diff --git a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
index 7339d02b8..c22ec6bbc 100644
--- a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
+++ b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
@@ -400,13 +400,9 @@ impl GpuDqnTrainer {
         let b = config.batch_size;
         let total_params = compute_total_params(&config);
 
-        // All GPU components share this single forked CudaStream. Disable cudarc's
-        // automatic event tracking: it injects cuStreamWaitEvent/cuEventRecord into
-        // every CudaSlice access, creating cross-stream event references that
-        // invalidate CUDA Graph capture and conflict with Candle tensor Drops on
-        // the default stream. Safe: single forked stream, all work serialized.
-        // SAFETY: single-stream, all GPU work is serialized on `stream`.
-        unsafe { stream.context().disable_event_tracking(); }
+        // Event tracking already disabled at stream fork time (constructor.rs).
+        // All CudaSlice allocations on this stream have no events — safe for
+        // CUDA Graph capture and single-stream serialized execution.
 
         // ── Compile all 6 training kernels from same module ──────────
         let (forward_loss_kernel, backward_kernel, grad_norm_kernel, adam_update_kernel, f32_to_bf16_kernel, bf16_to_f32_kernel) =
diff --git a/crates/ml/src/trainers/dqn/trainer/constructor.rs b/crates/ml/src/trainers/dqn/trainer/constructor.rs
index 2a4365526..d302e4915 100644
--- a/crates/ml/src/trainers/dqn/trainer/constructor.rs
+++ b/crates/ml/src/trainers/dqn/trainer/constructor.rs
@@ -136,7 +136,13 @@ impl DQNTrainer {
             if let candle_core::Device::Cuda(ref cuda_dev) = device {
                 let stream = cuda_dev.cuda_stream().fork()
                     .map_err(|e| anyhow::anyhow!("Failed to fork CUDA stream: {e}"))?;
-                info!("Forked dedicated CudaStream for all GPU components");
+                // Disable cudarc event tracking BEFORE any CudaSlice allocation.
+                // Must happen here — if delayed to GpuDqnTrainer::new(), earlier
+                // CudaSlice allocations (experience collector, replay buffer) will
+                // have stale events that cause hangs when tracking is later disabled.
+                // Safe: single forked stream, all GPU work is serialized.
+                unsafe { stream.context().disable_event_tracking(); }
+                info!("Forked dedicated CudaStream for all GPU components (event tracking disabled)");
                 Some(stream)
             } else {
                 None