Entail

About

Entail checks whether the configuration a model's files declare, such as RoPE base and scaling, soft-capping, sliding windows, chat templates, and prediction type, actually matches what the inference engine (transformers, vLLM, SGLang, diffusers) applies at runtime. It repairs mismatches before the first token when possible and logs exactly what broke otherwise; the project found that vLLM's launch-time rope_scaling override silently changes the RoPE base for 64 of the 180 most-downloaded Hugging Face models that use it, at roughly 1% added load time.