"""@module numeric-inference/ort-session — one ORT session, CUDA arena kept small.""" from __future__ import annotations import os from .gpu_exec import gpu_call def open_numeric_session(model_path: str): import onnxruntime as ort available = ort.get_available_providers() use_cuda = "CUDAExecutionProvider" in available options = ort.SessionOptions() options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL options.execution_mode = ort.ExecutionMode.ORT_SEQUENTIAL options.intra_op_num_threads = _threads(use_cuda) options.inter_op_num_threads = 1 options.enable_mem_pattern = True options.enable_cpu_mem_arena = not use_cuda providers: list = [] if use_cuda: providers.append(("CUDAExecutionProvider", { "arena_extend_strategy": "kSameAsRequested", "cudnn_conv_algo_search": "HEURISTIC", "do_copy_in_default_stream": True, })) if "CPUExecutionProvider" in available: providers.append("CPUExecutionProvider") return gpu_call(lambda: ort.InferenceSession(model_path, sess_options=options, providers=providers or None)) def _threads(use_cuda: bool) -> int: raw = os.getenv("AXIS_ORT_INTRA_OP", "").strip() if raw.isdigit() and int(raw) > 0: return int(raw) if use_cuda: return 1 return max(1, min(4, os.cpu_count() or 1))