Windows环境下TensorFlow 2.4.1 C++版本编译失败求助

I've run into this exact issue before when building TF 2.4.1 C++ on Windows, so let's break down what's going wrong and how to fix it:

First, Verify Your cuDNN File Structure #

The error tells us Bazel can't find cudnn_ops_infer.h, so start by manually checking if this file exists in your cuDNN path. Navigate to D:/code/sdk/cudnn-11.0-windows-x64-v8.0.4.30/cuda/include — you should see cudnn_ops_infer.h, cudnn.h, and all other cuDNN header files here. If they're missing:

• Double-check that you downloaded the correct cuDNN version (v8.0.4.30 for CUDA 11.0, Windows x64)
• Re-extract the cuDNN zip file fully (sometimes extraction tools skip files if there's permission issues)

Fix Environment Variables #

Even if you set paths in .tf_configure.bazelrc, Windows and Bazel sometimes rely on system environment variables to locate dependencies. Add these to your system environment variables:

• INCLUDE: Append D:/code/sdk/cudnn-11.0-windows-x64-v8.0.4.30/cuda/include
• LIB: Append D:/code/sdk/cudnn-11.0-windows-x64-v8.0.4.30/cuda/lib/x64
• PATH: Append D:/code/sdk/cudnn-11.0-windows-x64-v8.0.4.30/cuda/bin

After adding these, restart your command prompt/Bazel terminal — environment variables won't take effect until you do this.

Tweak Your Bazel Configuration #

Your .tf_configure.bazelrc has the right TF_CUDA_PATHS, but sometimes Bazel needs explicit guidance for cuDNN's include directory. Add this line to your .tf_configure.bazelrc:

build --action_env CUDNN_INCLUDE_DIR="D:/code/sdk/cudnn-11.0-windows-x64-v8.0.4.30/cuda/include"

Alternatively, re-run python configure.py and when prompted for the cuDNN path, make sure you input the full path to the cuda folder (with Linux-style slashes) and confirm that the script detects all cuDNN components correctly.

Clean Bazel Cache and Rebuild #

Bazel's cache can hold onto old path references that cause conflicts. Run this command to fully clear the cache:

bazel clean --expunge

Then re-run your build command. For the TensorFlow C++ API, the standard command is:

bazel build --config=cuda --config=opt //tensorflow:tensorflow_cc

A Quick Note on Windows Compatibility #

Yes, building TensorFlow 2.4.1 C++ API on Windows is absolutely possible — I've done it successfully with the exact same toolchain versions you're using (CUDA 11.0, cuDNN 8.0.4.30, Bazel 3.1.0, Python 3.6.8). The biggest pitfalls are path formatting (always use Linux-style slashes in Bazel configs) and ensuring all dependencies are properly linked via both Bazel config and system environment variables.

内容的提问来源于stack exchange，提问作者Alexander Soklev