From a6d621176cbfd8e91f78cb4e915d2de553734b84 Mon Sep 17 00:00:00 2001
From: Wenkai Du <wenkai.du@amd.com>
Date: Thu, 11 Jun 2020 13:02:45 +0800
Subject: [PATCH] Sender rank's opCount maybe ahead by one if it finishes
 earlier

---
 src/collectives/device/primitives.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/collectives/device/primitives.h b/src/collectives/device/primitives.h
index 023004cb3e..ce5c83c3e9 100644
--- a/src/collectives/device/primitives.h
+++ b/src/collectives/device/primitives.h
@@ -113,7 +113,7 @@ class ncclPrimitives {
     if (mismatch) {
       // In non-LL, we use _threadfence_system before incrementing opCount, yet we are still waiting for credits here, so there must be a size mismatch
       STORE(comm->fatalDevError, ncclDevAssertedMismatch);
-    } else if (conn && LOAD(conn->opCountRem) > opCount) {
+    } else if (conn && LOAD(conn->opCountRem) > opCount+1) {
       mismatch += 1;
     }
   }