Custom cuda version usage

[the-code-samurai-97]

My Laptop uses cuda 11.7 and i want to use Cuda 12 , Is there a way to use cuda 12.x without local installing can i download tar.gz using bazel

[cloudhan]

Sorry, I don't quite get you language. I am guessing that you want to extract the cuda toolkit to a non-standard path (that is, without instlalling), Then you can use call rules_cuda_dependencies(toolkit_path="<your_custom_cuda_path>") in your WORKSPACE file or you can use --repo_env=CUDA_PATH=<your_custom_cuda_path>

See https://github.com/bazel-contrib/rules_cuda/blob/5633f0c0f7/cuda/private/repositories.bzl#L31-L35

[sibi-venti]

can i download the cuda and compile using bazel?

[cloudhan]

@sibi-venti

Previous method change where the cuda_path point to, then in

rules_cuda/cuda/private/repositories.bzl

Lines 28 to 82 in 1a2ec3d

	def detect_cuda_toolkit(repository_ctx):
	"""Detect CUDA Toolkit.
	The path to CUDA Toolkit is determined as:
	- the value of `toolkit_path` passed to local_cuda as an attribute
	- taken from `CUDA_PATH` environment variable or
	- determined through 'which ptxas' or
	- defaults to '/usr/local/cuda'
	Args:
	repository_ctx: repository_ctx
	Returns:
	A struct contains the information of CUDA Toolkit.
	"""
	cuda_path = repository_ctx.attr.toolkit_path
	if cuda_path == "":
	cuda_path = repository_ctx.os.environ.get("CUDA_PATH", None)
	if cuda_path == None:
	ptxas_path = repository_ctx.which("ptxas")
	if ptxas_path:
	# ${CUDA_PATH}/bin/ptxas
	# Some distributions instead put CUDA binaries in a seperate path
	# Manually check and redirect there when necessary
	alternative = repository_ctx.path("/usr/lib/nvidia-cuda-toolkit/bin/nvcc")
	if str(ptxas_path) == "/usr/bin/ptxas" and alternative.exists:
	ptxas_path = alternative
	cuda_path = str(ptxas_path.dirname.dirname)
	if cuda_path == None and _is_linux(repository_ctx):
	cuda_path = "/usr/local/cuda"
	if cuda_path != None and not repository_ctx.path(cuda_path).exists:
	cuda_path = None
	bin_ext = ".exe" if _is_windows(repository_ctx) else ""
	nvlink = "@rules_cuda//cuda/dummy:nvlink"
	link_stub = "@rules_cuda//cuda/dummy:link.stub"
	bin2c = "@rules_cuda//cuda/dummy:bin2c"
	fatbinary = "@rules_cuda//cuda/dummy:fatbinary"
	if cuda_path != None:
	if repository_ctx.path(cuda_path + "/bin/nvlink" + bin_ext).exists:
	nvlink = str(Label("@local_cuda//:cuda/bin/nvlink{}".format(bin_ext)))
	if repository_ctx.path(cuda_path + "/bin/crt/link.stub").exists:
	link_stub = str(Label("@local_cuda//:cuda/bin/crt/link.stub"))
	if repository_ctx.path(cuda_path + "/bin/bin2c" + bin_ext).exists:
	bin2c = str(Label("@local_cuda//:cuda/bin/bin2c{}".format(bin_ext)))
	if repository_ctx.path(cuda_path + "/bin/fatbinary" + bin_ext).exists:
	fatbinary = str(Label("@local_cuda//:cuda/bin/fatbinary{}".format(bin_ext)))
	nvcc_version_major = -1
	nvcc_version_minor = -1
	if cuda_path != None:
	nvcc_version_major, nvcc_version_minor = _get_nvcc_version(repository_ctx, cuda_path)

_get_nvcc_version runs the nvcc to extract the version for your.

So the answer is yes and no. bazel does not compile, instead, it drives the compilation.

[cloudhan]

So by override the default cuda toolkit search path, and still relying the detecting machenism, you don't need do anything. Otherwise it should be treated as a bug.