AdvancedMIS Rollout Correction

MIS Rollout Correction

This page documents advanced/validate_mis_rollout_correction.py in mint-quickstart.

What this validation proves

  • create_model accepts a session-level Seq-MIS rollout_correction_config
  • a later forward_backward(..., loss_fn="importance_sampling") succeeds without resending rollout config per step
  • the response contains valid loss_fn_outputs
  • the temporary model is cleaned up by default

This is a narrow integration check, not a complete RL recipe. Today this flow requires a supported MoE / Megatron base model.

Recommended path today

Use the direct request / Tinker-compatible request path shown in the quickstart validation script. This page does not claim that a higher-level MinT SDK helper already exists for MIS rollout-correction wiring.

Command

export MINT_API_KEY=sk-...
export MINT_BASE_URL=<your-region-endpoint>
python advanced/validate_mis_rollout_correction.py --base-model Qwen/Qwen3-30B-A3B-Instruct-2507

Use the MinT endpoint that matches your region:

  • Mainland China: https://mint-cn.macaron.xin/
  • Outside Mainland China: https://mint.macaron.xin/

Supported inputs

  • --base-url
  • --api-key
  • --base-model
  • --lora-rank
  • --mis-threshold
  • --create-timeout-s
  • --forward-backward-timeout-s
  • --poll-interval-s
  • --skip-cleanup

MinT-style environment variables win over Tinker-compatible aliases:

  • MINT_BASE_URL over TINKER_BASE_URL
  • MINT_API_KEY over TINKER_API_KEY
  • MINT_BASE_MODEL over TINKER_MODEL

Expected output

[config] base_url=<your-region-endpoint> base_model=Qwen/Qwen3-30B-A3B-Instruct-2507 lora_rank=8 mis_threshold=1.1
[create_model] submitted session_id=validate-mis-1234abcd
[create_model] resolved model_id=model_...
[forward_backward] submitted model_id=model_... loss_fn=importance_sampling
[forward_backward] resolved outputs=1
PASS: MIS rollout_correction request succeeded and response was valid
[cleanup] deleted model_id=model_...

Common failure cases

  • FAIL [config]: missing API key
  • FAIL [create_model]: model creation rejected, timed out, or the model is unavailable
  • FAIL [forward_backward]: the training request failed after model creation
  • FAIL [malformed_response]: the server returned no loss_fn_outputs
  • [cleanup] warning: validation finished, but best-effort deletion failed

Status language

When recording results for this flow, use:

  • CONFIRMED: remote MinT run passed end to end
  • PARTIAL: docs/script exist, but a server-side blocker prevents confirmation
  • INSUFFICIENT_DATA: docs/script exist, but there is not yet reliable run evidence

See also Tinker Compatibility for compatibility status notes.

Resume