Windows - mamba-ssm installation fails with missing triton dependency and --no-build-isolation failure

[jimmypark4]

Here's the revised GitHub issue including the error logs you provided:

---

Title:

🚨 [Bug] Windows - mamba-ssm installation fails with missing triton dependency and --no-build-isolation failure

---

Environment:

• OS: Windows 11 Pro
• Python version: 3.11.5
• CUDA version: 12.4
• PyTorch version: 2.5.1+cu121
• Torchvision version: 0.20.1+cu121
• Torchaudio version: 2.5.1+cu121
• Visual Studio: VS 2022 Community 14.39.33519
• CUDA Compiler (nvcc): 12.4.131
• Conda Environment: MasterThesis

---

Description of the Issue:

I am attempting to install mamba-ssm on a Windows machine with a correctly configured CUDA and PyTorch environment. However, the installation fails with a missing triton dependency and build errors, even when using the --no-build-isolation and --no-deps flags.

---

Steps to Reproduce:

1. Create a fresh Conda environment:

   conda create -n test_env python=3.11
   conda activate test_env

2. Confirm CUDA installation:

   nvcc --version  # CUDA 12.4 confirmed

3. Install PyTorch with CUDA support:

   pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

4. Attempt to install mamba-ssm using --no-build-isolation:

   pip install mamba-ssm --no-build-isolation --no-deps

5. Clone and install directly from GitHub:

   git clone https://github.com/state-spaces/mamba.git
   cd mamba
   pip install . --no-build-isolation --no-deps

---

Observed Errors:

Error log during installation with --no-build-isolation:

Building wheels for collected packages: mamba-ssm
  Building wheel for mamba-ssm (pyproject.toml) ... error
  error: subprocess-exited-with-error

  × Building wheel for mamba-ssm (pyproject.toml) did not run successfully.
  │ exit code: 1
  ╰─> [75 lines of output]
      torch.__version__  = 2.5.1+cu121
      running bdist_wheel
      UserWarning: Attempted to use ninja as the BuildExtension backend but we could not find ninja.. Falling back to using the slow distutils backend.
      Guessing wheel URL:  https://github.com/state-spaces/mamba/releases/download/v2.2.4/mamba_ssm-2.2.4+cu12torch2.5cxx11abiFALSE-cp311-cp311-win_amd64.whl
      Precompiled wheel not found. Building from source...
      ...
      c1xx: fatal error C1083: Cannot open source file: 'csrc/selective_scan/selective_scan.cpp': No such file or directory
      error: command 'cl.exe' failed with exit code 2

---

Key Issues Noted:

• triton dependency could not be resolved.
• Source files (csrc/selective_scan/selective_scan.cpp) missing in the build.
• The --no-build-isolation flag does not prevent the error.
• Potential version mismatch:

  The detected CUDA version (12.4) has a minor version mismatch with the version that was used to compile PyTorch (12.1). 
  Most likely this shouldn't be a problem.
  

---

What I've Tried:

• ✅ Confirmed CUDA installation and compatibility (nvcc 12.4).
• ✅ Verified torch and torch.cuda.is_available() return True.
• ✅ Installed triton manually:

  pip install triton
  pip install triton==2.1.0

• ✅ Attempted installation from both PyPI and the GitHub repository.
• ✅ Upgraded setuptools and wheel:

  pip install --upgrade setuptools wheel

---

Expected Behavior:

• pip install mamba-ssm should correctly build the package without missing source files.
• The --no-build-isolation flag should prevent the CPU version of PyTorch from being reinstalled.

---

Possible Causes Identified:

• The csrc directory may be missing or improperly referenced in the pyproject.toml.
• triton might not be fully compatible with Windows or the latest CUDA version.
• The --no-build-isolation flag might not be functioning correctly in the setuptools configuration.

---

Request for Assistance:

• Is the mamba-ssm package officially supported for Windows?
• Should I consider downgrading CUDA or PyTorch for compatibility?
• Is there an alternative way to bypass the triton dependency for a CPU-only installation?

---

System Information (pip freeze):

filelock==3.13.1
fsspec==2024.2.0
Jinja2==3.1.3
MarkupSafe==2.1.5
mpmath==1.3.0
networkx==3.2.1
numpy==2.2.1
pillow==10.2.0
sympy==1.13.1
torch==2.5.1+cu121
torchaudio==2.5.1+cu121
torchvision==0.20.1+cu121
typing_extensions==4.9.0

---

🙏 Thank you for your time and assistance!

Please let me know if you need any additional information or test results.

---

[jimmypark4]

Problem-Solving Summary (for GitHub)

• Development Environment Change: Migrated from PyCharm to VS Code.
• WSL Usage: Utilized VS Code's WSL Remote feature to work in a Linux environment.
• Repository Cloning: Cloned the Mamba repository using git clone https://github.com/state-spaces/mamba.git.
• Package Installation: Installed the package with pip install . --no-cache-dir --verbose to avoid cache and enable verbose mode.

➡️ Successfully transitioned to VS Code + WSL and completed the Mamba package installation.