What problem does it solve?
When optimizing or patching model serving paths, engineers often repeat work that upstream frameworks already solved or miss known regression risks. This Skill provides a queryable knowledge base of PR-driven model optimization histories so agents can ground patch planning, fast-path selection, and validation choices in inspected PR diffs rather than guesswork.
Core Features & Use Cases
- Cross-framework PR history search: Query bilingual model-family history docs for SGLang, vLLM, TensorRT-LLM, and TokenSpeed using a local search helper with framework, model, language, and result-limit filters.
- Actionable evidence extraction: Surface implementation files, symbols, prior fusions, quantization and MoE changes, validation lanes, and regression risks from per-PR diff audit cards.
- SOTA loop integration: Feed history evidence into SGLang SOTA and Humanize optimization loops before patch planning, including competitor history when another framework is faster.
- Use Case: Before writing a fused QK-norm patch for Qwen3.5 on SGLang, query the knowledge base to discover that vLLM and TensorRT-LLM already landed similar fusions, then cite the relevant PRs and source files in the run's analysis notes.
Quick Start
Ask the agent to look up prior PR optimization history for a model family, for example by running the query helper with a model slug and search terms before planning any serving patch.
