KV event decoder drops group_idx/medium/lora_name, breaking prefix matching for hybrid and multi-tier models

[Jeffwan]

Describe the bug

Our ZMQ decoder reads only the first four positional fields of vLLM's BlockStored (block_hashes, parent_block_hash, token_ids, block_size) and drops the rest. Upstream BlockStored (vllm/distributed/kv_events.py) now also carries medium, lora_name, extra_keys, and group_idx. Two of these affect correctness:

• group_idx: hybrid-attention models (sliding-window + full layers — Gemma, Ministral, Llama-4) hash the same token range into different KV-cache groups. Collapsing all groups into one hash space can produce false prefix matches.
• medium: with KV offloading, blocks live on CPU/remote tiers, but we count every block as a GPU hit, so prefix routing over-credits pods that only hold the prefix on a slower tier.

lora_name is also the canonical adapter id (lora_id is deprecated upstream); prefix_cache.go currently passes loraID=-1 everywhere, so adapters aren't isolated in the index.

Relevant: pkg/cache/kvcache/msgpack_decoder.go (parseEventArray), pkg/plugins/gateway/algorithms/prefix_cache.go.

Steps to Reproduce

1. Serve a hybrid-attention model (e.g. google/gemma-2-9b-it) with --kv-events-config enabled.
2. Send shared-prefix traffic across ≥2 replicas with the kv-event prefix router.
3. Observe requests routed to pods that don't actually have a GPU prefix hit.

Expected behavior

Decoder parses group_idx/medium/lora_name; the index keys prefix entries by (model, lora_name, group_idx) and scores non-GPU mediums lower.

Environment

AIBrix: main · vLLM: main (kv-events enabled).

[pjdurden]

Picking this up. Plan to split it so the keying change can be reviewed on its own:

PR1 (decoder + types): parseEventArray reads lora_id/medium/lora_name/group_idx (msgspec positions 5-9, bounds-checked so older vLLM arrays still decode) and BlockStoredEvent carries them, with decoder tests against the raw vLLM array layout. No routing/keying change, just stops dropping the fields.

PR2 (keying + scoring): key prefix entries by (model, lora_name, group_idx) and score non-GPU mediums lower, plumbing the fields through the kvevent handler into the indexer. That changes the MatchPrefix/AddPrefix signature across the indexers, so I would rather land it after PR1 and once you have looked at the keying approach.

Will open PR1 shortly. Happy to do it as one PR if you prefer.