mamba-ssm fails to build on torch==2.7.0 #720
Comments
|
I'm running into this problem too when installing mamba-ssm==2.2.4 as part of a Ray environment. |
works for me |
|
|
|
@tridao might it be better to simply remove torch from the dependencies in the toml? Typically, users install torch on their own based on their environment, but having torch as a dependency leads to torch installing the latest torch during the build, and often results in conflicts. I believe this is why the --no-build-isolation works. By removing torch, I don't think there would be a need for this isolation flag, and a lot of headaches could be avoided. Does it make sense? What to others think? |
|
I was able to install it with |
|
I'm also having trouble installing mamba. |
|
Same issue 😢 |
davidxia
added a commit
to davidxia/vllm
that referenced
this issue
May 16, 2025
* make it apparent the set up instructions are for Nvidia CUDA * the `pip install` command requires the PyTorch index * show the workaround for state-spaces/mamba#720 * install the `python3-dev` Debian package that's required for building mamba-ssm wheel from source Signed-off-by: David Xia <[email protected]>
dtrifiro
added a commit
to dtrifiro/mamba
that referenced
this issue
May 19, 2025
the sdist build currently doesn't include the `csrc` directory, making the PyPi sdist impossible to build. fixes state-spaces#720
Open
|
I don't believe this has to do with the torch version. Testing out a build with uv: Results in the following logs: uv pip install -v mamba-ssm==2.2.4The short summary of this is that:
Resulting in:
So, merging #727 will fix this specific error by providing pre-built wheels for torch==2.7.0, although it will not solve the underlying issue, which is that the #729 fixes this. |
dtrifiro
added a commit
to dtrifiro/mamba
that referenced
this issue
May 19, 2025
the sdist build currently doesn't include the `csrc` directory, making the PyPi sdist impossible to build. fixes state-spaces#720
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
This causes the package to fail to install by default since the torch version is unpinned.
It works if i disable build isolation with torch==2.6.0 already installed i.e.
uv pip install mamba-ssm==2.2.4 --no-build-isolation, so we have to patch the vLLM installation manually vllm-project/vllm#17070Found in the vllm ci https://buildkite.com/vllm/ci/builds/18314/steps?jid=01966392-20ef-480d-af0d-9d14c5b0e699#01966392-20ef-480d-af0d-9d14c5b0e699/130-10994
The text was updated successfully, but these errors were encountered: