Use noinline and a few other fixes (#489)

* Use noinline and a few other fixes

* Tune collectives
This commit is contained in:
Wenkai Du
2022-01-11 16:51:06 -08:00
committed by GitHub
parent 7add135529
commit 6ff7690cb5
9 changed files with 36 additions and 39 deletions
+4 -4
View File
@@ -114,10 +114,10 @@ extern __device__ void NCCL_ONERANK_REDUCE_NAME(PreMulSum, double)(struct ncclWo
//#define ALLGATHER_CHUNKSTEPS (NCCL_STEPS/2)
//#define REDUCESCATTER_SLICESTEPS (NCCL_STEPS/4)
//#define REDUCESCATTER_CHUNKSTEPS (NCCL_STEPS/2)
#define ALLREDUCE_SLICESTEPS 1
#define ALLREDUCE_CHUNKSTEPS 1
#define ALLGATHER_SLICESTEPS 1
#define ALLGATHER_CHUNKSTEPS 1
#define ALLREDUCE_SLICESTEPS 2
#define ALLREDUCE_CHUNKSTEPS 4
#define ALLGATHER_SLICESTEPS 4
#define ALLGATHER_CHUNKSTEPS 4
#define REDUCESCATTER_SLICESTEPS 2
#define REDUCESCATTER_CHUNKSTEPS 4
#define BROADCAST_SLICESTEPS 1