The third iteration of my machine learning discord bot that uses the avernus API server which can be obtained here: https://github.com/Meatfucker/avernus
It has a variety of slash commands and an llm powered chatbot. See below for detailed descriptions of functionality.
- Create a python venv or conda environment
- Enter the environment.
- Install the requirements file via pip.
- Edit the example configs in configs and rename them to not have the example extension.
- Run metatron.py
configs/avernus.json:
| Key | Value | Description |
|---|---|---|
| ip | "127.0.0.1 | The ip address or hostname of the avernus API server |
| port | 6969 | The port of the avernus API server |
| llm_model | "Goekdeniz-Guelmez/Josiefied-Qwen2.5-7B-Instruct-abliterated-v2" | The Huggingface model repo to the model to use for the chat LLM |
| sdxl_model | "misri/zavychromaxl_v100" | The Huggingface model repo for the SDXL model to use |
| mtg_llm_model | "cognitivecomputations/Llama-3-8B-Instruct-abliterated-v2" | This is the Huggingface repo for the LLM to use for card titles and flavor text |
configs/discord.json
| Key | Value | Description |
|---|---|---|
| token | "yourtokenhere" | This is your discord bot token generated using the discord dev portal. |
| max_user_queue | 3 | The maximum amount of items any particular user can have queued. |
| max_user_history_message | 20 | This is the maximum amount of history to store per user. |
| mtg_gen_three_pack_send_link | false | Whether to send a hardcoded link with the three pack. You probably want this off and I plan on making it configurable in the future. Currently used for integration into my own website. |
configs/twitch.json
| Key | Value | Description |
|---|---|---|
| twitch_enabled | false | Whether to start the twitch client for twitch integration |
| token | "yourtwitchtokenhere | bot user twitch token |
| client_id | "yourclientid | bot user client id |
| client_secret | "yourclientsecret | bot user client secret |
| channel_name | "channelname" | The name of the twitch channel |
| channel_id | "channelid" | The twitch channel id of the channel |
| channel_token | "channeltoken" | The channel token that was created by the channel owner |
| refresh_token | "refreshtoken" | The channel refresh token that was created along the other channel token |
| reward_name | "Your Reward Name" | The name of the channel point redemption reward to watch for |
| token_expires_at | 1 | This is the unix timestamp the current token expires at. You generally do not need to set this as itll be managed by the twitch client if enabled |
| card_reward_channel | 435234532452345 | The discord channel id of the channel to post the twitch redeemed cards in |
configs/channels/somediscordchannelid.json
| Key | Value | Description |
|---|---|---|
| lora_name | "somelora.safetensors" | This is the name of the lora located on the avernus API server to use in this channel when generating magic cards |
| lora_prompt | "some prompt to add" | This prompt gets added to the mtg card gens that use the lora. |
Simply tag the bot with its name to interact with it. It keeps a per user history of the users conversation with it but cannot otherwise see chat.
The LLM model which is used can be configured in configs/avernus.json. It takes huggingface repo names. Note that the first time a model is used (such as the first run) there will be a large delay while the avernus server downloads the model.
There is also a /clear_chat_history command to empty your users chat history
This generates an image with the sdxl model specified in configs/avernus.json
- prompt: Is your prompt
- negative_prompt: The generated image will move away from this. Useful to avoid unwanted things in gens.
- width: The width in pixels of the image to generate. Defaults to 1024 Be careful of setting this too high as itll require more video memory and over a certain point cause deformed generations.
- height: Same as above but tall style
- batch_size: The number of images to generate at once. Combined with image resolution this can have a direct and large effect on video memory usage. If you are having lots of avernus Out Of Memory issues, try decreasing batch size and/or resolution
- lora_name: this is the filename(including safetensors extesion) of a lora located in appropriate avernus server lora directory.
- enhance_promt: This will send the user prompt to the default chat llm to be "improved"
This generates an image using the flux.dev model, which is higher quality than sdxl but slower and takes more memory while running.
- prompt: Is your prompt
- width: The width in pixels of the image to generate. Defaults to 1024 Be careful of setting this too high as itll require more video memory and over a certain point cause deformed generations.
- height: Same as above but tall style
- batch_size: The number of images to generate at once. Combined with image resolution this can have a direct and large effect on video memory usage. If you are having lots of avernus Out Of Memory issues, try decreasing batch size and/or resolution
-
- enhance_promt: This will send the user prompt to the default chat llm to be "improved"
This generates a satire magic card based on the users prompt using a combination of the chosen sdxl model, as well as the llm specified at "mtg_llm_model" in configs/avernus.json.
It will include the discord server icon as a set icon, a random artist for the artist name, and the user name and server name for the copyright.
THESE ARE SATIRE CARDS AND NOT MEANT TO BE USED COMMERCIALLY.
Same as above but generates three at a time, which is also what is triggered if twitch integration is enabled and the appropriate channel reward triggered.
