From ef92b4eb88905b7b9415548e060870a99d5b6104 Mon Sep 17 00:00:00 2001
From: Yi Zhang <1109276519@qq.com>
Date: Fri, 26 Dec 2025 17:14:37 +0800
Subject: [PATCH] [BUGFIX] fix edge case for qwen3-next (#14209)

---
 python/sglang/srt/managers/scheduler.py    |  5 ++++-
 python/sglang/srt/mem_cache/memory_pool.py | 11 +++++------
 2 files changed, 9 insertions(+), 7 deletions(-)

diff --git a/python/sglang/srt/managers/scheduler.py b/python/sglang/srt/managers/scheduler.py
index ef9df64e4..9d1f0450f 100644
--- a/python/sglang/srt/managers/scheduler.py
+++ b/python/sglang/srt/managers/scheduler.py
@@ -2583,7 +2583,10 @@ class Scheduler(
                 release_kv_cache(req, self.tree_cache)
 
             # For mamba radix cache
-            if req.mamba_pool_idx is not None:
+            if (
+                req.mamba_pool_idx is not None
+                and self.disaggregation_mode != DisaggregationMode.DECODE
+            ):
                 release_kv_cache(req, self.tree_cache, is_insert=False)
             logger.debug(f"Abort queued request. {req.rid=}")
 
diff --git a/python/sglang/srt/mem_cache/memory_pool.py b/python/sglang/srt/mem_cache/memory_pool.py
index 11446e102..b75d3be86 100644
--- a/python/sglang/srt/mem_cache/memory_pool.py
+++ b/python/sglang/srt/mem_cache/memory_pool.py
@@ -276,6 +276,11 @@ class MambaPool:
             self.mamba_cache.conv[i][:, select_index] = 0
         self.mamba_cache.temporal[:, select_index] = 0
 
+        # fill allocated slots with zeros
+        for i in range(len(self.mamba_cache.conv)):
+            self.mamba_cache.conv[i][:, select_index] = 0
+        self.mamba_cache.temporal[:, select_index] = 0
+
         return select_index
 
     def free(self, free_index: torch.Tensor):
@@ -284,12 +289,6 @@ class MambaPool:
         self.free_slots = torch.cat((self.free_slots, free_index))
 
     def clear(self):
-        # Zero the entire mamba cache before resetting free_slots
-        # This ensures that when slots are reallocated, they start with clean state
-        for i in range(len(self.mamba_cache.conv)):
-            self.mamba_cache.conv[i].zero_()
-        self.mamba_cache.temporal.zero_()
-
         self.free_slots = torch.arange(self.size, dtype=torch.int64, device=self.device)
 
     def copy_from(self, src_index: torch.Tensor, dst_index: torch.Tensor):