| 11 Nov 2025 |
 Daniel Fahey | That CPU has also got Intel AMX (Advanced Matrix Extensions)
We don't yet compile specifically for "XPU" in the vLLM derivation, but apparently would make CPU inference attractive. Maybe for another system without the H100s | 20:29:16 |
| Daniel Fahey | Next gen will have https://www.hwcooling.net/en/intel-and-amd-agree-on-future-of-x86-cpus-amx-and-ram-tagging/ (last time I checked) | 20:30:27 |
| Daniel Fahey | Hopefully NVIDIA see some disruption and all this gets a bit more accessible for us normal people | 20:30:58 |
 Ari Lotter | that's the dream :) | 20:31:05 |
| Daniel Fahey | And if you want to run Nixpkgs' vLLM on it, your colleagues will experience a slightly better CLI UX, lol | 20:40:40 |
| Ari Lotter | huh the jax checkPhase is sloooooow | 20:47:31 |
| Daniel Fahey | Famously | 20:53:41 |
| Daniel Fahey | Doesn't seem to work on Nixbuild.net either, lol https://github.com/NixOS/nixpkgs/issues/445824 | 20:55:16 |
| Daniel Fahey | Since it seems you're in the mood to watch computer go brrrrr, I use this command to monitor (and record stats) during compiles
nix-shell -p dool --run 'dool -tcms --swap --top-mem --top-cpu --display --output ~/vllm-build-stats.csv 10'
| 20:56:36 |
| Daniel Fahey | * Since it seems you're in the mood to watch computer go brrrrr, I use this command to monitor (and record stats) during compiles
nix-shell -p dool --run 'dool -tcms --swap --top-mem --top-cpu --display --output ~/vllm-build-stats.csv 10'
| 20:56:45 |
| Daniel Fahey | oh and if you're building JAX, that must mean you're not using Flox's new cache?
https://discourse.nixos.org/t/nix-flox-nvidia-opening-up-cuda-redistribution-on-nix/69189
extra-trusted-substituters = https://cache.flox.dev
extra-trusted-public-keys = flox-cache-public-1:7F4OyH7ZCnFhcze3fJdfyXYLQw/aV7GEed86nQ7IsOs=
| 20:58:20 |
| Daniel Fahey | * Oh, forgot, the command you'll want to post to the PR as a comment, use
nixpkgs-review pr 460701 --extra-nixpkgs-config '{ cudaSupport = true; allowUnfree = true;}' --post-result
``` thanks again!
| 20:59:12 |
| Daniel Fahey | * Oh, forgot, the command you'll want to post to the PR as a comment, use
nixpkgs-review pr 460701 --extra-nixpkgs-config '{ cudaSupport = true; allowUnfree = true;}' --post-result
```
thanks again!
| 20:59:23 |
| Daniel Fahey | * Oh, forgot, the command you'll want to post to the PR as a comment, use
nixpkgs-review pr 460701 --extra-nixpkgs-config '{ cudaSupport = true; allowUnfree = true;}' --post-result
thanks again!
| 20:59:34 |
| Ari Lotter | oh shoot, seems i'm not | 21:09:03 |
| Ari Lotter | gonna restart | 21:09:10 |
| Ari Lotter | lol | 21:09:11 |
| Daniel Fahey | Saves A LOT of time, absolute game changer | 21:09:28 |
| Ari Lotter | here we go again :) | 21:11:13 |
| Ari Lotter | nix-shell -p --run over nix run -- ? :p | 21:12:32 |
| Daniel Fahey | Old habits die hard! | 21:13:24 |
| Daniel Fahey | nix run nixpkgs#dool -- -tcms --swap --top-mem --top-cpu --display
nice
| 21:19:28 |
| Ari Lotter | lol it's still building jax for whatever reason | 21:27:49 |
| Ari Lotter | š«” | 21:27:53 |
| Daniel Fahey | master must be ahead of unstable-nixos (the target I think Flox build against) | 21:28:45 |
 GaƩtan Lepage | yes. | 21:47:41 |
| GaƩtan Lepage | We are following nixos-unstable-small on https://cache.nixos.org | 21:48:26 |
| Daniel Fahey | thanks, is that defined by https://github.com/NixOS/infra/blob/809088c05d04849e3660b22fa9e5bc895570c5fe/channels.nix#L32 ? | 22:16:35 |
| 12 Nov 2025 |
| Ari Lotter | holy moly. 2tb of ram was not enough for JAX. guess i ran too many jobs at once/ | 03:16:02 |
| Ari Lotter | seems to have perma-stalled at building python3.12-jax-0.8.0 (pytestCheckPhase): replacing crashed worker gw3 | 03:16:19 |
