etha.tensor_bus.bootstrap

etha.tensor_bus.bootstrap#

Bootstrap utilities.

Attributes#

logger

Classes#

BootstrapInfo

Information about the bootstrap process.

Functions#

bootstrap_client(...)

Bootstrap TensorBusClient with automatic agent rank resolution.

setup_ptrace()

Setup ptrace authorization for agent process.

Module Contents#

class etha.tensor_bus.bootstrap.BootstrapInfo#

Information about the bootstrap process.

Variables:

  • agent_rank – The agent rank this worker is connected to

  • global_rank – Global rank within the worker process group (from torchrun)

  • rank_offset – Offset used for calculation (if applicable)

  • device – CUDA device string (e.g., “cuda:0”)

  • command_queue_path – Path to Agent’s CommandQueue LMDB

  • state_path – Path to Agent’s State LMDB

  • method – How agent_rank was determined (“direct” or “offset”)

agent_rank: int#
command_queue_path: str#
global_rank: int#
method: str#
rank_offset: int | None#
state_path: str#
etha.tensor_bus.bootstrap.bootstrap_client(path_naming_fn: collections.abc.Callable[[int], tuple[str, str]] | None = None, connection_timeout: float = 30.0) → tuple[etha.tensor_bus.client.TensorBusClient, BootstrapInfo]#

Bootstrap TensorBusClient with automatic agent rank resolution.

This function encapsulates the entire Worker-side bootstrap process: 1. Determines agent_rank from environment variables (AGENT_RANK or LOCAL_RANK + OFFSET) 2. Resolves LMDB paths using naming convention 3. Creates and returns TensorBusClient

Environment Variables (priority order): 1. AGENT_RANK: Direct specification (highest priority) 2. LOCAL_RANK + AGENT_RANK_OFFSET: Offset-based calculation

Parameters:

  • path_naming_fn – Optional custom function to get (cmd_queue_path, state_path) from rank. If None, uses default convention: - /tmp/agent_rank{N}_command.lmdb - /tmp/agent_rank{N}_state.lmdb

  • connection_timeout – Max time to wait for Agent connection (seconds, default 30.0)

Returns:

Tuple of TensorBusClient and BootstrapInfo

Return type:

(client, info)

Raises:
etha.tensor_bus.bootstrap.setup_ptrace()#

Setup ptrace authorization for agent process.

For simplicity, we use PR_SET_PTRACER_ANY in this prototype. In production, you’d authorize specific agent PIDs.

etha.tensor_bus.bootstrap.logger#