[Bug]: what cuda version dlib supports so far #2832

Activity

1. flydragon2018

   added

   bug

   on Jul 20, 2023
2. freecraver commented on Nov 14, 2023

   

   freecraver

   on Nov 14, 2023

   More actions

   Faintly related - Stumbled upon this when building dlib with base image pytorch/pytorch:1.13.1-cuda11.6-cudnn8-devel. Workaround was: ENV CPATH=/usr/local/cuda/include:$CPATH. (Based on deepspeedai/DeepSpeed#2684 (comment))
3. lxt3 commented on Nov 10, 2025

   

   lxt3

   on Nov 10, 2025

   More actions

   I just installed it on WSL2 for CUDA 13.0 and cudnn 9.13.0
4. alexandreofbh commented on Nov 10, 2025

   

   alexandreofbh

   on Nov 10, 2025

   Last edited by alexandreofbh

   More actions

   DLIB v20.0, on Ubuntu 24.04
   I am using the 570 driver, cuda 12.8.0 and cudnn-linux-x86_64-9.13.1.26 and it worked perfectly.
   Cuda 13 on drive 580 did not work.
5. gregory112 commented on Jan 11, 2026

   

   gregory112

   on Jan 11, 2026

   More actions

   Also tried building python-dlib with CUDA 13 and driver 590, mine seems to fail to compile, with nvcc fatal : Unsupported gpu architecture 'sm_50'. Setting -DCMAKE_CUDA_ARCHITECTURES=86 does not work though, So I downgraded to CUDA 12.9. This builds, but with Enabling CUDA support for dlib. DLIB WILL USE CUDA, compute capabilities: 50.
6. davisking commented on Jan 11, 2026

   

   davisking

   on Jan 11, 2026 via email

   Owner

   More actions

   That's just cmake stuff, which you can adjust as a user to target whatever SM version you want. Although get the latest dlib from github, it has updated cuda cmake scripts that will probably Just Work automatically.

   …

   On Sun, Jan 11, 2026 at 3:10 AM Sergio Ryan ***@***.***> wrote: *gregory112* left a comment (davisking/dlib#2832) <#2832 (comment)> Also tried building python-dlib with CUDA 13 and driver 590, mine seems to fail to compile, with nvcc fatal : Unsupported gpu architecture 'sm_50'. Setting -DCMAKE_CUDA_ARCHITECTURES=86 does not work though, So I downgraded to CUDA 12.9. This builds, but with Enabling CUDA support for dlib. DLIB WILL USE CUDA, compute capabilities: 50. — Reply to this email directly, view it on GitHub <#2832 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABPYFR33WCIXH46DPNEUNLL4GIAOLAVCNFSM6AAAAACLTNB672VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTOMZUGIYTKNRTGI> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>