Sync some features with anaconda's recipe

recipe/meta.yaml

@@ -66,6 +66,7 @@ source:
     - patches/0014-Fix-FindOpenBLAS.patch
     # backport https://github.com/pytorch/pytorch/pull/138095
     - patches/0015-CD-Enable-Python-3.13-on-windows-138095.patch
+    - patches/0016-use-prefix-include-for-inductor.patch
 
 build:
   number: {{ build }}
@@ -381,6 +382,15 @@ outputs:
         - python -c "import torch; print(torch.__version__)"
         - python -c "import torch; assert torch.backends.mkldnn.m.is_available()"  # [x86 and cuda_compiler_version == "None"]
         - python -c "import torch; torch.tensor(1).to('cpu').numpy(); print('numpy support enabled!!!')"
+        # We have had issues with openmp .dylibs being doubly loaded in certain cases. These two tests catch the (observed) issue
+        - python -c "import torch; import numpy"
+        - python -c "import numpy; import torch"
+        # distributed support is enabled by default on linux; for mac, we enable it manually in build.sh
+        - python -c "import torch; assert torch.distributed.is_available()"        # [linux or osx]
+        - python -c "import torch; assert torch.backends.cuda.is_built()"          # [(cuda_compiler_version != "None")]
+        - python -c "import torch; assert torch.backends.cudnn.is_available()"     # [(cuda_compiler_version != "None")]
+        - python -c "import torch; assert torch.cuda.is_available()"               # [(cuda_compiler_version != "None")]
+        - python -c "import torch; assert torch.backends.cudnn.enabled"            # [(cuda_compiler_version != "None")]
         # At conda-forge, we target versions of OSX that are too old for MPS support
         # But if users install a newer version of OSX, they will have MPS support
         # https://github.com/conda-forge/pytorch-cpu-feedstock/pull/123#issuecomment-1186355073
@@ -424,6 +434,8 @@ outputs:
             "test/test_nn.py",
             "test/test_torch.py",
             "test/test_xnnpack_integration.py",
+            # The inductor tests test torch.compile
+            "test/inductor/test_torchinductor.py",                                      # [cuda_compiler_version != "None"]
             "-m \"not hypothesis\"",
         ]) %}
 

recipe/patches/0016-use-prefix-include-for-inductor.patch

@@ -0,0 +1,16 @@
+Currently inductor doesn't look in conda's includes and libs. This results in errors when it tries to compile, if system versions are being used of dependencies (e.g., sleef).
+
+author: [email protected]
+
+Index: pytorch/torch/_inductor/cpp_builder.py
+===================================================================
+--- pytorch.orig/torch/_inductor/cpp_builder.py	2024-12-16 15:16:47.074821258 -0600
++++ pytorch/torch/_inductor/cpp_builder.py	2024-12-16 15:17:33.922130106 -0600
+@@ -1055,6 +1055,7 @@
+         + python_include_dirs
+         + torch_include_dirs
+         + omp_include_dir_paths
++        + [os.getenv('CONDA_PREFIX') + '/include']
+     )
+     cflags = sys_libs_cflags + omp_cflags
+     ldflags = omp_ldflags