fix: normalize past_key_values across transformers DynamicCache API (#210)

[ousamabenyounes]

What does this PR do?

Fixes #210

llmlingua's iterative_compress_prompt relies on past_key_values being a tuple-of-(k,v) per layer: it slices past_key_values[0][0].shape[2] to read the cached length and iterates for k, v in past_key_values four separate times. Since transformers 4.36 the model returns a DynamicCache instead, which is why issue #210 shows AttributeError: 'list' object has no attribute 'get_seq_length' and why the existing baseline tests have been failing with ValueError: too many values to unpack (expected 2) at iterative_compress_prompt.

This PR normalizes past_key_values to the legacy tuple format after every model forward pass (via DynamicCache.to_legacy_cache() on older transformers, and via DynamicCache.layers on transformers >= 5.0 where to_legacy_cache was removed) and converts it back to a DynamicCache (via DynamicCache.from_legacy_cache() or the ddp_cache_data= constructor arg on newer versions) right before the next forward pass. The rest of the slicing/iteration code in iterative_compress_prompt is untouched.

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 — [Question]: error when compressing long prompt with LLMLingua1 & LongLLMLingua #210.
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests? — tests/test_issue_210.py

Verification

• Baseline (before this PR): 2 tests pass, 3 tests fail with ValueError: too many values to unpack (expected 2) in iterative_compress_prompt at llmlingua/prompt_compressor.py:1659.
• Post-fix: 13 tests pass (5 existing + 8 new regression tests in tests/test_issue_210.py), 0 regressions.

Tested against transformers 5.5.3 (latest on PyPI), which is what triggered the original issue. The helpers fall back to the legacy to_legacy_cache/from_legacy_cache methods when present, so older transformers releases keep working too.

Generated by Claude Code
Vibe coded by ousamabenyounes