src/pipeline.py

from huggingface_hub.constants import HF_HUB_CACHE
from diffusers import FluxPipeline
from PIL.Image import Image
from pipelines.models import TextToImageRequest
from torch import Generator
from diffusers import FluxTransformer2DModel

import torch
import torch._dynamo
import gc
import os

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

Pipeline = None
base_prompt = "insensible, timbale, pothery, electrovital, actinogram, taxis, intracerebellar, centrodesmus"

def load_pipeline() -> Pipeline:
    gc.collect()
    torch.cuda.empty_cache()
    torch.cuda.reset_max_memory_allocated()
    torch.cuda.reset_peak_memory_stats()

    transformer = FluxTransformer2DModel.from_pretrained(os.path.join(HF_HUB_CACHE, "models--fringuant--StreamCascade/snapshots/765016449ab8494685f030a7db03c67600cf4c55/transformer"), torch_dtype=torch.bfloat16, use_safetensors=False)
    pipeline = FluxPipeline.from_pretrained("fringuant/StreamCascade", revision="765016449ab8494685f030a7db03c67600cf4c55", transformer=transformer, local_files_only=True, torch_dtype=torch.bfloat16,)
    pipeline.to("cuda")
    pipeline.vae = torch.compile(pipeline.vae, mode="max-autotune", fullgraph=True, dynamic=True)
    
    for idx in range(3):
        pipeline(prompt=base_prompt, width=1024, height=1024, guidance_scale=5.0, num_inference_steps=4, max_sequence_length=256)
    
    return pipeline

@torch.no_grad()
def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
    prompt = getattr(request, 'prompt', base_prompt)
    return pipeline(
        prompt,
        generator=Generator(pipeline.device).manual_seed(request.seed),
        guidance_scale=6.5,
        num_inference_steps=4,
        max_sequence_length=256,
        height=request.height,
        width=request.width,
    ).images[0]