Fire resume_memory_occupation on all engines (non-blocking). Returns a list of Ray ObjectRefs. Skipped for groups that do not overla

Trigger post-process for int4/fp4 quantization on all rollout engines.

Remove vocab padding: param[:vocab_size] for embedding/output layers, else unchanged.

Reset internal offset to the start and return self.

Run a coroutine in the background event loop.

Parse ``--xxx-path`` CLI arguments, store as ``SLIME_CONTRACT_XXX_PATH`` env vars, then call pytest on *file*. Args: file: ``_

Block forward pre-hook for certain configurations.

Return True when a periodic action (eval/save/checkpoint) should run. Args: rollout_id: The current rollout index (0-based).

Create Ray actors, allocate ports, and fire ``engine.init()`` without waiting. Returns ``(init_handles, port_cursors)`` where *init_handles*

Get the text config, handling VLM nesting.

Return the server with ``update_weights=True``. When multiple updatable servers exist, returns the first one (multi-model weight upda

Detect whether W&B should run in offline mode. Priority order: 1) args.wandb_mode if provided 2) WANDB_MODE environment variable

Lock → broadcast → clear → unlock → pbar++. Lock prevents NCCL deadlock.

Add samples to the data source

Destroy all reloadable process groups.

Disable forward pre-hooks for provided DDP-wrapped model chunks. Args: model_chunks (Sequence[DDP]): Sequence of DDP modules to disable h

Extract the answer from inside a LaTeX \\boxed{} command

Gather per-rank metrics, reduce by mean on the DP source rank, and log. Expects `log_dict` to contain plain scalars. The DP source rank prin

Get current output queue size

Similar to itertools.groupby, but do not require iterable to be sorted

Normalize a final answer to a quantitative reasoning question. Args: final_answer: The answer string to normalize Returns: N

Extract PIL images (and videos) from the message list for training. Tries qwen_vl_utils first (Qwen VL family), falls back to generic extract

Reload all reloadable process groups.

Redistribute response tensors from allgather-CP layout to zigzag ring-attn layout. After allgather context parallelism, each rank holds a contigu

Build the configuration for Qwen2 models. Configures Qwen2-specific parameters such as QKV bias settings and layer normaliza

Arguments shared by both phases.

Compute 3D M-RoPE position IDs from input_ids in [bs, seq] format. Image regions are detected by looking for consecutive runs of ``im

Reconstruct full (global) input_ids from zigzag CP chunks. With zigzag CP, each CP rank r holds chunks [r] and [2*cp_size-1-r] for each seque

Calculate token delta for multi-turn conversations. Tokenization logic adapted from: https://verl.readthedocs.io/en/v0.4.1/s

Get valid groups including timeout-based groups