[Installation]: compile error C2719: 'unnamed-parameter': formal parameter with requested alignment of 128 won't be aligned

[sqweek]

Your current environment

Collecting environment information...
==============================
        System Info
==============================
OS                           : Could not collect
GCC version                  : Could not collect
Clang version                : Could not collect
CMake version                : version 4.2.3
Libc version                 : N/A

==============================
       PyTorch Info
==============================
PyTorch version              : 2.10.0+cu130
Is debug build               : False
CUDA used to build PyTorch   : 13.0
ROCM used to build PyTorch   : N/A

==============================
      Python Environment
==============================
Python version               : 3.12.11 (main, Aug 18 2025, 19:17:54) [MSC v.1944 64 bit (AMD64)] (64-bit runtime)
Python platform              : Windows-11-10.0.26200-SP0

==============================
       CUDA / GPU Info
==============================
Is CUDA available            : True
CUDA runtime version         : 13.1.115
CUDA_MODULE_LOADING set to   :
GPU models and configuration : GPU 0: NVIDIA GeForce RTX 5070 Ti
Nvidia driver version        : 595.71
cuDNN version                : Could not collect
HIP runtime version          : N/A
MIOpen runtime version       : N/A
Is XNNPACK available         : True

==============================
          CPU Info
==============================
'wmic' is not recognized as an internal or external command,
operable program or batch file.

==============================
Versions of relevant libraries
==============================
[pip3] numpy==2.4.1
[pip3] torch==2.10.0+cu130
[pip3] torchaudio==2.10.0+cu130
[pip3] torchsde==0.2.6
[pip3] torchvision==0.25.0+cu130
[pip3] transformers==5.0.0
[pip3] triton-windows==3.6.0.post25
[conda] Could not collect

==============================
         vLLM Info
==============================
ROCM Version                 : Could not collect
vLLM Version                 : 0.1.dev13975+g03f332b1c.d20260304 (git sha: 03f332b1c, date: 20260304)
vLLM Build Flags:
  CUDA Archs: Not Set; ROCm: Disabled
GPU Topology:
  Could not collect

==============================
     Environment Variables
==============================
CUDA_PATH=C:\Pkg\CUDA\13.1
CUDA_PATH_V13_1=C:\Pkg\CUDA\13.1
MAX_JOBS=6
VLLM_TARGET_DEVICE=cuda
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
TORCHINDUCTOR_CACHE_DIR=C:\Users\sqweek\AppData\Local\Temp\torchinductor_sqweek

How you are installing vllm

The failing command:

C:\Pkg\CUDA\13.1\bin\nvcc.exe -forward-unknown-to-host-compiler -DCUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1 -DPy_LIMITED_API=3 -DPy_NO_LINK_LIB -DQUTLASS_DISABLE_PYBIND=1 -DTARGET_CUDA_ARCH=120 -DTORCH_EXTENSION_NAME=_C -DUSE_C10D_GLOO -DUSE_DISTRIBUTED -D_C_EXPORTS -IC:\Env\vllm-windows\csrc -IC:\Env\vllm-windows\.deps\cutlass-src\include -IC:\Env\vllm-windows\.deps\cutlass-src\tools\util\include -IC:\Env\vllm-windows\.deps\qutlass-src -IC:\Env\vllm-windows\.deps\qutlass-src\qutlass -IC:\Env\vllm-windows\.deps\qutlass-src\qutlass\csrc\include -IC:\Env\vllm-windows\.deps\qutlass-src\qutlass\csrc\include\cutlass_extensions -isystem C:\Users\sqweek\AppData\Roaming\uv\python\cpython-3.12.11-windows-x86_64-none\Include -isystem C:\Env\cui\.venv\Lib\site-packages\torch\include -isystem C:\Env\cui\.venv\Lib\site-packages\torch\include\torch\csrc\api\include -isystem C:\Pkg\CUDA\13.1\include -D_WINDOWS -Xcompiler=" /EHsc" -DONNX_NAMESPACE=onnx_c2 --use-local-env -Xcudafe --diag_suppress=cc_clobber_ignored,--diag_suppress=field_without_dll_interface,--diag_suppress=base_class_has_different_dll_interface,--diag_suppress=dll_interface_conflict_none_assumed,--diag_suppress=dll_interface_conflict_dllexport_assumed,--diag_suppress=bad_friend_decl --Werror cross-execution-space-call --no-host-device-move-forward --expt-relaxed-constexpr --expt-extended-lambda -Xcompiler=" -O2 -Ob1" -DNDEBUG -Xcompiler /MD -std=c++17 -Xcompiler=-MD -Xcompiler=-Zi --expt-relaxed-constexpr -DENABLE_FP8 --threads=1 --compress-mode=size --disable-warnings -O2 -FS -Xptxas=-O2 -Xcompiler=/O2 -Xcompiler=/FS -Xcompiler=/Z7 -Xcompiler=/Zc:__cplusplus -Xcompiler=/DWIN32_LEAN_AND_MEAN -Xcompiler=/DUSE_CUDA -DENABLE_SCALED_MM_SM120=1 -DENABLE_SCALED_MM_C2X=1 -DENABLE_NVFP4_SM120=1 -DENABLE_CUTLASS_MOE_SM120=1 -DENABLE_CUTLASS_MLA=1 -gencode arch=compute_120f,code=sm_120f -MD -MT CMakeFiles\_C.dir\csrc\quantization\w8a8\cutlass\c3x\scaled_mm_sm120_fp8.cu.obj -MF CMakeFiles\_C.dir\csrc\quantization\w8a8\cutlass\c3x\scaled_mm_sm120_fp8.cu.obj.d -x cu -c C:\Env\vllm-windows\csrc\quantization\w8a8\cutlass\c3x\scaled_mm_sm120_fp8.cu -o CMakeFiles\_C.dir\csrc\quantization\w8a8\cutlass\c3x\scaled_mm_sm120_fp8.cu.obj -Xcompiler=-FdCMakeFiles\_C.dir\,-FS
nvcc warning : -Werror and -w are specified at the same time. NVCC will disable all warnings and not treat the warnings as errors.
scaled_mm_sm120_fp8.cu
cl : Command line warning D9025 : overriding '/Zi' with '/Z7'
scaled_mm_sm120_fp8.cu
cl : Command line warning D9025 : overriding '/Zi' with '/Z7'
scaled_mm_sm120_fp8.cu
MSVC/cl.exe with traditional preprocessor is used. This may lead to unexpected compilation errors. Please switch to the standard conforming preprocessor by passing `/Zc:preprocessor` to cl.exe. You can define CCCL_IGNORE_MSVC_TRADITIONAL_PREPROCESSOR_WARNING to suppress this warning.
cl : Command line warning D9025 : overriding '/Zi' with '/Z7'
tmpxft_00001374_00000000-7_scaled_mm_sm120_fp8.cudafe1.cpp
C:\Users\sqweek\AppData\Local\Temp\tmpxft_00001374_00000000-7_scaled_mm_sm120_fp8.cudafe1.stub.c(28): error C2719: 'unnamed-parameter': formal parameter with requested alignment of 128 won't be aligned
C:\Users\sqweek\AppData\Local\Temp\tmpxft_00001374_00000000-7_scaled_mm_sm120_fp8.cudafe1.stub.c(37): error C2719: 'unnamed-parameter': formal parameter with requested alignment of 128 won't be aligned
C:\Users\sqweek\AppData\Local\Temp\tmpxft_00001374_00000000-7_scaled_mm_sm120_fp8.cudafe1.stub.c(46): error C2719: 'unnamed-parameter': formal parameter with requested alignment of 128 won't be aligned
C:\Users\sqweek\AppData\Local\Temp\tmpxft_00001374_00000000-7_scaled_mm_sm120_fp8.cudafe1.stub.c(55): error C2719: 'unnamed-parameter': formal parameter with requested alignment of 128 won't be aligned
C:\Users\sqweek\AppData\Local\Temp\tmpxft_00001374_00000000-7_scaled_mm_sm120_fp8.cudafe1.stub.c(64): error C2719: 'unnamed-parameter': formal parameter with requested alignment of 128 won't be aligned
C:\Users\sqweek\AppData\Local\Temp\tmpxft_00001374_00000000-7_scaled_mm_sm120_fp8.cudafe1.stub.c(65): error C2719: 'unnamed-parameter': formal parameter with requested alignment of 128 won't be aligned
C:\Users\sqweek\AppData\Local\Temp\tmpxft_00001374_00000000-7_scaled_mm_sm120_fp8.cudafe1.stub.c(66): error C2719: 'unnamed-parameter': formal parameter with requested alignment of 128 won't be aligned
C:\Users\sqweek\AppData\Local\Temp\tmpxft_00001374_00000000-7_scaled_mm_sm120_fp8.cudafe1.stub.c(67): error C2719: 'unnamed-parameter': formal parameter with requested alignment of 128 won't be aligned

Adding -Xcompiler=/Zc:preprocessor to the nvcc arguments silences the MSVC/cl.exe "traditional preprocessor" warning, but compilation still fails with the alignment errors.

My build commands:

call c:\env\msvc\18\Community\VC\Auxiliary\Build\vcvarsall.bat x64 -vcvars_ver=14.44
call C:\env\cui\.venv\Scripts\activate.bat

## Ran these previously to setup the working directory & prepare the venv:
# python use_existing_torch.py
# pip install -r requirements/build.txt
# pip install -r requirements/windows.txt

set DISTUTILS_USE_SDK=1
set VLLM_TARGET_DEVICE=cuda
set MAX_JOBS=6

set VLLM_FORCE_FA3_WINDOWS_BUILD=1

python -mpip install . --no-build-isolation

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[sqweek]

The above error is while compiling csrc\quantization\w8a8\cutlass\c3x\scaled_mm_sm120_fp8.cu.
I also get similar errors against csrc\quantization\w8a8\cutlass\c3x\scaled_mm_blockwise_sm120_fp8.cu

[SystemPanic]

This is mostly a nvcc-generated stub compilation problem for the host part, with params aligned over MSVC limit of 64 bytes.

I can't test it as I does not have a RTX 50 (sm120). Can you try the following and build again?

In .deps\cutlass-src\include\cute\container\alignment.hpp, change:

template <class Child> struct CUTE_ALIGNAS(128) aligned_struct<128, Child> {};
template <class Child> struct CUTE_ALIGNAS(256) aligned_struct<256, Child> {};

to:

#if defined(_MSC_VER) && !defined(__CUDACC__)
template <class Child> struct CUTE_ALIGNAS(64) aligned_struct<128, Child> {};
template <class Child> struct CUTE_ALIGNAS(64) aligned_struct<256, Child> {};
#else
template <class Child> struct CUTE_ALIGNAS(128) aligned_struct<128, Child> {};
template <class Child> struct CUTE_ALIGNAS(256) aligned_struct<256, Child> {};
#endif

If that does not solves it, can you add --keep flag to the nvcc command and share the failing generated stub (for example, tmpxft_00001374_00000000-7_scaled_mm_sm120_fp8.cudafe1.stub.c) that MSVC is trying to compile?