diff --git a/aten/src/ATen/cudnn/Handles.cpp b/aten/src/ATen/cudnn/Handles.cpp
index 04a81d5a603..36f0beec55c 100644
--- a/aten/src/ATen/cudnn/Handles.cpp
+++ b/aten/src/ATen/cudnn/Handles.cpp
@@ -19,7 +19,16 @@ struct Handle {
   }
   ~Handle() {
     if (handle) {
+// this is because of something dumb in the ordering of
+// destruction. Sometimes atexit, the cuda context (or something)
+// would already be destroyed by the time this gets destroyed. It
+// happens in fbcode setting. @colesbury and I decided to not destroy
+// the handle as a workaround.
+//   - @soumith
+#ifdef NO_CUDNN_DESTROY_HANDLE
+#else
       cudnnDestroy(handle);
+#endif
     }
   }
 };
diff --git a/setup.py b/setup.py
index 8387f4d1a66..a6c8e939458 100644
--- a/setup.py
+++ b/setup.py
@@ -379,8 +379,8 @@ class build_deps(PytorchCommand):
 
         # Use copies instead of symbolic files.
         # Windows has very poor support for them.
-        sym_files = ['tools/shared/cwrap_common.py']
-        orig_files = ['aten/src/ATen/common_with_cwrap.py']
+        sym_files = ['tools/shared/cwrap_common.py', 'tools/shared/_utils_internal.py']
+        orig_files = ['aten/src/ATen/common_with_cwrap.py', 'torch/_utils_internal.py']
         for sym_file, orig_file in zip(sym_files, orig_files):
             if os.path.exists(sym_file):
                 os.remove(sym_file)
diff --git a/test/common.py b/test/common.py
index 859ca5aa98f..c87704fc259 100644
--- a/test/common.py
+++ b/test/common.py
@@ -27,6 +27,7 @@ import errno
 
 import torch
 import torch.cuda
+from torch._utils_internal import get_writable_path
 from torch._six import string_classes
 import torch.backends.cudnn
 import torch.backends.mkl
@@ -505,7 +506,7 @@ def download_file(url, binary=True):
         from urllib import request, error
 
     filename = os.path.basename(urlsplit(url)[2])
-    data_dir = os.path.join(os.path.dirname(__file__), 'data')
+    data_dir = get_writable_path(os.path.join(os.path.dirname(__file__), 'data'))
     path = os.path.join(data_dir, filename)
 
     if os.path.exists(path):
diff --git a/test/test_distributed.py b/test/test_distributed.py
index 543202a8076..5f50165f0b8 100644
--- a/test/test_distributed.py
+++ b/test/test_distributed.py
@@ -17,11 +17,12 @@ import torch.nn.functional as F
 from torch.autograd import Variable
 from common import TestCase
 
+from torch._utils_internal import TEST_MASTER_ADDR as MASTER_ADDR
+
 BACKEND = os.environ['BACKEND']
 TEMP_DIR = os.environ['TEMP_DIR']
 INIT_METHOD = os.getenv('INIT_METHOD', 'env://')
 MASTER_PORT = '29500'
-MASTER_ADDR = '127.0.0.1'
 
 DEFAULT_TIMEOUT = 15
 CUSTOMIZED_TIMEOUT = {'test_DistributedDataParallel': 25}
diff --git a/test/test_torch.py b/test/test_torch.py
index 4890907403e..e3020f53790 100644
--- a/test/test_torch.py
+++ b/test/test_torch.py
@@ -13,6 +13,7 @@ import unittest
 import warnings
 import pickle
 import gzip
+from torch._utils_internal import get_file_path, get_file_path_2
 from torch.utils.dlpack import from_dlpack, to_dlpack
 from torch._utils import _rebuild_tensor
 from itertools import product, combinations
@@ -6656,7 +6657,10 @@ class TestTorch(TestCase):
             return module
 
         with filecontext_lambda() as checkpoint:
-            fname = os.path.join(os.path.dirname(__file__), 'data/network1.py')
+            try:
+                fname = get_file_path_2(os.path.dirname(__file__), 'data', 'network1.py')
+            except IOError:
+                fname = get_file_path_2(os.path.dirname(__file__), 'data', 'network1.pyc')
             module = import_module(tmpmodule_name, fname)
             torch.save(module.Net(), checkpoint)
 
@@ -6669,7 +6673,10 @@ class TestTorch(TestCase):
                     self.assertEquals(len(w), 0)
 
             # Replace the module with different source
-            fname = os.path.join(os.path.dirname(__file__), 'data/network2.py')
+            try:
+                fname = get_file_path_2(os.path.dirname(__file__), 'data', 'network2.py')
+            except IOError:
+                fname = get_file_path_2(os.path.dirname(__file__), 'data', 'network2.pyc')
             module = import_module(tmpmodule_name, fname)
             checkpoint.seek(0)
             with warnings.catch_warnings(record=True) as w:
diff --git a/test/test_utils.py b/test/test_utils.py
index cde57e6f922..af936525627 100644
--- a/test/test_utils.py
+++ b/test/test_utils.py
@@ -530,6 +530,7 @@ class TestLuaReader(TestCase):
         return input, target.sub(1)
 
 
+@unittest.skipIf('SKIP_TEST_BOTTLENECK' in os.environ.keys(), 'SKIP_TEST_BOTTLENECK is set')
 class TestBottleneck(TestCase):
     def _run(self, command):
         """Returns (return-code, stdout, stderr)"""
diff --git a/tools/nnwrap/__init__.py b/tools/nnwrap/__init__.py
index 2a574f72992..d6457a58c48 100644
--- a/tools/nnwrap/__init__.py
+++ b/tools/nnwrap/__init__.py
@@ -1,5 +1 @@
-from .generate_wrappers import generate_wrappers, wrap_function
-try:
-    from .generate_wrappers import import_module
-except ImportError:
-    pass
+from .generate_wrappers import generate_wrappers, wrap_function, import_module
diff --git a/tools/nnwrap/generate_wrappers.py b/tools/nnwrap/generate_wrappers.py
index 1edbca928c1..db4caf6bef6 100644
--- a/tools/nnwrap/generate_wrappers.py
+++ b/tools/nnwrap/generate_wrappers.py
@@ -3,17 +3,16 @@ import sys
 from string import Template, ascii_lowercase
 from ..cwrap import cwrap
 from ..cwrap.plugins import NNExtension, NullableArguments, AutoGPU
+from ..shared import import_module
 
-BASE_PATH = os.path.realpath(os.path.join(__file__, '..', '..', '..'))
-WRAPPER_PATH = os.path.join(BASE_PATH, 'torch', 'csrc', 'nn')
-THNN_UTILS_PATH = os.path.join(BASE_PATH, 'torch', '_thnn', 'utils.py')
+from ..shared._utils_internal import get_file_path
 
+THNN_H_PATH = get_file_path('torch', 'lib', 'THNN.h')
+THCUNN_H_PATH = get_file_path('torch', 'lib', 'THCUNN.h')
 
-try:
-    from torch._thnn import utils as thnn_utils
-except ImportError:
-    from ..shared import import_module
-    thnn_utils = import_module('torch._thnn.utils', THNN_UTILS_PATH)
+THNN_UTILS_PATH = get_file_path('torch', '_thnn', 'utils.py')
+
+thnn_utils = import_module('torch._thnn.utils', THNN_UTILS_PATH)
 
 FUNCTION_TEMPLATE = Template("""\
 [[
@@ -105,7 +104,7 @@ def generate_wrappers(nn_root=None, install_dir=None, template_path=None):
 
 def wrap_nn(thnn_h_path, install_dir, template_path):
     wrapper = '#include <TH/TH.h>\n\n\n'
-    nn_functions = thnn_utils.parse_header(thnn_h_path or thnn_utils.THNN_H_PATH)
+    nn_functions = thnn_utils.parse_header(thnn_h_path or THNN_H_PATH)
     for fn in nn_functions:
         for t in ['Float', 'Double']:
             wrapper += wrap_function(fn.name, t, fn.arguments)
@@ -124,7 +123,7 @@ def wrap_nn(thnn_h_path, install_dir, template_path):
 def wrap_cunn(thcunn_h_path, install_dir, template_path):
     wrapper = '#include <TH/TH.h>\n'
     wrapper += '#include <THC/THC.h>\n\n\n'
-    cunn_functions = thnn_utils.parse_header(thcunn_h_path or thnn_utils.THCUNN_H_PATH)
+    cunn_functions = thnn_utils.parse_header(thcunn_h_path or THCUNN_H_PATH)
     for fn in cunn_functions:
         for t in ['CudaHalf', 'Cuda', 'CudaDouble']:
             wrapper += wrap_function(fn.name, t, fn.arguments)
diff --git a/tools/setup_helpers/generate_code.py b/tools/setup_helpers/generate_code.py
index c9779fb8697..d44e286d560 100644
--- a/tools/setup_helpers/generate_code.py
+++ b/tools/setup_helpers/generate_code.py
@@ -77,6 +77,7 @@ def generate_code(ninja_global=None,
     sys.path.insert(0, root)
     from tools.autograd.gen_autograd import gen_autograd
     from tools.jit.gen_jit_dispatch import gen_jit_dispatch
+
     from tools.nnwrap import generate_wrappers as generate_nn_wrappers
 
     # Build THNN/THCUNN.cwrap and then THNN/THCUNN.cpp. These are primarily
diff --git a/torch/CMakeLists.txt b/torch/CMakeLists.txt
index 0c5bcd3a2ca..ff26a2967f2 100644
--- a/torch/CMakeLists.txt
+++ b/torch/CMakeLists.txt
@@ -137,6 +137,10 @@ configure_file("${CAFFE2_PATH}/aten/src/ATen/common_with_cwrap.py"
                "${TOOLS_PATH}/shared/cwrap_common.py"
                COPYONLY)
 
+configure_file("${CAFFE2_PATH}/torch/_utils_internal.py"
+               "${TOOLS_PATH}/shared/_utils_internal.py"
+               COPYONLY)
+
 add_custom_command(
   OUTPUT
   "${TORCH_SRC_DIR}/csrc/nn/THNN.cpp"
diff --git a/torch/__init__.py b/torch/__init__.py
index 83b88c246ba..88706460d8e 100644
--- a/torch/__init__.py
+++ b/torch/__init__.py
@@ -8,9 +8,11 @@ It has a CUDA counterpart, that enables you to run your tensor computations
 on an NVIDIA GPU with compute capability >= 3.0.
 """
 
+import os
 import sys
 import platform
 from ._utils import _import_dotted_name
+from ._utils_internal import get_file_path, prepare_multiprocessing_environment
 from .version import __version__
 from ._six import string_classes as _string_classes
 
@@ -233,8 +235,8 @@ _tensor_classes = set()
 def manager_path():
     if platform.system() == 'Windows':
         return b""
-    import os
-    path = os.path.join(os.path.abspath(os.path.dirname(__file__)), 'lib', 'torch_shm_manager')
+    path = get_file_path('torch', 'lib', 'torch_shm_manager')
+    prepare_multiprocessing_environment(get_file_path('torch'))
     if not os.path.exists(path):
         raise RuntimeError("Unable to find torch_shm_manager at " + path)
     return path.encode('utf-8')
diff --git a/torch/_thnn/utils.py b/torch/_thnn/utils.py
index 84c93a81801..b3f09e46a4a 100644
--- a/torch/_thnn/utils.py
+++ b/torch/_thnn/utils.py
@@ -4,9 +4,15 @@ import importlib
 
 # in fbcode, this fails in some cases, but we don't need it, therefore the try-catch
 try:
-    THNN_H_PATH = os.path.join(os.path.dirname(__file__), '..', 'lib', 'THNN.h')
-    THCUNN_H_PATH = os.path.join(os.path.dirname(__file__), '..', 'lib', 'THCUNN.h')
-except Exception:
+    # when compiling a cffi extension, this works. When compiling
+    # torch itself, it doesn't work because the parent module can't
+    # yet be imported. However that's fine because we don't need it in
+    # that case.
+    from .._utils_internal import get_file_path
+
+    THNN_H_PATH = get_file_path('torch', 'lib', 'THNN.h')
+    THCUNN_H_PATH = get_file_path('torch', 'lib', 'THCUNN.h')
+except Exception as e:
     pass
 
 
diff --git a/torch/_utils_internal.py b/torch/_utils_internal.py
new file mode 100644
index 00000000000..611dc7827c9
--- /dev/null
+++ b/torch/_utils_internal.py
@@ -0,0 +1,32 @@
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import os
+
+# this arbitrary-looking assortment of functionality is provided here
+# to have a central place for overrideable behavior. The motivating
+# use is the FB build environment, where this source file is replaced
+# by an equivalent.
+
+if os.path.basename(os.path.dirname(__file__)) == 'shared':
+    torch_parent = os.path.dirname(os.path.dirname(os.path.dirname(__file__)))
+else:
+    torch_parent = os.path.dirname(os.path.dirname(__file__))
+
+
+def get_file_path(*path_components):
+    return os.path.join(torch_parent, *path_components)
+
+
+def get_file_path_2(*path_components):
+    return os.path.join(*path_components)
+
+
+def get_writable_path(path):
+    return path
+
+
+def prepare_multiprocessing_environment(path):
+    pass
+
+
+TEST_MASTER_ADDR = '127.0.0.1'