【Hackathon 8th No.1】add `lu_solve` api for paddle -part #71030

decade-afk · 2025-02-06T13:33:01Z

PR Category

User Experience

PR Types

New features

Description

【Hackathon 8th No.1】add lu_solve api for paddle
docs: PaddlePaddle/docs#7052
rfc: PaddlePaddle/community#962

paddle-bot · 2025-02-06T13:33:07Z

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

decade-afk · 2025-02-10T08:00:23Z

我想请问下dcu芯片的编译是不支持cuslover库的吗？

@jeff41404 @luotao1

jeff41404 · 2025-02-28T04:03:06Z

python/paddle/tensor/linalg.py

+
+        pivots (Tensor): Permutation matrix P of LU decomposition. It has shape :math:`(*, m)`, where :math:`*` is batch dimensions, that can be converted to a permutation matrix P, with data type int32.
+
+        trans (str): The transpose of the matrix A. It can be "N" , "T" or "C", default is "N".


The meaning of "N", "T" or "C" needs to be clearly described, and corresponding sample code should be added to help users better understand.

jeff41404 · 2025-02-28T04:06:49Z

python/paddle/tensor/linalg.py

+def lu_solve(
+    b: Tensor,
+    lu: Tensor,
+    pivots: Tensor,
+    trans: str = "N",


shall we support left to allow user to solve xA=b ?

No, to achieve xA=b , you need to change it to A^Tx^T=b^T , and then combine it with paddle.lu to implement it

jeff41404 · 2025-02-28T04:20:47Z

python/paddle/tensor/linalg.py

+    given LU decomposition :math:`A` and column vector :math:`b`.
+
+    Args:
+        b (Tensor): Column vector `b` in the above equation. It has shape :math:`(*, m, k)`, where :math:`*` is batch dimensions, with data type float32, float16.


also support float64?
and do we support type complex64/complex128?

The precision is aligned with the paddle.lu function. support float64 and float32

jeff41404 · 2025-02-28T04:21:08Z

test/legacy_test/test_lu_solve_op.py

+        self.A_shape = [10, 10]
+        self.b_shape = [10, 5]
+        self.trans = "T"
+        self.dtype = "float64"


Do we support type of complex64/complex128? If so, we need to add a single test.

complex64/complex128 are not supported

decade-afk · 2025-03-03T05:42:54Z

The CI tests have been fully completed, please review. @jeff41404

jeff41404

LGTM

decade-afk · 2025-03-04T06:34:04Z

请问这个一直waiting是正常情况吗 @sunzhongkai588

luotao1 · 2025-03-04T07:05:50Z

这个是新加的CI，先等孙师傅review

sunzhongkai588 · 2025-03-05T09:46:16Z

python/paddle/tensor/linalg.py

+        b (Tensor): Column vector `b` in the above equation. It has shape :math:`(*, m, k)`, where :math:`*` is batch dimensions, with data type float32, float64.
+
+        lu (Tensor): LU decomposition. It has shape :math:`(*, m, m)`, where :math:`*` is batch dimensions, that can be decomposed into an upper triangular matrix U and a lower triangular matrix L, with data type float32, float64.
+
+        pivots (Tensor): Permutation matrix P of LU decomposition. It has shape :math:`(*, m)`, where :math:`*` is batch dimensions, that can be converted to a permutation matrix P, with data type int32.
+
+        trans (str): The transpose of the matrix A. It can be "N" , "T" or "C", "N" means :math:`Ax=b`, "T" means :math:`A^Tx=b`, "C" means :math:`A^Hx=b`, default is "N".


Suggested change

b (Tensor): Column vector `b` in the above equation. It has shape :math:`(*, m, k)`, where :math:`*` is batch dimensions, with data type float32, float64.

lu (Tensor): LU decomposition. It has shape :math:`(*, m, m)`, where :math:`*` is batch dimensions, that can be decomposed into an upper triangular matrix U and a lower triangular matrix L, with data type float32, float64.

pivots (Tensor): Permutation matrix P of LU decomposition. It has shape :math:`(*, m)`, where :math:`*` is batch dimensions, that can be converted to a permutation matrix P, with data type int32.

trans (str): The transpose of the matrix A. It can be "N" , "T" or "C", "N" means :math:`Ax=b`, "T" means :math:`A^Tx=b`, "C" means :math:`A^Hx=b`, default is "N".

b (Tensor): Column vector `b` in the above equation. It has shape :math:`(*, m, k)`, where :math:`*` is batch dimensions, with data type float32, float64.

lu (Tensor): LU decomposition. It has shape :math:`(*, m, m)`, where :math:`*` is batch dimensions, that can be decomposed into an upper triangular matrix U and a lower triangular matrix L, with data type float32, float64.

pivots (Tensor): Permutation matrix P of LU decomposition. It has shape :math:`(*, m)`, where :math:`*` is batch dimensions, that can be converted to a permutation matrix P, with data type int32.

trans (str, optional)): The transpose of the matrix A. It can be "N" , "T" or "C", "N" means :math:`Ax=b`, "T" means :math:`A^Tx=b`, "C" means :math:`A^Hx=b`, default is "N".

另外，name 的参数描述也加一下

@decade-afk 注意一下这里，参数间不要留空行，可选参数需要添加说明，参考 https://www.paddlepaddle.org.cn/documentation/docs/zh/develop/dev_guides/api_contributing_guides/api_docs_guidelines_cn.html

@decade-afk 注意一下这里，参数间不要留空行，可选参数需要添加说明，参考 https://www.paddlepaddle.org.cn/documentation/docs/zh/develop/dev_guides/api_contributing_guides/api_docs_guidelines_cn.html

Done

sunzhongkai588 · 2025-03-05T11:29:40Z

python/paddle/tensor/linalg.py

+    name=None,
+):
+    r"""
+    Computes the solution y to the system of linear equations :math:`Ax = b` ,


y 是啥？好像下文都没有提及

y 是啥？好像下文都没有提及

已经都按照要求修改了

sunzhongkai588

LGTM

sunzhongkai588 · 2025-03-07T07:16:31Z

python/paddle/tensor/linalg.py

+    b: Tensor,
+    lu: Tensor,
+    pivots: Tensor,
+    trans: str = "N",
+    name=None,


这部分还有问题，辛苦 @SigureMo review一下

这个 PR 要不先合了吧，太大了

SigureMo · 2025-03-07T07:21:49Z

test/legacy_test/test_lu_solve_op.py

+
+    def test_static(self):
+        def run(place):
+            paddle.set_flags({'FLAGS_enable_pir_api': False})


为什么静态图只测老 IR？面向 PIR 新增 API 可以不写老 IR 分支，但不能完全不测 PIR

API 贡献指南有点过时了，我后面找时间更新下

为什么静态图只测老 IR？面向 PIR 新增 API 可以不写老 IR 分支，但不能完全不测 PIR

我是根据lu的测试代码写的，而且有测试到新pir的吧。从我另外一个pr那里就是目前ci他环境设置是 FLAGS_enable_pir_api=1

我是根据lu的测试代码写的，而且有测试到新pir的吧。

这是怎么确定的呢？那边代码可并没有禁掉 PIR

我是根据lu的测试代码写的，而且有测试到新pir的吧。

这是怎么确定的呢？那边代码可并没有禁掉 PIR

测试pir是需要设置export FLAGS_enable_pir_api=1对吧，我之前测试静态图的时候就弄过的，后面才改成export FLAGS_enable_pir_api=0，而且脚本那里没有关掉新pir功能。

之前测试静态图的时候就弄过的

不要之前，我们要的是监控，测过一遍并不能说明什么，你这里的代码能做到监控 PIR 么？后面 PIR 完全裸奔那这个 API 就有一半是白做的

测试pir是需要设置export FLAGS_enable_pir_api=1对吧

默认就是 PIR，不需要加 FLAG

之前测试静态图的时候就弄过的

不要之前，我们要的是监控，测过一遍并不能说明什么，你这里的代码能做到监控 PIR 么？后面 PIR 完全裸奔那这个 API 就有一半是白做的

那我是按照这个再写一个pir的测试代码？但是原本他就是pir模式的，我静态图代码之所以没有跑到就是因为，他跳到pir模式了

测试pir是需要设置export FLAGS_enable_pir_api=1对吧

默认就是 PIR，不需要加 FLAG

SigureMo · 2025-03-07T07:23:16Z

test/white_list/no_grad_set_white_list.py

@@ -63,4 +63,5 @@
    'smooth_l1_loss',
    'spectral_norm',
    'complex',
+    'lu_solve',


NEED_TO_FIX_OP_LIST 看起来是要修复的列表，为什么新增 API 要加到这里呢？是有什么问题吗？这个 OP 看起来是有反向的

NEED_TO_FIX_OP_LIST 看起来是要修复的列表，为什么新增 API 要加到这里呢？是有什么问题吗？这个 OP 看起来是有反向的

因为反向有一个算子 pivots 是不需要计算梯度的

pivots 不是输入么？

pivots 不是输入么？

但他是主元信息，他是矩阵进行行变换的位置信息。

我的问题是，如果将 API 加到这里，那反向是否有测到，因为代码明显实现了反向，但是是否有测试呢？

我只是说 pivots 应该是输入而不是 OP？

如果确认测到了，那就没问题

我的问题是，如果将 API 加到这里，那反向是否有测到，因为代码明显实现了反向，但是是否有测试呢？

我只是说 pivots 应该是输入而不是 OP？

如果确认测到了，那就没问题

有的有的，我在另一个pr那里写了的。就是写在impl那里检测不到
#71244
https://xly.bce.baidu.com/paddlepaddle/paddle/newipipe/detail/12424272/job/28789277

那这里我没问题了

SigureMo · 2025-03-07T07:24:38Z

python/paddle/tensor/linalg.py

+    else:
+        check_variable_and_dtype(b, 'dtype', ['float32', 'float64'], 'lu_solve')
+        check_variable_and_dtype(
+            lu, 'dtype', ['float32', 'float64'], 'lu_solve'
+        )
+        check_variable_and_dtype(pivots, 'dtype', ['int32'], 'lu_solve')
+        helper = LayerHelper('lu_solve', **locals())
+        out = helper.create_variable_for_type_inference(dtype=b.dtype)
+        helper.append_op(
+            type='lu_solve',
+            inputs={'B': b, 'Lu': lu, 'Pivots': pivots},
+            outputs={'Out': out},
+            attrs={'trans': trans},
+        )


如上所述，新增 API 这个分支可以不写，但不能为了过覆盖率直接把 PIR 静态图不测了

如上所述，新增 API 这个分支可以不写，但不能为了过覆盖率直接把 PIR 静态图不测了

那么意思就是说静态图的代码是可以不写的？新pir会自动动转静。另外我刚把另一个pr的报错去掉让他再跑了，让那个来检测我的静态代码是否有错误。
#71285

那么意思就是说静态图的代码是可以不写的？新pir会自动动转静。

并不是不写，_C_ops.lu_solve 里自动生成的代码会进行动静分发，只是不需要手动写而已，仍然是动态图走动态图逻辑静态图走静态图逻辑，本质没变

另外我刚把另一个pr的报错去掉让他再跑了，让那个来检测我的静态代码是否有错误。

为啥不在这个 PR 直接改

那么意思就是说静态图的代码是可以不写的？新pir会自动动转静。

并不是不写，_C_ops.lu_solve 里自动生成的代码会进行动静分发，只是不需要手动写而已，仍然是动态图走动态图逻辑静态图走静态图逻辑，本质没变

另外我刚把另一个pr的报错去掉让他再跑了，让那个来检测我的静态代码是否有错误。

为啥不在这个 PR 直接改

因为改的话，还需要加个export指令

SigureMo · 2025-03-07T07:25:50Z

python/paddle/tensor/linalg.py

+    b: Tensor,
+    lu: Tensor,
+    pivots: Tensor,
+    trans: str = "N",


如果是可枚举的值，使用 Literal 来标注类型

如果是可枚举的值，使用 Literal 来标注类型

Done

SigureMo · 2025-03-07T07:27:00Z

python/paddle/tensor/linalg.py

+    lu: Tensor,
+    pivots: Tensor,
+    trans: str = "N",
+    name=None,


Suggested change

name=None,

name: str | None = None,

decade-afk · 2025-03-07T07:48:35Z

那请问目前我是要修改哪里？还是要测试静态图代码？ @SigureMo

SigureMo

apply 下这个修改，如果 CI 没问题，那就没问题了

test/legacy_test/test_lu_solve_op.py

python/paddle/tensor/linalg.py

Co-authored-by: Nyakku Shigure <[email protected]>

GreatV · 2025-03-07T08:31:03Z

python/paddle/tensor/linalg.py

+        b (Tensor): Column vector `b` in the above equation. It has shape :math:`(*, m, k)`, where :math:`*` is batch dimensions, with data type float32, float64.
+        lu (Tensor): LU decomposition. It has shape :math:`(*, m, m)`, where :math:`*` is batch dimensions, that can be decomposed into an upper triangular matrix U and a lower triangular matrix L, with data type float32, float64.
+        pivots (Tensor): Permutation matrix P of LU decomposition. It has shape :math:`(*, m)`, where :math:`*` is batch dimensions, that can be converted to a permutation matrix P, with data type int32.
+        trans (str, optional): The transpose of the matrix A. It can be "N" , "T" or "C", "N" means :math:`Ax=b`, "T" means :math:`A^Tx=b`, "C" means :math:`A^Hx=b`, default is "N".


C 应该是为复数准备的吧，如果是实数 c 和 t 就意义一样了

C 应该是为复数准备的吧，如果是实数 c 和 t 就意义一样了

没错没错，因为我看lu那里也有预留复数的接口

decade-afk · 2025-03-07T08:45:03Z

apply 下这个修改，如果 CI 没问题，那就没问题了

好嘞

paddle-bot bot added the contributor External developers label Feb 6, 2025

luotao1 added the PaddlePaddle Hackathon label Feb 7, 2025

luotao1 assigned luotao1 and jeff41404 Feb 7, 2025

luotao1 mentioned this pull request Feb 8, 2025

【Hackathon 8th】开源贡献个人挑战赛（尝鲜版） #70746

Closed

add cpu

7a6ccd5

decade-afk force-pushed the new_solve branch from 4947137 to 7a6ccd5 Compare February 8, 2025 14:36

decade-afk added 4 commits February 9, 2025 11:38

add test

fdb233c

fix

6ed2302

fix

616c471

fix

276955d

decade-afk added 12 commits February 10, 2025 16:17

test

0c73108

test

29acb14

test

f4229ea

fix

3d069ca

test

6be155f

test

5c5097d

fix

20030cf

fix

033dd58

fix

f14c6ee

fix codestyle

bf9cce3

add cpu

c9b8212

Merge branch 'new_solve' of work:decade-afk/Paddle into new_solve

8f3a34b

decade-afk force-pushed the new_solve branch from 62cf76e to bce6474 Compare February 14, 2025 04:39

c++ test

c2d0d36

decade-afk force-pushed the new_solve branch from 887a4ea to c2d0d36 Compare February 15, 2025 05:03

Merge branch 'PaddlePaddle:develop' into new_solve

6c70f66

decade-afk force-pushed the new_solve branch from 0cec479 to bac95d3 Compare February 18, 2025 06:43

jeff41404 reviewed Feb 28, 2025

View reviewed changes

fix

8f19ce3

jeff41404 reviewed Feb 28, 2025

View reviewed changes

fix

07ff6f4

jeff41404 previously approved these changes Mar 3, 2025

View reviewed changes

luotao1 assigned sunzhongkai588 Mar 3, 2025

luotao1 mentioned this pull request Mar 5, 2025

【Hackathon 8th】开源贡献个人挑战赛 #71310

Open

sunzhongkai588 reviewed Mar 5, 2025

View reviewed changes

fix

9d3ae1e

decade-afk dismissed jeff41404’s stale review via 9d3ae1e March 5, 2025 13:02

fix example

f537bad

sunzhongkai588 previously approved these changes Mar 7, 2025

View reviewed changes

sunzhongkai588 reviewed Mar 7, 2025

View reviewed changes

SigureMo reviewed Mar 7, 2025

View reviewed changes

fix

b2784f7

decade-afk dismissed sunzhongkai588’s stale review via b2784f7 March 7, 2025 07:35

SigureMo reviewed Mar 7, 2025

View reviewed changes

test/legacy_test/test_lu_solve_op.py Outdated Show resolved Hide resolved

python/paddle/tensor/linalg.py Outdated Show resolved Hide resolved

decade-afk and others added 2 commits March 7, 2025 16:26

Update python/paddle/tensor/linalg.py

92ef570

Co-authored-by: Nyakku Shigure <[email protected]>

Update test/legacy_test/test_lu_solve_op.py

28275fd

Co-authored-by: Nyakku Shigure <[email protected]>

GreatV reviewed Mar 7, 2025

View reviewed changes

SigureMo approved these changes Mar 10, 2025

View reviewed changes

luotao1 merged commit 5911a3a into PaddlePaddle:develop Mar 10, 2025
31 checks passed

luotao1 changed the title ~~【Hackathon 8th No.1】add lu_solve api for paddle~~ 【Hackathon 8th No.1】add lu_solve api for paddle -part Mar 10, 2025


		pivots (Tensor): Permutation matrix P of LU decomposition. It has shape :math:`(, m)`, where :math:`` is batch dimensions, that can be converted to a permutation matrix P, with data type int32.

		trans (str): The transpose of the matrix A. It can be "N" , "T" or "C", default is "N".

【Hackathon 8th No.1】add lu_solve api for paddle -part #71030

【Hackathon 8th No.1】add lu_solve api for paddle -part #71030

Conversation

decade-afk commented Feb 6, 2025 • edited by zhwesky2010 Loading

PR Category

PR Types

Description

paddle-bot bot commented Feb 6, 2025

decade-afk commented Feb 10, 2025

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

decade-afk Feb 28, 2025 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

decade-afk commented Mar 3, 2025

jeff41404 left a comment

Choose a reason for hiding this comment

decade-afk commented Mar 4, 2025

luotao1 commented Mar 4, 2025

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

sunzhongkai588 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

decade-afk Mar 7, 2025 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

decade-afk commented Mar 7, 2025

SigureMo left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

decade-afk commented Mar 7, 2025

【Hackathon 8th No.1】add `lu_solve` api for paddle -part #71030

【Hackathon 8th No.1】add `lu_solve` api for paddle -part #71030

decade-afk commented Feb 6, 2025 •

edited by zhwesky2010

Loading

decade-afk Feb 28, 2025 •

edited

Loading

decade-afk Mar 7, 2025 •

edited

Loading