Skip to content

[tests] failed test_attention_no_look_ahead #22

@rggdmonk

Description

@rggdmonk

Summary

Failed on commit: 7736417

System MacOS.

Maybe make less strict?

    def test_attention_no_look_ahead():
        """Test that attention does not look ahead - causal masking is working correctly."""
        model, processor, collator = setup_model()
        model.eval()
    
        # Test sequences that share prefixes
        texts = ["a b c x y z", "a b d m"]
    
        # Force every word to predict a single byte (and EOS)
        # "a <eos>, b <eos>, c <eos>, <eos> <pad>" and "a <eos>, b <eos>, d <eos>, <eos> <pad>"
        processor.max_word_length = 1
    
        _, outputs = predict_dataset(texts, model, processor, collator)
        for text in texts:
            print(f"Loss for '{text}':", outputs[text].loss.cpu().numpy())
    
        # Check that the first 4 tokens have identical losses
        for i in range(4):
>           assert abs(outputs[texts[0]].loss[i] - outputs[texts[1]].loss[i]) < 1e-4, \
                f"Loss at position {i} should be identical: {outputs[texts[0]].loss[i]} vs {outputs[texts[1]].loss[i]}"
E           AssertionError: Loss at position 1 should be identical: 24.81259536743164 vs 24.813331604003906
E           assert tensor(0.0007) < 0.0001
E            +  where tensor(0.0007) = abs((tensor(24.8126) - tensor(24.8133)))

tests/test_model.py:89: AssertionError
FAILED tests/test_model.py::test_attention_no_look_ahead - AssertionError: Loss at position 1 should be identical: 24.81259536743164 vs 24.813331604003906

Reproduce

pytest

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions