| 22 Mar 2023 |
 SomeoneSerge (matrix works sometimes) | Oh | 01:31:56 |
 connor (he/him) | Also, looks like we might have torch.compile soon, even if it is only with the binary distributions: https://github.com/NixOS/nixpkgs/pull/221652 | 01:32:02 |
| connor (he/him) | Someone S: do you have an example of how one would use the env attribute? I've seen it used for things like env.NIX_C_FLAGS, but not sure what it should look like for a bunch of values. Is it okay to just declare it as an attribute set? | 01:49:28 |
| SomeoneSerge (matrix works sometimes) | I still haven't read any documentation on env, but it sure is an attribute set. AFAIU there's little magic to no in Nix-the-interpreter. Magic mostly belongs in nixpkgs 🙃 | 03:00:45 |
| SomeoneSerge (matrix works sometimes) | Pfew. The triton part is an absolute frankenstein, but it builds
https://github.com/NixOS/nixpkgs/pull/222273 | 03:36:04 |
| SomeoneSerge (matrix works sometimes) | Honestly, I'm surprised pytorch migrated to triton and so early | 03:36:41 |
| SomeoneSerge (matrix works sometimes) | So far, I don't see a way to build openai triton without cuda support -> our rocm build is going to be unfree | 03:37:17 |
| SomeoneSerge (matrix works sometimes) | In reply to @connorbaker:matrix.org
Also, looks like we might have torch.compile soon, even if it is only with the binary distributions: https://github.com/NixOS/nixpkgs/pull/221652 Which reminds me, that we still have not fixed meta.license in torch-bin. It should be unfreeRedistributable | 03:38:43 |
| SomeoneSerge (matrix works sometimes) | In reply to @connorbaker:matrix.org
Also, looks like we might have torch.compile soon, even if it is only with the binary distributions: https://github.com/NixOS/nixpkgs/pull/221652 * Which reminds me, that we still have not fixed meta.license in torch-bin. I think it should be unfreeRedistributable | 03:39:00 |
| connor (he/him) | I've been meaning to make some issues for torch-bin as well (like moving to autoPatchelfHook and autoAddOpenGLRunpathHook) | 04:32:38 |
|  mjlbach changed their display name from mjlbach to mjlbach (Old). | 05:14:32 |
| mjlbach changed their display name from mjlbach (Old) to mjlbach. | 05:51:05 |
| connor (he/him) | Someone S: fantastic work on https://github.com/NixOS/nixpkgs/pull/222273 by the way, kudos for handling that! 👏 | 14:50:33 |
| SomeoneSerge (matrix works sometimes) | Damn, nixpkgs is an inspiring place | 15:51:22 |
| SomeoneSerge (matrix works sometimes) | Btw, I see we still haven't added any cuda to maintainers/team-list.nix.
I think ROCM people get notifications about e.g. pytorch through:
maintainers = with maintainers; [ acowley lovesegfault ] ++ teams.rocm.members;
| 15:51:23 |
| SomeoneSerge (matrix works sometimes) | * Btw, I see we still haven't added any cuda to maintainers/team-list.nix.
I think ROCM people get notifications about e.g. pytorch through:
maintainers = with maintainers; [ acowley lovesegfault ] ++ teams.rocm.members;
| 15:51:32 |
| SomeoneSerge (matrix works sometimes) | FRidh: you mentioned that "we can use wrapCC to make our own cudaPackages.stdenv"
I'm not familiar with its purpose, nor the purpose of the arguments it takes.
Would it make sense to expose cuda_nvcc via cc-wrapper's extraTools?
Would it make sense to expose cuda_cudart in extraPackages?
It is my assumption, that people could then use such cudaPackages.stdenv to build cuda apps without having to add cuda_nvcc to nativeBuildInputs.
Are there examples of a similar use of wrapCC?
Is there some extra work (maybe some automated checks) that people would expect from cudaPackages.stdenv based just on its name?
| 16:07:26 |
| SomeoneSerge (matrix works sometimes) | Sorry if I misquote, I lost the original comment | 16:07:55 |
| SomeoneSerge (matrix works sometimes) | * FRidh: you mentioned that "we can use wrapCC to make our own cudaPackages.stdenv"
I'm not familiar with its purpose, nor the purpose of the arguments it takes.
Would it make sense to expose cuda_nvcc via cc-wrapper's extraTools?
Would it make sense to expose cuda_cudart in extraPackages?
It is my assumption, that people could then use such cudaPackages.stdenv to build cuda apps without having to add cuda_nvcc to nativeBuildInputs.
Are there examples of a similar use of wrapCC?
Is there some extra work (maybe some automated checks or passthru attributes) that people would expect from cudaPackages.stdenv based just on its name?
| 16:09:36 |
| SomeoneSerge (matrix works sometimes) | * FRidh: you mentioned that "we can use wrapCC to make our own cudaPackages.stdenv"
I'm not familiar with its purpose, nor the purpose of the arguments it takes.
Would it make sense to expose cuda_nvcc via cc-wrapper's extraTools?
Would it make sense to expose cuda_cudart in extraPackages?
It is my assumption, that people could then use such cudaPackages.stdenv to build cuda apps without having to add cuda_nvcc to nativeBuildInputs.
Are there examples of a similar use of wrapCC?
Is there some extra work (maybe some automated checks or passthru attributes) that people would expect from cudaPackages.stdenv based just on its name? I'm maybe not 100% certain what a "stdenv" actually is
| 16:10:10 |
| SomeoneSerge (matrix works sometimes) | In reply to @ss:someonex.net
hexa: do we have an established... approach to handling torch.utils.cpp_extension? Omg, now they use it to figure out linker flags for torch.compile | 23:26:14 |
| SomeoneSerge (matrix works sometimes) | And yes, this requires a compatible gcc in the PATH too | 23:27:00 |
| SomeoneSerge (matrix works sometimes) | At run time | 23:27:08 |
| mjlbach | I am trying to get cuda support for distrobox as an escape hatch for nix | 23:38:47 |
| mjlbach | I got it working on fedora but now to try on nixos :) | 23:39:15 |
| SomeoneSerge (matrix works sometimes) | Huh, what's distrobox? | 23:39:26 |
| mjlbach | It's like toolbx, it's a wrapper around podman/docker | 23:39:37 |
| 23 Mar 2023 |
| SomeoneSerge (matrix works sometimes) | It would be nice to open an issue at openai/triton and offer some solution to stabilize their paths'n'flags discovery: the way they deal with ptxas, the way they infer flags to link the cuda driver, etc. I can see they're still changing the related bits, so maybe we could pitch something we ourselves would be more comfortable with
Only idk what would that be
| 23:05:34 |
| 24 Mar 2023 |
|  htran joined the room. | 02:21:03 |
| 25 Mar 2023 |
|  valconius joined the room. | 23:46:14 |
