fix: use seq_length instead of padded_seq_length for topk output padding (#1929)

Signed-off-by: Zhaopeng Qiu <qiuzhaopeng@foxmail.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Author: zpqiu

nemo_rl/models/policy/workers/megatron_policy_worker.py

@@ -939,7 +939,7 @@ def collection_fn(_):
             for out in list_of_outputs:
                 tk = out["topk_logits"]
                 ti = out["topk_indices"]
-                pad_len = padded_seq_length - tk.shape[1]
+                pad_len = seq_length - tk.shape[1]
                 if pad_len > 0:
                     tk = torch.nn.functional.pad(tk, (0, 0, 0, pad_len), value=0.0)
                     ti = torch.nn.functional.pad(ti, (0, 0, 0, pad_len), value=0)