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

[refactor] refactor deepseek-related files (#2849)

Browse Source 
### What this PR does / why we need it?
This PR deletes ~2K lines of code about deepseek modeling. It falls back
CustomDeepseekV2 modules to original vllm implementations and adapts
some modifications in vllm about deepseek and moe.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
E2E  vllm serving with torchair graph mode and eager mode.

- vLLM version: v0.10.2
- vLLM main:
759ef49b15

---------

Signed-off-by: linfeng-yuan <1102311262@qq.com>
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
Co-authored-by: yiz-liu <136800916+yiz-liu@users.noreply.github.com>
Co-authored-by: Yizhou Liu <liu_yizhou@outlook.com>
This commit is contained in:
linfeng-yuan
2025-09-16 14:13:07 +08:00
committed by GitHub
parent 18ca7861f6
commit 1c5900327b
18 changed files with 295 additions and 1899 deletions
Download Patch File Download Diff File Expand all files Collapse all files

114
tests/ut/models/conftest.py Normal file
View File

@@ -0,0 +1,114 @@
from types import SimpleNamespace
from unittest.mock import MagicMock, Mock, patch
import pytest
import torch
from transformers import PretrainedConfig
from vllm.config import CacheConfig, EPLBConfig, ParallelConfig
from vllm.distributed.parallel_state import GroupCoordinator
@pytest.fixture
def base_config():
config = PretrainedConfig(
hidden_size=128,
num_attention_heads=8,
num_hidden_layers=2,
intermediate_size=256,
hidden_act="silu",
rms_norm_eps=1e-6,
rope_theta=10000.0,
max_position_embeddings=2048,
n_routed_experts=4,
n_shared_experts=1,
moe_intermediate_size=256,
num_experts_per_tok=2,
routed_scaling_factor=1.0,
first_k_dense_replace=0,
moe_layer_freq=1,
kv_lora_rank=16,
qk_nope_head_dim=16,
qk_rope_head_dim=16,
v_head_dim=32,
topk_method="noaux_tc",
scoring_func="softmax",
norm_topk_prob=True,
n_group=1,
topk_group=1,
vocab_size=10000,
)
return config
@pytest.fixture
def vllm_config(base_config):
model_config = SimpleNamespace(
hf_config=base_config,
tensor_parallel_size=1,
dtype=torch.float32,
use_mla=True,
quant_config=None,
max_model_len=2048,
)
parallel_config = MagicMock(spec=ParallelConfig)
eplb_config = MagicMock(spec=EPLBConfig)
eplb_config.num_redundant_experts = 0
parallel_config.eplb_config = eplb_config
cache_config = CacheConfig()
vllm_config = Mock()
vllm_config.model_config = model_config
vllm_config.cache_config = cache_config
vllm_config.quant_config = None
vllm_config.parallel_config = parallel_config
return vllm_config
@pytest.fixture
def mock_distributed():
tp_group = Mock(spec=GroupCoordinator)
tp_group.rank_in_group = 0
tp_group.world_size = 1
tp_group.device_group = Mock()
dp_group = Mock(spec=GroupCoordinator)
dp_group.rank_in_group = 0
dp_group.world_size = 1
ep_group = Mock(spec=GroupCoordinator)
ep_group.rank_in_group = 0
ep_group.world_size = 1
ep_group.device_group = Mock()
ep_group.device_group.rank.return_value = 0
ep_group.device_group.size.return_value = 1
pp_group = Mock(spec=GroupCoordinator)
pp_group.rank_in_group = 0
pp_group.world_size = 1
mock_vllm_config = Mock()
mock_vllm_config.scheduler_config = Mock(max_num_seqs=256)
mock_vllm_config.model_config = Mock(max_model_len=2048, quant_config=None)
with patch("vllm_ascend.models.deepseek_v2.get_tensor_model_parallel_rank", return_value=0), \
patch("vllm_ascend.models.deepseek_v2.get_tensor_model_parallel_world_size", return_value=1), \
patch("vllm_ascend.models.deepseek_v2.get_tp_group", return_value=tp_group), \
patch("vllm_ascend.models.deepseek_v2.get_pp_group", return_value=pp_group), \
patch("vllm_ascend.models.deepseek_v2.get_pp_group",
return_value=Mock(is_first_rank=False, is_last_rank=False)), \
patch("vllm_ascend.ops.fused_moe.get_current_vllm_config", return_value=mock_vllm_config), \
patch("vllm_ascend.ops.moe.token_dispatcher.torch.distributed.get_rank", return_value=0), \
patch("vllm_ascend.ops.moe.token_dispatcher.get_ascend_soc_version", return_value=None), \
patch.dict("vllm.distributed.parallel_state.__dict__", _TP=tp_group, _EP=ep_group, _DP=dp_group,
_PP=pp_group), \
patch.dict("vllm_ascend.distributed.parallel_state.__dict__", _MC2=ep_group), \
patch("torch.npu.current_device", return_value=0):
yield
@pytest.fixture
def mock_forward_context():
forward_context = Mock(in_profile_run=False, with_prefill=False)
with patch("vllm_ascend.models.deepseek_v2.get_forward_context",
return_value=forward_context):
yield

11
tests/ut/models/test_deepseek_mtp.py
View File

@@ -13,10 +13,13 @@ from vllm_ascend.models.deepseek_mtp import (
class TestCustomDeepSeekMultiTokenPredictorLayer(PytestBase):
@pytest.fixture
def setup_mtp_layer(self, mocker: MockerFixture):
def setup_mtp_layer(self, mocker: MockerFixture, vllm_config: VllmConfig,
mock_distributed):
config = PretrainedConfig(vocab_size=1000,
hidden_size=768,
rms_norm_eps=1e-5)
mocker.patch("vllm_ascend.models.deepseek_mtp.get_current_vllm_config",
return_value=vllm_config)
mocker.patch(
"vllm.model_executor.layers.vocab_parallel_embedding.VocabParallelEmbedding.__init__",
return_value=None)
@@ -29,15 +32,15 @@ class TestCustomDeepSeekMultiTokenPredictorLayer(PytestBase):
"vllm_ascend.models.deepseek_mtp.CustomDeepSeekShareHead.__init__",
return_value=None)
mocker_deepseek_v2_decode_layer = mocker.patch(
"vllm_ascend.models.deepseek_v2.CustomDeepseekV2DecoderLayer.__init__",
"vllm.model_executor.models.deepseek_v2.DeepseekV2DecoderLayer.__init__",
return_value=None)
mocker.patch(
"vllm_ascend.ops.vocab_parallel_embedding.AscendVocabParallelEmbedding.__init__",
return_value=None)
mocker.patch("vllm_ascend.utils.get_ascend_config",
mocker.patch("vllm_ascend.models.deepseek_v2.get_ascend_config",
return_value=mocker.Mock())
mtp_layer = CustomDeepSeekMultiTokenPredictorLayer(config, "", None)
mtp_layer = CustomDeepSeekMultiTokenPredictorLayer(config, "0", None)
mocker_deepseek_v2_decode_layer.assert_called_once()
return mtp_layer

181
tests/ut/models/test_deepseek_v2.py
View File

@@ -12,163 +12,19 @@
# limitations under the License.
# This file is a part of the vllm-ascend project.
#
from types import SimpleNamespace
from unittest.mock import Mock, patch
import pytest
import torch
from transformers import PretrainedConfig
from vllm.config import CacheConfig
from vllm.distributed.parallel_state import GroupCoordinator
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
from vllm_ascend.models.deepseek_v2 import (
CustomDeepseekV2MergedReplicatedLinear, CustomDeepseekV2MLAAttention,
CustomDeepseekV2MLP, CustomDeepseekV2RowParallelLinear,
CustomDeepseekV2RowParallelLinearReplaceAllreduce,
CustomDeepseekV2SiluAndMul, LogitsProcessor, ParallelLMHead)
from vllm_ascend.models.deepseek_v2 import (CustomDeepseekV2MLAAttention,
CustomDeepseekV2RowParallelLinear)
@pytest.fixture
def base_config():
config = PretrainedConfig(
hidden_size=128,
num_attention_heads=8,
num_hidden_layers=2,
intermediate_size=256,
hidden_act="silu",
rms_norm_eps=1e-6,
rope_theta=10000.0,
max_position_embeddings=2048,
n_routed_experts=4,
n_shared_experts=1,
moe_intermediate_size=256,
num_experts_per_tok=2,
routed_scaling_factor=1.0,
first_k_dense_replace=0,
moe_layer_freq=1,
kv_lora_rank=16,
qk_nope_head_dim=16,
qk_rope_head_dim=16,
v_head_dim=32,
topk_method="noaux_tc",
scoring_func="softmax",
norm_topk_prob=True,
n_group=1,
topk_group=1,
vocab_size=10000,
)
return config
@pytest.fixture
def vllm_config(base_config):
model_config = SimpleNamespace(
hf_config=base_config,
tensor_parallel_size=1,
dtype=torch.float32,
use_mla=False,
quant_config=None,
max_model_len=2048,
)
cache_config = CacheConfig()
vllm_config = Mock()
vllm_config.model_config = model_config
vllm_config.cache_config = cache_config
vllm_config.quant_config = None
return vllm_config
@pytest.fixture
def mock_distributed():
tp_group = Mock(spec=GroupCoordinator)
tp_group.rank_in_group = 0
tp_group.world_size = 1
tp_group.device_group = Mock()
dp_group = Mock(spec=GroupCoordinator)
dp_group.rank_in_group = 0
dp_group.world_size = 1
ep_group = Mock(spec=GroupCoordinator)
ep_group.rank_in_group = 0
ep_group.world_size = 1
pp_group = Mock(spec=GroupCoordinator)
pp_group.rank_in_group = 0
pp_group.world_size = 1
mock_vllm_config = Mock()
mock_vllm_config.scheduler_config = Mock(max_num_seqs=256)
mock_vllm_config.model_config = Mock(max_model_len=2048, quant_config=None)
with patch("vllm_ascend.models.deepseek_v2.get_tensor_model_parallel_rank", return_value=0), \
patch("vllm_ascend.models.deepseek_v2.get_tensor_model_parallel_world_size", return_value=1), \
patch("vllm_ascend.models.deepseek_v2.get_tp_group", return_value=tp_group), \
patch("vllm_ascend.models.deepseek_v2.get_ep_group", return_value=ep_group), \
patch("vllm_ascend.models.deepseek_v2.get_dp_group", return_value=dp_group), \
patch("vllm_ascend.models.deepseek_v2.get_pp_group", return_value=pp_group), \
patch("vllm_ascend.models.deepseek_v2.get_pp_group",
return_value=Mock(is_first_rank=False, is_last_rank=False)), \
patch("vllm_ascend.ops.fused_moe.get_current_vllm_config", return_value=mock_vllm_config), \
patch.dict("vllm.distributed.parallel_state.__dict__", _TP=tp_group, _EP=ep_group, _DP=dp_group,
_PP=pp_group), \
patch.dict("vllm_ascend.distributed.parallel_state.__dict__", _MC2=ep_group), \
patch("torch.npu.current_device", return_value=0):
yield
@pytest.fixture
def mock_forward_context():
forward_context = Mock(in_profile_run=False, with_prefill=False)
with patch("vllm_ascend.models.deepseek_v2.get_forward_context",
return_value=forward_context):
yield
def test_custom_deepseek_v2_silu_and_mul():
torch.set_default_device("cpu")
silu = CustomDeepseekV2SiluAndMul()
assert silu.weight_scale is None
x = torch.randn(2, 4)
output = silu.forward_oot(x)
assert output.shape == (2, 2)
weight_scale = Mock(return_value=torch.tensor(0.1))
silu = CustomDeepseekV2SiluAndMul(weight_scale=weight_scale)
quant_x = torch.randint(-128, 127, (2, 4), dtype=torch.int32)
dynamic_scale = torch.randn(2, 1)
with patch("torch_npu.npu_dequant_swiglu_quant",
return_value=torch.randn(2, 4)):
output = silu.forward_oot((quant_x, dynamic_scale))
assert output.shape == (2, 4)
def test_custom_deepseek_v2_merged_replicated_linear(mock_distributed):
linear = CustomDeepseekV2MergedReplicatedLinear(input_size=128,
output_sizes=[64, 64],
bias=False,
quant_config=None)
assert linear.output_sizes == [64, 64]
param = Mock()
param.data = torch.zeros(128, 128)
param.output_dim = 1
param.is_gguf_weight = False
param.is_gguf_weight_type = False
loaded_weight = torch.randn(128, 64)
linear.weight_loader(param, loaded_weight, loaded_shard_id=0)
with pytest.raises(AssertionError):
linear.weight_loader(param, torch.randn(128, 32), loaded_shard_id=0)
@pytest.mark.parametrize("cls", [
CustomDeepseekV2RowParallelLinearReplaceAllreduce,
CustomDeepseekV2RowParallelLinear
])
@pytest.mark.parametrize("cls", [CustomDeepseekV2RowParallelLinear])
def test_row_parallel_linear(cls, mock_distributed):
linear = cls(input_size=128, output_size=64, bias=False, quant_config=None)
linear.quant_method = Mock()
@@ -185,33 +41,6 @@ def test_row_parallel_linear(cls, mock_distributed):
assert output[0].shape == (2, 4, 64)
def test_custom_deepseek_v2_mlp(mock_distributed, base_config):
mlp = CustomDeepseekV2MLP(hidden_size=128,
intermediate_size=256,
hidden_act="silu",
quant_config=None)
assert isinstance(mlp.act_fn, CustomDeepseekV2SiluAndMul)
x = torch.randn(2, 4, 128)
output = mlp(x)
assert output.shape == (2, 4, 128)
with patch("vllm_ascend.models.deepseek_v2.QuantizationConfig"
) as mock_quant_config:
mock_quant_config.name = "w8a8dynamic"
with pytest.raises(NotImplementedError):
CustomDeepseekV2MLP(hidden_size=128,
intermediate_size=256,
hidden_act="silu",
quant_config=mock_quant_config,
force_replicate=False)
with pytest.raises(ValueError):
CustomDeepseekV2MLP(hidden_size=128,
intermediate_size=256,
hidden_act="relu",
quant_config=None)
@patch("torch_npu.npu_rms_norm")
def test_custom_deepseek_v2_mla_attention(mock_rms_norm, mock_distributed,
base_config):