viable/strict/1787738484: Re-enable expandable segments for CUDA Graph (#194767) (#194767)
- PyTorch 近 90 天出现 86 次
- PyTorch 近 90 天第 86 次发版
- 上一次：同一天稍早 · trunk/147aaef6de2917b07f797c344bb9ef7b26ec6135: [vllm hash update] update the pinned vllm hash (#194672)
发生了什么
Summary: Previously we temporarily disabled expandable segments for CUDA Graph due to some issues. Now, since the issues were fixed, we re-enable it Test Plan: baseline: f1125979746 QPS: 28.5k peak memory: 97.08% re-enabled: f1131430315 QPS: 28.7k peak memory: 89.28% Differential Revision: D117308651 Pull Request resolved: #194767 Approved by: https://github.com/Skylion007 , https://github.com/eqy
摘要按规则整理自下方来源原文