standalone-indexer: hash a vLLM cache_salt into its sequence's first block

[Shang-Pin]

Summary

vLLM mixes a request's cache_salt into the hash of the sequence's first block and publishes it in that block's BlockStored extra_keys. The normalizer kept only 64-hex MM ids, which caused two problems:

• Salted and unsalted chains with the same tokens shared tree edges while carrying different engine hashes. The tree keeps one engine hash per edge position, so the second chain's children were rejected as ParentBlockNotFound. This is the same failure plain image hashing caused in standalone-indexer: hash image blocks from tokens only #43/standalone-indexer (h24 line): hash image blocks from tokens only #44.
• Probes scored a salted prefix as reusable by any request.

This change:

• In a BlockStored with no parent, a bare string in block 0 that is neither an MM id nor the event's lora_name is the salt. vLLM's key order is lora + mm + [cache_salt only when start_token_idx == 0] + prompt_embeds.
• The salt joins block 0's MM ids as cache_salt_extra_hash(salt). A 64-hex salt parses like an MM id; any other salt hashes to its XXH3-64. A salted block therefore hashes as xxh3(tokens LE ++ sorted(mm ids + salt) LE, seed), with no new hash scheme.
• Non-root events, other blocks, and extra_keys_to_block_mm_infos (used by lib/llm) are unchanged. Unsalted traffic hashes exactly as before.
• The deepapi probe hashes the salt the same way, behind the per-model kv_probe_cache_salt flag (deepinfra/backend PR linked below). Tests pin hash vectors shared with deepapi's tests.

Known gap, same as for images: /kv_recover TreeDumps carry token-only hashes, so salted blocks restored at startup stay plain until evicted.

Validation

• cargo test -p dynamo-kv-router --lib: 555 passed on both lines (rustc 1.93.1). This includes 5 new zmq_wire tests:
  • salt hashed into the first block only;
  • ignored on non-root events;
  • a LoRA name is not taken as a salt;
  • salt plus image in the same block;
  • fixed vectors cross-checked with Python xxhash.

Rollout

Roll per model, together with the deepapi flag: indexer images first, then i models -n <model> feature-flags kv_probe_cache_salt.

🤖 Generated with Claude Code