Ported over from https://github.com/tensorflow/community/pull/352

The tfp-bindings crate mostly provides unsafe bindings to all TensorFlow C APIs required by plug-ins via libtensorflow_framework.so.2 library (yes, and currently in only supports Linux and Python 3.10). It also has some should-be-safe bindings to main functions required by compute kernels and their registration. Its safety mostly relies on validating data, checking result codes and asserts in debug mode to prevent common plug-in issues like unterminated strings.

Plugin itself (tfp-plugin) should link to that library via its build.rs script and implement SE_InitPlugin and TF_InitKernel functions with proper types (these are excluded from bindgen). It may also include optimizer, but it is not yet implemented in this repository.

• Linux, tested on Fedora 36, any should work fine.
• MSRV not yet specified, I use nightly. No strict limits are being set.
• Python 3 for test script. pip3 and venv to install TensorFlow
• clang for binding generation via bindgen

You can import the bindings crate by using such a dependency string. There are environment variables you can use to point the paths to TensorFlow library:

• TF_INCLUDE_PATH
• TF_LIBRARY_PATH

They could be set, for example, by Cargo project config

Cargo.toml example. Replace x by needed commit hash.

tfp-bindings = { git = "https://github.com/sh7dm/rust-tf-pluggabledevice", rev = "x" }

python3.10 -m venv venv
source venv/bin/activate
pip3 install --no-cache-dir tf-nightly-cpu # or tensorflow-cpu
# https://github.com/tensorflow/tensorflow/issues/55497
mv venv/lib64 venv/_lib64
cargo build
mkdir venv/lib/python3.10/site-packages/tensorflow-plugins/
cp target/debug/libtfp.so venv/lib/python3.10/site-packages/tensorflow-plugins/
python3 test.py

import tensorflow as tf
tf.config.list_physical_devices()

Specifying LD_LIBRARY_PATH manually is necessary as of https://github.com/rust-lang/cargo/issues/4044

LD_LIBRARY_PATH=$(pwd)/venv/lib/python3.10/site-packages/tensorflow cargo test