【深度学习】SDXL tensorRT 推理，Stable Diffusion 转onnx，转TensorRT

sdxl 转 diffusers

def convert_sdxl_to_diffusers(pretrained_ckpt_path, output_diffusers_path):
    import os
    os.environ["HF_ENDPOINT"] = "https://hf-mirror.com"  # 设置 HF 镜像源（国内用户使用）
    os.environ["CUDA_VISIBLE_DEVICES"] = "1"  # 设置 GPU 所使用的节点

    import torch
    from diffusers import StableDiffusionXLPipeline
    pipe = StableDiffusionXLPipeline.from_single_file(pretrained_ckpt_path, torch_dtype=torch.float16).to("cuda")
    pipe.save_pretrained(output_diffusers_path, variant="fp16")

转onnx

项目：https://huggingface.co/docs/diffusers/optimization/onnx

比如转sdxl模型：

optimum-cli export onnx --model stabilityai/stable-diffusion-xl-base-1.0 --task stable-diffusion-xl sd_xl_onnx/

optimum-cli export onnx --model frankjoshua/juggernautXL_version6Rundiffusion --task stable-diffusion-xl sdxl_onnx_juggernautXL_version6Rundiffusion

转TensorRT

stabilityai/stable-diffusion-xl-1.0-tensorrt

项目：https://huggingface.co/stabilityai/stable-diffusion-xl-1.0-tensorrt

TensorRT环境：

git clone https://github.com/rajeevsrao/TensorRT.git
cd TensorRT
git checkout release/9.2

stabilityai/stable-diffusion-xl-1.0-tensorrt项目

git lfs install 
git clone https://huggingface.co/stabilityai/stable-diffusion-xl-1.0-tensorrt
cd stable-diffusion-xl-1.0-tensorrt
git lfs pull
cd ..

进入容器：

docker run -it --gpus all -v $PWD:/workspace nvcr.io/nvidia/pytorch:23.11-py3 /bin/bash

安装环境：

cd demo/Diffusion
python3 -m pip install --upgrade pip
pip3 install -r requirements.txt
python3 -m pip install --pre --upgrade --extra-index-url https://pypi.nvidia.com tensorrt

执行SDXL推理：

python3 demo_txt2img_xl.py   "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"   --build-static-batch   --use-cuda-graph   --num-warmup-runs 1   --width 1024   --height 1024   --denoising-steps 30  --version=xl-1.0   --onnx-dir /workspace/stable-diffusion-xl-1.0-tensorrt/sdxl-1.0-base   --onnx-refiner-dir /workspace/stable-diffusion-xl-1.0-tensorrt/sdxl-1.0-refiner

python3 demo_txt2img_xl.py   "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"   --build-static-batch   --use-cuda-graph   --num-warmup-runs 1   --width 1024   --height 1024   --denoising-steps 30  --version=xl-1.0   --onnx-dir /workspace/sdxl_onnx_juggernautXL_version6Rundiffusion

这个py代码对终端解析有时候有点问题，直接在代码里改一下，直接指定一下：

3090速度：

SDXL-LCM

python3 demo_txt2img_xl.py \
  "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" \
  --version=xl-1.0 \
  --onnx-dir /workspace/stable-diffusion-xl-1.0-tensorrt/lcm \
  --engine-dir /workspace/stable-diffusion-xl-1.0-tensorrt/lcm/engine-sdxl-lcm-nocfg \
  --scheduler LCM \
  --denoising-steps 4 \
  --guidance-scale 0.0 \
  --seed 42

SDXL-LCMLORA

python3 demo_txt2img_xl.py \
  "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" \
  --version=xl-1.0 \
  --onnx-dir /workspace/stable-diffusion-xl-1.0-tensorrt/lcmlora \
  --engine-dir /workspace/stable-diffusion-xl-1.0-tensorrt/lcm/engine-sdxl-lcmlora-nocfg \
  --scheduler LCM \
  --lora-path latent-consistency/lcm-lora-sdxl \
  --lora-scale 1.0 \
  --denoising-steps 4 \
  --guidance-scale 0.0 \
  --seed 42

3090速度：