mirror of
https://github.com/unslothai/unsloth.git
synced 2026-10-21 22:12:39 +00:00
4dc8bf0bd2
## Summary The scheduled `Notebooks CI` run on main fails `static + transitive resolve (cron / dispatch only)`, in its `Diff Colab oracle vs committed snapshots (--strict on cron)` step. It has failed on every scheduled run since at least 2026-09-27, most recently 2026-10-18 (run 39104836103): ``` === pip-freeze.gpu.txt: upstream=720 snapshot=699 diff=226 (new=54 removed=33 changed=139) === === apt-list-gpu.txt: upstream=1163 snapshot=1160 diff=241 (new=60 removed=57 changed=124) === === os-info-gpu.txt: upstream=4 snapshot=4 diff=1 (new=0 removed=0 changed=1) === ##[error]A rule-bearing Colab oracle drifted from its committed snapshot; run `notebook_validator.py refresh-colab --all --snapshot-dir scripts/data` to acknowledge. ``` Colab moved to a new GPU runtime, and the step exists to make someone acknowledge that. Nothing in Unsloth is wrong. The committed snapshots are just behind. ## Change `python scripts/notebook_validator.py refresh-colab --all --snapshot-dir scripts/data`, as the step asks. The three snapshot files move to the current runtime. The parts the rules and the smoke seed read: | | before | after | |---|---|---| | torch / torchvision / torchaudio | 2.11.0 / 0.26.0 / 2.11.0 `+cu128` | same versions, `+cu130` | | torchcodec | 0.11.0+cu128 | 0.11.0+cu130 | | transformers | 5.16.1 | 5.18.0 | | huggingface_hub | 1.29.0 | 1.33.0 | | OS | Ubuntu 24.04.1 LTS | Ubuntu 24.04.4 LTS | ### CPU skip list (`colab_to_cpu_pin.json`) The new runtime is CUDA 13. RAPIDS wheels moved from `-cu12` to `-cu13`, and the CUDA 13 `nvidia-*` runtime wheels dropped the `-cu` suffix (`nvidia-cublas`, `nvidia-cufft`, ...). As a result, 28 of the old skip entries matched nothing, and the smoke seed would have installed 53 CUDA-only wheels on the CPU runners. - **Skip list:** re-derived for this freeze, 90 entries. - **Closure:** rebuilt from PyPI `Requires-Dist` for every pin (Linux x86_64, CPython 3.13, no extras). No retained pin requires a skipped one. torch is the only retained parent, and it is rewritten to the CPU index. ### Guards (`test_smoke_install_contract.py`) The old checks intersected a fixed `-cu12` name list with the freeze, so after the rotation they passed vacuously. Now: - **CUDA-only wheels** are classified by name shape: `-cuNN`, `cupy-cudaNNx`, `jax-cudaNN-*`, `nvidia-*` except `nvidia-ml-py`, `cuda-*`, and `triton`. Each shape has its own parametrized test cases. - **A skip entry missing from the freeze** fails, which is the signal that the map was not refreshed with the snapshot. - **The closure edge table** is the real cu13 graph, and a stale edge fails. The new freeze with the old map fails 3 of these. On main's old freeze the shape check also finds 3 CUDA wheels the old map already missed: `cuda-toolkit`, `librmm-cu12` and `libucx-cu12`. ## Testing - **Snapshot contract tests:** `test_smoke_install_contract.py`, `test_validator_fixtures.py` and `test_notebook_validator_shell_parsing.py` give 290 passed, 2 skipped on the refreshed files. After the skip-list fix, `test_smoke_install_contract.py` alone gives 52 passed. Those include the checks that every dev pin in the freeze is judged, the skip list is closed under the freeze's dependencies, and the CUDA-only wheels are skipped. - **`notebook_validator.py lint --colab-pin`:** I ran it against `unslothai/notebooks@main` with the old and the new freeze. Both give 86 errors and 6 warnings, on the same rules, notebooks and cells. Only version strings inside messages differ (`peft==0.20.0` to `0.21.1`, `+cu128` to `+cu130`). That lint step is advisory on cron (`continue-on-error`), and its backlog is unchanged. - **Seed resolve:** I replayed the seed step's filter on this freeze. It produced 3 torch pins and 626 others, and none of them is a CUDA-only name. Then `uv pip compile --python-version 3.13 --python-platform x86_64-manylinux_2_28 --only-binary :all:` (with the map's `no_binary` exceptions and the CPU torch index) resolved all of them to 630 packages, with `torch==2.11.0+cpu`. No CUDA wheel comes back in transitively: the only `nvidia` name in the resolve is the pure-Python `nvidia-ml-py`. - **The smoke-install matrix seeds from this freeze:** `Notebooks CI` with `include_smoke=true` on this branch: run 39194401235 was dispatched on the first commit, before the skip list was fixed, so I cancelled it. Run 39201002217 is running on cf2a5fbc8d, and I will add its result here. --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> |
||
|---|---|---|
| .. | ||
| colab_apt_list.gpu.txt | ||
| colab_os_info.gpu.txt | ||
| colab_pip_freeze.gpu.txt | ||
| colab_to_cpu_pin.json | ||