# Convert MHR TorchScript to CPU-safe file (Colab)

Run this in a GPU Colab (Runtime -> Change runtime type -> GPU). It loads the MHR torchscript saved in the original HF repo, moves it to CPU, and saves a CPU-safe file you can download.

1) Install dependencies
```bash
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118
pip install huggingface_hub
```

2) Run this Python snippet (paste into a notebook cell)
```python
import os
from huggingface_hub import snapshot_download
import torch
import pathlib

HF_TOKEN = os.environ.get('HF_TOKEN') or 'paste_your_token_here'
repo_id = 'facebook/sam-3d-body-vith'
local_dir = snapshot_download(repo_id=repo_id, use_auth_token=HF_TOKEN)
print('downloaded', local_dir)

mhr_path = pathlib.Path(local_dir) / 'assets' / 'mhr_model.pt'
print('mhr_path', mhr_path)

m = torch.jit.load(str(mhr_path), map_location='cuda' if torch.cuda.is_available() else 'cpu')
print('loaded')
try:
    m_cpu = m.to('cpu')
except Exception:
    m_cpu = m
out = '/content/mhr_model_cpu.pt'
torch.jit.save(m_cpu, out)
print('saved', out)
```

3) Download `mhr_model_cpu.pt` from the Colab files panel and copy it to your local machine. Then place it under the cached HF snapshot path or set the `LOCAL_MHR_PATH` env var to point at it when running the helper locally.

Local usage example (on your Mac):
```bash
export LOCAL_MHR_PATH=/path/to/mhr_model_cpu.pt
export LOCAL_CKPT_PATH=/path/to/model.ckpt  # optional, if you downloaded the checkpoint
python app/helper/model_loader.py
```

This will let your local helper load the MHR model without CUDA device pointers.