Do not ship nvrtc with LuxCore Wheels

[howetuft]

Overview

nvrtc stands for "NVidia Runtime Compiler". nvrtc is the external program that allows to compile Cuda kernels at runtime.
Till version 2.9, nvrtc is shipped together with LuxCoreRender, this encompassing BlendLuxCore, luxcoreui, luxcoreconsole etc.

For the Wheels, we propose not to ship nvrtc together.

Rationale

• nvrtc has become very heavy in last versions. Version 12.8 is 110Mb, thus embedding it would more than double Python wheel sizes, which are already big. It would not only be a drawback for Nvidia users but a real pain for a large proportion of users who are not concerned by this tool - macOS users, other GPU vendors (AMD, Intel...) users - to whom we would impose a significant overhead, with no gain.
• Embedding GPU vendor runtime compiler is not sustainable if we plan to address also other GPUs, like Metal, AMD, Intel... We won't be able add 100Mb for each vendor.
• nvrtc releasing activity is rather high, with nearly one new version per month (see here: https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/linux-x86_64/). We should delegate updating burden to standard package managers.
• to be optimal, nvrtc version should be fitted to user's GPU and nvidia driver. Either we target an old version, but users with recent GPUs won't benefit from that, or we target a recent version, but we may exclude uses users with old GPUs.
• In the mid term, shipping precompiled kernels would be more comfortable for user, with no lag at kernel compilation like today.

Solution

nvrtc is available as a pip package. We will just declare this package as a dependency for LuxCore Python wheels.

[CodeFHD]

@howetuft Reviewing if this Discussion can be closed or should remain open.

pyluxcore now relies on a pip package (as described in the original post) so the original point of this Discussion has been completed as far as I can tell.

The idea of switching to pre-compiled kernels is open, but this is mid-term and could be discussed, if needed, in a separate Discussion or via an Issue.

Do you agree to close, or do you want to keep it open with specific actions?

[howetuft]

Yes, we can close! I do it at once.