handler.py

# https://github.com/sayakpaul/diffusers-torchao

import os
from typing import Any, Dict

from diffusers import FluxPipeline, FluxTransformer2DModel, AutoencoderKL, TorchAoConfig
from PIL import Image
import torch
from torchao.quantization import quantize_, autoquant, int8_dynamic_activation_int8_weight, int8_dynamic_activation_int4_weight
from huggingface_hub import hf_hub_download

IS_COMPILE = False
IS_TURBO = False
IS_4BIT = True

if IS_COMPILE:
    import torch._dynamo
    torch._dynamo.config.suppress_errors = True

from huggingface_inference_toolkit.logging import logger

def load_pipeline_stable(repo_id: str, dtype: torch.dtype) -> Any:
    quantization_config = TorchAoConfig("int4dq" if IS_4BIT else "int8dq")
    vae = AutoencoderKL.from_pretrained(repo_id, subfolder="vae", torch_dtype=dtype)
    pipe = FluxPipeline.from_pretrained(repo_id, vae=vae, torch_dtype=dtype, quantization_config=quantization_config)
    pipe.transformer.fuse_qkv_projections()
    pipe.vae.fuse_qkv_projections()
    pipe.to("cuda")
    return pipe

def load_pipeline_compile(repo_id: str, dtype: torch.dtype) -> Any:
    quantization_config = TorchAoConfig("int4dq" if IS_4BIT else "int8dq")
    vae = AutoencoderKL.from_pretrained(repo_id, subfolder="vae", torch_dtype=dtype)
    pipe = FluxPipeline.from_pretrained(repo_id, vae=vae, torch_dtype=dtype, quantization_config=quantization_config)
    pipe.transformer.fuse_qkv_projections()
    pipe.vae.fuse_qkv_projections()
    pipe.transformer.to(memory_format=torch.channels_last)
    pipe.transformer = torch.compile(pipe.transformer, mode="reduce-overhead", fullgraph=False, dynamic=False)
    pipe.vae.to(memory_format=torch.channels_last)
    pipe.vae = torch.compile(pipe.vae, mode="reduce-overhead", fullgraph=False, dynamic=False)
    pipe.to("cuda")
    return pipe

def load_pipeline_autoquant(repo_id: str, dtype: torch.dtype) -> Any:
    pipe = FluxPipeline.from_pretrained(repo_id, torch_dtype=dtype).to("cuda")
    pipe.transformer.fuse_qkv_projections()
    pipe.vae.fuse_qkv_projections()
    pipe.transformer.to(memory_format=torch.channels_last)
    pipe.transformer = torch.compile(pipe.transformer, mode="max-autotune", fullgraph=True)
    pipe.vae.to(memory_format=torch.channels_last)
    pipe.vae = torch.compile(pipe.vae, mode="max-autotune", fullgraph=True)
    pipe.transformer = autoquant(pipe.transformer, error_on_unseen=False)
    pipe.vae = autoquant(pipe.vae, error_on_unseen=False)
    pipe.to("cuda")
    return pipe

def load_pipeline_turbo(repo_id: str, dtype: torch.dtype) -> Any:
    pipe = FluxPipeline.from_pretrained(repo_id, torch_dtype=dtype).to("cuda")
    pipe.load_lora_weights(hf_hub_download("ByteDance/Hyper-SD", "Hyper-FLUX.1-dev-8steps-lora.safetensors"), adapter_name="hyper-sd")
    pipe.set_adapters(["hyper-sd"], adapter_weights=[0.125])
    pipe.fuse_lora()
    pipe.transformer.fuse_qkv_projections()
    pipe.vae.fuse_qkv_projections()
    weight = int8_dynamic_activation_int4_weight() if IS_4BIT else int8_dynamic_activation_int8_weight()
    quantize_(pipe.transformer, weight, device="cuda")
    quantize_(pipe.vae, weight, device="cuda")
    quantize_(pipe.text_encoder_2, weight, device="cuda")
    pipe.to("cuda")
    return pipe

def load_pipeline_turbo_compile(repo_id: str, dtype: torch.dtype) -> Any:
    pipe = FluxPipeline.from_pretrained(repo_id, torch_dtype=dtype).to("cuda")
    pipe.load_lora_weights(hf_hub_download("ByteDance/Hyper-SD", "Hyper-FLUX.1-dev-8steps-lora.safetensors"), adapter_name="hyper-sd")
    pipe.set_adapters(["hyper-sd"], adapter_weights=[0.125])
    pipe.fuse_lora()
    pipe.transformer.fuse_qkv_projections()
    pipe.vae.fuse_qkv_projections()
    weight = int8_dynamic_activation_int4_weight() if IS_4BIT else int8_dynamic_activation_int8_weight()
    quantize_(pipe.transformer, weight, device="cuda")
    quantize_(pipe.vae, weight, device="cuda")
    quantize_(pipe.text_encoder_2, weight, device="cuda")
    pipe.transformer.to(memory_format=torch.channels_last)
    pipe.transformer = torch.compile(pipe.transformer, mode="reduce-overhead", fullgraph=False, dynamic=False)
    pipe.vae.to(memory_format=torch.channels_last)
    pipe.vae = torch.compile(pipe.vae, mode="reduce-overhead", fullgraph=False, dynamic=False)
    pipe.to("cuda")
    return pipe

class EndpointHandler:
    def __init__(self, path=""):
        repo_id = "NoMoreCopyrightOrg/flux-dev-8step" if IS_TURBO else "NoMoreCopyrightOrg/flux-dev"
        #dtype = torch.bfloat16
        dtype = torch.float16 # for older nVidia GPUs
        if IS_COMPILE: load_pipeline_compile(repo_id, dtype)
        else: self.pipeline = load_pipeline_stable(repo_id, dtype)

    def __call__(self, data: Dict[str, Any]) -> Image.Image:
        logger.info(f"Received incoming request with {data=}")

        if "inputs" in data and isinstance(data["inputs"], str):
            prompt = data.pop("inputs")
        elif "prompt" in data and isinstance(data["prompt"], str):
            prompt = data.pop("prompt")
        else:
            raise ValueError(
                "Provided input body must contain either the key `inputs` or `prompt` with the"
                " prompt to use for the image generation, and it needs to be a non-empty string."
            )

        parameters = data.pop("parameters", {})

        num_inference_steps = parameters.get("num_inference_steps", 8 if IS_TURBO else 28)
        width = parameters.get("width", 1024)
        height = parameters.get("height", 1024)
        guidance_scale = parameters.get("guidance_scale", 3.5)

        # seed generator (seed cannot be provided as is but via a generator)
        seed = parameters.get("seed", 0)
        generator = torch.manual_seed(seed)

        return self.pipeline(  # type: ignore
            prompt,
            height=height,
            width=width,
            guidance_scale=guidance_scale,
            num_inference_steps=num_inference_steps,
            generator=generator,
            output_type="pil",
        ).images[0]