Threads | justinniu | 沐曦开发者论坛

一、软硬件信息
1.服务器厂家:
2.沐曦GPU型号：C550
3.操作系统内核版本：5.14.0-3.1.8.kwai.x86_64
4.是否开启CPU虚拟化：是
5.mx-smi回显：

Quoted message:

mx-smi version: 2.2.9

=================== MetaX System Management Interface Log ===================
Timestamp : Wed Aug 5 16:33:27 2026

Attached GPUs : 1
+---------------------------------------------------------------------------------+
| MX-SMI 2.2.9 Kernel Mode Driver Version: 3.4.3 |
| MACA Version: 3.3.2.27 BIOS Version: 1.30.0.0 |
|------------------+-----------------+---------------------+----------------------|
| Board Name | GPU Persist-M | Bus-id | GPU-Util sGPU-M |
| Pwr:Usage/Cap | Temp Perf | Memory-Usage | GPU-State |
|==================+=================+=====================+======================|
| 0 MetaX C550 | 0 Off | 0000:c2:00.0 | 0% Disabled |
| 91W / 450W | 33C P0 | 858/65536 MiB | Available |
+------------------+-----------------+---------------------+----------------------+

+---------------------------------------------------------------------------------+
| Process: |
| GPU PID Process Name GPU Memory |
| Usage(MiB) |
|=================================================================================|
| no process found |
+---------------------------------------------------------------------------------+

End of Log

二、问题现象
请描述详细的问题现象日志。若日志过长，请上传附件(txt格式）。

代码 has written:

bool MxOnnxRuntimeBackend::Init(const ModelConfig& config) {
CleanupRuntimeResources();
config_ = config;
initialized_ = false;

options_ = MxRuntimeOptions();
options_.device_id = ReadEnvInt("MX_DEVICE_ID", 0, 0, INT_MAX);

std::string model_path = config_.LocalPath;
if (!model_path.empty() && model_path.front() != '/') {
    model_path = BuildRelativePath(model_path);
}
if (model_path.empty()) {
    LOG(ERROR) << LogPrefix("init:1/4") << " empty model path";
    return false;
}
LOG(INFO) << LogPrefix("init:1/4") << " LOAD-ONLY build (no warmup, no inference)"
          << " path=" << model_path
          << " MX_DEVICE_ID=" << options_.device_id;

try {
    LOG(INFO) << LogPrefix("init:2/4") << " acquiring shared Ort::Env";
    Ort::Env& env = SharedOrtEnv();

    // The documented minimum: name the device through the provider options
    // and let MacaEP do the rest. DisableMemPattern and ORT_DISABLE_ALL are
    // kept from the vendor sample because that configuration is known to
    // load successfully.
    Ort::SessionOptions session_options;
    OrtMACAProviderOptions maca_options;
    maca_options.device_id = options_.device_id;
    maca_options.maca_graph_switch = 0;
    session_options.AppendExecutionProvider_MACA(maca_options);
    session_options.DisableMemPattern();
    session_options.SetGraphOptimizationLevel(GraphOptimizationLevel::ORT_DISABLE_ALL);
    LOG(INFO) << LogPrefix("init:2/4") << " SessionOptions ready"
              << " device_id=" << maca_options.device_id;

    LOG(INFO) << LogPrefix("init:3/4") << " creating Ort::Session ..."
              << " (if this is the last line, session creation hung or crashed)";
    const auto begin = std::chrono::steady_clock::now();
    session_ = std::make_unique<Ort::Session>(env, model_path.c_str(), session_options);
    const auto end = std::chrono::steady_clock::now();
    LOG(INFO) << LogPrefix("init:3/4") << " Ort::Session created in "
              << std::chrono::duration_cast<std::chrono::milliseconds>(end - begin).count()
              << " ms";

    LogModelIo();

    initialized_ = true;
    LOG(INFO) << LogPrefix("init:done") << " model loaded"
              << " inputs=" << input_names_.size()
              << " outputs=" << output_names_.size()
              << " device_id=" << options_.device_id;
    return true;
} catch (const Ort::Exception& error) {
    LOG(ERROR) << LogPrefix("init:failed") << " ORT exception loading " << model_path
               << " code=" << error.GetOrtErrorCode() << ": " << error.what();
} catch (const std::exception& error) {
    LOG(ERROR) << LogPrefix("init:failed") << " exception loading " << model_path
               << ": " << error.what();
}
CleanupRuntimeResources();
return false;

}