google-deepmind / google-deepmind/gemma
main branch incompatible with published PyPI versions of kauldron and dialog
- Dominant language
- Python
- Stars
- 5.7k
- Forks
- 1k
- Avg merge
- 10h 33m
- Merged PRs (30d)
- 2
Description
The current main branch references APIs from kauldron and dialog that don't exist in
their latest published PyPI versions (kauldron==1.3.0, dialog==1.0.0). This makes it
impossible to run gemma from source after a fresh install.
Steps to reproduce
git clone https://github.com/google-deepmind/gemma.git
cd gemma
uv sync # or: pip install -e .
python -c "import gemma.gm as gm; gm.nn.Gemma3_4B()"
Errors
1. kauldron — InitTransform not found
Introduced in commit bd0dabf ("Rename AbstractPartialLoader to InitTransform").
File "gemma/gm/ckpts/_checkpoint.py", line 49, in
class LoadCheckpoint(kd.ckpts.InitTransform):
AttributeError: module 'kauldron.checkpoints' has no attribute 'InitTransform'
kauldron==1.3.0 exposes AbstractPartialLoader but not InitTransform.
Affected files:
- gemma/gm/ckpts/_checkpoint.py
- gemma/gm/ckpts/_lora.py
- gemma/gm/ckpts/_policy.py
2. dialog — Format not found
Introduced in commit 1ded5e5 ("Supports tool and dialog in ChatSampler").
File "gemma/gm/text/_tokenizer.py", line 166, in Tokenizer
FORMAT: ClassVar[dialog.Format] = dialog.Format.GEMMA3
AttributeError: module 'dialog' has no attribute 'Format'
dialog==1.0.0 does not have a Format class anywhere in the package.
Affected files:
- gemma/gm/text/_tokenizer.py
- gemma/gm/text/_sampler.py
Environment
- Python 3.13.2
- kauldron 1.3.0 (latest on PyPI)
- dialog 1.0.0 (latest on PyPI)
- macOS / Darwin
Possible fix
Add git source overrides in pyproject.toml so that uv sync / pip install pulls the
compatible versions:
[tool.uv.sources]
kauldron = { git = "https://github.com/google-research/kauldron.git" }
dialog = { git = "https://github.com/google-deepmind/dialog.git" }
Or publish updated versions of kauldron and dialog to PyPI that include InitTransform
and Format.
Contributor guide
Assessment
This issue has not been assessed yet.