better check

users/zeyer/experiments/exp2024_04_23_baselines/ctc_recog_ext.py

@@ -710,7 +710,6 @@ def finish(self, state: LMState):
         assert seq_len <= max_seq_len
         results = fl_decoder.decode(emissions_ptr, seq_len, model.wb_target_dim.dimension)
         hyps_per_batch = [result.tokens for result in results]
-        assert all(len(hyp) == seq_len for hyp in hyps_per_batch)
         scores_per_batch = [result.score for result in results]
         best_word_seq = [
             model.wb_target_dim.vocab.id_to_label(label_idx) if label_idx >= 0 else str(label_idx)
@@ -725,6 +724,9 @@ def finish(self, state: LMState):
             f" LM recalc whole seq count {fl_lm._count_recalc_whole_seq}"
             f" mem usage {dev_s}: {' '.join(_collect_mem_stats())}"
         )
+        assert all(
+            len(hyp) == seq_len for hyp in hyps_per_batch
+        ), f"seq_len {seq_len}, hyps lens {[len(hyp) for hyp in hyps_per_batch]}"
         if len(results) >= n_best:
             hyps_per_batch = hyps_per_batch[:n_best]
             scores_per_batch = scores_per_batch[:n_best]