ztxz16

Guide for adding Triton-backed CUDA operators to FastLLM. Use when modifying FastLLM CUDA op code to add, extend, debug, validate, or benchmark Triton-generated kernels through tools/fastllm_triton_server.py, src/devices/cuda/cudadevice.cpp, src/devices/cuda/fastllm-triton-cuda.cu, include/devices/cuda/fastllm-cuda.cuh, or related CMake wiring.

当用户要求提交代码、整理提交、准备 commit、拆分 commit、push，或指定提交与推送规范时使用。默认使用中文提交信息，将差异较大的改动拆分为多个提交；推送前先执行 fetch、stash、rebase、stash pop，再 push。