Author, debug, and optimize GPU kernels with Triton in Python.
npx skills add https://github.com/Shekswess/gpu-kernel-skills --skill triton
Please help me install this Agent Skill. Skill: triton Source: https://github.com/Shekswess/gpu-kernel-skills/tree/main/triton Command: npx skills add https://github.com/Shekswess/gpu-kernel-skills --skill triton
Triton enables developers to author, debug, and optimize GPU kernels for deep learning workloads in Python, simplifying kernel-level design and performance tuning.
Launch a tiny Triton kernel, e.g., a vector add, on small tensors to verify the environment and compile success.
High-intent search queries and answers about installing and using this skill.
Triton enables GPU kernel authoring for matmul or softmax using block-level SPMD programming with tl.program_id and tl.load/tl.store operations. You can achieve FP32 accumulation in tl.dot for deep learning workloads.
The best way to optimize GPU kernel performance across different hardware is using Triton's autotuning configurations. This allows cross-GPU performance optimization by automatically selecting the best block sizes and configurations for operations like GEMM and fused-ops.
Yes, you can integrate Triton kernels with PyTorch's torch.compile. The skill supports porting kernels to torch.compile, allowing block-level SPMD programming and Triton API integration to coexist within your PyTorch deep learning workflow.
Yes, Triton supports creating fused kernels for attention and layer norm operations. Developers can author, debug, and optimize fused GPU kernels using reference patterns for GEMM, fused-ops, and attention mechanisms in Python.
You need block-level SPMD programming when writing GPU kernels because it simplifies kernel-level design and performance tuning. Triton uses tl.program_id for block-level programming, enabling developers to author deep learning kernels in Python without low-level CUDA C.
sayalinvidiaInstall cuPyNumeric via conda or pip and verify in an isolated environment.
mindspore-aiImplement Triton-CUDA attention kernels with online Softmax for transformer workloads.
yo-stevenCreate verifiable Triton GPU kernels for deep learning compute patterns.
m0atDevelop, debug, and autotune Triton and Gluon GPU kernels for ML workloads.
