diff --git a/mineru/backend/pipeline/model_init.py b/mineru/backend/pipeline/model_init.py
index 2006402a..f696dc00 100644
--- a/mineru/backend/pipeline/model_init.py
+++ b/mineru/backend/pipeline/model_init.py
@@ -1,6 +1,7 @@
 # Copyright (c) Opendatalab. All rights reserved.
 import os
 import threading
+from contextlib import contextmanager
 
 import torch
 from loguru import logger
@@ -24,12 +25,39 @@ PIPELINE_LAYOUT_INFERENCE_LOCK = threading.RLock()
 PIPELINE_MFR_INFERENCE_LOCK = threading.RLock()
 PIPELINE_OCR_DET_INFERENCE_LOCK = threading.RLock()
 PIPELINE_OCR_REC_INFERENCE_LOCK = threading.RLock()
+# 推理准入门闩：清理线程等待静默区时，阻止新的推理调用继续插队进入阶段锁。
+PIPELINE_INFERENCE_ADMISSION_LOCK = threading.RLock()
+PIPELINE_INFERENCE_STAGE_LOCKS = (
+    PIPELINE_LAYOUT_INFERENCE_LOCK,
+    PIPELINE_MFR_INFERENCE_LOCK,
+    PIPELINE_OCR_DET_INFERENCE_LOCK,
+    PIPELINE_OCR_REC_INFERENCE_LOCK,
+)
 
 
 def _run_with_inference_lock(inference_lock, inference_callable, *args, **kwargs):
     """在指定推理锁内执行真实 native 模型调用。"""
-    with inference_lock:
+    with PIPELINE_INFERENCE_ADMISSION_LOCK:
+        inference_lock.acquire()
+    try:
         return inference_callable(*args, **kwargs)
+    finally:
+        inference_lock.release()
+
+
+@contextmanager
+def pipeline_inference_quiet_zone():
+    """进入 pipeline/hybrid 共享模型推理静默区，等待四类阶段推理全部退出。"""
+    acquired_locks = []
+    with PIPELINE_INFERENCE_ADMISSION_LOCK:
+        try:
+            for inference_lock in PIPELINE_INFERENCE_STAGE_LOCKS:
+                inference_lock.acquire()
+                acquired_locks.append(inference_lock)
+            yield
+        finally:
+            for inference_lock in reversed(acquired_locks):
+                inference_lock.release()
 
 
 def run_layout_inference(inference_callable, *args, **kwargs):
diff --git a/mineru/utils/model_utils.py b/mineru/utils/model_utils.py
index 6c1f486a..1c2e4f68 100644
--- a/mineru/utils/model_utils.py
+++ b/mineru/utils/model_utils.py
@@ -180,7 +180,15 @@ def get_res_list_from_layout_res(layout_res, overlap_threshold=0.8):
     return ocr_res_list, table_res_list, single_page_mfdetrec_res
 
 
-def clean_memory(device='cuda'):
+def _pipeline_inference_quiet_zone():
+    """懒加载推理静默区，避免 utils 模块导入时和 pipeline backend 形成循环依赖。"""
+    from mineru.backend.pipeline.model_init import pipeline_inference_quiet_zone
+
+    return pipeline_inference_quiet_zone()
+
+
+def _clean_memory_without_inference(device='cuda'):
+    """执行实际显存/内存清理动作，调用方负责进入推理静默区。"""
     if str(device).startswith("cuda"):
         if torch.cuda.is_available():
             torch.cuda.empty_cache()
@@ -205,6 +213,11 @@ def clean_memory(device='cuda'):
     gc.collect()
 
 
+def clean_memory(device='cuda'):
+    with _pipeline_inference_quiet_zone():
+        _clean_memory_without_inference(device)
+
+
 def clean_vram(device, vram_threshold=8):
     total_memory = get_vram(device)
     if total_memory and total_memory <= vram_threshold: