diff --git a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
index a89c2fb7d..6da91486c 100644
--- a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
+++ b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
@@ -2134,7 +2134,13 @@ impl GpuDqnTrainer {
             + self.config.branch_2_size + self.config.branch_3_size) * na;
 
         let d_val_norm = self.debug_buffer_norm_f32(self.d_value_logits_buf.raw_ptr(), b * na).unwrap_or(f32::NAN);
-        let d_adv_norm = self.debug_buffer_norm_f32(self.d_adv_logits_buf.raw_ptr(), b * tba).unwrap_or(f32::NAN);
+        // Guard: d_adv needs (b*tba+255)/256 blocks but grad_norm_partials only
+        // has grad_norm_blocks slots. Overflow corrupts adjacent GPU memory.
+        let d_adv_norm = if ((b * tba + 255) / 256) <= self.grad_norm_blocks {
+            self.debug_buffer_norm_f32(self.d_adv_logits_buf.raw_ptr(), b * tba).unwrap_or(f32::NAN)
+        } else {
+            f32::NAN // skip — would overflow partials buffer
+        };
         let grad_norm = self.debug_buffer_norm_f32(self.ptrs.grad_buf, self.total_params).unwrap_or(f32::NAN);
 
         tracing::warn!(
@@ -2160,10 +2166,11 @@ impl GpuDqnTrainer {
 
         // After graph replay: d_value/adv_logits contain BLENDED gradients,
         // MSE scratch contains MSE-only gradients, grad_buf contains backward output.
+        let adv_safe = ((b * tba + 255) / 256) <= self.grad_norm_blocks;
         let d_val_norm = self.debug_buffer_norm_f32(self.d_value_logits_buf.raw_ptr(), b * na).unwrap_or(f32::NAN);
-        let d_adv_norm = self.debug_buffer_norm_f32(self.d_adv_logits_buf.raw_ptr(), b * tba).unwrap_or(f32::NAN);
+        let d_adv_norm = if adv_safe { self.debug_buffer_norm_f32(self.d_adv_logits_buf.raw_ptr(), b * tba).unwrap_or(f32::NAN) } else { f32::NAN };
         let mse_val_norm = self.debug_buffer_norm_f32(self.d_value_logits_mse.raw_ptr(), b * na).unwrap_or(f32::NAN);
-        let mse_adv_norm = self.debug_buffer_norm_f32(self.d_adv_logits_mse.raw_ptr(), b * tba).unwrap_or(f32::NAN);
+        let mse_adv_norm = if adv_safe { self.debug_buffer_norm_f32(self.d_adv_logits_mse.raw_ptr(), b * tba).unwrap_or(f32::NAN) } else { f32::NAN };
         let grad_norm = self.debug_buffer_norm_f32(self.ptrs.grad_buf, self.total_params).unwrap_or(f32::NAN);
 
         tracing::warn!(