Skip to content

Commit

Permalink
Fix peristent mode comm->channels[c].workFifoSent counter
Browse files Browse the repository at this point in the history
we cannot increase comm->channels[c].workFifoSent when persistent
mode is on. Otherwise, comm->channels[c].workFifoSent can mismatch
ack
  • Loading branch information
KaimingOuyang committed Apr 15, 2024
1 parent e0c0a58 commit ee3d92b
Showing 1 changed file with 1 addition and 1 deletion.
2 changes: 1 addition & 1 deletion src/enqueue.cc
Original file line number Diff line number Diff line change
Expand Up @@ -1078,7 +1078,7 @@ static ncclResult_t uploadWork(struct ncclComm* comm, struct ncclKernelPlan* pla
// Tell channel to ack us back ix+1 indicating that all slots up to and
// including ix have been consumed.
q->work.header.doneAcks = ix+1;
comm->channels[c].workFifoSent = ix+1;
if (!persistent) comm->channels[c].workFifoSent = ix+1;
}
workHeap[ix & ixMask] = q->work; // C++ struct assignment
q = q->next;
Expand Down

0 comments on commit ee3d92b

Please sign in to comment.