docs/ZEROGPU.md

# Mythos-Coder on Hugging Face ZeroGPU

ZeroGPU only works on **Hugging Face Spaces** with the **Gradio** SDK. It does not speed up local `python scripts/train_lora_sft.py` on your machine.

## What was added

```
spaces/mythos-coder-zerogpu/
  app.py                      # Gradio UI with @spaces.GPU train + test
  mythos_lora_core.py         # shared train/test logic
  requirements.txt
  README.md
  data/train/mythos_sft_messages.jsonl
  data/eval/user_style_prompts.jsonl
```

## Deploy the Space

1. Create a new Space at https://huggingface.co/new-space
   - SDK: **Gradio**
   - Hardware: **ZeroGPU**
2. Upload the contents of `spaces/mythos-coder-zerogpu/` to the Space repo, **or** push with git:

```bash
cd spaces/mythos-coder-zerogpu
git init
git remote add origin https://huggingface.co/spaces/YOUR_USERNAME/mythos-coder-zerogpu
git add .
git commit -m "Add Mythos-Coder ZeroGPU LoRA pipeline"
git push
```

3. Open the Space and wait for the build to finish.

## Run on ZeroGPU

1. Open the **Train LoRA** tab → click **Train on ZeroGPU**
2. Open the **Test LoRA** tab → click **Run all eval prompts on ZeroGPU**
3. Or paste a messy prompt and click **Generate on ZeroGPU**

## Refresh training data in the Space

After adding more converted rows locally:

```bash
python scripts/build_sft_messages.py
copy data\train\mythos_sft_messages.jsonl spaces\mythos-coder-zerogpu\data\train\
```

Then recommit and push the Space.

## Local scripts

Local CPU/CUDA scripts still work if you want them:

```bash
python scripts/build_sft_messages.py
python scripts/train_lora_sft.py
python scripts/test_lora_model.py
```

Use ZeroGPU when local training or inference is too slow.