allow packaging tools to override CUDA settings

Summary: This makes sure circle ci builds work with cuda even on machines with no gpu.

Reviewed By: gkioxari

Differential Revision: D19543957

fbshipit-source-id: 9cbfcd4fca22ebe89434ffa71c25d75dd18d2eb6

Author: bottler

setup.py

@@ -23,21 +23,31 @@ def get_extensions():
     extra_compile_args = {"cxx": ["-std=c++17"]}
     define_macros = []
 
-    if torch.cuda.is_available() and CUDA_HOME is not None:
+    force_cuda = os.getenv("FORCE_CUDA", "0") == "1"
+    if (torch.cuda.is_available() and CUDA_HOME is not None) or force_cuda:
         extension = CUDAExtension
         sources += source_cuda
         define_macros += [("WITH_CUDA", None)]
-        extra_compile_args["nvcc"] = [
+        nvcc_args = [
             "-DCUDA_HAS_FP16=1",
             "-D__CUDA_NO_HALF_OPERATORS__",
             "-D__CUDA_NO_HALF_CONVERSIONS__",
             "-D__CUDA_NO_HALF2_OPERATORS__",
         ]
+        nvcc_flags_env = os.getenv("NVCC_FLAGS", "")
+        if nvcc_flags_env != "":
+            nvcc_args.extend(nvcc_flags_env.split(" "))
 
         # It's better if pytorch can do this by default ..
         CC = os.environ.get("CC", None)
         if CC is not None:
-            extra_compile_args["nvcc"].append("-ccbin={}".format(CC))
+            CC_arg = "-ccbin={}".format(CC)
+            if CC_arg not in nvcc_args:
+                if any(arg.startswith("-ccbin") for arg in nvcc_args):
+                    raise ValueError("Inconsistent ccbins")
+                nvcc_args.append(CC_arg)
+
+        extra_compile_args["nvcc"] = nvcc_args
 
     sources = [os.path.join(extensions_dir, s) for s in sources]