Triton
Also known as OpenAI Triton, Triton kernel language
In plain English
Triton is a Python-based language for writing custom accelerator kernels, letting ML engineers get near hand-tuned speed without writing low level code.
Technical definition
Triton is an open source kernel programming language and compiler that lets developers write high performance accelerator kernels in Python-like code, portable across vendors that maintain backends.
Engineering details
Writing peak performance kernels traditionally requires vendor specific expertise in CUDA or assembly level tuning. Triton raises the abstraction: the developer writes block level programs and the compiler handles memory coalescing, tiling, and scheduling. NVIDIA, AMD, and other vendors maintain backends, so one kernel source can target multiple architectures. Serving engines use it for fused operations, quantization paths, and MoE kernels where no vendor library fits. The name collides with NVIDIA Triton Inference Server, a separate model serving product.
Why it matters
Triton lowered the barrier between model researchers and hardware performance, and its cross vendor backends are strategically important because kernels written in it are not locked to one chip family. How well a vendor runs the Triton ecosystem has become part of its software story.
How to read it in InferenceX
Engines in InferenceX recipes ship substantial Triton kernel inventories alongside CUDA, CUTLASS, and AITER code, so compiler and backend maturity is one of the quiet forces behind curve movement between engine versions on both NVIDIA and AMD systems.
Source material
See the concept in real benchmarks
MI355X DeepSeek-V4-Pro on SGLang: 110.5x Throughput per GPU in 26 Days
The amd/deepseek_v4 side branch shipped TileLang attention indexer, Triton sparse MLA, fused RoPE/Hadamard, FlyDSL MoE, and FP4 weights across 31 performance optimizations PRs — lifting first-light 20 tok/s/GPU at 2.4 tok/s/user into 2,256 tok/s/GPU at 9.4 tok/s/user on 8K/1K, with both throughput and interactivity climbing together
InferenceX v2: NVIDIA Blackwell Vs AMD vs Hopper - Formerly InferenceMAX
GB300 NVL72, MI355X, B200, H100, Disaggregated Serving, Wide Expert Parallelism, Large Mixture of Experts, SGLang, vLLM, TRTLLM