Microsoft 與清華大學的研究分析 7 個 GPU kernel 函式庫的 300 筆已結案 GitHub issue,發現多數故障源自編譯器、建置系統或執行環境。
AI 伴隨問答
針對這篇報導有疑問?可直接點選快捷問題或向 AI 追問背景脈絡與技術細節:
New Microsoft and Tsinghua paper finds that most GPU kernel library failures in LLM systems come from compilers, builds, and environment mismatches, not kernel code.
LLM training and serving depend on GPU kernel libraries like FlashAttention, Triton, and CUTLASS. When they break, jobs can crash, hang, or silently produce wrong outputs, and the cause is hard to trace.
The team read 300 closed GitHub issue reports from 7 of these libraries, including the discussions and fixes. They labeled each one by what the user saw and what caused it.
Of the 300 issues, 60.33% traced to the compiler and build system or the runtime environment, and 39.67% to the kernel library. The largest single cause was software incompatibility among drivers, CUDA, frameworks, and packages, at 76 issues or 25.33%.