Merged
docs: 修正 MaskedSoftmaxWithRelPosBias 和 ExpandIntoJaggedPermute 输出非连续支持说明 #14401
Nice tryAdd on 8 days ago
docs: 修正 MaskedSoftmaxWithRelPosBias 和 ExpandIntoJaggedPermute 输出非连续支持说明 #14401
Merged
Merged fromNice_try/ops-nn-issuefix-1113intocann/ops-nnmaster
Total: 4 files changed+16-10
| @@ -127,11 +127,11 @@ aclnnStatus aclnnExpandIntoJaggedPermute( | |||
| 127 | <td>out</td> | 127 | <td>out</td> |
| 128 | <td>输出</td> | 128 | <td>输出</td> |
| 129 | <td>公式中的输出。</td> | 129 | <td>公式中的输出。</td> |
| 130 | - <td>shape为(outputSize)。</td> | 130 | + <td>shape为(outputSize),仅支持连续tensor。</td> |
| 131 | <td>与permute一致。</td> | 131 | <td>与permute一致。</td> |
| 132 | <td>ND</td> | 132 | <td>ND</td> |
| 133 | <td>-</td> | 133 | <td>-</td> |
| 134 | - <td>√</td> | 134 | + <td>×</td> |
| 135 | </tr> | 135 | </tr> |
| 136 | <tr> | 136 | <tr> |
| 137 | <td>workspaceSize</td> | 137 | <td>workspaceSize</td> |
| @@ -251,6 +251,8 @@ aclnnStatus aclnnExpandIntoJaggedPermute( | |||
| 251 | 251 | ||
| 252 | 4. 设P为permute的长度。P=0时outputSize必须为0;P>0时outputSize必须大于等于P。 | 252 | 4. 设P为permute的长度。P=0时outputSize必须为0;P>0时outputSize必须大于等于P。 |
| 253 | 253 | ||
| 254 | +5. out不支持非连续Tensor,仅支持连续Tensor。 | ||
| 255 | + | ||
| 254 | ## 调用示例 | 256 | ## 调用示例 |
| 255 | 257 | ||
| 256 | 示例代码如下,仅供参考,具体编译和执行过程请参考[编译与运行样例](../../../docs/zh/context/compile_and_run_sample.md)。 | 258 | 示例代码如下,仅供参考,具体编译和执行过程请参考[编译与运行样例](../../../docs/zh/context/compile_and_run_sample.md)。 |
| @@ -129,11 +129,11 @@ aclnnStatus aclnnExpandIntoJaggedPermute( | |||
| 129 | <td>out</td> | 129 | <td>out</td> |
| 130 | <td>Output</td> | 130 | <td>Output</td> |
| 131 | <td>Output in the formula.</td> | 131 | <td>Output in the formula.</td> |
| 132 | - <td>The shape is (outputSize).</td> | 132 | + <td>The shape is (outputSize). Only contiguous tensors are supported.</td> |
| 133 | <td>Consistent with permute.</td> | 133 | <td>Consistent with permute.</td> |
| 134 | <td>ND</td> | 134 | <td>ND</td> |
| 135 | <td>-</td> | 135 | <td>-</td> |
| 136 | - <td>√</td> | 136 | + <td>×</td> |
| 137 | </tr> | 137 | </tr> |
| 138 | <tr> | 138 | <tr> |
| 139 | <td>workspaceSize</td> | 139 | <td>workspaceSize</td> |
| @@ -255,6 +255,8 @@ aclnnStatus aclnnExpandIntoJaggedPermute( | |||
| 255 | 4. Let P be the length of permute. When P is 0, outputSize must be 0. When P is greater than 0, outputSize must be | 255 | 4. Let P be the length of permute. When P is 0, outputSize must be 0. When P is greater than 0, outputSize must be |
| 256 | greater than or equal to P. | 256 | greater than or equal to P. |
| 257 | 257 | ||
| 258 | +5. out does not support non-contiguous tensors and must be a contiguous tensor. | ||
| 259 | + | ||
| 258 | ## Example | 260 | ## Example |
| 259 | 261 | ||
| 260 | The following sample code is for reference only. For details about compilation and execution, see [Compile and Run Samples](../../../docs/en/context/compile_and_run_sample.md). | 262 | The following sample code is for reference only. For details about compilation and execution, see [Compile and Run Samples](../../../docs/en/context/compile_and_run_sample.md). |
| @@ -128,10 +128,10 @@ aclnnStatus aclnnMaskedSoftmaxWithRelPosBias( | |||
| 128 | <td>out</td> | 128 | <td>out</td> |
| 129 | <td>输出</td> | 129 | <td>输出</td> |
| 130 | <td>待计算出参,对应公式中的out。</td> | 130 | <td>待计算出参,对应公式中的out。</td> |
| 131 | - <td>shape与x相同。</td> | 131 | + <td>shape与x相同,仅支持连续tensor。</td> |
| 132 | <td>FLOAT、FLOAT16、BFLOAT16</td> | 132 | <td>FLOAT、FLOAT16、BFLOAT16</td> |
| 133 | <td>ND</td> | 133 | <td>ND</td> |
| 134 | - <td>√</td> | 134 | + <td>×</td> |
| 135 | </tr> | 135 | </tr> |
| 136 | <tr> | 136 | <tr> |
| 137 | <td>workspaceSize</td> | 137 | <td>workspaceSize</td> |
| @@ -246,7 +246,8 @@ aclnnStatus aclnnMaskedSoftmaxWithRelPosBias( | |||
| 246 | - innerPrecisionMode为保留字段,当前仅支持取值0,传入其他值时算子报错。 | 246 | - innerPrecisionMode为保留字段,当前仅支持取值0,传入其他值时算子报错。 |
| 247 | 247 | ||
| 248 | - 非连续Tensor支持: | 248 | - 非连续Tensor支持: |
| 249 | - - x、attenMaskOptional、relativePosBias、out均支持非连续Tensor(即strides不满足连续存储条件的Tensor),传入后由框架自动转为连续Tensor再参与计算。 | 249 | + - x、attenMaskOptional、relativePosBias支持非连续Tensor(即strides不满足连续存储条件的Tensor),传入后由框架自动转为连续Tensor再参与计算。 |
| 250 | + - out不支持非连续Tensor,仅支持连续Tensor。 | ||
| 250 | 251 | ||
| 251 | <!-- npu="310p" id8 --> | 252 | <!-- npu="310p" id8 --> |
| 252 | - <term>Atlas推理系列产品</term>:不支持入参x的最后一个维度S2非32Byte对齐的场景。 | 253 | - <term>Atlas推理系列产品</term>:不支持入参x的最后一个维度S2非32Byte对齐的场景。 |
| @@ -130,10 +130,10 @@ aclnnStatus aclnnMaskedSoftmaxWithRelPosBias( | |||
| 130 | <td>out</td> | 130 | <td>out</td> |
| 131 | <td>Output</td> | 131 | <td>Output</td> |
| 132 | <td>Output to be computed, corresponding to out in the formula.</td> | 132 | <td>Output to be computed, corresponding to out in the formula.</td> |
| 133 | - <td>Same shape as x.</td> | 133 | + <td>Same shape as x. Only contiguous tensors are supported.</td> |
| 134 | <td>FLOAT, FLOAT16, BFLOAT16</td> | 134 | <td>FLOAT, FLOAT16, BFLOAT16</td> |
| 135 | <td>ND</td> | 135 | <td>ND</td> |
| 136 | - <td>√</td> | 136 | + <td>×</td> |
| 137 | </tr> | 137 | </tr> |
| 138 | <tr> | 138 | <tr> |
| 139 | <td>workspaceSize</td> | 139 | <td>workspaceSize</td> |
| @@ -251,7 +251,8 @@ aclnnStatus aclnnMaskedSoftmaxWithRelPosBias( | |||
| 251 | - innerPrecisionMode is a reserved field. Currently, only the value 0 is supported. The operator reports an error if any other value is passed in. | 251 | - innerPrecisionMode is a reserved field. Currently, only the value 0 is supported. The operator reports an error if any other value is passed in. |
| 252 | 252 | ||
| 253 | - Non-contiguous tensor support: | 253 | - Non-contiguous tensor support: |
| 254 | - - x, attenMaskOptional, relativePosBias, and out support non-contiguous tensors (that is, tensors whose strides do not meet the contiguous storage condition). After being passed in, they are automatically converted to contiguous tensors by the framework before computation. | 254 | + - x, attenMaskOptional, and relativePosBias support non-contiguous tensors (that is, tensors whose strides do not meet the contiguous storage condition). After being passed in, they are automatically converted to contiguous tensors by the framework before computation. |
| 255 | + - out does not support non-contiguous tensors and must be a contiguous tensor. | ||
| 255 | 256 | ||
| 256 | <!-- npu="310p" id8 --> | 257 | <!-- npu="310p" id8 --> |
| 257 | - <term>Atlas inference products</term>: The scenario where the last dimension S2 of the input parameter x is not aligned to 32 bytes is not supported. | 258 | - <term>Atlas inference products</term>: The scenario where the last dimension S2 of the input parameter x is not aligned to 32 bytes is not supported. |