diff --git a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
index a6f6d3726..066fa7d82 100644
--- a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
+++ b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
@@ -1183,8 +1183,9 @@ impl GpuDqnTrainer {
         self.stream.synchronize()
             .map_err(|e| MLError::ModelError(format!("stream sync before capture: {e}")))?;
 
-        // Event tracking is already disabled at DQNTrainer construction (single-stream
-        // context — events are unnecessary with explicit sync barriers).
+        // Disable event tracking during capture — cudarc's device_ptr/device_ptr_mut
+        // record CudaEvents which are DISALLOWED inside CUDA Graph capture.
+        unsafe { self.stream.context().disable_event_tracking(); }
 
         // Begin stream capture — only work submitted from this thread on this
         // stream is captured (THREAD_LOCAL mode, safe for single-stream use).
@@ -1205,10 +1206,10 @@ impl GpuDqnTrainer {
             cudarc::driver::sys::CUgraphInstantiate_flags::CUDA_GRAPH_INSTANTIATE_FLAG_AUTO_FREE_ON_LAUNCH,
         );
 
-        // Keep event tracking disabled — graph capture wrote stale events to
-        // CudaSlices. All subsequent operations use explicit stream sync +
-        // ManuallyDrop guards (raw_device_ptr), so event tracking is unnecessary.
-        // Re-enabling would leave stale events that cause CUDA_ERROR_INVALID_VALUE.
+        // Re-enable event tracking after capture. With GPU-native cat/stack
+        // (no to_host roundtrip), stale events from capture don't cause issues
+        // because all post-capture GPU ops use DtoD copies, not DtoH.
+        unsafe { self.stream.context().enable_event_tracking(); }
 
         // Propagate submission error first
         submit_result?;