ONNX 后端、执行提供者与推理串行化 ​

/**
 * @file Handler file for choosing the correct version of ONNX Runtime, based on the environment.
 * Ideally, we could import the `onnxruntime-web` and `onnxruntime-node` packages only when needed,
 * but dynamic imports don't seem to work with the current webpack version and/or configuration.
 * This is possibly due to the experimental nature of top-level await statements.
 * So, we just import both packages, and use the appropriate one based on the environment:
 *   - When running in node, we use `onnxruntime-node`.
 *   - When running in the browser, we use `onnxruntime-web` (`onnxruntime-node` is not bundled).

/**
 * Currently, Transformers.js doesn't support simultaneous loading of sessions in WASM/WebGPU.
 * For this reason, we need to chain the loading calls.
 * @type {Promise<any>}
 */
let webInitChain = Promise.resolve();
...
export async function createInferenceSession(buffer_or_path, session_options, session_config) {
    await ensureWasmLoaded();
    const logSeverityLevel = getOnnxLogSeverityLevel(env.logLevel ?? LogLevel.WARNING);
    const load = () =>
        InferenceSession.create(buffer_or_path, {
            logSeverityLevel,
            ...session_options,
        });
    const session = await (apis.IS_WEB_ENV ? (webInitChain = webInitChain.then(load)) : load());
    session.config = session_config;
    return session;
}

/**
 * Currently, Transformers.js doesn't support simultaneous execution of sessions in WASM/WebGPU.
 * For this reason, we need to chain the inference calls (otherwise we get "Error: Session already started").
 * @type {Promise<any>}
 */
let webInferenceChain = Promise.resolve();

export async function runInferenceSession(session, ortFeed) {
    const run = () => session.run(ortFeed);
    return apis.IS_WEB_ENV ? (webInferenceChain = webInferenceChain.then(run)) : run();
}