Skip to content

[2/3] Optimize Slime Update Weights: Avoid GPU-to-CPU Device Sync when update expert weights #8753

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
ch-wan merged 2 commits into from
Aug 6, 2025
Merged

[2/3] Optimize Slime Update Weights: Avoid GPU-to-CPU Device Sync when update expert weights #8753

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
531368e
Avoid GPU-to-CPU Device Sync when update expert weights
hebiao064 Aug 4, 2025
3323503
avoid eplb
hebiao064 Aug 4, 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: 5 additions & 1 deletion python/sglang/srt/eplb/expert_location.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,6 +35,7 @@ class ExpertLocationMetadata:
physical_to_logical_map: torch.Tensor # (layers, num_physical_experts)
physical_to_logical_map_cpu: torch.Tensor
logical_to_all_physical_map: torch.Tensor # (layers, num_logical_experts, X)
logical_to_all_physical_map_cpu: torch.Tensor # CPU copy for performance
logical_to_all_physical_map_num_valid: torch.Tensor # (layers, num_logical_experts)
# (layers, num_logical_experts)
logical_to_rank_dispatch_physical_map: Optional[torch.Tensor]
Expand Down Expand Up @@ -221,6 +222,7 @@ def _init_raw(
physical_to_logical_map=physical_to_logical_map,
physical_to_logical_map_cpu=physical_to_logical_map.cpu(),
logical_to_all_physical_map=logical_to_all_physical_map_padded,
logical_to_all_physical_map_cpu=logical_to_all_physical_map_padded.cpu(),
logical_to_all_physical_map_num_valid=logical_to_all_physical_map_num_valid,
logical_to_rank_dispatch_physical_map=(
compute_logical_to_rank_dispatch_physical_map(
Expand Down Expand Up @@ -251,6 +253,7 @@ def update(
"physical_to_logical_map",
"physical_to_logical_map_cpu",
"logical_to_all_physical_map",
"logical_to_all_physical_map_cpu",
"logical_to_all_physical_map_num_valid",
"logical_to_rank_dispatch_physical_map",
]:
Expand All @@ -270,9 +273,10 @@ def update(
def logical_to_all_physical(
self, layer_id: int, logical_expert_id: int
) -> List[int]:
# Use CPU copy to avoid GPU→CPU sync on every call, which is expensive in update weights scenario
return [
physical_expert_id
for physical_expert_id in self.logical_to_all_physical_map[
for physical_expert_id in self.logical_to_all_physical_map_cpu[
layer_id, logical_expert_id
].tolist()
if physical_expert_id != -1
Expand Down
Loading