From 764ed01d6f5e525fb9a82007635aa951839e0b23 Mon Sep 17 00:00:00 2001 From: biondizzle Date: Thu, 28 May 2026 09:41:50 +0000 Subject: [PATCH] test: try M=64 in descriptor + idesc to debug 4x factor --- tests/unit/test_umma_qk.cu | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/tests/unit/test_umma_qk.cu b/tests/unit/test_umma_qk.cu index c6c37367..5544d03b 100644 --- a/tests/unit/test_umma_qk.cu +++ b/tests/unit/test_umma_qk.cu @@ -54,9 +54,9 @@ test_umma_qk_hd16( // Descriptors uint32_t sQ_smem = __cvta_generic_to_shared(sQ); uint32_t sK_smem = __cvta_generic_to_shared(sK); - uint64_t desc_q = make_umma_desc_kmajor_none(sQ_smem, 128); - uint64_t desc_k = make_umma_desc_kmajor_none(sK_smem, 128); - uint32_t idesc = 0; // Try idesc=0 (same as gau-nernst) + uint64_t desc_q = make_umma_desc_kmajor_none(sQ_smem, 64); // Try M=64 + uint64_t desc_k = make_umma_desc_kmajor_none(sK_smem, 64); // Try M=64 + uint32_t idesc = make_idesc(64, 128); // M=64, N=128 // Verify SMEM Q and K by reading back row 0 if (tid == 0) {