src/pipeline.py

from diffusers import FluxPipeline, AutoencoderKL, AutoencoderTiny
from diffusers.image_processor import VaeImageProcessor
from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
from huggingface_hub.constants import HF_HUB_CACHE
from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel
import torch
import torch._dynamo
import gc
from PIL import Image as img
from PIL.Image import Image
from pipelines.models import TextToImageRequest
from torch import Generator
import time
from diffusers import FluxTransformer2DModel, DiffusionPipeline
from torchao.quantization import quantize_, int8_weight_only, fpx_weight_only
import os

os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
torch._dynamo.config.suppress_errors = True

Pipeline = None

ckpt_id = "black-forest-labs/FLUX.1-schnell"
ckpt_revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"


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:
    empty_cache()

    transformer_path = os.path.join(
        HF_HUB_CACHE,
        "models--park234--Flux1-schnell-int8/snapshots/ba40332fa9ca3e25d439616fca8ac934a96b73f7",
    )
    transformer_model = FluxTransformer2DModel.from_pretrained(
        transformer_path, torch_dtype=torch.bfloat16, use_safetensors=False
    ).to(memory_format=torch.channels_last)
    
    vae_model = AutoencoderTiny.from_pretrained(
        "park234/flux1-schnell-vae32",
        revision="d7c57ba8ee0d581d67256219c152a8e69da853b5",
        torch_dtype=torch.float32,
    )
    
    text_encoder_2 = T5EncoderModel.from_pretrained(
        "city96/t5-v1_1-xxl-encoder-bf16",
        revision="1b9c856aadb864af93c1dcdc226c2774fa67bc86",
        torch_dtype=torch.bfloat16,
    ).to(memory_format=torch.channels_last)
    
    pipeline = FluxPipeline.from_pretrained(
        ckpt_id,
        vae=vae_model,
        revision=ckpt_revision,
        transformer=transformer_model,
        text_encoder_2=text_encoder_2,
        torch_dtype=torch.bfloat16,
    ).to("cuda")
    
    pipeline.transformer = torch.compile(pipeline.transformer, mode="reduce-overhead")

    empty_cache()
    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]