P-tuning v2 is a deep prompt tuning strategy designed as an alternative to full model fine-tuning.
Open source
24 alternativespeft
PEFT provides parameter-efficient fine-tuning methods for adapting pretrained machine learning models.
About peft
PEFT is an open-source Python library from Hugging Face for parameter-efficient fine-tuning, which adapts large pretrained models to new tasks. Instead of updating all of a model's parameters, PEFT methods train only a small number of added parameters. The project states that this reduces computational and storage costs.
A typical workflow installs the package with pip, wraps a base model and a PEFT configuration such as LoRA using get_peft_model, trains the model, saves it, and later loads it for inference with PeftModel. The project reports memory figures for one setup: an A100 80GB GPU, more than 64GB of CPU RAM, and the ought/raft/twitter_complaints dataset. The source does not say how these results apply to other hardware or tasks. It describes recent PEFT techniques as achieving performance comparable to full fine-tuning, not identical results.
Key features
- Trains a small number of added parameters instead of all model parameters
- LoRA configuration applied to a base model with get_peft_model
- Saving adapter weights with save_pretrained
- Loading adapters for inference with PeftModel.from_pretrained
- Reporting trainable parameter counts with print_trainable_parameters
- Integration with Transformers, Diffusers, and Accelerate
Best for: Machine learning practitioners and researchers who fine-tune large pretrained models and want to reduce compute and storage requirements.
Good for
- The project states that PEFT reduces computational and storage costs compared with full fine-tuning.
- The memory table lists lower GPU memory for LoRA than for full fine-tuning on the models shown, such as 14.4GB compared with 47.14GB for bigscience/T0_3B.
- It is integrated with Transformers, Diffusers, and Accelerate.
Watch out for
- The memory figures apply only to the hardware and dataset described in the source.
- The source describes PEFT performance as comparable to full fine-tuning, not identical to it.
- In the offloading example, LoRA with DeepSpeed CPU offloading uses 52GB of CPU memory for bigscience/mt0-xxl, so lower GPU memory can mean higher CPU memory use.
peft at a glance
- Website
- github.com/huggingface/peft
- Type
- Open-source project
- Pricing
- Not confirmed. Check the site
- Open source
- Yes · Source code
- Language
- English
- Last checked
- Oct 6, 2026 · Online when checked
- Listed
- Oct 5, 2026
peft alternatives
Tools and sites you could use instead, ranked by how closely they match.
oumi supports fine-tuning, evaluation, and deployment of open-weight language models.
Open source
24 alternatives
FEDOT automates machine learning model development through its modeling framework.
Open source
24 alternatives
Enables single machine or distributed training and evaluation of deep learning models.
Open source
24 alternatives
m2cgen converts machine learning models into native code for multiple programming languages without runtime dependencies.
Open source
24 alternatives
Hugging Face Accelerate launches and trains PyTorch models across devices and distributed configurations.
Open source
24 alternativesPopular in ML Training & Optimization
Well-known listings from the same topic.
Edwardedwardlib.orgAerosolveaerosolve.airbnb.techDataturedatature.ioTogether AItogether.aiOptunaoptuna.orgCatBoostcatboost.ai
peft FAQ
How do I install PEFT?
The source shows installation with pip install peft.
How is a PEFT model saved and loaded again?
The model is saved with save_pretrained, and a saved PEFT model is loaded for inference with PeftModel.from_pretrained applied to the base model.
Does PEFT train all of a model's parameters?
No. PEFT methods fine-tune only a small number of extra model parameters instead of all of the model's parameters.
Is something wrong or out of date?
Tell us and an editor will check it.
Own this website?
Claim this listing
Claim the listing and suggest updates yourself. An editor checks every claim first.