Utilities

Last updated: Jun 05, 2026 (API docstrings are auto-generated).

This section documents the small set of cross-cutting utilities provided by VeRL-Omni. Most other utilities (e.g. tracking, dataset collation) are inherited from upstream verl and are documented there.

File System Utilities

verl_omni.utils.fs.resolve_model_local_dir(path: str, use_shm: bool = False) → str[source]

Resolve path to an on-disk directory.

FSDP Utilities

VeRL-Omni reuses verl.utils.fsdp_utils verbatim and only overrides a single helper to support layered LoRA collection on diffusers transformer-block models (e.g. Qwen-Image).

verl_omni.utils.fsdp_utils.collect_lora_params(module, layered_summon: bool, base_sync_done: bool, is_diffusers: bool = False, adapter_name: str = 'default', layer_prefixes: Sequence[str] = ('transformer_blocks.',)) → OrderedDict[source]

Collect LoRA or base parameters for weight sync to the rollout worker.

Raises RuntimeError when no parameters were collected (e.g. mismatched layer_prefixes).

Parameters:

  • module – The FSDP-wrapped or plain module.

  • layered_summon – Summon one FSDP unit at a time instead of the full model.

  • base_sync_done – If True, collect only LoRA weights; else full base weights.

  • is_diffusers – Use the diffusers-specific layered summon helper.

  • adapter_name – LoRA adapter name (usually "default").

  • layer_prefixes – FSDP layer name prefixes (["transformer_blocks."]

Dataset Utilities

VeRL-Omni’s RLHF dataset class is a thin subclass of verl.utils.dataset.rl_dataset.RLHFDataset that adds an optional negative_prompt channel for classifier-free guidance. The get_collate_fn, get_dataset_class, create_rl_dataset and create_rl_sampler helpers keep callers importing dataset utilities from a single module.

class verl_omni.utils.dataset.rl_dataset.RLHFDataset(*args, config: DictConfig, **kwargs)[source]

Upstream RLHFDataset extended with negative_prompt support.

Diffusion models trained with classifier-free guidance need a paired negative prompt for every sample. We surface the raw negative prompt messages under raw_negative_prompt so the diffusion agent loop can encode them alongside the positive prompt.

__getitem__(item)[source]

For rollout, apply_chat_template has been moved to AgentLoop, so we only return raw_prompt here.

__init__(*args, config: DictConfig, **kwargs)[source]
verl_omni.utils.dataset.rl_dataset.get_collate_fn(data_config: DictConfig)[source]

Get a custom collate function from data config, falling back to upstream default.

verl_omni.utils.dataset.rl_dataset.get_dataset_class(data_config: DictConfig)[source]

Get RLHF dataset class.

Parameters:

data_config – The data config.

Returns:

The dataset class.

Return type:

dataset_cls

verl_omni.utils.dataset.rl_dataset.create_rl_dataset(data_paths, data_config, tokenizer, processor, is_train=True, max_samples: int = -1)[source]

Create a dataset.

Parameters:

  • data_paths – List of paths to data files.

  • data_config – The data config.

  • tokenizer (Tokenizer) – The tokenizer.

  • processor (Processor) – The processor.

Returns:

The dataset.

Return type:

dataset (Dataset)

Diffusion Padding Utilities

See verl_omni.workers.utils.padding.embeds_padding_2_no_padding() in the Workers Interface section.

vLLM-Omni LoRA Hooks

See verl_omni.utils.vllm_omni.utils.OmniTensorLoRARequest and verl_omni.utils.vllm_omni.utils.VLLMOmniHijack in the Rollout & Agent Loop section.