[NewIR]Support train model in new ir using Python API #57010
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
| core.DataType.INT16: 'int16', | ||
| core.DataType.INT32: 'int32', | ||
| core.DataType.INT64: 'int64', | ||
| core.DataType.UINT8: 'uint8', |
| if bias is not None: | ||
| return paddle._ir_ops.add(out, bias) | ||
| else: | ||
| return out |
| var, var.shape, float(self._value), var.dtype, place | ||
| ) | ||
| return None | ||
| else: |
There was a problem hiding this comment.
这里主要是动静态图场景语义不一致，这里在新ir静态图场景下需要创建一个OpResult，无法使用full_
| ) | ||
| for index, grad in enumerate(grads): |
There was a problem hiding this comment.
grad 有内部接口 caculate_gradients_helper 返回的就是前向和反向的对应dict 可考虑直接用内部接口的返回值
| @@ -300,6 +303,23 @@ def GenBuildOutputs( | |||
| VLOG(4) << "Builder construction meta_{name}"; | |||
| phi::MetaTensor meta_{name}(&ir_meta_tensor_{name}); | |||
| """ | |||
|
|
|||
| CREATE_OPTIONAL_INPUT_METATENSOR_TEMPLATE = """ | |||
There was a problem hiding this comment.
是否仅保留 CREATE_OPTIONAL_INPUT_METATENSOR_TEMPLATE即可？CREATE_INPUT_METATENSOR_TEMPLATE 可以删除了
There was a problem hiding this comment.
为了生成代码简洁性更好，这里建议保留CREATE_OPTIONAL_INPUT_METATENSOR_TEMPLATE
| @@ -39,5 +40,28 @@ pir::OpResult split_grad(std::vector<pir::OpResult> out_grads, int axis) { | |||
|
|
|||
| return split_grad_op.x_grad(); | |||
| } | |||
| pir::OpResult get_parameter(const std::string& name, | |||
| phi::DataType dtype, | |||
| const std::vector<int64_t>& shape) { | |||
There was a problem hiding this comment.
TODO：get_parameter的输入建议仅包括 name，type 的获取可从Program::parameters_获取
| if (defining_op->HasAttribute(attr_name)) { | ||
| auto attrs = defining_op->attribute(attr_name) | ||
| .dyn_cast<pir::ArrayAttribute>() | ||
| .AsVector(); |
There was a problem hiding this comment.
建议这里对 attr 属性添加判断，判断是否为 BoolAttribtue 后再 dyn_castpir::BoolAttribute
| phi::IntArray shape = CastPyArg2IntArray(shape_obj, "shape", 2); | ||
| // Call ir static api | ||
| auto static_api_out = | ||
| paddle::dialect::get_parameter(name, dtype, shape.GetData()); |
| @@ -1142,6 +1164,79 @@ def _create_optimization_pass( | |||
| end = len(target_block.ops) | |||
| return target_block._slice_ops(start, end) | |||
|
|
|||
| def _new_ir_create_optimization_pass( | |||
There was a problem hiding this comment.
| def _new_ir_create_optimization_pass( | |
| def _pir_create_optimization_pass( |
There was a problem hiding this comment.
python端命名后续会专门统一修改
PR types
Others
PR changes
Others
Description
Pcard-67164
实现在python端使用新IR进行模型训练功能：