File size: 1,660 Bytes
250eae8 |
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 |
from diffusers import AutoencoderKL
from diffusers.image_processor import VaeImageProcessor
import torch
import torch._dynamo
import gc
from PIL import Image
from pipelines.models import TextToImageRequest
from torch import Generator
from diffusers import DiffusionPipeline
from torchao.quantization import quantize_, int8_weight_only
Pipeline = None
MODEL_ID = "black-forest-labs/FLUX.1-schnell"
def clear():
gc.collect()
torch.cuda.empty_cache()
torch.cuda.reset_max_memory_allocated()
torch.cuda.reset_peak_memory_stats()
def load_pipeline() -> Pipeline:
clear()
dtype, device = torch.bfloat16, "cuda"
vae = AutoencoderKL.from_pretrained(
MODEL_ID, subfolder="vae", torch_dtype=torch.bfloat16
)
quantize_(vae, int8_weight_only(), device="cuda")
pipeline = DiffusionPipeline.from_pretrained(
MODEL_ID,
vae=vae,
torch_dtype=dtype,
)
pipeline.enable_sequential_cpu_offload()
pipeline(prompt="unpervaded, unencumber, froggish, groundneedle, transnatural, fatherhood, outjump, cinerator", width=1024, height=1024, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256)
clear()
return pipeline
@torch.inference_mode()
def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
clear()
if request.seed is None:
generator = None
else:
generator = Generator(device="cuda").manual_seed(request.seed)
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 |