Resource config for running everything in-process on the local machine. No slurm, no submitit. Generation uses all visible GPUs, training auto-

Log data to wandb Args: wandb_run: The wandb run to log to data: The data to log iteration: The iteration number

Return the lowercase file extension (including dot) and assert it's supported.

Template for the outer orchestrator job that runs the pipeline loop. A single long-running job with modest resources that coordinates the tra

Split list into exactly n_chunks chunks, distributing items as evenly as possible.

Returns data that can be logged to wandb. If wandb_run given then this data is logged inside this function.

Atomically write JSON to avoid NFS partial/corrupt writes.

Copy of results dict: task_id -> result.

Add string tasks. Returns the list of generated task_ids.

Polls parallel jobs and cleans up after finished jobs. For now we only fun eval jobs in parallel so this is all we need to do.

Copy of util_reports list.

Finetune a model on the train data. This function basically serves as a switcher between different trainign paths and submit it or not.

Returns a list of (state, start_ts, end_ts) per attempt (chronological). Timestamps are Unix seconds as floats. end_ts is None if not finishe

Launch a number of workers based on the resources config. This should be implemented by the subclass. If num_workers not giv

Send Lean code to the verifier and check if it passes. Args: verifier_address: The address of the verifier server. lean_code: A l

Block until all tasks are completed. Returns True if done, False if timeout.

Inputs: code (`str`): The code of formal proof. require_verification (`bool`, *optional*, defaults to True):

Evaluate a prover model. Args: prover_config: The config for the prover model. gen_resources_config: The resources config fo

Launches as many workers on master machine as there are GPUs available