Recursively convert config containers (e.g. LockableDict) to built-in Python types.

Yield (name, leaf) for every MM/LN leaf in the tree.

Check if a file path exists and raise an error if it doesn't. Args: path: The file path to check Raises: FileNotFoundEr

Create default tensors (device tensors and pinned memory tensors) based on the source weight device. Args: base_attrs: [(name, attr_name,

在指定维度上去除末尾的 padding 部分。 Args: padded_tensor: 已经经过 padding 的 PyTorch tensor。 depadding_len: 需要从末尾去除的长度。 dim: 需要去除 pad

根据模型名字自动检测量化精度 - 如果模型名字包含 "int8" → "int8" - 如果模型名字包含 "fp8" 且设备支持 → "fp8" - 否则返回 None（表示不使用量化）

Compute SNR between y_pred(tensor) and y_real(tensor) SNR can be calcualted as following equation: SNR(pred, real) = (pred - real)

发送/接收张量，根据是否使用 FP8 选择通信方式 参数: tensor: 要发送的张量 hidden_dims: 隐藏维度大小 comm: 通信器对象 use_fp8

Draw keypoints and connections representing hand pose on a given canvas. Args: canvas (np.ndarray): A 3D numpy array representing th

根据指定轴进行位移并保持维度一致