search

deeplearning4j / deeplearning4j

Suite of tools for deploying and training deep learning models using the JVM. Highlights include model import for keras, tensorflow, and onnx/pytorch, a modular and tiny c++ library for running math code and a java based math library on top of the core c++ library. Also includes samediff: a pytorch/tensorflow like library for running deep learning using automatic differentiation.
http://deeplearning4j.konduit.ai
Apache License 2.0
13.52k stars 3.82k forks source link

libnd4j/nd4j/dl4j Jetson builds #8378

Closed raver119 closed 3 years ago

raver119 commented 4 years ago

As discussed with @agibsonccc: Let's set up build pipeline for Jetson devices family. Basically ARM-targeted cross-compilation + 1 cc per device.

raver119 commented 4 years ago

cc @sshepel @saudet

saudet commented 4 years ago

BTW, Travis CI now supports ARM64 and the hardware apparently comes from here: https://www.packet.com/cloud/servers/

saudet commented 4 years ago

/cc @kwatters @vb216 @johanvos @peardox

AlexDBlack commented 4 years ago

Not sure if useful here, but Huawei Cloud also has ARM servers. https://www.huaweicloud.com/en-us/product/ecs.html "Kunpeng General Computing-plus KC1 ECS" https://en.wikichip.org/wiki/hisilicon/kunpeng/920-6426

kwatters commented 4 years ago

If I remember, from getting nd4j working with cuda on the jetson nano, it only supported Cuda 10.0 ... Also, there was no support for Intel MKL... I'm not sure if nVidia has updated support for the nano to cuda 10.1 yet... (jetson was cuda CC 5.3 ) ... my jetson nano has been collecting dust for the past few months... maybe it's time to dig it out again.. (I need to build against 1.0.0.-beta5 anyway for myrobotlab)

saudet commented 4 years ago

Yeah, CUDA on ARM is a mess, but they are supposed to fix this soon: https://nvidianews.nvidia.com/news/nvidia-brings-cuda-to-arm-enabling-new-path-to-exascale-supercomputing Maybe we should wait for that to happen...

kwatters commented 4 years ago

yeah, last i checked, latest version of jetpack for the jetson nano was 4.2 something and that included only cuda 10.0 ... so yeah, i agree, we should wait for a new release of cuda that hopefully is consistent across all platforms ( arm and x86 ) .. at that point it'll be easier to maintain.

sshepel commented 4 years ago

Jetson build pipeline has been added to CI, artifacts (snapshots) can be found at maven central.