Split a RayResourcePool into multiple SubRayResourcePool. resouce_pool can also be a SubRayResourcePool (have been splited) for multiple-time

Create a tool instance. Args: instance_id: The instance id of the tool. Returns: The instance id of the tool

Processing the raw synthetic data into unified formats. Input: raw synthetic dataset Output: structured dataset, eac

Synchronize on group commit forward.

Parsing configurations in self.config to the information required in synthesis, like task instruction, domain, parameters for model, ...

Tensor pop.

Tensor push.

Creates a decorator for bridging BatchMeta and DataProto. This decorator automatically handles conversions between `BatchMeta` and `DataProto

Union two tensordicts.

Union two dict. Will throw an error if there is an item not the same object with the same key. Args: dict1: dict2: Returns:

Update Prometheus configuration file with server addresses and reload on first node. server_addresses: vllm or sglang server addresses

torch.compile might be replaced by mindspeed on NPU, this contextmanager can revert torch.compile temporarily.

Validate the training configuration.

Validate an OmegaConf DictConfig. Args: config (DictConfig): The OmegaConf DictConfig to validate. use_reference_policy (bool): i

Probe all datasets and score them. Returns list of scored datasets sorted by score (highest first).

Phase 1: Probe all datasets and score them. Returns list of scored datasets sorted by score (highest first).

Verify if the solution is correct. Args: solution_str: The solution string to verify answer: The ground truth answer stri

verify the batch and save as ``acc`` tensor

verify the copy of src to dest by comparing their sizes and file structures. return: bool: True if the copy is verified, False other

Wake up all rollout replica instances.

Create and return a Ray actor with the configured options. Args: placement_group: Ray placement group for scheduling

Instantiate the stored class with the stored arguments.

We will expand this function gradually based on the available datasets

We will expand this function gradually based on the available datasets

We will expand this function gradually based on the available datasets

Make the manager callable like traditional reward managers. This method provides compatibility with the existing reward manager interface