r""" Instantiates a new model from a pretrained model from `transformers`. The pretrained model is loaded using the `from_pretrained`

move the batch to device Args: device (torch.device, str): torch device Returns: DataProto: the current Data

Return a tensor with the input `shape` as a view into the 1-D data starting at `start_index`.

r""" Push the pretrained model to the hub. This method is a wrapper around `transformers.PreTrainedModel.push_to_hub`. Please refer to

Pop a subset of the DataProto via `batch_keys` and `meta_info_keys` Args: batch_keys (list, optional): a list of strings indicati

Create a DataProto from a dict of tensors. This assumes that 1. All the tensor in tensors have the same dim0 2. Only dim0 is the batch

Generates model completions for the provided prompts. Args: prompts (`list[str]`): List of text prompts

r""" Return the state_dict of the pretrained model.

Select a subset of the DataProto via batch_keys and meta_info_keys Args: batch_keys (list, optional): a list of strings indicatin

Parse command-line args and config file into instances of the specified dataclass types. This method wraps [`transformers.HfArgument

Converts a sequence of ids in a string, using the tokenizer and vocabulary with options to remove special tokens and clean up tokeniz

Pads a list of tensors to the same shape along the first dimension. Args: tensors (`list[torch.Tensor]`): List of input

Log `logs` on the various objects watching training, including stored metrics. Args: logs (`dict[str, float]`):

r""" Save the pretrained model to a directory. This method is a wrapper around `transformers.PreTrainedModel.save_pretrained`. Please

Returns the autocast context manager

Returns the tokenizer used for tokenizing text inputs

Computes N-gram repetition penalty as described in Appendix C.2 of https://huggingface.co/papers/2502.03373. Reference implementation from: h

Generate a `ModelCard` from a template. Args: base_model (`str` or `None`): Base model name. model_name (`str`):

Updates running moments from batch's moments computed across ranks

Empties the cache of the available torch device. This function checks for the availability of different torch devices (XPU, MLU, NPU, CUDA) a

Split the batch among dim=0 into chunks. The meta_info is passed to each DataProto after split. Args: chunks (int): the number of

Copy src from hdfs to local if src is on hdfs or directly return src. If cache_dir is None, we will use the default cache dir of the system. Note

Compute the generalized Jensen-Shannon Divergence loss for knowledge distillation using F.kl_div. See Eq. (1) of https://huggingface.

Perform a single step of training. Args: epoch (int): The current epoch. global_step (int): The current glob

Converts a string to a sequence of ids (integer), using the tokenizer and vocabulary. Args: text (`str`, `List[str]` or

If Comet integration is enabled, return the URL of the current Comet experiment; otherwise, return `None`.

Repeat the batch data a specified number of times. Args: repeat_times (int): Number of times to repeat the data.

Run the environment on a list of queries. Args: queries (list[str]): A list of queries to run the model in the environme

A memory-efficient implementation of the common `log_softmax -> gather` operation. This function is equivalent to the following naive implem

Create a huggingface pretrained tokenizer. Args: name (str): The name of the tokenizer. correct_pad_token (bool): Whether to corr

r""" Check if the example is in a conversational format. Args: example (`dict[str, Any]`): A single data entry of a datas

Concat a list of DataProto. The batch is concatenated among dim=0. The meta_info is assumed to be identical and will use the first one.

Prepares the model for training with DeepSpeed (both for stage 2 and 3), configuring the appropriate settings based on the model and batch si

r""" Tokenize a row of the dataset. Args: features (`dict[str, str]`): Row of the dataset, should contain

Context manager to unwrap distributed or accelerated models for generation tasks. Args: model (`Union[DistributedDataParallel, DeepS

If Comet integration is enabled logs a table to the Comet experiment if it is currently running. Args: name (`str`): Tab

Append a new segment to the history. Args: text (`str`): The text of the new segment. tokens (`torch.LongTen

Creates a static reference copy of a model. Note that model will be in `.eval()` mode. Args: model (`PreTrainedModelWrapper`): The m

Computes the reward logits and the rewards for a given model and query responses. Args: model (`torch.nn.Module`): The m

Reward function that evaluates code snippets using a code execution provider. Assumes the dataset contains a `verification_info` column with test

Get ulysses sequence parallel process group.

Setup chat format by adding special tokens to the tokenizer, setting the correct format, and extending the embedding layer of the model based on

Split list `lst` into `n` evenly distributed sublists. Example: >>> chunk_list([1, 2, 3, 4, 5, 6], 2) [[1, 2, 3], [4, 5, 6]]

r""" Finds the correct formatting function based on the dataset structure. Currently supported datasets are: - `ChatML` with [{"role": str, "c

Compute mean of tensor with a masked values.

given TP,DP,PP rank to get the global rank.

Shift non-zero elements in the mask and corresponding tensors to the left. This function operates on a binary mask and any number of additio

Format reward function specifically for code responses. Args: language: Programming language supported by E2B https://e2b.dev/docs/code-i

Compute mean of tensor with a masked values.

Executes a batch of scripts in the sandbox environment. Args: scripts (list[str]): A list of code scripts to execute.

broadcast tensor from rank0 across mp_group

r""" Apply a chat template to a conversational example along with the schema for a list of functions in `tools`. For more details, see [`mayb

A func to sync embedding input with alltoall in sequence parallel gather sequence dimension and scatter head dim: e.g. seq_dim: 1, head_d

Return a tensor with the input `shape` as a view into the 1-D data starting at `start_index`.

Load a dataset or a mixture of datasets based on the configuration. Args: args (ScriptArguments): Script arguments containing dataset con

See: https://github.com/pytorch/pytorch/issues/563#issuecomment-330103591

r""" If the example is in a conversational format, apply a chat template to it. Args: example (`dict[str, list[dict[str, str]]`):

r""" Unpair a preference dataset if it is paired. Args: dataset (`Dataset` or `DatasetDict`): Preference dataset to unpai

pad a 2D tensors (e.g. responses, logprobs) in the last dim to max_seq_length. input shape: [bs, seq_length] output shape: [bs, max_seq_l