---
license: other
license_name: sapiens2-license
license_link: https://github.com/facebookresearch/sapiens2/blob/main/LICENSE.md
base_model: facebook/sapiens2-normal-0.8b
base_model_relation: quantized
library_name: sapiens
tags:
- sapiens
- sapiens2
- int4
- quantized
- safetensors
---

# sapiens2-normal-0.8b INT4-G128

Packed 4-bit derivative of [facebook/sapiens2-normal-0.8b](https://huggingface.co/facebook/sapiens2-normal-0.8b).

This artifact uses symmetric per-group INT4 packing with group size 128 for large floating-point weight tensors. Norms, biases, positional/rope tensors, and small tensors are kept in their source dtype. It is a storage/runtime-loader quant for the current official Sapiens2 code path, not an AWQ/GGUF/NVFP4 LLM artifact.

## Files

- `facebook__sapiens2-normal-0.8b-int4-g128.safetensors`: packed INT4 safetensors artifact.
- `load_sapiens2_int4.py`: loader that reconstructs a PyTorch state dict for the official Sapiens2 model code.
- `config.json` and `preprocessor_config.json`: copied from the source repo.
- `quantization_report.json`: build report.

## Quantization Report

- Source revision: `0b05360b013e0c0a6db020307a4476e1ea42790c`
- Group size: `128`
- Source bytes: `3538309852`
- Artifact bytes: `458381924`
- Compression ratio: `7.7191x`
- Tensors: `568`
- Quantized tensors: `202`
- Max tensor MAE during dequant smoke: `0.00762605`


## Fidelity Validation

The packed INT4 artifact was dequantized back to floating-point tensors and compared against the source checkpoint.

- Validation gate: global floating-tensor similarity >= `90.00%`
- Result: `PASS`
- Global floating-tensor similarity: `99.402250%`
- Minimum large-tensor cosine: `0.974280477`


## Loading

```python
from load_sapiens2_int4 import load_state_dict

state_dict = load_state_dict("facebook__sapiens2-normal-0.8b-int4-g128.safetensors", device="cpu")
# Then instantiate the matching official Sapiens2 architecture and load:
# model.load_state_dict(state_dict, strict=True)
```

## Limitations

This is a verified packed-weight artifact with a dequantizing loader. It does not claim native INT4 CUDA kernels for Sapiens2 yet. Runtime speedups require a Sapiens2-specific kernel/export path and should be benchmarked separately.