KataGo icon indicating copy to clipboard operation
KataGo copied to clipboard

CUDA 12.5 build produces visits/s = -nan(ind) even with compatible cuDNN and model

Open Immow opened this issue 5 months ago • 1 comments

Hi, I'm having trouble getting the CUDA 12.5 version of KataGo to work on Windows 11. The binary runs, but all benchmark tests fail with visits/s = -nan(ind) output — no matter what I try.

Setup OS: Windows 11 GPU: NVIDIA RTX 5070-class (Ada Lovelace) KataGo Version: v1.16.3-cuda12.5-cudnn8.9.7-windows-x64.zip CUDA: cuda_12.5.0_555.85_windows.exe (installed via NVIDIA site) cuDNN: cudnn-windows-x86_64-8.9.7.29_cuda12-archive.zip cudnn64_8.dll copied into same folder as katago.exe Model Used: g170e-b20c256x2-s5303129600-d1228401921.bin.gz

Steps Taken

Set PATH for CUDA 12.5 (nvcc --version confirms 12.5.40)

Placed the following DLLs in the KataGo folder:

  •     cudnn64_8.dll
    
  •     cudart64_120.dll
    
  •     cublas64_12.dll
    
  •     cublasLt64_12.dll
    
  •     and other common dependencies from CUDA\v12.5\bin
    

Generated config katago.exe genconfig -model g170e-b20c256x2-s5303129600-d1228401921.bin.gz -output default.cfg

Then ran katago.exe benchmark -model g170e-b20c256x2-s5303129600-d1228401921.bin.gz

Problem numSearchThreads = 5: 0 / 10 positions, visits/s = -nan(ind) (0.0 secs) No crash or error — just zero evaluations and NaN results.

What I've Tried

  • Verified model integrity and tried smaller networks
    
  • Rechecked that all DLLs match the CUDA version (12.5)
    
  • Ran from elevated CMD with full path access
    
  • Installed both Runtime and Development parts of CUDA 12.5
    

Questions

  • Is the 12.5 CUDA build tested with cuDNN 8.9.7?
    
  • Are there known compatibility issues with Ada GPUs?
    

Immow avatar Jul 25 '25 18:07 Immow