Skip to content

Phase 2: Add linear_logp strict-mode checks and fallback tests #21

Description

@inaniloquentee

Background

Split from #6 Phase 2: Productionize linear_logp.

This issue covers strict-mode validation and fallback behavior for linear_logp. Valid configurations should run without fallback under strict fast mode, while unsupported configurations should either fall back with a structured reason or fail clearly under strict policy.

Scope

Add strict checks for:

  • temperature compatibility
  • active mask validity
  • TP metadata completeness
  • dtype/downcast contract
  • unsupported CP redistribution
  • entropy requests
  • unsupported backend/dtype/hardware combinations
  • missing backward saved-state support
  • fallback behavior under --rlk-fast strict

Add tests for:

  • --rlk-fast auto
  • --rlk-fast strict
  • native fallback path
  • unsupported configuration with structured fallback reason
  • strict policy failure when fallback is not allowed

Acceptance Criteria

  • Validated configurations have fallback=0 under strict fast mode.
  • Unsupported configurations fall back with a structured reason under non-strict policy.
  • Unsupported configurations fail clearly under strict policy when fallback is not allowed.
  • Tests cover both successful strict execution and expected fallback/failure cases.
  • Logs explain every linear_logp backend decision.

Metadata

Metadata

Assignees

No one assigned

    Labels

    consistencyRollout-training consistencyphase-2Phase 2 workstreamrl-kernelRL-Kernel integration

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions