Litert_torch on Mac

Google AI Edge
1

Is litert_torch, formerly known as ai_edge_torch, currently NOT supported on Mac? I have an M4 chip. I’m able to successfully install litert_torch (v0.8.0) using uv, but when I try import litert_torch, I get the error:

ImportError: dlopen(/Users/brian.provost/work/test/.venv/lib/python3.12/site-packages/tensorflow/lite/python/metrics/_pywrap_tensorflow_lite_metrics_wrapper.so, 0x0002): Symbol not found: _Wrapped_PyInit__pywrap_tensorflow_lite_metrics_wrapper
  Referenced from: <B3DA96BD-AC59-3FF7-B6DD-A670BB6B54E1> /Users/brian.provost/work/test/.venv/lib/python3.12/site-packages/tensorflow/lite/python/metrics/_pywrap_tensorflow_lite_metrics_wrapper.so
  Expected in:     <562BFDD0-A53D-3FFD-A47C-7C2F81A1D900> /Users/brian.provost/work/test/.venv/lib/python3.12/site-packages/tensorflow/python/lib_pywrap_tensorflow_common.dylib
pyproject.toml 
[project]
name = "test"
version = "0.1.0"
description = "test repo"
readme = "README.md"
requires-python = ">=3.10,<3.13"
dependencies = [
    "ipykernel",
    "tensorflow",
    "litert_torch"
]

[tool.uv]
package = false
required-environments = ["sys_platform == 'darwin' and platform_machine == 'arm64'"]
Package versions 
Package                 Version
----------------------- ---------------------
absl-py                 2.4.0
ai-edge-litert          2.0.3
ai-edge-quantizer       0.4.0
ai-edge-tensorflow      2.21.0.dev20251110
anyio                   4.12.1
appnope                 0.1.4
asttokens               3.0.1
astunparse              1.6.3
backports-strenum       1.3.1
certifi                 2026.1.4
charset-normalizer      3.4.4
click                   8.3.1
comm                    0.2.3
debugpy                 1.8.20
decorator               5.2.1
executing               2.2.1
filelock                3.20.3
fire                    0.7.1
flatbuffers             25.12.19
fsspec                  2026.1.0
gast                    0.7.0
google-pasta            0.2.0
grpcio                  1.76.0
h11                     0.16.0
h5py                    3.14.0
hf-xet                  1.2.0
httpcore                1.0.9
httpx                   0.28.1
huggingface-hub         1.3.7
idna                    3.11
immutabledict           4.2.2
iniconfig               2.3.0
ipykernel               7.1.0
ipython                 9.10.0
ipython-pygments-lexers 1.1.1
jax                     0.9.0
jaxlib                  0.9.0
jaxtyping               0.3.7
jedi                    0.19.2
jinja2                  3.1.6
jupyter-client          8.8.0
jupyter-core            5.9.1
kagglehub               0.4.2
kagglesdk               0.1.15
keras                   3.13.2
keras-nightly           3.14.0.dev2026020104
libclang                18.1.1
litert-torch            0.8.0
markdown                3.10.1
markdown-it-py          4.0.0
markupsafe              3.0.3
matplotlib-inline       0.2.1
mdurl                   0.1.2
ml-dtypes               0.5.4
mpmath                  1.3.0
multipledispatch        1.0.0
namex                   0.1.0
nest-asyncio            1.6.0
networkx                3.6.1
numpy                   2.4.2
opt-einsum              3.4.0
optree                  0.18.0
packaging               26.0
parso                   0.8.5
pexpect                 4.9.0
pillow                  12.1.0
platformdirs            4.5.1
pluggy                  1.6.0
prompt-toolkit          3.0.52
protobuf                6.33.5
psutil                  7.2.2
ptyprocess              0.7.0
pure-eval               0.2.3
pygments                2.19.2
pytest                  8.4.2
python-dateutil         2.9.0.post0
pyyaml                  6.0.3
pyzmq                   27.1.0
regex                   2026.1.15
requests                2.32.5
rich                    14.3.2
safetensors             0.7.0
scipy                   1.17.0
setuptools              80.10.2
shellingham             1.5.4
six                     1.17.0
stack-data              0.6.3
sympy                   1.14.0
tabulate                0.9.0
tb-nightly              2.20.0a20250717
tensorboard             2.20.0
tensorboard-data-server 0.7.2
tensorflow              2.20.0
termcolor               3.3.0
tokenizers              0.22.2
torch                   2.9.0
torch-xla2              0.0.1.dev202412041639
torchao                 0.14.0
tornado                 6.5.4
tqdm                    4.67.2
traitlets               5.14.3
transformers            5.0.0
typer-slim              0.21.1
typing-extensions       4.15.0
urllib3                 2.6.3
wadler-lindig           0.1.7
wcwidth                 0.5.3
werkzeug                3.1.5
wheel                   0.46.3
wrapt                   2.1.0

Python version: 3.11.14

2

I was able to successfully import litert_torch using this pyproject.toml

[project]
name = "test"
version = "0.1.0"
description = "test repo"
readme = "README.md"
requires-python = ">=3.10,<3.13"
dependencies = [
    "ipykernel",
    "tensorflow; sys_platform != 'darwin' or platform_machine != 'arm64'",
    "tf_keras; sys_platform != 'darwin' or platform_machine != 'arm64'",
    "tf-nightly; sys_platform == 'darwin' and platform_machine == 'arm64'",
    "tflite",
    "litert_torch; sys_platform != 'darwin' or platform_machine != 'arm64'",
    "litert-torch-nightly; sys_platform == 'darwin' and platform_machine == 'arm64'",
]

[tool.uv]
package = false
required-environments = ["sys_platform == 'darwin' and platform_machine == 'arm64'"]
1 Like
3

Thank you for publishing that, @Brian_Provost !

Confirm that it works for my Apple M3. Successfully converted gemma-3-270m-it to .tflite format.

4

Hi @Brian_Provost, welcome here. Thanks for bringing this “binary ABI mismatch” issue for stable release v0.8.0 on apple silicon (M-series chipset). I spent some time and did this root cause analysis which community mightbe profited.


ImportError: ... Symbol not found: _Wrapped_PyInit__pywrap_tensorflow_lite_metrics_wrapper

This is the symbol mismatch. litert_torch relies on tf lite runtime for model exports and metrics. this error is comes from - macOS arm64 v0.8.0 tries to link it to std tensorflow pkg, but there were recent changes in TFLite C++ code set. the symbol names in std tensorflow wheel for macOS aren’t perfect match for what litert_torch v0.8.0 expecting, it tries to find the _Wrapped_PyInit… but that is not found in TFlite that pip/uv builder resolved.

I will look for more, also now checking your work around. Will come back soon.

5

Hi @Damyr_Hadiiev, big welcome here, Thanks for the info and trying the workaround. Great great !!

6

your working fix works best. Good sync as the nightly builds are built against exact same commit of tensorflow head as tf-nightly. every c++ symbol including that metrics will align just perfect.


sys_platform == 'darwin' and platform_machine == 'arm64'

this is helping fix macos without any disturbance to other OS. I can say you’re just supperb, stay at nightly until release of LiteRT Torch next stable release. Don’t forget to watch out for uv caching when you update any tf-nightly or litert-torch-nightly without updating the other, this error can give a surprise visit to you again. keep it sync. Thanks again.