WasmEdge-Tensorflow Dependencies

This repository provides pre-built dependencies for the WasmEdge with TensorFlow plug-in.

Motivation

The TensorFlow project no longer publishes the standalone libtensorflow C/C++ shared library packages — the official pre-built libraries only ship inside the Python pip wheels. Projects which link TensorFlow directly, such as the WasmEdge plug-ins, still need standalone shared libraries that run on the manylinux_2_28 (glibc >= 2.28) platforms.

But building TensorFlow takes lots of time. To reduce the compilation time, we create this project to compile and release the pre-built standalone TensorFlow and TensorFlow-Lite shared libraries for the manylinux_2_28, macOS, and Android platforms.

The headers vendored in this repository (tensorflow/, tsl/, xla/ at the repo root) are the include closure of the C API headers used by WasmEdge, extracted from the v2.21.0 source tree.

License

This project is under the Apache-2.0 License as the same as the TensorFlow project.

Credits

• TensorFlow

How to build the TensorFlow 2.21.0 dependencies

All libraries are built from the pristine v2.21.0 tag — no source patch, no ./configure, no local Python (hermetic Python via --repo_env=HERMETIC_PYTHON_VERSION=3.12).

The TensorFlow asset is the two-library pywrap pair (libtensorflow_cc + libtensorflow_framework, SONAME-form filenames) — the same libraries the official TensorFlow pip wheels ship, built by the same pipeline. Consumers must link both, libtensorflow_cc first (some C API symbols are defined in libtensorflow_framework).

Source, for every platform:

git clone https://github.com/tensorflow/tensorflow.git tensorflow_src
cd tensorflow_src && git checkout v2.21.0

Build on macOS

1. Environment settings

   • Full Xcode (the Command Line Tools alone are not enough), selected as the developer directory (xcode-select -p); ~50 GB free disk.
   • brew install bazelisk — it fetches bazel 7.7.0 from tensorflow/.bazelversion; always build with bazelisk, never a standalone bazel.

2. Build steps on darwin_arm64 (macOS 12+)

   # TensorFlow pair:
   bazelisk build --config=release_macos_arm64 \
     --repo_env=HERMETIC_PYTHON_VERSION=3.12 \
     //tensorflow/python:libtensorflow_cc.2.dylib \
     //tensorflow/python:libtensorflow_framework.2.dylib
   # -> bazel-bin/tensorflow/python/libtensorflow_cc.2.dylib, libtensorflow_framework.2.dylib
   
   # TensorFlow-Lite C library + flex delegate:
   bazelisk build --config=release_macos_arm64 --config=monolithic \
     --repo_env=HERMETIC_PYTHON_VERSION=3.12 \
     //tensorflow/lite/c:libtensorflowlite_c.dylib \
     //tensorflow/lite/delegates/flex:tensorflowlite_flex
   # -> bazel-bin/tensorflow/lite/c/libtensorflowlite_c.dylib,
   #    bazel-bin/tensorflow/lite/delegates/flex/libtensorflowlite_flex.dylib

3. Build steps on darwin_x86_64 (macOS 10.15+)

   The flag set below is release_macos_x86 without its host AVX flag, plus three additions: the oneDNN contraction-kernel opt-out and a target-only AVX2/FMA re-enable that forces RUY's x86 SIMD kernels on. Two things go wrong otherwise at v2.21.0:

   • The stock --config=release_macos_x86 fails to build: its --host_copt=-mavx is applied to the arm64 slice of gRPC's universal_binary codegen tools (built even on Intel), where -mavx is illegal; and macOS x86 has no real oneDNN, so the _with_mkl contraction kernel can't find dnnl.h.
   • Even once it builds, RUY (the TF-Lite / TF matmul kernels) disables all of its x86 SIMD paths on Apple — ruy/platform.h gates RUY_PLATFORM_X86_ENHANCEMENTS on __linux__ clang only. With no AVX2 kernel, RUY falls back to its incomplete SSE4.2 8-bit kernel stub and segfaults at runtime on quantized (uint8) inference. --copt=-DRUY_FORCE_ENABLE_X86_ENHANCEMENTS --copt=-mavx2 --copt=-mfma --copt=-mf16c restores RUY's real Kernel8bitAvx2. These are target-only --copts, so they do not reach the gRPC arm64 slice and do not reintroduce the build failure above.

   This makes the darwin_x86_64 libraries AVX2-baseline — safe on every Intel Mac (all Haswell+).

   # TensorFlow pair:
   bazelisk build --config=release_macos_base --cpu=darwin \
     --macos_minimum_os=10.15 --action_env=MACOSX_DEPLOYMENT_TARGET=10.15 \
     --define tensorflow_mkldnn_contraction_kernel=0 \
     --copt=-DRUY_FORCE_ENABLE_X86_ENHANCEMENTS --copt=-mavx2 --copt=-mfma --copt=-mf16c \
     --repo_env=HERMETIC_PYTHON_VERSION=3.12 \
     //tensorflow/python:libtensorflow_cc.2.dylib \
     //tensorflow/python:libtensorflow_framework.2.dylib
   # -> bazel-bin/tensorflow/python/libtensorflow_cc.2.dylib, libtensorflow_framework.2.dylib
   
   # TensorFlow-Lite C library + flex delegate:
   bazelisk build --config=release_macos_base --cpu=darwin \
     --macos_minimum_os=10.15 --action_env=MACOSX_DEPLOYMENT_TARGET=10.15 \
     --define tensorflow_mkldnn_contraction_kernel=0 --config=monolithic \
     --copt=-DRUY_FORCE_ENABLE_X86_ENHANCEMENTS --copt=-mavx2 --copt=-mfma --copt=-mf16c \
     --repo_env=HERMETIC_PYTHON_VERSION=3.12 \
     //tensorflow/lite/c:libtensorflowlite_c.dylib \
     //tensorflow/lite/delegates/flex:tensorflowlite_flex
   # -> bazel-bin/tensorflow/lite/c/libtensorflowlite_c.dylib,
   #    bazel-bin/tensorflow/lite/delegates/flex/libtensorflowlite_flex.dylib

Build on Linux (manylinux_2_28_x86_64 / manylinux_2_28_aarch64)

Measured floors: GLIBC 2.27, GLIBCXX 3.4.22, CXXABI 1.3.11 — the libraries run on any manylinux_2_28 (glibc ≥ 2.28) system. x86_64 baseline is AVX; aarch64 is armv8-a.

1. Environment settings

   Docker is required; the build runs in a manylinux_2_34 container (the hermetic clang prebuilts need build-host glibc ≥ 2.29, so manylinux_2_28 images cannot host the build — the artifacts still keep the GLIBC 2.27 floor); ~50 GB free disk. Start the container from the tensorflow_src checkout, then install xxd and bazelisk:

   docker run -it --rm -v "$PWD":/root/tensorflow_src -w /root/tensorflow_src \
     quay.io/pypa/manylinux_2_34_x86_64 bash
   # aarch64: quay.io/pypa/manylinux_2_34_aarch64
   
   # In the container:
   dnf -q -y install vim-common   # provides xxd
   curl -sLo /usr/local/bin/bazel \
     https://github.com/bazelbuild/bazelisk/releases/download/v1.26.0/bazelisk-linux-amd64 \
     && chmod +x /usr/local/bin/bazel
   # aarch64: .../bazelisk-linux-arm64

2. Build steps

   # x86_64 shown; aarch64: --config=release_arm64_linux
   bazel build --config=release_cpu_linux \
     --repo_env=HERMETIC_PYTHON_VERSION=3.12 \
     //tensorflow/python:libtensorflow_cc.so.2 \
     //tensorflow/python:libtensorflow_framework.so.2
   # -> bazel-bin/tensorflow/python/libtensorflow_cc.so.2, libtensorflow_framework.so.2
   
   # TensorFlow-Lite C library + flex delegate:
   bazel build --config=release_cpu_linux --config=monolithic \
     --repo_env=HERMETIC_PYTHON_VERSION=3.12 \
     //tensorflow/lite/c:libtensorflowlite_c.so \
     //tensorflow/lite/delegates/flex:tensorflowlite_flex
   # -> bazel-bin/tensorflow/lite/c/libtensorflowlite_c.so,
   #    bazel-bin/tensorflow/lite/delegates/flex/libtensorflowlite_flex.so

Build on Android (android_aarch64, TensorFlow-Lite only)

The Android asset contains only libtensorflowlite_c.so (arm64-v8a, minSdkVersion 23).

1. Environment settings

   • A macOS or Linux x86_64 host with bazelisk (as above).
   • The Android NDK r25c from https://dl.google.com/android/repository/android-ndk-r25c-<darwin|linux>.zip — no Android SDK, no Java.

2. Build steps

   bazelisk build \
     --config=android_arm64 \
     --repo_env=ANDROID_NDK_HOME=/path/to/android-ndk-r25c \
     --repo_env=ANDROID_NDK_API_LEVEL=23 \
     --repo_env=ANDROID_NDK_VERSION=25 \
     --repo_env=HERMETIC_PYTHON_VERSION=3.12 \
     //tensorflow/lite/c:tensorflowlite_c
   # -> bazel-bin/tensorflow/lite/c/libtensorflowlite_c.so

On every platform, copy the libraries out of bazel-bin with cp -L (dereferences bazel's symlinks) and chmod 755.

Legacy build guides

• LEGACY_BUILD_2.6.md — TensorFlow 2.6.0 (CentOS 7.6.1810, MacOS, Android)
• LEGACY_BUILD_2.12.md — TensorFlow 2.12.0 (CentOS 7.6.1810, MacOS, Android)

Minimum requirements of our pre-built shared libraries

Pre-built shared library	Platform	TensorFlow Tag	GLIBC	GLIBCXX	CXXABI
libtensorflow.so	manylinux2014_x86_64	2.6.0	2.17	3.4.19	1.3.7
libtensorflow_cc.so	manylinux2014_x86_64	2.6.0	2.17	3.4.19	1.3.7
libtensorflow_framework.so	manylinux2014_x86_64	2.6.0	2.16	3.4.19	1.3.7
libtensorflowlite_c.so	manylinux2014_x86_64	2.6.0	2.14	3.4.19	1.3.5
libtensorflow_cc.so	manylinux2014_x86_64	2.12.0	2.17	3.4.19	1.3.7
libtensorflow_framework.so	manylinux2014_x86_64	2.12.0	2.16	3.4.19	1.3.7
libtensorflowlite_c.so	manylinux2014_x86_64	2.12.0	2.14	3.4.19	1.3.5
libtensorflowlite_flex.so	manylinux2014_x86_64	2.12.0	2.17	3.4.19	1.3.7
libtensorflowlite_c.so	manylinux2014_aarch64	2.6.0	2.17	None	None
libtensorflow_cc.so	manylinux2014_aarch64	2.12.0	2.17	3.4.19	1.3.7
libtensorflow_framework.so	manylinux2014_aarch64	2.12.0	2.16	3.4.19	1.3.7
libtensorflowlite_c.so	manylinux2014_aarch64	2.12.0	2.17	None	None
libtensorflowlite_flex.so	manylinux2014_aarch64	2.12.0	2.17	3.4.19	1.3.7
libtensorflowlite_c.so	android_aarch64	2.6.0	None	None	None
libtensorflowlite_c.so	android_aarch64	2.12.0	None	None	None
libtensorflow_cc.so.2	manylinux_2_28_x86_64	2.21.0	2.27	3.4.22	1.3.11
libtensorflow_framework.so.2	manylinux_2_28_x86_64	2.21.0	2.27	3.4.22	1.3.11
libtensorflowlite_c.so	manylinux_2_28_x86_64	2.21.0	2.27	3.4.22	1.3.5
libtensorflowlite_flex.so	manylinux_2_28_x86_64	2.21.0	2.27	3.4.22	1.3.11
libtensorflow_cc.so.2	manylinux_2_28_aarch64	2.21.0	2.27	3.4.22	1.3.11
libtensorflow_framework.so.2	manylinux_2_28_aarch64	2.21.0	2.27	3.4.22	1.3.11
libtensorflowlite_c.so	manylinux_2_28_aarch64	2.21.0	2.27	3.4.22	1.3.5
libtensorflowlite_flex.so	manylinux_2_28_aarch64	2.21.0	2.27	3.4.22	1.3.11
libtensorflowlite_c.so	android_aarch64	2.21.0	None	None	None

Pre-built shared library	Platform	TensorFlow Tag	Minimum MacOS version
libtensorflow.dylib	darwin_x86_64	2.6.0	10.15
libtensorflow_cc.dylib	darwin_x86_64	2.6.0	10.15
libtensorflow_framework.dylib	darwin_x86_64	2.6.0	10.15
libtensorflowlite_c.dylib	darwin_x86_64	2.6.0	10.15
libtensorflow_cc.dylib	darwin_x86_64	2.12.0	10.15
libtensorflow_framework.dylib	darwin_x86_64	2.12.0	10.15
libtensorflowlite_c.dylib	darwin_x86_64	2.12.0	10.15
libtensorflowlite_flex.dylib	darwin_x86_64	2.12.0	10.15
libtensorflow_cc.dylib	darwin_arm64	2.12.0	12.0
libtensorflow_framework.dylib	darwin_arm64	2.12.0	12.0
libtensorflowlite_c.dylib	darwin_arm64	2.12.0	12.0
libtensorflowlite_flex.dylib	darwin_arm64	2.12.0	12.0
libtensorflow_cc.2.dylib	darwin_x86_64	2.21.0	10.15
libtensorflow_framework.2.dylib	darwin_x86_64	2.21.0	10.15
libtensorflowlite_c.dylib	darwin_x86_64	2.21.0	10.15
libtensorflowlite_flex.dylib	darwin_x86_64	2.21.0	10.15
libtensorflow_cc.2.dylib	darwin_arm64	2.21.0	12.0
libtensorflow_framework.2.dylib	darwin_arm64	2.21.0	12.0
libtensorflowlite_c.dylib	darwin_arm64	2.21.0	12.0
libtensorflowlite_flex.dylib	darwin_arm64	2.21.0	12.0

Active Releases

• TF-2.6.0: TensorFlow 2.6.0 C library
  • TensorFlow 2.6.0 C shared library for manylinux2014_x86_64 and darwin_x86_64.
  • TensorFlow-Lite 2.6.0 C shared library for manylinux2014_x86_64, manylinux2014_aarch64, darwin_x86_64, and android_aarch64.
• TF-2.6.0-CC: TensorFlow 2.6.0 C++ library
  • TensorFlow 2.6.0 C++ shared library for manylinux2014_x86_64 and darwin_x86_64.
  • TensorFlow-Lite 2.6.0 C shared library for manylinux2014_x86_64, manylinux2014_aarch64, darwin_x86_64, and android_aarch64.
• TF-2.12.0-CC: TensorFlow 2.12.0 C++ library
  • TensorFlow 2.12.0 C++ shared library for manylinux2014_x86_64, manylinux2014_aarch64, darwin_x86_64, and darwin_arm64.
  • TensorFlow-Lite 2.12.0 C shared library with Flex delegate shared library for manylinux2014_x86_64, manylinux2014_aarch64, darwin_x86_64, and darwin_arm64.
  • TensorFlow-Lite 2.12.0 C shared library for android_aarch64.
• TF-2.21.0-CC: TensorFlow 2.21.0 C++ library
  • TensorFlow 2.21.0 C++ shared library for manylinux_2_28_x86_64, manylinux_2_28_aarch64, darwin_x86_64, and darwin_arm64.
  • TensorFlow-Lite 2.21.0 C shared library with Flex delegate shared library for the same platforms.
  • TensorFlow-Lite 2.21.0 C shared library for android_aarch64.