[Bug]: UCM reports errors in multi-concurrency scenarios with long sequences (120K).

Your current environment

version:
ucm0.5.0rc1 
vllm-ascend 0.18.0rc1
GLM-4.7-w8a8-with-float-mtp

🐛 Describe the bug

1.vllm run：

node0：

#!/bin/sh

# this obtained through ifconfig
# nic_name is the network interface name corresponding to local_ip of the current node
nic_name="xxxx"
local_ip="xxxx"
nic_name="bond0"

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 Eco-Tech/GLM-4.7-W8A8-floatmtp \
  --host 0.0.0.0 \
  --port 8004 \
  --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 glm47 \
  --block_size  128  \
  --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":"Eco-Tech/GLM-4.7-W8A8-floatmtp", "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="xxxx"
local_ip="xxxx"
node0_ip="xxxx" # same as the local_IP address in node 0
nic_name="bond0"

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 Eco-Tech/GLM-4.7-W8A8-floatmtp \
  --host 0.0.0.0 \
  --port 8004 \
  --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 glm47 \
  --block_size  128  \
  --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":"Eco-Tech/GLM-4.7-W8A8-floatmtp", "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}}'

2.ucm config:

ucm_connectors:
  - ucm_connector_name: "UcmPipelineStore"
    ucm_connector_config:
      store_pipeline: "Cache|Posix"
      storage_backends: "/mnt/ucm"
      io_direct: false
      timeout_ms：60000

# When you use UcmNfsStore, you should set enable_event_sync to false.
enable_event_sync: true

3.question：

It runs fine under normal conditions, but timeouts occur when concurrency increases to 10+ with long sequences (120K)：

when add：
use_layerwise: true
the error is ：

the speed of disk read and write：

Start by reproducing the timeout with the provided two-node vLLM commands, UCMConnector configuration, concurrency of 10 or more, and 120K-token sequences. Compare the UcmPipelineStore behavior with and without use_layerwise: true, focusing on the timeout and storage I/O evidence shown in the report. Done means the multi-concurrency workload completes without UCM errors or timeouts.