Merged
[fix] index-add 确定性分支增加 #5415
L1919_snowAdd on Jul 7
[fix] index-add 确定性分支增加 #5415
Merged
Merged froml1919_snow/op-pluginindex-add-fixintoAscend/op-pluginmaster
Total: 1 files changed+39-4
| @@ -13,6 +13,7 @@ | |||
| 13 | // See the License for the specific language governing permissions and | 13 | // See the License for the specific language governing permissions and |
| 14 | // limitations under the License. | 14 | // limitations under the License. |
| 15 | 15 | ||
| 16 | +#include <ATen/ATen.h> | ||
| 16 | #include "op_plugin/AclOpsInterface.h" | 17 | #include "op_plugin/AclOpsInterface.h" |
| 17 | #include "op_plugin/OpApiInterface.h" | 18 | #include "op_plugin/OpApiInterface.h" |
| 18 | #include "op_plugin/utils/op_api_common.h" | 19 | #include "op_plugin/utils/op_api_common.h" |
| @@ -47,6 +48,23 @@ at::Tensor& index_add_out( | |||
| 47 | if (!result.is_same(self)) { | 48 | if (!result.is_same(self)) { |
| 48 | result.copy_(self); | 49 | result.copy_(self); |
| 49 | } | 50 | } |
| 51 | + | ||
| 52 | + if (at::globalContext().deterministicAlgorithms()) { | ||
| 53 | + auto wrapped_dim = at::maybe_wrap_dim(dim, self.dim()); | ||
| 54 | + c10::List<c10::optional<at::Tensor>> indices; | ||
| 55 | + indices.reserve(self.dim()); | ||
| 56 | + for (const auto i : c10::irange(self.dim())) { | ||
| 57 | + if (i == wrapped_dim) { | ||
| 58 | + indices.emplace_back(index.to(at::kLong)); | ||
| 59 | + } else { | ||
| 60 | + indices.emplace_back(c10::nullopt); | ||
| 61 | + } | ||
| 62 | + } | ||
| 63 | + | ||
| 64 | + result.index_put_(indices, source * alpha, true); | ||
| 65 | + return result; | ||
| 66 | + } | ||
| 67 | + | ||
| 50 | EXEC_NPU_CMD(aclnnIndexAdd, result, dim, index, source, alpha, result); | 68 | EXEC_NPU_CMD(aclnnIndexAdd, result, dim, index, source, alpha, result); |
| 51 | return result; | 69 | return result; |
| 52 | } | 70 | } |
| @@ -72,7 +90,25 @@ at::Tensor index_add( | |||
| 72 | source.sizes(), | 90 | source.sizes(), |
| 73 | OPS_ERROR(ErrCode::PARAM)); | 91 | OPS_ERROR(ErrCode::PARAM)); |
| 74 | at::Tensor result = at_npu::native::OpPreparation::apply_tensor_without_format(self.sizes(), self.options()); | 92 | at::Tensor result = at_npu::native::OpPreparation::apply_tensor_without_format(self.sizes(), self.options()); |
| 75 | - EXEC_NPU_CMD(aclnnIndexAdd, result.copy_(self), dim, index, source, alpha, result.copy_(self)); | 93 | + |
| 94 | + if (at::globalContext().deterministicAlgorithms()) { | ||
| 95 | + auto wrapped_dim = at::maybe_wrap_dim(dim, self.dim()); | ||
| 96 | + result.copy_(self); | ||
| 97 | + c10::List<c10::optional<at::Tensor>> indices; | ||
| 98 | + indices.reserve(self.dim()); | ||
| 99 | + for (const auto i : c10::irange(self.dim())) { | ||
| 100 | + if (i == wrapped_dim) { | ||
| 101 | + indices.emplace_back(index.to(at::kLong)); | ||
| 102 | + } else { | ||
| 103 | + indices.emplace_back(c10::nullopt); | ||
| 104 | + } | ||
| 105 | + } | ||
| 106 | + | ||
| 107 | + result.index_put_(indices, source * alpha, true); | ||
| 108 | + return result; | ||
| 109 | + } | ||
| 110 | + | ||
| 111 | + EXEC_NPU_CMD(aclnnIndexAdd, result.copy_(self), dim, index, source, alpha, result); | ||
| 76 | return result; | 112 | return result; |
| 77 | } | 113 | } |
| 78 | 114 | ||
| @@ -82,10 +118,9 @@ at::Tensor index_add( | |||
| 82 | at::Dimname dim, | 118 | at::Dimname dim, |
| 83 | const at::Tensor& index, | 119 | const at::Tensor& index, |
| 84 | const at::Tensor& source, | 120 | const at::Tensor& source, |
| 85 | - const at::Scalar& alpha) | 121 | + const at::Scalar& alpha) { |
| 86 | -{ | ||
| 87 | return op_api::index_add(self, dimname_to_position(self, dim), index, source, alpha); | 122 | return op_api::index_add(self, dimname_to_position(self, dim), index, source, alpha); |
| 88 | } | 123 | } |
 | |||
| 89 | #endif | 124 | #endif |
| 90 | 125 | ||
| 91 | -} | 126 | +} // namespace op_api |
🟠 High Priority
diff 删除了
index_add(Dimname)重载外层的#if !VERSION_BETWEEN(V2R13, VERSION_NEWEST)/#endif预处理器守卫（原第79行和第89行），使该重载在所有版本下均参与编译。但在 V2R13+ 上，
dimname_to_position函数被刻意不提供桩实现。NamedTensorCompat.h第52-55行注释明确说明："All Dimname / DimnameList operator overloads that need dimname_to_position / dimnames_to_positions are gated out on v2.13+ via #if !VERSION_BETWEEN(V2R13, VERSION_NEWEST), so those two helpers are not stubbed here."代码库中所有其他使用
dimname_to_position的 Dimname 重载（如CatKernelNpuOpApi.cpp、SortKernelNpuOpApi.cpp、SoftmaxKernelNpuOpApi.cpp、ScatterAddKernelNpuOpApi.cpp等）均保留该守卫。aclop 侧的index_add(Dimname)（op_plugin/ops/aclops/IndexAddKernelNpu.cpp第92行）也保留了该守卫。该守卫移除后，V2R13+ 构建将因
dimname_to_position符号未定义而失败。建议：恢复
index_add(Dimname)重载外层的#if !VERSION_BETWEEN(V2R13, VERSION_NEWEST)和#endif守卫，保持与代码库其余 Dimname 重载一致的版本门控策略。