FLUX.2-dev-MXFP4-RTN-AutoRound

Model Details

This is a MXFP4 (4-bit micro-scaling) quantization of black-forest-labs/FLUX.2-dev, a flagship text-to-image diffusion model. Generated by AutoRound with RTN (round-to-nearest, iters=0).

• Base model: black-forest-labs/FLUX.2-dev
• Quantization: MXFP4 (W4A4), group_size=32
• Method: AutoRound RTN
• Model size: ~62 GB (vs ~110 GB bf16)

Quantization Details

• Scheme: MXFP4 (data_type=mx_fp, bits=4, act_bits=4)
• Group size: 32
• Export format: auto_round (vllm-omni compatible)
• Calibration: coco2014, 28 steps, guidance 3.5

Evaluation

Evaluated with vllm-omni diffusion harness (28 steps, guidance 3.5, 1024×1024, seed 42).

Benchmark	BF16 Baseline	MXFP4 Quantized
DrawBench CLIP	32.48	32.44
DrawBench CLIP-IQA	71.35	71.01
DrawBench ImageReward	1.15	1.11
GenEval	0.844	0.835

MXFP4 quantization is nearly lossless vs the BF16 baseline (GenEval 0.835 vs 0.844, CLIP 32.44 vs 32.48).

Usage

from vllm_omni.entrypoints.omni import Omni
from vllm_omni.inputs.data import OmniDiffusionSamplingParams

omni = Omni(model="INCModel3/FLUX.2-dev-MXFP4-RTN-AutoRound", mode="text-to-image")
params = OmniDiffusionSamplingParams(
    height=1024, width=1024, seed=42,
    guidance_scale=3.5, num_inference_steps=28, num_outputs_per_prompt=1,
)
out = omni.generate("a red bench in a park", sampling_params_list=[params])

License

Please follow the license of the original model black-forest-labs/FLUX.2-dev.

Produced with autoquant-agent — agent-driven quantize + evaluate + self-heal.