`local-tokenizer` extra installs the full PyTorch/CUDA stack for SentencePiece-only models

[tiscovsa]

Description

Installing google-genai[local-tokenizer] installs Pillow, PyTorch, Torchvision, Transformers, Triton and the Linux NVIDIA CUDA packages for every locally tokenised model.

However, models such as gemini-2.5-flash use the lightweight SentencePiece path. The Hugging Face AutoProcessor path is selected only for the Gemma 4 model mappings.

This makes local tokenisation unexpectedly expensive for applications that do not use Gemma 4.

Reproduction

Environment:

• Linux x86_64
• Python 3.14.2
• uv
• google-genai==2.11.0

uv venv --python 3.14 base
uv pip install --python base/bin/python \
  'google-genai[aiohttp]==2.11.0'

uv venv --python 3.14 local
uv pip install --python local/bin/python \
  'google-genai[aiohttp,local-tokenizer]==2.11.0'

du -sh base local

Result:

45M  base
4.7G local

Largest installed components:

1.1G  torch
2.7G  nvidia
689M  triton
13M   torchvision

The dependency count increases from 33 packages to 83.

Expected behaviour

Applications using the SentencePiece-backed local tokenizer should not need to install the Gemma 4 Hugging Face and PyTorch stack.

Could the extras be split, for example:

local-tokenizer = [
    "sentencepiece>=0.2.0",
    "protobuf",
]

local-tokenizer-gemma4 = [
    "sentencepiece>=0.2.0",
    "protobuf",
    "pillow",
    "torch",
    "torchvision",
    "transformers",
]

The implementation already imports AutoProcessor lazily inside the Gemma 4 path, so the runtime boundary largely exists.

It may also be worth checking whether AutoTokenizer could replace AutoProcessor for this text-only use case, potentially removing some image-related dependencies. That requires validation against Gemma 4.

Downstream impact

This caused one downstream production service environment to increase from approximately 626 MiB to 5.1 GiB. We had to remove the local-tokenizer extra from our shared library's base installation and fall back to the remote count_tokens endpoint.

Relevant code:

• v2.11.0 optional dependencies
• Local tokenizer loader
• Gemma 4 implementation PR
• Dependency expansion PR

[Venkaiahbabuneelam]

Hi @tiscovsa, Thanks for reaching out us!

To help us understand this more efficiently, kindly provide a full reproducible code sample along with the error logs.

Thanks

[tiscovsa]

Hi, thanks for responding. This is a packaging dependency issue rather than a runtime error, so there is no application code or error log.

The full reproduction is already included:

uv venv --python 3.14 base
uv pip install --python base/bin/python \
  "google-genai[aiohttp]==2.11.0"

uv venv --python 3.14 local
uv pip install --python local/bin/python \
  "google-genai[aiohttp,local-tokenizer]==2.11.0"

du -sh base local

This produces approximately:

45M   base
4.7G  local

The cause is that local-tokenizer unconditionally includes the Transformers, PyTorch, Triton and NVIDIA CUDA dependency chain, even when tokenising a SentencePiece-backed model such as gemini-2.5-flash.

Could you please confirm whether the team would consider splitting the SentencePiece and Gemma 4 local-tokeniser dependencies into separate extras?