mirror of
https://github.com/PaddlePaddle/FastDeploy.git
synced 2025-12-24 13:28:13 +08:00
Sync v2.0 version of code to github repo
This commit is contained in:
62
docs/zh/quantization/wint2.md
Normal file
62
docs/zh/quantization/wint2.md
Normal file
@@ -0,0 +1,62 @@
|
||||
# WINT2量化
|
||||
|
||||
权重经过CCQ(Convolutional Coding Quantization)方法离线压缩。权重实际存储的数值类型是INT8,每个INT8数值中打包了4个权重,等价于每个权重2bits. 激活不做量化,计算时将权重实时地反量化、解码为BF16数值类型,并用BF16数值类型计算。
|
||||
- **支持硬件**:GPU
|
||||
- **支持结构**:MoE结构
|
||||
|
||||
CCQ WINT2一般用于资源受限的低门槛场景,以ERNIE-4.5-300B-A47B为例,将权重压缩到89GB,可支持141GB H20单卡部署。
|
||||
|
||||
## 启动WINT2推理服务
|
||||
|
||||
```
|
||||
python -m fastdeploy.entrypoints.openai.api_server \
|
||||
--model baidu/ERNIE-4.5-300B-A47B-2Bits-Paddle \
|
||||
--port 8180 --engine-worker-queue-port 8181 \
|
||||
--cache-queue-port 8182 --metrics-port 8182 \
|
||||
--tensor-parallel-size 1 \
|
||||
--max-model-len 32768 \
|
||||
--max-num-seqs 32
|
||||
```
|
||||
|
||||
通过指定 `--model baidu/ERNIE-4.5-300B-A47B-2Bits-Paddle` 可自动从AIStudio下载已离线量化好的WINT2模型,在该模型的config.json文件中,会有WINT2量化相关的配置信息,不用再在启动推理服务时设置 `--quantization`.
|
||||
|
||||
模型的config.json文件中的量化配置示例如下:
|
||||
|
||||
```
|
||||
"quantization_config": {
|
||||
"dense_quant_type": "wint8",
|
||||
"moe_quant_type": "w4w2",
|
||||
"quantization": "wint2",
|
||||
"moe_quant_config": {
|
||||
"moe_w4_quant_config": {
|
||||
"quant_type": "wint4",
|
||||
"quant_granularity": "per_channel",
|
||||
"quant_start_layer": 0,
|
||||
"quant_end_layer": 6
|
||||
},
|
||||
"moe_w2_quant_config": {
|
||||
"quant_type": "wint2",
|
||||
"quant_granularity": "pp_acc",
|
||||
"quant_group_size": 64,
|
||||
"quant_start_layer": 7,
|
||||
"quant_end_layer": 53
|
||||
}
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
- 更多部署教程请参考[get_started](../get_started/ernie-4.5.md);
|
||||
- 更多模型说明请参考[支持模型列表](https://console.cloud.baidu-int.com/devops/icode/repos/baidu/paddle_internal/FastDeploy/blob/feature%2Finference-refactor-20250528/docs/supported_models.md)。
|
||||
|
||||
|
||||
## WINT2效果
|
||||
|
||||
在ERNIE-4.5-300B-A47B模型上,WINT2与WINT4效果对比:
|
||||
|
||||
| 测试集 |数据集大小| WINT4 | WINT2 |
|
||||
|---------|---------|---------|---------|
|
||||
| IFEval |500|88.17 | 85.40 |
|
||||
|BBH|6511|94.43|92.02|
|
||||
|DROP|9536|91.17|89.97|
|
||||
|
||||
## WINT2推理性能
|
||||
Reference in New Issue
Block a user