Instructions for Mac?

[JohannesBertens]

Due diligence

• I have done my due diligence in trying to find the answer myself.

Topic

The PyTorch implementation

Question

~/projects/personaplex/ [main] pip install moshi/.           
Processing ./moshi
  Installing build dependencies ... done
  Getting requirements to build wheel ... done
  Preparing metadata (pyproject.toml) ... done
Requirement already satisfied: numpy<2.2,>=1.26 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from moshi-personaplex==0.1.0) (2.1.3)
Requirement already satisfied: safetensors<0.5,>=0.4.0 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from moshi-personaplex==0.1.0) (0.4.5)
Requirement already satisfied: huggingface-hub<0.25,>=0.24 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from moshi-personaplex==0.1.0) (0.24.7)
Requirement already satisfied: einops==0.7 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from moshi-personaplex==0.1.0) (0.7.0)
Requirement already satisfied: sentencepiece==0.2 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from moshi-personaplex==0.1.0) (0.2.0)
Requirement already satisfied: sounddevice==0.5 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from moshi-personaplex==0.1.0) (0.5.0)
Requirement already satisfied: sphn<0.2,>=0.1.4 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from moshi-personaplex==0.1.0) (0.1.12)
Collecting torch<2.5,>=2.2.0 (from moshi-personaplex==0.1.0)
  Using cached torch-2.4.1-cp312-none-macosx_11_0_arm64.whl.metadata (26 kB)
Requirement already satisfied: aiohttp<3.11,>=3.10.5 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from moshi-personaplex==0.1.0) (3.10.11)
Requirement already satisfied: CFFI>=1.0 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from sounddevice==0.5->moshi-personaplex==0.1.0) (2.0.0)
Requirement already satisfied: aiohappyeyeballs>=2.3.0 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from aiohttp<3.11,>=3.10.5->moshi-personaplex==0.1.0) (2.6.1)
Requirement already satisfied: aiosignal>=1.1.2 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from aiohttp<3.11,>=3.10.5->moshi-personaplex==0.1.0) (1.4.0)
Requirement already satisfied: attrs>=17.3.0 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from aiohttp<3.11,>=3.10.5->moshi-personaplex==0.1.0) (25.4.0)
Requirement already satisfied: frozenlist>=1.1.1 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from aiohttp<3.11,>=3.10.5->moshi-personaplex==0.1.0) (1.8.0)
Requirement already satisfied: multidict<7.0,>=4.5 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from aiohttp<3.11,>=3.10.5->moshi-personaplex==0.1.0) (6.7.0)
Requirement already satisfied: yarl<2.0,>=1.12.0 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from aiohttp<3.11,>=3.10.5->moshi-personaplex==0.1.0) (1.22.0)
Requirement already satisfied: filelock in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from huggingface-hub<0.25,>=0.24->moshi-personaplex==0.1.0) (3.20.2)
Requirement already satisfied: fsspec>=2023.5.0 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from huggingface-hub<0.25,>=0.24->moshi-personaplex==0.1.0) (2025.12.0)
Requirement already satisfied: packaging>=20.9 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from huggingface-hub<0.25,>=0.24->moshi-personaplex==0.1.0) (26.0)
Requirement already satisfied: pyyaml>=5.1 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from huggingface-hub<0.25,>=0.24->moshi-personaplex==0.1.0) (6.0.3)
Requirement already satisfied: requests in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from huggingface-hub<0.25,>=0.24->moshi-personaplex==0.1.0) (2.32.5)
Requirement already satisfied: tqdm>=4.42.1 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from huggingface-hub<0.25,>=0.24->moshi-personaplex==0.1.0) (4.67.1)
Requirement already satisfied: typing-extensions>=3.7.4.3 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from huggingface-hub<0.25,>=0.24->moshi-personaplex==0.1.0) (4.15.0)
Requirement already satisfied: sympy in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from torch<2.5,>=2.2.0->moshi-personaplex==0.1.0) (1.14.0)
Requirement already satisfied: networkx in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from torch<2.5,>=2.2.0->moshi-personaplex==0.1.0) (3.6.1)
Requirement already satisfied: jinja2 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from torch<2.5,>=2.2.0->moshi-personaplex==0.1.0) (3.1.6)
Requirement already satisfied: setuptools in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from torch<2.5,>=2.2.0->moshi-personaplex==0.1.0) (80.9.0)
Requirement already satisfied: idna>=2.0 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from yarl<2.0,>=1.12.0->aiohttp<3.11,>=3.10.5->moshi-personaplex==0.1.0) (3.11)
Requirement already satisfied: propcache>=0.2.1 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from yarl<2.0,>=1.12.0->aiohttp<3.11,>=3.10.5->moshi-personaplex==0.1.0) (0.4.1)
Requirement already satisfied: pycparser in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from CFFI>=1.0->sounddevice==0.5->moshi-personaplex==0.1.0) (3.0)
Requirement already satisfied: MarkupSafe>=2.0 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from jinja2->torch<2.5,>=2.2.0->moshi-personaplex==0.1.0) (3.0.3)
Requirement already satisfied: charset_normalizer<4,>=2 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from requests->huggingface-hub<0.25,>=0.24->moshi-personaplex==0.1.0) (3.4.4)
Requirement already satisfied: urllib3<3,>=1.21.1 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from requests->huggingface-hub<0.25,>=0.24->moshi-personaplex==0.1.0) (2.6.3)
Requirement already satisfied: certifi>=2017.4.17 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from requests->huggingface-hub<0.25,>=0.24->moshi-personaplex==0.1.0) (2026.1.4)
Requirement already satisfied: mpmath<1.4,>=1.1.0 in /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages (from sympy->torch<2.5,>=2.2.0->moshi-personaplex==0.1.0) (1.3.0)
Using cached torch-2.4.1-cp312-none-macosx_11_0_arm64.whl (62.1 MB)
Building wheels for collected packages: moshi-personaplex
  Building wheel for moshi-personaplex (pyproject.toml) ... done
  Created wheel for moshi-personaplex: filename=moshi_personaplex-0.1.0-py3-none-any.whl size=86111 sha256=0b5468c3ccd6aea73b7c84a8f57d030fb015083c4f1db1b75e8b5b0ec8d18283
  Stored in directory: /private/var/folders/dc/fd5s9xc12ndg0pqh1sg9_jnc0000gp/T/pip-ephem-wheel-cache-3iyvij17/wheels/b7/62/43/fb9ac425f43dacebce3ddae3bdef7dfea5a9b8da5be53b23c1
Successfully built moshi-personaplex
Installing collected packages: torch, moshi-personaplex
  Attempting uninstall: torch
    Found existing installation: torch 2.9.1
    Uninstalling torch-2.9.1:
      Successfully uninstalled torch-2.9.1
  Attempting uninstall: moshi-personaplex
    Found existing installation: moshi-personaplex 0.1.0
    Uninstalling moshi-personaplex-0.1.0:
      Successfully uninstalled moshi-personaplex-0.1.0
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
torchvision 0.24.1 requires torch==2.9.1, but you have torch 2.4.1 which is incompatible.
Successfully installed moshi-personaplex-0.1.0 torch-2.4.1
 ~/projects/personaplex/ [main] 

Seems this failes - how can I best install this on a mac? Any way to enable the metal backend?

[gennadyterekhov]

i have another issue, but i think it's also related to macos (i dont have cuda)

> SSL_DIR=$(mktemp -d); python -m moshi.server --ssl "$SSL_DIR"

AssertionError: Torch not compiled with CUDA enabled

[rajarshiroy-nvidia]

Officially, this model does not support Mac. Leaving this issue open for others who might enable it for Mac using mlx.

[eliliam]

@rajarshiroy-nvidia do you know if there is any way to run it, even on the CPU, on a mac, or is a CUDA capable GPU a hard requirement? Even when using the --cpu-offload flag it errors with the same error:

$ SSL_DIR=$(mktemp -d); python -m moshi.server --ssl "$SSL_DIR" --cpu-offload
2026-01-23 13:21:37,359 - __main__ - INFO - retrieving voice prompts
2026-01-23 13:21:37,498 - __main__ - INFO - voice_prompt_dir = /Users/eli/.cache/huggingface/hub/models--nvidia--personaplex-7b-v1/snapshots/f69e74f60b25bded21b08fd932c334e47e28f184/voices
2026-01-23 13:21:37,498 - __main__ - INFO - retrieving the static content
2026-01-23 13:21:37,608 - __main__ - INFO - static_path = /Users/eli/.cache/huggingface/hub/models--nvidia--personaplex-7b-v1/snapshots/f69e74f60b25bded21b08fd932c334e47e28f184/dist
2026-01-23 13:21:37,730 - __main__ - INFO - loading mimi
Traceback (most recent call last):
  File "<frozen runpy>", line 198, in _run_module_as_main
  File "<frozen runpy>", line 88, in _run_code
  File "/Users/eli/Source/personaplex/venv/lib/python3.12/site-packages/moshi/server.py", line 483, in <module>
    main()
  File "/Users/eli/Source/personaplex/venv/lib/python3.12/site-packages/moshi/server.py", line 434, in main
    mimi = loaders.get_mimi(args.mimi_weight, args.device)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/eli/Source/personaplex/venv/lib/python3.12/site-packages/moshi/models/loaders.py", line 134, in get_mimi
    encoder_transformer = transformer.ProjectedTransformer(
                          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/eli/Source/personaplex/venv/lib/python3.12/site-packages/moshi/modules/transformer.py", line 745, in __init__
    self.transformer = StreamingTransformer(d_model=d_model, **kwargs)
                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/eli/Source/personaplex/venv/lib/python3.12/site-packages/moshi/modules/transformer.py", line 681, in __init__
    layer_class(
  File "/Users/eli/Source/personaplex/venv/lib/python3.12/site-packages/moshi/modules/transformer.py", line 506, in __init__
    self.self_attn: StreamingMultiheadAttention = StreamingMultiheadAttention(
                                                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/eli/Source/personaplex/venv/lib/python3.12/site-packages/moshi/modules/transformer.py", line 362, in __init__
    in_proj = nn.Linear(embed_dim, mult * out_dim, bias=False, **factory_kwargs)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/eli/Source/personaplex/venv/lib/python3.12/site-packages/torch/nn/modules/linear.py", line 99, in __init__
    self.weight = Parameter(torch.empty((out_features, in_features), **factory_kwargs))
                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/eli/Source/personaplex/venv/lib/python3.12/site-packages/torch/cuda/__init__.py", line 305, in _lazy_init
    raise AssertionError("Torch not compiled with CUDA enabled")
AssertionError: Torch not compiled with CUDA enabled

ETA: The README.md file shows instruction for macOS in the "Prerequisites" section

# macOS
brew install opus

If macOS is not an option, it would probably be a good idea to remove that and mention specifically that macOS is not supported.

[rajarshiroy-nvidia]

I think we should keep it because:

1. MLX backend implementation might be worked out by someone else.
2. If you are using offline.py for research experiments, you could install cpu-only pytorch and give it a try. It would be slow but it should work.

[TengoMan]

Hi!
Thanks for the repo!

I'm still working on getting the proper Python module versions together
What I'm curious about is the Docker Compose file, which references a .env file in the yaml, but there is no example or documentation about the environment variables.
Any insights?
Thanks in advance!

[sunnyden]

Have you tried this command?

export PYTORCH_ENABLE_MPS_FALLBACK=1  
SSL_DIR=$(mktemp -d); python -m moshi.server --ssl "$SSL_DIR" --device mps

You need to specify the mps as device to inference. Also there are some ops that is not supported on Apply Silicon by torch so you need to set env variable PYTORCH_ENABLE_MPS_FALLBACK=1 to fallback to use CPU for these ops(aten::index_copy.out). (Alternatively you can use my branch in #25 which fixes this issue and have improved performance on macOS)

I tried on my side and this command is working. But running this on macOS is pretty slow 🤕

[100tomer]

+1

[lfbgit]

I got it to run to my Macbook Pro M4 Max by quickly patching the moshi.server warmup so that it doesn't call CUDA on Mac.

Step by step to patch:

1. figure out where moshi is installed:
   python3 -c "import moshi, inspect, os; print(os.path.dirname(inspect.getfile(moshi)))"
   For me it prints: /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages/moshi

2. Open the moshi server.py file:
   nano /Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages/moshi/server.py

3. Change the warmup synchronization by finding the line:
   torch.cuda.synchronize()
   replace it with:

if torch.cuda.is_available():
    torch.cuda.synchronize()
elif torch.backends.mps.is_available():
    torch.mps.synchronize()

4. Now run it with the export PYTORCH_ENABLE_MPS_FALLBACK=1 and --device mps flag.

export PYTORCH_ENABLE_MPS_FALLBACK=1  
SSL_DIR=$(mktemp -d); python3 -m moshi.server --ssl "$SSL_DIR" --device mps

5. (Optional): If you get an SSL error, and don't care about using SSL, you can also run it without --ssl "$SSL_DIR":

export PYTORCH_ENABLE_MPS_FALLBACK=1  
python3 -m moshi.server --device mps

However, this is still quite slow, leading to delays. This is, therefore, only a temporary workaround to get it to run at all.

[Slowqueso]

@rajarshiroy-nvidia Is there a way, I can make it service duplex instead of the default full duplex system on Apple Silicon.

Additionally, would that improve the performance of the model?

[pholz]

sorry but if there is no official mac support why do you lead the README with brew install steps for mac systems?

[rajarshiroy-nvidia]

@pholz removed it. Thanks for highlighting that.