The comm thread (used for etcd or listen-based metadata exchange) shares agent data structures (remoteSections_, remoteBackends_, …) with the caller.

Estimate the cost (duration) of a transfer operation.

Estimate the compute time of a batch, in seconds. Accounts for: - Attention FLOPs: O(S²) per layer - MLP FLOPs: O(S × H) per layer (assum

Run num_iter transfers using a sliding window of pipeline_depth in-flight requests. Depth=1 collapses to the original "one create, N posts, one releas

Extract the wheel to a temporary directory. Returns: Path to the temporary directory. The caller is responsible for cleaning up the d

Fetch metadata from etcd or wait for it to be available

Fetch a remote agent's metadata from etcd Once fetched, the metadata will be loaded and cached locally, enabling communication with the remote agent.

Helper function to fill GPU buffer with repeating pattern

Helper function to fill GPU buffer with repeating pattern

Helper function to fill GPU buffer with repeating pattern

* @brief Fill GPU buffer from a pre-filled host pattern buffer * @param gpu_buffer Pointer to the GPU buffer to fill * @param host_buffer Host buffe

Create a ModelConfig instance from a dictionary. Args: config_dict (Dict[str, Any]): Dictionary containing configuration

For now very naive, aggregate requests into batches until it exceeds max_batch_mem or batch_size Args: - num_user_requests: Number of

Args: - batches: List of batches

Generate GUSLI config file from device configurations

Helper function to generate timestamped filename

* @brief Generate a timestamped prefix for object keys * @param base_name Base name for the object key * @return Timestamped prefix string */

Returns the string at the given index

Safe immutable access to descriptor by index

* Asynchronously gets an object from Dell ObjectScale S3 using RDMA acceleration. * This method validates parameters, sets up the S3 request with RDM

* Get the overall status of all transfer requests. * Checks completion of all futures and returns the first error encountered, * or NIXL_SUC

System accelerator type getter

Get all active ranks in current phase.

Get the buffer capacity