chore: remove some clones in Indexer (#5612)

Signed-off-by: PeaBrane <yanrpei@gmail.com>

lib/llm/src/kv_router.rs

@@ -498,7 +498,7 @@ impl KvRouter {
         let isl_tokens = tokens.len();
 
         let block_hashes = compute_block_hash_for_seq(tokens, self.block_size, None);
-        let overlap_scores = self.indexer.find_matches(block_hashes.clone()).await?;
+        let overlap_scores = self.indexer.find_matches(block_hashes).await?;
 
         // Compute seq_hashes only if scheduler needs it for active blocks tracking
         let maybe_seq_hashes = self
@@ -596,7 +596,7 @@ impl KvRouter {
     pub async fn get_potential_loads(&self, tokens: &[u32]) -> Result<Vec<PotentialLoad>> {
         let isl_tokens = tokens.len();
         let block_hashes = compute_block_hash_for_seq(tokens, self.block_size, None);
-        let overlap_scores = self.indexer.find_matches(block_hashes.clone()).await?;
+        let overlap_scores = self.indexer.find_matches(block_hashes).await?;
 
         let maybe_seq_hashes = self
             .kv_router_config

lib/llm/src/kv_router/indexer.rs

@@ -979,13 +979,16 @@ impl KvIndexer {
 
                         Some(event) = event_rx.recv() => {
                             let event_type = KvIndexerMetrics::get_event_type(&event.event.data);
-                            let result = trie.apply_event(event.clone());
+                            // Only clone if we need the event for prune_manager afterward
+                            let event_for_prune = prune_manager.is_some().then(|| event.clone());
+                            let result = trie.apply_event(event);
                             let result_is_ok = result.is_ok();
                             metrics.increment_event_applied(event_type, result);
 
                             // Track blocks in PruneManager if TTL is enabled and event was stored successfully
                             let Some(ref mut pm) = prune_manager else { continue };
                             if !result_is_ok { continue };
+                            let Some(ref event) = event_for_prune else { continue };
                             let KvCacheEventData::Stored(ref store_data) = event.event.data else { continue };
 
                             let worker = WorkerWithDpRank::new(event.worker_id, event.event.dp_rank);
@@ -1707,13 +1710,16 @@ impl KvIndexerSharded {
 
                             Some(event) = shard_event_rx.recv() => {
                                 let event_type = KvIndexerMetrics::get_event_type(&event.event.data);
-                                let result = trie.apply_event(event.clone());
+                                // Only clone if we need the event for prune_manager afterward
+                                let event_for_prune = prune_manager.is_some().then(|| event.clone());
+                                let result = trie.apply_event(event);
                                 let result_is_ok = result.is_ok();
                                 metrics.increment_event_applied(event_type, result);
 
                                 // Track blocks in PruneManager if TTL is enabled and event was stored successfully
                                 let Some(ref mut pm) = prune_manager else { continue };
                                 if !result_is_ok { continue };
+                                let Some(ref event) = event_for_prune else { continue };
                                 let KvCacheEventData::Stored(ref store_data) = event.event.data else { continue };
 
                                 let worker = WorkerWithDpRank::new(event.worker_id, event.event.dp_rank);