Uninitialize the accumulation buffer to save some overhead (#27005)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/27005
Similar to https://github.com/pytorch/pytorch/pull/27002, we want to save some overhead.
ghstack-source-id: 91046563
Test Plan: CI
Differential Revision: D17641819
fbshipit-source-id: 9320919242a48f48532035e61d9844de671d39af