add Add/RmsNormDynamicMxQuant op

已合并

add Add/RmsNormDynamicMxQuant op_api #3176

刘琦创建于 3月25日从liiu-qii/add_rms_aclnn： ARND_aclnn 合入到 cann/ops-nn： master

add Add/RmsNormDynamicMxQuant op_api #3176

已合并

刘琦创建于 3月25日

从liiu-qii/add_rms_aclnn ARND_aclnn合入到cann/ops-nn master

收起

norm
- add_rms_norm_dynamic_mx_quant
  - README.md
  - docs
    - aclnnAddRmsNormDynamicMxQuant.md
  - op_host
    - CMakeLists.txt
    - op_api
      - aclnn_add_rms_norm_dynamic_mx_quant.cpp
      - aclnn_add_rms_norm_dynamic_mx_quant.h
      - add_rms_norm_dynamic_mx_quant.cpp
      - add_rms_norm_dynamic_mx_quant.h
- rms_norm_dynamic_mx_quant
  - CMakeLists.txt
  - README.md
  - docs
    - aclnnRmsNormDynamicMxQuant.md
  - examples
    - test_aclnn_rms_norm_dynamic_mx_quant.cpp
  - op_graph
    - CMakeLists.txt
    - rms_norm_dynamic_mx_quant_proto.h
  - op_host
    - CMakeLists.txt
    - arch35
      - rms_norm_dynamic_mx_quant_base_tiling.cpp
      - rms_norm_dynamic_mx_quant_full_load_tiling.cpp
      - rms_norm_dynamic_mx_quant_tiling_arch35.h
    - op_api
      - aclnn_rms_norm_dynamic_mx_quant.cpp
      - aclnn_rms_norm_dynamic_mx_quant.h
      - rms_norm_dynamic_mx_quant.cpp
      - rms_norm_dynamic_mx_quant.h
    - rms_norm_dynamic_mx_quant_def.cpp
    - rms_norm_dynamic_mx_quant_infershape.cpp
  - op_kernel
    - arch35
      - rms_norm_dynamic_mx_quant_common.h
      - rms_norm_dynamic_mx_quant_full_load.h
    - rms_norm_dynamic_mx_quant_apt.cpp
  - tests
    - CMakeLists.txt
    - ut / op_host
      - arch35
        test_rms_norm_dynamic_mx_quant_tiling_arch35.cpp
      - test_rms_norm_dynamic_mx_quant_infershape.cpp

共 29 个文件变更+745-88

Mnorm/add_rms_norm_dynamic_mx_quant/README.md+10-10

	@@ -26,7 +26,7 @@
26	y = \operatorname{RmsNorm}(x)=\frac{x}{\operatorname{Rms}(\mathbf{x})}\cdot gamma+beta, \quad \text { where } \operatorname{Rms}(\mathbf{x})=\sqrt{\frac{1}{n} \sum_{i=1}^n x_i^2+epsilon}	26	y = \operatorname{RmsNorm}(x)=\frac{x}{\operatorname{Rms}(\mathbf{x})}\cdot gamma+beta, \quad \text { where } \operatorname{Rms}(\mathbf{x})=\sqrt{\frac{1}{n} \sum_{i=1}^n x_i^2+epsilon}
27	$$	27	$$
28		28
29	- 当scaleAlg为0时：	29	+ 当scale_alg为0时：
30	- 将RmsNorm输出y在尾轴维度上按k = 32个数分组，一组k个数 $\{\{V_i\}_{i=1}^{k}\}$ 动态量化为 $\{mxscale,\{P_i\}_{i=1}^{k}\}$	30	- 将RmsNorm输出y在尾轴维度上按k = 32个数分组，一组k个数 $\{\{V_i\}_{i=1}^{k}\}$ 动态量化为 $\{mxscale,\{P_i\}_{i=1}^{k}\}$
31	$$	31	$$
32	shared\_exp = floor(log_2(max_i(\|V_i\|))) - emax	32	shared\_exp = floor(log_2(max_i(\|V_i\|))) - emax
	@@ -48,7 +48,7 @@
48	\| FLOAT8_E4M3FN \| 8 \|	48	\| FLOAT8_E4M3FN \| 8 \|
49	\| FLOAT8_E5M2 \| 15 \|	49	\| FLOAT8_E5M2 \| 15 \|
50		50
51	- 当scaleAlg为1时，只涉及FP8类型：	51	+ 当scale_alg为1时，只涉及FP8类型：
52	- 将长向量按块分，每块长度为k，对每块单独计算一个块缩放因子$S_{fp32}^b$，再把块内所有元素用同一个$S_{fp32}^b$映射到目标低精度类型FP8。	52	- 将长向量按块分，每块长度为k，对每块单独计算一个块缩放因子$S_{fp32}^b$，再把块内所有元素用同一个$S_{fp32}^b$映射到目标低精度类型FP8。
53	- 找到该块中数值的最大绝对值:	53	- 找到该块中数值的最大绝对值:
54	$$	54	$$
	@@ -124,28 +124,28 @@
124	<tr>	124	<tr>
125	<td>scale_alg</td>	125	<td>scale_alg</td>
126	<td>可选属性</td>	126	<td>可选属性</td>
127	- <td><ul><li>表示mxscaleOut的计算方法，对应公式中的scaleAlg。</li><li>支持取值0和1，取值为0表示Open Compute Project(OCP)实现，取值为1表示cuBLAS实现。当dstType为FLOAT4_E2M1/FLOAT4_E1M2时仅支持取值为0。</li><li>默认值为0。</li></ul></td>	127	+ <td><ul><li>表示mxscale的计算方法，对应公式中的scale_alg。</li><li>支持取值0和1，取值为0表示Open Compute Project(OCP)实现，取值为1表示cuBLAS实现。当dst_type为FLOAT4_E2M1/FLOAT4_E1M2时仅支持取值为0。</li><li>默认值为0。</li></ul></td>
128	<td>INT64</td>	128	<td>INT64</td>
129	<td>-</td>	129	<td>-</td>
130	</tr>	130	</tr>
131	<tr>	131	<tr>
132	<td>round_mode</td>	132	<td>round_mode</td>
133	<td>可选属性</td>	133	<td>可选属性</td>
134	- <td><ul><li>表示数据转换的模式，对应公式中的round_mode。</li><li>当dstType为40/41时，支持{"rint", "floor", "round"}。</li><li>当dstType为36/35时，仅支持{"rint"}。</li><li>默认值为"rint"。</li></ul></td>	134	+ <td><ul><li>表示数据转换的模式，对应公式中的round_mode。</li><li>当dst_type为40/41时，支持{"rint", "floor", "round"}。</li><li>当dst_type为36/35时，仅支持{"rint"}。</li><li>默认值为"rint"。</li></ul></td>
135	<td>STRING</td>	135	<td>STRING</td>
136	<td>-</td>	136	<td>-</td>
137	</tr>	137	</tr>
138	<tr>	138	<tr>
139	<td>dst_type</td>	139	<td>dst_type</td>
140	<td>可选属性</td>	140	<td>可选属性</td>
141	- <td><ul><li>表示指定数据转换后yOut的类型，对应公式中的DType。</li><li>输入范围为{35, 36, 40, 41}，分别对应{FLOAT8_E5M2, FLOAT8_E4M3FN, FLOAT4_E2M1, FLOAT4_E1M2}。</li><li>默认值为40。</li></ul></td>	141	+ <td><ul><li>表示指定数据转换后y的类型，对应公式中的DType。</li><li>输入范围为{35, 36, 40, 41}，分别对应{FLOAT8_E5M2, FLOAT8_E4M3FN, FLOAT4_E2M1, FLOAT4_E1M2}。</li><li>默认值为40。</li></ul></td>
142	<td>INT64</td>	142	<td>INT64</td>
143	<td>-</td>	143	<td>-</td>
144	</tr>	144	</tr>
145	<tr>	145	<tr>
146	<td>output_rstd</td>	146	<td>output_rstd</td>
147	<td>可选属性</td>	147	<td>可选属性</td>
148	- <td><ul><li>表示指定是否输出有效的rstdOut。</li><li>支持True和False。</li><li>默认值为False。</li><li>当outputRstd为False时，rstdOut最终输出为空Tensor，此时rstdOut为无效占位输出。</li></ul></td>	148	+ <td><ul><li>表示指定是否输出有效的rstd_out。</li><li>支持True和False。</li><li>默认值为False。</li><li>当output_rstd为False时，rstd为无效占位输出。</li></ul></td>
149	<td>BOOL</td>	149	<td>BOOL</td>
150	<td>-</td>	150	<td>-</td>
151	</tr>	151	</tr>
	@@ -173,7 +173,7 @@
173	<tr>	173	<tr>
174	<td>rstd</td>	174	<td>rstd</td>
175	<td>输出</td>	175	<td>输出</td>
176	- <td><ul><li>表示归一化后的标准差的倒数，对应公式中Rms(x)的倒数。</li><li>当output_rstd为True时，shape与入参x1的shape前几维保持一致，前几维指x1的维度减去gamma的维度，表示不需要norm的维度。</li><li>当output_rstd为False时，rstd_out输出空Tensor。</li></ul></td>	176	+ <td><ul><li>表示归一化后的标准差的倒数，对应公式中Rms(x)的倒数。</li><li>当output_rstd为True时，shape与入参x1的shape前几维保持一致，前几维指x1的维度减去gamma的维度，表示不需要norm的维度。</li><li>当output_rstd为False时，rstd为无效占位输出。</li></ul></td>
177	<td>FLOAT32</td>	177	<td>FLOAT32</td>
178	<td>ND</td>	178	<td>ND</td>
179	</tr>	179	</tr>
	@@ -189,13 +189,13 @@
189	- mxscale.shape[-1] = 2。	189	- mxscale.shape[-1] = 2。
190	- 其他维度与输入x1一致。	190	- 其他维度与输入x1一致。
191		191
192	-- 当输出yOut的数据类型为FLOAT4_E2M1或FLOAT4_E1M2，x1尾轴的值必须为偶数。	192	+- 当输出y的数据类型为FLOAT4_E2M1或FLOAT4_E1M2，x1尾轴的值必须为偶数。
193		193
194	- 输入gamma、可选输入beta的数据类型只能和x1的数据类型保持一致或者为FLOAT32。	194	- 输入gamma、可选输入beta的数据类型只能和x1的数据类型保持一致或者为FLOAT32。
195		195
196	- 边界值场景说明	196	- 边界值场景说明
197	- - 当输入是Inf时：1、输出yOut为0；2、输出xOut为Inf；3、输出mxscaleOut为255，偶数pad填充值为0；4、输出rstdOut为0。	197	+ - 当输入是Inf时：1、输出y为0；2、输出x为Inf；3、输出mxscale为255，偶数pad填充值为0；4、输出rstd为0。
198	- - 当输入是NaN时：1、输出yOut为0；2、输出xOut为Nan；3、输出mxscaleOut为255，偶数pad填充值为0；4、输出rstdOut为NaN。	198	+ - 当输入是NaN时：1、输出y为0；2、输出x为Nan；3、输出mxscale为255，偶数pad填充值为0；4、输出rstd为NaN。
199		199
200	## 调用说明	200	## 调用说明
201		201

Mnorm/add_rms_norm_dynamic_mx_quant/docs/aclnnAddRmsNormDynamicMxQuant.md+23-21

	@@ -80,7 +80,7 @@ aclnnStatus aclnnAddRmsNormDynamicMxQuantGetWorkspaceSize(
80	const aclTensor *beta,	80	const aclTensor *beta,
81	double epsilon,	81	double epsilon,
82	int64_t scaleAlg,	82	int64_t scaleAlg,
83	- char *roundModeOptional,	83	+ char *roundMode,
84	int64_t dstType,	84	int64_t dstType,
85	bool outputRstd,	85	bool outputRstd,
86	aclTensor *yOut,	86	aclTensor *yOut,
	@@ -129,7 +129,7 @@ aclnnStatus aclnnAddRmsNormDynamicMxQuant(
129	<td>x1（aclTensor*）</td>	129	<td>x1（aclTensor*）</td>
130	<td>输入</td>	130	<td>输入</td>
131	<td>表示标准化过程中的源数据张量。对应公式中的x1。</td>	131	<td>表示标准化过程中的源数据张量。对应公式中的x1。</td>
132	- <td><ul><li>不支持空Tensor。</li></ul></td>	132	+ <td><ul><li>不支持空Tensor。</li><li>具体约束详见约束说明。</li></ul></td>
133	<td>FLOAT16、BFLOAT16</td>	133	<td>FLOAT16、BFLOAT16</td>
134	<td>ND</td>	134	<td>ND</td>
135	<td>1-7</td>	135	<td>1-7</td>
	@@ -149,7 +149,7 @@ aclnnStatus aclnnAddRmsNormDynamicMxQuant(
149	<td>gamma（aclTensor*）</td>	149	<td>gamma（aclTensor*）</td>
150	<td>输入</td>	150	<td>输入</td>
151	<td>表示标准化过程中的权重张量。对应公式中的gamma。</td>	151	<td>表示标准化过程中的权重张量。对应公式中的gamma。</td>
152	- <td><ul><li>不支持空Tensor。</li><li>shape只能为一维，需要与x1最后一维维度匹配。</li></ul></td>	152	+ <td><ul><li>不支持空Tensor。</li><li>shape只能为一维，需要与x1最后一维维度匹配。</li><li>数据类型默认与输入x1一致；若不一致，则显示设为FLOAT32。</li></ul></td>
153	<td>FLOAT16、BFLOAT16、FLOAT32</td>	153	<td>FLOAT16、BFLOAT16、FLOAT32</td>
154	<td>ND</td>	154	<td>ND</td>
155	<td>1</td>	155	<td>1</td>
	@@ -209,7 +209,7 @@ aclnnStatus aclnnAddRmsNormDynamicMxQuant(
209	<td>outputRstd（bool）</td>	209	<td>outputRstd（bool）</td>
210	<td>输入</td>	210	<td>输入</td>
211	<td>表示指定是否输出有效的rstdOut。</td>	211	<td>表示指定是否输出有效的rstdOut。</td>
212	- <td><ul><li>支持True和False。</li><li>当outputRstd为False时，rstdOut最终输出为空Tensor，此时rstdOut为无效占位输出。</li></ul></td>	212	+ <td><ul><li>支持True和False。</li><li>当outputRstd为False时，rstdOut为无效输出。</li></ul></td>
213	<td>-</td>	213	<td>-</td>
214	<td>-</td>	214	<td>-</td>
215	<td>-</td>	215	<td>-</td>
	@@ -219,7 +219,7 @@ aclnnStatus aclnnAddRmsNormDynamicMxQuant(
219	<td>yOut（aclTensor*）</td>	219	<td>yOut（aclTensor*）</td>
220	<td>输出</td>	220	<td>输出</td>
221	<td>表示归一化并量化后的结果，对应公式中的Pi和di。</td>	221	<td>表示归一化并量化后的结果，对应公式中的Pi和di。</td>
222	- <td><ul><li>不支持空Tensor。</li><li>shape需要与输入x1一致。</li></ul></td>	222	+ <td><ul><li>不支持空Tensor。</li><li>shape需要与输入x1一致。</li><li>数据类型需要与dstType保持一致。</li></ul></td>
223	<td>FLOAT4_E2M1、FLOAT4_E1M2、FLOAT8_E4M3FN、FLOAT8_E5M2</td>	223	<td>FLOAT4_E2M1、FLOAT4_E1M2、FLOAT8_E4M3FN、FLOAT8_E5M2</td>
224	<td>ND</td>	224	<td>ND</td>
225	<td>1-7</td>	225	<td>1-7</td>
	@@ -239,7 +239,7 @@ aclnnStatus aclnnAddRmsNormDynamicMxQuant(
239	<td>mxscaleOut（aclTensor*）</td>	239	<td>mxscaleOut（aclTensor*）</td>
240	<td>输出</td>	240	<td>输出</td>
241	<td>表示每个分组对应的量化尺度，对应公式中的mxscale和Sb。</td>	241	<td>表示每个分组对应的量化尺度，对应公式中的mxscale和Sb。</td>
242	- <td><ul><li>不支持空Tensor。</li><li>shape在尾轴上为x对应值除以blocksize=32向上取整，并对其进行偶数pad，pad填充值为0，具体计算过程见约束说明。其余维度与入参x1的shape前几维保持一致，前几维指x1的维度减去gamma的维度，表示不需要norm的维度。</li></ul></td>	242	+ <td><ul><li>不支持空Tensor。</li><li>shape在尾轴上为x对应值除以blocksize=32向上取整，并对其进行偶数pad，pad填充值为0。</li><li>其余维度与入参x1的shape前几维保持一致，前几维指x1的维度减去gamma的维度，表示不需要norm的维度。</li><li>具体计算过程见约束说明。</li></ul></td>
243	<td>FLOAT8_E8M0</td>	243	<td>FLOAT8_E8M0</td>
244	<td>ND</td>	244	<td>ND</td>
245	<td>2-8</td>	245	<td>2-8</td>
	@@ -249,7 +249,7 @@ aclnnStatus aclnnAddRmsNormDynamicMxQuant(
249	<td>rstdOut（aclTensor*）</td>	249	<td>rstdOut（aclTensor*）</td>
250	<td>输出</td>	250	<td>输出</td>
251	<td>表示归一化后的标准差的倒数。对应公式中Rms(x)的倒数。</td>	251	<td>表示归一化后的标准差的倒数。对应公式中Rms(x)的倒数。</td>
252	- <td><ul><li>支持空Tensor。</li><li>当outputRstd为True时，shape与入参x1的shape前几维保持一致，前几维指x1的维度减去gamma的维度，表示不需要norm的维度。</li><li>当outputRstd为False时，rstdOut输出空Tensor。</li></ul></td>	252	+ <td><ul><li>不支持空Tensor。</li><li>当outputRstd为True时，shape与入参x1的shape前几维保持一致，前几维指x1的维度减去gamma的维度，表示不需要norm的维度。</li><li>当outputRstd为False时，rstdOut为无效输出。</li></ul></td>
253	<td>FLOAT32</td>	253	<td>FLOAT32</td>
254	<td>ND</td>	254	<td>ND</td>
255	<td>1-7</td>	255	<td>1-7</td>
	@@ -298,36 +298,38 @@ aclnnStatus aclnnAddRmsNormDynamicMxQuant(
298	</thead>	298	</thead>
299	<tbody>	299	<tbody>
300	<tr>	300	<tr>
301	- <td>ACLNN_ERR_PARAM_NULLPTR</td>	301	+ <td rowspan="2">ACLNN_ERR_PARAM_NULLPTR</td>
302	- <td>161001</td>	302	+ <td rowspan="2">161001</td>
303	- <td>传入的x1、x2、gamma、yOut、xOut、mxscaleOut和rstdOut是空指针。</td>	303	+ <td>传入的x1、x2、gamma、yOut、xOut和mxscaleOut是空指针。</td>
304	</tr>	304	</tr>
305	<tr>	305	<tr>
306	- <td>ACLNN_ERR_PARAM_INVALID</td>	306	+ <td>当outputRstd为True时，传入rstdOut是空指针。</td>
307	- <td>161002</td>
308	- <td>输入或输出的数据类型不在支持的范围之内。</td>
309	</tr>	307	</tr>
310	<tr>	308	<tr>
311	- <td rowspan="9">ACLNN_ERR_INNER_TILING_ERROR</td>	309	+ <td rowspan="8">ACLNN_ERR_INNER_TILING_ERROR</td>
312	- <td rowspan="9">561002</td>	310	+ <td rowspan="8">561002</td>
		311	+ <td>输入或输出参数的维度数和数据类型不在范围之内，dstType和yOut的数据类型不匹配。</td>
313	</tr>	312	</tr>
314	<tr>	313	<tr>
315	- <td>scaleAlg不是0或1，roundModeOptional(非空时)不是 {rint, floor, round}。</td>	314	+ <td>输入x1、x2和输出yOut、xOut的shape不是完全相同的shape。</td>
316	</tr>	315	</tr>
317	<tr>	316	<tr>
318	- <td>dstType为 fp8 时，roundModeOptional不是 rint。</td>	317	+ <td>gamma、beta(若存在)的shape不是完全相同的shape，或二者的维度数不为1，或轴长不等于x1的尾轴大小，或者类型不相同。</td>
319	</tr>	318	</tr>
320	<tr>	319	<tr>
321	- <td>输入x1、输出yOut的shape不是完全相同的shape。</td>	320	+ <td>gamma的维度和x1的需要做norm的维度不相同，或rstdOut的维度和x1的不需要norm的维度不相同，或rstdOut对应x1归一化维度的轴长不为1。</td>
322	</tr>	321	</tr>
323	<tr>	322	<tr>
324	- <td>mxscaleOut的维度数不等于输入x1的维度数+1。</td>	323	+ <td>scaleAlg不是0或1，roundMode不是 {rint, floor, round}。</td>
325	</tr>	324	</tr>
326	<tr>	325	<tr>
327	- <td>gamma、beta(若存在)的shape不是完全相同的shape，或者类型不相同。</td>	326	+ <td>dstType为 fp8 时，roundMode不是 rint。</td>
328	</tr>	327	</tr>
329	<tr>	328	<tr>
330	- <td>gamma的维度和x1的需要作norm的维度不相同，或rstdOut的维度和x1的不需要norm的维度不相同，或x1的需要norm的维度数不为1。</td>	329	+ <td>dstType为 fp4 时，scaleAlg不是0，或者输入x1的尾轴不能被2整除。</td>
		330	+ </tr>
		331	+ <tr>
		332	+ <td>mxscaleOut的维度数不等于输入x1的维度数+1，或轴长不符合约束说明。</td>
331	</tr>	333	</tr>
332	</tbody></table>	334	</tbody></table>
333		335

Mnorm/add_rms_norm_dynamic_mx_quant/op_host/CMakeLists.txt+1-1

	@@ -9,4 +9,4 @@
9	# -----------------------------------------------------------------------------------------------------------	9	# -----------------------------------------------------------------------------------------------------------
10		10
11	add_modules_sources(HOSTNAME ${OPHOST_NAME} MODE PRIVATE DIR ${CMAKE_CURRENT_SOURCE_DIR}	11	add_modules_sources(HOSTNAME ${OPHOST_NAME} MODE PRIVATE DIR ${CMAKE_CURRENT_SOURCE_DIR}
12	- OPTYPE add_rms_norm_dynamic_mx_quant ACLNNTYPE aclnn DEPENDENCIES norm_common)	12	+ OPTYPE add_rms_norm_dynamic_mx_quant ACLNNTYPE aclnn_exclude DEPENDENCIES norm_common)

Anorm/add_rms_norm_dynamic_mx_quant/op_host/op_api/aclnn_add_rms_norm_dynamic_mx_quant.cpp+157-0

@@ -0,0 +1,157 @@
	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	+#include "aclnn/aclnn_base.h"
	11	+#include "op_api/op_api_def.h"
	12	+
	13	+#include "opdev/common_types.h"
	14	+#include "opdev/data_type_utils.h"
	15	+#include "opdev/op_executor.h"
	16	+#include "opdev/op_log.h"
	17	+#include "opdev/format_utils.h"
	18	+#include "opdev/tensor_view_utils.h"
	19	+#include "opdev/op_dfx.h"
	20	+#include "opdev/shape_utils.h"
	21	+#include "opdev/platform.h"
	22	+
	23	+#include "aclnn_kernels/common/op_error_check.h"
	24	+#include "aclnn_kernels/cast.h"
	25	+#include "aclnn_kernels/contiguous.h"
	26	+#include "op_api/aclnn_util.h"
	27	+
	28	+#include "add_rms_norm_dynamic_mx_quant.h"
	29	+#include "aclnn_add_rms_norm_dynamic_mx_quant.h"
	30	+
	31	+using namespace op;
	32	+#ifdef __cplusplus
	33	+extern "C" {
	34	+#endif
	35	+
	36	+constexpr int IDX_0 = 0;
	37	+constexpr int IDX_1 = 1;
	38	+constexpr int IDX_2 = 2;
	39	+constexpr int IDX_3 = 3;
	40	+
	41	+static bool CheckNotNull(
	42	+ const aclTensor* x1, const aclTensor* x2, const aclTensor* gamma, aclTensor* yOut, aclTensor* xOut,
	43	+ aclTensor* mxscaleOut, bool outputRstd, aclTensor* rstdOut)
	44	+{
	45	+ OP_CHECK_NULL(x1, return false);
	46	+ OP_CHECK_NULL(x2, return false);
	47	+ OP_CHECK_NULL(gamma, return false);
	48	+ // beta 是可选输入，不校验nullptr
	49	+ OP_CHECK_NULL(yOut, return false);
	50	+ OP_CHECK_NULL(xOut, return false);
	51	+ OP_CHECK_NULL(mxscaleOut, return false);
	52	+ // rstdOut 根据 outputRstd 决定是否校验
	53	+ if (outputRstd) {
	54	+ OP_CHECK_NULL(rstdOut, return false);
	55	+ }
	56	+ return true;
	57	+}
	58	+
	59	+static aclnnStatus ComputeAddRmsNormDynamicMxQuant(
	60	+ const aclTensor* x1, const aclTensor* x2, const aclTensor* gamma, const aclTensor* beta, double epsilon,
	61	+ int64_t scaleAlg, char* roundMode, int64_t dstType, bool outputRstd, aclTensor* yOut, aclTensor* xOut,
	62	+ aclTensor* mxscaleOut, aclTensor* rstdOut, aclOpExecutor* executor)
	63	+{
	64	+ // 创建输出Tensor
	65	+ aclTensor* y_output = executor->AllocTensor(yOut->GetViewShape(), yOut->GetDataType(), yOut->GetViewFormat());
	66	+ aclTensor* x_output = executor->AllocTensor(xOut->GetViewShape(), xOut->GetDataType(), xOut->GetViewFormat());
	67	+ aclTensor* mxscale_output =
	68	+ executor->AllocTensor(mxscaleOut->GetViewShape(), mxscaleOut->GetDataType(), mxscaleOut->GetViewFormat());
	69	+ aclTensor* rstd_output =
	70	+ (outputRstd) ?
	71	+ executor->AllocTensor(rstdOut->GetViewShape(), rstdOut->GetDataType(), rstdOut->GetViewFormat()) :
	72	+ executor->AllocTensor(op::Shape({0}), op::DataType::DT_FLOAT, op::Format::FORMAT_ND);
	W W wangyu_ai4月8日此条代码评论区间+69至+72 可选场景能传空指针嘛？刘琦 4月9日评论： 4月9日评论：
	73	+
	74	+ auto addRmsNormDynamicMxQuantOuts = l0op::AddRmsNormDynamicMxQuant(
	75	+ x1, x2, gamma, beta, epsilon, scaleAlg, roundMode, dstType, outputRstd, y_output, x_output, mxscale_output,
	76	+ rstd_output, executor);
	77	+
	78	+ auto yComputeOut = std::get<IDX_0>(addRmsNormDynamicMxQuantOuts);
	79	+ auto xComputeOut = std::get<IDX_1>(addRmsNormDynamicMxQuantOuts);
	80	+ auto mxscaleComputeOut = std::get<IDX_2>(addRmsNormDynamicMxQuantOuts);
	81	+ auto rstdComputeOut = std::get<IDX_3>(addRmsNormDynamicMxQuantOuts);
	82	+
	83	+ // 校验输出不为空
	84	+ CHECK_RET(yComputeOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
	85	+ CHECK_RET(xComputeOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
	86	+ CHECK_RET(mxscaleComputeOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
	87	+
	88	+ // 将结果拷贝到输出tensor
	89	+ auto viewCopyYResult = l0op::ViewCopy(yComputeOut, yOut, executor);
	90	+ CHECK_RET(viewCopyYResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
	91	+
	92	+ auto viewCopyXResult = l0op::ViewCopy(xComputeOut, xOut, executor);
	93	+ CHECK_RET(viewCopyXResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
	94	+
	95	+ auto viewCopyMxscaleResult = l0op::ViewCopy(mxscaleComputeOut, mxscaleOut, executor);
	96	+ CHECK_RET(viewCopyMxscaleResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
	97	+
	98	+ if (outputRstd) {
	99	+ CHECK_RET(rstdComputeOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
	100	+ auto viewCopyRstdResult = l0op::ViewCopy(rstdComputeOut, rstdOut, executor);
	101	+ CHECK_RET(viewCopyRstdResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
	102	+ }
	103	+
	104	+ return ACLNN_SUCCESS;
	105	+}
	106	+
	107	+aclnnStatus aclnnAddRmsNormDynamicMxQuantGetWorkspaceSize(
	108	+ const aclTensor* x1, const aclTensor* x2, const aclTensor* gamma, const aclTensor* beta, double epsilon,
	109	+ int64_t scaleAlg, char* roundMode, int64_t dstType, bool outputRstd, aclTensor* yOut, aclTensor* xOut,
	110	+ aclTensor* mxscaleOut, aclTensor* rstdOut, uint64_t* workspaceSize, aclOpExecutor** executor)
	111	+{
	112	+ OP_LOGD("Enter aclnnAddRmsNormDynamicMxQuantGetWorkspaceSize.");
	113	+ L2_DFX_PHASE_1(
	114	+ aclnnAddRmsNormDynamicMxQuant, DFX_IN(x1, x2, gamma, beta, epsilon, scaleAlg, roundMode, dstType, outputRstd),
	115	+ DFX_OUT(yOut, xOut, mxscaleOut, rstdOut));
	116	+
	117	+ // 创建OpExecutor
	118	+ auto uniqueExecutor = CREATE_EXECUTOR();
	119	+ CHECK_RET(uniqueExecutor.get() != nullptr, ACLNN_ERR_INNER_CREATE_EXECUTOR);
	120	+
	121	+ // 检查必选输入/输出是否为空指针
	122	+ CHECK_RET(CheckNotNull(x1, x2, gamma, yOut, xOut, mxscaleOut, outputRstd, rstdOut), ACLNN_ERR_PARAM_NULLPTR);
	123	+
	124	+ // 固定写法，将输入转换成连续的tensor，可选输入不做判空校验
	125	+ auto x1Cont = l0op::Contiguous(x1, uniqueExecutor.get());
	126	+ auto x2Cont = l0op::Contiguous(x2, uniqueExecutor.get());
	127	+ auto gammaCont = l0op::Contiguous(gamma, uniqueExecutor.get());
	128	+
	129	+ CHECK_RET(x1Cont != nullptr, ACLNN_ERR_INNER_NULLPTR);
	130	+ CHECK_RET(x2Cont != nullptr, ACLNN_ERR_INNER_NULLPTR);
	131	+ CHECK_RET(gammaCont != nullptr, ACLNN_ERR_INNER_NULLPTR);
	132	+
	133	+ const aclTensor* betaCont = (beta == nullptr) ? nullptr : l0op::Contiguous(beta, uniqueExecutor.get());
	134	+
	135	+ auto ret = ComputeAddRmsNormDynamicMxQuant(
	136	+ x1Cont, x2Cont, gammaCont, betaCont, epsilon, scaleAlg, roundMode, dstType, outputRstd, yOut, xOut, mxscaleOut,
	137	+ rstdOut, uniqueExecutor.get());
	138	+ CHECK_RET(ret == ACLNN_SUCCESS, ret);
	139	+
	140	+ // 获取计算过程中需要使用的workspace大小
	141	+ *workspaceSize = uniqueExecutor->GetWorkspaceSize();
	142	+ uniqueExecutor.ReleaseTo(executor);
	143	+ OP_LOGD("Finish aclnnAddRmsNormDynamicMxQuantGetWorkspaceSize.");
	144	+ return ACLNN_SUCCESS;
	145	+}
	146	+
	147	+aclnnStatus aclnnAddRmsNormDynamicMxQuant(
	148	+ void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, aclrtStream stream)
	149	+{
	150	+ L2_DFX_PHASE_2(aclnnAddRmsNormDynamicMxQuant);
	151	+ // 固定写法，调用框架能力，完成计算
	152	+ return CommonOpExecutorRun(workspace, workspaceSize, executor, stream);
	153	+}
	154	+
	155	+#ifdef __cplusplus
	156	+}
	157	+#endif

Anorm/add_rms_norm_dynamic_mx_quant/op_host/op_api/aclnn_add_rms_norm_dynamic_mx_quant.h+97-0

@@ -0,0 +1,97 @@
	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	+
	11	+#ifndef OP_API_INC_LEVEL2_ADD_RMS_NORM_DYNAMIC_MX_QUANT_H_
	12	+#define OP_API_INC_LEVEL2_ADD_RMS_NORM_DYNAMIC_MX_QUANT_H_
	13	+
	14	+#include "aclnn/aclnn_base.h"
	15	+#include "aclnn_util.h"
	16	+
	17	+#ifdef __cplusplus
	18	+extern "C" {
	19	+#endif
	20	+
	21	+/**
	22	+ * @brief aclnnAddRmsNormDynamicMxQuant的第一段接口，根据具体的计算流程，计算workspace大小。
	23	+ * @domain aclnn_ops_infer
	24	+ *
	25	+ * 算子功能：Add + RmsNorm + 量化计算的融合算子，将加法的计算结果做层归一化计算后进行量化，
	26	+ * 并将归一化计算结果，加法的计算结果，量化尺度和表示归一化后的标准差的倒数返回。
	27	+ * 计算公式：
	28	+ * x = x1 + x2
	29	+ * RmsNorm(x) = (x/RMS(x))*gamma) + beta
	30	+ * y, mxscale_out = dynamicMxQuant(RmsNorm(x), round_mode, scale_alg, dst_type)
	31	+ * rstd = 1/RMS(x)
	32	+ *
	33	+ * @param [in] x1:
	34	+ * 公式中的输入x1，数据类型支持FLOAT16、BFLOAT16，shape维度支持1-7维。
	35	+ * 支持非连续的Tensor，数据格式支持ND。
	36	+ * @param [in] x2:
	37	+ * 公式中的输入x2，数据类型支持FLOAT16、BFLOAT16，shape维度支持1-7维。
	38	+ * 支持非连续的Tensor，数据格式支持ND。
	39	+ * @param [in] gamma:
	40	+ * 公式中的输入gamma，数据类型默认与输入x1一致，若不一致，则显示设为FLOAT32，shape维度支持1维。
	41	+ * 支持非连续的Tensor，数据格式支持ND。
	42	+ * @param [in] beta:
	43	+ * 公式中的输入beta，数据类型和shape维度与gamma保持一致。
	44	+ * 支持非连续的Tensor，数据格式支持ND。
	45	+ * @param [in] epsilon: double 类型，层归一化中用到的防止除0的参数。
	46	+ * @param [in] scaleAlg:
	47	+ * 公式中的scale_alg，int 类型，表示mxscale_out的计算方法。
	48	+ * 支持取值0:OCP计算方法,和取值1:cuBLAS计算方法，当dst_type为FLOAT4_E2M1/FLOAT4_E1M2时仅支持取值为0。
	49	+ * @param [in] roundMode:
	50	+ * 公式中的round_mode，string 类型，数据转换的模式。
	51	+ * 对应yOut数据类型为FLOAT4_E2M1/FLOAT4_E1M2时，支持{"rint", "floor", "round"}，
	52	+ * 对应yOut数据类型为FLOAT8_E4M3FN/FLOAT8_E5M2时，仅支持{"rint"}。
	53	+ * @param [in] dstType:
	54	+ * 公式中的dst_type，int 类型，表示指定数据转换后yOut的类型。
	55	+ * 输入范围为{35, 36, 40, 41}，分别对应{FLOAT8_E5M2, FLOAT8_E4M3FN, FLOAT4_E2M1, FLOAT4_E1M2}。
	56	+ * @param [in] outputRstd: bool 类型，表示指定是否输出有效的rstdOut，当为False时，rstdOut为无效输出。
	57	+ * @param [out] yOut:
	58	+ * 公式中的输出y，数据类型支持FLOAT4_E2M1、FLOAT4_E1M2、FLOAT8_E4M3FN、FLOAT8_E5M2，shape需要与x1一致。
	59	+ * 支持非连续的Tensor，数据格式支持ND。
	60	+ * @param [out] xOut:
	61	+ * 公式中的输出x，数据类型支持FLOAT16、BFLOAT16且需要与x1一致，shape需要与x1一致。
	62	+ * 支持非连续的Tensor，数据格式支持ND。
	63	+ * @param [out] mxscaleOut:
	64	+ * 公式中的输出mxscale_out，每个分组对应的量化尺度。数据类型支持FLOAT8_E8M0，shape维度支持2-8维。
	65	+ * 支持非连续的Tensor，数据格式支持ND。
	66	+ * @param [out] rstdOut:
	67	+ * 公式中的输出rstd，数据类型支持FLOAT，shape维度与输入x1保持一致。
	68	+ * 支持非连续的Tensor，数据格式支持ND。
	69	+ * @param [out] workspaceSize: 返回用户需要在npu device侧申请的workspace大小。
	70	+ * @param [out] executor: 返回op执行器，包含算子计算流程。
	71	+ * @return aclnnStatus: 返回状态码。
	72	+ */
	73	+ACLNN_API aclnnStatus aclnnAddRmsNormDynamicMxQuantGetWorkspaceSize(
	74	+ const aclTensor* x1, const aclTensor* x2, const aclTensor* gamma, const aclTensor* beta, double epsilon,
	75	+ int64_t scaleAlg, char* roundMode, int64_t dstType, bool outputRstd, aclTensor* yOut, aclTensor* xOut, aclTensor* mxscaleOut,
	76	+ aclTensor* rstdOut, uint64_t* workspaceSize, aclOpExecutor** executor);
	77	+
	78	+/**
	79	+ * @brief aclnnAddRmsNormDynamicMxQuant的第二段接口，用于执行计算。
	80	+ *
	81	+ * 算子功能：执行Add+RmsNorm+DynamicMxQuant计算。
	82	+ *
	83	+ * @param [in] workspace: 在npu device侧申请的workspace内存起址。
	84	+ * @param [in] workspaceSize: 在npu
	85	+ * device侧申请的workspace大小，由第一段接口aclnnAddRmsNormDynamicMxQuantGetWorkspaceSize获取。
	86	+ * @param [in] executor: op执行器，包含了算子计算流程。
	87	+ * @param [in] stream: acl stream流。
	88	+ * @return aclnnStatus: 返回状态码。
	89	+ */
	90	+ACLNN_API aclnnStatus
	91	+aclnnAddRmsNormDynamicMxQuant(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, aclrtStream stream);
	92	+
	93	+#ifdef __cplusplus
	94	+}
	95	+#endif
	96	+
	97	+#endif // OP_API_INC_LEVEL2_ADD_RMS_NORM_DYNAMIC_MX_QUANT_H_

Anorm/add_rms_norm_dynamic_mx_quant/op_host/op_api/add_rms_norm_dynamic_mx_quant.cpp+52-0

@@ -0,0 +1,52 @@
	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	+
	11	+/*!
	12	+ * \file add_rms_norm_dynamic_mx_quant.cpp
	13	+ * \brief
	14	+ */
	15	+#include "add_rms_norm_dynamic_mx_quant.h"
	16	+#include "opdev/data_type_utils.h"
	17	+#include "opdev/format_utils.h"
	18	+#include "opdev/make_op_executor.h"
	19	+#include "opdev/op_def.h"
	20	+#include "opdev/op_dfx.h"
	21	+#include "opdev/op_executor.h"
	22	+#include "opdev/op_log.h"
	23	+#include "opdev/shape_utils.h"
	24	+#include "opdev/common_types.h"
	25	+#include "opdev/platform.h"
	26	+#include "aclnn_kernels/cast.h"
	27	+
	28	+using namespace op;
	29	+
	30	+namespace l0op {
	31	+OP_TYPE_REGISTER(AddRmsNormDynamicMxQuant); // 完成算子原型的关联
	32	+
	33	+const std::array<aclTensor*, ADD_RMS_NORM_DYNAMIC_MX_QUANT_OUT_NUM> AddRmsNormDynamicMxQuant(
	34	+ const aclTensor* x1, const aclTensor* x2, const aclTensor* gamma, const aclTensor* beta, double epsilon,
	35	+ int64_t scaleAlg, char* roundMode, int64_t dstType, bool outputRstd, aclTensor* yOut, aclTensor* xOut,
	36	+ aclTensor* mxscaleOut, aclTensor* rstdOut, aclOpExecutor* executor)
	37	+{
	38	+ // 调用接口，将算子执行任务加到执行器executor中
	39	+ L0_DFX(AddRmsNormDynamicMxQuant, x1, x2, gamma, beta, epsilon, scaleAlg, roundMode, dstType, outputRstd);
	40	+
	41	+ // 调用device的AddRmsNormDynamicMxQuant算子
	42	+ auto ret = ADD_TO_LAUNCHER_LIST_AICORE(
	43	+ AddRmsNormDynamicMxQuant, OP_INPUT(x1, x2, gamma, beta), OP_OUTPUT(yOut, xOut, mxscaleOut, rstdOut),
	44	+ OP_ATTR(static_cast<float>(epsilon), scaleAlg, roundMode, dstType, outputRstd));
	45	+ if (ret != ACL_SUCCESS) {
	46	+ OP_LOGE(ACLNN_ERR_INNER_NULLPTR, "AddRmsNormDynamicMxQuant ADD_TO_LAUNCHER_LIST_AICORE failed.");
	47	+ return {nullptr, nullptr, nullptr, nullptr};
	48	+ }
	49	+ return {yOut, xOut, mxscaleOut, rstdOut};
	50	+}
	51	+
	52	+} // namespace l0op

Anorm/add_rms_norm_dynamic_mx_quant/op_host/op_api/add_rms_norm_dynamic_mx_quant.h+30-0

@@ -0,0 +1,30 @@
	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	+
	11	+/*!
	12	+ * \file add_rms_norm_dynamic_mx_quant.h
	13	+ * \brief
	14	+ */
	15	+
	16	+#ifndef OP_API_INC_LEVEL0_ADD_RMS_NORM_DYNAMIC_MX_QUANT_H_
	17	+#define OP_API_INC_LEVEL0_ADD_RMS_NORM_DYNAMIC_MX_QUANT_H_
	18	+
	19	+#include "opdev/op_executor.h"
	20	+
	21	+namespace l0op {
	22	+constexpr size_t ADD_RMS_NORM_DYNAMIC_MX_QUANT_OUT_NUM = 4;
	23	+
	24	+const std::array<aclTensor*, ADD_RMS_NORM_DYNAMIC_MX_QUANT_OUT_NUM> AddRmsNormDynamicMxQuant(
	25	+ const aclTensor* x1, const aclTensor* x2, const aclTensor* gamma, const aclTensor* beta, double epsilon,
	26	+ int64_t scaleAlg, char* roundMode, int64_t dstType, bool outputRstd, aclTensor* yOut, aclTensor* xOut,
	27	+ aclTensor* mxscaleOut, aclTensor* rstdOut, aclOpExecutor* executor);
	28	+} // namespace l0op
	29	+
	30	+#endif // OP_API_INC_LEVEL0_ADD_RMS_NORM_DYNAMIC_MX_QUANT_H_

Mnorm/rms_norm_dynamic_mx_quant/CMakeLists.txt+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 of	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").	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/rms_norm_dynamic_mx_quant/README.md+12-12

	@@ -22,7 +22,7 @@
22	y = \operatorname{RmsNorm}(x)=\frac{x}{\operatorname{Rms}(\mathbf{x})}\cdot gamma+beta, \quad \text { where } \operatorname{Rms}(\mathbf{x})=\sqrt{\frac{1}{n} \sum_{i=1}^n x_i^2+epsilon}	22	y = \operatorname{RmsNorm}(x)=\frac{x}{\operatorname{Rms}(\mathbf{x})}\cdot gamma+beta, \quad \text { where } \operatorname{Rms}(\mathbf{x})=\sqrt{\frac{1}{n} \sum_{i=1}^n x_i^2+epsilon}
23	$$	23	$$
24		24
25	- 当scaleAlg为0时：	25	+ 当scale_alg为0时：
26	- 将RmsNorm输出y在尾轴维度上按k = 32个数分组，一组k个数 $\{\{V_i\}_{i=1}^{k}\}$ 动态量化为 $\{mxscale,\{P_i\}_{i=1}^{k}\}$	26	- 将RmsNorm输出y在尾轴维度上按k = 32个数分组，一组k个数 $\{\{V_i\}_{i=1}^{k}\}$ 动态量化为 $\{mxscale,\{P_i\}_{i=1}^{k}\}$
27	$$	27	$$
28	shared\_exp = floor(log_2(max_i(\|V_i\|))) - emax	28	shared\_exp = floor(log_2(max_i(\|V_i\|))) - emax
	@@ -44,7 +44,7 @@
44	\| FLOAT8_E4M3FN \| 8 \|	44	\| FLOAT8_E4M3FN \| 8 \|
45	\| FLOAT8_E5M2 \| 15 \|	45	\| FLOAT8_E5M2 \| 15 \|
46		46
47	- 当scaleAlg为1时，只涉及FP8类型：	47	+ 当scale_alg为1时，只涉及FP8类型：
48	- 将长向量按块分，每块长度为k，对每块单独计算一个块缩放因子$S_{fp32}^b$，再把块内所有元素用同一个$S_{fp32}^b$映射到目标低精度类型FP8。	48	- 将长向量按块分，每块长度为k，对每块单独计算一个块缩放因子$S_{fp32}^b$，再把块内所有元素用同一个$S_{fp32}^b$映射到目标低精度类型FP8。
49	- 找到该块中数值的最大绝对值:	49	- 找到该块中数值的最大绝对值:
50	$$	50	$$
	@@ -85,56 +85,56 @@
85	<tr>	85	<tr>
86	<td>x</td>	86	<td>x</td>
87	<td>输入</td>	87	<td>输入</td>
88	- <td>表示标准化过程中的源数据张量，对应公式中的`x`。</td>	88	+ <td>表示标准化过程中的源数据张量，对应公式中的x。</td>
89	<td>FLOAT16、BFLOAT16</td>	89	<td>FLOAT16、BFLOAT16</td>
90	<td>ND</td>	90	<td>ND</td>
91	</tr>	91	</tr>
92	<tr>	92	<tr>
93	<td>gamma</td>	93	<td>gamma</td>
94	<td>输入</td>	94	<td>输入</td>
95	- <td>表示标准化过程中的权重张量，对应公式中的`gamma`。shape需要与x最后一维一致。</td>	95	+ <td>表示标准化过程中的权重张量，对应公式中的gamma。shape需要与x最后一维一致。</td>
96	<td>FLOAT16、BFLOAT16、FLOAT32</td>	96	<td>FLOAT16、BFLOAT16、FLOAT32</td>
97	<td>ND</td>	97	<td>ND</td>
98	</tr>	98	</tr>
99	<tr>	99	<tr>
100	<td>beta</td>	100	<td>beta</td>
101	<td>可选输入</td>	101	<td>可选输入</td>
102	- <td>表示标准化过程中的偏置项，对应公式中的`beta`。shape必须与gamma一致。</td>	102	+ <td>表示标准化过程中的偏置项，对应公式中的beta。shape必须与gamma一致。</td>
103	<td>FLOAT16、BFLOAT16、FLOAT32</td>	103	<td>FLOAT16、BFLOAT16、FLOAT32</td>
104	<td>ND</td>	104	<td>ND</td>
105	</tr>	105	</tr>
106	<tr>	106	<tr>
107	<td>epsilon</td>	107	<td>epsilon</td>
108	<td>可选属性</td>	108	<td>可选属性</td>
109	- <td><ul><li>表示添加到分母中的值，以确保数值稳定。对应公式中的`epsilon`。</li><li>默认值为1e-6。</li></ul></td>	109	+ <td><ul><li>表示添加到分母中的值，以确保数值稳定。对应公式中的epsilon。</li><li>默认值为1e-6。</li></ul></td>
110	<td>FLOAT32</td>	110	<td>FLOAT32</td>
111	<td>-</td>	111	<td>-</td>
112	</tr>	112	</tr>
113	<tr>	113	<tr>
114	<td>scale_alg</td>	114	<td>scale_alg</td>
115	<td>可选属性</td>	115	<td>可选属性</td>
116	- <td><ul><li>表示mxscaleOut的计算方法，对应公式中的scaleAlg。</li><li>支持取值0和1，取值为0表示OCP实现，取值为1表示cuBLAS实现。当dstType为FLOAT4_E2M1/FLOAT4_E1M2时仅支持取值为0。</li><li>默认值为0。</li></ul></td>	116	+ <td><ul><li>表示mxscale的计算方法，对应公式中的scale_alg。</li><li>支持取值0和1，取值为0表示Open Compute Project(OCP)实现，取值为1表示cuBLAS实现。当dst_type为FLOAT4_E2M1/FLOAT4_E1M2时仅支持取值为0。</li><li>默认值为0。</li></ul></td>
117	<td>INT64</td>	117	<td>INT64</td>
118	<td>-</td>	118	<td>-</td>
119	</tr>	119	</tr>
120	<tr>	120	<tr>
121	<td>round_mode</td>	121	<td>round_mode</td>
122	<td>可选属性</td>	122	<td>可选属性</td>
123	- <td><ul><li>表示数据转换的模式，对应公式中的round_mode。</li><li>当dstType为40/41时，支持{"rint", "floor", "round"}。</li><li>当dstType为36/35时，仅支持{"rint"}。</li><li>默认值为"rint"。</li></ul></td>	123	+ <td><ul><li>表示数据转换的模式，对应公式中的round_mode。</li><li>当dst_type为40/41时，支持{"rint", "floor", "round"}。</li><li>当dst_type为36/35时，仅支持{"rint"}。</li><li>默认值为"rint"。</li></ul></td>
124	<td>STRING</td>	124	<td>STRING</td>
125	<td>-</td>	125	<td>-</td>
126	</tr>	126	</tr>
127	<tr>	127	<tr>
128	<td>dst_type</td>	128	<td>dst_type</td>
129	<td>可选属性</td>	129	<td>可选属性</td>
130	- <td><ul><li>表示指定数据转换后yOut的类型，对应公式中的DType。</li><li>输入范围为{35, 36, 40, 41}，分别对应{FLOAT8_E5M2, FLOAT8_E4M3FN, FLOAT4_E2M1, FLOAT4_E1M2}。</li><li>默认值为40。</li></ul></td>	130	+ <td><ul><li>表示指定数据转换后y的类型，对应公式中的DType。</li><li>输入范围为{35, 36, 40, 41}，分别对应{FLOAT8_E5M2, FLOAT8_E4M3FN, FLOAT4_E2M1, FLOAT4_E1M2}。</li><li>默认值为40。</li></ul></td>
131	<td>INT64</td>	131	<td>INT64</td>
132	<td>-</td>	132	<td>-</td>
133	</tr>	133	</tr>
134	<tr>	134	<tr>
135	<td>output_rstd</td>	135	<td>output_rstd</td>
136	<td>可选属性</td>	136	<td>可选属性</td>
137	- <td><ul><li>表示指定是否输出有效的rstdOut。</li><li>支持True和False。</li><li>默认值为False。</li><li>当outputRstd为False时，算子不写出rstdOut，返回shape[0]的空Tensor，此时rstdOut为无效占位输出。</li></ul></td>	137	+ <td><ul><li>表示指定是否输出有效的rstd。</li><li>支持True和False。</li><li>默认值为False。</li><li>当output_rstd为False时，算子不写出rstd，此时rstd为无效占位输出。</li></ul></td>
138	<td>BOOL</td>	138	<td>BOOL</td>
139	<td>-</td>	139	<td>-</td>
140	</tr>	140	</tr>
	@@ -155,7 +155,7 @@
155	<tr>	155	<tr>
156	<td>rstd</td>	156	<td>rstd</td>
157	<td>输出</td>	157	<td>输出</td>
158	- <td><ul><li>表示归一化后的标准差的倒数，对应公式中`Rms(x)`的倒数。</li><li>shape为与入参`x`的shape前几维保持一致，前几维指`x`的维度减去`gamma`的维度，表示不需要norm的维度。</li></ul></td>	158	+ <td><ul><li>表示归一化后的标准差的倒数，对应公式中Rms(x)的倒数。</li><li>shape为与入参x的shape前几维保持一致，前几维指x的维度减去gamma的维度，表示不需要norm的维度。</li></ul></td>
159	<td>FLOAT32</td>	159	<td>FLOAT32</td>
160	<td>ND</td>	160	<td>ND</td>
161	</tr>	161	</tr>
	@@ -171,7 +171,7 @@
171	- mxscale.shape[-1] = 2。	171	- mxscale.shape[-1] = 2。
172	- 其他维度与输入x一致。	172	- 其他维度与输入x一致。
173		173
174	-- 当输出yOut的数据类型为FLOAT4_E2M1或FLOAT4_E1M2，x尾轴的值必须为偶数。	174	+- 当输出y的数据类型为FLOAT4_E2M1或FLOAT4_E1M2，x尾轴的值必须为偶数。
175		175
176	- 输入gamma、可选输入beta的数据类型只能和x的数据类型保持一致或者为FLOAT32。	176	- 输入gamma、可选输入beta的数据类型只能和x的数据类型保持一致或者为FLOAT32。
177		177

Mnorm/rms_norm_dynamic_mx_quant/docs/aclnnRmsNormDynamicMxQuant.md+28-27

	@@ -74,7 +74,7 @@ aclnnStatus aclnnRmsNormDynamicMxQuantGetWorkspaceSize(
74	const aclTensor *beta,	74	const aclTensor *beta,
75	double epsilon,	75	double epsilon,
76	int64_t scaleAlg,	76	int64_t scaleAlg,
77	- char *roundModeOptional,	77	+ char *roundMode,
78	int64_t dstType,	78	int64_t dstType,
79	bool outputRstd,	79	bool outputRstd,
80	aclTensor *yOut,	80	aclTensor *yOut,
	@@ -121,8 +121,8 @@ aclnnStatus aclnnRmsNormDynamicMxQuant(
121	<tr>	121	<tr>
122	<td>x（aclTensor*）</td>	122	<td>x（aclTensor*）</td>
123	<td>输入</td>	123	<td>输入</td>
124	- <td>表示标准化过程中的源数据张量。对应公式中的`x`。</td>	124	+ <td>表示标准化过程中的源数据张量。对应公式中的x。</td>
125	- <td><ul><li>不支持空Tensor。</li></ul></td>	125	+ <td><ul><li>不支持空Tensor。</li><li>具体约束详见约束说明。</li></ul></td>
126	<td>FLOAT16、BFLOAT16</td>	126	<td>FLOAT16、BFLOAT16</td>
127	<td>ND</td>	127	<td>ND</td>
128	<td>1-7</td>	128	<td>1-7</td>
	@@ -131,8 +131,8 @@ aclnnStatus aclnnRmsNormDynamicMxQuant(
131	<tr>	131	<tr>
132	<td>gamma（aclTensor*）</td>	132	<td>gamma（aclTensor*）</td>
133	<td>输入</td>	133	<td>输入</td>
134	- <td>表示标准化过程中的权重张量。对应公式中的`gamma`。</td>	134	+ <td>表示标准化过程中的权重张量。对应公式中的gamma。</td>
135	- <td><ul><li>不支持空Tensor。</li><li>shape只能为一维，需要与`x`最后一维维度匹配。</li></ul></td>	135	+ <td><ul><li>不支持空Tensor。</li><li>shape只能为一维，需要与x最后一维维度匹配。</li><li>数据类型默认与输入x一致；若不一致，则显示设为FLOAT32。</li></ul></td>
136	<td>FLOAT16、BFLOAT16、FLOAT32</td>	136	<td>FLOAT16、BFLOAT16、FLOAT32</td>
137	<td>ND</td>	137	<td>ND</td>
138	<td>1</td>	138	<td>1</td>
	@@ -141,8 +141,8 @@ aclnnStatus aclnnRmsNormDynamicMxQuant(
141	<tr>	141	<tr>
142	<td>beta（aclTensor*）</td>	142	<td>beta（aclTensor*）</td>
143	<td>输入</td>	143	<td>输入</td>
144	- <td>表示标准化过程中的偏置项。对应公式中的`beta`。</td>	144	+ <td>表示标准化过程中的偏置项。对应公式中的beta。</td>
145	- <td><ul><li>不支持空Tensor。</li><li>可选参数，支持传入空指针。</li><li>如果提供，shape和数据类型需要与`gamma`保持一致。</li></ul></td>	145	+ <td><ul><li>不支持空Tensor。</li><li>可选参数，支持传入空指针。</li><li>如果提供，shape和数据类型需要与gamma保持一致。</li></ul></td>
146	<td>FLOAT16、BFLOAT16、FLOAT32</td>	146	<td>FLOAT16、BFLOAT16、FLOAT32</td>
147	<td>ND</td>	147	<td>ND</td>
148	<td>1</td>	148	<td>1</td>
	@@ -151,7 +151,7 @@ aclnnStatus aclnnRmsNormDynamicMxQuant(
151	<tr>	151	<tr>
152	<td>epsilon（double）</td>	152	<td>epsilon（double）</td>
153	<td>输入</td>	153	<td>输入</td>
154	- <td>表示添加到分母中的值，以确保数值稳定。对应公式中的`epsilon`。</td>	154	+ <td>表示添加到分母中的值，以确保数值稳定。对应公式中的epsilon。</td>
155	<td><ul><li>建议传入较小正数，如1e-6。</li></ul></td>	155	<td><ul><li>建议传入较小正数，如1e-6。</li></ul></td>
156	<td>-</td>	156	<td>-</td>
157	<td>-</td>	157	<td>-</td>
	@@ -169,7 +169,7 @@ aclnnStatus aclnnRmsNormDynamicMxQuant(
169	<td>-</td>	169	<td>-</td>
170	</tr>	170	</tr>
171	<tr>	171	<tr>
172	- <td>roundModeOptional（char*）</td>	172	+ <td>roundMode（char*）</td>
173	<td>输入</td>	173	<td>输入</td>
174	<td>表示数据转换的模式，对应公式中的round_mode。</td>	174	<td>表示数据转换的模式，对应公式中的round_mode。</td>
175	<td><ul><li>当dstType为40/41时，支持{"rint", "floor", "round"}。</li><li>当dstType为36/35时，仅支持{"rint"}。</li></ul></td>	175	<td><ul><li>当dstType为40/41时，支持{"rint", "floor", "round"}。</li><li>当dstType为36/35时，仅支持{"rint"}。</li></ul></td>
	@@ -202,7 +202,7 @@ aclnnStatus aclnnRmsNormDynamicMxQuant(
202	<td>yOut（aclTensor*）</td>	202	<td>yOut（aclTensor*）</td>
203	<td>输出</td>	203	<td>输出</td>
204	<td>表示归一化并量化后的结果，对应公式中的Pi和di。</td>	204	<td>表示归一化并量化后的结果，对应公式中的Pi和di。</td>
205	- <td><ul><li>不支持空Tensor。</li><li>shape需要与输入`x`一致。</li></ul></td>	205	+ <td><ul><li>不支持空Tensor。</li><li>shape需要与输入x一致。</li><li>数据类型需要与dstType保持一致。</li></ul></td>
206	<td>FLOAT4_E2M1、FLOAT4_E1M2、FLOAT8_E4M3FN、FLOAT8_E5M2</td>	206	<td>FLOAT4_E2M1、FLOAT4_E1M2、FLOAT8_E4M3FN、FLOAT8_E5M2</td>
207	<td>ND</td>	207	<td>ND</td>
208	<td>1-7</td>	208	<td>1-7</td>
	@@ -212,7 +212,7 @@ aclnnStatus aclnnRmsNormDynamicMxQuant(
212	<td>mxscaleOut（aclTensor*）</td>	212	<td>mxscaleOut（aclTensor*）</td>
213	<td>输出</td>	213	<td>输出</td>
214	<td>表示每个分组对应的量化尺度，对应公式中的mxscale和Sb。</td>	214	<td>表示每个分组对应的量化尺度，对应公式中的mxscale和Sb。</td>
215	- <td><ul><li>不支持空Tensor。</li><li>shape的维度数为输入`x`的维度数加1。其中，前几维（`x`的维度数减去`gamma`的维度数）与`x`对应维度保持一致，表示不需要norm的维度；后两维由MX量化决定：倒数第二维为ceil(`x`尾轴 / blocksize=32)再向上取偶数对齐（pad填充值为0），最后一维固定为2。具体计算过程见约束说明。</li></ul></td>	215	+ <td><ul><li>不支持空Tensor。</li><li>shape在尾轴上为x对应值除以blocksize=32向上取整，并对其进行偶数pad，pad填充值为0。</li><li>其余维度与入参x的shape前几维保持一致，前几维指x的维度减去gamma的维度，表示不需要norm的维度。</li><li>具体计算过程见约束说明。</li></ul></td>
216	<td>FLOAT8_E8M0</td>	216	<td>FLOAT8_E8M0</td>
217	<td>ND</td>	217	<td>ND</td>
218	<td>2-8</td>	218	<td>2-8</td>
	@@ -221,7 +221,7 @@ aclnnStatus aclnnRmsNormDynamicMxQuant(
221	<tr>	221	<tr>
222	<td>rstdOut（aclTensor*）</td>	222	<td>rstdOut（aclTensor*）</td>
223	<td>输出</td>	223	<td>输出</td>
224	- <td>表示归一化后的标准差的倒数。对应公式中`Rms(x)`的倒数。</td>	224	+ <td>表示归一化后的标准差的倒数。对应公式中Rms(x)的倒数。</td>
225	<td><ul><li>不支持空Tensor。</li><li>当outputRstd为True时，shape与入参`x`的shape前几维保持一致，前几维指`x`的维度减去`gamma`的维度，表示不需要norm的维度，rstdOut的-1轴是1。</li><li>当outputRstd为False时，该参数的最终输出无效。</li></ul></td>	225	<td><ul><li>不支持空Tensor。</li><li>当outputRstd为True时，shape与入参`x`的shape前几维保持一致，前几维指`x`的维度减去`gamma`的维度，表示不需要norm的维度，rstdOut的-1轴是1。</li><li>当outputRstd为False时，该参数的最终输出无效。</li></ul></td>
226	<td>FLOAT32</td>	226	<td>FLOAT32</td>
227	<td>ND</td>	227	<td>ND</td>
	@@ -271,37 +271,38 @@ aclnnStatus aclnnRmsNormDynamicMxQuant(
271	</thead>	271	</thead>
272	<tbody>	272	<tbody>
273	<tr>	273	<tr>
274	- <td>ACLNN_ERR_PARAM_NULLPTR</td>	274	+ <td rowspan="2">ACLNN_ERR_PARAM_NULLPTR</td>
275	- <td>161001</td>	275	+ <td rowspan="2">161001</td>
276	- <td>如果传入参数是必选输入，输出或者必选属性，且是空指针，则返回161001。</td>	276	+ <td>传入的x、gamma、yOut和mxscaleOut是空指针。</td>
277	</tr>	277	</tr>
278	<tr>	278	<tr>
279	- <td rowspan="2">ACLNN_ERR_PARAM_INVALID</td>	279	+ <td>当outputRstd为True时，传入rstdOut是空指针。</td>
280	- <td rowspan="2">161002</td>
281	- <td>输入或输出的数据类型不在支持的范围之内。</td>
282	</tr>	280	</tr>
283	<tr>	281	<tr>
284	- <td>输入或输出的参数不满足参数说明的约束。</td>	282	+ <td rowspan="8">ACLNN_ERR_INNER_TILING_ERROR</td>
		283	+ <td rowspan="8">561002</td>
		284	+ <td>输入或输出参数的维度数和数据类型不在范围之内，dstType和yOut的数据类型不匹配。</td>
285	</tr>	285	</tr>
286	<tr>	286	<tr>
287	- <td rowspan="6">ACLNN_ERR_INNER_TILING_ERROR</td>	287	+ <td>输入x和输出yOut的shape不是完全相同的shape。</td>
288	- <td rowspan="6">561002</td>
289	- <td>scaleAlg不是0或1，roundModeOptional(非空时)不是 {rint, floor, round}。</td>
290	</tr>	288	</tr>
291	<tr>	289	<tr>
292	- <td>dstType为 fp8 时，roundModeOptional不是 rint。</td>	290	+ <td>gamma、beta(若存在)的shape不是完全相同的shape，或二者的维度数不为1，或轴长不等于x的尾轴大小，或者类型不相同。</td>
293	</tr>	291	</tr>
294	<tr>	292	<tr>
295	- <td>x、yOut的shape不是完全相同的shape。</td>	293	+ <td>gamma的维度和x的需要做norm的维度不相同，或rstdOut的维度和x的不需要norm的维度不相同，或rstdOut对应x归一化维度的轴长不为1。</td>
296	</tr>	294	</tr>
297	<tr>	295	<tr>
298	- <td>mxscaleOut的维度数不等于输入维度数+1。</td>	296	+ <td>scaleAlg不是0或1，roundMode不是 {rint, floor, round}。</td>
299	</tr>	297	</tr>
300	<tr>	298	<tr>
301	- <td>gamma、beta(若存在)的shape不是完全相同的shape，或者类型不是相同的类型。</td>	299	+ <td>dstType为 fp8 时，roundMode不是 rint。</td>
302	</tr>	300	</tr>
303	<tr>	301	<tr>
304	- <td>gamma的维度和x的需要作norm的维度不相同，或rstdOut的维度和x的不需要norm的维度不相同，或rstdOut的需要norm的维度不为1。</td>	302	+ <td>dstType为 fp4 时，scaleAlg不是0，或者输入x的尾轴不能被2整除。</td>
		303	+ </tr>
		304	+ <tr>
		305	+ <td>mxscaleOut的维度数不等于输入x的维度数+1，或轴长不符合约束说明。</td>
305	</tr>	306	</tr>
306	</tbody></table>	307	</tbody></table>
307		308

Mnorm/rms_norm_dynamic_mx_quant/examples/test_aclnn_rms_norm_dynamic_mx_quant.cpp+1-1文件内容审核中，请稍后刷新重试

Mnorm/rms_norm_dynamic_mx_quant/op_graph/CMakeLists.txt+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 of	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").	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/rms_norm_dynamic_mx_quant/op_graph/rms_norm_dynamic_mx_quant_proto.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 of	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").	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/rms_norm_dynamic_mx_quant/op_host/CMakeLists.txt+2-2

	@@ -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 of	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").	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.
	@@ -8,4 +8,4 @@
8	# See LICENSE in the root of the software repository for the full text of the License.	8	# See LICENSE in the root of the software repository for the full text of the License.
9	# ----------------------------------------------------------------------------	9	# ----------------------------------------------------------------------------
10		10
11	-add_modules_sources(HOSTNAME ${OPHOST_NAME} MODE PRIVATE DIR ${CMAKE_CURRENT_SOURCE_DIR} OPTYPE rms_norm_dynamic_mx_quant ACLNNTYPE aclnn COMPUTE_UNIT "ascend950" TILING_DIR "arch35" DISABLE_IN_OPP TRUE)	11	+add_modules_sources(HOSTNAME ${OPHOST_NAME} MODE PRIVATE DIR ${CMAKE_CURRENT_SOURCE_DIR} OPTYPE rms_norm_dynamic_mx_quant ACLNNTYPE aclnn_exclude COMPUTE_UNIT "ascend950" TILING_DIR "arch35" DISABLE_IN_OPP TRUE)

Mnorm/rms_norm_dynamic_mx_quant/op_host/arch35/rms_norm_dynamic_mx_quant_base_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 of	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").	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/rms_norm_dynamic_mx_quant/op_host/arch35/rms_norm_dynamic_mx_quant_full_load_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 of	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").	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/rms_norm_dynamic_mx_quant/op_host/arch35/rms_norm_dynamic_mx_quant_tiling_arch35.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 of	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").	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.

Anorm/rms_norm_dynamic_mx_quant/op_host/op_api/aclnn_rms_norm_dynamic_mx_quant.cpp+146-0

@@ -0,0 +1,146 @@
	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	+#include "aclnn/aclnn_base.h"
	11	+#include "op_api/op_api_def.h"
	12	+
	13	+#include "opdev/common_types.h"
	14	+#include "opdev/data_type_utils.h"
	15	+#include "opdev/op_executor.h"
	16	+#include "opdev/op_log.h"
	17	+#include "opdev/format_utils.h"
	18	+#include "opdev/tensor_view_utils.h"
	19	+#include "opdev/op_dfx.h"
	20	+#include "opdev/shape_utils.h"
	21	+#include "opdev/platform.h"
	22	+
	23	+#include "aclnn_kernels/common/op_error_check.h"
	24	+#include "aclnn_kernels/cast.h"
	25	+#include "aclnn_kernels/contiguous.h"
	26	+#include "op_api/aclnn_util.h"
	27	+
	28	+#include "rms_norm_dynamic_mx_quant.h"
	29	+#include "aclnn_rms_norm_dynamic_mx_quant.h"
	30	+
	31	+using namespace op;
	32	+#ifdef __cplusplus
	33	+extern "C" {
	34	+#endif
	35	+
	36	+constexpr int IDX_0 = 0;
	37	+constexpr int IDX_1 = 1;
	38	+constexpr int IDX_2 = 2;
	39	+
	40	+static bool CheckNotNull(
	41	+ const aclTensor* x, const aclTensor* gamma, aclTensor* yOut,
	42	+ aclTensor* mxscaleOut, bool outputRstd, aclTensor* rstdOut)
	43	+{
	44	+ OP_CHECK_NULL(x, return false);
	45	+ OP_CHECK_NULL(gamma, return false);
	46	+ // beta 是可选输入，不校验nullptr
	47	+ OP_CHECK_NULL(yOut, return false);
	48	+ OP_CHECK_NULL(mxscaleOut, return false);
	49	+ // rstdOut 根据 outputRstd 决定是否校验
	50	+ if (outputRstd) {
	51	+ OP_CHECK_NULL(rstdOut, return false);
	52	+ }
	53	+ return true;
	54	+}
	55	+
	56	+static aclnnStatus ComputeRmsNormDynamicMxQuant(
	57	+ const aclTensor* x, const aclTensor* gamma, const aclTensor* beta, double epsilon,
	58	+ int64_t scaleAlg, char* roundMode, int64_t dstType, bool outputRstd, aclTensor* yOut,
	59	+ aclTensor* mxscaleOut, aclTensor* rstdOut, aclOpExecutor* executor)
	60	+{
	61	+ // 创建输出Tensor
	62	+ aclTensor* y_output = executor->AllocTensor(yOut->GetViewShape(), yOut->GetDataType(), yOut->GetViewFormat());
	63	+ aclTensor* mxscale_output =
	64	+ executor->AllocTensor(mxscaleOut->GetViewShape(), mxscaleOut->GetDataType(), mxscaleOut->GetViewFormat());
	65	+ aclTensor* rstd_output =
	66	+ (outputRstd) ?
	67	+ executor->AllocTensor(rstdOut->GetViewShape(), rstdOut->GetDataType(), rstdOut->GetViewFormat()) :
	68	+ executor->AllocTensor(op::Shape({0}), op::DataType::DT_FLOAT, op::Format::FORMAT_ND);
	W W wangyu_ai4月8日此条代码评论区间+65至+68 同上刘琦 4月9日评论： 4月9日评论：
	69	+
	70	+ auto RmsNormDynamicMxQuantOuts = l0op::RmsNormDynamicMxQuant(
	71	+ x, gamma, beta, epsilon, scaleAlg, roundMode, dstType, outputRstd, y_output, mxscale_output,
	72	+ rstd_output, executor);
	73	+
	74	+ auto yComputeOut = std::get<IDX_0>(RmsNormDynamicMxQuantOuts);
	75	+ auto mxscaleComputeOut = std::get<IDX_1>(RmsNormDynamicMxQuantOuts);
	76	+ auto rstdComputeOut = std::get<IDX_2>(RmsNormDynamicMxQuantOuts);
	77	+
	78	+ // 校验输出不为空
	79	+ CHECK_RET(yComputeOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
	80	+ CHECK_RET(mxscaleComputeOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
	81	+
	82	+ // 将结果拷贝到输出tensor
	83	+ auto viewCopyYResult = l0op::ViewCopy(yComputeOut, yOut, executor);
	84	+ CHECK_RET(viewCopyYResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
	85	+
	86	+ auto viewCopyMxscaleResult = l0op::ViewCopy(mxscaleComputeOut, mxscaleOut, executor);
	87	+ CHECK_RET(viewCopyMxscaleResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
	88	+
	89	+ if (outputRstd) {
	90	+ CHECK_RET(rstdComputeOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
	91	+ auto viewCopyRstdResult = l0op::ViewCopy(rstdComputeOut, rstdOut, executor);
	92	+ CHECK_RET(viewCopyRstdResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
	93	+ }
	94	+
	95	+ return ACLNN_SUCCESS;
	96	+}
	97	+
	98	+aclnnStatus aclnnRmsNormDynamicMxQuantGetWorkspaceSize(
	99	+ const aclTensor* x, const aclTensor* gamma, const aclTensor* beta, double epsilon,
	100	+ int64_t scaleAlg, char* roundMode, int64_t dstType, bool outputRstd, aclTensor* yOut,
	101	+ aclTensor* mxscaleOut, aclTensor* rstdOut, uint64_t* workspaceSize, aclOpExecutor** executor)
	102	+{
	103	+ OP_LOGD("Enter aclnnRmsNormDynamicMxQuantGetWorkspaceSize.");
	104	+ L2_DFX_PHASE_1(
	105	+ aclnnRmsNormDynamicMxQuant, DFX_IN(x, gamma, beta, epsilon, scaleAlg, roundMode, dstType, outputRstd),
	106	+ DFX_OUT(yOut, mxscaleOut, rstdOut));
	107	+
	108	+ // 创建OpExecutor
	109	+ auto uniqueExecutor = CREATE_EXECUTOR();
	110	+ CHECK_RET(uniqueExecutor.get() != nullptr, ACLNN_ERR_INNER_CREATE_EXECUTOR);
	111	+
	112	+ // 检查必选输入/输出是否为空指针
	113	+ CHECK_RET(CheckNotNull(x, gamma, yOut, mxscaleOut, outputRstd, rstdOut), ACLNN_ERR_PARAM_NULLPTR);
	114	+
	115	+ // 固定写法，将输入转换成连续的tensor，可选输入不做判空校验
	116	+ auto xCont = l0op::Contiguous(x, uniqueExecutor.get());
	117	+ auto gammaCont = l0op::Contiguous(gamma, uniqueExecutor.get());
	118	+
	119	+ CHECK_RET(xCont != nullptr, ACLNN_ERR_INNER_NULLPTR);
	120	+ CHECK_RET(gammaCont != nullptr, ACLNN_ERR_INNER_NULLPTR);
	121	+
	122	+ const aclTensor* betaCont = (beta == nullptr) ? nullptr : l0op::Contiguous(beta, uniqueExecutor.get());
	123	+
	124	+ auto ret = ComputeRmsNormDynamicMxQuant(
	125	+ xCont, gammaCont, betaCont, epsilon, scaleAlg, roundMode, dstType, outputRstd, yOut, mxscaleOut,
	126	+ rstdOut, uniqueExecutor.get());
	127	+ CHECK_RET(ret == ACLNN_SUCCESS, ret);
	128	+
	129	+ // 获取计算过程中需要使用的workspace大小
	130	+ *workspaceSize = uniqueExecutor->GetWorkspaceSize();
	131	+ uniqueExecutor.ReleaseTo(executor);
	132	+ OP_LOGD("Finish aclnnRmsNormDynamicMxQuantGetWorkspaceSize.");
	133	+ return ACLNN_SUCCESS;
	134	+}
	135	+
	136	+aclnnStatus aclnnRmsNormDynamicMxQuant(
	137	+ void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, aclrtStream stream)
	138	+{
	139	+ L2_DFX_PHASE_2(aclnnRmsNormDynamicMxQuant);
	140	+ // 固定写法，调用框架能力，完成计算
	141	+ return CommonOpExecutorRun(workspace, workspaceSize, executor, stream);
	142	+}
	143	+
	144	+#ifdef __cplusplus
	145	+}
	146	+#endif

Anorm/rms_norm_dynamic_mx_quant/op_host/op_api/aclnn_rms_norm_dynamic_mx_quant.h+90-0

@@ -0,0 +1,90 @@
	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	+
	11	+#ifndef OP_API_INC_LEVEL2_RMS_NORM_DYNAMIC_MX_QUANT_H_
	12	+#define OP_API_INC_LEVEL2_RMS_NORM_DYNAMIC_MX_QUANT_H_
	13	+
	14	+#include "aclnn/aclnn_base.h"
	15	+#include "aclnn_util.h"
	16	+
	17	+#ifdef __cplusplus
	18	+extern "C" {
	19	+#endif
	20	+
	21	+/**
	22	+ * @brief aclnnRmsNormDynamicMxQuant的第一段接口，根据具体的计算流程，计算workspace大小。
	23	+ * @domain aclnn_ops_infer
	24	+ *
	25	+ * 算子功能：RmsNorm + 量化计算的融合算子，将加法的计算结果做层归一化计算后进行量化，
	26	+ * 并将归一化计算结果，加法的计算结果，量化尺度和表示归一化后的标准差的倒数返回。
	27	+ * 计算公式：
	28	+ * RmsNorm(x) = (x/RMS(x))*gamma) + beta
	29	+ * y, mxscale_out = dynamicMxQuant(RmsNorm(x), round_mode, scale_alg, dst_type)
	30	+ * rstd = 1/RMS(x)
	31	+ *
	32	+ * @param [in] x:
	33	+ * 公式中的输入x1，数据类型支持FLOAT16、BFLOAT16，shape维度支持1-7维。
	34	+ * 支持非连续的Tensor，数据格式支持ND。
	35	+ * @param [in] gamma:
	36	+ * 公式中的输入gamma，数据类型默认与输入x1一致，若不一致，则显示设为FLOAT32，shape维度支持1维。
	37	+ * 支持非连续的Tensor，数据格式支持ND。
	38	+ * @param [in] beta:
	39	+ * 公式中的输入beta，数据类型和shape维度与gamma保持一致。
	40	+ * 支持非连续的Tensor，数据格式支持ND。
	41	+ * @param [in] epsilon: double 类型，层归一化中用到的防止除0的参数。
	42	+ * @param [in] scaleAlg:
	43	+ * 公式中的scale_alg，int 类型，表示mxscale_out的计算方法。
	44	+ * 支持取值0:OCP计算方法,和取值1:cuBLAS计算方法，当dst_type为FLOAT4_E2M1/FLOAT4_E1M2时仅支持取值为0。
	45	+ * @param [in] roundMode:
	46	+ * 公式中的round_mode，string 类型，数据转换的模式。
	47	+ * 对应yOut数据类型为FLOAT4_E2M1/FLOAT4_E1M2时，支持{"rint", "floor", "round"}，
	48	+ * 对应yOut数据类型为FLOAT8_E4M3FN/FLOAT8_E5M2时，仅支持{"rint"}。
	49	+ * @param [in] dstType:
	50	+ * 公式中的dst_type，int 类型，表示指定数据转换后yOut的类型。
	51	+ * 输入范围为{35, 36, 40, 41}，分别对应{FLOAT8_E5M2, FLOAT8_E4M3FN, FLOAT4_E2M1, FLOAT4_E1M2}。
	52	+ * @param [in] outputRstd: bool 类型，表示指定是否输出有效的rstdOut，当为False时，rstdOut为无效输出。
	53	+ * @param [out] yOut:
	54	+ * 公式中的输出y，数据类型支持FLOAT4_E2M1、FLOAT4_E1M2、FLOAT8_E4M3FN、FLOAT8_E5M2，shape需要与x一致。
	55	+ * 支持非连续的Tensor，数据格式支持ND。
	56	+ * @param [out] mxscaleOut:
	57	+ * 公式中的输出mxscale_out，每个分组对应的量化尺度。数据类型支持FLOAT8_E8M0，shape维度支持2-8维。
	58	+ * 支持非连续的Tensor，数据格式支持ND。
	59	+ * @param [out] rstdOut:
	60	+ * 公式中的输出rstd，数据类型支持FLOAT，shape维度与输入x保持一致。
	61	+ * 支持空Tensor，数据格式支持ND。
	62	+ * @param [out] workspaceSize: 返回用户需要在npu device侧申请的workspace大小。
	63	+ * @param [out] executor: 返回op执行器，包含算子计算流程。
	64	+ * @return aclnnStatus: 返回状态码。
	65	+ */
	66	+ACLNN_API aclnnStatus aclnnRmsNormDynamicMxQuantGetWorkspaceSize(
	67	+ const aclTensor* x, const aclTensor* gamma, const aclTensor* beta, double epsilon,
	68	+ int64_t scaleAlg, char* roundMode, int64_t dstType, bool outputRstd, aclTensor* yOut,
	69	+ aclTensor* mxscaleOut, aclTensor* rstdOut, uint64_t* workspaceSize, aclOpExecutor** executor);
	70	+
	71	+/**
	72	+ * @brief aclnnRmsNormDynamicMxQuant的第二段接口，用于执行计算。
	73	+ *
	74	+ * 算子功能：执行RmsNorm+DynamicMxQuant计算。
	75	+ *
	76	+ * @param [in] workspace: 在npu device侧申请的workspace内存起址。
	77	+ * @param [in] workspaceSize: 在npu
	78	+ * device侧申请的workspace大小，由第一段接口aclnnRmsNormDynamicMxQuantGetWorkspaceSize获取。
	79	+ * @param [in] executor: op执行器，包含了算子计算流程。
	80	+ * @param [in] stream: acl stream流。
	81	+ * @return aclnnStatus: 返回状态码。
	82	+ */
	83	+ACLNN_API aclnnStatus
	84	+aclnnRmsNormDynamicMxQuant(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, aclrtStream stream);
	85	+
	86	+#ifdef __cplusplus
	87	+}
	88	+#endif
	89	+
	90	+#endif // OP_API_INC_LEVEL2_RMS_NORM_DYNAMIC_MX_QUANT_H_

Anorm/rms_norm_dynamic_mx_quant/op_host/op_api/rms_norm_dynamic_mx_quant.cpp+52-0

@@ -0,0 +1,52 @@
	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	+
	11	+/*!
	12	+ * \file rms_norm_dynamic_mx_quant.cpp
	13	+ * \brief
	14	+ */
	15	+#include "rms_norm_dynamic_mx_quant.h"
	16	+#include "opdev/data_type_utils.h"
	17	+#include "opdev/format_utils.h"
	18	+#include "opdev/make_op_executor.h"
	19	+#include "opdev/op_def.h"
	20	+#include "opdev/op_dfx.h"
	21	+#include "opdev/op_executor.h"
	22	+#include "opdev/op_log.h"
	23	+#include "opdev/shape_utils.h"
	24	+#include "opdev/common_types.h"
	25	+#include "opdev/platform.h"
	26	+#include "aclnn_kernels/cast.h"
	27	+
	28	+using namespace op;
	29	+
	30	+namespace l0op {
	31	+OP_TYPE_REGISTER(RmsNormDynamicMxQuant); // 完成算子原型的关联
	32	+
	33	+const std::array<aclTensor*, RMS_NORM_DYNAMIC_MX_QUANT_OUT_NUM> RmsNormDynamicMxQuant(
	34	+ const aclTensor* x, const aclTensor* gamma, const aclTensor* beta, double epsilon,
	35	+ int64_t scaleAlg, char* roundMode, int64_t dstType, bool outputRstd, aclTensor* yOut,
	36	+ aclTensor* mxscaleOut, aclTensor* rstdOut, aclOpExecutor* executor)
	37	+{
	38	+ // 调用接口，将算子执行任务加到执行器executor中
	39	+ L0_DFX(RmsNormDynamicMxQuant, x, gamma, beta, epsilon, scaleAlg, roundMode, dstType, outputRstd);
	40	+
	41	+ // 调用device的RmsNormDynamicMxQuant算子
	42	+ auto ret = ADD_TO_LAUNCHER_LIST_AICORE(
	43	+ RmsNormDynamicMxQuant, OP_INPUT(x, gamma, beta), OP_OUTPUT(yOut, mxscaleOut, rstdOut),
	44	+ OP_ATTR(static_cast<float>(epsilon), scaleAlg, roundMode, dstType, outputRstd));
	45	+ if (ret != ACL_SUCCESS) {
	46	+ OP_LOGE(ACLNN_ERR_INNER_NULLPTR, "RmsNormDynamicMxQuant ADD_TO_LAUNCHER_LIST_AICORE failed.");
	47	+ return {nullptr, nullptr, nullptr};
	48	+ }
	49	+ return {yOut, mxscaleOut, rstdOut};
	50	+}
	51	+
	52	+} // namespace l0op

Anorm/rms_norm_dynamic_mx_quant/op_host/op_api/rms_norm_dynamic_mx_quant.h+30-0

@@ -0,0 +1,30 @@
	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	+
	11	+/*!
	12	+ * \file rms_norm_dynamic_mx_quant.h
	13	+ * \brief
	14	+ */
	15	+
	16	+#ifndef OP_API_INC_LEVEL0_RMS_NORM_DYNAMIC_MX_QUANT_H_
	17	+#define OP_API_INC_LEVEL0_RMS_NORM_DYNAMIC_MX_QUANT_H_
	18	+
	19	+#include "opdev/op_executor.h"
	20	+
	21	+namespace l0op {
	22	+constexpr size_t RMS_NORM_DYNAMIC_MX_QUANT_OUT_NUM = 3;
	23	+
	24	+const std::array<aclTensor*, RMS_NORM_DYNAMIC_MX_QUANT_OUT_NUM> RmsNormDynamicMxQuant(
	25	+ const aclTensor* x, const aclTensor* gamma, const aclTensor* beta, double epsilon,
	26	+ int64_t scaleAlg, char* roundMode, int64_t dstType, bool outputRstd, aclTensor* yOut,
	27	+ aclTensor* mxscaleOut, aclTensor* rstdOut, aclOpExecutor* executor);
	28	+} // namespace l0op
	29	+
	30	+#endif // OP_API_INC_LEVEL0_RMS_NORM_DYNAMIC_MX_QUANT_H_

Mnorm/rms_norm_dynamic_mx_quant/op_host/rms_norm_dynamic_mx_quant_def.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 of	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").	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/rms_norm_dynamic_mx_quant/op_host/rms_norm_dynamic_mx_quant_infershape.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 of	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").	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/rms_norm_dynamic_mx_quant/op_kernel/arch35/rms_norm_dynamic_mx_quant_common.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 of	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").	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/rms_norm_dynamic_mx_quant/op_kernel/arch35/rms_norm_dynamic_mx_quant_full_load.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 of	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").	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/rms_norm_dynamic_mx_quant/op_kernel/rms_norm_dynamic_mx_quant_apt.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 of	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").	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/rms_norm_dynamic_mx_quant/tests/CMakeLists.txt+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 of	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").	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/rms_norm_dynamic_mx_quant/tests/ut/op_host/arch35/test_rms_norm_dynamic_mx_quant_tiling_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 of	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").	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/rms_norm_dynamic_mx_quant/tests/ut/op_host/test_rms_norm_dynamic_mx_quant_infershape.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 of	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").	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.