Skip to content

WINDOWS.md

Latest commit

 

History

History
75 lines (67 loc) · 5.68 KB

WINDOWS.md

File metadata and controls

75 lines (67 loc) · 5.68 KB
Raw

Maya on Windows

Back to the README.

The same installer sets Maya up natively on Windows 10/11 (64-bit), and the engine and the image encoder compile there with Visual Studio 2022 and CUDA 12.8 (CUDA 13 works for RTX 20 and newer). Maya is developed and measured on Linux; on Windows 11 a user runs Maya-L on an RTX 4090, built with CUDA 13.4 (the speed table above, #63). Tell us how it runs on yours.

  1. Install once:

    • The NVIDIA driver and CUDA Toolkit 12.8. CUDA 13 also works for RTX 20 and newer; V100 needs CUDA 12.x. The CUDA version shown by nvidia-smi is the driver's capability, not proof that the toolkit (nvcc) is installed.
    • Visual Studio 2022 Build Tools, with Desktop development with C++, MSVC v143 - VS 2022 C++ x64/x86 build tools and a Windows 10 or 11 SDK. The full Visual Studio IDE is optional. VS 2026 Build Tools work too with CUDA 13.2 or newer - which Visual Studio a CUDA release supports is in its installation guide (13.2 onward list VS 2026's MSVC 195x, 13.1 and every 12.x stop at VS 2022); with a CUDA 12.x toolkit, 2026 alone does not satisfy the installer's check; 2022 can be installed alongside it.
    • 64-bit Python 3.10+, for example winget install -e --id Python.Python.3.12 --scope user.
    • Git for Windows to clone and update the repository, or download and extract its source zip instead.

  2. Open a new PowerShell or Command Prompt after installing the prerequisites. Clone the repository (or open the extracted folder), then check the PC:

    git clone https://github.com/mw00/project-maya.git
cd project-maya
.\START-MAYA.bat --check

    --check reports the GPUs, toolkit, compiler, RAM and page file without downloading the model or building the engine. Maya finds Visual Studio's compiler environment itself; a Developer Command Prompt is not needed.

  3. Start the setup, putting the model on an NVMe SSD with enough free space for the chosen model. Change the drive and folder to suit your PC:

    .\START-MAYA.bat --setup --models-dir "D:\Maya-models"

    Choose the model and context when asked. The launcher creates its own .venv; setup installs its Python packages, CMake and Ninja there, builds the engine, and asks before downloading the model. Later, run .\START-MAYA.bat (or double-click it) to start the installed model, then open http://127.0.0.1:8080. All ./maya.sh options also work here; PowerShell needs the leading .\. Add --plain for plain-text setup.

  • Check the page-file size, even with "System managed" enabled. Windows' GPU allocations and Maya's pinned RAM tier consume commit capacity (RAM + page file). A small page file can keep the RAM tier below the free RAM available, leaving more experts on the SSD. --check recommends roughly the selected GPUs' total VRAM plus 9 GB of page-file capacity (about 57 GB for a 48 GB card). If it warns, set a custom size at least as large as it recommends, with enough free disk space: System > About > Advanced system settings > Performance > Settings > Advanced > Virtual memory. Restart Windows after changing it. "System managed" can remain too small on a large-RAM PC (#20). On an AMD APU (Strix Halo, Gorgon Halo) the page file needs to be at least about the GPU's share of the memory: see AMD_MAYA.md ("The page file on a Windows APU").
  • A GPU filled to the brim pages instead of failing. Windows' driver moves GPU memory to system RAM when the card runs out, and decode then falls to a few tokens per second. The expert pool checks the free VRAM again once it is written (--expert-cache auto, the default), and the engine log says at start how much is free on each GPU (LOW below 256 MiB). A drafter (DFlash2) takes a few hundred MB more on its first request. If it warns, or answers slow down after the first one, add "--vram-reserve-mib", "1500" (say) to the config's "args", or set up again with --vram-reserve-mib 1500 (the expert pool's size).
  • With several CUDA toolkits installed, setup chooses the newest compatible one. To select one explicitly, add --nvcc "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.8\bin\nvcc.exe" to the setup command.
  • After setup, stop any running Maya server and run .\START-MAYA.bat --calibrate --no-start to tune the CPU/PCIe split and thread count for your PC. Then start Maya normally. See Tuning for the optional RAM-resident tier and other settings, and Something went wrong? for --bench / --report.
  • Not WSL2: WSL2's GPU driver pins only about 1 GB of RAM, and Maya pins tens of GB. Run START-MAYA.bat in Windows itself.
  • A Tesla V100 runs it too, built with CUDA 12.4 (#57). Run --calibrate once: it tunes Maya for your PC, and a setup for another context keeps its settings.
  • AMD on Windows, Strix Halo / Gorgon Halo included: START-MAYA.bat --backend hip --gpu 0 --setup builds the HIP engine with AMD's HIP SDK, or offers AMD's ROCm SDK wheels in .venv when there is none (docs/AMD_MAYA.md). tools\hip\build_maya_windows.bat builds it by hand (an RX 7900 XTX, #54).