[SWE-Paddle] Complete task package for PaddlePaddle__Paddle-57741

swe-paddle/tasks/PaddlePaddle__Paddle-57741/README.md

@@ -0,0 +1,46 @@
+# PaddlePaddle__Paddle-57741
+
+This directory converts Paddle PR #57741 into a SWE-Paddle community task candidate (PIR: add `memcpy`).
+
+## Source
+
+| Field | Value |
+| --- | --- |
+| Repo | `PaddlePaddle/Paddle` |
+| Primary PR | [57741](https://github.com/PaddlePaddle/Paddle/pull/57741) |
+| PR title | `【PIR】add memcpy in PIR` |
+| Base commit | `f984ed1a56960aeee0059c67b965406984565356` |
+| Gold endpoint | `4288e25e07895e2fd9985b7a2ec94baedac39159` |
+| Merged at | `2023-10-23` |
+| Proposal | community `#1487` |
+| Task type | `feature_enhancement` |
+| Resource | CPU (source build required) |
+
+## Summary
+
+在 PIR 路径下补齐 `memcpy` 算子支持，使动转静场景中的 Tensor 设备间拷贝用例可在 PIR 下正确运行。
+
+## Why This Sample
+
+- **真实 PIR 迁移**：来自已合入的框架算子迁 PIR PR。
+- **基础算子路径**：覆盖 YAML / compat、kernel pass 与动转静测试开启。
+- **边界清晰**：目标集中在 `memcpy`；不扩展到其他设备管理 API。
+- **需 source build**：含 C++ / YAML / pass，不能只靠 wheel overlay。
+
+## Files
+
+- `proposal.md`: approved proposal (do not modify in this package PR).
+- `instruction.md`: self-contained problem statement for the coding agent.
+- `solution/code.patch`: gold production patch relative to base.
+- `tests/test.patch`: test patch exposing the target behavior.
+- `tests/test.sh`: minimal target test command.
+- `environment/README.md`: base commit and reproduction notes.
+
+## Verification
+
+```bash
+bash tests/test.sh
+```
+
+Expected: with `tests/test.patch` on `base_commit`, PIR memcpy related cases fail/error.
+After `solution/code.patch` and rebuild, target cases pass.

swe-paddle/tasks/PaddlePaddle__Paddle-57741/environment/README.md

@@ -0,0 +1,32 @@
+# Environment Notes
+
+SWE-Paddle task candidate for PaddlePaddle/Paddle PR #57741.
+
+## Expected Environment
+
+- Repository: `PaddlePaddle/Paddle`
+- Base commit: `f984ed1a56960aeee0059c67b965406984565356` (parent of #57741 merge `4288e25e`)
+- Gold endpoint: `4288e25e07895e2fd9985b7a2ec94baedac39159`
+- Resource: CPU
+- GPU required: no (CPU memcpy cases are the primary gate; GPU cases optional if CUDA is available)
+- Patch type: **source build required**. Production changes include PIR op YAML, compat, and `pd_op_to_kernel_pass` C++.
+- Paddle install: source checkout at `base_commit`, build/install so PIR dy2static paths are available.
+
+## Run Order (Run / Test / Fix)
+
+1. Check out `PaddlePaddle/Paddle` at the base commit and complete a source build/install.
+2. Apply `tests/test.patch`.
+3. Run `bash tests/test.sh`; PIR memcpy related cases should **fail / error** before the fix.
+4. Apply `solution/code.patch` and **rebuild** (YAML / C++ / pass changes).
+5. Run `bash tests/test.sh` again; all target cases should **pass**.
+
+## Minimal Test Command
+
+```bash
+bash tests/test.sh
+```
+
+## Known Risks
+
+- Historical PIR stack requires era-matched source build; newer trees may not apply patches cleanly.
+- GPU memcpy cases need a CUDA build; CPU-only environments should rely on the CPU test file.

swe-paddle/tasks/PaddlePaddle__Paddle-57741/instruction.md

@@ -0,0 +1,43 @@
+# 在 PIR 路径下补齐 memcpy 算子支持
+
+## 详细描述
+
+在动转静（`to_static`）场景中，Tensor 在不同 Place 之间拷贝（例如 CPU ↔ CUDA）依赖框架内部的 `memcpy` 能力。当前在 PIR 执行路径下，该能力缺失或不完整，导致相关拷贝用例无法在 PIR 下正确运行。
+
+需要达成的目标：
+
+- 在 PIR 路径下支持 `memcpy`，使动转静场景中的 Tensor 设备间拷贝可以完成。
+- 至少覆盖默认在 CPU 上的拷贝相关用例；若环境具备 CUDA，相关 GPU 用例也应可走通。
+- 拷贝后的 place 与数值结果应与非 PIR / 既有路径一致。
+- 不要求扩展到其他无关的设备管理或拷贝 API。
+
+## 问题复现（示意）
+
+在目标 `base_commit` 对应版本上：
+
+1. 构造简单动转静函数，对 Tensor 执行跨 Place 拷贝（例如拷到 CPU / CUDA）。
+2. 在 PIR 对比 / 开启路径下运行上述用例。
+3. 可观察到算子缺失、lowering 失败，或拷贝结果 / place 不符合预期。
+
+## 期望行为
+
+- PIR 下 `memcpy` 相关路径可运行。
+- 拷贝后的 place 与数值与既有非 PIR 行为一致（在任务覆盖范围内）。
+- 既有无关动转静行为不被破坏。
+
+## 验收说明
+
+- 上述目标行为可用，相关用例通过。
+- 已有无关动转静语义不被破坏。
+- 不允许通过删除测试、弱化断言或大范围绕过校验来「通过」任务。
+
+## 技术要求
+
+- 熟悉 Python 与 C++
+- 了解 Paddle PIR 算子声明、兼容配置与 kernel lowering 相关机制
+
+## Acceptance Criteria
+
+- The behavior described above should be implemented.
+- Existing valid behavior should remain unchanged.
+- Do not satisfy the task by deleting tests, weakening assertions, or bypassing validation broadly.

swe-paddle/tasks/PaddlePaddle__Paddle-57741/solution/code.patch

@@ -0,0 +1,73 @@
+diff --git a/paddle/fluid/pir/dialect/op_generator/ops_api_gen.py b/paddle/fluid/pir/dialect/op_generator/ops_api_gen.py
+--- a/paddle/fluid/pir/dialect/op_generator/ops_api_gen.py
++++ b/paddle/fluid/pir/dialect/op_generator/ops_api_gen.py
+@@ -101,6 +101,7 @@
+     'fused_bn_add_activation_',
+     'fused_feedforward',
+     'fused_scale_bias_relu_conv_bnstats',
++    'memcpy',
+     'print',
+     'recv_v2',
+     'rnn_',
+diff --git a/paddle/fluid/pir/dialect/operator/ir/ops.yaml b/paddle/fluid/pir/dialect/operator/ir/ops.yaml
+--- a/paddle/fluid/pir/dialect/operator/ir/ops.yaml
++++ b/paddle/fluid/pir/dialect/operator/ir/ops.yaml
+@@ -81,6 +81,16 @@
+   args : (Tensor[] x)
+   output : Tensor(out)
+
++- op : memcpy
++  args : (Tensor x, int dst_place_type)
++  output : Tensor(out)
++  infer_meta:
++    func: UnchangedInferMeta
++    param: [x]
++  kernel:
++    func : memcpy
++    param: [x, dst_place_type]
++
+ - op : print
+   args : (Tensor in, int first_n, str message, int summarize, bool print_tensor_name = true, bool print_tensor_type = true, bool print_tensor_shape = true, bool print_tensor_layout = true, bool print_tensor_lod = true, str print_phase = "BOTH", bool is_forward = true)
+   output : Tensor(out)
+diff --git a/paddle/fluid/pir/transforms/pd_op_to_kernel_pass.cc b/paddle/fluid/pir/transforms/pd_op_to_kernel_pass.cc
+--- a/paddle/fluid/pir/transforms/pd_op_to_kernel_pass.cc
++++ b/paddle/fluid/pir/transforms/pd_op_to_kernel_pass.cc
+@@ -678,6 +678,22 @@
+
+   phi::KernelKey res(kernel_backend, kernel_layout, kernel_data_type);
+
++  // kernel backend infered incorrectly from memcpy op operands,
++  // case that place from (not GPU) to GPU.
++  // We handle this special case by following code to fix up the problem.
++  // This could be further improved if we had another method.
++  if (!platform::is_gpu_place(place)) {
++    if (op->isa<paddle::dialect::MemcpyOp>()) {
++      VLOG(6) << "MemcpyOp need a special handle";
++      int dst_place_type = op->attribute("dst_place_type")
++                               .dyn_cast<pir::Int32Attribute>()
++                               .data();
++      if (dst_place_type == 1) {
++        res.set_backend(phi::Backend::GPU);
++      }
++    }
++  }
++
+   if (op->isa<paddle::dialect::LoadCombineOp>()) {
+     res.set_dtype(phi::DataType::FLOAT32);
+     VLOG(8) << "LoadCombineOp's kernel data type must be FLOAT32";
+diff --git a/paddle/phi/api/yaml/op_compat.yaml b/paddle/phi/api/yaml/op_compat.yaml
+--- a/paddle/phi/api/yaml/op_compat.yaml
++++ b/paddle/phi/api/yaml/op_compat.yaml
+@@ -3275,6 +3275,12 @@
+   attrs:
+     pivot : pivots
+
++- op: memcpy
++  inputs:
++    x: X
++  outputs:
++    out: Out
++
+ - op: memcpy_d2h
+   inputs :
+     x : X

swe-paddle/tasks/PaddlePaddle__Paddle-57741/tests/test.patch

@@ -0,0 +1,65 @@
+diff --git a/test/dygraph_to_static/test_tensor_memcpy_on_cpu.py b/test/dygraph_to_static/test_tensor_memcpy_on_cpu.py
+--- a/test/dygraph_to_static/test_tensor_memcpy_on_cpu.py
++++ b/test/dygraph_to_static/test_tensor_memcpy_on_cpu.py
+@@ -15,6 +15,7 @@
+ import unittest
+
+ import numpy as np
++from dygraph_to_static_util import test_and_compare_with_new_ir
+
+ import paddle
+
+@@ -47,6 +48,7 @@
+         x2 = tensor_copy_to_cpu(x1)
+         return x1.place, x2.place, x2.numpy()
+
++    @test_and_compare_with_new_ir(False)
+     def test_tensor_cpu_on_default_cpu(self):
+         paddle.base.framework._set_expected_place(paddle.CPUPlace())
+         dygraph_x1_place, dygraph_place, dygraph_res = self._run(
+@@ -67,6 +69,7 @@
+         x2 = tensor_copy_to_cuda(x1)
+         return x1.place, x2.place, x2.numpy()
+
++    @test_and_compare_with_new_ir(False)
+     def test_tensor_cuda_on_default_cpu(self):
+         if not paddle.base.is_compiled_with_cuda():
+             return
+diff --git a/test/dygraph_to_static/test_tensor_memcpy_on_gpu.py b/test/dygraph_to_static/test_tensor_memcpy_on_gpu.py
+--- a/test/dygraph_to_static/test_tensor_memcpy_on_gpu.py
++++ b/test/dygraph_to_static/test_tensor_memcpy_on_gpu.py
+@@ -16,6 +16,7 @@
+ import unittest
+
+ import numpy as np
++from dygraph_to_static_util import test_and_compare_with_new_ir
+
+ import paddle
+
+@@ -48,6 +49,7 @@
+         x2 = tensor_copy_to_cpu(x1)
+         return x1.place, x2.place, x2.numpy()
+
++    @test_and_compare_with_new_ir(False)
+     def test_tensor_cpu_on_default_gpu(self):
+         if paddle.base.is_compiled_with_cuda():
+             place = paddle.CUDAPlace(
+@@ -74,6 +76,7 @@
+         x2 = tensor_copy_to_cuda(x1)
+         return x1.place, x2.place, x2.numpy()
+
++    @test_and_compare_with_new_ir(False)
+     def test_tensor_cuda_on_default_gpu(self):
+         if paddle.base.is_compiled_with_cuda():
+             place = paddle.CUDAPlace(
+diff --git a/test/white_list/new_ir_op_test_white_list b/test/white_list/new_ir_op_test_white_list
+--- a/test/white_list/new_ir_op_test_white_list
++++ b/test/white_list/new_ir_op_test_white_list
+@@ -136,6 +136,7 @@
+ test_matrix_power_op
+ test_maxout_op
+ test_mean_op
++test_memcpy_op
+ test_mode_op
+ test_multi_dot_op
+ test_multiplex_op

swe-paddle/tasks/PaddlePaddle__Paddle-57741/tests/test.sh

@@ -0,0 +1,11 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# Target tests for PaddlePaddle__Paddle-57741.
+# Run from the root of a built PaddlePaddle/Paddle source checkout.
+python -m pytest \
+  test/dygraph_to_static/test_tensor_memcpy_on_cpu.py \
+  -q
+
+# Optional if CUDA is available:
+# python -m pytest test/dygraph_to_static/test_tensor_memcpy_on_gpu.py -q