Skip to content

[Dev] recover FusedMLP API #2978

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

Closed
HaochenYuan wants to merge 2 commits into from
Closed

[Dev] recover FusedMLP API #2978

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
125451a
refactor padding_mask pass
HaochenYuan Jan 16, 2026
3f9734f
fix linting
HaochenYuan Jan 16, 2026
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 megatron/core/extensions/transformer_engine.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2139,7 +2139,7 @@ def forward_post_hook(module, *_) -> None:
"TEFusedMLP module does not support submodules with post-backward hooks"
)

def forward(self, hidden_states: torch.Tensor, **kwargs) -> Tuple[Tensor, Optional[Tensor]]:
def forward(self, hidden_states: torch.Tensor) -> Tuple[Tensor, Optional[Tensor]]:
"""Forward."""

# Construct fused impl if needed
Expand Down
10 changes: 5 additions & 5 deletions megatron/core/transformer/transformer_layer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -653,6 +653,8 @@ def _forward_mlp(self, hidden_states, inference_context=None, padding_mask=None)

# Residual connection.
residual = hidden_states
# Only MoE layers need padding_mask
mlp_kwargs = {"padding_mask": padding_mask} if self.is_moe_layer else {}

if self.offload_mlp_norm:
hidden_states = fine_grained_offloading_group_start(hidden_states, name="mlp_norm")
Expand Down Expand Up @@ -687,13 +689,11 @@ def _forward_mlp(self, hidden_states, inference_context=None, padding_mask=None)
tensor_parallel.random.get_cuda_rng_tracker,
self.pg_collection.tp,
pre_mlp_layernorm_output,
padding_mask=padding_mask,
**mlp_kwargs,
)
else:
mlp_output_with_bias = tensor_parallel.checkpoint(
functools.partial(self.mlp, padding_mask=padding_mask),
False,
pre_mlp_layernorm_output,
functools.partial(self.mlp, **mlp_kwargs), False, pre_mlp_layernorm_output
)
elif should_chunk_mlp_for_prefill:
# Chunk input along sequence dimension
Expand All @@ -709,7 +709,7 @@ def _forward_mlp(self, hidden_states, inference_context=None, padding_mask=None)
bias_output = torch.stack(bias_chunks, dim=0).sum(dim=0) if bias_chunks else None
mlp_output_with_bias = (mlp_output, bias_output)
else:
mlp_output_with_bias = self.mlp(pre_mlp_layernorm_output, padding_mask=padding_mask)
mlp_output_with_bias = self.mlp(pre_mlp_layernorm_output, **mlp_kwargs)

if self.recompute_pre_mlp_layernorm:
# discard the output of the pre-mlp layernorm and register the recompute
Expand Down
Loading