build(deps): unpin torch and remove torchvision from CPU requirements

[pctablet505]

Problem

The CPU torch CI leg was resolving a CUDA-build torch from PyPI alongside a CPU-build torchvision from the PyTorch extra index, causing an ABI mismatch:

RuntimeError: operator torchvision::nms does not exist

This also broke the lazy Qwen2ForCausalLM import from transformers at collection time.

Fix

Remove the torchvision dependency entirely and leave torch unpinned in requirements.txt. The Keras Hub test suite does not need torchvision, so dropping it avoids the ABI mismatch without requiring a pinned matched pair.

Only the CPU leg (requirements.txt) changes; the CUDA requirements file is untouched.

[gemini-code-assist]

Code Review

This pull request updates requirements.txt to pin torch and torchvision to specific versions. However, the specified versions (2.13.0 and 0.28.0) do not exist. The reviewer pointed out this issue, identifying it as hallucinated 'AI slop' in violation of the repository style guide, and suggested using valid, existing versions such as torch==2.6.0 and torchvision==0.21.0.

[gemini-code-assist]

The specified versions torch==2.13.0 and torchvision==0.28.0 do not exist. PyTorch's latest stable release is 2.6.0 (with torchvision 0.21.0). This appears to be hallucinated 'AI slop' (violating Repository Style Guide Rule 2). Please use a valid, existing matched pair of PyTorch and torchvision versions, such as torch==2.6.0 and torchvision==0.21.0.

torch==2.6.0
torchvision==0.21.0

References

1. Rule 2: Call out 'AI Slop' - Actively look for and identify 'AI slop'—generic, overly verbose, or hallucinated code that lacks context or violates best practices. ^(link)

[pctablet505]

This is Gemini hallucinated and spitting wrong facts, even the CI logs show it got installed and passed the tests.

[laxmareddyp]

Hi @pctablet505 , Thanks for the PR..!

This was always a latent issue.I would prefer not to hard-pin torch and torch vision, as it creates an ongoing maintenance burden for every future PyTorch release.
The CI only crashes because newer transformers versions check for torchvision, triggering a broken C++ extension load due to pip version mismatches.
Instead of pinning, let's just remove torchvision from requirements.txt entirely, since our only usage is a lazy, safely-guarded import in other models such as SAM3 . This makes us to zero future maintenance.

[laxmareddyp]

@pctablet505 Please resolve the conflicts and as i mentioned earlier no hard pinning for torch version. Thanks

[laxmareddyp]

LGTM, Thanks..!

[laxmareddyp]

@pctablet505 ,

Apologies , my earlier suggestion to remove torchvision. I missed to check the SAM3 test files that run in the torch CI job. These tests exercise the full model forward pass which hits roi_align.py: import torchvision at runtime. So torchvision is a hard runtime dependency for SAM3 on the torch backend, not just an optional one.

Your original approach of pinning matched versions is the correct fix( We need to keep eye on to update these hard pinned versions time to time)

Thanks..!