Dose any game company consider integrating DeepSeek R1 into gaming environments?

[kkk]

I wish you guys capture the chance,,,

[Glorfunk]

i ran deepseek locally on ollama, tbh its not really at the lwevel to be used or this type o ga,me yet. Someone needs to train it for this type of thing first

[agrimes]

There is some hope for the deepseek distilled Qwen but the issue is that it burns through its token alotment doing its thinking and doesn't produce output, the max_tokens limit must be removed so it can take the time it needs to come up with a good response.

[AI Roguelite: AI-generated RPG]

In the latest beta there is better max_tokens override support for OpenAI/Oobabooga API in the config files under "model-params" (details in the README), and HTML tags are automatically removed from responses

[AI Roguelite: AI-generated RPG]

Addendum: It just occurred to me the text surrounded by <think> isn't actually removed yet. I'll add the functionality in a later update soon

[agrimes]

LM studio has been updated and now separates out the <think> content to a different segment of the reply packet. The issue of context length remains.

[AI Roguelite: AI-generated RPG]

Thanks. The configurable context length should be working in the beta as detailed in my other comment

[agrimes]

You want me to add something to some file in model-params? I'm looking at it but I need an example or detailed guide. =\

[AI Roguelite: AI-generated RPG]

Originally posted by agrimes:

You want me to add something to some file in model-params? I'm looking at it but I need an example or detailed guide. =\

Go to model-params, openai-common-params.txt, and try replacing the json with:

{
"max_tokens": 2000
}

(or whatever the desired number may be)

Be aware changes to the default folder could get overwritten in updates, so best practice is to create a mod and use a modded folder