| 23 May 2024 |
 hexa | this is on release-24.05 fwiw | 23:07:06 |
| 24 May 2024 |
 aidalgol | SomeoneSerge (UTC+3): I didn't quite understand your "deploy this impurely" comment on the libXNVCtrl patch. | 18:53:47 |
| aidalgol | * SomeoneSerge (UTC+3): I didn't quite understand your "deploy this impurely" comment on the libXNVCtrl PR. | 18:53:49 |
 SomeoneSerge (matrix works sometimes) | In reply to @aidalgol:matrix.org
SomeoneSerge (UTC+3): I didn't quite understand your "deploy this impurely" comment on the libXNVCtrl PR. /run/opengl-driver instead of rpath | 18:56:55 |
| aidalgol | Ah, right. I'm not certain whether it needs to match the driver version or not. | 18:57:58 |
| SomeoneSerge (matrix works sometimes) | In reply to @aidalgol:matrix.org
Ah, right. I'm not certain whether it needs to match the driver version or not. Same. I suggest that we test | 19:00:03 |
 connor (he/him) | Okay, my absolute favorite new functionality in the PR I've been working on is being able to do this:
nix build --impure -L .#cudaPackages_11_8.pkgs.xgboost
and build xgboost with cudaPackages_11_8.
| 20:56:50 |
| connor (he/him) | It makes it so much easier to see if something is broken because of the CUDA version | 20:57:06 |
| connor (he/him) | Relevant change https://github.com/NixOS/nixpkgs/pull/306172/commits/98906659f42a774afc6d43bfae795cf9bbf4768b | 21:06:17 |
| SomeoneSerge (matrix works sometimes) | In reply to @connorbaker:matrix.org
Okay, my absolute favorite new functionality in the PR I've been working on is being able to do this:
nix build --impure -L .#cudaPackages_11_8.pkgs.xgboost
and build xgboost with cudaPackages_11_8.
This looks nice but also looks like a way to shoot yourself in the foot and/or cause infinite recursion | 21:06:22 |
| connor (he/him) | Has anyone had the opportunity to do anything with Intel Optane's enterprise drives? | 21:08:07 |
| connor (he/him) | Unrelated: has anyone encountered a similar error? Looks like xgboost doesn't build with CUDA 12.4.
xgboost> /build/source/src/common/quantile.cu(184): error: no instance of function template "cuda::std::__4::make_tuple" matches the argument list
xgboost> argument types are: (bool, size_t)
xgboost> return thrust::make_tuple<uint64_t, uint64_t>(!ind, ind);
xgboost> ^
xgboost> /nix/store/n5qpzxhxivwp6nxhlrf5synfxjzqqbnh-cuda12.4-cudatoolkit/include/cuda/std/detail/libcxx/include/tuple(1106): note #3326-D: function template "cuda::std::__4::make_tuple" does not match because argument #2 does not match parameter
xgboost> make_tuple(_Tp &&...__t) {
xgboost> ^
| 21:09:32 |
| connor (he/him) | :f https://gitlab.com/petsc/petsc/-/merge_requests/7354 | 21:11:03 |
| SomeoneSerge (matrix works sometimes) | !ind hmmhmm was that meant to be ~ind | 21:15:18 |
 Philip Taron (UTC-8) | In reply to @connorbaker:matrix.org
Has anyone had the opportunity to do anything with Intel Optane's enterprise drives? I have two of them I bought off ebay that I stuck in my machine. They're fast and high iop, but I haven't felt like they changed the game in any meaningful way. | 21:15:47 |
| connor (he/him) | I've seen a lot of random IO on my builders, so I was thinking about maybe getting one | 21:16:41 |
| Philip Taron (UTC-8) | They go for ~$55-75 USD each on ebay. | 21:20:01 |
| Philip Taron (UTC-8) | Not a lot of downsides if you have the spare M.2 space in your machines. | 21:20:58 |
| SomeoneSerge (matrix works sometimes) | Where do you mount these? | 21:21:50 |
| SomeoneSerge (matrix works sometimes) | As in: how are they more useful than tmpfs? | 21:22:26 |
| Philip Taron (UTC-8) | They're basically not, other than: stable storage and doesn't compete for RAM | 21:23:18 |
| SomeoneSerge (matrix works sometimes) | Ok how do you use them? | 21:23:52 |
| SomeoneSerge (matrix works sometimes) | * Ok how do you use them? | 21:24:06 |
| Philip Taron (UTC-8) | https://wiki.archlinux.org/title/Bcachefs -- yeah, I live on the edge 🙈
They're the "SSD" in front of my actual SSDs | 21:28:12 |
| Philip Taron (UTC-8) | Again, I have done no benchmarking and perceptually, there's no difference. | 21:28:50 |
| aidalgol | SomeoneSerge (UTC+3): Also, what should I do about backporting that PR since the automatic process for doing that failed? | 21:32:42 |
| connor (he/him) | SomeoneSerge (UTC+3): they're basically (slow) non-volatile RAM. Intel initially tried to position it in the datacenter as a tier between fast disks and RAM but it kind of flopped :l
the 4k random read/write is still the best on the market though | 21:33:19 |
| connor (he/him) | Philip Taron (UTC-8): I was looking at the p5800x, which unfortunately doesn't go for that amount :f | 21:35:05 |
| Philip Taron (UTC-8) | Oh yeah, those don't need the crazy software hacks to be made use of like the raw "here's a tiny bit of optane" ones do. | 21:36:21 |
| SomeoneSerge (matrix works sometimes) | In reply to @aidalgol:matrix.org
SomeoneSerge (UTC+3): Also, what should I do about backporting that PR since the automatic process for doing that failed? You can just follow the instructions posted by the bot and ping me for the merge | 21:39:28 |
