Backend optimization#

Triton launch options and useful tile sizes depend on the device and compiler. Start with the architecture and compiler notes for your backend, then measure representative shapes with the FlagGems tuning workflow.

• Ascend: AI Core memory and compute paths, Triton-Ascend options, and NPU tuning.
• Hygon: DCU/HIP memory hierarchy, Triton launch options, and tile tuning.

Treat the configurations in these guides as candidates. Check numerical results and latency on the target device and installed compiler before adopting them.