Skip to content

Commit 16491e4

Browse files
zhangxinyuehfadzhangxinyuehfad
committed
[Test] Add test for deepseek-v3.2-exp-w8a8
Signed-off-by: hfadzxy <starmoon_zhang@163.com>
1 parent 5df5219 commit 16491e4

File tree

6 files changed

+203
-152
lines changed

6 files changed

+203
-152
lines changed

‎.github/workflows/_e2e_nightly_single_node.yaml‎

Lines changed: 16 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -33,6 +33,9 @@ on:
3333
tests:
3434
required: true
3535
type: string
36+
name:
37+
required: false
38+
type: string
3639

3740
# Bash shells do not use ~/.profile or ~/.bashrc so these shells need to be explicitly
3841
# declared as "shell: bash -el {0}" on steps that need to be properly activated.
@@ -94,6 +97,19 @@ jobs:
9497
pip install -r requirements-dev.txt
9598
pip install -v -e .
9699
100+
- name: Install custom-ops (for DeepSeek-V3.2-Exp)
101+
if: ${{ inputs.name == 'deepseek3_2-exp-w8a8' }}
102+
shell: bash -l {0}
103+
run: |
104+
wget https://vllm-ascend.obs.cn-north-4.myhuaweicloud.com/vllm-ascend/a3/CANN-custom_ops-sfa-linux.aarch64.run
105+
chmod +x ./CANN-custom_ops-sfa-linux.aarch64.run
106+
./CANN-custom_ops-sfa-linux.aarch64.run --quiet
107+
export ASCEND_CUSTOM_OPP_PATH=/usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize:${ASCEND_CUSTOM_OPP_PATH}
108+
export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/op_api/lib/:${LD_LIBRARY_PATH}
109+
wget https://vllm-ascend.obs.cn-north-4.myhuaweicloud.com/vllm-ascend/a3/custom_ops-1.0-cp311-cp311-linux_aarch64.whl
110+
pip install custom_ops-1.0-cp311-cp311-linux_aarch64.whl
111+
. /usr/local/Ascend/ascend-toolkit/set_env.sh
112+
97113
- name: Checkout aisbench repo and Install aisbench
98114
run: |
99115
git clone https://gitee.com/aisbench/benchmark.git

‎.github/workflows/vllm_ascend_test_nightly_a3.yaml‎

Lines changed: 7 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -84,12 +84,16 @@ jobs:
8484
- name: qwen3-235b-w8a8
8585
os: linux-aarch64-a3-16
8686
tests: tests/e2e/nightly/models/test_qwen3_235b_w8a8.py
87+
- name: deepseek3_2-exp-w8a8
88+
os: linux-aarch64-a3-16
89+
tests: tests/e2e/nightly/models/test_deepseek_v3_2_exp_w8a8.py
8790
uses: ./.github/workflows/_e2e_nightly_single_node.yaml
8891
with:
8992
vllm: v0.11.0
9093
runner: ${{ matrix.test_config.os }}
9194
image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.3.rc1-a3-ubuntu22.04-py3.11
9295
tests: ${{ matrix.test_config.tests }}
96+
name: ${{ matrix.test_config.name }}
9397

9498
multi-node-tests:
9599
name: multi-node
@@ -106,9 +110,6 @@ jobs:
106110
- name: multi-node-qwen3-dp
107111
config_file_path: Qwen3-235B-A3B.yaml
108112
size: 2
109-
- name: multi-node-deepseek3.2-exp-dp
110-
config_file_path: tests/e2e/nightly/multi_node/config/models/DeepSeek3_2-Exp-W8A8.yaml
111-
size: 2
112113
- name: multi-node-dpsk-4node-pd
113114
config_file_path: DeepSeek-R1-W8A8.yaml
114115
size: 4
@@ -118,6 +119,9 @@ jobs:
118119
- name: multi-node-glm-2node
119120
config_file_path: GLM-4_5.yaml
120121
size: 2
122+
- name: multi-node-dpsk3.2-exp-2node
123+
config_file_path: DeepSeek-V3_2-Exp-bf16.yaml
124+
size: 2
121125
uses: ./.github/workflows/_e2e_nightly_multi_node.yaml
122126
with:
123127
soc_version: a3

‎tests/e2e/nightly/models/test_deepseek_v3_2_exp_w8a8.py‎

Lines changed: 120 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,120 @@
1+
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
2+
# Copyright 2023 The vLLM team.
3+
#
4+
# Licensed under the Apache License, Version 2.0 (the "License");
5+
# you may not use this file except in compliance with the License.
6+
# You may obtain a copy of the License at
7+
#
8+
# http://www.apache.org/licenses/LICENSE-2.0
9+
#
10+
# Unless required by applicable law or agreed to in writing, software
11+
# distributed under the License is distributed on an "AS IS" BASIS,
12+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13+
# See the License for the specific language governing permissions and
14+
# limitations under the License.
15+
# This file is a part of the vllm-ascend project.
16+
#
17+
from typing import Any
18+
19+
import openai
20+
import pytest
21+
from vllm.utils import get_open_port
22+
23+
from tests.e2e.conftest import RemoteOpenAIServer
24+
from tools.aisbench import run_aisbench_cases
25+
26+
MODELS = [
27+
"vllm-ascend/DeepSeek-V3.2-Exp-W8A8",
28+
]
29+
30+
TENSOR_PARALLELS = [8]
31+
DATA_PARALLELS = [2]
32+
FULL_GRAPH = [True, False]
33+
34+
prompts = [
35+
"San Francisco is a",
36+
]
37+
38+
api_keyword_args = {
39+
"max_tokens": 10,
40+
}
41+
42+
aisbench_cases = [{
43+
"case_type": "accuracy",
44+
"dataset_path": "vllm-ascend/gsm8k-lite",
45+
"request_conf": "vllm_api_general_chat",
46+
"dataset_conf": "gsm8k/gsm8k_gen_0_shot_cot_chat_prompt",
47+
"max_out_len": 4096,
48+
"batch_size": 8,
49+
"baseline": 95,
50+
"threshold": 5
51+
}, {
52+
"case_type": "performance",
53+
"dataset_path": "vllm-ascend/GSM8K-in3500-bs400",
54+
"request_conf": "vllm_api_stream_chat",
55+
"dataset_conf": "gsm8k/gsm8k_gen_0_shot_cot_str_perf",
56+
"num_prompts": 16,
57+
"max_out_len": 1500,
58+
"batch_size": 8,
59+
"request_rate": 0,
60+
"baseline": 1,
61+
"threshold": 0.97
62+
}]
63+
64+
65+
@pytest.mark.asyncio
66+
@pytest.mark.parametrize("model", MODELS)
67+
@pytest.mark.parametrize("tp_size", TENSOR_PARALLELS)
68+
@pytest.mark.parametrize("dp_size", DATA_PARALLELS)
69+
@pytest.mark.parametrize("full_graph", FULL_GRAPH)
70+
async def test_models(model: str, tp_size: int, dp_size: int,
71+
full_graph: bool) -> None:
72+
port = get_open_port()
73+
env_dict = {"HCCL_BUFFSIZE": "1024", "VLLM_ASCEND_ENABLE_MLAPO": "0"}
74+
server_args = [
75+
"--no-enable-prefix-caching",
76+
"--enable-expert-parallel",
77+
"--tensor-parallel-size",
78+
str(tp_size),
79+
"--data-parallel-size",
80+
str(dp_size),
81+
"--port",
82+
str(port),
83+
"--max-model-len",
84+
"16384",
85+
"--max-num-batched-tokens",
86+
"16384",
87+
"--block-size",
88+
"16",
89+
"--trust-remote-code",
90+
"--quantization",
91+
"ascend",
92+
"--gpu-memory-utilization",
93+
"0.9",
94+
"--additional-config",
95+
'{"ascend_scheduler_config":{"enabled":true},'
96+
'"torchair_graph_config":{"enabled":true,"graph_batch_sizes":[16]}}',
97+
]
98+
if full_graph:
99+
server_args += [
100+
"--compilation-config",
101+
'{"cudagraph_capture": [16], "cudagraph_model":"FULL_DECODE_ONLY"}'
102+
]
103+
request_keyword_args: dict[str, Any] = {
104+
**api_keyword_args,
105+
}
106+
with RemoteOpenAIServer(model,
107+
server_args,
108+
server_port=port,
109+
env_dict=env_dict,
110+
auto_port=False) as server:
111+
client = server.get_async_client()
112+
batch = await client.completions.create(
113+
model=model,
114+
prompt=prompts,
115+
**request_keyword_args,
116+
)
117+
choices: list[openai.types.CompletionChoice] = batch.choices
118+
assert choices[0].text, "empty response"
119+
# aisbench test
120+
run_aisbench_cases(model, port, aisbench_cases)

‎tests/e2e/nightly/multi_node/config/models/DeepSeek-V3_2-Exp-bf16.yaml‎

Lines changed: 55 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,55 @@
1+
test_name: "test DeepSeek-V3.2-Exp-bf16 multi-dp"
2+
model: "Yanguan/DeepSeek-V3.2-Exp-bf16"
3+
num_nodes: 2
4+
npu_per_node: 16
5+
env_common:
6+
VLLM_USE_MODELSCOPE: true
7+
OMP_PROC_BIND: false
8+
OMP_NUM_THREADS: 100
9+
HCCL_BUFFSIZE: 1024
10+
SERVER_PORT: 8080
11+
VLLM_ASCEND_ENABLE_MLAPO: 0
12+
13+
deployment:
14+
-
15+
server_cmd: >
16+
vllm serve Yanguan/DeepSeek-V3.2-Exp-bf16 \
17+
--host 0.0.0.0
18+
--port $SERVER_PORT
19+
--data-parallel-address $LOCAL_IP
20+
--data-parallel-size 2
21+
--data-parallel-size-local 1
22+
--data-parallel-rpc-port 13389
23+
--tensor-parallel-size 16
24+
--seed 1024
25+
--enable-expert-parallel
26+
--max-num-seqs 16
27+
--max-model-len 17450
28+
--max-num-batched-tokens 17450
29+
--trust-remote-code
30+
--no-enable-prefix-caching
31+
--gpu-memory-utilization 0.9
32+
--additional-config '{"ascend_scheduler_config":{"enabled":true},"torchair_graph_config":{"enabled":true,"graph_batch_sizes":[16]}}'
33+
34+
-
35+
server_cmd: >
36+
vllm serve Yanguan/DeepSeek-V3.2-Exp-bf16 \
37+
--host 0.0.0.0
38+
--port $SERVER_PORT
39+
--headless
40+
--data-parallel-size 2
41+
--data-parallel-size-local 1
42+
--data-parallel-start-rank 1
43+
--data-parallel-address $MASTER_IP
44+
--data-parallel-rpc-port 13389
45+
--tensor-parallel-size 16
46+
--seed 1024
47+
--max-num-seqs 16
48+
--max-model-len 17450
49+
--max-num-batched-tokens 17450
50+
--enable-expert-parallel
51+
--trust-remote-code
52+
--no-enable-prefix-caching
53+
--gpu-memory-utilization 0.92
54+
--additional-config '{"ascend_scheduler_config":{"enabled":true},"torchair_graph_config":{"enabled":true,"graph_batch_sizes":[16]}}'
55+
benchmarks:

‎tests/e2e/nightly/multi_node/config/models/DeepSeek3_2-Exp-W8A8.yaml‎

Lines changed: 0 additions & 86 deletions
This file was deleted.

0 commit comments

Comments
 (0)