github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-06-01 20:05:27 +00:00
Code Issues Packages Projects Releases Wiki Activity

[shardformer] fix emerged bugs after updating transformers (#4526)

Browse Source
This commit is contained in:
Baizhou Zhang 2023-08-29 11:25:05 +08:00 committed by GitHub
parent c554b7f559
commit 0387a47e63
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 9 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
colossalai/pipeline/schedule/_utils.py
View File

@ -123,6 +123,9 @@ def merge_batch(data: List[Any]) -> Any:
merged_data = [] merged_data = []
for elem_batch in zip(*flattened_data): for elem_batch in zip(*flattened_data):
if isinstance(elem_batch[0], torch.Tensor): if isinstance(elem_batch[0], torch.Tensor):
if len(elem_batch[0].shape) == 0: # set loss to None in pipeline outputs
merged_data.append(None)
else:
merged_data.append(torch.cat(elem_batch, dim=0)) merged_data.append(torch.cat(elem_batch, dim=0))
else: else:
merged_data.append(list(elem_batch)) merged_data.append(list(elem_batch))

6
tests/test_shardformer/test_model/_utils.py
View File

@ -195,7 +195,11 @@ def check_output_hidden_state(org_output: Tensor,
sharded_hidden_state = sharded_output.last_hidden_state sharded_hidden_state = sharded_output.last_hidden_state
if stage_manager and stage_manager.is_last_stage(): if stage_manager and stage_manager.is_last_stage():
sharded_hidden_state = torch.cat([output.last_hidden_state for output in sharded_output['outputs']], dim=dim) pipeline_output = sharded_output['outputs']
if isinstance(pipeline_output, List):
sharded_hidden_state = torch.cat([output.last_hidden_state for output in pipeline_output], dim=dim)
else:
sharded_hidden_state = pipeline_output.last_hidden_state
assert torch.allclose(org_hidden_state.float(), sharded_hidden_state.float(), atol=atol, rtol=rtol), \ assert torch.allclose(org_hidden_state.float(), sharded_hidden_state.float(), atol=atol, rtol=rtol), \
f"shard model's output hidden state is not equal to origin model's last hidden state\n{org_hidden_state}\n{sharded_hidden_state}" f"shard model's output hidden state is not equal to origin model's last hidden state\n{org_hidden_state}\n{sharded_hidden_state}"