FLUX.2-klein-9B-MXFP8-RTN-AutoRound

Model Details

This is a MXFP8 (8-bit micro-scaling) quantization of black-forest-labs/FLUX.2-klein-9B, a lightweight distilled text-to-image model. Generated by AutoRound with RTN (round-to-nearest, iters=0).

• Base model: black-forest-labs/FLUX.2-klein-9B
• Quantization: MXFP8 (W8A8), group_size=32
• Method: AutoRound RTN
• Model size: ~10 GB (vs ~18 GB bf16)

Quantization Details

• Scheme: MXFP8 (data_type=mx_fp, bits=8, act_bits=8)
• Group size: 32
• Export format: auto_round (vllm-omni compatible)
• Calibration: coco2014, 4 steps, guidance 1.0

Evaluation

Evaluated with vllm-omni diffusion harness (4 steps, guidance 1.0, 1024×1024, seed 42).

Benchmark	BF16 Baseline	MXFP8 Quantized
DrawBench CLIP	32.81	32.93
DrawBench CLIP-IQA	69.64	69.63
DrawBench ImageReward	1.23	1.21
GenEval	0.856	0.853

MXFP8 quantization is essentially lossless vs the BF16 baseline (GenEval 0.853 vs 0.856, CLIP 32.93 vs 32.81).

Usage

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

omni = Omni(model="INCModel3/FLUX.2-klein-9B-MXFP8-RTN-AutoRound", mode="text-to-image")
params = OmniDiffusionSamplingParams(
    height=1024, width=1024, seed=42,
    guidance_scale=1.0, num_inference_steps=4, 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-klein-9B.

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