From 4ac65e3c7d31b4b347f65fd253aa68a1fab81d92 Mon Sep 17 00:00:00 2001
From: huangtingwei <141888744+huangtingwei9988@users.noreply.github.com>
Date: Fri, 14 Nov 2025 03:31:53 +0800
Subject: [PATCH] [PD / HiCache]fix deocde kvcache offload manager memory leak
 (#12774)

---
 .../decode_kvcache_offload_manager.py         |  5 +++--
 python/sglang/srt/server_args.py              | 19 ++++++++++++-------
 2 files changed, 15 insertions(+), 9 deletions(-)

diff --git a/python/sglang/srt/disaggregation/decode_kvcache_offload_manager.py b/python/sglang/srt/disaggregation/decode_kvcache_offload_manager.py
index 40ed3be1f..bf1d22cdf 100644
--- a/python/sglang/srt/disaggregation/decode_kvcache_offload_manager.py
+++ b/python/sglang/srt/disaggregation/decode_kvcache_offload_manager.py
@@ -184,14 +184,15 @@ class DecodeKVCacheOffloadManager:
             finish_count -= 1
 
     def _release_finished_req(self, req: Req, prefill_offloaded_len: int):
-        # FIXME: not sure which length to use here: kv_allocated_len or kv_committed_len
+        kv_committed_len = req.pop_committed_kv_cache()
         kv_indices = self.req_to_token_pool.req_to_token[
-            req.req_pool_idx, prefill_offloaded_len : req.kv_allocated_len
+            req.req_pool_idx, prefill_offloaded_len:kv_committed_len
         ]
 
         # Free the incremental part of the request
         self.token_to_kv_pool_allocator.free(kv_indices)
         self.req_to_token_pool.free(req.req_pool_idx)
+        self.tree_cache.protected_size_ -= len(req.prefix_indices)
 
     def _check_backup_progress(self, finish_count):
         """Check the progress of backup from host to storage."""
diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py
index 036748f7c..6711ae5f2 100644
--- a/python/sglang/srt/server_args.py
+++ b/python/sglang/srt/server_args.py
@@ -1762,13 +1762,18 @@ class ServerArgs:
                 "and cannot be used at the same time. Please use only one of them."
             )
 
-        if (
-            self.disaggregation_decode_enable_offload_kvcache
-            and self.disaggregation_mode != "decode"
-        ):
-            raise ValueError(
-                "The argument disaggregation-decode-enable-offload-kvcache is only supported for decode side."
-            )
+        if self.disaggregation_decode_enable_offload_kvcache:
+            if self.disaggregation_mode != "decode":
+                raise ValueError(
+                    "The argument disaggregation-decode-enable-offload-kvcache is only supported for decode side."
+                )
+            if (
+                self.disaggregation_mode == "decode"
+                and envs.SGLANG_ENABLE_SPEC_V2.get()
+            ):
+                raise ValueError(
+                    "Spec v2 and decode offload kv cache are incompatible and cannot be enabled together."
+                )
 
     def _handle_metrics_labels(self):
         if (