Veomni Uv Update

Verify the lockfile diff is reasonable (git diff uv.lock — should only show version changes, not wholesale rewrites).

Scenario 2: Update a Regular Dependency

1. Edit version constraint in pyproject.toml under [project.dependencies] or the relevant [project.optional-dependencies] extra.
2. Regenerate lockfile and sync:

uv lock
uv sync --extra gpu --dev

3. Run tests: pytest tests/
4. Commit both pyproject.toml and uv.lock together.

Scenario 3: Update torch / CUDA Stack

This is the most complex update. torch versions are pinned in multiple places:

For GPU (gpu extra):

• pyproject.toml -> [project.optional-dependencies] -> gpu list
• pyproject.toml -> [tool.uv] -> override-dependencies (the extra == 'gpu' entries)
• pyproject.toml -> [tool.uv.sources] -> torch (direct wheel URL — must update to matching wheel)
• Related packages: torchvision, torchaudio, torchcodec, nvidia-cudnn-cu12

For NPU (npu / npu_aarch64 extras):

• Same pattern but with +cpu suffix or no suffix

Steps:

1. Identify the target torch version and matching wheel URLs from https://download.pytorch.org/whl/
2. Update all pinned versions in pyproject.toml (extras, overrides, sources)
3. Check flash-attn / flash-attn-3 wheel compatibility — these are tied to specific torch versions via direct URLs in [tool.uv.sources]
4. Update torchcodec version if needed (compatibility note in pyproject.toml)
5. Regenerate lockfile:

uv lock
uv sync --extra gpu --dev

6. Run tests: pytest tests/
7. Update Docker images if torch version changed

Scenario 4: Update transformers Version

transformers uses a dual-track setup: default 4.57.3 (group transformers-stable) and experimental 5.2.0 (extra transformers5-exp), declared as conflicts in [tool.uv.conflicts].

Bump within a track (e.g. 4.57.3 → 4.58.0, or 5.2.0 → 5.3.0):

1. Edit the pinned version in the relevant section of pyproject.toml:
   • Stable: [dependency-groups] -> transformers-stable
   • Experimental: [project.optional-dependencies] -> transformers5-exp
2. Regenerate lockfile and sync:

uv lock
# Stable:
uv sync --extra gpu --dev
# Or v5:
uv sync --no-group transformers-stable --extra transformers5-exp --extra gpu --dev

3. Check for API breakage — key v4→v5 differences:
   • AutoModelForVision2Seq removed (use AutoModelForImageTextToText)
   • no_init_weights moved from transformers.modeling_utils to transformers.initialization
   • Model __init__.py version gates: is_transformers_version_greater_or_equal_to("5.0.0") chooses generated/ (v5) vs upstream + apply_*_patch() (v4)
   • Some models (e.g. qwen3_5, glm_moe_dsa) only register on >= 5.2.0
4. Run tests: pytest tests/models/ tests/e2e/
5. Regenerate model patches if needed: make patchgen (with the target transformers installed)

Scenario 5: Regenerate Lockfile Only

When uv.lock is out of sync or corrupt:

uv lock
uv sync --extra gpu --dev

If uv lock fails due to version conflicts, check:

• [tool.uv] -> conflicts declarations
• override-dependencies markers
• Direct wheel URL availability

Common Pitfalls

• Forgetting to update Docker: uv version and torch version changes must be reflected in docker/ Dockerfiles, otherwise CI builds will fail.
• Partial torch updates: updating torch but not torchvision/torchaudio/torchcodec to matching versions causes import errors.
• flash-attn wheel mismatch: flash-attn wheels are built for specific torch+CUDA combinations. A torch version bump requires finding or building new wheels.
• Committing only pyproject.toml: always commit uv.lock together. Docker builds use --locked which requires the lockfile to match.
• override-dependencies markers: the extra == 'gpu' markers in overrides are critical. Removing them causes uv to download wrong torch variants from PyPI.
• no-build-isolation: flash-attn and flash-attn-3 are listed under no-build-isolation-package. They require torch to be installed first. If sync fails, try uv sync without these extras first, then add them.