[CI] Install torchvision for the torchtitan integration tests (#194702)
The torchtitan model integration job has been failing for every PR, not just
one. Two of the three failing flavors die at model-build time:
muse_glimmer_mm_fsdp+tp+sp ModuleNotFoundError: No module named 'torchvision'
kimi_k2_5_muon_pp+fsdp+ep ModuleNotFoundError: No module named 'torchvision'
test_torchtitan installs torchao, torchcomms and helion, but not torchvision,
and torchtitan's muse_glimmer and kimi_k2_7 models import it while building.
Use the existing install_torchvision helper, which builds from the pinned
vision commit against the CI-built torch rather than pulling a PyPI wheel that
would drag in its own torch -- the same reason the function already guards
against helion and torchtitan overwriting the CI-built packages.
The third failing flavor, qwen3_5_fsdp+tp+varlen_attn+per_op_sac, is a separate
problem and is not addressed here: it needs flash_attn_interface (FA3), which
torchtitan's own tests/integration_tests/features.py already notes "the core
integration CI does not install", and the job runs on A10G (sm86) where FA3 is
unavailable regardless. That belongs in torchtitan.
Test Plan:
Reproduced against the pinned torchtitan commit
(87eca1ac050e88143bbd61e0f123ed976c4e67c9) on 4 GPUs. With torchvision present,
both flavors that fail in CI pass:
```
muse_glimmer_debugmodel_mm_fsdp2_tp2 exit=0 PASS
kimi_k2_5_debugmodel_muon_fsdp2_pp2_ep2 exit=0 PASS
```
Without it they reproduce the CI failure exactly.
Authored with assistance from Claude Code (Anthropic).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/194702
Approved by: https://github.com/anijain2305, https://github.com/fegin W
Wei Feng committed
d397b88c7a282ba41efd9198d0d669fd65c344e6
Parent: 6b1f3af
Committed by PyTorch MergeBot <pytorchmergebot@users.noreply.github.com>
on 8/26/2026, 3:42:55 AM