PatrickStar分布式深度學(xué)習(xí)訓(xùn)練工具

PatrickStar 是一款騰訊開(kāi)發(fā)的分布式深度學(xué)習(xí)訓(xùn)練工具，它的設(shè)計(jì)目標(biāo)是支持以 GPT、Bert 為代表的超大預(yù)訓(xùn)練模型訓(xùn)練。

用法

PatrickStar 基于 PyTorch，這使得遷移 pytorch 項(xiàng)目變得容易。以下是 PatrickStar 的示例：

from patrickstar.runtime import initialize_engine

config = {
    "optimizer": {
        "type": "Adam",
        "params": {
            "lr": 0.001,
            "betas": (0.9, 0.999),
            "eps": 1e-6,
            "weight_decay": 0,
            "use_hybrid_adam": True,
        },
    },
    "fp16": {  # loss scaler params
        "enabled": True,
        "loss_scale": 0,
        "initial_scale_power": 2 ** 3,
        "loss_scale_window": 1000,
        "hysteresis": 2,
        "min_loss_scale": 1,
    },
    "default_chunk_size": 64 * 1024 * 1024,
    "release_after_init": True,
    "use_cpu_embedding": False,
}

def model_func():
    # MyModel is a derived class for torch.nn.Module
    return MyModel(...)

model, optimizer = initialize_engine(model_func=model_func, local_rank=0, config=config)

...

for data in dataloader:
    optimizer.zero_grad()

    loss = model(data)
    model.backward(loss)
    optimizer.step()

使用與 DeepSpeed 配置 JSON 相同的config格式，主要包括優(yōu)化器、損失縮放器和一些 PatrickStar 特定配置的參數(shù)。

引用我們

@article{fang2021patrickstar,
  title={PatrickStar: Parallel Training of Pre-trained Models via a Chunk-based Memory Management},
  author={Fang, Jiarui and Yu, Yang and Zhu, Zilin and Li, Shenggui and You, Yang and Zhou, Jie},
  journal={arXiv preprint arXiv:2108.05818},
  year={2021}
}