5a7181569c
### What this PR does / why we need it?
This PR introduces Oproj matrix tensor model parallel to achieve
decreasing of memory consumption. It only support graph mode in pure DP
scenario.
In deepseek r1 w8a8 PD disagregated Decode instance, using pure DP, with
oproj_tensor_parallel_size = 8, we have 1 ms TPOT increasing, saved 5.8
GB NPU memory per RANK. We got best performance when
oproj_tensor_parallel_size=4 without TPOT increasing.
performance data:
<img width="1442" height="442" alt="image"
src="https://github.com/user-attachments/assets/83270fc5-868a-4387-b0a9-fac29b4a376d"
/>
### Does this PR introduce _any_ user-facing change?
This PR introduces one new config in `additional_config`.
| Name | Effect | Required | Type | Constraints |
| :---------------------------- |
:--------------------------------------- | :------- | :--- |
:----------------- |
| oproj_tensor_parallel_size | Split the o_proj matrix along the row
dimension (head num * head dim) into oproj_tensor_parallel_size pieces.
| No | int | default value is None, once this value is set, the feature
will be enabled, head num * head dim must be divisible by this value. |
example
`--additional_config={"oproj_tensor_parallel_size": 8}`
### How was this patch tested?
- vLLM version: v0.10.1.1
- vLLM main:
eddaafc1c7
---------
Signed-off-by: zzhx1 <zzh_201018@outlook.com>
Co-authored-by: zzh <zzh_201018@outlook.com>
70 lines
3.2 KiB
Python
70 lines
3.2 KiB
Python
from unittest.mock import Mock, patch
|
|
|
|
import torch
|
|
|
|
from tests.ut.base import TestBase
|
|
from vllm_ascend.quantization.w8a8_dynamic import \
|
|
AscendW8A8DynamicFusedMoEMethod
|
|
|
|
|
|
class TestAscendW8A8FusedMoEMethod(TestBase):
|
|
num_experts = 8
|
|
hidden_size = 128
|
|
intermediate_size = 128
|
|
|
|
@patch("torch.distributed.get_rank")
|
|
@patch("vllm_ascend.quantization.w8a8_dynamic.get_mc2_group")
|
|
@patch("vllm_ascend.quantization.w8a8_dynamic.get_ascend_config")
|
|
@patch("vllm_ascend.quantization.w8a8_dynamic.get_ep_group")
|
|
def setUp(self, mock_get_ep_group, mock_get_ascend_config,
|
|
mock_get_mc2_group, mock_get_rank):
|
|
with patch(
|
|
'vllm_ascend.quantization.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": 256})
|
|
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()
|
|
|
|
# 创建一个具有具体属性的 Mock 对象来表示 ascend_scheduler_config
|
|
mock_ascend_scheduler_config = Mock()
|
|
mock_ascend_scheduler_config.enabled = False
|
|
mock_ascend_scheduler_config.max_num_batched_tokens = 1024
|
|
mock_ascend_scheduler_config.max_model_len = 2048
|
|
mock_ascend_config.ascend_scheduler_config = mock_ascend_scheduler_config
|
|
|
|
mock_ascend_config.torchair_graph_config = Mock(enabled=False)
|
|
mock_ascend_config.enable_chunked_prefill = False
|
|
mock_get_ascend_config.return_value = mock_ascend_config
|
|
mock_mc2_group = Mock(device_group=0)
|
|
mock_get_mc2_group.return_value = mock_mc2_group
|
|
mock_rank = Mock()
|
|
mock_get_rank.return_value = mock_rank
|
|
|
|
self.quant_method = AscendW8A8DynamicFusedMoEMethod()
|
|
|
|
def test_get_weight(self):
|
|
param_dict = self.quant_method.get_weight(self.num_experts,
|
|
self.intermediate_size,
|
|
self.hidden_size,
|
|
torch.bfloat16)
|
|
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))
|
|
|
|
def test_get_dynamic_quant_param(self):
|
|
param_dict = self.quant_method.get_dynamic_quant_param(
|
|
self.num_experts, self.intermediate_size, self.hidden_size,
|
|
torch.bfloat16)
|
|
self.assertEqual(param_dict["w13_weight_scale"].dtype, torch.bfloat16)
|
|
self.assertEqual(param_dict["w13_weight_scale"].shape,
|
|
(self.num_experts, 2 * self.intermediate_size, 1))
|