Skip to content

Commit 48cc77a

Browse files
wangxiyuanzhangxinyuehfad
wangxiyuan
authored and
zhangxinyuehfad
committed
Refactor e2e CI (vllm-project#2276)
Refactor E2E CI to make it clear and faster 1. remove some uesless e2e test 2. remove some uesless function 3. Make sure all test runs with VLLMRunner to avoid oom error 4. Make sure all ops test end with torch.empty_cache to avoid oom error 5. run the test one by one to avoid resource limit error - vLLM version: v0.10.1.1 - vLLM main: vllm-project/vllm@a344a5a Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com> Signed-off-by: hfadzxy <starmoon_zhang@163.com>
1 parent 9f1e054 commit 48cc77a

File tree

16 files changed

+63
-74
lines changed

16 files changed

+63
-74
lines changed

‎.github/workflows/vllm_ascend_test.yaml‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -201,7 +201,7 @@ jobs:
201201
pytest -sv tests/e2e/singlecard/test_embedding.py
202202
pytest -sv tests/e2e/singlecard/test_guided_decoding.py
203203
# TODO: Fix lora accuracy error
204-
pytest -sv tests/e2e/singlecard/test_ilama_lora.py
204+
# pytest -sv tests/e2e/singlecard/test_ilama_lora.py
205205
pytest -sv tests/e2e/singlecard/test_profile_execute_duration.py
206206
pytest -sv tests/e2e/singlecard/test_quantization.py
207207
pytest -sv tests/e2e/singlecard/test_sampler.py
@@ -280,7 +280,7 @@ jobs:
280280
# external_launcher test is not stable enough. Fix it later
281281
# pytest -sv tests/e2e/multicard/test_external_launcher.py
282282
pytest -sv tests/e2e/multicard/test_fused_moe_allgather_ep.py
283-
pytest -sv tests/e2e/multicard/test_ilama_lora_tp2.py
283+
# pytest -sv tests/e2e/multicard/test_ilama_lora_tp2.py
284284
285285
# To avoid oom, we need to run the test in a single process.
286286
pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_QwQ

‎examples/disaggregated_prefill_v1/gen_ranktable.py‎

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -4,7 +4,8 @@
44

55
import torch.distributed as dist
66

7-
from vllm_ascend.utils import AscendSocVersion, init_ascend_soc_version, get_ascend_soc_version
7+
from setup import AscendSocVersion
8+
from vllm_ascend._build_info import __ascend_soc_version__
89

910
parser = argparse.ArgumentParser(
1011
description="Arguments of rank table generator", )
@@ -34,8 +35,7 @@
3435
# and is different from WORLD_SIZE in gen_rank_table.sh.
3536
world_size = os.environ.get("WORLD_SIZE")
3637

37-
init_ascend_soc_version()
38-
soc_info = get_ascend_soc_version()
38+
soc_info = __ascend_soc_version__
3939

4040

4141
def get_cmd_stdout(cmd):

‎requirements-dev.txt‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -5,7 +5,7 @@ openai
55
pytest >= 6.0
66
pytest-asyncio
77
pytest-mock
8-
lm-eval==0.4.8
8+
lm-eval==0.4.9.1
99
types-jsonschema
1010
xgrammar
1111
zmq

‎setup.py‎

Lines changed: 31 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -22,6 +22,8 @@
2222
import os
2323
import subprocess
2424
import sys
25+
import torch_npu
26+
from enum import Enum
2527
from sysconfig import get_paths
2628
from typing import Dict, List
2729

@@ -32,6 +34,19 @@
3234
from setuptools.command.install import install
3335
from setuptools_scm import get_version
3436

37+
# Supported SOC_VERSION codes
38+
ASCEND_A2_SOC_VERSION_CODES = range(220, 226)
39+
ASCEND_A3_SOC_VERSION_CODES = range(250, 256)
40+
ASCEND_310P_SOC_VERSION_CODE = 202
41+
42+
# TODO(zzzzwwjj): Currently there is no clear SOC_VERSION policy for A2 and A3 in CANN.
43+
# So we get the version dynamically. In the future, we should get the version info from _build_info like 310p does.
44+
class AscendSocVersion(Enum):
45+
A2 = 0
46+
A3 = 1
47+
P3 = 2
48+
UNDEFINED = 3
49+
3550

3651
def load_module_from_path(module_name, path):
3752
spec = importlib.util.spec_from_file_location(module_name, path)
@@ -79,27 +94,38 @@ def __init__(self,
7994
self.cmake_lists_dir = os.path.abspath(cmake_lists_dir)
8095

8196

97+
8298
class custom_build_info(build_py):
8399

84100
def run(self):
85-
soc_version = envs.SOC_VERSION
86-
if not soc_version:
101+
soc_version_name = envs.SOC_VERSION
102+
if not soc_version_name:
87103
raise ValueError(
88104
"SOC version is not set. Please set SOC_VERSION environment variable."
89105
)
90-
if "310" in soc_version and not envs.COMPILE_CUSTOM_KERNELS:
106+
if "310" in soc_version_name and not envs.COMPILE_CUSTOM_KERNELS:
91107
raise ValueError(
92108
"SOC version 310 only supports custom kernels. Please set COMPILE_CUSTOM_KERNELS=1 to enable custom kernels."
93109
)
110+
soc_version_code = torch_npu.npu.get_soc_version()
111+
if soc_version_code in ASCEND_A2_SOC_VERSION_CODES:
112+
ascend_soc_version = AscendSocVersion.A2
113+
elif soc_version_code in ASCEND_A3_SOC_VERSION_CODES:
114+
ascend_soc_version = AscendSocVersion.A3
115+
elif soc_version_code == ASCEND_310P_SOC_VERSION_CODE:
116+
ascend_soc_version = AscendSocVersion.P3
117+
else:
118+
ascend_soc_version = AscendSocVersion.UNDEFINED
94119

95120
package_dir = os.path.join(ROOT_DIR, "vllm_ascend", "_build_info.py")
96121
with open(package_dir, "w+") as f:
97122
f.write('# Auto-generated file\n')
98-
f.write(f"__soc_version__ = '{soc_version}'\n")
123+
f.write(f"__soc_version__ = '{soc_version_name}'\n")
124+
f.write(f"__ascend_soc_version__ = '{ascend_soc_version}'\n")
99125
f.write(
100126
f"__sleep_mode_enabled__ = {envs.COMPILE_CUSTOM_KERNELS}\n")
101127
logging.info(
102-
f"Generated _build_info.py with SOC version: {soc_version}")
128+
f"Generated _build_info.py with SOC version: {soc_version_name}")
103129
super().run()
104130

105131

‎tests/ut/ops/test_fused_ops.py‎

Lines changed: 4 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -30,7 +30,7 @@
3030
AscendUnquantizedFusedMoEMethod)
3131
from vllm_ascend.ops.layers.experts_selector import select_experts
3232
from vllm_ascend.ops.layers.moe_mlp import unified_apply_mlp
33-
from vllm_ascend.utils import AscendSocVersion, adapt_patch
33+
from setup import AscendSocVersion, adapt_patch
3434

3535
adapt_patch(True)
3636

@@ -157,7 +157,6 @@ def capture_register(dispatcher_instance):
157157
scheduler_config=MagicMock(max_num_seqs=4),
158158
model_config=MagicMock(max_model_len=2048)
159159
)), \
160-
patch("vllm_ascend.utils.get_ascend_soc_version", return_value=AscendSocVersion.A3), \
161160
patch.object(token_dispatcher_module, 'setup_token_dispatchers', mock_setup_token_dispatchers), \
162161
patch('vllm_ascend.ops.layers.moe_mlp.get_forward_context',
163162
return_value=mock_forward_context_obj):
@@ -465,8 +464,9 @@ def test_apply_with_expert_map(self, moe_method, mock_dist_env,
465464
with_quant=False,
466465
token_dispatcher=selected_token_dispatcher)
467466

468-
with patch("vllm_ascend.ops.fused_moe.get_forward_context", return_value=forward_context), \
469-
patch("vllm_ascend.utils.get_ascend_soc_version", return_value=AscendSocVersion.A3):
467+
with patch("vllm_ascend.ops.fused_moe.MOE_ALL2ALL_BUFFER",
468+
alltoall_buffer), \
469+
patch("vllm_ascend.ops.fused_moe.get_forward_context", return_value=forward_context):
470470

471471
expert_map = torch.tensor([0, 1, 2, -1, -1, -1, -1, -1])
472472
moe_method.ep_size = ep_size

‎tests/ut/ops/test_token_dispatcher.py‎

Lines changed: 0 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -50,12 +50,6 @@ def setUp(self):
5050
return_value=self.forward_context)
5151
self.forward_context_patch.start()
5252

53-
# Mock get_ascend_soc_version()
54-
self.ascend_soc_version_patch = patch(
55-
"vllm_ascend.ops.moe_dispatcher.token_dispatcher.get_ascend_soc_version",
56-
return_value=AscendSocVersion.A3)
57-
self.ascend_soc_version_patch.start()
58-
5953
kwargs = {"with_quant": False, "top_k": 8, "num_experts": 128}
6054
self.dispatcher = TokenDispatcherWithMC2(**kwargs)
6155
self.row_idx = torch.arange(10, dtype=torch.int32)

‎tests/ut/torchair/ops/test_torchair_fused_moe.py‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -27,7 +27,7 @@
2727
from vllm_ascend.quantization.quantizer import W8A8Quantizer
2828
from vllm_ascend.torchair.ops.torchair_fused_moe import (
2929
TorchairAscendFusedMoE, TorchairAscendUnquantizedFusedMoEMethod)
30-
from vllm_ascend.utils import AscendSocVersion, adapt_patch # noqa E402
30+
from setup import AscendSocVersion, adapt_patch # noqa E402
3131

3232
adapt_patch(True)
3333

‎tests/ut/worker/test_worker_v1.py‎

Lines changed: 0 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -41,7 +41,6 @@ def setUp(self):
4141
@patch("vllm_ascend.worker.worker_v1._register_atb_extensions")
4242
@patch("vllm_ascend.worker.worker_v1.register_ascend_customop")
4343
@patch("vllm_ascend.worker.worker_v1.init_ascend_config")
44-
@patch("vllm_ascend.worker.worker_v1.init_ascend_soc_version")
4544
@patch("vllm_ascend.worker.worker_v1.try_register_lib")
4645
@patch("vllm.utils.init_cached_hf_modules")
4746
@patch("vllm_ascend.worker.worker_v1.NPUWorker._init_profiler")
@@ -50,7 +49,6 @@ def test_init_npu_worker_normal_case(
5049
mock_init_profiler,
5150
mock_init_cached_hf_modules,
5251
mock_try_register_lib,
53-
mock_init_ascend_soc_version,
5452
mock_init_ascend_config,
5553
mock_register_ascend_customop,
5654
mock_register_atb_extensions,
@@ -78,7 +76,6 @@ def test_init_npu_worker_normal_case(
7876
mock_register_atb_extensions.assert_called_once()
7977
mock_register_ascend_customop.assert_called_once()
8078
mock_init_ascend_config.assert_called_once_with(self.vllm_config_mock)
81-
mock_init_ascend_soc_version.assert_called_once()
8279

8380
# Verify try_register_lib call
8481
mock_try_register_lib.assert_called_once_with(
@@ -98,7 +95,6 @@ def test_init_npu_worker_normal_case(
9895
@patch("vllm_ascend.worker.worker_v1._register_atb_extensions")
9996
@patch("vllm_ascend.worker.worker_v1.register_ascend_customop")
10097
@patch("vllm_ascend.worker.worker_v1.init_ascend_config")
101-
@patch("vllm_ascend.worker.worker_v1.init_ascend_soc_version")
10298
@patch("vllm_ascend.worker.worker_v1.try_register_lib")
10399
@patch("vllm.utils.init_cached_hf_modules")
104100
@patch("vllm_ascend.worker.worker_v1.NPUWorker._init_profiler")
@@ -107,7 +103,6 @@ def test_init_npu_worker_with_trust_remote_code(
107103
mock_init_profiler,
108104
mock_init_cached_hf_modules,
109105
mock_try_register_lib,
110-
mock_init_ascend_soc_version,
111106
mock_init_ascend_config,
112107
mock_register_ascend_customop,
113108
mock_register_atb_extensions,
@@ -138,7 +133,6 @@ def test_init_npu_worker_with_trust_remote_code(
138133
@patch("vllm_ascend.worker.worker_v1._register_atb_extensions")
139134
@patch("vllm_ascend.worker.worker_v1.register_ascend_customop")
140135
@patch("vllm_ascend.worker.worker_v1.init_ascend_config")
141-
@patch("vllm_ascend.worker.worker_v1.init_ascend_soc_version")
142136
@patch("vllm_ascend.worker.worker_v1.try_register_lib")
143137
@patch("vllm.utils.init_cached_hf_modules")
144138
@patch("vllm_ascend.worker.worker_v1.NPUWorker._init_profiler")
@@ -147,7 +141,6 @@ def test_init_npu_worker_with_custom_cache_dtype(
147141
mock_init_profiler,
148142
mock_init_cached_hf_modules,
149143
mock_try_register_lib,
150-
mock_init_ascend_soc_version,
151144
mock_init_ascend_config,
152145
mock_register_ascend_customop,
153146
mock_register_atb_extensions,

‎vllm_ascend/distributed/llmdatadist_c_mgr_connector.py‎

Lines changed: 3 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -28,7 +28,7 @@
2828
from vllm.v1.request import Request, RequestStatus
2929

3030
import vllm_ascend.envs as envs_ascend
31-
from vllm_ascend.utils import AscendSocVersion, get_ascend_soc_version
31+
from setup import AscendSocVersion
3232

3333
TORCH_DTYPE_TO_NPU_DTYPE = {
3434
torch.half: llm_datadist.DataType.DT_FLOAT16,
@@ -336,7 +336,8 @@ def __init__(self, vllm_config: VllmConfig):
336336
self.local_agent_metadata.cluster_id)
337337
self.init_llm_datadist()
338338
self.finished_reqs: set[str] = set()
339-
self.soc_info = get_ascend_soc_version()
339+
from vllm_ascend._build_info import __ascend_soc_version__
340+
self.soc_info = __ascend_soc_version__
340341
# Set hccl deterministic for model execute
341342
os.environ["HCCL_DETERMINISTIC"] = "true"
342343
self.done_receiving_counts: defaultdict[str,

‎vllm_ascend/distributed/moe_comm_method.py‎

Lines changed: 3 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -14,7 +14,7 @@
1414
from vllm_ascend.distributed.communication_op import \
1515
data_parallel_reduce_scatter
1616
from vllm_ascend.distributed.parallel_state import get_mc2_group
17-
from vllm_ascend.utils import AscendSocVersion, get_ascend_soc_version
17+
from setup import AscendSocVersion
1818

1919

2020
class MoECommMethod(ABC):
@@ -311,7 +311,8 @@ def __init__(self, moe_config: Optional[FusedMoEConfig]):
311311
# Feature flags
312312
self.enable_dispatch_v2 = hasattr(torch_npu,
313313
"npu_moe_distribute_dispatch_v2")
314-
self.is_ascend_a3 = get_ascend_soc_version() == AscendSocVersion.A3
314+
from vllm_ascend._build_info import __ascend_soc_version__ # type: ignore
315+
self.is_ascend_a3 = __ascend_soc_version__ == AscendSocVersion.A3
315316
self.need_extra_args = self.is_ascend_a3
316317
self._restore_tp_across_dp()
317318

0 commit comments

Comments
 (0)