vllm.model_executor.model_loader.reload.lifecycle

Begin and complete a checkpoint-format weight update in the reload mode VLLM_RELOAD_MODE selects (fixed for the process).

Functions:

• finish_reload –

  Complete the weight update start_reload began.

• start_reload –

  Begin a weight update in the configured mode.

• weight_update_active –

  Whether a weight update is running, when one load_weights call may hold

finish_reload(model, model_config, device)

Complete the weight update start_reload began.

Source code in vllm/model_executor/model_loader/reload/lifecycle.py

def finish_reload(
    model: torch.nn.Module, model_config: ModelConfig, device: torch.device
) -> None:
    """Complete the weight update `start_reload` began."""
    global _weight_update_active
    _weight_update_active = False
    if direct_reload_enabled():
        finish_direct_reload(model, model_config, device)
    else:
        finalize_layerwise_reload(model, model_config)

start_reload(model)

Begin a weight update in the configured mode.

Source code in vllm/model_executor/model_loader/reload/lifecycle.py

def start_reload(model: torch.nn.Module) -> None:
    """Begin a weight update in the configured mode."""
    global _weight_update_active
    if direct_reload_enabled():
        begin_direct_reload(model)
    else:
        initialize_layerwise_reload(model)
    _weight_update_active = True

weight_update_active()

Whether a weight update is running, when one load_weights call may hold only part of the checkpoint. Process-wide, so a module nested in the model passed to start_reload sees it too.

Source code in vllm/model_executor/model_loader/reload/lifecycle.py

def weight_update_active() -> bool:
    """Whether a weight update is running, when one ``load_weights`` call may hold
    only part of the checkpoint. Process-wide, so a module nested in the model
    passed to `start_reload` sees it too."""
    return _weight_update_active