构建Windows移植版Faster-RCNN时遇nvcc报错：'--ptxas-options=-v': expected a number

DevOps 小满

2026-5-13

Fixing "nvcc fatal : '--ptxas-options=-v': expected a number" When Building Faster-RCNN on Windows

I’ve run into this exact nvcc error before when porting Faster-RCNN to Windows with CUDA 10.1, so let’s walk through the fixes step by step:

Root Cause

The --ptxas-options=-v flag works smoothly on Linux, but Windows’ nvcc parser can misinterpret the shorthand -v as a numerical argument instead of the verbose flag. This is a quirk of how CUDA’s compiler handles parameter formatting across operating systems.

Step-by-Step Solutions

1. Fix the Compiler Flag in Your Python Config Script

Locate the build configuration script (usually setup.py or a related .py config file for Faster-RCNN) and find the section where nvcc arguments are defined.

Before (problematic line):

extra_compile_args = {'nvcc': ['--ptxas-options=-v']}

After (fixed version):

extra_compile_args = {'nvcc': ['--ptxas-options=--verbose']}

Alternatively, you can use Windows-style parameter formatting:

extra_compile_args = {'nvcc': ['--ptxas-options=/v']}

2. Use the Correct Command Prompt Environment

Never run the build command in a regular Windows Command Prompt. Instead:

Open the Developer Command Prompt for VS 2019 (search for it in the Start Menu)
Activate your Python 3.7 virtual environment (if using one)
Run your build command from here

This ensures VS 2019’s C++ build tools and CUDA 10.1’s environment variables are properly loaded, avoiding weird parsing issues.

3. Verify CUDA + VS 2019 Compatibility

Double-check these details to rule out environment mismatches:

Confirm CUDA 10.1 is fully installed with all components (run nvcc --version in the dev prompt to verify)
Ensure VS 2019 has the Desktop development with C++ workload installed (this includes the required compilers and libraries CUDA needs)
Check that your CUDA_PATH environment variable points to the correct CUDA 10.1 installation directory