Skip to content

LunaVox Windows CUDA Dependency Guide

Covers both CUDA 12 and CUDA 13 runtime sets on Windows x64. Pick the column that matches the binaries you downloaded with lunavox build libs --platform win_cuda12 (or win_cuda13).

1. Verified Combinations

Component CUDA 12 CUDA 13
ONNX Runtime 1.24.4 (GPU) 1.24.4 (GPU, CUDA 13 build)
llama.cpp — b8470 (CUDA 13.1)
CUDA Toolkit 12.4.x 13.2.0
cuDNN 9.1.1 9.20.0
GPU under test RTX 3090 (CC 8.6) RTX 3090

2. cuDNN 9 Is Not a Single File

On Windows, cudnn64_9.dll is only an entry shim. The full set of eight DLLs must be present, regardless of CUDA version:

cudnn64_9.dll
cudnn_adv64_9.dll
cudnn_cnn64_9.dll
cudnn_engines_precompiled64_9.dll
cudnn_engines_runtime_compiled64_9.dll
cudnn_graph64_9.dll
cudnn_heuristic64_9.dll
cudnn_ops64_9.dll

Copying only cudnn64_9.dll is the most common cause of "CUDA acceleration silently disabled".

3. Minimum DLL Set

Core (LunaVox-built)

  • onnxruntime.dll (GPU build)
  • onnxruntime_providers_cuda.dll
  • onnxruntime_providers_shared.dll
  • ggml-cuda.dll

CUDA / Math

CUDA 12 CUDA 13
Runtime cudart64_12.dll cudart64_13.dll
cuBLAS cublas64_12.dll, cublasLt64_12.dll cublas64_13.dll, cublasLt64_13.dll
cuFFT cufft64_11.dll cufft64_12.dll
cuRAND curand64_10.dll curand64_10.dll
cuDNN full 8-DLL set (Section 2) full 8-DLL set (Section 2)

MSVC runtime

vcruntime140.dll, vcruntime140_1.dll, msvcp140.dll

4. Deployment

  • Conda (recommended): install cuda-toolkit + cudnn of the matching major into a dedicated env and ensure Library\bin is on PATH.
  • Portable: drop every DLL listed above next to lunavox-cli.exe.
  • After switching CUDA major version, run lunavox build libs --platform win_cuda13 (or win_cuda12) and rebuild — leftover binaries from the other major will fail with error 126 because they reference the wrong cublasLt64_*.dll.

5. Troubleshooting

  1. Error 126 (module not found): a DLL is referencing the other CUDA major. Re-download the libs for the correct platform.
  2. Provider silently CPU: run with --verbose; decoder should report CUDAExecutionProvider.
  3. llama.cpp falls back to CPU: ggml-cuda.dll does not match the CUDA major. Replace it.
  4. Version skew: onnxruntime_providers_cuda.dll and onnxruntime.dll must come from the same build.