Instructions to use dphn/dolphin-2.9.2-mixtral-8x22b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use dphn/dolphin-2.9.2-mixtral-8x22b with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="dphn/dolphin-2.9.2-mixtral-8x22b") messages = [ {"role": "user", "content": "Who are you?"}, ] pipe(messages)# pip install -U transformers accelerate # Load model directly from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("dphn/dolphin-2.9.2-mixtral-8x22b") model = AutoModelForCausalLM.from_pretrained("dphn/dolphin-2.9.2-mixtral-8x22b", device_map="auto") messages = [ {"role": "user", "content": "Who are you?"}, ] inputs = tokenizer.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=256) print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:])) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use dphn/dolphin-2.9.2-mixtral-8x22b with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "dphn/dolphin-2.9.2-mixtral-8x22b" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "dphn/dolphin-2.9.2-mixtral-8x22b", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/dphn/dolphin-2.9.2-mixtral-8x22b
- SGLang
How to use dphn/dolphin-2.9.2-mixtral-8x22b 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 "dphn/dolphin-2.9.2-mixtral-8x22b" \ --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": "dphn/dolphin-2.9.2-mixtral-8x22b", "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 "dphn/dolphin-2.9.2-mixtral-8x22b" \ --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": "dphn/dolphin-2.9.2-mixtral-8x22b", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use dphn/dolphin-2.9.2-mixtral-8x22b with Docker Model Runner:
docker model run hf.co/dphn/dolphin-2.9.2-mixtral-8x22b
Congrats!
This fine tune is a work of art. It's super smart and super obedient to the system message, way better than 2.9.1.
I think we are getting closer and closer to close source with open source thanks to your great work! :)
I'd say we already beat them in a lot of use cases.
1 week with 8xH100's is crazy too, thats a lot of compute for a finetune. This seems like the real deal certainly!
How much does that cost? I wouldn't mind a WizardLM2-8x22b finetune like this
to rent 8xH100's for a week is roughly around 5 grand USD give or take average pricing
Possibly less, I guess it depends but the quotes im looking at are around there
1 week with 8xH100's is crazy too, thats a lot of compute for a finetune. This seems like the real deal certainly!
We have some new techniques for FFT we'll share soon - but in total this model took 3 days 22 hours to train.
oops, I think I forgot to update the model card there
I had no idea it was so expensive. I thought maybe a few hundred bucks...
Thanks for releasing these finetunes ehartford
The H100 is probably within the top 3 most powerful gpu's in the world right now. The H200 is king IIRC and I know AMD has something out to compete. Thus why i think its probably within the top 3 or 4.
I had no idea it was so expensive. I thought maybe a few hundred bucks...
Thanks for releasing these finetunes ehartford
We have a compute sponsor for most of these models, so while yes it’s very expensive - it’s not coming out of our pocket.
This fine tune is a work of art. It's super smart and super obedient to the system message, way better than 2.9.1.
I think we are getting closer and closer to close source with open source thanks to your great work! :)
I'd say we already beat them in a lot of use cases.
How smart actually?
How smart actually?
I am wondering if it would top the newest qwen model that just came out
Qwen2 is not yet released.
I really enjoy Dolphin 2.9.2 Mixtral 8x22b. For now it's my favorite Dolphin that's ever been released.
But there will absolutely be a Dolphin trained on Qwen2.
Ah, I thought I saw that it had been released on Reddit but I must have read it wrong. I tried quill which is supposedly an early version and it was decent.
Qwen2 is not yet released.
I really enjoy Dolphin 2.9.2 Mixtral 8x22b. For now it's my favorite Dolphin that's ever been released.
But there will absolutely be a Dolphin trained on Qwen2.
Will it be follow systems prompt good like this finetune?
And Qwen It's quite bad to often insert Chinese into answers, I hope Qwen 2 will fix it.
I hope this model be hosted somewhere so i can try it.
And Qwen It's quite bad to often insert Chinese into answers
Yes I have also witnessed this issue. It seems to plague the qwen models as I have tried other chinese made models and they do not do this.