1 Aug 2024 |
yorickvp | I'll try to repro with torch on nixos-unstable | 15:57:23 |
yorickvp | yeah, ${python3.pkgs.torchWithCuda.lib}/lib/libtorch_cuda.so links to gcc-12.4.0-lib | 16:16:14 |
SomeoneSerge (utc+3) | Wow | 16:40:20 |
SomeoneSerge (utc+3) | This looks like a regression | 16:40:27 |
SomeoneSerge (utc+3) | Well the first obvious leak (the one we see in the wrapper) is https://github.com/NixOS/nixpkgs/blob/fc27807b85986bb26a8f28e590e01fae742e6b53/pkgs/build-support/cc-wrapper/default.nix#L596-L606 | 16:53:54 |
SomeoneSerge (utc+3) | Notably, cudaPackages.saxpy works fine at that commit | 16:54:12 |
SomeoneSerge (utc+3) | I'm running github:NixOS/nixpkgs/c66e984bda09e7230ea7b364e677c5ba4f0d36d0#opencv4.tests.no-libstdcxx-errors now (only defined for cudaSupport = true) | 16:54:41 |
SomeoneSerge (utc+3) | Going to take a while | 16:54:45 |
SomeoneSerge (utc+3) | But it might be the regression is somehow magically torch specific | 16:54:59 |
SomeoneSerge (utc+3) | No idea why https://github.com/NixOS/nixpkgs/blame/fc27807b85986bb26a8f28e590e01fae742e6b53/pkgs/build-support/cc-wrapper/default.nix#L605-L606 uses cc_solib honestly | 16:55:53 |
yorickvp | you know, I blame cmake | 17:00:59 |
yorickvp | * you know, I blame cmake :) | 17:01:03 |
yorickvp | looking at 36 megabytes of cmake logs, it obviously parses it out of some gcc output (together with the correct one, which it puts first in the path). I'm not sure what it does with it after | 17:02:50 |
SomeoneSerge (utc+3) | Waiting for opencv, but so far I'm leaning towards "maybe pytorch devs replaced some of the cmake logic with an unnecessary gcc -print-search-paths" | 17:06:46 |
yorickvp | I'm looking at https://github.com/Kitware/CMake/blob/master/Modules/CMakeParseImplicitLinkInfo.cmake | 17:08:01 |
SomeoneSerge (utc+3) | saxpy and opencv are built using cmake too | 17:08:34 |
SomeoneSerge (utc+3) | At least one of them has been shown to still work (whatever the cost) | 17:08:56 |
SomeoneSerge (utc+3) | gy skimage.transform skimage.util skimage.segmentation
python3-3.11.9-env> building '/nix/store/4rqjcjk4h2mnfwsbvcgf3igjnmpxhxwf-python3-3.11.9-env.drv'
python3-3.11.9-env> created 521 symlinks in user environment
opencv-4.9.0-libstdcxx-test> building '/nix/store/2gh11xabzlxbfgvydhcln0qbfiharw32-opencv-4.9.0-libstdcxx-test.drv'
┏━ Dependency Graph:
┃ ┌─ ✔ opencv-4.9.0 ⏱ 17m40s
┃ ┌─ ✔ python3.11-pillow-heif-0.16.0 ⏱ 2m0s
┃ ┌─ ✔ python3.11-imageio-2.34.2 ⏱ 11s
┃ ┌─ ✔ python3.11-scikit-image-0.22.0 ⏱ 1m37s
┃ ┌─ ✔ python3-3.11.9-env ⏱ 1s
┃ ✔ opencv-4.9.0-libstdcxx-test
┣━━━ Builds
┗━ ∑ ⏵ 0 │ ✔ 6 │ ⏸ 0 │ Finished at 17:11:37 after 21m35s
| 17:12:13 |
SomeoneSerge (utc+3) | So ugh at least opencv4's python extension must be linking the right libstdc++ | 17:13:11 |
SomeoneSerge (utc+3) | Hmm the last torch update was almost two months ago https://github.com/NixOS/nixpkgs/pull/317576 | 17:14:41 |
SomeoneSerge (utc+3) | * Hmm the last merged torch update was almost two months ago https://github.com/NixOS/nixpkgs/pull/317576 | 17:14:45 |
SomeoneSerge (utc+3) | yorickvp would you volunteer to run the bisection? 🫠 | 17:15:40 |
yorickvp | sure, do you have a known working commit? | 17:15:47 |
SomeoneSerge (utc+3) | Well, I got a workstation sat
Revision: b2852eb9365c6de48ffb0dc2c9562591f652242a
Last modified: 2024-06-27 16:44:53
Let me check if torch actually works there | 17:16:31 |
SomeoneSerge (utc+3) | ❯ nix-shell -p 'python3.withPackages (ps: [ ps.torch ])'
trace: warning: cudaPackages.autoAddDriverRunpath is deprecated, use pkgs.autoAddDriverRunpath instead
trace: warning: cudaPackages.autoFixElfFiles is deprecated, use pkgs.autoFixElfFiles instead
trace: warning: cudaPackages.autoAddOpenGLRunpathHook is deprecated, use pkgs.autoAddDriverRunpathHook instead
this derivation will be built:
/nix/store/qmmz2hxinp65zsprb3g92my7wqvbwncm-python3-3.11.9-env.drv
building '/nix/store/qmmz2hxinp65zsprb3g92my7wqvbwncm-python3-3.11.9-env.drv'...
created 516 symlinks in user environment
[WARN] - (starship::utils): Executing command "/home/ss/.nix-profile/bin/git" timed out.
[WARN] - (starship::utils): You can set command_timeout in your config to a higher value to allow longer-running commands to keep executing.
ss in 🌐 cs-338 in triton on openai-triton [$] via ❄️ impure (shell)
❯ python
Python 3.11.9 (main, Apr 2 2024, 08:25:04) [GCC 13.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> torch.cuda.is_available()
True
>>>
| 17:17:20 |
yorickvp | it probably works but still secretly links gcc-12.4.0, which isn't always fatal | 17:17:19 |
SomeoneSerge (utc+3) | No, it shouldn't | 17:17:33 |
SomeoneSerge (utc+3) | It definitely was the case that gcc-12 reference was not retained in the output path | 17:17:56 |
yorickvp | patchelf --print-rpath $(nix-build python3.pkgs.torch.lib)/lib/libtorch_cuda.so ? | 17:18:22 |
SomeoneSerge (utc+3) | Yup, it's 12.3 in this commit and it happens to work, but yet again, this is a regression | 17:19:46 |