Python Script to run it below 16GB vram
Unfortunately i dont recall from which github gist i got most of the code, but with this i am able to run FLUX.1-dev in q8 on less than 16GB vram.
Might help some of you.
On my 22GB RTX 2080Ti it takes approx. 7,5min per image.
I have 2 cards, i am still looking how i can split it somehow on both of the cards (seems that diffusion image models are far far behind llm's in quantization AND in parallelisation of vram [multi gpu for single image output sharing])
import torch
from optimum.quanto import freeze, qfloat8, quantize
from diffusers import FlowMatchEulerDiscreteScheduler, AutoencoderKL
from diffusers.models.transformers.transformer_flux import FluxTransformer2DModel
from diffusers.pipelines.flux.pipeline_flux import FluxPipeline
from transformers import CLIPTextModel, CLIPTokenizer,T5EncoderModel, T5TokenizerFast
dtype = torch.bfloat16
bfl_repo = "black-forest-labs/FLUX.1-dev"
revision = "refs/heads/main"
scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(bfl_repo, subfolder="scheduler", revision=revision)
text_encoder = CLIPTextModel.from_pretrained("openai/clip-vit-large-patch14", torch_dtype=dtype)
tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14", torch_dtype=dtype)
text_encoder_2 = T5EncoderModel.from_pretrained(bfl_repo, subfolder="text_encoder_2", torch_dtype=dtype, revision=revision)
tokenizer_2 = T5TokenizerFast.from_pretrained(bfl_repo, subfolder="tokenizer_2", torch_dtype=dtype, revision=revision)
vae = AutoencoderKL.from_pretrained(bfl_repo, subfolder="vae", torch_dtype=dtype, revision=revision)
transformer = FluxTransformer2DModel.from_pretrained(bfl_repo, subfolder="transformer", torch_dtype=dtype, revision=revision)
quantize(transformer, weights=qfloat8)
freeze(transformer)
quantize(text_encoder_2, weights=qfloat8)
freeze(text_encoder_2)
pipe = FluxPipeline(
scheduler=scheduler,
text_encoder=text_encoder,
tokenizer=tokenizer,
text_encoder_2=None,
tokenizer_2=tokenizer_2,
vae=vae,
transformer=None,
)
pipe.text_encoder_2 = text_encoder_2
pipe.transformer = transformer
pipe.enable_model_cpu_offload()
generator = torch.Generator().manual_seed(42)
image = pipe(
prompt='Photorealistic nekomusume cat girl selfie, university graduation gown, high heels, holding "Richard" card, cat ears, cinematic lighting, shallow depth of field, campus back'
width=1024,
height=1024,
num_inference_steps=50,
generator=generator,
guidance_scale=3.5,
max_sequence_length=512,
).images[0]
image.save('test_flux_distilled.png')
What is less than 16GB VRAM? Would this work on 8GB?
Hold my beer: 3070 mobile 8 GB VRAM: https://github.com/InServiceOfX/InServiceOfX/blob/master/PythonLibraries/HuggingFace/MoreDiffusers/morediffusers/Applications/terminal_only_finite_loop_flux.py Note that you use a .yml file to configure the input values dynamically and will prompt you for the prompt.
An image couple walking on the road
Hold my beer: 3070 mobile 8 GB VRAM: https://github.com/InServiceOfX/InServiceOfX/blob/master/PythonLibraries/HuggingFace/MoreDiffusers/morediffusers/Applications/terminal_only_finite_loop_flux.py Note that you use a .yml file to configure the input values dynamically and will prompt you for the prompt.
how do i get this running? would i clone the entire repo?
-Thanks!
I found that as with SD3, it is possible to get FLUX.1-schnell and FLUX.1-dev running on 12GB Nvidia RTX3060 just by 8bit quantizing the big T5 text encoder in conjunction with using accelerate's device_map. For FLUX it is text_encoder_2. Note that I have accelerate and bitsandbytes installed in my Python environment (pip install)
import torch
from diffusers import DiffusionPipeline
from transformers import T5EncoderModel, BitsAndBytesConfig
#https://huggingface.co/black-forest-labs/FLUX.1-dev
#https://huggingface.co/docs/diffusers/en/api/pipelines/flux
#https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/flux/pipeline_flux.py
quantization_config = BitsAndBytesConfig(load_in_8bit=True)
model_id = "black-forest-labs/FLUX.1-schnell" #needs 4 steps only - it is faster than the dev version as the name implies
#model_id = "black-forest-labs/FLUX.1-dev" #needs a lot more steps - eg 50
text_encoder = T5EncoderModel.from_pretrained(
model_id,
subfolder="text_encoder_2",
quantization_config=quantization_config,
torch_dtype=torch.bfloat16 #bfloat16 and normal float16 both work - former gives a warning but seems to work
)
pipe = DiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16, #bfloat16 and float16 both work, must match the T5
text_encoder_2=text_encoder,
device_map="balanced", max_memory={0:"11GiB", "cpu":"20GiB"})
pipe.vae.enable_tiling() #less memory usage at VAE time
prompt = "A cat holding a sign that says hello world"
prompt2 = "A photorealistic image of a relaxed looking tabby cat holding up a brown cardboard sign that says 'hello world' written in blue felt marker pen, ensure that the paws look correct"
image = pipe(
prompt,
prompt_2=prompt2,
num_images_per_prompt=1,
guidance_scale=0.0, #must be 0.0 for schnell version, dev version can be as per SD
num_inference_steps=4, #only need 4 for schnell version, dev version needs 50 or so
max_sequence_length=256, #relates to the T5 encoder - text_encoder_2 - max 256 for schnell
generator=torch.Generator("cpu").manual_seed(0)
).images[0]
image.save("flux-schnell.png")
You may need to tweak the to suit your hardware. Note that adding a second GPU involves adding additional entries to the dict, eg: 1:"24GiB" or whatever.
This code works for me for both schnell and dev versions on a 12GB Nvidia card. I don't know if it will support smaller cards.
