feat: update compatibility for transformers>=4.57.1 and vLLM 0.14

[avinsongw]

Summary

• Update dependency versions to support transformers>=4.57.1 and vLLM>=0.11.0
• Fix deprecated DynamicCache.from_legacy_cache() API in HuggingFace backend
• Major refactor of vLLM backend for V1 engine compatibility (vLLM 0.11+ removed V0 engine)

Changes

pyproject.toml

• transformers>=4.57.1 (was >=4.36.0)
• vllm>=0.11.0 (was >=0.6.6,<=0.10.0)

genlm/backend/llm/hf.py

• Replace deprecated DynamicCache.from_legacy_cache(pasts) with DynamicCache(pasts)

genlm/backend/llm/vllm.py

• Remove PassThroughLogitsProcessor class (V1 doesn't support per-request logits processors)
• Use max_logprobs parameter to get full vocabulary logprobs instead
• Update Counter import path (vllm.utils.counter)
• Remove deprecated disable_async_output_proc and disable_log_requests params
• Update tokenizer access for V1 API (async_llm_engine.tokenizer)
• Update cleanup method (async_engine.shutdown())

tests/conftest.py

• Add max_logprobs to ReferenceVirtualLM for V1 compatibility
• Update tokenizer access
• Simplify cleanup

Test plan

• All HF backend tests pass
• Cache/Trie/Vocabulary tests pass (47 passed, 3 skipped)
• vLLM integration verified with standalone test (full vocab logprobs, properly normalized)

🤖 Generated with Claude Code