black-forest-labs/FLUX.1-dev quantized to INT8 using Optimum Quanto.
pip install diffusers optimum-quanto
import json
import torch
import diffusers
import transformers
from optimum.quanto import requantize
from safetensors.torch import load_file
from huggingface_hub import hf_hub_download
def load_quanto_transformer(repo_path):
with open(hf_hub_download(repo_path, "transformer/quantization_map.json"), "r") as f:
quantization_map = json.load(f)
with torch.device("meta"):
transformer = diffusers.FluxTransformer2DModel.from_config(hf_hub_download(repo_path, "transformer/config.json")).to(torch.bfloat16)
state_dict = load_file(hf_hub_download(repo_path, "transformer/diffusion_pytorch_model.safetensors"))
requantize(transformer, state_dict, quantization_map, device=torch.device("cuda"))
return transformer
def load_quanto_text_encoder_2(repo_path):
with open(hf_hub_download(repo_path, "text_encoder_2/quantization_map.json"), "r") as f:
quantization_map = json.load(f)
with open(hf_hub_download(repo_path, "text_encoder_2/config.json")) as f:
t5_config = transformers.T5Config(**json.load(f))
with torch.device("meta"):
text_encoder_2 = transformers.T5EncoderModel(t5_config).to(torch.bfloat16)
state_dict = load_file(hf_hub_download(repo_path, "text_encoder_2/model.safetensors"))
requantize(text_encoder_2, state_dict, quantization_map, device=torch.device("cuda"))
return text_encoder_2
pipe = diffusers.AutoPipelineForText2Image.from_pretrained("Disty0/FLUX.1-dev-qint8", transformer=None, text_encoder_2=None, torch_dtype=torch.bfloat16)
pipe.transformer = load_quanto_transformer("Disty0/FLUX.1-dev-qint8")
pipe.text_encoder_2 = load_quanto_text_encoder_2("Disty0/FLUX.1-dev-qint8")
pipe = pipe.to("cuda", dtype=torch.bfloat16)
prompt = "A cat holding a sign that says hello world"
image = pipe(
prompt,
height=1024,
width=1024,
guidance_scale=3.5,
num_inference_steps=50,
max_sequence_length=512,
generator=torch.Generator("cpu").manual_seed(0)
).images[0]
image.save("flux-dev.png")