diff --git a/tests/unit/test_fmha_hd64_debug.cu b/tests/unit/test_fmha_hd64_debug.cu index 79288dd1..b3ac07fe 100644 --- a/tests/unit/test_fmha_hd64_debug.cu +++ b/tests/unit/test_fmha_hd64_debug.cu @@ -242,8 +242,8 @@ int main() { printf("Cosine: %.8f\n", cs_reg); // Compare MMA PV vs reference - printf("\n--- MMA PV vs FP32 reference ---\n"); - for (int d=0;d<8;d++) printf(" MMA[%d]=%.6f ref[%d]=%.6f ratio=%.4f\n", + printf("\n--- MMA PV vs FP32 reference (all 64 elements) ---\n"); + for (int d=0;d1e-6f ? bf16_to_f32_host(h_o_mma[d])/h_o_ref[d] : 0); float cs_mma = 0, na_mma = 0, nb_mma = 0;