Usability issues when using BladeDISC to compile GPT3

[tanyokwok]

Describe We meet some usability issues when using BladeDISC to compile GPT3. Track them here:

• [ ] Performance issue @tanyokwok @qiuxiafei @JamesTheZ
• [ ] BladeDISC compilation's time is too long @wyzero @tanyokwok
• [ ] The model is too large to success save to the file @tanyokwok @qiuxiafei

[qiuxiafei]

Another: model startup time is too long