juliagpu / nccl.jl

A Julia wrapper for the NVIDIA Collective Communications Library.

NCCL.jl

A Julia wrapper for the NVIDIA Collective Communications Library (NCCL).

nccl.jl's Issues

TagBot trigger issue

This issue is used to trigger TagBot; feel free to unsubscribe.

If you haven't already, you should update your TagBot.yml to include issue comment triggers.
Please see this post on Discourse for instructions and more details.

If you'd like for me to do this for you, comment TagBot fix on this issue.
I'll open a PR within a few hours, please be patient!

Complex number wrapper

NCCL does not support complex numbers directly and does not plan to (see issue). Are we willing to add a wrapper to NCCL.jl to make using complex numbers more convienient? Alternatively, the wrapper could be put in a higher level package (ex. Lux.jl, see issue). I am happy to start working on this but would like some feedback if possible. My primary motivation is using neural networks with complex valued weights and this feature would greatly simplify things.

Tests assume 2 GPUs

Should probably also work on a single-GPU device (we can force some of the CI jobs on a multigpu system to ensure coverage), and also on cyclops which has 8 devices.

status

Hi @maleadt Could you please let me know the status of nccl.jl package. I would like to contribute and finish this package.
Thank you very much!

Best
Raj

Tests failed

Do I need to configure anything to pass the test?

(This is a fresh new installation based on the pytorch:24.01-py3 image)

     Testing Running tests...
┌ Info: CUDA information:
│ CUDA runtime 12.5, artifact installation
│ CUDA driver 12.5
│ NVIDIA driver 535.161.8, originally for CUDA 12.2
│
│ CUDA libraries:
│ - CUBLAS: 12.5.3
│ - CURAND: 10.3.6
│ - CUFFT: 11.2.3
│ - CUSOLVER: 11.6.3
│ - CUSPARSE: 12.5.1
│ - CUPTI: 2024.2.1 (API 23.0.0)
│ - NVML: 12.0.0+535.161.8
│
│ Julia packages:
│ - CUDA: 5.4.3
│ - CUDA_Driver_jll: 0.9.1+1
│ - CUDA_Runtime_jll: 0.14.1+0
│
│ Toolchain:
│ - Julia: 1.10.4
│ - LLVM: 15.0.7
│
│ 8 devices:
│   0: NVIDIA H100 80GB HBM3 (sm_90, 79.106 GiB / 79.647 GiB available)
│   1: NVIDIA H100 80GB HBM3 (sm_90, 79.106 GiB / 79.647 GiB available)
│   2: NVIDIA H100 80GB HBM3 (sm_90, 79.106 GiB / 79.647 GiB available)
│   3: NVIDIA H100 80GB HBM3 (sm_90, 79.106 GiB / 79.647 GiB available)
│   4: NVIDIA H100 80GB HBM3 (sm_90, 79.106 GiB / 79.647 GiB available)
│   5: NVIDIA H100 80GB HBM3 (sm_90, 79.106 GiB / 79.647 GiB available)
│   6: NVIDIA H100 80GB HBM3 (sm_90, 79.106 GiB / 79.647 GiB available)
└   7: NVIDIA H100 80GB HBM3 (sm_90, 79.106 GiB / 79.647 GiB available)
[ Info: NCCL version: 2.19.4
Communicator: Error During Test at /....../NCCL.jl/test/runtests.jl:11
  Got exception outside of a @test
  NCCLError(code ncclUnhandledCudaError, a call to a CUDA function failed)
  Stacktrace:
  Stacktrace:
    [1] check
      @ /....../NCCL.jl/src/libnccl.jl:17 [inlined]
    [2] ncclCommInitAll
      @ ~/.julia/packages/CUDA/Tl08O/lib/utils/call.jl:34 [inlined]
    [3] Communicators(deviceids::Vector{Int32})
      @ NCCL /....../NCCL.jl/src/communicator.jl:70
    [4] Communicators(devices::CUDA.DeviceIterator)
      @ NCCL /....../NCCL.jl/src/communicator.jl:80
    [5] macro expansion
      @ /....../NCCL.jl/test/runtests.jl:13 [inlined]
    [6] macro expansion
      @ ~/.julia/juliaup/julia-1.10.4+0.x64.linux.gnu/share/julia/stdlib/v1.10/Test/src/Test.jl:1577 [inlined]
    [7] macro expansion
      @ /....../NCCL.jl/test/runtests.jl:13 [inlined]
    [8] macro expansion
      @ ~/.julia/juliaup/julia-1.10.4+0.x64.linux.gnu/share/julia/stdlib/v1.10/Test/src/Test.jl:1577 [inlined]
    [9] top-level scope
      @ /....../NCCL.jl/test/runtests.jl:11
   [10] include(fname::String)
      @ Base.MainInclude ./client.jl:489
   [11] top-level scope
      @ none:6
   [12] eval
      @ ./boot.jl:385 [inlined]
   [13] exec_options(opts::Base.JLOptions)
      @ Base ./client.jl:291
   [14] _start()
      @ Base ./client.jl:552

Register?

It seems like NCCL has been functional for a month now. Is there a plan/timeline to register it?