[BUG]: cub/detail/detect_cuda_runtime.cuh should not include cuda_runtime_api.h when compiled by NVRTC

[leofang]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this bug and that I agree to the Code of Conduct

Type of Bug

Compile-time Error

Component

CUB

Describe the bug

CuPy is in the process of reducing its Jitify usage to reduce JIT overhead and moving closer to CCCL head (cupy/cupy#8412), and this long-time issue gets exposed.

The header cuda_runtime_api.h is currently included unconditionally by CUB:

cccl/cub/cub/detail/detect_cuda_runtime.cuh

Line 47 in 8a185fe

#include <cuda_runtime_api.h>

However it should not be. In Python it is possible that CUB & CuPy are deployed without the presence of the full set of CUDA headers (since they are not redistributable), and so when JIT'ing CUB headers using NVRTC a compile-time "file not found" error would be raised.

With Jitify, it defaults to commenting out headers that are not found and retrying the compilation, therefore this issue has been guarded for a long time until now.

How to Reproduce

n/a

Expected behavior

Mimicking Jitify behavior and making the CUDA headers optional for NVRTC.

Reproduction link

No response

Operating System

No response

nvidia-smi output

No response

NVCC version

No response

[leofang]

FYI this simple patch did the job for CuPy:

diff --git a/cub/cub/detail/detect_cuda_runtime.cuh b/cub/cub/detail/detect_cuda_runtime.cuh
index 44ee81119..2c2398b56 100644
--- a/cub/cub/detail/detect_cuda_runtime.cuh
+++ b/cub/cub/detail/detect_cuda_runtime.cuh
@@ -44,7 +44,10 @@
 #  pragma system_header
 #endif // no system header
 
+// CUDA headers might not be present when using NVRTC, see NVIDIA/cccl#2095 for detail
+#ifndef __CUDACC_RTC__
 #include <cuda_runtime_api.h>
+#endif
 
 #ifdef DOXYGEN_SHOULD_SKIP_THIS // Only parse this during doxygen passes: