quickstart/README.md

# Quickstart

Load the FELA LLM 1.5 code model and run a real fill in the middle completion on CPU.
Uses the self contained loader in `../modeling.py`.

## Steps

1. Install the pinned requirements (CPU PyTorch):

   ```bash
   pip install -r requirements.txt
   ```

2. Point the loader at the directory that holds the weights (this repo). The model
   ships as `model.safetensors` next to `config.json` and `tokenizer.json`:

   ```bash
   export FELA_LLM_WEIGHTS=/path/to/this/repo
   python run.py
   ```

## Load from Python

```python
from modeling import load_model
m = load_model("/path/to/this/repo")

# Fill in the middle: give the code before and after the cursor, get the middle
r = m.complete("def add(a, b):\n    ", suffix="\nresult = add(2, 3)\n")
print(r["middle"])          # e.g. " return a + b"

# Plain autocomplete: continue a single line
print(m.complete("import numpy as ")["middle"])   # e.g. " np"
```

`complete` is greedy by default, so the same prompt gives the same real output every
time. When you pass a non empty `suffix` and the tokenizer has the FIM tokens, it uses
the fill in the middle layout `<|fim_prefix|> P <|fim_suffix|> S <|fim_middle|>` and
writes the code that belongs between them; otherwise it continues the prefix.

## What to expect

This is the final fill in the middle model. Single line patterns land well (imports,
obvious returns, boilerplate); multi line blocks and novel logic are outside what it is
built for. Every completion is a genuine forward of the model, never a lookup. The
reference load path here runs on CPU with no GPU. For fast production serving the same
weights run int8 on the CPU native FELA server.