llm-replicate 0.3.1

LLM plugin for models hosted on Replicate

llm-replicate

LLM plugin for models hosted on Replicate

Installation

First, install the LLM command-line utility.

Now install this plugin in the same environment as LLM.

llm install llm-replicate

Configuration

You will need an API key from Replicate. You can obtain one here.

You can set that as an environment variable called REPLICATE_API_TOKEN, or add it to the llm set of saved keys using:

llm keys set replicate

Enter key: <paste key here>

To fetch and save details of the default collection of language models hosted on Replicate, run this:

llm replicate fetch-models

To add specific models that aren't listed in that collection, use the llm replicate add command.

For the Llama 2 model from a16z-infra/llama13b-v2-chat run this:

llm replicate add a16z-infra/llama13b-v2-chat \
  --chat --alias llama2

The --chat flag indicates that this is a chat model, which means it will be able to work with -c continue mode.

Here's the 70b version of that model:

llm replicate add \
  replicate/llama70b-v2-chat \
  --chat --alias llama70b

Usage

To run a prompt against a model, pass its name or an alias to llm -m:

llm -m llama2 "Ten great names for a pet pelican"

Sure, here are ten great names for a pet pelican:

1. Pelty
2. Peanut
3. Puddles
4. Nibbles
5. Fuzzy
6. Gizmo
7. Hank
8. Luna
9. Scooter
10. Splishy

I hope these suggestions help you find the perfect name for your pet pelican! Do you have any other questions?

Chat models can support continuing conversations, for example:

llm -c "Five more and make them more nautical"

Ahoy matey! Here be five more nautical-themed names for yer pet pelican:

1. Captain Hook
2. Anchoryn
3. Seadog
4. Plunder
5. Pointe Pelican

I hope these suggestions help ye find the perfect name for yer feathered friend! Do ye have any other questions, matey?

Run llm models list to see the full list of models:

llm models list

You should see something like this:

Replicate: replicate-flan-t5-xl
Replicate: replicate-llama-7b
Replicate: replicate-gpt-j-6b
Replicate: replicate-dolly-v2-12b
Replicate: replicate-oasst-sft-1-pythia-12b
Replicate: replicate-stability-ai-stablelm-tuned-alpha-7b
Replicate: replicate-vicuna-13b
Replicate: replicate-replit-code-v1-3b
Replicate: replicate-replit-replit-code-v1-3b
Replicate: replicate-joehoover-falcon-40b-instruct (aliases: falcon)
Replicate (chat): replicate-a16z-infra-llama13b-v2-chat (aliases: llama2)

Then run a prompt through a specific model like this:

llm -m replicate-vicuna-13b "Five great names for a pet llama"

Registering extra models

To register additional models that are not included in the default Language models collection, find their ID on Replicate and use the llm replicate add command.

For example, to add the joehoover/falcon-40b-instruct model, run this:

llm replicate add joehoover/falcon-40b-instruct \
  --alias falcon

This adds the model with the alias falcon - you can have 0 or more aliases for a model.

Now you can run it like this:

llm -m replicate-joehoover-falcon-40b-instruct \
  "Three reasons to get a pet falcon"

Or using the alias like this:

llm -m falcon "Three reasons to get a pet falcon"

You can edit the list of models you have registered using the default $EDITOR like this:

llm replicate edit-models

If you register a model using the --chat option that model will be treated slightly differently. Prompts sent to the model will be formatted like this:

User: user input here
Assistant:

If you use -c conversation mode the prompt will include previous messages in the conversation, like this:

User: Ten great names for a pet pelican
Assistant: Sure, here are ten great names for a pet pelican:

1. Pelty
2. Peanut
3. Puddles
4. Nibbles
5. Fuzzy
6. Gizmo
7. Hank
8. Luna
9. Scooter
10. Splishy

I hope these suggestions help you find the perfect name for your pet pelican! Do you have any other questions?
User: Five more and make them more nautical
Assistant:

Fetching all Replicate predictions

Replicate logs all predictions made against models. You can fetch all of these predictions using the llm replicate fetch-predictions command:

llm replicate fetch-predictions

This will create or populate a table in your LLM logs.db database called replicate_predictions.

The data in this table will cover ALL Replicate models, not just language models that have been queried using this tool.

Running llm replicate fetch-predictions multiple times will only fetch predictions that have been created since the last time the command was run.

To browse the resulting data in Datasette, run this:

datasette "$(llm logs path)"

The schema for that table will look like this:

CREATE TABLE [replicate_predictions] (
   [id] TEXT PRIMARY KEY,
   [_model_guess] TEXT,
   [completed_at] TEXT,
   [created_at] TEXT,
   [error] TEXT,
   [input] TEXT,
   [logs] TEXT,
   [metrics] TEXT,
   [output] TEXT,
   [started_at] TEXT,
   [status] TEXT,
   [urls] TEXT,
   [version] TEXT,
   [webhook_completed] TEXT
)

This schema may change if the Replicate API adds new fields in the future.

Development

To set up this plugin locally, first checkout the code. Then create a new virtual environment:

cd llm-replicate
python3 -m venv venv
source venv/bin/activate

Now install the dependencies and test dependencies:

pip install -e '.[test]'

To run the tests:

pytest