From a7f81331f85bb0eb0de2e4378b5731bbaad1adab Mon Sep 17 00:00:00 2001
From: biondizzle <biondizzle@gmail.com>
Date: Thu, 28 May 2026 09:25:37 +0000
Subject: [PATCH] test: skip TMEM zeroing again, alloc+dealloc only

---
 tests/unit/test_umma_qk.cu | 17 ++++++++---------
 1 file changed, 8 insertions(+), 9 deletions(-)

diff --git a/tests/unit/test_umma_qk.cu b/tests/unit/test_umma_qk.cu
index 903243d7..e014e53c 100644
--- a/tests/unit/test_umma_qk.cu
+++ b/tests/unit/test_umma_qk.cu
@@ -71,15 +71,14 @@ test_umma_qk_hd16(
     __syncthreads();
     uint32_t tmem_base = *sTmemBase;
 
-    // Zero TMEM — test: only zero first 32 columns (min power of 2)
-    // Note: 128 columns might be too many for tmem_store in a loop
-    if (wid == 0) {
-        for (int col = 0; col < 32; col++) {
-            tmem_store(tmem_base + col, 0, 0, 0, 0);
-        }
-        tmem_fence_store();
-    }
-    __syncthreads();
+    // Zero TMEM — skip for now
+    // if (wid == 0) {
+    //     for (int col = 0; col < 32; col++) {
+    //         tmem_store(tmem_base + col, 0, 0, 0, 0);
+    //     }
+    //     tmem_fence_store();
+    // }
+    // __syncthreads();
 
     // ================================================================
     // Load Q and K into SMEM in canonical layout