Fix validate consistency #803

LittleHeroZZZX · 2026-01-13T07:13:30Z

PR Docs

本 PR 修复差异文档与 PaConvert 中映射规则不一致的问题，具体描述见：PaddlePaddle/docs#7679

PR APIs

Fix validate consistency

- Update torch.nn.MultiheadAttention mapping to use ChangeAPIMatcher with paddle.compat.nn.MultiheadAttention - Update torch.nn.functional.scaled_dot_product_attention mapping to use ChangeAPIMatcher with paddle.compat.nn.functional.scaled_dot_product_attention - Add torch.Tensor.shape to global variable manager - Add test case for torch.Tensor.shape with empty tensor - Remove unsupport flags from MultiheadAttention tests and add parameter initialization - Simplify API mapping configurations by removing detailed argument specifications

Changed the tolerance parameter from `rtol` (relative tolerance) to `atol` (absolute tolerance) in all test cases within `test_nn_MultiheadAttention.py`. This adjustment ensures more consistent and reliable test comparisons, particularly for values near zero where relative tolerance can be overly strict.

… attention tests - Changed tolerance parameter from `rtol` to `atol` in all test cases for more consistent numerical comparisons - Added new test cases (10-13) for GQA (Grouped Query Attention) and different sequence length scenarios - Maintains same tolerance values but uses absolute tolerance instead of relative tolerance

paddle-bot · 2026-01-13T07:13:36Z

Thanks for your contribution!

zhwesky2010 · 2026-01-13T12:09:55Z

paconvert/api_mapping.json

      "inplace"
    ]
  },
-  "torch.nn.Module": {


这些配置先不用删，no_need_convert会覆盖api_mapping.json。后面会将no_need_convert列表统一迁移到api_mapping.json中来配置。no_need_convert名单里不用动。

zhwesky2010 · 2026-01-13T12:11:11Z

paconvert/api_matcher.py

        return code


+class CompositeMatcher(BaseMatcher):


这个的功能是啥，和BaseMatcher一致？

zhwesky2010 · 2026-01-13T12:12:21Z

tests/torchvision_tests/test_inception_v3.py



-def test_case_8():
+# PaConvert can't tell if args in **kawrgs are supported by paddle or not.


这个功能退化了？之前是可以加未转换标记的

这个应该是符合预期的，这个类的init 参数里只有**kwargs，不需要看用户到底传入了些啥参数，PaConvert把它原封不动传进去就行

这个应该是符合预期的，这个类的init 参数里只有**kwargs，不需要看用户到底传入了些啥参数，PaConvert把它原封不动传进去就行

这个case之前是可以跑通的，针对torchvision.models.inception_v3的转换做了修改吗

zhwesky2010 · 2026-01-13T12:12:50Z

paconvert/global_var.py

+        "torch.nn",
+        "torch.double",
+        "torch.cfloat",
+        "transformers.PreTrainedTokenizer",


这个名单里可能还得加几个，baddbmm、round这些

zhwesky2010 · 2026-01-13T12:14:10Z

paconvert/attribute_mapping.json

    "Matcher": "GenericMatcher",
    "paddle_api": "paddle.newaxis"
  },
-  "torch.nn": {


这些留着吧，非必要不删，有用户报过这种case：

if hasattr(torch.nn, 'xxx')

zhwesky2010

目前NO_NEED_CONVERT这一套还是在globar_var里维护的，如果在api_mapping.json里配置。需要保证globar_var配置好，因为官网脚本目前还是会从globar_var里抓数据。

后面再开展：全量NO_NEED_CONVERT迁移到api_mapping.json+更新官网脚本。

zhwesky2010 · 2026-01-15T08:54:01Z

paconvert/api_mapping.json

+    "Matcher": "GenericMatcher",
+    "paddle_api": "paddle.Tensor.log_normal_",
+    "kwargs_change": {
+      "generator": ""


NoNeedConvertMatcher会处理generator、memory_format，可以计作NoNeedConvertMatcher

zhwesky2010 · 2026-01-15T08:59:50Z

tests/torchvision_tests/test_inception_v3.py



-def test_case_8():
+# PaConvert can't tell if args in **kawrgs are supported by paddle or not.


这个应该是符合预期的，这个类的init 参数里只有**kwargs，不需要看用户到底传入了些啥参数，PaConvert把它原封不动传进去就行

这个case之前是可以跑通的，针对torchvision.models.inception_v3的转换做了修改吗

LittleHeroZZZX · 2026-01-16T07:41:38Z

这个case之前是可以跑通的，针对torchvision.models.inception_v3的转换做了修改吗

这个 case 测试的是当存在不支持的参数时 PaConvert 提示不能转换。torchvision.models.inception_v3 的转换规则做了修改，把 **kargs 直接转给 paddle 对应的类，所以没办法识别出不支持的转换，符合预期，不影响对于参数都支持的情况的转换

zhwesky2010 · 2026-01-16T08:16:09Z

paconvert/api_mapping.json

  "torchvision.models.inception_v3": {
    "Matcher": "WeightsMatcher",
    "paddle_api": "paddle.vision.models.inception_v3",
    "args_list": [


这些参数顺序是不是不对？文档中是

由于测试没有覆盖位置参数的用法，导致这里没有发现问题，后面单测还得补充增强。

这里确实没有对齐，校验工具需要检查参数顺序。但是就这个 API 来说，看了下它的源码，他的第一个位置参数是同时支持是 pretained（bool）或者为 weights 参数的

这个 API 后面改文档吧？pretrained 参数是弃用的用法，把 weights 放前面更合适

zhwesky2010

后面看下这一系列API的重载形式，确定docs和paconvert的最终版本。

zhwesky2010 · 2026-01-19T08:22:41Z

paconvert/api_mapping.json

    "Matcher": "WeightsMatcher",
    "paddle_api": "paddle.vision.models.inception_v3",
    "args_list": [
      "weights",


现在看起来不是前后问题，是overload的问题，可以按torch.std_mean的文档形式来，两个签名要二选一的话，优先适配其中公开推荐的用法。

LittleHeroZZZX added 9 commits November 27, 2025 02:26

Merge branch 'master' into apis/log

7b71fc3

Skip sdpa test on cpu and bf16 test on sm<80 gpu

038c856

Set rtol=1e-3 for test_nn_MHA

dc5e782

Remove bf16 skip on v100

b5210d9

Fix validate consistency

6f20e4d

Merge remote-tracking branch 'origin/master' into fix/docs

c8b53c3

LittleHeroZZZX added 2 commits January 13, 2026 09:08

Clear

b9cb247

Fix noneed converter

2adc83f

zhwesky2010 reviewed Jan 13, 2026

View reviewed changes

LittleHeroZZZX added 4 commits January 14, 2026 03:42

Fix ci

4f35680

Fix qwen baseline

bb3dd87

fix

074033c

Add Tensor.round_

6ed3ea6

zhwesky2010 reviewed Jan 15, 2026

View reviewed changes

update torch.Tensor.log_normal_ to NoNeedConvert

b65bd87

zhwesky2010 reviewed Jan 16, 2026

View reviewed changes

zhwesky2010 approved these changes Jan 19, 2026

View reviewed changes

zhwesky2010 merged commit c44691d into PaddlePaddle:master Jan 19, 2026
8 checks passed



		def test_case_8():
		# PaConvert can't tell if args in **kawrgs are supported by paddle or not.

Fix validate consistency #803

Fix validate consistency #803

Uh oh!

Conversation

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

Uh oh!

PR Docs

PR APIs

Uh oh!

paddle-bot bot commented Jan 13, 2026

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!

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!

Choose a reason for hiding this comment

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!

LittleHeroZZZX commented Jan 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

LittleHeroZZZX Jan 19, 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

Uh oh!

zhwesky2010 left a comment

Choose a reason for hiding this comment

Uh oh!

zhwesky2010 Jan 19, 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!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

LittleHeroZZZX commented Jan 13, 2026 •

edited

Loading

zhwesky2010 left a comment •

edited

Loading

LittleHeroZZZX commented Jan 16, 2026 •

edited

Loading

LittleHeroZZZX Jan 19, 2026 •

edited

Loading

zhwesky2010 Jan 19, 2026 •

edited

Loading