[Paddle TensorRT No.19] Add pd_op.cumsum converter #69330
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
|
Sorry to inform you that da03c88's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually. |
| dtype = input_tensor.dtype | ||
| axis = paddle_op.operands()[1].source().get_defining_op().attrs()["value"] | ||
| input_shape = input_tensor.shape | ||
| rank = len(input_shape) |
There was a problem hiding this comment.
现在支持rank=0了，和旧ir的converter保持对齐吧
There was a problem hiding this comment.
现在rank=0的情况似乎通过zero_dims_to_one_dims自动转为rank = 1了
|
对
发自我的iPhone
…------------------ Original ------------------
From: Hanyonggong ***@***.***>
Date: Thu,Dec 26,2024 2:51 PM
To: PaddlePaddle/Paddle ***@***.***>
Cc: lizexu123 ***@***.***>, Comment ***@***.***>
Subject: Re: [PaddlePaddle/Paddle] [Paddle TensorRT No.19] Add pd_op.cumsumconverter (PR #69330)
@Hanyonggong commented on this pull request.
In python/paddle/tensorrt/impls/math.py:
> @@ -227,6 +228,96 @@ def all_converter(network, paddle_op, inputs): ) ***@***.***_registry.register("pd_op.cumsum", trt_version="8.x") +def cumsum_converter(network, paddle_op, inputs): + input_tensor = inputs[0] + dtype = input_tensor.dtype + axis = paddle_op.operands()[1].source().get_defining_op().attrs()["value"] + input_shape = input_tensor.shape + rank = len(input_shape)
现在rank=0的情况似乎通过zero_dims_to_one_dims自动转为rank = 1了
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you commented.Message ID: ***@***.***>
|
那应该就不用加rank =0 的情况了吧 |
python/paddle/tensorrt/impls/math.py
Outdated
| # Obtain the number of cycles | ||
| if input_shape[axis] > 0: | ||
| axis_tensor = np.array(input_shape[axis], dtype=np.int32) | ||
| trip_limit = network.add_constant((), axis_tensor) |
There was a problem hiding this comment.
add_1d_constant_layer(scalar也要区分是否为标量),
python/paddle/tensorrt/impls/math.py
Outdated
| axis_tensor = np.array(input_shape[axis], dtype=np.int32) | ||
| trip_limit = network.add_constant((), axis_tensor) | ||
| else: | ||
| dynamic_shape = network.add_shape(input_tensor).get_output(0) |
There was a problem hiding this comment.
trt_reshape和get_shape_tensor_element复用
| # Obtain the slice shape | ||
| shape_list = [] | ||
| for i in range(rank): | ||
| if i == axis: |
There was a problem hiding this comment.
你这块代码和旧ir的差距较大，我建议你也写一个getAxisLength函数，这边怎么输入维度<0，又要reshape维度
| start = [0] * rank | ||
| size = [1] * rank | ||
| stride = [1] * rank | ||
| input_sliced = network.add_slice(input_tensor, start, size, stride) |
There was a problem hiding this comment.
如果axis=0，sizes_tensor为add_1d_constant_layer(network,1)，否则get_axis_length，这段代码在哪，strides_tensor呢
PR Category
Inference
PR Types
New features
Description
card-71500
添加了pd_op.cumsum Marker 和Converter