Skip to content

0.4.0 release verification

Release source: 2edf707ac2040a2dc7a55dfe26dfa9f427b8fc45, annotated tag v0.4.0. Baseline was 053ce9e115b213e0ded3634e00d564ef97561f9b and published 0.3.0. Implementation and confirmed delivery fixes were integrated through PRs 11, 12, 13 and 14. The eight-job source CI passed on the exact release commit.

Confirmed problems and implementation

The old recipe had divergent TRL/raw-data and pretokenized paths, first-turn formatting, a BF16-to-FP16 load fallback without matching training precision, and broad TypeError retries that removed requested options. One package-backed Transformers runtime, shared validated configuration and dataset processing now serve generated recipes and benchmarks. The default remains all-token loss. System and multi-turn history are preserved; assistant-only supervision requires verified native masks or completion offsets. Causal-shift supervision, EOS/PAD, malformed rows, truncation and staged saves have behavioral regressions.

The actual old tiny recipe completed one update after a Windows GBK UnicodeDecodeError prevented TRL import and selected the legacy fallback. This does not establish that all modern TRL versions fail. The new runtime removes the confirmed semantic split independently of that environment-specific result.

Actual Transformers 5.8.1 configuration merging also retained a checkpoint's 4-bit configuration despite a requested int8 configuration. Declared native pre-quantized bases are now rejected before weight loading, including metadata discovery. This config-only probe is not a quantization accuracy experiment. The shared guard is covered without downloading weights. Doctor installation advice now matches qualified constraints, and the PyPI README's main links use absolute versioned repository URLs after a published relative link returned 404.

Key implementation: src/canifinetune/configuration.py, training/data.py, training/runtime.py, bench/runner.py, estimator/memory.py, utils/gpu.py, evidence.py, demo.py, packaged recipe/web resources, and .github/workflows/{ci,release}.yml. See the implementation record and training contract.

Source tests and installed candidate

Commands used in isolated project environments:

python -m pytest -q -m "not training" --cov=canifinetune --cov-fail-under=70
python -m pytest -q -m training
ruff check .
ruff format --check .
mypy src
python scripts/check_generated.py
python -m build --outdir .cache/release-0.4.0
python -m twine check .cache/release-0.4.0/*.whl .cache/release-0.4.0/*.tar.gz
python scripts/verify_candidate.py --dist .cache/release-0.4.0 --tag v0.4.0 --commit 2edf707ac2040a2dc7a55dfe26dfa9f427b8fc45

Core: 102 passed, 71.44% coverage; 14 training tests selected separately. Each minimum/recommended local stack passed all 14 real CPU integration tests. CI covers core Python 3.10–3.14, Linux minimum/recommended training and Windows recommended training on Python 3.12. The original 70% coverage gate is retained. Optional torch runtime execution is verified separately rather than represented as torch-free core coverage. Ruff, format, mypy and generated Python checks passed.

The isolated build produced an sdist first and built the wheel from that sdist. Twine, metadata, README, license and packaged-resource checks passed. The exact two files below were installed non-editably, qualified outside the checkout and uploaded as release assets. Private earlier candidates were excluded. No rebuild or file substitution occurred after this qualification.

Distribution SHA-256
canifinetune-0.4.0-py3-none-any.whl 080b1bb450c8abb8cf8213f15717131fefe90544e79b06cdd8416a1e5e07cd58
canifinetune-0.4.0.tar.gz 6294418f881271483e7aeaba8bcd9094c323d0a8c4ce635b68a54c1bf6c6fcdb

Package source fingerprint: b0cdb9ac5cc747e284e584d5de7c3312442ac1d1e905fffffc991a44c969b10c. The candidate manifest binds the commit, files and qualification receipt. The receipt's SHA-256 is 1cc0924f3bebc046a0a2777fd66fdf916d76343d5e24954abe762c3656749478.

Core qualification ran without torch: installed import/version, CLI, offline estimate/recommend, recipe generation/AST and demo/core equality. The README estimate is 8.420 GiB, yes against 16 GiB, heuristic medium. Missing-torch bench exits 1 and records a dependency failure with absent peaks; recipe input/dependency errors exit 2, and genuine recipe OOM exits 3.

Minimum CPU stack: Python 3.12.13, torch 2.6.0+cu124 (CPU execution), Transformers 4.57.6, PEFT 0.18.1, Accelerate 1.12.0, bitsandbytes 0.49.2. Recommended stack: Transformers 5.8.1, PEFT 0.19.1, Accelerate 1.13.0; the same Python/torch/bitsandbytes. CPU full/LoRA each made two actual updates, changed 26/8 saved parameter tensors, saved and reloaded for generation.

Candidate CUDA qualification used one native Windows RTX 4080, total 15.992 GiB, driver 596.49, CUDA 12.4, BF16 and SDPA. LoRA and NF4 double-quant QLoRA each completed two updates, changed eight saved adapter tensors, saved and reloaded. QLoRA used paged AdamW8bit, rank 128 to exercise real uint8 optimizer state, 14 quantized linear layers and verified BF16 compute after updates. Allocated/reserved peaks were 0.017112/0.021484 GiB for tiny LoRA and 0.017678/0.021484 GiB for tiny QLoRA. These cover load through final update. Free budgets used the lower CUDA-driver/nvidia-smi view, approximately 14 GiB. Waiting under earlier unrelated GPU pressure did not alter gates, stop another process or count an unrun attempt as a pass.

Prospective memory comparison

Four predeclared Qwen2.5-0.5B-Instruct observations at pinned revision 7ae557604adf67be50417f59c2c2f167def9a775 cover LoRA/QLoRA and sequence 256/512, three updates each, one GPU. New weights were 988,097,824 bytes, within the 1.5 GB budget. Their original source fingerprints and driver-view free-memory fields remain unchanged; they are distinct from final wheel receipts.

Published 0.3.0 and 0.4.0 predictions on the same observations are identical: MAPE 46.2%, median signed error +46.0%, range +29.3% to +63.7%, four overestimates and zero underestimates. No accuracy improvement is demonstrated. The process planning proxy excludes the separate safety allowance and is compared with reserved allocator peaks; device free memory is a different measurement scope. False-feasible is 0/4 for the recorded short workload/budget. False-infeasible is undefined because there is no predicted-no sample. These are descriptive counts, not calibrated OOM probabilities. No new cohort observations were fitted.

See raw records and paired analysis and full validation report. Historical raw measurements are preserved. A resource-pressure repeat was interrupted and excluded; it is neither a successful observation nor an OOM with a fabricated peak. Tiny/random and short real-model runs demonstrate execution, not useful fine-tuned language quality.

Public release and verification

GitHub v0.4.0 was published at 2026-10-04 00:41:55 UTC. The release workflow completed successfully: candidate verification, Trusted Publishing and public installation all passed. Official PyPI 0.4.0 exposes exactly the wheel and sdist above. Both files were downloaded from files.pythonhosted.org and hashed again; both match the qualified release assets. PyPI displays publishing attestations for both files, bound to the release commit and the existing .github/workflows/release.yml publisher. The no-environment OIDC binding is preserved; only the upload job has id-token: write.

Post-release commands used fresh isolated Python 3.12 environments:

uv pip install --python PUBLIC_CORE_PYTHON --index-url https://pypi.org/simple --no-cache canifinetune==0.4.0
uv pip install --python PUBLIC_TRAIN_PYTHON --index-url https://pypi.org/simple --no-cache -c constraints/train-recommended.txt "canifinetune[train]==0.4.0"
python scripts/verify_public.py --version 0.4.0 --candidate .cache/release-0.4.0
PUBLIC_CORE_PYTHON /absolute/path/scripts/qualify_candidate.py --receipt public-core-qualification.json
PUBLIC_TRAIN_PYTHON /absolute/path/scripts/qualify_candidate.py --training --cuda --receipt public-training-qualification.json

PUBLIC_*_PYTHON denotes each environment's interpreter, not a shell-independent literal command. Torch 2.6.0+cu124 and recommended train dependencies were installed from their official indexes before the train-extra command. Qualification commands ran with the working directory outside the checkout. Imports resolve to installed site-packages, direct_url.json is absent for the index installs, author/version metadata matches, and both public packages have the candidate source fingerprint. The core environment has no torch. Public CPU full/LoRA and CUDA LoRA/QLoRA each completed two updates, changed saved tensors and reloaded for generation; QLoRA again allocated actual uint8 optimizer state. The README CLI estimate returned 8.420 GiB/YES/medium. All ten versioned README documentation/license/constraint links returned HTTP 200, and the live PyPI badge displayed 0.4.0.

A fresh public-package pinned 0.5B QLoRA recipe at sequence 256 also completed two updates, changed 96 saved adapter tensors and reloaded for eight-token generation. It used rank 16, paged AdamW8bit with float32/uint8 state, BF16 compute, 168 quantized linear layers, four instruction rows and no dropped or truncated rows. Allocated/reserved peaks were 1.419052/1.521484 GiB; cached weights were reused, with no additional weight download. This is a short quickstart execution check, not a new accuracy-cohort observation or quality test.

The public qualification receipt contains the official distribution hashes, installation checks and redacted run records. The local demo was restarted from the public core installation; its visible 8.42 GiB result and generated commands were checked again. No external release blocker remains. Documentation completion commits do not move the tag or replace the published package files.

Product entry, compatibility and attribution

The PyPI-only quickstart separates native Windows and Linux/WSL environment commands. canifinetune demo serves the same estimator at http://127.0.0.1:8765 using stdlib and packaged resources. Desktop and 375px mobile layouts, free-budget inputs, errors, stale-result clearing and copied commands were checked. Repository Pages is unconfigured; this is a shipped local application, with no claim of a hosted service. Community measurement export is manual, explicit opt-in, redacted by allowlist and unreviewed by default; it does not automatically enter calibration or authoritative accuracy statistics.

Core supports Python 3.10–3.14; training qualification is Python 3.12 with the minimum/recommended constraints above. Regenerate recipes after upgrading, remove unknown YAML keys, and deliberately opt into right truncation if needed. CPU requires fp32; unsupported BF16/full-FP16 requests fail without silent fallback. Use raw unquantized bases and request quantization explicitly. Old calibration caches require regeneration. Saved inference artifacts are not full training resume. WSL GPU, other GPUs, Flash Attention, Liger, int8/fp4 accuracy, long training and language quality remain unqualified; no new cross-hardware claim is made.

All new Git commit authors and package author metadata are Daoyuan Li. No agent/bot authors, coauthors or hand-maintained contributors were added. GitHub's merge committer, workflow actors and package attestations remain platform provenance. Legitimate historical third-party attribution and licenses were preserved.