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<<<dim3(1,1,1), 192, smem_size>>>(
         d_q_nope_fp8, d_q_nope_scale, d_q_rope_bf16,
         d_k_nope_fp8, d_k_nope_scale, d_k_rope_bf16,