Wu Yi

Paddle v1 pserver uses OMPI_COMM_WORLD_SIZE for cost averaging: `paddle/pserver/ParameterServer2.cpp:101`, so paddle cloud should set this env for v1 job to work.

Do we need paddlectl client once we have the kubernetes custom controller?

15

Once we have TPR/CRD declared resource: ```yaml apiVersion: paddlepaddle.org/v1 kind: TrainingJob metadata: name: job-1 spec: image: "paddlepaddle/paddlecloud-job" trainer: entrypoint: "python train.py" workspace: "/home/job-1/" min-instance: 3 max-instance: 6 resources: limits: alpha.kubernetes.io/nvidia-gpu:...

"publish" command doesn't check whether the file exists in the cloud, causing error published files can not download.

Add the web pages basic feature design for paddle cloud.

[Feature] Add serve command to start a serverless predict serving port

2

Can run `paddlecloud serve -model-path xxx -scale 100 -cpu 1 -memory 8Gi -entry "infer.py" to start a serverless URL endpoint for serve the model.

1. Configure submit backend in `~/.paddle/config`, like kubernetes, docker-copose, etc. 1. User run `paddlecloud submit` to submit a training job. 1. Paddle cloud server will dispatch the job to corresponding...

Like in https://github.com/PaddlePaddle/cloud/issues/297 We have many envs defined for jobs to use internally, need to document about that.

Support gpu-manager vGPU resource for notebook

8

Support [gpu-manager](https://github.com/tkestack/gpu-manager) vGPU resource for jupyter notebook. Will add vGPU support for Kubeflow Pipeline later.

feat: add `memoryVendors` and arbitrary counts for notebook GPU resource limits

4

Add support for configuring arbitrary GPU resource types like https://github.com/4paradigm/k8s-vgpu-scheduler, https://github.com/tkestack/gpu-manager etc. Take using https://github.com/tkestack/gpu-manager as an example, after install `gpu-manager` the cluster should provide virtual GPU resource types `tencent.com/vcuda-core`...