src/pipeline.py

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 FluxPipeline
from torchao.quantization import quant_api
from deps import f

Pipeline = None
MODEL_ID = "black-forest-labs/FLUX.1-schnell"
DTYPE = torch.bfloat16
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()
    pipeline = FluxPipeline.from_pretrained(MODEL_ID,
                                        torch_dtype=DTYPE)
    torch.backends.cudnn.benchmark = True
    torch.backends.cuda.matmul.allow_tf32 = True
    torch.cuda.set_per_process_memory_fraction(0.99)
    # quant_api.swap_conv2d_1x1_to_linear(pipeline.vae, f)
    pipeline.text_encoder.to(memory_format=torch.channels_last)
    pipeline.text_encoder_2.to(memory_format=torch.channels_last)
    pipeline.transformer.to(memory_format=torch.channels_last)
    pipeline.vae.to(memory_format=torch.channels_last)
    pipeline.vae = torch.compile(pipeline.vae)
    pipeline._exclude_from_cpu_offload = ["vae"]
    pipeline.enable_sequential_cpu_offload()
    for _ in range(1):
        clear()
        with torch.backends.cuda.sdp_kernel(enable_flash=False, enable_math=True, enable_mem_efficient=True):
            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)
    return pipeline


sample = True
@torch.inference_mode()
def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
    global sample
    if sample:
        clear()
        sample = None
    # torch.cuda.reset_peak_memory_stats()
    generator = Generator("cuda").manual_seed(request.seed)
    image = None
    with torch.backends.cuda.sdp_kernel(enable_flash=False, enable_math=True, enable_mem_efficient=True):
        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)