File size: 2,572 Bytes
f366336 |
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 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 |
from diffusers import AutoencoderKL, FluxPipeline, FluxTransformer2DModel
from huggingface_hub.constants import HF_HUB_CACHE
from transformers import T5EncoderModel
from PIL import Image
from pipelines.models import TextToImageRequest
from torch import Generator
from typing import Type
from torchao.quantization import quantize_, int8_weight_only, fpx_weight_only
import torch
import torch._dynamo
import os
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
os.environ["TOKENIZERS_PARALLELISM"] = "True"
torch._dynamo.config.suppress_errors = True
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.enabled = True
Pipeline = None
def load_pipeline() -> Pipeline:
ckpt_id = "intensity809/FLUX-encoder-VAE"
ckpt_revision = "4dc7cf59f64cedde94422acff2ea8ef2b91a37d4"
vae = AutoencoderKL.from_pretrained(
ckpt_id,
revision=ckpt_revision,
subfolder="vae",
local_files_only=True,
torch_dtype=torch.bfloat16,
)
quantize_(vae, int8_weight_only())
text_encoder_2 = T5EncoderModel.from_pretrained(
"intensity809/FLUX1-transformer",
revision="390b27551542e3b39c89ea96267c3a11f2e00980",
subfolder="text_encoder_2",
torch_dtype=torch.bfloat16,
)
path = os.path.join(
HF_HUB_CACHE,
"models--intensity809--FLUX1-transformer/snapshots/390b27551542e3b39c89ea96267c3a11f2e00980/transformer",
)
transformer = FluxTransformer2DModel.from_pretrained(
path, torch_dtype=torch.bfloat16, use_safetensors=False
)
pipeline = FluxPipeline.from_pretrained(
ckpt_id,
revision=ckpt_revision,
transformer=transformer,
text_encoder_2=text_encoder_2,
torch_dtype=torch.bfloat16,
)
pipeline.to("cuda")
pipeline.to(memory_format=torch.channels_last)
for _ in range(1):
pipeline(
prompt="insensible, timbale, pothery, electrovital, actinogram, taxis, intracerebellar, centrodesmus",
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, 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]
|