LM Buddy¶

[!IMPORTANT]

The lm-buddy repo is being archived and its functionality is being folded into Lumigator. For more on the context and decisions behind this, please read here.

LM Buddy is a collection of jobs for finetuning and evaluating open-source (large) language models. The library makes use of YAML-based configuration files as inputs to CLI commands for each job, and tracks input/output artifacts on Weights & Biases.

The package currently exposes two types of jobs:

finetuning job using HuggingFace model/training implementations and Ray Train for compute scaling, or an
evaluation job using lm-evaluation-harness with inference performed via an in-process HuggingFace model or an externally-hosted vLLM server.

Installation¶

LM Buddy is available on PyPI and can be installed as follows:

pip install lm-buddy

Minimum Python version¶

LM Buddy is intended to be used in production on a Ray cluster (see section below on Ray job submission). Currently, we are utilizing Ray clusters running Python 3.11.9. In order to avoid dependency/syntax errors when executing LM Buddy on Ray, installation of this package requires Python between [3.11, 3.12).

CLI usage¶

LM Buddy exposes a CLI with a few commands, one for each type of job. You can explore the CLI options by running lm-buddy --help.

Once LM Buddy is installed in your local Python environment, usage is as follows:

# LLM finetuning
lm_buddy finetune --config finetuning_config.yaml

# LLM evaluation
lm_buddy evaluate lm-harness --config lm_harness_config.yaml
lm_buddy evaluate prometheus --config prometheus_config.yaml

See the examples/configs folder for examples of the job configuration structure. For a full end-to-end interactive workflow for using the package, see the example notebooks.

Ray job submission¶

Although the LM Buddy CLI can be used as a standalone tool, its commands are intended to be used as the entrypoints for jobs on a Ray compute cluster. The suggested method for submitting an LM Buddy job to Ray is by using the Ray Python SDK within a local Python driver script. This requires you to specify a Ray runtime environment containing:

A working_dir for the local directory containing your job config YAML file, and
A pip dependency for your desired version of lm-buddy.

Additionally, if your job requires GPU resources on the Ray entrypoint worker (e.g., for loading large/quantized models), you should specify the entrypoint_num_gpus parameter upon submission.

An example of the submission process is as follows:

from ray.job_submission import JobSubmissionClient

# If using a remote cluster, replace 127.0.0.1 with the head node's IP address.
client = JobSubmissionClient("http://127.0.0.1:8265")

runtime_env = {
    "working_dir": "/path/to/working/directory",
    "pip": ["lm-buddy==X.X.X"]
    
}

# Assuming 'config.yaml' is present in the working directory
client.submit_job(
    entrypoint="lm_buddy finetune <job-name> --config config.yaml", 
    runtime_env=runtime_env,
    entrypoint_num_gpus=1
)

See the examples/ folder for more examples of submitting Ray jobs.

Development¶

See the contributing guide for more information on development workflows and/or building locally.