README.md

Scripts

This directory contains development and utility scripts for the LLM framework.

Usage

All scripts can be run with uv run:

uv run scripts/<script_name>.py [options]

Available Scripts

Script	Description
e2e_pipeline.py	End-to-end smoke test (Train → Evaluate → Inference)
train_simple_decoder.py	Train a decoder model on text files
benchmark_inference.py	Benchmark inference performance

Script Details

e2e_pipeline.py

Automated end-to-end pipeline for validating the framework:

# Quick test with defaults
uv run scripts/e2e_pipeline.py

# Custom configuration
uv run scripts/e2e_pipeline.py --epochs 5 --hidden-size 128 --num-layers 4

Options:

• --hidden-size: Model dimension (default: 64)
• --num-layers: Number of transformer layers (default: 2)
• --epochs: Training epochs (default: 3)
• --num-samples: Synthetic samples to generate (default: 200)
• --device: Device to use: 'cpu', 'cuda', or 'auto' (default: auto)

train_simple_decoder.py

Train a decoder model on your own text data:

uv run scripts/train_simple_decoder.py \
  --file-path data/train.txt \
  --val-file-path data/val.txt \
  --epochs 10 \
  --batch-size 32

benchmark_inference.py

Benchmark inference throughput and latency:

uv run scripts/benchmark_inference.py