Problem
#451 adds vLLM cache artifact transfer so warm replicas can share runtime-generated caches before engine init. We should make the same path work for SGLang instead of leaving artifact transfer as vLLM-only.
MVP
- identify SGLang cache artifacts worth publishing/installing,
- install compatible artifacts before SGLang model init,
- publish after the source is ready/warm,
- keep it gated by
MX_ARTIFACT_TRANSFER=1 + P2P metadata,
- add a small SGLang e2e/CI check.
Problem
#451 adds vLLM cache artifact transfer so warm replicas can share runtime-generated caches before engine init. We should make the same path work for SGLang instead of leaving artifact transfer as vLLM-only.
MVP
MX_ARTIFACT_TRANSFER=1+ P2P metadata,