From a9d71ff6ab843ef7eb5f3912ca97043703686e51 Mon Sep 17 00:00:00 2001 From: biondizzle Date: Thu, 28 May 2026 08:38:08 +0000 Subject: [PATCH] debug: print TMEM values after MMA --- dsv4/kernels/attention/fmha_qk_verify.cuh | 14 +++++++++++++- 1 file changed, 13 insertions(+), 1 deletion(-) diff --git a/dsv4/kernels/attention/fmha_qk_verify.cuh b/dsv4/kernels/attention/fmha_qk_verify.cuh index b260c392..db146b42 100644 --- a/dsv4/kernels/attention/fmha_qk_verify.cuh +++ b/dsv4/kernels/attention/fmha_qk_verify.cuh @@ -82,7 +82,19 @@ fmha_qk_verify( if (wid == 0 && lane == 0) tmem_fence_store(); __syncthreads(); - // Read S row 0 from TMEM + // Read S row 0 from TMEM — try multiple column offsets + if (tid == 0) { + printf("[qk] tmem_base=%u, reading S...\n", tmem_base); + for (int col = 0; col < 4; col++) { + uint32_t u0, u1, u2, u3; + tmem_load(tmem_base + col, u0, u1, u2, u3); + printf("[qk] S[0,%d] raw: %f %f %f %f\n", col, + u32_to_f32(u0), u32_to_f32(u1), u32_to_f32(u2), u32_to_f32(u3)); + } + } + __syncthreads(); + + // Also read using the output array (same as before) if (wid == 0) { for (int col = lane; col < 128; col += WARP) { uint32_t u0, u1, u2, u3;