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

drop ascend scheduler (#4498)

Browse Source 
Ascend scheduler was added for non chunk prefill case before, since that
the npu ops didn't work well with chunked prefill.

Now the ops with chunked prefill work better, it's time to remove the
ascend scheduler to use vLLM default scheduler.

- vLLM version: v0.11.2

---------

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
wangxiyuan
2025-11-29 16:18:34 +08:00
committed by GitHub
parent 53a52d6614
commit f10acddb78
52 changed files with 85 additions and 2948 deletions
Download Patch File Download Diff File Expand all files Collapse all files

134
tests/ut/core/test_schedule_config.py
View File

@@ -1,134 +0,0 @@
#
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
#
# 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.
from vllm.config import SchedulerConfig
from tests.ut.base import TestBase
from vllm_ascend.core.schedule_config import AscendSchedulerConfig
class TestAscendSchedulerConfig(TestBase):
def setUp(self):
self.basic_scheduler_config = SchedulerConfig(
max_num_batched_tokens=8192,
max_model_len=8192,
is_multimodal_model=False,
send_delta_data=False,
)
def test_initialize_from_config_with_default(self):
# No additional config given, check the default value here.
ascend_config = AscendSchedulerConfig.initialize_from_config(
self.basic_scheduler_config, {})
self.assertEqual(ascend_config.enable_chunked_prefill, False)
self.assertEqual(ascend_config.policy, "fcfs")
self.assertEqual(ascend_config.scheduler_cls,
"vllm_ascend.core.scheduler.AscendScheduler")
self.assertEqual(ascend_config.max_num_encoder_input_tokens, 8192)
self.assertEqual(ascend_config.encoder_cache_size, 8192)
def test_initialize_from_config_with_override(self):
# test override
ascend_config = AscendSchedulerConfig.initialize_from_config(
self.basic_scheduler_config,
AscendSchedulerConfig(
enable_chunked_prefill=False,
policy="fcfs",
scheduler_cls="vllm_ascend.core.scheduler.AscendScheduler",
max_num_batched_tokens=8192,
max_model_len=2048,
max_long_partial_prefills=1,
long_prefill_token_threshold=512,
),
)
self.assertEqual(ascend_config.enable_chunked_prefill, False)
self.assertEqual(ascend_config.policy, "fcfs")
self.assertEqual(ascend_config.scheduler_cls,
"vllm_ascend.core.scheduler.AscendScheduler")
self.assertEqual(ascend_config.max_num_batched_tokens, 8192)
self.assertEqual(ascend_config.encoder_cache_size, 8192)
self.assertEqual(ascend_config.max_long_partial_prefills, 1)
self.assertEqual(ascend_config.long_prefill_token_threshold, 512)
def test_not_implemented_policy(self):
with self.assertRaises(NotImplementedError) as context:
AscendSchedulerConfig.initialize_from_config(
self.basic_scheduler_config,
AscendSchedulerConfig(
policy="custom_policy",
max_num_batched_tokens=8192,
max_model_len=2048,
),
)
self.assertIn(
"currently AscendScheduler only supports fcfs policy",
str(context.exception),
)
def test_no_override(self):
ascend_config = AscendSchedulerConfig.initialize_from_config(
self.basic_scheduler_config, {})
self.assertEqual(ascend_config.max_num_encoder_input_tokens, 8192)
self.assertEqual(ascend_config.encoder_cache_size, 8192)
def test_valid_config_with_multimodal(self):
config = AscendSchedulerConfig.initialize_from_config(
SchedulerConfig(is_multimodal_model=True,
max_num_batched_tokens=8192), {})
self.assertTrue(config.is_multimodal_model)
def test_valid_config_with_chunked_prefill(self):
ascend_config = AscendSchedulerConfig.initialize_from_config(
self.basic_scheduler_config,
AscendSchedulerConfig(
enable_chunked_prefill=True,
max_num_batched_tokens=8192,
max_model_len=8192,
),
)
self.assertEqual(ascend_config.max_num_batched_tokens, 8192)
self.assertEqual(ascend_config.max_model_len, 8192)
self.assertTrue(ascend_config.enable_chunked_prefill)
def test_invalid_config_without_chunked_prefill(self):
with self.assertRaises(ValueError) as context:
AscendSchedulerConfig.initialize_from_config(
self.basic_scheduler_config,
AscendSchedulerConfig(
enable_chunked_prefill=False,
max_num_batched_tokens=2048,
max_model_len=8192,
),
)
self.assertIn(
"Ascend scheduler is enabled without chunked prefill feature",
str(context.exception),
)
self.assertIn("max_num_batched_tokens (2048)", str(context.exception))
self.assertIn("max_model_len (8192)", str(context.exception))
def test_initialize_from_config_with_pd_transfer(self):
ascend_config = AscendSchedulerConfig.initialize_from_config(
self.basic_scheduler_config,
AscendSchedulerConfig(
enable_pd_transfer=True,
decode_max_num_seqs=48,
max_num_batched_tokens=8192,
max_model_len=4096,
),
)
self.assertEqual(ascend_config.enable_pd_transfer, True)
self.assertEqual(ascend_config.decode_max_num_seqs, 48)

1473
tests/ut/core/test_scheduler.py
View File

File diff suppressed because it is too large Load Diff

1
tests/ut/ops/test_linear.py
View File

@@ -99,7 +99,6 @@ class TestAscendRowParallelLinear(BaseLinearTest):
ascend_config._ASCEND_CONFIG = MagicMock()
ascend_config._ASCEND_CONFIG.oproj_tensor_parallel_size = 2
ascend_config._ASCEND_CONFIG.ascend_scheduler_config.enabled = False
linear = AscendRowParallelLinear(
input_size=16,

7
tests/ut/ops/test_vocab_parallel_embedding.py
View File

@@ -209,12 +209,7 @@ class TestAscendLogitsProcessor(unittest.TestCase):
return_value=torch.randn(1, self.vocab_size)),
patch(
"vllm_ascend.ops.vocab_parallel_embedding.get_lmhead_tp_group.all_gather",
return_value=torch.randn(1, self.vocab_size)),
patch(
"vllm_ascend.core.schedule_config.AscendSchedulerConfig.initialize_from_config",
return_value=MagicMock(max_num_batched_tokens=1000,
max_model_len=512,
enable_chunked_prefill=False))
return_value=torch.randn(1, self.vocab_size))
]
for p in self.patches:

7
tests/ut/quantization/test_w8a8_dynamic.py
View File

@@ -33,13 +33,6 @@ class TestAscendW8A8FusedMoEMethod(TestBase):
mock_get_ep_group.return_value = mock_ep_group
mock_ascend_config = Mock()
# 创建一个具有具体属性的 Mock 对象来表示 ascend_scheduler_config
mock_ascend_scheduler_config = Mock()
mock_ascend_scheduler_config.enabled = False
mock_ascend_scheduler_config.max_num_batched_tokens = 1024
mock_ascend_scheduler_config.max_model_len = 2048
mock_ascend_config.ascend_scheduler_config = mock_ascend_scheduler_config
mock_ascend_config.torchair_graph_config = Mock(enabled=False)
mock_ascend_config.enable_chunked_prefill = False
mock_get_ascend_config.return_value = mock_ascend_config

9
tests/ut/test_ascend_config.py
View File

@@ -56,9 +56,6 @@ class TestAscendConfig(TestBase):
self.assertTrue(torchair_graph_config.enable_frozen_parameter)
self.assertFalse(torchair_graph_config.enable_kv_nz)
ascend_scheduler_config = ascend_config.ascend_scheduler_config
self.assertFalse(ascend_scheduler_config.enabled)
@_clean_up_ascend_config
def test_init_ascend_config_with_additional_config(self):
test_vllm_config = VllmConfig()
@@ -74,9 +71,6 @@ class TestAscendConfig(TestBase):
"enable_kv_nz": True
},
"multistream_overlap_shared_expert": True,
"ascend_scheduler_config": {
"enabled": True
},
"expert_map_path": "test_expert_map_path",
"refresh": True,
}
@@ -94,9 +88,6 @@ class TestAscendConfig(TestBase):
self.assertTrue(torchair_graph_config.enable_frozen_parameter)
self.assertTrue(torchair_graph_config.enable_kv_nz)
ascend_scheduler_config = ascend_config.ascend_scheduler_config
self.assertTrue(ascend_scheduler_config.enabled)
@_clean_up_ascend_config
def test_init_ascend_config_with_refresh(self):
test_vllm_config = VllmConfig()

26
tests/ut/test_platform.py
View File

@@ -32,7 +32,6 @@ class TestNPUPlatform(TestBase):
def mock_vllm_ascend_config():
mock_ascend_config = MagicMock()
mock_ascend_config.torchair_graph_config.enabled = False
mock_ascend_config.ascend_scheduler_config.enabled = False
mock_ascend_config.enable_shared_expert_dp = False
return mock_ascend_config
@@ -522,31 +521,6 @@ class TestNPUPlatform(TestBase):
self.platform.check_and_update_config(vllm_config)
self.assertEqual(vllm_config.compilation_config.custom_ops, [])
@patch('vllm_ascend.utils.get_ascend_device_type',
return_value=AscendDeviceType._910_93)
@patch("vllm_ascend.ascend_config.check_ascend_config")
@patch("vllm_ascend.ascend_config.init_ascend_config")
@patch(
"vllm_ascend.core.recompute_schedule_config.RecomputeSchedulerConfig.initialize_from_config"
)
def test_check_and_update_config_ascend_scheduler_config(
self, mock_init_recompute, mock_init_ascend, mock_check_ascend,
mock_soc_version):
mock_ascend_config = TestNPUPlatform.mock_vllm_ascend_config()
mock_ascend_config.ascend_scheduler_config.enabled = True
mock_init_ascend.return_value = mock_ascend_config
vllm_config = TestNPUPlatform.mock_vllm_config()
vllm_config.parallel_config.tensor_parallel_size = 1
mock_init_recompute.return_value = MagicMock()
with patch("vllm_ascend.core.schedule_config.AscendSchedulerConfig"
) as mock_scheduler:
from vllm_ascend import platform
importlib.reload(platform)
self.platform.check_and_update_config(vllm_config)
mock_scheduler.initialize_from_config.assert_called_once()
@patch('vllm_ascend.platform.get_ascend_config')
def test_get_attn_backend_cls_use_v1_and_mla(self, mock_get_ascend_config):
mock_config = MagicMock()

4
tests/ut/test_utils.py
View File

@@ -253,12 +253,10 @@ class TestUtils(TestBase):
model_path = os.path.join(os.path.dirname(__file__), "fake_weight")
test_model_config = ModelConfig(model=model_path, enforce_eager=True)
test_parallel_config = ParallelConfig()
ascend_config = {"ascend_scheduler_config": {"enabled": False}}
test_vllm_config = VllmConfig(
model_config=test_model_config,
compilation_config=test_compilation_config,
parallel_config=test_parallel_config,
additional_config=ascend_config)
parallel_config=test_parallel_config)
utils.update_aclgraph_sizes(test_vllm_config)
os.environ['HCCL_OP_EXPANSION_MODE'] = 'AIV'
utils.update_aclgraph_sizes(test_vllm_config)

2
tests/ut/torchair/models/test_torchair_deepseek_v2.py
View File

@@ -235,8 +235,6 @@ def test_torchair_deepseek_v2_mlp(mock_distributed, base_config):
hidden_act="silu",
quant_config=None)
assert isinstance(mlp.act_fn, TorchairDeepseekV2SiluAndMul)
ascend_config = MagicMock()
ascend_config._ASCEND_CONFIG.ascend_scheduler_config.enabled = False
with patch(
"vllm_ascend.torchair.models.torchair_deepseek_v2.QuantizationConfig"
) as mock_quant_config: