[NCCL over MSCCL++] Mscclpp failure: InvalidUsage 'This NVLS connection cannot map the requested devBuffSize' when test single node H100 (8GPUs)

[hustymf]

I build libmscclpp_nccl.so and nccl-tests with docker images ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda12.4

compile libmscclpp_nccl.so with the command:
cd mscclpp/build && cmake -DCMAKE_BUILD_TYPE=Release .. -DMSCCLPP_BUILD_TESTS=OFF -DMSCCLPP_BUILD_PYTHON_BINDINGS=OFF && make -j

compile nccl-tests with the command:
make MPI=1 MPI_HOME=/usr/local/mpi NCCL_HOME=/usr/lib/x86_64-linux-gnu/

And run nccl-tests/all_reduce_perf using libmscclpp_nccl.so on single node H100 (8GPUs) as follows,
mpirun -np 8 --bind-to numa --allow-run-as-root -x LD_PRELOAD=$MSCCLPP_BUILD/build/apps/nccl/libmscclpp_nccl.so ./build/all_reduce_perf -b 1k -e 16G -d bfloat16 -f 2 -g 1 -t 1 -c 0 -n 20 -w 20

And I got the error:
what(): terminate called after throwing an instance of 'mscclpp::Error'
what(): This NVLS connection cannot map the requested devBuffSize (Mscclpp failure: InvalidUsage)terminate called after throwing an instance of 'mscclpp::Error'
what(): This NVLS connection cannot map the requested devBuffSize (Mscclpp failure: InvalidUsage)
terminate called after throwing an instance of 'mscclpp::Error'
This NVLS connection cannot map the requested devBuffSize (Mscclpp failure: InvalidUsage)

When I use -x MSCCLPP_FORCE_DISABLE_NVLS=true to disbale NVLS, the test run successfully on single node H100 8GPUs,
but if test on single node H100 4GPUs, I got an error as follows,
Test CUDA failure common.cu:306 'an illegal memory access was encountered'
terminate called after throwing an instance of 'mscclpp::CudaError'
what(): Call to cudaThreadExchangeStreamCaptureMode(&mode_) failed. /mscclpp/src/gpu_utils.cc:11 (Cuda failure: driver shutting down)

[Binyang2014]

Let me review the NVLS implementation.
BTW, you can try this command: mpirun -np 8 --bind-to numa --allow-run-as-root -x LD_PRELOAD=/repo/mscclpp/build/apps/nccl/libmscclpp_nccl.so ./all_reduce_perf -b 1K -e 1G -f 2 -G 1 -n 20 -w 5 -c 1 -t 1 Change max message size to 1GB

I think our implement limit the max message size to 1GB for NVLS algo

For the second problem we hard coding to support 8 GPUs per node. Will remove this limitation in future release

[hustymf]

@Binyang2014 Thanks!
After changing the -e 16G to -e 1G, the test run successfully, but the result is float, not bfloat16, I got this

invalid type bfloat16, defaulting to float ...

but if use libnccl.so, I can get bfloat16 result

[Binyang2014]

Pls try with this branch: binyli/nccl. It's due to nccl-test use NCCL_VERSION to enable/disable some test for specific datatype, and MSCCLPP return incorrect NCCL version. This doesn't impact the real workload

[hustymf]

@Binyang2014 Got it, thank you so much!