apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-12-24 13:28:13 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

更新文档 (#3975)

Browse Source
This commit is contained in:
yangjianfengo1
2025-09-08 16:53:37 +08:00
committed by GitHub
parent 571ddc677b
commit 9ead10e1bc
13 changed files with 429 additions and 130 deletions
Download Patch File Download Diff File Expand all files Collapse all files

82
docs/supported_models.md
View File

@@ -2,9 +2,9 @@
FastDeploy currently supports the following models, which can be downloaded automatically during FastDeploy deployment.Specify the ``model`` parameter as the model name in the table below to automatically download model weights (all supports resumable downloads). The following three download sources are supported:
- 1. Search for corresponding Paddle-version ERNIE models on [AIStudio/PaddlePaddle](https://aistudio.baidu.com/modelsoverview), e.g., `ERNIE-4.5-0.3B-Paddle`
- 2. Download Paddle-version ERNIE models from [HuggingFace/baidu/models](https://huggingface.co/baidu/models), e.g., `baidu/ERNIE-4.5-0.3B-Paddle`
- 3. Search for corresponding Paddle-version ERNIE models on [ModelScope/PaddlePaddle](https://www.modelscope.cn/models?name=PaddlePaddle&page=1&tabKey=task), e.g., `ERNIE-4.5-0.3B-Paddle`
- [AIStudio](https://aistudio.baidu.com/modelsoverview)
- [ModelScope](https://www.modelscope.cn/models)
- [HuggingFace](https://huggingface.co/models)
When using automatic download, the default download source is AIStudio. Users can modify the default download source by setting the ``FD_MODEL_SOURCE`` environment variable, which can be set to “AISTUDIO”, ‘MODELSCOPE’ or “HUGGINGFACE”. The default download path is ``~/`` (i.e., the user's home directory). Users can modify the default download path by setting the ``FD_MODEL_CACHE`` environment variable, e.g.:
@@ -13,25 +13,61 @@ export FD_MODEL_SOURCE=AISTUDIO # "AISTUDIO", "MODELSCOPE" or "HUGGINGFACE"
export FD_MODEL_CACHE=/ssd1/download_models
```
| Model Name | Context Length | Quantization | Minimum Deployment Resources | Notes |
| :------------------------------------------ | :------------- | :----------- | :--------------------------- | :----------------------------------------------------------------------------------------- |
| baidu/ERNIE-4.5-VL-424B-A47B-Paddle | 32K/128K | WINT4 | 4*80G GPU VRAM/1T RAM | Chunked Prefill required for 128K |
| baidu/ERNIE-4.5-VL-424B-A47B-Paddle | 32K/128K | WINT8 | 8*80G GPU VRAM/1T RAM | Chunked Prefill required for 128K |
| baidu/ERNIE-4.5-300B-A47B-Paddle | 32K/128K | WINT4 | 4*64G GPU VRAM/600G RAM | Chunked Prefill required for 128K |
| baidu/ERNIE-4.5-300B-A47B-Paddle | 32K/128K | WINT8 | 8*64G GPU VRAM/600G RAM | Chunked Prefill required for 128K |
| baidu/ERNIE-4.5-300B-A47B-2Bits-Paddle | 32K/128K | WINT2 | 1*141G GPU VRAM/600G RAM | Chunked Prefill required for 128K |
| baidu/ERNIE-4.5-300B-A47B-W4A8C8-TP4-Paddle | 32K/128K | W4A8C8 | 4*64G GPU VRAM/160G RAM | Fixed 4-GPU setup, Chunked Prefill recommended |
| baidu/ERNIE-4.5-300B-A47B-FP8-Paddle | 32K/128K | FP8 | 8*64G GPU VRAM/600G RAM | Chunked Prefill recommended, only supports PD Disaggragated Deployment with EP parallelism |
| baidu/ERNIE-4.5-300B-A47B-Base-Paddle | 32K/128K | WINT4 | 4*64G GPU VRAM/600G RAM | Chunked Prefill recommended |
| baidu/ERNIE-4.5-300B-A47B-Base-Paddle | 32K/128K | WINT8 | 8*64G GPU VRAM/600G RAM | Chunked Prefill recommended |
| baidu/ERNIE-4.5-VL-28B-A3B-Paddle | 32K | WINT4 | 1*24G GPU VRAM/128G RAM | Chunked Prefill required |
| baidu/ERNIE-4.5-VL-28B-A3B-Paddle | 128K | WINT4 | 1*48G GPU VRAM/128G RAM | Chunked Prefill required |
| baidu/ERNIE-4.5-VL-28B-A3B-Paddle | 32K/128K | WINT8 | 1*48G GPU VRAM/128G RAM | Chunked Prefill required |
| baidu/ERNIE-4.5-21B-A3B-Paddle | 32K/128K | WINT4 | 1*24G GPU VRAM/128G RAM | Chunked Prefill required for 128K |
| baidu/ERNIE-4.5-21B-A3B-Paddle | 32K/128K | WINT8 | 1*48G GPU VRAM/128G RAM | Chunked Prefill required for 128K |
| baidu/ERNIE-4.5-21B-A3B-Base-Paddle | 32K/128K | WINT4 | 1*24G GPU VRAM/128G RAM | Chunked Prefill required for 128K |
| baidu/ERNIE-4.5-21B-A3B-Base-Paddle | 32K/128K | WINT8 | 1*48G GPU VRAM/128G RAM | Chunked Prefill required for 128K |
| baidu/ERNIE-4.5-0.3B-Paddle | 32K/128K | BF16 | 1*6G/12G GPU VRAM/2G RAM | |
| baidu/ERNIE-4.5-0.3B-Base-Paddle | 32K/128K | BF16 | 1*6G/12G GPU VRAM/2G RAM | |
> ⭐ **Note**: Models marked with an asterisk can directly use **HuggingFace Torch weights** and support **FP8/WINT8/WINT4** as well as **BF16**. When running inference, you need to enable **`--load_choices "default_v1"`**.
> Example launch Command using baidu/ERNIE-4.5-21B-A3B-PT:
```
python -m fastdeploy.entrypoints.openai.api_server \
--model baidu/ERNIE-4.5-0.3B-PT \
--port 8180 \
--metrics-port 8181 \
--engine-worker-queue-port 8182 \
--max-model-len 32768 \
--max-num-seqs 32 \
--load_choices "default_v1"
```
## Large Language Models
These models accept text input.
|Models|DataType|Example HF Model|
|-|-|-|
|⭐ERNIE|BF16\WINT4\WINT8\W4A8C8\WINT2\FP8|baidu/ERNIE-4.5-VL-424B-A47B-Paddle;<br>baidu/ERNIE-4.5-300B-A47B-Paddle<br>&emsp;[quick start](./get_started/ernie-4.5.md) &emsp; [best practice](./best_practices/ERNIE-4.5-300B-A47B-Paddle.md);<br>baidu/ERNIE-4.5-300B-A47B-2Bits-Paddle;<br>baidu/ERNIE-4.5-300B-A47B-W4A8C8-TP4-Paddle;<br>baidu/ERNIE-4.5-300B-A47B-FP8-Paddle;<br>baidu/ERNIE-4.5-300B-A47B-Base-Paddle;<br>[baidu/ERNIE-4.5-21B-A3B-Paddle](./best_practices/ERNIE-4.5-21B-A3B-Paddle.md);<br>baidu/ERNIE-4.5-21B-A3B-Base-Paddle;<br>baidu/ERNIE-4.5-0.3B-Paddle<br>&emsp;[quick start](./get_started/quick_start.md) &emsp; [best practice](./best_practices/ERNIE-4.5-0.3B-Paddle.md);<br>baidu/ERNIE-4.5-0.3B-Base-Paddle, etc.|
|⭐QWEN3-MOE|BF16/WINT4/WINT8/FP8|Qwen/Qwen3-235B-A22B;<br>Qwen/Qwen3-30B-A3B, etc.|
|⭐QWEN3|BF16/WINT8/FP8|Qwen/qwen3-32B;<br>Qwen/qwen3-14B;<br>Qwen/qwen3-8B;<br>Qwen/qwen3-4B;<br>Qwen/qwen3-1.7B;<br>[Qwen/qwen3-0.6B](./get_started/quick_start_qwen.md), etc.|
|⭐QWEN2.5|BF16/WINT8/FP8|Qwen/qwen2.5-72B;<br>Qwen/qwen2.5-32B;<br>Qwen/qwen2.5-14B;<br>Qwen/qwen2.5-7B;<br>Qwen/qwen2.5-3B;<br>Qwen/qwen2.5-1.5B;<br>Qwen/qwen2.5-0.5B, etc.|
|⭐QWEN2|BF16/WINT8/FP8|Qwen/Qwen/qwen2-72B;<br>Qwen/Qwen/qwen2-7B;<br>Qwen/qwen2-1.5B;<br>Qwen/qwen2-0.5B;<br>Qwen/QwQ-32, etc.|
|DEEPSEEK|BF16/WINT4|unsloth/DeepSeek-V3.1-BF16;<br>unsloth/DeepSeek-V3-0324-BF16;<br>unsloth/DeepSeek-R1-BF16, etc.|
## Multimodal Language Models
These models accept multi-modal inputs (e.g., images and text).
|Models|DataType|Example HF Model|
|-|-|-|
| ERNIE-VL |BF16/WINT4/WINT8| baidu/ERNIE-4.5-VL-424B-A47B-Paddle<br>&emsp;[quick start](./get_started/ernie-4.5-vl.md) &emsp; [best practice](./best_practices/ERNIE-4.5-VL-424B-A47B-Paddle.md) ;<br>baidu/ERNIE-4.5-VL-28B-A3B-Paddle<br>&emsp;[quick start](./get_started/quick_start_vl.md) &emsp; [best practice](./best_practices/ERNIE-4.5-VL-28B-A3B-Paddle.md) ;|
| QWEN-VL |BF16/WINT4/FP8| Qwen/Qwen2.5-VL-72B-Instruct;<br>Qwen/Qwen2.5-VL-32B-Instruct;<br>Qwen/Qwen2.5-VL-7B-Instruct;<br>Qwen/Qwen2.5-VL-3B-Instruct|
## Minimum Resource Deployment Instruction
There is no universal formula for minimum deployment resources; it depends on both context length and quantization method. We recommend estimating the required GPU memory using the following formula:
```
Required GPU Memory = Number of Parameters × Quantization Byte factor
```
> (The factor list is provided below.)
And the final number of GPUs depends on:
```
Number of GPUs = Total Memory Requirement ÷ Memory per GPU
```
| Quantization Method | Bytes per Parameter factor |
| :--- | :--- |
|BF16 |2 |
|FP8 |1 |
|WINT8 |1 |
|WINT4 |0.5 |
|W4A8C8 |0.5 |
More models are being supported. You can submit requests for new model support via [Github Issues](https://github.com/PaddlePaddle/FastDeploy/issues).