mirror of
https://github.com/unslothai/unsloth.git
synced 2026-08-23 15:53:46 +00:00
484a0d9f64
`Notebooks CI :: static + transitive resolve` has failed all 20 of its last cron runs on one step, `colab-diff --strict`. Nothing is broken. That step is a drift alarm: it compares googlecolab/backend-info against scripts/data/colab_*.txt and fails when Colab rotates, and the `Refresh Colab oracle` step below it writes only into the runner's checkout. So no refresh is ever committed back, and the alarm re-fires every night until someone acknowledges it, which is what this does. The step's own comment names the action: "refresh-colab --all acknowledges a report." The drift is 127 entries: 5 new, 6 removed, 116 version bumps. Why this is safe to take wholesale ------------------------------------------------------------------------ The snapshot is the oracle `lint --colab-pin` resolves R-INST-002/003/004/005 against, so refreshing it could in principle move a verdict. It does not. Exactly one rule-relevant pin drifted, peft 0.19.1 -> 0.20.0, and both sides sit above the 0.19 trigger in the peft/torchao floor table; torchao itself does not appear in the diff at all. Confirmed rather than argued, by linting unslothai/notebooks@main against the old and new oracles and comparing the findings: old 6 R-API-003, 8 R-INST-001, 36 R-INST-003, 1 R-INST-005 new 6 R-API-003, 8 R-INST-001, 36 R-INST-003, 1 R-INST-005 Identical, so this commit changes what the alarm says and nothing about what the linter concludes. Worth reading before the next rotation ------------------------------------------------------------------------ os-info drift is printed but never fails, and nothing reads that file, so this is not a CI matter. It is still the most interesting line in the diff: Colab has moved from Python 3.12.13 to 3.13.15 (and R 4.6.0 to 4.6.1). Flagging it here because it is the kind of change that surfaces in a notebook long before it surfaces in a lint rule. The 36 R-INST-003 findings are pre-existing and remain open. They are real: peft's is_torchao_available() raises ImportError when torchao is installed below 0.16.0, which is a runtime check inside peft rather than a declared dependency, so it is invisible in PyPI metadata and reachable from any get_peft_model call. That backlog belongs to unslothai/notebooks and the lint step is continue-on-error until it clears. Verified: colab-diff --strict exits 0; tests/notebooks 32 passed, 2 skipped. Co-authored-by: danielhanchen <unslothai@gmail.com> |
||
|---|---|---|
| .. | ||
| colab_apt_list.gpu.txt | ||
| colab_os_info.gpu.txt | ||
| colab_pip_freeze.gpu.txt | ||
| colab_to_cpu_pin.json | ||