Disable NET transport for intra-node communication by setting the env to 1
It provides an option to error out instead of falling back to NET when superior intra-node transports (P2P and SHM) are unavailable
[ROCm/rccl commit: c88c9f873f]
Display hints of cause so that it would be easier for user to debug.
Also change the error type from InternalError to InvalidUsage as most
of time this is caused by a mismatch in collective size or env settings.
[ROCm/rccl commit: f589932130]
I noticed when I enabled `NCCL_DEBUG_SUBSYS=ALLOC` that this function is
called thousands of times, making the log output unintelligible.
Fortunately, this function can be implemented without heap allocations.
[ROCm/rccl commit: 8cf7325d69]
* Fixing cmake_install_prefix search to include /opt/rocm-xxxx
* Removing all hard references to /opt/rocm with ROCM_PATH
* Setting ROCM_PATH CMake variable in install script
[ROCm/rccl commit: 7b8b54955b]
* gtest: dynamically generate tests based on test machine's GPU count
* Adjust test element size and bfloat16 threshold for up to 16 GPUs
[ROCm/rccl commit: 03a830293c]
* Fixing cmake_install_prefix search to include /opt/rocm-xxxx
* Removing all hard references to /opt/rocm with ROCM_PATH
[ROCm/rccl commit: a6dba6b9dd]
* Initial commit of all_reduce_only support
* Working AllReduce only build
* Removing printfs and restoring release build
* Restore P2P index
* Updates to build_allreduce_only mode.
* cleaning up macro ifdefs
[ROCm/rccl commit: 7e55b211c5]
* [TransferBench] Adding a p2p benchmark mode
* [TransferBench] Switching to using single sync mode by default (USE_SINGLE_SYNC=1)
[ROCm/rccl commit: 550d732d6c]
* Use relaxed atomics and add sleep and wakeup in barrier loop
* atomicAdd in ROCm 4.3 only support unsigned long long
* Switch to atomicAdd and atomicExch in more places
* Restore LOAD/STORE define to __ATOMIC_SEQ_CST
* Restore atomic for sizes FIFO
[ROCm/rccl commit: 020484bf40]
Add new API for creating a reduction operation which multiplies the input by a rank-specific scalar before doing an inter-rank summation (see: ncclRedOpCreatePreMulSum).
Improve CollNet (SHARP) performance of ncclAllReduce when captured in a CUDA Graph via user buffer registration.
Add environment variable NCCL_NET_PLUGIN="<suffix>" to allow user to choose among multiple NCCL net plugins by substituting into "libnccl-net-<suffix>.so".
Fix memory leak of NVB connections.
Fix topology detection of IB Virtual Functions (SR-IOV).
[ROCm/rccl commit: e11238b302]
Ke Wen
Wenkai Du
gilbertlee-amd
Chris Jones
Stanley Tsang
Wenkai Du
rachanaramanna
pavahora
Saad Rahim
Roopa Malavally
John Bachan
Sylvain Jeaugey
Ke Wen