Add Scheduler::running_len

mistralrs-core/src/dummy_paged_attention/scheduler.rs

@@ -345,7 +345,10 @@ impl Scheduler for PagedAttentionScheduler {
         }
     }
     fn waiting_len(&self) -> usize {
-        self.waiting.len()
+        self.waiting.len() + self.swapped_out.len()
+    }
+    fn running_len(&self) -> usize {
+        self.running.len()
     }
     fn block_tables(&self) -> Option<&BlockTables> {
         Some(&self.block_engine.block_tables)

mistralrs-core/src/paged_attention/scheduler.rs

@@ -346,7 +346,10 @@ impl Scheduler for PagedAttentionScheduler {
         }
     }
     fn waiting_len(&self) -> usize {
-        self.waiting.len()
+        self.waiting.len() + self.swapped_out.len()
+    }
+    fn running_len(&self) -> usize {
+        self.running.len()
     }
     fn block_tables(&self) -> Option<&BlockTables> {
         Some(&self.block_engine.block_tables)

mistralrs-core/src/scheduler/default_scheduler.rs

@@ -308,6 +308,9 @@ impl Scheduler for DefaultScheduler<VecDeque<Sequence>> {
     fn waiting_len(&self) -> usize {
         self.waiting.len()
     }
+    fn running_len(&self) -> usize {
+        self.running.len()
+    }
     fn add_seq(&mut self, seq: Sequence) {
         if seq.is_running() {
             // prefill case

mistralrs-core/src/scheduler/mod.rs

@@ -48,6 +48,7 @@ pub enum SchedulerOutput<'a> {
 pub trait Scheduler {
     fn schedule(&mut self) -> SchedulerOutput<'_>;
     fn waiting_len(&self) -> usize;
+    fn running_len(&self) -> usize;
     fn add_seq(&mut self, seq: Sequence);
     /// This may do nothing. It depends on the implementation
     fn free_finished_sequence_groups(&mut self);