model-migrate-flagos

To migrate a vLLM model into the vllm-plugin-FL project, invoke this skill with the model name to trigger a 13-step copy-then-patch workflow that applies compatibility patches, registers the model, and runs end-to-end verification against upstream ground truth.

Backporting upstream vLLM models to an older plugin version involves a copy-then-patch workflow utilizing a config-bridge approach to avoid upgrading transformers, ensuring compatibility while preserving the original upstream implementations during migration.

End-to-end verification for migrated vLLM models works by running provided scripts like e2e_eval.py to perform token-level comparison against upstream ground truth, ensuring the migrated model in the FL plugin produces repeatable and accurate results.

Yes, you can use a config-bridge approach to avoid upgrading transformers when migrating vLLM models into the vllm-plugin-FL project pinned to v0.13.0, allowing you to apply necessary compatibility patches without forcing a full dependency upgrade.

The scripts needed for vLLM model migration, benchmarking, and serving include validate_migration.py, benchmark.sh, serve.sh, request.sh, and e2e_eval.py, which collectively orchestrate the migration pipeline and validate the end-to-end results.

Yes, the vllm-plugin-FL migration workflow supports automatic model registration, automatically registering and validating the migrated models within the plugin as part of its comprehensive 13-step migration pipeline.