src/pipeline.py

import os
import gc
import torch
from torch import Generator
from PIL.Image import Image
from diffusers import AutoencoderKL, FluxPipeline
from diffusers.image_processor import VaeImageProcessor
from pipelines.models import TextToImageRequest
from transformers import T5EncoderModel
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:False,garbage_collection_threshold:0.001"
torch.set_float32_matmul_precision("medium")
os.environ["TOKENIZERS_PARALLELISM"] = "True"
ckpt_id = "black-forest-labs/FLUX.1-schnell"
dtype = torch.bfloat16
Pipeline = None
# Configure CUDA settings
torch.backends.cudnn.benchmark = True
torch.backends.cuda.matmul.allow_tf32 = True
torch.cuda.set_per_process_memory_fraction(0.999)

class BasicQuantization:
    def __init__(self, bits=16):
        self.bits = bits
        self.qmin = -(2**(bits-1))
        self.qmax = 2**(bits-1) - 1

    def quantize_tensor(self, tensor):
        scale = (tensor.max() - tensor.min()) / (self.qmax - self.qmin)
        zero_point = self.qmin - torch.round(tensor.min() / scale)
        qtensor = torch.round(tensor / scale + zero_point)
        qtensor = torch.clamp(qtensor, self.qmin, self.qmax)
        return (qtensor - zero_point) * scale, scale, zero_point

class ModelQuantization:
    def __init__(self, model, bits=16):
        self.model = model
        self.quant = BasicQuantization(bits)

    def quantize_model(self):
        for name, module in self.model.named_modules():
            if isinstance(module, torch.nn.Linear):
                if hasattr(module, 'weightML'):
                    quantized_weight, _, _ = self.quant.quantize_tensor(module.weight)
                    module.weight = torch.nn.Parameter(quantized_weight)
                if hasattr(module, 'bias') and module.bias is not None:
                    quantized_bias, _, _ = self.quant.quantize_tensor(module.bias)
                    module.bias = torch.nn.Parameter(quantized_bias)

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()
    
    # Load and quantize VAE
    vae = AutoencoderKL.from_pretrained(ckpt_id, subfolder="vae", torch_dtype=dtype)
    quantizer = ModelQuantization(vae)
    quantizer.quantize_model()
    
    text_encoder_2 = T5EncoderModel.from_pretrained(
        "city96/t5-v1_1-xxl-encoder-bf16", torch_dtype=torch.bfloat16
    )

    # Initialize pipeline
    pipeline = FluxPipeline.from_pretrained(
        ckpt_id,
        text_encoder_2=text_encoder_2,
        vae=vae,
        torch_dtype=dtype
    )


    # Optimize memory format
    for component in [pipeline.text_encoder, pipeline.text_encoder_2, pipeline.transformer, pipeline.vae]:
        component.to(memory_format=torch.channels_last)

    # Compile and configure pipeline
    pipeline.vae = torch.compile(pipeline.vae, fullgraph=True, dynamic=False, mode="max-autotune")
    pipeline._exclude_from_cpu_offload = ["vae"]
    pipeline.enable_sequential_cpu_offload()

    # Warmup run
    empty_cache()
    for _ in range(3):
        pipeline(
            prompt="posteroexternal, eurythmical, inspection, semicotton, specification, Mercatorial, ethylate, misprint",
            width=1024,
            height=1024,
            guidance_scale=0.0,
            num_inference_steps=4,
            max_sequence_length=256
        )
    
    return pipeline

_inference_count = 0

@torch.inference_mode()
def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
    global _inference_count
    
    # Clear on first inference
    if _inference_count == 0:
        empty_cache()
    
    # Increment counter and empty cache every 4 inferences
    _inference_count += 1
    if _inference_count >= 4:
        empty_cache()
        _inference_count = 0
    
    torch.cuda.reset_peak_memory_stats()
    generator = Generator("cuda").manual_seed(request.seed)
    return pipeline(
            prompt=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]