Skip to content

Commit 996d350

Browse files
liziyu179liziyu179
committed
unregister mooncake layerwise connector
Signed-off-by: liziyu <[email protected]>
1 parent 28a1529 commit 996d350

File tree

2 files changed

+1
-261
lines changed

2 files changed

+1
-261
lines changed

‎docs/source/tutorials/multi_node_pd_disaggregation_mooncake.md‎

Lines changed: 1 addition & 256 deletions
Original file line numberDiff line numberDiff line change
@@ -102,239 +102,6 @@ make install
102102

103103
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.
104104

105-
### Layerwise
106-
107-
:::::{tab-set}
108-
109-
::::{tab-item} Prefiller node 1
110-
111-
```shell
112-
unset ftp_proxy
113-
unset https_proxy
114-
unset http_proxy
115-
export HCCL_IF_IP=192.0.0.1
116-
export GLOO_SOCKET_IFNAME="eth0" # network card name
117-
export TP_SOCKET_IFNAME="eth0"
118-
export HCCL_SOCKET_IFNAME="eth0"
119-
export VLLM_USE_V1=1
120-
export HCCL_BUFFSIZE=1024
121-
export OMP_PROC_BIND=false
122-
export OMP_NUM_THREADS=10
123-
export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/python/site-packages:$LD_LIBRARY_PATH
124-
125-
vllm serve /model/Qwen3-235B-A22B-W8A8 \
126-
--host 0.0.0.0 \
127-
--port 8004 \
128-
--api-server-count 1 \
129-
--data-parallel-size 2 \
130-
--data-parallel-size-local 2 \
131-
--data-parallel-address 192.0.0.1 \
132-
--data-parallel-rpc-port 13389 \
133-
--tensor-parallel-size 8 \
134-
--enable-expert-parallel \
135-
--seed 1024 \
136-
--enforce-eager \
137-
--distributed-executor-backend mp \
138-
--served-model-name qwen3-moe \
139-
--max-model-len 32768 \
140-
--max-num-batched-tokens 32768 \
141-
--trust-remote-code \
142-
--gpu-memory-utilization 0.9 \
143-
--kv-transfer-config \
144-
'{"kv_connector": "MooncakeLayerwiseConnector",
145-
"kv_role": "kv_producer",
146-
"kv_port": "30000",
147-
"engine_id": "0",
148-
"kv_connector_module_path": "vllm_ascend.distributed.mooncake_layerwise_connector",
149-
"kv_connector_extra_config": {
150-
"prefill": {
151-
"dp_size": 2,
152-
"tp_size": 8
153-
},
154-
"decode": {
155-
"dp_size": 32,
156-
"tp_size": 1
157-
}
158-
}
159-
}'
160-
```
161-
162-
::::
163-
164-
::::{tab-item} Prefiller node 2
165-
166-
```shell
167-
unset ftp_proxy
168-
unset https_proxy
169-
unset http_proxy
170-
export HCCL_IF_IP=192.0.0.2
171-
export GLOO_SOCKET_IFNAME="eth0" # network card name
172-
export TP_SOCKET_IFNAME="eth0"
173-
export HCCL_SOCKET_IFNAME="eth0"
174-
export VLLM_USE_V1=1
175-
export HCCL_BUFFSIZE=1024
176-
export OMP_PROC_BIND=false
177-
export OMP_NUM_THREADS=10
178-
export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/python/site-packages:$LD_LIBRARY_PATH
179-
180-
vllm serve /model/Qwen3-235B-A22B-W8A8 \
181-
--host 0.0.0.0 \
182-
--port 8004 \
183-
--api-server-count 1 \
184-
--data-parallel-size 2 \
185-
--data-parallel-size-local 2 \
186-
--data-parallel-address 192.0.0.2 \
187-
--data-parallel-rpc-port 13389 \
188-
--tensor-parallel-size 8 \
189-
--enable-expert-parallel \
190-
--seed 1024 \
191-
--enforce-eager \
192-
--distributed-executor-backend mp \
193-
--served-model-name qwen3-moe \
194-
--max-model-len 32768 \
195-
--max-num-batched-tokens 32768 \
196-
--trust-remote-code \
197-
--gpu-memory-utilization 0.9 \
198-
--kv-transfer-config \
199-
'{"kv_connector": "MooncakeLayerwiseConnector",
200-
"kv_role": "kv_producer",
201-
"kv_port": "30100",
202-
"engine_id": "1",
203-
"kv_connector_module_path": "vllm_ascend.distributed.mooncake_layerwise_connector",
204-
"kv_connector_extra_config": {
205-
"prefill": {
206-
"dp_size": 2,
207-
"tp_size": 8
208-
},
209-
"decode": {
210-
"dp_size": 32,
211-
"tp_size": 1
212-
}
213-
}
214-
}'
215-
```
216-
217-
::::
218-
219-
::::{tab-item} Decoder node 1 (master node)
220-
221-
```shell
222-
unset ftp_proxy
223-
unset https_proxy
224-
unset http_proxy
225-
export HCCL_IF_IP=192.0.0.3
226-
export GLOO_SOCKET_IFNAME="eth0" # network card name
227-
export TP_SOCKET_IFNAME="eth0"
228-
export HCCL_SOCKET_IFNAME="eth0"
229-
export VLLM_USE_V1=1
230-
export HCCL_BUFFSIZE=2048
231-
export OMP_PROC_BIND=false
232-
export OMP_NUM_THREADS=10
233-
export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/python/site-packages:$LD_LIBRARY_PATH
234-
235-
vllm serve /model/Qwen3-235B-A22B-W8A8 \
236-
--host 0.0.0.0 \
237-
--port 8004 \
238-
--api-server-count 1 \
239-
--data-parallel-size 32 \
240-
--data-parallel-size-local 16 \
241-
--data-parallel-address 192.0.0.3 \
242-
--data-parallel-rpc-port 5964 \
243-
--tensor-parallel-size 1 \
244-
--enable-expert-parallel \
245-
--seed 1024 \
246-
--distributed-executor-backend mp \
247-
--served-model-name qwen3-moe \
248-
--max-model-len 32768 \
249-
--max-num-batched-tokens 512 \
250-
--max-num_seqs 16 \
251-
--trust-remote-code \
252-
--no-enable-prefix-caching \
253-
--gpu-memory-utilization 0.9 \
254-
--compilation-config '{"cudagraph_capture_sizes":[16]}' \
255-
--kv-transfer-config \
256-
'{"kv_connector": "MooncakeLayerwiseConnector",
257-
"kv_role": "kv_consumer",
258-
"kv_port": "30200",
259-
"engine_id": "2",
260-
"kv_connector_module_path": "vllm_ascend.distributed.mooncake_layerwise_connector",
261-
"kv_connector_extra_config": {
262-
"prefill": {
263-
"dp_size": 2,
264-
"tp_size": 8
265-
},
266-
"decode": {
267-
"dp_size": 32,
268-
"tp_size": 1
269-
}
270-
}
271-
}'
272-
```
273-
274-
::::
275-
276-
::::{tab-item} Decoder node 2 (primary node)
277-
278-
```shell
279-
unset ftp_proxy
280-
unset https_proxy
281-
unset http_proxy
282-
export HCCL_IF_IP=192.0.0.4
283-
export GLOO_SOCKET_IFNAME="eth0" # network card name
284-
export TP_SOCKET_IFNAME="eth0"
285-
export HCCL_SOCKET_IFNAME="eth0"
286-
export VLLM_USE_V1=1
287-
export HCCL_BUFFSIZE=2048
288-
export OMP_PROC_BIND=false
289-
export OMP_NUM_THREADS=10
290-
export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/python/site-packages:$LD_LIBRARY_PATH
291-
292-
vllm serve /model/Qwen3-235B-A22B-W8A8 \
293-
--host 0.0.0.0 \
294-
--port 8004 \
295-
--headless \
296-
--data-parallel-size 32 \
297-
--data-parallel-size-local 16 \
298-
--data-parallel-start-rank 16 \
299-
--data-parallel-address 192.0.0.3 \
300-
--data-parallel-rpc-port 5964 \
301-
--tensor-parallel-size 1 \
302-
--enable-expert-parallel \
303-
--seed 1024 \
304-
--distributed-executor-backend mp \
305-
--served-model-name qwen3-moe \
306-
--max-model-len 32768 \
307-
--max-num-batched-tokens 512 \
308-
--max-num_seqs 16 \
309-
--trust-remote-code \
310-
--no-enable-prefix-caching \
311-
--gpu-memory-utilization 0.9 \
312-
--compilation-config '{"cudagraph_capture_sizes":[16]}' \
313-
--kv-transfer-config \
314-
'{"kv_connector": "MooncakeLayerwiseConnector",
315-
"kv_role": "kv_consumer",
316-
"kv_port": "30200",
317-
"engine_id": "2",
318-
"kv_connector_module_path": "vllm_ascend.distributed.mooncake_layerwise_connector",
319-
"kv_connector_extra_config": {
320-
"prefill": {
321-
"dp_size": 2,
322-
"tp_size": 8
323-
},
324-
"decode": {
325-
"dp_size": 32,
326-
"tp_size": 1
327-
}
328-
}
329-
}'
330-
```
331-
332-
::::
333-
334-
:::::
335-
336-
### Non-layerwise
337-
338105
:::::{tab-set}
339106

340107
::::{tab-item} Prefiller node 1
@@ -566,25 +333,7 @@ vllm serve /model/Qwen3-235B-A22B-W8A8 \
566333

567334
## Example Proxy for Deployment
568335

569-
Run a proxy server on the same node with the 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)
570-
571-
:::::{tab-set}
572-
573-
::::{tab-item} Layerwise
574-
575-
```shell
576-
python load_balance_proxy_layerwise_server_example.py \
577-
--host 192.0.0.1 \
578-
--port 8080 \
579-
--prefiller-hosts 192.0.0.1 192.0.0.2\
580-
--prefiller-port 8004 8004\
581-
--decoder-hosts 192.0.0.3\
582-
--decoder-ports 8004
583-
```
584-
585-
::::
586-
587-
::::{tab-item} Non-layerwise
336+
Run a proxy server on the same node with the prefiller service instance. You can get the proxy program in the repository's examples: [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)
588337

589338
```shell
590339
python load_balance_proxy_server_example.py \
@@ -596,10 +345,6 @@ python load_balance_proxy_server_example.py \
596345
--decoder-ports 8004
597346
```
598347

599-
::::
600-
601-
:::::
602-
603348
## Verification
604349

605350
Check service health using the proxy server endpoint.

‎vllm_ascend/distributed/__init__.py‎

Lines changed: 0 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -33,8 +33,3 @@ def register_connector():
3333
"MooncakeConnectorStoreV1",
3434
"vllm_ascend.distributed.mooncake.mooncake_store_connector_v1",
3535
"MooncakeConnectorV1")
36-
37-
KVConnectorFactory.register_connector(
38-
"MooncakeLayerwiseConnector",
39-
"vllm_ascend.distributed.mooncake_layerwise_connector",
40-
"MooncakeLayerwiseConnector")

0 commit comments

Comments
 (0)