| 8 Mar 2023 |
 connor (he/him) | Direct link to the project: https://github.com/orgs/NixOS/projects/27
It's a private project because that seems to be the default; I gave the CUDA maintainers team admin access so it could be changed to public if that's of interest.
I spent a little bit of time going through open issues on Nixpkgs which are at all related to CUDA and added them to the swim lane named "New." If any don't belong on this project (for example, because it falls outside the scope of this maintainer group), please feel free to remove it.
I haven't figured out a way to have new issues tagged with the CUDA topic automatically added to the project, but I figure this is a good start to get some visibility into what's currently open. | 00:57:41 |
 SomeoneSerge (matrix works sometimes) | connor (he/him): that's so great! | 00:59:52 |
| SomeoneSerge (matrix works sometimes) | Yes, can see it | 00:59:56 |
| SomeoneSerge (matrix works sometimes) | https://github.blog/changelog/2023-01-19-github-issues-january-19th-update/
Seems that we need this special "auto-add to project" workflow | 01:04:50 |
| connor (he/him) | Hopefully the "by assignee" and "by status" views will be helpful too (tabular instead of Kanban) | 01:05:19 |
| connor (he/him) | Damn enterprise-grade features ;( | 01:06:36 |
| connor (he/him) | * Damn the auto-add to project is an enterprise-grade feature ;( | 01:06:53 |
| SomeoneSerge (matrix works sometimes) | NixOS and GH do have some communication channel, I guess we could ask | 01:07:24 |
| SomeoneSerge (matrix works sometimes) | https://github.com/NixOS/nixos-foundation/issues/25#issuecomment-1459113224 | 01:13:27 |
| SomeoneSerge (matrix works sometimes) | We do have the automation we don't need xD https://github.com/NixOS/nixpkgs/issues/217913#event-fa1e4cc9-b97c-53b4-bf98-595b6708f0fa | 01:18:10 |
| connor (he/him) | 
Download Screenshot 2023-03-07 at 8.20.12 PM.png | 01:20:20 |
| connor (he/him) | We only have a few of them; the auto-add to project one would be nice to make sure everything which is tagged with CUDA makes its way into the project. These are the ones I see / have enabled so far: | 01:20:22 |
| connor (he/him) | Okay I turned on a few more (reopened items are sent to the backlog to be reprioritized; code changes do the same; code review approvals move it to Ready so it can be merged) | 01:23:00 |
| SomeoneSerge (matrix works sometimes) | I think we should prevent issues being closed when their tickets are just moved around on the board | 01:30:27 |
| SomeoneSerge (matrix works sometimes) | In reply to @ss:someonex.net
We do have the automation we don't need xD https://github.com/NixOS/nixpkgs/issues/217913#event-fa1e4cc9-b97c-53b4-bf98-595b6708f0fa We haven't merged https://github.com/NixOS/nixpkgs/pull/218035 yet | 01:31:01 |
| connor (he/him) | Ah, would you rather I have deleted that issue from the project? It was closed as a duplicate, so to me it's "done" in the sense that we no longer need to worry about it | 01:34:39 |
| SomeoneSerge (matrix works sometimes) |
was closed as a duplicate,
oh, ok
| 01:35:29 |
| SomeoneSerge (matrix works sometimes) | *
was closed as a duplicate,
oh, ok
| 01:35:33 |
| connor (he/him) | Also just learned that if you accidentally reset your commit and force push it turning your PR into an empty diff it's closed automatically | 01:35:55 |
| 9 Mar 2023 |
 hexa | https://opensource.googleblog.com/2023/03/openxla-is-ready-to-accelerate-and-simplify-ml-development.html?m=1 | 08:36:41 |
| SomeoneSerge (matrix works sometimes) | https://github.com/openxla/xla/issues/1 | 13:21:07 |
| connor (he/him) | Is there a centralized location for docs for CUDA-maintainer related stuff?
Containing answers to (for example):
- Do we have any infrastructure (like CI) besides cachix?
- What populates our cachix?
- What's the storage limit for our cachix (meaning, is the number of derivations we host a result of limited compute, storage, or both)?
- If it's not CI populating the cache, what's the process for getting permissions to push to it?
| 15:47:43 |
| SomeoneSerge (matrix works sometimes) | I'd CUDA page on NixOS Wiki should be one | 15:59:20 |
| SomeoneSerge (matrix works sometimes) | * I'd say CUDA page on NixOS Wiki should be one | 15:59:23 |
| SomeoneSerge (matrix works sometimes) | https://nixos.wiki/wiki/CUDA | 15:59:31 |
| SomeoneSerge (matrix works sometimes) | Hasn't been maintained for a while | 15:59:44 |
| connor (he/him) | Is there any sort of VCS or approval/review process for that? | 19:50:45 |
| 10 Mar 2023 |
| connor (he/him) | These results are preliminary but it looks like using -Xfatbin=-compress-always nearly cut the size of the magma NAR in half (from 429.4M to 233.6M) when building for just 8.6: https://github.com/NixOS/nixpkgs/pull/220402
That's super impressive, so I'm excited to see what it looks like when targeting multiple capabilities.
| 00:47:45 |
| SomeoneSerge (matrix works sometimes) | Just a wiki | 02:39:29 |
| connor (he/him) | Looks like PyTorch nightlies are using a much newer version of Triton relative to what they were a month or two ago. They’ve got their own branch which closely tracks master: https://github.com/openai/triton/tree/torch-inductor-stable.
Packaging that could be difficult given they pull in their own build of MLIR (based on LLVM 17: https://github.com/openai/triton/blob/2c32f4399986045ff25cae201ed3b16d922a9d3b/python/setup.py#L72) and unconditionally grab NVCC from conda (https://github.com/openai/triton/blob/2c32f4399986045ff25cae201ed3b16d922a9d3b/python/setup.py#L107).
We don’t have MLIR packaged yet: https://github.com/NixOS/nixpkgs/pull/163878 (although I think we do build it for ROCm?)
Thoughts? | 14:29:53 |
