fix(modelopt_fp4): skip NVFP4 swiglu-fusion interleave for shared experts with swiglu_limit (#37378)

This commit is contained in:
Xinyuan Tong
2026-09-04 22:42:29 -07:00
committed by GitHub
parent ae3205ba28
commit 32a1d55431
+1
View File
@@ -787,6 +787,7 @@ class DeepseekV2MoE(nn.Module):
ModelOptFp4LinearMethod,
)
and fc1_n % 128 == 0
and self.shared_experts.swiglu_limit is None
and not check_cuda_graph_backend(Phase.PREFILL, Backend.TC_PIECEWISE)
):
self.shared_experts.gate_up_proj._interleave_for_swiglu_fusion = True