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

[CI] Add unit test framework (#1201)

Browse Source 
This PR added the unit test framework to enable ut for vLLM Ascend. Unit
test runs on CPU machines. It'll be ran once lint check is passed the
same as e2e test.

For unit test, this PR created a new folder called `ut` under `tests`
module. All the test file in `ut` should keep the same with the code in
`vllm-ascend`. The file name should be start with `test_` prefix. For
example, in this PR. the `test_ascend_config.py` is added for
`ascend_config.py` test.

A new fille `worker/test_worker_v1.py` is also added as the placeholder.
This file should be the unit test for `vllm-ascend/worker/worker_v1.py`.

Additional, a new `fake_weight` folder is added, it contains the
config.json from `facebook/opt-125m`, so that the test will not always
visit huggingface.

TODO:
We should add all the unit test file one by one in the future.

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
wangxiyuan
2025-06-16 18:32:28 +08:00
committed by GitHub
parent 966557a2a3
commit 69b817ed65
57 changed files with 396 additions and 267 deletions
Download Patch File Download Diff File Expand all files Collapse all files

57
tests/multicard/test_dynamic_npugraph_batchsize.py
View File

@@ -1,57 +0,0 @@
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
# Copyright 2023 The vLLM team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# This file is a part of the vllm-ascend project.
#
import pytest
import torch
from vllm import LLM, SamplingParams
MODELS = [
"Qwen/Qwen2.5-0.5B-Instruct",
]
TENSOR_PARALLELS = [2]
prompts = [
"Hello, my name is",
"The future of AI is",
]
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("tp_size", TENSOR_PARALLELS)
@pytest.mark.parametrize("max_tokens", [64])
@pytest.mark.parametrize("temperature", [0.0])
@pytest.mark.parametrize("ignore_eos", [True])
def test_models(model: str, tp_size: int, max_tokens: int, temperature: int,
ignore_eos: bool) -> None:
# Create an LLM.
llm = LLM(
model=model,
tensor_parallel_size=tp_size,
)
# Prepare sampling_parames
sampling_params = SamplingParams(
max_tokens=max_tokens,
temperature=temperature,
ignore_eos=ignore_eos,
)
# Generate texts from the prompts.
# The output is a list of RequestOutput objects
outputs = llm.generate(prompts, sampling_params)
torch.npu.synchronize()
# The output length should be equal to prompts length.
assert len(outputs) == len(prompts)

21
tests/multicard/test_ilama_lora_tp2.py
View File

@@ -1,21 +0,0 @@
import pytest
from tests.conftest import VllmRunner
from tests.singlecard.test_ilama_lora import (EXPECTED_LORA_OUTPUT, MODEL_PATH,
do_sample)
@pytest.mark.parametrize("distributed_executor_backend", ["mp"])
def test_ilama_lora_tp2(distributed_executor_backend, ilama_lora_files):
with VllmRunner(model_name=MODEL_PATH,
enable_lora=True,
max_loras=4,
max_model_len=1024,
max_num_seqs=16,
tensor_parallel_size=2,
distributed_executor_backend=distributed_executor_backend
) as vllm_model:
output = do_sample(vllm_model.model, ilama_lora_files, lora_id=2)
for i in range(len(EXPECTED_LORA_OUTPUT)):
assert output[i] == EXPECTED_LORA_OUTPUT[i]

114
tests/multicard/test_offline_inference_distributed.py
View File

@@ -1,114 +0,0 @@
#
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
# Copyright 2023 The vLLM team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# This file is a part of the vllm-ascend project.
# Adapted from vllm/tests/basic_correctness/test_basic_correctness.py
#
"""Compare the short outputs of HF and vLLM when using greedy sampling.
Run `pytest tests/test_offline_inference.py`.
"""
import os
from unittest.mock import patch
from modelscope import snapshot_download # type: ignore
from vllm import SamplingParams
from tests.conftest import VllmRunner
os.environ["PYTORCH_NPU_ALLOC_CONF"] = "max_split_size_mb:256"
def test_models_distributed_QwQ():
example_prompts = [
"Hello, my name is",
]
dtype = "half"
max_tokens = 5
with VllmRunner(
"Qwen/QwQ-32B",
dtype=dtype,
tensor_parallel_size=4,
distributed_executor_backend="mp",
) as vllm_model:
vllm_model.generate_greedy(example_prompts, max_tokens)
def test_models_distributed_DeepSeek():
example_prompts = [
"Hello, my name is",
]
dtype = "half"
max_tokens = 5
with VllmRunner(
"deepseek-ai/DeepSeek-V2-Lite",
dtype=dtype,
tensor_parallel_size=4,
distributed_executor_backend="mp",
) as vllm_model:
vllm_model.generate_greedy(example_prompts, max_tokens)
@patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_TOPK_OPTIMIZE": "1"})
def test_models_distributed_topk() -> None:
example_prompts = [
"vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs.",
"Briefly describe the major milestones in the development of artificial intelligence from 1950 to 2020.",
"Compare and contrast artificial intelligence with human intelligence in terms of processing information.",
]
dtype = "half"
sampling_params = SamplingParams(max_tokens=5,
temperature=0.0,
top_k=50,
top_p=0.9)
with VllmRunner(
"deepseek-ai/DeepSeek-V2-Lite",
dtype=dtype,
tensor_parallel_size=4,
distributed_executor_backend="mp",
) as vllm_model:
vllm_model.generate(example_prompts, sampling_params)
@patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_DBO": "1"})
def test_models_distributed_DeepSeek_dbo():
example_prompts = ["The president of the United States is"] * 41
dtype = "half"
sampling_params = SamplingParams(max_tokens=100, temperature=0.0)
with VllmRunner(
"deepseek-ai/DeepSeek-V2-Lite",
dtype=dtype,
tensor_parallel_size=4,
distributed_executor_backend="mp",
) as vllm_model:
vllm_model.generate(example_prompts, sampling_params)
def test_models_distributed_DeepSeek_W8A8():
example_prompts = [
"Hello, my name is",
]
max_tokens = 5
with VllmRunner(
snapshot_download("vllm-ascend/DeepSeek-V2-Lite-W8A8"),
max_model_len=8192,
enforce_eager=True,
dtype="auto",
tensor_parallel_size=4,
quantization="ascend",
) as vllm_model:
vllm_model.generate_greedy(example_prompts, max_tokens)

110
tests/multicard/test_pyhccl_distributed.py
View File

@@ -1,110 +0,0 @@
#
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
# This file is a part of the vllm-ascend project.
# Adapted from vllm/tests/basic_correctness/test_basic_correctness.py
# Copyright 2023 The vLLM team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import multiprocessing
import os
import torch
from vllm.distributed.parallel_state import (get_world_group,
init_distributed_environment)
from vllm.utils import update_environment_variables
from vllm_ascend.distributed.device_communicators.pyhccl import \
PyHcclCommunicator
def distributed_run(fn, world_size):
number_of_processes = world_size
processes: list[multiprocessing.Process] = []
for i in range(number_of_processes):
env: dict[str, str] = {}
env['RANK'] = str(i)
env['LOCAL_RANK'] = str(i)
env['WORLD_SIZE'] = str(number_of_processes)
env['LOCAL_WORLD_SIZE'] = str(number_of_processes)
env['MASTER_ADDR'] = 'localhost'
env['MASTER_PORT'] = '12345'
p = multiprocessing.Process(target=fn, args=(env, ))
processes.append(p)
p.start()
for p in processes:
p.join()
for p in processes:
assert p.exitcode == 0
def worker_fn_wrapper(fn):
# `multiprocessing.Process` cannot accept environment variables directly
# so we need to pass the environment variables as arguments
# and update the environment variables in the function
def wrapped_fn(env):
update_environment_variables(env)
local_rank = os.environ['LOCAL_RANK']
device = torch.device(f"npu:{local_rank}")
torch.npu.set_device(device)
init_distributed_environment(backend="hccl")
fn()
return wrapped_fn
@worker_fn_wrapper
def worker_fn():
pynccl_comm = PyHcclCommunicator(get_world_group().cpu_group,
device=get_world_group().device)
tensor = torch.ones(16, 1024, 1024,
dtype=torch.float32).npu(pynccl_comm.rank)
tensor = pynccl_comm.all_reduce(tensor)
torch.npu.synchronize()
assert torch.all(tensor == pynccl_comm.world_size).cpu().item()
# def test_pyhccl():
# distributed_run(worker_fn, 2)
@worker_fn_wrapper
def broadcast_worker_fn():
# Test broadcast for every root rank.
# Essentially this is an all-gather operation.
pyhccl_comm = PyHcclCommunicator(get_world_group().cpu_group,
device=get_world_group().device)
recv_tensors = [
torch.empty(16,
1024,
1024,
dtype=torch.float32,
device=pyhccl_comm.device)
for i in range(pyhccl_comm.world_size)
]
recv_tensors[pyhccl_comm.rank] = torch.ones(
16, 1024, 1024, dtype=torch.float32,
device=pyhccl_comm.device) * pyhccl_comm.rank
for i in range(pyhccl_comm.world_size):
pyhccl_comm.broadcast(recv_tensors[i], src=i)
# the broadcast op might be launched in a different stream
# need to synchronize to make sure the tensor is ready
torch.npu.synchronize()
assert torch.all(recv_tensors[i] == i).cpu().item()
# def test_pyhccl_broadcast():
# distributed_run(broadcast_worker_fn, 4)

80
tests/multicard/test_torchair_graph_mode.py
View File

@@ -1,80 +0,0 @@
#
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
# Copyright 2023 The vLLM team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# This file is a part of the vllm-ascend project.
#
"""Compare the short outputs of HF and vLLM when using greedy sampling.
Run `pytest tests/multicard/test_torchair_graph_mode.py`.
"""
import os
import pytest
from tests.conftest import VllmRunner
os.environ["PYTORCH_NPU_ALLOC_CONF"] = "max_split_size_mb:256"
@pytest.mark.skipif(os.getenv("VLLM_USE_V1") == "0",
reason="torchair graph is not supported on v0")
def test_e2e_deepseekv3_with_torchair(monkeypatch: pytest.MonkeyPatch):
with monkeypatch.context() as m:
m.setenv("VLLM_USE_MODELSCOPE", "True")
m.setenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
example_prompts = [
"Hello, my name is",
"The president of the United States is",
"The capital of France is",
"The future of AI is",
]
dtype = "half"
max_tokens = 5
# torchair is only work without chunked-prefill now
with VllmRunner(
"vllm-ascend/DeepSeek-V3-Pruning",
dtype=dtype,
tensor_parallel_size=4,
distributed_executor_backend="mp",
additional_config={
"torchair_graph_config": {
"enabled": True,
},
"ascend_scheduler_config": {
"enabled": True,
},
"refresh": True,
},
enforce_eager=False,
) as vllm_model:
# use greedy sampler to make sure the generated results are fix
vllm_output = vllm_model.generate_greedy(example_prompts,
max_tokens)
# NOTE: vllm-ascend/DeepSeek-V3-Pruning is a random weight of
# DeepSeek-V3 with 2 hidden layers, thus the golden results seems
# inaccurate. This will only change if accuracy improves with the
# official weights of DeepSeek-V3.
golden_results = [
'Hello, my name is feasibility伸 spazio debtor添',
'The president of the United States is begg"""\n杭州风和 bestimm',
'The capital of France is frequentlyশามalinkAllowed',
'The future of AI is deleting俯احت怎么样了حراف',
]
assert len(golden_results) == len(vllm_output)
for i in range(len(vllm_output)):
assert golden_results[i] == vllm_output[i][1]
print(f"Generated text: {vllm_output[i][1]!r}")