-
-
Notifications
You must be signed in to change notification settings - Fork 4.6k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[Bug]: Mistral Large Instruct 2407 tool calling leakage #8301
Comments
I originally built and tested tool calling with the chat template provided in the repo and Adding Is there a way to force the use of |
For what it's worth, I have found that Mistral's tool calling behavior is also VERY sensitive to system prompts. Generally, I have to give it very explicit instructions about what the tools are, when it should/shouldn't use them, I have to tell it that it's an AI agent that can call tools OR generate a text response; etc. You can see this in |
I'll try and see if I can get a token comparison between the prompt after the chat template is applied for autotokenizer and mistral_commons on mistral-large.
It sounds like we need to compare AnyTokenizer to HF's autotokenizer. |
So The |
Looks like #7739 is the source of this change. Can you try using |
@patrickvonplaten could you possibly help out here? It would be nice if tool calling in vLLM worked easily with the mistral tokenizer too |
Thanks for flagging - I'll look into this! |
PR to enable function calling for "mistral" formatted models is here: #8515 (should actually even work for Pixtral!) |
Your current environment
When using vllm 0.6.0, the mistral tool call parser does not work as expected for Mistral Large 2407 https://huggingface.co/mistralai/Mistral-Large-Instruct-2407 @K-Mistele
🐛 Describe the bug
It used to work fine when using Autotokenizer to instantiate the tokenizer, but not with MistralTokenizer from mistral_commons.
Basically, when you run the model and give it tools, the model thinks it is the tool.
Aka, if I give Mistral Large 2407 a tool for looking up movie information on IMDB, the model responds to "Who are you" with "I am a movie database lookup bot" instead of "I am an AI trained by Mistral AI"
Before submitting a new issue...
The text was updated successfully, but these errors were encountered: