This is a nano4M multimodal foundation model trained as part of the COM-304 course at EPFL.

## Model Description

nano4M is a small-scale reimplementation of 4M (Massively Multimodal Masked Modeling), a unified multimodal foundation model trained to predict any modality from any other.

- **Architecture**: Encoder-decoder transformer, **6 encoder + 6 decoder layers**, width **512**.
- **Modalities**:
  - `tok_rgb@256` — RGB image tokens (Cosmos tokenizer)
  - `tok_depth@256` — depth tokens
  - `tok_normal@256` — surface-normal tokens
  - `tok_rgb_flextok@256` — RGB tokens from the FlexTok tokenizer
  - `scene_desc` — scene description text tokens
- **Training data**: MultiCLEVR
- **Training objective**: `span`

## Variants

This model is one of three variants trained in this project:

| Variant | Objective                                        |
|---------|--------------------------------------------------|
| naive   | Standard random-token masked prediction          |
| vlpa    | Variable-length partial-autoregressive masking   |
| span    | Span-based masking                               |

## Usage

```python
from nanofm.models import load_model_from_safetensors  # adjust import to your repo
from huggingface_hub import hf_hub_download
import torch

device = "cuda" if torch.cuda.is_available() else "cpu"

ckpt_path = hf_hub_download(
    repo_id="cdparker/4M_with_span_flextok",
    filename="checkpoint-final.safetensors",
)
model = load_model_from_safetensors(ckpt_path, device=device)
```

## Reference

- Mizrahi et al., *4M: Massively Multimodal Masked Modeling*, NeurIPS 2023.
- Original 4M codebase: https://github.com/apple/ml-4m