Retrieve object from cache

Print a message in the main process.

Perform all-reduce operation on a tensor cross all ranks and return the mean.

r""" Gets the current available device.

Load outputs from cache

Check if the current process is the main process.

Log a dictionary of scalars to the logger backend.

Load pre-trained model and tokenizer from a given path.

Evaluate the model on the evaluation dataset.

Get benchmark configs from yaml file in benchmark folder Args: benchmark_name (str): benchmark name Returns:

This function is used to convert the custom configurations to dict.

Set global random seed for reproducibility.

Display single benchmark results in command line. Args: benchmark_name (str): benchmark name result (Dict[str, Dict[st

r""" Sets the device for PyTorch.

Get the dataloaders based on data_dtype.

Make API requests with caching to avoid duplicate calls. Args: messages: List of message dictionaries to send to the API mod

Return a new dataset containing only the given indices.

Return the dict used for model inference

Compute the mean of a tensor with a mask.

r""" Frees up unused memory on supported hardware accelerators (GPU, NPU, MPS, etc.). This function performs two key steps: 1. Invokes Py

Gather log probabilities of the given labels from the logits.

Save model and tokenizer in Hugging Face format.

Override save_pretrained to ensure proper saving of configuration.

e.g. `a basketball` or `an icecream`

e.g. `a basketball` or `an icecream`

Save outputs to cache

Set logger.

Train the model.

Check if two tokenizers are the same.

Set training mode for all models.

Check if the string ends with any of the substrings. Args: s (str): The string to check. substrings (list[str]): The list of subs

Evaluate the model on the evaluation dataset.

Evaluate the model on the evaluation dataset.

Prepare the DeepSpeed config for training.

Build a default conversation from a text prompt.

Initialize DeepSpeed engines.

Save transformers model and tokenizer in Hugging Face format.

Compute log probabilities of given sequences.

Save model and tokenizer in Hugging Face format.

NOTE: It may be much less efficient to `include_receptacle_info` than to not. :param object_id: :param include_receptacle_in

e.g. basketball

e.g. basketball

Resize tokenizer and embedding. Note: This is the unoptimized version that may make your embedding size not be divisible by 64.

Train the model.

Check if the response shows signs of over-refusal to legitimate technical questions

Evaluate the model on the evaluation dataset.

Perform all-reduce operation on a tensor cross all ranks and return the max.