Update submodule NCCL to v2.18.3 (pytorch#104993)

Update NCCL submodule to v2.18.3 which fixes numerous bugs and performance issues, particularly on newer GPUs: https://docs.nvidia.com/deeplearning/nccl/release-notes/rel_2-18-3.html#rel_2-18-3 Pull Request resolved: pytorch#104993 Approved by: https://github.com/malfet
rui-ren · Aug 18, 2023 · 93f2a64 · 93f2a64
1 parent 64e02de
commit 93f2a64
Show file tree

Hide file tree

Showing 2 changed files with 2 additions and 1 deletion.
diff --git a/cmake/External/nccl.cmake b/cmake/External/nccl.cmake
@@ -50,6 +50,7 @@ if(NOT __NCCL_INCLUDED)
         "NVCC_GENCODE=${NVCC_GENCODE}"
         "BUILDDIR=${__NCCL_BUILD_DIR}"
         "VERBOSE=0"
+        "DEBUG=0"
       BUILD_BYPRODUCTS "${__NCCL_BUILD_DIR}/lib/libnccl_static.a"
       INSTALL_COMMAND ""
       )

diff --git a/third_party/nccl/nccl b/third_party/nccl/nccl
+5 −0		makefiles/common.mk
+2 −2		makefiles/version.mk
+1 −1		src/Makefile
+70 −2		src/bootstrap.cc
+118 −23		src/channel.cc
+7 −7		src/collectives/device/all_gather.h
+201 −54		src/collectives/device/all_reduce.h
+2 −3		src/collectives/device/common.h
+57 −193		src/collectives/device/common_kernel.h
+2 −1		src/collectives/device/functions.cu
+1 −1		src/collectives/device/onerank_reduce.cu
+39 −4		src/collectives/device/op128.h
+8 −16		src/collectives/device/primitives.h
+5 −5		src/collectives/device/prims_ll.h
+5 −5		src/collectives/device/prims_ll128.h
+57 −74		src/collectives/device/prims_simple.h
+92 −24		src/collectives/device/reduce_kernel.h
+5 −5		src/collectives/device/reduce_scatter.h
+6 −7		src/collectives/device/sendrecv.h
+2 −0		src/debug.cc
+73 −55		src/enqueue.cc
+166 −64		src/graph/connect.cc
+24 −25		src/graph/paths.cc
+187 −97		src/graph/search.cc
+111 −13		src/graph/topo.cc
+11 −4		src/graph/topo.h
+101 −53		src/graph/tuning.cc
+3 −9		src/graph/xml.cc
+8 −33		src/group.cc
+3 −0		src/include/align.h
+93 −4		src/include/alloc.h
+1 −0		src/include/bootstrap.h
+3 −1		src/include/channel.h
+29 −29		src/include/checks.h
+5 −4		src/include/collectives.h
+56 −19		src/include/comm.h
+4 −0		src/include/cudawrap.h
+15 −6		src/include/devcomm.h
+1 −1		src/include/gdrwrap.h
+9 −7		src/include/graph.h
+28 −1		src/include/group.h
+1,043 −0		src/include/ibvcore.h
+44 −0		src/include/ibvsymbols.h
+7 −1,031		src/include/ibvwrap.h
+3 −2		src/include/info.h
+0 −19		src/include/net.h
+21 −12		src/include/nvtx3/nvtx3.hpp
+43 −28		src/include/nvtx3/nvtxExtDetail/nvtxExtPayloadTypeInfo.h
+17 −0		src/include/p2p.h
+32 −19		src/include/proxy.h
+39 −7		src/include/transport.h
+616 −262		src/init.cc
+42 −3		src/misc/cudawrap.cc
+158 −0		src/misc/ibvsymbols.cc
+54 −182		src/misc/ibvwrap.cc
+6 −8		src/misc/shmutils.cc
+4 −1		src/misc/socket.cc
+14 −1		src/nccl.h.in
+31 −25		src/net.cc
+362 −293		src/proxy.cc
+35 −24		src/transport.cc
+102 −93		src/transport/coll_net.cc
+223 −180		src/transport/net.cc
+59 −24		src/transport/net_ib.cc
+183 −138		src/transport/nvls.cc
+283 −102		src/transport/p2p.cc
+22 −20		src/transport/shm.cc