PaddlePaddle · co63oc · Aug 12, 2026
diff --git a/swe-paddle/tasks/PaddlePaddle__Paddle-73880/README.md b/swe-paddle/tasks/PaddlePaddle__Paddle-73880/README.md
@@ -0,0 +1,51 @@
+# PaddlePaddle__Paddle-73880
+
+This directory converts Paddle PR #73880 into a SWE-Paddle community task candidate.
+
+## Source
+
+| Field | Value |
+| --- | --- |
+| Repo | `PaddlePaddle/Paddle` |
+| PR | [73880](https://github.com/PaddlePaddle/Paddle/pull/73880) |
+| PR title | `[0-size Tensor Job2 No.66] Add 0-size Tensor support for paddle.nn.functional.softmax_with_cross_entropy` |
+| Base commit | `e1842d4ce364b6e8334c39a8807b256682185d23` |
+| Gold commit | `4b55ef8495c5bdb91e6e96892f31d9cee63ed681` |
+| Merged at | `2025-07-16` |
+| Task type | `bug_fix` |
+| Resource | CPU |
+| Scope | C++ operator kernel (CPU/GPU/XPU) |
+
+## Summary
+
+Fix `paddle.nn.functional.softmax_with_cross_entropy` to correctly handle 0-size tensors in CPU/GPU/XPU kernels.
+
+## 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 C++ kernel level and covers multiple backends (CPU, GPU, XPU).
+- The failure is deterministic: the base revision fails when processing 0-size tensors in cross entropy kernels.
+- The task has clear regression coverage for existing non-zero-size 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.
+- `tests/test.patch`: tests exposing the target behavior.
+- `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) | softmax_with_cross_entropy F2P |
+| --- | ---: | ---: |
+| Base + `tests/test.patch` | PASS | FAIL |
+| Base + `tests/test.patch` + `solution/code.patch` | PASS | PASS |
diff --git a/swe-paddle/tasks/PaddlePaddle__Paddle-73880/environment/README.md b/swe-paddle/tasks/PaddlePaddle__Paddle-73880/environment/README.md
@@ -0,0 +1,52 @@
+# Environment Notes
+
+## Expected Environment
+
+- Repository: `PaddlePaddle/Paddle`
+- Base commit: `e1842d4ce364b6e8334c39a8807b256682185d23`
+- Gold commit: `4b55ef8495c5bdb91e6e96892f31d9cee63ed681`
+- Resource: CPU
+- GPU required: no
+- Patch type: C++ kernel (CPU/GPU/XPU backends)
+- Python dependencies: PaddlePaddle (source build), NumPy, pytest
+
+The verifier should execute against the Paddle source revision represented by the selected patch state. A source build is required since the patch modifies C++ kernel code.
+
+## 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 non-zero-size behavior should pass.
+5. Run the 0-size tensor tests; the target case should fail before the fix.
+6. Apply `solution/code.patch`.
+7. Rebuild Paddle from source.
+8. Reinstall Paddle package.
+9. Run `bash tests/test.sh`; all target tests should pass.
+
+## Minimal Test Command
+
+```bash
+bash tests/test.sh
+```
+
+## Expected Matrix
+
+| Revision state | P2P | softmax_with_cross_entropy 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-73880/instruction.md b/swe-paddle/tasks/PaddlePaddle__Paddle-73880/instruction.md
@@ -0,0 +1,45 @@
+# 修复 `paddle.nn.functional.softmax_with_cross_entropy` 对 0-size Tensor 的处理
+
+## 详细描述
+
+当 `paddle.nn.functional.softmax_with_cross_entropy` 的输入 logits 中存在大小为 `0` 的 dimension（即 `softmax->numel() == 0`）时，当前 CPU/GPU/XPU kernel 实现会直接进入后续计算逻辑，导致 kernel 内部对空数据执行计算时出错或产生未定义行为。
+
+典型表现包括：
+
+- kernel 在处理 0-size tensor 时崩溃或报错
+- 调用失败并抛出与 shape 或内存访问相关的错误
+
+例如：
+
+```python
+import numpy as np
+import paddle
+
+paddle.disable_static()
+# shape [0, 10], soft_label=False
+logits = paddle.to_tensor(np.random.uniform(0.1, 1.0, (0, 10)).astype('float32'))
+label = paddle.to_tensor(np.random.randint(0, 10, (0, 1), dtype='int64'))
+loss, softmax = paddle.nn.functional.softmax_with_cross_entropy(logits, label, return_softmax=True)
+```
+
+上述调用中 `logits` 的 shape 为 `[0, 10]`，不包含任何元素。当 `soft_label` 为 False 时，axis 所在列不能为 0，其他列相同，因此 softmax 和 loss 的 numel 都为 0。按照 API semantics，该调用应正常完成并返回正确 shape 的空 tensor。
+
+需要在 cross entropy kernel 的入口添加 0-size 早期返回处理：
+- 当 `softmax->numel() == 0` 时，分配输出内存后直接返回
+- 当 `soft_label` 为 True 时，loss 需要填充为 0
+- 梯度 kernel 也需要类似处理：当 `logits_grad->numel() == 0` 时分配内存后直接返回
+
+## 验收说明
+
+- 当输入 logits 的 numel 为 0 时，`softmax_with_cross_entropy` 的前向和反向计算应正常完成
+- 返回的 softmax 和 loss tensor 应保持正确 shape 和 dtype
+- 非 0-size tensor 输入下的 cross entropy 行为不得退化
+- CPU、GPU、XPU 后端均应正确处理 0-size 输入
+
+## 技术要求
+
+- 熟悉 C++/CUDA 和 Paddle PHI kernel 开发
+- 了解 Tensor shape、0-size Tensor 和 kernel 执行路径
+- 了解 cross entropy 算子的前向和反向计算语义
+- 了解 Paddle CPU/GPU/XPU kernel 的实现模式
+- 需要从源码编译 Paddle 以验证修改
diff --git a/swe-paddle/tasks/PaddlePaddle__Paddle-73880/proposal.md b/swe-paddle/tasks/PaddlePaddle__Paddle-73880/proposal.md
@@ -0,0 +1,58 @@
+# SWE-Paddle Task Proposal: PaddlePaddle__Paddle-73880
+
+## 1. 来源信息
+
+- Instance ID: `PaddlePaddle__Paddle-73880`
+- PR 链接: https://github.com/PaddlePaddle/Paddle/pull/73880
+- PR 标题: `[0-size Tensor Job2 No.66] Add 0-size Tensor support for paddle.nn.functional.softmax_with_cross_entropy`
+- Base commit: `e1842d4ce364b6e8334c39a8807b256682185d23`
+- Gold commit: `4b55ef8495c5bdb91e6e96892f31d9cee63ed681`
+- Merged at: 2025-07-16
+- 你的身份: contributor
+
+## 2. 问题一句话
+
+`paddle.nn.functional.softmax_with_cross_entropy` 在输入 logits 为 0-size Tensor 时，CPU/GPU/XPU kernel 未处理 0-size 边界情况导致出错，需要在 kernel 入口添加 0-size 早期返回逻辑。
+
+## 3. 为什么适合作为 SWE-Paddle 样本
+
+- **真实性**: 来自 Paddle「0-size Tensor 机制建设」系列任务，是真实研发需求。
+- **代表性**: 覆盖 C++ kernel 层面的 0-size Tensor 边界处理，涉及 CPU、GPU、XPU 三个后端的前向和梯度 kernel。
+- **边界清楚**: 目标仅限 `softmax->numel() == 0` 时的 kernel 早期返回；正向非零尺寸输入不应受影响。
+- **非平凡性**: 修复需要在 `cross_entropy_kernel.cc`、`cross_entropy_grad_kernel.cc`（CPU）、`cross_entropy_kernel.cu`、`cross_entropy_grad_kernel.cu`（GPU）、`cross_entropy_kernel.cc`、`cross_entropy_grad_kernel.cc`（XPU）中分别添加 `numel() == 0` 的早期返回，涉及对 soft_label 为 True 时 loss 填充 0 的特殊处理。
+- **回归护栏明确**: 目标 F2P 可覆盖 0-size Tensor 输入的 `softmax_with_cross_entropy` 算子测试；同文件中已有的 `TestSoftmaxWithCrossEntropyOp` 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: `[operator_kernel, cross_entropy, softmax_with_cross_entropy, 0-size_tensor, cpu_kernel, gpu_kernel, xpu_kernel]`
+
+## 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_softmax_with_cross_entropy_op.py`（`TestSoftmaxWithCrossEntropyOp_ZeroSize`、`TestSoftmaxWithCrossEntropyOp_ZeroSize2`）
+- P2P 候选: 同文件中已有的 `TestSoftmaxWithCrossEntropyOp` 标准算子测试用例。
+- 修复前预期: `base_commit` + `tests/test.patch` 后，0-size Tensor 输入的算子测试失败（kernel 报错或产生错误结果）。
+- 修复后预期: 继续应用 `solution/code.patch` 并重新编译后，0-size Tensor 输入正常返回空 Tensor，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。
+- 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: 含 C++ kernel 修改（CPU/GPU/XPU 端），需要重新编译 Paddle。
+- 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「softmax_with_cross_entropy 对 0-size Tensor 输入的行为异常」，不指出具体 `numel() == 0` 分支逻辑或具体代码位置。
+- 环境风险: 中。任务涉及 C++ kernel 修改，需要源码编译 Paddle，编译时间较长。
+- flaky 风险: 低。测试使用固定的 0-size 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 cross entropy kernel 的 0-size 早期返回，测试明确指向新增的 ZeroSize 测试类，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 在 `base_commit` 编译后确实失败。GPU 和 XPU 的修改在 CPU 环境下无法验证，但 CPU kernel 的修改足以让测试通过。
diff --git a/swe-paddle/tasks/PaddlePaddle__Paddle-73880/solution/code.patch b/swe-paddle/tasks/PaddlePaddle__Paddle-73880/solution/code.patch
@@ -0,0 +1,143 @@
+diff --git a/paddle/phi/kernels/cpu/cross_entropy_grad_kernel.cc b/paddle/phi/kernels/cpu/cross_entropy_grad_kernel.cc
+index 7f755227cb7a3..f9b3daee2571a 100644
+--- a/paddle/phi/kernels/cpu/cross_entropy_grad_kernel.cc
++++ b/paddle/phi/kernels/cpu/cross_entropy_grad_kernel.cc
+@@ -34,6 +34,11 @@ void CrossEntropyWithSoftmaxGradCPUKernel(const CPUContext& dev_ctx,
+                                           int ignore_index,
+                                           int axis,
+                                           DenseTensor* logits_grad) {
++  if (logits_grad->numel() == 0) {
++    dev_ctx.template Alloc<T>(logits_grad);
++    return;
++  }
++
+   const DenseTensor* out_grad = &loss_grad;
+   DenseTensor* logit_grad = logits_grad;
+
+diff --git a/paddle/phi/kernels/cpu/cross_entropy_kernel.cc b/paddle/phi/kernels/cpu/cross_entropy_kernel.cc
+index fa2728257f883..e170f642f8792 100644
+--- a/paddle/phi/kernels/cpu/cross_entropy_kernel.cc
++++ b/paddle/phi/kernels/cpu/cross_entropy_kernel.cc
+@@ -17,6 +17,7 @@ limitations under the License. */
+ #include "paddle/phi/backends/cpu/cpu_context.h"
+ #include "paddle/phi/core/kernel_registry.h"
+ #include "paddle/phi/core/tensor_utils.h"
++#include "paddle/phi/kernels/full_kernel.h"
+ #include "paddle/phi/kernels/funcs/axis_utils.h"
+ #include "paddle/phi/kernels/funcs/cross_entropy.h"
+ #include "paddle/phi/kernels/funcs/math_function.h"
+@@ -79,6 +80,19 @@ void CrossEntropyWithSoftmaxKernel(const Context& dev_ctx,
+                                    int axis,
+                                    DenseTensor* softmax,
+                                    DenseTensor* loss) {
++  if (softmax->numel() == 0) {
++    // When soft_label is False, the axis column cannot be 0. Other dimensions
++    // are the same, so the numel of softmax and loss are both 0.
++    dev_ctx.template Alloc<T>(softmax);
++    dev_ctx.template Alloc<T>(loss);
++
++    // When soft_label is True, the axis column is 1.
++    if (soft_label) {
++      phi::Full<T, Context>(
++          dev_ctx, phi::IntArray(common::vectorize(loss->dims())), 0, loss);
++    }
++    return;
++  }
+   // do not with softmax op, and input is softmax
+   if (!use_softmax) {
+     CrossEntropy<T>(
+diff --git a/paddle/phi/kernels/gpu/cross_entropy_grad_kernel.cu b/paddle/phi/kernels/gpu/cross_entropy_grad_kernel.cu
+index ea3be7c95480e..fed6a9778092a 100644
+--- a/paddle/phi/kernels/gpu/cross_entropy_grad_kernel.cu
++++ b/paddle/phi/kernels/gpu/cross_entropy_grad_kernel.cu
+@@ -241,6 +241,10 @@ void CrossEntropyWithSoftmaxGradKernel(const Context& dev_ctx,
+                                        int ignore_index,
+                                        int axis,
+                                        DenseTensor* logits_grad) {
++  if (logits_grad->numel() == 0) {
++    dev_ctx.template Alloc<T>(logits_grad);
++    return;
++  }
+   auto dtype = label.dtype();
+   if (soft_label) {
+     PADDLE_ENFORCE_EQ(
+diff --git a/paddle/phi/kernels/gpu/cross_entropy_kernel.cu b/paddle/phi/kernels/gpu/cross_entropy_kernel.cu
+index 790e701b2e57f..7a1be5e4b8d04 100644
+--- a/paddle/phi/kernels/gpu/cross_entropy_kernel.cu
++++ b/paddle/phi/kernels/gpu/cross_entropy_kernel.cu
+@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+ #include "paddle/phi/kernels/cross_entropy_kernel.h"
++#include "paddle/phi/kernels/full_kernel.h"
+
+ #include "glog/logging.h"
+
+@@ -1402,6 +1403,20 @@ void CrossEntropyWithSoftmaxKernel(const Context& dev_ctx,
+                                    int axis,
+                                    DenseTensor* softmax,
+                                    DenseTensor* loss) {
++  if (softmax->numel() == 0) {
++    // When soft_label is False, the axis column cannot be 0. Other dimensions
++    // are the same, so the numel of softmax and loss are both 0.
++    dev_ctx.template Alloc<T>(softmax);
++    dev_ctx.template Alloc<T>(loss);
++
++    // When soft_label is True, the axis column is 1.
++    if (soft_label) {
++      phi::Full<T, Context>(
++          dev_ctx, phi::IntArray(common::vectorize(loss->dims())), 0, loss);
++    }
++    return;
++  }
++
+   auto dtype = label.dtype();
+   if (soft_label) {
+     PADDLE_ENFORCE_EQ(
+diff --git a/paddle/phi/kernels/xpu/cross_entropy_grad_kernel.cc b/paddle/phi/kernels/xpu/cross_entropy_grad_kernel.cc
+index a4572fea4187c..e660f64b876bc 100644
+--- a/paddle/phi/kernels/xpu/cross_entropy_grad_kernel.cc
++++ b/paddle/phi/kernels/xpu/cross_entropy_grad_kernel.cc
+@@ -33,6 +33,9 @@ void CrossEntropyWithSoftmaxGradKernel(const Context& dev_ctx,
+                                        DenseTensor* logit_grad) {
+   using XPUType = typename XPUTypeTrait<T>::Type;
+   dev_ctx.template Alloc<T>(logit_grad);
++  if (logit_grad->numel() == 0) {
++    return;
++  }
+
+   const int rank = logit_grad->dims().size();
+   const int axis = phi::funcs::CanonicalAxis(axis_in, rank);
+diff --git a/paddle/phi/kernels/xpu/cross_entropy_kernel.cc b/paddle/phi/kernels/xpu/cross_entropy_kernel.cc
+index a574100165aac..a6be20843ed61 100644
+--- a/paddle/phi/kernels/xpu/cross_entropy_kernel.cc
++++ b/paddle/phi/kernels/xpu/cross_entropy_kernel.cc
+@@ -16,6 +16,7 @@ limitations under the License. */
+
+ #include "paddle/phi/backends/xpu/enforce_xpu.h"
+ #include "paddle/phi/core/kernel_registry.h"
++#include "paddle/phi/kernels/full_kernel.h"
+ #include "paddle/phi/kernels/funcs/axis_utils.h"
+
+ namespace phi {
+@@ -31,6 +32,20 @@ void CrossEntropyWithSoftmaxKernel(const Context& dev_ctx,
+                                    int axis_in,
+                                    DenseTensor* softmax,
+                                    DenseTensor* loss) {
++  if (softmax->numel() == 0) {
++    // When soft_label is False, the axis column cannot be 0. Other dimensions
++    // are the same, so the numel of softmax and loss are both 0.
++    dev_ctx.template Alloc<T>(softmax);
++    dev_ctx.template Alloc<T>(loss);
++
++    // When soft_label is True, the axis column is 1.
++    if (soft_label) {
++      phi::Full<T, Context>(
++          dev_ctx, phi::IntArray(common::vectorize(loss->dims())), 0, loss);
++    }
++    return;
++  }
++
+   using XPUType = typename XPUTypeTrait<T>::Type;
+   const int rank = logits.dims().size();
+   const int axis = phi::funcs::CanonicalAxis(axis_in, rank);