Skip to content

feat(moe): support fused permute when capacity is not none #424

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
wants to merge 2 commits into
base: develop
Choose a base branch
from
Open

feat(moe): support fused permute when capacity is not none #424

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
b08b17b
fix capacity mask corruption for fused permute
blankde Apr 9, 2025
4af54f6
relax fused kernel usage
blankde Apr 9, 2025
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
6 changes: 2 additions & 4 deletions internlm/model/moe/dropless_layer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -229,9 +229,7 @@ def __init__(
if self.token_dispatch_policy == "alltoall":
self.token_permutation_func = self.token_permutation_by_alltoall
self.token_unpermutation_func = self.token_unpermutation_by_alltoall
self.enable_fused_permute = (
GEMM_INSTALLED and enable_fused_permute and not drop_and_pad and capacity_factor is None
)
self.enable_fused_permute = GEMM_INSTALLED and enable_fused_permute and not drop_and_pad
self.input_splits = None
self.output_splits = None
self.num_global_tokens_per_local_expert_cpu = None
Expand Down Expand Up @@ -342,7 +340,7 @@ def topk_softmax_with_capacity(self, gates):
exceed_mask = torch.gather(drop_mask, 1, indices)
# shape: [num_token, topk]
final_expert_weights = expert_weights * torch.logical_not(exceed_mask)
final_indices = indices.clone().masked_fill_(exceed_mask, torch.iinfo(torch.long).max)
final_indices = indices.clone().masked_fill_(exceed_mask, torch.iinfo(torch.int32).max)

tokens_per_expert_before_capacity = topk_mask.sum(dim=0)

Expand Down
Loading