chatlearn/models/megatron/memory_manager/inference.py

# Copyright 2024 Alibaba Group Holding Limited. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Inference Memery manager for Megatron."""
from typing import Optional, List

from chatlearn.utils.flat_tensors import BucketizedFlatTensors
from chatlearn.utils.logger import log_rank_0
from chatlearn.utils.megatron_import_helper import DistributedDataParallel


class InferenceMemoryManager:
    """
    Memory manager for Megatron inference modules which provides utilities to free memory when unused.
    """

    def __init__(self, model, bucket_size_mb=0):
        self._model = model

        assert not isinstance(
            model, (DistributedDataParallel,)
        ), f'Only support model type non-DistributedDataParallel, current type is {str(type(model))}.'

        self._weights_offloaded = False
        self._group_flat_weights: Optional[List[BucketizedFlatTensors]] = None
        self._bucket_size_mb = bucket_size_mb

    def offload_weights(self):
        """
        offload weights
        """
        if self._weights_offloaded:
            log_rank_0('Call offload_weights when already offloaded. Ignore it.')
            return

        if self._group_flat_weights is None:
            dtype_to_params = {}
            for p in self._model.parameters():
                dtype = p.dtype
                if dtype not in dtype_to_params:
                    dtype_to_params[dtype] = []
                dtype_to_params[dtype].append(p)

            self._group_flat_weights = []
            for params in dtype_to_params.values():
                self._group_flat_weights.append(
                    BucketizedFlatTensors(params, primary_store_device='cpu', bucket_size_mb=self._bucket_size_mb)
                )

        for flat_weights in self._group_flat_weights:
            flat_weights.copy_to_primary_store()

        self._weights_offloaded = True

    def onload_weights(self):
        """
        onload weights
        """
        if not self._weights_offloaded:
            log_rank_0('Call onload_weights when already onloaded. Ignore it.')
            return

        for flat_weights in self._group_flat_weights:
            flat_weights.copy_to_gpu_buffer()

        self._weights_offloaded = False