Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "4360cfc6a850ee2431cc7b21fdba4fdc6bec4d0f"
Commit d824b9d3 authored by zhuwenwen's avatar zhuwenwen
Browse files

update vllm0.5.0

parent 8d5187cb
Show whitespace changes
Inline Side-by-side
Showing with 4084 additions and 53 deletions
.gitmodules
View file @ d824b9d3
[submodule "vllm"]
path = vllm
url = http://developer.hpccube.com/codes/OpenDAS/vllm.git
branch = vllm-v0.3.3-dtk24.04
\ No newline at end of file
Dockerfile
View file @ d824b9d3
FROM image.sourcefind.cn:5000/dcu/admin/base/pytorch:2.1.0-ubuntu20.04-dtk24.04.1-py3.10 FROM image.sourcefind.cn:5000/dcu/admin/base/pytorch:2.1.0-ubuntu20.04-dtk24.04.2-py3.10
ENV LANG C.UTF-8 ENV LANG C.UTF-8
\ No newline at end of file
RUN pip install aiohttp==3.9.1 outlines==0.0.37 openai==1.23.3 -i http://mirrors.aliyun.com/pypi/simple/ --trusted-host mirrors.aliyun.com
\ No newline at end of file
README.md
View file @ d824b9d3
...@@ -2,7 +2,7 @@ ...@@ -2,7 +2,7 @@
* @Author: zhuww * @Author: zhuww
* @email: zhuww@sugon.com * @email: zhuww@sugon.com
* @Date: 2024-05-24 14:15:07 * @Date: 2024-05-24 14:15:07
* @LastEditTime: 2024-08-23 14:40:01 * @LastEditTime: 2024-09-30 08:30:01
--> -->
# Qwen1.5 # Qwen1.5
...@@ -27,15 +27,13 @@ Qwen1.5是阿里云开源大型语言模型系列，是Qwen2.0的beta版本。 ...@@ -27,15 +27,13 @@ Qwen1.5是阿里云开源大型语言模型系列，是Qwen2.0的beta版本。
提供[光源](https://www.sourcefind.cn/#/image/dcu/custom)拉取推理的docker镜像： 提供[光源](https://www.sourcefind.cn/#/image/dcu/custom)拉取推理的docker镜像：
``` ```
docker pull image.sourcefind.cn:5000/dcu/admin/base/pytorch:2.1.0-ubuntu20.04-dtk24.04.1-py3.10 docker pull image.sourcefind.cn:5000/dcu/admin/base/pytorch:2.1.0-ubuntu20.04-dtk24.04.2-py3.10
# <Image ID>用上面拉取docker镜像的ID替换 # <Image ID>用上面拉取docker镜像的ID替换
# <Host Path>主机端路径 # <Host Path>主机端路径
# <Container Path>容器映射路径 # <Container Path>容器映射路径
docker run -it --name qwen1.5_vllm --privileged --shm-size=64G --device=/dev/kfd --device=/dev/dri/ --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --ulimit memlock=-1:-1 --ipc=host --network host --group-add video -v /opt/hyhal:/opt/hyhal -v <Host Path>:<Container Path> <Image ID> /bin/bash docker run -it --name qwen1.5_vllm --privileged --shm-size=64G --device=/dev/kfd --device=/dev/dri/ --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --ulimit memlock=-1:-1 --ipc=host --network host --group-add video -v /opt/hyhal:/opt/hyhal -v <Host Path>:<Container Path> <Image ID> /bin/bash
pip install aiohttp==3.9.1 outlines==0.0.37 openai==1.23.3 -i http://mirrors.aliyun.com/pypi/simple/ --trusted-host mirrors.aliyun.com
``` ```
`Tips：若在K100/Z100L上使用，使用镜像docker pull image.sourcefind.cn:5000/dcu/admin/base/custom:vllm0.3.3-dtk24.04.1-ubuntu20.4-py310-zk-v1` `Tips：若在K100/Z100L上使用，使用定制镜像docker pull image.sourcefind.cn:5000/dcu/admin/base/custom:vllm0.5.0-dtk24.04.1-ubuntu20.04-py310-zk-v1,K100/Z100L不支持awq量化`
### Dockerfile（方法二） ### Dockerfile（方法二）
``` ```
...@@ -44,71 +42,56 @@ pip install aiohttp==3.9.1 outlines==0.0.37 openai==1.23.3 -i http://mirrors.ali ...@@ -44,71 +42,56 @@ pip install aiohttp==3.9.1 outlines==0.0.37 openai==1.23.3 -i http://mirrors.ali
docker build -t qwen1.5:latest . docker build -t qwen1.5:latest .
docker run -it --name qwen1.5_vllm --privileged --shm-size=64G --device=/dev/kfd --device=/dev/dri/ --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --ulimit memlock=-1:-1 --ipc=host --network host --group-add video -v /opt/hyhal:/opt/hyhal:ro -v <Host Path>:<Container Path> qwen1.5:latest /bin/bash docker run -it --name qwen1.5_vllm --privileged --shm-size=64G --device=/dev/kfd --device=/dev/dri/ --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --ulimit memlock=-1:-1 --ipc=host --network host --group-add video -v /opt/hyhal:/opt/hyhal:ro -v <Host Path>:<Container Path> qwen1.5:latest /bin/bash
``` ```
`Tips：若在K100/Z100L上使用，需要替换flash_attn，下载链接：https://forum.hpccube.com/thread/515`
### Anaconda（方法三） ### Anaconda（方法三）
``` ```
conda create -n qwen1.5_vllm python=3.10 conda create -n qwen1.5_vllm python=3.10
pip install aiohttp==3.9.1 outlines==0.0.37 openai==1.23.3
``` ```
关于本项目DCU显卡所需的特殊深度学习库可从[光合](https://developer.hpccube.com/tool/)开发者社区下载安装。 关于本项目DCU显卡所需的特殊深度学习库可从[光合](https://developer.hpccube.com/tool/)开发者社区下载安装。
* DTK驱动：dtk24.04.1 * DTK驱动：dtk24.04.2
* Pytorch: 2.1.0 * Pytorch: 2.1.0
* triton:2.1.0 * triton:2.1.0
* vllm: 0.3.3 * lmslim: 0.1.0
* xformers: 0.0.25 * xformers: 0.0.25
* flash_attn: 2.0.4 * flash_attn: 2.0.4
* vllm: 0.5.0
* python: python3.10 * python: python3.10
`Tips：需先安装相关依赖，最后安装vllm包`
`Tips：若在K100/Z100L上使用，需要替换flash_attn，下载链接：https://forum.hpccube.com/thread/515`
## 数据集 ## 数据集
无 无
## 推理 ## 推理
### 源码编译安装
```
# 若使用光源的镜像，可以跳过源码编译安装，镜像中已安装vllm。
git clone http://developer.hpccube.com/codes/modelzoo/qwen1.5_vllm.git
cd qwen1.5_vllm
git submodule init && git submodule update
cd vllm
pip install wheel
python setup.py bdist_wheel
cd dist && pip install vllm*
```
### 模型下载 ### 模型下载
| 基座模型 | chat模型 | GPTQ模型 | | 基座模型 | chat模型 | GPTQ模型 | AWQ模型 |
| ------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------- | | ------- | ------- | ------- | ------- |
| [Qwen-7B](https://huggingface.co/Qwen/Qwen-7B) | [Qwen-7B-Chat](http://113.200.138.88:18080/aimodels/Qwen-7B-Chat) | | [Qwen-7B](https://huggingface.co/Qwen/Qwen-7B) | [Qwen-7B-Chat](http://113.200.138.88:18080/aimodels/Qwen-7B-Chat) | [Qwen-7B-Chat-GPTQ-Int4](https://huggingface.co/Qwen/Qwen-7B-Chat-Int4) |
| [Qwen-14B](https://huggingface.co/Qwen/Qwen-14B) | [Qwen-14B-Chat](http://113.200.138.88:18080/aimodels/Qwen-14B-Chat) | | [Qwen-14B](https://huggingface.co/Qwen/Qwen-14B) | [Qwen-14B-Chat](http://113.200.138.88:18080/aimodels/Qwen-14B-Chat) | [Qwen-14B-Chat-GPTQ-Int4](https://huggingface.co/Qwen/Qwen-14B-Chat-Int4) |
| [Qwen-72B](http://113.200.138.88:18080/aimodels/qwen/Qwen-72B) | [Qwen-72B-Chat](http://113.200.138.88:18080/aimodels/Qwen-72B-Chat) | | [Qwen-72B](http://113.200.138.88:18080/aimodels/qwen/Qwen-72B) | [Qwen-72B-Chat](http://113.200.138.88:18080/aimodels/Qwen-72B-Chat) | [Qwen-72B-Chat-GPTQ-Int4](https://huggingface.co/Qwen/Qwen-72B-Chat-Int4) |
| [Qwen1.5-7B](https://huggingface.co/Qwen/Qwen1.5-7B) | [Qwen1.5-7B-Chat](https://huggingface.co/Qwen/Qwen1.5-7B-Chat) | [Qwen1.5-7B-Chat-GPTQ-Int4](https://huggingface.co/Qwen/Qwen1.5-7B-Chat-GPTQ-Int4) | | [Qwen1.5-7B](https://huggingface.co/Qwen/Qwen1.5-7B) | [Qwen1.5-7B-Chat](https://huggingface.co/Qwen/Qwen1.5-7B-Chat) | [Qwen1.5-7B-Chat-GPTQ-Int4](https://huggingface.co/Qwen/Qwen1.5-7B-Chat-GPTQ-Int4) | [Qwen1.5-7B-Chat-AWQ-Int4](http://113.200.138.88:18080/aimodels/qwen/Qwen1.5-7B-Chat-AWQ) |
| [Qwen1.5-14B](https://huggingface.co/Qwen/Qwen1.5-14B) | [Qwen1.5-14B-Chat](http://113.200.138.88:18080/aimodels/qwen/Qwen1.5-14B-Chat) | [Qwen1.5-14B-Chat-GPTQ-Int4](https://huggingface.co/Qwen/Qwen1.5-14B-Chat-GPTQ-Int4) | | [Qwen1.5-14B](https://huggingface.co/Qwen/Qwen1.5-14B) | [Qwen1.5-14B-Chat](http://113.200.138.88:18080/aimodels/qwen/Qwen1.5-14B-Chat) | [Qwen1.5-14B-Chat-GPTQ-Int4](https://huggingface.co/Qwen/Qwen1.5-14B-Chat-GPTQ-Int4) | [Qwen1.5-14B-Chat-AWQ-Int4](http://113.200.138.88:18080/aimodels/qwen/Qwen1.5-14B-Chat-AWQ) |
| [Qwen1.5-32B](http://113.200.138.88:18080/aimodels/Qwen1.5-32B) | [Qwen1.5-32B-Chat](http://113.200.138.88:18080/aimodels/Qwen1.5-32B-Chat) | [Qwen1.5-32B-Chat-GPTQ-Int4](http://113.200.138.88:18080/aimodels/Qwen1.5-32B-Chat-GPTQ-Int4) | | [Qwen1.5-32B](http://113.200.138.88:18080/aimodels/Qwen1.5-32B) | [Qwen1.5-32B-Chat](http://113.200.138.88:18080/aimodels/Qwen1.5-32B-Chat) | [Qwen1.5-32B-Chat-GPTQ-Int4](http://113.200.138.88:18080/aimodels/Qwen1.5-32B-Chat-GPTQ-Int4) | [Qwen1.5-32B-Chat-AWQ-Int4](https://huggingface.co/Qwen/Qwen1.5-32B-Chat-AWQ) |
| [Qwen1.5-72B](http://113.200.138.88:18080/aimodels/Qwen1.5-72B) | [Qwen1.5-72B-Chat](http://113.200.138.88:18080/aimodels/Qwen1.5-72B-Chat) | [Qwen1.5-72B-Chat-GPTQ-Int4](https://huggingface.co/Qwen/Qwen1.5-72B-Chat-GPTQ-Int4) | | [Qwen1.5-72B](http://113.200.138.88:18080/aimodels/Qwen1.5-72B) | [Qwen1.5-72B-Chat](http://113.200.138.88:18080/aimodels/Qwen1.5-72B-Chat) | [Qwen1.5-72B-Chat-GPTQ-Int4](https://huggingface.co/Qwen/Qwen1.5-72B-Chat-GPTQ-Int4) | [Qwen1.5-72B-Chat-AWQ-Int4](http://113.200.138.88:18080/aimodels/qwen/Qwen1.5-72B-Chat-AWQ) |
| [Qwen1.5-110B](http://113.200.138.88:18080/aimodels/Qwen1.5-110B) | [Qwen1.5-110B-Chat](http://113.200.138.88:18080/aimodels/Qwen1.5-110B-Chat) | [Qwen1.5-110B-Chat-GPTQ-Int4](https://huggingface.co/Qwen/Qwen1.5-110B-Chat-GPTQ-Int4) | | [Qwen1.5-110B](http://113.200.138.88:18080/aimodels/Qwen1.5-110B) | [Qwen1.5-110B-Chat](http://113.200.138.88:18080/aimodels/Qwen1.5-110B-Chat) | [Qwen1.5-110B-Chat-GPTQ-Int4](https://huggingface.co/Qwen/Qwen1.5-110B-Chat-GPTQ-Int4) | [Qwen1.5-110B-Chat-AWQ-Int4](http://113.200.138.88:18080/aimodels/qwen/Qwen1.5-110B-Chat-AWQ) |
| [Qwen2-7B](http://113.200.138.88:18080/aimodels/Qwen2-7B) | [Qwen2-7B-Instruct](http://113.200.138.88:18080/aimodels/Qwen2-7B-Instruct) | [Qwen2-7B-Instruct-GPTQ-Int4](https://huggingface.co/Qwen/Qwen2-7B-Instruct-GPTQ-Int4) | | [Qwen2-7B](http://113.200.138.88:18080/aimodels/Qwen2-7B) | [Qwen2-7B-Instruct](http://113.200.138.88:18080/aimodels/Qwen2-7B-Instruct) | [Qwen2-7B-Instruct-GPTQ-Int4](https://huggingface.co/Qwen/Qwen2-7B-Instruct-GPTQ-Int4) | [Qwen2-7B-Instruct-AWQ-Int4](http://113.200.138.88:18080/aimodels/qwen/Qwen2-7B-Instruct-AWQ)
| [Qwen2-72B](http://113.200.138.88:18080/aimodels/Qwen2-72B) | [Qwen2-72B-Instruct](http://113.200.138.88:18080/aimodels/Qwen2-72B-Instruct) | [Qwen2-72B-Instruct-GPTQ-Int4](https://huggingface.co/Qwen/Qwen2-72B-Instruct-GPTQ-Int4) | | [Qwen2-72B](http://113.200.138.88:18080/aimodels/Qwen2-72B) | [Qwen2-72B-Instruct](http://113.200.138.88:18080/aimodels/Qwen2-72B-Instruct) | [Qwen2-72B-Instruct-GPTQ-Int4](https://huggingface.co/Qwen/Qwen2-72B-Instruct-GPTQ-Int4) |[Qwen2-72B-Instruct-AWQ-Int4](http://113.200.138.88:18080/aimodels/qwen/Qwen2-72B-Instruct-AWQ) |
### 离线批量推理 ### 离线批量推理
```bash ```bash
python vllm/examples/offline_inference.py python examples/offline_inference.py
``` ```
其中，`prompts`为提示词；`temperature`为控制采样随机性的值，值越小模型生成越确定，值变高模型生成更随机，0表示贪婪采样，默认为1；`max_tokens=16`为生成长度，默认为1； 其中，`prompts`为提示词；`temperature`为控制采样随机性的值，值越小模型生成越确定，值变高模型生成更随机，0表示贪婪采样，默认为1；`max_tokens=16`为生成长度，默认为1；
`model`为模型路径；`tensor_parallel_size=1`为使用卡数，默认为1；`dtype="float16"`为推理数据类型，如果模型权重是bfloat16,需要修改为float16推理,`quantization="gptq"`为使用gptq量化进行推理,需下载以上GPTQ模型。 `model`为模型路径；`tensor_parallel_size=1`为使用卡数，默认为1；`dtype="float16"`为推理数据类型，如果模型权重是bfloat16,需要修改为float16推理,`quantization="gptq"`为使用gptq量化进行推理,需下载以上GPTQ模型。`quantization="awq"`为使用awq量化进行推理,需下载以上AWQ模型。
### 离线批量推理性能测试 ### 离线批量推理性能测试
`Tips：若测试qwen1.5-7b/qwen1.5-72b/qwen1.5-72b,添加环境变量LLAMA_NN=1`
1、指定输入输出 1、指定输入输出
```bash ```bash
python vllm/benchmarks/benchmark_throughput.py --num-prompts 1 --input-len 32 --output-len 128 --model Qwen/Qwen1.5-7B-Chat -tp 1 --trust-remote-code --enforce-eager --dtype float16 python benchmarks/benchmark_throughput.py --num-prompts 1 --input-len 32 --output-len 128 --model Qwen/Qwen1.5-7B-Chat -tp 1 --trust-remote-code --enforce-eager --dtype float16
``` ```
其中`--num-prompts`是batch数，`--input-len`是输入seqlen，`--output-len`是输出token长度，`--model`为模型路径，`-tp`为使用卡数，`dtype="float16"`为推理数据类型，如果模型权重是bfloat16,需要修改为float16推理。若指定`--output-len 1`即为首字延迟。`-q gptq`为使用gptq量化模型进行推理。 其中`--num-prompts`是batch数，`--input-len`是输入seqlen，`--output-len`是输出token长度，`--model`为模型路径，`-tp`为使用卡数，`dtype="float16"`为推理数据类型，如果模型权重是bfloat16,需要修改为float16推理。若指定`--output-len 1`即为首字延迟。`-q gptq`为使用gptq量化模型进行推理。
...@@ -119,7 +102,7 @@ wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/r ...@@ -119,7 +102,7 @@ wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/r
``` ```
```bash ```bash
python benchmark_throughput.py --num-prompts 1 --model Qwen/Qwen1.5-7B-Chat --dataset ShareGPT_V3_unfiltered_cleaned_split.json -tp 1 --trust-remote-code --enforce-eager --dtype float16 python benchmarks/benchmark_throughput.py --num-prompts 1 --model Qwen/Qwen1.5-7B-Chat --dataset ShareGPT_V3_unfiltered_cleaned_split.json -tp 1 --trust-remote-code --enforce-eager --dtype float16
``` ```
其中`--num-prompts`是batch数，`--model`为模型路径，`--dataset`为使用的数据集，`-tp`为使用卡数，`dtype="float16"`为推理数据类型，如果模型权重是bfloat16,需要修改为float16推理。`-q gptq`为使用gptq量化模型进行推理。 其中`--num-prompts`是batch数，`--model`为模型路径，`--dataset`为使用的数据集，`-tp`为使用卡数，`dtype="float16"`为推理数据类型，如果模型权重是bfloat16,需要修改为float16推理。`-q gptq`为使用gptq量化模型进行推理。
...@@ -132,9 +115,9 @@ python -m vllm.entrypoints.api_server --model Qwen/Qwen1.5-7B-Chat --dtype flo ...@@ -132,9 +115,9 @@ python -m vllm.entrypoints.api_server --model Qwen/Qwen1.5-7B-Chat --dtype flo
2、启动客户端： 2、启动客户端：
```bash ```bash
python vllm/benchmarks/benchmark_serving.py --model Qwen/Qwen1.5-7B-Chat --dataset ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts 1 --trust-remote-code python benchmarks/benchmark_serving.py --model Qwen/Qwen1.5-7B-Chat --dataset ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts 1 --trust-remote-code
``` ```
参数同使用数据集，离线批量推理性能测试，具体参考[vllm/benchmarks/benchmark_serving.py] 参数同使用数据集，离线批量推理性能测试，具体参考[benchmarks/benchmark_serving.py](benchmarks/benchmark_serving.py)
### OpenAI兼容服务 ### OpenAI兼容服务
...@@ -142,7 +125,7 @@ python vllm/benchmarks/benchmark_serving.py --model Qwen/Qwen1.5-7B-Chat --datas ...@@ -142,7 +125,7 @@ python vllm/benchmarks/benchmark_serving.py --model Qwen/Qwen1.5-7B-Chat --datas
```bash ```bash
python -m vllm.entrypoints.openai.api_server --model Qwen/Qwen1.5-7B-Chat --enforce-eager --dtype float16 --trust-remote-code python -m vllm.entrypoints.openai.api_server --model Qwen/Qwen1.5-7B-Chat --enforce-eager --dtype float16 --trust-remote-code
``` ```
这里`--model`为加载模型路径，`--dtype`为数据类型：float16，默认情况使用tokenizer中的预定义聊天模板，`--chat-template`可以添加新模板覆盖默认模板,`-q gptq`为使用gptq量化模型进行推理。 这里`--model`为加载模型路径，`--dtype`为数据类型：float16，默认情况使用tokenizer中的预定义聊天模板，`--chat-template`可以添加新模板覆盖默认模板,`-q gptq`为使用gptq量化模型进行推理,`-q awqq`为使用awq量化模型进行推理。
列出模型型号： 列出模型型号：
```bash ```bash
...@@ -160,7 +143,7 @@ curl http://localhost:8000/v1/completions \ ...@@ -160,7 +143,7 @@ curl http://localhost:8000/v1/completions \
"temperature": 0 "temperature": 0
}' }'
``` ```
或者使用[vllm/examples/openai_completion_client.py](https://developer.hpccube.com/codes/OpenDAS/vllm/-/blob/df6349c78b49a5b8f6f600d0d9490791cd1d32ee/examples/openai_completion_client.py) 或者使用[examples/openai_completion_client.py](examples/openai_completion_client.py)
### OpenAI Chat API和vllm结合使用 ### OpenAI Chat API和vllm结合使用
...@@ -175,7 +158,7 @@ curl http://localhost:8000/v1/chat/completions \ ...@@ -175,7 +158,7 @@ curl http://localhost:8000/v1/chat/completions \
] ]
}' }'
``` ```
或者使用[vllm/examples/openai_chatcompletion_client.py](https://developer.hpccube.com/codes/OpenDAS/vllm/-/blob/df6349c78b49a5b8f6f600d0d9490791cd1d32ee/examples/openai_chatcompletion_client.py) 或者使用[examples/openai_chatcompletion_client.py](examples/openai_chatcompletion_client.py)
## result ## result
......
benchmarks/README.md 0 → 100644
View file @ d824b9d3
# Benchmarking vLLM
## Downloading the ShareGPT dataset
You can download the dataset by running:
```bash
wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
```
benchmarks/backend_request_func.py 0 → 100644
View file @ d824b9d3
import json
import os
import sys
import time
import traceback
from dataclasses import dataclass, field
from typing import List, Optional
import aiohttp
from tqdm.asyncio import tqdm
AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=6 * 60 * 60)
@dataclass
class RequestFuncInput:
prompt: str
api_url: str
prompt_len: int
output_len: int
model: str
best_of: int = 1
use_beam_search: bool = False
@dataclass
class RequestFuncOutput:
generated_text: str = ""
success: bool = False
latency: float = 0.0
ttft: float = 0.0 # Time to first token
itl: List[float] = field(
default_factory=list) # List of inter-token latencies
prompt_len: int = 0
error: str = ""
async def async_request_tgi(
request_func_input: RequestFuncInput,
pbar: Optional[tqdm] = None,
) -> RequestFuncOutput:
api_url = request_func_input.api_url
assert api_url.endswith("generate_stream")
async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
assert not request_func_input.use_beam_search
params = {
"best_of": request_func_input.best_of,
"max_new_tokens": request_func_input.output_len,
"do_sample": True,
"temperature": 0.01, # TGI does not accept 0.0 temperature.
"top_p": 0.99, # TGI does not accept 1.0 top_p.
}
payload = {
"inputs": request_func_input.prompt,
"parameters": params,
}
output = RequestFuncOutput()
output.prompt_len = request_func_input.prompt_len
ttft = 0.0
st = time.perf_counter()
most_recent_timestamp = st
try:
async with session.post(url=api_url, json=payload) as response:
if response.status == 200:
async for chunk_bytes in response.content:
chunk_bytes = chunk_bytes.strip()
if not chunk_bytes:
continue
chunk = remove_prefix(chunk_bytes.decode("utf-8"),
"data:")
data = json.loads(chunk)
timestamp = time.perf_counter()
# First token
if ttft == 0.0:
ttft = time.perf_counter() - st
output.ttft = ttft
# Decoding phase
else:
output.itl.append(timestamp -
most_recent_timestamp)
most_recent_timestamp = timestamp
output.latency = most_recent_timestamp - st
output.success = True
output.generated_text = data["generated_text"]
else:
output.error = response.reason or ""
output.success = False
except Exception:
output.success = False
exc_info = sys.exc_info()
output.error = "".join(traceback.format_exception(*exc_info))
if pbar:
pbar.update(1)
return output
async def async_request_trt_llm(
request_func_input: RequestFuncInput,
pbar: Optional[tqdm] = None,
) -> RequestFuncOutput:
api_url = request_func_input.api_url
assert api_url.endswith("generate_stream")
async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
assert not request_func_input.use_beam_search
assert request_func_input.best_of == 1
payload = {
"accumulate_tokens": True,
"text_input": request_func_input.prompt,
"temperature": 0.0,
"top_p": 1.0,
"max_tokens": request_func_input.output_len,
"stream": True,
}
output = RequestFuncOutput()
output.prompt_len = request_func_input.prompt_len
ttft = 0.0
st = time.perf_counter()
most_recent_timestamp = st
try:
async with session.post(url=api_url, json=payload) as response:
if response.status == 200:
async for chunk_bytes in response.content:
chunk_bytes = chunk_bytes.strip()
if not chunk_bytes:
continue
chunk = remove_prefix(chunk_bytes.decode("utf-8"),
"data:")
data = json.loads(chunk)
output.generated_text += data["text_output"]
timestamp = time.perf_counter()
# First token
if ttft == 0.0:
ttft = time.perf_counter() - st
output.ttft = ttft
# Decoding phase
else:
output.itl.append(timestamp -
most_recent_timestamp)
most_recent_timestamp = timestamp
output.latency = most_recent_timestamp - st
output.success = True
else:
output.error = response.reason or ""
output.success = False
except Exception:
output.success = False
exc_info = sys.exc_info()
output.error = "".join(traceback.format_exception(*exc_info))
if pbar:
pbar.update(1)
return output
async def async_request_deepspeed_mii(
request_func_input: RequestFuncInput,
pbar: Optional[tqdm] = None,
) -> RequestFuncOutput:
async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
assert request_func_input.best_of == 1
assert not request_func_input.use_beam_search
payload = {
"prompt": request_func_input.prompt,
"max_tokens": request_func_input.output_len,
"temperature": 0.01, # deepspeed-mii does not accept 0.0 temp.
"top_p": 1.0,
}
output = RequestFuncOutput()
output.prompt_len = request_func_input.prompt_len
# NOTE: DeepSpeed-MII doesn't support streaming as of Jan 28 2024,
# will use 0 as placeholder.
# See https://github.com/microsoft/DeepSpeed-MII/pull/311
output.ttft = 0
st = time.perf_counter()
try:
async with session.post(url=request_func_input.api_url,
json=payload) as response:
if response.status == 200:
parsed_resp = await response.json()
output.latency = time.perf_counter() - st
output.generated_text = parsed_resp["text"][0]
output.success = True
else:
output.error = response.reason or ""
output.success = False
except Exception:
output.success = False
exc_info = sys.exc_info()
output.error = "".join(traceback.format_exception(*exc_info))
if pbar:
pbar.update(1)
return output
async def async_request_openai_completions(
request_func_input: RequestFuncInput,
pbar: Optional[tqdm] = None,
) -> RequestFuncOutput:
api_url = request_func_input.api_url
assert api_url.endswith(
"v1/completions"
), "OpenAI Completions API URL must end with 'v1/completions'."
async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
assert not request_func_input.use_beam_search
payload = {
"model": request_func_input.model,
"prompt": request_func_input.prompt,
"temperature": 0.0,
"best_of": request_func_input.best_of,
"max_tokens": request_func_input.output_len,
"stream": True,
}
headers = {
"Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"
}
output = RequestFuncOutput()
output.prompt_len = request_func_input.prompt_len
generated_text = ""
ttft = 0.0
st = time.perf_counter()
most_recent_timestamp = st
try:
async with session.post(url=api_url, json=payload,
headers=headers) as response:
if response.status == 200:
async for chunk_bytes in response.content:
chunk_bytes = chunk_bytes.strip()
if not chunk_bytes:
continue
chunk = remove_prefix(chunk_bytes.decode("utf-8"),
"data: ")
if chunk == "[DONE]":
latency = time.perf_counter() - st
else:
data = json.loads(chunk)
if data["choices"][0]["text"]:
timestamp = time.perf_counter()
# First token
if ttft == 0.0:
ttft = time.perf_counter() - st
output.ttft = ttft
# Decoding phase
# NOTE: Some completion API might have a last
# usage summary response without a token so we
# do not want to include as inter-token-latency
elif data.get("usage", None) is None:
output.itl.append(timestamp -
most_recent_timestamp)
most_recent_timestamp = timestamp
generated_text += data["choices"][0]["text"]
output.generated_text = generated_text
output.success = True
output.latency = latency
else:
output.error = response.reason or ""
output.success = False
except Exception:
output.success = False
exc_info = sys.exc_info()
output.error = "".join(traceback.format_exception(*exc_info))
if pbar:
pbar.update(1)
return output
async def async_request_openai_chat_completions(
request_func_input: RequestFuncInput,
pbar: Optional[tqdm] = None,
) -> RequestFuncOutput:
api_url = request_func_input.api_url
assert api_url.endswith(
"v1/chat/completions"
), "OpenAI Chat Completions API URL must end with 'v1/chat/completions'."
async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
assert not request_func_input.use_beam_search
payload = {
"model": request_func_input.model,
"messages": [
{
"role": "user",
"content": request_func_input.prompt,
},
],
"temperature": 0.0,
"max_tokens": request_func_input.output_len,
"stream": True,
}
headers = {
"Content-Type": "application/json",
"Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}",
}
output = RequestFuncOutput()
output.prompt_len = request_func_input.prompt_len
generated_text = ""
ttft = 0.0
st = time.perf_counter()
most_recent_timestamp = st
try:
async with session.post(url=api_url, json=payload,
headers=headers) as response:
if response.status == 200:
async for chunk_bytes in response.content:
chunk_bytes = chunk_bytes.strip()
if not chunk_bytes:
continue
chunk = remove_prefix(chunk_bytes.decode("utf-8"),
"data: ")
if chunk == "[DONE]":
latency = time.perf_counter() - st
else:
timestamp = time.perf_counter()
data = json.loads(chunk)
delta = data["choices"][0]["delta"]
if delta.get("content", None):
# First token
if ttft == 0.0:
ttft = time.perf_counter() - st
output.ttft = ttft
# Decoding phase
else:
output.itl.append(timestamp -
most_recent_timestamp)
generated_text += delta["content"]
most_recent_timestamp = timestamp
output.generated_text = generated_text
output.success = True
output.latency = latency
else:
output.error = response.reason or ""
output.success = False
except Exception:
output.success = False
exc_info = sys.exc_info()
output.error = "".join(traceback.format_exception(*exc_info))
if pbar:
pbar.update(1)
return output
# Since vllm must support Python 3.8, we can't use str.removeprefix(prefix)
# introduced in Python 3.9
def remove_prefix(text: str, prefix: str) -> str:
if text.startswith(prefix):
return text[len(prefix):]
return text
ASYNC_REQUEST_FUNCS = {
"tgi": async_request_tgi,
"vllm": async_request_openai_completions,
"lmdeploy": async_request_openai_completions,
"deepspeed-mii": async_request_deepspeed_mii,
"openai": async_request_openai_completions,
"openai-chat": async_request_openai_chat_completions,
"tensorrt-llm": async_request_trt_llm,
}
benchmarks/benchmark_latency.py 0 → 100644
View file @ d824b9d3
"""Benchmark the latency of processing a single batch of requests."""
import argparse
import json
import time
from pathlib import Path
from typing import List, Optional
import numpy as np
import torch
from tqdm import tqdm
from vllm import LLM, SamplingParams
from vllm.inputs import PromptStrictInputs
from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
def main(args: argparse.Namespace):
print(args)
# NOTE(woosuk): If the request cannot be processed in a single batch,
# the engine will automatically process the request in multiple batches.
llm = LLM(model=args.model,
speculative_model=args.speculative_model,
num_speculative_tokens=args.num_speculative_tokens,
tokenizer=args.tokenizer,
quantization=args.quantization,
tensor_parallel_size=args.tensor_parallel_size,
trust_remote_code=args.trust_remote_code,
dtype=args.dtype,
enforce_eager=args.enforce_eager,
kv_cache_dtype=args.kv_cache_dtype,
quantization_param_path=args.quantization_param_path,
device=args.device,
ray_workers_use_nsight=args.ray_workers_use_nsight,
use_v2_block_manager=args.use_v2_block_manager,
enable_chunked_prefill=args.enable_chunked_prefill,
download_dir=args.download_dir,
block_size=args.block_size,
gpu_memory_utilization=args.gpu_memory_utilization,
distributed_executor_backend=args.distributed_executor_backend)
sampling_params = SamplingParams(
n=args.n,
temperature=0.0 if args.use_beam_search else 1.0,
top_p=1.0,
use_beam_search=args.use_beam_search,
ignore_eos=True,
max_tokens=args.output_len,
)
print(sampling_params)
dummy_prompt_token_ids = np.random.randint(10000,
size=(args.batch_size,
args.input_len))
dummy_inputs: List[PromptStrictInputs] = [{
"prompt_token_ids": batch
} for batch in dummy_prompt_token_ids.tolist()]
def run_to_completion(profile_dir: Optional[str] = None):
if profile_dir:
with torch.profiler.profile(
activities=[
torch.profiler.ProfilerActivity.CPU,
torch.profiler.ProfilerActivity.CUDA,
],
on_trace_ready=torch.profiler.tensorboard_trace_handler(
str(profile_dir))) as p:
llm.generate(dummy_inputs,
sampling_params=sampling_params,
use_tqdm=False)
print(p.key_averages())
else:
start_time = time.perf_counter()
llm.generate(dummy_inputs,
sampling_params=sampling_params,
use_tqdm=False)
end_time = time.perf_counter()
latency = end_time - start_time
return latency
print("Warming up...")
for _ in tqdm(range(args.num_iters_warmup), desc="Warmup iterations"):
run_to_completion(profile_dir=None)
if args.profile:
profile_dir = args.profile_result_dir
if not profile_dir:
profile_dir = Path(
"."
) / "vllm_benchmark_result" / f"latency_result_{time.time()}"
print(f"Profiling (results will be saved to '{profile_dir}')...")
run_to_completion(profile_dir=profile_dir)
return
# Benchmark.
latencies = []
for _ in tqdm(range(args.num_iters), desc="Profiling iterations"):
latencies.append(run_to_completion(profile_dir=None))
latencies = np.array(latencies)
percentages = [10, 25, 50, 75, 90]
percentiles = np.percentile(latencies, percentages)
print(f'Avg latency: {np.mean(latencies)} seconds')
for percentage, percentile in zip(percentages, percentiles):
print(f'{percentage}% percentile latency: {percentile} seconds')
# Output JSON results if specified
if args.output_json:
results = {
"avg_latency": np.mean(latencies),
"latencies": latencies.tolist(),
"percentiles": dict(zip(percentages, percentiles.tolist())),
}
with open(args.output_json, "w") as f:
json.dump(results, f, indent=4)
if __name__ == '__main__':
parser = argparse.ArgumentParser(
description='Benchmark the latency of processing a single batch of '
'requests till completion.')
parser.add_argument('--model', type=str, default='facebook/opt-125m')
parser.add_argument('--speculative-model', type=str, default=None)
parser.add_argument('--num-speculative-tokens', type=int, default=None)
parser.add_argument('--tokenizer', type=str, default=None)
parser.add_argument('--quantization',
'-q',
choices=[*QUANTIZATION_METHODS, None],
default=None)
parser.add_argument('--tensor-parallel-size', '-tp', type=int, default=1)
parser.add_argument('--input-len', type=int, default=32)
parser.add_argument('--output-len', type=int, default=128)
parser.add_argument('--batch-size', type=int, default=8)
parser.add_argument('--n',
type=int,
default=1,
help='Number of generated sequences per prompt.')
parser.add_argument('--use-beam-search', action='store_true')
parser.add_argument('--num-iters-warmup',
type=int,
default=10,
help='Number of iterations to run for warmup.')
parser.add_argument('--num-iters',
type=int,
default=30,
help='Number of iterations to run.')
parser.add_argument('--trust-remote-code',
action='store_true',
help='trust remote code from huggingface')
parser.add_argument(
'--dtype',
type=str,
default='auto',
choices=['auto', 'half', 'float16', 'bfloat16', 'float', 'float32'],
help='data type for model weights and activations. '
'The "auto" option will use FP16 precision '
'for FP32 and FP16 models, and BF16 precision '
'for BF16 models.')
parser.add_argument('--enforce-eager',
action='store_true',
help='enforce eager mode and disable CUDA graph')
parser.add_argument(
'--kv-cache-dtype',
type=str,
choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
default="auto",
help='Data type for kv cache storage. If "auto", will use model '
'data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. '
'ROCm (AMD GPU) supports fp8 (=fp8_e4m3)')
parser.add_argument(
'--quantization-param-path',
type=str,
default=None,
help='Path to the JSON file containing the KV cache scaling factors. '
'This should generally be supplied, when KV cache dtype is FP8. '
'Otherwise, KV cache scaling factors default to 1.0, which may cause '
'accuracy issues. FP8_E5M2 (without scaling) is only supported on '
'cuda version greater than 11.8. On ROCm (AMD GPU), FP8_E4M3 is '
'instead supported for common inference criteria.')
parser.add_argument(
'--profile',
action='store_true',
help='profile the generation process of a single batch')
parser.add_argument(
'--profile-result-dir',
type=str,
default=None,
help=('path to save the pytorch profiler output. Can be visualized '
'with ui.perfetto.dev or Tensorboard.'))
parser.add_argument(
"--device",
type=str,
default="cuda",
choices=["cuda", "cpu"],
help='device type for vLLM execution, supporting CUDA and CPU.')
parser.add_argument('--block-size',
type=int,
default=16,
help='block size of key/value cache')
parser.add_argument(
'--enable-chunked-prefill',
action='store_true',
help='If True, the prefill requests can be chunked based on the '
'max_num_batched_tokens')
parser.add_argument('--use-v2-block-manager', action='store_true')
parser.add_argument(
"--ray-workers-use-nsight",
action='store_true',
help="If specified, use nsight to profile ray workers",
)
parser.add_argument('--download-dir',
type=str,
default=None,
help='directory to download and load the weights, '
'default to the default cache dir of huggingface')
parser.add_argument(
'--output-json',
type=str,
default=None,
help='Path to save the latency results in JSON format.')
parser.add_argument('--gpu-memory-utilization',
type=float,
default=0.9,
help='the fraction of GPU memory to be used for '
'the model executor, which can range from 0 to 1.'
'If unspecified, will use the default value of 0.9.')
parser.add_argument(
'--distributed-executor-backend',
choices=['ray', 'mp'],
default=None,
help='Backend to use for distributed serving. When more than 1 GPU '
'is used, will be automatically set to "ray" if installed '
'or "mp" (multiprocessing) otherwise.')
args = parser.parse_args()
main(args)
benchmarks/benchmark_prefix_caching.py 0 → 100644
View file @ d824b9d3
import argparse
import time
from vllm import LLM, SamplingParams
PROMPT = "You are a helpful assistant in recognizes the content of tables in markdown format. Here is a table as fellows. You need to answer my question about the table.\n# Table\n|Opening|Opening|Sl. No.|Film|Cast|Director|Music Director|Notes|\n|----|----|----|----|----|----|----|----|\n|J A N|9|1|Agni Pushpam|Jayabharathi, Kamalahasan|Jeassy|M. K. Arjunan||\n|J A N|16|2|Priyamvada|Mohan Sharma, Lakshmi, KPAC Lalitha|K. S. Sethumadhavan|V. Dakshinamoorthy||\n|J A N|23|3|Yakshagaanam|Madhu, Sheela|Sheela|M. S. Viswanathan||\n|J A N|30|4|Paalkkadal|Sheela, Sharada|T. K. Prasad|A. T. Ummer||\n|F E B|5|5|Amma|Madhu, Srividya|M. Krishnan Nair|M. K. Arjunan||\n|F E B|13|6|Appooppan|Thikkurissi Sukumaran Nair, Kamal Haasan|P. Bhaskaran|M. S. Baburaj||\n|F E B|20|7|Srishti|Chowalloor Krishnankutty, Ravi Alummoodu|K. T. Muhammad|M. S. Baburaj||\n|F E B|20|8|Vanadevatha|Prem Nazir, Madhubala|Yusufali Kechery|G. Devarajan||\n|F E B|27|9|Samasya|Madhu, Kamalahaasan|K. Thankappan|Shyam||\n|F E B|27|10|Yudhabhoomi|K. P. Ummer, Vidhubala|Crossbelt Mani|R. K. Shekhar||\n|M A R|5|11|Seemantha Puthran|Prem Nazir, Jayabharathi|A. B. Raj|M. K. Arjunan||\n|M A R|12|12|Swapnadanam|Rani Chandra, Dr. Mohandas|K. G. George|Bhaskar Chandavarkar||\n|M A R|19|13|Thulavarsham|Prem Nazir, sreedevi, Sudheer|N. Sankaran Nair|V. Dakshinamoorthy||\n|M A R|20|14|Aruthu|Kaviyoor Ponnamma, Kamalahasan|Ravi|G. Devarajan||\n|M A R|26|15|Swimming Pool|Kamal Haasan, M. G. Soman|J. Sasikumar|M. K. Arjunan||\n\n# Question\nWhat' s the content in the (1,1) cells\n" # noqa: E501
def test_prefix(llm=None, sampling_params=None, prompts=None):
start_time = time.time()
llm.generate(prompts, sampling_params=sampling_params)
end_time = time.time()
print(f"cost time {end_time - start_time}")
def main(args):
llm = LLM(model=args.model,
tokenizer_mode='auto',
trust_remote_code=True,
enforce_eager=True,
use_v2_block_manager=args.use_v2_block_manager,
tensor_parallel_size=args.tensor_parallel_size,
enable_prefix_caching=args.enable_prefix_caching)
num_prompts = 100
prompts = [PROMPT] * num_prompts
sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len)
print("------warm up------")
test_prefix(
llm=llm,
prompts=prompts,
sampling_params=sampling_params,
)
print("------start generating------")
test_prefix(
llm=llm,
prompts=prompts,
sampling_params=sampling_params,
)
if __name__ == "__main__":
parser = argparse.ArgumentParser(
description='Benchmark the performance with or without automatic '
'prefix caching.')
parser.add_argument('--model',
type=str,
default='baichuan-inc/Baichuan2-13B-Chat')
parser.add_argument('--tensor-parallel-size', '-tp', type=int, default=1)
parser.add_argument('--output-len', type=int, default=10)
parser.add_argument('--enable-prefix-caching',
action='store_true',
help='enable prefix caching')
parser.add_argument('--use-v2-block-manager',
action='store_true',
help='Use BlockSpaceMangerV2')
args = parser.parse_args()
main(args)
benchmarks/benchmark_serving.py 0 → 100644
View file @ d824b9d3
"""Benchmark online serving throughput.
On the server side, run one of the following commands:
vLLM OpenAI API server
python -m vllm.entrypoints.openai.api_server \
--model <your_model> --swap-space 16 \
--disable-log-requests
(TGI backend)
./launch_tgi_server.sh <your_model> <max_batch_total_tokens>
On the client side, run:
python benchmarks/benchmark_serving.py \
--backend <backend> \
--model <your_model> \
--dataset-name sharegpt \
--dataset-path <path to dataset> \
--request-rate <request_rate> \ # By default <request_rate> is inf
--num-prompts <num_prompts> # By default <num_prompts> is 1000
when using tgi backend, add
--endpoint /generate_stream
to the end of the command above.
"""
import argparse
import asyncio
import json
import os
import random
import time
import warnings
from dataclasses import dataclass
from datetime import datetime
from typing import AsyncGenerator, List, Optional, Tuple
import numpy as np
from backend_request_func import (ASYNC_REQUEST_FUNCS, RequestFuncInput,
RequestFuncOutput)
from tqdm.asyncio import tqdm
from transformers import PreTrainedTokenizerBase
from vllm.transformers_utils.tokenizer import get_tokenizer
@dataclass
class BenchmarkMetrics:
completed: int
total_input: int
total_output: int
request_throughput: float
input_throughput: float
output_throughput: float
mean_ttft_ms: float
median_ttft_ms: float
p99_ttft_ms: float
mean_tpot_ms: float
median_tpot_ms: float
p99_tpot_ms: float
mean_itl_ms: float
median_itl_ms: float
p99_itl_ms: float
def sample_sharegpt_requests(
dataset_path: str,
num_requests: int,
tokenizer: PreTrainedTokenizerBase,
fixed_output_len: Optional[int] = None,
) -> List[Tuple[str, int, int]]:
if fixed_output_len is not None and fixed_output_len < 4:
raise ValueError("output_len too small")
# Load the dataset.
with open(dataset_path) as f:
dataset = json.load(f)
# Filter out the conversations with less than 2 turns.
dataset = [data for data in dataset if len(data["conversations"]) >= 2]
# Only keep the first two turns of each conversation.
dataset = [(data["conversations"][0]["value"],
data["conversations"][1]["value"]) for data in dataset]
# Shuffle the dataset.
random.shuffle(dataset)
# Filter out sequences that are too long or too short
filtered_dataset: List[Tuple[str, int, int]] = []
for i in range(len(dataset)):
if len(filtered_dataset) == num_requests:
break
# Tokenize the prompts and completions.
prompt = dataset[i][0]
prompt_token_ids = tokenizer(prompt).input_ids
completion = dataset[i][1]
completion_token_ids = tokenizer(completion).input_ids
prompt_len = len(prompt_token_ids)
output_len = len(completion_token_ids
) if fixed_output_len is None else fixed_output_len
if prompt_len < 4 or output_len < 4:
# Prune too short sequences.
continue
if prompt_len > 1024 or prompt_len + output_len > 2048:
# Prune too long sequences.
continue
filtered_dataset.append((prompt, prompt_len, output_len))
return filtered_dataset
def sample_sonnet_requests(
dataset_path: str,
num_requests: int,
input_len: int,
output_len: int,
prefix_len: int,
tokenizer: PreTrainedTokenizerBase,
) -> List[Tuple[str, str, int, int]]:
assert (
input_len > prefix_len
), "'args.sonnet-input-len' must be greater than 'args.prefix-input-len'."
# Load the dataset.
with open(dataset_path) as f:
poem_lines = f.readlines()
# Tokenize the poem lines.
poem_token_ids = tokenizer(poem_lines).input_ids
average_poem_len = sum(
len(token_ids) for token_ids in poem_token_ids) / len(poem_token_ids)
# Base prefix for all requests.
base_prompt = "Pick as many lines as you can from these poem lines:\n"
base_message = [{
"role": "user",
"content": base_prompt,
}]
base_prompt_formatted = tokenizer.apply_chat_template(
base_message, add_generation_prompt=True, tokenize=False)
base_prompt_offset = len(tokenizer(base_prompt_formatted).input_ids)
assert (
input_len > base_prompt_offset
), f"Please set 'args.sonnet-input-len' higher than {base_prompt_offset}."
num_input_lines = round(
(input_len - base_prompt_offset) / average_poem_len)
# First approximately `prefix_len` number of tokens in the
# prompt are fixed poem lines.
assert (
prefix_len > base_prompt_offset
), f"Please set 'args.sonnet-prefix-len' higher than {base_prompt_offset}."
num_prefix_lines = round(
(prefix_len - base_prompt_offset) / average_poem_len)
prefix_lines = poem_lines[:num_prefix_lines]
# Sample the rest of lines per request.
sampled_requests: List[Tuple[str, int, int]] = []
for _ in range(num_requests):
sampled_lines = "".join(
prefix_lines +
random.sample(poem_lines, num_input_lines - num_prefix_lines))
prompt = f"{base_prompt}{sampled_lines}"
message = [
{
"role": "user",
"content": prompt,
},
]
prompt_formatted = tokenizer.apply_chat_template(
message, add_generation_prompt=True, tokenize=False)
prompt_len = len(tokenizer(prompt_formatted).input_ids)
sampled_requests.append(
(prompt, prompt_formatted, prompt_len, output_len))
return sampled_requests
async def get_request(
input_requests: List[Tuple[str, int, int]],
request_rate: float,
) -> AsyncGenerator[Tuple[str, int, int], None]:
input_requests = iter(input_requests)
for request in input_requests:
yield request
if request_rate == float("inf"):
# If the request rate is infinity, then we don't need to wait.
continue
# Sample the request interval from the exponential distribution.
interval = np.random.exponential(1.0 / request_rate)
# The next request will be sent after the interval.
await asyncio.sleep(interval)
def calculate_metrics(
input_requests: List[Tuple[str, int, int]],
outputs: List[RequestFuncOutput],
dur_s: float,
tokenizer: PreTrainedTokenizerBase,
) -> Tuple[BenchmarkMetrics, List[int]]:
actual_output_lens = []
total_input = 0
completed = 0
itls = []
tpots = []
ttfts = []
for i in range(len(outputs)):
if outputs[i].success:
# We use the tokenizer to count the number of output tokens for all
# serving backends instead of looking at len(outputs[i].itl) since
# multiple output tokens may be bundled together
# Note: this may inflate the output token count slightly
output_len = len(
tokenizer(outputs[i].generated_text,
add_special_tokens=False).input_ids)
actual_output_lens.append(output_len)
total_input += input_requests[i][1]
if output_len > 1:
tpots.append(
(outputs[i].latency - outputs[i].ttft) / (output_len - 1))
itls += outputs[i].itl
ttfts.append(outputs[i].ttft)
completed += 1
else:
actual_output_lens.append(0)
if completed == 0:
warnings.warn(
"All requests failed. This is likely due to a misconfiguration "
"on the benchmark arguments.",
stacklevel=2)
metrics = BenchmarkMetrics(
completed=completed,
total_input=total_input,
total_output=sum(actual_output_lens),
request_throughput=completed / dur_s,
input_throughput=total_input / dur_s,
output_throughput=sum(actual_output_lens) / dur_s,
mean_ttft_ms=np.mean(ttfts or 0) *
1000, # ttfts is empty if streaming is not supported by backend
median_ttft_ms=np.median(ttfts or 0) * 1000,
p99_ttft_ms=np.percentile(ttfts or 0, 99) * 1000,
mean_tpot_ms=np.mean(tpots or 0) * 1000,
median_tpot_ms=np.median(tpots or 0) * 1000,
p99_tpot_ms=np.percentile(tpots or 0, 99) * 1000,
mean_itl_ms=np.mean(itls or 0) * 1000,
median_itl_ms=np.median(itls or 0) * 1000,
p99_itl_ms=np.percentile(itls or 0, 99) * 1000,
)
return metrics, actual_output_lens
async def benchmark(
backend: str,
api_url: str,
model_id: str,
tokenizer: PreTrainedTokenizerBase,
input_requests: List[Tuple[str, int, int]],
best_of: int,
use_beam_search: bool,
request_rate: float,
disable_tqdm: bool,
):
if backend in ASYNC_REQUEST_FUNCS:
request_func = ASYNC_REQUEST_FUNCS.get(backend)
else:
raise ValueError(f"Unknown backend: {backend}")
print("Starting initial single prompt test run...")
test_prompt, test_prompt_len, test_output_len = input_requests[0]
test_input = RequestFuncInput(
model=model_id,
prompt=test_prompt,
api_url=api_url,
prompt_len=test_prompt_len,
output_len=test_output_len,
best_of=best_of,
use_beam_search=use_beam_search,
)
test_output = await request_func(request_func_input=test_input)
if not test_output.success:
raise ValueError(
"Initial test run failed - Please make sure benchmark arguments "
f"are correctly specified. Error: {test_output.error}")
else:
print("Initial test run completed. Starting main benchmark run...")
print(f"Traffic request rate: {request_rate}")
pbar = None if disable_tqdm else tqdm(total=len(input_requests))
benchmark_start_time = time.perf_counter()
tasks = []
async for request in get_request(input_requests, request_rate):
prompt, prompt_len, output_len = request
request_func_input = RequestFuncInput(
model=model_id,
prompt=prompt,
api_url=api_url,
prompt_len=prompt_len,
output_len=output_len,
best_of=best_of,
use_beam_search=use_beam_search,
)
tasks.append(
asyncio.create_task(
request_func(request_func_input=request_func_input,
pbar=pbar)))
outputs: List[RequestFuncOutput] = await asyncio.gather(*tasks)
if not disable_tqdm:
pbar.close()
benchmark_duration = time.perf_counter() - benchmark_start_time
metrics, actual_output_lens = calculate_metrics(
input_requests=input_requests,
outputs=outputs,
dur_s=benchmark_duration,
tokenizer=tokenizer,
)
print("{s:{c}^{n}}".format(s=' Serving Benchmark Result ', n=50, c='='))
print("{:<40} {:<10}".format("Successful requests:", metrics.completed))
print("{:<40} {:<10.2f}".format("Benchmark duration (s):",
benchmark_duration))
print("{:<40} {:<10}".format("Total input tokens:", metrics.total_input))
print("{:<40} {:<10}".format("Total generated tokens:",
metrics.total_output))
print("{:<40} {:<10.2f}".format("Request throughput (req/s):",
metrics.request_throughput))
print("{:<40} {:<10.2f}".format("Input token throughput (tok/s):",
metrics.input_throughput))
print("{:<40} {:<10.2f}".format("Output token throughput (tok/s):",
metrics.output_throughput))
print("{s:{c}^{n}}".format(s='Time to First Token', n=50, c='-'))
print("{:<40} {:<10.2f}".format("Mean TTFT (ms):", metrics.mean_ttft_ms))
print("{:<40} {:<10.2f}".format("Median TTFT (ms):",
metrics.median_ttft_ms))
print("{:<40} {:<10.2f}".format("P99 TTFT (ms):", metrics.p99_ttft_ms))
print("{s:{c}^{n}}".format(s='Time per Output Token (excl. 1st token)',
n=50,
c='-'))
print("{:<40} {:<10.2f}".format("Mean TPOT (ms):", metrics.mean_tpot_ms))
print("{:<40} {:<10.2f}".format("Median TPOT (ms):",
metrics.median_tpot_ms))
print("{:<40} {:<10.2f}".format("P99 TPOT (ms):", metrics.p99_tpot_ms))
print("{s:{c}^{n}}".format(s='Inter-token Latency', n=50, c='-'))
print("{:<40} {:<10.2f}".format("Mean ITL (ms):", metrics.mean_itl_ms))
print("{:<40} {:<10.2f}".format("Median ITL (ms):", metrics.median_itl_ms))
print("{:<40} {:<10.2f}".format("P99 ITL (ms):", metrics.p99_itl_ms))
print("=" * 50)
result = {
"duration": benchmark_duration,
"completed": metrics.completed,
"total_input_tokens": metrics.total_input,
"total_output_tokens": metrics.total_output,
"request_throughput": metrics.request_throughput,
"input_throughput": metrics.input_throughput,
"output_throughput": metrics.output_throughput,
"mean_ttft_ms": metrics.mean_ttft_ms,
"median_ttft_ms": metrics.median_ttft_ms,
"p99_ttft_ms": metrics.p99_ttft_ms,
"mean_tpot_ms": metrics.mean_tpot_ms,
"median_tpot_ms": metrics.median_tpot_ms,
"p99_tpot_ms": metrics.p99_tpot_ms,
"mean_itl_ms": metrics.mean_itl_ms,
"median_itl_ms": metrics.median_itl_ms,
"p99_itl_ms": metrics.p99_itl_ms,
"input_lens": [output.prompt_len for output in outputs],
"output_lens": actual_output_lens,
"ttfts": [output.ttft for output in outputs],
"itls": [output.itl for output in outputs],
"generated_texts": [output.generated_text for output in outputs],
"errors": [output.error for output in outputs],
}
return result
def main(args: argparse.Namespace):
print(args)
random.seed(args.seed)
np.random.seed(args.seed)
backend = args.backend
model_id = args.model
tokenizer_id = args.tokenizer if args.tokenizer is not None else args.model
if args.base_url is not None:
api_url = f"{args.base_url}{args.endpoint}"
else:
api_url = f"http://{args.host}:{args.port}{args.endpoint}"
tokenizer = get_tokenizer(tokenizer_id,
trust_remote_code=args.trust_remote_code)
if args.dataset is not None:
warnings.warn(
"The '--dataset' argument will be deprecated in the next "
"release. Please use '--dataset-name' and "
"'--dataset-path' in the future runs.",
stacklevel=2)
input_requests = sample_sharegpt_requests(
dataset_path=args.dataset,
num_requests=args.num_prompts,
tokenizer=tokenizer,
fixed_output_len=args.sharegpt_output_len,
)
elif args.dataset_name == "sharegpt":
input_requests = sample_sharegpt_requests(
dataset_path=args.dataset_path,
num_requests=args.num_prompts,
tokenizer=tokenizer,
fixed_output_len=args.sharegpt_output_len,
)
elif args.dataset_name == "sonnet":
# Do not format the prompt, pass to message directly
if args.backend == "openai-chat":
input_requests = sample_sonnet_requests(
dataset_path=args.dataset_path,
num_requests=args.num_prompts,
input_len=args.sonnet_input_len,
output_len=args.sonnet_output_len,
prefix_len=args.sonnet_prefix_len,
tokenizer=tokenizer,
)
input_requests = [(prompt, prompt_len, output_len)
for prompt, prompt_formatted, prompt_len,
output_len in input_requests]
else:
assert (
tokenizer.chat_template or tokenizer.default_chat_template
), "Tokenizer/model must have chat template for sonnet dataset."
input_requests = sample_sonnet_requests(
dataset_path=args.dataset_path,
num_requests=args.num_prompts,
input_len=args.sonnet_input_len,
output_len=args.sonnet_output_len,
prefix_len=args.sonnet_prefix_len,
tokenizer=tokenizer,
)
input_requests = [(prompt_formatted, prompt_len, output_len)
for prompt, prompt_formatted, prompt_len,
output_len in input_requests]
else:
raise ValueError(f"Unknown dataset: {args.dataset_name}")
benchmark_result = asyncio.run(
benchmark(
backend=backend,
api_url=api_url,
model_id=model_id,
tokenizer=tokenizer,
input_requests=input_requests,
best_of=args.best_of,
use_beam_search=args.use_beam_search,
request_rate=args.request_rate,
disable_tqdm=args.disable_tqdm,
))
# Save config and results to json
if args.save_result:
result_json = {}
# Setup
current_dt = datetime.now().strftime("%Y%m%d-%H%M%S")
result_json["date"] = current_dt
result_json["backend"] = backend
result_json["model_id"] = model_id
result_json["tokenizer_id"] = tokenizer_id
result_json["best_of"] = args.best_of
result_json["use_beam_search"] = args.use_beam_search
result_json["num_prompts"] = args.num_prompts
# Metadata
if args.metadata:
for item in args.metadata:
if "=" in item:
kvstring = item.split("=")
result_json[kvstring[0].strip()] = kvstring[1].strip()
else:
raise ValueError(
"Invalid metadata format. Please use KEY=VALUE format."
)
# Traffic
result_json["request_rate"] = (
args.request_rate if args.request_rate < float("inf") else "inf")
# Merge with benchmark result
result_json = {**result_json, **benchmark_result}
# Save to file
base_model_id = model_id.split("/")[-1]
file_name = f"{backend}-{args.request_rate}qps-{base_model_id}-{current_dt}.json" #noqa
if args.result_dir:
file_name = os.path.join(args.result_dir, file_name)
with open(file_name, "w") as outfile:
json.dump(result_json, outfile)
if __name__ == "__main__":
parser = argparse.ArgumentParser(
description="Benchmark the online serving throughput.")
parser.add_argument(
"--backend",
type=str,
default="vllm",
choices=list(ASYNC_REQUEST_FUNCS.keys()),
)
parser.add_argument(
"--base-url",
type=str,
default=None,
help="Server or API base url if not using http host and port.",
)
parser.add_argument("--host", type=str, default="localhost")
parser.add_argument("--port", type=int, default=8000)
parser.add_argument(
"--endpoint",
type=str,
default="/v1/completions",
help="API endpoint.",
)
parser.add_argument(
"--dataset",
type=str,
default=None,
help="Path to the ShareGPT dataset, will be deprecated in the "
"next release.",
)
parser.add_argument(
"--dataset-name",
type=str,
default="sharegpt",
choices=["sharegpt", "sonnet"],
help="Name of the dataset to benchmark on.",
)
parser.add_argument("--dataset-path",
type=str,
default=None,
help="Path to the dataset.")
parser.add_argument(
"--model",
type=str,
required=True,
help="Name of the model.",
)
parser.add_argument(
"--tokenizer",
type=str,
help=
"Name or path of the tokenizer, if not using the default tokenizer.",
)
parser.add_argument(
"--best-of",
type=int,
default=1,
help="Generates `best_of` sequences per prompt and "
"returns the best one.",
)
parser.add_argument("--use-beam-search", action="store_true")
parser.add_argument(
"--num-prompts",
type=int,
default=1000,
help="Number of prompts to process.",
)
parser.add_argument(
"--sharegpt-output-len",
type=int,
default=None,
help="Output length for each request. Overrides the output length "
"from the ShareGPT dataset.")
parser.add_argument(
"--sonnet-input-len",
type=int,
default=550,
help=
"Number of input tokens per request, used only for sonnet dataset.",
)
parser.add_argument(
"--sonnet-output-len",
type=int,
default=150,
help=
"Number of output tokens per request, used only for sonnet dataset.",
)
parser.add_argument(
"--sonnet-prefix-len",
type=int,
default=200,
help=
"Number of prefix tokens per request, used only for sonnet dataset.",
)
parser.add_argument(
"--request-rate",
type=float,
default=float("inf"),
help="Number of requests per second. If this is inf, "
"then all the requests are sent at time 0. "
"Otherwise, we use Poisson process to synthesize "
"the request arrival times.",
)
parser.add_argument("--seed", type=int, default=0)
parser.add_argument(
"--trust-remote-code",
action="store_true",
help="Trust remote code from huggingface",
)
parser.add_argument(
"--disable-tqdm",
action="store_true",
help="Specify to disable tqdm progress bar.",
)
parser.add_argument(
"--save-result",
action="store_true",
help="Specify to save benchmark results to a json file",
)
parser.add_argument(
"--metadata",
metavar="KEY=VALUE",
nargs="*",
help="Key-value pairs (e.g, --metadata version=0.3.3 tp=1) "
"for metadata of this run to be saved in the result JSON file "
"for record keeping purposes.",
)
parser.add_argument(
"--result-dir",
type=str,
default=None,
help="Specify directory to save benchmark json results."
"If not specified, results are saved in the current directory.",
)
args = parser.parse_args()
main(args)
benchmarks/benchmark_throughput.py 0 → 100644
View file @ d824b9d3
"""Benchmark offline inference throughput."""
import argparse
import json
import random
import time
from typing import List, Optional, Tuple
import numpy as np
import torch
from tqdm import tqdm
from transformers import (AutoModelForCausalLM, AutoTokenizer,
PreTrainedTokenizerBase)
from vllm.inputs import PromptStrictInputs
from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
def sample_requests(
dataset_path: str,
num_requests: int,
tokenizer: PreTrainedTokenizerBase,
fixed_output_len: Optional[int],
) -> List[Tuple[str, int, int]]:
if fixed_output_len is not None and fixed_output_len < 4:
raise ValueError("output_len too small")
# Load the dataset.
with open(dataset_path) as f:
dataset = json.load(f)
# Filter out the conversations with less than 2 turns.
dataset = [data for data in dataset if len(data["conversations"]) >= 2]
# Only keep the first two turns of each conversation.
dataset = [(data["conversations"][0]["value"],
data["conversations"][1]["value"]) for data in dataset]
# Shuffle the dataset.
random.shuffle(dataset)
# Filter out sequences that are too long or too short
filtered_dataset: List[Tuple[str, int, int]] = []
for i in range(len(dataset)):
if len(filtered_dataset) == num_requests:
break
# Tokenize the prompts and completions.
prompt = dataset[i][0]
prompt_token_ids = tokenizer(prompt).input_ids
completion = dataset[i][1]
completion_token_ids = tokenizer(completion).input_ids
prompt_len = len(prompt_token_ids)
output_len = len(completion_token_ids
) if fixed_output_len is None else fixed_output_len
if prompt_len < 4 or output_len < 4:
# Prune too short sequences.
continue
if prompt_len > 1024 or prompt_len + output_len > 2048:
# Prune too long sequences.
continue
filtered_dataset.append((prompt, prompt_len, output_len))
return filtered_dataset
def run_vllm(
warmup_requests: List[Tuple[str, int, int]],
requests: List[Tuple[str, int, int]],
model: str,
tokenizer: str,
quantization: Optional[str],
tensor_parallel_size: int,
seed: int,
n: int,
use_beam_search: bool,
trust_remote_code: bool,
dtype: str,
max_model_len: Optional[int],
enforce_eager: bool,
kv_cache_dtype: str,
quantization_param_path: Optional[str],
device: str,
enable_prefix_caching: bool,
enable_chunked_prefill: bool,
max_num_batched_tokens: int,
distributed_executor_backend: Optional[str],
gpu_memory_utilization: float = 0.9,
download_dir: Optional[str] = None,
) -> float:
from vllm import LLM, SamplingParams
llm = LLM(
model=model,
tokenizer=tokenizer,
quantization=quantization,
tensor_parallel_size=tensor_parallel_size,
seed=seed,
trust_remote_code=trust_remote_code,
dtype=dtype,
max_model_len=max_model_len,
gpu_memory_utilization=gpu_memory_utilization,
enforce_eager=enforce_eager,
kv_cache_dtype=kv_cache_dtype,
quantization_param_path=quantization_param_path,
device=device,
enable_prefix_caching=enable_prefix_caching,
download_dir=download_dir,
enable_chunked_prefill=enable_chunked_prefill,
max_num_batched_tokens=max_num_batched_tokens,
distributed_executor_backend=distributed_executor_backend,
)
# Add the requests to the engine.
prompts = []
sampling_params = []
for prompt, _, output_len in requests:
prompts.append(prompt)
sampling_params.append(
SamplingParams(
n=n,
temperature=0.0 if use_beam_search else 1.0,
top_p=1.0,
use_beam_search=use_beam_search,
ignore_eos=True,
max_tokens=output_len,
))
# warmup
warmup_prompts = []
warmup_sampling_params = []
for prompt, _, output_len in warmup_requests:
warmup_prompts.append(prompt)
warmup_sampling_params.append(
SamplingParams(
n=n,
temperature=0.0 if use_beam_search else 1.0,
top_p=1.0,
use_beam_search=use_beam_search,
ignore_eos=True,
max_tokens=output_len,
))
print("Warming up...")
for _ in tqdm(range(args.num_iters_warmup), desc="Warmup iterations"):
llm.generate(warmup_prompts, warmup_sampling_params, use_tqdm=True)
# dummy_prompt_token_ids = np.random.randint(10000,
# size=(args.num_prompts,
# args.input_len))
# dummy_inputs: List[PromptStrictInputs] = [{
# "prompt_token_ids": batch
# } for batch in dummy_prompt_token_ids.tolist()]
# def run_to_completion():
# llm.generate(dummy_inputs,
# sampling_params=sampling_params,
# use_tqdm=False)
# print("Warming up...")
# for _ in tqdm(range(args.num_iters_warmup), desc="Warmup iterations"):
# run_to_completion()
start = time.perf_counter()
llm.generate(prompts, sampling_params, use_tqdm=True)
end = time.perf_counter()
return end - start
def run_hf(
requests: List[Tuple[str, int, int]],
model: str,
tokenizer: PreTrainedTokenizerBase,
n: int,
use_beam_search: bool,
max_batch_size: int,
trust_remote_code: bool,
) -> float:
assert not use_beam_search
llm = AutoModelForCausalLM.from_pretrained(
model, torch_dtype=torch.float16, trust_remote_code=trust_remote_code)
if llm.config.model_type == "llama":
# To enable padding in the HF backend.
tokenizer.pad_token = tokenizer.eos_token
llm = llm.cuda()
pbar = tqdm(total=len(requests))
start = time.perf_counter()
batch: List[str] = []
max_prompt_len = 0
max_output_len = 0
for i in range(len(requests)):
prompt, prompt_len, output_len = requests[i]
# Add the prompt to the batch.
batch.append(prompt)
max_prompt_len = max(max_prompt_len, prompt_len)
max_output_len = max(max_output_len, output_len)
if len(batch) < max_batch_size and i != len(requests) - 1:
# Check if we can add more requests to the batch.
_, next_prompt_len, next_output_len = requests[i + 1]
if (max(max_prompt_len, next_prompt_len) +
max(max_output_len, next_output_len)) <= 2048:
# We can add more requests to the batch.
continue
# Generate the sequences.
input_ids = tokenizer(batch, return_tensors="pt",
padding=True).input_ids
llm_outputs = llm.generate(
input_ids=input_ids.cuda(),
do_sample=not use_beam_search,
num_return_sequences=n,
temperature=1.0,
top_p=1.0,
use_cache=True,
max_new_tokens=max_output_len,
)
# Include the decoding time.
tokenizer.batch_decode(llm_outputs, skip_special_tokens=True)
pbar.update(len(batch))
# Clear the batch.
batch = []
max_prompt_len = 0
max_output_len = 0
end = time.perf_counter()
return end - start
def run_mii(
requests: List[Tuple[str, int, int]],
model: str,
tensor_parallel_size: int,
output_len: int,
) -> float:
from mii import client, serve
llm = serve(model, tensor_parallel=tensor_parallel_size)
prompts = [prompt for prompt, _, _ in requests]
start = time.perf_counter()
llm.generate(prompts, max_new_tokens=output_len)
end = time.perf_counter()
client = client(model)
client.terminate_server()
return end - start
def main(args: argparse.Namespace):
print(args)
random.seed(args.seed)
# Sample the requests.
tokenizer = AutoTokenizer.from_pretrained(
args.tokenizer, trust_remote_code=args.trust_remote_code)
warmup_prompt = "hi" * 10
warmup_requests = [(warmup_prompt, 10, 10)
for _ in range(1)]
if args.dataset is None:
# Synthesize a prompt with the given input length.
prompt = "hi" * (args.input_len - 1)
requests = [(prompt, args.input_len, args.output_len)
for _ in range(args.num_prompts)]
else:
requests = sample_requests(args.dataset, args.num_prompts, tokenizer,
args.output_len)
if args.backend == "vllm":
elapsed_time = run_vllm(
warmup_requests, requests, args.model, args.tokenizer, args.quantization,
args.tensor_parallel_size, args.seed, args.n, args.use_beam_search,
args.trust_remote_code, args.dtype, args.max_model_len,
args.enforce_eager, args.kv_cache_dtype,
args.quantization_param_path, args.device,
args.enable_prefix_caching, args.enable_chunked_prefill,
args.max_num_batched_tokens, args.distributed_executor_backend,
args.gpu_memory_utilization, args.download_dir)
elif args.backend == "hf":
assert args.tensor_parallel_size == 1
elapsed_time = run_hf(requests, args.model, tokenizer, args.n,
args.use_beam_search, args.hf_max_batch_size,
args.trust_remote_code)
elif args.backend == "mii":
elapsed_time = run_mii(requests, args.model, args.tensor_parallel_size,
args.output_len)
else:
raise ValueError(f"Unknown backend: {args.backend}")
total_num_tokens = sum(prompt_len + output_len
for _, prompt_len, output_len in requests)
if args.dataset is None:
total_out_tokens = args.output_len * args.num_prompts
else:
total_out_tokens = sum(output_len for _, _, output_len in requests)
print(f"Latency: {elapsed_time:.2f} s")
print(f"All Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
f"{total_num_tokens / elapsed_time:.2f} tokens/s")
print(f"Generate Throughput: {total_out_tokens / elapsed_time:.2f} tokens/s")
# Output JSON results if specified
if args.output_json:
results = {
"elapsed_time": elapsed_time,
"num_requests": len(requests),
"total_num_tokens": total_num_tokens,
"requests_per_second": len(requests) / elapsed_time,
"tokens_per_second": total_num_tokens / elapsed_time,
}
with open(args.output_json, "w") as f:
json.dump(results, f, indent=4)
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Benchmark the throughput.")
parser.add_argument("--backend",
type=str,
choices=["vllm", "hf", "mii"],
default="vllm")
parser.add_argument("--dataset",
type=str,
default=None,
help="Path to the dataset.")
parser.add_argument("--input-len",
type=int,
default=None,
help="Input prompt length for each request")
parser.add_argument("--output-len",
type=int,
default=None,
help="Output length for each request. Overrides the "
"output length from the dataset.")
parser.add_argument("--model", type=str, default="facebook/opt-125m")
parser.add_argument("--tokenizer", type=str, default=None)
parser.add_argument('--quantization',
'-q',
choices=[*QUANTIZATION_METHODS, None],
default=None)
parser.add_argument("--tensor-parallel-size", "-tp", type=int, default=1)
parser.add_argument("--n",
type=int,
default=1,
help="Number of generated sequences per prompt.")
parser.add_argument("--use-beam-search", action="store_true")
parser.add_argument('--num-iters-warmup',
type=int,
default=1,
help='Number of iterations to run for warmup.')
parser.add_argument("--num-prompts",
type=int,
default=1000,
help="Number of prompts to process.")
parser.add_argument("--seed", type=int, default=0)
parser.add_argument("--hf-max-batch-size",
type=int,
default=None,
help="Maximum batch size for HF backend.")
parser.add_argument('--trust-remote-code',
action='store_true',
help='trust remote code from huggingface')
parser.add_argument(
'--max-model-len',
type=int,
default=None,
help='Maximum length of a sequence (including prompt and output). '
'If None, will be derived from the model.')
parser.add_argument(
'--dtype',
type=str,
default='auto',
choices=['auto', 'half', 'float16', 'bfloat16', 'float', 'float32'],
help='data type for model weights and activations. '
'The "auto" option will use FP16 precision '
'for FP32 and FP16 models, and BF16 precision '
'for BF16 models.')
parser.add_argument('--gpu-memory-utilization',
type=float,
default=0.9,
help='the fraction of GPU memory to be used for '
'the model executor, which can range from 0 to 1.'
'If unspecified, will use the default value of 0.9.')
parser.add_argument("--enforce-eager",
action="store_true",
help="enforce eager execution")
parser.add_argument(
'--kv-cache-dtype',
type=str,
choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
default="auto",
help='Data type for kv cache storage. If "auto", will use model '
'data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. '
'ROCm (AMD GPU) supports fp8 (=fp8_e4m3)')
parser.add_argument(
'--quantization-param-path',
type=str,
default=None,
help='Path to the JSON file containing the KV cache scaling factors. '
'This should generally be supplied, when KV cache dtype is FP8. '
'Otherwise, KV cache scaling factors default to 1.0, which may cause '
'accuracy issues. FP8_E5M2 (without scaling) is only supported on '
'cuda version greater than 11.8. On ROCm (AMD GPU), FP8_E4M3 is '
'instead supported for common inference criteria.')
parser.add_argument(
"--device",
type=str,
default="cuda",
choices=["cuda", "cpu"],
help='device type for vLLM execution, supporting CUDA and CPU.')
parser.add_argument(
"--enable-prefix-caching",
action='store_true',
help="enable automatic prefix caching for vLLM backend.")
parser.add_argument("--enable-chunked-prefill",
action='store_true',
help="enable chunked prefill for vLLM backend.")
parser.add_argument('--max-num-batched-tokens',
type=int,
default=None,
help='maximum number of batched tokens per '
'iteration')
parser.add_argument('--download-dir',
type=str,
default=None,
help='directory to download and load the weights, '
'default to the default cache dir of huggingface')
parser.add_argument(
'--output-json',
type=str,
default=None,
help='Path to save the throughput results in JSON format.')
parser.add_argument(
'--distributed-executor-backend',
choices=['ray', 'mp'],
default=None,
help='Backend to use for distributed serving. When more than 1 GPU '
'is used, will be automatically set to "ray" if installed '
'or "mp" (multiprocessing) otherwise.')
args = parser.parse_args()
if args.tokenizer is None:
args.tokenizer = args.model
if args.dataset is None:
assert args.input_len is not None
assert args.output_len is not None
else:
assert args.input_len is None
if args.backend == "vllm":
if args.hf_max_batch_size is not None:
raise ValueError("HF max batch size is only for HF backend.")
elif args.backend == "hf":
if args.hf_max_batch_size is None:
raise ValueError("HF max batch size is required for HF backend.")
if args.quantization is not None:
raise ValueError("Quantization is only for vLLM backend.")
elif args.backend == "mii":
if args.dtype != "auto":
raise ValueError("dtype must be auto for MII backend.")
if args.n != 1:
raise ValueError("n must be 1 for MII backend.")
if args.use_beam_search:
raise ValueError("Beam search is not supported for MII backend.")
if args.quantization is not None:
raise ValueError("Quantization is only for vLLM backend.")
if args.hf_max_batch_size is not None:
raise ValueError("HF max batch size is only for HF backend.")
if args.tokenizer != args.model:
raise ValueError("Tokenizer must be the same as the model for MII "
"backend.")
main(args)
benchmarks/cutlass_benchmarks/w8a8_benchmarks.py 0 → 100644
View file @ d824b9d3
import argparse
import copy
import itertools
import pickle as pkl
import time
from typing import Callable, Iterable, List, Tuple
import torch
import torch.utils.benchmark as TBenchmark
from torch.utils.benchmark import Measurement as TMeasurement
from weight_shapes import WEIGHT_SHAPES
from vllm import _custom_ops as ops
DEFAULT_MODELS = list(WEIGHT_SHAPES.keys())[1:]
DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]
DEFAULT_TP_SIZES = [1]
# helpers
def to_fp8(tensor: torch.tensor) -> torch.tensor:
finfo = torch.finfo(torch.float8_e4m3fn)
return torch.round(tensor.clamp(
min=finfo.min, max=finfo.max)).to(dtype=torch.float8_e4m3fn)
def to_int8(tensor: torch.tensor) -> torch.tensor:
return torch.round(tensor.clamp(min=-128, max=127)).to(dtype=torch.int8)
def make_rand_tensors(dtype: torch.dtype, m: int, n: int,
k: int) -> Tuple[torch.tensor, torch.tensor]:
a = torch.randn((m, k), device='cuda') * 5
b = torch.randn((n, k), device='cuda').t() * 5
if dtype == torch.int8:
return to_int8(a), to_int8(b)
if dtype == torch.float8_e4m3fn:
return to_fp8(a), to_fp8(b)
raise ValueError("unsupported dtype")
# impl
def pytorch_i8_impl(a: torch.tensor, b: torch.tensor, scale_a: torch.tensor,
scale_b: torch.tensor,
out_dtype: torch.dtype) -> torch.tensor:
return torch.mm(a, b)
def pytorch_fp8_impl(a: torch.tensor, b: torch.tensor, scale_a: torch.tensor,
scale_b: torch.tensor,
out_dtype: torch.dtype) -> torch.tensor:
return torch._scaled_mm(a,
b,
scale_a=scale_a,
scale_b=scale_b,
out_dtype=out_dtype)
def pytorch_fp8_impl_fast_accum(a: torch.tensor, b: torch.tensor,
scale_a: torch.tensor, scale_b: torch.tensor,
out_dtype: torch.dtype) -> torch.tensor:
return torch._scaled_mm(a,
b,
scale_a=scale_a,
scale_b=scale_b,
out_dtype=out_dtype,
use_fast_accum=True)
def cutlass_impl(a: torch.tensor, b: torch.tensor, scale_a: torch.tensor,
scale_b: torch.tensor,
out_dtype: torch.dtype) -> torch.tensor:
return ops.cutlass_scaled_mm_dq(a,
b,
scale_a,
scale_b,
out_dtype=out_dtype)
# bench
def bench_fn(a: torch.tensor, b: torch.tensor, scale_a: torch.tensor,
scale_b: torch.tensor, out_dtype: torch.dtype, label: str,
sub_label: str, fn: Callable, description: str) -> TMeasurement:
min_run_time = 1
globals = {
"a": a,
"b": b,
"scale_a": scale_a,
"scale_b": scale_b,
"out_dtype": out_dtype,
"fn": fn,
}
return TBenchmark.Timer(
stmt="fn(a, b, scale_a, scale_b, out_dtype)",
globals=globals,
label=label,
sub_label=sub_label,
description=description,
).blocked_autorange(min_run_time=min_run_time)
def bench_int8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
sub_label: str) -> Iterable[TMeasurement]:
assert dtype == torch.int8
a, b = make_rand_tensors(torch.int8, m, n, k)
scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
timers = []
# pytorch impl
timers.append(
bench_fn(a.to(dtype=torch.bfloat16, device="cuda"),
b.to(dtype=torch.bfloat16, device="cuda"), scale_a, scale_b,
torch.bfloat16, label, sub_label, pytorch_i8_impl,
"pytorch_bf16_bf16_bf16_matmul-no-scales"))
# cutlass impl
timers.append(
bench_fn(a, b, scale_a.to(device="cpu"), scale_b.to(device="cpu"),
torch.bfloat16, label, sub_label, cutlass_impl,
"cutlass_i8_i8_bf16_scaled_mm"))
return timers
def bench_fp8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
sub_label: str) -> Iterable[TMeasurement]:
assert dtype == torch.float8_e4m3fn
a, b = make_rand_tensors(torch.float8_e4m3fn, m, n, k)
scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
timers = []
# pytorch impl: bf16 output, without fp8 fast accum
timers.append(
bench_fn(a, b, scale_a, scale_b, torch.bfloat16, label, sub_label,
pytorch_fp8_impl, "pytorch_fp8_fp8_bf16_scaled_mm"))
# pytorch impl: bf16 output, with fp8 fast accum
timers.append(
bench_fn(a, b, scale_a, scale_b, torch.bfloat16, label, sub_label,
pytorch_fp8_impl_fast_accum,
"pytorch_fp8_fp8_bf16_scaled_mm_fast_accum"))
# pytorch impl: fp16 output, without fp8 fast accum
timers.append(
bench_fn(a, b, scale_a, scale_b, torch.float16, label, sub_label,
pytorch_fp8_impl, "pytorch_fp8_fp8_fp16_scaled_mm"))
# pytorch impl: fp16 output, with fp8 fast accum
timers.append(
bench_fn(a, b, scale_a, scale_b, torch.float16, label, sub_label,
pytorch_fp8_impl_fast_accum,
"pytorch_fp8_fp8_fp16_scaled_mm_fast_accum"))
# cutlass impl: bf16 output
timers.append(
bench_fn(a, b, scale_a.to(device="cpu"), scale_b.to(device="cpu"),
torch.bfloat16, label, sub_label, cutlass_impl,
"cutlass_fp8_fp8_bf16_scaled_mm"))
# cutlass impl: fp16 output
timers.append(
bench_fn(a, b, scale_a.to(device="cpu"), scale_b.to(device="cpu"),
torch.float16, label, sub_label, cutlass_impl,
"cutlass_fp8_fp8_fp16_scaled_mm"))
return timers
def bench(dtype: torch.dtype, m: int, k: int, n: int, label: str,
sub_label: str) -> Iterable[TMeasurement]:
if dtype == torch.int8:
return bench_int8(dtype, m, k, n, label, sub_label)
if dtype == torch.float8_e4m3fn:
return bench_fp8(dtype, m, k, n, label, sub_label)
raise ValueError("unsupported type")
# runner
def print_timers(timers: Iterable[TMeasurement]):
compare = TBenchmark.Compare(timers)
compare.print()
def run(dtype: torch.dtype,
MKNs: Iterable[Tuple[int, int, int]]) -> Iterable[TMeasurement]:
results = []
for m, k, n in MKNs:
timers = bench(dtype, m, k, n, f"scaled-{dtype}-gemm",
f"MKN=({m}x{k}x{n})")
print_timers(timers)
results.extend(timers)
return results
# output makers
def make_output(data: Iterable[TMeasurement],
MKNs: Iterable[Tuple[int, int, int]],
base_description: str,
timestamp=None):
print(f"== All Results {base_description} ====")
print_timers(data)
# pickle all the results
timestamp = int(time.time()) if timestamp is None else timestamp
with open(f"{base_description}-{timestamp}.pkl", "wb") as f:
pkl.dump(data, f)
# argparse runners
def run_square_bench(args):
dim_sizes = list(
range(args.dim_start, args.dim_end + 1, args.dim_increment))
MKNs = list(zip(dim_sizes, dim_sizes, dim_sizes))
data = run(args.dtype, MKNs)
make_output(data, MKNs, f"square_bench-{args.dtype}")
def run_range_bench(args):
dim_sizes = list(range(args.dim_start, args.dim_end, args.dim_increment))
n = len(dim_sizes)
Ms = [args.m_constant] * n if args.m_constant is not None else dim_sizes
Ks = [args.k_constant] * n if args.k_constant is not None else dim_sizes
Ns = [args.n_constant] * n if args.n_constant is not None else dim_sizes
MKNs = list(zip(Ms, Ks, Ns))
data = run(args.dtype, MKNs)
make_output(data, MKNs, f"range_bench-{args.dtype}")
def run_model_bench(args):
print("Benchmarking models:")
for i, model in enumerate(args.models):
print(f"[{i}] {model}")
def model_shapes(model_name: str, tp_size: int) -> List[Tuple[int, int]]:
KNs = []
for KN, tp_split_dim in copy.deepcopy(WEIGHT_SHAPES[model_name]):
KN[tp_split_dim] = KN[tp_split_dim] // tp_size
KNs.append(KN)
return KNs
model_bench_data = []
models_tps = list(itertools.product(args.models, args.tp_sizes))
for model, tp_size in models_tps:
Ms = args.batch_sizes
KNs = model_shapes(model, tp_size)
MKNs = []
for m in Ms:
for k, n in KNs:
MKNs.append((m, k, n))
data = run(args.dtype, MKNs)
model_bench_data.append(data)
# Print all results
for data, model_tp in zip(model_bench_data, models_tps):
model, tp_size = model_tp
print(f"== Results {args.dtype} {model}-TP{tp_size} ====")
print_timers(data)
timestamp = int(time.time())
all_data = []
for d in model_bench_data:
all_data.extend(d)
# pickle all data
with open(f"model_bench-{args.dtype}-{timestamp}.pkl", "wb") as f:
pkl.dump(all_data, f)
if __name__ == '__main__':
def to_torch_dtype(dt):
if dt == "int8":
return torch.int8
if dt == "fp8":
return torch.float8_e4m3fn
raise ValueError("unsupported dtype")
parser = argparse.ArgumentParser(
description="""
Benchmark Cutlass GEMM.
To run square GEMMs:
python3 ./benchmarks/cutlass_benchmarks/w8a8_benchmarks.py --dtype fp8 square_bench --dim-start 128 --dim-end 512 --dim-increment 64
To run constant N and K and sweep M:
python3 ./benchmarks/cutlass_benchmarks/w8a8_benchmarks.py --dtype fp8 range_bench --dim-start 128 --dim-end 512 --dim-increment 64 --n-constant 16384 --k-constant 16384
To run dimensions from a model:
python3 ./benchmarks/cutlass_benchmarks/w8a8_benchmarks.py --dtype fp8 model_bench --models meta-llama/Llama-2-7b-hf --batch-sizes 16 --tp-sizes 1
Output:
- a .pkl file, that is a list of raw torch.benchmark.utils.Measurements for the pytorch and cutlass implementations for the various GEMMs.
""", # noqa: E501
formatter_class=argparse.RawTextHelpFormatter)
parser.add_argument("--dtype",
type=to_torch_dtype,
required=True,
help="Available options are ['int8', 'fp8']")
subparsers = parser.add_subparsers(dest="cmd")
square_parser = subparsers.add_parser("square_bench")
square_parser.add_argument("--dim-start", type=int, required=True)
square_parser.add_argument("--dim-end", type=int, required=True)
square_parser.add_argument("--dim-increment", type=int, required=True)
square_parser.set_defaults(func=run_square_bench)
range_parser = subparsers.add_parser("range_bench")
range_parser.add_argument("--dim-start", type=int, required=True)
range_parser.add_argument("--dim-end", type=int, required=True)
range_parser.add_argument("--dim-increment", type=int, required=True)
range_parser.add_argument("--m-constant", type=int, default=None)
range_parser.add_argument("--n-constant", type=int, default=None)
range_parser.add_argument("--k-constant", type=int, default=None)
range_parser.set_defaults(func=run_range_bench)
model_parser = subparsers.add_parser("model_bench")
model_parser.add_argument("--models",
nargs="+",
type=str,
default=DEFAULT_MODELS,
choices=WEIGHT_SHAPES.keys())
model_parser.add_argument("--tp-sizes",
nargs="+",
type=int,
default=DEFAULT_TP_SIZES)
model_parser.add_argument("--batch-sizes",
nargs="+",
type=int,
default=DEFAULT_BATCH_SIZES)
model_parser.set_defaults(func=run_model_bench)
args = parser.parse_args()
args.func(args)
benchmarks/cutlass_benchmarks/weight_shapes.py 0 → 100644
View file @ d824b9d3
# Weight Shapes are in the format
# ([K, N], TP_SPLIT_DIM)
# Example:
# A shape of ([14336, 4096], 0) indicates the following GEMM shape,
# - TP1 : K = 14336, N = 4096
# - TP2 : K = 7168, N = 4096
# A shape of ([4096, 6144], 1) indicates the following GEMM shape,
# - TP1 : K = 4096, N = 6144
# - TP4 : K = 4096, N = 1536
# TP1 shapes
WEIGHT_SHAPES = {
"mistralai/Mistral-7B-v0.1": [
([4096, 6144], 1),
([4096, 4096], 0),
([4096, 28672], 1),
([14336, 4096], 0),
],
"meta-llama/Llama-2-7b-hf": [
([4096, 12288], 1),
([4096, 4096], 0),
([4096, 22016], 1),
([11008, 4096], 0),
],
"meta-llama/Llama-2-13b-hf": [
([5120, 15360], 1),
([5120, 5120], 0),
([5120, 27648], 1),
([13824, 5120], 0),
],
"meta-llama/Llama-2-70b-hf": [
([8192, 10240], 1),
([8192, 8192], 0),
([8192, 57344], 1),
([28672, 8192], 0),
],
}
benchmarks/kernels/benchmark_aqlm.py 0 → 100644
View file @ d824b9d3
import argparse
import os
import sys
from typing import Optional
import torch
import torch.nn.functional as F
from vllm import _custom_ops as ops
from vllm.model_executor.layers.quantization.aqlm import (
dequantize_weight, generic_dequantize_gemm, get_int_dtype,
optimized_dequantize_gemm)
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
def torch_mult(
input: torch.Tensor, # [..., in_features]
weights: torch.Tensor,
scales: torch.Tensor, # [num_out_groups, 1, 1, 1]
) -> torch.Tensor:
output = F.linear(input, weights)
return output
def dequant_out_scale(
input: torch.Tensor, # [..., in_features]
codes: torch.IntTensor, # [num_out_groups, num_in_groups, num_codebooks]
codebooks: torch.
Tensor, # [num_codebooks, codebook_size, out_group_size, in_group_size]
scales: torch.Tensor, # [num_out_groups, 1, 1, 1]
output_partition_sizes: torch.IntTensor,
bias: Optional[torch.Tensor],
) -> torch.Tensor:
weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes)
if bias is None:
output = F.linear(input, weights, bias)
orig_shape = output.shape
flattened_output = output.view(-1, output.size(-1))
f_scales = scales.view(-1, scales.shape[0])
b_scales = f_scales.expand(flattened_output.shape[0], -1)
flattened_output *= b_scales
return flattened_output.view(orig_shape)
else:
b_scales = scales.view(scales.shape[:-3] + (-1, )).expand(
-1, weights.shape[1])
weights *= b_scales
return F.linear(input, weights, bias)
def dequant_weight_scale(
input: torch.Tensor, # [..., in_features]
codes: torch.IntTensor, # [num_out_groups, num_in_groups, num_codebooks]
codebooks: torch.
Tensor, # [num_codebooks, codebook_size, out_group_size, in_group_size]
scales: torch.Tensor, # [num_out_groups, 1, 1, 1]
output_partition_sizes: torch.IntTensor,
bias: Optional[torch.Tensor],
) -> torch.Tensor:
weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes)
b_scales = scales.view(scales.shape[:-3] + (-1, )).expand(
-1, weights.shape[1])
weights *= b_scales
return F.linear(input, weights, bias)
def dequant_no_scale(
input: torch.Tensor, # [..., in_features]
codes: torch.IntTensor, # [num_out_groups, num_in_groups, num_codebooks]
codebooks: torch.
Tensor, # [num_codebooks, codebook_size, out_group_size, in_group_size]
scales: torch.Tensor, # [num_out_groups, 1, 1, 1]
output_partition_sizes: torch.IntTensor,
bias: Optional[torch.Tensor],
) -> torch.Tensor:
weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes)
return F.linear(input, weights, bias)
# Compare the optimized 1x16 and 2x8 cuda decompression/dequant kernels against
# the generic pytorch version.
# Just visual comparison.
def dequant_test(k: int, parts: torch.tensor, nbooks: int, bits: int) -> None:
n = parts.sum().item()
device = torch.device('cuda:0')
code_range = (1 << bits) // 2
ingroups = 8
codes = torch.randint(-code_range,
code_range,
size=(n, k // ingroups, nbooks),
dtype=get_int_dtype(bits),
device=device)
codebooks = torch.randn(size=(parts.shape[0] * nbooks, 1 << bits, 1, 8),
dtype=torch.float16,
device=device)
count = 0
for index in range(16):
for i in range(8):
for book in range(nbooks):
codebooks[book, index, 0, i] = count * (10**book)
count += 1
print("codes shape", codes.shape)
for i in range(16):
for book in range(nbooks):
codes[0, i, book] = i
codes[0, -i, book] = i
weights = dequantize_weight(codes, codebooks, None)
weights2 = ops.aqlm_dequant(codes, codebooks, parts)
print("weights shape:", weights.shape)
print("weights2 shape:", weights2.shape)
print("weights are:", weights)
print("weights2 are:", weights2)
print("first 128 weights are", weights[0, 0:128].to(torch.int32))
print("first 128 weights2 are:", weights2[0, 0:128].to(torch.int32))
print("last 128 weights are", weights[0, -128:])
print("last 128 weights2 are:", weights2[0, -128:])
def main():
parser = argparse.ArgumentParser(description="Benchmark aqlm performance.")
# Add arguments
parser.add_argument("--nbooks",
type=int,
default=1,
help="Number of codebooks (default: 1)")
parser.add_argument("--bits",
type=int,
default=16,
help="Number of bits per code element (default: 16)")
parser.add_argument(
"--test",
type=bool,
default=False,
help="Run the decompression/dequant tester rather than benchmarking "
"(default: False)")
# Parse the arguments
args = parser.parse_args()
# Extract values
nbooks = args.nbooks
bits = args.bits
if args.test:
dequant_test(4096, torch.tensor((4096, )), nbooks, bits)
return
# Otherwise, benchmark.
methods = [
ops.aqlm_gemm,
dequant_out_scale,
generic_dequantize_gemm,
optimized_dequantize_gemm,
dequant_weight_scale,
torch_mult,
dequant_no_scale,
]
filename = f"./aqlm_benchmark_{nbooks}x{bits}.csv"
print(f"writing benchmarks to file {filename}")
with open(filename, "w") as f:
sys.stdout = f
print('m | k | n | n parts', end='')
for method in methods:
print(f" | {method.__name__.replace('_', ' ')} (µs)", end='')
print('')
# These are reasonable prefill sizes.
ksandpartions = ((4096, (4096, 4096, 4096)), (4096, (4096, )),
(4096, (11008, 11008)), (11008, (4096, )))
# reasonable ranges for m.
for m in [
1, 2, 4, 8, 10, 12, 14, 16, 24, 32, 48, 52, 56, 64, 96, 112,
128, 256, 512, 1024, 1536, 2048, 3072, 4096
]:
print(f'{m}', file=sys.__stdout__)
for ksp in ksandpartions:
run_grid(m, ksp[0], torch.tensor(ksp[1]), nbooks, bits,
methods)
sys.stdout = sys.__stdout__
def run_grid(m: int, k: int, parts: torch.tensor, nbooks: int, bits: int,
methods):
# I didn't see visible improvements from increasing these, but feel free :)
num_warmup_trials = 1
num_trials = 1
num_calls = 100
# warmup.
for method in methods:
for _ in range(num_warmup_trials):
run_timing(
num_calls=num_calls,
m=m,
k=k,
parts=parts,
nbooks=nbooks,
bits=bits,
method=method,
)
n = parts.sum().item()
print(f'{m} | {k} | {n} | {parts.tolist()}', end='')
for method in methods:
best_time_us = 1e20
for _ in range(num_trials):
kernel_dur_ms = run_timing(
num_calls=num_calls,
m=m,
k=k,
parts=parts,
nbooks=nbooks,
bits=bits,
method=method,
)
kernel_dur_us = 1000 * kernel_dur_ms
if kernel_dur_us < best_time_us:
best_time_us = kernel_dur_us
print(f' | {kernel_dur_us:.0f}', end='')
print('')
def run_timing(num_calls: int, m: int, k: int, parts: torch.tensor,
nbooks: int, bits: int, method) -> float:
n = parts.sum().item()
device = torch.device('cuda:0')
input = torch.randn((1, m, k), dtype=torch.float16, device=device)
code_range = (1 << bits) // 2
ingroups = 8
codes = torch.randint(-code_range,
code_range,
size=(n, k // ingroups, nbooks),
dtype=get_int_dtype(bits),
device=device)
codebooks = torch.randn(size=(parts.shape[0] * nbooks, 1 << bits, 1, 8),
dtype=torch.float16,
device=device)
scales = torch.randn(size=(n, 1, 1, 1), dtype=torch.float16, device=device)
# for comparison to just a pytorch mult.
weights = torch.randn((n, k), dtype=torch.float16, device=device)
start_event = torch.cuda.Event(enable_timing=True)
end_event = torch.cuda.Event(enable_timing=True)
start_event.record()
if method is torch_mult:
for i in range(num_calls):
torch_mult(input, weights, scales)
else:
for i in range(num_calls):
method(input, codes, codebooks, scales, parts, None)
end_event.record()
end_event.synchronize()
dur_ms = start_event.elapsed_time(end_event) / num_calls
return dur_ms
if __name__ == "__main__":
sys.exit(main())
benchmarks/kernels/benchmark_marlin.py 0 → 100644
View file @ d824b9d3
import argparse
import torch
import torch.utils.benchmark as benchmark
from benchmark_shapes import WEIGHT_SHAPES
from vllm import _custom_ops as ops
from vllm.model_executor.layers.quantization.gptq_marlin import (
GPTQ_MARLIN_MAX_PARALLEL, GPTQ_MARLIN_MIN_THREAD_N,
GPTQ_MARLIN_SUPPORTED_GROUP_SIZES, GPTQ_MARLIN_SUPPORTED_NUM_BITS)
from vllm.model_executor.layers.quantization.gptq_marlin_24 import (
GPTQ_MARLIN_24_MAX_PARALLEL, GPTQ_MARLIN_24_MIN_THREAD_N,
GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES, GPTQ_MARLIN_24_SUPPORTED_NUM_BITS)
from vllm.model_executor.layers.quantization.utils.marlin_utils import (
MarlinWorkspace, marlin_24_quantize, marlin_quantize)
from vllm.model_executor.layers.quantization.utils.quant_utils import (
gptq_pack, quantize_weights, sort_weights)
DEFAULT_MODELS = ["meta-llama/Llama-2-7b-hf/TP1"]
DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]
ACT_ORDER_OPTS = [False, True]
K_FULL_OPTS = [False, True]
def bench_run(results, model, act_order, is_k_full, num_bits, group_size,
size_m, size_k, size_n):
label = "Quant Matmul"
sub_label = ("{}, act={} k_full={}, b={}, g={}, "
"MKN=({}x{}x{})".format(model, act_order, is_k_full, num_bits,
group_size, size_m, size_k, size_n))
print(f"Testing: {sub_label}")
a = torch.randn(size_m, size_k).to(torch.half).cuda()
b = torch.rand(size_k, size_n).to(torch.half).cuda()
a_tmp = (torch.zeros(size_m, size_k).to(torch.half).cuda())
# Marlin quant
(
marlin_w_ref,
marlin_q_w,
marlin_s,
marlin_g_idx,
marlin_sort_indices,
marlin_rand_perm,
) = marlin_quantize(b, num_bits, group_size, act_order)
# Marlin_24 quant
(marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta,
marlin_24_s) = marlin_24_quantize(b, num_bits, group_size)
# GPTQ quant
(w_ref, q_w, s, g_idx,
rand_perm) = quantize_weights(b, num_bits, group_size, act_order)
q_w_gptq = gptq_pack(q_w, num_bits, size_k, size_n)
# For act_order, sort the "weights" and "g_idx"
# so that group ids are increasing
repack_sort_indices = torch.empty(0, dtype=torch.int, device=b.device)
if act_order:
(q_w, g_idx, repack_sort_indices) = sort_weights(q_w, g_idx)
# Prepare
marlin_workspace = MarlinWorkspace(size_n, GPTQ_MARLIN_MIN_THREAD_N,
GPTQ_MARLIN_MAX_PARALLEL)
marlin_24_workspace = MarlinWorkspace(size_n, GPTQ_MARLIN_24_MIN_THREAD_N,
GPTQ_MARLIN_24_MAX_PARALLEL)
globals = {
# Gen params
"num_bits": num_bits,
"group_size": group_size,
"size_m": size_m,
"size_n": size_n,
"size_k": size_k,
"a": a,
"a_tmp": a_tmp,
# Marlin params
"marlin_w_ref": marlin_w_ref,
"marlin_q_w": marlin_q_w,
"marlin_s": marlin_s,
"marlin_g_idx": marlin_g_idx,
"marlin_sort_indices": marlin_sort_indices,
"marlin_rand_perm": marlin_rand_perm,
"marlin_workspace": marlin_workspace,
"is_k_full": is_k_full,
# Marlin_24 params
"marlin_24_w_ref": marlin_24_w_ref,
"marlin_24_q_w_comp": marlin_24_q_w_comp,
"marlin_24_meta": marlin_24_meta,
"marlin_24_s": marlin_24_s,
"marlin_24_workspace": marlin_24_workspace,
# GPTQ params
"q_w_gptq": q_w_gptq,
"repack_sort_indices": repack_sort_indices,
# Kernels
"gptq_marlin_gemm": ops.gptq_marlin_gemm,
"gptq_marlin_24_gemm": ops.gptq_marlin_24_gemm,
"gptq_marlin_repack": ops.gptq_marlin_repack,
}
min_run_time = 1
# Warmup pytorch
for i in range(5):
torch.matmul(a, marlin_w_ref)
results.append(
benchmark.Timer(
stmt="torch.matmul(a, marlin_w_ref)",
globals=globals,
label=label,
sub_label=sub_label,
description="pytorch_gemm",
).blocked_autorange(min_run_time=min_run_time))
results.append(
benchmark.Timer(
stmt=
"output = gptq_marlin_gemm(a, marlin_q_w, marlin_s, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, num_bits, size_m, size_n, size_k, is_k_full)", # noqa: E501
globals=globals,
label=label,
sub_label=sub_label,
description="gptq_marlin_gemm",
).blocked_autorange(min_run_time=min_run_time))
if (num_bits in GPTQ_MARLIN_24_SUPPORTED_NUM_BITS
and group_size in GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES):
results.append(
benchmark.Timer(
stmt=
"output = gptq_marlin_24_gemm(a, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s, marlin_24_workspace.scratch, num_bits, size_m, size_n, size_k)", # noqa: E501
globals=globals,
label=label,
sub_label=sub_label,
description="gptq_marlin_24_gemm",
).blocked_autorange(min_run_time=min_run_time))
results.append(
benchmark.Timer(
stmt=
"q_res = gptq_marlin_repack(q_w_gptq, repack_sort_indices, size_k, size_n, num_bits)", # noqa: E501
globals=globals,
label=label,
sub_label=sub_label,
description="gptq_marlin_repack",
).blocked_autorange(min_run_time=min_run_time))
def main(args):
print("Benchmarking models:")
for i, model in enumerate(args.models):
print(f"[{i}] {model}")
results = []
for model in args.models:
for layer in WEIGHT_SHAPES[model]:
size_k = layer[0]
size_n = layer[1]
if len(args.limit_k) > 0 and size_k not in args.limit_k:
continue
if len(args.limit_n) > 0 and size_n not in args.limit_n:
continue
for act_order in ACT_ORDER_OPTS:
if len(args.limit_act_order
) > 0 and act_order not in args.limit_act_order:
continue
for is_k_full in K_FULL_OPTS:
if len(args.limit_k_full
) > 0 and is_k_full not in args.limit_k_full:
continue
for num_bits in GPTQ_MARLIN_SUPPORTED_NUM_BITS:
if len(args.limit_num_bits
) > 0 and num_bits not in args.limit_num_bits:
continue
for group_size in GPTQ_MARLIN_SUPPORTED_GROUP_SIZES:
if len(
args.limit_group_size
) > 0 and group_size not in args.limit_group_size:
continue
# For act_order, the group_size must be less than
# size_k
if act_order and (group_size == size_k
or group_size == -1):
continue
for size_m in args.batch_sizes:
bench_run(results, model, act_order, is_k_full,
num_bits, group_size, size_m, size_k,
size_n)
compare = benchmark.Compare(results)
compare.print()
# For quick benchmarking use:
# python benchmark_marlin.py --batch-sizes 1 16 32 --limit-k 4096 --limit-n 4096 --limit-group-size 128 --limit-num-bits 4 --limit-act-order 0 --limit-k-full 1 # noqa E501
#
if __name__ == "__main__":
parser = argparse.ArgumentParser(
description="Benchmark Marlin across specified models/shapes/batches")
parser.add_argument(
"--models",
nargs="+",
type=str,
default=DEFAULT_MODELS,
choices=WEIGHT_SHAPES.keys(),
)
parser.add_argument("--batch-sizes",
nargs="+",
type=int,
default=DEFAULT_BATCH_SIZES)
parser.add_argument("--limit-k", nargs="+", type=int, default=[])
parser.add_argument("--limit-n", nargs="+", type=int, default=[])
parser.add_argument("--limit-group-size", nargs="+", type=int, default=[])
parser.add_argument("--limit-num-bits", nargs="+", type=int, default=[])
parser.add_argument("--limit-act-order", nargs="+", type=int, default=[])
parser.add_argument("--limit-k-full", nargs="+", type=int, default=[])
args = parser.parse_args()
main(args)
benchmarks/kernels/benchmark_moe.py 0 → 100644
View file @ d824b9d3
import argparse
import time
from datetime import datetime
from typing import Any, Dict, List, Tuple
import ray
import torch
import triton
from ray.experimental.tqdm_ray import tqdm
from transformers import AutoConfig
from vllm.model_executor.layers.fused_moe.fused_moe import *
def benchmark_config(
config: Dict[str, int],
num_tokens: int,
num_experts: int,
shard_intermediate_size: int,
hidden_size: int,
topk: int,
dtype: torch.dtype,
use_fp8: bool,
num_iters: int = 100,
) -> float:
init_dtype = torch.float16 if use_fp8 else dtype
x = torch.randn(num_tokens, hidden_size, dtype=dtype)
w1 = torch.randn(num_experts,
shard_intermediate_size,
hidden_size,
dtype=init_dtype)
w2 = torch.randn(num_experts,
hidden_size,
shard_intermediate_size // 2,
dtype=init_dtype)
gating_output = torch.randn(num_iters,
num_tokens,
num_experts,
dtype=torch.float32)
w1_scale = None
w2_scale = None
a1_scale = None
a2_scale = None
if use_fp8:
w1_scale = torch.randn(num_experts, dtype=torch.float32)
w2_scale = torch.randn(num_experts, dtype=torch.float32)
a1_scale = torch.randn(1, dtype=torch.float32)
a2_scale = torch.randn(1, dtype=torch.float32)
w1 = w1.to(torch.float8_e4m3fn)
w2 = w2.to(torch.float8_e4m3fn)
input_gating = torch.empty(num_tokens, num_experts, dtype=torch.float32)
def prepare(i: int):
input_gating.copy_(gating_output[i])
def run():
fused_moe(
x,
w1,
w2,
input_gating,
topk,
renormalize=True,
inplace=True,
override_config=config,
use_fp8=use_fp8,
w1_scale=w1_scale,
w2_scale=w2_scale,
a1_scale=a1_scale,
a2_scale=a2_scale,
)
# JIT compilation & warmup
run()
torch.cuda.synchronize()
# Capture 10 invocations with CUDA graph
graph = torch.cuda.CUDAGraph()
with torch.cuda.graph(graph):
for _ in range(10):
run()
torch.cuda.synchronize()
# Warmup
for _ in range(5):
graph.replay()
torch.cuda.synchronize()
start_event = torch.cuda.Event(enable_timing=True)
end_event = torch.cuda.Event(enable_timing=True)
latencies = []
for i in range(num_iters):
prepare(i)
torch.cuda.synchronize()
start_event.record()
graph.replay()
end_event.record()
end_event.synchronize()
latencies.append(start_event.elapsed_time(end_event))
avg = sum(latencies) / (num_iters * 10) * 1000 # us
graph.reset()
return avg
def get_configs_compute_bound() -> List[Dict[str, int]]:
# Reduced search space for faster tuning.
# TODO(woosuk): Increase the search space and use a performance model to
# prune the search space.
configs = []
for num_stages in [2, 3, 4, 5]:
for block_m in [16, 32, 64, 128, 256]:
for block_k in [64, 128, 256]:
for block_n in [32, 64, 128, 256]:
for num_warps in [4, 8]:
for group_size in [1, 16, 32, 64]:
configs.append({
"BLOCK_SIZE_M": block_m,
"BLOCK_SIZE_N": block_n,
"BLOCK_SIZE_K": block_k,
"GROUP_SIZE_M": group_size,
"num_warps": num_warps,
"num_stages": num_stages,
})
return configs
@ray.remote(num_gpus=1)
class BenchmarkWorker:
def __init__(self, seed: int) -> None:
torch.set_default_device("cuda")
torch.cuda.manual_seed_all(seed)
self.seed = seed
def benchmark(
self,
num_tokens: int,
num_experts: int,
shard_intermediate_size: int,
hidden_size: int,
topk: int,
dtype: torch.dtype,
use_fp8: bool,
) -> Tuple[Dict[str, int], float]:
torch.cuda.manual_seed_all(self.seed)
dtype_str = "float8" if use_fp8 else None
# NOTE(woosuk): The current naming convention uses w2.shape[2], which
# is the intermediate size after silu_and_mul.
op_config = get_moe_configs(num_experts, shard_intermediate_size // 2,
dtype_str)
if op_config is None:
config = get_default_config(num_tokens, num_experts,
shard_intermediate_size, hidden_size,
topk, dtype_str)
else:
config = op_config[min(op_config.keys(),
key=lambda x: abs(x - num_tokens))]
kernel_time = benchmark_config(config, num_tokens, num_experts,
shard_intermediate_size, hidden_size,
topk, dtype, use_fp8)
return config, kernel_time
def tune(
self,
num_tokens: int,
num_experts: int,
shard_intermediate_size: int,
hidden_size: int,
topk: int,
dtype: torch.dtype,
use_fp8: bool,
search_space: List[Dict[str, int]],
) -> Dict[str, int]:
best_config = None
best_time = float("inf")
for config in tqdm(search_space):
try:
kernel_time = benchmark_config(config,
num_tokens,
num_experts,
shard_intermediate_size,
hidden_size,
topk,
dtype,
use_fp8,
num_iters=10)
except triton.runtime.autotuner.OutOfResources:
# Some configurations may be invalid and fail to compile.
continue
if kernel_time < best_time:
best_time = kernel_time
best_config = config
now = datetime.now()
print(f"{now.ctime()}] Completed tuning for batch_size={num_tokens}")
return best_config
def sort_config(config: Dict[str, int]) -> Dict[str, int]:
return {
"BLOCK_SIZE_M": config["BLOCK_SIZE_M"],
"BLOCK_SIZE_N": config["BLOCK_SIZE_N"],
"BLOCK_SIZE_K": config["BLOCK_SIZE_K"],
"GROUP_SIZE_M": config["GROUP_SIZE_M"],
"num_warps": config["num_warps"],
"num_stages": config["num_stages"],
}
def save_configs(
configs: Dict[int, Dict[str, int]],
num_experts: int,
shard_intermediate_size: int,
hidden_size: int,
topk: int,
dtype: torch.dtype,
use_fp8: bool,
) -> None:
dtype_str = "float8" if use_fp8 else None
# NOTE(woosuk): The current naming convention uses w2.shape[2], which
# is the intermediate size after silu_and_mul.
filename = get_config_file_name(num_experts, shard_intermediate_size // 2,
dtype_str)
print(f"Writing best config to {filename}...")
with open(filename, "w") as f:
json.dump(configs, f, indent=4)
f.write("\n")
def main(args: argparse.Namespace):
print(args)
config = AutoConfig.from_pretrained(args.model)
if config.architectures[0] == "DbrxForCausalLM":
E = config.ffn_config.moe_num_experts
topk = config.ffn_config.moe_top_k
intermediate_size = config.ffn_config.ffn_hidden_size
shard_intermediate_size = 2 * intermediate_size // args.tp_size
else:
# Default: Mixtral.
E = config.num_local_experts
topk = config.num_experts_per_tok
intermediate_size = config.intermediate_size
shard_intermediate_size = 2 * intermediate_size // args.tp_size
hidden_size = config.hidden_size
dtype = config.torch_dtype
use_fp8 = args.dtype == "fp8"
if args.batch_size is None:
batch_sizes = [
1, 2, 4, 8, 16, 24, 32, 48, 64, 96, 128, 256, 512, 1024, 1536,
2048, 3072, 4096
]
else:
batch_sizes = [args.batch_size]
ray.init()
num_gpus = int(ray.available_resources()["GPU"])
workers = [BenchmarkWorker.remote(args.seed) for _ in range(num_gpus)]
def _distribute(method: str, inputs: List[Any]) -> List[Any]:
outputs = []
worker_idx = 0
for input_args in inputs:
worker = workers[worker_idx]
worker_method = getattr(worker, method)
output = worker_method.remote(*input_args)
outputs.append(output)
worker_idx = (worker_idx + 1) % num_gpus
return ray.get(outputs)
if args.tune:
search_space = get_configs_compute_bound()
print(f"Start tuning over {len(search_space)} configurations...")
start = time.time()
configs = _distribute(
"tune", [(batch_size, E, shard_intermediate_size, hidden_size,
topk, dtype, use_fp8, search_space)
for batch_size in batch_sizes])
best_configs = {
M: sort_config(config)
for M, config in zip(batch_sizes, configs)
}
save_configs(best_configs, E, shard_intermediate_size, hidden_size,
topk, dtype, use_fp8)
end = time.time()
print(f"Tuning took {end - start:.2f} seconds")
else:
outputs = _distribute("benchmark",
[(batch_size, E, shard_intermediate_size,
hidden_size, topk, dtype, use_fp8)
for batch_size in batch_sizes])
for batch_size, (config, kernel_time) in zip(batch_sizes, outputs):
print(f"Batch size: {batch_size}, config: {config}")
print(f"Kernel time: {kernel_time:.2f} us")
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--model",
type=str,
default="mistralai/Mixtral-8x7B-Instruct-v0.1")
parser.add_argument("--tp-size", "-tp", type=int, default=2)
parser.add_argument("--dtype",
type=str,
choices=["auto", "fp8"],
default="auto")
parser.add_argument("--seed", type=int, default=0)
parser.add_argument("--batch-size", type=int, required=False)
parser.add_argument("--tune", action="store_true")
args = parser.parse_args()
main(args)
benchmarks/kernels/benchmark_paged_attention.py 0 → 100644
View file @ d824b9d3
import argparse
import random
import time
from typing import Optional
import torch
from vllm import _custom_ops as ops
from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, create_kv_caches_with_random
NUM_BLOCKS = 1024
PARTITION_SIZE = 512
@torch.inference_mode()
def main(
version: str,
num_seqs: int,
seq_len: int,
num_query_heads: int,
num_kv_heads: int,
head_size: int,
use_alibi: bool,
block_size: int,
dtype: torch.dtype,
seed: int,
do_profile: bool,
device: str = "cuda",
kv_cache_dtype: Optional[str] = None,
) -> None:
random.seed(seed)
torch.random.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
scale = float(1.0 / (head_size**0.5))
query = torch.empty(num_seqs,
num_query_heads,
head_size,
dtype=dtype,
device=device)
query.uniform_(-scale, scale)
assert num_query_heads % num_kv_heads == 0
alibi_slopes = None
if use_alibi:
alibi_slopes = torch.randn(num_query_heads,
dtype=torch.float,
device=device)
seq_lens = [seq_len for _ in range(num_seqs)]
max_seq_len = max(seq_lens)
seq_lens = torch.tensor(seq_lens, dtype=torch.int, device=device)
# Create the block tables.
max_num_blocks_per_seq = (max_seq_len + block_size - 1) // block_size
block_tables = []
for _ in range(num_seqs):
block_table = [
random.randint(0, NUM_BLOCKS - 1)
for _ in range(max_num_blocks_per_seq)
]
block_tables.append(block_table)
block_tables = torch.tensor(block_tables, dtype=torch.int, device=device)
# Create the KV cache.
key_caches, value_caches = create_kv_caches_with_random(NUM_BLOCKS,
block_size,
1,
num_kv_heads,
head_size,
kv_cache_dtype,
dtype,
device=device)
key_cache, value_cache = key_caches[0], value_caches[0]
# Prepare for the paged attention kernel.
output = torch.empty_like(query)
if version == "v2":
num_partitions = ((max_seq_len + PARTITION_SIZE - 1) // PARTITION_SIZE)
tmp_output = torch.empty(
size=(num_seqs, num_query_heads, num_partitions, head_size),
dtype=output.dtype,
device=output.device,
)
exp_sums = torch.empty(
size=(num_seqs, num_query_heads, num_partitions),
dtype=torch.float32,
device=output.device,
)
max_logits = torch.empty_like(exp_sums)
def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float:
torch.cuda.synchronize()
if profile:
torch.cuda.cudart().cudaProfilerStart()
start_time = time.perf_counter()
# Using default kv_scale
kv_scale = 1.0
for _ in range(num_iters):
if version == "v1":
ops.paged_attention_v1(
output,
query,
key_cache,
value_cache,
num_kv_heads,
scale,
block_tables,
seq_lens,
block_size,
max_seq_len,
alibi_slopes,
kv_cache_dtype,
kv_scale,
)
elif version == "v2":
ops.paged_attention_v2(
output,
exp_sums,
max_logits,
tmp_output,
query,
key_cache,
value_cache,
num_kv_heads,
scale,
block_tables,
seq_lens,
block_size,
max_seq_len,
alibi_slopes,
kv_cache_dtype,
kv_scale,
)
else:
raise ValueError(f"Invalid version: {version}")
torch.cuda.synchronize()
end_time = time.perf_counter()
if profile:
torch.cuda.cudart().cudaProfilerStart()
return (end_time - start_time) / num_iters
# Warmup.
print("Warming up...")
run_benchmark = run_cuda_benchmark
run_benchmark(num_iters=3, profile=False)
# Benchmark.
if do_profile:
latency = run_benchmark(num_iters=1, profile=True)
else:
latency = run_benchmark(num_iters=100, profile=False)
print(f"Kernel running time: {latency * 1000000:.3f} us")
if __name__ == '__main__':
parser = argparse.ArgumentParser(
description="Benchmark the paged attention kernel.")
parser.add_argument("--version",
type=str,
choices=["v1", "v2"],
default="v2")
parser.add_argument("--batch-size", type=int, default=8)
parser.add_argument("--seq_len", type=int, default=4096)
parser.add_argument("--num-query-heads", type=int, default=64)
parser.add_argument("--num-kv-heads", type=int, default=8)
parser.add_argument("--head-size",
type=int,
choices=[64, 80, 96, 112, 128, 192, 256],
default=128)
parser.add_argument("--block-size", type=int, choices=[16, 32], default=16)
parser.add_argument("--use-alibi", action="store_true")
parser.add_argument("--dtype",
type=str,
choices=["half", "bfloat16", "float"],
default="half")
parser.add_argument("--seed", type=int, default=0)
parser.add_argument("--profile", action="store_true")
parser.add_argument(
"--kv-cache-dtype",
type=str,
choices=["auto", "fp8", "fp8_e5m2", "fp8_e4m3"],
default="auto",
help="Data type for kv cache storage. If 'auto', will use model "
"data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. "
"ROCm (AMD GPU) supports fp8 (=fp8_e4m3)")
args = parser.parse_args()
print(args)
if args.num_query_heads % args.num_kv_heads != 0:
raise ValueError("num_query_heads must be divisible by num_kv_heads")
main(
version=args.version,
num_seqs=args.batch_size,
seq_len=args.seq_len,
num_query_heads=args.num_query_heads,
num_kv_heads=args.num_kv_heads,
head_size=args.head_size,
block_size=args.block_size,
use_alibi=args.use_alibi,
dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype],
seed=args.seed,
do_profile=args.profile,
kv_cache_dtype=args.kv_cache_dtype,
)
benchmarks/kernels/benchmark_rope.py 0 → 100644
View file @ d824b9d3
import argparse
from itertools import accumulate
from typing import Optional
import nvtx
import torch
from vllm.model_executor.layers.rotary_embedding import get_rope
def benchmark_rope_kernels_multi_lora(
is_neox_style: bool,
batch_size: int,
seq_len: int,
num_heads: int,
head_size: int,
rotary_dim: Optional[int],
dtype: torch.dtype,
seed: int,
device: str,
max_position: int = 8192,
base: int = 10000,
) -> None:
torch.random.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
torch.set_default_device(device)
if rotary_dim is None:
rotary_dim = head_size
# silulating serving 4 LoRAs
scaling_factors = [1, 2, 4, 8]
# batched RoPE can take multiple scaling factors
batched_rope = get_rope(head_size, rotary_dim, max_position, base,
is_neox_style, {
"type": "linear",
"factor": tuple(scaling_factors)
})
# non-batched RoPE takes only one scaling factor, we create multiple
# instances to simulate the same behavior
non_batched_ropes = []
for scaling_factor in scaling_factors:
non_batched_ropes.append(
get_rope(head_size, rotary_dim, max_position, base, is_neox_style,
{
"type": "linear",
"factor": (scaling_factor, )
}))
positions = torch.randint(0, max_position, (batch_size, seq_len))
query = torch.randn(batch_size,
seq_len,
num_heads * head_size,
dtype=dtype)
key = torch.randn_like(query)
# create query offsets for batched RoPE, we concat multiple kv cache
# together and each query needs to find the right kv cache of its type
offset_map = torch.tensor(
list(
accumulate([0] + [
max_position * scaling_factor * 2
for scaling_factor in scaling_factors[:-1]
])))
query_types = torch.randint(0,
len(scaling_factors), (batch_size, seq_len),
device=device)
# map query types to offsets
query_offsets = offset_map[query_types]
# the kernel takes flattened offsets
flatten_offsets = query_offsets.flatten()
# batched queries of the same type together for non-batched RoPE
queries = [query[query_types == i] for i in range(len(scaling_factors))]
keys = [key[query_types == i] for i in range(len(scaling_factors))]
packed_qkr = zip(queries, keys, non_batched_ropes)
# synchronize before start timing
torch.cuda.synchronize()
with nvtx.annotate("non-batched", color="yellow"):
for q, k, r in packed_qkr:
r.forward(positions, q, k)
torch.cuda.synchronize()
with nvtx.annotate("batched", color="green"):
batched_rope.forward(positions, query, key, flatten_offsets)
torch.cuda.synchronize()
if __name__ == '__main__':
parser = argparse.ArgumentParser(
description="Benchmark the rotary embedding kernels.")
parser.add_argument("--is-neox-style", type=bool, default=True)
parser.add_argument("--batch-size", type=int, default=16)
parser.add_argument("--seq-len", type=int, default=512)
parser.add_argument("--num-heads", type=int, default=8)
parser.add_argument("--head-size",
type=int,
choices=[64, 80, 96, 112, 128, 192, 256],
default=128)
parser.add_argument("--rotary-dim", type=int, choices=[16, 32], default=32)
parser.add_argument("--dtype",
type=str,
choices=["bfloat16", "float"],
default="float")
parser.add_argument("--seed", type=int, default=0)
parser.add_argument("--device",
type=str,
choices=["cuda:0", "cuda:1"],
default="cuda:0")
args = parser.parse_args()
print(args)
benchmark_rope_kernels_multi_lora(
is_neox_style=args.is_neox_style,
batch_size=args.batch_size,
seq_len=args.seq_len,
num_heads=args.num_heads,
head_size=args.head_size,
rotary_dim=args.rotary_dim,
dtype=getattr(torch, args.dtype),
seed=args.seed,
device=args.device,
)
benchmarks/kernels/benchmark_shapes.py 0 → 100644
View file @ d824b9d3
WEIGHT_SHAPES = {
"ideal": [[4 * 256 * 32, 256 * 32]],
"mistralai/Mistral-7B-v0.1/TP1": [
[4096, 6144],
[4096, 4096],
[4096, 28672],
[14336, 4096],
],
"mistralai/Mistral-7B-v0.1/TP2": [
[4096, 3072],
[2048, 4096],
[4096, 14336],
[7168, 4096],
],
"mistralai/Mistral-7B-v0.1/TP4": [
[4096, 1536],
[1024, 4096],
[4096, 7168],
[3584, 4096],
],
"meta-llama/Llama-2-7b-hf/TP1": [
[4096, 12288],
[4096, 4096],
[4096, 22016],
[11008, 4096],
],
"meta-llama/Llama-2-7b-hf/TP2": [
[4096, 6144],
[2048, 4096],
[4096, 11008],
[5504, 4096],
],
"meta-llama/Llama-2-7b-hf/TP4": [
[4096, 3072],
[1024, 4096],
[4096, 5504],
[2752, 4096],
],
"meta-llama/Llama-2-13b-hf/TP1": [
[5120, 15360],
[5120, 5120],
[5120, 27648],
[13824, 5120],
],
"meta-llama/Llama-2-13b-hf/TP2": [
[5120, 7680],
[2560, 5120],
[5120, 13824],
[6912, 5120],
],
"meta-llama/Llama-2-13b-hf/TP4": [
[5120, 3840],
[1280, 5120],
[5120, 6912],
[3456, 5120],
],
"meta-llama/Llama-2-70b-hf/TP1": [
[8192, 10240],
[8192, 8192],
[8192, 57344],
[28672, 8192],
],
"meta-llama/Llama-2-70b-hf/TP2": [
[8192, 5120],
[4096, 8192],
[8192, 28672],
[14336, 8192],
],
"meta-llama/Llama-2-70b-hf/TP4": [
[8192, 2560],
[2048, 8192],
[8192, 14336],
[7168, 8192],
],
}
benchmarks/launch_tgi_server.sh 0 → 100644
View file @ d824b9d3
#!/bin/bash
PORT=8000
MODEL=$1
TOKENS=$2
docker run -e HF_TOKEN=$HF_TOKEN --gpus all --shm-size 1g -p $PORT:80 \
-v $PWD/data:/data \
ghcr.io/huggingface/text-generation-inference:1.4.0 \
--model-id $MODEL \
--sharded false \
--max-input-length 1024 \
--max-total-tokens 2048 \
--max-best-of 5 \
--max-concurrent-requests 5000 \
--max-batch-total-tokens $TOKENS
benchmarks/overheads/benchmark_hashing.py 0 → 100644
View file @ d824b9d3
import argparse
import cProfile
import pstats
from vllm import LLM, SamplingParams
# A very long prompt, total number of tokens is about 15k.
LONG_PROMPT = ["You are an expert in large language models, aren't you?"
] * 1000
LONG_PROMPT = ' '.join(LONG_PROMPT)
def main(args):
llm = LLM(
model=args.model,
enforce_eager=True,
enable_prefix_caching=True,
tensor_parallel_size=args.tensor_parallel_size,
use_v2_block_manager=args.use_v2_block_manager,
)
sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len)
profiler = cProfile.Profile()
print("------warm up------")
for i in range(3):
output = llm.generate(LONG_PROMPT, sampling_params)
print(output[0].outputs[0].text)
print("------start generating------")
for i in range(3):
profiler.runctx('llm.generate(LONG_PROMPT, sampling_params)',
globals(), locals())
# analyze the runtime of hashing function
stats = pstats.Stats(profiler)
stats.sort_stats('cumulative')
total_time = 0
total_calls = 0
for func in stats.stats:
if 'hash_of_block' in func[2]:
total_time = stats.stats[func][3]
total_calls = stats.stats[func][0]
percentage = (total_time / stats.total_tt) * 100
print(f"Hashing took {total_time:.2f} seconds,"
f"{percentage:.2f}% of the total runtime.")
if __name__ == "__main__":
parser = argparse.ArgumentParser(
description='Benchmark the performance of hashing function in'
'automatic prefix caching.')
parser.add_argument('--model', type=str, default='lmsys/longchat-7b-16k')
parser.add_argument('--tensor-parallel-size', '-tp', type=int, default=1)
parser.add_argument('--output-len', type=int, default=10)
parser.add_argument('--enable-prefix-caching',
action='store_true',
help='enable prefix caching')
parser.add_argument('--use-v2-block-manager',
action='store_true',
help='Use BlockSpaceMangerV2')
args = parser.parse_args()
main(args)
benchmarks/sonnet.txt 0 → 100644
View file @ d824b9d3
FROM fairest creatures we desire increase,
That thereby beauty's rose might never die,
But as the riper should by time decease,
His tender heir might bear his memory:
But thou, contracted to thine own bright eyes,
Feed'st thy light'st flame with self-substantial fuel,
Making a famine where abundance lies,
Thyself thy foe, to thy sweet self too cruel.
Thou that art now the world's fresh ornament
And only herald to the gaudy spring,
Within thine own bud buriest thy content
And, tender churl, makest waste in niggarding.
Pity the world, or else this glutton be,
To eat the world's due, by the grave and thee.
When forty winters shall beseige thy brow,
And dig deep trenches in thy beauty's field,
Thy youth's proud livery, so gazed on now,
Will be a tatter'd weed, of small worth held:
Then being ask'd where all thy beauty lies,
Where all the treasure of thy lusty days,
To say, within thine own deep-sunken eyes,
Were an all-eating shame and thriftless praise.
How much more praise deserved thy beauty's use,
If thou couldst answer 'This fair child of mine
Shall sum my count and make my old excuse,'
Proving his beauty by succession thine!
This were to be new made when thou art old,
And see thy blood warm when thou feel'st it cold.
Look in thy glass, and tell the face thou viewest
Now is the time that face should form another;
Whose fresh repair if now thou not renewest,
Thou dost beguile the world, unbless some mother.
For where is she so fair whose unear'd womb
Disdains the tillage of thy husbandry?
Or who is he so fond will be the tomb
Of his self-love, to stop posterity?
Thou art thy mother's glass, and she in thee
Calls back the lovely April of her prime:
So thou through windows of thine age shall see
Despite of wrinkles this thy golden time.
But if thou live, remember'd not to be,
Die single, and thine image dies with thee.
Unthrifty loveliness, why dost thou spend
Upon thyself thy beauty's legacy?
Nature's bequest gives nothing but doth lend,
And being frank she lends to those are free.
Then, beauteous niggard, why dost thou abuse
The bounteous largess given thee to give?
Profitless usurer, why dost thou use
So great a sum of sums, yet canst not live?
For having traffic with thyself alone,
Thou of thyself thy sweet self dost deceive.
Then how, when nature calls thee to be gone,
What acceptable audit canst thou leave?
Thy unused beauty must be tomb'd with thee,
Which, used, lives th' executor to be.
Those hours, that with gentle work did frame
The lovely gaze where every eye doth dwell,
Will play the tyrants to the very same
And that unfair which fairly doth excel:
For never-resting time leads summer on
To hideous winter and confounds him there;
Sap cheque'd with frost and lusty leaves quite gone,
Beauty o'ersnow'd and bareness every where:
Then, were not summer's distillation left,
A liquid prisoner pent in walls of glass,
Beauty's effect with beauty were bereft,
Nor it nor no remembrance what it was:
But flowers distill'd though they with winter meet,
Leese but their show; their substance still lives sweet.
Then let not winter's ragged hand deface
In thee thy summer, ere thou be distill'd:
Make sweet some vial; treasure thou some place
With beauty's treasure, ere it be self-kill'd.
That use is not forbidden usury,
Which happies those that pay the willing loan;
That's for thyself to breed another thee,
Or ten times happier, be it ten for one;
Ten times thyself were happier than thou art,
If ten of thine ten times refigured thee:
Then what could death do, if thou shouldst depart,
Leaving thee living in posterity?
Be not self-will'd, for thou art much too fair
To be death's conquest and make worms thine heir.
Lo! in the orient when the gracious light
Lifts up his burning head, each under eye
Doth homage to his new-appearing sight,
Serving with looks his sacred majesty;
And having climb'd the steep-up heavenly hill,
Resembling strong youth in his middle age,
yet mortal looks adore his beauty still,
Attending on his golden pilgrimage;
But when from highmost pitch, with weary car,
Like feeble age, he reeleth from the day,
The eyes, 'fore duteous, now converted are
From his low tract and look another way:
So thou, thyself out-going in thy noon,
Unlook'd on diest, unless thou get a son.
Music to hear, why hear'st thou music sadly?
Sweets with sweets war not, joy delights in joy.
Why lovest thou that which thou receivest not gladly,
Or else receivest with pleasure thine annoy?
If the true concord of well-tuned sounds,
By unions married, do offend thine ear,
They do but sweetly chide thee, who confounds
In singleness the parts that thou shouldst bear.
Mark how one string, sweet husband to another,
Strikes each in each by mutual ordering,
Resembling sire and child and happy mother
Who all in one, one pleasing note do sing:
Whose speechless song, being many, seeming one,
Sings this to thee: 'thou single wilt prove none.'
Is it for fear to wet a widow's eye
That thou consumest thyself in single life?
Ah! if thou issueless shalt hap to die.
The world will wail thee, like a makeless wife;
The world will be thy widow and still weep
That thou no form of thee hast left behind,
When every private widow well may keep
By children's eyes her husband's shape in mind.
Look, what an unthrift in the world doth spend
Shifts but his place, for still the world enjoys it;
But beauty's waste hath in the world an end,
And kept unused, the user so destroys it.
No love toward others in that bosom sits
That on himself such murderous shame commits.
For shame! deny that thou bear'st love to any,
Who for thyself art so unprovident.
Grant, if thou wilt, thou art beloved of many,
But that thou none lovest is most evident;
For thou art so possess'd with murderous hate
That 'gainst thyself thou stick'st not to conspire.
Seeking that beauteous roof to ruinate
Which to repair should be thy chief desire.
O, change thy thought, that I may change my mind!
Shall hate be fairer lodged than gentle love?
Be, as thy presence is, gracious and kind,
Or to thyself at least kind-hearted prove:
Make thee another self, for love of me,
That beauty still may live in thine or thee.
As fast as thou shalt wane, so fast thou growest
In one of thine, from that which thou departest;
And that fresh blood which youngly thou bestowest
Thou mayst call thine when thou from youth convertest.
Herein lives wisdom, beauty and increase:
Without this, folly, age and cold decay:
If all were minded so, the times should cease
And threescore year would make the world away.
Let those whom Nature hath not made for store,
Harsh featureless and rude, barrenly perish:
Look, whom she best endow'd she gave the more;
Which bounteous gift thou shouldst in bounty cherish:
She carved thee for her seal, and meant thereby
Thou shouldst print more, not let that copy die.
When I do count the clock that tells the time,
And see the brave day sunk in hideous night;
When I behold the violet past prime,
And sable curls all silver'd o'er with white;
When lofty trees I see barren of leaves
Which erst from heat did canopy the herd,
And summer's green all girded up in sheaves
Borne on the bier with white and bristly beard,
Then of thy beauty do I question make,
That thou among the wastes of time must go,
Since sweets and beauties do themselves forsake
And die as fast as they see others grow;
And nothing 'gainst Time's scythe can make defence
Save breed, to brave him when he takes thee hence.
O, that you were yourself! but, love, you are
No longer yours than you yourself here live:
Against this coming end you should prepare,
And your sweet semblance to some other give.
So should that beauty which you hold in lease
Find no determination: then you were
Yourself again after yourself's decease,
When your sweet issue your sweet form should bear.
Who lets so fair a house fall to decay,
Which husbandry in honour might uphold
Against the stormy gusts of winter's day
And barren rage of death's eternal cold?
O, none but unthrifts! Dear my love, you know
You had a father: let your son say so.
Not from the stars do I my judgment pluck;
And yet methinks I have astronomy,
But not to tell of good or evil luck,
Of plagues, of dearths, or seasons' quality;
Nor can I fortune to brief minutes tell,
Pointing to each his thunder, rain and wind,
Or say with princes if it shall go well,
By oft predict that I in heaven find:
But from thine eyes my knowledge I derive,
And, constant stars, in them I read such art
As truth and beauty shall together thrive,
If from thyself to store thou wouldst convert;
Or else of thee this I prognosticate:
Thy end is truth's and beauty's doom and date.
When I consider every thing that grows
Holds in perfection but a little moment,
That this huge stage presenteth nought but shows
Whereon the stars in secret influence comment;
When I perceive that men as plants increase,
Cheered and cheque'd even by the self-same sky,
Vaunt in their youthful sap, at height decrease,
And wear their brave state out of memory;
Then the conceit of this inconstant stay
Sets you most rich in youth before my sight,
Where wasteful Time debateth with Decay,
To change your day of youth to sullied night;
And all in war with Time for love of you,
As he takes from you, I engraft you new.
But wherefore do not you a mightier way
Make war upon this bloody tyrant, Time?
And fortify yourself in your decay
With means more blessed than my barren rhyme?
Now stand you on the top of happy hours,
And many maiden gardens yet unset
With virtuous wish would bear your living flowers,
Much liker than your painted counterfeit:
So should the lines of life that life repair,
Which this, Time's pencil, or my pupil pen,
Neither in inward worth nor outward fair,
Can make you live yourself in eyes of men.
To give away yourself keeps yourself still,
And you must live, drawn by your own sweet skill.
Who will believe my verse in time to come,
If it were fill'd with your most high deserts?
Though yet, heaven knows, it is but as a tomb
Which hides your life and shows not half your parts.
If I could write the beauty of your eyes
And in fresh numbers number all your graces,
The age to come would say 'This poet lies:
Such heavenly touches ne'er touch'd earthly faces.'
So should my papers yellow'd with their age
Be scorn'd like old men of less truth than tongue,
And your true rights be term'd a poet's rage
And stretched metre of an antique song:
But were some child of yours alive that time,
You should live twice; in it and in my rhyme.
Shall I compare thee to a summer's day?
Thou art more lovely and more temperate:
Rough winds do shake the darling buds of May,
And summer's lease hath all too short a date:
Sometime too hot the eye of heaven shines,
And often is his gold complexion dimm'd;
And every fair from fair sometime declines,
By chance or nature's changing course untrimm'd;
But thy eternal summer shall not fade
Nor lose possession of that fair thou owest;
Nor shall Death brag thou wander'st in his shade,
When in eternal lines to time thou growest:
So long as men can breathe or eyes can see,
So long lives this and this gives life to thee.
Devouring Time, blunt thou the lion's paws,
And make the earth devour her own sweet brood;
Pluck the keen teeth from the fierce tiger's jaws,
And burn the long-lived phoenix in her blood;
Make glad and sorry seasons as thou fleets,
And do whate'er thou wilt, swift-footed Time,
To the wide world and all her fading sweets;
But I forbid thee one most heinous crime:
O, carve not with thy hours my love's fair brow,
Nor draw no lines there with thine antique pen;
Him in thy course untainted do allow
For beauty's pattern to succeeding men.
Yet, do thy worst, old Time: despite thy wrong,
My love shall in my verse ever live young.
A woman's face with Nature's own hand painted
Hast thou, the master-mistress of my passion;
A woman's gentle heart, but not acquainted
With shifting change, as is false women's fashion;
An eye more bright than theirs, less false in rolling,
Gilding the object whereupon it gazeth;
A man in hue, all 'hues' in his controlling,
Much steals men's eyes and women's souls amazeth.
And for a woman wert thou first created;
Till Nature, as she wrought thee, fell a-doting,
And by addition me of thee defeated,
By adding one thing to my purpose nothing.
But since she prick'd thee out for women's pleasure,
Mine be thy love and thy love's use their treasure.
So is it not with me as with that Muse
Stirr'd by a painted beauty to his verse,
Who heaven itself for ornament doth use
And every fair with his fair doth rehearse
Making a couplement of proud compare,
With sun and moon, with earth and sea's rich gems,
With April's first-born flowers, and all things rare
That heaven's air in this huge rondure hems.
O' let me, true in love, but truly write,
And then believe me, my love is as fair
As any mother's child, though not so bright
As those gold candles fix'd in heaven's air:
Let them say more than like of hearsay well;
I will not praise that purpose not to sell.
My glass shall not persuade me I am old,
So long as youth and thou are of one date;
But when in thee time's furrows I behold,
Then look I death my days should expiate.
For all that beauty that doth cover thee
Is but the seemly raiment of my heart,
Which in thy breast doth live, as thine in me:
How can I then be elder than thou art?
O, therefore, love, be of thyself so wary
As I, not for myself, but for thee will;
Bearing thy heart, which I will keep so chary
As tender nurse her babe from faring ill.
Presume not on thy heart when mine is slain;
Thou gavest me thine, not to give back again.
As an unperfect actor on the stage
Who with his fear is put besides his part,
Or some fierce thing replete with too much rage,
Whose strength's abundance weakens his own heart.
So I, for fear of trust, forget to say
The perfect ceremony of love's rite,
And in mine own love's strength seem to decay,
O'ercharged with burden of mine own love's might.
O, let my books be then the eloquence
And dumb presagers of my speaking breast,
Who plead for love and look for recompense
More than that tongue that more hath more express'd.
O, learn to read what silent love hath writ:
To hear with eyes belongs to love's fine wit.
Mine eye hath play'd the painter and hath stell'd
Thy beauty's form in table of my heart;
My body is the frame wherein 'tis held,
And perspective it is the painter's art.
For through the painter must you see his skill,
To find where your true image pictured lies;
Which in my bosom's shop is hanging still,
That hath his windows glazed with thine eyes.
Now see what good turns eyes for eyes have done:
Mine eyes have drawn thy shape, and thine for me
Are windows to my breast, where-through the sun
Delights to peep, to gaze therein on thee;
Yet eyes this cunning want to grace their art;
They draw but what they see, know not the heart.
Let those who are in favour with their stars
Of public honour and proud titles boast,
Whilst I, whom fortune of such triumph bars,
Unlook'd for joy in that I honour most.
Great princes' favourites their fair leaves spread
But as the marigold at the sun's eye,
And in themselves their pride lies buried,
For at a frown they in their glory die.
The painful warrior famoused for fight,
After a thousand victories once foil'd,
Is from the book of honour razed quite,
And all the rest forgot for which he toil'd:
Then happy I, that love and am beloved
Where I may not remove nor be removed.
Lord of my love, to whom in vassalage
Thy merit hath my duty strongly knit,
To thee I send this written embassage,
To witness duty, not to show my wit:
Duty so great, which wit so poor as mine
May make seem bare, in wanting words to show it,
But that I hope some good conceit of thine
In thy soul's thought, all naked, will bestow it;
Till whatsoever star that guides my moving
Points on me graciously with fair aspect
And puts apparel on my tatter'd loving,
To show me worthy of thy sweet respect:
Then may I dare to boast how I do love thee;
Till then not show my head where thou mayst prove me.
Weary with toil, I haste me to my bed,
The dear repose for limbs with travel tired;
But then begins a journey in my head,
To work my mind, when body's work's expired:
For then my thoughts, from far where I abide,
Intend a zealous pilgrimage to thee,
And keep my drooping eyelids open wide,
Looking on darkness which the blind do see
Save that my soul's imaginary sight
Presents thy shadow to my sightless view,
Which, like a jewel hung in ghastly night,
Makes black night beauteous and her old face new.
Lo! thus, by day my limbs, by night my mind,
For thee and for myself no quiet find.
How can I then return in happy plight,
That am debarr'd the benefit of rest?
When day's oppression is not eased by night,
But day by night, and night by day, oppress'd?
And each, though enemies to either's reign,
Do in consent shake hands to torture me;
The one by toil, the other to complain
How far I toil, still farther off from thee.
I tell the day, to please them thou art bright
And dost him grace when clouds do blot the heaven:
So flatter I the swart-complexion'd night,
When sparkling stars twire not thou gild'st the even.
But day doth daily draw my sorrows longer
And night doth nightly make grief's strength seem stronger.
When, in disgrace with fortune and men's eyes,
I all alone beweep my outcast state
And trouble deal heaven with my bootless cries
And look upon myself and curse my fate,
Wishing me like to one more rich in hope,
Featured like him, like him with friends possess'd,
Desiring this man's art and that man's scope,
With what I most enjoy contented least;
Yet in these thoughts myself almost despising,
Haply I think on thee, and then my state,
Like to the lark at break of day arising
From sullen earth, sings hymns at heaven's gate;
For thy sweet love remember'd such wealth brings
That then I scorn to change my state with kings.
When to the sessions of sweet silent thought
I summon up remembrance of things past,
I sigh the lack of many a thing I sought,
And with old woes new wail my dear time's waste:
Then can I drown an eye, unused to flow,
For precious friends hid in death's dateless night,
And weep afresh love's long since cancell'd woe,
And moan the expense of many a vanish'd sight:
Then can I grieve at grievances foregone,
And heavily from woe to woe tell o'er
The sad account of fore-bemoaned moan,
Which I new pay as if not paid before.
But if the while I think on thee, dear friend,
All losses are restored and sorrows end.
Thy bosom is endeared with all hearts,
Which I by lacking have supposed dead,
And there reigns love and all love's loving parts,
And all those friends which I thought buried.
How many a holy and obsequious tear
Hath dear religious love stol'n from mine eye
As interest of the dead, which now appear
But things removed that hidden in thee lie!
Thou art the grave where buried love doth live,
Hung with the trophies of my lovers gone,
Who all their parts of me to thee did give;
That due of many now is thine alone:
Their images I loved I view in thee,
And thou, all they, hast all the all of me.
If thou survive my well-contented day,
When that churl Death my bones with dust shall cover,
And shalt by fortune once more re-survey
These poor rude lines of thy deceased lover,
Compare them with the bettering of the time,
And though they be outstripp'd by every pen,
Reserve them for my love, not for their rhyme,
Exceeded by the height of happier men.
O, then vouchsafe me but this loving thought:
'Had my friend's Muse grown with this growing age,
A dearer birth than this his love had brought,
To march in ranks of better equipage:
But since he died and poets better prove,
Theirs for their style I'll read, his for his love.'
Full many a glorious morning have I seen
Flatter the mountain-tops with sovereign eye,
Kissing with golden face the meadows green,
Gilding pale streams with heavenly alchemy;
Anon permit the basest clouds to ride
With ugly rack on his celestial face,
And from the forlorn world his visage hide,
Stealing unseen to west with this disgrace:
Even so my sun one early morn did shine
With all triumphant splendor on my brow;
But out, alack! he was but one hour mine;
The region cloud hath mask'd him from me now.
Yet him for this my love no whit disdaineth;
Suns of the world may stain when heaven's sun staineth.
Why didst thou promise such a beauteous day,
And make me travel forth without my cloak,
To let base clouds o'ertake me in my way,
Hiding thy bravery in their rotten smoke?
'Tis not enough that through the cloud thou break,
To dry the rain on my storm-beaten face,
For no man well of such a salve can speak
That heals the wound and cures not the disgrace:
Nor can thy shame give physic to my grief;
Though thou repent, yet I have still the loss:
The offender's sorrow lends but weak relief
To him that bears the strong offence's cross.
Ah! but those tears are pearl which thy love sheds,
And they are rich and ransom all ill deeds.
No more be grieved at that which thou hast done:
Roses have thorns, and silver fountains mud;
Clouds and eclipses stain both moon and sun,
And loathsome canker lives in sweetest bud.
All men make faults, and even I in this,
Authorizing thy trespass with compare,
Myself corrupting, salving thy amiss,
Excusing thy sins more than thy sins are;
For to thy sensual fault I bring in sense--
Thy adverse party is thy advocate--
And 'gainst myself a lawful plea commence:
Such civil war is in my love and hate
That I an accessary needs must be
To that sweet thief which sourly robs from me.
Let me confess that we two must be twain,
Although our undivided loves are one:
So shall those blots that do with me remain
Without thy help by me be borne alone.
In our two loves there is but one respect,
Though in our lives a separable spite,
Which though it alter not love's sole effect,
Yet doth it steal sweet hours from love's delight.
I may not evermore acknowledge thee,
Lest my bewailed guilt should do thee shame,
Nor thou with public kindness honour me,
Unless thou take that honour from thy name:
But do not so; I love thee in such sort
As, thou being mine, mine is thy good report.
As a decrepit father takes delight
To see his active child do deeds of youth,
So I, made lame by fortune's dearest spite,
Take all my comfort of thy worth and truth.
For whether beauty, birth, or wealth, or wit,
Or any of these all, or all, or more,
Entitled in thy parts do crowned sit,
I make my love engrafted to this store:
So then I am not lame, poor, nor despised,
Whilst that this shadow doth such substance give
That I in thy abundance am sufficed
And by a part of all thy glory live.
Look, what is best, that best I wish in thee:
This wish I have; then ten times happy me!
\ No newline at end of file
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment