Bring your project.
Use 32 GB of VRAM.

What's installed

PyTorch 2.9.1 with CUDA 12.8, torchvision, torchaudio, NumPy, Pillow, SciPy, pandas, scikit-learn, matplotlib, seaborn, OpenCV (headless), Transformers, Datasets, Accelerate, huggingface_hub, tokenizers, safetensors, sentencepiece, timm, einops, PEFT (LoRA), bitsandbytes (4/8-bit) and Diffusers.

Good fits

Fine-tunes (LoRA or full), training runs, batch inference and evaluation: anything that runs as a script, saves its results and ends. You're responsible for your program and its outputs.

Prepare your upload

• Upload a Python file or an archive with your code, data and model weights.
• Choose the script path inside that upload, for example project/train.py.
• Jobs can't download packages or weights from the internet, so include anything that isn't installed above.
• Write returned files to the output folder. Temporary files are discarded.

Inspect what happened

The job page can show your program log and supported output previews. Download produced files before their expiry. An execution error or a runtime limit can end the job before your project is complete.

A small preparation example

Illustrative code.

from pathlib import Path
import pandas as pd

# Read data beside your uploaded script.
project = Path(__file__).resolve().parent
data = pd.read_csv(project / "data.csv")

# The current folder is the returned output folder.
data.describe().to_csv("summary.csv")

Start with a small input, inspect the log and results, then decide whether to run a larger project.