First full MNode image for Decode PoC (MiniMax): testing #1689
A full MLNode image for Decode PoC on MiniMax, with the current implementation verified. Collecting coefficients or thresholds is not acceptance. Self-validation and cross-device validation must pass under every scenario below. They must not break when PoC, Decode PoC, and inference are combined. Related: integration / design — #1688. DeepSeek thresholds and seeding — #1690. Confirm live on-chain models from poc_params.models if the MiniMax id has drifted. Target model for this image: MiniMaxAI/MiniMax-M2.7.
Process — Experiments completed and written up (hardware, models, what passed, what did not) — MLNode starts and reaches a working state (does not crash on boot) — Current Decode PoC implementation is tested against the matrix below, not only coefficients — Image rechecked before handover — If a scenario is skipped: notes before handover, with the reason
Six steps
- Self-validation on one machine Decode PoC validates against itself on a single host / GPU
- Cross-device validation Decode PoC validates across different hardware
- Chain cycle: PoC starts, then PoC stops — Cycle runs with no inference in the window — self-validation — Cycle runs with no inference in the window — cross-device — Cycle runs with inference in the window — self-validation — Cycle runs with inference in the window — cross-device
- Decode cycle: Decode PoC starts, then Decode PoC stops, then inference runs — Decode ON → Decode OFF → inference — self-validation — Decode ON → Decode OFF → inference — cross-device
- Backward compatibility: old (prefill) PoC is turned on — Old PoC still works after Decode PoC has been used — self-validation — Old PoC still works after Decode PoC has been used — cross-device
- Mix matrix: Decode PoC, old PoC, and inference, every combination
Static combinations
- 1. Old PoC on, Decode PoC off, Inference no — self
- 1. Old PoC on, Decode PoC off, Inference no — cross-device
- 2. Old PoC on, Decode PoC off, Inference yes — self
- 2. Old PoC on, Decode PoC off, Inference yes — cross-device
- 3. Old PoC off, Decode PoC on, Inference no — self
- 3. Old PoC off, Decode PoC on, Inference no — cross-device
- 4. Old PoC off, Decode PoC on, Inference yes — self
- 4. Old PoC off, Decode PoC on, Inference yes — cross-device
- 5. Old PoC on, Decode PoC on, Inference no — self
- 5. Old PoC on, Decode PoC on, Inference no — cross-device
- 6. Old PoC on, Decode PoC on, Inference yes — self
- 6. Old PoC on, Decode PoC on, Inference yes — cross-device
- 7. Old PoC off, Decode PoC off, Inference yes — self
- 7. Old PoC off, Decode PoC off, Inference yes — cross-device
- 8. Old PoC off, Decode PoC off, Inference no — self
- 8. Old PoC off, Decode PoC off, Inference no — cross-device
Each row is a scenario. Both self-validation and cross-device must pass.
Scenario 8 is the idle baseline (both PoCs off, no inference). If it is not a meaningful run, that is stated in notes, and the row is still marked.
Cycles (order matters)
- 9. Old PoC ON → Old PoC OFF, no inference in the window — self
- 9. Old PoC ON → Old PoC OFF, no inference in the window — cross-device
- 10. Old PoC ON → Old PoC OFF, inference in the window — self
- 10. Old PoC ON → Old PoC OFF, inference in the window — cross-device
- 11. Decode PoC ON → Decode PoC OFF → inference after — self
- 11. Decode PoC ON → Decode PoC OFF → inference after — cross-device
- 12. Decode PoC ON → Decode PoC OFF, no inference after — self
- 12. Decode PoC ON → Decode PoC OFF, no inference after — cross-device
- 13. Decode PoC used, then old PoC ON (backward compatibility) — self
- 13. Decode PoC used, then old PoC ON (backward compatibility) — cross-device
- 14. Old PoC ON → Decode PoC ON — self
- 14. Old PoC ON → Decode PoC ON — cross-device
- 15. Decode PoC ON → old PoC ON — self
- 15. Decode PoC ON → old PoC ON — cross-device
- 16. Both ON → Decode PoC OFF (old PoC remains) — self
- 16. Both ON → Decode PoC OFF (old PoC remains) — cross-device
- 17. Both ON → old PoC OFF (Decode PoC remains) — self
- 17. Both ON → old PoC OFF (Decode PoC remains) — cross-device
- 18. Inference → old PoC ON → old PoC OFF → inference — self
- 18. Inference → old PoC ON → old PoC OFF → inference — cross-device
- 19. Inference → Decode PoC ON → Decode PoC OFF → inference — self
- 19. Inference → Decode PoC ON → Decode PoC OFF → inference — cross-device
This image (MiniMax Decode PoC)
- Full MLNode image tag published
- PoC (old and/or decode, as in the matrix) stays within the allowed memory budget
- Inference works on MiniMaxAI/MiniMax-M2.7 where the scenario includes inference
- Experiment notes attached (hardware for self, hardware pair for cross-device, logs / notebooks)
- Every scenario above is pass, fail, or skipped-with-reason
Handover
- Image tag published
- Matrix filled
- Limitations, if any, written in notes before handover
- The image is treated as delivered when this checklist is complete.
💬 Comments (5)
ty
Dependency: blocked by #1688. @tcharchian — this changes the ordering.
Why
- Acceptance here is the acceptance of a frozen algorithm: 19 scenarios, each in self-validation and cross-device validation. Cross-device results depend on every consensus constant (expert seeding, ladder base, artifact payload, admission behaviour) — a change after the run invalidates the whole matrix.
- Two items on #1688 would still change those constants: whether the old prefill PoC stays behind a flag, and sign-off on the constants themselves (@axeltec-gonka, not given).
- The image is built from gonka-ai/vllm#100 and gonka-ai/gonka-vllm-plugins#8; neither has landed in the intake branches.
Can run before that, as smoke rather than acceptance: image build, boot/crash checks, old PoC after Decode PoC (scenarios 13-17), ON/OFF cycles without inference.
Next: smoke set starts after the gonka-ai/gonka-vllm-plugins#8 refresh, 2026-09-05. The acceptance matrix is scheduled here the day #1688 closes. Status update every Monday, next 2026-09-07.
Status: paused on our side.
@vbgd0 built the decode-PoC image himself (tag decode-poc-int, #1743) and handed testing to another team, so we are not working on this issue right now.
We may come back to it if the images turn out to need building per hardware configuration.
Status: paused on our side, unchanged since 09-11.
The image and its testing are with @vbgd0 and the other team; nothing for us to run. Next update on change.
🔄 Auto-synced from Issue #1689 every hour.
TY!