llava-onevision2-consistency

To validate HuggingFace and Megatron checkpoint consistency for LLaVA-OneVision2, run the targeted pytest suite and legacy script workflow via run_consistency_tests.sh to verify behavioral parity across TP and PP configurations.

Running HF vs Megatron consistency tests requires configuring HF_MODEL_PATH, MCORE_CHECKPOINT_PATH, CONSISTENCY_TEST_TP, and CONSISTENCY_TEST_PP in your pytest environment to execute cross-framework validation.

Cross-framework checkpoint consistency testing for LLaVA models covers weight, encoder, and end-to-end LLM checks to ensure reliable behavioral parity between HuggingFace and Megatron/MCore deployments.

You can use a pre-converted MCore artifact for Megatron consistency validation, or the skill guides you through HF to mcore conversion when needed to verify identical downstream behavior.

HuggingFace to Megatron consistency validation works across different TP and PP configurations, enabling reproducible cross-checks in varied hardware setups using both the fast pytest suite and legacy scripts.