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

[Fix][AMD] Qwen3.5 MoE: disable global-slot shared-expert fusion under per-rank EP backends (MoRI + dp-attention init crash) (#31793)

Browse Source
This commit is contained in:
jacky.cheng
2026-07-25 23:59:47 -07:00
committed by GitHub
parent 72e415dfc8
commit 833e1bc601
1 changed files with 15 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/qwen2_moe.py
+15
View File
@@ -68,6 +68,7 @@ from sglang.srt.layers.moe.utils import (
RoutingMethodType, RoutingMethodType,
filter_moe_weight_param_global_expert, filter_moe_weight_param_global_expert,
is_deepep_class_backend, is_deepep_class_backend,
uses_per_rank_fused_shared_slots,
) )
from sglang.srt.layers.quantization.base_config import QuantizationConfig from sglang.srt.layers.quantization.base_config import QuantizationConfig
from sglang.srt.layers.radix_attention import RadixAttention from sglang.srt.layers.radix_attention import RadixAttention
@@ -249,6 +250,20 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
self.num_shared_experts > 0 self.num_shared_experts > 0
and can_fuse_shared_expert(config, quant_config) and can_fuse_shared_expert(config, quant_config)
) )
if (
self.enable_shared_expert_fusion
and uses_per_rank_fused_shared_slots()
and get_parallel().moe_ep_size > 1
):
logger.warning_once(
"Disabling Qwen shared-expert fusion: it uses a single global "
"shared slot with a per-token gate, which is incompatible with "
"per-rank EP shared-slot backends (e.g. DeepEP/MoRI) at "
"moe_ep_size=%d. Using the separate shared-expert MLP instead.",
get_parallel().moe_ep_size,
)
self.enable_shared_expert_fusion = False
if self.enable_shared_expert_fusion: if self.enable_shared_expert_fusion:
self.num_fused_shared_experts = self.num_shared_experts self.num_fused_shared_experts = self.num_shared_experts