EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Fix missing a2a backend init of GLM4.5 MoE Block (#11692)

Browse Source 
Signed-off-by: Shangming Cai <csmthu@gmail.com>
This commit is contained in:
Shangming Cai
2025-10-16 10:13:08 +08:00
committed by GitHub
parent 3289da5b41
commit 476c67d7fc
1 changed files with 4 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/models/glm4_moe.py
View File

@@ -467,7 +467,7 @@ class Glm4MoeSparseMoeBlock(DeepseekV2MoE):
self.top_k = config.num_experts_per_tok
if get_moe_a2a_backend().is_deepep():
if get_moe_a2a_backend().is_deepep() or get_moe_a2a_backend().is_mooncake():
# TODO: we will support tp < ep in the future
self.ep_size = get_moe_expert_parallel_world_size()
self.num_experts = (
@@ -496,7 +496,9 @@ class Glm4MoeSparseMoeBlock(DeepseekV2MoE):
return_recv_hook=True,
)
self._enable_deepep_moe = get_moe_a2a_backend().is_deepep()
self._enable_a2a_moe = (
get_moe_a2a_backend().is_deepep() or get_moe_a2a_backend().is_mooncake()
)
def forward_normal_dual_stream(
self,