pytorch: 0.2.0 → 0.3.1 with CUDA and cuDNN (#38530)

* pytorch-0.3 with optional cuda and cudnn

* pytorch tests reenabled if compiling without cuda

* pytorch: Conditionalize cudnn dependency on cudaSupport

Signed-off-by: Anders Kaseorg <andersk@mit.edu>

* pytorch: Compile with the same GCC version used by CUDA if cudaSupport

Fixes this error:

In file included from /nix/store/gv7w3c71jg627cpcff04yi6kwzpzjyap-cudatoolkit-9.1.85.1/include/host_config.h:50:0,
                 from /nix/store/gv7w3c71jg627cpcff04yi6kwzpzjyap-cudatoolkit-9.1.85.1/include/cuda_runtime.h:78,
                 from <command-line>:0:
/nix/store/gv7w3c71jg627cpcff04yi6kwzpzjyap-cudatoolkit-9.1.85.1/include/crt/host_config.h:121:2: error: #error -- unsupported GNU version! gcc versions later than 6 are not supported!
 #error -- unsupported GNU version! gcc versions later than 6 are not supported!
  ^~~~~

Signed-off-by: Anders Kaseorg <andersk@mit.edu>

* pytorch: Build with joined cudatoolkit

Similar to #30058 for TensorFlow.

Signed-off-by: Anders Kaseorg <andersk@mit.edu>

* pytorch: 0.3.0 -> 0.3.1

Signed-off-by: Anders Kaseorg <andersk@mit.edu>

* pytorch: Patch for “refcounted file mapping not supported” failure

Signed-off-by: Anders Kaseorg <andersk@mit.edu>

* pytorch: Skip distributed tests

Signed-off-by: Anders Kaseorg <andersk@mit.edu>

* pytorch: Use the stub libcuda.so from cudatoolkit for running tests

Signed-off-by: Anders Kaseorg <andersk@mit.edu>

pkgs/development/python-modules/pytorch/default.nix

@@ -1,8 +1,31 @@
-{ buildPythonPackage, fetchFromGitHub, lib, numpy, pyyaml, cffi, cmake,
+{ buildPythonPackage,
-  git, stdenv }:
+  cudaSupport ? false, cudatoolkit ? null, cudnn ? null,
+  fetchFromGitHub, fetchpatch, lib, numpy, pyyaml, cffi, cmake,
+  git, stdenv, linkFarm, symlinkJoin,
+  utillinux, which }:
 
-buildPythonPackage rec {
+assert cudnn == null || cudatoolkit != null;
-  version = "0.2.0";
+assert !cudaSupport || cudatoolkit != null;
+
+let
+  cudatoolkit_joined = symlinkJoin {
+    name = "${cudatoolkit.name}-unsplit";
+    paths = [ cudatoolkit.out cudatoolkit.lib ];
+  };
+
+  # Normally libcuda.so.1 is provided at runtime by nvidia-x11 via
+  # LD_LIBRARY_PATH=/run/opengl-driver/lib.  We only use the stub
+  # libcuda.so from cudatoolkit for running tests, so that we don’t have
+  # to recompile pytorch on every update to nvidia-x11 or the kernel.
+  cudaStub = linkFarm "cuda-stub" [{
+    name = "libcuda.so.1";
+    path = "${cudatoolkit}/lib/stubs/libcuda.so";
+  }];
+  cudaStubEnv = lib.optionalString cudaSupport
+    "LD_LIBRARY_PATH=${cudaStub}\${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH} ";
+
+in buildPythonPackage rec {
+  version = "0.3.1";
   pname = "pytorch";
   name = "${pname}-${version}";
 
@@ -10,18 +33,41 @@ buildPythonPackage rec {
     owner  = "pytorch";
     repo   = "pytorch";
     rev    = "v${version}";
-    sha256 = "1s3f46ga1f4lfrcj3lpvvhgkdr1pi8i2hjd9xj9qiz3a9vh2sj4n";
+    fetchSubmodules = true;
+    sha256 = "1k8fr97v5pf7rni5cr2pi21ixc3pdj3h3lkz28njbjbgkndh7mr3";
   };
 
-  checkPhase = ''
+  patches = [
-    ${stdenv.shell} test/run_test.sh
+    (fetchpatch {
+      # make sure stdatomic.h is included when checking for ATOMIC_INT_LOCK_FREE
+      # Fixes this test failure:
+      # RuntimeError: refcounted file mapping not supported on your system at /tmp/nix-build-python3.6-pytorch-0.3.0.drv-0/source/torch/lib/TH/THAllocator.c:525
+      url = "https://github.com/pytorch/pytorch/commit/502aaf39cf4a878f9e4f849e5f409573aa598aa9.patch";
+      stripLen = 3;
+      extraPrefix = "torch/lib/";
+      sha256 = "1miz4lhy3razjwcmhxqa4xmlcmhm65lqyin1czqczj8g16d3f62f";
+    })
+  ];
+
+  postPatch = ''
+    substituteInPlace test/run_test.sh --replace \
+      "INIT_METHOD='file://'\$TEMP_DIR'/shared_init_file' \$PYCMD ./test_distributed.py" \
+      "echo Skipped for Nix package"
+  '';
+
+  preConfigure = lib.optionalString cudaSupport ''
+    export CC=${cudatoolkit.cc}/bin/gcc
+  '' + lib.optionalString (cudaSupport && cudnn != null) ''
+    export CUDNN_INCLUDE_DIR=${cudnn}/include
   '';
 
   buildInputs = [
      cmake
      git
      numpy.blas
-  ];
+     utillinux
+     which
+  ] ++ lib.optionals cudaSupport [cudatoolkit_joined cudnn];
 
   propagatedBuildInputs = [
     cffi
@@ -29,8 +75,8 @@ buildPythonPackage rec {
     pyyaml
   ];
 
-  preConfigure = ''
+  checkPhase = ''
-    export NO_CUDA=1
+    ${cudaStubEnv}${stdenv.shell} test/run_test.sh
   '';
 
   meta = {

pkgs/top-level/python-packages.nix

@@ -5623,7 +5623,17 @@ in {
     };
   };
 
-  pytorch = callPackage ../development/python-modules/pytorch { };
+  pytorch = callPackage ../development/python-modules/pytorch {
+    cudaSupport = pkgs.config.cudaSupport or false;
+  };
+
+  pytorchWithCuda = self.pytorch.override {
+    cudaSupport = true;
+  };
+
+  pytorchWithoutCuda = self.pytorch.override {
+    cudaSupport = false;
+  };
 
   python2-pythondialog = buildPythonPackage rec {
     name = "python2-pythondialog-${version}";