrainingPhase.FORWARD)