triton-kernels: add vendor-neutral Triton skill

[jaygala223]

Related issue

Closes #302

What does this PR do?

Adds a new triton-kernels skill to kernel-builder/skills/, providing vendor-neutral guidance for writing portable Triton kernels that run on NVIDIA and AMD GPUs without modification.

Motivation

The existing skills are either raw CUDA (cuda-kernels) or backend-specific Triton (rocm-kernels, xpu-kernels). There's no shared base covering portable Triton patterns that the backend-specific skills can build on. This was identified in #302 and the skills/cuda-kernels link in the original issue is now dead since skills moved to kernel-builder/skills/ in #456.

Changes

• Add kernel-builder/skills/triton-kernels/SKILL.md: core DSL patterns (program_id, masked loads, 2D tiling, reductions, tl.dot), autotune guidance, numerics (fp32 accumulation, safe softmax, dtype handling), reference kernel implementations (softmax, matmul, rmsnorm), benchmarking patterns, get_kernel integration, transformers patching, and common pitfalls table
• Add references/autotune-guide.md: deep dive on configs, key parameter, num_warps/num_stages selection, when NOT to autotune, practical config sets
• Add references/kernel-patterns.md: dropout (in-SRAM mask), fused add+rmsnorm, SiLU, SwiGLU, group-major PID ordering, pointer sliding
• Add references/benchmarking-guide.md: GB/s vs TFLOPS, roofline model, do_bench usage, when to expect gains over PyTorch
• Add scripts/benchmark_template.py: reusable benchmark harness with correctness check
• Add scripts/correctness_template.py: multi-shape, multi-dtype test suite template
• Add examples/fused_softmax.py: complete working kernel with test + benchmark
• Add manifest.txt and CHANGELOG.md

Testing

• The example kernel (examples/fused_softmax.py) passes correctness tests against torch.softmax on V100 with irregular shapes (1823x781) at atol=1e-3
• Benchmark template runs and produces GB/s measurements
• Verified the skill follows the same directory structure as rocm-kernels (SKILL.md, manifest.txt, CHANGELOG.md, references/, scripts/, examples/)

Checklist

• This PR is linked to an issue that was discussed and approved
• I have tested these changes locally
• New/changed functionality has test coverage
• LLM disclosure:
  • I did not use an LLM to create this PR.
  • I used an LLM for assistance while creating this PR.
  • This PR was mostly or completely generated by an LLM.

[github-actions]

Hi @jaygala223, thanks for your interest in contributing!

This project requires that pull request authors are vouched, and you are not in the list of vouched users.

This PR will be closed automatically. See https://github.com/huggingface/kernels/blob/main/CONTRIBUTING.md for more details.

[sayakpaul]

Can you provide an example kernel that you wrote with this skill and tested? Feel free to apply to be able to publish kernel repo types on the Hub.

[jaygala223]

Hi @sayakpaul I tested the skill by using it as context for an AI agent and asking it to write a LayerNorm kernel (which is not one of the reference kernels already in the skill). The agent followed the patterns from the SKILL.md: row-wise reduction, fp32 accumulation, masked loads, contiguity assert, num_warps heuristic.

Results on V100 (fp32, M=4096):

Correctness: All shapes pass against torch.nn.functional.layer_norm at atol=1e-3, including irregular shapes (1823x781), single element, and bf16.

Performance: ~1.45x faster than PyTorch LayerNorm across all tested N values:

   N    Triton    PyTorch
 256    497       365 GB/s
1024    727       548 GB/s
2048    771       534 GB/s

The generated kernel + test is here: examples/skill_test_layernorm.py (8d7da2e)

I also applied for kernel repo access on the Hub.

[sayakpaul]

What is your Hub username?

[jaygala223]

jaygala223

[sayakpaul]

You have access to publish the kernels as proper kernel repos on the Hub. Could we please see the kernels generated with this skill on the Hub?

[jaygala223]

Yes. Published the layernorm kernel generated using the skill as a kernel repo: https://huggingface.co/kernels/jaygala223/triton-layernorm

[sayakpaul]

Are the results in #675 (comment) for https://huggingface.co/kernels/jaygala223/triton-layernorm?

[jaygala223]

Yes they are from the same kernel

…

On Wed, 24 Jun, 2026, 19:35 Sayak Paul, ***@***.***> wrote: *sayakpaul* left a comment (huggingface/kernels#675) <#675 (comment)> Are the results in #675 (comment) <#675 (comment)> for https://huggingface.co/kernels/jaygala223/triton-layernorm? — Reply to this email directly, view it on GitHub <#675?email_source=notifications&email_token=ANS4OMRPRM3JADW74F3SVPD5BPN2PA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINZZGAYTCMZZGI42M4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-4790113929>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ANS4OMTYIT2M3U2LQPLLIQL5BPN2PAVCNFSNUABFKJSXA33TNF2G64TZHM4DSNRRGI3DINBTHNEXG43VMU5TINZTGIYDONZXGQ42C5QC> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/ANS4OMQF2UCAN5HOVUDG4NT5BPN2PA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINZZGAYTCMZZGI42M4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJKTGN5XXIZLSL5UW64Y> and Android <https://github.com/notifications/mobile/android/ANS4OMSKXG2ZRRNERZM3ALT5BPN2PA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINZZGAYTCMZZGI42M4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLTGN5XXIZLSL5QW4ZDSN5UWI>. Download it today! You are receiving this because you were mentioned.Message ID: ***@***.***>

[sayakpaul]

We should also add skill to our docs.

[jaygala223]

Hi @sayakpaul added triton-kernels to the supported skills list in docs/source/cli-skills.md.

[sayakpaul]

There should be a couple of changes skin to #614.

[jaygala223]

Okay done!

Added a tip and example kernel.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.