scripts/smoke_test.py

# /// script
# requires-python = ">=3.10"
# dependencies = [
#     "trl>=1.8.0",
#     "transformers>=4.45.0",
#     "accelerate>=0.34.0",
#     "scikit-learn",
#     "datasets",
# ]
# ///
"""CPU smoke test: verifies XRPOTrainer runs end-to-end on a tiny model/dataset
before spending GPU budget. Not a real accuracy test."""

import sys
import os

sys.path.insert(0, os.path.dirname(__file__))

from datasets import Dataset
from transformers import AutoTokenizer
import transformers.core_model_loading as _cml
_cml.GLOBAL_WORKERS = 1  # macOS ThreadPoolExecutor segfault workaround for local smoke test
from trl import GRPOConfig

from xrpo_lib import XRPOTrainer, build_prompt

MODEL = "trl-internal-testing/tiny-Qwen3ForCausalLM"

tokenizer = AutoTokenizer.from_pretrained(MODEL)
if not hasattr(tokenizer, "apply_chat_template") or tokenizer.chat_template is None:
    tokenizer.chat_template = (
        "{% for message in messages %}{{ message['role'] }}: {{ message['content'] }}\n{% endfor %}"
        "{% if add_generation_prompt %}assistant:{% endif %}"
    )

questions = [
    ("What is 2+3?", "5"),
    ("What is 10-4?", "6"),
    ("What is 6+1?", "7"),
    ("What is 9-2?", "7"),
]

rows = []
for i, (q, a) in enumerate(questions * 4):
    try:
        prompt = build_prompt(tokenizer, q)
    except Exception:
        prompt = f"user: {q}\nassistant:"
    rows.append({"prompt": prompt, "question": q, "answer": a, "qid": q})

dataset = Dataset.from_list(rows)


def dummy_reward(completions, **kwargs):
    return [0.0 for _ in completions]


config = GRPOConfig(
    output_dir="/tmp/xrpo_smoke",
    per_device_train_batch_size=4,
    num_generations=4,
    gradient_accumulation_steps=1,
    max_steps=2,
    max_completion_length=16,
    report_to=[],
    logging_steps=1,
    beta=0.001,
    push_to_hub=False,
    save_strategy="no",
    dataloader_num_workers=0,
    disable_tqdm=False,
)

print("Building trainer...", flush=True)

trainer = XRPOTrainer(
    model=MODEL,
    reward_funcs=dummy_reward,
    train_dataset=dataset,
    args=config,
    use_novelty=True,
    use_icl=True,
    tokenizer_for_icl=tokenizer,
)

print("Trainer built, starting train()...", flush=True)
trainer.train()
print("SMOKE TEST OK")
print("icl bank size:", len(trainer.solved_bank))
print("question_stats sample:", dict(list(trainer.question_stats.items())[:2]))