Fix the _chunked_req_scheduled_last_iter flag with a content-based stash gate (#26938)

python/sglang/srt/managers/scheduler.py

@@ -936,15 +936,6 @@ class Scheduler(
         elif self.chunked_prefill_size is not None and self.chunked_prefill_size <= 0:
             self.chunked_prefill_size = None
         self.chunked_req = None
-        # Tracks whether the current self.chunked_req was actually scheduled
-        # into last iteration's batch (i.e., in can_run_list -> got a fresh
-        # req_pool_idx from prepare_for_extend). Used to gate the
-        # stash_chunked_request call at the top of get_next_batch_to_run:
-        # if add_chunked_req early-returned under hybrid-SWA pressure,
-        # the req_pool_idx was already freed and the full_untruncated_fill_ids
-        # was rebuilt by init_next_round_input, so running stash would
-        # double-free and corrupt prefix_indices.
-        self._chunked_req_scheduled_last_iter = False
         self.is_mixed_chunk = (
             self.chunked_prefill_size is not None
             and self.server_args.enable_mixed_chunk
@@ -2443,7 +2434,11 @@ class Scheduler(
             # only finished requests to running_batch.
             chunked_req_to_exclude.add(self.chunked_req)
 
-            if self._chunked_req_scheduled_last_iter:
+            # Stash (cache) the previous chunk only when it produced new KV
+            # beyond what is already cached. A parked chunk (add_chunked_req
+            # hybrid-SWA early-return) leaves fill_len == len(prefix_indices),
+            # so there is nothing new to cache and stashing would be a no-op.
+            if self.chunked_req.fill_len > len(self.chunked_req.prefix_indices):
                 self.stash_chunked_request(self.chunked_req)
 
         # HiSparse has its own prefill-to-decode transition; skip last_batch merge.
@@ -2645,11 +2640,6 @@ class Scheduler(
         if self.chunked_req is not None:
             self.chunked_req.init_next_round_input()
             self.chunked_req = adder.add_chunked_req(self.chunked_req)
-            self._chunked_req_scheduled_last_iter = (
-                self.chunked_req in adder.can_run_list
-            )
-        else:
-            self._chunked_req_scheduled_last_iter = False
 
         if self.enable_lora:
             running_loras = {req.lora_id for req in self.running_batch.reqs}
@@ -2749,9 +2739,6 @@ class Scheduler(
             # Update chunked prefill
             assert self.chunked_req is None
             self.chunked_req = adder.new_chunked_req
-            # new_chunked_req is added to can_run_list by add_one_req,
-            # so it will be scheduled this iter -> stash is needed next iter.
-            self._chunked_req_scheduled_last_iter = True
 
         if self.chunked_req is not None:
             self.chunked_req.inflight_middle_chunks += 1