Star886
1788
代码
代码
Issues234
Pull Requests322
Actions
项目讨论
Wiki
分析
Star886
1788
已合并
BNInfer,InplaceApplyRMSProp以及ApplyCamePart1算子检视意见闭环 #9453
tianyu52创建于 1 天前
BNInfer,InplaceApplyRMSProp以及ApplyCamePart1算子检视意见闭环 #9453
已合并
tianyu52创建于 1 天前
从tianyu52/ops-nnmaster合入到cann/ops-nnmaster
共 30 个文件变更+295-279
Mnorm/bn_infer/CMakeLists.txt+1-1
@@ -1,5 +1,5 @@
1# This program is free software, you can redistribute it and/or modify.1# This program is free software, you can redistribute it and/or modify.
2-# Copyright (c) 2025 Huawei Technologies Co., Ltd.2+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
3# This file is a part of the CANN Open Software.3# This file is a part of the CANN Open Software.
4# Licensed under CANN Open Software License Agreement Version 2.0 (the "License").4# Licensed under CANN Open Software License Agreement Version 2.0 (the "License").
5# Please refer to the License for details. You may not use this file except in compliance with the License.5# Please refer to the License for details. You may not use this file except in compliance with the License.
Mnorm/bn_infer/README.md+29-21
@@ -13,7 +13,7 @@
13 13 
14## 功能说明14## 功能说明
15 15 
16-- 算子功能：推理场景下对输入张量x执行批量归一化，使用已给定的均值mean和方差variance计算输出y。16+- 算子功能：在推理场景下，使用给定的均值`mean`和方差`variance`对输入`x`进行批归一化，得到输出`y`。
17 17 
18- 计算公式：18- 计算公式：
19 19 
@@ -21,7 +21,7 @@
21 y = scale \times \frac{x - mean}{\sqrt{variance + epsilon}} + offset21 y = scale \times \frac{x - mean}{\sqrt{variance + epsilon}} + offset
22 $$22 $$
23 23 
24- scale、offset、mean、variance均为一维张量，长度等于输入x的通道维。24+ `scale`、`offset`、`mean`和`variance`均为一维张量，长度等于输入`x`的通道数。
25 25 
26## 参数说明26## 参数说明
27 27 
@@ -44,62 +44,70 @@
44 <tr>44 <tr>
45 <td>x</td>45 <td>x</td>
46 <td>输入</td>46 <td>输入</td>
47- <td>被归一化的输入张量。A2/A3兼容声明包含NC1HWC0、NULL及其未知形状映射；Ascend950物理执行路径按950实现收窄为ND、NCHW、NCDHW、NHWC和NDHWC，未提供NC1HWC0和NULL路径。ND格式下通道维为第1维，NCHW/NCDHW格式下通道维为C维，NHWC/NDHWC格式下通道维为最后一维。</td>47+ <td>待归一化的输入张量。不同数据格式对应的通道维见约束说明。</td>
48 <td>FLOAT、FLOAT16、BFLOAT16</td>48 <td>FLOAT、FLOAT16、BFLOAT16</td>
49- <td>ND、NCHW、NCDHW、NHWC、NDHWC</td>49+ <td>NC1HWC0、NULL、ND、NCHW、NCDHW、NHWC、NDHWC</td>
C
Cchenjiao22 小时前

NULL数据格式？有这种format?

likedislike
50 </tr>50 </tr>
51 <tr>51 <tr>
52 <td>scale</td>52 <td>scale</td>
53 <td>输入</td>53 <td>输入</td>
54- <td>缩放参数，一维张量，shape为(C)。</td>54+ <td>缩放参数，一维张量，shape为[C]。</td>
55 <td>FLOAT</td>55 <td>FLOAT</td>
56- <td>ND</td>56+ <td>ND、NC1HWC0、NULL、NCDHW</td>
57 </tr>57 </tr>
58 <tr>58 <tr>
59 <td>offset</td>59 <td>offset</td>
60 <td>输入</td>60 <td>输入</td>
61- <td>偏置参数，一维张量，shape为(C)。</td>61+ <td>偏置参数，一维张量，shape为[C]。</td>
62 <td>FLOAT</td>62 <td>FLOAT</td>
63- <td>ND</td>63+ <td>ND、NC1HWC0、NULL、NCDHW</td>
64 </tr>64 </tr>
65 <tr>65 <tr>
66 <td>mean</td>66 <td>mean</td>
67 <td>输入</td>67 <td>输入</td>
68- <td>推理使用的均值，一维张量，shape为(C)。</td>68+ <td>推理时使用的均值，一维张量，shape为[C]。</td>
69 <td>FLOAT</td>69 <td>FLOAT</td>
70- <td>ND</td>70+ <td>ND、NC1HWC0、NULL、NCDHW</td>
71 </tr>71 </tr>
72 <tr>72 <tr>
73 <td>variance</td>73 <td>variance</td>
74 <td>输入</td>74 <td>输入</td>
75- <td>推理使用的方差，一维张量，shape为(C)。</td>75+ <td>推理时使用的方差，一维张量，shape为[C]。</td>
76 <td>FLOAT</td>76 <td>FLOAT</td>
77- <td>ND</td>77+ <td>ND、NC1HWC0、NULL、NCDHW</td>
78 </tr>78 </tr>
79 <tr>79 <tr>
80 <td>epsilon</td>80 <td>epsilon</td>
81 <td>必选属性</td>81 <td>必选属性</td>
82- <td>添加到方差中的小值，用于避免除0。</td>82+ <td>添加到方差中的数值稳定性常数，用于避免除零。</td>
83 <td>FLOAT</td>83 <td>FLOAT</td>
84 <td>-</td>84 <td>-</td>
85 </tr>85 </tr>
86 <tr>86 <tr>
87 <td>y</td>87 <td>y</td>
88 <td>输出</td>88 <td>输出</td>
89- <td>归一化后的输出张量，声明格式、数据类型和shape与x一致；Ascend950物理执行格式与x相同。</td>89+ <td>归一化后的输出张量，数据类型、数据格式和shape与x一致。</td>
90 <td>FLOAT、FLOAT16、BFLOAT16</td>90 <td>FLOAT、FLOAT16、BFLOAT16</td>
91- <td>ND、NCHW、NCDHW、NHWC、NDHWC</td>91+ <td>NC1HWC0、NULL、ND、NCHW、NCDHW、NHWC、NDHWC</td>
92 </tr>92 </tr>
93 </tbody></table>93 </tbody></table>
94 94 
95## 约束说明95## 约束说明
96 96 
97-- scale、offset、mean、variance的数据类型必须为FLOAT。97+- scale、offset、mean和variance的数据类型必须为FLOAT，且均为shape为[C]的一维张量，其中C为x的通道数。
98-- scale、offset、mean、variance必须为一维张量，长度必须等于x的通道维大小。98+- <term>Ascend 950PR/Ascend 950DT</term>：
99-- ND格式下x维度数必须大于等于2；NCHW和NHWC格式下x必须为4维；NCDHW和NDHWC格式下x必须为5维。99+ - x和y支持ND、NCHW、NCDHW、NHWC和NDHWC格式，不支持NC1HWC0和NULL格式。
100-- Ascend950物理执行路径不支持x的任一参与维度为0的空Tensor；host tiling返回`GRAPH_FAILED`，不会进入kernel。100+ - scale、offset、mean和variance仅支持ND格式。
101-- A2/A3原型契约保留NC1HWC0和NULL；Ascend950 AscendC物理执行面按950实现收窄为ND、NCHW、NCDHW、NHWC和NDHWC。由于950 Vector kernel/tiling未实现NC1HWC0存储布局，且已知形状NULL无法映射到950物理路径，host tiling在kernel前结构化拒绝这两类格式；该差异通过950独立OpDef/config和tiling隔离，不修改A2/A3实现。101+ - ND格式下，x的rank不小于2，通道维为第1维；NCHW和NCDHW格式下，通道维为C维；NHWC和NDHWC格式下，通道维为最后一维。
102-- 本算子为GE图内部算子，不提供公开aclnnBNInfer接口；aclnn/torch单算子接口不涉及。102+ - NCHW和NHWC格式下，x必须为4维；NCDHW和NDHWC格式下，x必须为5维。
103+ - 不支持空Tensor，x的所有维度大小均必须大于0。
104+- <term>Atlas A3 训练系列产品/Atlas A3 推理系列产品</term>、<term>Atlas A2 训练系列产品/Atlas A2 推理系列产品</term>：
105+ - x和y支持NC1HWC0、NULL和NCDHW格式；动态shape场景下，NULL格式支持NCHW和NHWC映射。
106+ - scale、offset、mean和variance支持NC1HWC0、NULL和NCDHW格式。
107+ - x和y不支持ND和NDHWC格式。
108+- <term>Atlas 200I/500 A2 推理产品</term>、<term>Atlas 推理系列产品</term>、<term>Atlas 训练系列产品</term>：
109+ - x和y不支持BFLOAT16。
110+- 本算子支持GE图模式和TensorFlow Parser调用，不提供公开的aclnn接口。
103 111 
104## 调用说明112## 调用说明
105 113 
Mnorm/bn_infer/op_graph/CMakeLists.txt+1-1
@@ -1,5 +1,5 @@
1# This program is free software, you can redistribute it and/or modify.1# This program is free software, you can redistribute it and/or modify.
2-# Copyright (c) 2025 Huawei Technologies Co., Ltd.2+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
3# This file is a part of the CANN Open Software.3# This file is a part of the CANN Open Software.
4# Licensed under CANN Open Software License Agreement Version 2.0 (the "License").4# Licensed under CANN Open Software License Agreement Version 2.0 (the "License").
5# Please refer to the License for details. You may not use this file except in compliance with the License.5# Please refer to the License for details. You may not use this file except in compliance with the License.
Mnorm/bn_infer/op_graph/bn_infer_proto.h+5-4
@@ -18,18 +18,19 @@ namespace ge {
18 * @brief Performs batch normalization inference.18 * @brief Performs batch normalization inference.
19 *19 *
20 * @par Inputs:20 * @par Inputs:
21- * @li x: A 4D tensor of type float16 or float32 or bfloat16, with format NHWC or NCHW.21+ * @li x: A tensor of type float16 or float32 or bfloat16, with format NC1HWC0, NULL, ND, NCHW, NCDHW,
22+ * NHWC or NDHWC.
22 * @li scale: A 1D tensor of type float32, for the scale factor, the shape is same as dim C of input x. \n23 * @li scale: A 1D tensor of type float32, for the scale factor, the shape is same as dim C of input x. \n
23 * @li offset: A 1D tensor of type float32, for the offset, the shape is same as dim C of input x. \n24 * @li offset: A 1D tensor of type float32, for the offset, the shape is same as dim C of input x. \n
24 * @li mean: A 1D tensor of type float32, for the mean, the shape is same as dim C of input x. \n25 * @li mean: A 1D tensor of type float32, for the mean, the shape is same as dim C of input x. \n
25 * @li variance: A 1D tensor of type float32, for the variance, the shape is same as dim C of input x. \n26 * @li variance: A 1D tensor of type float32, for the variance, the shape is same as dim C of input x. \n
26 *27 *
27 * @par Attributes:28 * @par Attributes:
28- * epsilon: An optional float32, specifying the small value added to variance to29+ * epsilon: A required float32 attribute, specifying the small value added to variance to
29- * avoid dividing by zero. Defaults to "0.0001" . \n30+ * avoid dividing by zero. \n
30 *31 *
31 * @par Outputs:32 * @par Outputs:
32- * y: A 4D tensor of type float16 or float32 or bfloat16 for the normalized "x", with format NHWC or NCHW. \n33+ * y: A tensor of type float16 or float32 or bfloat16 for the normalized "x", with the same format as x. \n
33 *34 *
34 * @attention Constraints:35 * @attention Constraints:
35 * For Atlas 200/300/500 Inference Product, the result accuracy fails to reach 1/1000 due to the36 * For Atlas 200/300/500 Inference Product, the result accuracy fails to reach 1/1000 due to the
Mnorm/bn_infer/op_host/CMakeLists.txt+1-1
@@ -1,5 +1,5 @@
1# This program is free software, you can redistribute it and/or modify.1# This program is free software, you can redistribute it and/or modify.
2-# Copyright (c) 2025 Huawei Technologies Co., Ltd.2+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
3# This file is a part of the CANN Open Software.3# This file is a part of the CANN Open Software.
4# Licensed under CANN Open Software License Agreement Version 2.0 (the "License").4# Licensed under CANN Open Software License Agreement Version 2.0 (the "License").
5# Please refer to the License for details. You may not use this file except in compliance with the License.5# Please refer to the License for details. You may not use this file except in compliance with the License.
Mnorm/bn_infer/op_host/arch35/bn_infer_tiling.cpp+1-1
@@ -1,5 +1,5 @@
1/**1/**
2- * Copyright (c) 2025 Huawei Technologies Co., Ltd.2+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3 * This program is free software, you can redistribute it and/or modify it under the terms and conditions of3 * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4 * CANN Open Software License Agreement Version 2.0 (the "License").4 * CANN Open Software License Agreement Version 2.0 (the "License").
5 * Please refer to the License for details. You may not use this file except in compliance with the License.5 * Please refer to the License for details. You may not use this file except in compliance with the License.
Mnorm/bn_infer/op_host/arch35/bn_infer_tiling.h+1-1
@@ -1,5 +1,5 @@
1/**1/**
2- * Copyright (c) 2025 Huawei Technologies Co., Ltd.2+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3 * This program is free software, you can redistribute it and/or modify it under the terms and conditions of3 * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4 * CANN Open Software License Agreement Version 2.0 (the "License").4 * CANN Open Software License Agreement Version 2.0 (the "License").
5 * Please refer to the License for details. You may not use this file except in compliance with the License.5 * Please refer to the License for details. You may not use this file except in compliance with the License.
Mnorm/bn_infer/op_host/arch35/bn_infer_tiling_infer_arch35.cpp+1-1
@@ -1,5 +1,5 @@
1/**1/**
2- * Copyright (c) 2025 Huawei Technologies Co., Ltd.2+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3 * This program is free software, you can redistribute it and/or modify it under the terms and conditions of3 * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4 * CANN Open Software License Agreement Version 2.0 (the "License").4 * CANN Open Software License Agreement Version 2.0 (the "License").
5 * Please refer to the License for details. You may not use this file except in compliance with the License.5 * Please refer to the License for details. You may not use this file except in compliance with the License.
Mnorm/bn_infer/op_host/arch35/bn_infer_tiling_infer_last_channel_arch35.cpp+8-1
@@ -1,5 +1,5 @@
1/**1/**
2- * Copyright (c) 2025 Huawei Technologies Co., Ltd.2+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3 * This program is free software, you can redistribute it and/or modify it under the terms and conditions of3 * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4 * CANN Open Software License Agreement Version 2.0 (the "License").4 * CANN Open Software License Agreement Version 2.0 (the "License").
5 * Please refer to the License for details. You may not use this file except in compliance with the License.5 * Please refer to the License for details. You may not use this file except in compliance with the License.
@@ -443,6 +443,13 @@ ge::graphStatus BNInferLastChannelTiling::FillLastChannelTilingForBSplit(int64_t
443 bInner > 1) {443 bInner > 1) {
444 bInner--;444 bInner--;
445 }445 }
446+ const int64_t requiredUbBytes = paramBytes + cacheBytes +
447+ bInner * fusedALen * INPUT_OUTPUT_NUM * DOUBLE_BUFFER * bytesPerElement;
448+ OP_CHECK_IF(requiredUbBytes > static_cast<int64_t>(aicoreParams_.ubSize),
449+ OP_LOGE_FOR_INVALID_VALUES_WITH_REASON(context_->GetNodeName(), "requiredUbBytes",
450+ std::to_string(requiredUbBytes).c_str(),
451+ "last-channel tiling exceeds available UB"),
452+ return ge::GRAPH_FAILED);
446 int64_t bOuter = Ops::Base::CeilDiv(fusedBLen, bInner);453 int64_t bOuter = Ops::Base::CeilDiv(fusedBLen, bInner);
447 int64_t bTail = fusedBLen % bInner;454 int64_t bTail = fusedBLen % bInner;
448 int64_t tileBlockBTail = bTail == 0 ? bInner : bTail;455 int64_t tileBlockBTail = bTail == 0 ? bInner : bTail;
Mnorm/bn_infer/op_kernel/arch35/bn_infer_infer.h+1-1
@@ -1,5 +1,5 @@
1/**1/**
2- * Copyright (c) 2025 Huawei Technologies Co., Ltd.2+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3 * This program is free software, you can redistribute it and/or modify it under the terms and conditions of3 * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4 * CANN Open Software License Agreement Version 2.0 (the "License").4 * CANN Open Software License Agreement Version 2.0 (the "License").
5 * Please refer to the License for details. You may not use this file except in compliance with the License.5 * Please refer to the License for details. You may not use this file except in compliance with the License.
Mnorm/bn_infer/op_kernel/arch35/bn_infer_infer_last_channel.h+1-1
@@ -1,5 +1,5 @@
1/**1/**
2- * Copyright (c) 2025 Huawei Technologies Co., Ltd.2+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3 * This program is free software, you can redistribute it and/or modify it under the terms and conditions of3 * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4 * CANN Open Software License Agreement Version 2.0 (the "License").4 * CANN Open Software License Agreement Version 2.0 (the "License").
5 * Please refer to the License for details. You may not use this file except in compliance with the License.5 * Please refer to the License for details. You may not use this file except in compliance with the License.
Mnorm/bn_infer/op_kernel/arch35/bn_infer_infer_last_channel_continuous_a.h+1-1
@@ -1,5 +1,5 @@
1/**1/**
2- * Copyright (c) 2025 Huawei Technologies Co., Ltd.2+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3 * This program is free software, you can redistribute it and/or modify it under the terms and conditions of3 * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4 * CANN Open Software License Agreement Version 2.0 (the "License").4 * CANN Open Software License Agreement Version 2.0 (the "License").
5 * Please refer to the License for details. You may not use this file except in compliance with the License.5 * Please refer to the License for details. You may not use this file except in compliance with the License.
Mnorm/bn_infer/op_kernel/arch35/bn_infer_infer_last_channel_small_a.h+1-1
@@ -1,5 +1,5 @@
1/**1/**
2- * Copyright (c) 2025 Huawei Technologies Co., Ltd.2+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3 * This program is free software, you can redistribute it and/or modify it under the terms and conditions of3 * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4 * CANN Open Software License Agreement Version 2.0 (the "License").4 * CANN Open Software License Agreement Version 2.0 (the "License").
5 * Please refer to the License for details. You may not use this file except in compliance with the License.5 * Please refer to the License for details. You may not use this file except in compliance with the License.
Mnorm/bn_infer/op_kernel/arch35/bn_infer_infer_small_ab1.h+1-1
@@ -1,5 +1,5 @@
1/**1/**
2- * Copyright (c) 2025 Huawei Technologies Co., Ltd.2+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3 * This program is free software, you can redistribute it and/or modify it under the terms and conditions of3 * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4 * CANN Open Software License Agreement Version 2.0 (the "License").4 * CANN Open Software License Agreement Version 2.0 (the "License").
5 * Please refer to the License for details. You may not use this file except in compliance with the License.5 * Please refer to the License for details. You may not use this file except in compliance with the License.
Mnorm/bn_infer/tests/assets/golden.py+3-3
@@ -25,9 +25,9 @@ __spec__ = {
25}25}
26 26 
27_TOL = {27_TOL = {
28- "float32": {"standard": "stat_rel_err", "threshold": 1e-2},28+ "float32": {"standard": "cross_check", "threshold": 1e-2},
29- "float16": {"standard": "stat_rel_err", "threshold": 2e-2},29+ "float16": {"standard": "cross_check", "threshold": 2e-2},
30- "bfloat16": {"standard": "stat_rel_err", "threshold": 4e-2},30+ "bfloat16": {"standard": "cross_check", "threshold": 4e-2},
31}31}
32 32 
33 33 
Mnorm/bn_infer/tests/ut/op_host/arch35/test_bn_infer_tiling.cpp+6-0
@@ -209,6 +209,12 @@ TEST_F(BNInferTilingTest, nhwcFloatContinuousATilingKey901000)
209 EXPECT_EQ(tilingKey, 901000);209 EXPECT_EQ(tilingKey, 901000);
210}210}
211 211 
212+TEST_F(BNInferTilingTest, rejectLastChannelWhenChannelCacheExceedsUb)
213+{
214+ auto status = RunBNInferTiling({1, 1, 1, 4096}, ge::DT_FLOAT, ge::FORMAT_NHWC);
215+ EXPECT_EQ(status, ge::GRAPH_FAILED);
216+}
217+ 
212TEST_F(BNInferTilingTest, nhwcSmallATilingKey902000)218TEST_F(BNInferTilingTest, nhwcSmallATilingKey902000)
213{219{
214 uint64_t tilingKey = 0;220 uint64_t tilingKey = 0;
Mnorm/bn_infer/tests/ut/op_host/test_bn_infer_infershape.cpp+4-0
@@ -73,6 +73,10 @@ TEST_F(BNInferInferShapeTest, staticShape) { CheckInferShape({2, 3, 4}); }
73 73 
74TEST_F(BNInferInferShapeTest, dynamicShape) { CheckInferShape({2, -1, 16}); }74TEST_F(BNInferInferShapeTest, dynamicShape) { CheckInferShape({2, -1, 16}); }
75 75 
76+TEST_F(BNInferInferShapeTest, nchwLikeShape) { CheckInferShape({2, 3, 4, 5}); }
77+ 
78+TEST_F(BNInferInferShapeTest, ndhwcLikeShape) { CheckInferShape({2, 3, 4, 5, 6}); }
79+ 
76TEST_F(BNInferInferShapeTest, unknownRank) { CheckInferShape({-2}); }80TEST_F(BNInferInferShapeTest, unknownRank) { CheckInferShape({-2}); }
77 81 
78void CheckInferDataType(ge::DataType xDataType)82void CheckInferDataType(ge::DataType xDataType)
Moptim/apply_came_part1/README.md+53-42
@@ -13,26 +13,28 @@
13 13 
14## 功能说明14## 功能说明
15 15 
16-- **算子功能**：计算CAME（Confidence-guided Adaptive Memory Efficient）优化器第一阶段的平方梯度归约值。16+- 算子功能：计算CAME（Confidence-guided Adaptive Memory Efficient）优化器第一阶段的平方梯度归约值。
17 17 
18-- **计算公式**：18+- 计算公式：
19 19 
20- 对二维输入张量 `grad` 中的元素 $g_{i,j}$，以及标量张量 `eps`，先计算：20+ 对输入`grad`最后两维$[N, M]$上的元素$g_{b,\dots,i,j}$，先计算：
21 21 
22 $$22 $$
23- x_{i,j} = g_{i,j}^{2} + eps23+ x_{b,\dots,i,j} = g_{b,\dots,i,j}^{2} + eps
24 $$24 $$
25 25 
26- 再分别进行行、列和全局归约：26+ 再对每个批次切片分别进行行、列和二维归约：
27 27 
28 $$28 $$
29 \begin{aligned}29 \begin{aligned}
30- sum\_grad\_r_i &= \sum_j x_{i,j} \\30+ sum\_grad\_r_{b,\dots,i} &= \sum_j x_{b,\dots,i,j} \\
31- sum\_grad\_c_j &= \sum_i x_{i,j} \\31+ sum\_grad\_c_{b,\dots,j} &= \sum_i x_{b,\dots,i,j} \\
32- sum\_grad\_rc &= \sum_i \sum_j x_{i,j}32+ sum\_grad\_rc_{b,\dots} &= \sum_i \sum_j x_{b,\dots,i,j}
33 \end{aligned}33 \end{aligned}
34 $$34 $$
35 35 
36+ 其中，$b,\dots$表示批次维度索引，$i$和$j$分别表示最后两维的索引。
37+ 
36## 参数说明38## 参数说明
37 39 
38<table style="table-layout: fixed; width: 100%">40<table style="table-layout: fixed; width: 100%">
@@ -53,49 +55,58 @@
53 </tr>55 </tr>
54</thead>56</thead>
55<tbody>57<tbody>
56- <tr>58+ <tr>
57- <td>grad</td>59+ <td>grad</td>
58- <td>输入</td>60+ <td>输入</td>
59- <td>梯度张量，公式中的<code>grad</code>，rank不小于2；最后两维为<code>[N, M]</code>，前导维为batch维。</td>61+ <td>梯度张量，最后两维为<code>[N, M]</code>，具体shape要求见约束说明。</td>
60- <td>FLOAT16、FLOAT32、BFLOAT16</td>62+ <td>FLOAT16、FLOAT、BFLOAT16</td>
61- <td>ND</td>63+ <td>ND</td>
62- </tr>64+ </tr>
63- <tr>65+ <tr>
64- <td>eps</td>66+ <td>eps</td>
65- <td>输入</td>67+ <td>输入</td>
66- <td>数值稳定项，公式中的<code>eps</code>，为标量或单元素一维张量。</td>68+ <td>数值稳定性常数，具体shape要求见约束说明。</td>
67- <td>FLOAT32</td>69+ <td>FLOAT</td>
68- <td>ND</td>70+ <td>ND</td>
69- </tr>71+ </tr>
70- <tr>72+ <tr>
71- <td>sum_grad_r</td>73+ <td>sum_grad_r</td>
72- <td>输出</td>74+ <td>输出</td>
73- <td>按列归约后的行结果，shape为<code>[batch..., N]</code>，对应公式中的<code>sum_grad_r</code>。</td>75+ <td>对最后一维进行归约得到的结果，具体shape见约束说明。</td>
74- <td>FLOAT32</td>76+ <td>FLOAT</td>
75- <td>ND</td>77+ <td>ND</td>
76- </tr>78+ </tr>
77 <tr>79 <tr>
78 <td>sum_grad_c</td>80 <td>sum_grad_c</td>
79 <td>输出</td>81 <td>输出</td>
80- <td>按行归约后的列结果，shape为<code>[batch..., M]</code>，对应公式中的<code>sum_grad_c</code>。</td>82+ <td>对倒数第二维进行归约得到的结果，具体shape见约束说明。</td>
81- <td>FLOAT32</td>83+ <td>FLOAT</td>
82 <td>ND</td>84 <td>ND</td>
83 </tr>85 </tr>
84- <tr>86+ <tr>
85- <td>sum_grad_rc</td>87+ <td>sum_grad_rc</td>
86- <td>输出</td>88+ <td>输出</td>
87- <td>全局归约结果，shape为<code>[batch...]</code>，对应公式中的<code>sum_grad_rc</code>。</td>89+ <td>对最后两维进行归约得到的结果，具体shape见约束说明。</td>
88- <td>FLOAT32</td>90+ <td>FLOAT</td>
89- <td>ND</td>91+ <td>ND</td>
90- </tr>92+ </tr>
91-</tbody>93+ </tbody>
92</table>94</table>
93 95 
94## 约束说明96## 约束说明
95 97 
96-- `grad` rank不小于2，数据格式为ND，所有维度都必须大于0；最后两维为<code>[N, M]</code>，前导维作为batch维。98+- `grad`必须为非空ND张量，所有维度的大小均必须大于0。
97-- `eps` 为FLOAT32类型、ND格式的标量或单元素一维张量。99+- `eps`必须为仅含一个FLOAT类型元素的ND张量。
98-- `sum_grad_r`、`sum_grad_c`和`sum_grad_rc`的数据类型均为FLOAT32。100+- <term>Ascend 950PR/Ascend 950DT</term>：
101+ - `grad`的rank不小于2，最后两维为`[N, M]`，其余维度为批次维度。
102+ - `grad`所有维度大小的乘积不能超过INT64的最大值。
103+ - `eps`支持标量或shape为`[1]`的一维张量。
104+ - `sum_grad_r`、`sum_grad_c`和`sum_grad_rc`的shape分别为`[batch..., N]`、`[batch..., M]`和`[batch...]`；当`grad`的rank为2时，`sum_grad_rc`为标量。
105+- <term>Atlas A3 训练系列产品/Atlas A3 推理系列产品</term>、<term>Atlas A2 训练系列产品/Atlas A2 推理系列产品</term>：
106+ - `grad`仅支持shape为`[N, M]`的二维张量。
107+ - `eps`仅支持标量。
108+ - `sum_grad_r`、`sum_grad_c`和`sum_grad_rc`的shape分别为`[N]`、`[M]`和`[1]`。
109+- 本算子仅支持GE图模式调用，不提供公开的aclnn接口。
99 110 
100## 调用说明111## 调用说明
101 112 
Moptim/apply_came_part1/op_graph/apply_came_part1_proto.h+5-5
@@ -18,14 +18,14 @@ namespace ge {
18 *18 *
19 * @par Inputs:19 * @par Inputs:
20 * including:20 * including:
21- * @li grad: A mutable Tensor with rank 2, such as [n, m] , support types:21+ * @li grad: A mutable Tensor with rank not less than 2, such as [batch..., n, m] , support types:
22 * float16, float32, bfloat16.22 * float16, float32, bfloat16.
23- * @li eps: A scalar, support types: float32. \n23+ * @li eps: A scalar or 1-element Tensor, support types: float32. \n
24 *24 *
25 * @par Outputs:25 * @par Outputs:
26- * @li sum_grad_r: A 1-dimensional Tensor, such as [n], support types: float32.26+ * @li sum_grad_r: A Tensor with shape [batch..., n], support types: float32.
27- * @li sum_grad_c: A 1-dimensional Tensor, such as [m], support types: float32.27+ * @li sum_grad_c: A Tensor with shape [batch..., m], support types: float32.
28- * @li sum_grad_rc: A 1-dimensional Tensor, such as [1], support28+ * @li sum_grad_rc: A Tensor with shape [batch...], and a scalar when grad rank is 2, support
29 * types: float32. \n29 * types: float32. \n
30 *30 *
31 * @par Restrictions:31 * @par Restrictions:
Doptim/apply_came_part1/op_host/apply_came_part1_infershape_datatype.cpp+0-38
@@ -1,38 +0,0 @@
1-/**
2- * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3- * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4- * CANN Open Software License Agreement Version 2.0 (the "License").
5- * Please refer to the License for details. You may not use this file except in compliance with the License.
6- * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7- * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8- * See LICENSE in the root of the software repository for the full text of the License.
9- *
10- * Generated By CANNBot
11- */
12- 
13-/**
14- * NOTE: Portions of this code were AI-generated and have been
15- * technically reviewed for functional accuracy and security
16- */
17- 
18-#include "register/op_impl_registry.h"
19- 
20-using namespace ge;
21- 
22-namespace ops {
23-namespace {
24-constexpr size_t kSumGradRIndex = 0;
25-constexpr size_t kSumGradCIndex = 1;
26-constexpr size_t kSumGradRCIndex = 2;
27-} // namespace
28- 
29-static ge::graphStatus InferDataType4ApplyCamePart1(gert::InferDataTypeContext* context)
30-{
31- context->SetOutputDataType(kSumGradRIndex, ge::DT_FLOAT);
32- context->SetOutputDataType(kSumGradCIndex, ge::DT_FLOAT);
33- context->SetOutputDataType(kSumGradRCIndex, ge::DT_FLOAT);
34- return ge::GRAPH_SUCCESS;
35-}
36- 
37-IMPL_OP(ApplyCamePart1).InferDataType(InferDataType4ApplyCamePart1);
38-} // namespace ops
Moptim/apply_came_part1/op_kernel/arch35/apply_came_part1_common.h+30-0文件内容审核中，请稍后刷新重试
Moptim/apply_came_part1/op_kernel/arch35/apply_came_part1_fp16.h+3-22文件内容审核中，请稍后刷新重试
Moptim/apply_came_part1/op_kernel/arch35/apply_came_part1_fp32.h+3-22文件内容审核中，请稍后刷新重试
Moptim/apply_came_part1/tests/ut/op_host/test_apply_came_part1_infershape.cpp+48-40文件内容审核中，请稍后刷新重试
Moptim/apply_came_part3/README.md+0-1
@@ -159,7 +159,6 @@
159- `u`和输入`m`必须为形状相同的二维ND张量，两个维度都必须大于0，且每个维度不超过`INT32_MAX`。159- `u`和输入`m`必须为形状相同的二维ND张量，两个维度都必须大于0，且每个维度不超过`INT32_MAX`。
160- `eps`、`beta1`、`clip_threshold`和`sum_square_u`必须为FLOAT32类型的标量或单元素一维张量。160- `eps`、`beta1`、`clip_threshold`和`sum_square_u`必须为FLOAT32类型的标量或单元素一维张量。
161- `global_shape`为可选INT64类型输入，必须是一维长度为2的张量`[global_n, global_m]`；未提供时使用输入`u`的二维形状进行归约计算。161- `global_shape`为可选INT64类型输入，必须是一维长度为2的张量`[global_n, global_m]`；未提供时使用输入`u`的二维形状进行归约计算。
162-- 本仓新增Ascend950/arch35实现；Atlas A2/A3产品沿用canndev既有实现。当前未提供aclnn接口，使用图模式调用。
163 162 
164## 调用说明163## 调用说明
165 164 
Moptim/inplace_apply_centered_rms_prop/README.md+44-44
@@ -5,29 +5,27 @@
5| 产品 | 是否支持 |5| 产品 | 是否支持 |
6| :----------------------------------------- | :------:|6| :----------------------------------------- | :------:|
7| <term>Ascend 950PR/Ascend 950DT</term> | √ |7| <term>Ascend 950PR/Ascend 950DT</term> | √ |
8-| <term>Atlas A3 训练系列产品/Atlas A3 推理系列产品</term> | √ |8+| <term>Atlas A3 训练系列产品/Atlas A3 推理系列产品</term> | × |
9-| <term>Atlas A2 训练系列产品/Atlas A2 推理系列产品</term> | √ |9+| <term>Atlas A2 训练系列产品/Atlas A2 推理系列产品</term> | × |
10| <term>Atlas 200I/500 A2 推理产品</term> | × |10| <term>Atlas 200I/500 A2 推理产品</term> | × |
11| <term>Atlas 推理系列产品</term> | × |11| <term>Atlas 推理系列产品</term> | × |
12| <term>Atlas 训练系列产品</term> | × |12| <term>Atlas 训练系列产品</term> | × |
13 13 
14## 功能说明14## 功能说明
15 15 
16-- 算子功能：InplaceApplyCenteredRMSProp是Centered RMSProp优化器步进的in-place实现，对标TensorFlow `ResourceApplyCenteredRMSProp`，在模型训练的每步迭代中根据梯度、学习率和衰减率就地更新权重变量（var）、梯度移动平均（mg）、梯度平方移动平均（ms）和动量缓冲区（mom）。16+- 算子功能：实现Centered RMSProp优化器的参数更新。在模型训练过程中，根据当前梯度、学习率和衰减系数，原地更新模型参数`var`、梯度一阶矩`mg`、梯度二阶矩`ms`和动量`mom`。
17- 计算公式：17- 计算公式：
18 18 
19-$$19+ $$
20-\begin{aligned}20+ \begin{aligned}
21-\text{mg}_t &= \rho \cdot \text{mg}_{t-1} + (1 - \rho) \cdot \text{grad} \\21+ mg_t &= \rho \cdot mg_{t-1} + (1 - \rho) \cdot g_t \\
22-\text{ms}_t &= \rho \cdot \text{ms}_{t-1} + (1 - \rho) \cdot \text{grad}^2 \\22+ ms_t &= \rho \cdot ms_{t-1} + (1 - \rho) \cdot g_t^2 \\
23-\text{mom}_t &= \text{momentum} \cdot \text{mom}_{t-1} + \frac{\text{lr} \cdot \text{grad}}{\sqrt{\text{ms}_t - \text{mg}_t^2 + \epsilon}} \\23+ mom_t &= \mu \cdot mom_{t-1} + lr \cdot \frac{g_t}{\sqrt{ms_t - mg_t^2 + \epsilon}} \\
24-\text{var}_t &= \text{var}_{t-1} - \text{mom}_t24+ var_t &= var_{t-1} - mom_t
25-\end{aligned}25+ \end{aligned}
26-$$26+ $$
27 27 
28-其中：var为权重，mg为梯度移动平均，ms为梯度平方移动平均，mom为动量缓冲区，grad为梯度，lr为学习率，rho为衰减率，momentum为动量系数，epsilon为数值稳定性常数。28+ 其中，$var$表示模型参数，$mg$表示梯度一阶矩，$ms$表示梯度二阶矩，$mom$表示动量，$g_t$表示当前步梯度，$lr$表示学习率，$\rho$表示衰减系数，$\mu$表示动量系数，$\epsilon$表示数值稳定性常数。$\epsilon$位于平方根内部，与TensorFlow `ResourceApplyCenteredRMSProp`的计算方式一致。
29- 
30-epsilon加在sqrt内部（`sqrt(ms - mg² + epsilon)`），对标TensorFlow。PyTorch的epsilon在sqrt之外，不作为对标对象。
31 29 
32## 参数说明30## 参数说明
33 31 
@@ -49,99 +47,99 @@ epsilon加在sqrt内部（`sqrt(ms - mg² + epsilon)`），对标TensorFlow。Py
49<tbody>47<tbody>
50 <tr>48 <tr>
51 <td>var</td>49 <td>var</td>
52- <td>输入 / 输出(inplace)</td>50+ <td>输入</td>
53- <td>模型权重张量。Kernel内inplace更新，GE IR输出视图与输入var共享Device内存。</td>51+ <td>待更新的模型参数，shape与mg、ms、mom和grad一致。</td>
54 <td>FLOAT、FLOAT16</td>52 <td>FLOAT、FLOAT16</td>
55 <td>ND</td>53 <td>ND</td>
56 </tr>54 </tr>
57 <tr>55 <tr>
58 <td>mg</td>56 <td>mg</td>
59- <td>输入(inplace更新)</td>57+ <td>输入</td>
60- <td>梯度移动平均。shape/dtype必须与var一致；Kernel内显式写回输入GM地址。</td>58+ <td>梯度一阶矩的指数移动平均，shape与var一致。</td>
61 <td>FLOAT、FLOAT16</td>59 <td>FLOAT、FLOAT16</td>
62 <td>ND</td>60 <td>ND</td>
63 </tr>61 </tr>
64 <tr>62 <tr>
65 <td>ms</td>63 <td>ms</td>
66- <td>输入(inplace更新)</td>64+ <td>输入</td>
67- <td>梯度平方移动平均。shape/dtype必须与var一致；Kernel内显式写回输入GM地址。</td>65+ <td>梯度二阶矩的指数移动平均，shape与var一致。</td>
68 <td>FLOAT、FLOAT16</td>66 <td>FLOAT、FLOAT16</td>
69 <td>ND</td>67 <td>ND</td>
70 </tr>68 </tr>
71 <tr>69 <tr>
72 <td>mom</td>70 <td>mom</td>
73- <td>输入(inplace更新)</td>71+ <td>输入</td>
74- <td>动量缓冲区。shape/dtype必须与var一致；Kernel内显式写回输入GM地址。</td>72+ <td>动量，shape与var一致。</td>
75 <td>FLOAT、FLOAT16</td>73 <td>FLOAT、FLOAT16</td>
76 <td>ND</td>74 <td>ND</td>
77 </tr>75 </tr>
78 <tr>76 <tr>
79 <td>lr</td>77 <td>lr</td>
80 <td>输入</td>78 <td>输入</td>
81- <td>学习率（0-d tensor）。dtype必须与var一致。</td>79+ <td>学习率。必须为标量或仅含一个元素的一维张量。</td>
82 <td>FLOAT、FLOAT16</td>80 <td>FLOAT、FLOAT16</td>
83 <td>ND</td>81 <td>ND</td>
84 </tr>82 </tr>
85 <tr>83 <tr>
86 <td>rho</td>84 <td>rho</td>
87 <td>输入</td>85 <td>输入</td>
88- <td>衰减率（0-d tensor）。dtype必须与var一致。</td>86+ <td>衰减系数。必须为标量或仅含一个元素的一维张量。</td>
89 <td>FLOAT、FLOAT16</td>87 <td>FLOAT、FLOAT16</td>
90 <td>ND</td>88 <td>ND</td>
91 </tr>89 </tr>
92 <tr>90 <tr>
93 <td>momentum</td>91 <td>momentum</td>
94 <td>输入</td>92 <td>输入</td>
95- <td>动量系数（0-d tensor）。dtype必须与var一致。</td>93+ <td>动量系数。必须为标量或仅含一个元素的一维张量。</td>
96 <td>FLOAT、FLOAT16</td>94 <td>FLOAT、FLOAT16</td>
97 <td>ND</td>95 <td>ND</td>
98 </tr>96 </tr>
99 <tr>97 <tr>
100 <td>epsilon</td>98 <td>epsilon</td>
101 <td>输入</td>99 <td>输入</td>
102- <td>数值稳定性常数（0-d tensor），加在sqrt内部。dtype必须与var一致。</td>100+ <td>数值稳定性常数。必须为标量或仅含一个元素的一维张量。</td>
103 <td>FLOAT、FLOAT16</td>101 <td>FLOAT、FLOAT16</td>
104 <td>ND</td>102 <td>ND</td>
105 </tr>103 </tr>
106 <tr>104 <tr>
107 <td>grad</td>105 <td>grad</td>
108 <td>输入</td>106 <td>输入</td>
109- <td>当前步梯度。shape/dtype必须与var一致。</td>107+ <td>当前步梯度，shape与var一致。</td>
110 <td>FLOAT、FLOAT16</td>108 <td>FLOAT、FLOAT16</td>
111 <td>ND</td>109 <td>ND</td>
112 </tr>110 </tr>
113 <tr>111 <tr>
114 <td>use_locking</td>112 <td>use_locking</td>
115 <td>属性</td>113 <td>属性</td>
116- <td>是否在更新时加锁。默认false。当前实现不强制互斥锁，仅作语义占位。</td>114+ <td>可选兼容属性，默认值为false，取值不影响计算结果。</td>
117 <td>BOOL</td>115 <td>BOOL</td>
118 <td>-</td>116 <td>-</td>
119 </tr>117 </tr>
120 <tr>118 <tr>
121- <td>var (output)</td>119+ <td>var</td>
122 <td>输出</td>120 <td>输出</td>
123- <td>更新后的var Tensor，与输入var共享Device内存（inplace）。</td>121+ <td>更新后的模型参数，与输入var共享存储。</td>
124 <td>FLOAT、FLOAT16</td>122 <td>FLOAT、FLOAT16</td>
125 <td>ND</td>123 <td>ND</td>
126 </tr>124 </tr>
127 <tr>125 <tr>
128- <td>mg (output)</td>126+ <td>mg</td>
129 <td>输出</td>127 <td>输出</td>
130- <td>更新后的mg Tensor，与输入mg共享Device内存（inplace）。</td>128+ <td>更新后的梯度一阶矩，与输入mg共享存储。</td>
131 <td>FLOAT、FLOAT16</td>129 <td>FLOAT、FLOAT16</td>
132 <td>ND</td>130 <td>ND</td>
133 </tr>131 </tr>
134 <tr>132 <tr>
135- <td>ms (output)</td>133+ <td>ms</td>
136 <td>输出</td>134 <td>输出</td>
137- <td>更新后的ms Tensor，与输入ms共享Device内存（inplace）。</td>135+ <td>更新后的梯度二阶矩，与输入ms共享存储。</td>
138 <td>FLOAT、FLOAT16</td>136 <td>FLOAT、FLOAT16</td>
139 <td>ND</td>137 <td>ND</td>
140 </tr>138 </tr>
141 <tr>139 <tr>
142- <td>mom (output)</td>140+ <td>mom</td>
143 <td>输出</td>141 <td>输出</td>
144- <td>更新后的mom Tensor，与输入mom共享Device内存（inplace）。</td>142+ <td>更新后的动量，与输入mom共享存储。</td>
145 <td>FLOAT、FLOAT16</td>143 <td>FLOAT、FLOAT16</td>
146 <td>ND</td>144 <td>ND</td>
147 </tr>145 </tr>
@@ -149,14 +147,16 @@ epsilon加在sqrt内部（`sqrt(ms - mg² + epsilon)`），对标TensorFlow。Py
149 147 
150## 约束说明148## 约束说明
151 149 
152-- **数据类型一致性**：var、mg、ms、mom、grad五个tensor的数据类型必须一致；lr、rho、momentum、epsilon四个scalar的数据类型必须与tensor一致。150+- var、mg、ms、mom和grad的shape必须完全一致。
153-- **Shape约束**：var、mg、ms、mom、grad的shape必须完全相同；lr、rho、momentum、epsilon必须为scalar（0-d tensor）；tensor维度范围0-8维。151+- lr、rho、momentum和epsilon必须为标量或仅含一个元素的一维张量。
154-- **In-place语义**：var、mg、ms、mom四个参数为就地更新，执行后原始数据被覆盖。152+- 所有输入和输出的数据类型必须相同，支持FLOAT和FLOAT16。
155-- **epsilon位置**：epsilon必须加在sqrt内部（`sqrt(ms - mg² + ε)`），禁止实现为 `sqrt(ms - mg²) + ε`。153+- var、mg、ms和mom均为原地更新，输出与对应输入共享存储。
156-- **标量精度**：lr/rho/momentum/epsilon统一在FP32域读取，禁止用FP16标量参与计算。154+- <term>Ascend 950PR/Ascend 950DT</term>：
157-- **FP16计算**：FP16输入在kernel内部提升到FP32计算，结果cast回FP16（CAST_NONE输入 / CAST_RINT输出）。155+ - var、mg、ms、mom、grad及其对应输出仅支持ND格式，rank不超过8。
158-- **大值溢出**：当 `ms - mg² + ε` 接近零时，`lr * grad / sqrt(denom)` 会产生大值溢出，kernel保持数学正确性（不截断、不饱和），输出可能为Inf/NaN。156+ - 支持空Tensor。
159-- **空Tensor**：支持空Tensor（numel=0），kernel跳过计算。157+ - FLOAT16输入在内部使用FLOAT精度进行计算，结果转换为FLOAT16后输出。
158+- `use_locking`仅用于接口兼容，当前实现不提供额外的互斥锁。
159+- 本算子仅支持GE图模式调用，不提供公开的aclnn接口。
160 160 
161## 调用说明161## 调用说明
162 162 
Moptim/inplace_apply_rms_prop/README.md+8-8文件内容审核中，请稍后刷新重试
Moptim/inplace_apply_rms_prop/op_graph/inplace_apply_rms_prop_proto.h+2-7文件内容审核中，请稍后刷新重试
Moptim/inplace_apply_rms_prop/op_kernel/arch35/inplace_apply_rms_prop.h+8-3文件内容审核中，请稍后刷新重试