Awesome
llm-manager
A utility written in Python 3 for querying and managing keys and values in ~/.config/llm-manager/llm.conf
and /etc/llm.conf
.
The idea is that if ie. an editor needs to use the best available LLM model for code completion, it can use this utility to get the user-configured models.
$ llm-manager code-completion
deepseek-coder-v2:236b
This way, the user can install and configure the best Ollama model that their laptop and/or Ollama server (defined with OLLAMA_HOST
) can support.
Example use
Set the default text-generation
model to gemma2:2b
:
$ llm-manager set text-generation gemma2:2b
Get the current text-generation
model
$ llm-manager get text-generation
gamma2:2b
A shortcut:
$ llm-manager text-generation
gemma2:2b
Changing the text-generation
model to llama3.2
and the 3b
tag:
$ llm-manager set text-generation llama3.2:3b
Get the current text-generation
model:
$ llm-manager text-generation
llama3.2:3b
Default values
The default values for the current version of llm-manager
and /etc/llm.conf
are:
Task | Model |
---|---|
chat | llama3.2:3b |
code | deepseek-coder:1.3b |
code-completion | deepseek-coder:1.3b |
test | tinyllama:1b |
text-generation | gemma2:2b |
tool-use | llama3.2:3b |
translation | mixtral:8x7b |
vision | llava:7b |
The default configuration may change over time as better models become available.
For now, relatively small models are chosen, so that more people can use them, even without a GPU.
Here is the default configuration file, llm.conf
:
# For chatting
chat=llama3.2:3b
# For analyzing or generating code
code=deepseek-coder-v2:latest
# For code completion / tab autocompletion
code-completion=deepseek-coder:1.3b
# A small model, for quick tests
test=tinyllama:1b
# Text generation
text-generation=gemma2:2b
# Tool use and function calling
tool-use=llama3.2:3b
# For translating text (not single words, though)
translation=mixtral:8x7b
# Vision and image detection
vision=llava:7b
Related projects
- The usermodel package, for Go.
General info
- Version: 1.2.1
- License: BSD-3
- Author: Alexander F. Rødseth <xyproto@archlinux.org>