Fix flaky Qwen3-Next KL divergence tests by reverting mamba slot release (#18910)

python/sglang/srt/managers/scheduler.py

@@ -2091,8 +2091,6 @@ class Scheduler(
                 running_loras.add(req.lora_id)
 
             if res != AddReqResult.CONTINUE:
-                self.maybe_release_mamba_cache(req)
-
                 if res == AddReqResult.NO_TOKEN:
                     if self.enable_hierarchical_cache:
                         # Set batch_is_full after making sure there are requests that can be served
@@ -2190,25 +2188,6 @@ class Scheduler(
 
         return new_batch
 
-    def maybe_release_mamba_cache(self, req: Req) -> None:
-        """Release mamba slot if allocated via COW but scheduling failed.
-
-        Without this, the slot remains held by a waiting request, causing
-        check_memory() to detect a "memory leak" and crash the server.
-        The next schedule round will re-allocate safely via match_prefix().
-
-        Note: In disaggregation DECODE mode, mamba state is transferred from PREFILL and
-        is not recoverable if freed, so we do not free it here. To avoid false-positive
-        leak checks in this situation, self_check_during_idle skips memory checking when
-        the waiting queue is not empty.
-        """
-        if (
-            req.mamba_pool_idx is not None
-            and self.disaggregation_mode != DisaggregationMode.DECODE
-        ):
-            self.req_to_token_pool.mamba_pool.free(req.mamba_pool_idx.unsqueeze(-1))
-            req.mamba_pool_idx = None
-
     def update_running_batch(self, batch: ScheduleBatch) -> Optional[ScheduleBatch]:
         """Update the current running decoding batch."""
         initial_bs = batch.batch_size()