# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# /// script
# dependencies = [
#     "trl",
#     "peft",
#     "trackio",
#     "kernels",
# ]
# ///

import argparse
import importlib
import os
import sys
from dataclasses import dataclass, field

from trl import ScriptArguments


@dataclass
class GRPOScriptArguments(ScriptArguments):
    """
    Script arguments for the GRPO training script.

    Args:
        reward_model_name_or_path (`str`, *optional*):
            Reward model id of a pretrained model hosted inside a model repo on huggingface.co or local path to a
            directory containing model weights saved using [`~transformers.PreTrainedModel.save_pretrained`].
        reward_funcs (`list[str]`, *optional*):
            Reward functions to use. Supported values are:
                - `"accuracy_reward"`
                - `"reasoning_accuracy_reward"`
                - `"think_format_reward"`
                - `"get_soft_overlong_punishment"` (used value are `max_completion_len=1280`, `soft_punish_cache=256`)
                - any dotted import path " (e.g., `'my_lib.rewards.custom_reward'`).
    """

    reward_model_name_or_path: str | None = field(
        default=None,
        metadata={
            "help": "Reward model id of a pretrained model hosted inside a model repo on huggingface.co or "
            "local path to a directory containing model weights saved using `PreTrainedModel.save_pretrained`."
        },
    )
    reward_funcs: list[str] | None = field(
        default=None,
        metadata={
            "help": "Reward functions to use. Supported values are: `accuracy_reward`,  `reasoning_accuracy_reward`, `think_format_reward`, "
            "`get_soft_overlong_punishment` (used values are `max_completion_len=1280`, `soft_punish_cache=256`), or "
            "any dotted import path (e.g., `'my_lib.rewards.custom_reward'`)."
        },
    )


def main(script_args, training_args, model_args, dataset_args):
    import torch
    from accelerate import logging
    from datasets import load_dataset

    from trl import GRPOTrainer, get_dataset, get_kbit_device_map, get_peft_config, get_quantization_config
    from trl.rewards import (
        accuracy_reward,
        get_soft_overlong_punishment,
        reasoning_accuracy_reward,
        think_format_reward,
    )

    logger = logging.get_logger(__name__)

    reward_funcs_registry = {
        "accuracy_reward": accuracy_reward,
        "reasoning_accuracy_reward": reasoning_accuracy_reward,
        "think_format_reward": think_format_reward,
        "get_soft_overlong_punishment": get_soft_overlong_punishment(max_completion_len=1280, soft_punish_cache=256),
    }

    # Get the reward models and functions
    reward_funcs = []
    if script_args.reward_model_name_or_path:
        reward_funcs.append(script_args.reward_model_name_or_path)

    if script_args.reward_funcs:
        for func_name in script_args.reward_funcs:
            if func_name in reward_funcs_registry:
                reward_funcs.append(reward_funcs_registry[func_name])
            elif "." in func_name:
                module_path, func_name = func_name.rsplit(".", 1)
                sys.path.insert(0, os.getcwd())
                module = importlib.import_module(module_path)
                reward_func = getattr(module, func_name)
                reward_funcs.append(reward_func)
            else:
                raise ValueError(
                    f"Could not load reward function '{func_name}'. Expected one of "
                    f"{list(reward_funcs_registry.keys())} or a valid import path."
                )
    dtype = model_args.dtype if model_args.dtype in ["auto", None] else getattr(torch, model_args.dtype)

    model_kwargs = dict(
        revision=model_args.model_revision,
        attn_implementation=model_args.attn_implementation,
        dtype=dtype,
    )
    quantization_config = get_quantization_config(model_args)

    if quantization_config is not None:
        # Passing None would not be treated the same as omitting the argument, so we include it only when valid.
        model_kwargs["device_map"] = get_kbit_device_map()
        model_kwargs["quantization_config"] = quantization_config

    training_args.model_init_kwargs = model_kwargs

    # Load the dataset
    if dataset_args.datasets and script_args.dataset_name:
        logger.warning(
            "Both `datasets` and `dataset_name` are provided. The `datasets` argument will be used to load the "
            "dataset and `dataset_name` will be ignored."
        )
        dataset = get_dataset(dataset_args)
    elif dataset_args.datasets and not script_args.dataset_name:
        dataset = get_dataset(dataset_args)
    elif not dataset_args.datasets and script_args.dataset_name:
        dataset = load_dataset(
            script_args.dataset_name, name=script_args.dataset_config, streaming=script_args.dataset_streaming
        )
    else:
        raise ValueError("Either `datasets` or `dataset_name` must be provided.")

    # Initialize the GRPO trainer
    trainer = GRPOTrainer(
        model=model_args.model_name_or_path,
        reward_funcs=reward_funcs,
        args=training_args,
        train_dataset=dataset[script_args.dataset_train_split],
        eval_dataset=dataset[script_args.dataset_test_split] if training_args.eval_strategy != "no" else None,
        peft_config=get_peft_config(model_args),
    )

    # Train the model
    trainer.train()

    # Log training complete
    trainer.accelerator.print("✅ Training completed.")

    # Save and push to Hub
    trainer.save_model(training_args.output_dir)
    trainer.accelerator.print(f"💾 Model saved to {training_args.output_dir}.")

    if training_args.push_to_hub:
        trainer.push_to_hub(dataset_name=script_args.dataset_name)
        trainer.accelerator.print(f"🤗 Model pushed to the Hub in https://huggingface.co/{trainer.hub_model_id}.")


def make_parser(subparsers: argparse._SubParsersAction | None = None, prog: str | None = None):
    from trl import DatasetMixtureConfig, GRPOConfig, ModelConfig, TrlParser

    dataclass_types = (GRPOScriptArguments, GRPOConfig, ModelConfig, DatasetMixtureConfig)
    if subparsers is not None:
        parser = subparsers.add_parser("grpo", help="Run the GRPO training script", dataclass_types=dataclass_types)
    else:
        parser = TrlParser(dataclass_types, prog=prog)
    return parser


if __name__ == "__main__":
    parser = make_parser()
    script_args, training_args, model_args, dataset_args = parser.parse_args_and_config(fail_with_unknown_args=False)
    main(script_args, training_args, model_args, dataset_args)