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

Fix MoE no_combine: skip router weight in down projection (#23545)

Browse Source
This commit is contained in:
Lianmin Zheng
2026-04-23 16:47:58 -07:00
committed by GitHub
parent cf88fdcc9c
commit bb962b0046
2 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/moe/moe_runner/triton_utils/fused_moe.py
+1 -1
View File
@@ -618,7 +618,7 @@ def _fused_moe_kernel_sequence(
sorted_token_ids, sorted_token_ids,
expert_ids, expert_ids,
num_tokens_post_padded, num_tokens_post_padded,
not apply_router_weight_on_input, not apply_router_weight_on_input and not no_combine,
1, 1,
down_config or config, down_config or config,
compute_type=compute_type, compute_type=compute_type,
python/sglang/srt/layers/moe/token_dispatcher/base.py
+1 -1
View File
@@ -262,7 +262,7 @@ class BaseDispatcher(ABC):
"""Base class for dispatchers.""" """Base class for dispatchers."""
def __init__(self): def __init__(self):
self.quant_config: Optional[dict] = None self.quant_config: dict = {}
# Overlap args # Overlap args
self.overlap_args: Optional[CombineOverlapArgs] = None self.overlap_args: Optional[CombineOverlapArgs] = None