LLM-Based Triton Kernel Optimization with Compiler-Grounded Diagnosis
A recent study published on arXiv (2607.23089) presents a hierarchical diagnosis framework grounded in compilers aimed at enhancing Triton kernels through large language models. Current optimization methods utilizing LLMs depend on superficial indicators such as compilation feedback and profiling metrics, which reveal performance issues but fail to clarify backend compiler shortcomings, particularly with new accelerators like NPUs. This innovative method treats kernel optimization as a layered diagnosis challenge, connecting runtime issues to intermediate representation (IR) structure and compiler actions prior to source code alterations. The system progresses from basic pattern assessment and profiling to more in-depth IR analysis and compiler-based evaluations as necessary, ultimately suggesting source-level modifications supported by evidence. It has been implemented on Triton for Ascend NPUs.
Key facts
- arXiv paper 2607.23089 proposes a compiler-grounded hierarchical diagnosis framework for Triton kernel optimization.
- Existing LLM-based approaches use surface signals like compilation feedback and profiling metrics.
- Surface signals reveal kernel slowness but not why the backend compiler fails to optimize.
- The framework targets emerging accelerators such as NPUs.
- Kernel optimization is formulated as a progressive cross-layer diagnosis problem.
- The system links runtime symptoms to IR structure and compiler behavior.
- It escalates from lightweight pattern triage to deeper analysis only when needed.
- Implementation is on Triton for Ascend NPUs.
Entities
Institutions
- arXiv