Pad the DataProto by concating with padding_candidate.repeat(padding_size) Args: padding_size (int): the number of repeated paddi

Choose the forward function based on the model and backend. Args: model (PreTrainedModel): The model to apply the monkey patch.

The generation_config created from model config may be different to the pretrained model, this may lead to error when generating: htt

Monkey patch to fix a bug in transformers 4.57.3 where Qwen3VLMoeTextSparseMoeBlock.forward incorrectly uses torch.zeros_like(hidden_states)

Performs a single batch search for multiple queries (original search tool behavior). Args: retrieval_service_url: The URL of the ret

Process tokenizer outputs to consistent shapes via padding/truncation. Args: input_ids: Token indices [batch_size, seq_len] atten

Gather local router maps from all PP ranks into a global router map. Args: local_layers_router_map (torch.Tensor): Local router map

Prepare the sources where data will be acquired. (local corpora / web dataset / teacher model) Input: self.config Ou

Prepare training and validation data for GRPO (Group Relative Policy Optimization). Converts JSONL files to Parquet format expected by verl

Print information about adapter parameters in the model.

Probe a dataset with 1 sample to validate it has image + QA fields. Uses LLM to identify which fields map to image/input/output.

Phase 1: Probe a dataset with 1 sample to validate quality. Returns dataset with score if valid, None otherwise.

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

FP8 quantization based on parameter name Args: weights: Generator of (name, tensor) pairs quant_config: Quantization configuratio

Records the topk indices.

Release the tool instance. Args: instance_id: The instance id of the tool.

This disables various destructive functions and prevents the generated code from interfering with the test (e.g. fork bomb, killing other pro

Note that this operation is in-place

Reorder and merge per-VPP layer blocks into a contiguous layer dimension. Given a tensor shaped as [bs*vpp_size, max_token_len, layer_num_pe

Reset event queue for a new SSE stream (new phase).

Resolve agent loop configuration file path. In multi-node Ray training, relative paths may not resolve correctly because the working director

A context manager to add attributes to a trace for the configured backend. Args: sample_index: Sample index for the trace. step:

Run the pipeline with two-stage process: 1. Generate initial dataset (logged in task executor) 2. Initial dataset evaluating

Run coroutine in the MCP event loop.

Run GRPO training with log streaming.

Initialize Ray cluster and run distributed PPO training process. Args: config: Training configuration object containing all necessary par

Run SFT training with log streaming.

if test(generated_code) is not None it'll try to run the code. otherwise it'll just return an input and output pair.

Check concatenated input_ids of apply_chat_template to each turn equals apply_chat_template to whole messages.

Save only adapter parameters to checkpoint. This is much more efficient than saving the full model when using PEFT, as adapters typically rep

Save lora adapter to safetensors. Returns: lora_path: str, the path to the lora adapter. None if no lora adapter found.

Search image datasets from HuggingFace based on task keywords. Returns a list of dataset metadata (with splits info).

Search datasets from HuggingFace based on task keywords. Returns a list of dataset metadata (without rows).

Search for image-text datasets on HuggingFace Hub by query keyword. Results are filtered to datasets with both image and text modalit

This function sets the global logging format and level. It will be called when import verl

Set input tensor to be used instead of forward()'s input. When doing pipeline parallelism the input from the previous stage comes fro

Set input tensor to be used instead of forward()'s input. When doing pipeline parallelism the input from the previous stage comes fro

Set pad_token_id to eos_token_id if it is None. Args: tokenizer (transformers.PreTrainedTokenizer): The tokenizer to be set.

Scatter the packed router top-k indices back to sequence-parallel ranks and update each local RouterReplay instance with target indices for r

Mark job as running.

Sets up embedding layer in first stage and output layer in last stage. This function initializes word embeddings in the final stage when we a

This is a convenience method to surface the language model's word embeddings, which is necessary for `finalize_model_grads._allreduce_word_emb

Determine whether to quantize to FP8 based on parameter name Quantization rules: - Must end with .weight (exclude bias) - Exclude embeddi

Determine if checkpoint should be saved based on capacity esi expiration. Args: max_steps_duration: Max estimated time (seconds) req

Create a dictionary of worker groups for fused workers. Args: prefix_set: Set of prefixes to create worker groups for Re