File size: 2,114 Bytes
402cc8d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 | 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("jade012/FLUX.1-schnell1", revision = "124d80794ed0cf70cac48ecd204dbe04b00dff94", subfolder="text_encoder_2",torch_dtype=torch.bfloat16)
path = os.path.join(HF_HUB_CACHE, "models--jade012--FLUX.1-schnell1/snapshots/124d80794ed0cf70cac48ecd204dbe04b00dff94/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)
#quantize_(pipeline.vae, int8_weight_only())
pipeline(prompt="logomancy, afterglow, aetheric, chondrichthyes, calibrator, pomeous, lackwit, presentness", 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:
image=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]
return(image)
|