kscope 0.8.0

A user toolkit for analyzing and interfacing with Large Language Models (LLMs)

Kaleidoscope-SDK

A user toolkit for analyzing and interfacing with Large Language Models (LLMs)

Overview

kaleidoscope-sdk is a Python module used to interact with large language models hosted via the Kaleidoscope service available at: https://github.com/VectorInstitute/kaleidoscope. It provides a simple interface to launch LLMs on an HPC cluster, asking them to perform basic features like text generation, but also retrieve intermediate information from inside the model, such as log probabilities and activations. These features are exposed via a few high-level APIs, namely:

• model_instances - Shows a list of all active LLMs instantiated by the model service
• load_model - Loads an LLM via the model service
• generate - Returns an LLM text generation based on prompt input
• module_names - Returns all modules names in the LLM neural network
• get_activations - Retrieves all activations for a set of modules

Getting Started

Requires Python version >= 3.8

Install

python3 -m pip install kscope

or install from source:

pip install git+https://github.com/VectorInstitute/kaleidoscope-sdk.git

Authentication

In order to submit generation jobs, a designated Vector Institute cluster account is required. Please contact the AI Engineering Team in charge of Kaleidoscope for more information.

Sample Workflow

The following workflow shows how to load and interact with an OPT-175B model on the Vector Institute Vaughan cluster.

#!/usr/bin/env python3
import kscope
import time

# Establish a client connection to the Kaleidoscope service
# If you have not previously authenticated with the service, you will be prompted to now
client = kscope.Client(gateway_host="llm.cluster.local", gateway_port=3001)

# See which models are supported
client.models

# See which models are instantiated and available to use
client.model_instances

# Get a handle to a model. If this model is not actively running, it will get launched in the background.
# In this example we want to use the OPT-175B model
opt_model = client.load_model("opt-175b")

# If the model was not actively running, this it could take several minutes to load. Wait for it come online.
while opt_model.state != "ACTIVE":
    time.sleep(1)

# Sample text generation w/ input parameters
text_gen = opt_model.generate("What is the answer to life, the universe, and everything?", {'max_tokens': 5, 'top_k': 4, 'temperature': 0.5})
dir(text_gen) # display methods associated with generated text object
text_gen.generation['sequences'] # display only text
text_gen.generation['logprobs'] # display logprobs
text_gen.generation['tokens'] # display tokens

# Now let's retrieve some activations from the model
# First, show a list of modules in the neural network
print(opt_model.module_names)

# Setup a request for module acivations for a certain module layer
requested_activations = ['decoder.layers.0']
activations = opt_model.get_activations("What are activations?", requested_activations)
print(activations)

# Next, let's manipulate the activations in the model. First, we need to import a few more modules.
import cloudpickle
import codecs
import torch
from torch import Tensor
from typing import Callable, Dict

# Define a function to manipulate the activations
def replace_with_ones(act: Tensor) -> Tensor:
    """Replace an activation with an activation filled with ones."""
    out = torch.ones_like(act, dtype=act.dtype).cuda()
    return out

# Now send the edit request
editing_fns: Dict[str, Callable] = {}
editing_fns['decoder.layers.0'] = replace_with_ones
edited_activations = opt_model.edit_activations("Testing activation editing", editing_fns)
print(edited_activations)

Documentation

Full documentation and API reference are available at: http://kaleidoscope-sdk.readthedocs.io.

Contributing

Contributing to kaleidoscope is welcomed. See Contributing for guidelines.

License

MIT

Citation

Reference to cite when you use Kaleidoscope in a project or a research paper:

Sivaloganathan, J., Coatsworth, M., Willes, J., Choi, M., & Shen, G. (2022). Kaleidoscope. http://VectorInstitute.github.io/kaleidoscope. computer software, Vector Institute for Artificial Intelligence. Retrieved from https://github.com/VectorInstitute/kaleidoscope-sdk.git.