feat: hybrid modeling

[tianhaox]

Overview:

• rename sol mode of database to silicon mode.
• add empirical mode, where the latency = sol / empirical factor. Current emprical factor is a constant number, could be more complicated in future. Add each op should design its own empirical factors.
• add hybrid mode, when missing silicon data, in this mode, allow to fallback to empirical mode.

So in future, when you add a new op, you can support running configuration without actual data if you use sol/empirical/hybrid mode. The hybrid mode should give you the best guess of the overal model performance.

This is the agg mode of running qwen3 32b on h200. Pure emprical mode has much smaller diff than sol mode, comparing with silicon mode. Imagine if part of the model is missing, the hybrid mode can give you a very close result unless the missing part is taking the major portion.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[Arsene12358]

Also, it seems we only enable this via webapp, what about users using the CLI?

[tianhaox]

Get the hybrid time, hybrid math and hybrid mem

should be enabled afterwards when we do some refactoring of the config

[davilu-nvidia]

I'll approve first to unblock merging.

[tianhaox]

Also, it seems we only enable this via webapp, what about users using the CLI?

added cli.
aiconfigurator cli default --model QWEN3_480B --total_gpus 32 --system b200_sxm --database_mode HYBRID
this command shall work while silicon mode cannot.