ByteDance

/

Hyper-SD

like 1.34k

Follow

ByteDance 4.57k

import torch
from diffusers import FluxTransformer2DModel, FluxPipeline
from transformers import T5EncoderModel, CLIPTextModel
from optimum.quanto import freeze, qfloat8, quantize

bfl_repo = "black-forest-labs/FLUX.1-dev"
dtype = torch.bfloat16

transformer = FluxTransformer2DModel.from_single_file("https://huggingface.co/Kijai/flux-fp8/blob/main/flux1-dev-fp8.safetensors", torch_dtype=dtype)
quantize(transformer, weights=qfloat8)
freeze(transformer)

from diffusers import FluxPipeline
import torch

pipe = FluxPipeline.from_pretrained(bfl_repo, 
                                    transformer=transformer, 
                                    torch_dtype=dtype)
pipe.enable_model_cpu_offload()

repo_name = "ByteDance/Hyper-SD"
# Take 8-steps lora as an example
ckpt_8steps_name = "Hyper-FLUX.1-dev-8steps-lora.safetensors"
pipe.load_lora_weights(hf_hub_download(repo_name, ckpt_8steps_name), 
                       adapter_name="default")

Not work if i quantize the transformer first. I want to switch to another lora even after quantize the transformer. Any advice?