# Fine-tune models on GPUs

Fine-tune LLMs and diffusion models on serverless H100s without touching infrastructure. Decorate your training function, run the file, and Beam handles the GPUs, storage, and scaling.

## Fine-tuning without wrangling GPUs

Beam runs the compute, just bring your code. LoRA, QLoRA, and more with zero infrastructure to manage.

### GPUs on-demand
Grab a B300 the moment you kick off a run and release it the second training ends.

### Runs like local Python
Decorate your training function and call .remote(). Beam syncs your working directory and streams logs back to your shell.

### Volumes for weights & datasets
Persistent storage volumes hold model weights, datasets, and checkpoints between runs. Storage is included, free of charge.

### Secrets for gated models
Store your Hugging Face token once and pull gated models like Llama and Gemma directly inside the container.

### No timeouts
Run tasks as long as you need to.

### Train and deploy, all in one platform
Serve your fine-tuned weights on a managed inference endpoint.

## Launch a training job

01
### Upload weights and data to a volume
Create a volume and download your base model and dataset.

02
### Run training with one decorator
Wrap your training loop in @function, declare the GPU inline, and run the file. Beam streams logs to your terminal.

03
### Serve the result
Point an endpoint at the fine-tuned weights in the same volume and deploy it as an autoscaling API.

```bash
# 1. Create a volume and stage weights + data
$ beam volume create gemma-ft
$ python upload.py   # snapshot_download() into the volume

# 2. Kick off training on an H100
$ python finetune.py
=> Running function: <finetune:fine_tune>
{'loss': 1.617, 'epoch': 0.01}
...
=> Function complete

# 3. Inspect the artifacts
$ beam ls gemma-ft/gemma-2b-finetuned
adapter_config.json
adapter_model.safetensors
```

## Frequently asked questions

### Which fine-tuning methods work?
Anything that runs in PyTorch: LoRA and QLoRA with PEFT, Unsloth, TRL, or your own training loop. Beam provides the managed GPU runtime — the framework is up to you.

### Do I pay when I'm not training?
No. For serverless, billing is per-second and stops when the run ends. Storage for weights and checkpoints is included, free of charge.

### Where do my checkpoints go?
To a persistent storage volume mounted into the container. Inspect files with beam ls, download them, or serve them directly from an inference endpoint.

### Can I fine-tune gated models like Llama or Gemma?
Yes. Save your Hugging Face token with beam secret create HF_TOKEN, and the container can pull gated repos securely.

### What GPUs can I train on?
T4s and A10Gs up through B200s and B300s. Pick the GPU in the decorator and change it any time with one line.