[Hardware][Intel] Integrate CPU backend build procedure

[bigPYJ1151]

Hi, vLLM genius @WoosukKwon @simon-mo @zhuohan123 , we'd like to integrate our code from #2378 #1028 gradually to enable vLLM on Intel GPU/CPU devices.

In this PR, we made some changes to make the vLLM built for non-CUDA devices, using CPU as the example.

Changes to vLLM:

• Added an environment variable VLLM_TARGET_DEVICE to specify which vLLM build procedure will be triggered. By default, cuda is specified.
• Forward the non-CUDA extension build to isolated CMake scripts to avoid potential breaks.

Please take review and looking forward your comments.

[bigPYJ1151]

Hi @WoosukKwon , please take a look and let me know whether this change is acceptable, thanks!

[simon-mo]

@bigPYJ1151, we are going to migrate from setup.py to CMake. Can you take a look at #2830 and see how can Intel GPU/CPU fit in here?

[bigPYJ1151]

@bigPYJ1151, we are going to migrate from setup.py to CMake. Can you take a look at #2830 and see how can Intel GPU/CPU fit in here?

Hi @simon-mo , thanks for your comment! #2830 makes setup.py clearer and more easily to maintain. For Intel GPU/CPU, I think it is not difficult to fit in. We can still use BuildExtension from torch cpp_extension, because our build procedure is not very complex. Or we can provide another CMakeLists.txt based on the one in #2830 .

Which one do you prefer? And will #2830 be merged soon? So I can rebase this PR to it.

[bigPYJ1151]

Hi @simon-mo @WoosukKwon @zhuohan123, we have updated the CPU backend build procedure with the latest vLLM CMake build system, please take a look and let me know whether this change is acceptable, thanks!