[Operator Mechanism] Add scatter_nd_add double gradient support

[leeleolay]

PR Category

Operator Mechanism

PR Types

Bug fixes

Description

Fixes #79493

Motivation

Molecular force models compute forces as coordinate gradients of energy and differentiate a force loss again during training. paddle.scatter_nd_add currently raises:

The Op scatter_nd_add_grad doesn't have any grad op.

when the first gradient is requested with create_graph=True.

Changes

• Register scatter_nd_add_double_grad in backward.yaml.
• Implement the double gradient as grad_x_grad + scatter_nd(index, grad_updates_grad).
• Handle either optional incoming second gradient.
• Add scatter_nd_add_grad to the PIR VJP blacklist because its double gradient is composite-only, consistent with index_put_grad and put_along_axis_grad.
• Add CPU dygraph coverage for the default create_graph=True path and all three optional-input combinations under eager prim.

Validation

• Full CPU paddle_python build completed successfully (1495/1495).
• python -m unittest test_scatter_nd_op.TestScatterNdAddDoubleGrad: 2 tests passed.
• Full test_scatter_nd_op.py: 29 tests passed, 4 device-specific tests skipped.
• Pre-commit passed for all four changed files.
• Duplicate-index second gradients are covered for only grad_x_grad, only grad_updates_grad, and both inputs.

是否引起精度变化

否

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[risemeup1111]

未发现需要阻塞合入的问题。关于 optional double-grad 分支的测试覆盖有一条非阻塞建议，详情见行级评论；Pre Commit 已通过，其余 CI 以最终结果为准。

^{Powered by Nyanpasu with gpt-5.5 xhigh, please check the suggestions carefully.}

[risemeup1111]

优先级：P2
处理要求：请针对该评论进行回复（同意并已修改请回复 Done，不同意请说明理由）。

当前用例在第一次 paddle.grad 中同时请求 [x, updates]，因此后续 double grad 会同时给 x_grad 和 updates_grad 传入 cotangent；但本 PR 的 YAML/实现新增了 grad_x_grad、grad_updates_grad 两个 optional 分支，且 #79493 的最小复现只请求 updates 的一阶梯度。建议补一个只请求单个一阶梯度输出的断言，避免 optional 输出映射或空分支回归时漏测。

可参考把计算封装成 helper 后增加类似用例：

updates_grad = paddle.grad(loss, updates, create_graph=True)[0]
updates_double_grad = paddle.grad(updates_grad.sum(), updates)[0]
np.testing.assert_allclose(
    updates_double_grad.numpy(),
    np.array([[4.0, 4.0], [4.0, 4.0], [2.0, 2.0]]),
)

若认为现有 [x, updates] 用例已能覆盖单输出可选分支，请说明依据。

[leeleolay]

Done. The test now covers only grad_x_grad, only grad_updates_grad, and both optional inputs with duplicate indices. I also added a separate default-eager assertion for the original create_graph=True reproduction. The composite second-gradient numerical checks run with eager prim enabled, consistent with other composite-only higher-gradient tests. Full CPU build and the complete test_scatter_nd_op.py pass.

[risemeup1111]

已确认当前提交已覆盖原建议：test_create_graph 保留默认 eager 下的 create_graph=True 复现路径，test_double_grad 在 eager prim 下分别覆盖 only grad_x_grad、only grad_updates_grad 和两者同时存在的分支，重复 index 的期望值也对齐。该线程我会标记为已解决。

[risemeup1111]

已复查当前提交，之前关于 optional 分支测试覆盖的建议已补齐，未发现新的阻塞问题。细节见对应的行级讨论。

^{Powered by Nyanpasu with gpt-5.5 xhigh, please check the suggestions carefully.}

[PaddlePaddle-bot]

🤖 Paddle-CI-Agent | pr_review | 2026-07-17 01:46:21

📋 Review 摘要

PR 概述：为 scatter_nd_add 注册并实现二阶梯度，补充动态图 create_graph/二阶梯度回归测试
变更范围：backward.yaml、composite double backward、PIR VJP 黑名单、legacy 单测
影响面 Tag：[Operator Mechanism] [Execute Infrastructure]

问题

未发现阻塞性问题。PR 规范问题在下面章节报,不要在这里重复

历史 Findings 修复情况

Finding	问题	状态
F1	历史指出测试只覆盖两个 optional 同时传入；当前新增 "x"/"updates"/"both" 三个子用例，已分别覆盖 grad_x_grad 与 grad_updates_grad 单独传入。	✅ 已修复

📝 PR 规范检查

标题缺少官方 Tag；描述结构符合 checklist §D2 模板。

标题建议（可直接复制）：

• [Operator Mechanism] Add scatter_nd_add double gradient support

总体评价

新增 double grad 公式与现有 scatter_nd_add_grad 的数学关系一致，测试覆盖了 duplicate index 的累加和两个 optional 梯度分支。本轮未拉取 CI，未进行编译/测试运行；完整编译和 CI 仍以流水线为准。

[leeleolay]

/re-run all-failed

[leeleolay]

The approval gate now passes the review-bot check and reports two remaining code-owner requirements. Could @wanghuancoder please review the backward.yaml / paddle/phi changes, and could @xiaoguoguo626807 please review the new composite registration?

Validation completed locally: full CPU paddle_python build 1495/1495, generated PIR VJP targets 80/80, complete test_scatter_nd_op.py 29 tests passed (4 skipped), and all pre-commit hooks passed. The new tests cover default eager create_graph=True plus only-x, only-updates, and combined optional double-gradient branches. Additional paddle/phi gate checks on the same build also pass: zero-size trailing dimensions (out, x double grad, and updates double grad all shape [2, 0]) and analytical precision checks for both float32 and float64. BigTensor does not exercise new code here because this PR changes no kernel or index arithmetic; the composite double gradient delegates to the existing scatter_nd_add implementation.

[leeleolay]

CI status note: the only non-successful job, Mac-CPU / Build and test, was cancelled by the workflow's 60-minute job timeout. The macOS build completed successfully, and 357/1952 tests had passed with no test failure when GitHub emitted 'The operation was canceled' exactly at the timeout. Coverage, CI-Windows, CI-Build, and CI-H all passed. Please rerun the Mac job if a green aggregate CI result is required.

[PaddlePaddle-bot]

🤖 Paddle-CI-Agent | ci_status_monitor | 2026-07-17 10:46:26 UTC+08:00

CI报告基于以下代码生成（30分钟更新一次）:
PR commit: 414b60f | Merge base: f92f172 (branch: develop)

---

1 Required任务 : 47/49 通过

总执行（rerun次数）	总任务	✅ 通过	❌ 失败	⏳ 运行中	⏸️ 等待中	跳过
88(4)	84	82	1	0	0	0

任务	错误类型	置信度	日志
Check approval	需要 Approval	高	Job

2 失败详情

🔴 Check approval — 需要 Approval（置信度: 高）

该 Job 需要人工 Approval，完成审批后 CI 才会继续执行。