liger-autopatch

You can automate Liger Kernel monkey-patching by running a pipeline that analyzes the target HuggingFace model architecture, generates lce_forward implementations and patch functions, and updates tests. It uses a decision matrix to determine model norms, MLP types, and RoPE variants.

Adding a new HuggingFace model to Liger Kernel involves analyzing the modeling file, generating lce_forward code and monkey-patch registrations, and updating convergence tests. The workflow requires human checkpoints between the analysis, generation, and validation stages.

To fix instance-level patching bugs or upstream HuggingFace changes, you can use a modify workflow that updates existing lce_forward files and monkey-patch functions. It then runs instance-patching and convergence validations with retry logic to ensure correctness.

Yes, Liger Kernel patching supports different MLP types, RoPE variants, MoE structures, and norm variants. The model analysis stage reads the modeling code and uses a 12-question decision matrix to identify these specific architecture details before generating patches.

The automation requires mandatory human checkpoints before code generation and after validation to ensure correctness. While it generates code, updates tests, and runs convergence checks with retry logic, it cannot bypass these manual review stages.

The pipeline validates generated monkey-patches by running instance-patching tests, convergence tests, and lint checks. It incorporates retry logic for failures and requires a final human checkpoint after the validation stage completes.