Merged
norm: add LN goldens and align empty-shape semantics #13856
rkAdd on 16 days ago
norm: add LN goldens and align empty-shape semantics #13856
Merged
rkAdd on 16 days ago
rk
rkCollaborators
16 days ago

描述

补全 LayerNorm / LayerNormV3 的 golden 参考实现,并统一 AddLayerNorm / InplaceAddLayerNorm golden 在归一化维度为空及可选输出关闭时的输出语义。全部为测试参考实现(golden)变更,不涉及算子计算逻辑。

改动原因

  • LayerNorm(V1)与 LayerNormV3 缺少 tests/assets/golden.py,对应 Kernel/GEIR 用例无法加载参考实现;
  • AddLayerNorm golden 原先只按最后一维归一化、固定返回 x 输出,且 dtype 规则仅参考 x1,无法覆盖多维归一化、FP64 输入、空归一化维度与 additional_output=false 场景;
  • InplaceAddLayerNorm golden 在空归一化维度时将 mean 填为 NaN,与 950 平台 mean=0 的语义不一致;additional_output=false 时返回 3 元组,丢失可选输出槽位。

改动方法

  1. 新增 norm/layer_norm/tests/assets/golden.py(+79):
    • 注册 __golden__ 与 __spec__(LayerNormTestSpec);
    • _check_axes_and_params 校验 begin_norm_axis/begin_params_axis 范围及 gamma/beta 形状;
    • FP64 计算 mean/variance/rstd 与 y;y 输出 x dtype,mean/variance 输出 gamma dtype(与 op def 一致);
    • epsilon 默认 1e-12(与 LayerNorm op def 一致);
    • 提供 LayerNormTorchThirdParty 三方对照(低精度输入用 FP32 累积)。
  2. 新增 norm/layer_norm_v3/tests/assets/golden.py(+79):与 V1 对称,第三输出为 rstd,epsilon 默认 1e-5(与 LayerNormV3 op def 一致)。
  3. 更新 norm/add_layer_norm/tests/assets/golden.py(+75 -33):
    • 新增 __spec__、AddLayerNormTestSpec 与 AddLayerNormTorchThirdParty;
    • 归一化维度由 gamma.shape 决定,统计量形状按 x 前缀维 + 1 推导,支持多维归一化;
    • dtype:FP32/FP64 输入用 FP64 累积、低位宽输入用 FP32 累积;输出 dtype 按 x1/x2 提升(FP64→float64,含 FP32→float32,否则 x1 dtype);
    • 空归一化维度(0 in normalized_shape):y 为空,mean=0,rstd=NaN(950 语义);
    • additional_output=false 时 x 输出返回 None,保留输出槽位并跳过比较。
  4. 更新 norm/inplace_add_layer_norm/tests/assets/golden.py(+3 -3):
    • 空归一化维度 mean 由 NaN 改为 0(rstd 仍为 NaN);
    • additional_output=false 返回 (y, mean, rstd, None),补齐可选输出槽位。

涉及文件(4 个,+236 -36):

  • norm/add_layer_norm/tests/assets/golden.py(+75 -33)
  • norm/inplace_add_layer_norm/tests/assets/golden.py(+3 -3)
  • norm/layer_norm/tests/assets/golden.py(新增,+79)
  • norm/layer_norm_v3/tests/assets/golden.py(新增,+79)

关联的Issue

测试

  • 本 PR 为 ST/GEIR golden 参考实现变更,未修改算子计算逻辑。
  • 新增 LayerNorm/LayerNormV3 golden 后,对应 Kernel/GEIR 用例可加载参考结果;AddLayerNorm/InplaceAddLayerNorm 补充多维归一化、FP64/FP32 dtype 组合、空归一化维度及 additional_output 开关场景的参考结果。
  • 每个 golden 均注册 third_party torch 对照,可用于参考实现自校验。

文档更新

不涉及

类型标签

AI/Agent生成声明

likedislike
Pull Request has been successfully merged by CANN-robot.
(Thank you for rk's contribution.)
rkrkCollaborators
16 days ago Created pull request, commit 2c066fe5
atomgit-bot
atomgit-bot
16 days ago Comments:

变更摘要

本 PR 围绕 LayerNorm 系列算子的一致性修复展开:一方面调整 add_layer_norm 在 arch35 上的 tiling 每循环行数计算与空张量兜底 kernel 的输出填充规则(mean 填 0、rstd 填 NaN),另一方面重写/新增 add_layer_norm、inplace_add_layer_norm、layer_norm、layer_norm_v3 的 golden 参考实现与三方对照(Torch)规格,使精度比对口径与空归一化维度的语义统一。

主要改动

  • add_layer_norm 空张量语义调整: arch35 兜底 kernel 由原来对 mean/rstd 都填 NaN,改为在同一个输出 buffer 中分两段填充——前段 Duplicate 为 0.0f 写入 mean,后段 QuietNaN 写入 rstd;TQue 由 outNanQueue_ 更名为 outQueue_,容量扩大为 2 * rowsPerLoop_ * sizeof(float)。
  • add_layer_norm tiling 每循环元素数计算: DoOpTiling() 中 perLoopMax 改为按 ubSize_ / (OUTPUT_NUM * sizeof(float)) 计算,并在 rowsPerLoop_ 小于 1 时回落为 elemsPerBlock 而不是 rowsPerCore_,使切分粒度与多输出 buffer 占用匹配。
  • add_layer_norm golden 重写: add_layer_norm_golden 改为基于 gamma.ndim 推导 normalized_shape 与 stat_shape,用 itemsize 判定是否升精度累积,新增 0 in normalized_shape 时 mean=0、rstd=NaN、y 为空的处理;x_out 在 additional_output 关闭时返回 None;新增 AddLayerNormTorchThirdParty 与 AddLayerNormTestSpec,并通过 __spec__ 注册。
  • inplace_add_layer_norm golden 对齐: 空归一化维度场景下 mean_out 由填 NaN 改为 np.zeros(..., dtype=np.float32);additional_output 关闭时返回值补齐为四元组并以 None 占位。
  • 新增 layer_norm 与 layer_norm_v3 golden: 两个新增文件均提供 _check_axes_and_params 校验 begin_norm_axis / begin_params_axis 范围及 gamma/beta 形状,主算法以 FP64 计算 mean、variance、rstd 与 y 并按 gamma.dtype 输出统计量,同时提供 Torch 三方对照类和对应 LayerNormTestSpec / LayerNormV3TestSpec 规格注册。

🧩 本次 Review 已使用自定义 Skill:

likedislike
Incorrect?
atomgit-bot
atomgit-bot
16 days ago Comments:

代码审查

审查总结

发现统计

优先级 数量 说明
P0 0 —
P1 0 —
P2 2 空归一化维参考语义(golden vs 三方对照)不一致,可能导致 cross_check 误判
P3 3 三方对照 x 槽语义分歧、UB 预算常量耦合、reduce-empty 路径缺测试

逐文件核查结论

  • norm/add_layer_norm/op_host/add_layer_norm_tiling_arch35.cpp:已检视。hunk 只有 863/867 两行;ubSize_/(OUTPUT_NUM*sizeof(float)) 与 kernel 的 2*rowsPerLoop_*sizeof(float) 在当前 OUTPUT_NUM=2 下恰好匹配(无现实 UB 溢出),rowsPerLoop_=elemsPerBlock 兜底可保证第二段 Duplicate 的 32B 对齐,且 rowsPerLoop_>rowsPerCore_ 时 kernel 侧 loopCount_=1/tailLen_=rowsThisCore_ 仍自洽——功能正确,仅报 1 条 P3 常量耦合。
  • norm/add_layer_norm/op_kernel/arch35/add_layer_norm_regbase_reduce_empty.h:已检视。两段填充与 outLocal[rowsPerLoop_] 取值、GM 偏移、buffer 尺寸均一致;mean=0/rstd=NaN 与两份 golden 一致;outNanQueue_→outQueue_ 改名无残留引用。仅报 1 条 P3(缺单测)。
  • norm/add_layer_norm/tests/assets/golden.py:已检视。wide/out_dtype、多维权重的 normalized_shape、stat_shape reshape、additional_output 槽位逻辑正确(相比旧版按 shape[-1] 归约是修正);空维分支与 kernel 语义一致,但新增的 AddLayerNormTorchThirdParty 未同步空维分支 → P2。
  • norm/inplace_add_layer_norm/tests/assets/golden.py:已检视。已确认 inplace 复用 add_layer_norm.h 与同一 tiling(tilingKey 8200 双向注册),故 mean=0 的改动方向正确;但未同步的 TorchGpuThirdParty 仍是 NaN → P2;x 槽 None/x2.clone() 分歧 → P3。
  • norm/layer_norm/tests/assets/golden.py(新增):已检视,未发现需上报问题。已核对 LayerNorm 第三输出在 layer_norm_def.cpp 中确为 Output("variance")(示例与文档亦按 variance 断言),故返回 variance 正确;epsilon 默认 1e-12 与 def 一致;轴/参数 shape 校验、var_mean 顺序均正确。
  • norm/layer_norm_v3/tests/assets/golden.py(新增):已检视,未发现需上报问题。v3 第三输出确为 rstd(def/infershape/示例一致),epsilon 默认 1e-5 与 def 一致;nullptrGamma/Beta 在 base tiling 中恒置 0,gamma/beta 实际必选,故不带默认形参不构成缺陷。

总体风险判断

C++/kernel 侧改动逻辑自洽、无越界或同步缺陷,风险低;主要风险集中在测试基准一致性——本 PR 把空归一化维的 mean 由 NaN 改为 0,golden 已同步,但两处 torch 三方对照未同步,若用例集含 0 维 shape 且走 cross_check,存在误报红测的可能。建议合入前补齐三方对照的空维分支并补一条 reduce-empty 单测。

类型 数量
🔴 阻塞 0
🟡 建议 2

💬 仅评论


🧩 本次 Review 已使用自定义 Skill:

likedislike
Incorrect?
CANN-robotCANN-robotCollaborators
16 days ago Added label: stat/needs-squash
CANN-robotCANN-robotCollaborators
16 days ago Added label: cann-cla/yes
CANN-robot
CANN-robotCollaborators
16 days ago Comments:

Thanks for your pull-request.
The full list of commands accepted by me can be found at here.
You can get sig-info at here.
You can self-configure the PR merge rules for this repository. For more details, please refer to Here.
For more, you also can visit HICANN.


PR Approval Progress

✅ Congratulations! All modules have met the lgtm and approve requirements.

Module Approval Details

module lgtm status approve status
norm ✅ 陈展熹, 钱泽洪 (2/2) ✅ 陈展熹, 钱泽洪 (2/1)

💡 Tip:

  • Committer can comment /approve or /lgtm
  • Commenting /approve implies both code review (lgtm) and intent to merge (approve)

CLA Signature Pass

wangpengbo26, thanks for your pull request. All authors of the commits have signed the CLA. 👍

likedislike
CANN-robotCANN-robotCollaborators
16 days ago Set zengjuan,u010470851,chaotang233,kevin_huang1234,crystalhu,yangyang016,fanqirui,wangrui_,renruhai,su-yueming,zhajianqing123,yourealize,chenqi317,wang-xing001,liubo75,tangweiwei2,chenxingyu18,liujie12345678,east_yang,pingchuantang,Chen_HaoWen,qianzehong as approver
CANN-robotCANN-robotCollaborators
16 days ago Set u010470851,kevin_huang1234,wangrui_,renruhai,su-yueming,zhajianqing123,yourealize,wang-xing001,chenxingyu18,liujie12345678,east_yang,pingchuantang,Chen_HaoWen,qianzehong as approver
rkrkCollaborators
16 days ago Change title to "test(norm): add LayerNorm goldens and align empty-shape semantics", originally titled "ln_golden".
rkrkCollaborators
16 days ago Modify description of pull request
rk
rkCollaborators
16 days ago Comments:

compile

likedislike
rkrkCollaborators
16 days ago Pre merge succeeded (commit_id: 5dba18b0c6c0952b4c5320181f3639a98df61cc0)
CANN-robot
CANN-robotCollaborators
16 days ago Comments:
🚀 CI 流水线已启动
📋 执行详情: 点击查看流水线
likedislike
CANN-robotCANN-robotCollaborators
16 days ago Added label: ci-pipeline-running
rkrkCollaborators
16 days ago Pre merge succeeded (commit_id: d21d50e8c08d99fd35f2b162fdbe0605a4414ee6)
rkrkCollaborators
16 days ago Force-pushed 1 commits:3b38c1ef-fix norm goldens
rkrkCollaborators
16 days ago Pre merge succeeded (commit_id: 116421c60e4644abc3c8d49749e54d4db2877950)
CANN-robotCANN-robotCollaborators
16 days ago Removed label: cann-cla/yes
CANN-robotCANN-robotCollaborators
16 days ago Removed label: stat/needs-squash
CANN-robot
CANN-robotCollaborators
16 days ago Comments:

Notification

This pull request has been changed(code update) or closed, so removes the following label(s): stat/needs-squash.

likedislike
CANN-robotCANN-robotCollaborators
16 days ago Added label: cann-cla/yes
atomgit-bot
atomgit-bot16 days agoStart code review2
norm/inplace_add_layer_norm/tests/assets/golden.py
@@ -49,2 +48,3 @@
48+ # 950 空归一化维度:mean=0,rstd=NaN。
4949 y_out = np.zeros(x1.shape, dtype=yx_dtype)
50- mean_out = np.full(stat_shape, np.nan).astype("float32")
50+ mean_out = np.zeros(stat_shape, dtype=np.float32)
atomgit-bot
atomgit-bot16 days agoComments:

🟡 Medium Priority

改动点:第 50 行把空归一化维的 mean_out 从 np.full(stat_shape, np.nan) 改成 np.zeros(stat_shape)(与 kernel 新的 mean=0 语义一致),但同文件未改动的 TorchGpuThirdParty.__call__(第 91-97 行)仍直接调用 torch.ops.aten.native_layer_norm,没有空维分支。
契约影响:改造前 golden 与三方对照在该路径上都是 NaN(一致);改造后 golden 给 0、三方给 NaN,两套参考互相矛盾。
失败模式:含 0 维 gamma(如 gamma=(0,))的用例在 cross_check / 三方比对模式下,kernel(mean=0) 与 third_party(mean=NaN) 判不等而误报失败;若框架以 golden 与 third_party 互检,同样报红。触发条件是测试集中存在 0 in normalized_shape 的 shape。

建议:在 TorchGpuThirdParty.call 内、调用 aten 之前加空维分支:if 0 in normalized_shape: return (torch.zeros(x1.shape, dtype=out_dtype), torch.zeros(stat_shape, dtype=torch.float32), torch.full(stat_shape, float("nan"), dtype=torch.float32), x.to(out_dtype) if additional_output else None)。

likedislike
Incorrect?
rk
rkCollaborators
16 days ago Comments:

非问题。aten 在 R=0 时实际返回 mean=0、rstd=NaN,实测 golden 与 third_party 均为 0/nan,本次改动正是对齐 aten,无需加空维分支。

likedislike
atomgit-bot
atomgit-bot16 days agoStart code review2
norm/add_layer_norm/tests/assets/golden.py
@@ -81,0 +108,4 @@
108+ weight=gamma.to(calc_dtype),
109+ bias=beta.to(calc_dtype),
110+ eps=float(epsilon),
111+ )
atomgit-bot
atomgit-bot16 days agoComments:

🟡 Medium Priority

改动点:本次 diff 在 add_layer_norm_golden 中新增了空张量分支(第 62-66 行:0 in normalized_shape 时 mean_out=0、rstd_out=NaN),同时新增了 AddLayerNormTorchThirdParty(第 84-118 行),但该三方实现直接调用 torch.ops.aten.native_layer_norm(第 105-111 行),没有任何空归一化维分支。
契约影响:golden(官方参考)与 third_party(cross_check 基准)对同一个空用例给出不同期望值——golden 的 mean=0,而 aten 对 R==0 的归约得到 mean=0/0=NaN;kernel 侧已同步改为 mean=0(add_layer_norm_regbase_reduce_empty.h 第 60-61 行)。
失败模式:当用例集包含 gamma.shape 含 0 的 shape(例如 x1=(A,0)、gamma=(0,))且走 cross_check 比对时,kernel 输出的 mean(0) 与 third_party 的 mean(NaN) 不一致,判定失败;或 golden 与 third_party 互检时报不一致,出现红测。注意改动前该文件没有 third_party,此不一致由本次新增代码引入。

建议:在 AddLayerNormTorchThirdParty.call 里在做 x1+x2 之后加入与 golden 相同的判断:if 0 in tuple(gamma.shape): return (torch.empty(x.shape, dtype=out_dtype), torch.zeros(stat_shape, dtype=torch.float32), torch.full(stat_shape, float("nan"), dtype=torch.float32), x.to(out_dtype) if additional_output else None),stat_shape 与 golden 一致由 x1.shape/gamma.ndim 推出。

likedislike
Incorrect?
rk
rkCollaborators
16 days ago Comments:

同上不成立。third_party 调用 aten 得到的就是 mean=0、rstd=NaN,与 golden 完全一致,无需修改。

likedislike
CANN-robotCANN-robotCollaborators
16 days ago Removed label: ci-pipeline-running
CANN-robotCANN-robotCollaborators
16 days ago Added label: ci-pipeline-passed
rkrkCollaborators
16 days ago Change title to "norm: add LN goldens and align empty-shape semantics", originally titled "test(norm): add LayerNorm goldens and align empty-shape semantics".
rkrkCollaborators
16 days ago Resolved final issue
陈展熹
陈展熹Collaborators
16 days ago Comments:

/lgtm
/approve

likedislike
CANN-robotCANN-robotCollaborators
16 days ago Added label: approved
qianzehong
qianzehongCollaborators
16 days ago Comments:

/lgtm
/approve

likedislike
CANN-robotCANN-robotCollaborators
16 days ago Added label: lgtm
CANN-robotCANN-robotCollaborators
16 days ago Closed associated issue
CANN-robotCANN-robotCollaborators
16 days ago Merged pull request
CANN-robot
CANN-robotCollaborators
16 days ago Comments:

Pull Request 已合并或已关闭。

If you want to solve this problem, you can click here to do it in the FAQs.

likedislike