InterviewsVector

TensorFlow Not Detecting GPU / CUDA Installed via Conda: The Version-Match Fix

TensorFlow & GPUsDebuggingcudagpu

Quick answer

TensorFlow ships built against specific CUDA and cuDNN versions, so tf.config.list_physical_devices('GPU') returns [] when the installed versions don't match, or when a system CUDA install shadows the conda one. The modern fix on Linux is pip install tensorflow[and-cuda], which bundles matching CUDA/cuDNN wheels. Otherwise install the exact cudatoolkit and cudnn versions TensorFlow expects into the same conda env and don't mix conda with a system CUDA. The tensorflow-gpu package is deprecated.

Short answer: TensorFlow is built against specific CUDA/cuDNN versions, so tf.config.list_physical_devices('GPU') returns [] when the installed versions don't match — or when a system CUDA install shadows the conda one. On Linux, the modern fix is pip install tensorflow[and-cuda] (bundles matching CUDA/cuDNN). Otherwise, install the exact cudatoolkit/cudnn TensorFlow expects into the same conda env. tensorflow-gpu is deprecated.

You installed CUDA through conda, nvidia-smi shows the GPU, but TensorFlow still runs on CPU. This is almost never "CUDA isn't installed" — it's a version mismatch or a conda-vs-system conflict.

Step 1 — confirm what TensorFlow actually sees

import tensorflow as tf
print(tf.config.list_physical_devices("GPU"))   # []  -> not detected
# What CUDA/cuDNN this TF build expects:
info = tf.sysconfig.get_build_info()
print(info["cuda_version"], info["cudnn_version"])

Compare that against the driver and toolkit actually present:

nvidia-smi          # driver + max CUDA the driver supports
nvcc --version      # toolkit version on PATH (may differ from conda's)

If cuda_version from TensorFlow doesn't match what's installed, that's your bug. TensorFlow built against CUDA 11.8 will not use a CUDA 12 install.

The root cause

Conda installs CUDA inside the environment prefix ($CONDA_PREFIX), not a system directory. TensorFlow may not look there, or a system CUDA install may be found first and shadow the conda one. Mixing the two is the classic trap.

Fix 1 — let pip bundle CUDA (recommended, Linux)

The simplest reliable path is to stop matching versions by hand and let the wheel do it:

pip install 'tensorflow[and-cuda]'

This pulls the exact CUDA/cuDNN pip wheels TensorFlow was built against, into the environment — no manual cudatoolkit/cudnn install needed.

Fix 2 — matched conda environment

If you prefer conda, install the exact versions TensorFlow expects into the same env (check TensorFlow's tested-build compatibility table):

conda create -n tf python=3.11
conda activate tf
conda install -c conda-forge cudatoolkit=11.8 cudnn=8.6   # match your TF build
pip install tensorflow

Then make sure the env's libraries are on the path:

export LD_LIBRARY_PATH="$CONDA_PREFIX/lib:$LD_LIBRARY_PATH"

Common pitfalls

  • Mixing conda and system CUDA — pick one; a system install often wins on LD_LIBRARY_PATH.
  • Installing a CPU-only build — pip install tensorflow on macOS, or a mismatched wheel, yields no GPU.
  • Using tensorflow-gpu — deprecated; use tensorflow / tensorflow[and-cuda].
  • Native Windows — TensorFlow dropped native-Windows GPU support after 2.10; use WSL2 with tensorflow[and-cuda].

Related GPU/CUDA fixes

Sources

Key takeaways

  • •Confirm the problem with tf.config.list_physical_devices('GPU') — an empty list means no GPU detected.
  • •The usual cause is a CUDA/cuDNN version mismatch: TF is built against exact versions (CUDA 11.8 build won't use CUDA 12).
  • •Modern fix on Linux: pip install tensorflow[and-cuda] bundles the matching CUDA/cuDNN wheels — no manual install.
  • •For conda, install the exact cudatoolkit + cudnn TF expects into the same env; don't mix conda and system CUDA.
  • •tensorflow-gpu is deprecated (merged into tensorflow); native Windows GPU support ended at TF 2.10 — use WSL2.

Frequently asked questions

How do I check if TensorFlow sees my GPU?

Run import tensorflow as tf; print(tf.config.list_physical_devices('GPU')). A non-empty list means the GPU is detected. If it's empty, TensorFlow is not finding a compatible CUDA/cuDNN, or you installed a CPU-only build.

Why doesn't TensorFlow use the CUDA I installed with conda?

TensorFlow is compiled against specific CUDA and cuDNN versions. If the conda-installed versions don't match, or a system CUDA install is found first, TensorFlow silently falls back to CPU. Match the versions to what your TensorFlow build expects, or install tensorflow[and-cuda] which bundles them.

What is the difference between tensorflow and tensorflow-gpu?

There is no longer a difference — tensorflow-gpu is deprecated. Since TensorFlow 2.x, the standard tensorflow package includes GPU support. On Linux, pip install tensorflow[and-cuda] additionally pulls the matching CUDA/cuDNN libraries.

By Mohammad Wasi

Software Engineering Leader & Technical Author · Updated August 26, 2026

See all posts

Related Posts