[SWE-Paddle] Add task PaddlePaddle__Paddle-56705

[Echo-Nie]

关联

• SWE-Paddle 共建 issue: [Call for Bench] SWE-Paddle 社区共建 Paddle#79447
• 来源 PR: [BugFix]Fix memory leak in mplayers Paddle#56705

说明

• 新增 SWE-Paddle 任务 PaddlePaddle__Paddle-56705

@sunzhongkai588

验证结果

状态	Communication contracts P2P	_c_identity PyLayer reuse F2P	_mp_allreduce PyLayer reuse F2P
Base + tests/test.patch	PASS	FAIL，符合预期	FAIL，符合预期
Base + test patch + solution/code.patch	PASS	PASS	PASS

Base P2P：Existing communication contracts

test_c_identity_forward_backward_contract:
.                                                                        [100%]
1 passed in 0.03s
Base c_identity communication-contract P2P exit code: 0

test_mp_allreduce_forward_backward_contract:
.                                                                        [100%]
1 passed in 0.02s
Base mp_allreduce communication-contract P2P exit code: 0

Base F2P：_c_identity PyLayer class reuse

F                                                                        [100%]
================================== FAILURES ===================================
______________ test_c_identity_reuses_pylayer_class_across_calls ______________

    def test_c_identity_reuses_pylayer_class_across_calls(monkeypatch):
        namespace, _ = _load_checkout_overlay(monkeypatch)
        group = FakeGroup(group_id=37)

        for _ in range(4):
            namespace["_c_identity"](object(), group=group)

        dispatched_classes = [item[0] for item in FakePyLayer.invocations]
>       assert len(set(dispatched_classes)) == 1, (
            "Repeated _c_identity calls created distinct PyLayer runtime types; "
            "this deterministic type growth is a proxy for the reported long-run "
            "Python-side memory accumulation."
        )
E       AssertionError: Repeated _c_identity calls created distinct PyLayer runtime types; this deterministic type growth is a proxy for the reported long-run Python-side memory accumulation.
E       assert 4 == 1
E        +  where 4 = len({<class 'c_identity_eager'>, <class 'c_identity_eager'>, <class 'c_identity_eager'>, <class 'c_identity_eager'>})

FAILED test/swe_paddle/test_pr56705_mp_ops_pylayer_lifecycle.py::test_c_identity_reuses_pylayer_class_across_calls
1 failed in 0.16s
Base c_identity PyLayer-reuse F2P exit code: 1

Base F2P：_mp_allreduce PyLayer class reuse

F                                                                        [100%]
================================== FAILURES ===================================
_____________ test_mp_allreduce_reuses_pylayer_class_across_calls _____________

    def test_mp_allreduce_reuses_pylayer_class_across_calls(monkeypatch):
        namespace, _ = _load_checkout_overlay(monkeypatch)
        group = FakeGroup(group_id=41)

        for _ in range(4):
            namespace["_mp_allreduce"](
                object(),
                op=ReduceOp.SUM,
                group=group,
                use_calc_stream=True,
                use_model_parallel=True,
            )

        dispatched_classes = [item[0] for item in FakePyLayer.invocations]
>       assert len(set(dispatched_classes)) == 1, (
            "Repeated _mp_allreduce calls created distinct PyLayer runtime types; "
            "this deterministic type growth is a proxy for the reported long-run "
            "Python-side memory accumulation."
        )
E       AssertionError: Repeated _mp_allreduce calls created distinct PyLayer runtime types; this deterministic type growth is a proxy for the reported long-run Python-side memory accumulation.
E       assert 4 == 1
E        +  where 4 = len({<class 'mp_allreduce_eager'>, <class 'mp_allreduce_eager'>, <class 'mp_allreduce_eager'>, <class 'mp_allreduce_eager'>})

FAILED test/swe_paddle/test_pr56705_mp_ops_pylayer_lifecycle.py::test_mp_allreduce_reuses_pylayer_class_across_calls
1 failed in 0.19s
Base mp_allreduce PyLayer-reuse F2P exit code: 1

Solution 测试

Solution c_identity communication-contract P2P:
1 passed in 0.02s

Solution mp_allreduce communication-contract P2P:
1 passed in 0.02s

Solution c_identity PyLayer-reuse F2P:
1 passed in 0.02s

Solution mp_allreduce PyLayer-reuse F2P:
1 passed in 0.02s

tests/test.sh

....                                                                     [100%]
4 passed in 0.04s
Solution tests/test.sh exit code: 0

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请检查PR提交格式和内容是否完备，具体请参考示例和模版。
Your PR has been submitted. Thanks for your contribution!
Please check its format and content. For this, you can refer to Template and Demo.