green-peft
Constraint-aware PEFT strategy recommender. Wraps the surrogate models trained in Cell 13 of the GreenPEFT notebook so you can ask "which method + model size should I use?" from the command line, without running anything.
Install
pip install -e .
Export artifacts from your notebook run
The CLI needs two things out of your Kaggle/Colab peft_bench/ folder:
peft_bench/results/surrogate_models.joblib # written by Cell 13
peft_bench/configs/backbones.yaml # written by Cell 3 (has model_zoo:)
peft_bench/configs/methods/*.yaml # written by Cell 3
Zip results/ and configs/ together, download from Kaggle, unzip locally into
e.g. ./peft_bench_export/, and point --artifacts-dir at it.
This repository includes a ready-to-use export at:
../../model/artifacts_export/
├── configs/backbones.yaml
├── configs/methods/*.yaml
├── configs/tasks.yaml
├── results/surrogate_models.joblib
├── results/surrogate_cv_metrics.json
└── results/surrogate_dataset.csv
From the repository root, install the package and use that export directly:
cd green_peft_cli/green_peft_pkg
python -m pip install -e .
green-peft list-zoo --artifacts-dir ../../model/artifacts_export
Usage
# Basic recommendation under a VRAM + accuracy constraint
green-peft recommend --artifacts-dir ./peft_bench_export \
--vram 16 --accuracy 0.90 --profile balanced
# Strict carbon budget, custom weight profile, JSON output for scripting
green-peft recommend --artifacts-dir ./peft_bench_export \
--vram 24 --carbon 0.003 --weights 0.3,0.2,0.4,0.1 --top-k 5 --json
# See every backbone x method combination the engine can currently score
green-peft list-zoo --artifacts-dir ./peft_bench_export
What it actually does
- Builds a candidate table crossing every backbone in your
model_zoo(including ones you never ran) with every configured method. - Runs the four surrogate models (feasibility classifier + accuracy / peak-VRAM / energy / wall-clock regressors) over every candidate.
- Drops candidates the feasibility classifier doesn't trust, or that violate your
--vram/--carbon/--accuracy/--timeconstraints. - Computes the Pareto front and a GEI score (same formula as the paper: weighted sum of normalized accuracy/memory/carbon/time scores) over the surviving candidates.
- Prints the top-k ranked options with a plain-English explanation, or an explicit breakdown of why nothing survived if your constraints are infeasible together.
Predictions are only as reliable as the surrogate's leave-one-tier-out validation
(results/surrogate_cv_metrics.json from Cell 13) says they are for that target --
check regression.<target>.best_model R2 there before trusting a specific number.
Current validation and limitations
The included export reports leave-one-tier-out feasibility accuracy of 0.7833.
Selected regression models report the following validation results:
| Target | Model | MAE | R2 |
|---|---|---|---|
| Accuracy | Gradient boosting | 0.0182 | -0.4573 |
| Peak VRAM | Gradient boosting | 3.9703 GB | -0.5154 |
| Energy | Ridge | 0.000563 kWh | 0.3740 |
| Wall-clock time | Gradient boosting | 33.12 s | 0.4077 |
Use the CLI for experiment planning, shortlist generation, and explicit budget checks. It does not run fine-tuning, measure the target hardware, or guarantee performance. The benchmark evidence is primarily SST-2 classification on a Tesla T4; recommendations for unbenchmarked catalog entries are extrapolations. Validate the selected option with a real run and retain the JSON output alongside the measured results.
Metadata
Release files for green-peft 0.1.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| green_peft-0.1.0.tar.gz | 12.7 kB | Without provenance | Details for green_peft-0.1.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| green_peft-0.1.0-py3-none-any.whl | Python 3 | none | any | Details for green_peft-0.1.0-py3-none-any.whl |
Total release size: 25.0 kB
Release files / green_peft-0.1.0.tar.gz
| Download URL | green_peft-0.1.0.tar.gz |
|---|---|
| Size | 12.7 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
50f991cf7194aeec3402e31fdf8ddb63d711a06092e5f35e1069ec33b6aae094
|
|
BLAKE2b-256 checksum How to use checksums |
9dc7db1111bbb237b84c73ed3a02fbe1ac56caf2c5d902b08db7bdd090e3d42e
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.12.10
|
Release files / green_peft-0.1.0-py3-none-any.whl
| Download URL | green_peft-0.1.0-py3-none-any.whl |
|---|---|
| Size | 12.3 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
93d5e158e53fe231893120117b9719721a090cb3df2d16ead6ba1547a63ddbdb
|
|
BLAKE2b-256 checksum How to use checksums |
31c5c1d4e7dc9da269ef66beceafb0a9773d4a64bcbc16ba4a931659c3f69d96
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.12.10
|
Security for green-peft 0.1.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.