don't depend on cuda-nvcc; use cuda-nvcc-impl to avoid pulling in GCC

[h-vetinari]

Fixes #37

[conda-forge-admin]

Hi! This is the friendly automated conda-forge-linting service.

I just wanted to let you know that I linted all conda-recipes in your PR (recipe/recipe.yaml) and found it was in an excellent condition.

[conda-forge-admin]

Hi! This is the friendly conda-forge automerge bot!

I considered the following status checks when analyzing this PR:

• linter: passed
• azure: failed

Thus the PR was not passing and not merged.

[h-vetinari]

Hm, seems like we got some apparently new incompatibility:

 $PREFIX/targets/x86_64-linux/include/generated_cuda_meta.h:754:5: error: 'CUmemcpyAttributes' does not name a type; did you mean 'cudaMemcpyAttributes'?
 │ │     754 |     CUmemcpyAttributes *attrs;
 │ │         |     ^~~~~~~~~~~~~~~~~~
 │ │         |     cudaMemcpyAttributes

[jakirkham]

Thanks Axel! 🙏

Have a couple questions below

[h-vetinari]

This is ready except for dealing with CUDA 12.8. I'm waiting for a response in triton-lang/triton#5737 (even though that's for the upcoming 3.2, the same principle would apply to 3.1 as well if my proposed backport is deemed OK)

Otherwise we can add some <12.8 constraint to unblock this.

[jakirkham]

Let's stick to CUDA 12.6 for the moment

As great as it would be to have CUDA 12.8 here, there is some prep work needed

Will bring this up for discussion later today

[h-vetinari]

Let's stick to CUDA 12.6 for the moment

As great as it would be to have CUDA 12.8 here, there is some prep work needed

Will bring this up for discussion later today

I'm not talking about building anything with 12.8. But currently triton does not have a runtime constraint on the CUDA version (and I'd like to keep it that way). However, that means that we end up pulling in cuda-cupti 12.8 into the test environment, and triton doesn't know which ptx format to use and fails.

It's IMO triton's bug to have an == 6 instead of >=6 in the following

triton-feedstock/recipe/patches/0009-CODEGEN-Support-CUDA-12.6-4588.patch

Lines 20 to 26 in 1cb58a3

	if major == 12:
	- return 80 + minor
	+ if minor < 6:
	+ return 80 + minor
	+ elif minor == 6:
	+ return 85
	if major == 11:

This is the change I'd much rather make, but I wanted to give a bit of time for feedback to arrive in triton-lang/triton#5737.

[jakirkham]

Ok for more context CUDA 12.8 adds 2 new architectures. So am thinking about how we roll that out

In any event will raise an issue to discuss and link it here

[h-vetinari]

Ok for more context CUDA 12.8 adds 2 new architectures.

I don't see how that matters, as long as builds don't ask to build for those architectures. The function I referenced literally says

     def ptx_get_version(cuda_version) -> int:
         '''
         Get the highest PTX version supported by the current CUDA driver.
         '''

Note "highest". My point is that using ptx 85 should be fine regardless of whether the toolchain is 12.6 or 12.8.