Gather a tensor across a process group and optionally unpad its padded elements. Args: x (Tensor): Input tensor to gather. g

Get the value associated with the given key. If the key does not exist, return the default value. Args: key (str): The attribute

Load and initialize a reward manager based on the configuration. Args: config: PPO trainer configuration object containing reward_mo

Compute per-token log-probabilities for the given labels. Uses a Flash-Attention–based cross-entropy (if available) for efficient backward,

Postprocess packed sequences

broadcast tensor across mp_group

broadcast tensor across mp_group

model: AutoModelForCausalLM

Apply monkey patch to the models for ulysses sequence parallel and fused kernel. In the end of this function forward function of the model i

from https://github.com/sgl-project/sglang/blob/844e2f227ab0cce6ef818a719170ce37b9eb1e1b/python/sglang/srt/utils.py#L905 Broadcast inputs from sr

Check and disable incompatible configurations for older Megatron version. Args: original_config (dict): The original model configura

Compute reward for a batch of data. Args: data: DataProto object containing the input data. reward_fn: Reward function to com

Convert Hydra configs and other special types to regular Python types.

r"""Works like shutil.copy() for file, and shutil.copytree for dir, and supports hdfs. Copy data and mode bits ("cp src dst"). Return the file's

Estimate the FLOPS based on the number of valid tokens in the current batch and the time taken. Args: batch_seqlens (Lis

Fallback implementation for older transformers versions. Returns False to disable features that require this function.

Loads the full state dict (could be only on rank 0) into the sharded model. This is done by broadcasting the parameters from rank 0 to all ot

Get FSDP wrap policy for the module. Args: module: The module to get wrap policy for config: Configuration for wrap policy

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],

Pad a DataProto to size divisible by size_divisor Args: size_divisor (int): size divisor Returns: data: (DataProto): the pad

Applies a monkey patch to the forward method of a given model class to enable Ulysses sequence parallelism input slicing.

Reduces a dictionary of metric lists by computing the mean of each list. Args: metrics: A dictionary mapping metric names to lists o

Resume GPU occupation.

Unpad the data proto with pad_size. i.e. `data[:-pad_size]`

Reorder the data on single controller such that each dp rank gets similar total tokens

broadcast tensor in tp shards across mp_group

broadcast tensor in tp shards across mp_group

broadcast tensor in tp shards across mp_group

Computes information about prompts and responses from a batch. This is an internal helper function that extracts masks and lengths for promp

Inner function that handles the core timing logic. Args: name (str): The name/identifier for this timing measurement. timing_raw

Update the input_ids, attention_mask, position_ids, and loss_mask of the request in additive manner.

Create a dataset. Arguments: data_paths: List of paths to data files. data_config: The data config. tokenizer (Tokenizer)

r""" Copy paste Qwen2VL's forward https://github.com/linkedin/Liger-Kernel/blob/main/src/liger_kernel/transformers/model/qwen2_vl.py ```

torch.nn.utils.clip_grad_norm_ cann't run on cpu parameter DTensor

Create a schedule with a learning rate that decreases following the values of the cosine function between the initial lr set in the optimizer

Build the model.

Gets the position ids for Qwen2-VL, it should be generated before sharding the sequence. The batch dim has been removed and the input_ids sho

Convert huggingface PretrainedConfig to mcore TransformerConfig. Args: hf_config: The huggingface PretrainedConfig. dtype: The dt

Load weights for mcore GPT model.

Separate original prompt from partial response that may be included in the prompt. This method handles cases where the prompt contains parti

Generate data using the provided data generation strategy. Note: This method is intended to change the dataset after each training ba

Prepare a batch for dynamic batching. Args: data (DataProto): The input data. max_token_len (int): The maximum token length

Process validation metrics into a structured format with statistical analysis. This function organizes validation metrics by data source and

Reduce timing information across all processes. This function uses distributed communication to gather and sum the timing information from al

Release GPU occupation temporarily.

Rotates half the hidden dims of the input.

Rotates half the hidden dims of the input.

Execute the main PPO training workflow. This method sets up the distributed training environment, initializes workers, datasets, and

Repeat each row of the batch data a specified number of times. Args: repeat_times (torch.tensor, list, tuple, ndarray):