Extract tool calls from the responses. Args: responses_ids (List[int]): The ids of the responses. Returns: T

The training loop of PPO. The driver process only need to call the compute functions of the worker group through RPC to const

Format tool response for gpt-oss model. Args: tool_response: Tool response string tool_call_name: Name of the tool that was called

We assume: - The model takes input: (input_ids, attention_mask, position_ids). No rmpad for the input - The communication sha

Create a worker group from existing detached workers. Args: name_prefix: Prefix for worker names worker_names: Names

Parse configuration from dictionary

Gather the label from data. The value in label should be [0, vocab_size) Args: data: (..., vocab_size) label (torch.IntTensor) :

Generate synthetic training data from a YAML configuration file.

Generate text using the SGLang server asynchronously.

Generate sequence with token-in-token-out.

Extract general patterns from task instructions and optional demonstration examples. Args: demo_examples: Optional list

Generates a response for the current turn of interaction. Returns a tuple containing: - should_terminate_sequence (bool): Tru

Generate responses for input prompts. This method handles three output formats: 1. Single prompt, n=1: Returns string

Extract only adapter parameters from a model. Args: model: PyTorch model (possibly wrapped in DDP/Float16Module) Returns: Di

Return the local layer range/count for the current process and the full assignment table. Args: tf_config: Configuration object used by c

Get the theoretical FLOPS (Floating Point Operations Per Second) capacity of the current device. Args: unit (str): The unit to return the

Calculate the pipeline sharding configuration for Megatron-LM. Args: layer_num: Total number of layers in the model. pp_size: Num

Factory method to create appropriate executor based on config. Modality structure: - text.local -> LocalTaskExecutor

Get the full state dict from an FSDP model. Args: model (torch.nn.Module): The FSDP model to get state dict from offload_to_

Dynamically import and return the interaction class.

Generate a unique local cache path for an HDFS resource. Creates a MD5-hashed subdirectory in cache_dir to avoid name conflicts, then returns

Get logs from position. Returns (new_logs, new_position).

Get the forward function for given model architecture.

Lazily create event loop and thread for MCP tools.

Get detailed GPU memory information

Get client for various providers. Supported providers: openai, anthropic, deepseek, qwen, ollama Environment variables:

Get the number of gpus in this cluster.

Group tensors into buckets based on a specified size in megabytes. Args: iterable: An iterator of tuples containing tensor names and

Get a generator that yields per-tensor parameters and optional peft config. Returns: Generator[tuple[str, torch.Tensor]]

A filter function to return the PPO Ray runtime environment. To avoid repeat of some environment variables that are already set.

Note that here we only asks execute_all and execute_rank_zero to be implemented Leave the choice of how these two functions handle argument '

Returns the recorded topk indices.

end of sentence token can be int or list: 1 or [1, 2] e.g. response_id = torch.tensor([[20, 10, 34, 1, 0, 0, 0],

Get the reward loop manager class with a given name. Args: name: `(str)` The name of the reward loop manager. Returns:

Get the reward manager class with a given name. Args: name: `(str)` The name of the reward manager. Returns: `(t

collect rng state across data parallel ranks

Calculate the 3D rope index based on image and video's temporal, height and width in LLM. Explanation: Each embedding sequence cont

Get rope scaling args.

Determine the appropriate sharding strategy based on the number of dimensions of the device mesh. Args: device_mesh (torch.distribut

Get trajectory info. Args: step (int): global steps in the trainer. index (list): form datastore extra_info.index column.

Get the transformer layer specification. https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/core/models/gpt/gpt_layer_specs.py