minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

[NPU][Fix Issue]: Send expert weights contiguous tensor across cards during EPLB rebalance (#32001)

Browse Source
This commit is contained in:
icarus_zh
2026-07-27 09:20:28 +08:00
committed by GitHub
parent 3863612023
commit a358374ae9
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/hardware_backend/npu/quantization/moe_methods.py
+1 -1
View File
@@ -332,7 +332,7 @@ class NPUW8A8Int8MoEMethod(_NPUMoEMethodBase):
# Process weight # Process weight
weight: torch.Tensor = getattr(layer, f"{weight_prefix}_weight") weight: torch.Tensor = getattr(layer, f"{weight_prefix}_weight")
weight.data = npu_format_cast(weight.data.transpose(1, 2)) weight.data = npu_format_cast(weight.data.transpose(1, 2).contiguous())
# Set dispatcher output dtype # Set dispatcher output dtype
if weight_prefix == "w13": if weight_prefix == "w13":