Skip to content

fix: topkgating major bug #7986

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

Merged
delock merged 3 commits into from
Apr 30, 2026
+6 −7
Merged

fix: topkgating major bug #7986

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
36809e5
fix: topkgating major bug
excepshenal Apr 28, 2026
d334266
fix: further cleanup
excepshenal Apr 28, 2026
2d024e9
Merge branch 'master' into dshen-fix-topkgating
delock Apr 30, 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
13 changes: 6 additions & 7 deletions deepspeed/moe/sharded_moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -384,14 +384,12 @@ def topkgating(
"""Implements TopKGating on logits."""

# everything is in fp32 in this function
# get topk gates
top_gate, top_idx = torch.topk(logits, k=k, dim=1)
# gating decisions
gates = F.softmax(logits, dim=1)
num_experts = int(gates.shape[1])

# get topk mask
topk_masked_gates = torch.zeros_like(logits).scatter(1, top_idx, top_gate)
# get topk gates
top_gate, top_idx = torch.topk(gates, k=k, dim=1)
Comment thread
delock marked this conversation as resolved.

mask = torch.zeros_like(gates, dtype=torch.bool).scatter_(1, top_idx, 1)

Expand All @@ -408,9 +406,10 @@ def topkgating(
# update mask and locations by capacity

if drop_policy == 'probs':
capacity_probs, capacity_indices = torch.topk(topk_masked_gates, k=capacity, dim=0, sorted=False)
capacity_mask = torch.zeros_like(logits).scatter(0, capacity_indices, 1)
mask = torch.logical_and(mask, capacity_mask)
topk_masked_gates = torch.zeros_like(gates).scatter(1, top_idx, top_gate)
_, capacity_indices = torch.topk(topk_masked_gates, k=capacity, dim=0, sorted=False)
capacity_mask = torch.zeros_like(gates, dtype=torch.bool).scatter_(0, capacity_indices, True)
mask &= capacity_mask
locations = torch.cumsum(mask, dim=0) - 1

elif drop_policy == "position":
Expand Down
Loading