[Serve] Support xPyD Disaggregation in multiple nodes

examples/qwen/conf/config_qwen2.5_7b_disagg_xpyd.yaml

@@ -0,0 +1,36 @@
+defaults:
+- _self_
+- serve: serve_qwen2.5_7b
+
+experiment:
+  exp_name: qwen2.5_7b
+  exp_dir: outputs/${experiment.exp_name}
+  task:
+    type: serve
+  deploy:
+    port: 10001
+    use_fs_serve: false
+    prefill_decode_disaggregation: true
+    prefill_num: 2
+    prefill_address: x.x.x.x # optional, default "auto"
+    decode_num: 2
+    decode_address: x.x.x.x # optional, default "auto"
+  runner:
+    hostfile: examples/qwen/conf/hostfile.txt
+    docker: fr-v2
+  envs:
+    CUDA_DEVICE_MAX_CONNECTIONS: 1
+    VLLM_USE_V1: 0
+    FLAGCX_SOCKET_IFNAME: bond0
+    FLAGCX_PATH: /path/to/FlagCX/
+    FLAGCX_DEBUG: TRACE
+    FLAGCX_DEBUG_SUBSYS: ALL
+    USE_FLAGCX: true
+  cmds:
+    before_start: source /root/miniconda3/bin/activate flagscale-inference
+
+action: run
+
+hydra:
+  run:
+    dir: ${experiment.exp_dir}/hydra

examples/qwen/conf/hostfile.txt

@@ -0,0 +1,5 @@
+# ip slots type=xxx[optional]
+# master node
+x.x.x.x slots=8 type=gpu
+# worker nodes
+x.x.x.x slots=8 type=gpu

examples/qwen/conf/serve/serve_qwen2.5_7b.yaml

@@ -2,6 +2,7 @@
   engine: vllm
   engine_args:
     model: /models/Qwen2.5-7B-Instruct
+    host: 0.0.0.0
     tensor_parallel_size: 1
     pipeline_parallel_size: 1
     gpu_memory_utilization: 0.9