CZWin32768 / trlx

A repo for distributed training of language models with Reinforcement Learning via Human Feedback (RLHF)

Welcome to Transformer Reinforcement Learning X (trlX)

A repo for distributed training of language models with Reinforcement Learning via Human Feedback (RLHF)

Documentation

Overview

Inspired by the popular trl library, the trlX repo allows you to fine-tune Huggingface supported language models up to 20B parameters via either reinforcement learning using a provided scoring function or reward-labeled dataset. We aim to support a range of both online and offline RL algorithms including Proximal Policy Optimization (PPO), Natural Language Policy Optimization (NLPO), Actor Critic (A2C), and Implicit Q Learning (ILQL).

The library supports gpt2 and gptj with plans to include GPT-NeoX, T5 and more. PPO and ILQL algorithms are implemented. Disibtributed training has been implemented via HF Accelerate and tested up to two nodes, each with 8 gpus.

Structure

The training pipeline is broken into four pieces:

• Prompt pipeline: Handles loading of prompts/text used to prompt model for exploration in online methods
• Rollout pipeline: Handles loading and storage of reward labeled data used
• Orchestrator: Handles exploration/rollout collection of online methods. Pushes collected rollouts to the rollout pipeline.
• Model: Wraps the supplied base model (ex: gpt2) and implements the desired training method loss (ex: PPO).

Adding a task for RLHF training depends on the desired training method and pre-existing data. If we are online and have no reward labeled data this is as simple as writing a new prompt pipeline, which supplies prompts for exploration, and a new reward function to be passed into the PPOOrchestrator class.

Installation

git clone https://github.com/CarperAI/trlx.git
cd trlx
pip install -e ".[dev]"
pre-commit install # see .pre-commit-config.yaml

Example: How to add a task

In the below we implement a sentiment learning task.

Configure accelerate

accelerate config

Implement a prompt pipeline

@register_datapipeline
class PPOPipeline(BasePipeline):
    def __init__(self, tokenizer, config, prompt_dataset_path=None):
        super().__init__()

        ds = load_dataset("imdb", split="test")
        ds = ds.rename_columns({"text": "review", "label": "sentiment"})
        ds = ds.filter(lambda x: len(x["review"]) < 500, batched=False)

        self.tokens = [
            tokenizer(
                text,
                truncation=True,
                padding="max_length",
                max_length=config.train.input_size,
                return_tensors="pt",
            )["input_ids"]
            .long()
            .flatten()
            for text in ds["review"]
        ]
        self.text = [tokenizer.decode(tokens.tolist()) for tokens in self.tokens]

    def __getitem__(self, index: int) -> PromptElement:
        return PromptElement(self.text[index], self.tokens[index])

    def __len__(self) -> int:
        return len(self.text)

    def create_loader(
        self,
        batch_size: int,
        shuffle: bool,
        prep_fn: Callable = None,
        num_workers: int = 0,
    ) -> DataLoader:
        # TODO(dahoas): Decide how to support varying sizes of prompts without having to tokenize on fly
        def collate_fn(elems: Iterable[PromptElement]) -> PromptElement:
            return PromptBatch(
                [elem.text for elem in elems],
                torch.stack(
                    [elem.tokens for elem in elems]
                ),  # Assumes token tensors all same size
            )

        return DataLoader(
            self, batch_size, shuffle, collate_fn=collate_fn, num_workers=num_workers
        )

Launch training

from typing import List

import torch
from transformers import pipeline

import wandb
from trlx.data.configs import TRLConfig
from trlx.model.accelerate_ppo_model import AcceleratePPOModel
from trlx.orchestrator.ppo_orchestrator import PPOOrchestrator
from trlx.pipeline.ppo_pipeline import PPOPipeline
from trlx.utils.loading import get_model, get_orchestrator, get_pipeline

if __name__ == "__main__":
    cfg = TRLConfig.load_yaml("configs/ppo_config.yml")

    sentiment_pipe = pipeline(
        "sentiment-analysis", "lvwerra/distilbert-imdb", device=-1
    )

    def reward_fn(samples: List[str]):
        sent_kwargs = {
            "return_all_scores": True,
            "function_to_apply": None,
            "batch_size": cfg.method.chunk_size,
        }
        pipe_outputs = sentiment_pipe(samples, **sent_kwargs)
        scores = torch.tensor([output[1]["score"] for output in pipe_outputs])
        return scores

    model: AcceleratePPOModel = get_model(cfg.model.model_type)(cfg)
    if model.accelerator.is_main_process:
        wandb.watch(model.model)

    pipeline: PPOPipeline = get_pipeline(cfg.train.pipeline)(model.tokenizer, cfg)
    orch: PPOOrchestrator = get_orchestrator(cfg.train.orchestrator)(
        model, pipeline, reward_fn=reward_fn, chunk_size=cfg.method.chunk_size
    )
    orch.make_experience(cfg.method.num_rollouts)
    model.learn()

    print("DONE!")

And run accelerate launch my_script.py

References

Proximal Policy Optimisation

The PPO implementation largely follows the structure introduced in the paper "Fine-Tuning Language Models from Human Preferences" by D. Ziegler et al. [paper, code].

Language models

The language models utilize the transformers library by 🤗 Hugging Face.