I have tried everything, using copilot and grok and at least 15 hours, i've been trying to install xformers and simple can't. One thing or another is either too old, or too new. Got triton installed but then something else is bad. Tried to recompile it, no...
I had copilot write this - because none of it makes sense. The program that makes the AI is easy. But the part that makes it fast was designed by the devil.
I’ve spent over 15 hours trying to compile xFormers from source on Windows 11 using Torch 2.5.1+cu121, Python 3.11, and Visual Studio Build Tools 2022. My goal is to enable CUDA extensions for Easy Diffusion with full memory-efficient attention support. I installed all required components: MSVC v143, Windows 11 SDK, and confirmed cl.exe works. I also installed setuptools, wheel, cmake, and ninja, and verified that Torch 2.5.1+cu121 is active.
I used both the Developer Command Prompt and the x64 Native Tools Command Prompt for VS 2022. I set the required environment variables (DISTUTILS_USE_SDK=1 and PIP_NO_BUILD_ISOLATION=1) to prevent MSVC reinitialization and Torch version override. Despite this, every time I ran pip install -e ., pip reinstalled Torch 2.9.0 due to pyproject.toml constraints, breaking compatibility with my CUDA stack. I confirmed that xFormers expects Torch 2.0.1 or 2.8.0, and no prebuilt wheel exists for Torch 2.5.1+cu121 on Python 3.11 for Windows.
The build consistently failed with a compiler error: error C3861: '_addcarry_u64': identifier not found. This intrinsic requires x64 compilation and MSVC 14.29+, which I confirmed I have. The error persisted even in the x64 Native Tools shell, suggesting xFormers is still compiling some components in x86 mode or the source is not guarded for platform-specific intrinsics. I also confirmed that the build isolation environment was re-triggering downloads of large packages (Torch, Sympy, etc.) every time, despite the environment variables.
At this point, I’ve exhausted all clean build paths. I’m requesting either (1) a prebuilt xFormers wheel for Torch 2.5.1+cu121 on Python 3.11 for Windows, (2) a patch to bypass _addcarry_u64 or guard it for x64-only builds, or (3) guidance on disabling xFormers in Easy Diffusion without breaking core functionality. I’m happy to test any workaround or validate a patched build. Let me know what’s needed to move this forward.
Any ideas, hope that made sense. It had me install and put in so many forms of 2.5-3.5gb files I think I transferred around a 100gigs and I gave up, i'm going in circles now. In a loop of this works and then that fails or its in but parts don't work.
I would much rather this was integrated in there and just working. Thanks.
I have tried everything, using copilot and grok and at least 15 hours, i've been trying to install xformers and simple can't. One thing or another is either too old, or too new. Got triton installed but then something else is bad. Tried to recompile it, no...
I had copilot write this - because none of it makes sense. The program that makes the AI is easy. But the part that makes it fast was designed by the devil.
I’ve spent over 15 hours trying to compile xFormers from source on Windows 11 using Torch 2.5.1+cu121, Python 3.11, and Visual Studio Build Tools 2022. My goal is to enable CUDA extensions for Easy Diffusion with full memory-efficient attention support. I installed all required components: MSVC v143, Windows 11 SDK, and confirmed cl.exe works. I also installed setuptools, wheel, cmake, and ninja, and verified that Torch 2.5.1+cu121 is active.
I used both the Developer Command Prompt and the x64 Native Tools Command Prompt for VS 2022. I set the required environment variables (DISTUTILS_USE_SDK=1 and PIP_NO_BUILD_ISOLATION=1) to prevent MSVC reinitialization and Torch version override. Despite this, every time I ran pip install -e ., pip reinstalled Torch 2.9.0 due to pyproject.toml constraints, breaking compatibility with my CUDA stack. I confirmed that xFormers expects Torch 2.0.1 or 2.8.0, and no prebuilt wheel exists for Torch 2.5.1+cu121 on Python 3.11 for Windows.
The build consistently failed with a compiler error: error C3861: '_addcarry_u64': identifier not found. This intrinsic requires x64 compilation and MSVC 14.29+, which I confirmed I have. The error persisted even in the x64 Native Tools shell, suggesting xFormers is still compiling some components in x86 mode or the source is not guarded for platform-specific intrinsics. I also confirmed that the build isolation environment was re-triggering downloads of large packages (Torch, Sympy, etc.) every time, despite the environment variables.
At this point, I’ve exhausted all clean build paths. I’m requesting either (1) a prebuilt xFormers wheel for Torch 2.5.1+cu121 on Python 3.11 for Windows, (2) a patch to bypass _addcarry_u64 or guard it for x64-only builds, or (3) guidance on disabling xFormers in Easy Diffusion without breaking core functionality. I’m happy to test any workaround or validate a patched build. Let me know what’s needed to move this forward.
Any ideas, hope that made sense. It had me install and put in so many forms of 2.5-3.5gb files I think I transferred around a 100gigs and I gave up, i'm going in circles now. In a loop of this works and then that fails or its in but parts don't work.
I would much rather this was integrated in there and just working. Thanks.