src/pipeline.py

import gc
import os
from typing import TypeAlias

import torch
from PIL.Image import Image
from diffusers import FluxPipeline, FluxTransformer2DModel, AutoencoderKL, AutoencoderTiny
from huggingface_hub.constants import HF_HUB_CACHE
from pipelines.models import TextToImageRequest
from torch import Generator
from torchao.quantization import quantize_, int8_weight_only
from transformers import T5EncoderModel, CLIPTextModel


Pipeline: TypeAlias = FluxPipeline
torch.backends.cudnn.benchmark = True

CHECKPOINT = "jokerbit/flux.1-schnell-Robert-int8wo"
REVISION = "5ef0012f11a863e5111ec56540302a023bc8587b"

TinyVAE = "madebyollin/taef1"
TinyVAE_REV = "2d552378e58c9c94201075708d7de4e1163b2689"


def load_pipeline() -> Pipeline:
    text_encoder = CLIPTextModel.from_pretrained(
        CHECKPOINT,
        revision=REVISION,
        subfolder="text_encoder",
        local_files_only=True,
        torch_dtype=torch.bfloat16,
    )

    text_encoder_2 = T5EncoderModel.from_pretrained(
        CHECKPOINT,
        revision=REVISION,
        subfolder="text_encoder_2",
        local_files_only=True,
        torch_dtype=torch.bfloat16,
    )

    vae = AutoencoderTiny.from_pretrained(
        TinyVAE,
        revision=TinyVAE_REV,
        local_files_only=True,
        torch_dtype=torch.bfloat16,
    )

    path = os.path.join(HF_HUB_CACHE, "models--jokerbit--flux.1-schnell-Robert-int8wo/snapshots/5ef0012f11a863e5111ec56540302a023bc8587b/transformer")

    transformer = FluxTransformer2DModel.from_pretrained(
        path,
        torch_dtype=torch.bfloat16,
        use_safetensors=False,
    )
    pipeline = FluxPipeline.from_pretrained(
        CHECKPOINT,
        revision=REVISION,
        local_files_only=True,
        text_encoder=text_encoder,
        text_encoder_2=text_encoder_2,
        transformer=transformer,
        vae=vae,
        torch_dtype=torch.bfloat16,
    ).to("cuda") 
#    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) 
    for _ in range(2):
        pipeline("cat", num_inference_steps=4)

    return pipeline

def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
    gc.collect()
    torch.cuda.empty_cache()
    torch.cuda.reset_peak_memory_stats()

    generator = Generator(pipeline.device).manual_seed(request.seed)

    return pipeline(
        request.prompt,
        generator=generator,
        guidance_scale=0.0,
        num_inference_steps=4,
        max_sequence_length=256,
        height=request.height,
        width=request.width,
    ).images[0]


if __name__ == "__main__":
    from time import perf_counter
    PROMPT = 'martyr, semiconformity, peregrination, quip, twineless, emotionless, tawa, depickle'
    request = TextToImageRequest(prompt=PROMPT, 
            height=None,
            width=None,
            seed=666)
    start_time = perf_counter()
    pipe_ = load_pipeline()
    stop_time = perf_counter()
    print(f"Pipeline is loaded in {stop_time - start_time}s")
    for _ in range(4):
        start_time = perf_counter()
        infer(request, pipe_)
        stop_time = perf_counter()
        print(f"Request in {stop_time - start_time}s")