[XPU]Add PD ep4tp4 CudaGraph CI Case

fastdeploy/worker/xpu_model_runner.py

@@ -1364,12 +1364,8 @@ def get_input_length_list(self, num_tokens: int, batch_size: int, expected_decod
         """
         max_dec_len = expected_decode_len + 1
         input_length = min(num_tokens // batch_size, self.model_config.max_model_len - max_dec_len)
-
-        # NOTE(wanglongzhi): When the full length is too large, DeepEP's buffer size will not be enough to cause the result to appear nan.
-        # TODO(wanglongzhi): Figure out the accurate buffer size of DeepEP.
         if self.fd_config.parallel_config.enable_expert_parallel:
-            input_length = min(input_length, 32)
-
+            input_length = min(input_length, 4)
         block_num = (
             input_length + self.cache_config.block_size - 1
         ) // self.cache_config.block_size + self.cache_config.enc_dec_block_num

start_pd_ep_cudagraph.sh

@@ -0,0 +1,219 @@
+#!/bin/bash
+# FastDeploy 服务启动脚本
+# 支持两种模式：集中式EP服务 / PD分离服务
+
+set -e
+
+#============================================================
+# 公共配置
+#============================================================
+export PYTHONPATH=/opt/output/work_dir/ssd1/yinwei06/FastDeploy
+# export FD_DEBUG=1
+
+# 取消代理设置
+unset http_proxy
+unset https_proxy
+
+# 模型路径
+MODEL_PATH="/opt/output/work_dir/ssd1/yinwei06/FD_slice_merge/10.63.230.221:8891//ERNIE-4.5-21B-A3B-Paddle"
+
+#============================================================
+# 第一部分：清理残余进程和日志
+#============================================================
+cleanup() {
+    echo "[1/3] 清理残余进程和日志..."
+
+    # 清理日志目录
+    rm -rf log
+    rm -rf log_router
+    mkdir -p log_router
+
+    # 杀死残余 Python 进程
+    local patterns=("python -m" "python -u" "python -c" "splitwise_role")
+    for pattern in "${patterns[@]}"; do
+        pids=$(ps aux | grep "$pattern" | grep -v grep | awk '{print $2}' || true)
+        if [ -n "$pids" ]; then
+            echo "  - 终止进程: $pattern (PIDs: $pids)"
+            echo "$pids" | xargs kill -9 2>/dev/null || true
+        fi
+    done
+
+    # 杀掉占用的端口
+    local ports=(8188 8189 8199 8209 8210 8211 8212)
+    for port in "${ports[@]}"; do
+        pids=$(lsof -t -i :${port} 2>/dev/null || true)
+        if [ -n "$pids" ]; then
+            echo "  - 杀掉端口 ${port} 占用进程 (PIDs: $pids)"
+            echo "$pids" | xargs kill -9 2>/dev/null || true
+        fi
+    done
+    sleep 3
+    xpu-smi -r -i "0,1,2,3,4,5,6,7"
+    echo "  清理完成"
+}
+
+#============================================================
+# 第二部分：集中式 EP 服务配置
+#============================================================
+start_centralized_ep() {
+    echo "[2/3] 启动集中式 EP 服务..."
+
+    # XPU 设备配置
+    export XPU_VISIBLE_DEVICES="0,1,2,3"
+
+    # BKCL 通信配置
+    export BKCL_ENABLE_XDR=1
+    export BKCL_RDMA_NICS=eth1,eth1,eth2,eth2,eth3,eth3,eth4,eth4
+    export BKCL_TRACE_TOPO=1
+    export BKCL_PCIE_RING=1
+    export BKCL_RDMA_VERBS=1
+
+    # 共享内存配置
+    export XSHMEM_MODE=1
+    export XSHMEM_QP_NUM_PER_RANK=32
+
+    # 启动 Router
+    local router_port=8188
+    nohup python -m fastdeploy.router.launch \
+        --port ${router_port} \
+        2>&1 >./log_router/nohup &
+    sleep 1
+
+    # 启动 API Server
+    python -m fastdeploy.entrypoints.openai.multi_api_server \
+        --ports 8189 \
+        --num-servers 1 \
+        --metrics-ports 8199 \
+        --args \
+        --model ${MODEL_PATH} \
+        --engine-worker-queue-port 8209 \
+        --max-model-len 32768 \
+        --max-num-seqs 64 \
+        --data-parallel-size 1 \
+        --tensor-parallel-size 4 \
+        --enable-expert-parallel \
+        --quantization wint4 \
+        --enable-prefix-caching \
+        --router 0.0.0.0:${router_port} \
+        --graph-optimization-config '{"use_cudagraph":true}' \
+        --disable-sequence-parallel-moe
+
+    echo "  集中式 EP 服务启动完成"
+}
+
+#============================================================
+# 第三部分：PD 分离服务配置
+#============================================================
+start_pd_disaggregated() {
+    echo "[3/3] 启动 PD 分离服务..."
+
+    # KVCache 传输配置
+    export KVCACHE_GDRCOPY_FLUSH_ENABLE=1
+    export CUDA_ENABLE_P2P_NO_UVA=1
+    export KVCACHE_RDMA_NICS=mlx5_1,mlx5_1,mlx5_2,mlx5_2,mlx5_3,mlx5_3,mlx5_4,mlx5_4
+
+    # BKCL 通信配置
+    export BKCL_ENABLE_XDR=1
+    export BKCL_RDMA_NICS=eth1,eth1,eth2,eth2
+    export BKCL_TRACE_TOPO=1
+    export BKCL_PCIE_RING=1
+    export BKCL_RDMA_VERBS=1
+
+    # 共享内存配置
+    export XSHMEM_MODE=1
+    export XSHMEM_QP_NUM_PER_RANK=32
+
+    # Router 配置
+    local router_port=8188
+    local prefill_port=8199
+    local decode_port=8209
+
+    # 启动 Router (splitwise 模式)
+    python -m fastdeploy.router.launch \
+        --port ${router_port} \
+        --splitwise \
+        2>&1 >./log_router/nohup &
+    sleep 1
+
+    # # 启动 Prefill 服务 (GPU 0-3)
+    export XPU_VISIBLE_DEVICES="0,1,2,3"
+
+    nohup python -m fastdeploy.entrypoints.openai.multi_api_server \
+        --port ${prefill_port} \
+        --num-servers 1 \
+        --args \
+        --model ${MODEL_PATH} \
+        --tensor-parallel-size 4 \
+        --data-parallel-size 1 \
+        --max-model-len 32768 \
+        --max-num-seqs 64 \
+        --quantization wint4 \
+        --splitwise-role prefill \
+        --cache-transfer-protocol rdma \
+        --enable-expert-parallel \
+        --graph-optimization-config '{"use_cudagraph":false}' \
+        --router 0.0.0.0:${router_port} \
+        --disable-sequence-parallel-moe \
+        2>&1 >./log_router/nohup_prefill &
+
+    # echo "  Prefill 服务启动中，等待 10 秒..."
+    # sleep 10
+
+    # 启动 Decode 服务 (GPU 4-7)
+    export XPU_VISIBLE_DEVICES="4,5,6,7"
+    nohup python -m fastdeploy.entrypoints.openai.multi_api_server \
+        --port ${decode_port} \
+        --num-servers 1 \
+        --args \
+        --model ${MODEL_PATH} \
+        --tensor-parallel-size 4 \
+        --data-parallel-size 1 \
+        --max-model-len 32768 \
+        --max-num-seqs 64 \
+        --quantization wint4 \
+        --splitwise-role decode \
+        --cache-transfer-protocol rdma \
+        --enable-expert-parallel \
+        --graph-optimization-config '{"use_cudagraph":true}' \
+        --router 0.0.0.0:${router_port} \
+        --disable-sequence-parallel-moe \
+        2>&1 >./log_router/nohup_decode &
+
+    echo "  PD 分离服务启动完成"
+    echo "  - Router:  port ${router_port}"
+    echo "  - Prefill: port ${prefill_port}"
+    echo "  - Decode:  port ${decode_port}"
+}
+
+#============================================================
+# 主入口：选择启动模式
+#============================================================
+# 设置启动模式: "centralized" 或 "pd_disaggregated"
+MODE="${1:-pd_disaggregated}"
+
+echo "============================================"
+echo "FastDeploy 服务启动脚本"
+echo "模式: ${MODE}"
+echo "============================================"
+
+# 执行清理
+cleanup
+
+# 根据模式启动服务
+case "${MODE}" in
+    centralized)
+        start_centralized_ep
+        ;;
+    pd_disaggregated)
+        start_pd_disaggregated
+        ;;
+    *)
+        echo "错误: 未知模式 '${MODE}'"
+        echo "用法: $0 [centralized|pd_disaggregated]"
+        exit 1
+        ;;
+esac
+
+echo "============================================"
+echo "服务启动完成！"
+echo "============================================"