File size: 2,579 Bytes
ae6f4f4 |
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 52 53 54 55 56 57 58 |
import torch
import torch._dynamo
import gc
import os
from huggingface_hub.constants import HF_HUB_CACHE
from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel
from diffusers import FluxPipeline, AutoencoderKL, AutoencoderTiny
from PIL.Image import Image
from pipelines.models import TextToImageRequest
from torch import Generator
from diffusers import FluxTransformer2DModel, DiffusionPipeline
from torchao.quantization import quantize_, int8_weight_only, fpx_weight_only
os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
os.environ["TOKENIZERS_PARALLELISM"] = "True"
torch._dynamo.config.suppress_errors = True
Pipeline = None
ids = "black-forest-labs/FLUX.1-schnell"
Revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"
def load_pipeline() -> Pipeline:
ttimagemodel = "edgetensor/max-schnell"
ttimagerevision = "13313e1f7d5b9d655554366d7394c42fb4e765d4"
vae = AutoencoderKL.from_pretrained(ids,revision=Revision, subfolder="vae", local_files_only=True, torch_dtype=torch.bfloat16,)
quantize_(vae, int8_weight_only())
text_encoder_2 = T5EncoderModel.from_pretrained("edgetensor/edgetensor-t5-v1_1-xxl-encoder-bf16", revision = "bfed5213335c2ead9c9b5aff657680db420a7c7d", torch_dtype=torch.bfloat16).to(memory_format=torch.channels_last)
path = os.path.join(HF_HUB_CACHE, "models--edgetensor--edgetensor-FLUX.1-schnell-int8wo/snapshots/6b4c594d4c510da13a40f8c3b483789eb82d36df")
transformer = FluxTransformer2DModel.from_pretrained(path, torch_dtype=torch.bfloat16, use_safetensors=False).to(memory_format=torch.channels_last)
pipeline = DiffusionPipeline.from_pretrained(ids, revision=Revision, transformer=transformer, text_encoder_2=text_encoder_2, torch_dtype=torch.bfloat16,)
pipeline.to("cuda")
for _ in range(2):
pipeline(prompt="satiety, unwitherable, Pygmy, ramlike, Curtis, fingerstone, rewhisper", width=1024, height=1024, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256)
return pipeline
@torch.no_grad()
def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
generator = Generator(pipeline.device).manual_seed(request.seed)
try:
prompt = request.prompt
except Exception as e:
prompt = "satiety, unwitherable, Pygmy, ramlike, Curtis, fingerstone, rewhisper"
return pipeline(
prompt,
generator=generator,
guidance_scale=0.0,
num_inference_steps=4,
max_sequence_length=256,
height=request.height,
width=request.width,
).images[0]
|