diff --git a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
index 066fa7d82..182c14c82 100644
--- a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
+++ b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
@@ -869,12 +869,26 @@ impl GpuDqnTrainer {
             .map_err(|e| MLError::ModelError(format!("stream sync before readback: {e}")))?;
 
         // ── Gather 2 scalars to host: total_loss + grad_norm ────────
+        // These buffers were written by the CUDA Graph (which runs with event
+        // tracking disabled). Their CudaSlice write events are stale, so
+        // cudarc's memcpy_dtoh would fail with CUDA_ERROR_INVALID_VALUE.
+        // Use raw_device_ptr (ManuallyDrop guard) + synchronous cuMemcpyDtoH.
         let mut loss_host = [0.0_f32; 1];
-        self.stream.memcpy_dtoh(&self.total_loss_buf, &mut loss_host)
-            .map_err(|e| MLError::ModelError(format!("DtoH total_loss: {e}")))?;
         let mut norm_host = [0.0_f32; 1];
-        self.stream.memcpy_dtoh(&self.grad_norm_buf, &mut norm_host)
-            .map_err(|e| MLError::ModelError(format!("DtoH grad_norm: {e}")))?;
+        let loss_ptr = raw_device_ptr(&self.total_loss_buf, &self.stream);
+        let norm_ptr = raw_device_ptr(&self.grad_norm_buf, &self.stream);
+        let r1 = unsafe { cudarc::driver::sys::cuMemcpyDtoH_v2(
+            loss_host.as_mut_ptr().cast(), loss_ptr, 4,
+        )};
+        if r1 != cudarc::driver::sys::CUresult::CUDA_SUCCESS {
+            return Err(MLError::ModelError(format!("DtoH total_loss: {r1:?}")));
+        }
+        let r2 = unsafe { cudarc::driver::sys::cuMemcpyDtoH_v2(
+            norm_host.as_mut_ptr().cast(), norm_ptr, 4,
+        )};
+        if r2 != cudarc::driver::sys::CUresult::CUDA_SUCCESS {
+            return Err(MLError::ModelError(format!("DtoH grad_norm: {r2:?}")));
+        }
         self.scalars_readback_host = [loss_host[0], norm_host[0]];
 
         Ok(FusedTrainScalars {