From 2ecee7571cdfe10a124ae313fe020166ad5c56b8 Mon Sep 17 00:00:00 2001 From: Junrong Lin <33685709+ocss884@users.noreply.github.com> Date: Fri, 5 Dec 2025 09:55:07 +0800 Subject: [PATCH] [Bug] fix not desired disable fused share experts caused by rocm logic (#14432) --- python/sglang/srt/models/deepseek_v2.py | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/python/sglang/srt/models/deepseek_v2.py b/python/sglang/srt/models/deepseek_v2.py index 30d7224dc..c2213cf95 100644 --- a/python/sglang/srt/models/deepseek_v2.py +++ b/python/sglang/srt/models/deepseek_v2.py @@ -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():