Make transformers an opt-in dependency (conserver-local group)

[howethomas]

Summary

transformers is the single largest installed package in the project (~48MB on its own; ~67MB once you count tokenizers, huggingface_hub, and hf_xet, plus shared numpy). Despite that weight, its only consumer is conserver/links/hugging_llm_link, and only the LocalHuggingFaceLLM path uses it:

• That path is off by default (use_local_model: false).
• It calls transformers.pipeline("text-generation", ...), which needs a model backend (torch). torch is not a project dependency, so the local path cannot run today regardless.
• The default path is a plain requests.post() to the HuggingFace inference API and needs none of transformers.
• LLM analysis is already covered by the analyze / analyze_and_label / analyze_vcon links (OpenAI/Groq).

This makes transformers a large dependency carried by every conserver image for an off-by-default, non-functional path. This PR makes it opt-in.

Changes

• pyproject.toml: move transformers>=4.48.0 out of the base conserver group into a new opt-in conserver-local group.
• hugging_llm_link/__init__.py: defer import transformers into LocalHuggingFaceLLM.__init__, raising a clear ImportError (with the install command) if the optional group isn't present. Importing the link no longer loads transformers.
• hugging_llm_link/main.py: lazy-import pipeline inside HuggingLLMLink.__init__.
• README: document that local inference is opt-in and additionally needs a backend (torch).
• uv.lock: regenerated — the transformers node is unchanged, just regrouped.

Impact

Removes ~67MB from the base conserver install with no functional change to the default API-based path. Users who want on-device inference install it explicitly:

uv sync --group conserver --group conserver-local   # + a backend, e.g. torch

Testing

conserver/tests/test_link_metrics.py: 25/25 pass (incl. both TestHuggingLLMMetrics cases). Verified that importing links.hugging_llm_link (and its main) no longer pulls transformers into sys.modules.

Related

Same theme as #160 (Redis Stack): heavy/specialized dependency with shallow or absent usage. A follow-up lever: pandas (48MB) + grpcio (37MB) ride in transitively via pymilvus with zero direct use — candidate for a storage-milvus extra split.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR makes the transformers dependency opt-in for the conserver service by moving it into a new optional dependency group (conserver-local) and deferring transformers imports until the local-model paths are actually used. This reduces the default conserver install/image size while keeping the default HuggingFace HTTP API path unchanged.

Changes:

• Move transformers>=4.48.0 from the base conserver dependency group to a new optional conserver-local group (and regroup in uv.lock).
• Lazy-import transformers/pipeline within hugging_llm_link code paths so importing the link no longer pulls ML dependencies.
• Document that local inference is opt-in and requires installing conserver-local plus a backend (e.g. torch).

Reviewed changes

Copilot reviewed 4 out of 5 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
uv.lock	Regroups transformers under a new conserver-local dependency group.
pyproject.toml	Adds conserver-local optional group and removes transformers from base conserver.
conserver/links/hugging_llm_link/README.md	Documents opt-in local inference and required install command/backend.
conserver/links/hugging_llm_link/main.py	Defers pipeline import to initialization time.
conserver/links/hugging_llm_link/__init__.py	Defers transformers import to LocalHuggingFaceLLM.__init__ with a custom ImportError.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 5 changed files in this pull request and generated 3 comments.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 5 changed files in this pull request and generated 1 comment.