2.18.1-1

Add support for IB SHARP to NVLS (NVLink SHARP algorithm).
Add NVLS+Tree algorithm.
Add support for memory management using cuMem* functions.
Use all NICs for Send/Receive operations on systems with more than
one NIC per GPU (NVIDIA#804).
Add ncclCommSplit primitive, with resource sharing option in config.
Fix alltoallv hang (NVIDIA#788)
Increase number of channels on H100 when we're not limited by NVLink.
Improve error reporting in case of IB failure, printing local and
remote ID (NVIDIA#779).
Add build option to allow compilation against RDMA includes instead
of dynamically loading IB verbs symbols (NVIDIA#802).
Fix context creation for progress thread (NVIDIA#803).
NET/IB: add option to use multiple QPs in round-robin mode.
Fix tree performance issue when NVB is disabled on HCM topologies.

Author: sjeaugey

makefiles/common.mk

@@ -12,6 +12,7 @@ DEBUG ?= 0
 TRACE ?= 0
 PROFAPI ?= 1
 NVTX ?= 1
+RDMA_CORE ?= 0
 
 NVCC = $(CUDA_HOME)/bin/nvcc
 
@@ -106,3 +107,7 @@ endif
 ifneq ($(PROFAPI), 0)
 CXXFLAGS += -DPROFAPI
 endif
+
+ifneq ($(RDMA_CORE), 0)
+CXXFLAGS += -DNCCL_BUILD_RDMA_CORE=1
+endif

makefiles/version.mk

@@ -1,6 +1,6 @@
 ##### version
 NCCL_MAJOR   := 2
-NCCL_MINOR   := 17
+NCCL_MINOR   := 18
 NCCL_PATCH   := 1
 NCCL_SUFFIX  :=
 PKG_REVISION := 1