Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.
Python cliconsumer-gpudpofine-tuningggufhuggingfacellmllmopslocal-ailocal-llmloralow-vrammachine-learningollamapeftpythonpytorchqlorasfttransformers
"""Ordered optional batch generation shared by the pure evaluation scorers."""
from typing import Callable, Optional
def generate_outputs(gen: Callable[[str], str], prompts: list[str]) -> list[Optional[str]]:
"""Return one output per prompt; ``None`` marks a failed generation.
A callable ``batch(prompts) -> list[str]`` is optional. If a batch raises,
scalar retries isolate failures to their items. A malformed result cannot
be aligned safely and fails the batch; non-string items fail individually.
Empty strings remain valid outputs, distinct from generation failures.
"""
if not callable(gen):
raise TypeError("gen must be callable")
if not prompts:
return []
batch = getattr(gen, "batch", None)
if callable(batch):
try:
outputs = batch(prompts)
except Exception: # noqa: BLE001 — retry scalars to isolate failed items
pass
else:
if not isinstance(outputs, list) or len(outputs) != len(prompts):
return [None] * len(prompts)
return [output if isinstance(output, str) else None for output in outputs]
outputs = []
for prompt in prompts:
try:
output = gen(prompt)
except Exception: # noqa: BLE001 — a generation error is a failed item
output = None
outputs.append(output if isinstance(output, str) else None)
return outputs