add cross stack profiler to profile super ernie

[lw921014]

PR types

New features

PR changes

Others

Describe

We develop a Cross Stack Profile (CSP) tool to help those SOTA models obtain best performance, such as super ernie, GPT-3 and so on, especially when leveraging many sophisticated distributed strategies inside a large-scale cluster which may has thousands of GPUs.

CSP is the first profile toolkit in the industry, which

1. can collect cross stack data, include system resources, framework runtime info and user-defined description;
2. can convert gigebytes raw data from thousands of nodes into useful knowledge with little overhead and code intrusion;
3. can achieve nearly real-time display by making full use of multi-core processor to generate friendly tracing file.

Following is an example, and we try to make sure that our model run the same mode as that mentioned in the paper.

figure 1 running model mentioned in paper

figure 2 running model from CSP

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[gongweibao]

LGTM