src/pipeline.py

from diffusers import DiffusionPipeline, AutoencoderKL
from transformers import T5EncoderModel
import torch
import gc
from PIL import Image
from pipelines.models import TextToImageRequest  # Assuming this defines your request object
import os
from torch import Generator

Pipeline = None
# Consistent environment variable setting
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:128"  # More robust memory management

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

def load_pipeline() -> Pipeline:
    gc.collect()
    # torch.cuda.empty_cache()

    dtype = torch.bfloat16
    text_encoder_2 = T5EncoderModel.from_pretrained(
        "city96/t5-v1_1-xxl-encoder-bf16", torch_dtype=dtype
    ).to(memory_format=torch.channels_last)
    vae = AutoencoderKL.from_pretrained(ckpt_id, subfolder="vae", torch_dtype=dtype).to(memory_format=torch.channels_last)


    pipeline = DiffusionPipeline.from_pretrained(
        ckpt_id,
        vae=vae,
        text_encoder_2=text_encoder_2,
        torch_dtype=dtype,
    )#.to("cuda")
    pipeline.transformer.to(memory_format=torch.channels_last)
    # pipeline.text_encoder.to(memory_format=torch.channels_last)
    

    # Optimize after moving to GPU
    pipeline.vae = torch.compile(pipeline.vae) # compile after moving to device

    # It's unclear if offloading helped in the originals. Test with and without!
    pipeline._exclude_from_cpu_offload = ["vae"]
    pipeline.enable_sequential_cpu_offload()

    # Warmup on GPU
    for _ in range(2):
        pipeline(prompt="onomancy, aftergo, spirantic, Platyhelmia, modificator, drupaceous, jobbernowl, hereness", width=1024, height=1024, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256)
    
    return pipeline



@torch.inference_mode()
def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
    torch.cuda.reset_peak_memory_stats()
    generator = Generator("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)