fix(htm): set NON_PORTABLE_CLUSTER_SIZE_ALLOWED=1 for cluster_size=16

overlay/htm_rust/src/gpu/fused.rs

@@ -244,6 +244,18 @@ impl FusedState {
             result::module::get_function(module, CString::new("htm_fused_step_batched").unwrap())
         }?;
 
+        // Cluster size 16 on Hopper is "non-portable" (> 8 requires opt-in).
+        // Must set CU_FUNC_ATTRIBUTE_NON_PORTABLE_CLUSTER_SIZE_ALLOWED=1 on
+        // every launched kernel function, otherwise cuLaunchKernelEx rejects
+        // the cluster dim with CUDA_ERROR_INVALID_CLUSTER_SIZE.
+        unsafe {
+            let attr = sys::CUfunction_attribute::CU_FUNC_ATTRIBUTE_NON_PORTABLE_CLUSTER_SIZE_ALLOWED;
+            // Ignore errors: older CUDA may lack the attribute, in which case
+            // only portable sizes (<= 8) work — plan_fused_launch caps at 8.
+            let _ = sys::lib().cuFuncSetAttribute(function, attr, 1);
+            let _ = sys::lib().cuFuncSetAttribute(function_batched, attr, 1);
+        }
+
         // Probe SM count.
         let sm_count = match dev.attribute(
             cudarc::driver::sys::CUdevice_attribute::CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT,