[RobotSail]

Summary

Distills operational learnings from merged PRs since the last AGENTS.md update (June 3, 2026).

New entries

Working With This Codebase Specifically:

• Newer NVIDIA drivers (580+/610+) crash vLLM's torch.compile pipeline — use enforce_eager=True
• Unsloth import order requirement in ART subprocess (import unsloth before import art)

Past Incidents:

• vLLM torch.compile crash on newer drivers — training_hub PR #98
• trl GuidedDecodingParams import failure in ART subprocess — training_hub PR #98

OSFT Specific:

• osft_output_dtype removed, now follows train_dtype — mini_trainer PR #108

PRs analyzed (merged since June 3, 2026)

• training_hub #98 — Fix ART backend compatibility with newer NVIDIA drivers
• training_hub #83 — Add ITS Hub rollout adapter for GRPO training
• training_hub #77 — Add GEPA prompt optimization algorithm
• training_hub #75 — Add templated OSFT notebook
• mini_trainer #108 — Remove osft_output_dtype parameter
• instructlab/training #704 — Rename checkpoint trigger file (no new learnings — already documented)

Not captured (and why)

• GEPA algorithm (PR Add GEPA prompt optimization algorithm #77), ITS Hub rollout adapter (PR Add ITS Hub rollout adapter for GRPO training #83), templated OSFT notebook (PR Add templated OSFT notebook (#69) #75): new features with code visible in source — architecture/usage derivable by reading the code, no operational surprises to document.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Documentation
  • Added troubleshooting guidance for runtime failures in the training pipeline, including vLLM torch.compile crashes and workarounds.
  • Documented subprocess import-order requirements to avoid vLLM shim issues.
  • Updated precision control guidance for OSFT, clarifying available configuration options.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: e3257b5e-1cd4-49d2-ba0c-ffd21f3036f0

📥 Commits

Reviewing files that changed from the base of the PR and between ca6352d and 598d6ed.

📒 Files selected for processing (1)

• AGENTS.md

---

📝 Walkthrough

Walkthrough

Documentation-only update to AGENTS.md adding two runtime failure modes with workarounds (vLLM torch.compile crash and unsloth/art import order) in both the operational guidance and Past Incidents sections, plus a note that osft_output_dtype was removed and OSFT output dtype now follows train_dtype.

Changes

AGENTS.md Runtime Troubleshooting and OSFT Dtype Docs

Layer / File(s)	Summary
vLLM and import-order failure guidance AGENTS.md	Adds two operational bullets (lines 299–300) for vLLM torch.compile crash on newer NVIDIA drivers with enforce_eager=True workaround, and required unsloth-before-art import ordering. Adds two corresponding Past Incidents entries (lines 322–323) with expanded detail on each failure.
OSFT output dtype change note AGENTS.md	Adds a single line (360) noting that osft_output_dtype was removed and OSFT output dtype now follows train_dtype, with osft_upcast_dtype, train_dtype, and save_dtype as the remaining precision controls.

Estimated code review effort

🎯 1 (Trivial) | ⏱️ ~2 minutes

Possibly related PRs

• Red-Hat-AI-Innovation-Team/training_hub#37: Established the baseline AGENTS.md agent guidelines that this PR extends with new runtime troubleshooting bullets and OSFT precision notes.
• Red-Hat-AI-Innovation-Team/training_hub#98: Implemented the ART/vLLM fixes (enforce_eager=True propagation and unsloth-before-art import ordering) that this PR documents as past incidents and operational guidance.

Poem

🐇 A rabbit once trained with a vLLM,
But torch.compile crashed — what a problem!
"enforce_eager=True!"
And import unsloth too,
Now the docs keep us out of the bedlam! 🥕

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and specifically summarizes the main change: updating AGENTS.md documentation with learnings from recent work in June 2026, matching the core purpose of the changeset.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch agent/mini-trainer-learner/b79e6499

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}