Split transport connect phase into multiple steps to avoid port
exhaustion when connecting alltoall at large scale. Defaults to 128
peers per round.
Fix memory leaks on CUDA graph capture.
Fix alltoallv crash on self-sendrecv.
Make topology detection more deterministic when PCI speeds are not
available (fix issue #1020).
Properly close shared memory in NVLS resources.
Revert proxy detach after 5 seconds.
Add option to print progress during transport connect.
Add option to set NCCL_DEBUG to INFO on first WARN.
Add local user buffer registration for NVLink SHARP.
Add tuning plugin support.
Increase net API to v7 to allow for device-side packet reordering;
remove support for v4 plugins.
Add support for RoCE ECE.
Add support for C2C links.
Better detect SHM allocation failures to avoid crash with Bus Error.
Fix missing thread unlocks in bootstrap (Fixes#936).
Disable network flush by default on H100.
Move device code from src/collectives/device to src/device.
We are trying to use the copy result of first cudaMemcpyAsync in the
second cudaMemcpyAsync without sync in between. This patch fixes it
by allocating a CPU side array to cache device side addr so that we
can avoid this consecutive cuda mem copy.
Fixes#957
We use CMake to determine if we're compiling against a version of ROCm that supports gcnArchName and handles architecture checking appropriately. It includes a few helper functions as drop ins for the functionality we used gcnArch for before; sometimes to enable flags, and sometimes to set frequencies.
* initial checkin
* enable msccl when hip graph is on
* remove the commented out code of msccl enable check
* clean up the code
* remove the msccl HighestTransportType check logic
Fix data corruption with Tree/LL128 on systems with 1GPU:1NIC.
Fix hang with Collnet on bfloat16 on systems with less than one NIC
per GPU.
Fix long initialization time.
Fix data corruption with Collnet when mixing multi-process and
multi-GPU per process.
Fix crash when shared memory creation fails.
Fix Avg operation with Collnet/Chain.
Fix performance of alltoall at scale with more than one NIC per GPU.
Fix performance for DGX H800.
Fix race condition in connection progress causing a crash.
Fix network flush with Collnet.
Fix performance of aggregated allGather/reduceScatter operations.
Fix PXN operation when CUDA_VISIBLE_DEVICES is set.
Fix NVTX3 compilation issues on Debian 10.
* Fix build on additional architectures
Instead of directly wrapping a platform-specific operation with a
preprocessor check against a gfx macro, it can be more flexible to
check a macro that can be overriden by the user. The gfx macro can then
just provide the default value for the macro, resulting in the same
default behaviour as if the gfx macro was checked directly but with
more control at build-time.
For example, to build rccl without using buffer_wbinvl1_vol on
gfx902, but still use the default on other archs, a user could
export CXXFLAGS='-Xarch_gfx902 -DRCCL_USE_WBINVL1_VOL=1' before
configuring the build. This flexibility isn't always necessary, but
it's nicer to have it and not need it than to need it and not have it.
* Define WARP_SIZE using warpSize builtin
BertanDogancay
Bertan Dogancay
Wenkai Du
Nilesh M Negi
Pedram Alizadeh
akolliasAMD
Ziyue Yang
Wen-Heng (Jack) Chung
Sylvain Jeaugey
Kaiming Ouyang
Audrey MP
Andy li
Nusrat Islam
David Pagan
Wenkai Du
Cory Bloor