From 9b37e28b83bb5a01d3d6395adb70852ac7a504bd Mon Sep 17 00:00:00 2001 From: Alison Shao <54658187+alisonshao@users.noreply.github.com> Date: Fri, 9 Oct 2026 17:56:20 -0700 Subject: [PATCH] Disable DualGemm fusion for Inkling dense MLP (#43447) --- python/sglang/srt/models/inkling_common/dense_mlp.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/python/sglang/srt/models/inkling_common/dense_mlp.py b/python/sglang/srt/models/inkling_common/dense_mlp.py index 2c269fb873..522e359f3a 100644 --- a/python/sglang/srt/models/inkling_common/dense_mlp.py +++ b/python/sglang/srt/models/inkling_common/dense_mlp.py @@ -120,6 +120,8 @@ class InklingDenseMLP(LlamaMLP): fused = fused and not lora_compatible_layout_enabled() self.layer_id = layer_id self.act_fn = InklingSwiglu(interleaved=fused) + # DualGemm assumes [gate||up] and SiluAndMul; it would bypass act_fn. + self.dual_gemm.mode = None self.scattered_sconv = get_exec().comm.enable_scattered_sconv def forward(