接上文 不使用 Docker 构建 Triton 服务器并在 Google Colab 平台上部署 HuggingFace 模型
MultiGPU && Multi Instance
Config
追加
bash
·
·
·
instance_group [
{
count: 4
kind: KIND_GPU
gpus: [ 0, 1 ]
}
]
Python Backend
Triton 会根据配置信息启动四个实例,model_instance_device_id
可以获取到 Triton 给每个实例自动分配的 GPU,模型加载到GPU时使用 .to(f"cuda:{gpu}"时指定 GPU 的 id 即可。
python
import numpy as np
import triton_python_backend_utils as pb_utils
from transformers import ViTImageProcessor, ViTModel
from diffusers import DiffusionPipeline
import torch
import time
import os
import shutil
import json
import numpy as np
class TritonPythonModel:
def initialize(self, args):
gpu = json.loads(args["model_instance_device_id"])
self.model = DiffusionPipeline.from_pretrained(
"playgroundai/playground-v2.5-1024px-aesthetic",
torch_dtype=torch.float16,
variant="fp16"
).to(f"cuda:{gpu}")
·
·
·
Dynamic Batch
开启此配置 Triton 会将一段时间间隔内的请求组成一个batch交给模型批处理
Config
bash
·
·
·
dynamic_batching {
max_queue_delay_microseconds: 100
}
Python Backend
此时需要对后端代码进行改造,之前的代码每次只能处理一个请求,GPU仅仅占用30G,对于80G的 GPU 来说实在是浪费资源。SDXL 依次是可以处理多个 Prompt 生成多个图片的,只要不把现存撑爆就行。现在我们要在最大 Batch 的限制内处理多个请求,假设客户端每个请求只包含一个 Prompt。
我们获取到一个请求后不直接输入到模型,而是都添加到 Prompts 列表里,然后统一生成图片,然后把生成的图片和请求一一对应上,最后响应给客户端就OK了。
python
·
·
·
def execute(self, requests):
responses = []
prompts = []
for request in requests:
inp = pb_utils.get_input_tensor_by_name(request, "prompt")
for i in inp.as_numpy():
prompts.append(i[0].decode())
images = self.model(prompt=prompts, num_inference_steps=50, guidance_scale=3).images
pixel_values = []
for image in images:
pixel_values.append(np.asarray(image))
inference_response = pb_utils.InferenceResponse(
output_tensors=[
pb_utils.Tensor(
"generated_image",
np.array(pixel_values),
)
]
)
responses.append(inference_response)
return responses
ONNX_RUNTIME(失败)
转换 playgroundai/playground-v2.5-1024px-aesthetic pipeline 为 onnx 格式
bash
pip3 install diffusers>=0.27.0 transformers accelerate safetensors optimum["onnxruntime"]
optimum-cli export onnx --model playgroundai/playground-v2.5-1024px-aesthetic --task stable-diffusion-xl playground-v2.5_onnx/
分享已经转换好的文件:
测试一下
python
from optimum.onnxruntime import ORTStableDiffusionXLPipeline
model_id = "playground-v2.5_onnx"
pipeline = ORTStableDiffusionXLPipeline.from_pretrained(model_id)
prompt = "sailing ship in storm by Leonardo da Vinci"
image = pipeline(prompt).images[0]
# pipeline.save_pretrained("playground-v2.5-onnx/")
成功加载,发现使用的是 CPU 推理,参考Accelerated inference on NVIDIA GPUs发现默认是用CPU,需要卸载 optimum["onnxruntime"] 安装 optimum[onnxruntime-gpu]。
python
pipeline = ORTStableDiffusionXLPipeline.from_pretrained(model_id, provider="CUDAExecutionProvider", device=f"cuda:{gpu}")
FAQ
CPU 推理正常,GPU推理报错,类似这种错误:
- [E:onnxruntime:Default, provider_bridge_ort.cc:1480 TryGetProviderInfo_CUDA] /onnxruntime_src/onnxruntime/core/session/provider_bridge_ort.cc:1193 onnxruntime::Provider& onnxruntime::ProviderLibrary::Get() [ONNXRuntimeError] : 1 : FAIL : Failed to load library libonnxruntime_providers_cuda.so with error: libcublasLt.so.12: cannot open shared object file: No such file or directory
解决方法: 安装CUDNN,官方手册
bash
apt install libcudnn8=8.9.2.26-1+cuda12.1 -y
apt install libcudnn8-dev=8.9.2.26-1+cuda12.1 -y
apt install libcudnn8-samples=8.9.2.26-1+cuda12.1 -y
- 2024-04-08 15:21:39.497586288 [E:onnxruntime:, sequential_executor.cc:514 ExecuteKernel] Non-zero status code returned while running Add node. Name:'/down_blocks.1/attentions.0/Add' Status Message: /down_blocks.1/attentions.0/Add: left operand cannot broadcast on dim 3 LeftShape: {2,64,4096,10}, RightShape: {2,640,64,64}
解决方法:
未解决,参考 Issue
TENSORRT_RUNTIME
Doing...