Update README.md (#1)
Browse files- Update README.md (2b3c45757d4c7686d115fda501bddd5383e5e116)
Co-authored-by: Suen.ZK <[email protected]>
README.md
CHANGED
@@ -32,6 +32,9 @@ library_name: transformers
|
|
32 |
|
33 |
# ERNIE-4.5-VL-424B-A47B
|
34 |
|
|
|
|
|
|
|
35 |
## ERNIE 4.5 Highlights
|
36 |
|
37 |
The advanced capabilities of the ERNIE 4.5 models, particularly the MoE-based A47B and A3B series, are underpinned by several key technical innovations:
|
@@ -61,60 +64,6 @@ ERNIE-4.5-VL-424B-A47B is a multimodal MoE Chat model based on ERNIE-4.5-VL-424B
|
|
61 |
|
62 |
## Quickstart
|
63 |
|
64 |
-
### FastDeploy Inference
|
65 |
-
|
66 |
-
Quickly deploy services using FastDeploy as shown below. For more detailed usage, refer to the [FastDeploy GitHub Repository](https://github.com/PaddlePaddle/FastDeploy).
|
67 |
-
|
68 |
-
**Note**: 80GB x 8 GPU resources are required. The `--quantization` parameter supports specifying `wint4` or `wint8` for deployment with 4-bit or 8-bit quantization, respectively.
|
69 |
-
|
70 |
-
```bash
|
71 |
-
python -m fastdeploy.entrypoints.openai.api_server \
|
72 |
-
--model baidu/ERNIE-4.5-VL-424B-A47B-Paddle \
|
73 |
-
--port 8180 \
|
74 |
-
--metrics-port 8181 \
|
75 |
-
--engine-worker-queue-port 8182 \
|
76 |
-
--tensor-parallel-size 8 \
|
77 |
-
--quantization wint4 \
|
78 |
-
--max-model-len 32768 \
|
79 |
-
--enable-mm \
|
80 |
-
--reasoning-parser ernie-45-vl \
|
81 |
-
--max-num-seqs 32
|
82 |
-
```
|
83 |
-
|
84 |
-
The ERNIE-4.5-VL model supports enabling or disabling thinking mode through request parameters.
|
85 |
-
|
86 |
-
#### Enable Thinking Mode
|
87 |
-
|
88 |
-
```bash
|
89 |
-
curl -X POST "http://0.0.0.0:8180/v1/chat/completions" \
|
90 |
-
-H "Content-Type: application/json" \
|
91 |
-
-d '{
|
92 |
-
"messages": [
|
93 |
-
{"role": "user", "content": [
|
94 |
-
{"type": "image_url", "image_url": {"url": "https://paddlenlp.bj.bcebos.com/datasets/paddlemix/demo_images/example2.jpg"}},
|
95 |
-
{"type": "text", "text": "Descript this image"}
|
96 |
-
]}
|
97 |
-
],
|
98 |
-
"metadata": {"enable_thinking": true}
|
99 |
-
}'
|
100 |
-
```
|
101 |
-
|
102 |
-
#### Disable Thinking Mode
|
103 |
-
|
104 |
-
```bash
|
105 |
-
curl -X POST "http://0.0.0.0:8180/v1/chat/completions" \
|
106 |
-
-H "Content-Type: application/json" \
|
107 |
-
-d '{
|
108 |
-
"messages": [
|
109 |
-
{"role": "user", "content": [
|
110 |
-
{"type": "image_url", "image_url": {"url": "https://paddlenlp.bj.bcebos.com/datasets/paddlemix/demo_images/example2.jpg"}},
|
111 |
-
{"type": "text", "text": "Descript this image"}
|
112 |
-
]}
|
113 |
-
],
|
114 |
-
"metadata": {"enable_thinking": false}
|
115 |
-
}'
|
116 |
-
```
|
117 |
-
|
118 |
### vLLM inference
|
119 |
|
120 |
We are working with the community to fully support ERNIE4.5 models, stay tuned.
|
|
|
32 |
|
33 |
# ERNIE-4.5-VL-424B-A47B
|
34 |
|
35 |
+
> [!NOTE]
|
36 |
+
> Note: "**-Paddle**" models use [PaddlePaddle](https://github.com/PaddlePaddle/Paddle) weights, while "**-PT**" models use Transformer-style PyTorch weights.
|
37 |
+
|
38 |
## ERNIE 4.5 Highlights
|
39 |
|
40 |
The advanced capabilities of the ERNIE 4.5 models, particularly the MoE-based A47B and A3B series, are underpinned by several key technical innovations:
|
|
|
64 |
|
65 |
## Quickstart
|
66 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
67 |
### vLLM inference
|
68 |
|
69 |
We are working with the community to fully support ERNIE4.5 models, stay tuned.
|