[API Compatibility] Implement paddle.addcmul API -part by Manfredss · Pull Request #77333 · PaddlePaddle/Paddle

Manfredss · 2026-01-13T06:18:15Z

PR Category

User Experience

PR Types

New features

Description

This PR implements the addcmul operator for PaddlePaddle, which performs element-wise multiplication of two tensors, multiplies the result by a scalar value, and adds it to an input tensor.

Formula: output = input + value * tensor1 * tensor2

This operator provides users with a convenient operation for combined multiply-add computations.

Implementation Details

Core Components

C++ Kernels (paddle/phi/kernels/)
- Forward kernels: addcmul_kernel.h and implementations for CPU/GPU
- Backward kernels: addcmul_grad_kernel.h and implementations for CPU/GPU
- Implementation files in impl/ directory with templated functions for different ranks (0-6D)
Operator Configuration (paddle/phi/ops/yaml/)
- Added addcmul operator definition in ops.yaml
- Added addcmul_grad backward operator in backward.yaml
- Configured with proper infer_meta and kernel functions
Shape Inference (paddle/phi/infermeta/)
- Implemented AddcmulInferMeta in ternary.cc/h
- Handles broadcasting for three input tensors
- Validates dimension compatibility
PIR Support (paddle/fluid/pir/dialect/operator/interface/infer_symbolic_shape/)
- Added AddcmulOpInferSymbolicShape for new IR system
- Handles symbolic shape inference with broadcasting
Python API (python/paddle/tensor/)
- Added paddle.addcmul() function in math.py
- Registered Tensor.addcmul() method in __init__.py
- Supports both dynamic and static graph modes
Testing (test/legacy_test/)
- Comprehensive test suite with 52 test cases
- Tests multiple data types: float16, float32, float64, bfloat16
- Tests various tensor shapes and broadcasting scenarios
- Tests gradient computation for all inputs
- Tests zero-size tensors and error conditions
- Tests both OpTest framework and high-level API
Configuration (test/white_list/)
- Added addcmul to FP64 gradient threshold whitelist

Features

Multi-device support: CPU and GPU (CUDA)
Multiple data types: float16, float32, float64, bfloat16
Broadcasting: Full NumPy-style broadcasting support
Gradient support: Automatic differentiation for all three inputs
Tensor dimensions: Supports 0D to 6D tensors
API compatibility: Similar interface to PyTorch's torch.addcmul
Zero-size tensors: Properly handles edge cases

Testing Results

All 52 tests pass successfully:

(paddle) D:\Xue\ML\Paddle\PaddleDebug>python test/legacy_test/test_addcmul.py
WARNING: Logging before InitGoogleLogging() is written to STDERR
W0113 14:12:09.673414 20628 gpu_resources.cc:116] Please NOTE: device: 0, GPU Compute Capability: 12.0, Driver API Version: 13.1, Runtime API Version: 12.9
....I0113 14:12:09.695261 20628 pir_interpreter.cc:1529] New Executor is Running ...
I0113 14:12:09.695261 20628 pir_interpreter.cc:1552] pir interpreter is running by multi-thread mode ...
..I0113 14:12:09.702877 20628 program_interpreter.cc:255] New Executor is Running.
I0113 14:12:09.704878 20628 interpreter_util.cc:624] Standalone Executor is Used.
W0113 14:12:09.744876 20628 eager_utils.cc:3584] Paddle static graph(PIR) not support input out tensor for now!!!!!
C:\Users\***\anaconda3\envs\paddle\Lib\site-packages\paddle\pir\math_op_patch.py:241: UserWarning: Tensor do not have 'place' interface for pir graph mode, try not to use it. None will be returned.
  warnings.warn(
..............................................
----------------------------------------------------------------------
Ran 52 tests in 16.889s

OK

Test coverage includes:

Basic functionality with various shapes (1D, 2D, 3D, large tensors)
Different value parameters (positive, negative, default)
Multiple data types (FP16, FP32, FP64, BF16)
Broadcasting scenarios
Gradient checks for all inputs
Zero-size tensor edge cases
Error handling for invalid inputs
Both static and dynamic graph modes
Tensor method (tensor.addcmul())

API Examples

Dynamic Graph Mode

import paddle

input = paddle.ones([2, 2])
tensor1 = paddle.ones([2, 2]) * 2
tensor2 = paddle.ones([2, 2]) * 3

# Using function API
out = paddle.addcmul(input, tensor1, tensor2, value=0.5)
# Result: [[4., 4.], [4., 4.]]

# Using tensor method
out = input.addcmul(tensor1, tensor2, value=0.5)

Static Graph Mode

import paddle

paddle.enable_static()
input = paddle.static.data('input', shape=[2, 2], dtype='float32')
tensor1 = paddle.static.data('tensor1', shape=[2, 2], dtype='float32')
tensor2 = paddle.static.data('tensor2', shape=[2, 2], dtype='float32')
out = paddle.addcmul(input, tensor1, tensor2, value=0.5)

Broadcasting

input = paddle.ones([3, 4])
tensor1 = paddle.randn([1, 4])
tensor2 = paddle.randn([3, 1])
out = paddle.addcmul(input, tensor1, tensor2, value=2.0)

Backward Compatibility

This PR adds new functionality without modifying existing APIs or behaviors. It is fully backward compatible.

Checklist

Related Issues

【启航计划】PaddlePaddle API兼容性增强 No.354

Additional Notes

The operator uses Eigen for efficient computation with automatic vectorization
Mixed precision computation is handled via MPTypeTrait for numerical stability
Broadcasting follows NumPy semantics
Gradient computation is mathematically verified and tested

Files Changed

New Files (9):

paddle/phi/kernels/addcmul_kernel.h
paddle/phi/kernels/addcmul_grad_kernel.h
paddle/phi/kernels/impl/addcmul_kernel_impl.h
paddle/phi/kernels/impl/addcmul_grad_kernel_impl.h
paddle/phi/kernels/cpu/addcmul_kernel.cc
paddle/phi/kernels/cpu/addcmul_grad_kernel.cc
paddle/phi/kernels/gpu/addcmul_kernel.cu
paddle/phi/kernels/gpu/addcmul_grad_kernel.cu
test/legacy_test/test_addcmul.py

Modified Files (9):

paddle/phi/ops/yaml/ops.yaml
paddle/phi/ops/yaml/backward.yaml
paddle/phi/infermeta/ternary.h
paddle/phi/infermeta/ternary.cc
paddle/fluid/pir/dialect/operator/interface/infer_symbolic_shape/multiary_infer_sym.h
paddle/fluid/pir/dialect/operator/interface/infer_symbolic_shape/multiary_infer_sym.cc
python/paddle/tensor/__init__.py
python/paddle/tensor/math.py
test/white_list/op_threshold_white_list.py

是否引起精度变化

否

…rnels for CPU/GPU (fp16, fp32, fp64, bf16) - Add operator configuration in ops.yaml and backward.yaml - Implement AddcmulInferMeta for shape inference - Add PIR symbolic shape inference support - Add Python API: paddle.addcmul() and Tensor.addcmul() - Add comprehensive test suite (52 tests, all passing) - Add to FP64 gradient threshold whitelist - Formula: output = input + value * tensor1 * tensor2 - Supports broadcasting and multiple dtypes.

… ApiEnhance354

paddle-bot · 2026-01-13T06:18:24Z

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.

zhwesky2010

看下覆盖率，确保都能测到。

同时提前进行下PaConvert测试，确保与torch计算结果一致。截图下PaConvert的case计算结果。

zhwesky2010 · 2026-01-13T10:33:03Z

python/paddle/tensor/math.py

        return _C_ops.addmm_(input, x, y, beta, alpha)


+def addcmul(


新增API直接采取C++下沉的方法吧，这个可以不加

… ApiEnhance354

Manfredss · 2026-01-14T22:05:38Z

/re-run all-failed

… ApiEnhance354

zhwesky2010

这个PR看怎么减小下大小

zhwesky2010 · 2026-01-19T04:55:12Z

python/paddle/_paddle_docs.py


 add_doc_and_signature(
-    "i1",
+    "addcmul",


不要删掉别的，改完后自己先check下所有改动是否符合预期

… ApiEnhance354

zhwesky2010 · 2026-01-21T10:36:06Z

python/paddle/tensor/math.py

        return _C_ops.addmm_(input, x, y, beta, alpha)


+# def addcmul(


这个PR压缩下行数，这些删除掉

codecov-commenter · 2026-01-21T13:49:53Z

Codecov Report

❌ Patch coverage is 34.70149% with 175 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@b79c618). Learn more about missing BASE report.

⚠️ Current head a3ebc16 differs from pull request most recent head 55f8f54

Please upload reports for the commit 55f8f54 to get more accurate results.

Files with missing lines	Patch %	Lines
paddle/phi/kernels/impl/addcmul_grad_kernel_impl.h	0.00%	138 Missing ⚠️
paddle/phi/kernels/impl/addcmul_kernel_impl.h	50.00%	30 Missing ⚠️
...terface/infer_symbolic_shape/multiary_infer_sym.cc	87.17%	5 Missing ⚠️
paddle/phi/kernels/funcs/common_shape.h	0.00%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop   #77333   +/-   ##
==========================================
  Coverage           ?   34.70%           
==========================================
  Files              ?        8           
  Lines              ?      268           
  Branches           ?        0           
==========================================
  Hits               ?       93           
  Misses             ?      175           
  Partials           ?        0

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

Manfredss · 2026-01-21T21:24:07Z

/re-run all-failed

Manfredss · 2026-01-22T23:15:50Z

/re-run all-failed

Manfredss · 2026-01-26T01:43:02Z

/re-run all-failed

zhwesky2010 · 2026-02-12T07:53:07Z

示例代码还是没过，是不是要导入到 _init_.py里去

SigureMo · 2026-02-12T07:59:22Z

示例代码还是没过，是不是要导入到 __init__.py 里去

对的，需要的，要在 paddle/__init__.py 加一下

Manfredss · 2026-02-12T22:53:03Z

/re-run all-failed

zhwesky2010

Coverage无法通过的原因是test_addcmul无法运行通过，本地在GPU上测下addcmul吧。
同时尽可能减少skipif和atol/rtol的修改。

zhwesky2010 · 2026-02-26T09:59:28Z

test/legacy_test/test_addcmul_op.py

+
+        cinn_loss = net_cinn(x, t1, t2)
+        np.testing.assert_allclose(
+            cinn_loss.numpy(), dy_loss.numpy(), rtol=1e-5, atol=1e-5


这几个地方建议精简下case，避免atol/rtol的使用。

zhwesky2010 · 2026-02-26T09:59:51Z

test/legacy_test/test_addcmul_op.py

+
+        cinn_loss = net_cinn(x, t1, t2)
+        np.testing.assert_allclose(
+            cinn_loss.numpy(), dy_loss.numpy(), rtol=1e-5, atol=1e-5


zhwesky2010 · 2026-02-26T10:00:28Z

test/legacy_test/test_addcmul_op.py

+        dy_out = fn(*inputs)
+
+        np.testing.assert_allclose(
+            cinn_out.numpy(), dy_out.numpy(), rtol=1e-5, atol=1e-5


cinn的计算结果与常规的应该是一样的，这里降低阈值的原因是

zhwesky2010 · 2026-02-26T10:01:37Z

test/legacy_test/test_addcmul_op.py

+        paddle.enable_static()
+
+
+@unittest.skipIf(


这里如果需要大量skipif，还是从CMakeLists.txt里面来控制吧。需尽量避免skipif的使用。

Manfredss · 2026-03-05T06:26:12Z

/re-run all-failed

zhwesky2010 · 2026-03-05T11:01:19Z

单测还是运行失败了，看起来是静态图测试InferSymbolicShape的case没跑过，你本地运行能跑过吗？

… ApiEnhance354

Manfredss · 2026-03-09T06:46:28Z

/re-run all-failed

zhwesky2010 · 2026-03-09T10:50:44Z

test/legacy_test/test_addcmul_op.py

+        main = paddle.static.Program()
+        startup = paddle.static.Program()
+        with base.program_guard(main, startup):
+            x = paddle.static.data(name="x", shape=self.shape, dtype=self.dtype)


看这个报错信息是 PIR下创建OP时，创建完一个DataOP，在Insert到block时越界错误。看起来和addcmul自身的OP逻辑无关，是不是单测本身不对。本地复现下问题调试看看。

主要我这边用 python 测单测一直是通过的

你有没有Linux GPU的运行环境

zhwesky2010 · 2026-03-24T03:39:41Z

@Manfredss 这个PR的问题复现了吗

zhwesky2010 · 2026-03-31T04:04:50Z

@Manfredss 这个尽快调试吧

Manfredss · 2026-04-01T19:59:44Z

@Manfredss 这个PR的问题复现了吗

我 linux 下好像也没问题？

Manfredss · 2026-04-03T03:11:38Z

/re-run all-failed

… ApiEnhance354

Manfredss added 2 commits January 13, 2026 14:01

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

b7f40a1

… ApiEnhance354

Manfredss requested review from juncaipeng and zhangting2020 as code owners January 13, 2026 06:18

paddle-bot bot added the contributor External developers label Jan 13, 2026

luotao1 mentioned this pull request Jan 13, 2026

【启航计划】PaddlePaddle API兼容性增强 #76301

Open

zhwesky2010 added the API Compatibility label Jan 13, 2026

zhwesky2010 reviewed Jan 13, 2026

View reviewed changes

Manfredss added 2 commits January 14, 2026 14:33

cpp sink for addcmul

285af67

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

ca45edd

… ApiEnhance354

Manfredss force-pushed the ApiEnhance354 branch from 1179fb5 to ca45edd Compare January 14, 2026 06:44

Manfredss added 2 commits January 15, 2026 13:51

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

8e8d4e4

… ApiEnhance354

fix for linux

45b967f

zhwesky2010 reviewed Jan 19, 2026

View reviewed changes

Manfredss added 2 commits January 21, 2026 09:38

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

8773dc9

… ApiEnhance354

add converage for addcmul

a20a003

Manfredss force-pushed the ApiEnhance354 branch from a9a8f4b to a20a003 Compare January 21, 2026 06:51

zhwesky2010 reviewed Jan 21, 2026

View reviewed changes

Manfredss added 2 commits January 22, 2026 06:34

del commented code and refine tests for addcmul

6919d31

fix test

b82d790

Manfredss added 4 commits January 23, 2026 16:48

fix

5736202

fix

6f3b32c

improve

a582a5a

coverage test

7e24908

fix

8144956

Manfredss dismissed zhwesky2010’s stale review via 8144956 February 12, 2026 04:03

zhwesky2010 previously approved these changes Feb 12, 2026

View reviewed changes

fix

55f8f54

Manfredss dismissed zhwesky2010’s stale review via 55f8f54 February 12, 2026 09:25

zhwesky2010 previously approved these changes Feb 12, 2026

View reviewed changes

zhwesky2010 requested a review from zyfncg February 13, 2026 01:09

SigureMo previously approved these changes Feb 13, 2026

View reviewed changes

zhwesky2010 reviewed Feb 26, 2026

View reviewed changes

fix

d530867

Manfredss dismissed stale reviews from SigureMo and zhwesky2010 via d530867 March 2, 2026 06:11

luotao1 changed the title ~~[API Compatibility No.354] Implement paddle.addcmul API -part~~ [API Compatibility] Implement paddle.addcmul API -part Mar 4, 2026

Manfredss added 3 commits March 6, 2026 07:46

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

6c67637

… ApiEnhance354

fix

954ceec

fix

b331c0d

zhwesky2010 reviewed Mar 9, 2026

View reviewed changes

Manfredss added 2 commits April 4, 2026 06:10

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

5ce27ed

… ApiEnhance354

align api

d82eb97

Manfredss force-pushed the ApiEnhance354 branch from 1d41693 to d82eb97 Compare April 4, 2026 03:57

		return _C_ops.addmm_(input, x, y, beta, alpha)


		# def addcmul(

		paddle.enable_static()


		@unittest.skipIf(

Conversation

Manfredss commented Jan 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

PR Category

PR Types

Description

Implementation Details

Core Components

Features

Testing Results

API Examples

Dynamic Graph Mode

Static Graph Mode

Broadcasting

Backward Compatibility

Checklist

Related Issues

Additional Notes

Files Changed

是否引起精度变化

Uh oh!

paddle-bot bot commented Jan 13, 2026

Uh oh!

zhwesky2010 left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Manfredss commented Jan 14, 2026

Uh oh!

zhwesky2010 left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

codecov-commenter commented Jan 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

Manfredss commented Jan 21, 2026

Uh oh!

Manfredss commented Jan 22, 2026

Uh oh!

Manfredss commented Jan 26, 2026

Uh oh!

zhwesky2010 commented Feb 12, 2026

Uh oh!

SigureMo commented Feb 12, 2026

Uh oh!

Manfredss commented Feb 12, 2026

Uh oh!

zhwesky2010 left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Manfredss commented Mar 5, 2026

Uh oh!

zhwesky2010 commented Mar 5, 2026

Uh oh!

Manfredss commented Mar 9, 2026

Uh oh!

zhwesky2010 Mar 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Manfredss Mar 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Manfredss commented Jan 13, 2026 •

edited

Loading

zhwesky2010 left a comment •

edited

Loading

codecov-commenter commented Jan 21, 2026 •

edited

Loading

zhwesky2010 Mar 9, 2026 •

edited

Loading

Manfredss Mar 11, 2026 •

edited

Loading

Manfredss commented Apr 1, 2026 •

edited

Loading