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

[Feature] model_runner refactor (#4764)

Browse Source 
### What this PR does / why we need it?
refactor npu_modelrunner， we should be close to gpu_modelrunner 

### Does this PR introduce _any_ user-facing change?
NO

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

---------

Signed-off-by: zhenwenqi2024 <zhenwenqi_2022@qq.com>
Signed-off-by: zhenwenqi2024 <155598497+zhenwenqi2024@users.noreply.github.com>
This commit is contained in:
zhenwenqi2024
2025-12-12 17:27:09 +08:00
committed by GitHub
parent 5b12c068f9
commit f708d919f8
10 changed files with 676 additions and 1815 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
tests/ut/worker/test_input_batch.py
View File

@@ -24,6 +24,7 @@ from vllm.utils.torch_utils import make_tensor_with_pad
from vllm.v1.pool.metadata import PoolingMetadata
from vllm.v1.sample.logits_processor import LogitsProcessors
from vllm.v1.sample.metadata import SamplingMetadata
from vllm.v1.utils import CpuGpuBuffer
from vllm_ascend.worker.block_table import BlockTable, MultiGroupBlockTable
from vllm_ascend.worker.npu_input_batch import CachedRequestState, InputBatch
@@ -67,6 +68,8 @@ def _compare_objs(obj1,
is_same = True # if we make it here must be same
elif a == b:
is_same = True
elif isinstance(a, CpuGpuBuffer):
is_same = np.allclose(a.np, b.np) and torch.allclose(a.gpu, b.gpu)
assert is_same, f"Attribute {attr_name} is different"\
f" in {obj1} and {obj2}: {a} != {b}"

113
tests/ut/worker/test_model_runner_v1.py
View File

@@ -1,113 +0,0 @@
# 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.
from unittest.mock import MagicMock, patch
import pytest
from vllm_ascend.ascend_forward_context import MoECommType
from vllm_ascend.utils import AscendDeviceType
from vllm_ascend.worker.model_runner_v1 import NPUModelRunner
# yapf: disable
@pytest.mark.parametrize(
"soc_version, enable_expert_parallel, world_size, pipeline_size, num_tokens, mc2_tokens_capacity, quant_type, expected_method",
[
# Case 1: Expert parallel is disabled, should always be 'allgather'
(AscendDeviceType._910B, False, 8, 2, 100, 256, None, MoECommType.ALLGATHER),
(AscendDeviceType._910_93, False, 16, 2, 500, 256, None, MoECommType.ALLGATHER),
# Case 2: A2 SOC with w4a8_dynamic -> use alltoall when not mc2
(AscendDeviceType._910B, True, 8, 1, 100, 256, "w4a8_dynamic", MoECommType.ALLTOALL),
(AscendDeviceType._910B, True, 16, 1, 257, 256, "w4a8_dynamic", MoECommType.ALLTOALL),
(AscendDeviceType._910B, True, 16, 1, 100, 256, "w4a8_dynamic", MoECommType.MC2), # meets mc2 condition
# Case 3: A2 SOC without w4a8_dynamic -> fallback to allgather
(AscendDeviceType._910B, True, 8, 2, 100, 256, None, MoECommType.ALLGATHER),
(AscendDeviceType._910B, True, 16, 2, 257, 256, None, MoECommType.ALLGATHER),
# Case 4: A3 SOC
(AscendDeviceType._910_93, True, 8, 2, 100, 256, None, MoECommType.MC2),
(AscendDeviceType._910_93, True, 8, 2, 257, 256, None, MoECommType.ALLTOALL),
])
# yapf: enable
def test_select_moe_comm_method(soc_version, enable_expert_parallel,
world_size, pipeline_size, num_tokens,
mc2_tokens_capacity, quant_type,
expected_method):
"""
Tests the _select_moe_comm_method with various configurations including quant_type.
"""
# Mock the NPUModelRunner instance and its dependencies
mock_runner = MagicMock(spec=NPUModelRunner)
mock_runner.parallel_config = MagicMock()
mock_runner.parallel_config.enable_expert_parallel = enable_expert_parallel
mock_runner.parallel_config.world_size_across_dp = world_size
mock_runner.parallel_config.pipeline_parallel_size = pipeline_size
mock_runner.mc2_tokens_capacity = mc2_tokens_capacity
# Add vllm_config.model_config.hf_config mock with moe_quantize
mock_hf_config = MagicMock()
mock_hf_config.moe_quantize = quant_type
mock_model_config = MagicMock()
mock_model_config.hf_config = mock_hf_config
mock_vllm_config = MagicMock()
mock_vllm_config.model_config = mock_model_config
mock_runner.vllm_config = mock_vllm_config
# Patch the helper functions
with patch('vllm_ascend.worker.model_runner_v1.get_ascend_device_type',
return_value=soc_version), \
patch('vllm_ascend.worker.model_runner_v1.is_global_first_rank',
return_value=True), \
patch('vllm_ascend.worker.model_runner_v1.is_moe_model',
return_value=True):
# Bind the real method to the mock object
method = NPUModelRunner._select_moe_comm_method(
mock_runner, num_tokens)
# Assert the result
assert method == expected_method
def test_select_moe_comm_method_unsupported_soc():
"""
Tests that _select_moe_comm_method raises ValueError for an unsupported SOC.
"""
mock_runner = MagicMock(spec=NPUModelRunner)
mock_runner.parallel_config = MagicMock()
mock_runner.parallel_config.enable_expert_parallel = True
mock_runner.mc2_tokens_capacity = 256
# Add vllm_config.model_config.hf_config mock with moe_quantize
mock_hf_config = MagicMock()
mock_hf_config.moe_quantize = None
mock_model_config = MagicMock()
mock_model_config.hf_config = mock_hf_config
mock_vllm_config = MagicMock()
mock_vllm_config.model_config = mock_model_config
mock_runner.vllm_config = mock_vllm_config
unsupported_soc = "UnsupportedSOC"
with patch('vllm_ascend.worker.model_runner_v1.get_ascend_device_type',
return_value=unsupported_soc), \
patch('vllm_ascend.worker.model_runner_v1.is_global_first_rank',
return_value=True), \
patch('vllm_ascend.worker.model_runner_v1.is_moe_model',
return_value=True), \
pytest.raises(ValueError, match=f"Unsupported soc_version: {unsupported_soc}"):
NPUModelRunner._select_moe_comm_method(mock_runner, 100)