build without conda

[ShivaKothuru]

📚 Documentation

Can I have this build without conda? If yes, please specify the steps.

Thank you.

[gyeongin]

Thank you for the question!
We implemented Nimble on PyTorch (we inserted our code into PyTorch codebase), so you need to build PyTorch from source to install Nimble.
According to the official guideline of PyTorch, the PyTorch team highly recommend building PyTorch with Anaconda environment, and there is no documentation about how to build PyTorch without Anaconda.

So I also recommend you to build PyTorch (and Nimble) using conda environment.
We haven't tried building PyTorch without conda so far.

[ShivaKothuru]

Thank you @gyeongin

I would like to implement Nimble on Jetson device(e.g. Jeston nano, TX2) for faster inference.
So, I would like to know if we can build without conda.
Also, Can we build with simplified version just for inference?

[gyeongin]

For the first question about Jetson devices, there is a document for building PyTorch from source on Jetson devices.
I think you can also build Nimble if you follow this instruction. (I cannot verify this by myself because I don't have Jetson 😢)

For the second question, our team is currently discussing about inference-only version of Nimble, which does not modify PyTorch codebase so can be installed without building PyTorch from source.
But for now, there is no inference-only version available.

[eash3010]

Hi @ShivaKothuru,
Were you able to run Nimble on Jetson devices? If so, could you please share the steps to do so.
Thank you!

[WandererCU]

For the second question, our team is currently discussing about inference-only version of Nimble, which does not modify PyTorch codebase so can be installed without building PyTorch from source. But for now, there is no inference-only version available.

Hi @gyeongin , could you talk about how to implement the inference-only version without modifying the PyTorch codebase? Does that mean you will implement it in Python and run the tasks by calling some Pytorch functions? Will you use the Pytorch CUDA graph API? Thanks!

[umairjavaid]

Waiting for the answer for the above question