app.py

import torch
from datasets import load_dataset
from peft import LoraConfig, get_peft_model
from transformers import AutoModelForCausalLM, AutoTokenizer
from trl import GRPOConfig, GRPOTrainer

import wandb

wandb.login(key="2fa14e3cc1db3ff6c0d83973c3b7b9d152a73b70")

dataset = load_dataset("mlabonne/smoltldr")
print(dataset)

import os
os.environ["FLASH_ATTENTION_FORCE_DISABLED"] = "1"

model_id = "HuggingFaceTB/SmolLM-135M-Instruct"
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(model_id)

# Load LoRA
lora_config = LoraConfig(
    task_type="CAUSAL_LM",
    r=16,
    lora_alpha=32,
    target_modules="all-linear",
)
model = get_peft_model(model, lora_config)
print(model.print_trainable_parameters())

# Reward function
ideal_length = 50


def reward_len(completions, **kwargs):
    return [-abs(ideal_length - len(completion)) for completion in completions]

training_args = GRPOConfig(
    output_dir="GRPO",
    learning_rate=2e-5,
    per_device_train_batch_size=8,
    gradient_accumulation_steps=2,
    max_prompt_length=512,
    max_completion_length=96,
    num_generations=8,
    num_train_epochs=1,
    report_to=["wandb"],
    remove_unused_columns=False,
    logging_steps=1,
    bf16=False,
    fp16=True,                     # если есть GPU
    optim="adamw_torch_fused",     # НЕ "adamw_8bit"
)

trainer = GRPOTrainer(
    model=model,
    reward_funcs=[reward_len],
    args=training_args,
    train_dataset=dataset["train"],
)

# Train model
wandb.init(project="GRPO")
trainer.train()