Get all registered Domains

Get the global user simulator guidelines. Args: use_tools: Whether to use the tools guidelines. Returns: The global use

Get the hash of the database.

Returns information about the registry.

Compute per-token log-probabilities (and optionally entropy) on responses. For each sample, extracts response-aligned logits and tokens, then com

Compute log probabilities and entropy. Parameters: logits: Model output logits with shape [seq_len, vocab_size] target_tokens: Ta

Convert the results to a dataframe and add a column for success. Checks that all simulations have the same number of trials. Returns the

Return the next micro-batch for the requested keys. - If `micro_batch_indices` is provided, selects rows according to the current i

Yield response-aligned `(logits_chunk, tokens_chunk)` pairs per sample. After squeezing batch dimension and applying temperature scaling, this

Get the statistics of the database.

Get the statistics of the database.

Get the statistics of the database.

Get the statistics of the database.

Get cache statistics.

Check if the message has text content.

Helper function to initialize or reset a session

Invoke the fused MOE backward kernels to compute gradients. Args: grad_output: Gradient of output, shape (num_tokens * topk, N) or (

Return whether a port is available.

Loads the tasks for the given domain.

Load tasks from a data file, could be json, yaml or toml file.

ref: https://github.com/volcengine/verl/blob/main/verl/utils/fsdp_utils.py

Fire resume_memory_occupation on all engines (non-blocking). Returns a list of Ray ObjectRefs.

Pause health checking. Called when engines are offloaded.

Decorator to register a new User implementation

Resume health checking. Called when engines are onloaded.

Run a list of environment function calls. If the function call is an assertion, an assertion check will be performed.

Search for forecasts for a specific location, optionally filtered by validity window and model. Args: location_id: The l

Set the solo mode of the environment.

Split the elements in a tuple/interval, while handling well-formatted commas in large numbers

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

Health check endpoint

Broadcast `values` (from critic) and optionally `log_probs`/`ref_log_probs` (from actor) across PP ranks to align data dependencies. - V

Convert a list of Tau2 messages to a list of litellm messages.

Run training over a rollout consisting of multiple steps. The model is switched to train mode, training hooks are configured, and ``train_one

Unpack sequences from a packed batch. Args: packed_batch: Packed batch Returns: List of unpacked batches

Update the database of the ToolKit.

Send metadata (Ray), broadcast tensors (NCCL rank 0 → engines).

Use a tool available to the assistant of the domain.

Validate the tool call in solo mode.