diff --git a/crates/ml-dqn/src/gpu_replay_buffer.rs b/crates/ml-dqn/src/gpu_replay_buffer.rs
index 607caf604..c6c1caf9c 100644
--- a/crates/ml-dqn/src/gpu_replay_buffer.rs
+++ b/crates/ml-dqn/src/gpu_replay_buffer.rs
@@ -256,6 +256,12 @@ impl GpuReplayBuffer {
         let update_batch_max = a32f(stream, 1, "update_batch_max")?;
         let update_max_merge = a32f(stream, 1, "update_max_merge")?;
 
+        // Pre-initialize cast kernels while CUDA context is clean.
+        // OnceLock + cuModuleLoadData requires a current CUDA context.
+        // If deferred to first PER update (after graph capture), the context
+        // may be in a state that blocks cuModuleLoadData on H100 CUDA 13.
+        get_cast_kernels(stream)?;
+
         Ok(Self {
             config, stream: Arc::clone(stream), kernels: k,
             states: s, next_states: ns, actions: a, rewards: r, dones: d, priorities: p,