Update EOL components (python, cuda, ubuntu) #49

Activity

1. andy778

   added

   enhancementNew feature or requestNew feature or request

   on Mar 22, 2026
2. andy778 commented on Apr 4, 2026

   

   andy778

   on Apr 4, 2026

   Last edited by andy778

   ContributorAuthor

   More actions

   This old pytorch need to be updated to support newer hardware (This is the latest released docker https://hub.docker.com/r/airiksarkivet/htrflow)

   htrflow-1  | /app/.venv/lib/python3.10/site-packages/torch/cuda/__init__.py:235: UserWarning:
   htrflow-1  | NVIDIA GeForce RTX 5060 Ti with CUDA capability sm_120 is not compatible with the current PyTorch installation.
   htrflow-1  | The current PyTorch install supports CUDA capabilities sm_50 sm_60 sm_70 sm_75 sm_80 sm_86 sm_90.
   htrflow-1  | If you want to use the NVIDIA GeForce RTX 5060 Ti GPU with PyTorch, please check the instructions at https://pytorch.org/get-started/locally/
   
3. andy778 commented on Apr 5, 2026

   

   andy778

   on Apr 5, 2026

   Last edited by andy778

   ContributorAuthor

   More actions

   The pytest are failing also and Node.js 20 are soon EOL
4. andy778 commented on Jul 30, 2026

   

   andy778

   on Jul 30, 2026

   ContributorAuthor

   More actions

   Here a PR for python 3.10, after october we can remove it
   #59
5. ikeago commented on Oct 12, 2026

   

   ikeago

   on Oct 12, 2026

   More actions

   I tested the current HTRflow stack on an RTX 5090 after seeing the sm_120 report above. A native smoke test passed with driver 591.86, PyTorch 2.14.1+cu130, CUDA 13.0, and cuDNN 9.24: the wheel included sm_120, a 4096×4096 CUDA matrix operation completed in 0.216 s, and Riksarkivet/trocr-base-handwritten-hist-swe-2 produced a nonempty transcription in 4.599 s (about 1.62 GB peak allocated VRAM).

   I also checked the exact lock in PR #57: on Python 3.12 it resolves PyTorch 2.13.0+cu130, whose compiled architecture flags include sm_120. Building on that contributor's work, I prepared one commit that updates the GPU image defaults to CUDA 13.0.2 / Ubuntu 24.04 / Python 3.12, pins uv to the runtime's system interpreter, and adds an in-image CUDA/cuDNN smoke script:

   • branch: https://github.com/ikeago/htrflow/tree/build/blackwell-container
   • commit: ikeago@35e4b36

   I have not opened this as a pull request because the branch is intentionally one commit on top of #57, and I do not want to duplicate or obscure Andy's unmerged work. I am happy to rebase it onto main after #57 lands, or adjust the submission order if the maintainers prefer.

   The native GPU test passed, but I could not build the Docker image locally because this WSL account cannot access the Docker socket, so I am not claiming that the image build itself was verified.

   OpenAI Codex assisted with implementation and testing, and Claude Sonnet independently reviewed the patch. I reviewed the final changes and take responsibility for them.

   👍React with 👍1Reacted by Andy B
6. andy778 commented on Oct 12, 2026

   

   andy778

   on Oct 12, 2026

   ContributorAuthor

   More actions

   It would be quite fast to update this to modern quality if there is someone that would merge the pull requests. I had much more in the pipe but stopped as there was no mergers...
7. andy778 commented on Oct 14, 2026

   

   andy778

   on Oct 14, 2026

   ContributorAuthor

   More actions

   @viklofg is there some plans to pickup the work on this or should it be seen as and EOL project ?
8. viklofg commented on Oct 15, 2026

   

   viklofg

   on Oct 15, 2026

   Contributor

   More actions

   Hi @andy778 . Thank you for your interest and contributions! I am sorry we have not yet reviewed your work. This repo has been dormant for a while but we're planning a new release soon, so we'll make sure to review your contributions in the process.