mindspeed-fsdp2-model-migration

To migrate a model to MindSpeed-MM FSDP2 while preserving source semantics, you create a model plugin under mindspeed_mm/fsdp/models, register it via @model_register, and validate loading signatures to ensure forward compatibility.

The model_contract_checklist and signature_compat_report are optional validation checks that verify from_pretrained and _from_config path compatibility, outputting compliance reports to ensure consistent model loading behavior during FSDP2 migration.

To handle token and embedding updates during MindSpeed-MM FSDP2 migration, you perform token displacement, resetting, or initialization as needed to adapt embeddings, ensuring inference consistency and forward compatibility within the model plugin.

Yes, MindSpeed-MM FSDP2 migration requires compatible from_pretrained and _from_config loading paths to be established within the model plugin, ensuring the registered model executes correctly and maintains behavioral consistency across the loading chain.

Behavior consistency is lost when model plugins lack proper registration contracts or compatible loading signatures, which this FSDP2 migration framework resolves by enforcing @model_register contracts and validating from_pretrained paths.

Yes, you can create FSDP2-compatible model plugins for target architectures under mindspeed_mm/fsdp/models, provided you satisfy the registration contracts, loading signatures, and embedding/token updates required for forward compatibility.