Skip to content

Online request profiling and risk estimation #1510

Description

@kms12425-ctrl

Background

Accurate request profiling is required for admission, scheduling, and memory residency decisions.

Scope

  • Build a cost estimator from prompt length, output length, history, and live runtime signals.
  • Estimate both latency risk and memory residency cost.
  • Support calibration against experimental traces.

Deliverables

  • Profiling module.
  • Estimation formula documentation.
  • Calibration and offline analysis scripts.

Acceptance Criteria

  • Output: expected_prefill_ms, expected_decode_ms, expected_kv_growth_bytes, expected_weight_residency_cost, queue_slack_ms, deadline_miss_risk.
  • Profiling outputs are consumable by scheduler and admission logic.
  • Calibration workflow is reproducible on saved traces.

Dependencies

Notes

Supports informed admission control under dynamic multi-model workloads.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions