pd with nixl backend

Dockerfile.nixl

@@ -0,0 +1,89 @@
+FROM nvcr.io/nvidia/tritonserver:25.04-py3-min as base
+ARG PYTORCH_VERSION=2.6.0
+ARG PYTHON_VERSION=3.9
+ARG CUDA_VERSION=12.4
+ARG MAMBA_VERSION=23.1.0-1
+ARG TARGETPLATFORM
+ARG INSTALL_NIXL=true
+
+ENV PATH=/opt/conda/bin:$PATH \
+    CONDA_PREFIX=/opt/conda
+
+RUN chmod 777 -R /tmp && apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
+    ca-certificates \
+    libssl-dev \
+    curl \
+    g++ \
+    make \
+    git && \
+    rm -rf /var/lib/apt/lists/*
+
+RUN case ${TARGETPLATFORM} in \
+    "linux/arm64")  MAMBA_ARCH=aarch64  ;; \
+    *)              MAMBA_ARCH=x86_64   ;; \
+    esac && \
+    curl -fsSL -o ~/mambaforge.sh -v "https://github.com/conda-forge/miniforge/releases/download/${MAMBA_VERSION}/Mambaforge-${MAMBA_VERSION}-Linux-${MAMBA_ARCH}.sh" && \
+    bash ~/mambaforge.sh -b -p /opt/conda && \
+    rm ~/mambaforge.sh
+
+RUN case ${TARGETPLATFORM} in \
+    "linux/arm64")  exit 1 ;; \
+    *)              /opt/conda/bin/conda update -y conda &&  \
+    /opt/conda/bin/conda install -y "python=${PYTHON_VERSION}" ;; \
+    esac && \
+    /opt/conda/bin/conda clean -ya
+
+
+WORKDIR /root
+
+COPY ./requirements.txt /lightllm/requirements.txt
+RUN --mount=type=cache,target=/root/.cache/pip pip install -r /lightllm/requirements.txt --ignore-installed --extra-index-url https://download.pytorch.org/whl/cu124
+
+RUN pip install --no-cache-dir https://github.com/ModelTC/flash-attn-3-build/releases/download/v2.7.4.post1/flash_attn-3.0.0b1-cp39-cp39-linux_x86_64.whl
+
+RUN --mount=type=cache,target=/root/.cache/pip pip install nvidia-nccl-cu12==2.25.1  # for allreduce hang issues in multinode H100
+
+RUN if [ "$INSTALL_NIXL" == "true" ]; then \
+        apt-get update && apt-get install -y cmake automake autotools-dev  libtool libz-dev && \
+        DEBIAN_FRONTEND=noninteractive apt-get -y install --reinstall libibverbs-dev rdma-core ibverbs-utils libibumad-dev; \
+        rm -rf /usr/lib/ucx && \
+        rm -rf /opt/hpcx/ucx && \
+        cd /usr/local/src && \
+        git clone https://github.com/openucx/ucx.git && \
+        cd ucx && 			     \
+        git checkout v1.19.x &&	     \
+        ./autogen.sh && ./configure     \
+        --enable-shared             \
+        --disable-static            \
+        --disable-doxygen-doc       \
+        --enable-optimizations      \
+        --enable-cma                \
+        --enable-devel-headers      \
+        --with-cuda=/usr/local/cuda \
+        --with-verbs=yes                \
+        --with-dm                   \
+        --with-gdrcopy=/usr/local   \
+        --with-efa                  \
+        --enable-mt &&              \
+        make -j &&                      \
+        make -j install-strip &&        \
+        ldconfig;    \
+    fi
+
+RUN if [ "$INSTALL_NIXL" == "true" ]; then \
+        apt-get update && apt-get install -y  pkg-config tmux net-tools;  \
+        cd /usr/local/src; \
+        pip install --upgrade meson pybind11 patchelf; \
+        git clone https://github.com/ai-dynamo/nixl.git -b main && \
+        cd nixl && \
+        rm -rf build && \
+        mkdir build && \
+        meson setup build/ --prefix=/usr/local/nixl --buildtype=release && \
+        cd build && \
+        ninja && \
+        ninja install && \
+        cd .. && pip install . --no-deps; \
+    fi
+
+COPY . /lightllm
+RUN pip install -e /lightllm --no-cache-dir

lightllm/common/basemodel/basemodel.py

@@ -174,6 +174,10 @@ def _init_kv_move_buffer(self):
         # p d 分离的推理模式下才需要做这一步初始化
         if self.run_mode in ["prefill", "decode"]:
             self.mem_manager.alloc_kv_move_buffer(self.mem_manager.size)
+        elif self.run_mode in ["nixl_prefill", "nixl_decode"]:
+            page_num = int(os.getenv("PD_NIXL_MOVE_PAGE_NUM", 32))
+            page_size = int(os.getenv("PD_NIXL_MOVE_PAGE_SIZE", 1024))
+            self.mem_manager.alloc_paged_kv_move_buffer(page_num, page_size)
 
     def _check_mem_size(self):
         self.max_total_token_num = self.mem_manager.size

lightllm/common/deepseek2_mem_manager.py

@@ -36,6 +36,12 @@ def alloc_kv_move_buffer(self, max_req_total_len):
         self.token_dim_size = self.kv_move_buffer.shape[-1] * self.kv_move_buffer.shape[-2]
         return
 
+    def alloc_paged_kv_move_buffer(self, page_num, page_size):
+        self.kv_move_buffer = torch.empty(
+            (page_num, page_size, self.layer_num, self.head_num, self.head_dim), dtype=self.dtype, device="cuda"
+        )
+        return
+
     def send_to_decode_node(
         self,
         move_tasks: List[KVMoveTask],

lightllm/common/mem_manager.py

@@ -96,6 +96,14 @@ def alloc_kv_move_buffer(self, max_req_total_len):
         self.token_dim_size = self.kv_move_buffer.shape[-2] * self.kv_move_buffer.shape[-1]
         return
 
+    def alloc_paged_kv_move_buffer(self, page_num, page_size):
+        if isinstance(self, MemoryManager) and type(self) != MemoryManager:
+            raise NotImplementedError("subclass need reimpl this method")
+        self.kv_move_buffer = torch.empty(
+            (page_num, page_size, self.layer_num, 2 * self.head_num, self.head_dim), dtype=self.dtype, device="cuda"
+        )
+        return
+
     def send_to_decode_node(
         self,
         move_tasks: List[KVMoveTask],

lightllm/server/api_cli.py

@@ -7,7 +7,7 @@ def make_argument_parser() -> argparse.ArgumentParser:
     parser.add_argument(
         "--run_mode",
         type=str,
-        choices=["normal", "prefill", "decode", "pd_master", "config_server"],
+        choices=["normal", "prefill", "decode", "nixl_prefill", "nixl_decode", "pd_master", "config_server"],
         default="normal",
         help="""set run mode, normal is started for a single server, prefill decode pd_master is for pd split run mode,
                 config_server is for pd split mode used to register pd_master node, and get pd_master node list,
@@ -54,6 +54,20 @@ def make_argument_parser() -> argparse.ArgumentParser:
         default=None,
         help="The port number for the config server in config_server mode.",
     )
+    parser.add_argument(
+        "--pd_nixl_remote_prefill_http_port",
+        type=int,
+        default=42001,
+        help="nixl pd mode, prefill node used for triggering prefill http port.",
+    )
+
+    parser.add_argument(
+        "--pd_nixl_remote_prefill_port",
+        type=int,
+        default=42002,
+        help="nixl pd mode, prefill and decode used for meta exchange.",
+    )
+
     parser.add_argument(
         "--model_name",
         type=str,

lightllm/server/api_start.py

@@ -67,7 +67,7 @@ def normal_or_p_d_start(args):
 
         enable_mps()
 
-    if args.run_mode not in ["normal", "prefill", "decode"]:
+    if args.run_mode not in ["normal", "prefill", "decode", "nixl_prefill", "nixl_decode"]:
         return
 
     assert args.zmq_mode in ["tcp://", "ipc:///tmp/"]

lightllm/server/core/objs/__init__.py

@@ -1,5 +1,5 @@
 from .sampling_params import SamplingParams
-from .req import Req, FinishStatus
+from .req import Req, FinishStatus, PDNIXLChunkedPrefillReq
 from .shm_req_manager import ShmReqManager
 from .rpc_shm import RpcShmParams, RpcShmResults, ShmSyncStatusArray
 from .start_args_type import StartArgs

lightllm/server/core/objs/req.py

@@ -106,6 +106,7 @@ def get_str(self):
             f"shm_cur_kv_len:{self.shm_cur_kv_len},"
             f"shm_cur_output_len:{self.shm_cur_output_len},"
             f"finish_status:{self.finish_status.is_finished()}"
+            f"group_id: {self.group_req_id}"
         )
 
     def init(
@@ -354,3 +355,55 @@ def post_init(
         # 错误问题。
         self.sample_params.max_new_tokens = self.sample_params.max_new_tokens + self.prefix_token_ids.size + 6
         return
+
+
+class PdNixlReqState(ctypes.Structure):
+    _pack_ = 4
+    _MAX_TP_SIZE = 32
+    _fields_ = [("dp_world_size", ctypes.c_int), ("state", ctypes.c_int * _MAX_TP_SIZE)]
+
+    def __init__(self):
+        self.dp_world_size = 0
+        self.state = (ctypes.c_int * self._MAX_TP_SIZE)(*([0] * self._MAX_TP_SIZE))
+
+    def set_dp_world_size(self, size: int):
+        self.dp_world_size = size
+
+    def set_tp_state(self, tp_id: int, state: int):
+        assert (
+            self.dp_world_size > 0 and tp_id >= 0 and tp_id < self.dp_world_size
+        ), f"tp_id {tp_id} out of range [0, {self.dp_world_size})"
+        self.state[tp_id] = state
+
+    def set_state(self):
+        assert self.dp_world_size > 0, "dp_world_size should be set before calling this"
+        unique_state = np.unique(self.state[: self.dp_world_size])
+        self.state[self.dp_world_size] = unique_state[0]
+
+    def get_state(self):
+        assert self.dp_world_size > 0, "dp_world_size should be set before calling this"
+        return self.state[self.dp_world_size]
+
+
+class PDNIXLChunkedPrefillReq(ChunkedPrefillReq):
+    _pack_ = 4
+    _fields_ = ChunkedPrefillReq._fields_ + [
+        # 用于pd nixl状态同步
+        ("pd_nixl_req_state", PdNixlReqState)
+    ]
+
+    def set_dp_world_size(self, dp_world_size):
+        self.pd_nixl_req_state.dp_world_size = dp_world_size
+
+    # called by each tp rank, no contention
+    def set_pd_req_rank_state(self, tp_id: int, state: int):
+        self.pd_nixl_req_state.set_tp_state(tp_id, state)
+
+    # state: -1 for failed, 0 for in progress, 1 for success
+    # set by router
+    def set_pd_req_state(self):
+        self.pd_nixl_req_state.set_state()
+
+    # read by all rank
+    def get_pd_req_state(self):
+        return self.pd_nixl_req_state.get_state()

lightllm/server/core/objs/shm_req_manager.py

@@ -3,7 +3,7 @@
 from lightllm.utils.envs_utils import get_unique_server_name
 from multiprocessing import shared_memory
 from lightllm.utils.log_utils import init_logger
-from .req import Req, NormalReq, ChunkedPrefillReq, TokenHealingReq
+from .req import Req, NormalReq, ChunkedPrefillReq, TokenHealingReq, PDNIXLChunkedPrefillReq
 from .shm_array import ShmArray
 from .atomic_array_lock import AtomicShmArrayLock, AtomicLockItem
 from .atomic_lock import AtomicShmLock
@@ -33,6 +33,9 @@ def get_req_class_type(self):
         if args.token_healing_mode:
             return TokenHealingReq
 
+        if args.run_mode in ["nixl_prefill", "nixl_decode"]:
+            return PDNIXLChunkedPrefillReq
+
         if args.disable_chunked_prefill:
             return NormalReq
         else:

lightllm/server/core/objs/start_args_type.py

@@ -6,7 +6,10 @@
 
 @dataclass
 class StartArgs:
-    run_mode: str = field(default="normal", metadata={"choices": ["normal", "prefill", "decode", "pd_master"]})
+    run_mode: str = field(
+        default="normal",
+        metadata={"choices": ["normal", "prefill", "decode", "pd_master", "nixl_prefill", "nixl_decode"]},
+    )
     host: str = field(default="127.0.0.1")
     port: int = field(default=8000)
     zmq_mode: str = field(

lightllm/server/httpserver/manager.py

@@ -101,7 +101,7 @@ def __init__(
         self.metric_client = MetricClient(metric_port)
 
         self.pd_mode: NodeRole = NodeRole(self.args.run_mode)
-        assert self.pd_mode in [NodeRole.P, NodeRole.D, NodeRole.NORMAL]
+        assert self.pd_mode in [NodeRole.P, NodeRole.D, NodeRole.NORMAL, NodeRole.NP, NodeRole.ND]
         self.id_gen = ReqIDGenerator()
         self.first_time_costs = MovingAverage()
         self.per_token_costs = MovingAverage()
@@ -228,7 +228,7 @@ def alloc_req_id(self, sampling_params, is_health_req: bool = False):
         # health 请求 request_id 为负数，直接返回
         if is_health_req:
             return sampling_params.group_request_id
-        if self.pd_mode == NodeRole.NORMAL:
+        if self.pd_mode.is_normal():
             if not self.is_multinode_tp:
                 group_request_id = self.id_gen.generate_id()
             else:
@@ -238,7 +238,7 @@ def alloc_req_id(self, sampling_params, is_health_req: bool = False):
                     assert sampling_params.group_request_id != -1
                     group_request_id = sampling_params.group_request_id
             sampling_params.group_request_id = group_request_id
-        elif self.pd_mode == NodeRole.P or self.pd_mode == NodeRole.D:
+        elif self.pd_mode.is_P_or_D():
             assert sampling_params.group_request_id is not None, "p d mode, group_request_id must be setting"
             group_request_id = sampling_params.group_request_id
         else:
@@ -419,7 +419,7 @@ async def transfer_to_next_module_or_node(
         if self.is_multinode_tp_master:
             async with self.transfer_lock:
                 for sender in self.multinode_req_manager:
-                    sender.send_pyobj(
+                    await sender.send_pyobj(
                         (prompt, sampling_params, original_multimodal_params),
                         protocol=pickle.HIGHEST_PROTOCOL,
                     )
@@ -448,35 +448,37 @@ async def transfer_to_next_module(
         group_req_objs: Optional[GroupReqObjs] = None,
     ):
 
-        if self.pd_mode == NodeRole.P:
+        if self.pd_mode.is_P():
             if self.enable_multimodal:
-                self.send_to_visual.send_pyobj(
+                await self.send_to_visual.send_pyobj(
                     group_req_objs.to_group_req_index(),
                     protocol=pickle.HIGHEST_PROTOCOL,
                 )
             else:
-                self.send_to_router.send_pyobj(
+
+                # P 模式下，直接将请求发送到路由器
+                await self.send_to_router.send_pyobj(
                     group_req_objs.to_group_req_index(),
                     protocol=pickle.HIGHEST_PROTOCOL,
                 )
             return
 
-        if self.pd_mode == NodeRole.D:
+        if self.pd_mode.is_D():
             # 在 D 模式下，不需要传输真的多模态参数，因为其已经被 P 处理好了, 传输一个空的即可
-            self.send_to_router.send_pyobj(
+            await self.send_to_router.send_pyobj(
                 group_req_objs.to_group_req_index(),
                 protocol=pickle.HIGHEST_PROTOCOL,
             )
             return
 
-        if self.pd_mode == NodeRole.NORMAL:
+        if self.pd_mode.is_normal():
             if self.enable_multimodal:
-                self.send_to_visual.send_pyobj(
+                await self.send_to_visual.send_pyobj(
                     group_req_objs.to_group_req_index(),
                     protocol=pickle.HIGHEST_PROTOCOL,
                 )
             else:
-                self.send_to_router.send_pyobj(
+                await self.send_to_router.send_pyobj(
                     group_req_objs.to_group_req_index(),
                     protocol=pickle.HIGHEST_PROTOCOL,
                 )
@@ -521,7 +523,7 @@ async def _wait_to_token_package(
                     # pd master 节点需要这个做统计信息， 所以放在元数据中返回给 pd master 节点
                     metadata["prompt_tokens"] = prompt_tokens
                     # p 节点返回 prompt_ids 信息，防止 d 节点重新 encode
-                    if self.pd_mode == NodeRole.P and is_first_token:
+                    if self.pd_mode.is_P() and is_first_token:
                         metadata["prompt_ids"] = prompt_ids
 
                     prompt_cache_len = metadata.pop("prompt_cache_len", 0)
@@ -623,7 +625,7 @@ async def recycle_resource_loop(self):
                 pre_time_mark = time.time()
                 for req_status in self.req_id_to_out_inf.values():
                     logger.info(
-                        f"left req id {req_status.group_req_objs.group_req_id}"
+                        f"left req id {req_status.group_req_objs.group_req_id} "
                         f"can release {req_status.group_req_objs.shm_req_objs[0].can_released_mark} "
                         f"refcount {req_status.group_req_objs.shm_req_objs[0].ref_count}"
                     )