jax-ml/jax · error · std::runtime_error

cuDNN not found.

Error message

cuDNN not found.

What it means

jaxlib queries cuDNN's version at runtime via cudnnGetVersion(); a 0 return means TSL's stub loader could not locate or load libcudnn, so a dummy stub answered. jaxlib refuses to report a bogus version and raises instead.

Source

Thrown at jaxlib/cuda/versions_helpers.cc:98

int CusparseGetVersion() {
  // cusparseGetVersion is unhappy if passed a null library handle. But
  // cusparseGetProperty doesn't require one.
  int major, minor, patch;
  JAX_THROW_IF_ERROR(JAX_AS_STATUS(cusparseGetProperty(MAJOR_VERSION, &major)));
  JAX_THROW_IF_ERROR(JAX_AS_STATUS(cusparseGetProperty(MINOR_VERSION, &minor)));
  JAX_THROW_IF_ERROR(JAX_AS_STATUS(cusparseGetProperty(PATCH_LEVEL, &patch)));
  ABSL_ANNOTATE_MEMORY_IS_INITIALIZED(&major, sizeof major);
  ABSL_ANNOTATE_MEMORY_IS_INITIALIZED(&minor, sizeof minor);
  ABSL_ANNOTATE_MEMORY_IS_INITIALIZED(&patch, sizeof patch);
  return major * 1000 + minor * 100 + patch;
}
size_t CudnnGetVersion() {
  size_t version = ::cudnnGetVersion();
  // If the cudnn stub in TSL can't find the library, it will use a dummy stub
  // that returns 0, since cudnnGetVersion() cannot fail.
  if (version == 0) {
    throw std::runtime_error("cuDNN not found.");
  }
  ABSL_ANNOTATE_MEMORY_IS_INITIALIZED(&version, sizeof version);
  return version;
}
int CudaComputeCapability(int device) {
  int major, minor;
  {
    absl::MutexLock lock(driver_initialization_mutex);
    if (!driver_initialized) {
      JAX_THROW_IF_ERROR(JAX_AS_STATUS(cuInit(0)));
      driver_initialized = true;
    }
  }
  JAX_THROW_IF_ERROR(JAX_AS_STATUS(gpuDeviceGetAttribute(
      &major, GPU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, device)));
  JAX_THROW_IF_ERROR(JAX_AS_STATUS(gpuDeviceGetAttribute(
      &minor, GPU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, device)));
  ABSL_ANNOTATE_MEMORY_IS_INITIALIZED(&major, sizeof major);

View on GitHub (pinned to 1e1c6a8fc0)

Solutions

1. pip install -U nvidia-cudnn-cu12 matching your jaxlib CUDA version
2. Set LD_LIBRARY_PATH to the directory containing libcudnn.so (e.g. the pip nvidia/cudnn/lib dir)
3. Verify with python -c "import ctypes; ctypes.CDLL('libcudnn.so')" and fix loader errors
4. Ensure cuDNN major version matches the one your jaxlib wheel was built against

Example fix

# before: pip install jax
# after:  pip install -U "jax[cuda12]"

import ctypes
try:
    ctypes.CDLL('libcudnn.so')
    ok = True
except OSError as e:
    ok = False; print(e)

When it happens

Trigger: Importing/initializing jax CUDA support when LD_LIBRARY_PATH/PIP cuDNN does not contain a loadable libcudnn (e.g. nvidia-cudnn-cu12 wheel missing or wrong major version for your CUDA).

Common situations: pip-installing jax without the matching nvidia-cudnn-cu12 package; custom CUDA envs where cuDNN is not on the loader path; upgrading jaxlib but not the CUDA extras wheels.

Related errors

• Nonsymmetric eigendecomposition requires cusolver 11.7.1 or
• cuDNN doesn't support right window: {r_window} when causal m
• bfloat16 support not implemented for LSTM
• Unknown GPU platform for __dlpack__: {platform_version}
• Couldn't get local_hardware_id for __dlpack__

AI-assisted analysis of jax-ml/jax@1e1c6a8fc0 (2026-08-27). Data as JSON: /api/errors/a3c68f2d219faf7d. Report an issue: GitHub.