PaddlePaddle · sunzhongkai588 · Jul 17, 2026 · Jul 16, 2026
diff --git a/swe-paddle/tasks/PaddlePaddle__Paddle-78301/README.md b/swe-paddle/tasks/PaddlePaddle__Paddle-78301/README.md
@@ -0,0 +1,49 @@
+# PaddlePaddle__Paddle-78301
+
+This directory converts Paddle PR #78301 into a SWE-Paddle community task candidate.
+
+## Source
+
+| Field | Value |
+| --- | --- |
+| Repository | `PaddlePaddle/Paddle` |
+| PR | [78301](https://github.com/PaddlePaddle/Paddle/pull/78301) |
+| PR title | `[API Compatibility] align paddle.nn.Layer.to/paddle.Tensor.to -part` |
+| Base commit | `fd041ffe2d941d7219090cb12f6ffb10860dc851` |
+| Squash commit | `d6e41b70154ba52525884afdc15c2a9d763a2cae` |
+| Merged at | `2026-04-03T07:08:36Z` |
+| Track | `C` (feature_or_api) |
+| Task type | `feature_enhancement` |
+| Resource | CPU |
+
+## Summary
+
+Align the public conversion call forms of `paddle.nn.Layer.to` and `paddle.Tensor.to`. The task covers positional and keyword dtype/device/reference-Tensor arguments, blocking options, `copy`, no-argument calls, invalid-input errors, and recursive in-place layer conversion.
+
+## Why This Task
+
+- It comes from a merged API-compatibility fix with a clear user-visible contract.
+- It requires consistent overload parsing across two public APIs rather than a signature-only change.
+- It exercises success paths, argument conflicts, invalid input, layer identity, and recursive sublayer conversion.
+- Its Python-only scope is suitable for CPU verification while remaining non-trivial.
+
+Only the behavior merged in PR #78301 is in scope. Follow-up PRs referenced by `proposal.md` are intentionally excluded from the gold patch and acceptance criteria.
+
+## Files
+
+- `proposal.md`: approved candidate proposal and scope.
+- `instruction.md`: self-contained problem statement for the coding agent.
+- `solution/code.patch`: gold patch for non-test source files.
+- `tests/test.patch`: target behavior tests, including a same-dtype positional `copy=True` case that requires a distinct result.
+- `tests/test.sh`: minimal F2P/P2P test command.
+- `environment/README.md`: base revision, run order, and verification risks.
+
+## Verification
+
+From a Paddle source checkout at the base commit, apply `tests/test.patch` and run:
+
+```bash
+bash tests/test.sh
+```
+
+The target tests should fail before `solution/code.patch` is applied and pass afterward. The existing `TestLayerTo::test_main` node is included as a P2P regression guard.
diff --git a/swe-paddle/tasks/PaddlePaddle__Paddle-78301/environment/README.md b/swe-paddle/tasks/PaddlePaddle__Paddle-78301/environment/README.md
@@ -0,0 +1,37 @@
+# Environment Notes
+
+## Expected Environment
+
+- Repository: `PaddlePaddle/Paddle`
+- Base commit: `fd041ffe2d941d7219090cb12f6ffb10860dc851`
+- Resource: CPU
+- GPU required: no
+- Build path: Paddle source checkout at the base commit with a base-compatible Paddle Python installation. The task changes Python sources only, so the verifier may use a Python overlay on an existing compatible build; otherwise build and install Paddle from the base checkout.
+
+## Run Order
+
+1. Check out `PaddlePaddle/Paddle` at the base commit.
+2. Apply `tests/test.patch`.
+3. Run `bash tests/test.sh`; at least the target behavior should fail before the fix.
+4. Apply `solution/code.patch`.
+5. Run `bash tests/test.sh` again; the target behavior and P2P guard should pass.
+
+## Minimal Test Command
+
+```bash
+bash tests/test.sh
+```
+
+## Verification Scope
+
+- F2P targets: `TestLayerAndTensorToAPI` and `TestTensorToCopyCompatibility::test_copy_as_positional_argument` in `test/legacy_test/test_api_compatibility_part3.py`.
+- P2P guard: `TestLayerTo::test_main` in `test/legacy_test/test_base_layer.py`.
+- The standalone copy case passes `copy=True` as the third positional value to a same-dtype conversion and requires a distinct result. The base parser ignores that positional value, so this case is expected to fail before the fix and pass afterward.
+
+## Risks
+
+- A compatible Paddle build is required even though the patch itself is Python-only.
+- The source test class changes Paddle's global dynamic/static mode in `tearDown`. Run the target class and the P2P guard in separate pytest processes so their global modes cannot leak into one another.
+- The legacy P2P guard uses eager-only gradient internals and must run with `FLAGS_enable_pir_api=0`; otherwise it fails before exercising `Layer.to`.
+- Later follow-up fixes to these APIs are outside this task. Verification must use the stated base commit and only the supplied patches.
+- Verification was completed with a base-compatible local Paddle build and Python-source overlays. Maintainers should reproduce the Run/Test/Fix sequence in the designated verifier environment.
diff --git a/swe-paddle/tasks/PaddlePaddle__Paddle-78301/instruction.md b/swe-paddle/tasks/PaddlePaddle__Paddle-78301/instruction.md
@@ -0,0 +1,23 @@
+# 对齐 Tensor 与 Layer 的转换调用行为
+
+Paddle 的 Tensor 和神经网络 Layer 转换接口目前无法一致地支持常见调用形式。请调整它们的可观察行为，使下列用法能够可靠工作。
+
+## 行为要求
+
+- 接受 dtype、device 或参考 Tensor 作为第一个位置参数。
+- 在适用场景中，支持同时以位置参数传入 device 和 dtype。
+- 支持以关键字参数传入 `device` 和 `dtype`。
+- 支持 `blocking`、`non_blocking` 和 `copy` 的合法位置参数或关键字参数形式。
+- 允许不传参数的转换调用。
+- 同时指定 `blocking` 和 `non_blocking` 时抛出 `TypeError`。
+- 位置参数过多或存在未知关键字参数时抛出 `TypeError`。
+- 第一个位置参数无法识别时抛出 `ValueError`。
+- Layer 转换应原地执行、返回同一个 Layer 对象，并递归应用于子层。
+- 传入参考 Tensor 时，目标对象应采用其转换属性。
+
+## 验收标准
+
+- Tensor 转换支持 dtype、device、参考 Tensor、阻塞选项和复制语义。
+- Layer 转换支持对应的 dtype、device、参考 Tensor 和阻塞选项调用，并保持对象身份不变。
+- 嵌套 Layer 和已注册 buffer 均得到一致转换；按照本任务目标行为，dtype 转换也适用于整数 buffer。
+- 现有 Layer 转换回归测试继续通过。
diff --git a/swe-paddle/tasks/PaddlePaddle__Paddle-78301/proposal.md b/swe-paddle/tasks/PaddlePaddle__Paddle-78301/proposal.md
@@ -37,24 +37,31 @@
 - 目标测试文件：
   - `test/legacy_test/test_api_compatibility_part3.py`
   - `test/legacy_test/test_base_layer.py`
-- 目标测试类：`TestLayerAndTensorToAPI`；P2P 护栏为 `TestLayerTo::test_main`
+- 目标测试类：`TestLayerAndTensorToAPI`、`TestTensorToCopyCompatibility`；P2P 护栏为 `TestLayerTo::test_main`
 - 目标测试命令：
 
   ```bash
+  export PYTHONPATH="$(pwd)/test/legacy_test:$(pwd)/test/dygraph_to_static${PYTHONPATH:+:$PYTHONPATH}"
+  export FLAGS_enable_pir_api=0
   python -m pytest \
     test/legacy_test/test_api_compatibility_part3.py::TestLayerAndTensorToAPI \
+    -q
+  python -m pytest \
+    test/legacy_test/test_api_compatibility_part3.py::TestTensorToCopyCompatibility::test_copy_as_positional_argument \
+    -q
+  python -m pytest \
     test/legacy_test/test_base_layer.py::TestLayerTo::test_main \
     -q
   ```
 
 - 关键 F2P 场景：
   - `Layer.to` 和 `Tensor.to` 接受 dtype、device、参考 Tensor 的位置参数，以及对应的关键字参数组合。
-  - 支持 `blocking`、`non_blocking` 和 `copy` 的合法调用形式；测试补丁应至少增加一个 `Tensor.to(dtype='float64', copy=True)` 或等价的 `copy` smoke case。新增 benchmark 测试可以补充原 PR 未覆盖的边界，但不能引入 #78593 之后的语义。
+  - 支持 `blocking`、`non_blocking` 和 `copy` 的合法调用形式；测试补丁增加同 dtype Tensor 的位置参数用例 `tensor.to(paddle.float64, True, True)`，并验证 `copy=True` 返回不同对象。该用例不得引入 #78593 之后的语义。
   - 同时设置互相冲突的阻塞参数时抛出 `TypeError`；无参数调用保持合法；参数过多或未知关键字抛出 `TypeError`；无法识别的首个位置参数抛出 `ValueError`。
   - `Layer.to` 返回原对象，并对子层执行一致的转换。
   - 按 #78301 合入时的行为，使用 dtype 转换 Layer 时，测试中的整数 buffer 也会转换到目标 dtype。后续 #78839 对这一行为的修正不纳入本任务。
 - 修复前预期：在 `base_commit` 上应用围绕 #78301 目标行为构造的 `tests/test.patch` 后，新增兼容性用例应出现失败或错误。例如 dtype 作为首个位置参数时会被错误解释为 device，部分位置参数组合、无参数调用或阻塞参数冲突无法按目标行为执行，且新增的 Layer/Tensor 统一调用场景不能全部通过。
-- 修复后预期：继续应用仅来自 #78301 非测试文件的 `solution/code.patch` 后，`TestLayerAndTensorToAPI` 及指定的既有 Layer 回归测试全部通过，并满足 #78301 合入时的参数解析和转换行为。
+- 修复后预期：继续应用仅来自 #78301 非测试文件的 `solution/code.patch` 后，`TestLayerAndTensorToAPI`、`TestTensorToCopyCompatibility::test_copy_as_positional_argument` 及指定的既有 Layer 回归测试全部通过，并满足 #78301 合入时的参数解析和转换行为。
 - P2P 候选：`test/legacy_test/test_base_layer.py::TestLayerTo::test_main` 在 #78301 中未被修改，可作为 Layer.to 既有行为回归护栏；完整任务包阶段还应从同模块中挑选未被 `test.patch` 修改、且在 base 和修复后都通过的存量用例。
 
 ## 6. 环境与资源