| 4 Nov 2025 |
 SomeoneSerge (back on matrix) | It's frustrating | 15:08:03 |
| SomeoneSerge (back on matrix) | Elaborated on github, but here for redundancy: the reference in onnxruntime only appears when nvcc is propagated by all these cuda libs, https://github.com/NixOS/nixpkgs/pull/457424#issuecomment-3475736738 | 15:11:32 |
 Gaétan Lepage | TIL: python3Packages.torchWithRocm is apprently sensitive to config.cudaSupport. | 20:11:25 |
 Ari Lotter | ugh i wish we could compile packages with cudaCapabilities individually per-capability and merge them later, it's such a nightmare adding one new capability level and it causing a huge 8-hour recompile.. | 20:40:40 |
 connor (burnt/out) (UTC-8) | These aliases must die, they make my life so difficult | 21:45:22 |
| connor (burnt/out) (UTC-8) | Join the club
And it’s not even like we could do a mega-build in an intermediate derivation and then prune unused capabilities according to whatever the user requested because the amount of generated device code is so large linking will fail lmao | 21:46:17 |
| connor (burnt/out) (UTC-8) | Gaétan Lepage are any of SomeoneSerge (back on matrix)’s comments on https://github.com/NixOS/nixpkgs/pull/457803 actionable or is it good to merge? | 21:48:00 |
| connor (burnt/out) (UTC-8) | Also, would you mind reviewing https://github.com/NixOS/nixpkgs/pull/458619? | 21:48:09 |
 hacker1024 | This is most likely due to a dependency, but I will also point out that all torch variants are at the moment due to an unconditional version access
https://github.com/NixOS/nixpkgs/blob/b3d51a0365f6695e7dd5cdf3e180604530ed33b4/pkgs/development/python-modules/torch/source/default.nix#L458
| 21:48:19 |
| Gaétan Lepage | I quadruppled check.
Both commits of my PR are actually necessary to get a nvcc-free onnxruntime. | 21:48:42 |
| Gaétan Lepage | Let me change one comment to mention the bisection | 21:48:57 |
| Gaétan Lepage | connor (burnt/out) (UTC-8), I reviewed nccl-tests. Feel free to merge | 22:08:11 |
| Ari Lotter | i'm trying to fix this exact linker error right now 😭 trying to get flash-attn built for cuda capabilities 7.5 thru 12.0a, and i'm so stuck, and every rebuild with an attempted fix takes ~2 hours... any ideas? 😭 | 22:17:28 |
| Ari Lotter | maybe we're just screwed :) | 22:20:25 |
 Robbie Buxton | Which flash attention version | 22:24:21 |
| Robbie Buxton | V2 or v3 | 22:24:27 |
| Robbie Buxton | And from what got tag? | 22:24:51 |
| Robbie Buxton | * And from what git tag? | 22:24:59 |
| Ari Lotter | v2, from tag v2.8.2 | 22:29:50 |
| Robbie Buxton | I think there is currently a pr open in nixpkgs to add this, is that the one you’re building? | 22:30:41 |
| Ari Lotter | oh neat, no | 22:31:37 |
| Ari Lotter | let me compare my derivation with that one | 22:31:40 |
| Ari Lotter | ok yeah, decently similar. difference is i'm building against cutlass 4.0 instead of 4.1, and.. somehow my deps list is wayy simpler, yet the build works (on previous versions of my derivation, pre updating CUDA)? very strange.. | 22:35:13 |
| Ari Lotter | but yeah i just smash into
> build/lib.linux-x86_64-cpython-312/flash_attn_2_cuda.cpython-312-x86_64-linux-gnu.so: PC-relative offset overflow in PLT entry for `_ZNK3c1010TensorImpl4sizeEl'
``` 🤷
| 22:35:28 |
| Ari Lotter | i'm so tired of CUDA nightmares 😭 im so close to giving up and building dockerized devenvs, i just really don't want to give in..... :( | 22:37:57 |
| Gaétan Lepage | (It's a secret, but you might want to add https://cache.nixos-cuda.org as a substituter, it is slowly getting more and more artifacts)
Public key: cache.nixos-cuda.org:74DUi4Ye579gUqzH4ziL9IyiJBlDpMRn9MBN8oNan9M= | 22:44:02 |
| Gaétan Lepage | connor (burnt/out) (UTC-8), Serge and I got #457803 ready.
We are waiting for nixpkgs's CI to get fixed (https://github.com/NixOS/nixpkgs/pull/458647).
Let's merge ASAP | 23:38:07 |
| Robbie Buxton | For flash attention you should use the version of cutlass in the repo | 23:54:57 |
| Robbie Buxton | They have a hash | 23:55:06 |
| Robbie Buxton | In csrc/cutlass | 23:56:01 |
