perf: lazy-import litellm, faiss, torch to speed up import lotus

[01luyicheng]

Purpose

Closes #143.

import lotus currently takes ~4s (per issue #143), and up to ~10s in environments where all optional backends (torch, sentence_transformers, faiss) are installed. The root cause is that lotus/models/lm.py, lotus/pricing.py, lotus/types.py, lotus/models/__init__.py, and lotus/vector_store/__init__.py all eagerly import heavy dependencies (litellm ~4s, torch, sentence_transformers, faiss) at import lotus time, even when the user never instantiates an LM or a vector store.

This PR defers all of those imports until they are actually needed.

Test Plan

# 1. Verify heavy deps are no longer loaded at import time
python -c "
import lotus, sys
assert 'litellm' not in sys.modules
assert 'torch' not in sys.modules
assert 'sentence_transformers' not in sys.modules
assert 'faiss' not in sys.modules
print('lazy ok')
"

# 2. Verify import time dropped
python -c "import time; t0=time.perf_counter(); import lotus; print(f'{time.perf_counter()-t0:.3f}s')"

# 3. Verify LM still works (litellm loads lazily on first use)
python -c "
from lotus.models import LM
lm = LM(model='gpt-4o-mini')
print(lm.count_tokens('hello world'))
"

# 4. Run the new + existing tests
pytest tests/test_import_lazy.py -v
pytest tests/test_lm.py::TestLM::test_lm_rate_limiting_with_mock -v

Test Results

Import time (same environment, all backends installed):

	Before (main)	After (this PR)
import lotus	10.55s	1.10s (9.6x faster)
litellm loaded	yes	no (deferred)
torch loaded	yes	no (deferred)
sentence_transformers loaded	yes	no (deferred)
faiss loaded	yes	no (deferred)

In an environment with only litellm installed (no torch/st/faiss, matching the issue reporter's setup), import lotus drops from ~4s to ~1s.

pytest:

• tests/test_import_lazy.py: 8 passed (asserts litellm/torch/st/faiss absent from sys.modules after import lotus; verifies public API still resolves; covers __getattr__ cache + AttributeError contract)
• tests/test_lm.py::TestLM::test_lm_rate_limiting_with_mock: passed (mock path updated from lotus.models.lm.batch_completion to litellm.batch_completion)
• tests/test_lm.py: 4 remaining failures are pre-existing and unrelated to this PR — they require OPENAI_API_KEY credentials (Missing credentials error) and fail on main as well.
• ruff check / ruff format --check: clean on all changed files.

(Optional) Documentation Update

No public API changes. from lotus.models import LM, from lotus.models import SentenceTransformersRM, from lotus.vector_store import FaissVS, etc. all continue to work unchanged (now resolved lazily via PEP 562 __getattr__).

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Performance improvement
• Refactoring (no functional changes)

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code, updating docstrings
• I have made corresponding changes to the documentation
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

BEFORE SUBMITTING, PLEASE READ https://github.com/lotus-data/lotus/blob/main/CONTRIBUTING.md

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.