megengine.core¶

"""
Users should never --
    import megengine.core
But MegEngine developers should know what is inside.
"""

警告

我们不承诺 core 模块中 API 的兼容性和稳定性。

tesnor¶

Store metadata for quantize dtype.

`dtype.get_dtype_bit`
`dtype.is_lowbit`
`dtype.is_bfloat16`
`dtype.create_quantized_dtype`	Get quantized dtype with metadata attribute according to _metadata_dict.
`dtype.quint8`	Consturct a quantized unsigned int8 data type with `scale` (float) and `zero_point` (uint8).
`dtype.qint8`	Construct a quantized int8 data type with `scale` (float).
`dtype.qint32`	Construct a quantized int32 data type with `scale` (float).
`dtype.quint4`	Consturct a quantized unsigned int4 data type with `scale` (float) and `zero_point` (uint8).
`dtype.qint4`	Construct a quantized int4 data type with `scale` (float).
`dtype.convert_to_quint8`	Quantize a float NumPy ndarray into a quint8 one with specified params.
`dtype.convert_to_qint8`	Quantize a float NumPy ndarray into a qint8 one with specified params.
`dtype.convert_from_qint8`	Dequantize a qint8 NumPy ndarray into a float one.
`dtype.convert_to_qint32`	Quantize a float NumPy ndarray into a qint32 one with specified params.
`dtype.convert_from_qint32`	Dequantize a qint32 NumPy ndarray into a float one.
`dtype.convert_to_quint4`	Quantize a float NumPy ndarray into a quint4 one with specified params.
`dtype.convert_from_quint4`	Dequantize a quint4 NumPy ndarray into a float one.
`dtype.convert_to_qint4`	Quantize a float NumPy ndarray into a qint4 one with specified params.
`dtype.convert_from_qint4`	Dequantize a qint4 NumPy ndarray into a float one.
`dtype.convert_from_quint8`	Dequantize a quint8 NumPy ndarray into a float one.

`megbrain_graph.Graph`
`megbrain_graph.VarNode`
`megbrain_graph.OpNode`
`megbrain_graph.InputNode`
`megbrain_graph.OutputNode`
`megbrain_graph.ValueOutputNode`
`megbrain_graph.TensorAttr`
`megbrain_graph.AttrOutputNode`
`megbrain_graph.VirtualDepNode`

`megbrain_graph.set_priority_to_id`	For all oprs in the subgraph constructed by dest_vars, sets its priority to id if its original priority is zero.
`megbrain_graph.optimize_for_inference`	Applies optimize_for_inference pass for computing graph.
`megbrain_graph.modify_opr_algo_strategy_inplace`	C++ graph version of `set_execution_strategy`.
`megbrain_graph.dump_graph`	serialize the computing graph of output_vars and get byte result.
`megbrain_graph.load_graph`	Load a serialized computing graph from file.
`megbrain_graph.apply_normal_varnode`
`megbrain_graph.input_callback`
`megbrain_graph.output_callback`

megengine.Parameter.transpose

megengine.core.tensor.array_method.ArrayMethodMixin