[Question] Proposal: Add Train–Inference Consistency Data Collection to slime
Maintainers usually reply within 1 day
Nobody has claimed this yet.
Assessment
- Difficulty
- 5/5
- Estimated time
- Over a week
- Newbie friendliness
- 25/100
- Issue type
- Feature
- Clarity
- Needs clarification
- Activity status
- Quiet
- Tech stack
- python
- Domain
- machine-learning
Research direction
Start with CONTRIBUTING.md and the linked documentation and FAQ, then seek maintainer confirmation because the issue names no source files, tests, or entry points. The proposal requires a detailed design and integration scope before implementation. Done would mean an accepted, default-off, parameter-controlled capability that does not change the existing training path.
Written by the indexing model from the issue text.
Description
Your Question
Background
We appreciate the slime team’s work. As an RL framework, slime is strong in engineering readiness and issue diagnosis, especially in key stages such as rollout and training updates.
In practice, we find that slime currently lacks train–inference consistency data collection. We would like to discuss with the authors whether this capability can be upstreamed.
Why This Matters
In MoE-based RL workloads, train–inference consistency is critical. Systematic forward-pass divergence can distort reward / KL signals, or in severe cases cause training divergence or policy collapse. Root causes are often complex (e.g., operator numerics, data types, framework adaptation). With train–inference consistency data collection, differences can be localized quickly, reducing precision debugging time.
What We Propose
We have already implemented a train–inference consistency data collection feature, and propose upstreaming it to slime in a parameter-controlled, default-off manner that does not affect the existing training path.
Request for Confirmation
Please confirm:
- Whether “train–inference consistency data collection / comparison” is accepted as a complementary capability for slime.
- If accepted, whether a default-off, flag-controlled, non-intrusive integration approach is acceptable.
If the direction is approved, we will follow up with a more detailed design and integration scope. We also welcome any constraints or guidance from the authors.
We would appreciate your reply at your earliest convenience. Thank you.
@zhuzilin
What I've Tried
We have already implemented a train–inference consistency data collection feature, and propose upstreaming it to slime in a parameter-controlled, default-off manner that does not affect the existing training path.
Environment (if relevant)
- slime version:
- Python version:
- PyTorch version:
- CUDA/ROCm version:
- GPU type and count:
- OS:
Additional Context
No response
Pre-submission Checklist
- I have read the CONTRIBUTING.md and understand the collaboration scope.
- I have read the documentation and FAQ and my question is not answered there.
- I have searched for existing issues and my question has not been asked before.
- Dominant language
- Python
- Stars
- 8.8k
- Forks
- 1.3k
- Avg merge
- 19h
- Merged PRs (30d)
- 26
Getting set up
- No Dockerfile or Docker Compose file
- No pull request template
- Read the contributing guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
More from THUDM/slime
-
[Bug] Math reward accepts distinct large integers after float normalizationPossibly taken @lll-hhh claimed this 2 days ago. Open
Difficulty 2/5 1-3 hours Newbie friendliness 78/100
Maintainers usually reply within 1 day
-
gpqa: when no option letter can be extracted, a single-letter gold label is substring-matched against the whole reply, so a reply that never answers can score 1.0Possibly taken @shaurya416 claimed this 38 days ago. Open
Difficulty 2/5 1-3 hours Newbie friendliness 88/100
Maintainers usually reply within 1 day
-
[Bug] shipped dynamic-sampling filters crash on fan-out groups (list[list[Sample]]) from multi-turn agent rolloutsPossibly taken @QIU-Guanzong claimed this 50 days ago. Open
Difficulty 2/5 1-3 hours Newbie friendliness 84/100
Maintainers usually reply within 1 day
-
[Bug] W&B initialization fails with wandb 0.28.2 because wandb.util.generate_id was removedPossibly taken @cyyecao-lappland claimed this 68 days ago. Openbug
Difficulty 1/5 1-3 hours Newbie friendliness 90/100
THUDM/slime#2332 · 1 comment ·
Maintainers usually reply within 1 day
-
[Bug] solution_str = sample.prompt + sample.responsePossibly taken @mvanhorn claimed this 132 days ago. Openbug
Difficulty 2/5 1-3 hours Newbie friendliness 68/100
THUDM/slime#1829 · 1 comment ·
Maintainers usually reply within 1 day
Similar issues
-
URLValidator ignores an explicit empty scheme allowlistPossibly taken @IlyaKhalafi claimed this today. Open
Difficulty 2/5 1-3 hours Newbie friendliness 85/100
tortoise/tortoise-orm#2393 ·
Maintainers usually reply within 2 days
-
bug
Difficulty 2/5 1-3 hours Newbie friendliness 62/100
frappe/frappe_docker#2046 ·
Maintainers usually reply within 1 day
-
Difficulty 2/5 1-3 hours Newbie friendliness 62/100
-
complete_box_iou_loss / complete_box_iou: NaN gradients for zero-height boxes, NaN for point boxesOpen
Difficulty 2/5 1-3 hours Newbie friendliness 62/100
Maintainers usually reply within 1 day
-
AXI SMU SPEC
Difficulty 2/5 1-3 hours Newbie friendliness 62/100
tenstorrent/tt-oca-harness#3556 ·
Maintainers usually reply within 1 day