File size: 3,157 Bytes
2961be3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 | # Copyright 2020-2026 The HuggingFace Team. 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.
"""Distributed training backend abstraction."""
from contextlib import contextmanager
class DistributedBackend:
"""Abstracts distributed backend specifics (DeepSpeed ZeRO, FSDP) behind a uniform API.
Detects the active backend once at construction from ``accelerator.state``, then provides context managers that are
no-ops on backends where the operation is not needed.
Args:
accelerator ([`~accelerate.Accelerator`]):
The accelerator instance managing the distributed state.
Example:
```python
>>> dist = DistributedBackend(accelerator)
>>> with dist.gather_params(list(model.parameters())):
... model.merge_adapter()
>>> with dist.summon_full_params(model, recurse=False):
... outputs = model.generate(inputs)
```
"""
def __init__(self, accelerator):
ds_plugin = accelerator.state.deepspeed_plugin
fsdp_plugin = getattr(accelerator.state, "fsdp_plugin", None)
self.zero_stage = ds_plugin.zero_stage if ds_plugin else 0
self.fsdp_version = getattr(fsdp_plugin, "fsdp_version", None) if fsdp_plugin else None
@property
def is_zero3(self) -> bool:
"""Whether DeepSpeed ZeRO Stage 3 is active."""
return self.zero_stage == 3
@property
def is_fsdp(self) -> bool:
"""Whether FSDP (any version) is active."""
return self.fsdp_version is not None
@contextmanager
def gather_params(self, params):
"""Gather sharded parameters under DeepSpeed ZeRO-3; no-op otherwise.
Args:
params (iterable of `torch.nn.Parameter`):
Parameters to gather.
"""
if self.is_zero3:
import deepspeed
with deepspeed.zero.GatheredParameters(params):
yield
else:
yield
@contextmanager
def summon_full_params(self, module, **kwargs):
"""Materialize full FSDP v1 parameters; no-op for FSDP v2 or non-FSDP backends.
FSDP v2 parameters are always accessible and require no explicit materialization.
Args:
module (`torch.nn.Module`):
The FSDP-wrapped module.
**kwargs:
Forwarded to ``FSDP.summon_full_params`` (e.g. ``recurse``, ``writeback``).
"""
if self.fsdp_version == 1:
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
with FSDP.summon_full_params(module, **kwargs):
yield
else:
yield
|