sunzhongkai588 · Aug 17, 2026 · Aug 7, 2026
diff --git a/swe-paddle/tasks/PaddlePaddle__Paddle-73535/README.md b/swe-paddle/tasks/PaddlePaddle__Paddle-73535/README.md
@@ -0,0 +1,51 @@
+# PaddlePaddle__Paddle-73535
+
+This directory converts Paddle PR #73535 into a SWE-Paddle community task candidate.
+
+## Source
+
+| Field | Value |
+| --- | --- |
+| Repo | `PaddlePaddle/Paddle` |
+| PR | [73535](https://github.com/PaddlePaddle/Paddle/pull/73535) |
+| PR title | `[Accuracy diff No.112] Fix accuracy diff for paddle.nn.functional.conv1d API` |
+| Base commit | `9c1900ce422e3398bfccf95d3d33ba2cfa91faed` |
+| Gold commit | `f8e6a83f6cb3725902082e7fbb011d7c1e8f6406` |
+| Merged at | `2025-06-24` |
+| Task type | `bug_fix` |
+| Resource | CPU |
+| Scope | Python API (conv1d float16 on CPU) |
+
+## Summary
+
+Fix `paddle.nn.functional.conv1d` to correctly handle float16 weight/bias on CPU by converting them to float32 before computation and casting the result back to float16.
+
+## Why This Is A Good SWE-Paddle Candidate
+
+- It is derived from a merged Paddle bug-fix PR rather than a synthetic issue.
+- The target behavior is isolated to the Python API layer (`conv1d`) and does not require C++ kernel changes.
+- The failure is deterministic: the base revision fails when `conv1d` receives float16 weight/bias on CPU because CPU does not support float16 conv operations.
+- The task has clear regression coverage for existing float32/float64 conv1d behavior.
+- The task runs on CPU and does not require distributed execution, external services, or additional datasets.
+
+## Files
+
+- `proposal.md`: candidate proposal for maintainer triage.
+- `instruction.md`: self-contained problem statement for the coding agent.
+- `solution/code.patch`: gold implementation patch (Python API changes in `conv.py`).
+- `tests/test.patch`: tests exposing the target behavior (adds `TestFunctionalConv1D_CPU_FP16`).
+- `tests/test.sh`: minimal target test command.
+- `environment/README.md`: environment and reproduction notes.
+
+## Verification
+
+```bash
+bash tests/test.sh
+```
+
+Expected behavior:
+
+| Revision state | Existing behavior (P2P) | conv1d F2P |
+| --- | ---: | ---: |
+| Base + `tests/test.patch` | PASS | FAIL |
+| Base + `tests/test.patch` + `solution/code.patch` | PASS | PASS |
diff --git a/swe-paddle/tasks/PaddlePaddle__Paddle-73535/environment/README.md b/swe-paddle/tasks/PaddlePaddle__Paddle-73535/environment/README.md
@@ -0,0 +1,57 @@
+# Environment Notes
+
+## Expected Environment
+
+- Repository: `PaddlePaddle/Paddle`
+- Base commit: `9c1900ce422e3398bfccf95d3d33ba2cfa91faed`
+- Gold commit: `f8e6a83f6cb3725902082e7fbb011d7c1e8f6406`
+- Resource: CPU
+- GPU required: no
+- Patch type: Python API (no C++ kernel changes)
+- Python dependencies: PaddlePaddle (source build or pre-built), NumPy
+
+The verifier should execute against the Paddle source revision represented by the selected patch state. Since the patch only modifies Python code, no recompilation is needed — only the Python files change.
+
+## Build Instructions
+
+1. Check out `PaddlePaddle/Paddle` at the base commit.
+2. Apply `tests/test.patch`.
+3. Build Paddle from source (CPU-only build is sufficient):
+```bash
+mkdir build && cd build
+cmake .. -DWITH_GPU=OFF -DWITH_TESTING=ON -DCMAKE_BUILD_TYPE=Release
+make -j$(nproc)
+```
+4. Install the built Paddle package.
+
+## Run Order
+
+1. Check out `PaddlePaddle/Paddle` at the base commit.
+2. Build and install Paddle from source.
+3. Apply `tests/test.patch`.
+4. Run the P2P tests; existing conv1d behavior should pass.
+5. Run the CPU float16 test; the target case should fail before the fix.
+6. Apply `solution/code.patch`.
+7. Repackage and reinstall the Python files so the modified `conv.py` takes effect:
+```bash
+cd python
+python setup.py bdist_wheel
+pip install --force-reinstall dist/paddlepaddle-*.whl
+cd ..
+```
+8. Run `bash tests/test.sh`; all target tests should pass.
+
+## Minimal Test Command
+
+```bash
+bash tests/test.sh
+```
+
+## Expected Matrix
+
+| Revision state | P2P | conv1d F2P |
+| --- | ---: | ---: |
+| Base + test patch | PASS | FAIL |
+| Base + test patch + solution patch | PASS | PASS |
+
+No GPU, distributed runtime, external service, or additional dataset is required.
diff --git a/swe-paddle/tasks/PaddlePaddle__Paddle-73535/instruction.md b/swe-paddle/tasks/PaddlePaddle__Paddle-73535/instruction.md
@@ -0,0 +1,43 @@
+# 修复 `paddle.nn.functional.conv1d` 在 CPU 上对 float16 weight/bias 的支持
+
+## 详细描述
+
+当 `paddle.nn.functional.conv1d(x, weight, bias, ...)` 在 CPU 设备上运行时，如果 `weight` 或 `bias` 的 dtype 为 `float16`，当前实现会直接调用底层 `conv2d` 算子，但 CPU 的 `conv2d` 算子不支持 `float16` 输入，导致运行时报错或精度异常。
+
+典型表现包括：
+
+- CPU 环境下传入 float16 的 weight/bias 时，conv1d 计算报错或产生错误结果
+- PaddleAPITest 精度对比测试中 `paddle.nn.functional.conv1d` 在 CPU 上出现 accuracy diff
+
+例如：
+
+```python
+import paddle
+import paddle.nn.functional as F
+
+with paddle.base.device_guard("cpu"):
+    x = paddle.ones([1, 1, 1])
+    w = paddle.ones([1, 1, 1]).astype(paddle.float16)
+    b = paddle.ones([1]).astype(paddle.float16)
+    y = F.conv1d(x, w, b)
+    # 期望: y 的值为 [[[2]]]，dtype 为 float16
+```
+
+上述调用中 weight 和 bias 均为 float16，在 CPU 上应能正确执行并返回 float16 类型的结果。
+
+当前 `conv1d` 函数在将 1D 卷积转换为 2D 卷积之前，没有对 CPU 设备上的 float16 输入进行类型转换处理。需要在调用底层 conv2d 算子之前，检测 CPU 设备并将 float16 的 weight 和 bias 临时转换为 x 的 dtype（通常为 float32），完成计算后再将结果转换回 float16。
+
+## 验收说明
+
+- 当 CPU 设备上 weight 或 bias 为 float16 时，`paddle.nn.functional.conv1d` 应正常完成计算
+- 输出的 dtype 应保持为 float16（与输入 weight 的 dtype 一致）
+- 输出的数值应与 GPU 上 float16 conv1d 的结果一致（精度在可接受范围内）
+- 非 float16 输入（float32、float64 等）下的 conv1d 行为不得退化
+- 非 CPU 设备（GPU）上的 conv1d 行为不受影响
+
+## 技术要求
+
+- 熟悉 Python 和 Paddle 动态图 API 开发
+- 了解 Paddle 设备（device）检测和 dtype 转换机制
+- 了解 conv1d 到 conv2d 的转换实现
+- 不需要修改 C++ kernel，仅需修改 Python 层代码
diff --git a/swe-paddle/tasks/PaddlePaddle__Paddle-73535/proposal.md b/swe-paddle/tasks/PaddlePaddle__Paddle-73535/proposal.md
@@ -0,0 +1,58 @@
+# SWE-Paddle Task Proposal: PaddlePaddle__Paddle-73535
+
+## 1. 来源信息
+
+- Instance ID: `PaddlePaddle__Paddle-73535`
+- PR 链接: https://github.com/PaddlePaddle/Paddle/pull/73535
+- PR 标题: `[Accuracy diff No.112] Fix accuracy diff for paddle.nn.functional.conv1d API`
+- Base commit: `9c1900ce422e3398bfccf95d3d33ba2cfa91faed`
+- Gold commit: `f8e6a83f6cb3725902082e7fbb011d7c1e8f6406`
+- Merged at: 2025-06-24
+- 你的身份: contributor
+
+## 2. 问题一句话
+
+`paddle.nn.functional.conv1d` 在 CPU 上不支持 float16 的 weight/bias，需要在 Python 层将 float16 临时转为 float32 计算后再转回 float16。
+
+## 3. 为什么适合作为 SWE-Paddle 样本
+
+- **真实性**: 来自 PaddleAPITest 精度对比测试发现的真实问题（PFCCLab/PaddleAPITest#306）。
+- **代表性**: 覆盖 Python API 层的设备/dtype 兼容性修复，涉及设备检测、dtype 转换和结果还原。
+- **边界清楚**: 目标仅限 CPU 设备上 float16 weight/bias 的 conv1d 调用；其他设备和 dtype 不受影响。
+- **非平凡性**: 修复需要理解 conv1d 到 conv2d 的转换流程，在正确的位置插入 dtype 转换逻辑，并确保输出 dtype 正确还原。
+- **回归护栏明确**: 目标 F2P 可覆盖 CPU float16 conv1d 测试；同文件中已有的 `TestFunctionalConv1DError` 等标准测试用例可作为 P2P 护栏。
+
+## 4. 任务类型和标签
+
+- 任务类型: `bug_fix`
+- 执行后端: `cpu`
+- 设备范围: `cpu_only`
+- 模块标签: `[python_api, conv1d, float16, cpu_compatibility, dtype_conversion]`
+
+## 5. 验证思路
+
+- 目标测试命令: `bash tests/test.sh`
+- 目标测试文件:
+  - `test/legacy_test/test_functional_conv1d.py`（`TestFunctionalConv1D_CPU_FP16`）
+- P2P 候选: 同文件中已有的 `TestFunctionalConv1DError`（含 `Case1`、`Case2`）等标准 conv1d 测试用例。
+- 修复前预期: `base_commit` + `tests/test.patch` 后，CPU float16 conv1d 测试失败（CPU conv2d 不支持 float16）。
+- 修复后预期: 继续应用 `solution/code.patch` 后，CPU float16 conv1d 正常返回正确结果，P2P 存量测试仍然通过。
+
+## 6. 环境与资源
+
+- 是否能提供 Docker: 无
+- Dockerfile 或镜像地址: 暂无
+- Paddle 来源: `PaddlePaddle/Paddle` source checkout at `base_commit`，需要源码编译（或安装对应版本的 Paddle）。
+- OS / Python / CUDA / cuDNN / 其他关键依赖: Linux CPU + Python + numpy；编译需要 CMake、GCC；不要求 CUDA/cuDNN（CPU 编译即可验证）。
+- 硬件: CPU 即可（编译和测试均不需要 GPU）。
+- patch 类型: 仅 Python 代码修改（`python/paddle/nn/functional/conv.py`），无需重新编译。
+- 最小测试命令: `bash tests/test.sh`
+- 是否有 oracle 日志: 无
+
+## 7. 风险自查
+
+- 泄露风险: 正式 `instruction.md` 只描述「conv1d 在 CPU 上对 float16 weight/bias 的行为异常」，不指出具体的转换代码位置或实现细节。
+- 环境风险: 中。任务需要 Paddle 源码环境，但 patch 仅涉及 Python 文件，无需重新编译。
+- flaky 风险: 低。测试使用固定的小 tensor 构造，不依赖随机数差异或多设备同步。
+- 拆分风险: 低。该 PR 目标集中在 `conv1d` 函数的 CPU float16 兼容处理，测试明确指向 `TestFunctionalConv1D_CPU_FP16`，适合作为一个独立样本。
+- 其他不确定点: 完整任务包阶段应确认新增 F2P（`TestFunctionalConv1D_CPU_FP16`）在 `base_commit` 上确实失败。
diff --git a/swe-paddle/tasks/PaddlePaddle__Paddle-73535/solution/code.patch b/swe-paddle/tasks/PaddlePaddle__Paddle-73535/solution/code.patch
@@ -0,0 +1,46 @@
+diff --git a/python/paddle/nn/functional/conv.py b/python/paddle/nn/functional/conv.py
+index 5e5cfed220..067de2c6cf 100644
+--- a/python/paddle/nn/functional/conv.py
++++ b/python/paddle/nn/functional/conv.py
+@@ -16,6 +16,7 @@ from __future__ import annotations
+
+ from typing import TYPE_CHECKING
+
++import paddle
+ from paddle import _C_ops, _legacy_C_ops, get_flags, in_dynamic_mode, pir
+ from paddle.base.framework import _global_flags, in_dynamic_or_pir_mode
+ from paddle.device import (
+@@ -458,6 +459,15 @@ def conv1d(
+     dilation = [1, *convert_to_list(dilation, 1, "dilation")]
+     from ...tensor.creation import assign as paddle_assign
+
++    # cpu not support float16, need to convert dtype.
++    float16_convert = False
++    if paddle.device.get_device() == "cpu":
++        if weight.dtype == paddle.float16:
++            float16_convert = True
++            weight = weight.astype(x.dtype)
++        if bias is not None and bias.dtype == paddle.float16:
++            float16_convert = True
++            bias = bias.astype(x.dtype)
+     weight = paddle_assign(weight)
+     weight = unsqueeze(weight, axis=[-2])
+
+@@ -475,7 +485,6 @@ def conv1d(
+
+     squeeze_axis = -3 if channel_last else -2
+     x = unsqueeze(x, axis=[squeeze_axis])
+-
+     if in_dynamic_or_pir_mode():
+         if l_type == 'conv2d':
+             out = _C_ops.conv2d(
+@@ -530,6 +539,9 @@ def conv1d(
+         if bias is not None:
+             out = _add_with_axis(out, bias, axis=channel_dim)
+     out = squeeze(out, axis=[squeeze_axis])
++    if float16_convert:
++        # out is float16
++        out = out.astype(paddle.float16)
+     return out
+
+
diff --git a/swe-paddle/tasks/PaddlePaddle__Paddle-73535/tests/test.patch b/swe-paddle/tasks/PaddlePaddle__Paddle-73535/tests/test.patch
@@ -0,0 +1,36 @@
+diff --git a/test/legacy_test/test_functional_conv1d.py b/test/legacy_test/test_functional_conv1d.py
+index 751b35b255..760e1a2f2b 100644
+--- a/test/legacy_test/test_functional_conv1d.py
++++ b/test/legacy_test/test_functional_conv1d.py
+@@ -82,5 +82,31 @@ class TestFunctionalConv1DErrorCase2(TestFunctionalConv1DError):
+         self.data_format = "NCL"
+
+
++class TestFunctionalConv1D_CPU_FP16(TestCase):
++    def setUp(self):
++        self.padding = 0
++        self.stride = 1
++        self.dilation = 1
++        self.groups = 1
++        self.data_format = "NCL"
++
++    def test_cpu_fp16(self):
++        with dg.guard(paddle.CPUPlace()):
++            x = paddle.ones([1, 1, 1])
++            w = paddle.ones([1, 1, 1]).astype(paddle.float16)
++            b = paddle.ones([1]).astype(paddle.float16)
++            y = F.conv1d(
++                x,
++                w,
++                b,
++                padding=self.padding,
++                stride=self.stride,
++                dilation=self.dilation,
++                groups=self.groups,
++                data_format=self.data_format,
++            )
++            np.testing.assert_allclose(y.numpy(), [[[2]]])
++
++
+ if __name__ == "__main__":
+     unittest.main()
diff --git a/swe-paddle/tasks/PaddlePaddle__Paddle-73535/tests/test.sh b/swe-paddle/tasks/PaddlePaddle__Paddle-73535/tests/test.sh
@@ -0,0 +1,8 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# P2P tests (pass-to-pass)
+python -m pytest test/legacy_test/test_functional_conv1d.py::TestFunctionalConv1DError -q
+
+# F2P tests (fail-to-pass)
+python -m pytest test/legacy_test/test_functional_conv1d.py::TestFunctionalConv1D_CPU_FP16 -q