Skip to content

[CUDA] cudastack lists incompatible TensorRT and nccl for Jetpack 5 #1677

@stylish-llama

Description

@stylish-llama

Search before asking

  • I have searched the jetson-containers issues and found no similar feature requests.

jetson-containers Component

CUDA

Bug

The current configuration lists incompatible TensorRT and nccl in cudastack for Jetpack 5:

else: # JetPack 5
cudnn_ver = '8.6.0'
cudnn_url = "https://repo.download.nvidia.com/jetson/common/pool/main/c/cudnn/libcudnn8_8.6.0.166-1+cuda11.4_arm64.deb"
cudnn_packages = "libcudnn8 libcudnn8-dev"
tensorrt_ver = '8.6.0'
tensorrt_url = "https://nvidia.box.com/shared/static/hmwr57hm88bxqrycvlyma34c3k4c53t9.deb"
nccl_ver = '2.21.5'
nccl_url = ''

The latest Jetpack 5 (version 5.1.6) comes with the following:

  • CUDA 11.4.19
  • TensorRT 8.5.2
  • cuDNN 8.6.0
  • DLA 3.12.1

The latest nccl version supporting cuda 11.4 is verion 2.11.4 (see release notes).

The line tensorrt_url = "https://nvidia.box.com/shared/static/hmwr57hm88bxqrycvlyma34c3k4c53t9.deb" seems to be an external url, with no clear replacement for the correct version.

Environment

┌───────────────────────┬────────────────────────┐
│ L4T_VERSION   35.6.4  │ JETPACK_VERSION  5.1.6 │
│ CUDA_VERSION  11.4    │ PYTHON_VERSION   3.8   │
│ SYSTEM_ARCH   aarch64 │ LSB_RELEASE      20.04 │
└───────────────────────┴────────────────────────┘

Additional

No response

Are you willing to submit a PR?

  • Yes I'd like to help by submitting a PR!

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions