[FEATURE] Triton Skill

[L-Reichardt]

I believe a Triton skill would be complementary to the cuda-kernel skill, which would allow for more vendor independent development.
I do not currently have the expertise to implement this myself, so I am submitting this as a feature request to express interest and gauge demand.

[akshan-main]

I would be interested in working on this @sayakpaul , do you mind if I take this

[sayakpaul]

Go ahead. And tag @burtenshaw once done!

[akshan-main]

Thanks @sayakpaul. Before I start drafting, want to lock in scope since rocm-kernels already seems to be a Triton skill (AMD-flavored).

Proposed scope for triton-kernels (vendor-neutral):

• Triton DSL patterns, block/stage sizing, tl.constexpr idioms
• @triton.autotune configs that work across NVIDIA + AMD
• Numerics / dtype handling (fp16/bf16/fp8), masked loads, reductions
• Portability notes between CUDA and ROCm backends
• Integration with kernels hub (get_kernel), transformers, diffusers
• Benchmarking harness + correctness checks vs. a reference

Does that split make sense, or would you rather triton-kernels also cover the NVIDIA-Triton tuning that's currently missing (since cuda-kernels is raw-CUDA oriented)? Just want to pick one before drafting.

cc @burtenshaw for visibility on scope.