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

[quantization] Add w8a16 quantization support (#4541)

Browse Source 
### What this PR does / why we need it?
related to https://github.com/vllm-project/vllm-ascend/issues/4267

### Does this PR introduce _any_ user-facing change?
support w8a16 quantization now

### How was this patch tested?

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

### Test
tested using [aisbench](https://gitee.com/aisbench/benchmark/) with tp2
#### Precision
  | ceval | mmlu | gsm8k
-- | -- | -- | --
bf16 | 90.46 | 89.17 | 96.21
w8a16 | 89.51 | 89.29 | 95.98

#### Performance
  | input_len | output_len | concurrency | TTFT (ms) | TPOT (ms) | TPS
(Total) (tokens/s)
-- | -- | -- | -- | -- | -- | --
bf16 | 2048 | 2048 | 10 | 1911.7136 | 77.988 | 253.9866
w8a16 | 2048 | 2048 | 10 | 2128.6334 | 67.1633 | 293.9117
bf16 | 3500 | 1024 | 10 | 3076.2509 | 84.3525 | 506.949
w8a16 | 3500 | 1024 | 10 | 2685.2031 | 73.015 | 585.4717

---------

Signed-off-by: yyt <yangyit139@gmail.com>
Signed-off-by: TmacAaron <yangyit139@gmail.com>
Co-authored-by: realliujiaxu <realliujiaxu@163.com>
This commit is contained in:
TmacAaron
2025-12-24 19:49:32 +08:00
committed by GitHub
parent 515267de22
commit 5018f2d8fd
4 changed files with 229 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

46
tests/e2e/singlecard/test_quantization.py
View File

@@ -18,6 +18,7 @@
from modelscope import snapshot_download # type: ignore[import-untyped]
from tests.e2e.conftest import VllmRunner
from tests.e2e.model_utils import check_outputs_equal
def test_qwen3_w8a8_quant():
@@ -25,10 +26,53 @@ def test_qwen3_w8a8_quant():
example_prompts = [
"vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs."
]
vllm_target_outputs = [([
85, 4086, 44, 374, 264, 1550, 42747, 628, 323, 4938, 72816, 44378, 323,
13480, 4712, 369, 444, 10994, 82, 13, 1084, 374, 6188, 311, 387
], 'vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. It is designed to be'
)]
with VllmRunner(
snapshot_download("vllm-ascend/Qwen3-0.6B-W8A8"),
max_model_len=8192,
gpu_memory_utilization=0.7,
quantization="ascend",
) as vllm_model:
vllm_model.generate_greedy(example_prompts, max_tokens)
vllm_quant_w8a8_outputs = vllm_model.generate_greedy(
example_prompts, max_tokens)
check_outputs_equal(
outputs_0_lst=vllm_target_outputs,
outputs_1_lst=vllm_quant_w8a8_outputs,
name_0="vllm_target_outputs",
name_1="vllm_w8a16_outputs",
)
def test_qwen3_dense_w8a16():
max_tokens = 5
example_prompts = [
"vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs."
]
vllm_target_outputs = [([
85, 4086, 44, 374, 264, 1550, 42747, 628, 323, 4938, 72816, 44378, 323,
13480, 4712, 369, 444, 10994, 82, 13, 1084, 374, 6188, 311, 387
], 'vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. It is designed to be'
)]
with VllmRunner(
snapshot_download("vllm-ascend/Qwen3-0.6B-W8A16"),
max_model_len=8192,
enforce_eager=False,
gpu_memory_utilization=0.7,
quantization="ascend",
) as vllm_model:
vllm_quant_w8a16_outputs = vllm_model.generate_greedy(
example_prompts, max_tokens)
check_outputs_equal(
outputs_0_lst=vllm_target_outputs,
outputs_1_lst=vllm_quant_w8a16_outputs,
name_0="vllm_target_outputs",
name_1="vllm_w8a16_outputs",
)

91
tests/ut/quantization/test_w8a16.py Normal file
View File

@@ -0,0 +1,91 @@
import os
from unittest.mock import MagicMock, patch
import torch
from tests.ut.base import TestBase
from vllm_ascend.quantization.w8a16 import AscendW8A16LinearMethod
class TestAscendW8A16LinearMethod(TestBase):
def setUp(self):
self.method = AscendW8A16LinearMethod()
def test_get_weight(self):
weight = self.method.get_weight(10, 20)
self.assertEqual(weight['weight'].dtype, torch.int8)
self.assertEqual(weight['weight'].shape, (20, 10))
@patch("torch_npu.npu_weight_quant_batchmatmul")
def test_apply_with_x_is_int8(self, mock_npu_weight_quant_batchmatmul):
layer = MagicMock()
layer.weight.data = torch.randn(128, 256)
layer.weight_scale.data = torch.randn(128, 1)
layer.weight_offset.data = torch.randn(128, 1)
x = torch.randn(32, 128)
bias = torch.randn(256)
expected_y_output = torch.randn(32, 256)
mock_npu_weight_quant_batchmatmul.return_value = expected_y_output
output = self.method.apply(layer, x, bias)
expected_y_output += bias
self.assertTrue(torch.equal(output, expected_y_output))
@patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_NZ": "0"})
@patch('torch_npu.npu_format_cast')
def test_process_weights_after_loading_with_nz0(self,
mock_npu_format_cast):
layer = MagicMock()
layer.weight.data = torch.randint(-127,
128, (128, 256),
dtype=torch.int8)
layer.weight_scale.data = torch.randn(128, 1)
layer.weight_offset.data = torch.randn(128, 1)
mock_npu_format_cast.return_value = MagicMock
self.method.process_weights_after_loading(layer)
self.assertEqual(layer.weight_scale.data.shape, (128, ))
self.assertEqual(layer.weight_offset.data.shape, (128, ))
mock_npu_format_cast.assert_not_called()
@patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_NZ": "1"})
@patch('torch_npu.npu_format_cast')
def test_process_weights_after_loading_with_nz1(self,
mock_npu_format_cast):
layer = MagicMock()
layer.weight.data = torch.randint(-127,
128, (128, 256),
dtype=torch.int8)
layer.weight_scale.data = torch.randn(128, 1)
layer.weight_offset.data = torch.randn(128, 1)
mock_npu_format_cast.return_value = MagicMock
self.method.process_weights_after_loading(layer)
self.assertEqual(layer.weight_scale.data.shape, (128, ))
self.assertEqual(layer.weight_offset.data.shape, (128, ))
mock_npu_format_cast.assert_called_once()
@patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_NZ": "2"})
@patch('torch_npu.npu_format_cast')
def test_process_weights_after_loading_with_nz2(self,
mock_npu_format_cast):
layer = MagicMock()
layer.weight.data = torch.randint(-127,
128, (128, 256),
dtype=torch.int8)
layer.weight_scale.data = torch.randn(128, 1)
layer.weight_offset.data = torch.randn(128, 1)
mock_npu_format_cast.return_value = MagicMock
self.method.process_weights_after_loading(layer)
self.assertEqual(layer.weight_scale.data.shape, (128, ))
self.assertEqual(layer.weight_offset.data.shape, (128, ))
mock_npu_format_cast.assert_called_once()

4
vllm_ascend/quantization/utils.py
View File

@@ -14,6 +14,7 @@ from .w8a8_dynamic import (AscendW8A8DynamicFusedMoEMethod,
AscendW8A8DynamicLinearMethod)
from .w8a8_pdmix import (AscendW8A8PDMixFusedMoeMethod,
AscendW8A8PDMixLinearMethod)
from .w8a16 import AscendW8A16LinearMethod
ASCEND_QUANTIZATION_METHOD_MAP: Dict[str, Dict[str, Type[Any]]] = {
"W4A16": {
@@ -36,6 +37,9 @@ ASCEND_QUANTIZATION_METHOD_MAP: Dict[str, Dict[str, Type[Any]]] = {
"W8A8_MIX": {
"linear": AscendW8A8PDMixLinearMethod,
"moe": AscendW8A8PDMixFusedMoeMethod,
},
"W8A16": {
"linear": AscendW8A16LinearMethod,
}
}

89
vllm_ascend/quantization/w8a16.py Normal file
View File

@@ -0,0 +1,89 @@
#
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
# This file is a part of the vllm-ascend project.
#
# 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 typing import Any, Dict, Optional
import torch
import torch_npu
from vllm_ascend.utils import maybe_trans_nz
class AscendW8A16LinearMethod:
"""Linear method for Ascend W8A16.
"""
def __init__(self) -> None:
pass
@staticmethod
def get_weight(
input_size: int,
output_size: int,
params_dtype: torch.dtype = torch.bfloat16,
) -> Dict[str, Any]:
params_dict = {
"weight": torch.empty(output_size, input_size, dtype=torch.int8)
}
return params_dict
@staticmethod
def get_pertensor_param(params_dtype: torch.dtype) -> Dict[str, Any]:
return {}
@staticmethod
def get_perchannel_param(
output_size: int,
params_dtype: torch.dtype,
) -> Dict[str, Any]:
params_dict = {}
params_dict["weight_scale"] = torch.empty(output_size,
1,
dtype=params_dtype)
params_dict["weight_offset"] = torch.empty(output_size,
1,
dtype=params_dtype)
return params_dict
def get_pergroup_param(self,
input_size: int,
output_size: int,
params_dtype: torch.dtype,
layer_type: Optional[str] = None) -> Dict[str, Any]:
return {}
@staticmethod
def apply(
layer: torch.nn.Module,
x: torch.Tensor,
bias: Optional[torch.Tensor] = None,
tp_rank: Optional[int] = 0,
) -> torch.Tensor:
output = torch_npu.npu_weight_quant_batchmatmul(
x=x,
weight=layer.weight,
antiquant_scale=layer.weight_scale,
antiquant_offset=layer.weight_offset,
bias=bias)
return output
def process_weights_after_loading(self, layer):
layer.weight.data = layer.weight.data.transpose(0, 1).contiguous()
layer.weight.data = maybe_trans_nz(layer.weight.data)
layer.weight_scale.data = torch.flatten(layer.weight_scale.data)
layer.weight_offset.data = torch.flatten(layer.weight_offset.data)