使用 RunPod GPU

RunPod 是提供 GPU 资源的一个平台，可以在平台上租 GPU 按小时计费。支持所有主流的 GPU，价格也还好。本文将介绍如何在 RunPod 启动 Notebook 运行模型，模型使用 Flux UpScaler，显存需求大于 24 G。登陆网站，注册并充值，https://www.runpod.io/

创建 POD

选择 GPU 和 Docker 镜像，镜像选择 runpod/pytorch:2.4.0-py3.11-cuda12.4.1-devel-ubuntu22.04。

默认硬盘只有 20 G，如果模型很大，需要进行修改存储尺寸，点击 Edit Template，存储是要收费的，公用就可以。

启动Pod

安装需要的依赖

!pip install transformers diffusers accelerate protobuf sentencepiece matplotlib

UpScale

代码来自huggging face 的例子，https://huggingface.co/jasperai/Flux.1-dev-Controlnet-Upscaler

import torch
from diffusers.utils import load_image
from diffusers import FluxControlNetModel
from diffusers.pipelines import FluxControlNetPipeline
from huggingface_hub import login

login(token = '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')
from huggingface_hub import snapshot_download
flux = snapshot_download("black-forest-labs/FLUX.1-dev",local_dir="./flux")
upscaler = snapshot_download("jasperai/Flux.1-dev-Controlnet-Upscaler",local_dir="./flux")
# Load pipeline
controlnet = FluxControlNetModel.from_pretrained(
  upscaler,
  torch_dtype=torch.bfloat16
)
pipe = FluxControlNetPipeline.from_pretrained(
 flux,
  controlnet=controlnet,
  torch_dtype=torch.bfloat16
)
pipe.to("cuda")

# Load a control image
control_image = load_image(
  "https://huggingface.co/jasperai/Flux.1-dev-Controlnet-Upscaler/resolve/main/examples/input.jpg"
)

w, h = control_image.size

# Upscale x4
control_image = control_image.resize((w * 4, h * 4))

image = pipe(
    prompt="", 
    control_image=control_image,
    controlnet_conditioning_scale=0.6,
    num_inference_steps=28, 
    guidance_scale=3.5,
    height=control_image.size[1],
    width=control_image.size[0]
).images[0]
image

原始图片

import requests
from PIL import Image
from io import BytesIO
import matplotlib.pyplot as plt

# URL of the image
url = "https://huggingface.co/jasperai/Flux.1-dev-Controlnet-Upscaler/resolve/main/examples/input.jpg"

# Send a GET request to the URL
response = requests.get(url)

# Check if the request was successful
if response.status_code == 200:
    # Open the image from the response content
    image = Image.open(BytesIO(response.content))
    
    # Display the image
    plt.imshow(image)
    plt.axis('off')  # Hide axes
    plt.show()
else:
    print(f"Failed to retrieve image. Status code: {response.status_code}")

总结

RunPod 可以支持大显存需求的模型，将 Runpod 镜像下载到本地先调试好，例如需要的类库等等，这样可以省时间，有很多平台的都 提供 GPU 租赁，原理都是相似的，就看哪家更便宜些，国内的算力平台更便宜，就是英伟达的卡不好租。