It's simply a pain to build it regardless of which version of CUDA, gcc compiler, etc. you have

[korvin2000]

I have tried a lot of options and version in order to build it, but...
It's simply a pain to build it regardless of which version of CUDA you have
I you want versatility for any platform - Don't use mamba, looks at mambaout or other project, because compiling mamba it's a real PAIN

99% of issues - people can't install it, people can't build it, people can't use it
I am using arch linux
I have downgraded my cuda from 13.1 to 12.8
I have switched to python 3.10
I have switched gcc compiler to 14
I have downgraded my pytoch to 2.4
I have tried different build options, nothing helps and nothing works

[nimuh]

Building for arch linux is a pain but possible. I used:
CUDA 12.4
gcc 13

The main thing on my end was having all the right system arguments when building.

Maybe try it with the above versions and tell us what the error is :)

[korvin2000]

yeah, thanks, I have successfully build it,
but have still some other dep's incompatibilities
it's makes my crazy, sorry

now i have:
/lib/python3.10/site-packages/mamba_ssm/ops/selective_scan_interface.py", line 20, in
import selective_scan_cuda

/lib/python3.10/site-packages/selective_scan_cuda.cpython-310-x86_64-linux-gnu.so: undefined symbol: _ZN3c104cuda29c10_cuda_check_implementationEiPKcS2_jb

I hope sometimes anyone who will use mamba in it's arches / projects will be banned.
sorry, but ... i already hate this project

[juneleung]

Traceback (most recent call last):
  File "/home/juneleung/Desktop/vol_juneleung/test_mamba.py", line 2, in <module>
    from mamba_ssm import Mamba
  File "/vol/grid-solar/sgeusers/juneleung/env/miniforge3/envs/jl310/lib/python3.10/site-packages/mamba_ssm/__init__.py", line 3, in <module>
    from mamba_ssm.ops.selective_scan_interface import selective_scan_fn, mamba_inner_fn
  File "/vol/grid-solar/sgeusers/juneleung/env/miniforge3/envs/jl310/lib/python3.10/site-packages/mamba_ssm/ops/selective_scan_interface.py", line 20, in <module>
    import selective_scan_cuda
ImportError: /vol/grid-solar/sgeusers/juneleung/env/miniforge3/envs/jl310/lib/python3.10/site-packages/selective_scan_cuda.cpython-310-x86_64-linux-gnu.so: undefined symbol: _ZN3c104cuda29c10_cuda_check_implementationEiPKcS2_jb

same error, also on Arch Linux...

[ibitec7]

Seems like you guys aren't installing the right wheel. The error of undefined symbol you get is because the environment the wheel was compiled for is different from the environment that Pytorch was compiled on your local machine.

You guys are probably using pip install mamba_ssm, but the thing is that pip does not see the compile environment of your machine like the CUDA version and whether Pytorch was compiled using an ABI or not?

I have downloaded mamba dozens of times and works all the time. Go to the release page: releases

pip install the correct Wheel for your environment from there. for e.g:

mamba_ssm-2.2.5+cu12torch2.6cxx11abiTRUE-cp311-cp311-linux_x86_64.whl

1. "..cu12torch2.6..": make sure you have correct Pytorch version installed with CUDA 12.0+. Make sure you compile it with the right version, you can't have cuda-toolkit 12.8, and have Pytorch 2.6 compiled with 12.4 or 12.6.
2. "..cxx11abiTRUE..": check if your Pytorch was compiled with C++11 ABI or not. Check the output of running this in the terminal: python -c "import torch; print(torch._C._GLIBCXX_USE_CXX11_ABI)", if it outputs True, then choose TRUE, else FALSE.
3. "..cp311..": CPython/Python version, make sure you have the right version, check by running python --version

Once you know what wheel to download, run something like this:

pip install --no-cache-dir https://github.com/state-spaces/mamba/releases/download/v2.2.5/<YOUR_WHEEL>

Once you install it this way and import a function, you won't get the undefined symbol error.

Hope this helps.

[luckyfanfanfan]

看来你们安装的 wheel 文件不对。出现“未定义符号”的错误是因为 wheel 文件编译时的环境与你们本地机器上 PyTorch 的编译环境不同。

你们可能正在使用pip install mamba_ssm，但问题是 pip 无法识别你们机器的编译环境，例如 CUDA 版本以及 PyTorch 是否使用 ABI 编译？

我下载过Mamba几十次了，每次都能正常运行。请访问发布页面：releases

从那里使用 pip 安装适合您环境的 Wheel 包。例如：

mamba_ssm-2.2.5+cu12torch2.6cxx11abiTRUE-cp311-cp311-linux_x86_64.whl

1. "..cu12torch2.6..": 请确保您已安装正确的 PyTorch 版本，并且 CUDA 版本为 12.0 或更高。请确保您使用正确的版本进行编译，您不能同时拥有 cuda-toolkit 12.8 和使用 12.4 或 12.6 版本编译的 PyTorch 2.6。
2. "..cxx11abiTRUE..": 检查你的 PyTorch 是否使用 C++11 ABI 编译。在终端运行以下命令并检查输出：python -c "import torch; print(torch._C._GLIBCXX_USE_CXX11_ABI)"，如果输出 True，则选择 TRUE，否则选择 FALSE。
3. "..cp311.": ​​CPython/Python 版本，请确保您使用的是正确的版本，可通过运行以下命令进行检查：python --version

确定要下载哪个轮毂后，运行类似这样的命令：

pip install --no-cache-dir https://github.com/state-spaces/mamba/releases/download/v2.2.5/<YOUR_WHEEL>

按照这种方式安装并导入函数后，就不会出现未定义符号错误了。

希望这能帮到你。

同学我按照这种方法安装，还是出现了“未定义符号”的错误，令人抓狂

[tridao]

can you try the latest version 2.3.1?

[Biobiofrost]

Seems like you guys aren't installing the right wheel. The error of undefined symbol you get is because the environment the wheel was compiled for is different from the environment that Pytorch was compiled on your local machine.

You guys are probably using pip install mamba_ssm, but the thing is that pip does not see the compile environment of your machine like the CUDA version and whether Pytorch was compiled using an ABI or not?

I have downloaded mamba dozens of times and works all the time. Go to the release page: releases

pip install the correct Wheel for your environment from there. for e.g:

mamba_ssm-2.2.5+cu12torch2.6cxx11abiTRUE-cp311-cp311-linux_x86_64.whl

1. "..cu12torch2.6..": make sure you have correct Pytorch version installed with CUDA 12.0+. Make sure you compile it with the right version, you can't have cuda-toolkit 12.8, and have Pytorch 2.6 compiled with 12.4 or 12.6.
2. "..cxx11abiTRUE..": check if your Pytorch was compiled with C++11 ABI or not. Check the output of running this in the terminal: python -c "import torch; print(torch._C._GLIBCXX_USE_CXX11_ABI)", if it outputs True, then choose TRUE, else FALSE.
3. "..cp311..": CPython/Python version, make sure you have the right version, check by running python --version

Once you know what wheel to download, run something like this:

pip install --no-cache-dir https://github.com/state-spaces/mamba/releases/download/v2.2.5/<YOUR_WHEEL>

Once you install it this way and import a function, you won't get the undefined symbol error.

Hope this helps.

root@b6b6ae7d791e:/# python - <<EOF
import torch
import mamba_ssm

print("torch:", torch.version)
print("cuda available:", torch.cuda.is_available())
print("ABI:", torch._C._GLIBCXX_USE_CXX11_ABI)
print("mamba imported successfully")
EOF
Traceback (most recent call last):
File "", line 2, in
File "/opt/conda/lib/python3.11/site-packages/mamba_ssm/init.py", line 3, in
from mamba_ssm.ops.selective_scan_interface import selective_scan_fn, mamba_inner_fn
File "/opt/conda/lib/python3.11/site-packages/mamba_ssm/ops/selective_scan_interface.py", line 20, in
import selective_scan_cuda
ImportError: /opt/conda/lib/python3.11/site-packages/selective_scan_cuda.cpython-311-x86_64-linux-gnu.so: undefined symbol: _ZN3c107WarningC1ESt7variantIJNS0_11UserWarningENS0_18DeprecationWarningEEERKNS_14SourceLocationENSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEEb
root@b6b6ae7d791e:/# python - <<EOF
import torch

print("torch:", torch.version)
print("cuda available:", torch.cuda.is_available())
print("ABI:", torch._C._GLIBCXX_USE_CXX11_ABI)
EOF

torch: 2.6.0+cu124
cuda available: True
ABI: False
root@b6b6ae7d791e:/#
root@b6b6ae7d791e:/# exit
exit
root@localhost:~/eegmamba_build# cat Dockerfile

For build

FROM pytorch/pytorch:2.6.0-cuda12.4-cudnn9-devel

WORKDIR /app
ENV CUDA_HOME=/usr/local/cuda
ENV LD_LIBRARY_PATH=/usr/local/cuda/lib64:${LD_LIBRARY_PATH}

RUN sed -i 's/archive.ubuntu.com/mirrors.tuna.tsinghua.edu.cn/g' /etc/apt/sources.list &&
sed -i 's/security.ubuntu.com/mirrors.tuna.tsinghua.edu.cn/g' /etc/apt/sources.list

RUN apt-get update && apt-get install -y build-essential python3-dev git ninja-build cmake &&
rm -rf /var/lib/apt/lists/*

Install mamba, ensure correct versions

Pytorch: 2.6.0

Cuda: 12.0+

Python/CPython: 3.11

RUN pip install --no-cache-dir
https://github.com/state-spaces/mamba/releases/download/v2.2.5/mamba_ssm-2.2.5+cu12torch2.6cxx11abiFALSE-cp311-cp311-linux_x86_64.whl

COPY requirements.txt requirements.txt

RUN pip install --no-cache-dir -r requirements.txt --prefer-binary

COPY . .

This is confusing. I copied your dockerfile.The pulled base image has been tested, and its cxxabi test is FALSE.Therefore, I modified the whl of the cxxabi TRUE version that needs to be installed as the FALSE version. In the end, even if the environment and precompiled package requirements correspond, the installation failed. Could you please provide more dockerfile details about the successful installation?