mirror of
https://github.com/PaddlePaddle/FastDeploy.git
synced 2026-04-23 00:17:25 +08:00
[Docs] Add docs for disaggregated deployment (#6700)
* add docs for disaggregated deployment * pre-commit run for style check * update docs
This commit is contained in:
Jingfeng Wu
@@ -0,0 +1,254 @@
|
|||||||
|
[简体中文](../zh/best_practices/Disaggregated.md)
|
||||||
|
|
||||||
|
# PD Disaggregated Deployment Best Practices
|
||||||
|
|
||||||
|
This document provides a comprehensive guide to FastDeploy's PD (Prefill-Decode) disaggregated deployment solution, covering both single-machine and cross-machine deployment modes with support for Tensor Parallelism (TP), Data Parallelism (DP), and Expert Parallelism (EP).
|
||||||
|
|
||||||
|
## 1. Deployment Overview and Environment Preparation
|
||||||
|
|
||||||
|
This guide demonstrates deployment practices using the ERNIE-4.5-300B-A47B-Paddle model on H100 80GB GPUs. Below are the minimum GPU requirements for different deployment configurations:
|
||||||
|
|
||||||
|
**Single-Machine Deployment (8 GPUs, Single Node)**
|
||||||
|
|
||||||
|
| Configuration | TP | DP | EP | GPUs Required |
|
||||||
|
|---------|----|----|----|---------|
|
||||||
|
| P：TP4DP1<br>D：TP4DP1 | 4 | 1 | - | 8 |
|
||||||
|
| P：TP1DP4EP4 <br> D：TP1DP4EP4| 1 | 4 | ✓ | 8 |
|
||||||
|
|
||||||
|
**Multi-Machine Deployment (16 GPUs, Cross-Node)**
|
||||||
|
|
||||||
|
| Configuration | TP | DP | EP | GPUs Required |
|
||||||
|
|---------|----|----|----|---------|
|
||||||
|
| P：TP8DP1<br>D：TP8DP1 | 8 | 1 | - | 16 |
|
||||||
|
| P：TP4DP2<br>D：TP4DP2 | 4 | 2 | - | 16 |
|
||||||
|
| P：TP1DP8EP8<br>D：TP1DP8EP8 | 1 | 8 | ✓ | 16 |
|
||||||
|
|
||||||
|
**Important Notes**:
|
||||||
|
1. **Quantization**: All configurations above use WINT4 quantization, specified via `--quantization wint4`
|
||||||
|
2. **EP Limitations**: When Expert Parallelism (EP) is enabled, only TP=1 is currently supported; multi-TP scenarios are not yet available
|
||||||
|
3. **Cross-Machine Network**: Cross-machine deployment requires RDMA network support for high-speed KV Cache transmission
|
||||||
|
4. **GPU Calculation**: Total GPUs = TP × DP × 2, with identical configurations for both Prefill and Decode instances
|
||||||
|
5. **CUDA Graph Capture**: Decode instances enable CUDA Graph capture by default for inference acceleration, while Prefill instances do not
|
||||||
|
|
||||||
|
### 1.1 Installing FastDeploy
|
||||||
|
|
||||||
|
Please refer to the [FastDeploy Installation Guide](https://paddlepaddle.github.io/FastDeploy/zh/install/) to set up your environment.
|
||||||
|
|
||||||
|
For model downloads, please check the [Supported Models List](https://paddlepaddle.github.io/FastDeploy/zh/model_summary/).
|
||||||
|
|
||||||
|
### 1.2 Deployment Topology
|
||||||
|
|
||||||
|
**Single-Machine Deployment Topology**
|
||||||
|
|
||||||
|
```
|
||||||
|
┌──────────────────────────────┐
|
||||||
|
│ Single Machine 8×H100 80GB │
|
||||||
|
│ ┌──────────────┐ │
|
||||||
|
│ │ Router │ │
|
||||||
|
│ │ 0.0.0.0:8109│ │
|
||||||
|
│ └──────────────┘ │
|
||||||
|
│ │ │
|
||||||
|
│ ┌────┴────┐ │
|
||||||
|
│ ▼ ▼ │
|
||||||
|
│ ┌─────────┐ ┌─────────┐ │
|
||||||
|
│ │Prefill │ │Decode │ │
|
||||||
|
│ │GPU 0-3 │ │GPU 4-7 │ │
|
||||||
|
│ └─────────┘ └─────────┘ │
|
||||||
|
└──────────────────────────────┘
|
||||||
|
```
|
||||||
|
|
||||||
|
**Cross-Machine Deployment Topology**
|
||||||
|
|
||||||
|
```
|
||||||
|
┌─────────────────────┐ ┌─────────────────────┐
|
||||||
|
│ Prefill Machine │ RDMA Network │ Decode Machine │
|
||||||
|
│ 8×H100 80GB │◄────────────────────►│ 8×H100 80GB │
|
||||||
|
│ │ │ │
|
||||||
|
│ ┌──────────────┐ │ │ │
|
||||||
|
│ │ Router │ │ │ │
|
||||||
|
│ │ 0.0.0.0:8109 │───┼──────────────────────┼────────── │
|
||||||
|
│ └──────────────┘ │ │ │ │
|
||||||
|
│ │ │ │ │ │
|
||||||
|
│ ▼ │ │ ▼ │
|
||||||
|
│ ┌──────────────┐ │ │ ┌──────────────┐ │
|
||||||
|
│ │Prefill Nodes │ │ │ │Decode Nodes │ │
|
||||||
|
│ │GPU 0-7 │ │ │ │GPU 0-7 │ │
|
||||||
|
│ └──────────────┘ │ │ └──────────────┘ │
|
||||||
|
└─────────────────────┘ └─────────────────────┘
|
||||||
|
```
|
||||||
|
|
||||||
|
---
|
||||||
|
## 2. Single-Machine PD Disaggregated Deployment
|
||||||
|
|
||||||
|
### 2.1 Test Scenarios and Parallelism Configuration
|
||||||
|
|
||||||
|
This chapter demonstrates the **TP4DP1｜D：TP4DP1** configuration test scenario:
|
||||||
|
- **Tensor Parallelism (TP)**: 4 — Each 4 GPUs independently load complete model parameters
|
||||||
|
- **Data Parallelism (DP)**: 1 — Each GPU forms a data parallelism group
|
||||||
|
- **Expert Parallelism (EP)**: Not enabled
|
||||||
|
|
||||||
|
**To test other parallelism configurations, adjust parameters as follows:**
|
||||||
|
1. **TP Adjustment**: Modify `--tensor-parallel-size`
|
||||||
|
2. **DP Adjustment**: Modify `--data-parallel-size`, ensuring `--ports` and `--num-servers` remain consistent with DP
|
||||||
|
3. **EP Toggle**: Add or remove `--enable-expert-parallel`
|
||||||
|
4. **GPU Allocation**: Control GPUs used by Prefill and Decode instances via `CUDA_VISIBLE_DEVICES`
|
||||||
|
|
||||||
|
### 2.2 Startup Scripts
|
||||||
|
|
||||||
|
#### Start Router
|
||||||
|
|
||||||
|
```bash
|
||||||
|
python -m fastdeploy.router.launch \
|
||||||
|
--port 8109 \
|
||||||
|
--splitwise
|
||||||
|
```
|
||||||
|
|
||||||
|
Note: This uses the Python version of the router. If needed, you can also use the high-performance [Golang version router](../online_serving/router.md).
|
||||||
|
|
||||||
|
#### Start Prefill Nodes
|
||||||
|
|
||||||
|
```bash
|
||||||
|
export CUDA_VISIBLE_DEVICES=0,1,2,3
|
||||||
|
|
||||||
|
python -m fastdeploy.entrypoints.openai.api_server \
|
||||||
|
--model /path/to/ERNIE-4.5-300B-A47B-Paddle \
|
||||||
|
--port 8188 \
|
||||||
|
--splitwise-role "prefill" \
|
||||||
|
--cache-transfer-protocol "rdma,ipc" \
|
||||||
|
--router "0.0.0.0:8109" \
|
||||||
|
--quantization wint4 \
|
||||||
|
--tensor-parallel-size 4 \
|
||||||
|
--data-parallel-size 1 \
|
||||||
|
--max-model-len 8192 \
|
||||||
|
--max-num-seqs 64
|
||||||
|
```
|
||||||
|
|
||||||
|
#### Start Decode Nodes
|
||||||
|
|
||||||
|
```bash
|
||||||
|
export CUDA_VISIBLE_DEVICES=4,5,6,7
|
||||||
|
|
||||||
|
python -m fastdeploy.entrypoints.openai.multi_api_server \
|
||||||
|
--model /path/to/ERNIE-4.5-300B-A47B-Paddle \
|
||||||
|
--ports 8200,8201 \
|
||||||
|
--splitwise-role "decode" \
|
||||||
|
--cache-transfer-protocol "rdma,ipc" \
|
||||||
|
--router "0.0.0.0:8109" \
|
||||||
|
--quantization wint4 \
|
||||||
|
--tensor-parallel-size 2 \
|
||||||
|
--data-parallel-size 2 \
|
||||||
|
--max-model-len 8192 \
|
||||||
|
--max-num-seqs 64
|
||||||
|
```
|
||||||
|
|
||||||
|
### 2.3 Key Parameter Descriptions
|
||||||
|
|
||||||
|
| Parameter | Description |
|
||||||
|
|-----|------|
|
||||||
|
| `--splitwise` | Enable PD disaggregated mode |
|
||||||
|
| `--splitwise-role` | Node role: `prefill` or `decode` |
|
||||||
|
| `--cache-transfer-protocol` | KV Cache transfer protocol: `rdma` or `ipc` |
|
||||||
|
| `--router` | Router service address |
|
||||||
|
| `--quantization` | Quantization strategy (wint4/wint8/fp8, etc.) |
|
||||||
|
| `--tensor-parallel-size` | Tensor parallelism degree (TP) |
|
||||||
|
| `--data-parallel-size` | Data parallelism degree (DP) |
|
||||||
|
| `--max-model-len` | Maximum sequence length |
|
||||||
|
| `--max-num-seqs` | Maximum concurrent sequences |
|
||||||
|
| `--num-gpu-blocks-override` | GPU KV Cache block count override |
|
||||||
|
|
||||||
|
---
|
||||||
|
|
||||||
|
## 3. Cross-Machine PD Disaggregated Deployment
|
||||||
|
|
||||||
|
### 3.1 Deployment Principles
|
||||||
|
|
||||||
|
Cross-machine PD disaggregation deploys Prefill and Decode instances on different physical machines:
|
||||||
|
- **Prefill Machine**: Runs the Router and Prefill nodes, responsible for processing input sequence prefill computation
|
||||||
|
- **Decode Machine**: Runs Decode nodes, communicates with the Prefill machine via RDMA network, responsible for autoregressive decoding generation
|
||||||
|
|
||||||
|
### 3.2 Test Scenarios and Parallelism Configuration
|
||||||
|
|
||||||
|
This chapter demonstrates the **TP1DP8EP8｜D：TP1DP8EP8** cross-machine configuration (16 GPUs total):
|
||||||
|
- **Tensor Parallelism (TP)**: 1
|
||||||
|
- **Data Parallelism (DP)**: 8 — 8 GPUs per machine, totaling 8 Prefill instances and 8 Decode instances
|
||||||
|
- **Expert Parallelism (EP)**: Enabled — MoE layer shared experts are distributed across 8 GPUs for parallel computation
|
||||||
|
|
||||||
|
**To test other cross-machine parallelism configurations, adjust parameters as follows:**
|
||||||
|
1. **Inter-Machine Communication**: Ensure RDMA network connectivity between machines; Prefill machine needs `KVCACHE_RDMA_NICS` environment variable configured
|
||||||
|
2. **Router Address**: The `--router` parameter on the Decode machine must point to the actual IP address of the Prefill machine
|
||||||
|
3. **Port Configuration**: The number of ports in the `--ports` list must match `--num-servers` and `--data-parallel-size`
|
||||||
|
4. **GPU Visibility**: Each machine specifies its local GPUs via `CUDA_VISIBLE_DEVICES`
|
||||||
|
|
||||||
|
### 3.3 Prefill Machine Startup Scripts
|
||||||
|
|
||||||
|
#### Start Router
|
||||||
|
|
||||||
|
```bash
|
||||||
|
unset http_proxy && unset https_proxy
|
||||||
|
|
||||||
|
python -m fastdeploy.router.launch \
|
||||||
|
--port 8109 \
|
||||||
|
--splitwise
|
||||||
|
```
|
||||||
|
|
||||||
|
#### Start Prefill Nodes
|
||||||
|
|
||||||
|
```bash
|
||||||
|
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
|
||||||
|
|
||||||
|
python -m fastdeploy.entrypoints.openai.multi_api_server \
|
||||||
|
--ports 8198,8199,8200,8201,8202,8203,8204,8205 \
|
||||||
|
--num-servers 8 \
|
||||||
|
--args --model /path/to/ERNIE-4.5-300B-A47B-Paddle \
|
||||||
|
--splitwise-role "prefill" \
|
||||||
|
--cache-transfer-protocol "rdma,ipc" \
|
||||||
|
--router "<ROUTER_MACHINE_IP>:8109" \
|
||||||
|
--quantization wint4 \
|
||||||
|
--tensor-parallel-size 1 \
|
||||||
|
--data-parallel-size 8 \
|
||||||
|
--enable-expert-parallel \
|
||||||
|
--max-model-len 8192 \
|
||||||
|
--max-num-seqs 64
|
||||||
|
```
|
||||||
|
|
||||||
|
### 3.4 Decode Machine Startup Scripts
|
||||||
|
|
||||||
|
#### Start Decode Nodes
|
||||||
|
|
||||||
|
```bash
|
||||||
|
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
|
||||||
|
|
||||||
|
python -m fastdeploy.entrypoints.openai.multi_api_server \
|
||||||
|
--ports 8198,8199,8200,8201,8202,8203,8204,8205 \
|
||||||
|
--num-servers 8 \
|
||||||
|
--args --model /path/to/ERNIE-4.5-300B-A47B-Paddle \
|
||||||
|
--splitwise-role "decode" \
|
||||||
|
--cache-transfer-protocol "rdma,ipc" \
|
||||||
|
--router "<PREFILL_MACHINE_IP>:8109" \
|
||||||
|
--quantization wint4 \
|
||||||
|
--tensor-parallel-size 1 \
|
||||||
|
--data-parallel-size 8 \
|
||||||
|
--enable-expert-parallel \
|
||||||
|
--max-model-len 8192 \
|
||||||
|
--max-num-seqs 64
|
||||||
|
```
|
||||||
|
|
||||||
|
**Note**: Please replace `<PREFILL_MACHINE_IP>` with the actual IP address of the Prefill machine.
|
||||||
|
|
||||||
|
## 4. Sending Test Requests
|
||||||
|
|
||||||
|
```bash
|
||||||
|
curl -X POST "http://localhost:8109/v1/chat/completions" \
|
||||||
|
-H "Content-Type: application/json" \
|
||||||
|
-d '{
|
||||||
|
"messages": [
|
||||||
|
{"role": "user", "content": "你好，请介绍一下自己。"}
|
||||||
|
],
|
||||||
|
"max_tokens": 100,
|
||||||
|
"stream": false
|
||||||
|
}'
|
||||||
|
```
|
||||||
|
|
||||||
|
## 5. Frequently Asked Questions (FAQ)
|
||||||
|
|
||||||
|
If you encounter issues during use, please refer to [FAQ](./FAQ.md) for solutions.
|
||||||
@@ -9,3 +9,4 @@
|
|||||||
- [ERNIE-4.5-VL-28B-A3B-Paddle](ERNIE-4.5-VL-28B-A3B-Paddle.md)
|
- [ERNIE-4.5-VL-28B-A3B-Paddle](ERNIE-4.5-VL-28B-A3B-Paddle.md)
|
||||||
- [ERNIE-4.5-VL-424B-A47B-Paddle](ERNIE-4.5-VL-424B-A47B-Paddle.md)
|
- [ERNIE-4.5-VL-424B-A47B-Paddle](ERNIE-4.5-VL-424B-A47B-Paddle.md)
|
||||||
- [PaddleOCR-VL-0.9B.md](PaddleOCR-VL-0.9B.md)
|
- [PaddleOCR-VL-0.9B.md](PaddleOCR-VL-0.9B.md)
|
||||||
|
- [Disaggregated.md](Disaggregated.md)
|
||||||
|
|||||||
@@ -1,5 +1,7 @@
|
|||||||
[简体中文](../zh/features/disaggregated.md)
|
[简体中文](../zh/features/disaggregated.md)
|
||||||
|
|
||||||
|
[Best Practice](../best_practices/Disaggregated.md)
|
||||||
|
|
||||||
# Disaggregated Deployment
|
# Disaggregated Deployment
|
||||||
|
|
||||||
Large Language Model (LLM) inference is divided into two phases: **Prefill** and **Decode**, which are compute-intensive and memory-bound, respectively.
|
Large Language Model (LLM) inference is divided into two phases: **Prefill** and **Decode**, which are compute-intensive and memory-bound, respectively.
|
||||||
|
|||||||
@@ -0,0 +1,253 @@
|
|||||||
|
[English](../../best_practices/Disaggregated.md)
|
||||||
|
|
||||||
|
# PD分离部署最佳实践
|
||||||
|
|
||||||
|
本文档详细介绍 FastDeploy 的 PD（Prefill-Decode）分离式部署方案，涵盖单机部署与跨机部署两种模式，支持张量并行（TP）、数据并行（DP）和专家并行（EP）。
|
||||||
|
|
||||||
|
## 一、部署方案概览与环境准备
|
||||||
|
|
||||||
|
本文以 ERNIE-4.5-300B-A47B-Paddle 模型为例进行部署实践，硬件环境采用 H100 80GB GPU。下面例举了不同部署模式下的最小 GPU 卡数需求：
|
||||||
|
|
||||||
|
**单机部署（8卡单节点）**
|
||||||
|
|
||||||
|
| 配置方案 | TP | DP | EP | 所需卡数 |
|
||||||
|
|---------|----|----|----|---------|
|
||||||
|
| P：TP4DP1<br>D：TP4DP1 | 4 | 1 | - | 8 |
|
||||||
|
| P：TP1DP4EP4 <br> D：TP1DP4EP4| 1 | 4 | ✓ | 8 |
|
||||||
|
|
||||||
|
**多机部署（16卡跨节点）**
|
||||||
|
|
||||||
|
| 配置方案 | TP | DP | EP | 所需卡数 |
|
||||||
|
|---------|----|----|----|---------|
|
||||||
|
| P：TP8DP1<br>D：TP8DP1 | 8 | 1 | - | 16 |
|
||||||
|
| P：TP4DP2<br>D：TP4DP2 | 4 | 2 | - | 16 |
|
||||||
|
| P：TP1DP8EP8<br>D：TP1DP8EP8 | 1 | 8 | ✓ | 16 |
|
||||||
|
|
||||||
|
**重要说明**：
|
||||||
|
1. **量化精度**：以上所有配置均采用 WINT4 量化，通过 `--quantization wint4` 参数指定
|
||||||
|
2. **EP 限制**：开启专家并行（EP）后，当前仅支持 TP=1，暂不支持多 TP 场景
|
||||||
|
3. **跨机网络**：跨机部署依赖 RDMA 网络实现 KV Cache 的高速传输
|
||||||
|
4. **卡数计算**：总卡数 = TP × DP × 2（Prefill 实例与 Decode 实例配置相同）
|
||||||
|
5. **CUDA Graph 捕获**：Decode 实例默认启用 CUDA Graph 捕获以加速推理，Prefill 实例默认不启用
|
||||||
|
|
||||||
|
### 1.1 安装 FastDeploy
|
||||||
|
|
||||||
|
请参考 [FastDeploy 安装指南](https://paddlepaddle.github.io/FastDeploy/zh/install/) 完成环境搭建。
|
||||||
|
|
||||||
|
模型下载请参考 [支持模型列表](https://paddlepaddle.github.io/FastDeploy/zh/model_summary/)。
|
||||||
|
|
||||||
|
### 1.2 部署拓扑结构
|
||||||
|
|
||||||
|
**单机部署拓扑**
|
||||||
|
|
||||||
|
```
|
||||||
|
┌──────────────────────────────┐
|
||||||
|
│ 单机 8×H100 80GB │
|
||||||
|
│ ┌──────────────┐ │
|
||||||
|
│ │ Router │ │
|
||||||
|
│ │ 0.0.0.0:8109│ │
|
||||||
|
│ └──────────────┘ │
|
||||||
|
│ │ │
|
||||||
|
│ ┌────┴────┐ │
|
||||||
|
│ ▼ ▼ │
|
||||||
|
│ ┌─────────┐ ┌─────────┐ │
|
||||||
|
│ │Prefill │ │Decode │ │
|
||||||
|
│ │GPU 0-3 │ │GPU 4-7 │ │
|
||||||
|
│ └─────────┘ └─────────┘ │
|
||||||
|
└──────────────────────────────┘
|
||||||
|
```
|
||||||
|
|
||||||
|
**跨机部署拓扑**
|
||||||
|
|
||||||
|
```
|
||||||
|
┌─────────────────────┐ ┌─────────────────────┐
|
||||||
|
│ Prefill Machine │ RDMA Network │ Decode Machine │
|
||||||
|
│ 8×H100 80GB │◄────────────────────►│ 8×H100 80GB │
|
||||||
|
│ │ │ │
|
||||||
|
│ ┌──────────────┐ │ │ │
|
||||||
|
│ │ Router │ │ │ │
|
||||||
|
│ │ 0.0.0.0:8109 │───┼──────────────────────┼────────── │
|
||||||
|
│ └──────────────┘ │ │ │ │
|
||||||
|
│ │ │ │ │ │
|
||||||
|
│ ▼ │ │ ▼ │
|
||||||
|
│ ┌──────────────┐ │ │ ┌──────────────┐ │
|
||||||
|
│ │Prefill Nodes │ │ │ │Decode Nodes │ │
|
||||||
|
│ │GPU 0-7 │ │ │ │GPU 0-7 │ │
|
||||||
|
│ └──────────────┘ │ │ └──────────────┘ │
|
||||||
|
└─────────────────────┘ └─────────────────────┘
|
||||||
|
```
|
||||||
|
|
||||||
|
---
|
||||||
|
## 二、单机 PD 分离部署
|
||||||
|
|
||||||
|
### 2.1 测试场景与并行度配置
|
||||||
|
|
||||||
|
本节演示的测试场景为 **P：TP4DP1｜D：TP4DP1** 配置：
|
||||||
|
- **张量并行度（TP）**：4 —— 每4张 GPU 独立加载完整模型参数
|
||||||
|
- **数据并行度（DP）**：1 —— 每张 GPU 组成一个数据并行组
|
||||||
|
- **专家并行（EP）**：不启用
|
||||||
|
|
||||||
|
**若需测试其他并行度配置，请按以下方式调整参数：**
|
||||||
|
1. **TP 调整**：修改 `--tensor-parallel-size`
|
||||||
|
2. **DP 调整**：修改 `--data-parallel-size`，同时确保 `--ports` 和 `--num-servers` 与 DP 保持一致
|
||||||
|
3. **EP 开关**：添加或移除 `--enable-expert-parallel`
|
||||||
|
4. **GPU 分配**：通过 `CUDA_VISIBLE_DEVICES` 控制 Prefill 和 Decode 实例使用的 GPU
|
||||||
|
|
||||||
|
### 2.2 启动脚本
|
||||||
|
|
||||||
|
#### 启动 Router
|
||||||
|
|
||||||
|
```bash
|
||||||
|
python -m fastdeploy.router.launch \
|
||||||
|
--port 8109 \
|
||||||
|
--splitwise
|
||||||
|
```
|
||||||
|
|
||||||
|
注意：这里使用的是python版本router，如果有需要也可以使用高性能的[Golang版本router](../online_serving/router.md)
|
||||||
|
#### 启动 Prefill 节点
|
||||||
|
|
||||||
|
```bash
|
||||||
|
export CUDA_VISIBLE_DEVICES=0,1,2,3
|
||||||
|
|
||||||
|
python -m fastdeploy.entrypoints.openai.api_server \
|
||||||
|
--model /path/to/ERNIE-4.5-300B-A47B-Paddle \
|
||||||
|
--port 8188 \
|
||||||
|
--splitwise-role "prefill" \
|
||||||
|
--cache-transfer-protocol "rdma,ipc" \
|
||||||
|
--router "0.0.0.0:8109" \
|
||||||
|
--quantization wint4 \
|
||||||
|
--tensor-parallel-size 4 \
|
||||||
|
--data-parallel-size 1 \
|
||||||
|
--max-model-len 8192 \
|
||||||
|
--max-num-seqs 64
|
||||||
|
```
|
||||||
|
|
||||||
|
#### 启动 Decode 节点
|
||||||
|
|
||||||
|
```bash
|
||||||
|
export CUDA_VISIBLE_DEVICES=4,5,6,7
|
||||||
|
|
||||||
|
python -m fastdeploy.entrypoints.openai.multi_api_server \
|
||||||
|
--model /path/to/ERNIE-4.5-300B-A47B-Paddle \
|
||||||
|
--ports 8200,8201 \
|
||||||
|
--splitwise-role "decode" \
|
||||||
|
--cache-transfer-protocol "rdma,ipc" \
|
||||||
|
--router "0.0.0.0:8109" \
|
||||||
|
--quantization wint4 \
|
||||||
|
--tensor-parallel-size 2 \
|
||||||
|
--data-parallel-size 2 \
|
||||||
|
--max-model-len 8192 \
|
||||||
|
--max-num-seqs 64
|
||||||
|
```
|
||||||
|
|
||||||
|
### 2.3 关键参数说明
|
||||||
|
|
||||||
|
| 参数 | 说明 |
|
||||||
|
|-----|------|
|
||||||
|
| `--splitwise` | 开启 PD 分离模式 |
|
||||||
|
| `--splitwise-role` | 节点角色：`prefill` 或 `decode` |
|
||||||
|
| `--cache-transfer-protocol` | KV Cache 传输协议：`rdma` 或 `ipc` |
|
||||||
|
| `--router` | Router 服务地址 |
|
||||||
|
| `--quantization` | 量化策略（wint4/wint8/fp8 等） |
|
||||||
|
| `--tensor-parallel-size` | 张量并行度（TP） |
|
||||||
|
| `--data-parallel-size` | 数据并行度（DP） |
|
||||||
|
| `--max-model-len` | 最大序列长度 |
|
||||||
|
| `--max-num-seqs` | 最大并发序列数 |
|
||||||
|
| `--num-gpu-blocks-override` | GPU KV Cache 块数量覆盖值 |
|
||||||
|
|
||||||
|
---
|
||||||
|
|
||||||
|
## 三、跨机 PD 分离部署
|
||||||
|
|
||||||
|
### 3.1 部署原理
|
||||||
|
|
||||||
|
跨机 PD 分离将 Prefill 和 Decode 实例部署在不同物理机器上：
|
||||||
|
- **Prefill 机器**：运行 Router 和 Prefill 节点，负责处理输入序列的预填充计算
|
||||||
|
- **Decode 机器**：运行 Decode 节点，通过 RDMA 网络与 Prefill 机器通信，负责自回归解码生成
|
||||||
|
|
||||||
|
### 3.2 测试场景与并行度配置
|
||||||
|
|
||||||
|
本章节演示的测试场景为 **P：TP1DP8EP8 ｜ D：P：TP1DP8EP8** 跨机配置（共 16 张 GPU）：
|
||||||
|
- **张量并行度（TP）**：1
|
||||||
|
- **数据并行度（DP）**：8 —— 每机 8 张 GPU，共 8 个 Prefill 实例和 8 个 Decode 实例
|
||||||
|
- **专家并行（EP）**：启用—— MoE 层的共享专家分布在8张 GPU 上并行计算
|
||||||
|
|
||||||
|
**若需测试其他跨机并行度配置，请按以下方法调整参数：**
|
||||||
|
1. **机器间通信**：确保两机之间 RDMA 网络连通，Prefill 机器需配置 `KVCACHE_RDMA_NICS` 环境变量
|
||||||
|
2. **Router 地址**：Decode 机器的 `--router` 参数需指向 Prefill 机器的实际 IP 地址
|
||||||
|
3. **端口配置**：`--ports` 列表的端口数量必须与 `--num-servers` 和 `--data-parallel-size` 保持一致
|
||||||
|
4. **GPU 可见性**：每机通过 `CUDA_VISIBLE_DEVICES` 指定本机使用的 GPU
|
||||||
|
|
||||||
|
### 3.3 Prefill 机器启动脚本
|
||||||
|
|
||||||
|
#### 启动 Router
|
||||||
|
|
||||||
|
```bash
|
||||||
|
unset http_proxy && unset https_proxy
|
||||||
|
|
||||||
|
python -m fastdeploy.router.launch \
|
||||||
|
--port 8109 \
|
||||||
|
--splitwise
|
||||||
|
```
|
||||||
|
|
||||||
|
#### 启动 Prefill 节点
|
||||||
|
|
||||||
|
```bash
|
||||||
|
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
|
||||||
|
|
||||||
|
python -m fastdeploy.entrypoints.openai.multi_api_server \
|
||||||
|
--ports 8198,8199,8200,8201,8202,8203,8204,8205 \
|
||||||
|
--num-servers 8 \
|
||||||
|
--args --model /path/to/ERNIE-4.5-300B-A47B-Paddle \
|
||||||
|
--splitwise-role "prefill" \
|
||||||
|
--cache-transfer-protocol "rdma,ipc" \
|
||||||
|
--router "<ROUTER_MACHINE_IP>:8109" \
|
||||||
|
--quantization wint4 \
|
||||||
|
--tensor-parallel-size 1 \
|
||||||
|
--data-parallel-size 8 \
|
||||||
|
--enable-expert-parallel \
|
||||||
|
--max-model-len 8192 \
|
||||||
|
--max-num-seqs 64
|
||||||
|
```
|
||||||
|
|
||||||
|
### 3.4 Decode 机器启动脚本
|
||||||
|
|
||||||
|
#### 启动 Decode 节点
|
||||||
|
|
||||||
|
```bash
|
||||||
|
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
|
||||||
|
|
||||||
|
python -m fastdeploy.entrypoints.openai.multi_api_server \
|
||||||
|
--ports 8198,8199,8200,8201,8202,8203,8204,8205 \
|
||||||
|
--num-servers 8 \
|
||||||
|
--args --model /path/to/ERNIE-4.5-300B-A47B-Paddle \
|
||||||
|
--splitwise-role "decode" \
|
||||||
|
--cache-transfer-protocol "rdma,ipc" \
|
||||||
|
--router "<PREFILL_MACHINE_IP>:8109" \
|
||||||
|
--quantization wint4 \
|
||||||
|
--tensor-parallel-size 1 \
|
||||||
|
--data-parallel-size 8 \
|
||||||
|
--enable-expert-parallel \
|
||||||
|
--max-model-len 8192 \
|
||||||
|
--max-num-seqs 64
|
||||||
|
```
|
||||||
|
|
||||||
|
**注意**：请将 `<PREFILL_MACHINE_IP>` 替换为 Prefill 机器的实际 IP 地址。
|
||||||
|
|
||||||
|
## 四、发送测试请求
|
||||||
|
|
||||||
|
```bash
|
||||||
|
curl -X POST "http://localhost:8109/v1/chat/completions" \
|
||||||
|
-H "Content-Type: application/json" \
|
||||||
|
-d '{
|
||||||
|
"messages": [
|
||||||
|
{"role": "user", "content": "你好，请介绍一下自己。"}
|
||||||
|
],
|
||||||
|
"max_tokens": 100,
|
||||||
|
"stream": false
|
||||||
|
}'
|
||||||
|
```
|
||||||
|
|
||||||
|
## 五、常见问题 FAQ
|
||||||
|
|
||||||
|
如果您在使用过程中遇到问题，可以在 [FAQ](./FAQ.md) 中查阅解决方案。
|
||||||
@@ -9,3 +9,4 @@
|
|||||||
- [ERNIE-4.5-VL-28B-A3B-Paddle](ERNIE-4.5-VL-28B-A3B-Paddle.md)
|
- [ERNIE-4.5-VL-28B-A3B-Paddle](ERNIE-4.5-VL-28B-A3B-Paddle.md)
|
||||||
- [ERNIE-4.5-VL-424B-A47B-Paddle](ERNIE-4.5-VL-424B-A47B-Paddle.md)
|
- [ERNIE-4.5-VL-424B-A47B-Paddle](ERNIE-4.5-VL-424B-A47B-Paddle.md)
|
||||||
- [PaddleOCR-VL-0.9B.md](PaddleOCR-VL-0.9B.md)
|
- [PaddleOCR-VL-0.9B.md](PaddleOCR-VL-0.9B.md)
|
||||||
|
- [Disaggregated.md](Disaggregated.md)
|
||||||
|
|||||||
@@ -1,5 +1,7 @@
|
|||||||
[English](../../features/disaggregated.md)
|
[English](../../features/disaggregated.md)
|
||||||
|
|
||||||
|
[最佳实践](../best_practices/Disaggregated.md)
|
||||||
|
|
||||||
# 分离式部署
|
# 分离式部署
|
||||||
|
|
||||||
LLM大模型推理分为Prefill和Decode两个阶段，分别为计算密集型和访存密集型。
|
LLM大模型推理分为Prefill和Decode两个阶段，分别为计算密集型和访存密集型。
|
||||||
|
|||||||
Reference in New Issue
Block a user