From c755e6fddea7f4a7ecab46409a1ae08a32b4f0e7 Mon Sep 17 00:00:00 2001 From: biondizzle Date: Fri, 29 May 2026 18:40:24 +0000 Subject: [PATCH] fix: TMEM read/dealloc for 128-thread kernel --- tests/unit/test_qk_minimal.cu | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tests/unit/test_qk_minimal.cu b/tests/unit/test_qk_minimal.cu index 646e26b7..09b84f94 100644 --- a/tests/unit/test_qk_minimal.cu +++ b/tests/unit/test_qk_minimal.cu @@ -91,7 +91,7 @@ test_qk_minimal_kernel(float* __restrict__ out_s, __syncthreads(); // Read from TMEM - if (wid == 0) { + if (lane < T) { for (int n = 0; n < NUM_READS; n++) { float tmp[8]; asm volatile("tcgen05.ld.sync.aligned.32x32b.x8.b32 {%0,%1,%2,%3,%4,%5,%6,%7},[%8];" @@ -108,7 +108,7 @@ test_qk_minimal_kernel(float* __restrict__ out_s, } } __syncthreads(); - if (is_mma_warp) tmem_dealloc(tb, TMEM_N); + tmem_dealloc(tb, TMEM_N); } int main() {