Views
No views yet
pip install bitsandbytes1from diffusers import FluxKontextPipeline
2import torch
3pipeline = FluxKontextPipeline.from_pretrained("eramth/flux-kentext-4bit-fp4",torch_dtype=torch.float16).to("cuda")
4# This allows you to generate higher resolution images without much extra VRAM usage.
5pipeline.vae.enable_tiling()
61from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig
2from transformers import BitsAndBytesConfig as TransformersBitsAndBytesConfig
3from diffusers import FluxKontextPipeline,FluxTransformer2DModel
4from transformers import T5EncoderModel
5import torch
6
7token = ""
8repo_id = ""
9
10quant_config = TransformersBitsAndBytesConfig(load_in_4bit=True,bnb_4bit_compute_dtype=torch.float16,bnb_4bit_quant_type="fp4")
11
12text_encoder_2_4bit = T5EncoderModel.from_pretrained(
13 "black-forest-labs/FLUX.1-Kontext-dev",
14 subfolder="text_encoder_2",
15 quantization_config=quant_config,
16 torch_dtype=torch.float16,
17 token=token
18)
19
20quant_config = DiffusersBitsAndBytesConfig(load_in_4bit=True,bnb_4bit_compute_dtype=torch.float16,bnb_4bit_quant_type="fp4")
21
22transformer_4bit = FluxTransformer2DModel.from_pretrained(
23 "black-forest-labs/FLUX.1-Kontext-dev",
24 subfolder="transformer",
25 quantization_config=quant_config,
26 torch_dtype=torch.float16,
27 token=token
28)
29
30pipe = FluxKontextPipeline.from_pretrained(
31 "black-forest-labs/FLUX.1-Kontext-dev",
32 transformer=transformer_4bit,
33 text_encoder_2=text_encoder_2_4bit,
34 torch_dtype=torch.float16,
35 token=token
36)
37
38pipe.push_to_hub(repo_id,token=token)