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

[refact] unified soc_version code (#4359)

Browse Source 
### What this PR does / why we need it?

Currently, there are two paths to judge the chip type in code,
`get_ascend_soc_version` use `get_soc_version` api in torch_npu, and
`is_310p` `use _build_info.__soc_version__`, which generate when
install. We need to unify the two paths.

We need to unify these codes based on the following points:

1. We need to ensure consistency in chip type judgment between compiling
and running states;
2. In compiling state, we need chip type to complete op's compilation,
but in running state, we only need device
type(910B/910_93/310P/910_95/etc) to make code branch judgement;
3. In compiling state, torch_npu may not have been installed yet, so we
can't use torch_npu's api.

Based on the above points, we have made the following changes:

1. When user set env `SOC_VERSION`, use it; when not set, query
soc_version by `npu-smi`;
2. generate device_type based on soc_version when compiling, and write
`__device_type__` instead of `__soc_version__` in `_build_info.py`;
3. In running state, use `__device_type__` to judge code branch.

### Does this PR introduce _any_ user-facing change?

When not set env `SOC_VERSION`, it will not be `ASCEND910B1` by default,
we will query soc_version by `npu-smi`. And env `SOC_VERSION` must be in
the list `soc_to_device` in `setup.py`.

- vLLM version: v0.11.0
- vLLM main:
2918c1b49c

Signed-off-by: zzzzwwjj <1183291235@qq.com>
This commit is contained in:
zzzzwwjj
2025-11-26 14:28:55 +08:00
committed by GitHub
parent a91e76cd84
commit 136ea9ff56
42 changed files with 361 additions and 243 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
tests/ut/ops/test_activation.py
View File

@@ -19,6 +19,8 @@ import pytest
import torch
from vllm.model_executor.layers.activation import QuickGELU, SiluAndMul
from vllm_ascend.utils import AscendDeviceType
@pytest.fixture
def dummy_tensor():
@@ -36,20 +38,22 @@ def test_QuickGELU_forward(mock_gelu, dummy_tensor):
mock_gelu.assert_called_once()
@pytest.mark.parametrize("is_310p_return", [True, False])
@pytest.mark.parametrize("is_310p", [True, False])
@patch("torch_npu.npu_swiglu", side_effect=lambda x: x + 1)
@patch("torch.ops.vllm.maybe_wait_prefetch_done", side_effect=lambda x: None)
@patch("torch.ops.vllm.maybe_prefetch_mlp_down_proj",
side_effect=lambda x: None)
def test_SiluAndMul_forward(mock_maybe_prefetch_mlp_down_proj,
mock_maybe_wait_prefetch_done, mock_swiglu,
is_310p_return, dummy_tensor):
is_310p, dummy_tensor):
with patch("vllm_ascend.utils.is_310p", return_value=is_310p_return):
with patch("vllm_ascend.utils.get_ascend_device_type",
return_value=AscendDeviceType._310P
if is_310p else AscendDeviceType._910_93):
layer = SiluAndMul()
out = layer.forward(dummy_tensor)
if is_310p_return:
if is_310p:
expected_arg = dummy_tensor.to(torch.float32)
else:
expected_arg = dummy_tensor

27
tests/ut/ops/test_fused_moe.py
View File

@@ -29,7 +29,7 @@ from vllm_ascend.ops.fused_moe.fused_moe import (
AscendFusedMoE, AscendUnquantizedFusedMoEMethod)
from vllm_ascend.ops.fused_moe.moe_mlp import (cumsum_group_list,
unified_apply_mlp)
from vllm_ascend.utils import AscendSocVersion, adapt_patch
from vllm_ascend.utils import AscendDeviceType, adapt_patch
adapt_patch(True)
@@ -129,7 +129,7 @@ def mock_dist_env(mocker: MockerFixture):
return_value=mock_forward_context_obj), \
patch('vllm_ascend.ops.fused_moe.prepare_finalize.get_forward_context',
return_value=mock_forward_context_obj), \
patch("vllm_ascend.utils.get_ascend_soc_version", return_value=AscendSocVersion.A3), \
patch("vllm_ascend.utils.get_ascend_device_type", return_value=AscendDeviceType._910_93), \
patch('vllm_ascend.ops.fused_moe.moe_mlp.get_forward_context',
return_value=mock_forward_context_obj), \
patch('vllm_ascend.ops.fused_moe.moe_comm_method.MC2CommImpl._get_token_dispatcher',
@@ -323,22 +323,21 @@ class TestCumsumGroupList(TestBase):
class TestUnifiedApplyMLP(TestBase):
@patch('vllm_ascend.ops.fused_moe.moe_mlp.get_forward_context')
@patch('vllm_ascend.ops.fused_moe.moe_mlp.is_310p')
@patch('vllm_ascend.utils.get_ascend_device_type',
return_value=AscendDeviceType._910_93)
@patch('torch_npu.npu_grouped_matmul')
@patch('torch_npu.npu_dynamic_quant')
@patch('torch_npu.npu_dequant_swiglu_quant')
def test_unified_apply_mlp_with_quantization_mc2(self, mock_npu_dequant,
mock_npu_dynamic_quant,
mock_npu_grouped_matmul,
mock_is_310p,
mock_soc_version,
mock_get_forward_context):
mock_forward_context = MagicMock()
mock_forward_context.moe_comm_type = MoECommType.MC2
mock_get_forward_context.return_value = mock_forward_context
mock_is_310p.return_value = False
mock_npu_dynamic_quant.return_value = (torch.randint(-128,
127, (10, 20),
dtype=torch.int8),
@@ -387,7 +386,8 @@ class TestUnifiedApplyMLP(TestBase):
self.assertEqual(result.dtype, torch.bfloat16)
@patch('vllm_ascend.ops.fused_moe.moe_mlp.is_310p')
@patch('vllm_ascend.utils.get_ascend_device_type',
return_value=AscendDeviceType._910_93)
@patch('torch_npu.npu_grouped_matmul')
@patch('torch_npu.npu_swiglu')
@patch('torch_npu.npu_dynamic_quant')
@@ -395,9 +395,7 @@ class TestUnifiedApplyMLP(TestBase):
mock_npu_dynamic_quant,
mock_npu_swiglu,
mock_npu_grouped_matmul,
mock_is_310p):
mock_is_310p.return_value = False
mock_soc_version):
mock_npu_grouped_matmul.side_effect = [[
torch.randn(10, 40, dtype=torch.float16)
], [torch.randn(10, 20, dtype=torch.float16)]]
@@ -490,15 +488,14 @@ class TestUnifiedApplyMLP(TestBase):
self.assertEqual(result.shape, hidden_states_shape)
self.assertEqual(result.dtype, torch.bfloat16)
@patch('vllm_ascend.ops.fused_moe.moe_mlp.is_310p')
@patch('vllm_ascend.utils.get_ascend_device_type',
return_value=AscendDeviceType._310P)
@patch('torch_npu.npu_grouped_matmul')
@patch('torch_npu.npu_swiglu')
@patch('torch_npu.npu_dynamic_quant')
def test_unified_apply_mlp_without_quantization_310p(
self, mock_npu_dynamic_quant, mock_npu_swiglu,
mock_npu_grouped_matmul, mock_is_310p):
mock_is_310p.return_value = True
mock_npu_grouped_matmul, mock_soc_version):
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],
@@ -527,8 +524,6 @@ class TestUnifiedApplyMLP(TestBase):
topk_scales=topk_scales,
with_quant=False)
mock_is_310p.assert_called_once()
self.assertEqual(mock_npu_grouped_matmul.call_count, 2)
mock_npu_swiglu.assert_called_once()

6
tests/ut/ops/test_layernorm.py
View File

@@ -7,6 +7,7 @@ from vllm.model_executor.layers.layernorm import RMSNorm
from tests.ut.base import PytestBase
from vllm_ascend.quantization.w8a8 import AscendW8A8LinearMethod
from vllm_ascend.utils import AscendDeviceType
def mock_rms_norm(x, weight, eps):
@@ -60,8 +61,9 @@ class TestAscendRMSNorm(PytestBase):
# Test case for addrmsnorm + w8a8 quant fusion
def test_forward_oot_with_quant_fusion(self, mocker: MockerFixture):
mock_is_310p = mocker.patch("vllm_ascend.utils.is_310p")
mock_is_310p.return_value = False
mock_soc_version = mocker.patch(
"vllm_ascend.utils.get_ascend_device_type")
mock_soc_version.return_value = AscendDeviceType._910_93
mock_get_forward_context = mocker.patch(
"vllm_ascend.ops.layernorm.get_forward_context")

8
tests/ut/ops/test_rotary_embedding.py
View File

@@ -12,6 +12,7 @@ from vllm.platforms import CpuArchEnum
from tests.ut.base import TestBase
from vllm_ascend.ascend_forward_context import set_ascend_forward_context
from vllm_ascend.ops.rotary_embedding import _custom_rotary_embedding_enabled
from vllm_ascend.utils import AscendDeviceType
MODEL = "Qwen3-0.6B"
MODEL_VL = "Qwen/Qwen2.5-VL-3B-Instruct"
@@ -97,7 +98,8 @@ class TestAscendRotaryEmbedding(unittest.TestCase):
self.mock_self.is_neox_style = self.is_neox_style
@patch('torch.ops._C_ascend')
@patch('vllm_ascend.ops.rotary_embedding.is_310p', return_value=False)
@patch('vllm_ascend.utils.get_ascend_device_type',
return_value=AscendDeviceType._910_93)
@patch('vllm_ascend.ops.rotary_embedding._custom_rotary_embedding_enabled',
return_value=True)
@patch('torch.ops._npu_rotary_embedding')
@@ -106,8 +108,8 @@ class TestAscendRotaryEmbedding(unittest.TestCase):
@patch('vllm.distributed.parallel_state._DP', MagicMock(world_size=1))
@patch('vllm.distributed.parallel_state._TP', MagicMock(world_size=1))
def test_rope_forward_oot_custom_kernel(self, mock_rotary_embedding,
mock_custom_enabled, mock_is_310p,
mock__c):
mock_custom_enabled,
mock_soc_version, mock__c):
mock_config = MagicMock()
mock_config.torchair_graph_config.enabled = False

8
tests/ut/ops/test_token_dispatcher.py
View File

@@ -22,7 +22,7 @@ import torch
from tests.ut.base import TestBase
from vllm_ascend.ops.fused_moe.token_dispatcher import ( # isort: skip
AscendSocVersion, TokenDispatcherWithAll2AllV,
AscendDeviceType, TokenDispatcherWithAll2AllV,
TokenDispatcherWithAllGather, TokenDispatcherWithMC2)
@@ -50,10 +50,10 @@ class TestTokenDispatcherWithMC2(TestBase):
return_value=self.forward_context)
self.forward_context_patch.start()
# Mock get_ascend_soc_version()
# Mock get_ascend_device_type()
self.ascend_soc_version_patch = patch(
"vllm_ascend.ops.fused_moe.token_dispatcher.get_ascend_soc_version",
return_value=AscendSocVersion.A3)
"vllm_ascend.ops.fused_moe.token_dispatcher.get_ascend_device_type",
return_value=AscendDeviceType._910_93)
self.ascend_soc_version_patch.start()
kwargs = {"with_quant": False, "top_k": 8, "num_experts": 128}