src/pipeline.py

import os
import gc
import time
import torch
from PIL import Image as img
from PIL.Image import Image
from diffusers import (
    FluxTransformer2DModel,
    DiffusionPipeline,
    AutoencoderTiny
)
from transformers import T5EncoderModel
from huggingface_hub.constants import HF_HUB_CACHE
from torchao.quantization import quantize_, int8_weight_only
from first_block_cache.diffusers_adapters import apply_cache_on_pipe
from pipelines.models import TextToImageRequest
from torch import Generator

os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"

Pipeline = None
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = True

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


def are_two_tensors_similar(t1, t2, *, threshold, parallelized=False):
    diff_norm = torch.norm(t1 - t2, p='fro')
    base_norm = torch.norm(t1, p='fro')
    relative_diff = diff_norm / (base_norm + 1e-8)
    return relative_diff.item() < 0.5


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()

    dtype, device = torch.bfloat16, "cuda"

    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)

    path = os.path.join(HF_HUB_CACHE, "models--RobertML--FLUX.1-schnell-int8wo/snapshots/307e0777d92df966a3c0f99f31a6ee8957a9857a")
    model = FluxTransformer2DModel.from_pretrained(
        path, 
        torch_dtype=dtype, 
        use_safetensors=False
    ).to(memory_format=torch.channels_last)
    
    pipeline = DiffusionPipeline.from_pretrained(
        ckpt_id,
        revision=ckpt_revision,
        transformer=model,
        text_encoder_2=text_encoder_2,
        torch_dtype=dtype,
    ).to(device)
    
    #quantize_(pipeline.vae, int8_weight_only())
    apply_cache_on_pipe(pipeline)

    for _ in range(3):
        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.no_grad()
def infer(request: TextToImageRequest, pipeline: Pipeline, generator: Generator) -> Image:
    try:
        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]
    except:
        image = img.open("./RobertML.png")
    return image