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

[CI]Add Disaggregated PD Nightly Test for Qwen3-235B and Qwen3-VL-235B (#5502)

Browse Source 
### What this PR does / why we need it?
This PR adds online **Disaggregated Prefill/Decode** performance and
accuracy tests for the **Qwen3-235B-A22B** and
**Qwen3-VL-235B-A22B-Instruct** models to the Nightly test suite.

These test configurations simulate the deployment of massive MoE and
Vision-Language models in **a dual-node (32 NPU)** environment,
utilizing Mooncake (KVCache Transfer) technology to achieve efficient KV
cache transfer between the Prefill node and the Decode node.

#### Test Configuration
**Qwen3-235B-A22B**
- Model: Qwen/Qwen3-235B-A22B
- Hardware: A3, 2 Nodes (32 NPUs total, 16 NPUs per node)
- Architecture: Disaggregated Prefill & Decode
- Node 0 (Producer/Prefill): **DP2 + TP8 + EP + FLASHCOMM1 +
FUSED_MC2**.
- Node 1 (Consumer/Decode): **DP4 + TP4 + EP + FLASHCOMM1 + FUSED_MC2 +
FULL_DECODE_ONLY**.
- Benchmarks:
  - Performance: vllm-ascend/GSM8K-in3500-bs2800.
  - Accuracy: vllm-ascend/gsm8k-lite.

**Qwen3-VL-235B-A22B-Instruct**
- Model: Qwen/Qwen3-VL-235B-A22B-Instruct
- Hardware: A3, 2 Nodes (32 NPUs total, 16 NPUs per node)
- Architecture: Disaggregated Prefill & Decode
  - Node 0 (Producer/Prefill): **DP2 + TP8 + EP**.
  - Node 1 (Consumer/Decode): **DP4 + TP4 + EP + FULL_DECODE_ONLY**.
- Benchmarks:
  - Performance: vllm-ascend/textvqa-perf-1080p.
  - Accuracy: vllm-ascend/textvqa-lite.

### How was this patch tested?
Nightly test action on CI

- vLLM version: v0.13.0
- vLLM main:
45c1ca1ca1

---------

Signed-off-by: MrZ20 <2609716663@qq.com>
This commit is contained in:
SILONG ZENG
2026-01-09 16:25:20 +08:00
committed by GitHub
parent f63c1341d9
commit 09b3f9d91b
4 changed files with 241 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
.github/workflows/nightly_test_a3.yaml vendored
View File

@@ -77,6 +77,12 @@ jobs:
- name: multi-node-qwenw8a8-2node-longseq
config_file_path: Qwen3-235B-W8A8-longseq.yaml
size: 2
- name: multi-node-qwen-disagg-pd
config_file_path: Qwen3-235B-disagg-pd.yaml
size: 2
- name: multi-node-qwen-vl-disagg-pd
config_file_path: Qwen3-VL-235B-disagg-pd.yaml
size: 2
uses: ./.github/workflows/_e2e_nightly_multi_node.yaml
with:
soc_version: a3

121
tests/e2e/nightly/multi_node/config/Qwen3-235B-disagg-pd.yaml Normal file
View File

@@ -0,0 +1,121 @@
test_name: "test Qwen3-235B-A22B disaggregated_prefill"
model: "Qwen/Qwen3-235B-A22B"
num_nodes: 2
npu_per_node: 16
env_common:
VLLM_USE_MODELSCOPE: true
PYTORCH_NPU_ALLOC_CONF: expandable_segments:True
HCCL_BUFFSIZE: 1024
HCCL_OP_EXPANSION_MODE: "AIV"
OMP_PROC_BIND: false
OMP_NUM_THREADS: 1
VLLM_ASCEND_ENABLE_FLASHCOMM1: 1
VLLM_ASCEND_ENABLE_FUSED_MC2: 2
TASK_QUEUE_ENABLE: 1
SERVER_PORT: 8080
disaggregated_prefill:
enabled: true
prefiller_host_index: [0]
decoder_host_index: [1]
deployment:
-
server_cmd: >
vllm serve "Qwen/Qwen3-235B-A22B"
--host 0.0.0.0
--port $SERVER_PORT
--data-parallel-size 2
--data-parallel-size-local 2
--data-parallel-start-rank 0
--data-parallel-address $LOCAL_IP
--data-parallel-rpc-port 13389
--tensor-parallel-size 8
--seed 1024
--max-num-seqs 32
--max-model-len 8192
--max-num-batched-tokens 8192
--enable-expert-parallel
--trust-remote-code
--gpu-memory-utilization 0.9
--no-enable-prefix-caching
--kv-transfer-config
'{"kv_connector": "MooncakeConnectorV1",
"kv_role": "kv_producer",
"kv_port": "30000",
"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": 4,
"tp_size": 4
}
}
}'
-
server_cmd: >
vllm serve "Qwen/Qwen3-235B-A22B"
--host 0.0.0.0
--port $SERVER_PORT
--data-parallel-size 4
--data-parallel-size-local 4
--data-parallel-start-rank 0
--data-parallel-address $LOCAL_IP
--data-parallel-rpc-port 13389
--tensor-parallel-size 4
--seed 1024
--max-num-seqs 32
--max-model-len 8192
--max-num-batched-tokens 8192
--enable-expert-parallel
--trust-remote-code
--gpu-memory-utilization 0.9
--no-enable-prefix-caching
--compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
--async-scheduling
--kv-transfer-config
'{"kv_connector": "MooncakeConnectorV1",
"kv_role": "kv_consumer",
"kv_port": "30100",
"engine_id": "1",
"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": 4,
"tp_size": 4
}
}
}'
benchmarks:
perf:
case_type: performance
dataset_path: vllm-ascend/GSM8K-in3500-bs2800
request_conf: vllm_api_stream_chat
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
num_prompts: 2800
max_out_len: 1500
batch_size: 700
request_rate: 11.2
baseline: 1
threshold: 0.97
acc:
case_type: accuracy
dataset_path: vllm-ascend/gsm8k-lite
request_conf: vllm_api_general_chat
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_chat_prompt
max_out_len: 7680
batch_size: 512
baseline: 97
threshold: 3

108
tests/e2e/nightly/multi_node/config/Qwen3-VL-235B-disagg-pd.yaml Normal file
View File

@@ -0,0 +1,108 @@
test_name: "test Qwen3-VL-235B-A22B disaggregated_prefill"
model: "Qwen/Qwen3-VL-235B-A22B-Instruct"
num_nodes: 2
npu_per_node: 16
env_common:
VLLM_USE_MODELSCOPE: true
HCCL_BUFFSIZE: 1024
SERVER_PORT: 8080
OMP_PROC_BIND: false
OMP_NUM_THREADS: 1
HCCL_OP_EXPANSION_MODE: "AIV"
TASK_QUEUE_ENABLE: 1
PYTORCH_NPU_ALLOC_CONF: expandable_segments:True
disaggregated_prefill:
enabled: true
prefiller_host_index: [0]
decoder_host_index: [1]
deployment:
-
server_cmd: >
vllm serve "Qwen/Qwen3-VL-235B-A22B-Instruct"
--host 0.0.0.0
--port $SERVER_PORT
--data-parallel-size 2
--data-parallel-size-local 2
--tensor-parallel-size 8
--seed 1024
--enable-expert-parallel
--max-num-seqs 32
--max-model-len 8192
--max-num-batched-tokens 8192
--trust-remote-code
--no-enable-prefix-caching
--gpu-memory-utilization 0.9
--kv-transfer-config
'{"kv_connector": "MooncakeConnectorV1",
"kv_role": "kv_producer",
"kv_port": "30000",
"engine_id": "0",
"kv_connector_module_path": "vllm_ascend.distributed.mooncake_connector",
"kv_connector_extra_config": {
"prefill": {
"dp_size": 2,
"tp_size": 8
},
"decode": {
"dp_size": 4,
"tp_size": 4
}
}
}'
-
server_cmd: >
vllm serve "Qwen/Qwen3-VL-235B-A22B-Instruct"
--host 0.0.0.0
--port $SERVER_PORT
--data-parallel-size 4
--data-parallel-size-local 4
--tensor-parallel-size 4
--seed 1024
--enable-expert-parallel
--max-num-seqs 32
--max-model-len 8192
--max-num-batched-tokens 8192
--trust-remote-code
--no-enable-prefix-caching
--gpu-memory-utilization 0.9
--compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
--kv-transfer-config
'{"kv_connector": "MooncakeConnectorV1",
"kv_role": "kv_consumer",
"kv_port": "30200",
"engine_id": "1",
"kv_connector_module_path": "vllm_ascend.distributed.mooncake_connector",
"kv_connector_extra_config": {
"prefill": {
"dp_size": 2,
"tp_size": 8
},
"decode": {
"dp_size": 4,
"tp_size": 4
}
}
}'
benchmarks:
perf:
case_type: performance
dataset_path: vllm-ascend/textvqa-perf-1080p
request_conf: vllm_api_stream_chat
dataset_conf: textvqa/textvqa_gen_base64
num_prompts: 2800
max_out_len: 1500
batch_size: 64
request_rate: 11.2
baseline: 1
threshold: 0.97
acc:
case_type: accuracy
dataset_path: vllm-ascend/textvqa-lite
request_conf: vllm_api_stream_chat
dataset_conf: textvqa/textvqa_gen_base64
max_out_len: 7680
batch_size: 64
baseline: 85
threshold: 5

8
vllm_ascend/ascend_config.py
View File

@@ -108,8 +108,12 @@ class AscendConfig:
decode_tp_size = min(decode_tp_size, num_kv_head)
self.pd_head_ratio = prefill_tp_size // decode_tp_size
except Exception:
raise AssertionError(
"Can not get num_key_value_heads from model_config")
raise ValueError(
"The text_config extracted from the model config does not have "
"`num_key_value_heads` attribute. This indicates a mismatch "
"between the model config and vLLM's expectations. Please "
"ensure that the model config is compatible with vLLM."
)
if self.pd_tp_ratio == 0:
raise AssertionError(