[Bug Fix]Fix undefined symbol : cuDriverGetVersion error after updating CUTLASS

[HydraQYH]

I saw that #2108 updated CUTLASS.
After the CUTLASS update, FA3 encountered the following issues:

The reason is that new CUTLASS uses the cuDriverGetVersion API:
https://github.com/NVIDIA/cutlass/blame/853ad93d60b23b4f87bc46dfbc3c9ce757773ed7/include/cute/atom/copy_traits_sm90_tma.hpp#L1056
This symbol is located in libcuda.so, but this library is not linked. So I added a extra_link_args to fix this bug.
@tridao

[HydraQYH]

@tridao @drisspg @johnnynunez Hello, can you take a look?

[johnnynunez]

@tridao @drisspg @johnnynunez Hello, can you take a look?

@drisspg @tridao maybe is related with this issue also apart from tvm ffi

sgl-project/sglang#15182
@Qiaolin-Yu @Fridge003

[johnnynunez]

I saw that #2108 updated CUTLASS. After the CUTLASS update, FA3 encountered the following issues: The reason is that new CUTLASS uses the cuDriverGetVersion API: https://github.com/NVIDIA/cutlass/blame/853ad93d60b23b4f87bc46dfbc3c9ce757773ed7/include/cute/atom/copy_traits_sm90_tma.hpp#L1056 This symbol is located in libcuda.so, but this library is not linked. So I added a extra_link_args to fix this bug. @tridao

Is this occuring in fa cute dsl also? @HydraQYH flash-attention/cute/

[drisspg]

I would need to look deeper at this issue first, but we should not link against driver Apis. Cutlass has a mechanism for dynamic loading at runtime the unresolved symbols and if something new was added that broke this that is a bug in cutlass.

[johnnynunez]

thanks @drisspg reported to cutlass team

[johnnynunez]

I saw that #2108 updated CUTLASS. After the CUTLASS update, FA3 encountered the following issues: The reason is that new CUTLASS uses the cuDriverGetVersion API: https://github.com/NVIDIA/cutlass/blame/853ad93d60b23b4f87bc46dfbc3c9ce757773ed7/include/cute/atom/copy_traits_sm90_tma.hpp#L1056 This symbol is located in libcuda.so, but this library is not linked. So I added a extra_link_args to fix this bug. @tridao

A fix is coming in v4.4.0.dev @drisspg @tridao @HydraQYH
You can close it

[depaulmillz]

This PR to CUTLASS should fix the issue. It switches to the runtime API instead of the driver API - NVIDIA/cutlass#2928

[HydraQYH]

I saw that #2108 updated CUTLASS. After the CUTLASS update, FA3 encountered the following issues: The reason is that new CUTLASS uses the cuDriverGetVersion API: https://github.com/NVIDIA/cutlass/blame/853ad93d60b23b4f87bc46dfbc3c9ce757773ed7/include/cute/atom/copy_traits_sm90_tma.hpp#L1056 This symbol is located in libcuda.so, but this library is not linked. So I added a extra_link_args to fix this bug. @tridao

A fix is coming in v4.4.0.dev @drisspg @tridao @HydraQYH You can close it

@johnnynunez Ok. I will close this PR after @depaulmillz 's PR is merged and FA updates CUTLASS.

[HydraQYH]

Move to #2142