数据接口
最后更新: 05/19/2025 (API docstrings 自动生成)。
DataProto 是数据交换的接口。
verl.DataProto 类包含两个关键成员：
batch: 一个 tensordict.TensorDict 对象，用于存放实际数据
meta_info: 一个 Dict，包含额外的元信息
TensorDict
DataProto.batch 构建在 tensordict 之上，这是一个 PyTorch 生态系统中的项目。 TensorDict 是一个类似字典的容器，用于存放张量（tensors）。要实例化一个 TensorDict，你必须指定键值对以及批次大小（batch size）。
>>> import torch
>>> from tensordict import TensorDict
>>> tensordict = TensorDict({"zeros": torch.zeros(2, 3, 4), "ones": torch.ones(2, 3, 5)}, batch_size=[2,])
>>> tensordict["twos"] = 2 * torch.ones(2, 5, 6)
>>> zeros = tensordict["zeros"]
>>> tensordict
TensorDict(
fields={
ones: Tensor(shape=torch.Size([2, 3, 5]), device=cpu, dtype=torch.float32, is_shared=False),
twos: Tensor(shape=torch.Size([2, 5, 6]), device=cpu, dtype=torch.float32, is_shared=False),
zeros: Tensor(shape=torch.Size([2, 3, 4]), device=cpu, dtype=torch.float32, is_shared=False)},
batch_size=torch.Size([2]),
device=None,
is_shared=False)
也可以沿着 batch_size 对 tensordict 进行索引。TensorDict 的内容也可以被集体操作。
>>> tensordict[..., :1]
TensorDict(
fields={
ones: Tensor(shape=torch.Size([1, 3, 5]), device=cpu, dtype=torch.float32, is_shared=False),
twos: Tensor(shape=torch.Size([1, 5, 6]), device=cpu, dtype=torch.float32, is_shared=False),
zeros: Tensor(shape=torch.Size([1, 3, 4]), device=cpu, dtype=torch.float32, is_shared=False)},
batch_size=torch.Size([1]),
device=None,
is_shared=False)
>>> tensordict = tensordict.to("cuda:0")
>>> tensordict = tensordict.reshape(6)
有关 tensordict.TensorDict 用法的更多信息，请参阅官方 tensordict 文档。
核心 API
- class verl.DataProto(batch: ~tensordict._td.TensorDict = None, non_tensor_batch: dict = <factory>, meta_info: dict = <factory>)[source]
A DataProto is a data structure that aims to provide a standard protocol for data exchange between functions. It contains a batch (TensorDict) and a meta_info (Dict). The batch is a TensorDict https://pytorch.org/tensordict/. TensorDict allows you to manipulate a dictionary of Tensors like a single Tensor. Ideally, the tensors with the same batch size should be put inside batch.
- static concat(data: list[DataProto]) DataProto[source]
Concat a list of DataProto. The batch is concatenated among dim=0. The meta_info is merged, with special handling for metrics from different workers.
- make_iterator(mini_batch_size, epochs, seed=None, dataloader_kwargs=None)[source]
Make an iterator from the DataProto. This is built upon that TensorDict can be used as a normal Pytorch dataset. See https://pytorch.org/tensordict/tutorials/data_fashion for more details.
- Parameters:
mini_batch_size (int) – mini-batch size when iterating the dataset. We require that
batch.batch_size[0] % mini_batch_size == 0.epochs (int) – number of epochs when iterating the dataset.
dataloader_kwargs (Any) – internally, it returns a DataLoader over the batch. The dataloader_kwargs is the kwargs passed to the DataLoader.
- Returns:
- an iterator that yields a mini-batch data at a time. The total number of iteration
steps is
self.batch.batch_size * epochs // mini_batch_size
- Return type:
Iterator
- select(batch_keys=None, non_tensor_batch_keys=None, meta_info_keys=None, deepcopy=False) DataProto[source]
Select a subset of the DataProto via batch_keys and meta_info_keys
- Parameters:
batch_keys (list, optional) – a list of strings indicating the keys in batch to select
meta_info_keys (list, optional) – a list of keys indicating the meta info to select
- Returns:
the DataProto with the selected batch_keys and meta_info_keys
- Return type:
- to(device) DataProto[source]
move the batch to device
- Parameters:
device (torch.device, str) – torch device
- Returns:
the current DataProto
- Return type: