This model uses per layer fine grained quantization.
What dtype to use for a layer is selected dynamically by trial and error until the std normalized mse loss is lower than the selected threshold.

Minimum allowed dtype is set to uint4 and std normalized mse loss threshold is set to 1e-2.
This created a mixed precision model with uint4 and int5 dtypes.
SVD quantization is disabled.

Usage:

pip install sdnq

import torch
import diffusers
from sdnq import SDNQConfig # import sdnq to register it into diffusers and transformers
from sdnq.common import use_torch_compile as triton_is_available
from sdnq.loader import apply_sdnq_options_to_model

pipe = diffusers.Flux2KleinPipeline.from_pretrained("Disty0/FLUX.2-klein-4B-SDNQ-4bit-dynamic", torch_dtype=torch.bfloat16)

# Enable INT8 MatMul for AMD, Intel ARC and Nvidia GPUs:
if triton_is_available and (torch.cuda.is_available() or torch.xpu.is_available()):
    pipe.transformer = apply_sdnq_options_to_model(pipe.transformer, use_quantized_matmul=True)
    pipe.text_encoder = apply_sdnq_options_to_model(pipe.text_encoder, use_quantized_matmul=True)
    # pipe.transformer = torch.compile(pipe.transformer) # optional for faster speeds

pipe.enable_model_cpu_offload()

prompt = "A cat holding a sign that says hello world"
image = pipe(
    prompt=prompt,
    height=1024,
    width=1024,
    guidance_scale=1.0,
    num_inference_steps=4,
    generator=torch.manual_seed(0)
).images[0]

image.save("flux-klein-sdnq-4bit-dynamic.png")

Original BF16 vs SDNQ quantization comparison:

Quantization	Model Size	Visualization
Original BF16	7.8 GB	Original BF16
SDNQ 4 Bit	2.5 GB	SDNQ UINT4

SamuelTallet/FLUX.2-klein-4B-SDNQ-4bit-dynamic

SamuelTallet/
FLUX.2-klein-4B-SDNQ-4bit-dynamic