|
|
from huggingface_hub.constants import HF_HUB_CACHE |
|
|
from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel |
|
|
import torch |
|
|
import torch._dynamo |
|
|
import gc |
|
|
from PIL import Image as img |
|
|
from PIL.Image import Image |
|
|
from pipelines.models import TextToImageRequest |
|
|
from torch import Generator |
|
|
from diffusers import FluxTransformer2DModel, DiffusionPipeline |
|
|
import os |
|
|
|
|
|
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True" |
|
|
|
|
|
Pipeline = None |
|
|
|
|
|
ckpt_id = "black-forest-labs/FLUX.1-schnell" |
|
|
ckpt_revision = "741f7c3ce8b383c54771c7003378a50191e9efe9" |
|
|
|
|
|
|
|
|
def empty_cache(): |
|
|
gc.collect() |
|
|
torch.cuda.empty_cache() |
|
|
torch.cuda.reset_max_memory_allocated() |
|
|
torch.cuda.reset_peak_memory_stats() |
|
|
|
|
|
|
|
|
def load_pipeline() -> Pipeline: |
|
|
empty_cache() |
|
|
|
|
|
dtype, device = torch.bfloat16, "cuda" |
|
|
text_encoder_2 = T5EncoderModel.from_pretrained( |
|
|
"escort321/FLUX.1-schnell1-up", |
|
|
revision="d7e70e3a8fbc36ec3c47e78913e9c7142bc87b7b", |
|
|
subfolder="text_encoder_2", |
|
|
torch_dtype=torch.bfloat16, |
|
|
) |
|
|
|
|
|
path = os.path.join( |
|
|
HF_HUB_CACHE, |
|
|
"models--escort321--FLUX.1-schnell1-up/snapshots/d7e70e3a8fbc36ec3c47e78913e9c7142bc87b7b/transformer", |
|
|
) |
|
|
transformer = FluxTransformer2DModel.from_pretrained( |
|
|
path, torch_dtype=torch.bfloat16, use_safetensors=False |
|
|
) |
|
|
pipeline = DiffusionPipeline.from_pretrained( |
|
|
ckpt_id, |
|
|
revision=ckpt_revision, |
|
|
transformer=transformer, |
|
|
text_encoder_2=text_encoder_2, |
|
|
torch_dtype=dtype, |
|
|
).to(device) |
|
|
|
|
|
pipeline( |
|
|
prompt="wordcraft, radiance, ethereal, cartilaginous, tuner, fruity, dullard, existence", |
|
|
width=1024, |
|
|
height=1024, |
|
|
guidance_scale=0.0, |
|
|
num_inference_steps=4, |
|
|
max_sequence_length=256, |
|
|
) |
|
|
|
|
|
empty_cache() |
|
|
return pipeline |
|
|
|
|
|
|
|
|
@torch.no_grad() |
|
|
def infer( |
|
|
request: TextToImageRequest, pipeline: Pipeline, generator: Generator |
|
|
) -> Image: |
|
|
return pipeline( |
|
|
request.prompt, |
|
|
generator=generator, |
|
|
guidance_scale=0.0, |
|
|
num_inference_steps=4, |
|
|
max_sequence_length=256, |
|
|
height=request.height, |
|
|
width=request.width, |
|
|
output_type="pil", |
|
|
).images[0] |
|
|
|
