src/pipeline.py

import os
import gc
import torch
import torch._dynamo
from PIL import Image
from typing import Type
from torch import Generator
from transformers import T5EncoderModel
from huggingface_hub.constants import HF_HUB_CACHE
from pipelines.models import TextToImageRequest
from torchao.quantization import quantize_, int8_weight_only
from diffusers import AutoencoderKL, FluxPipeline, FluxTransformer2DModel

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
torch.backends.cudnn.benchmark = True

Pipeline = None

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

def load_pipeline() -> Pipeline:
    ckpt_id = "black-forest-labs/FLUX.1-schnell"
    ckpt_revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"
    
    text_encoder_2 = T5EncoderModel.from_pretrained("db900/encode-stream", revision = "8a6b7bd09dc66733fa582900186f929353f63619", subfolder="text_encoder_2", torch_dtype=torch.bfloat16).to(memory_format=torch.channels_last)
    path = os.path.join(HF_HUB_CACHE, "models--db900--encode-stream/snapshots/8a6b7bd09dc66733fa582900186f929353f63619/transformer")
    transformer = FluxTransformer2DModel.from_pretrained(path, torch_dtype=torch.bfloat16, use_safetensors=False).to(memory_format=torch.channels_last)
    quantize_(AutoencoderKL.from_pretrained(ckpt_id, revision=ckpt_revision, subfolder="vae", local_files_only=True, torch_dtype=torch.bfloat16,), int8_weight_only())
    pipeline = FluxPipeline.from_pretrained(ckpt_id, revision=ckpt_revision, transformer=transformer, text_encoder_2=text_encoder_2, torch_dtype=torch.bfloat16,)
    pipeline.to("cuda")
    with torch.inference_mode():
        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]