ReadonlygammaThe trained block width — the server pins numDraftTokens to this so the group never asks for more positions than the block was trained for.
ReadonlygroupedReadonlyidHuman-readable id (registry id / path tail) for logs + cache namespacing.
ReadonlyweightsStaticload
Server-lifetime owner of the draft machinery (the loaded draft model). Its grouped rows are how every request drafts, one request or several.