From dd1cbe1faa2aa080f52b83abc5ae81338d1d5ae8 Mon Sep 17 00:00:00 2001 From: biondizzle Date: Wed, 3 Jun 2026 03:47:01 +0000 Subject: [PATCH] Fix smem size for prefill debug test --- tests/unit/test_prefill_t2_debug.cu | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/tests/unit/test_prefill_t2_debug.cu b/tests/unit/test_prefill_t2_debug.cu index d34acebf..e8b55eb2 100644 --- a/tests/unit/test_prefill_t2_debug.cu +++ b/tests/unit/test_prefill_t2_debug.cu @@ -530,7 +530,8 @@ int main() { // Launch debug kernel printf("\n=== GPU Kernel Execution ===\n"); - int smem_size = 256 * 1024; // generous + int smem_size = 200 * 1024; // ~149KB needed, stay under 232KB limit + cudaFuncSetAttribute(prefill_t2_debug_kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, smem_size); prefill_t2_debug_kernel<<>>( d_q_nope_fp8, d_q_nope_scale, d_q_rope_bf16, d_k_nope_fp8, d_k_nope_scale, d_k_rope_bf16,