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

[Feat] 310p support MoE W8A8 quantizaition (#6641)

Browse Source 
### What this PR does / why we need it?
This PR introduces support for W8A8 dynamic quantization for
Mixture-of-Experts (MoE) models on Ascend 310P devices. This is achieved
by:
- Implementing a new quantization scheme
`AscendW8A8DynamicFusedMoEMethod310`.
- Adding a unified MLP implementation (`unified_apply_mlp`) for 310P
that handles both quantized and unquantized paths.
- Refactoring the MoE and quantization configuration logic to correctly
route to the new 310P-specific implementations.
- Adding new e2e and unit tests to verify the functionality of MoE W8A8
quantization.

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

### How was this patch tested?
- Added a new e2e test `test_qwen3_moe_tp2_w8a8` to test MoE W8A8
quantization in a multi-card setup.
- Added several new unit tests for the 310P-specific MoE components,
including `experts_selector`, `fused_moe`, `moe_comm_method`, `moe_mlp`,
and the new `w8a8_dynamic` quantization method.

- vLLM version: v0.15.0
- vLLM main:
d7e17aaacd

---------

Signed-off-by: pu-zhe <zpuaa@outlook.com>
This commit is contained in:
pu-zhe
2026-02-10 17:17:44 +08:00
committed by GitHub
parent 1eb07986bf
commit 02886e2641
15 changed files with 695 additions and 157 deletions
Download Patch File Download Diff File Expand all files Collapse all files

42
tests/ut/_310p/fused_moe/test_experts_selector_310.py Normal file
View File

@@ -0,0 +1,42 @@
#
# Copyright (c) 2026 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.
import pytest
import torch
from vllm_ascend._310p.fused_moe.experts_selector import select_experts
class TestExpertsSelector310:
@pytest.mark.parametrize("global_num_experts", [256, 128])
def test_select_experts(self, global_num_experts):
x = torch.randn(8, 2)
router_logits = torch.randn(8, 2)
topk_weights, topk_ids = select_experts(
hidden_states=x,
router_logits=router_logits,
top_k=2,
use_grouped_topk=False,
renormalize=True,
topk_group=None,
num_expert_group=None,
custom_routing_function=None,
scoring_func="softmax",
e_score_correction_bias=None,
global_num_experts=global_num_experts,
)
assert topk_weights.shape == (8, 2)
assert topk_ids.shape == (8, 2)

132
tests/ut/_310p/fused_moe/test_moe_mlp_310.py Normal file
View File

@@ -0,0 +1,132 @@
#
# Copyright (c) 2026 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 unittest.mock import call, patch
import torch
from tests.ut.base import TestBase
from vllm_ascend._310p.fused_moe.moe_mlp import unified_apply_mlp
class TestUnifiedApplyMLP310(TestBase):
@patch("torch_npu.npu_grouped_matmul")
@patch("torch_npu.npu_swiglu")
def test_unified_apply_mlp_without_quantization_310(self, mock_npu_swiglu, mock_npu_grouped_matmul):
mock_gmm1_out = torch.randn(10, 40, dtype=torch.float16)
mock_gmm2_out = torch.randn(10, 20, dtype=torch.float16)
mock_npu_grouped_matmul.side_effect = [[mock_gmm1_out], [mock_gmm2_out]]
mock_npu_swiglu_output = torch.randn(10, 40, dtype=torch.float16)
mock_npu_swiglu.return_value = mock_npu_swiglu_output
hidden_states = torch.randn(10, 20, dtype=torch.float16)
w1 = torch.randn(5, 20, 40, dtype=torch.float16)
w2 = torch.randn(5, 40, 20, dtype=torch.float16)
group_list = torch.tensor([2, 4, 6, 8, 10], dtype=torch.int64)
result = unified_apply_mlp(
hidden_states=hidden_states,
w1=w1,
w1_scale=None,
w2=w2,
w2_scale=None,
group_list=group_list,
group_list_type=1,
with_quant=False,
)
self.assertEqual(mock_npu_grouped_matmul.call_count, 2)
mock_npu_grouped_matmul.assert_has_calls(
[
call(
x=[hidden_states], weight=[w1], split_item=2, group_list_type=1, group_type=0, group_list=group_list
),
call(
x=[mock_npu_swiglu_output],
weight=[w2],
split_item=2,
group_list_type=1,
group_type=0,
group_list=group_list,
),
],
any_order=True,
)
mock_npu_swiglu.assert_called_once()
mock_npu_swiglu.assert_called_with(mock_gmm1_out)
self.assertEqual(result.shape, hidden_states.shape)
self.assertEqual(result.dtype, torch.float16)
@patch("torch.cumsum")
@patch("torch_npu.npu_quant_grouped_matmul_dequant")
@patch("torch_npu.npu_swiglu")
def test_unified_apply_mlp_with_quantization_310(
self, mock_npu_swiglu, mock_npu_quant_grouped_matmul_dequant, mock_cumsum
):
mock_cumsum_out = torch.arange(0, 10, dtype=torch.int64)
mock_cumsum.return_value = mock_cumsum_out
mock_gmm1_out = torch.randn(10, 40, dtype=torch.float16)
mock_gmm2_out = torch.randn(10, 20, dtype=torch.float16)
mock_npu_quant_grouped_matmul_dequant.side_effect = [mock_gmm1_out, mock_gmm2_out]
mock_npu_swiglu_output = torch.randn(10, 40, dtype=torch.float16)
mock_npu_swiglu.return_value = mock_npu_swiglu_output
hidden_states = torch.randn(10, 20, dtype=torch.float16)
w1 = torch.randn(5, 20, 40, dtype=torch.float16)
w1_scale = torch.rand(5, 40, dtype=torch.float32)
w2 = torch.randn(5, 40, 20, dtype=torch.float16)
w2_scale = torch.rand(5, 40, dtype=torch.float32)
group_list = torch.tensor([2, 4, 6, 8, 10], dtype=torch.int64)
result = unified_apply_mlp(
hidden_states=hidden_states,
w1=w1,
w1_scale=w1_scale,
w2=w2,
w2_scale=w2_scale,
group_list=group_list,
group_list_type=1,
with_quant=True,
)
mock_cumsum.assert_called_once()
self.assertEqual(mock_npu_quant_grouped_matmul_dequant.call_count, 2)
mock_npu_quant_grouped_matmul_dequant.assert_has_calls(
[
call(
x=hidden_states,
quantized_weight=w1,
weight_scale=w1_scale,
group_list=mock_cumsum_out,
quant_mode="pertoken",
),
call(
x=mock_npu_swiglu_output,
quantized_weight=w2,
weight_scale=w2_scale,
group_list=mock_cumsum_out,
quant_mode="pertoken",
),
],
any_order=True,
)
mock_npu_swiglu.assert_called_once()
mock_npu_swiglu.assert_called_with(mock_gmm1_out)
self.assertEqual(result.shape, hidden_states.shape)
self.assertEqual(result.dtype, torch.float16)

47
tests/ut/_310p/quantization/test_modelslim_config_310.py
View File

@@ -1,10 +1,26 @@
#
# Copyright (c) 2026 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 unittest.mock import MagicMock, patch
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.fused_moe.config import FusedMoEConfig
from vllm.model_executor.layers.fused_moe.config import FusedMoEConfig, FusedMoEParallelConfig
from vllm.model_executor.layers.linear import LinearBase
from tests.ut.base import TestBase
from vllm_ascend._310p.fused_moe.fused_moe import AscendUnquantizedFusedMoEMethod310
from vllm_ascend._310p.quantization.modelslim_config import AscendModelSlimConfig310
from vllm_ascend.ops.linear import AscendUnquantizedLinearMethod
@@ -31,7 +47,7 @@ class TestAscendModelSlimConfig310(TestBase):
# Test skipped layer
with (
patch("vllm_ascend._310p.quantization.modelslim_config.get_current_vllm_config", return_value=mock_config),
patch.object(self.ascend_config, "is_layer_skipped_ascend", return_value=True)
patch.object(self.ascend_config, "is_layer_skipped_ascend", return_value=True),
):
method = self.ascend_config.get_quant_method(linear_layer, ".attn")
self.assertIsInstance(method, AscendUnquantizedLinearMethod)
@@ -54,14 +70,35 @@ class TestAscendModelSlimConfig310(TestBase):
fused_moe_layer = MagicMock(spec=FusedMoE)
fused_moe_layer.moe = MagicMock(spec=FusedMoEConfig)
fused_moe_layer.moe_config = MagicMock(spec=FusedMoEConfig)
fused_moe_layer.moe_config.moe_parallel_config = MagicMock(spec=FusedMoEParallelConfig)
fused_moe_layer.moe_config.moe_parallel_config.use_ep = True
fused_moe_layer.moe_config.moe_parallel_config.dp_size = 1
mock_config = MagicMock()
mock_config.model_config.hf_config.model_type = None
mock_config.compilation_config.custom_ops = ["all"]
mock_scheme = MagicMock()
# Test skipped layer
with (
patch("vllm.config.vllm.get_current_vllm_config", return_value=mock_config),
patch("vllm_ascend._310p.quantization.modelslim_config.get_current_vllm_config", return_value=mock_config),
patch("vllm_ascend.quantization.modelslim_config.get_current_vllm_config", return_value=mock_config),
patch.object(self.ascend_config, "is_layer_skipped_ascend", return_value=True),
):
method = self.ascend_config.get_quant_method(fused_moe_layer, ".moe")
self.assertIsInstance(method, AscendUnquantizedFusedMoEMethod310)
# Test quantized layer
mock_scheme = MagicMock()
with (
patch.object(self.ascend_config, "is_layer_skipped_ascend", return_value=False),
patch("vllm.config.vllm.get_current_vllm_config", return_value=mock_config),
patch("vllm_ascend._310p.quantization.modelslim_config.get_current_vllm_config", return_value=mock_config),
patch("vllm_ascend.quantization.modelslim_config.get_current_vllm_config", return_value=mock_config),
patch("vllm_ascend._310p.quantization.modelslim_config.create_scheme_for_layer", return_value=mock_scheme),
patch("vllm_ascend._310p.quantization.modelslim_config.AscendLinearMethod", return_value=MagicMock()),
self.assertRaises(NotImplementedError),
patch(
"vllm_ascend._310p.quantization.modelslim_config.AscendFusedMoEMethod", return_value=MagicMock()
) as fused_moe_method,
):
self.ascend_config.get_quant_method(fused_moe_layer, "moe_layer")
method = self.ascend_config.get_quant_method(fused_moe_layer, ".moe")
self.assertIs(method, fused_moe_method.return_value)
fused_moe_method.assert_called_once_with(mock_scheme, fused_moe_layer.moe_config)

66
tests/ut/_310p/quantization/test_w8a8_dynamic_310.py Normal file
View File

@@ -0,0 +1,66 @@
#
# Copyright (c) 2026 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 unittest.mock import Mock, patch
import torch
from tests.ut.base import TestBase
from vllm_ascend._310p.quantization.methods.w8a8_dynamic import AscendW8A8DynamicFusedMoEMethod310
class TestAscendW8A8FusedMoEMethod310(TestBase):
num_experts = 8
hidden_size = 128
intermediate_size = 128
@patch("vllm_ascend._310p.quantization.methods.w8a8_dynamic.get_ep_group")
def setUp(self, mock_get_ep_group):
with patch(
"vllm_ascend._310p.quantization.methods.w8a8_dynamic.get_current_vllm_config"
) as mock_get_current_vllm_config:
mock_vllm_config = Mock()
mock_vllm_config.quant_config = Mock(quant_description={"group_size": 0})
mock_vllm_config.scheduler_config = Mock(
max_num_batched_tokens=2048, max_model_len=2048, enable_chunked_prefill=False
)
mock_get_current_vllm_config.return_value = mock_vllm_config
mock_ep_group = Mock()
mock_get_ep_group.return_value = mock_ep_group
mock_ascend_config = Mock()
mock_ascend_config.enable_chunked_prefill = False
self.quant_method = AscendW8A8DynamicFusedMoEMethod310()
def test_get_weight_310(self):
param_dict = self.quant_method.get_weight(
self.num_experts, self.intermediate_size, self.hidden_size, torch.float16
)
self.assertEqual(param_dict["w13_weight"].dtype, torch.int8)
self.assertEqual(
param_dict["w13_weight"].shape, (self.num_experts, 2 * self.intermediate_size, self.hidden_size)
)
self.assertEqual(param_dict["w2_weight"].dtype, torch.int8)
self.assertEqual(param_dict["w2_weight"].shape, (self.num_experts, self.hidden_size, self.intermediate_size))
def test_get_dynamic_quant_param_310(self):
param_dict = self.quant_method.get_dynamic_quant_param(
self.num_experts, self.intermediate_size, self.hidden_size, torch.float16
)
self.assertEqual(param_dict["w13_weight_scale"].dtype, torch.float32)
self.assertEqual(param_dict["w13_weight_scale"].shape, (self.num_experts, 2 * self.intermediate_size, 1))
self.assertEqual(param_dict["w2_weight_scale"].dtype, torch.float32)
self.assertEqual(param_dict["w2_weight_scale"].shape, (self.num_experts, self.hidden_size, 1))

27
tests/ut/_310p/quantization/test_w8a8_310.py → tests/ut/_310p/quantization/test_w8a8_static_310.py
View File

@@ -1,3 +1,18 @@
#
# Copyright (c) 2026 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 unittest.mock import MagicMock, patch
import torch
@@ -16,19 +31,19 @@ class TestAscendW8A8LinearMethod310(TestBase):
self.assertEqual(weight["weight"].shape, (20, 10))
def test_get_pertensor_param_310(self):
params = self.method.get_pertensor_param(torch.bfloat16)
self.assertEqual(params["input_scale"].dtype, torch.bfloat16)
params = self.method.get_pertensor_param(torch.float16)
self.assertEqual(params["input_scale"].dtype, torch.float16)
self.assertEqual(params["input_offset"].dtype, torch.int8)
self.assertEqual(params["input_scale"].shape, (1,))
self.assertEqual(params["input_offset"].shape, (1,))
def test_get_perchannel_param_310(self):
params = self.method.get_perchannel_param(10, torch.bfloat16)
params = self.method.get_perchannel_param(10, torch.float16)
self.assertEqual(params["quant_bias"].dtype, torch.int32)
self.assertEqual(params["deq_scale"].dtype, torch.float32)
self.assertEqual(params["weight_scale"].dtype, torch.bfloat16)
self.assertEqual(params["weight_offset"].dtype, torch.bfloat16)
self.assertEqual(params["deq_scale"].dtype, torch.int64)
self.assertEqual(params["weight_scale"].dtype, torch.float16)
self.assertEqual(params["weight_offset"].dtype, torch.float16)
self.assertEqual(params["quant_bias"].shape, (10,))
self.assertEqual(params["deq_scale"].shape, (10,))
self.assertEqual(params["weight_scale"].shape, (10, 1))