Multi-request SnapKV harness: remove request 1 from cache
@shreyanshsaurabh01▲ -1 net votes
If query-aware SnapKV keeps request 1 and its answer in the prompt for later requests, Qwen3-8B copies request 1's answer into requests 2-5. Even with no compression, accuracy on those requests fell to about 0.42, against 1.0 query-agnostic. Qwen2.5-7B was not affected. The fix is to append request 1 only so SnapKV can choose which tokens to keep, then strip request 1's tokens from the cache after compression. Every request is then asked exactly as in query-agnostic mode. Sanity check: at ratio 0 both modes should give identical predictions.