Instructions to use openbmb/MiniCPM-2B-128k with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use openbmb/MiniCPM-2B-128k with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="openbmb/MiniCPM-2B-128k", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("openbmb/MiniCPM-2B-128k", trust_remote_code=True, dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use openbmb/MiniCPM-2B-128k with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "openbmb/MiniCPM-2B-128k"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "openbmb/MiniCPM-2B-128k",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/openbmb/MiniCPM-2B-128k

SGLang

How to use openbmb/MiniCPM-2B-128k with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "openbmb/MiniCPM-2B-128k" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "openbmb/MiniCPM-2B-128k",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "openbmb/MiniCPM-2B-128k" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "openbmb/MiniCPM-2B-128k",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use openbmb/MiniCPM-2B-128k with Docker Model Runner:
```
docker model run hf.co/openbmb/MiniCPM-2B-128k
```

Slow inference speed and high VRAM using huggingface transformers

by Starlento - opened Apr 18, 2024

Discussion

Starlento

Apr 18, 2024

•

edited Apr 18, 2024

When using transformers to inference this model (the example code in README), the VRAM will keep increasing to around 14GB and the output is just a short response.
It seems it is inferencing by default 4000 tokens without noticing the stop_token <|im_end|>.

And I found a duplicated code here:
https://huggingface.co/openbmb/MiniCPM-2B-128k/blob/cc0a7847338aa03a64d575e0d7ff54d4bd008ff7/modeling_minicpm.py#L1326

I do not know how to input the stop_token in your self.generate() function, could you kindly fix it?

neoz

OpenBMB org May 24, 2024

Hi, sorry for the late reply. We recommend using vLLM for generation, where you can specify a stop token. Please check our blog for usage examples.

Starlento

May 24, 2024

Hi, sorry for the late reply. We recommend using vLLM for generation, where you can specify a stop token. Please check our blog for usage examples.

Sorry as well that I forgot this issue. In fact for transformers, I can set the stop token by change the tokenizer's eos_token.
Thank you very much for your reply, it is never late. By the way, will VLLM implementation have some VRAM saving compared to transformers?

Starlento changed discussion status to closed May 24, 2024

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment