Replace cudaGetDriverEntryPoint with cudaGetDriverEntryPointByVersion

[wujingyue]

The former will be deprecated

[wujingyue]

test_driver_api.cc already has a similar test

[wujingyue]

!test

[github-actions]

Description

• Replace cudaGetDriverEntryPoint with cudaGetDriverEntryPointByVersion

• Update driver API wrapper macros to include version parameter

• Remove deprecated CUDA driver API tests

---

Changes walkthrough 📝

	Relevant files
Enhancement	driver_api.cpp Update driver API wrapper with version parameter                  csrc/driver_api.cpp Update DEFINE_DRIVER_API_WRAPPER macro to include version parameter Use cudaGetDriverEntryPointByVersion instead of cudaGetDriverEntryPoint +28/-25  driver_api.h Update driver API declarations with version parameter        csrc/driver_api.h Update DECLARE_DRIVER_API_WRAPPER macro to include version parameter Specify version for each driver API in ALL_DRIVER_API_WRAPPER Add conditional compilation for CUDA_VERSION >= 12000 to include newer APIs +43/-22
Tests	test_multidevice_ipc.cpp Remove deprecated CUDA driver API tests                                    tests/cpp/test_multidevice_ipc.cpp Remove tests for deprecated CUDA driver APIs cuStreamWriteValue32 and cuStreamWaitValue32 +0/-21

---

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🧪 No relevant tests

⚡ Recommended focus areas for review Possible Issue The macro DEFINE_DRIVER_API_WRAPPER now requires a version argument, but the usage of this macro in the code does not provide this argument. This will likely cause a compilation error. decltype(::funcName)* funcName = \ Documentation The comment about the magic happening in the csrc/driver_api.h file could be improved for clarity, especially since the macro usage has changed. It should explain the new version argument and its significance. // How to lazily load a driver API and invoke it? Just forget about lazy loading // and write code as if you are using the driver API directly. Magic will // happen. To understand how the magic works, please refer to the cpp file's doc Removed Test The test StreamWriteValue32 has been removed. It should be verified if this test is no longer needed or if it should be replaced with a similar test that uses the new cudaGetDriverEntryPointByVersion function.

[xwang233]

cc @jjsjann123

The new api cudaGetDriverEntryPointByVersion doesn't exist in cuda 11.8, and this is causing nvfuser with cuda 11.8 build to fail. Since pytorch still has cuda 11.8 wheel, do we want to support nvfuser with cuda 11.8?

[wujingyue]

The new api cudaGetDriverEntryPointByVersion doesn't exist in cuda 11.8, and this is causing nvfuser with cuda 11.8 build to fail. Since pytorch still has cuda 11.8 wheel, do we want to support nvfuser with cuda 11.8?

Can you verify #4447?

[jjsjann123]

cc @jjsjann123

The new api cudaGetDriverEntryPointByVersion doesn't exist in cuda 11.8, and this is causing nvfuser with cuda 11.8 build to fail. Since pytorch still has cuda 11.8 wheel, do we want to support nvfuser with cuda 11.8?

out of curiosity, build against cuda variants is only done in nightly CI? github CI only does the latest cuda I assume?