"""Common loader for audio, multimodal and native speech inference.""" def load_swa_checkpoint(run_dir, checkpoint_step, launch_config, device): import os import torch budget = float(os.environ.get("HFD_INFER_MEMORY_BUDGET_GIB", "0")) if budget > 0 and str(device).startswith("cuda"): total = torch.cuda.get_device_properties(device).total_memory torch.cuda.set_per_process_memory_fraction(min(1.0, budget * 2**30 / total), device) architecture = getattr(launch_config.model, "interaction_architecture", "legacy") if architecture == "a1_c_deferred": from .runtime import load_inference_model return load_inference_model(run_dir, checkpoint_step, launch_config, device) if architecture != "legacy": raise ValueError(f"Unknown interaction architecture: {architecture}") from psi.models.full_duplex_swa import FullDuplexSwaModel return FullDuplexSwaModel.from_streaming_checkpoint( run_dir, checkpoint_step, launch_config, device)