EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

[Bugfix] multi_node_pd_disaggregation_mooncake.md update (#3400)

Browse Source 
### What this PR does / why we need it?
multi_node_pd_disaggregation_mooncake.md update. Fix issues encountered
during service startup.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
By ci


- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

Signed-off-by: wangxiaoteng@huawei.com <wangxiaoteng@huawei.com>
This commit is contained in:
wangxiaoteng888
2025-10-14 09:29:35 +08:00
committed by GitHub
parent 49b850270f
commit 19b85ef1bc
1 changed files with 71 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files

108
docs/source/tutorials/multi_node_pd_disaggregation_mooncake.md
View File

@@ -58,10 +58,11 @@ git clone -b pooling_async_memecpy_v1 https://github.com/AscendTransport/Mooncak
Update and install Python
```shell
apt-get update
apt-get install python3
```
Install the relevant dependencies.
Install the relevant dependencies. The installation of Go is not required.
```shell
cd Mooncake
@@ -89,12 +90,11 @@ make -j
make install
cp mooncake-transfer-engine/src/transport/ascend_transport/hccl_transport/ascend_transport_c/libascend_transport_mem.so /usr/local/Ascend/ascend-toolkit/latest/python/site-packages/
cp mooncake-transfer-engine/src/libtransfer_engine.so /usr/local/Ascend/ascend-toolkit/latest/python/site-packages/
export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/python/site-packages:$LD_LIBRARY_PATH
```
## Prefiller / Decoder Deployment
We can run the following scripts to launch a server on the prefiller/decoder node respectively.
We can run the following scripts to launch a server on the prefiller/decoder node respectively. Please note that each P/D node will occupy ports ranging from kv_port to kv_port + num_chips to initialize socket listeners. To avoid any issues, port conflicts should be prevented. Additionally, ensure that each node's engine_id is uniquely assigned to avoid conflicts.
### layerwise
@@ -118,6 +118,8 @@ export ASCEND_AGGREGATE_ENABLE=1
export ASCEND_TRANSPORT_PRINT=0
export ACL_OP_INIT_MODE=1
export ASCEND_A3_ENABLE=1
export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/python/site-packages:$LD_LIBRARY_PATH
vllm serve /model/Qwen3-235B-A22B-W8A8 \
--host 0.0.0.0 \
--port 8004 \
@@ -133,7 +135,7 @@ vllm serve /model/Qwen3-235B-A22B-W8A8 \
--distributed-executor-backend mp \
--served-model-name qwen3-moe \
--max-model-len 32768 \
--max-num-batched-tokens 4096 \
--max-num-batched-tokens 32768 \
--trust-remote-code \
--no-enable-prefix-caching \
--gpu-memory-utilization 0.9 \
@@ -144,7 +146,6 @@ vllm serve /model/Qwen3-235B-A22B-W8A8 \
"engine_id": "0",
"kv_connector_module_path": "vllm_ascend.distributed.mooncake_layerwise_connector",
"kv_connector_extra_config": {
"use_ascend_direct": true,
"prefill": {
"dp_size": 2,
"tp_size": 8
@@ -177,6 +178,8 @@ export ASCEND_AGGREGATE_ENABLE=1
export ASCEND_TRANSPORT_PRINT=0
export ACL_OP_INIT_MODE=1
export ASCEND_A3_ENABLE=1
export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/python/site-packages:$LD_LIBRARY_PATH
vllm serve /model/Qwen3-235B-A22B-W8A8 \
--host 0.0.0.0 \
--port 8004 \
@@ -192,7 +195,7 @@ vllm serve /model/Qwen3-235B-A22B-W8A8 \
--distributed-executor-backend mp \
--served-model-name qwen3-moe \
--max-model-len 32768 \
--max-num-batched-tokens 4096 \
--max-num-batched-tokens 32768 \
--trust-remote-code \
--no-enable-prefix-caching \
--gpu-memory-utilization 0.9 \
@@ -217,7 +220,7 @@ vllm serve /model/Qwen3-235B-A22B-W8A8 \
::::
::::{tab-item} Decoder node 1
::::{tab-item} Decoder node 1 (master Node)
```shell
unset ftp_proxy
@@ -235,6 +238,8 @@ export ASCEND_AGGREGATE_ENABLE=1
export ASCEND_TRANSPORT_PRINT=0
export ACL_OP_INIT_MODE=1
export ASCEND_A3_ENABLE=1
export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/python/site-packages:$LD_LIBRARY_PATH
vllm serve /model/Qwen3-235B-A22B-W8A8 \
--host 0.0.0.0 \
--port 8004 \
@@ -276,7 +281,7 @@ vllm serve /model/Qwen3-235B-A22B-W8A8 \
::::
::::{tab-item} Decoder node 2
::::{tab-item} Decoder node 2 (primary Node)
```shell
unset ftp_proxy
@@ -294,6 +299,8 @@ export ASCEND_AGGREGATE_ENABLE=1
export ASCEND_TRANSPORT_PRINT=0
export ACL_OP_INIT_MODE=1
export ASCEND_A3_ENABLE=1
export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/python/site-packages:$LD_LIBRARY_PATH
vllm serve /model/Qwen3-235B-A22B-W8A8 \
--host 0.0.0.0 \
--port 8004 \
@@ -318,8 +325,8 @@ vllm serve /model/Qwen3-235B-A22B-W8A8 \
--kv-transfer-config \
'{"kv_connector": "MooncakeLayerwiseConnector",
"kv_role": "kv_consumer",
"kv_port": "30300",
"engine_id": "3",
"kv_port": "30200",
"engine_id": "2",
"kv_connector_module_path": "vllm_ascend.distributed.mooncake_layerwise_connector",
"kv_connector_extra_config": {
"prefill": {
@@ -338,7 +345,7 @@ vllm serve /model/Qwen3-235B-A22B-W8A8 \
:::::
### mooncake
### non-layerwise
:::::{tab-set}
@@ -356,6 +363,8 @@ export VLLM_USE_V1=1
export HCCL_BUFFSIZE=1024
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=10
export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/python/site-packages:$LD_LIBRARY_PATH
vllm serve /model/Qwen3-235B-A22B-W8A8 \
--host 0.0.0.0 \
--port 8004 \
@@ -371,7 +380,7 @@ vllm serve /model/Qwen3-235B-A22B-W8A8 \
--distributed-executor-backend mp \
--served-model-name qwen3-moe \
--max-model-len 32768 \
--max-num-batched-tokens 4096 \
--max-num-batched-tokens 32768 \
--trust-remote-code \
--no-enable-prefix-caching \
--gpu-memory-utilization 0.9 \
@@ -382,14 +391,13 @@ vllm serve /model/Qwen3-235B-A22B-W8A8 \
"engine_id": "0",
"kv_connector_module_path": "vllm_ascend.distributed.mooncake_connector",
"kv_connector_extra_config": {
"use_ascend_direct": true,
"prefill": {
"dp_size": 2,
"tp_size": 8
},
"decode": {
"dp_size": 32,
"tp_size": 1
"dp_size": 4,
"tp_size": 8
}
}
}'
@@ -411,7 +419,9 @@ export VLLM_USE_V1=1
export HCCL_BUFFSIZE=1024
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=10
vllm serve /mnt/weight/Qwen3-235B-A22B-W8A8 \
export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/python/site-packages:$LD_LIBRARY_PATH
vllm serve /model/Qwen3-235B-A22B-W8A8 \
--host 0.0.0.0 \
--port 8004 \
--api-server-count 2 \
@@ -426,7 +436,7 @@ vllm serve /mnt/weight/Qwen3-235B-A22B-W8A8 \
--distributed-executor-backend mp \
--served-model-name qwen3-moe \
--max-model-len 32768 \
--max-num-batched-tokens 4096 \
--max-num-batched-tokens 32768 \
--trust-remote-code \
--no-enable-prefix-caching \
--gpu-memory-utilization 0.9 \
@@ -442,8 +452,8 @@ vllm serve /mnt/weight/Qwen3-235B-A22B-W8A8 \
"tp_size": 8
},
"decode": {
"dp_size": 32,
"tp_size": 1
"dp_size": 4,
"tp_size": 8
}
}
}'
@@ -451,7 +461,7 @@ vllm serve /mnt/weight/Qwen3-235B-A22B-W8A8 \
::::
::::{tab-item} Decoder node 1
::::{tab-item} Decoder node 1 (master Node)
```shell
unset ftp_proxy
@@ -465,16 +475,17 @@ export VLLM_USE_V1=1
export HCCL_BUFFSIZE=2048
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=10
export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/python/site-packages:$LD_LIBRARY_PATH
vllm serve /model/Qwen3-235B-A22B-W8A8 \
--host 0.0.0.0 \
--port 8004 \
--api-server-count 4 \
--data-parallel-size 32 \
--data-parallel-size-local 16 \
--data-parallel-size 4 \
--data-parallel-size-local 2 \
--data-parallel-address 192.0.0.3 \
--data-parallel-rpc-port 5964 \
--tensor-parallel-size 1 \
--tensor-parallel-size 8 \
--enable-expert-parallel \
--seed 1024 \
--distributed-executor-backend mp \
@@ -498,8 +509,8 @@ vllm serve /model/Qwen3-235B-A22B-W8A8 \
"tp_size": 8
},
"decode": {
"dp_size": 32,
"tp_size": 1
"dp_size": 4,
"tp_size": 8
}
}
}'
@@ -507,7 +518,7 @@ vllm serve /model/Qwen3-235B-A22B-W8A8 \
::::
::::{tab-item} Decoder node 2
::::{tab-item} Decoder node 2 (primary Node)
```shell
unset ftp_proxy
@@ -521,17 +532,18 @@ export VLLM_USE_V1=1
export HCCL_BUFFSIZE=2048
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=10
export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/python/site-packages:$LD_LIBRARY_PATH
vllm serve /model/Qwen3-235B-A22B-W8A8 \
--host 0.0.0.0 \
--port 8004 \
--headless \
--data-parallel-size 32 \
--data-parallel-size-local 16 \
--data-parallel-start-rank 16 \
--data-parallel-size 4 \
--data-parallel-size-local 2 \
--data-parallel-start-rank 2 \
--data-parallel-address 192.0.0.3 \
--data-parallel-rpc-port 5964 \
--tensor-parallel-size 1 \
--tensor-parallel-size 8 \
--enable-expert-parallel \
--seed 1024 \
--distributed-executor-backend mp \
@@ -546,8 +558,8 @@ vllm serve /model/Qwen3-235B-A22B-W8A8 \
--kv-transfer-config \
'{"kv_connector": "MooncakeConnector",
"kv_role": "kv_consumer",
"kv_port": "30300",
"engine_id": "3",
"kv_port": "30200",
"engine_id": "2",
"kv_connector_module_path": "vllm_ascend.distributed.mooncake_connector",
"kv_connector_extra_config": {
"prefill": {
@@ -555,8 +567,8 @@ vllm serve /model/Qwen3-235B-A22B-W8A8 \
"tp_size": 8
},
"decode": {
"dp_size": 32,
"tp_size": 1
"dp_size": 4,
"tp_size": 8
}
}
}'
@@ -568,7 +580,11 @@ vllm serve /model/Qwen3-235B-A22B-W8A8 \
## Example proxy for Deployment
Run a proxy server on the same node with prefiller service instance. You can get the proxy program in the repository's examples: [load\_balance\_proxy\_layerwise\_server\_example.py](https://github.com/vllm-project/vllm-ascend/blob/main/examples/disaggregated_prefill_v1/load_balance_proxy_server_example.py)
Run a proxy server on the same node with prefiller service instance. You can get the proxy program in the repository's examples: [load\_balance\_proxy\_layerwise\_server\_example.py](https://github.com/vllm-project/vllm-ascend/blob/main/examples/disaggregated_prefill_v1/load_balance_proxy_layerwise_server_example.py) or [load\_balance\_proxy\_server\_example.py](https://github.com/vllm-project/vllm-ascend/blob/main/examples/disaggregated_prefill_v1/load_balance_proxy_server_example.py)
:::::{tab-set}
::::{tab-item} layerwise
```shell
python load_balance_proxy_layerwise_server_example.py \
@@ -576,10 +592,28 @@ python load_balance_proxy_layerwise_server_example.py \
--port 8080 \
--prefiller-hosts 192.0.0.1 192.0.0.2\
--prefiller-port 8004 8004\
--decoder-hosts 192.0.0.3 192.0.0.4\
--decoder-ports 8004 8004
--decoder-hosts 192.0.0.3\
--decoder-ports 8004
```
::::
::::{tab-item} non-layerwise
```shell
python load_balance_proxy_server_example.py \
--host 192.0.0.1 \
--port 8080 \
--prefiller-hosts 192.0.0.1 192.0.0.2\
--prefiller-port 8004 8004\
--decoder-hosts 192.0.0.3\
--decoder-ports 8004
```
::::
:::::
## Verification
Check service health using the proxy server endpoint.