diff --git a/tests/unit/test_fmha_hd64_smem_p.cu b/tests/unit/test_fmha_hd64_smem_p.cu index 06ccb152..43b1dcd6 100644 --- a/tests/unit/test_fmha_hd64_smem_p.cu +++ b/tests/unit/test_fmha_hd64_smem_p.cu @@ -114,8 +114,8 @@ test_fmha_hd64_smem_p(const bf16_t* q, const bf16_t* k, const bf16_t* v, : "r"(tb + n*8)); asm volatile("tcgen05.wait::ld.sync.aligned;"); if (lane == 0) for (int c=0;c<8;c++) { - s_vals[n*8+c] = tmp[c] * scale * 2.0f; // ×2 to undo QK 0.5 scale - row_max = fmaxf(row_max, tmp[c] * scale * 2.0f); + s_vals[n*8+c] = tmp[c] * scale; // No ×2 correction needed for QK MMA + row_max = fmaxf(row_max, tmp[c] * scale); } } row_max = wmax(row_max);