Add model card
Browse files
README.md
ADDED
|
@@ -0,0 +1,64 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
base_model:
|
| 3 |
+
- black-forest-labs/FLUX.2-dev
|
| 4 |
+
pipeline_tag: text-to-image
|
| 5 |
+
license: other
|
| 6 |
+
tags:
|
| 7 |
+
- quantized
|
| 8 |
+
- mxfp4
|
| 9 |
+
- autoround
|
| 10 |
+
- diffusion
|
| 11 |
+
- text-to-image
|
| 12 |
+
- autoquant-agent
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
# FLUX.2-dev-MXFP4-RTN-AutoRound
|
| 16 |
+
|
| 17 |
+
## Model Details
|
| 18 |
+
|
| 19 |
+
This is a **MXFP4** (4-bit micro-scaling) quantization of [black-forest-labs/FLUX.2-dev](https://huggingface.co/black-forest-labs/FLUX.2-dev), a flagship text-to-image diffusion model. Generated by [AutoRound](https://github.com/intel/auto-round) with RTN (round-to-nearest, iters=0).
|
| 20 |
+
|
| 21 |
+
- **Base model:** [black-forest-labs/FLUX.2-dev](https://huggingface.co/black-forest-labs/FLUX.2-dev)
|
| 22 |
+
- **Quantization:** MXFP4 (W4A4), group_size=32
|
| 23 |
+
- **Method:** AutoRound RTN
|
| 24 |
+
- **Model size:** ~62 GB (vs ~110 GB bf16)
|
| 25 |
+
|
| 26 |
+
## Quantization Details
|
| 27 |
+
|
| 28 |
+
- **Scheme:** MXFP4 (data_type=mx_fp, bits=4, act_bits=4)
|
| 29 |
+
- **Group size:** 32
|
| 30 |
+
- **Export format:** auto_round (vllm-omni compatible)
|
| 31 |
+
- **Calibration:** coco2014, 28 steps, guidance 3.5
|
| 32 |
+
|
| 33 |
+
## Evaluation
|
| 34 |
+
|
| 35 |
+
Evaluated with vllm-omni diffusion harness (28 steps, guidance 3.5, 1024×1024, seed 42).
|
| 36 |
+
|
| 37 |
+
| Benchmark | BF16 Baseline | MXFP4 Quantized |
|
| 38 |
+
|---|---|---|
|
| 39 |
+
| DrawBench CLIP | 32.48 | 32.44 |
|
| 40 |
+
| DrawBench CLIP-IQA | 71.35 | 71.01 |
|
| 41 |
+
| DrawBench ImageReward | 1.15 | 1.11 |
|
| 42 |
+
| GenEval | 0.844 | 0.835 |
|
| 43 |
+
|
| 44 |
+
MXFP4 quantization is nearly lossless vs the BF16 baseline (GenEval 0.835 vs 0.844, CLIP 32.44 vs 32.48).
|
| 45 |
+
|
| 46 |
+
## Usage
|
| 47 |
+
|
| 48 |
+
```python
|
| 49 |
+
from vllm_omni.entrypoints.omni import Omni
|
| 50 |
+
from vllm_omni.inputs.data import OmniDiffusionSamplingParams
|
| 51 |
+
|
| 52 |
+
omni = Omni(model="INCModel3/FLUX.2-dev-MXFP4-RTN-AutoRound", mode="text-to-image")
|
| 53 |
+
params = OmniDiffusionSamplingParams(
|
| 54 |
+
height=1024, width=1024, seed=42,
|
| 55 |
+
guidance_scale=3.5, num_inference_steps=28, num_outputs_per_prompt=1,
|
| 56 |
+
)
|
| 57 |
+
out = omni.generate("a red bench in a park", sampling_params_list=[params])
|
| 58 |
+
```
|
| 59 |
+
|
| 60 |
+
## License
|
| 61 |
+
|
| 62 |
+
Please follow the license of the original model [black-forest-labs/FLUX.2-dev](https://huggingface.co/black-forest-labs/FLUX.2-dev).
|
| 63 |
+
|
| 64 |
+
_Produced with [autoquant-agent](https://github.com/) — agent-driven quantize + evaluate + self-heal._
|