🚀 The feature, motivation and pitch
Liger-Kernel provides a simple way to improve training throughput, but we need to manually implement logging to see these results while training.
A lightweight monitoring interface would:
- Make Liger's benefits visible in real experiments
- Help debug performance regressions across release versions
The integration can feature a common liger_kernel.monitor module with a backend agnostic interface that can enable adding external logging backends such as Weights and Biases, Trackio, Comet, etc. in a seamless manner. The external backends would be decoupled such that the backends would only be imported if they are installed in the user environment. This would also ensure that no external dependencies are added to the core dependency list.
Feel free to give feedback on this RFC on the scope of the API and its design.
Alternatives
No response
Additional context
No response
🚀 The feature, motivation and pitch
Liger-Kernel provides a simple way to improve training throughput, but we need to manually implement logging to see these results while training.
A lightweight monitoring interface would:
The integration can feature a common
liger_kernel.monitormodule with a backend agnostic interface that can enable adding external logging backends such as Weights and Biases, Trackio, Comet, etc. in a seamless manner. The external backends would be decoupled such that the backends would only be imported if they are installed in the user environment. This would also ensure that no external dependencies are added to the core dependency list.Feel free to give feedback on this RFC on the scope of the API and its design.
Alternatives
No response
Additional context
No response