Problem
Comet needs to order batch and interactive work before large request bodies enter shared queues, but SMG must remain the sole authority for scheduler capacity. A Comet admission decision therefore needs a short-lived, single-use SMG capacity credit that cannot be forged, replayed, or rebound to another request.
Design context: LLM360/comet#576 and LLM360/comet#577.
First slice
Add an unwired scheduler-local credit registry that:
- binds a credit to allocator generation, policy epoch, partition, canonical model, tenant, request attempt, and output-token estimate
- uses an unguessable opaque token with redacted debug output
- returns the same active credit for an idempotent duplicate issue
- rejects a conflicting duplicate request attempt
- atomically redeems a matching credit at most once
- rejects binding mismatches without consuming the rightful credit
- expires and cancels credits while dropping owned capacity payloads outside the registry lock
- retains bounded terminal tombstones long enough to reject immediate replay
The first slice has no endpoint, config, CLI, Python binding, adaptive admission change, SlotPool change, or request-path wiring.
Follow-up
- store an acquired
SchedulerPermit as the credit payload
- add an authenticated local Comet to SMG issue/cancel API
- add proxy-only redemption and strip all client copies of the internal credit header
- validate generation handoff and expiry during blue/green rollout
Problem
Comet needs to order batch and interactive work before large request bodies enter shared queues, but SMG must remain the sole authority for scheduler capacity. A Comet admission decision therefore needs a short-lived, single-use SMG capacity credit that cannot be forged, replayed, or rebound to another request.
Design context: LLM360/comet#576 and LLM360/comet#577.
First slice
Add an unwired scheduler-local credit registry that:
The first slice has no endpoint, config, CLI, Python binding, adaptive admission change, SlotPool change, or request-path wiring.
Follow-up
SchedulerPermitas the credit payload