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

Refactor tensor_parallel and comm_utils (#2814)

Browse Source 
### What this PR does / why we need it?
1. Move ops/comm_utils to ops/moe/comm_utils
2. Move distributed/tensor_parallel/gather_from_sequence_parallel_region
to ops/moe/comm_utils
3. Delete distributed/tensor_parallel

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

### How was this patch tested?
e2e & ut

- vLLM version: main
- vLLM main:
a1213fae5f

---------

Signed-off-by: wuweiqiang24 <1005334931@qq.com>
Signed-off-by: wuweiqiang24 <wuweiqiang11@huawei.com>
This commit is contained in:
wuweiqiang24
2025-09-11 21:26:36 +08:00
committed by GitHub
parent 0005479b9c
commit 9615dea3a7
6 changed files with 153 additions and 392 deletions
Download Patch File Download Diff File Expand all files Collapse all files

139
tests/ut/distributed/test_distributed_tensor_parallel.py
View File

@@ -1,139 +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 importlib
import pytest
import torch
from pytest_mock import MockerFixture
from tests.ut.base import PytestBase
from vllm_ascend.distributed.tensor_parallel import (
_gather_along_first_dim, _gather_along_last_dim,
_reduce_scatter_along_first_dim, _reduce_scatter_along_last_dim,
all_to_all_hp2sp, all_to_all_sp2hp)
class TestDistributedCommunication(PytestBase):
@pytest.fixture(autouse=True)
def context(self, mocker: MockerFixture):
mocker.patch("torch.npu.current_device", return_value="cpu")
mocker.patch("torch.distributed.get_world_size", return_value=4)
mocker.patch("torch.distributed.get_rank", return_value=0)
@pytest.mark.parametrize("world_size, test_tensor, expected",
[(1, torch.randn(8, 16), (8, 16)),
(4, torch.randn(8, 16), (32, 16))])
def test_gather_along_first_dim(self, test_tensor, expected, world_size,
mocker: MockerFixture):
"""test _gather_along_first_dim"""
mocker.patch("torch.distributed.get_world_size",
return_value=world_size)
result = _gather_along_first_dim(test_tensor, mocker.MagicMock())
assert result.shape == expected
@pytest.mark.parametrize("test_tensor, output_split_sizes, expected", [
(torch.randn(8, 16), [5, 10, 15, 2], (32, 16)),
])
def test_gather_along_first_dim_unequal_split(self, test_tensor, expected,
output_split_sizes,
mocker: MockerFixture):
"""test _gather_along_first_dim"""
result = _gather_along_first_dim(test_tensor, mocker.MagicMock(),
output_split_sizes)
assert result.shape == expected
@pytest.mark.parametrize("world_size, test_tensor, expected",
[(1, torch.randn(8, 16, 32), (8, 16, 32)),
(4, torch.randn(8, 16, 32), (8, 16, 32 * 4))])
def test_gather_along_last_dim(self, test_tensor, expected, world_size,
mocker: MockerFixture):
"""test _gather_along_last_dim"""
mocker.patch("torch.distributed.get_world_size",
return_value=world_size)
result = _gather_along_last_dim(test_tensor, mocker.MagicMock())
assert result.shape == expected
@pytest.mark.parametrize("input_shape,expected_shape", [
((32, 16), (8, 16)),
((40, 10), (10, 10)),
])
def test_reduce_scatter_along_first_dim(self, input_shape, expected_shape,
mocker: MockerFixture):
input_tensor = torch.randn(*input_shape)
result = _reduce_scatter_along_first_dim(input_tensor,
mocker.MagicMock())
assert result.shape == expected_shape
@pytest.mark.parametrize("input_shape,expected_shape", [
((8, 16, 32), (8, 16, 8)),
])
def test_reduce_scatter_along_last_dim(self, input_shape, expected_shape,
mocker: MockerFixture):
input_tensor = torch.randn(*input_shape)
result = _reduce_scatter_along_last_dim(input_tensor,
mocker.MagicMock())
assert result.shape == expected_shape
@pytest.mark.parametrize("func,input_shape,expected_shape", [
("all_gather_last_dim_from_tensor_parallel_region", (8, 16, 32),
(8, 16, 128)),
("reduce_scatter_to_sequence_parallel_region", (32, 16), (8, 16)),
("reduce_scatter_last_dim_to_tensor_parallel_region", (8, 16, 32),
(8, 16, 8)),
("gather_from_sequence_parallel_region", (8, 16), (32, 16)),
])
def test_wrapper_functions(self, func, input_shape, expected_shape,
mocker: MockerFixture):
"""test wrapper funcs"""
mod = importlib.import_module(
'vllm_ascend.distributed.tensor_parallel')
globals = mod.__dict__
test_func = globals[func]
input_tensor = torch.randn(*input_shape)
result = test_func(input_tensor, mocker.MagicMock())
assert result.shape == expected_shape
@pytest.mark.parametrize(
"input_shape,output_shape",
[
((8, 16), (32, 4)), # [num_tokens/TP, H] -> [num_tokens, H/TP]
])
def test_all_to_all_sp2hp(self, input_shape, output_shape,
mocker: MockerFixture):
input_tensor = torch.randn(*input_shape)
result = all_to_all_sp2hp(input_tensor, mocker.MagicMock())
assert result.shape == output_shape
@pytest.mark.parametrize(
"input_shape,output_shape",
[
((32, 4), (8, 16)), # [num_tokens, H/TP] -> [num_tokens/TP, H]
])
def test_all_to_all_hp2sp(self, input_shape, output_shape,
mocker: MockerFixture):
input_tensor = torch.randn(*input_shape)
result = all_to_all_hp2sp(input_tensor, mocker.MagicMock())
assert result.shape == output_shape

98
tests/ut/ops/test_comm_utils.py Normal file
View File

@@ -0,0 +1,98 @@
#
# 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 pytest_mock import MockerFixture
from tests.ut.base import PytestBase
from vllm_ascend.ops.moe.comm_utils import (
_gather_along_first_dim, async_all_to_all,
gather_from_sequence_parallel_region)
class TestDistributedCommunication(PytestBase):
@pytest.fixture(autouse=True)
def context(self, mocker: MockerFixture):
mocker.patch("torch.npu.current_device", return_value="cpu")
mocker.patch("torch.distributed.get_world_size", return_value=4)
mocker.patch("torch.distributed.get_rank", return_value=0)
@pytest.mark.parametrize(
"input_tensor, output_split_sizes, input_split_sizes",
[(torch.randn(8, 16), [2, 2, 2, 2], [2, 2, 2, 2]),
(torch.randn(16, 32), None, None)])
def test_async_all_to_all(self, input_tensor, output_split_sizes,
input_split_sizes, mocker: MockerFixture):
"""Test async_all_to_all"""
mock_group = mocker.MagicMock()
mocker.patch("torch.distributed.all_to_all_single",
return_value=mocker.MagicMock())
_, a2a_out, handle = async_all_to_all(input_tensor, output_split_sizes,
input_split_sizes, mock_group)
# Check if the output tensor is created properly
if output_split_sizes is None:
assert a2a_out.shape == input_tensor.shape
else:
total_output_size = sum(output_split_sizes)
expected_shape = [total_output_size] + list(
input_tensor.size())[1:]
assert a2a_out.shape == torch.Size(expected_shape)
# Ensure handle is returned from async operation
assert handle is not None
assert isinstance(handle, mocker.MagicMock)
@pytest.mark.parametrize("world_size, test_tensor, expected",
[(1, torch.randn(8, 16), (8, 16)),
(4, torch.randn(8, 16), (32, 16))])
def test_gather_along_first_dim(self, test_tensor, expected, world_size,
mocker: MockerFixture):
"""Test _gather_along_first_dim"""
mocker.patch("torch.distributed.get_world_size",
return_value=world_size)
result = _gather_along_first_dim(test_tensor, mocker.MagicMock())
assert result.shape == expected
@pytest.mark.parametrize("input_tensor, output_split_sizes",
[(torch.randn(8, 16), None),
(torch.randn(8, 16), [2, 2, 2, 2])])
def test_gather_from_sequence_parallel_region(self, input_tensor,
output_split_sizes,
mocker: MockerFixture):
"""Test gather_from_sequence_parallel_region"""
mock_group = mocker.MagicMock()
result = gather_from_sequence_parallel_region(input_tensor, mock_group,
output_split_sizes)
# If output_split_sizes is not provided, result should have expanded first dimension by world size
if output_split_sizes is None:
expected_shape = [input_tensor.shape[0] * 4] + list(
input_tensor.shape[1:])
assert result.shape == torch.Size(expected_shape)
else:
# If output_split_sizes is provided, result shape is dictated by sum of output_split_sizes
expected_shape = [sum(output_split_sizes)] + list(
input_tensor.shape[1:])
assert result.shape == torch.Size(expected_shape)

2
tests/ut/ops/test_token_dispatcher.py
View File

@@ -348,7 +348,7 @@ class TestTokenDispatcherWithAll2AllV(TestBase):
self.mock_npu_moe_token_unpermute.return_value = torch.randn(8, 16)
# Mock async_all_to_all
patcher6 = patch('vllm_ascend.ops.comm_utils.async_all_to_all')
patcher6 = patch('vllm_ascend.ops.moe.comm_utils.async_all_to_all')
self.mock_async_all_to_all = patcher6.start()
self.addCleanup(patcher6.stop)
self.mock_async_all_to_all.return_value = (None, torch.randn(16, 16),