Look for nvcc in CUDA_HOME

bindings/torch/setup.py

@@ -80,26 +80,64 @@ def find_cl_path():
 cpp_standard = 14
 
 # Get CUDA version and make sure the targeted compute capability is compatible
-if os.system("nvcc --version") == 0:
-	nvcc_out = subprocess.check_output(["nvcc", "--version"]).decode()
-	cuda_version = re.search(r"release (\S+),", nvcc_out)
-
-	if cuda_version:
-		cuda_version = parse_version(cuda_version.group(1))
-		print(f"Detected CUDA version {cuda_version}")
-		if cuda_version >= parse_version("11.0"):
-			cpp_standard = 17
-
-		supported_compute_capabilities = [
-			cc for cc in compute_capabilities if cc >= min_supported_compute_capability(cuda_version) and cc <= max_supported_compute_capability(cuda_version)
-		]
-
-		if not supported_compute_capabilities:
-			supported_compute_capabilities = [max_supported_compute_capability(cuda_version)]
-
-		if supported_compute_capabilities != compute_capabilities:
-			print(f"WARNING: Compute capabilities {compute_capabilities} are not all supported by the installed CUDA version {cuda_version}. Targeting {supported_compute_capabilities} instead.")
-			compute_capabilities = supported_compute_capabilities
+def _maybe_find_nvcc():
+	# Try PATH first
+	maybe_nvcc = shutil.which("nvcc")
+
+	if maybe_nvcc is not None:
+		return maybe_nvcc
+
+	# Then try CUDA_HOME from torch (cpp_extension.CUDA_HOME is undocumented, which is why we only use
+	# it as a fallback)
+	try:
+		from torch.utils.cpp_extension import CUDA_HOME
+	except ImportError:
+		return None
+
+	if not CUDA_HOME:
+		return None
+
+	return os.path.join(CUDA_HOME, "bin", "nvcc")
+
+def _maybe_nvcc_version():
+	maybe_nvcc = _maybe_find_nvcc()
+
+	if maybe_nvcc is None:
+		return None
+
+	nvcc_version_result = subprocess.run(
+		[maybe_nvcc, "--version"],
+		text=True,
+		check=False,
+		stdout=subprocess.PIPE,
+	)
+
+	if nvcc_version_result.returncode != 0:
+		return None
+
+	cuda_version = re.search(r"release (\S+),", nvcc_version_result.stdout)
+
+	if not cuda_version:
+		return None
+
+	return parse_version(cuda_version.group(1))
+
+cuda_version = _maybe_nvcc_version()
+if cuda_version is not None:
+	print(f"Detected CUDA version {cuda_version}")
+	if cuda_version >= parse_version("11.0"):
+		cpp_standard = 17
+
+	supported_compute_capabilities = [
+		cc for cc in compute_capabilities if cc >= min_supported_compute_capability(cuda_version) and cc <= max_supported_compute_capability(cuda_version)
+	]
+
+	if not supported_compute_capabilities:
+		supported_compute_capabilities = [max_supported_compute_capability(cuda_version)]
+
+	if supported_compute_capabilities != compute_capabilities:
+		print(f"WARNING: Compute capabilities {compute_capabilities} are not all supported by the installed CUDA version {cuda_version}. Targeting {supported_compute_capabilities} instead.")
+		compute_capabilities = supported_compute_capabilities
 
 min_compute_capability = min(compute_capabilities)