Skip to content

Fix GPT-2 FlashAttention mask preparation without pipeline #6434

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
sdjasj wants to merge 1 commit into
base: main
Choose a base branch
from
+12 −2
Open

Fix GPT-2 FlashAttention mask preparation without pipeline #6434

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion colossalai/shardformer/modeling/gpt2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -43,7 +43,7 @@ def _get_attention_mask(
# Received input is already split for non-first pipeline stages,
# but attn mask isn't
batch_size = hidden_states.size(0)
seq_len = attention_mask.size(-1)
seq_len = attention_mask.size(-1) if attention_mask is not None else hidden_states.size(1)

sp_mode = shard_config.sequence_parallelism_mode
# If a 2D or 3D attention mask is provided for the cross-attention
Expand Down
4 changes: 3 additions & 1 deletion colossalai/shardformer/policies/gpt2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -277,7 +277,9 @@ def module_policy(self):
target_key=GPT2Attention,
)

if not self.shard_config.pipeline_stage_manager and self.shard_config.enable_sequence_parallelism:
if not self.shard_config.pipeline_stage_manager and (
self.shard_config.enable_sequence_parallelism or use_flash_attention
):
policy[GPT2Model].method_replacement = {
"forward": partial(GPT2PipelineForwards.gpt2_model_forward, shard_config=self.shard_config)
}
Expand Down
8 changes: 8 additions & 0 deletions tests/test_shardformer/test_model/test_shard_gpt2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -138,6 +138,14 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
@parameterize(
"test_config",
[
{
"tp_size": 1,
"pp_size": 1,
"enable_all_optimization": True,
"use_lazy_init": False,
"precision": "fp32",
"initial_scale": 1,
},
{
"sp_size": 2,
"tp_size": 1,
Expand Down