Skip to content

Commit e02d49f

Browse files
authored
Merge pull request #1 from zhangtaolab/dev
Dev
2 parents 61682c0 + 184acc3 commit e02d49f

1 file changed

Lines changed: 12 additions & 1 deletion

File tree

docs/faq/models_troubleshooting.md

Lines changed: 12 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,3 +1,14 @@
11
# Models Troubleshooting
22

3-
Please see [models](../user_guide/models.md) section for detailed usage, any troubleshooting not recorded will be added here.
3+
#### Mamba models on macOS (Apple Silicon)
4+
5+
`mamba-ssm` relies on [Triton](https://github.com/triton-lang/triton), which only provides pre-built wheels for Linux. As a result, the `[mamba]` extra cannot be installed on macOS.
6+
7+
**Workaround:** HuggingFace `transformers` includes a pure-PyTorch Mamba implementation (`MambaModel`) that works on macOS without `mamba-ssm`. To use it:
8+
9+
```bash
10+
# Install without the [mamba] extra
11+
uv pip install -e '.[base,mcp,dev,notebook]'
12+
```
13+
14+
Then load a Mamba model through `transformers` (cpu only) as usual. Note that this fallback path is significantly slower than the optimized `mamba-ssm` kernels (which require Linux + CUDA).

0 commit comments

Comments
 (0)