host-opt

Device profilers only cover device-side kernels, so host overhead like encoding, transfer, disk writes, and loading is invisible to them. Build a fixed-overhead account with in-process timing points or server-side timing plus log lines, splitting each segment's share of same-caliber end-to-end latency.

Decompose delivery into four segments: encoding (critical path, codec parameters, hardware encoder, repeated encoding), transfer path (hop count times bytes per hop), disk writes (per-request writes, duplicate copies), and async offloading. Async moves are only valid when return timing, error propagation, and retry semantics stay unchanged.

Warmup silently fails when the warmup input shapes or dtypes differ from production, creating a separate set of cache entries. Verify warmup with hit criteria such as cache entry counts, graph hits, or log lines rather than total elapsed time.

No. Concurrency, throughput, batch scheduling, queue depth, and rate limiting are explicitly out of scope because they change all performance baselines and SLO semantics. This Skill only reduces fixed overhead without changing the request concurrency model.

Environment installation handles getting weights in place, including download, mounting, and path verification. This Skill optimizes the loading process itself: fewer reads, fewer copies, parallelism where possible, and avoiding duplicate loads across processes or instances.

Use same-window A/B comparison and re-issue a same-caliber overhead account showing which line improved. Print actually-read values like environment variables and encoding parameters rather than trusting log strings, and confirm delivered artifacts are byte-identical when claiming results are unchanged.