parity-testing

To verify numerical parity for Hugging Face to Megatron weight conversions, run the specified example conversion verification scripts to check exact round-trip state-dict matches and forward-pass logit equivalence within documented tolerances.

Debugging HF to Megatron checkpoint mismatches involves an ordered triage path using built-in tensor comparison utilities to localize whether the failure stems from state-dict mapping errors, TP/PP/EP partitioning, or architectural configuration.

Forward-pass logit equivalence testing runs GPU forward passes on converted checkpoints to validate output similarity, serving as the second validation level after confirming exact state-dict mapping matches.

Yes, parity testing validates HF to Megatron checkpoint round-trips across TP, PP, and EP settings by verifying multi-GPU sharding and aggregation to ensure weights map correctly across different parallelism configurations.

Missing or extra keys in Hugging Face to Megatron state-dict mappings indicate mapping failures during conversion, which you can interpret and diagnose using the built-in reusable tensor and state-dict comparison patterns.

You should run a lightweight training parity check after validating exact state-dict mapping and forward-pass logit equivalence when you need to verify numerical stability and identical gradients during actual model training iterations.