What problem does it solve?
Shipping Day-0 support for a new LLM, VLM, MoE, or speculative-decoding model in SGLang involves far more than registering a model class: teams must lock immutable revisions, map architecture gaps, split work into reviewable PRs, validate seven correctness gates, and sanitize private evidence before publishing. This Skill turns that ad-hoc effort into an auditable, repeatable program.
Core Features & Use Cases
- Eight-file Day-0 bundle: Generates and validates a scope contract, architecture gap map, PR DAG, validation matrix, release lock, PR body, follow-up ledger, and sanitization report.
- Evidence discipline: Enforces machine-checkable public PR evidence records with immutable heads, states, and limitations, plus a validation hierarchy that prevents upgrading unit tests into end-to-end claims.
- Case-study guidance: Ships manually reviewed Kimi K3 and DeepSeek V4 public case studies covering hybrid attention, speculative state, quantization, platform ports, and revert handling.
- Use Case: When a new model like a hybrid KDA/MLA VLM is announced, use this Skill to produce the full support plan, classify each patch as day0-required versus performance-only, and validate the final public PR bundle with the included scripts.
Quick Start
Use the sglang-model-day0-support skill to design and validate an auditable SGLang Day-0 support bundle for this model.
