github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-08 12:30:42 +00:00
Code Issues Packages Projects Releases Wiki Activity

[shardformer, pipeline] add gradient_checkpointing_ratio and heterogenous shard policy for llama (#5508)

Browse Source 
* feat: add `GradientCheckpointConfig` and `PipelineGradientCheckpointConfig`

* feat: apply `GradientCheckpointConfig` to policy and llama_forward

* feat: move `distribute_layer` and `get_stage_index` to PipelineStageManager

* fix: add optional args for `distribute_layer` and `get_stage_index`

* fix: fix changed API calls

* test: update llama tests

* style: polish `GradientCheckpointConfig`

* fix: fix pipeline utils tests
This commit is contained in:
Wenhao Chen
2024-04-01 11:34:58 +08:00
committed by GitHub
parent df5e9c53cf
commit e614aa34f3
28 changed files with 396 additions and 213 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
colossalai/booster/plugin/hybrid_parallel_plugin.py
View File

@@ -26,7 +26,7 @@ from colossalai.cluster import ProcessGroupMesh
from colossalai.interface import AMPModelMixin, ModelWrapper, OptimizerWrapper
from colossalai.pipeline.schedule import InterleavedSchedule, OneForwardOneBackwardSchedule
from colossalai.pipeline.stage_manager import PipelineStageManager
from colossalai.shardformer import ShardConfig, ShardFormer
from colossalai.shardformer import GradientCheckpointConfig, ShardConfig, ShardFormer
from colossalai.shardformer.layer.utils import SeqParallelUtils
from colossalai.shardformer.policies.base_policy import Policy
from colossalai.tensor.d_tensor.api import is_distributed_tensor
@@ -930,6 +930,7 @@ class HybridParallelPlugin(PipelinePluginBase):
custom_policy (Policy, optional): Custom policy for Shardformer. Defaults to None.
pp_style (str, optional): The style for pipeline parallelism. Defaults to '1f1b'.
num_model_chunks (int, optional): The number of model chunks for interleaved pipeline parallelism. Defaults to 1.
gradient_checkpoint_config (GradientCheckpointConfig, optional): Configuration for gradient checkpointing. Defaults to None.
enable_metadata_cache (bool, optional): Whether to enable metadata cache for pipeline parallelism. Defaults to True.
"""
@@ -969,6 +970,7 @@ class HybridParallelPlugin(PipelinePluginBase):
custom_policy: Policy = None,
pp_style: str = "1f1b",
num_model_chunks: int = 1,
gradient_checkpoint_config: Optional[GradientCheckpointConfig] = None,
enable_metadata_cache: bool = True,
) -> None:
super().__init__()
@@ -1043,6 +1045,7 @@ class HybridParallelPlugin(PipelinePluginBase):
enable_sequence_parallelism=enable_sequence_parallelism,
enable_sequence_overlap=enable_sequence_overlap,
parallel_output=parallel_output,
gradient_checkpoint_config=gradient_checkpoint_config,
)
self.amp_config = dict(
initial_scale=initial_scale,