rmusr/sd-3.5-nf4

from diffusers import SD3Transformer2DModel, AutoencoderKL
from transformers import T5EncoderModel, CLIPTextModelWithProjection

model_id = "stabilityai/stable-diffusion-3.5-medium" # or "large" or "large-turbo" 
ckpt_nf4_id = "rmusr/sd-3.5-nf4"

text_encoder_3_nf4 = T5EncoderModel.from_pretrained(
    ckpt_nf4_id, subfolder="text_encoder_3", torch_dtype=torch.float16
)

transformer_nf4 = SD3Transformer2DModel.from_pretrained(
    ckpt_nf4_id, subfolder="transformer-medium", torch_dtype=torch.float16
)

# you can also load nf4 checkpoint of text_encoder, text_encoder_2, vae

pipeline = StableDiffusion3Pipeline.from_pretrained(
    model_id,
    text_encoder_3=text_encoder_3_nf4,
    transformer=transformer_nf4,
    torch_dtype=torch.float16,
).to("cuda")
pipeline.enable_attention_slicing()