fix: route compress_prompt to llmlingua2 path when use_slingua=True (#235)

[ousamabenyounes]

What does this PR do?

Fixes #235

When PromptCompressor is created with use_slingua=True, calling compress_prompt() raised:

TypeError: XLMRobertaForTokenClassification.forward() got an unexpected keyword argument 'past_key_values'

Root cause: compress_prompt() only routed to compress_prompt_llmlingua2() when self.use_llmlingua2 was True. __init__ already had the correct if use_llmlingua2 or use_slingua: guard (line 92) to call init_llmlingua2, but the dispatch in compress_prompt was missing the or self.use_slingua half.

Fix: one-line change at line 536 of llmlingua/prompt_compressor.py:

# Before
if self.use_llmlingua2:
# After
if self.use_llmlingua2 or self.use_slingua:

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Was this discussed/approved via a Github issue? Please add a link to it if that's the case.
  Issue: [Bug]: When calling slingua through PromptCompressor, use_llmlingua2 should also be set to true. #235
• Did you make sure to update the documentation with your changes? (not applicable — no new API)
• Did you write any new necessary tests?
  Added tests/test_slingua_routing.py with 3 unit tests covering the fixed routing, existing llmlingua2 routing, and the unchanged neither-flag path. Tests use sys.modules stubs so they run without ML dependencies.

Who can review?

@iofu728 @QianhuiWu @XufangLuo @mydmdm

Generated by Ora Studio
Vibe coded by ousamabenyounes

---

Vibe Coded by Ousama Ben Younes
Developed With Ora Studio (Claude Code)

[ousamabenyounes]

@microsoft-github-policy-service agree

[ousamabenyounes]

Auto-review (no human reviews yet)

Summary: This PR fixes a one-line routing bug: compress_prompt was not delegating to compress_prompt_llmlingua2 when use_slingua=True, even though the slingua path requires the LLMLingua-2 compression pipeline. The change is minimal and correct. A new test file exercises the three routing states (use_slingua, use_llmlingua2, neither) using sys.modules patching to avoid heavy ML dependency installation. Overall risk is low; the production change is trivial and well-motivated.

Verdict: comment

Findings (3):

• [duplication/warning] tests/test_slingua_routing.py:62 — Near-identical test bodies for use_slingua and use_llmlingua2
  test_use_slingua_routes_to_llmlingua2 (line 62) and test_use_llmlingua2_routes_to_llmlingua2 (line 75) share an identical 10-line body differing only in the flag passed to _make_compressor. These should be collapsed into a single parameterised test (e.g. @unittest.parameterized or subTest) so future changes to the assertion pattern don't require two synchronised edits.
  confidence: 0.85

• [bugs/warning] tests/test_slingua_routing.py:93 — Bare except swallows errors that occur before the routing branch
  test_neither_flag_does_not_route_to_llmlingua2 wraps the entire compress_prompt call in except Exception: pass. If an AttributeError or similar fires before the routing condition is even evaluated (e.g. a missing attribute introduced by a future refactor), the exception is silently discarded and the test still passes, giving a false green. Limit the swallowed exceptions to the expected stub-related ones (e.g. AttributeError, TypeError) or assert on the exception type with assertRaises.
  confidence: 0.80

• [tests/info] tests/test_slingua_routing.py:17 — sys.modules patching relies on undocumented pytest --dist=loadfile assumption
  _install_mock_modules is called at import time and mutates the global sys.modules dict. The docstring says isolation is guaranteed by pytest --dist=loadfile, but there is no Makefile or pytest.ini in this diff enforcing that flag. If tests run in the default in-process mode, MagicMock stubs for torch/transformers/numpy could bleed into other test modules imported afterward, silently breaking integration tests. Add a conftest.py fixture or a pytest.ini marker to enforce the worker-isolation assumption, or move _install_mock_modules inside a setUp/tearDown pair with proper cleanup.
  confidence: 0.72

---

Generated by review-all-pr self-review pass. The fixer will attempt to address these in the same run.