minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Fix nan in global scaling factor for large scale nvfp4 EP (#13162)

Browse Source
This commit is contained in:
Shu Wang
2025-11-12 15:32:23 -08:00
committed by GitHub
parent 4eda9969e8
commit 7aa443903d
2 changed files with 17 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/eplb/expert_location.py
+13 -9
View File
@@ -284,9 +284,17 @@ class ExpertLocationMetadata:
# -------------------------------- usage ------------------------------------ # -------------------------------- usage ------------------------------------
def logical_to_all_physical( def logical_to_all_physical(
self, layer_id: int, logical_expert_id: int self,
layer_id: int,
logical_expert_id: int,
require_global_experts: bool = False,
) -> List[int]: ) -> List[int]:
# Use CPU copy to avoid GPU→CPU sync on every call, which is expensive in update weights scenario # Use CPU copy to avoid GPU→CPU sync on every call, which is expensive in update weights scenario
if require_global_experts:
num_physical_experts = self.logical_to_all_physical_map_cpu[layer_id].shape[
-1
]
return list(torch.arange(0, num_physical_experts))
return [ return [
physical_expert_id physical_expert_id
for physical_expert_id in self.logical_to_all_physical_map_cpu[ for physical_expert_id in self.logical_to_all_physical_map_cpu[
@@ -355,14 +363,10 @@ def _compute_logical_to_all_physical_map(
) )
# Replace by the nearest physical expert # Replace by the nearest physical expert
mapped_physical_experts = logical_to_all_physical_map[layer_id][ if nearest_expert != -1:
logical_expert_id logical_to_all_physical_map[layer_id][logical_expert_id] = [
] nearest_expert
if ( ]
nearest_expert != -1
and nearest_expert not in mapped_physical_experts
):
mapped_physical_experts[0] = nearest_expert
logical_to_all_physical_map = _pad_nested_array( logical_to_all_physical_map = _pad_nested_array(
logical_to_all_physical_map, pad_value=-1 logical_to_all_physical_map, pad_value=-1
python/sglang/srt/layers/moe/fused_moe_triton/layer.py
+4 -1
View File
@@ -517,9 +517,12 @@ class FusedMoE(torch.nn.Module):
# This is a shared expert. # This is a shared expert.
physical_expert_ids = [expert_id] physical_expert_ids = [expert_id]
else: else:
require_global_experts = getattr(
param, "_sglang_require_global_experts", False
)
physical_expert_ids = ( physical_expert_ids = (
global_expert_location_metadata.logical_to_all_physical( global_expert_location_metadata.logical_to_all_physical(
self.layer_id, expert_id self.layer_id, expert_id, require_global_experts
) )
) )