DeepSeek/DeepSeek-VL2
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepSeek-VL2 synced 2025-06-26 18:25:56 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

update

Browse Source 
Gradio Demo Example, Incremental Prefilling and VLMEvalKit Support
This commit is contained in:
StevenLiuWen
2024-12-26 22:37:57 +08:00
parent 8bde1c1ae1
commit faf18023f2
38 changed files with 1369 additions and 168 deletions
Download Patch File Download Diff File Expand all files Collapse all files

225
README.md
View File

@@ -69,6 +69,7 @@ Zhiyu Wu*, Xiaokang Chen*, Zizheng Pan*, Xingchao Liu*, Wen Liu**, Damai Dai, Hu
![](./images/vl2_teaser.jpeg)
## 2. Release
✅ <b>2024-12-25</b>: Gradio Demo Example, Incremental Prefilling and VLMEvalKit Support.
✅ <b>2024-12-13</b>: DeepSeek-VL2 family released, including <code>DeepSeek-VL2-tiny</code>, <code>DeepSeek-VL2-small</code>, <code>DeepSeek-VL2</code>.
## 3. Model Download
@@ -96,7 +97,9 @@ On the basis of `Python >= 3.8` environment, install the necessary dependencies
pip install -e .
```
### Simple Inference Example
### Simple Inference Example with One Image
**Note: You may need 80GB GPU memory to run this script with deepseek-vl2-small and even larger for deepseek-vl2.**
```python
import torch
@@ -107,7 +110,7 @@ from deepseek_vl2.utils.io import load_pil_images
# specify the path to the model
model_path = "deepseek-ai/deepseek-vl2-small"
model_path = "deepseek-ai/deepseek-vl2-tiny"
vl_chat_processor: DeepseekVLV2Processor = DeepseekVLV2Processor.from_pretrained(model_path)
tokenizer = vl_chat_processor.tokenizer
@@ -119,23 +122,78 @@ conversation = [
{
"role": "<|User|>",
"content": "<image>\n<|ref|>The giraffe at the back.<|/ref|>.",
"images": ["./images/visual_grounding.jpeg"],
"images": ["./images/visual_grounding_1.jpeg"],
},
{"role": "<|Assistant|>", "content": ""},
]
# load images and prepare for inputs
pil_images = load_pil_images(conversation)
prepare_inputs = vl_chat_processor(
conversations=conversation,
images=pil_images,
force_batchify=True,
system_prompt=""
).to(vl_gpt.device)
# run image encoder to get the image embeddings
inputs_embeds = vl_gpt.prepare_inputs_embeds(**prepare_inputs)
# run the model to get the response
outputs = vl_gpt.language.generate(
inputs_embeds=inputs_embeds,
attention_mask=prepare_inputs.attention_mask,
pad_token_id=tokenizer.eos_token_id,
bos_token_id=tokenizer.bos_token_id,
eos_token_id=tokenizer.eos_token_id,
max_new_tokens=512,
do_sample=False,
use_cache=True
)
answer = tokenizer.decode(outputs[0].cpu().tolist(), skip_special_tokens=False)
print(f"{prepare_inputs['sft_format'][0]}", answer)
```
And the output is something like:
```
<|User|>: <image>
<|ref|>The giraffe at the back.<|/ref|>.
<|Assistant|>: <|ref|>The giraffe at the back.<|/ref|><|det|>[[580, 270, 999, 900]]<|/det|><｜end▁of▁sentence｜>
```
### Simple Inference Example with Multiple Images
**Note: You may need 80GB GPU memory to run this script with deepseek-vl2-small and even larger for deepseek-vl2.**
```python
import torch
from transformers import AutoModelForCausalLM
from deepseek_vl2.models import DeepseekVLV2Processor, DeepseekVLV2ForCausalLM
from deepseek_vl2.utils.io import load_pil_images
# specify the path to the model
model_path = "deepseek-ai/deepseek-vl2-tiny"
vl_chat_processor: DeepseekVLV2Processor = DeepseekVLV2Processor.from_pretrained(model_path)
tokenizer = vl_chat_processor.tokenizer
vl_gpt: DeepseekVLV2ForCausalLM = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True)
vl_gpt = vl_gpt.to(torch.bfloat16).cuda().eval()
# multiple images/interleaved image-text
conversation_multi_images = [
conversation = [
{
"role": "<|User|>",
"content": "This is image_1: <image>\n"
"This is image_2: <image>\n"
"This is image_3: <image>\n If I am a vegetarian, what can I cook with these ingredients?",
"This is image_3: <image>\n Can you tell me what are in the images?",
"images": [
"images/multi_image_1.png",
"images/multi_image_2.jpg",
"images/multi_image_3.jpg",
"images/multi_image_1.jpeg",
"images/multi_image_2.jpeg",
"images/multi_image_3.jpeg",
],
},
{"role": "<|Assistant|>", "content": ""}
@@ -169,12 +227,151 @@ answer = tokenizer.decode(outputs[0].cpu().tolist(), skip_special_tokens=False)
print(f"{prepare_inputs['sft_format'][0]}", answer)
```
### Gradio Demo (TODO)
And the output is something like:
```
<|User|>: This is image_1: <image>
This is image_2: <image>
This is image_3: <image>
Can you tell me what are in the images?
<|Assistant|>: The images show three different types of vegetables. Image_1 features carrots, which are orange with green tops. Image_2 displays corn cobs, which are yellow with green husks. Image_3 contains raw pork ribs, which are pinkish-red with some marbling.<｜end▁of▁sentence｜>
```
### Simple Inference Example with Incremental Prefilling
**Note: We use incremental prefilling to inference within 40GB GPU using deepseek-vl2-small.**
```python
import torch
from transformers import AutoModelForCausalLM
from deepseek_vl2.models import DeepseekVLV2Processor, DeepseekVLV2ForCausalLM
from deepseek_vl2.utils.io import load_pil_images
### Demo
This figure present some examples of DeepSeek-VL2.
![](./images/github_demo.png)
# specify the path to the model
model_path = "deepseek-ai/deepseek-vl2-small"
vl_chat_processor: DeepseekVLV2Processor = DeepseekVLV2Processor.from_pretrained(model_path)
tokenizer = vl_chat_processor.tokenizer
vl_gpt: DeepseekVLV2ForCausalLM = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True)
vl_gpt = vl_gpt.to(torch.bfloat16).cuda().eval()
# multiple images/interleaved image-text
conversation = [
{
"role": "<|User|>",
"content": "This is image_1: <image>\n"
"This is image_2: <image>\n"
"This is image_3: <image>\n Can you tell me what are in the images?",
"images": [
"images/multi_image_1.jpeg",
"images/multi_image_2.jpeg",
"images/multi_image_3.jpeg",
],
},
{"role": "<|Assistant|>", "content": ""}
]
# load images and prepare for inputs
pil_images = load_pil_images(conversation)
prepare_inputs = vl_chat_processor(
conversations=conversation,
images=pil_images,
force_batchify=True,
system_prompt=""
).to(vl_gpt.device)
with torch.no_grad():
# run image encoder to get the image embeddings
inputs_embeds = vl_gpt.prepare_inputs_embeds(**prepare_inputs)
# incremental_prefilling when using 40G GPU for vl2-small
inputs_embeds, past_key_values = vl_gpt.incremental_prefilling(
input_ids=prepare_inputs.input_ids,
images=prepare_inputs.images,
images_seq_mask=prepare_inputs.images_seq_mask,
images_spatial_crop=prepare_inputs.images_spatial_crop,
attention_mask=prepare_inputs.attention_mask,
chunk_size=512 # prefilling size
)
# run the model to get the response
outputs = vl_gpt.generate(
inputs_embeds=inputs_embeds,
input_ids=prepare_inputs.input_ids,
images=prepare_inputs.images,
images_seq_mask=prepare_inputs.images_seq_mask,
images_spatial_crop=prepare_inputs.images_spatial_crop,
attention_mask=prepare_inputs.attention_mask,
past_key_values=past_key_values,
pad_token_id=tokenizer.eos_token_id,
bos_token_id=tokenizer.bos_token_id,
eos_token_id=tokenizer.eos_token_id,
max_new_tokens=512,
do_sample=False,
use_cache=True,
)
answer = tokenizer.decode(outputs[0][len(prepare_inputs.input_ids[0]):].cpu().tolist(), skip_special_tokens=False)
print(f"{prepare_inputs['sft_format'][0]}", answer)
```
And the output is something like:
```
<|User|>: This is image_1: <image>
This is image_2: <image>
This is image_3: <image>
Can you tell me what are in the images?
<|Assistant|>: The first image contains carrots. The second image contains corn. The third image contains meat.<｜end▁of▁sentence｜>
```
### Full Inference Example
```shell
# without incremental prefilling
CUDA_VISIBLE_DEVICES=0 python inference.py --model_patn "deepseek-ai/deepseek-vl2"
# with incremental prefilling, when using 40G GPU for vl2-small
CUDA_VISIBLE_DEVICES=0 python inference.py --model_patn "deepseek-ai/deepseek-vl2-small" --chunck_size 512
```
### Gradio Demo
* Install the necessary dependencies:
```shell
pip install -e .[gradio]
```
* then run the following command:
```shell
# vl2-tiny, 3.37B-MoE in total, activated 1B, can be run on a single GPU < 40GB
CUDA_VISIBLE_DEVICES=2 python web_demo.py \
--model_name "deepseek-ai/deepseek-vl2-tiny" \
--port 37914
# vl2-small, 16.1B-MoE in total, activated 2.4B
# If run on A100 40GB GPU, you need to set the `--chunk_size 512` for incremental prefilling for saving memory and it might be slow.
# If run on > 40GB GPU, you can ignore the `--chunk_size 512` for faster response.
CUDA_VISIBLE_DEVICES=2 python web_demo.py \
--model_name "deepseek-ai/deepseek-vl2-small" \
--port 37914 \
--chunk_size 512
# # vl27.5-MoE in total, activated 4.2B
CUDA_VISIBLE_DEVICES=2 python web_demo.py \
--model_name "deepseek-ai/deepseek-vl2" \
--port 37914
```
* **Important**: This is a basic and native demo implementation without any deployment optimizations, which may result in slower performance. For production environments, consider using optimized deployment solutions, such as vllm, sglang, lmdeploy, etc. These optimizations will help achieve faster response times and better cost efficiency.
## 5. License
@@ -184,13 +381,13 @@ This code repository is licensed under [MIT License](./LICENSE-CODE). The use of
```
@misc{wu2024deepseekvl2mixtureofexpertsvisionlanguagemodels,
title={DeepSeek-VL2: Mixture-of-Experts Vision-Language Models for Advanced Multimodal Understanding},
title={DeepSeek-VL2: Mixture-of-Experts Vision-Language Models for Advanced Multimodal Understanding},
author={Zhiyu Wu and Xiaokang Chen and Zizheng Pan and Xingchao Liu and Wen Liu and Damai Dai and Huazuo Gao and Yiyang Ma and Chengyue Wu and Bingxuan Wang and Zhenda Xie and Yu Wu and Kai Hu and Jiawei Wang and Yaofeng Sun and Yukun Li and Yishi Piao and Kang Guan and Aixin Liu and Xin Xie and Yuxiang You and Kai Dong and Xingkai Yu and Haowei Zhang and Liang Zhao and Yisong Wang and Chong Ruan},
year={2024},
eprint={2412.10302},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2412.10302},
url={https://arxiv.org/abs/2412.10302},
}
```