diff --git a/swe-paddle/tasks/PaddlePaddle__Paddle-74491/README.md b/swe-paddle/tasks/PaddlePaddle__Paddle-74491/README.md
new file mode 100644
index 000000000..7b8145b14
--- /dev/null
+++ b/swe-paddle/tasks/PaddlePaddle__Paddle-74491/README.md
@@ -0,0 +1,44 @@
+# PaddlePaddle__Paddle-74491
+
+This directory converts Paddle PR #74491 into a SWE-Paddle community task candidate.
+
+## Source
+
+| Field | Value |
+| --- | --- |
+| Repo | `PaddlePaddle/Paddle` |
+| PR | [#74491](https://github.com/PaddlePaddle/Paddle/pull/74491) |
+| PR title | [API compatibility] add new API `paddle.Tensor.requires_grad` |
+| Base commit | `01666a6667e744874d7f7c379b2649d8bae67f09` |
+| Merged at | `2025-08-13` |
+| Task type | `feature_enhancement` |
+| Resource | CPU |
+
+## Summary
+
+Add a `requires_grad` compatibility property for Paddle Tensor-like values across dynamic, static, and PIR execution modes while keeping existing gradient-control behavior intact.
+
+## Why This Is A Good SWE-Paddle Candidate
+
+- The task comes from a merged Paddle API-compatibility PR with a narrow, well-defined behavior contract.
+- The observable behavior can be verified deterministically without a source build, GPU, network access, or distributed runtime.
+- The change spans three execution modes, so the task checks consistency rather than a single isolated code path.
+- Existing Tensor metadata behavior can be protected independently with a P2P test.
+
+## Files
+
+- `proposal.md`: candidate proposal for maintainer triage.
+- `instruction.md`: self-contained problem statement for the coding agent.
+- `solution/code.patch`: gold patch from the merged PR, limited to production files.
+- `tests/test.patch`: test patch exposing the target behavior.
+- `tests/test.sh`: minimal target test command.
+- `environment/README.md`: environment notes for reproduction.
+- `README.md`: task overview and verification entrypoint.
+
+## Verification
+
+```bash
+bash tests/test.sh
+```
+
+Expected behavior: applying `tests/test.patch` to `base_commit` should fail on the new `requires_grad` behavior while the P2P remains valid; applying both `tests/test.patch` and `solution/code.patch` should pass all target tests.
diff --git a/swe-paddle/tasks/PaddlePaddle__Paddle-74491/environment/README.md b/swe-paddle/tasks/PaddlePaddle__Paddle-74491/environment/README.md
new file mode 100644
index 000000000..a3eced8d7
--- /dev/null
+++ b/swe-paddle/tasks/PaddlePaddle__Paddle-74491/environment/README.md
@@ -0,0 +1,27 @@
+# Environment Notes
+
+This candidate is part of the SWE-Paddle community task set.
+
+## Expected Environment
+
+- Repository: `PaddlePaddle/Paddle`
+- Base commit: `01666a6667e744874d7f7c379b2649d8bae67f09`
+- Resource: CPU
+- GPU required: no
+- Build path: Python-only AST overlay; no Paddle source build is required.
+
+## Run Order
+
+1. Check out `PaddlePaddle/Paddle` at the base commit.
+2. Apply `tests/test.patch`.
+3. Run `bash tests/test.sh`; the target behavior should fail before the fix.
+4. Apply `solution/code.patch`.
+5. Run `bash tests/test.sh` again; the target behavior should pass after the gold patch.
+
+## Minimal Test Command
+
+```bash
+bash tests/test.sh
+```
+
+The verifier is responsible for deriving stable F2P and P2P node IDs from repeated runs.
diff --git a/swe-paddle/tasks/PaddlePaddle__Paddle-74491/instruction.md b/swe-paddle/tasks/PaddlePaddle__Paddle-74491/instruction.md
new file mode 100644
index 000000000..253f3adb8
--- /dev/null
+++ b/swe-paddle/tasks/PaddlePaddle__Paddle-74491/instruction.md
@@ -0,0 +1,26 @@
+# 为 Tensor 增加 `requires_grad` 属性
+
+## 详细描述
+
+Paddle 目前通过 `stop_gradient` 控制 Tensor 是否参与梯度计算。为了提升 API 兼容性和用户体验，还需要提供语义对应的 `requires_grad` 属性。
+
+需要为动态图 Tensor、传统静态图 Variable，以及 PIR 静态图中的 `paddle.pir.Value` 增加 `requires_grad`。其中，`paddle.pir.Value` 是 PIR 计算图中表示算子输入和输出数据的对象，可以理解为 PIR 静态图中的 Tensor-like 符号对象。
+
+- `requires_grad=True` 时，`stop_gradient=False`
+- `requires_grad=False` 时，`stop_gradient=True`
+
+`requires_grad` 应支持读取和设置，并且只能赋布尔值。传入其他类型时应抛出 `TypeError`。
+
+## 验收说明
+
+- 动态图 Tensor、静态图 Variable 和 PIR Value 均支持 `requires_grad`
+- `requires_grad` 的值始终与 `stop_gradient` 相反
+- 设置 `requires_grad` 后，`stop_gradient` 应同步更新
+- 为 `requires_grad` 赋非布尔值时应抛出 `TypeError`
+- 现有 `stop_gradient` 的行为保持不变
+
+## 技术要求
+
+- 熟悉 Python
+- 了解 Paddle 动态图、静态图和 PIR
+- 了解 `stop_gradient` 与梯度计算的关系
diff --git a/swe-paddle/tasks/PaddlePaddle__Paddle-74491/proposal.md b/swe-paddle/tasks/PaddlePaddle__Paddle-74491/proposal.md
new file mode 100644
index 000000000..8eaa5f1e5
--- /dev/null
+++ b/swe-paddle/tasks/PaddlePaddle__Paddle-74491/proposal.md
@@ -0,0 +1,53 @@
+# Task Proposal: PaddlePaddle__Paddle-74491
+
+## 1. 来源信息
+
+* Instance ID：`PaddlePaddle__Paddle-74491`
+* PR 链接：https://github.com/PaddlePaddle/Paddle/pull/74491
+* PR 标题：`[API compatibility] add new API paddle.Tensor.requires_grad`
+* `base_commit`：`01666a6667e744874d7f7c379b2649d8bae67f09`
+* merged 时间：`2025-08-13`
+* 你的身份：熟悉该模块的 contributor
+* 后续联系人：TBD
+
+## 2. 问题一句话
+
+为动态图、静态图和 PIR 模式下的 Tensor 增加统一的 `requires_grad` 属性，使调用方能够通过兼容接口读取和设置 Tensor 是否参与梯度计算。
+
+## 3. 为什么适合作为 SWE-Paddle 样本
+
+* **真实性**：该任务来自已合入 Paddle `develop` 分支的 API compatibility PR，不是人工构造的需求。
+* **代表性**：该任务涉及常用的 Tensor 梯度控制接口，并要求动态图、静态图和 PIR 模式保持一致的属性行为。
+* **边界清楚**：目标行为集中在 `requires_grad` 的读取、赋值、类型校验，以及它与现有 `stop_gradient` 状态之间的对应关系。
+* **非平凡性**：任务需要在三类 Tensor 接口中提供一致的属性语义，同时保证现有梯度控制和 Tensor 元数据行为不受影响。
+
+## 4. 任务类型和标签
+
+* 任务类型：`feature_enhancement`
+* 执行后端：`cpu`
+* 设备范围：`cpu_only`
+* 模块标签：`[tensor, autograd, api_compatibility, dynamic_graph, static_graph, pir]`
+
+## 5. 验证思路
+
+* 目标测试命令：`bash tests/test.sh`
+* 目标测试文件：`test/swe_paddle/test_pr74491_requires_grad.py`
+* 修复前预期：现有 Tensor 元数据和 `stop_gradient` 相关回归测试应通过；动态图、静态图和 PIR 模式下的 `requires_grad` 属性测试应失败。
+* 修复后预期：继续应用 `solution/code.patch` 后，三种模式下的 Tensor 均应满足 `requires_grad == (not stop_gradient)`；布尔赋值应正确更新梯度状态，非布尔赋值应抛出 `TypeError`，全部目标测试应通过。
+* P2P 候选：三种 Tensor 路径中现有的 `stop_gradient` 状态和 `dim` 等元数据行为保持不变。
+
+## 6. 环境与资源
+
+* 资源需求：CPU
+* Paddle 来源：`PaddlePaddle/Paddle` source checkout at `base_commit`
+* 是否能提供 Docker：暂无
+* patch 类型：Python-only
+* 最小测试命令：`bash tests/test.sh`
+* 是否有 oracle 日志：由 SWE-Paddle verifier 结果另行维护
+
+## 7. 风险自查
+
+* 泄露风险：`instruction.md` 只描述 `requires_grad` 的公开属性行为、类型要求和兼容性约束，不透露属性的注册位置、具体函数名称或实现步骤。
+* 环境风险：通过 AST overlay 隔离 source checkout 与当前运行环境之间的版本差异，避免依赖历史 Paddle native runtime。
+* flaky 风险：测试只验证确定性的布尔属性、异常类型和对象状态变化，不依赖随机数、GPU、并发或外部资源。
+* 拆分风险：动态图、静态图和 PIR 三条路径共同实现同一个跨执行模式 Tensor API，适合作为一个完整任务。
diff --git a/swe-paddle/tasks/PaddlePaddle__Paddle-74491/solution/code.patch b/swe-paddle/tasks/PaddlePaddle__Paddle-74491/solution/code.patch
new file mode 100644
index 000000000..4a320e152
--- /dev/null
+++ b/swe-paddle/tasks/PaddlePaddle__Paddle-74491/solution/code.patch
@@ -0,0 +1,159 @@
+diff --git a/python/paddle/base/dygraph/math_op_patch.py b/python/paddle/base/dygraph/math_op_patch.py
+index 6aa14dc8013470b2695374e8ab31302e8aa8c0a1..1571b7e0de27e65dbb51e81748d6aabc61deeaac 100644
+--- a/python/paddle/base/dygraph/math_op_patch.py
++++ b/python/paddle/base/dygraph/math_op_patch.py
+@@ -286,6 +286,40 @@ def monkey_patch_math_tensor():
+         out = _C_ops.transpose(var, perm)
+         return out
+ 
++    @property
++    def requires_grad(self: Tensor) -> bool:
++        """
++        Whether this Tensor requires gradient computation.
++
++        This is a convenience property that returns the opposite of stop_gradient.
++        Setting requires_grad=True is equivalent to setting stop_gradient=False.
++
++        Examples:
++            .. code-block:: python
++
++                >>> import paddle
++                >>> x = paddle.randn([2, 3])
++                >>> print(x.requires_grad)  # False by default
++                >>>
++                >>> x.requires_grad = False
++                >>> print(x.stop_gradient)  # True
++        """
++        return not self.stop_gradient
++
++    @requires_grad.setter
++    def requires_grad(self: Tensor, value: bool) -> None:
++        """
++        Set whether this Tensor requires gradient computation.
++
++        Args:
++            value (bool): True to enable gradient computation, False to disable.
++        """
++        if not isinstance(value, bool):
++            raise TypeError(
++                f"requires_grad must be bool, but got {type(value)}"
++            )
++        self.stop_gradient = not value
++
+     eager_methods = [
+         ('__neg__', _neg_),
+         ('__abs__', _abs_),
+@@ -305,6 +339,7 @@ def monkey_patch_math_tensor():
+         ('size', _size_),
+         ('T', _T_),
+         ('mT', _mT_),
++        ("requires_grad", requires_grad),
+         # for logical compare
+         ('__array_ufunc__', None),
+     ]
+diff --git a/python/paddle/base/layers/math_op_patch.py b/python/paddle/base/layers/math_op_patch.py
+index 5e22912256b0f2f50acc36b68085077f7fec4a1a..8239d53c535e7770f98f074085df886e6bb46a44 100644
+--- a/python/paddle/base/layers/math_op_patch.py
++++ b/python/paddle/base/layers/math_op_patch.py
+@@ -564,6 +564,40 @@ def monkey_patch_variable():
+         """
+         return len(self.shape)
+ 
++    @property
++    def requires_grad(self) -> bool:
++        """
++        Whether this Tensor requires gradient computation.
++
++        This is a convenience property that returns the opposite of stop_gradient.
++        Setting requires_grad=True is equivalent to setting stop_gradient=False.
++
++        Examples:
++            .. code-block:: python
++
++                >>> import paddle
++                >>> x = paddle.randn([2, 3])
++                >>> print(x.requires_grad)  # False by default
++                >>>
++                >>> x.requires_grad = False
++                >>> print(x.stop_gradient)  # True
++        """
++        return not self.stop_gradient
++
++    @requires_grad.setter
++    def requires_grad(self, value: bool) -> None:
++        """
++        Set whether this Tensor requires gradient computation.
++
++        Args:
++            value (bool): True to enable gradient computation, False to disable.
++        """
++        if not isinstance(value, bool):
++            raise TypeError(
++                f"requires_grad must be bool, but got {type(value)}"
++            )
++        self.stop_gradient = not value
++
+     def _scalar_add_(var, value):
+         return _scalar_op_(var, 1.0, value)
+ 
+@@ -814,6 +848,7 @@ def monkey_patch_variable():
+         ('dim', dim),
+         ('ndimension', ndimension),
+         ('ndim', _ndim),
++        ("requires_grad", requires_grad),
+         (
+             '__add__',
+             _binary_creator_('__add__', 'elementwise_add', False, _scalar_add_),
+diff --git a/python/paddle/pir/math_op_patch.py b/python/paddle/pir/math_op_patch.py
+index f51cfd916428a134004a3a6f0abc6cb65124afe3..b114c100ff284f4e6f9078aeda3b4a8abc3c9c15 100644
+--- a/python/paddle/pir/math_op_patch.py
++++ b/python/paddle/pir/math_op_patch.py
+@@ -633,6 +633,40 @@ def monkey_patch_value():
+ 
+         return _C_ops.transpose(self, perm)
+ 
++    @property
++    def requires_grad(self) -> bool:
++        """
++        Whether this Tensor requires gradient computation.
++
++        This is a convenience property that returns the opposite of stop_gradient.
++        Setting requires_grad=True is equivalent to setting stop_gradient=False.
++
++        Examples:
++            .. code-block:: python
++
++                >>> import paddle
++                >>> x = paddle.randn([2, 3])
++                >>> print(x.requires_grad)  # False by default
++                >>>
++                >>> x.requires_grad = False
++                >>> print(x.stop_gradient)  # True
++        """
++        return not self.stop_gradient
++
++    @requires_grad.setter
++    def requires_grad(self, value: bool) -> None:
++        """
++        Set whether this Tensor requires gradient computation.
++
++        Args:
++            value (bool): True to enable gradient computation, False to disable.
++        """
++        if not isinstance(value, bool):
++            raise TypeError(
++                f"requires_grad must be bool, but got {type(value)}"
++            )
++        self.stop_gradient = not value
++
+     def _int_(self):
+         error_msg = """\
+             int(Tensor) is not supported in static graph mode. Because it's value is not available during the static mode.
+@@ -1182,6 +1216,7 @@ def monkey_patch_value():
+         ('size', _size_),
+         ('T', _T_),
+         ('mT', _mT_),
++        ("requires_grad", requires_grad),
+         ('clone', clone),
+         ('clear_gradient', clear_gradient),
+         ('append', append),
diff --git a/swe-paddle/tasks/PaddlePaddle__Paddle-74491/tests/test.patch b/swe-paddle/tasks/PaddlePaddle__Paddle-74491/tests/test.patch
new file mode 100644
index 000000000..a9bf2a0c3
--- /dev/null
+++ b/swe-paddle/tasks/PaddlePaddle__Paddle-74491/tests/test.patch
@@ -0,0 +1,133 @@
+diff --git a/test/swe_paddle/test_pr74491_requires_grad.py b/test/swe_paddle/test_pr74491_requires_grad.py
+new file mode 100644
+index 0000000..25b888c
+--- /dev/null
++++ b/test/swe_paddle/test_pr74491_requires_grad.py
+@@ -0,0 +1,127 @@
++import ast
++import copy
++from pathlib import Path
++
++import pytest
++
++
++MODE_FILES = {
++    "dynamic": Path("python/paddle/base/dygraph/math_op_patch.py"),
++    "static": Path("python/paddle/base/layers/math_op_patch.py"),
++    "pir": Path("python/paddle/pir/math_op_patch.py"),
++}
++
++
++def _walk_functions(tree, name):
++    return [
++        node
++        for node in ast.walk(tree)
++        if isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef))
++        and node.name == name
++    ]
++
++
++def _compile_function(path, name):
++    tree = ast.parse(path.read_text(encoding="utf-8"), filename=str(path))
++    matches = _walk_functions(tree, name)
++    assert matches, f"{path} must provide {name}"
++    node = copy.deepcopy(matches[0])
++    node.decorator_list = []
++    node.name = f"_swe_{name}"
++    module = ast.Module(body=[node], type_ignores=[])
++    ast.fix_missing_locations(module)
++    namespace = {"Tensor": object}
++    exec(compile(module, str(path), "exec"), namespace)
++    return namespace[node.name]
++
++
++def _requires_grad_property(path):
++    tree = ast.parse(path.read_text(encoding="utf-8"), filename=str(path))
++    candidates = _walk_functions(tree, "requires_grad")
++
++    getter = None
++    setter = None
++    for node in candidates:
++        for decorator in node.decorator_list:
++            if isinstance(decorator, ast.Name) and decorator.id == "property":
++                getter = node
++            if (
++                isinstance(decorator, ast.Attribute)
++                and decorator.attr == "setter"
++                and isinstance(decorator.value, ast.Name)
++                and decorator.value.id == "requires_grad"
++            ):
++                setter = node
++
++    assert getter is not None and setter is not None, (
++        f"{path} must expose a readable and writable requires_grad property"
++    )
++
++    getter = copy.deepcopy(getter)
++    setter = copy.deepcopy(setter)
++    getter.name = "_swe_requires_grad_getter"
++    setter.name = "_swe_requires_grad_setter"
++    getter.decorator_list = []
++    setter.decorator_list = []
++
++    module = ast.Module(body=[getter, setter], type_ignores=[])
++    ast.fix_missing_locations(module)
++    namespace = {"Tensor": object}
++    exec(compile(module, str(path), "exec"), namespace)
++    return property(
++        namespace[getter.name],
++        namespace[setter.name],
++    )
++
++
++def test_existing_dim_metadata_behavior_remains_compatible():
++    class FakeTensor:
++        shape = [2, 3, 4, 5]
++
++    for path in MODE_FILES.values():
++        dim = _compile_function(path, "dim")
++        assert dim(FakeTensor()) == 4
++
++
++def test_requires_grad_tracks_stop_gradient_across_tensor_modes():
++    for mode, path in MODE_FILES.items():
++        descriptor = _requires_grad_property(path)
++
++        class FakeTensor:
++            requires_grad = descriptor
++
++            def __init__(self, stop_gradient):
++                self.stop_gradient = stop_gradient
++
++        tensor = FakeTensor(True)
++        assert tensor.requires_grad is False, mode
++
++        tensor.requires_grad = True
++        assert tensor.requires_grad is True, mode
++        assert tensor.stop_gradient is False, mode
++
++        tensor.requires_grad = False
++        assert tensor.requires_grad is False, mode
++        assert tensor.stop_gradient is True, mode
++
++        tensor.stop_gradient = False
++        assert tensor.requires_grad is True, mode
++
++
++def test_requires_grad_rejects_non_boolean_assignments():
++    invalid_values = ["true", 1, 0, None, [], {}]
++
++    for mode, path in MODE_FILES.items():
++        descriptor = _requires_grad_property(path)
++
++        class FakeTensor:
++            requires_grad = descriptor
++
++            def __init__(self):
++                self.stop_gradient = True
++
++        tensor = FakeTensor()
++        for value in invalid_values:
++            with pytest.raises(TypeError, match="requires_grad must be bool"):
++                tensor.requires_grad = value
++            assert tensor.stop_gradient is True, mode
diff --git a/swe-paddle/tasks/PaddlePaddle__Paddle-74491/tests/test.sh b/swe-paddle/tasks/PaddlePaddle__Paddle-74491/tests/test.sh
new file mode 100644
index 000000000..ed5a9b576
--- /dev/null
+++ b/swe-paddle/tasks/PaddlePaddle__Paddle-74491/tests/test.sh
@@ -0,0 +1,4 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+python -m pytest test/swe_paddle/test_pr74491_requires_grad.py -q