Conversation
- Add the `download_torch_versions.py` script to download Torch wheels. We can use this to grab PyTorch RC versions and upload them to HF Hub (for URLs with stable hashes). - Extend `generate_torch_versions.py` to handle the new `torchTesting` key in `torch_versions.json` entries. The presence of this key indicates that we want to download an RC version and what RC version. These versions are then downloaded from HF Hub.
Also use 12.9.4 for CUDA 12.6 and 12.8 because Torch requires it.
Currently only CUDA. ROCm, XPU, Metal, etc. will be added in subsequent PRs.
In nixpkgs `cudaPackages.backendStdenv` is not a wrapper anymore and calling `override` applies it to what used to be the underlying stdenv. Adjust our use of `backendStdenv` to fix this. Unfortunately, this also entails vendoring `backendStdenv`, since we cannot override it anymore.
|
|
||
| cuda_nvcc = cudaPackages.cuda_nvcc.override { | ||
| backendStdenv = cudaPackages.backendStdenv.override { | ||
| backendStdenv = import ../../pkgs/cuda/backendStdenv { |
There was a problem hiding this comment.
backendStdenv itself is not overridable anymore in upstream nixpkgs (.override is on the actual stdenv and not what used to be the wrapper). So I decided the easiest is to vendor this functionality until I come up with a better idea :).
| @@ -0,0 +1,273 @@ | |||
| # Vendored from nixpkgs, since we need to be able to pass another stdenv. | |||
There was a problem hiding this comment.
This file is vendored, so does not need a review.
There was a problem hiding this comment.
just adding a pinned link to the original file in this comment for future reference
https://github.com/NixOS/nixpkgs/blob/f8ce89e3edbc488a5b17c559ad55f083282420e9/pkgs/development/cuda-modules/backendStdenv/default.nix
| "12.6" = cuda_12; | ||
| "12.8" = cuda_12; | ||
| "12.9" = cuda_12; |
There was a problem hiding this comment.
Turns out we always need the latest, as required by Torch now. However, the latest does not build against older CUDA versions, so we use a binary wheel instead.
2 participants
Add support for Torch 2.10. Related changes in this PR:
xpuPackages_2025_3: init at 2025.3.1rocmPackages_7_0: 7.0.1 -> 7.0.2rocmPackages_7_1: init at 7.1.1