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.Image import Image
from pipelines.models import TextToImageRequest
from torch import Generator
from torchao.quantization import quantize_, int8_weight_only
from time import perf_counter


HOME = os.environ["HOME"]
QUANTIZED_MODEL = ["text_encoder_2", "vae"]
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:False,garbage_collection_threshold:0.01"
FLUX_CHECKPOINT = "jokerbit/flux.1-schnell-Robert-int8wo"
FLUX_CACHE = os.path.join(HOME, ".cache/huggingface/hub/models--jokerbit--flux.1-schnell-Robert-int8wo/snapshots/5ef0012f11a863e5111ec56540302a023bc8587b")
torch.backends.cudnn.benchmark = True
torch.backends.cuda.matmul.allow_tf32 = True
torch.cuda.set_per_process_memory_fraction(0.99)

QUANT_CONFIG = int8_weight_only()
DTYPE = torch.bfloat16
NUM_STEPS = 4
PROMPT = 'martyr, semiconformity, peregrination, quip, twineless, emotionless, tawa, depickle'


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


def quantize(pipe, config):
    if "text_encoder" in QUANTIZED_MODEL:
        quantize_(pipe.text_encoder, config)
    if "text_encoder_2" in QUANTIZED_MODEL:
        quantize_(pipe.text_encoder_2, config)
    if "transformer" in QUANTIZED_MODEL:
        quantize_(pipe.transformer, config, device="cuda")
    if "vae" in QUANTIZED_MODEL:
        quantize_(pipe.vae, config)
    return pipe


def load_pipeline() -> FluxPipeline:
    empty_cache()
    transformer = FluxTransformer2DModel.from_pretrained(os.path.join(FLUX_CACHE, "transformer"), use_safetensors=False, torch_dtype=DTYPE)
    pipe = FluxPipeline.from_pretrained(FLUX_CHECKPOINT,
                                        transformer=transformer,
                                        torch_dtype=DTYPE) 
    pipe.vae.enable_tiling()
    pipe.vae.enable_slicing()
    quantize(pipe, QUANT_CONFIG)
    pipe.to("cuda")
    request = TextToImageRequest(prompt=PROMPT, height=1024, width=1024, seed=666)
    infer(request, pipe)
    # pipe.enable_model_cpu_offload()
    return pipe


def encode_prompt(_pipeline, prompt: str):
    pipeline = FluxPipeline.from_pipe(
            _pipeline,
            transformer=None,
            vae=None,
            ).to("cuda")
    with torch.no_grad():
        outputs = pipeline.encode_prompt(
                prompt=prompt,
                prompt_2=None,
                max_sequence_length=256)
    del pipeline 
    empty_cache()
    return outputs


def infer_latents(_pipeline, prompt_embeds, pooled_prompt_embeds, width: int | None, height: int | None, seed: int | None):
    pipeline = FluxPipeline.from_pipe(
            _pipeline,
            text_encoder=None,
            text_encoder_2=None,
            tokenizer=None,
            tokenizer_2=None,
            vae=None,
            ).to("cuda")

    if seed is None:
        generator = None
    else:
        generator = Generator(pipeline.device).manual_seed(seed)
    outputs = pipeline(
            prompt_embeds=prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            num_inference_steps=4,
            guidance_scale=0.0,
            width=width,
            height=height,
            generator=generator,
            output_type="latent",
            ).images
    del pipeline
    empty_cache()
    return outputs


def decode_latents(vae, latents, width, height):
    vae.to("cuda")
    vae_scale_factor = 2 ** (len(vae.config.block_out_channels))
    width = width or 64 * vae_scale_factor
    height = height or 64 * vae_scale_factor
    image_processor = VaeImageProcessor(vae_scale_factor=vae_scale_factor)
    with torch.no_grad():
        latents = FluxPipeline._unpack_latents(latents, height, width, vae_scale_factor)
        latents = (latents / vae.config.scaling_factor) + vae.config.shift_factor
        image = vae.decode(latents, return_dict=False)[0] 
    return image_processor.postprocess(image, output_type="pil")[0]


# def infer(request: TextToImageRequest, _pipeline: FluxPipeline) -> Image: 
#     prompt_embeds, pooled_prompt_embeds, text_ids = encode_prompt(_pipeline, request.prompt) 
#     latents = infer_latents(_pipeline, prompt_embeds, pooled_prompt_embeds, request.width, request.height, request.seed)
#     del prompt_embeds
#     del pooled_prompt_embeds
#     del text_ids
#     # _pipeline.transformer.to("cpu")
#     image = decode_latents(_pipeline.vae, latents, request.width, request.height)
#     torch.cuda.reset_peak_memory_stats()
#     return image


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

if __name__ == "__main__":
    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")

    # pipe("cat holding a womboai sign", num_inference_steps=4, guidance_scale=0, generator=torch.Generator(pipe.device).manual_seed(666)).images[0].save("sample.png")