[Fix] Lock PP dynamic-chunk profiling requests before releasing through the tree cache (#37509)

python/sglang/srt/managers/schedule_batch.py

@@ -1371,9 +1371,8 @@ class Req(ReqDllmMixin):
         appending only the new output tokens.
 
         Falls back to a full rebuild when the in-place append is invalid:
-        - aliasing: scheduler_pp_mixin assigns full_untruncated_fill_ids =
+        - aliasing: full_untruncated_fill_ids is origin_input_ids itself, so
-          origin_input_ids directly, so extending in place would write output
+          extending in place would write output tokens into the origin;
-          tokens into the origin;
         - lengths disagree: fresh req (array still empty), retraction
           (output_ids reset to empty), or set_finish_with_abort (origin
           replaced by a 1-token stub).

python/sglang/srt/managers/scheduler.py

@@ -1253,7 +1253,7 @@ class Scheduler(
                 self.profile_and_init_predictor()
             except Exception as e:
                 logger.warning(
-                    f"[PP Dynamic Chunk] Failed to profile prefill latency: {e}. "
+                    f"[PP Dynamic Chunk] Failed to profile prefill latency: {e!r}. "
                     "Dynamic chunking will be disabled."
                 )
                 self.enable_dynamic_chunking = False

python/sglang/srt/managers/scheduler_pp_mixin.py

@@ -634,8 +634,11 @@ class SchedulerPPMixin:
                     origin_input_ids=input_ids,
                     sampling_params=sampling_params,
                 )
-                req.full_untruncated_fill_ids = req.origin_input_ids
+                # Walk the same match -> lock -> alloc lifecycle as a scheduled
-                req.logprob_start_len = -1
+                # request so release_kv_cache can release it symmetrically.
+                req.init_next_round_input(self.tree_cache)
+                lock = self.tree_cache.inc_lock_ref(req.last_node)
+                req.swa_uuid_for_lock = lock.swa_uuid_for_lock
                 req.set_extend_range(
                     len(req.prefix_indices), len(req.full_untruncated_fill_ids)
                 )