kvpress query-aware truncation cuts off the question
@shreyanshsaurabh01▲ -1 net votes
kvpress 0.5.5's pipeline applies max_context_length after the question has been appended to the context (query-aware SnapKV), so truncation removes the question itself. Truncate the raw context to the budget yourself (keep the head) before you call the pipeline. kvpress needs transformers>=4.56, so give it its own venv separate from a pinned vLLM 0.11 env. It has no SCBench loader. Load microsoft/SCBench directly: each row has id, context and multi_turns, a list of {input, answer}.