How can I change the build config for the platform natives to mkl_aarch64

tensorflow / java

Java bindings for TensorFlow

Apache License 2.0

832 stars 202 forks source link

How can I change the build config for the platform natives to mkl_aarch64 #409

Open ashesfall opened 2 years ago

ashesfall commented 2 years ago

System information

OS Platform and Distribution (e.g., Linux Ubuntu 16.04 x86_64): Aarch64 (Apple Sillicon)
TensorFlow installed from (source or binary): Source
TensorFlow version: 2.7
Java version (i.e., the output of java -version): 17
Java command line flags (e.g., GC parameters):
Installed from Maven Central?: No
Bazel version (if compiling from source): 3.7.2
GCC/Compiler version (if compiling from source): Apple clang version 13.0.0
CUDA/cuDNN version: N/A
GPU model and memory: Apple M1

When building core-api I can see from the logs it is targeting x86_64. I wish to target Aarch64.

./build.sh in core-api

Craigacp commented 2 years ago

Do you want to cross-compile from an x86 machine or compile natively on an M1?

I've got a branch where native compilation works, but you need to run the bazel build as the superuser due to some library discovery issue I've not figured out. https://github.com/tensorflow/java/issues/394#issuecomment-981722302

I don't think we know how to cross-compile it from an x86 Mac.

ashesfall commented 2 years ago

The goal was compiling directly on the M1.

And yes I eventually found your old discussion and have succeeded.

When will this become more official?

Craigacp commented 2 years ago

When I or someone else figures out how to compile it without needing to run bazel as root we'll merge it into master. We won't be able to do builds for it due to a lack of appropriate build resources unless we manage to figure out cross compiling.

ashesfall commented 2 years ago

Okay. You can't lose this ticket.

ashesfall commented 2 years ago

Oh one last thing. Is integration with tensorflow-metal on the road map either?

Craigacp commented 2 years ago

We don't currently expose TF_LoadPluggableDeviceLibrary which is the entry point for the pluggable device infrastructure. If we did, then I think you should be able to download the tensorflow-metal whl, unzip it and then load it with that function. As far as I can tell tensorflow-metal is closed source, so I don't think we'd be able to repackage it for Java.

Apple's docs do say "V1 TensorFlow Networks" are unsupported, but I'm not sure what they mean by that.

ashesfall commented 2 years ago

Yea it can't be included, but it would be great if I could load the pluggable device. You can close the ticket.