Description

CnnDailymail.compute_score (tensorrt_llm/evaluate/cnn_dailymail.py) assigns rouge1 only inside the beam loop:

for beam_idx in range(len(outputs[0].outputs)):
    ...
    if beam_idx == 0:
        rouge1 = metrics["rouge1"] * 100
return rouge1   # UnboundLocalError if outputs[0].outputs is empty

If outputs[0].outputs is empty the loop never runs and return rouge1 raises UnboundLocalError instead of yielding a score.

Change

Initialize rouge1 = 0.0 before the loop.

Test

New CPU-only unit test in tests/unittest/others/test_cnn_dailymail_score.py: compute_score with an empty beam list returns 0.0 (constructed via __new__ to skip the heavy metric-loading __init__).

Summary by CodeRabbit

• Bug Fixes

  • Fixed CNN/DailyMail evaluation when no beam outputs are available.
  • The evaluator now returns a score of 0.0 instead of failing in this scenario.

• Tests

  • Added coverage for scoring requests with empty outputs.