Solving Incompatibility Issue of Tensorflow 2.21.0 and Newest CUDA 13 (Blackwell GPU)

tf_install_bw_cu13.py

# Install TensorFlow + CUDA runtime wheels
uv add "tensorflow[and-cuda]==2.21.0"

# Set up Tensorflow
import os
import site
import ctypes
from pathlib import Path

# Optional: reduce TensorFlow startup logs
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"

# Add NVIDIA wheel library folders to LD_LIBRARY_PATH
site_packages = next(p for p in site.getsitepackages() if p.endswith("site-packages"))
nvidia_root = Path(site_packages) / "nvidia"
lib_dirs = sorted(
    {str(p.parent) for p in nvidia_root.rglob("*.so*") if p.parent.name == "lib"}
)

if lib_dirs:
    old = os.environ.get("LD_LIBRARY_PATH", "")
    os.environ["LD_LIBRARY_PATH"] = ":".join(lib_dirs + ([old] if old else []))

# Preload common CUDA libs in this process so TensorFlow can resolve symbols
preload = [
    "cudart/lib/libcudart.so*",
    "cublas/lib/libcublas.so*",
    "cublas/lib/libcublasLt.so*",
    "cudnn/lib/libcudnn.so*",
    "cufft/lib/libcufft.so*",
    "curand/lib/libcurand.so*",
    "cusolver/lib/libcusolver.so*",
    "cusparse/lib/libcusparse.so*",
    "nccl/lib/libnccl.so*",
    "nvjitlink/lib/libnvJitLink.so*",
    "cuda_nvrtc/lib/libnvrtc.so*",
    "cuda_cupti/lib/libcupti.so*",
]

for pattern in preload:
    matches = sorted(nvidia_root.glob(pattern))
    if matches:
        ctypes.CDLL(str(matches[-1]), mode=ctypes.RTLD_GLOBAL)

# Verify installation
import tensorflow as tf

print("TF version:", tf.__version__)
print("CUDA build:", tf.sysconfig.get_build_info().get("cuda_version"))
print("cuDNN:", tf.sysconfig.get_build_info().get("cudnn_version"))
print("GPUs:", tf.config.list_physical_devices("GPU"))

# OUTPUT: 
# TF version: 2.21.0
# CUDA build: 12.5.1
# cuDNN: 9
# GPUs: [PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

Some relevant discussions from tensorflow/tensorflow#106653. This gist at least worked in my NVIDIA RTX Pro 500 Blackwell with CUDA 13.2 driver.