From 7e17dee8b77e027b07e750516e6a9c227b2cafa2 Mon Sep 17 00:00:00 2001 From: Yuhao Yang <47235274+yhyang201@users.noreply.github.com> Date: Sat, 12 Sep 2026 02:53:59 +0800 Subject: [PATCH] Keep wo_a small-batch on its validated G=2 shape --- python/sglang/srt/models/deepseek_v4.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/python/sglang/srt/models/deepseek_v4.py b/python/sglang/srt/models/deepseek_v4.py index 18cfd3b0a2..bc2775b925 100644 --- a/python/sglang/srt/models/deepseek_v4.py +++ b/python/sglang/srt/models/deepseek_v4.py @@ -506,7 +506,7 @@ def _apply_wo_a_bf16_matmul( ): if is_decode and o.shape[0] == 1: return wo_a_bf16_gemv(o, wo_a) - if 2 <= o.shape[0] <= 8: + if 2 <= o.shape[0] <= 8 and o.shape[1] == 2: if fuse_mxfp8_quant: return Mxfp8SwizzledInput(*wo_a_bf16_small_batch_mxfp8(o, wo_a)) return wo_a_bf16_small_batch(o, wo_a)