diff --git a/test/registered/amd/accuracy/mi30x/test_vlms_mmmu_eval_amd.py b/test/registered/amd/accuracy/mi30x/test_vlms_mmmu_eval_amd.py index 13eb87ff4f..fd2006e841 100644 --- a/test/registered/amd/accuracy/mi30x/test_vlms_mmmu_eval_amd.py +++ b/test/registered/amd/accuracy/mi30x/test_vlms_mmmu_eval_amd.py @@ -229,7 +229,10 @@ class TestNightlyVLMMmmuEvalAMD(unittest.TestCase): eval_name="mmmu", num_examples=100, num_threads=64, - max_tokens=30, + # The MMMU prompt asks for step-by-step reasoning ending in + # "Answer: X" (#21841); 30 tokens truncated it before the + # answer, so the parser fell back to "A" for every model. + max_tokens=1024, ) # Run evaluation with timing