fix

Author: wangzaijun

README.md

@@ -22,12 +22,11 @@ LightLLM is a Python-based LLM (Large Language Model) inference and serving fram
 
 ## Tech Blogs
 - [2025/11] 🚀 Prefix KV Cache Transfer between DP rankers is now supported! Check out the technical deep dive in our [blog post](https://light-ai.top/lightllm-blog/2025/11/18/dp_kv_fetch.html).
-- [2025/05] LightLLM paper on constrained decoding accepted by [ACL2025](https://arxiv.org/pdf/2506.03887) (Pre $^3$: Enabling Deterministic Pushdown Automata for Faster Structured LLM Generation). For a more accessible overview of the research with key insights and examples, check out our blog post: [LightLLM Blog](https://www.light-ai.top/lightllm-blog/2025/06/15/pre3.html)
 
 ## News
-
 - [2025/09] 🔥 LightLLM [v1.1.0](https://www.light-ai.top/lightllm-blog/2025/09/03/lightllm.html) release!
 - [2025/08] Pre $^3$ achieves the outstanding paper award of [ACL2025](https://2025.aclweb.org/program/awards/).
+- [2025/05] LightLLM paper on constrained decoding accepted by [ACL2025](https://arxiv.org/pdf/2506.03887) (Pre $^3$: Enabling Deterministic Pushdown Automata for Faster Structured LLM Generation). For a more accessible overview of the research with key insights and examples, check out our blog post: [LightLLM Blog](https://www.light-ai.top/lightllm-blog/2025/06/15/pre3.html)
 - [2025/04] LightLLM paper on request scheduler published in [ASPLOS’25](https://dl.acm.org/doi/10.1145/3676641.3716011) (Past-Future Scheduler for LLM Serving under SLA Guarantees)
 - [2025/02] 🔥 LightLLM v1.0.0 release, achieving the **fastest DeepSeek-R1** serving performance on single H200 machine.
 

lightllm/server/router/model_infer/mode_backend/base_backend.py

@@ -213,6 +213,8 @@ def init_model(self, kvargs):
             self.args.run_mode in ["nixl_prefill", "nixl_decode", "prefill", "decode"]
             or self.args.enable_dp_prompt_cache_fetch
         ):
+            # 如果存在需要跨进程使用mem manger的特性，则将mem manager写入到 shm中，方便
+            # 读取
             self.model.mem_manager.write_to_shm(req_manager=self.model.req_manager)
             dist.barrier(group=self.node_nccl_group)
 
@@ -229,9 +231,6 @@ def init_model(self, kvargs):
         if self.args.mtp_mode:
             self.init_mtp_draft_model(kvargs)
 
-        # 如果存在需要跨进程使用mem manger的特性，则将mem manager写入到 shm中，方便
-        # 读取
-
         # 启动infer_loop_thread, 启动两个线程进行推理，对于具备双batch推理折叠得场景
         # 可以降低 cpu overhead，大幅提升gpu得使用率。
         self.infer_loop_thread = threading.Thread(target=self.infer_loop, daemon=True)

lightllm/server/router/model_infer/mode_backend/dp_backend/dp_shared_kv_trans.py

@@ -4,7 +4,7 @@
 import dataclasses
 import torch
 from typing import List
-from lightllm.common.mem_manager import MemoryManager
+from lightllm.common.kv_cache_mem_manager import MemoryManager
 from lightllm.utils.envs_utils import get_unique_server_name, get_env_start_args
 from lightllm.utils.dist_utils import get_dp_rank_in_node
 from lightllm.server.core.objs.shm_array import ShmArray

lightllm/server/router/model_infer/mode_backend/dp_backend/impl.py

@@ -1,11 +1,8 @@
 import torch
 import time
-import numpy as np
-import os
 import torch.nn.functional as F
 import torch.distributed as dist
 from typing import List, Tuple, Optional, Callable
-from lightllm.common.kv_trans_kernel.kv_trans_v2 import kv_trans_for_dp
 from lightllm.server.router.model_infer.mode_backend.base_backend import ModeBackend
 from lightllm.common.basemodel.batch_objs import ModelOutput, ModelInput
 from lightllm.server.router.model_infer.infer_batch import InferSamplingParams, g_infer_context, InferReq
@@ -26,8 +23,6 @@
 from lightllm.server.router.model_infer.pin_mem_manager import g_pin_mem_manager
 from lightllm.common.basemodel.triton_kernel.mtp_utils import mtp_scatter_next_token_ids
 from .control_state import DPControlState
-from lightllm.common.mem_manager import MemoryManager
-from .dp_shared_kv_trans import DPKVSharedMoudle
 
 
 class DPChunkedPrefillBackend(ModeBackend):