fix understand_audio: clone tensors for inference mode, working on GPU (52s)

train_engine.py

@@ -3414,6 +3414,7 @@ def understand_audio(
     past_think = False
 
     for step in range(max_new_tokens):
+        logits = logits.clone()
         # After </think>: block audio codes so the LM only generates text
         if past_think:
             logits[0, _AUDIO_CODE_BASE:] = float("-inf")
@@ -3422,9 +3423,8 @@ def understand_audio(
         if temperature <= 0:
             next_id = int(logits[0].argmax().item())
         else:
-            scaled = logits[0] / temperature
+            scaled = logits[0].clone() / temperature
             if top_k > 0:
-                # Zero out everything below top_k
                 topk_vals, _ = torch.topk(scaled, min(top_k, scaled.shape[0]))
                 scaled[scaled < topk_vals[-1]] = float("-inf")
             if top_p > 0 and top_p < 1.0: