Fix a FIFO correctness bug (#549)

* Add a FIFO test code that reproduced a correctness issue
* Fix the correctness issue by using pinned memory instead of cudaMemcpy

---------

Co-authored-by: Binyang Li <binyli@microsoft.com>
This commit is contained in:
Changho Hwang
2025-07-11 16:53:59 -07:00
committed by GitHub
parent 9b71d524b3
commit 20eca28942
18 changed files with 96 additions and 155 deletions

View File

@@ -51,7 +51,7 @@ class MyProxyService {
semaphores_[nghr]->signal();
connections_[nghr]->flush();
}
return mscclpp::ProxyHandlerResult::FlushFifoTailAndContinue;
return mscclpp::ProxyHandlerResult::Continue;
}
void start() { proxy_.start(); }