diff --git a/projects/rccl/src/include/collectives.h b/projects/rccl/src/include/collectives.h index d52d597063..518643cd14 100644 --- a/projects/rccl/src/include/collectives.h +++ b/projects/rccl/src/include/collectives.h @@ -123,11 +123,11 @@ extern __device__ void NCCL_ONERANK_REDUCE_NAME(PreMulSum, double)(struct ncclWo #define REDUCESCATTER_SLICESTEPS 2 #define REDUCESCATTER_CHUNKSTEPS 4 #define BROADCAST_SLICESTEPS 1 -#define BROADCAST_CHUNKSTEPS 2 +#define BROADCAST_CHUNKSTEPS 4 #define REDUCE_SLICESTEPS 1 -#define REDUCE_CHUNKSTEPS 2 +#define REDUCE_CHUNKSTEPS 4 #define SENDRECV_SLICEFACTOR 1 -#define NCCL_MAX_SLICE_PER_CHUNK 2 // max value for CHUNKSTEPS/SLICESTEPS, must accord with above +#define NCCL_MAX_SLICE_PER_CHUNK 4 // max value for CHUNKSTEPS/SLICESTEPS, must accord with above #define ALLTOALL_PIVOT_SLICESTEPS 2 #define ALLTOALL_PIVOT_CHUNKSTEPS 4