src/pipeline.py

import os
from diffusers import FluxPipeline, AutoencoderKL, FluxTransformer2DModel
from diffusers.image_processor import VaeImageProcessor
from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel, CLIPTextConfig, T5Config
import torch
import gc
from PIL import Image
from pipelines.models import TextToImageRequest
from torch import Generator
from time import perf_counter


os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"

CHECKPOINT = "black-forest-labs/FLUX.1-schnell"

DTYPE = torch.bfloat16
NUM_STEPS = 4


def empty_cache():
    gc.collect()
    torch.cuda.empty_cache()
    torch.cuda.reset_max_memory_allocated()
    torch.cuda.reset_peak_memory_stats()


def load_pipeline() -> FluxPipeline:
    empty_cache()
    
    pipe = FluxPipeline.from_pretrained(CHECKPOINT, torch_dtype=DTYPE)
    
    pipe.text_encoder_2.to(memory_format=torch.channels_last)
    pipe.transformer.to(memory_format=torch.channels_last)
    
    pipe.vae.to(memory_format=torch.channels_last)
    pipe.vae = torch.compile(pipe.vae)
    pipe._exclude_from_cpu_offload = ["vae"]
    
    pipe.enable_sequential_cpu_offload()
    
    empty_cache()
    pipe("dog", guidance_scale=0.0, max_sequence_length=256, num_inference_steps=4)
    return pipe


@torch.inference_mode()
def infer(request: TextToImageRequest, _pipeline: FluxPipeline) -> Image:
    torch.cuda.reset_peak_memory_stats()
    
    if request.seed is None:
        generator = None
    else:
        generator = Generator(device="cuda").manual_seed(request.seed)

    empty_cache()
    image = _pipeline(prompt=request.prompt,
                      width=request.width,
                      height=request.height,
                      guidance_scale=0.0,
                      generator=generator,
                      output_type="pil",
                      max_sequence_length=256,
                      num_inference_steps=NUM_STEPS).images[0]
    return image