Keep wo_a small-batch on its validated G=2 shape

This commit is contained in:
Yuhao Yang
2026-10-20 10:24:26 +08:00
committed by minke.yu
parent 0090df5d64
commit 7e17dee8b7
+1 -1
View File
@@ -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)