mirror of
https://github.com/vllm-project/vllm-ascend.git
synced 2025-10-20 13:43:53 +08:00
f69a83b7ba
Support flash comm v1(Sequence Parallelism) for Allgather EP. - vLLM version: v0.11.0rc3 - vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0 --------- Signed-off-by: realliujiaxu <realliujiaxu@163.com> Co-authored-by: zhaozx-cn <zhaozx2116@163.com>
131 lines
5.5 KiB
Python
131 lines
5.5 KiB
Python
#
|
|
# 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, Mock, patch
|
|
|
|
import pytest
|
|
import torch
|
|
from vllm.config import CacheConfig
|
|
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
|
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
|
|
|
|
from vllm_ascend import ascend_config
|
|
from vllm_ascend.models.deepseek_v2 import (CustomDeepseekV2MLAAttention,
|
|
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()
|
|
linear.quant_method.apply.return_value = torch.randn(2, 4, 64)
|
|
input_ = torch.randn(2, 4, 128)
|
|
with patch("vllm_ascend.models.deepseek_v2.split_tensor_along_last_dim",
|
|
return_value=[torch.randn(2, 4, 64)]):
|
|
linear.input_is_parallel = False
|
|
output = linear(input_, is_prefill=True)
|
|
assert output[0].shape == (2, 4, 64)
|
|
|
|
linear.input_is_parallel = True
|
|
output = linear(input_, is_prefill=False)
|
|
assert output[0].shape == (2, 4, 64)
|
|
|
|
|
|
@patch("vllm_ascend.models.layers.mla.get_forward_context")
|
|
@patch("torch.ops.vllm.mla_forward")
|
|
@patch("torch_npu.npu_rms_norm")
|
|
def test_custom_deepseek_v2_mla_attention(mock_rms_norm, mock_mla_forward,
|
|
mock_forward_context,
|
|
mock_distributed, base_config):
|
|
mock_rms_norm.return_value = (torch.randn(2, 128), torch.randn(2, 128))
|
|
# Make a fake ascend config because of the AscendLinearBase
|
|
vllm_config = MagicMock()
|
|
vllm_config.additional_config = None
|
|
vllm_config.parallel_config.enable_expert_parallel = False
|
|
vllm_config.parallel_config.tensor_parallel_size = 1
|
|
vllm_config.kv_transfer_config = None
|
|
ascend_config.init_ascend_config(vllm_config)
|
|
dummy_forward_context = MagicMock()
|
|
dummy_forward_context.sp_enabled = False
|
|
mock_forward_context.return_value = dummy_forward_context
|
|
|
|
attn = CustomDeepseekV2MLAAttention(config=base_config,
|
|
hidden_size=128,
|
|
num_heads=8,
|
|
qk_nope_head_dim=16,
|
|
qk_rope_head_dim=16,
|
|
v_head_dim=32,
|
|
q_lora_rank=16,
|
|
kv_lora_rank=16,
|
|
cache_config=CacheConfig(),
|
|
quant_config=None,
|
|
prefix="layers.0.self_attn")
|
|
assert attn.debug_layer_idx == 0
|
|
|
|
x = torch.randn(2, 4, 128)
|
|
positions = torch.arange(4).repeat(2, 1)
|
|
with patch.object(attn.mla_attn,
|
|
"__call__",
|
|
return_value=torch.randn(2, 4, 128)):
|
|
attn(positions, x)
|
|
mock_mla_forward.assert_called_once()
|
|
|
|
attn = CustomDeepseekV2MLAAttention(config=base_config,
|
|
hidden_size=128,
|
|
num_heads=8,
|
|
qk_nope_head_dim=16,
|
|
qk_rope_head_dim=16,
|
|
v_head_dim=32,
|
|
q_lora_rank=None,
|
|
kv_lora_rank=16,
|
|
prefix="layers.1.self_attn")
|
|
assert hasattr(attn, "q_proj")
|
|
ascend_config._ASCEND_CONFIG = None
|
|
|
|
|
|
def test_deepseek_v2_lmhead(mock_distributed, vllm_config):
|
|
# 创建一个简单的配置对象
|
|
class SimpleConfig:
|
|
|
|
def __init__(self):
|
|
self.vocab_size = 10000
|
|
self.hidden_size = 128
|
|
|
|
config = SimpleConfig()
|
|
|
|
# Make a fake ascend config because of the AscendLinearBase
|
|
vllm_config = MagicMock()
|
|
vllm_config.additional_config = None
|
|
vllm_config.parallel_config.enable_expert_parallel = False
|
|
vllm_config.parallel_config.tensor_parallel_size = 1
|
|
vllm_config.kv_transfer_config = None
|
|
ascend_config.init_ascend_config(vllm_config)
|
|
|
|
# 直接创建lmhead和logits_processor
|
|
lmhead = ParallelLMHead(config.vocab_size, config.hidden_size)
|
|
logits_processor = LogitsProcessor(config.vocab_size)
|
|
|
|
# 创建模拟输出
|
|
mock_output = torch.randn(2, 4, config.hidden_size)
|
|
mock_logits = torch.randn(2, 4, config.vocab_size)
|
|
|
|
# 直接测试logits_processor
|
|
with patch.object(lmhead.quant_method, "apply", return_value=mock_logits):
|
|
with patch.object(logits_processor,
|
|
"_gather_logits",
|
|
return_value=mock_logits):
|
|
logits = logits_processor(lmhead, mock_output)
|
|
assert logits.shape == (2, 4, config.vocab_size)
|
|
ascend_config._ASCEND_CONFIG = None
|