debug: add TMEM read diagnostic

2026-05-28 07:46:15 +00:00
parent a12607b0bd
commit b50f6a8512
1 changed files with 3 additions and 0 deletions
--- a/dsv4/kernels/attention/fmha_epilogue_sm100.cuh
+++ b/dsv4/kernels/attention/fmha_epilogue_sm100.cuh
@@ -270,6 +270,9 @@ fmha_decode_tmem(
            float r2 = u32_to_f32(u2) * inv_sum;
            float r3 = u32_to_f32(u3) * inv_sum;

+            // Debug: print first few values from lane 0
+            if (lane == 0 && col == 0) printf("[tmem] read: r0=%f r1=%f r2=%f r3=%f inv_sum=%f\n", r0, r1, r2, r3, inv_sum);
+
            // Step 4: Cast to BF16 and write to GMEM
            int base = col * 128;
            int d0 = base + lane * 4 + 0;