Star857
1350
代码
代码
Issues1663
Pull Requests2219
项目讨论
Wiki
分析
Star857
1350
已合并
modify document #45358
m0_68000167创建于 12 天前
modify document #45358
已合并
m0_68000167创建于 12 天前
从m0_68000167/pytorchv2.7.1-26.1.0合入到Ascend/pytorchv2.7.1-26.1.0
共 6 个文件变更+34-34
Mdocs/en/framework_feature_guide_pytorch/adaptation_development_opplugin.md+8-8
@@ -14,9 +14,9 @@
14│ │ └── opapi # aclnn operator adapter directory14│ │ └── opapi # aclnn operator adapter directory
15│ │ ├── AbsKernelNpuOpApi.cpp15│ │ ├── AbsKernelNpuOpApi.cpp
16│ │ └── ...16│ │ └── ...
17-| ├── python17+│ ├── python
18-| │ └── meta18+│ │ └── meta
19-| │ └── _meta_registrations.py # Operator meta implementation registration file19+│ │ └── _meta_registrations.py # Operator meta implementation registration file
20│ ├── OpInterface.h # Auto-generated header file for op_plugin external interfaces (for framework-side operator invocation)20│ ├── OpInterface.h # Auto-generated header file for op_plugin external interfaces (for framework-side operator invocation)
21│ ├── OpInterface.cpp # Auto-generated routing implementation for op_plugin external interfaces (internal branch selection logic for different operator types)21│ ├── OpInterface.cpp # Auto-generated routing implementation for op_plugin external interfaces (internal branch selection logic for different operator types)
22│ ├── AclOpsInterface.h # Auto-generated header file for aclop operator plugin adapter 22│ ├── AclOpsInterface.h # Auto-generated header file for aclop operator plugin adapter
@@ -28,8 +28,8 @@
28├── test28├── test
29│ ├── allowlist_for_publicAPI.json # Public interface allowlist29│ ├── allowlist_for_publicAPI.json # Public interface allowlist
30│ ├── core_tests30│ ├── core_tests
31-│ | └── torch_npu_OpApi_schema_all.json # Operator interface schema configuration file31+│ │ └── torch_npu_OpApi_schema_all.json # Operator interface schema configuration file
32-| └── test_custom_ops # Operator developer test directory32+│ └── test_custom_ops # Operator developer test directory
33│ └── ...33│ └── ...
34```34```
35 35 
@@ -72,7 +72,7 @@ Parameters
72 72 
73- `all_version`: Indicates all versions supported by the current PyTorch. The version list will be adjusted as torch_npu evolves, and the actual code shall prevail. You can use `[]` to specify the version range supported by an operator. For example, `[v2.1, newest]` indicates that the operator supports versions from v2.1 to the latest.73- `all_version`: Indicates all versions supported by the current PyTorch. The version list will be adjusted as torch_npu evolves, and the actual code shall prevail. You can use `[]` to specify the version range supported by an operator. For example, `[v2.1, newest]` indicates that the operator supports versions from v2.1 to the latest.
74- `official` and `custom`: Respectively indicate that operators under these fields are PyTorch native operators and custom operators. The `symint` field indicates that the operator supports symint-type input parameters. For such operators, refer to [Symint Operator Adaptation](#symint-operator-adaptation).74- `official` and `custom`: Respectively indicate that operators under these fields are PyTorch native operators and custom operators. The `symint` field indicates that the operator supports symint-type input parameters. For such operators, refer to [Symint Operator Adaptation](#symint-operator-adaptation).
75-- `func`: Defines the operator schema (operator description specification). Its content fully follows the schema definition rules of PyTorch native Aten IR operators and completely describes the operator calling interface and semantic constraints through the structured form of "operator name + input parameter list + return parameter". For specific rules, refer to [PyTorch schema rules](reference.md#section001).75+- `func`: Defines the operator schema (operator description specification). Its content fully follows the schema definition rules of PyTorch native Aten IR operators and completely describes the operator calling interface and semantic constraints through the structured form of "operator name + input parameter list + return parameter". For specific rules, refer to [PyTorch schema rules](reference.md#pytorch-schema-rules).
76- `acl_op`: Indicates that acl_op calls are supported in this version. If the supported versions are the same as the versions indicated by `all_version`, you can use `all_version` instead. This is an optional field.76- `acl_op`: Indicates that acl_op calls are supported in this version. If the supported versions are the same as the versions indicated by `all_version`, you can use `all_version` instead. This is an optional field.
77- `op_api`: Indicates that op_api calls are supported in this version. If the supported versions are the same as the versions indicated by `all_version`, you can use `all_version` instead. This is an optional field.77- `op_api`: Indicates that op_api calls are supported in this version. If the supported versions are the same as the versions indicated by `all_version`, you can use `all_version` instead. This is an optional field.
78- `gen_opapi`: For operators that support op_api calls, if the adaptation code is simple and the underlying operator can be called directly without additional adaptation, you can consider using structured adaptation to automatically generate the adaptation code. For details, see [(Optional) Structured Adaptation](#optional-structured-adaptation).78- `gen_opapi`: For operators that support op_api calls, if the adaptation code is simple and the underlying operator can be called directly without additional adaptation, you can consider using structured adaptation to automatically generate the adaptation code. For details, see [(Optional) Structured Adaptation](#optional-structured-adaptation).
@@ -224,7 +224,7 @@ The adaptation file path for aclnn operators is `op_plugin/ops/opapi/AbsKernelNp
224 ```224 ```
225 225 
226 > [!NOTE] 226 > [!NOTE]
227- > For more common API interfaces used in operator adaptation, see [LINK](reference.md#section002). For more common macro definitions, see [LINK](reference.md#section003).227+ > For more common API interfaces used in operator adaptation, see [LINK](reference.md#common-api-interfaces-for-operator-adaptation). For more common macro definitions, see [LINK](reference.md#common-macro-definitions-for-operator-adaptation).
228 228 
2293. If the interface contains multiple variants, such as a variant with an `out` parameter or an in-place operation variant, you need to add the corresponding adaptation code. Refer to the following for adaptation:2293. If the interface contains multiple variants, such as a variant with an `out` parameter or an in-place operation variant, you need to add the corresponding adaptation code. Refer to the following for adaptation:
230 230 
@@ -455,7 +455,7 @@ aclop operators are an earlier operator implementation method and are not recomm
455 if (!npu_utils::check_match(&result)) { 455 if (!npu_utils::check_match(&result)) {
456 // If result is non-contiguous, create a contiguous tensor (contig_tensor) to receive the output of the ACLOP operator (abs). Then copy contig_tensor to the original output result456 // If result is non-contiguous, create a contiguous tensor (contig_tensor) to receive the output of the ACLOP operator (abs). Then copy contig_tensor to the original output result
457 at::Tensor contiguous_result = npu_utils::format_contiguous(result); 457 at::Tensor contiguous_result = npu_utils::format_contiguous(result);
458- abs_out_nocheck(contigTensor, self); 458+ abs_out_nocheck(ccontiguous_result, self);
459 npu_utils::format_fresh_view(result, contiguous_result); 459 npu_utils::format_fresh_view(result, contiguous_result);
460 } else { 460 } else {
461 // If result is contiguous, directly call the ACLOP operator461 // If result is contiguous, directly call the ACLOP operator
Mdocs/en/framework_feature_guide_pytorch/sample_call_opplugin.md+6-6
@@ -34,7 +34,7 @@ After completing the development of the adaptation plugin for the PyTorch framew
34 Copy the following information to the `custom` node in `op_plugin_functions.yaml`.34 Copy the following information to the `custom` node in `op_plugin_functions.yaml`.
35 35 
36 ```yaml36 ```yaml
37- - func: npu_add_custom(Tensor x, Tensor y, *, Scalar alpha=1) -> Tensor 37+ - func: npu_add_custom(Tensor x, Tensor y, *, Scalar alpha=1) -> Tensor
38 op_api: v2.738 op_api: v2.7
39 - func: npu_add_custom_backward(Tensor grad) -> (Tensor, Tensor)39 - func: npu_add_custom_backward(Tensor grad) -> (Tensor, Tensor)
40 op_api: v2.740 op_api: v2.7
@@ -44,7 +44,7 @@ After completing the development of the adaptation plugin for the PyTorch framew
44 44 
45 ```yaml45 ```yaml
46 custom: 46 custom:
47- - func: npu_add_custom(Tensor x, Tensor y, *, Scalar alpha=1) -> Tensor 47+ - func: npu_add_custom(Tensor x, Tensor y, *, Scalar alpha=1) -> Tensor
48 op_api: v2.748 op_api: v2.7
49 - func: npu_add_custom_backward(Tensor grad) -> (Tensor, Tensor)49 - func: npu_add_custom_backward(Tensor grad) -> (Tensor, Tensor)
50 op_api: v2.750 op_api: v2.7
@@ -81,7 +81,7 @@ After completing the development of the adaptation plugin for the PyTorch framew
81 81 
82 ```yaml82 ```yaml
83 custom: 83 custom:
84- - func: npu_add_custom(Tensor x, Tensor y, *, Scalar alpha=1) -> Tensor 84+ - func: npu_add_custom(Tensor x, Tensor y, *, Scalar alpha=1) -> Tensor
85 op_api: v2.785 op_api: v2.7
86 gen_opapi:86 gen_opapi:
87 out:87 out:
@@ -102,8 +102,8 @@ After completing the development of the adaptation plugin for the PyTorch framew
102 2. After completing the operator adaptation, the complete `AddCustomKernelNpuOpApi.cpp` file is as follows.102 2. After completing the operator adaptation, the complete `AddCustomKernelNpuOpApi.cpp` file is as follows.
103 103 
104 ```cpp104 ```cpp
105- #include "op_plugin/OpApiInterface.h" 105+ #include "op_plugin/OpApiInterface.h"
106- #include "op_plugin/utils/op_api_common.h" 106+ #include "op_plugin/utils/op_api_common.h"
107 107
108 namespace op_api { 108 namespace op_api {
109 using npu_preparation = at_npu::native::OpPreparation;109 using npu_preparation = at_npu::native::OpPreparation;
@@ -117,7 +117,7 @@ After completing the development of the adaptation plugin for the PyTorch framew
117 // Call the EXEC_NPU_CMD interface to compute the output result117 // Call the EXEC_NPU_CMD interface to compute the output result
118 // The format of the first input parameter is aclnn+Optype, and the subsequent parameters are the inputs and outputs respectively118 // The format of the first input parameter is aclnn+Optype, and the subsequent parameters are the inputs and outputs respectively
119 EXEC_NPU_CMD(aclnnAdd, x, y, alpha, result); 119 EXEC_NPU_CMD(aclnnAdd, x, y, alpha, result);
120- return result; 120+ return result;
121 }121 }
122 122
123 // Backward interface123 // Backward interface
Mdocs/zh/environment_variable_reference/TORCHINDUCTOR_ENABLE_MFUSION.md+4-4
@@ -4,10 +4,10 @@
4 4 
5通过此环境变量可控制是否启用MFusion融合优化功能。MFusion是一种针对NPU平台的图融合优化技术，能够自动融合多个算子为单一kernel，从而减少数据传输开销，提升整体计算性能。5通过此环境变量可控制是否启用MFusion融合优化功能。MFusion是一种针对NPU平台的图融合优化技术，能够自动融合多个算子为单一kernel，从而减少数据传输开销，提升整体计算性能。
6 6 
7-- 配置为"0"或未配置时：禁用MFusion融合优化。7+- 配置为“0”或未配置时：禁用MFusion融合优化。
8-- 配置为"1"时：启用MFusion融合优化。8+- 配置为“1”时：启用MFusion融合优化。
9 9 
10-此环境变量默认配置为"0"。10+此环境变量默认配置为“0”。
11 11 
12## 配置示例12## 配置示例
13 13 
@@ -25,7 +25,7 @@ export TORCHINDUCTOR_ENABLE_MFUSION="1"
25 25 
26## 使用约束26## 使用约束
27 27 
28-- 该功能仅在torch.compile图编译后端为"Inductor"生效。28+- 该功能仅在torch.compile图编译后端为“Inductor”生效。
29- 该功能仅在PyTorch2.7.1和2.9.0版本生效。29- 该功能仅在PyTorch2.7.1和2.9.0版本生效。
30 30 
31## 支持的型号31## 支持的型号
Mdocs/zh/environment_variable_reference/TORCHINDUCTOR_NPU_BACKEND.md+4-4
@@ -4,11 +4,11 @@
4 4 
5通过此环境变量可配置图模式（Inductor）下的后端优化策略，支持在Triton、MLIR、DVM等模式之间切换。5通过此环境变量可配置图模式（Inductor）下的后端优化策略，支持在Triton、MLIR、DVM等模式之间切换。
6 6 
7-- 配置为"default"或未配置时：使用默认的Triton模式。7+- 配置为“default”或未配置时：使用默认的Triton模式。
8-- 配置为"mlir"时：使用MLIR模式。8+- 配置为“mlir”时：使用MLIR模式。
9-- 配置为"dvm"时：使用DVM模式。9+- 配置为“dvm”时：使用DVM模式。
10 10 
11-此环境变量默认配置为"default"。11+此环境变量默认配置为“default”。
12 12 
13## 配置示例13## 配置示例
14 14 
Mdocs/zh/framework_feature_guide_pytorch/adaptation_development_opplugin.md+6-6
@@ -14,9 +14,9 @@
14│ │ └── opapi # aclnn算子适配目录14│ │ └── opapi # aclnn算子适配目录
15│ │ ├── AbsKernelNpuOpApi.cpp15│ │ ├── AbsKernelNpuOpApi.cpp
16│ │ └── ...16│ │ └── ...
17-| ├── python17+│ ├── python
18-| │ └── meta18+│ │ └── meta
19-| │ └── _meta_registrations.py # 算子meta实现注册文件19+│ │ └── _meta_registrations.py # 算子meta实现注册文件
20│ ├── OpInterface.h # 编译自动生成OpPlugin对外接口的头文件，用于框架侧调用算子20│ ├── OpInterface.h # 编译自动生成OpPlugin对外接口的头文件，用于框架侧调用算子
21│ ├── OpInterface.cpp # 编译自动生成OpPlugin对外接口路由实现，内部实现不同类型算子分支选择代码21│ ├── OpInterface.cpp # 编译自动生成OpPlugin对外接口路由实现，内部实现不同类型算子分支选择代码
22│ ├── AclOpsInterface.h # 编译自动生成aclop算子插件适配所对应头文件 22│ ├── AclOpsInterface.h # 编译自动生成aclop算子插件适配所对应头文件
@@ -28,8 +28,8 @@
28├── test28├── test
29│ ├── allowlist_for_publicAPI.json # 对外公开接口白名单29│ ├── allowlist_for_publicAPI.json # 对外公开接口白名单
30│ ├── core_tests30│ ├── core_tests
31-│ | └── torch_npu_OpApi_schema_all.json # 算子接口schema配置文件31+│ │ └── torch_npu_OpApi_schema_all.json # 算子接口schema配置文件
32-| └── test_custom_ops # 算子开发者测试目录32+│ └── test_custom_ops # 算子开发者测试目录
33│ └── ...33│ └── ...
34```34```
35 35 
@@ -455,7 +455,7 @@ aclop算子是早期的算子实现方式，不推荐使用。适配文件路径
455 if (!npu_utils::check_match(&result)) { 455 if (!npu_utils::check_match(&result)) {
456 // 若result非连续，创建连续tensor(contig_tensor)，接收ACLOP算子(abs)的输出。再将contig_tensor拷贝到原始输出result。456 // 若result非连续，创建连续tensor(contig_tensor)，接收ACLOP算子(abs)的输出。再将contig_tensor拷贝到原始输出result。
457 at::Tensor contiguous_result = npu_utils::format_contiguous(result); 457 at::Tensor contiguous_result = npu_utils::format_contiguous(result);
458- abs_out_nocheck(contigTensor, self); 458+ abs_out_nocheck(contiguous_result, self);
459 npu_utils::format_fresh_view(result, contiguous_result); 459 npu_utils::format_fresh_view(result, contiguous_result);
460 } else { 460 } else {
461 // 若result连续，直接调用ACLOP算子。461 // 若result连续，直接调用ACLOP算子。
Mdocs/zh/framework_feature_guide_pytorch/sample_call_opplugin.md+6-6
@@ -34,7 +34,7 @@
34 将如下信息拷贝至op\_plugin\_functions.yaml中的custom节点下。34 将如下信息拷贝至op\_plugin\_functions.yaml中的custom节点下。
35 35 
36 ```yaml36 ```yaml
37- - func: npu_add_custom(Tensor x, Tensor y, *, Scalar alpha=1) -> Tensor 37+ - func: npu_add_custom(Tensor x, Tensor y, *, Scalar alpha=1) -> Tensor
38 op_api: v2.738 op_api: v2.7
39 - func: npu_add_custom_backward(Tensor grad) -> (Tensor, Tensor)39 - func: npu_add_custom_backward(Tensor grad) -> (Tensor, Tensor)
40 op_api: v2.740 op_api: v2.7
@@ -44,7 +44,7 @@
44 44 
45 ```yaml45 ```yaml
46 custom: 46 custom:
47- - func: npu_add_custom(Tensor x, Tensor y, *, Scalar alpha=1) -> Tensor 47+ - func: npu_add_custom(Tensor x, Tensor y, *, Scalar alpha=1) -> Tensor
48 op_api: v2.748 op_api: v2.7
49 - func: npu_add_custom_backward(Tensor grad) -> (Tensor, Tensor)49 - func: npu_add_custom_backward(Tensor grad) -> (Tensor, Tensor)
50 op_api: v2.750 op_api: v2.7
@@ -81,7 +81,7 @@
81 81 
82 ```yaml82 ```yaml
83 custom: 83 custom:
84- - func: npu_add_custom(Tensor x, Tensor y, *, Scalar alpha=1) -> Tensor 84+ - func: npu_add_custom(Tensor x, Tensor y, *, Scalar alpha=1) -> Tensor
85 op_api: v2.785 op_api: v2.7
86 gen_opapi:86 gen_opapi:
87 out:87 out:
@@ -102,8 +102,8 @@
102 2. 完成算子适配，完整的AddCustomKernelNpuOpApi.cpp文件如下。102 2. 完成算子适配，完整的AddCustomKernelNpuOpApi.cpp文件如下。
103 103 
104 ```cpp104 ```cpp
105- #include "op_plugin/OpApiInterface.h" 105+ #include "op_plugin/OpApiInterface.h"
106- #include "op_plugin/utils/op_api_common.h" 106+ #include "op_plugin/utils/op_api_common.h"
107 107
108 namespace op_api { 108 namespace op_api {
109 using npu_preparation = at_npu::native::OpPreparation;109 using npu_preparation = at_npu::native::OpPreparation;
@@ -117,7 +117,7 @@
117 // 调用EXEC_NPU_CMD接口，完成输出结果的计算117 // 调用EXEC_NPU_CMD接口，完成输出结果的计算
118 // 第一个入参格式为aclnn+Optype，之后的参数分别为输入输出118 // 第一个入参格式为aclnn+Optype，之后的参数分别为输入输出
119 EXEC_NPU_CMD(aclnnAdd, x, y, alpha, result); 119 EXEC_NPU_CMD(aclnnAdd, x, y, alpha, result);
120- return result; 120+ return result;
121 }121 }
122 122
123 // 反向接口123 // 反向接口