[GH-ISSUE #397] vllm+mooncake store,pd分离多rdma卡场景bench卡住 #136

Closed
opened 2026-10-29 21:09:58 +00:00 by bsakaguchi · 10 comments
Owner

Originally created by @qyzhaoxun on GitHub (May 24, 2025).
Original GitHub issue: https://github.com/kvcache-ai/Mooncake/issues/397

诉求

vllm+mooncake-store怎么在有多张rdma卡的机器上跑pd分离

问题描述

两台机器,每台机器8张gpu和rdma卡。按照教程部署了pd分离服务,bench时会一直卡住

prefill日志报错

INFO 05-23 06:50:39 [engine.py:310] Added request cmpl-ca60681e919c4882a450a6d5c5c23538-0.
I0523 06:50:40.052094 223 transfer_metadata_dump.cpp:51] Failed to get segment descriptor for segment 10.32.5.183:12728 address 0x7f3cf6020000--0x7f3cf6020094
I0523 06:50:40.052116 223 transfer_metadata_dump.cpp:63] -----------------------------------------------------------
I0523 06:50:40.052117 223 transfer_metadata_dump.cpp:64] TransferMetadata::dumpMetadataContent
I0523 06:50:40.052119 223 transfer_metadata_dump.cpp:65] -----------------------------------------------------------
I0523 06:50:40.052120 223 transfer_metadata_dump.cpp:66] === Cached Segment Descriptors ===
I0523 06:50:40.052120 223 transfer_metadata_dump.cpp:72] segment id: 1, ref object 0x2e464850
I0523 06:50:40.052124 223 transfer_metadata_dump.cpp:20] segment name: 10.32.5.183:12728
I0523 06:50:40.052125 223 transfer_metadata_dump.cpp:21] protocol: rdma
I0523 06:50:40.052126 223 transfer_metadata_dump.cpp:22] topology: {
"cpu:0" :
[
[
"mlx5_bond_0",
"mlx5_bond_1",
"mlx5_bond_2",
"mlx5_bond_3",
"mlx5_bond_4",
"mlx5_bond_5",
"mlx5_bond_6",
"mlx5_bond_7"
],
[]
]
}
I0523 06:50:40.052135 223 transfer_metadata_dump.cpp:23] devices:
I0523 06:50:40.052136 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_0, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:8a
I0523 06:50:40.052137 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_1, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:b2
I0523 06:50:40.052138 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_2, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:a6
I0523 06:50:40.052139 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_3, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:8e
I0523 06:50:40.052140 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_4, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:da
I0523 06:50:40.052141 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_5, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:d6
I0523 06:50:40.052141 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_6, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:de
I0523 06:50:40.052142 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_7, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:ba
I0523 06:50:40.052143 223 transfer_metadata_dump.cpp:28] buffers:
I0523 06:50:40.052145 223 transfer_metadata_dump.cpp:30] buffer type cpu:0, address 0x7f9637000000--0x7f9677000000
I0523 06:50:40.052145 223 transfer_metadata_dump.cpp:30] buffer type cpu:0, address 0x7f956e000000--0x7f9636000000
I0523 06:50:40.052151 223 transfer_metadata_dump.cpp:34] nvmeof buffers: 0 items
I0523 06:50:40.052153 223 transfer_metadata_dump.cpp:35] timestamp: 2025-05-23 06:36:16.122743
I0523 06:50:40.052155 223 transfer_metadata_dump.cpp:72] segment id: 0, ref object 0x14549f30
"cpu:0" :
[
[
"mlx5_bond_0",
"mlx5_bond_1",
"mlx5_bond_2",
"mlx5_bond_3",
"mlx5_bond_4",
"mlx5_bond_5",
"mlx5_bond_6",
"mlx5_bond_7"
],
[]
]
}
I0523 06:50:40.052165 223 transfer_metadata_dump.cpp:23] devices:
I0523 06:50:40.052166 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_0, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:82
I0523 06:50:40.052168 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_1, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:a2
I0523 06:50:40.052170 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_2, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:9e
I0523 06:50:40.052173 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_3, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:96
I0523 06:50:40.052175 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_4, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:c6
I0523 06:50:40.052178 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_5, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:be
I0523 06:50:40.052181 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_6, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:ce
I0523 06:50:40.052181 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_7, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:b6
I0523 06:50:40.052182 223 transfer_metadata_dump.cpp:28] buffers:
I0523 06:50:40.052184 223 transfer_metadata_dump.cpp:30] buffer type cpu:0, address 0x7f4637000000--0x7f4677000000
I0523 06:50:40.052186 223 transfer_metadata_dump.cpp:30] buffer type cpu:0, address 0x7f456e000000--0x7f4636000000
I0523 06:50:40.052188 223 transfer_metadata_dump.cpp:34] nvmeof buffers: 0 items
I0523 06:50:40.052191 223 transfer_metadata_dump.cpp:35] timestamp:
I0523 06:50:40.052192 223 transfer_metadata_dump.cpp:77] === Local RPC Route ===
I0523 06:50:40.052193 223 transfer_metadata_dump.cpp:78] location: 10.32.5.203:15337
I0523 06:50:40.052196 223 transfer_metadata_dump.cpp:80] === Remote RPC Routes ===
E0523 06:50:40.054558 223 client.cpp:444] Transfer failed for task0
W0523 06:50:40.054589 223 client.cpp:450] Transfer incomplete, retrying... (attempt 1/3)
E0523 06:50:40.064651 223 client.cpp:444] Transfer failed for task0
W0523 06:50:40.064657 223 client.cpp:450] Transfer incomplete, retrying... (attempt 2/3)
E0523 06:50:40.074718 223 client.cpp:444] Transfer failed for task0
W0523 06:50:40.074728 223 client.cpp:450] Transfer incomplete, retrying... (attempt 3/3)
E0523 06:50:40.084792 223 client.cpp:460] transfer_incomplete max_attempts=3
E0523 06:50:40.085733 223 store_py.cpp:367] Put operation failed with error: TRANSFER_FAIL
E0523 06:50:40.087774 223 client.cpp:444] Transfer failed for task0
W0523 06:50:40.087782 223 client.cpp:450] Transfer incomplete, retrying... (attempt 1/3)
E0523 06:50:40.097843 223 client.cpp:444] Transfer failed for task0
W0523 06:50:40.097848 223 client.cpp:450] Transfer incomplete, retrying... (attempt 2/3)
E0523 06:50:40.107908 223 client.cpp:444] Transfer failed for task0
W0523 06:50:40.107914 223 client.cpp:450] Transfer incomplete, retrying... (attempt 3/3)
E0523 06:50:40.117973 223 client.cpp:460] transfer_incomplete max_attempts=3
E0523 06:50:40.118625 223 store_py.cpp:367] Put operation failed with error: TRANSFER_FAIL

decode日志报错

WARNING 05-23 06:39:54 [mooncake_store_connector.py:181] [rank0]: Failed to receive all KVs and hidden states, redo model forwarding.
(VllmWorkerProcess pid=217) WARNING 05-23 06:39:54 [mooncake_store_connector.py:181] [rank1]: Failed to receive all KVs and hidden states, redo model forwarding.
(VllmWorkerProcess pid=218) WARNING 05-23 06:39:54 [mooncake_store_connector.py:181] [rank2]: Failed to receive all KVs and hidden states, redo model forwarding.
(VllmWorkerProcess pid=219) WARNING 05-23 06:39:54 [mooncake_store_connector.py:181] [rank3]: Failed to receive all KVs and hidden states, redo model forwarding.
(VllmWorkerProcess pid=221) WARNING 05-23 06:39:54 [mooncake_store_connector.py:181] [rank5]: Failed to receive all KVs and hidden states, redo model forwarding.
(VllmWorkerProcess pid=223) WARNING 05-23 06:39:54 [mooncake_store_connector.py:181] [rank7]: Failed to receive all KVs and hidden states, redo model forwarding.
(VllmWorkerProcess pid=220) WARNING 05-23 06:39:54 [mooncake_store_connector.py:181] [rank4]: Failed to receive all KVs and hidden states, redo model forwarding.
(VllmWorkerProcess pid=222) WARNING 05-23 06:39:54 [mooncake_store_connector.py:181] [rank6]: Failed to receive all KVs and hidden states, redo model forwarding.

mooncake配置文件

prefill mooncake配置

mooncake.json: |
{
"local_hostname": "10.32.5.203",
"prefill_url": "10.32.5.203:13003",
"decode_url": "10.32.5.183:13003",
"master_server_address": "10.32.5.215:50051",
"metadata_server": "etcd-headless.vllm-mooncake.svc.cluster.local:2379",
"metadata_backend": "etcd",
"protocol": "rdma",
"device_name": "mlx5_bond_0,mlx5_bond_1,mlx5_bond_2,mlx5_bond_3,mlx5_bond_4,mlx5_bond_5,mlx5_bond_6,mlx5_bond_7"
}

decode mooncake配置

mooncake.json: |
{
"local_hostname": "10.32.5.183",
"prefill_url": "10.32.5.203:13003",
"decode_url": "10.32.5.183:13003",
"master_server_address": "10.32.5.215:50051",
"metadata_server": "etcd-headless.vllm-mooncake.svc.cluster.local:2379",
"metadata_backend": "etcd",
"protocol": "rdma",
"device_name": "mlx5_bond_0,mlx5_bond_1,mlx5_bond_2,mlx5_bond_3,mlx5_bond_4,mlx5_bond_5,mlx5_bond_6,mlx5_bond_7"
}

补充说明

  1. 同样的节点上使用一张gpu卡和rdma卡就很稳定,不会出现卡主的问题。
  2. 另外使用transfer_engine_bench测试发现target起来后会记住第一个客户端使用的rdma设备,虽然etcd中数据会变,但是客户端使用同一个rdma设备就可以测试通过,使用不同的设备测试就会报错,以下是具体失败日志
    E0524 12:53:38.269760 773 rdma_endpoint.cpp:149] Invalid argument: received packet mismatch, local.local_nic_path: 10.32.5.183@mlx5_bond_2, local.peer_nic_path: 10.32.5.203@mlx5_bond_2, peer.local_nic_path: , peer.peer_nic_path:
    E0524 12:53:38.269788 773 worker_pool.cpp:227] Worker: Cannot make connection for endpoint: 10.32.5.203@mlx5_bond_2, mark it inactive
    I0524 12:53:38.276801 779 transfer_engine_bench.cpp:215] FAILED
    W0524 12:53:38.276849 779 transport.h:148] detected slice leak: allocated 128 freed 0
    I0524 12:53:38.276801 776 transfer_engine_bench.cpp:215] FAILED
    W0524 12:53:38.276886 776 transport.h:148] detected slice leak: allocated 128 freed 0
Originally created by @qyzhaoxun on GitHub (May 24, 2025). Original GitHub issue: https://github.com/kvcache-ai/Mooncake/issues/397 ### 诉求 vllm+mooncake-store怎么在有多张rdma卡的机器上跑pd分离 ### 问题描述 两台机器,每台机器8张gpu和rdma卡。按照[教程](https://kvcache-ai.github.io/Mooncake/getting_started/examples/vllm-integration-v1.html)部署了pd分离服务,bench时会一直卡住 #### prefill日志报错 INFO 05-23 06:50:39 [engine.py:310] Added request cmpl-ca60681e919c4882a450a6d5c5c23538-0. I0523 06:50:40.052094 223 transfer_metadata_dump.cpp:51] Failed to get segment descriptor for segment 10.32.5.183:12728 address 0x7f3cf6020000--0x7f3cf6020094 I0523 06:50:40.052116 223 transfer_metadata_dump.cpp:63] ----------------------------------------------------------- I0523 06:50:40.052117 223 transfer_metadata_dump.cpp:64] TransferMetadata::dumpMetadataContent I0523 06:50:40.052119 223 transfer_metadata_dump.cpp:65] ----------------------------------------------------------- I0523 06:50:40.052120 223 transfer_metadata_dump.cpp:66] === Cached Segment Descriptors === I0523 06:50:40.052120 223 transfer_metadata_dump.cpp:72] segment id: 1, ref object 0x2e464850 I0523 06:50:40.052124 223 transfer_metadata_dump.cpp:20] segment name: 10.32.5.183:12728 I0523 06:50:40.052125 223 transfer_metadata_dump.cpp:21] protocol: rdma I0523 06:50:40.052126 223 transfer_metadata_dump.cpp:22] topology: { "cpu:0" : [ [ "mlx5_bond_0", "mlx5_bond_1", "mlx5_bond_2", "mlx5_bond_3", "mlx5_bond_4", "mlx5_bond_5", "mlx5_bond_6", "mlx5_bond_7" ], [] ] } I0523 06:50:40.052135 223 transfer_metadata_dump.cpp:23] devices: I0523 06:50:40.052136 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_0, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:8a I0523 06:50:40.052137 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_1, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:b2 I0523 06:50:40.052138 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_2, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:a6 I0523 06:50:40.052139 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_3, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:8e I0523 06:50:40.052140 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_4, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:da I0523 06:50:40.052141 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_5, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:d6 I0523 06:50:40.052141 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_6, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:de I0523 06:50:40.052142 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_7, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:ba I0523 06:50:40.052143 223 transfer_metadata_dump.cpp:28] buffers: I0523 06:50:40.052145 223 transfer_metadata_dump.cpp:30] buffer type cpu:0, address 0x7f9637000000--0x7f9677000000 I0523 06:50:40.052145 223 transfer_metadata_dump.cpp:30] buffer type cpu:0, address 0x7f956e000000--0x7f9636000000 I0523 06:50:40.052151 223 transfer_metadata_dump.cpp:34] nvmeof buffers: 0 items I0523 06:50:40.052153 223 transfer_metadata_dump.cpp:35] timestamp: 2025-05-23 06:36:16.122743 I0523 06:50:40.052155 223 transfer_metadata_dump.cpp:72] segment id: 0, ref object 0x14549f30 "cpu:0" : [ [ "mlx5_bond_0", "mlx5_bond_1", "mlx5_bond_2", "mlx5_bond_3", "mlx5_bond_4", "mlx5_bond_5", "mlx5_bond_6", "mlx5_bond_7" ], [] ] } I0523 06:50:40.052165 223 transfer_metadata_dump.cpp:23] devices: I0523 06:50:40.052166 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_0, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:82 I0523 06:50:40.052168 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_1, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:a2 I0523 06:50:40.052170 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_2, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:9e I0523 06:50:40.052173 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_3, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:96 I0523 06:50:40.052175 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_4, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:c6 I0523 06:50:40.052178 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_5, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:be I0523 06:50:40.052181 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_6, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:ce I0523 06:50:40.052181 223 transfer_metadata_dump.cpp:25] device name mlx5_bond_7, lid 0, 00:00:00:00:00:00:00:00:00:00:ff:ff:1d:c2:1e:b6 I0523 06:50:40.052182 223 transfer_metadata_dump.cpp:28] buffers: I0523 06:50:40.052184 223 transfer_metadata_dump.cpp:30] buffer type cpu:0, address 0x7f4637000000--0x7f4677000000 I0523 06:50:40.052186 223 transfer_metadata_dump.cpp:30] buffer type cpu:0, address 0x7f456e000000--0x7f4636000000 I0523 06:50:40.052188 223 transfer_metadata_dump.cpp:34] nvmeof buffers: 0 items I0523 06:50:40.052191 223 transfer_metadata_dump.cpp:35] timestamp: I0523 06:50:40.052192 223 transfer_metadata_dump.cpp:77] === Local RPC Route === I0523 06:50:40.052193 223 transfer_metadata_dump.cpp:78] location: 10.32.5.203:15337 I0523 06:50:40.052196 223 transfer_metadata_dump.cpp:80] === Remote RPC Routes === E0523 06:50:40.054558 223 client.cpp:444] Transfer failed for task0 W0523 06:50:40.054589 223 client.cpp:450] Transfer incomplete, retrying... (attempt 1/3) E0523 06:50:40.064651 223 client.cpp:444] Transfer failed for task0 W0523 06:50:40.064657 223 client.cpp:450] Transfer incomplete, retrying... (attempt 2/3) E0523 06:50:40.074718 223 client.cpp:444] Transfer failed for task0 W0523 06:50:40.074728 223 client.cpp:450] Transfer incomplete, retrying... (attempt 3/3) E0523 06:50:40.084792 223 client.cpp:460] transfer_incomplete max_attempts=3 E0523 06:50:40.085733 223 store_py.cpp:367] Put operation failed with error: TRANSFER_FAIL E0523 06:50:40.087774 223 client.cpp:444] Transfer failed for task0 W0523 06:50:40.087782 223 client.cpp:450] Transfer incomplete, retrying... (attempt 1/3) E0523 06:50:40.097843 223 client.cpp:444] Transfer failed for task0 W0523 06:50:40.097848 223 client.cpp:450] Transfer incomplete, retrying... (attempt 2/3) E0523 06:50:40.107908 223 client.cpp:444] Transfer failed for task0 W0523 06:50:40.107914 223 client.cpp:450] Transfer incomplete, retrying... (attempt 3/3) E0523 06:50:40.117973 223 client.cpp:460] transfer_incomplete max_attempts=3 E0523 06:50:40.118625 223 store_py.cpp:367] Put operation failed with error: TRANSFER_FAIL #### decode日志报错 WARNING 05-23 06:39:54 [mooncake_store_connector.py:181] [rank0]: Failed to receive all KVs and hidden states, redo model forwarding. (VllmWorkerProcess pid=217) WARNING 05-23 06:39:54 [mooncake_store_connector.py:181] [rank1]: Failed to receive all KVs and hidden states, redo model forwarding. (VllmWorkerProcess pid=218) WARNING 05-23 06:39:54 [mooncake_store_connector.py:181] [rank2]: Failed to receive all KVs and hidden states, redo model forwarding. (VllmWorkerProcess pid=219) WARNING 05-23 06:39:54 [mooncake_store_connector.py:181] [rank3]: Failed to receive all KVs and hidden states, redo model forwarding. (VllmWorkerProcess pid=221) WARNING 05-23 06:39:54 [mooncake_store_connector.py:181] [rank5]: Failed to receive all KVs and hidden states, redo model forwarding. (VllmWorkerProcess pid=223) WARNING 05-23 06:39:54 [mooncake_store_connector.py:181] [rank7]: Failed to receive all KVs and hidden states, redo model forwarding. (VllmWorkerProcess pid=220) WARNING 05-23 06:39:54 [mooncake_store_connector.py:181] [rank4]: Failed to receive all KVs and hidden states, redo model forwarding. (VllmWorkerProcess pid=222) WARNING 05-23 06:39:54 [mooncake_store_connector.py:181] [rank6]: Failed to receive all KVs and hidden states, redo model forwarding. ### mooncake配置文件 #### prefill mooncake配置 mooncake.json: | { "local_hostname": "10.32.5.203", "prefill_url": "10.32.5.203:13003", "decode_url": "10.32.5.183:13003", "master_server_address": "10.32.5.215:50051", "metadata_server": "etcd-headless.vllm-mooncake.svc.cluster.local:2379", "metadata_backend": "etcd", "protocol": "rdma", "device_name": "mlx5_bond_0,mlx5_bond_1,mlx5_bond_2,mlx5_bond_3,mlx5_bond_4,mlx5_bond_5,mlx5_bond_6,mlx5_bond_7" } #### decode mooncake配置 mooncake.json: | { "local_hostname": "10.32.5.183", "prefill_url": "10.32.5.203:13003", "decode_url": "10.32.5.183:13003", "master_server_address": "10.32.5.215:50051", "metadata_server": "etcd-headless.vllm-mooncake.svc.cluster.local:2379", "metadata_backend": "etcd", "protocol": "rdma", "device_name": "mlx5_bond_0,mlx5_bond_1,mlx5_bond_2,mlx5_bond_3,mlx5_bond_4,mlx5_bond_5,mlx5_bond_6,mlx5_bond_7" } ### 补充说明 1. 同样的节点上使用一张gpu卡和rdma卡就很稳定,不会出现卡主的问题。 2. 另外使用transfer_engine_bench测试发现target起来后会记住第一个客户端使用的rdma设备,虽然etcd中数据会变,但是客户端使用同一个rdma设备就可以测试通过,使用不同的设备测试就会报错,以下是具体失败日志 E0524 12:53:38.269760 773 rdma_endpoint.cpp:149] Invalid argument: received packet mismatch, local.local_nic_path: 10.32.5.183@mlx5_bond_2, local.peer_nic_path: 10.32.5.203@mlx5_bond_2, peer.local_nic_path: , peer.peer_nic_path: E0524 12:53:38.269788 773 worker_pool.cpp:227] Worker: Cannot make connection for endpoint: 10.32.5.203@mlx5_bond_2, mark it inactive I0524 12:53:38.276801 779 transfer_engine_bench.cpp:215] FAILED W0524 12:53:38.276849 779 transport.h:148] detected slice leak: allocated 128 freed 0 I0524 12:53:38.276801 776 transfer_engine_bench.cpp:215] FAILED W0524 12:53:38.276886 776 transport.h:148] detected slice leak: allocated 128 freed 0
bsakaguchi 2026-10-29 21:09:59 +00:00
Author
Owner

@jiafuzha commented on GitHub (May 26, 2025):

You can try to increase listening backlog size. Please check #388.

<!-- gh-comment-id:2908258774 --> @jiafuzha commented on GitHub (May 26, 2025): You can try to increase listening backlog size. Please check #388.
Author
Owner

@qyzhaoxun commented on GitHub (May 26, 2025):

@jiafuzha 非常感谢,那我应该怎么验证这个mr可以符合预期呢?我看vllm应该还没有适配0.3.2,vllm我是直接用的官方的镜像,没有自己编译vllm,是直接通过pip3 install --upgrade mooncake-transfer-engine==0.3.2就搞定了吗?另外这里有方式通过transfer_engine_bench工具验证吗?

<!-- gh-comment-id:2908420564 --> @qyzhaoxun commented on GitHub (May 26, 2025): @jiafuzha 非常感谢,那我应该怎么验证这个mr可以符合预期呢?我看vllm应该还没有适配0.3.2,vllm我是直接用的官方的镜像,没有自己编译vllm,是直接通过pip3 install --upgrade mooncake-transfer-engine==0.3.2就搞定了吗?另外这里有方式通过transfer_engine_bench工具验证吗?
Author
Owner

@xiaguan commented on GitHub (May 26, 2025):

你可以从388 pr的ci里下载whl包,比如,https://github.com/kvcache-ai/Mooncake/suites/39108435316/artifacts/3194443374

<!-- gh-comment-id:2908530331 --> @xiaguan commented on GitHub (May 26, 2025): 你可以从388 pr的ci里下载whl包,比如,https://github.com/kvcache-ai/Mooncake/suites/39108435316/artifacts/3194443374
Author
Owner

@jiafuzha commented on GitHub (May 26, 2025):

#388 里面有错误描述。如果你没有发现这些错误,应该就好了。当然也可能你的错误本来就是别的问题造成的。

<!-- gh-comment-id:2908536038 --> @jiafuzha commented on GitHub (May 26, 2025): #388 里面有错误描述。如果你没有发现这些错误,应该就好了。当然也可能你的错误本来就是别的问题造成的。
Author
Owner

@mugglew commented on GitHub (Jul 17, 2025):

@qyzhaoxun hi i met same problem, have you solve that?

<!-- gh-comment-id:3083379464 --> @mugglew commented on GitHub (Jul 17, 2025): @qyzhaoxun hi i met same problem, have you solve that?
Author
Owner

@mgw2168-1 commented on GitHub (Sep 5, 2025):

@mugglew 你好,我也遇到了这个问题,有啥解决办法吗?我mooncake v0.3.5

<!-- gh-comment-id:3256828670 --> @mgw2168-1 commented on GitHub (Sep 5, 2025): @mugglew 你好,我也遇到了这个问题,有啥解决办法吗?我mooncake v0.3.5
Author
Owner

@thqq479 commented on GitHub (Mar 3, 2026):

The same problem occurs in sglang+mooncake

<!-- gh-comment-id:3990731948 --> @thqq479 commented on GitHub (Mar 3, 2026): The same problem occurs in sglang+mooncake
Author
Owner

@thqq479 commented on GitHub (Mar 3, 2026):

I0303 12:07:47.329501 159911 transfer_metadata_dump.cpp:51] Failed to get segment descriptor for segment 10.92.241.56:15917 address 0x7f7454c94000--0x7f7454ca4000
@alogfans I encountered this error when deploying DeepSeek PD in separation mode, especially when the QPS value was high

<!-- gh-comment-id:3990735290 --> @thqq479 commented on GitHub (Mar 3, 2026): I0303 12:07:47.329501 159911 transfer_metadata_dump.cpp:51] Failed to get segment descriptor for segment 10.92.241.56:15917 address 0x7f7454c94000--0x7f7454ca4000 @alogfans I encountered this error when deploying DeepSeek PD in separation mode, especially when the QPS value was high
Author
Owner

@github-actions[bot] commented on GitHub (Jul 11, 2026):

This issue has had no activity for 90 days and will be closed in 7 days if there is no further activity. Please comment or react if it should stay open.

<!-- gh-comment-id:4942239868 --> @github-actions[bot] commented on GitHub (Jul 11, 2026): This issue has had no activity for 90 days and will be closed in 7 days if there is no further activity. Please comment or react if it should stay open.
Author
Owner

@github-actions[bot] commented on GitHub (Jul 19, 2026):

Closing due to 3 months of inactivity. If this is still relevant, please comment and we can reopen.

<!-- gh-comment-id:5014198941 --> @github-actions[bot] commented on GitHub (Jul 19, 2026): Closing due to 3 months of inactivity. If this is still relevant, please comment and we can reopen.
Sign in to join this conversation.
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
kvcache-ai/Mooncake#136
No description provided.