pythonPackages.tensorflow: Don't change the rpath to point to gcc4.9 #26213
Conversation
|
It was added in #19701. No idea why. |
|
@teh, do you recall? |
| @@ -106,7 +106,7 @@ buildPythonPackage rec { | |||
| postFixup = let | |||
| rpath = stdenv.lib.makeLibraryPath | |||
| (if cudaSupport then | |||
| [ gcc49.cc.lib zlib cudatoolkit cudnn | |||
| [ gcc.cc.lib zlib cudatoolkit cudnn | |||
There was a problem hiding this comment.
stdenv.cc.cc.lib should be more idiomatic.
| @@ -106,7 +106,7 @@ buildPythonPackage rec { | |||
| postFixup = let | |||
| rpath = stdenv.lib.makeLibraryPath | |||
| (if cudaSupport then | |||
| [ gcc49.cc.lib zlib cudatoolkit cudnn | |||
| [ gcc.cc.lib zlib cudatoolkit cudnn | |||
| linuxPackages.nvidia_x11 ] | |||
| else | |||
| [ gcc.cc.lib zlib ] | |||
There was a problem hiding this comment.
You can change it here as well.
|
Actually, I think I recall. Tensorflow requires the compiler during runtime. What should be done, is hardcoding the paths to the compiler. See also #18273 (comment) |
|
cuda's nvcc is tied to specific gcc versions (at least it was, maybe it's no longer?), that's why we needed to pass through a specific gcc library. Edit: See table here: http://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#axzz3x7mwvZrG |
|
@teh: The CUDA compatibility table seems to allow for many different versions of GCC, but it does not mention version 4.9 (which is in place before the PR). It seems to me that CUDA is in fact compatible with most GCC versions, so I'd say to go for the default GCC of nixpkgs until tensorflow break. (With the added benefit to make tensorflow compatible with the rest of nixpkgs --- before this PR tensorflow is eg. incompatible with matplotlib) Furthermore, if it is CUDA that requires a specifc GCC version, shouldn't the version-pinning happen in the CUDA package instead of tensorflow? @FRidh: I do not know how to hardcode the path the compiler. Also, it does not seem that tensorflow requires the GCC at runtime in the current version. LLVM is used, according to https://www.tensorflow.org/performance/xla/ |
When using cuda, the rpath was set to include GCC lib version 4.9. I am not sure what this was attempting to do, but an effect was to prevent certain python libraries to find the correct (newer) version of the std lib. Also avoid mentions of any specifc version in the propagatedBuildInputs
|
@dezgeg I've applied the changes as you suggested. In fact I've removed all explicit mentions of gcc. Let me know if the code is now idiomatic or not. :) |
When using cuda, the rpath was set to include GCC lib version 4.9.
I am not sure what this was attempting to do, but an effect was to
prevent certain python libraries to find the correct (newer) version
of the std lib.
Motivation for this change
Things done
(nix.useSandbox on NixOS,
or option
build-use-sandboxinnix.confon non-NixOS)
nix-shell -p nox --run "nox-review wip"./result/bin/)