We have not released anything for python3.9.

Release process is as follows:

1. Ensure all of our dependencies support python3.9 (not yet there)
2. Ensure our code works with all these dependencies on all supported versions of python (at the moment, at least gast needs additional support work).
3. Release nightly version supporting the new python version
4. Release a full release.

As we are already doing the 2.4 release, python3.9 support for TF will come in 2.5 release at the earliest.

@ravikyram, @amahendrakar, @rmothukuru let's deduplicate to this issue all py39 requests/issues. I'll monitor this one and provide updates as we start adding support.

@mihaimaruseac,
Sure

I may be wrong, but I believe this is the case:

python3.8 -m pipdeptree -p tensorflow | grep "^  -" | sed "s/  - //" | sed "s/ \[.*]//" | xargs python3.9 -m pip install

which is essentially

python3.9 -m pip install absl-py astunparse flatbuffers gast google-pasta grpcio h5py keras-preprocessing numpy opt-einsum protobuf six tensorboard tensorflow-estimator termcolor typing-extensions wheel wrapt

works fine for me. Am I missing anything?

(from #40840 (comment))

The main issue is that we also have C++ code and we need the needed dependencies to also upgrade for their C++ code. By this time, likely most are solved, so 2.5 release will have py3.9 support.

We should get py3.9 in nightly soon.

So, all of our depdencies have a release for py3.9 but our setup.py upper bounds to versions which are below that release. We will need to increase these bounds while making sure we don't break existing usecases/tests. Will slowly happen over December.

I've built tf 2.4 for python 3.9.1 on windows 10 and ubuntu 18.04 and they work fine, no problem at all

I've built tf 2.4 for python 3.9.1 on windows 10 and ubuntu 18.04 and they work fine, no problem at all

Curious to know how you did that since you cannot use it because of this “stupid” dependency check. Oh you mean you built it from source? If that’s so, would you mind sharing the commands/options you used?

thanks!

@willemavjc I can share my build from source setup. Win 10 and works with python 3.9.1, TF 2.4.
Python 3.9 installed in c:\python39\ dir.

# Copy https://github.com/bazelbuild/bazel/releases/download/3.7.2/bazel-3.7.2-windows-x86_64.exe to somewhere in PATH and rename to bazel.exe 
# Make sure you have Visual Studio 2019 Build Tools installed (located under C:\Program Files (x86)\Microsoft Visual Studio\2019\BuildTools\VC)
# Make sure cudnn-8.0.5.39 is unpacked into C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.0 respective dirs.
# Ensure your env vars contain something like this:
# CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.0
# CUDA_TOOLKIT_PATH=%CUDA_PATH%
# PATH=%CUDA_PATH%\bin;%PATH%
# PATH=%CUDA_PATH%\extras\CUPTI\libx64;%PATH%

# Using PowerShell 7.1.0
git clone https://github.com/tensorflow/tensorflow.git tf24build
cd tf24build
git checkout v2.4.0
\python39\python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -U pip

# Now, depending on how fresh is your Windows installation, you need either numpy==1.19.3 or or 1.19.4.
# See https://github.com/numpy/numpy/issues/16744 for details
pip install six numpy==1.19.4 wheel keras_applications keras_preprocessing --no-deps

$Env:BAZEL_VC = "C:\Program Files (x86)\Microsoft Visual Studio\2019\BuildTools\VC"
# Adjust the Compute Capabilities to match your target GPU!
$Env:TF_CUDA_COMPUTE_CAPABILITIES = "7.5"
$Env:CC_OPT_FLAGS = "/arch:AVX2"
$Env:TF_NEED_CUDA = "1"

# Check the prompts but should be just enter for everything
.\configure.cmd

bazel build --config=opt --define=no_tensorflow_py_deps=true //tensorflow/tools/pip_package:build_pip_package

# Generates tensorflow_2_4_py39.whl
bazel-bin\tensorflow\tools\pip_package\build_pip_package tensorflow_2_4_py39

# Create a new venv and install the new wheel, test it's health with
# python -c "import tensorflow as tf;print(tf.reduce_sum(tf.random.normal([1000, 1000])));print(tf.config.list_physical_devices('GPU'))" 

Sorry but AFAIK protobuf does not support Python 3.9 properly with their Linux wheels. See protocolbuffers/protobuf#7978.

Ah, I think I also left out a few bits here...
Because grpcio version that tensorflow depends on is not having the python 3.9 wheels (but newer version has), my setup to install the wheels is twofold:

# Requirements from https://github.com/tensorflow/tensorflow/blob/master/tensorflow/tools/pip_package/setup.py
pip install "grpcio==1.34.0" "absl-py~=0.10" "astunparse~=1.6.3" "flatbuffers~=1.12.0" "google_pasta~=0.2" "h5py~=3.1.0" "keras_preprocessing~=1.1.2" "numpy~=1.19.2" "opt_einsum~=3.3.0" "protobuf>=3.9.2" "six~=1.15.0" "termcolor~=1.1.0" "typing_extensions~=3.7.4" "wheel~=0.35" "wrapt~=1.12.1" "gast==0.3.3"

pip install --no-deps C:\tmp\tensorflow-2.4.0-cp39-cp39-win_amd64.whl

This makes it easier to also manage the protobuf installation separately. At least for Windows there is an unofficial Python 3.9 protobuf wheel at https://www.lfd.uci.edu/~gohlke/pythonlibs/#protobuf (not the latest but good enough, 3.12.3, ).

protobuf and gast are current blockers to have 100% tests passing

protobuf and gast are current blockers to have 100% tests passing

gast 0.4.0 seems to support Python 3.9. Haven't tested it yet though.

@willemavjc
only problem blocked me was grpcio 1.32.0, which I ignored using --no-dependency option

I've built tensorflow from source of course

cuda 11.2
cudnn 8.0.5
branch - r2.4

protobuf and gast are current blockers to have 100% tests passing

gast 0.4.0 seems to support Python 3.9. Haven't tested it yet though.

It does, but I think @mdanatg has a change to update and there are a few things that need to change to make all tests pass.

The change to upgrade to 0.4 is ready and just needs one final approval. We should be able to resolve this shortly after the approver is back in office, on Jan 5.

The change to upgrade to 0.4 is ready and just needs one final approval. We should be able to resolve this shortly after the approver is back in office, on Jan 5.

This kind of info is related, more in general, to a quite old open topic (since 2016) as you can see #46039 (comment)

And this specific issue was exposed from October (#44146) so we had two months to be exposed to concurrent contributions conflicts.

protocolbuffers/protobuf#8121

can you share the generated tensorflow-2.4.0-cp39-cp39-win_amd64.whl file?

When building for mac Big Sur it seems that you need to rename the whl from tensorflow-2.4.0-cp39-cp39-macosx_11_0_x86_64.whl to tensorflow-2.4.0-cp39-cp39-macosx_10_16_x86_64.whl, but after installing gast==0.4, hypy==3.1.0 and grpcio==1.34.0 everything is working perfectly (so far ;))

Any progress here? I would really appreciate being able to use Python 3.9.

Not yet. We should support py3.9 in nightly in about a few weeks.

Python 3.9.2 release is scheduled to happen in two weeks according to https://www.python.org/dev/peps/pep-0596

Perhaps work should start now on testing against Python 3.10a4 so the community does not face yet another 7-month delay.

Python 3.9.2 release is scheduled to happen in two weeks according to https://www.python.org/dev/peps/pep-0596

Perhaps work should start now on testing against Python 3.10a4 so the community does not face yet another 7-month delay.

Python 3.9 is released at October 5th, 2020. Or are you referring to something else by “7-month delay”?

TF has a large number of dependencies. We cannot make progress until all dependencies are updated.

Right now protobuf is still lacking a py39 wheel. This means TF is unable to make real progress anyway (as we need the py39 wheel due to the C++ protobuf interface)

We have to wait until protocolbuffers/protobuf#8121 lands and a new protobuf containing it (or similar) is released.

Now that protocolbuffers/protobuf#8121 landed, we just need to wait for the release.