diff --git a/scripts/generate_torch_ops.py b/scripts/generate_torch_ops.py
index 9e45a5f4e..6b73a7c87 100644
--- a/scripts/generate_torch_ops.py
+++ b/scripts/generate_torch_ops.py
@@ -1118,11 +1118,11 @@ def _generate_base_header(name: str, ops: list[Op]) -> str:
             init_pieces.append(f"        has_{api_name}_{{{api_name}.has_value()}}")
             init_pieces.append(
                 f"        {api_name}_shape_{{{api_name} ? "
-                f"{api_name}->shape() : Tensor::Shape{{}}}}"
+                f"Tensor::Shape{{{api_name}->shape()}} : Tensor::Shape{{}}}}"
             )
             init_pieces.append(
                 f"        {api_name}_strides_{{{api_name} ? "
-                f"{api_name}->strides() : Tensor::Strides{{}}}}"
+                f"Tensor::Strides{{{api_name}->strides()}} : Tensor::Strides{{}}}}"
             )
             init_pieces.append(
                 f"        {api_name}_type_{{{api_name} ? "
diff --git a/src/base/add.h b/src/base/add.h
index 985e6ed03..d21ecdc16 100644
--- a/src/base/add.h
+++ b/src/base/add.h
@@ -4,7 +4,6 @@
 #include <algorithm>
 #include <cstddef>
 #include <cstdint>
-#include <vector>
 
 #include "operator.h"
 
@@ -68,7 +67,7 @@ class Add : public Operator<Add> {
     auto input_shape = input.shape();
     auto other_shape = other.shape();
     auto ndim = std::max(input_shape.size(), other_shape.size());
-    std::vector<std::size_t> out_shape(ndim, 1);
+    typename TensorLike::Shape out_shape(ndim, 1);
 
     for (std::size_t i = 0; i < ndim; ++i) {
       auto input_dim = i < ndim - input_shape.size()
diff --git a/src/base/batch_norm_elemt.h b/src/base/batch_norm_elemt.h
index c5e406e2a..5da418ebe 100644
--- a/src/base/batch_norm_elemt.h
+++ b/src/base/batch_norm_elemt.h
@@ -25,12 +25,15 @@ class BatchNormElemt : public Operator<BatchNormElemt> {
         out_strides_{out.strides()},
         out_type_{out.dtype()},
         has_weight_{weight.has_value()},
-        weight_shape_{weight ? weight->shape() : Tensor::Shape{}},
-        weight_strides_{weight ? weight->strides() : Tensor::Strides{}},
+        weight_shape_{weight ? Tensor::Shape{weight->shape()}
+                             : Tensor::Shape{}},
+        weight_strides_{weight ? Tensor::Strides{weight->strides()}
+                               : Tensor::Strides{}},
         weight_type_{weight ? weight->dtype() : DataType::kFloat32},
         has_bias_{bias.has_value()},
-        bias_shape_{bias ? bias->shape() : Tensor::Shape{}},
-        bias_strides_{bias ? bias->strides() : Tensor::Strides{}},
+        bias_shape_{bias ? Tensor::Shape{bias->shape()} : Tensor::Shape{}},
+        bias_strides_{bias ? Tensor::Strides{bias->strides()}
+                           : Tensor::Strides{}},
         bias_type_{bias ? bias->dtype() : DataType::kFloat32},
         eps_{eps},
         device_index_{out.device().index()} {}
diff --git a/src/base/binary_cross_entropy.h b/src/base/binary_cross_entropy.h
index 6218f2ee0..5856fed00 100644
--- a/src/base/binary_cross_entropy.h
+++ b/src/base/binary_cross_entropy.h
@@ -22,8 +22,10 @@ class BinaryCrossEntropy : public Operator<BinaryCrossEntropy> {
         out_strides_{out.strides()},
         out_type_{out.dtype()},
         has_weight_{weight.has_value()},
-        weight_shape_{weight ? weight->shape() : Tensor::Shape{}},
-        weight_strides_{weight ? weight->strides() : Tensor::Strides{}},
+        weight_shape_{weight ? Tensor::Shape{weight->shape()}
+                             : Tensor::Shape{}},
+        weight_strides_{weight ? Tensor::Strides{weight->strides()}
+                               : Tensor::Strides{}},
         weight_type_{weight ? weight->dtype() : DataType::kFloat32},
         reduction_{reduction},
         device_index_{out.device().index()} {}
diff --git a/src/base/binary_cross_entropy_backward.h b/src/base/binary_cross_entropy_backward.h
index 2764896fd..3d0d00948 100644
--- a/src/base/binary_cross_entropy_backward.h
+++ b/src/base/binary_cross_entropy_backward.h
@@ -26,8 +26,10 @@ class BinaryCrossEntropyBackward : public Operator<BinaryCrossEntropyBackward> {
         grad_input_strides_{grad_input.strides()},
         grad_input_type_{grad_input.dtype()},
         has_weight_{weight.has_value()},
-        weight_shape_{weight ? weight->shape() : Tensor::Shape{}},
-        weight_strides_{weight ? weight->strides() : Tensor::Strides{}},
+        weight_shape_{weight ? Tensor::Shape{weight->shape()}
+                             : Tensor::Shape{}},
+        weight_strides_{weight ? Tensor::Strides{weight->strides()}
+                               : Tensor::Strides{}},
         weight_type_{weight ? weight->dtype() : DataType::kFloat32},
         reduction_{reduction},
         device_index_{grad_input.device().index()} {}
diff --git a/src/base/clamp.h b/src/base/clamp.h
index 5b1665d9a..d1b1bde7e 100644
--- a/src/base/clamp.h
+++ b/src/base/clamp.h
@@ -30,12 +30,12 @@ class Clamp : public Operator<Clamp> {
         out_strides_{out.strides()},
         out_type_{out.dtype()},
         has_min_{min.has_value()},
-        min_shape_{min ? min->shape() : Tensor::Shape{}},
-        min_strides_{min ? min->strides() : Tensor::Strides{}},
+        min_shape_{min ? Tensor::Shape{min->shape()} : Tensor::Shape{}},
+        min_strides_{min ? Tensor::Strides{min->strides()} : Tensor::Strides{}},
         min_type_{min ? min->dtype() : DataType::kFloat32},
         has_max_{max.has_value()},
-        max_shape_{max ? max->shape() : Tensor::Shape{}},
-        max_strides_{max ? max->strides() : Tensor::Strides{}},
+        max_shape_{max ? Tensor::Shape{max->shape()} : Tensor::Shape{}},
+        max_strides_{max ? Tensor::Strides{max->strides()} : Tensor::Strides{}},
         max_type_{max ? max->dtype() : DataType::kFloat32},
         device_index_{out.device().index()} {}
 
diff --git a/src/base/clip.h b/src/base/clip.h
index b20bd6b6e..72337ac91 100644
--- a/src/base/clip.h
+++ b/src/base/clip.h
@@ -30,12 +30,12 @@ class Clip : public Operator<Clip> {
         out_strides_{out.strides()},
         out_type_{out.dtype()},
         has_min_{min.has_value()},
-        min_shape_{min ? min->shape() : Tensor::Shape{}},
-        min_strides_{min ? min->strides() : Tensor::Strides{}},
+        min_shape_{min ? Tensor::Shape{min->shape()} : Tensor::Shape{}},
+        min_strides_{min ? Tensor::Strides{min->strides()} : Tensor::Strides{}},
         min_type_{min ? min->dtype() : DataType::kFloat32},
         has_max_{max.has_value()},
-        max_shape_{max ? max->shape() : Tensor::Shape{}},
-        max_strides_{max ? max->strides() : Tensor::Strides{}},
+        max_shape_{max ? Tensor::Shape{max->shape()} : Tensor::Shape{}},
+        max_strides_{max ? Tensor::Strides{max->strides()} : Tensor::Strides{}},
         max_type_{max ? max->dtype() : DataType::kFloat32},
         device_index_{out.device().index()} {}
 
diff --git a/src/base/conv_infinilm.h b/src/base/conv_infinilm.h
index 6b84a5993..ca4db7b2a 100644
--- a/src/base/conv_infinilm.h
+++ b/src/base/conv_infinilm.h
@@ -23,8 +23,10 @@ class ConvInfinilm : public Operator<ConvInfinilm> {
         weight_strides_{weight.strides()},
         out_shape_{out.shape()},
         out_strides_{out.strides()},
-        bias_shape_{bias.has_value() ? bias->shape() : Tensor::Shape{}},
-        bias_strides_{bias.has_value() ? bias->strides() : Tensor::Strides{}},
+        bias_shape_{bias.has_value() ? Tensor::Shape{bias->shape()}
+                                     : Tensor::Shape{}},
+        bias_strides_{bias.has_value() ? Tensor::Strides{bias->strides()}
+                                       : Tensor::Strides{}},
         input_type_{input.dtype()},
         weight_type_{weight.dtype()},
         out_type_{out.dtype()},
diff --git a/src/base/diff.h b/src/base/diff.h
index 43b79f572..48d0c02a6 100644
--- a/src/base/diff.h
+++ b/src/base/diff.h
@@ -19,12 +19,16 @@ class Diff : public Operator<Diff> {
         out_strides_{out.strides()},
         out_type_{out.dtype()},
         has_prepend_{prepend.has_value()},
-        prepend_shape_{prepend ? prepend->shape() : Tensor::Shape{}},
-        prepend_strides_{prepend ? prepend->strides() : Tensor::Strides{}},
+        prepend_shape_{prepend ? Tensor::Shape{prepend->shape()}
+                               : Tensor::Shape{}},
+        prepend_strides_{prepend ? Tensor::Strides{prepend->strides()}
+                                 : Tensor::Strides{}},
         prepend_type_{prepend ? prepend->dtype() : DataType::kFloat32},
         has_append_{append.has_value()},
-        append_shape_{append ? append->shape() : Tensor::Shape{}},
-        append_strides_{append ? append->strides() : Tensor::Strides{}},
+        append_shape_{append ? Tensor::Shape{append->shape()}
+                             : Tensor::Shape{}},
+        append_strides_{append ? Tensor::Strides{append->strides()}
+                               : Tensor::Strides{}},
         append_type_{append ? append->dtype() : DataType::kFloat32},
         n_{n},
         dim_{dim},
diff --git a/src/base/embedding.h b/src/base/embedding.h
index c3dd16b0e..8d10243bb 100644
--- a/src/base/embedding.h
+++ b/src/base/embedding.h
@@ -101,7 +101,7 @@ class Embedding : public Operator<Embedding> {
       const double /*norm_type*/ = 2.0,
       const bool /*scale_grad_by_freq*/ = false,
       const bool /*sparse*/ = false) {
-    auto out_shape = input.shape();
+    typename TensorLike::Shape out_shape{input.shape()};
     out_shape.push_back(weight.size(1));
 
     return TensorLike::Empty(out_shape, weight.dtype(), weight.device());
diff --git a/src/base/flash_attn_with_kvcache.h b/src/base/flash_attn_with_kvcache.h
index b775e2d73..2c6144a85 100644
--- a/src/base/flash_attn_with_kvcache.h
+++ b/src/base/flash_attn_with_kvcache.h
@@ -90,46 +90,59 @@ class FlashAttnWithKvcache : public Operator<FlashAttnWithKvcache> {
       : q_shape_{q.shape()},
         k_cache_shape_{k_cache.shape()},
         v_cache_shape_{v_cache.shape()},
-        k_shape_{k.has_value() ? k->shape() : Tensor::Shape{}},
-        v_shape_{v.has_value() ? v->shape() : Tensor::Shape{}},
-        rotary_cos_shape_{rotary_cos.has_value() ? rotary_cos->shape()
-                                                 : Tensor::Shape{}},
-        rotary_sin_shape_{rotary_sin.has_value() ? rotary_sin->shape()
-                                                 : Tensor::Shape{}},
-        cache_seqlens_shape_{cache_seqlens.has_value() ? cache_seqlens->shape()
-                                                       : Tensor::Shape{}},
+        k_shape_{k.has_value() ? Tensor::Shape{k->shape()} : Tensor::Shape{}},
+        v_shape_{v.has_value() ? Tensor::Shape{v->shape()} : Tensor::Shape{}},
+        rotary_cos_shape_{rotary_cos.has_value()
+                              ? Tensor::Shape{rotary_cos->shape()}
+                              : Tensor::Shape{}},
+        rotary_sin_shape_{rotary_sin.has_value()
+                              ? Tensor::Shape{rotary_sin->shape()}
+                              : Tensor::Shape{}},
+        cache_seqlens_shape_{cache_seqlens.has_value()
+                                 ? Tensor::Shape{cache_seqlens->shape()}
+                                 : Tensor::Shape{}},
         cache_batch_idx_shape_{cache_batch_idx.has_value()
-                                   ? cache_batch_idx->shape()
+                                   ? Tensor::Shape{cache_batch_idx->shape()}
                                    : Tensor::Shape{}},
-        cache_leftpad_shape_{cache_leftpad.has_value() ? cache_leftpad->shape()
-                                                       : Tensor::Shape{}},
-        block_table_shape_{block_table.has_value() ? block_table->shape()
-                                                   : Tensor::Shape{}},
-        alibi_slopes_shape_{alibi_slopes.has_value() ? alibi_slopes->shape()
-                                                     : Tensor::Shape{}},
+        cache_leftpad_shape_{cache_leftpad.has_value()
+                                 ? Tensor::Shape{cache_leftpad->shape()}
+                                 : Tensor::Shape{}},
+        block_table_shape_{block_table.has_value()
+                               ? Tensor::Shape{block_table->shape()}
+                               : Tensor::Shape{}},
+        alibi_slopes_shape_{alibi_slopes.has_value()
+                                ? Tensor::Shape{alibi_slopes->shape()}
+                                : Tensor::Shape{}},
         out_shape_{out.shape()},
         q_strides_{q.strides()},
         k_cache_strides_{k_cache.strides()},
         v_cache_strides_{v_cache.strides()},
-        k_strides_{k.has_value() ? k->strides() : Tensor::Strides{}},
-        v_strides_{v.has_value() ? v->strides() : Tensor::Strides{}},
-        rotary_cos_strides_{rotary_cos.has_value() ? rotary_cos->strides()
-                                                   : Tensor::Strides{}},
-        rotary_sin_strides_{rotary_sin.has_value() ? rotary_sin->strides()
-                                                   : Tensor::Strides{}},
+        k_strides_{k.has_value() ? Tensor::Strides{k->strides()}
+                                 : Tensor::Strides{}},
+        v_strides_{v.has_value() ? Tensor::Strides{v->strides()}
+                                 : Tensor::Strides{}},
+        rotary_cos_strides_{rotary_cos.has_value()
+                                ? Tensor::Strides{rotary_cos->strides()}
+                                : Tensor::Strides{}},
+        rotary_sin_strides_{rotary_sin.has_value()
+                                ? Tensor::Strides{rotary_sin->strides()}
+                                : Tensor::Strides{}},
         cache_seqlens_strides_{cache_seqlens.has_value()
-                                   ? cache_seqlens->strides()
+                                   ? Tensor::Strides{cache_seqlens->strides()}
                                    : Tensor::Strides{}},
-        cache_batch_idx_strides_{cache_batch_idx.has_value()
-                                     ? cache_batch_idx->strides()
-                                     : Tensor::Strides{}},
+        cache_batch_idx_strides_{
+            cache_batch_idx.has_value()
+                ? Tensor::Strides{cache_batch_idx->strides()}
+                : Tensor::Strides{}},
         cache_leftpad_strides_{cache_leftpad.has_value()
-                                   ? cache_leftpad->strides()
+                                   ? Tensor::Strides{cache_leftpad->strides()}
                                    : Tensor::Strides{}},
-        block_table_strides_{block_table.has_value() ? block_table->strides()
-                                                     : Tensor::Strides{}},
-        alibi_slopes_strides_{alibi_slopes.has_value() ? alibi_slopes->strides()
-                                                       : Tensor::Strides{}},
+        block_table_strides_{block_table.has_value()
+                                 ? Tensor::Strides{block_table->strides()}
+                                 : Tensor::Strides{}},
+        alibi_slopes_strides_{alibi_slopes.has_value()
+                                  ? Tensor::Strides{alibi_slopes->strides()}
+                                  : Tensor::Strides{}},
         out_strides_{out.strides()},
         q_dtype_{q.dtype()},
         k_cache_dtype_{k_cache.dtype()},
diff --git a/src/base/histogram.h b/src/base/histogram.h
index 787c459c6..4181646c2 100644
--- a/src/base/histogram.h
+++ b/src/base/histogram.h
@@ -26,8 +26,10 @@ class Histogram : public Operator<Histogram> {
         bin_edges_strides_{bin_edges.strides()},
         bin_edges_type_{bin_edges.dtype()},
         has_weight_{weight.has_value()},
-        weight_shape_{weight ? weight->shape() : Tensor::Shape{}},
-        weight_strides_{weight ? weight->strides() : Tensor::Strides{}},
+        weight_shape_{weight ? Tensor::Shape{weight->shape()}
+                             : Tensor::Shape{}},
+        weight_strides_{weight ? Tensor::Strides{weight->strides()}
+                               : Tensor::Strides{}},
         weight_type_{weight ? weight->dtype() : DataType::kFloat32},
         density_{density},
         device_index_{hist.device().index()} {}
@@ -46,8 +48,10 @@ class Histogram : public Operator<Histogram> {
         bin_edges_strides_{bin_edges.strides()},
         bin_edges_type_{bin_edges.dtype()},
         has_weight_{weight.has_value()},
-        weight_shape_{weight ? weight->shape() : Tensor::Shape{}},
-        weight_strides_{weight ? weight->strides() : Tensor::Strides{}},
+        weight_shape_{weight ? Tensor::Shape{weight->shape()}
+                             : Tensor::Shape{}},
+        weight_strides_{weight ? Tensor::Strides{weight->strides()}
+                               : Tensor::Strides{}},
         weight_type_{weight ? weight->dtype() : DataType::kFloat32},
         density_{density},
         bins_{bins},
diff --git a/src/base/internal_batch_norm_with_update.h b/src/base/internal_batch_norm_with_update.h
index 4a59647b0..b1288c3e8 100644
--- a/src/base/internal_batch_norm_with_update.h
+++ b/src/base/internal_batch_norm_with_update.h
@@ -36,12 +36,15 @@ class BatchNormWithUpdate : public Operator<BatchNormWithUpdate> {
         reserve_strides_{reserve.strides()},
         reserve_type_{reserve.dtype()},
         has_weight_{weight.has_value()},
-        weight_shape_{weight ? weight->shape() : Tensor::Shape{}},
-        weight_strides_{weight ? weight->strides() : Tensor::Strides{}},
+        weight_shape_{weight ? Tensor::Shape{weight->shape()}
+                             : Tensor::Shape{}},
+        weight_strides_{weight ? Tensor::Strides{weight->strides()}
+                               : Tensor::Strides{}},
         weight_type_{weight ? weight->dtype() : DataType::kFloat32},
         has_bias_{bias.has_value()},
-        bias_shape_{bias ? bias->shape() : Tensor::Shape{}},
-        bias_strides_{bias ? bias->strides() : Tensor::Strides{}},
+        bias_shape_{bias ? Tensor::Shape{bias->shape()} : Tensor::Shape{}},
+        bias_strides_{bias ? Tensor::Strides{bias->strides()}
+                           : Tensor::Strides{}},
         bias_type_{bias ? bias->dtype() : DataType::kFloat32},
         momentum_{momentum},
         eps_{eps},
diff --git a/src/base/internal_conv_depthwise2d.h b/src/base/internal_conv_depthwise2d.h
index 8f61092b9..291dae48c 100644
--- a/src/base/internal_conv_depthwise2d.h
+++ b/src/base/internal_conv_depthwise2d.h
@@ -26,8 +26,9 @@ class ConvDepthwise2d : public Operator<ConvDepthwise2d> {
         out_strides_{out.strides()},
         out_type_{out.dtype()},
         has_bias_{bias.has_value()},
-        bias_shape_{bias ? bias->shape() : Tensor::Shape{}},
-        bias_strides_{bias ? bias->strides() : Tensor::Strides{}},
+        bias_shape_{bias ? Tensor::Shape{bias->shape()} : Tensor::Shape{}},
+        bias_strides_{bias ? Tensor::Strides{bias->strides()}
+                           : Tensor::Strides{}},
         bias_type_{bias ? bias->dtype() : DataType::kFloat32},
         kernel_size_{kernel_size},
         stride_{stride},
diff --git a/src/base/internal_native_batch_norm_legit.h b/src/base/internal_native_batch_norm_legit.h
index d3aa4c309..4da0da085 100644
--- a/src/base/internal_native_batch_norm_legit.h
+++ b/src/base/internal_native_batch_norm_legit.h
@@ -33,12 +33,15 @@ class NativeBatchNormLegit : public Operator<NativeBatchNormLegit> {
         save_invstd_strides_{save_invstd.strides()},
         save_invstd_type_{save_invstd.dtype()},
         has_weight_{weight.has_value()},
-        weight_shape_{weight ? weight->shape() : Tensor::Shape{}},
-        weight_strides_{weight ? weight->strides() : Tensor::Strides{}},
+        weight_shape_{weight ? Tensor::Shape{weight->shape()}
+                             : Tensor::Shape{}},
+        weight_strides_{weight ? Tensor::Strides{weight->strides()}
+                               : Tensor::Strides{}},
         weight_type_{weight ? weight->dtype() : DataType::kFloat32},
         has_bias_{bias.has_value()},
-        bias_shape_{bias ? bias->shape() : Tensor::Shape{}},
-        bias_strides_{bias ? bias->strides() : Tensor::Strides{}},
+        bias_shape_{bias ? Tensor::Shape{bias->shape()} : Tensor::Shape{}},
+        bias_strides_{bias ? Tensor::Strides{bias->strides()}
+                           : Tensor::Strides{}},
         bias_type_{bias ? bias->dtype() : DataType::kFloat32},
         training_{training},
         momentum_{momentum},
@@ -62,12 +65,15 @@ class NativeBatchNormLegit : public Operator<NativeBatchNormLegit> {
         save_invstd_strides_{save_invstd.strides()},
         save_invstd_type_{save_invstd.dtype()},
         has_weight_{weight.has_value()},
-        weight_shape_{weight ? weight->shape() : Tensor::Shape{}},
-        weight_strides_{weight ? weight->strides() : Tensor::Strides{}},
+        weight_shape_{weight ? Tensor::Shape{weight->shape()}
+                             : Tensor::Shape{}},
+        weight_strides_{weight ? Tensor::Strides{weight->strides()}
+                               : Tensor::Strides{}},
         weight_type_{weight ? weight->dtype() : DataType::kFloat32},
         has_bias_{bias.has_value()},
-        bias_shape_{bias ? bias->shape() : Tensor::Shape{}},
-        bias_strides_{bias ? bias->strides() : Tensor::Strides{}},
+        bias_shape_{bias ? Tensor::Shape{bias->shape()} : Tensor::Shape{}},
+        bias_strides_{bias ? Tensor::Strides{bias->strides()}
+                           : Tensor::Strides{}},
         bias_type_{bias ? bias->dtype() : DataType::kFloat32},
         training_{training},
         momentum_{momentum},
diff --git a/src/base/internal_slow_conv2d_forward.h b/src/base/internal_slow_conv2d_forward.h
index 8c60f4137..1a77c565d 100644
--- a/src/base/internal_slow_conv2d_forward.h
+++ b/src/base/internal_slow_conv2d_forward.h
@@ -25,8 +25,9 @@ class SlowConv2dForward : public Operator<SlowConv2dForward> {
         output_strides_{output.strides()},
         output_type_{output.dtype()},
         has_bias_{bias.has_value()},
-        bias_shape_{bias ? bias->shape() : Tensor::Shape{}},
-        bias_strides_{bias ? bias->strides() : Tensor::Strides{}},
+        bias_shape_{bias ? Tensor::Shape{bias->shape()} : Tensor::Shape{}},
+        bias_strides_{bias ? Tensor::Strides{bias->strides()}
+                           : Tensor::Strides{}},
         bias_type_{bias ? bias->dtype() : DataType::kFloat32},
         kernel_size_{kernel_size},
         stride_{stride},
diff --git a/src/base/linalg_matrix_rank.h b/src/base/linalg_matrix_rank.h
index 4e29d230b..c34d51edf 100644
--- a/src/base/linalg_matrix_rank.h
+++ b/src/base/linalg_matrix_rank.h
@@ -30,12 +30,14 @@ class MatrixRank : public Operator<MatrixRank> {
         out_strides_{out.strides()},
         out_type_{out.dtype()},
         has_atol_{atol.has_value()},
-        atol_shape_{atol ? atol->shape() : Tensor::Shape{}},
-        atol_strides_{atol ? atol->strides() : Tensor::Strides{}},
+        atol_shape_{atol ? Tensor::Shape{atol->shape()} : Tensor::Shape{}},
+        atol_strides_{atol ? Tensor::Strides{atol->strides()}
+                           : Tensor::Strides{}},
         atol_type_{atol ? atol->dtype() : DataType::kFloat32},
         has_rtol_{rtol.has_value()},
-        rtol_shape_{rtol ? rtol->shape() : Tensor::Shape{}},
-        rtol_strides_{rtol ? rtol->strides() : Tensor::Strides{}},
+        rtol_shape_{rtol ? Tensor::Shape{rtol->shape()} : Tensor::Shape{}},
+        rtol_strides_{rtol ? Tensor::Strides{rtol->strides()}
+                           : Tensor::Strides{}},
         rtol_type_{rtol ? rtol->dtype() : DataType::kFloat32},
         hermitian_{hermitian},
         device_index_{out.device().index()} {}
diff --git a/src/base/linalg_pinv.h b/src/base/linalg_pinv.h
index 3d84d0e86..0b45e4a4a 100644
--- a/src/base/linalg_pinv.h
+++ b/src/base/linalg_pinv.h
@@ -29,12 +29,14 @@ class Pinv : public Operator<Pinv> {
         out_strides_{out.strides()},
         out_type_{out.dtype()},
         has_atol_{atol.has_value()},
-        atol_shape_{atol ? atol->shape() : Tensor::Shape{}},
-        atol_strides_{atol ? atol->strides() : Tensor::Strides{}},
+        atol_shape_{atol ? Tensor::Shape{atol->shape()} : Tensor::Shape{}},
+        atol_strides_{atol ? Tensor::Strides{atol->strides()}
+                           : Tensor::Strides{}},
         atol_type_{atol ? atol->dtype() : DataType::kFloat32},
         has_rtol_{rtol.has_value()},
-        rtol_shape_{rtol ? rtol->shape() : Tensor::Shape{}},
-        rtol_strides_{rtol ? rtol->strides() : Tensor::Strides{}},
+        rtol_shape_{rtol ? Tensor::Shape{rtol->shape()} : Tensor::Shape{}},
+        rtol_strides_{rtol ? Tensor::Strides{rtol->strides()}
+                           : Tensor::Strides{}},
         rtol_type_{rtol ? rtol->dtype() : DataType::kFloat32},
         hermitian_{hermitian},
         device_index_{out.device().index()} {}
diff --git a/src/base/moe_sum.h b/src/base/moe_sum.h
index d5b5259de..bb2eba5ce 100644
--- a/src/base/moe_sum.h
+++ b/src/base/moe_sum.h
@@ -60,8 +60,7 @@ class MoeSum : public Operator<MoeSum> {
         (hidden_size_ == 0 || num_tokens_ <= kMaxSignedIndex / hidden_size_) &&
         "`MoeSum` output size must fit signed index arithmetic");
 
-    const auto offsets_fit = [](const Tensor::Shape& shape,
-                                const Tensor::Strides& strides) {
+    const auto offsets_fit = [](const auto& shape, const auto& strides) {
       uint64_t max_offset = 0;
       constexpr auto kMaxOffset =
           static_cast<uint64_t>(std::numeric_limits<int64_t>::max());
diff --git a/src/base/multi_margin_loss.h b/src/base/multi_margin_loss.h
index 984ac13a6..f51b14394 100644
--- a/src/base/multi_margin_loss.h
+++ b/src/base/multi_margin_loss.h
@@ -22,8 +22,10 @@ class MultiMarginLoss : public Operator<MultiMarginLoss> {
         out_strides_{out.strides()},
         out_type_{out.dtype()},
         has_weight_{weight.has_value()},
-        weight_shape_{weight ? weight->shape() : Tensor::Shape{}},
-        weight_strides_{weight ? weight->strides() : Tensor::Strides{}},
+        weight_shape_{weight ? Tensor::Shape{weight->shape()}
+                             : Tensor::Shape{}},
+        weight_strides_{weight ? Tensor::Strides{weight->strides()}
+                               : Tensor::Strides{}},
         weight_type_{weight ? weight->dtype() : DataType::kFloat32},
         p_{p},
         margin_{margin},
diff --git a/src/base/multi_margin_loss_backward.h b/src/base/multi_margin_loss_backward.h
index ad05e9820..26209699a 100644
--- a/src/base/multi_margin_loss_backward.h
+++ b/src/base/multi_margin_loss_backward.h
@@ -27,8 +27,10 @@ class MultiMarginLossBackward : public Operator<MultiMarginLossBackward> {
         grad_input_strides_{grad_input.strides()},
         grad_input_type_{grad_input.dtype()},
         has_weight_{weight.has_value()},
-        weight_shape_{weight ? weight->shape() : Tensor::Shape{}},
-        weight_strides_{weight ? weight->strides() : Tensor::Strides{}},
+        weight_shape_{weight ? Tensor::Shape{weight->shape()}
+                             : Tensor::Shape{}},
+        weight_strides_{weight ? Tensor::Strides{weight->strides()}
+                               : Tensor::Strides{}},
         weight_type_{weight ? weight->dtype() : DataType::kFloat32},
         p_{p},
         margin_{margin},
diff --git a/src/base/native_batch_norm.h b/src/base/native_batch_norm.h
index f88f2e5fd..0c69ec9d5 100644
--- a/src/base/native_batch_norm.h
+++ b/src/base/native_batch_norm.h
@@ -28,25 +28,30 @@ class NativeBatchNorm : public Operator<NativeBatchNorm> {
         save_invstd_strides_{save_invstd.strides()},
         save_invstd_type_{save_invstd.dtype()},
         has_weight_{weight.has_value()},
-        weight_shape_{weight ? weight->shape() : Tensor::Shape{}},
-        weight_strides_{weight ? weight->strides() : Tensor::Strides{}},
+        weight_shape_{weight ? Tensor::Shape{weight->shape()}
+                             : Tensor::Shape{}},
+        weight_strides_{weight ? Tensor::Strides{weight->strides()}
+                               : Tensor::Strides{}},
         weight_type_{weight ? weight->dtype() : DataType::kFloat32},
         has_bias_{bias.has_value()},
-        bias_shape_{bias ? bias->shape() : Tensor::Shape{}},
-        bias_strides_{bias ? bias->strides() : Tensor::Strides{}},
+        bias_shape_{bias ? Tensor::Shape{bias->shape()} : Tensor::Shape{}},
+        bias_strides_{bias ? Tensor::Strides{bias->strides()}
+                           : Tensor::Strides{}},
         bias_type_{bias ? bias->dtype() : DataType::kFloat32},
         has_running_mean_{running_mean.has_value()},
-        running_mean_shape_{running_mean ? running_mean->shape()
+        running_mean_shape_{running_mean ? Tensor::Shape{running_mean->shape()}
                                          : Tensor::Shape{}},
-        running_mean_strides_{running_mean ? running_mean->strides()
-                                           : Tensor::Strides{}},
+        running_mean_strides_{running_mean
+                                  ? Tensor::Strides{running_mean->strides()}
+                                  : Tensor::Strides{}},
         running_mean_type_{running_mean ? running_mean->dtype()
                                         : DataType::kFloat32},
         has_running_var_{running_var.has_value()},
-        running_var_shape_{running_var ? running_var->shape()
+        running_var_shape_{running_var ? Tensor::Shape{running_var->shape()}
                                        : Tensor::Shape{}},
-        running_var_strides_{running_var ? running_var->strides()
-                                         : Tensor::Strides{}},
+        running_var_strides_{running_var
+                                 ? Tensor::Strides{running_var->strides()}
+                                 : Tensor::Strides{}},
         running_var_type_{running_var ? running_var->dtype()
                                       : DataType::kFloat32},
         training_{training},
diff --git a/src/base/nll_loss.h b/src/base/nll_loss.h
index b5dfc5c56..8ceb8fb96 100644
--- a/src/base/nll_loss.h
+++ b/src/base/nll_loss.h
@@ -25,8 +25,10 @@ class NllLoss : public Operator<NllLoss> {
         out_strides_{out.strides()},
         out_type_{out.dtype()},
         has_weight_{weight.has_value()},
-        weight_shape_{weight ? weight->shape() : Tensor::Shape{}},
-        weight_strides_{weight ? weight->strides() : Tensor::Strides{}},
+        weight_shape_{weight ? Tensor::Shape{weight->shape()}
+                             : Tensor::Shape{}},
+        weight_strides_{weight ? Tensor::Strides{weight->strides()}
+                               : Tensor::Strides{}},
         weight_type_{weight ? weight->dtype() : DataType::kFloat32},
         reduction_{
             ReductionFromPythonArguments(size_average, reduce, reduction)},
@@ -48,8 +50,10 @@ class NllLoss : public Operator<NllLoss> {
         out_strides_{out.strides()},
         out_type_{out.dtype()},
         has_weight_{weight.has_value()},
-        weight_shape_{weight ? weight->shape() : Tensor::Shape{}},
-        weight_strides_{weight ? weight->strides() : Tensor::Strides{}},
+        weight_shape_{weight ? Tensor::Shape{weight->shape()}
+                             : Tensor::Shape{}},
+        weight_strides_{weight ? Tensor::Strides{weight->strides()}
+                               : Tensor::Strides{}},
         weight_type_{weight ? weight->dtype() : DataType::kFloat32},
         reduction_{reduction},
         ignore_index_{ignore_index},
diff --git a/src/base/nll_loss2d.h b/src/base/nll_loss2d.h
index 3c144b157..591199493 100644
--- a/src/base/nll_loss2d.h
+++ b/src/base/nll_loss2d.h
@@ -22,8 +22,10 @@ class NllLoss2d : public Operator<NllLoss2d> {
         out_strides_{out.strides()},
         out_type_{out.dtype()},
         has_weight_{weight.has_value()},
-        weight_shape_{weight ? weight->shape() : Tensor::Shape{}},
-        weight_strides_{weight ? weight->strides() : Tensor::Strides{}},
+        weight_shape_{weight ? Tensor::Shape{weight->shape()}
+                             : Tensor::Shape{}},
+        weight_strides_{weight ? Tensor::Strides{weight->strides()}
+                               : Tensor::Strides{}},
         weight_type_{weight ? weight->dtype() : DataType::kFloat32},
         reduction_{reduction},
         ignore_index_{ignore_index},
diff --git a/src/base/nll_loss2d_backward.h b/src/base/nll_loss2d_backward.h
index 704b207ec..81527295a 100644
--- a/src/base/nll_loss2d_backward.h
+++ b/src/base/nll_loss2d_backward.h
@@ -29,8 +29,10 @@ class NllLoss2dBackward : public Operator<NllLoss2dBackward> {
         grad_input_strides_{grad_input.strides()},
         grad_input_type_{grad_input.dtype()},
         has_weight_{weight.has_value()},
-        weight_shape_{weight ? weight->shape() : Tensor::Shape{}},
-        weight_strides_{weight ? weight->strides() : Tensor::Strides{}},
+        weight_shape_{weight ? Tensor::Shape{weight->shape()}
+                             : Tensor::Shape{}},
+        weight_strides_{weight ? Tensor::Strides{weight->strides()}
+                               : Tensor::Strides{}},
         weight_type_{weight ? weight->dtype() : DataType::kFloat32},
         reduction_{reduction},
         ignore_index_{ignore_index},
diff --git a/src/base/nll_loss2d_forward.h b/src/base/nll_loss2d_forward.h
index 82730c03e..696030ea1 100644
--- a/src/base/nll_loss2d_forward.h
+++ b/src/base/nll_loss2d_forward.h
@@ -26,8 +26,10 @@ class NllLoss2dForward : public Operator<NllLoss2dForward> {
         total_weight_strides_{total_weight.strides()},
         total_weight_type_{total_weight.dtype()},
         has_weight_{weight.has_value()},
-        weight_shape_{weight ? weight->shape() : Tensor::Shape{}},
-        weight_strides_{weight ? weight->strides() : Tensor::Strides{}},
+        weight_shape_{weight ? Tensor::Shape{weight->shape()}
+                             : Tensor::Shape{}},
+        weight_strides_{weight ? Tensor::Strides{weight->strides()}
+                               : Tensor::Strides{}},
         weight_type_{weight ? weight->dtype() : DataType::kFloat32},
         reduction_{reduction},
         ignore_index_{ignore_index},
diff --git a/src/base/nll_loss_backward.h b/src/base/nll_loss_backward.h
index 067f91900..7ee3a8290 100644
--- a/src/base/nll_loss_backward.h
+++ b/src/base/nll_loss_backward.h
@@ -29,8 +29,10 @@ class NllLossBackward : public Operator<NllLossBackward> {
         grad_input_strides_{grad_input.strides()},
         grad_input_type_{grad_input.dtype()},
         has_weight_{weight.has_value()},
-        weight_shape_{weight ? weight->shape() : Tensor::Shape{}},
-        weight_strides_{weight ? weight->strides() : Tensor::Strides{}},
+        weight_shape_{weight ? Tensor::Shape{weight->shape()}
+                             : Tensor::Shape{}},
+        weight_strides_{weight ? Tensor::Strides{weight->strides()}
+                               : Tensor::Strides{}},
         weight_type_{weight ? weight->dtype() : DataType::kFloat32},
         reduction_{reduction},
         ignore_index_{ignore_index},
diff --git a/src/base/nll_loss_forward.h b/src/base/nll_loss_forward.h
index 17397662b..30119cc16 100644
--- a/src/base/nll_loss_forward.h
+++ b/src/base/nll_loss_forward.h
@@ -25,8 +25,10 @@ class NllLossForward : public Operator<NllLossForward> {
         total_weight_strides_{total_weight.strides()},
         total_weight_type_{total_weight.dtype()},
         has_weight_{weight.has_value()},
-        weight_shape_{weight ? weight->shape() : Tensor::Shape{}},
-        weight_strides_{weight ? weight->strides() : Tensor::Strides{}},
+        weight_shape_{weight ? Tensor::Shape{weight->shape()}
+                             : Tensor::Shape{}},
+        weight_strides_{weight ? Tensor::Strides{weight->strides()}
+                               : Tensor::Strides{}},
         weight_type_{weight ? weight->dtype() : DataType::kFloat32},
         reduction_{reduction},
         ignore_index_{ignore_index},
diff --git a/src/base/rotary_embedding.h b/src/base/rotary_embedding.h
index f405faa43..01073bf6f 100644
--- a/src/base/rotary_embedding.h
+++ b/src/base/rotary_embedding.h
@@ -16,11 +16,13 @@ class RotaryEmbedding : public Operator<RotaryEmbedding> {
                   int64_t rope_dim_offset = 0, bool inverse = false)
       : positions_shape_{positions.shape()},
         query_shape_{query.shape()},
-        key_shape_{key.has_value() ? key->shape() : Tensor::Shape{}},
+        key_shape_{key.has_value() ? Tensor::Shape{key->shape()}
+                                   : Tensor::Shape{}},
         cos_sin_cache_shape_{cos_sin_cache.shape()},
         positions_strides_{positions.strides()},
         query_strides_{query.strides()},
-        key_strides_{key.has_value() ? key->strides() : Tensor::Strides{}},
+        key_strides_{key.has_value() ? Tensor::Strides{key->strides()}
+                                     : Tensor::Strides{}},
         cos_sin_cache_strides_{cos_sin_cache.strides()},
         positions_type_{positions.dtype()},
         query_type_{query.dtype()},
diff --git a/src/base/scaled_dot_product_attention.h b/src/base/scaled_dot_product_attention.h
index ace5d5cde..40d60f65d 100644
--- a/src/base/scaled_dot_product_attention.h
+++ b/src/base/scaled_dot_product_attention.h
@@ -18,14 +18,16 @@ class ScaledDotProductAttention : public Operator<ScaledDotProductAttention> {
       : query_shape_{query.shape()},
         key_shape_{key.shape()},
         value_shape_{value.shape()},
-        attn_mask_shape_{attn_mask.has_value() ? attn_mask->shape()
-                                               : Tensor::Shape{}},
+        attn_mask_shape_{attn_mask.has_value()
+                             ? Tensor::Shape{attn_mask->shape()}
+                             : Tensor::Shape{}},
         out_shape_{out.shape()},
         query_strides_{query.strides()},
         key_strides_{key.strides()},
         value_strides_{value.strides()},
-        attn_mask_strides_{attn_mask.has_value() ? attn_mask->strides()
-                                                 : Tensor::Strides{}},
+        attn_mask_strides_{attn_mask.has_value()
+                               ? Tensor::Strides{attn_mask->strides()}
+                               : Tensor::Strides{}},
         out_strides_{out.strides()},
         query_type_{query.dtype()},
         attn_mask_type_{attn_mask.has_value() ? attn_mask->dtype()
@@ -43,7 +45,7 @@ class ScaledDotProductAttention : public Operator<ScaledDotProductAttention> {
            "dtypes");
     assert(query.size(-1) == key.size(-1) && key.size(-2) == value.size(-2) &&
            "`ScaledDotProductAttention` input dimensions are incompatible");
-    auto expected_out_shape = query.shape();
+    Tensor::Shape expected_out_shape{query.shape()};
     expected_out_shape.back() = value.size(-1);
     assert(out.shape() == expected_out_shape &&
            "`ScaledDotProductAttention` output shape is incorrect");
@@ -83,7 +85,7 @@ class ScaledDotProductAttention : public Operator<ScaledDotProductAttention> {
     (void)scale;
     (void)enable_gqa;
 
-    auto out_shape = query.shape();
+    typename TensorLike::Shape out_shape{query.shape()};
     out_shape.back() = value.size(-1);
     return TensorLike::Empty(out_shape, query.dtype(), query.device());
   }
diff --git a/src/base/searchsorted.h b/src/base/searchsorted.h
index 481811946..f76575260 100644
--- a/src/base/searchsorted.h
+++ b/src/base/searchsorted.h
@@ -24,8 +24,10 @@ class Searchsorted : public Operator<Searchsorted> {
         out_strides_{out.strides()},
         out_type_{out.dtype()},
         has_sorter_{sorter.has_value()},
-        sorter_shape_{sorter ? sorter->shape() : Tensor::Shape{}},
-        sorter_strides_{sorter ? sorter->strides() : Tensor::Strides{}},
+        sorter_shape_{sorter ? Tensor::Shape{sorter->shape()}
+                             : Tensor::Shape{}},
+        sorter_strides_{sorter ? Tensor::Strides{sorter->strides()}
+                               : Tensor::Strides{}},
         sorter_type_{sorter ? sorter->dtype() : DataType::kFloat32},
         out_int32_{out_int32},
         right_{right},
@@ -43,8 +45,10 @@ class Searchsorted : public Operator<Searchsorted> {
         out_strides_{out.strides()},
         out_type_{out.dtype()},
         has_sorter_{sorter.has_value()},
-        sorter_shape_{sorter ? sorter->shape() : Tensor::Shape{}},
-        sorter_strides_{sorter ? sorter->strides() : Tensor::Strides{}},
+        sorter_shape_{sorter ? Tensor::Shape{sorter->shape()}
+                             : Tensor::Shape{}},
+        sorter_strides_{sorter ? Tensor::Strides{sorter->strides()}
+                               : Tensor::Strides{}},
         sorter_type_{sorter ? sorter->dtype() : DataType::kFloat32},
         out_int32_{out_int32},
         right_{right},
diff --git a/src/base/silu_and_mul.h b/src/base/silu_and_mul.h
index 603c132d2..19a60c115 100644
--- a/src/base/silu_and_mul.h
+++ b/src/base/silu_and_mul.h
@@ -37,7 +37,7 @@ class SiluAndMul : public Operator<SiluAndMul> {
 
   template <typename TensorLike>
   static auto MakeReturnValue(const TensorLike& input) {
-    auto out_shape = input.shape();
+    typename TensorLike::Shape out_shape{input.shape()};
     assert(!out_shape.empty() && out_shape.back() % 2 == 0 &&
            "`SiluAndMul` requires an even input last dimension");
     out_shape.back() /= 2;
diff --git a/src/base/slow_conv3d.h b/src/base/slow_conv3d.h
index 3f55f8425..a33025f9f 100644
--- a/src/base/slow_conv3d.h
+++ b/src/base/slow_conv3d.h
@@ -25,8 +25,9 @@ class SlowConv3d : public Operator<SlowConv3d> {
         out_strides_{out.strides()},
         out_type_{out.dtype()},
         has_bias_{bias.has_value()},
-        bias_shape_{bias ? bias->shape() : Tensor::Shape{}},
-        bias_strides_{bias ? bias->strides() : Tensor::Strides{}},
+        bias_shape_{bias ? Tensor::Shape{bias->shape()} : Tensor::Shape{}},
+        bias_strides_{bias ? Tensor::Strides{bias->strides()}
+                           : Tensor::Strides{}},
         bias_type_{bias ? bias->dtype() : DataType::kFloat32},
         kernel_size_{kernel_size},
         stride_{stride},
diff --git a/src/base/slow_conv3d_forward.h b/src/base/slow_conv3d_forward.h
index 75a42d082..0786cd8c0 100644
--- a/src/base/slow_conv3d_forward.h
+++ b/src/base/slow_conv3d_forward.h
@@ -25,8 +25,9 @@ class SlowConv3dForward : public Operator<SlowConv3dForward> {
         output_strides_{output.strides()},
         output_type_{output.dtype()},
         has_bias_{bias.has_value()},
-        bias_shape_{bias ? bias->shape() : Tensor::Shape{}},
-        bias_strides_{bias ? bias->strides() : Tensor::Strides{}},
+        bias_shape_{bias ? Tensor::Shape{bias->shape()} : Tensor::Shape{}},
+        bias_strides_{bias ? Tensor::Strides{bias->strides()}
+                           : Tensor::Strides{}},
         bias_type_{bias ? bias->dtype() : DataType::kFloat32},
         kernel_size_{kernel_size},
         stride_{stride},
diff --git a/src/base/slow_conv_transpose2d.h b/src/base/slow_conv_transpose2d.h
index cc561cba3..59ffc8293 100644
--- a/src/base/slow_conv_transpose2d.h
+++ b/src/base/slow_conv_transpose2d.h
@@ -27,8 +27,9 @@ class SlowConvTranspose2d : public Operator<SlowConvTranspose2d> {
         out_strides_{out.strides()},
         out_type_{out.dtype()},
         has_bias_{bias.has_value()},
-        bias_shape_{bias ? bias->shape() : Tensor::Shape{}},
-        bias_strides_{bias ? bias->strides() : Tensor::Strides{}},
+        bias_shape_{bias ? Tensor::Shape{bias->shape()} : Tensor::Shape{}},
+        bias_strides_{bias ? Tensor::Strides{bias->strides()}
+                           : Tensor::Strides{}},
         bias_type_{bias ? bias->dtype() : DataType::kFloat32},
         kernel_size_{kernel_size},
         stride_{stride},
diff --git a/src/base/slow_conv_transpose3d.h b/src/base/slow_conv_transpose3d.h
index f0ce6dfe6..209bca8bf 100644
--- a/src/base/slow_conv_transpose3d.h
+++ b/src/base/slow_conv_transpose3d.h
@@ -27,8 +27,9 @@ class SlowConvTranspose3d : public Operator<SlowConvTranspose3d> {
         out_strides_{out.strides()},
         out_type_{out.dtype()},
         has_bias_{bias.has_value()},
-        bias_shape_{bias ? bias->shape() : Tensor::Shape{}},
-        bias_strides_{bias ? bias->strides() : Tensor::Strides{}},
+        bias_shape_{bias ? Tensor::Shape{bias->shape()} : Tensor::Shape{}},
+        bias_strides_{bias ? Tensor::Strides{bias->strides()}
+                           : Tensor::Strides{}},
         bias_type_{bias ? bias->dtype() : DataType::kFloat32},
         kernel_size_{kernel_size},
         stride_{stride},
diff --git a/src/base/thnn_conv2d.h b/src/base/thnn_conv2d.h
index 810a799f1..f56bc22f5 100644
--- a/src/base/thnn_conv2d.h
+++ b/src/base/thnn_conv2d.h
@@ -25,8 +25,9 @@ class ThnnConv2d : public Operator<ThnnConv2d> {
         out_strides_{out.strides()},
         out_type_{out.dtype()},
         has_bias_{bias.has_value()},
-        bias_shape_{bias ? bias->shape() : Tensor::Shape{}},
-        bias_strides_{bias ? bias->strides() : Tensor::Strides{}},
+        bias_shape_{bias ? Tensor::Shape{bias->shape()} : Tensor::Shape{}},
+        bias_strides_{bias ? Tensor::Strides{bias->strides()}
+                           : Tensor::Strides{}},
         bias_type_{bias ? bias->dtype() : DataType::kFloat32},
         kernel_size_{kernel_size},
         stride_{stride},
diff --git a/src/pybind11_utils.h b/src/pybind11_utils.h
index 292a2c00b..fbdf4b61e 100644
--- a/src/pybind11_utils.h
+++ b/src/pybind11_utils.h
@@ -143,17 +143,31 @@ inline Device DeviceFromPybind11HandleImpl(py::handle obj) {
   return Device{DeviceTypeFromString(device_type_str), device_index};
 }
 
+template <typename Metadata>
+Metadata MetadataFromPybind11Sequence(py::handle obj) {
+  auto values{py::reinterpret_borrow<py::sequence>(obj)};
+  Metadata metadata;
+  metadata.reserve(values.size());
+
+  for (const auto& value : values) {
+    metadata.push_back(value.cast<typename Metadata::value_type>());
+  }
+
+  return metadata;
+}
+
 inline Tensor TensorFromPybind11HandleImpl(py::handle obj) {
   auto data{
       reinterpret_cast<void*>(obj.attr("data_ptr")().cast<std::uintptr_t>())};
 
-  auto shape{obj.attr("shape").cast<typename Tensor::Shape>()};
+  auto shape{MetadataFromPybind11Sequence<Tensor::Shape>(obj.attr("shape"))};
 
   auto dtype{DataTypeFromPybind11HandleImpl(obj.attr("dtype"))};
 
   auto device{DeviceFromPybind11HandleImpl(obj)};
 
-  auto strides{obj.attr("stride")().cast<typename Tensor::Strides>()};
+  auto strides{
+      MetadataFromPybind11Sequence<Tensor::Strides>(obj.attr("stride")())};
 
   return Tensor{data, std::move(shape), dtype, device, std::move(strides)};
 }
diff --git a/src/torch/tensor_.h b/src/torch/tensor_.h
index e199bf704..556eed5cd 100644
--- a/src/torch/tensor_.h
+++ b/src/torch/tensor_.h
@@ -5,6 +5,7 @@
 #include <torch/version.h>
 
 #include <optional>
+#include <vector>
 
 #include "tensor.h"
 #include "torch/device_.h"
@@ -83,9 +84,9 @@ inline at::ScalarType ToAtenDataType(DataType dtype) {
 // Build an ATen tensor from explicit metadata. Use this instead of reading
 // shape/strides from the `Tensor` parameter, which may have been moved-from
 // by the `Call()` dispatch path (see `operator.h`).
-template <Device::Type kDev>
+template <Device::Type kDev, typename ShapeRange, typename StridesRange>
 inline at::Tensor ToAtenTensor(
-    void* data, const Tensor::Shape& shape, const Tensor::Strides& strides,
+    void* data, const ShapeRange& shape, const StridesRange& strides,
     DataType dtype, int device_index = 0,
     std::optional<at::ScalarType> dtype_override = std::nullopt) {
   std::vector<int64_t> at_shape(shape.begin(), shape.end());
diff --git a/tests/test_cpp_api.py b/tests/test_cpp_api.py
index 02933b8bf..1cc60fe34 100644
--- a/tests/test_cpp_api.py
+++ b/tests/test_cpp_api.py
@@ -58,6 +58,33 @@ def test_cpp_returning_call_smoke(tmp_path):
     _run([str(binary)])
 
 
+@pytest.mark.parametrize(
+    "header",
+    (
+        "base/clamp.h",
+        "base/moe_sum.h",
+        "base/scaled_dot_product_attention.h",
+    ),
+)
+def test_cpp_base_headers_compile_with_metadata_views(tmp_path, header):
+    install_prefix = _install_prefix()
+    include_dir = install_prefix / "include"
+    source = tmp_path / f"{Path(header).stem}_metadata_view.cc"
+    source.write_text(f"#include <{header}>\n\nint main() {{ return 0; }}\n")
+
+    _run(
+        [
+            _compiler("CXX", "c++"),
+            "-std=c++17",
+            "-Werror",
+            "-UNDEBUG",
+            "-fsyntax-only",
+            f"-I{include_dir}",
+            str(source),
+        ]
+    )
+
+
 def _install_prefix():
     prefix = os.environ.get("INFINI_OPS_INSTALL_PREFIX")
 
@@ -146,10 +173,15 @@ def _run(command):
     r"""
     #include <infini/ops.h>
 
+    #include <base/embedding.h>
+    #include <base/scaled_dot_product_attention.h>
+    #include <base/silu_and_mul.h>
+
     #include <cmath>
     #include <functional>
     #include <numeric>
     #include <stdexcept>
+    #include <type_traits>
     #include <utility>
     #include <vector>
 
@@ -179,9 +211,32 @@ class OwningTensor {
 
       const void* data() const { return data_.data(); }
 
-      const Shape& shape() const { return shape_; }
+      template <typename ShapeAccess = decltype(
+                    std::declval<const infini::ops::Tensor&>().shape())>
+      decltype(auto) shape() const {
+        if constexpr (std::is_reference_v<ShapeAccess>) {
+          return (shape_);
+        } else {
+          return ShapeAccess{shape_.data(), shape_.size()};
+        }
+      }
+
+      template <typename StridesAccess = decltype(
+                    std::declval<const infini::ops::Tensor&>().strides())>
+      decltype(auto) strides() const {
+        if constexpr (std::is_reference_v<StridesAccess>) {
+          return (strides_);
+        } else {
+          return StridesAccess{strides_.data(), strides_.size()};
+        }
+      }
 
-      const Strides& strides() const { return strides_; }
+      Shape::value_type size(std::ptrdiff_t dim) const {
+        const auto index = dim < 0
+                               ? static_cast<std::ptrdiff_t>(shape_.size()) + dim
+                               : dim;
+        return shape_[static_cast<std::size_t>(index)];
+      }
 
       infini::ops::DataType dtype() const { return dtype_; }
 
@@ -249,6 +304,30 @@ class OwningTensor {
         return 1;
       }
 
+      OwningTensor silu_input(std::vector<float>(8), {2, 4});
+      auto silu_output = infini::ops::SiluAndMul::MakeReturnValue(silu_input);
+      if (silu_output.shape() != OwningTensor::Shape{2, 2}) {
+        return 1;
+      }
+
+      OwningTensor embedding_input(std::vector<float>(6), {2, 3});
+      OwningTensor embedding_weight(std::vector<float>(40), {10, 4});
+      auto embedding_output = infini::ops::Embedding::MakeReturnValue(
+          embedding_input, embedding_weight);
+      if (embedding_output.shape() != OwningTensor::Shape{2, 3, 4}) {
+        return 1;
+      }
+
+      OwningTensor query(std::vector<float>(24), {1, 2, 3, 4});
+      OwningTensor key(std::vector<float>(40), {1, 2, 5, 4});
+      OwningTensor value(std::vector<float>(60), {1, 2, 5, 6});
+      auto attention_output =
+          infini::ops::ScaledDotProductAttention::MakeReturnValue(query, key,
+                                                                  value);
+      if (attention_output.shape() != OwningTensor::Shape{1, 2, 3, 6}) {
+        return 1;
+      }
+
       return 0;
     }
     """
diff --git a/tests/test_generate_torch_ops.py b/tests/test_generate_torch_ops.py
index 165511180..bd1cfa862 100644
--- a/tests/test_generate_torch_ops.py
+++ b/tests/test_generate_torch_ops.py
@@ -80,6 +80,14 @@ def test_optional_tensor_params_are_exposed_and_forwarded_to_aten():
     assert "std::optional<Tensor> bias" in base
     assert "bool has_weight_" in base
     assert "bool has_bias_" in base
+    assert (
+        "weight_shape_{weight ? Tensor::Shape{weight->shape()} : Tensor::Shape{}}"
+        in base
+    )
+    assert (
+        "weight_strides_{weight ? Tensor::Strides{weight->strides()} "
+        ": Tensor::Strides{}}" in base
+    )
     assert "c10::optional<at::Tensor> at_weight" in source
     assert "c10::optional<at::Tensor> at_bias" in source
     assert "weight->shape()" in source