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

[Bug] fix not desired disable fused share experts caused by rocm logic (#14432)

Browse Source
This commit is contained in:
Junrong Lin
2025-12-05 09:55:07 +08:00
committed by GitHub
parent 6d37e70883
commit 2ecee7571c
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/deepseek_v2.py
View File

@@ -3270,8 +3270,10 @@ class DeepseekV2ForCausalLM(nn.Module):
"Only Deepseek V3/R1 on NV-platform with capability >= 80 "
"or AMD-platform with capability >= gfx942(MI30x) can use shared experts fusion optimization."
)
elif get_moe_expert_parallel_world_size() > 1 and (
not _is_hip or torch.cuda.get_device_capability("cuda") < (9, 4)
elif (
get_moe_expert_parallel_world_size() > 1
and _is_hip
and torch.cuda.get_device_capability("cuda") < (9, 4)
):
disable_reason = "Only Deepseek V3/R1 on AMD-platform with capability >= gfx942(MI30x) can use shared experts fusion optimization under expert parallelism."
elif disable_reason is None and get_moe_a2a_backend().is_deepep():