[Bug]: ucm0.5.0rc1对于双节点的共享存储报错

[ppzze]

Your current environment

Glm4.7 vllm-ascend 0.17.0rc1 ucm 0.5.0rc1 910b4双节点 单节点8×64

🐛 Describe the bug

用ucm的example两种模式都试过了，还是报错，现在两个节点用的是共享存储。详细参数请见上个usage

[sumingZero]

看下你两个节点起服务的命令呢，你开PP并行了吗

[ppzze]

您好 下面是我的启动命令  node0： #!/bin/sh # this obtained through ifconfig # nic_name is the network interface name corresponding to local_ip of the current node nic_name="bond0" local_ip="27.196.186.244" export HCCL_IF_IP=$local_ip export GLOO_SOCKET_IFNAME=$nic_name export TP_SOCKET_IFNAME=$nic_name export HCCL_SOCKET_IFNAME=$nic_name export HCCL_BUFFSIZE=512 export OMP_PROC_BIND=false export OMP_NUM_THREADS=1 export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True export HCCL_OP_EXPANSION_MODE=AIV export VLLM_ASCEND_BALANCE_SCHEDULING=1 export VLLM_ASCEND_ENABLE_TOPK_OPTIMIZE=1 export VLLM_ASCEND_ENABLE_FLASHCOMM1=1 export VLLM_ASCEND_ENABLE_FUSED_MC2=1 vllm serve /data/models/glm4.7-w8a8 \   --host 0.0.0.0 \   --port 18118  \   --data-parallel-size 2 \   --data-parallel-size-local 1 \   --data-parallel-start-rank 0 \   --data-parallel-address $local_ip \   --data-parallel-rpc-port 13389 \   --tensor-parallel-size 8 \   --enable-expert-parallel \   --seed 1024 \   --max-model-len 140000 \   --max-num-batched-tokens 8192 \   --max-num-seqs 16 \   --async-scheduling \   --quantization ascend \   --trust-remote-code \   --gpu-memory-utilization 0.9 \   --enable-auto-tool-choice \   --reasoning-parser glm45 \   --tool-call-parser glm47 \   --served-model-name GLM-4.7-w8a8 \   --no-enable-prefix-caching  \   --kv-transfer-config '{"kv_connector": "UCMConnector", "kv_connector_module_path": "ucm.integration.vllm.ucm_connector", "kv_role": "kv_both", "kv_connector_extra_config": {"UCM_CONFIG_FILE": "/workspace/unified-cache-management/examples/ucm_config_example.yaml"}}'  \   --speculative-config '{"num_speculative_tokens": 3, "model":"/data/models/glm4.7-w8a8", "method":"mtp"}' \   --compilation-config '{"cudagraph_capture_sizes": [1,2,4,8,16,32,64,128,256,512], "cudagraph_mode": "FULL_DECODE_ONLY"}' \   --additional-config '{"enable_shared_expert_dp": true, "ascend_fusion_config": {"fusion_ops_gmmswigluquant": false}}' node1： #!/bin/sh # this obtained through ifconfig # nic_name is the network interface name corresponding to local_ip of the current node nic_name="bond0" local_ip="27.196.186.239" node0_ip="27.196.186.244" # same as the local_IP address in node 0 export HCCL_IF_IP=$local_ip export GLOO_SOCKET_IFNAME=$nic_name export TP_SOCKET_IFNAME=$nic_name export HCCL_SOCKET_IFNAME=$nic_name export HCCL_BUFFSIZE=512 export OMP_PROC_BIND=false export OMP_NUM_THREADS=1 export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True export HCCL_OP_EXPANSION_MODE=AIV export VLLM_ASCEND_BALANCE_SCHEDULING=1 export VLLM_ASCEND_ENABLE_TOPK_OPTIMIZE=1 export VLLM_ASCEND_ENABLE_FLASHCOMM1=1 export VLLM_ASCEND_ENABLE_FUSED_MC2=1 vllm serve /data/models/glm4.7-w8a8 \   --host 0.0.0.0 \   --port 18118 \   --headless \   --data-parallel-size 2 \   --data-parallel-size-local 1 \   --data-parallel-start-rank 1 \   --data-parallel-address $node0_ip \   --data-parallel-rpc-port 13389 \   --tensor-parallel-size 8 \   --enable-expert-parallel \   --seed 1024 \   --max-model-len 140000 \   --max-num-batched-tokens 8192 \   --max-num-seqs 16 \   --async-scheduling \   --quantization ascend \   --trust-remote-code \   --gpu-memory-utilization 0.9 \   --enable-auto-tool-choice \   --reasoning-parser glm45 \   --tool-call-parser glm47 \   --served-model-name GLM-4.7-w8a8 \   --no-enable-prefix-caching  \   --kv-transfer-config '{"kv_connector": "UCMConnector", "kv_connector_module_path": "ucm.integration.vllm.ucm_connector", "kv_role": "kv_both", "kv_connector_extra_config": {"UCM_CONFIG_FILE": "/workspace/unified-cache-management/examples/ucm_config_example.yaml"}}' \   --speculative-config '{"num_speculative_tokens": 3, "model":"/data/models/glm4.7-w8a8", "method":"mtp"}' \   --compilation-config '{"cudagraph_capture_sizes": [1,2,4,8,16,32,64,128,256,512], "cudagraph_mode": "FULL_DECODE_ONLY"}' \   --additional-config '{"enable_shared_expert_dp": true, "ascend_fusion_config": {"fusion_ops_gmmswigluquant": false}}'

…

------------------ 原始邮件 ------------------ 发件人: ***@***.***>; 发送时间: 2026年4月3日(星期五) 下午5:30 收件人: ***@***.***>; 抄送: ***@***.***>; ***@***.***>; 主题: Re: [ModelEngine-Group/unified-cache-management] [Bug]: ucm0.5.0rc1对于双节点的共享存储报错 (Issue #893) sumingZero left a comment (ModelEngine-Group/unified-cache-management#893) 看下你两个节点起服务的命令呢，你开PP并行了吗 — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>