已合并
fix smoothl1loss blending accuracy #2640
simbaliuxx创建于 2025年6月17日
fix smoothl1loss blending accuracy #2640
已合并
从refs/pull/2640/head合入到master
共 2 个文件变更+5-2
| @@ -36,7 +36,8 @@ at::Tensor smooth_l1_loss(const at::Tensor &self, const at::Tensor &target, int6 | |||
| 36 | { | 36 | { |
| 37 | DO_COMPATIBILITY(aclnnSmoothL1Loss, acl_op::smooth_l1_loss(self, target, reduction, beta)); | 37 | DO_COMPATIBILITY(aclnnSmoothL1Loss, acl_op::smooth_l1_loss(self, target, reduction, beta)); |
| 38 | auto outputSize = op_infer::smooth_l1_loss_npu_output_size(self, reduction); | 38 | auto outputSize = op_infer::smooth_l1_loss_npu_output_size(self, reduction); |
| 39 | - at::Tensor result = npu_preparation::apply_tensor_without_format(self, outputSize); | 39 | + at::ScalarType high_type = at::native::result_type(self, target); |
| 40 | + at::Tensor result = npu_preparation::apply_tensor_without_format(outputSize, self.options().dtype(high_type)); | ||
| 40 | float sigma = static_cast<float>(beta); | 41 | float sigma = static_cast<float>(beta); |
| 41 | EXEC_NPU_CMD(aclnnSmoothL1Loss, self, target, reduction, sigma, result); | 42 | EXEC_NPU_CMD(aclnnSmoothL1Loss, self, target, reduction, sigma, result); |
| 42 | return result; | 43 | return result; |
| @@ -15,6 +15,7 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | + | ||
| 18 | 19 | ||
| 19 | 20 | ||
| 20 | 21 | ||
| @@ -580,8 +581,9 @@ ctc_loss_npu_output_size(const at::Tensor &log_probs, int64_t max_length) | |||
| 580 | } | 581 | } |
| 581 | c10::SmallVector<int64_t, SIZE> neg_log_likelihood_size = {batch_size}; | 582 | c10::SmallVector<int64_t, SIZE> neg_log_likelihood_size = {batch_size}; |
| 582 | int64_t alpha_tail_size = 2 * max_length + 1; | 583 | int64_t alpha_tail_size = 2 * max_length + 1; |
| 584 | + static bool isRegBaseSoc = c10_npu::GetSocVersion() >= c10_npu::SocVersion::Ascend910_95; | ||
| 583 | // Apply for a 32 byte aligned space to avoid address shifting in the OP. | 585 | // Apply for a 32 byte aligned space to avoid address shifting in the OP. |
| 584 | - int64_t alpha_tail_size_align = (alpha_tail_size + 7) / 8 * 8; | 586 | + int64_t alpha_tail_size_align = isRegBaseSoc ? alpha_tail_size : (alpha_tail_size + 7) / 8 * 8; |
| 585 | c10::SmallVector<int64_t, SIZE> log_alpha_size = {batch_size, time_size, alpha_tail_size_align}; | 587 | c10::SmallVector<int64_t, SIZE> log_alpha_size = {batch_size, time_size, alpha_tail_size_align}; |
| 586 | 588 | ||
| 587 | if (log_probs.dim() == dim_num_two) { | 589 | if (log_probs.dim() == dim_num_two) { |
定义加个static