Sparse-vLLM

Review Sparse-vLLM diffs for correctness, sparse-runtime architecture, scheduling semantics, reproducibility, public documentation hygiene, performance, and tests. Use when reviewing PRs, git diffs, sparse method integrations, cache-manager or scheduler changes, benchmark/evaluation scripts, docs changes, or when the user asks for a code review; if no range is specified, diff the current branch against main.

Add or refactor a first-class Sparse-vLLM sparse method alongside vanilla, SnapKV, OmniKV, QuEST, and DeltaKV. Use when Codex needs to introduce a new `vllm_sparse_method`, move method logic out of `attention.py` or `utils/`, add method-specific cache metadata or decode-time view building, wire config and registration, and preserve the repo's cache-manager-first architecture.

Add or refactor a first-class Sparse-vLLM sparse method alongside vanilla, SnapKV, OmniKV, QuEST, and DeltaKV. Use when Codex needs to introduce a new `vllm_sparse_method`, move method logic out of `attention.py` or `utils/`, add method-specific cache metadata or decode-time view building, wire config and registration, and preserve the repo's cache-manager-first architecture.