_summary_ Use Megatron dist_checkpointing to load the model state dicts from the checkpoint directory. Args: model_ckpt_p

Log rollout data to disk. Args: batch (DataProto): The batch containing rollout data reward_extra_infos_dict (dict): A

Make causal mask used for bi-directional self-attention.

Make causal mask used for bi-directional self-attention.

Make causal mask used for bi-directional self-attention.

Log a table of validation samples to the configured logger (wandb or swanlab)

Clamp, validate, and potentially resize a bounding box. This function ensures the final bounding box is within image bounds and meet

Calculate the mapping of global layer_idx to local layer_idx Returns: layer_map (Dict: int -> tuple(int, int, int)): mapping f

Calculate the mapping of global layer_idx to local layer_idx Returns: layer_map (Dict: int -> tuple(int, int, int)): mapping f

Calculate the mapping of global layer_idx to local layer_idx Returns: layer_map (Dict: int -> tuple(int, int, int)): mapping f

Calculate the mapping of global layer_idx to local layer_idx Returns: layer_map (Dict: int -> tuple(int, int, int)): mapping f

Calculate the mapping of global layer_idx to local layer_idx Returns: layer_map (Dict: int -> tuple(int, int, int)): mapping f

Calculate the mapping of global layer_idx to local layer_idx Returns: layer_map (Dict: int -> tuple(int, int, int)): mapping f

Calculate the mapping of global layer_idx to local layer_idx Returns: layer_map (Dict: int -> tuple(int, int, int)): mapping f

Calculate the mapping of global layer_idx to local layer_idx Returns: layer_map (Dict: int -> tuple(int, int, int)): mapping f

Merges a list of tensors based on their DTensor placement

hdfs mkdir

Copy images to output_dir/images/ with sequential naming. Args: source_images: List of source image paths source_cont

Parse VLM response containing JSON array of samples.

Parse LLM response to extract list of keywords. Args: response: LLM response text Returns: List of

Attempts to parse latex to an expression sympy can read.

Parse pairwise evaluation results to determine winners. Args: evaluation_result: Result from evaluate() call Re

Convert retrieval results to formatted string.

Patched version of topk_routing_with_score_function that supports router replay.

transform samples in dataset to the samples which rewriting demands, according to the evaluations

Helper function to process a single test case.

Process a single message and return its tokenized representation. Args: index: turn index in the conversation

Query the collect info for a given mesh name. Args: mesh_name (str): Name of the mesh to query collect info for.

Query the dispatch info for a given mesh name. Args: mesh_name (str): Name of the mesh to query dispatch info for

Deduplicate consecutive image tokens in prompt_ids for Qwen2.5-VL, since vLLM will replicate the <|image_pad|> token by image_data. For examp

Register worker with router with error handling. This method attempts to register the current worker with a router service. If regist

Mark worker as free after request completes.

Save pairwise evaluation results to JSON file.

Select the least-loaded worker (simple round-robin by request count).

Recursively split text using separator hierarchy. Args: text: Text to split separators: List of separators t

Splits a tensor into multiple tensors based on the name. This is used to handle qkv and gate_up tensors.

Start profiling for all worker groups if profiling is enabled.

Stop profiling for all worker groups if profiling is enabled.

Parses an expression with sympy.

Best-effort: convert arbitrary input into a 1-D numpy array; fallback to object dtype.

Translate a single sample. Translates the 'input' and 'output' fields of the sample.

Update the multi_modal_inputs of the request in additive manner.

Validate config options not implemented for Megatron backend

Validate config options not implemented for Megatron backend

Validate that passages_dir exists and contains JSONL files. Args: passages_dir: Directory containing JSONL passage files

Validate and improve samples using majority voting (always enabled). Processes samples in batches for efficiency. Args:

Compares the merged Megatron state_dict against a reference safetensors model. Applies necessary name mappings from Megatron to Huggi

Add actor rollout worker based on the actor strategy.

Add critic worker to role mapping.

Add generation prompt for gpt-oss model. Args: message_content: Message content string Returns: Message content string with g

Add reference policy worker if KL loss or KL reward is used.

Add reward model worker if enabled.

allgather a dict into a dict of list Args: data: a dict group: the process group to allgather Returns: dict containing a lis

TODO: optimize this. - We can use async ops - We can use only one allgather Args: tensors: size: group:

Decorate a Worker member function to profile the current rank in the current training step. Requires the target function to be a member funct

Apply KL penalty to the token-level rewards. This function computes the KL divergence between the reference policy and current policy, then a

BYPASS MODE: Use rollout_log_probs as old_log_probs Skips expensive actor forward pass for old_log_prob computation Two sub-modes (contr