[SCU][Paddle TensorRT No.58] Add pd_op.unbind converter #70066
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
|
|
||
| if (op->HasAttribute("axis")) { | ||
| int axis = op->attribute<pir::Int32Attribute>("axis").data(); | ||
| if (axis == 0) { |
There was a problem hiding this comment.
?我如果第一维度不是动态shape,二三维度是动态shape呢
There was a problem hiding this comment.
已修改，但是我发现如果碰到动态shape，报错在marker pass之前，测试时不等进入这里的限制条件判断就报错崩溃了
There was a problem hiding this comment.
那这里就删除吧，convert先按照旧ir的写，如果还是出错，加一个TODO
There was a problem hiding this comment.
已经提交按旧ir写的converter的commit了，还是出错。请问TODO是加在测试还是marker
| axis = paddle_op.attrs().get("axis", 0) | ||
| if axis < 0: | ||
| axis += input_rank | ||
| axis_tensor = add_1D_constant_layer(network, axis) |
There was a problem hiding this comment.
以下代码和旧ir逻辑不同，请参考旧ir-trt
…rker test due to no path to get the false result.
| shuffle_layer = network.add_shuffle(sliced_tensor) | ||
| shuffle_layer.set_input(1, newDims_tensor) | ||
| reshaped_tensor = shuffle_layer.get_output(0) | ||
| output_tensors.append(reshaped_tensor) |
There was a problem hiding this comment.
这里应该是直接返回reshaped_tensor吧，为什么要返回list
There was a problem hiding this comment.
查阅paddle文档以及示例用法，unbind api应该返回一个list，list里面是一个个切割得到的tensor
|
Sorry to inform you that 1840c33's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually. |
PR Category
Inference
PR Types
New features
Description
新增了
pd_op.unbindMarker 和 Converter