fix: correct prefill/decode block defaults when no overlaps (#3811)

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

lib/llm/src/kv_router/sequence.rs

@@ -826,17 +826,22 @@ impl ActiveSequencesMultiWorker {
         let token_sequence_shared = token_sequence.map(Arc::new);
         let mut receivers = Vec::new();
 
-        // Iterate through overlaps to process each WorkerWithDpRank
-        for (worker, overlap) in overlaps.scores.iter() {
-            // Check if the worker has a sender
-            if let Some(sender) = self.senders.get(worker) {
+        // Iterate through all workers, not just those with overlap
+        // This ensures we properly account for active tokens/blocks on all workers
+        for sender_entry in self.senders.iter() {
+            let worker = *sender_entry.key();
+            let sender = sender_entry.value();
+
+            // Get overlap for this worker (defaults to 0 if not in overlaps)
+            let overlap = *overlaps.scores.get(&worker).unwrap_or(&0);
+
             let (resp_tx, resp_rx) = tokio::sync::oneshot::channel();
-                receivers.push((*worker, resp_rx));
+            receivers.push((worker, resp_rx));
 
             if let Err(e) = sender.send(UpdateSequences::PotentialBlocksAndTokens {
                 token_sequence: token_sequence_shared.clone(),
                 isl,
-                    overlap: *overlap,
+                overlap,
                 resp_tx,
             }) {
                 tracing::error!(
@@ -846,7 +851,6 @@ impl ActiveSequencesMultiWorker {
                 );
             }
         }
-        }
 
         // Collect results from all workers
         for (worker, receiver) in receivers {