Optim deformable detr

src/transformers/models/deformable_detr/modeling_deformable_detr.py

@@ -523,8 +523,8 @@ def __init__(self, embedding_dim=64, temperature=10000, normalize=False, scale=N
     def forward(self, pixel_values, pixel_mask):
         if pixel_mask is None:
             raise ValueError("No pixel mask provided")
-        y_embed = pixel_mask.cumsum(1, dtype=torch.float32)
-        x_embed = pixel_mask.cumsum(2, dtype=torch.float32)
+        y_embed = pixel_mask.cumsum(1, dtype=torch.float16)
+        x_embed = pixel_mask.cumsum(2, dtype=torch.float16)
         if self.normalize:
             eps = 1e-6
             y_embed = (y_embed - 0.5) / (y_embed[:, -1:, :] + eps) * self.scale
@@ -580,11 +580,14 @@ def build_position_encoding(config):
 
 
 def multi_scale_deformable_attention(
-    value: Tensor, value_spatial_shapes: Tensor, sampling_locations: Tensor, attention_weights: Tensor
+    value: Tensor,
+    value_spatial_shapes: Union[Tensor, List[Tuple]],
+    sampling_locations: Tensor,
+    attention_weights: Tensor,
 ) -> Tensor:
     batch_size, _, num_heads, hidden_dim = value.shape
     _, num_queries, num_heads, num_levels, num_points, _ = sampling_locations.shape
-    value_list = value.split([height.item() * width.item() for height, width in value_spatial_shapes], dim=1)
+    value_list = value.split([height * width for height, width in value_spatial_shapes], dim=1)
     sampling_grids = 2 * sampling_locations - 1
     sampling_value_list = []
     for level_id, (height, width) in enumerate(value_spatial_shapes):
@@ -695,6 +698,7 @@ def forward(
         position_embeddings: Optional[torch.Tensor] = None,
         reference_points=None,
         spatial_shapes=None,
+        spatial_shapes_list=None,
         level_start_index=None,
         output_attentions: bool = False,
     ):
@@ -704,7 +708,8 @@ def forward(
 
         batch_size, num_queries, _ = hidden_states.shape
         batch_size, sequence_length, _ = encoder_hidden_states.shape
-        if (spatial_shapes[:, 0] * spatial_shapes[:, 1]).sum() != sequence_length:
+        total_elements = sum(shape[0] * shape[1] for shape in spatial_shapes_list)
+        if total_elements != sequence_length:
             raise ValueError(
                 "Make sure to align the spatial shapes with the sequence length of the encoder hidden states"
             )
@@ -739,9 +744,11 @@ def forward(
         else:
             raise ValueError(f"Last dim of reference_points must be 2 or 4, but got {reference_points.shape[-1]}")
 
-        if self.disable_custom_kernels:
+        if self.disable_custom_kernels or MultiScaleDeformableAttention is None:
             # PyTorch implementation
-            output = multi_scale_deformable_attention(value, spatial_shapes, sampling_locations, attention_weights)
+            output = multi_scale_deformable_attention(
+                value, spatial_shapes_list, sampling_locations, attention_weights
+            )
         else:
             try:
                 # custom kernel
@@ -755,7 +762,9 @@ def forward(
                 )
             except Exception:
                 # PyTorch implementation
-                output = multi_scale_deformable_attention(value, spatial_shapes, sampling_locations, attention_weights)
+                output = multi_scale_deformable_attention(
+                    value, spatial_shapes_list, sampling_locations, attention_weights
+                )
         output = self.output_proj(output)
 
         return output, attention_weights
@@ -900,6 +909,7 @@ def forward(
         position_embeddings: torch.Tensor = None,
         reference_points=None,
         spatial_shapes=None,
+        spatial_shapes_list=None,
         level_start_index=None,
         output_attentions: bool = False,
     ):
@@ -932,6 +942,7 @@ def forward(
             position_embeddings=position_embeddings,
             reference_points=reference_points,
             spatial_shapes=spatial_shapes,
+            spatial_shapes_list=spatial_shapes_list,
             level_start_index=level_start_index,
             output_attentions=output_attentions,
         )
@@ -997,6 +1008,7 @@ def forward(
         position_embeddings: Optional[torch.Tensor] = None,
         reference_points=None,
         spatial_shapes=None,
+        spatial_shapes_list=None,
         level_start_index=None,
         encoder_hidden_states: Optional[torch.Tensor] = None,
         encoder_attention_mask: Optional[torch.Tensor] = None,
@@ -1048,6 +1060,7 @@ def forward(
             position_embeddings=position_embeddings,
             reference_points=reference_points,
             spatial_shapes=spatial_shapes,
+            spatial_shapes_list=spatial_shapes_list,
             level_start_index=level_start_index,
             output_attentions=output_attentions,
         )
@@ -1219,6 +1232,7 @@ def forward(
         attention_mask=None,
         position_embeddings=None,
         spatial_shapes=None,
+        spatial_shapes_list=None,
         level_start_index=None,
         valid_ratios=None,
         output_attentions=None,
@@ -1260,7 +1274,8 @@ def forward(
         hidden_states = inputs_embeds
         hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
 
-        reference_points = self.get_reference_points(spatial_shapes, valid_ratios, device=inputs_embeds.device)
+        spatial_shapes_tuple = tuple(spatial_shapes_list)
+        reference_points = self.get_reference_points(spatial_shapes_tuple, valid_ratios, device=inputs_embeds.device)
 
         encoder_states = () if output_hidden_states else None
         all_attentions = () if output_attentions else None
@@ -1275,6 +1290,7 @@ def forward(
                     position_embeddings,
                     reference_points,
                     spatial_shapes,
+                    spatial_shapes_list,
                     level_start_index,
                     output_attentions,
                 )
@@ -1285,6 +1301,7 @@ def forward(
                     position_embeddings=position_embeddings,
                     reference_points=reference_points,
                     spatial_shapes=spatial_shapes,
+                    spatial_shapes_list=spatial_shapes_list,
                     level_start_index=level_start_index,
                     output_attentions=output_attentions,
                 )
@@ -1341,6 +1358,7 @@ def forward(
         position_embeddings=None,
         reference_points=None,
         spatial_shapes=None,
+        spatial_shapes_list=None,
         level_start_index=None,
         valid_ratios=None,
         output_attentions=None,
@@ -1416,6 +1434,7 @@ def forward(
                     position_embeddings,
                     reference_points_input,
                     spatial_shapes,
+                    spatial_shapes_list,
                     level_start_index,
                     encoder_hidden_states,
                     encoder_attention_mask,
@@ -1428,6 +1447,7 @@ def forward(
                     encoder_hidden_states=encoder_hidden_states,
                     reference_points=reference_points_input,
                     spatial_shapes=spatial_shapes,
+                    spatial_shapes_list=spatial_shapes_list,
                     level_start_index=level_start_index,
                     encoder_attention_mask=encoder_attention_mask,
                     output_attentions=output_attentions,
@@ -1735,11 +1755,11 @@ def forward(
         source_flatten = []
         mask_flatten = []
         lvl_pos_embed_flatten = []
-        spatial_shapes = []
+        spatial_shapes_list = []
         for level, (source, mask, pos_embed) in enumerate(zip(sources, masks, position_embeddings_list)):
             batch_size, num_channels, height, width = source.shape
             spatial_shape = (height, width)
-            spatial_shapes.append(spatial_shape)
+            spatial_shapes_list.append(spatial_shape)
             source = source.flatten(2).transpose(1, 2)
             mask = mask.flatten(1)
             pos_embed = pos_embed.flatten(2).transpose(1, 2)
@@ -1750,7 +1770,7 @@ def forward(
         source_flatten = torch.cat(source_flatten, 1)
         mask_flatten = torch.cat(mask_flatten, 1)
         lvl_pos_embed_flatten = torch.cat(lvl_pos_embed_flatten, 1)
-        spatial_shapes = torch.as_tensor(spatial_shapes, dtype=torch.long, device=source_flatten.device)
+        spatial_shapes = torch.as_tensor(spatial_shapes_list, dtype=torch.long, device=source_flatten.device)
         level_start_index = torch.cat((spatial_shapes.new_zeros((1,)), spatial_shapes.prod(1).cumsum(0)[:-1]))
         valid_ratios = torch.stack([self.get_valid_ratio(m, dtype=source_flatten.dtype) for m in masks], 1)
 
@@ -1762,6 +1782,7 @@ def forward(
                 attention_mask=mask_flatten,
                 position_embeddings=lvl_pos_embed_flatten,
                 spatial_shapes=spatial_shapes,
+                spatial_shapes_list=spatial_shapes_list,
                 level_start_index=level_start_index,
                 valid_ratios=valid_ratios,
                 output_attentions=output_attentions,
@@ -1819,6 +1840,7 @@ def forward(
             encoder_attention_mask=mask_flatten,
             reference_points=reference_points,
             spatial_shapes=spatial_shapes,
+            spatial_shapes_list=spatial_shapes_list,
             level_start_index=level_start_index,
             valid_ratios=valid_ratios,
             output_attentions=output_attentions,

src/transformers/models/grounding_dino/modeling_grounding_dino.py

@@ -583,11 +583,14 @@ def build_position_encoding(config):
 
 # Copied from transformers.models.deformable_detr.modeling_deformable_detr.multi_scale_deformable_attention
 def multi_scale_deformable_attention(
-    value: Tensor, value_spatial_shapes: Tensor, sampling_locations: Tensor, attention_weights: Tensor
+    value: Tensor,
+    value_spatial_shapes: Union[Tensor, List[Tuple]],
+    sampling_locations: Tensor,
+    attention_weights: Tensor,
 ) -> Tensor:
     batch_size, _, num_heads, hidden_dim = value.shape
     _, num_queries, num_heads, num_levels, num_points, _ = sampling_locations.shape
-    value_list = value.split([height.item() * width.item() for height, width in value_spatial_shapes], dim=1)
+    value_list = value.split([height * width for height, width in value_spatial_shapes], dim=1)
     sampling_grids = 2 * sampling_locations - 1
     sampling_value_list = []
     for level_id, (height, width) in enumerate(value_spatial_shapes):
@@ -699,6 +702,7 @@ def forward(
         position_embeddings: Optional[torch.Tensor] = None,
         reference_points=None,
         spatial_shapes=None,
+        spatial_shapes_list=None,
         level_start_index=None,
         output_attentions: bool = False,
     ):
@@ -708,6 +712,7 @@ def forward(
 
         batch_size, num_queries, _ = hidden_states.shape
         batch_size, sequence_length, _ = encoder_hidden_states.shape
+        # Ignore copy
         if (spatial_shapes[:, 0] * spatial_shapes[:, 1]).sum() != sequence_length:
             raise ValueError(
                 "Make sure to align the spatial shapes with the sequence length of the encoder hidden states"
@@ -743,7 +748,7 @@ def forward(
         else:
             raise ValueError(f"Last dim of reference_points must be 2 or 4, but got {reference_points.shape[-1]}")
 
-        if self.disable_custom_kernels:
+        if self.disable_custom_kernels or MultiScaleDeformableAttention is None:
             # PyTorch implementation
             output = multi_scale_deformable_attention(value, spatial_shapes, sampling_locations, attention_weights)
         else:

src/transformers/models/mask2former/modeling_mask2former.py

@@ -17,7 +17,7 @@
 import math
 import warnings
 from dataclasses import dataclass
-from typing import Dict, List, Optional, Tuple
+from typing import Dict, List, Optional, Tuple, Union
 
 import numpy as np
 import torch
@@ -800,11 +800,14 @@ def get_num_masks(self, class_labels: torch.Tensor, device: torch.device) -> tor
 
 # Copied from transformers.models.deformable_detr.modeling_deformable_detr.multi_scale_deformable_attention
 def multi_scale_deformable_attention(
-    value: Tensor, value_spatial_shapes: Tensor, sampling_locations: Tensor, attention_weights: Tensor
+    value: Tensor,
+    value_spatial_shapes: Union[Tensor, List[Tuple]],
+    sampling_locations: Tensor,
+    attention_weights: Tensor,
 ) -> Tensor:
     batch_size, _, num_heads, hidden_dim = value.shape
     _, num_queries, num_heads, num_levels, num_points, _ = sampling_locations.shape
-    value_list = value.split([height.item() * width.item() for height, width in value_spatial_shapes], dim=1)
+    value_list = value.split([height * width for height, width in value_spatial_shapes], dim=1)
     sampling_grids = 2 * sampling_locations - 1
     sampling_value_list = []
     for level_id, (height, width) in enumerate(value_spatial_shapes):

src/transformers/models/oneformer/modeling_oneformer.py

@@ -18,7 +18,7 @@
 import math
 import warnings
 from dataclasses import dataclass
-from typing import Dict, List, Optional, Tuple
+from typing import Dict, List, Optional, Tuple, Union
 
 import numpy as np
 import torch
@@ -63,11 +63,14 @@ def _get_clones(module, N):
 
 # Copied from transformers.models.deformable_detr.modeling_deformable_detr.multi_scale_deformable_attention
 def multi_scale_deformable_attention(
-    value: Tensor, value_spatial_shapes: Tensor, sampling_locations: Tensor, attention_weights: Tensor
+    value: Tensor,
+    value_spatial_shapes: Union[Tensor, List[Tuple]],
+    sampling_locations: Tensor,
+    attention_weights: Tensor,
 ) -> Tensor:
     batch_size, _, num_heads, hidden_dim = value.shape
     _, num_queries, num_heads, num_levels, num_points, _ = sampling_locations.shape
-    value_list = value.split([height.item() * width.item() for height, width in value_spatial_shapes], dim=1)
+    value_list = value.split([height * width for height, width in value_spatial_shapes], dim=1)
     sampling_grids = 2 * sampling_locations - 1
     sampling_value_list = []
     for level_id, (height, width) in enumerate(value_spatial_shapes):

src/transformers/models/rt_detr/modeling_rt_detr.py

@@ -733,13 +733,14 @@ def forward(self, hidden_state):
 
 # Copied from transformers.models.deformable_detr.modeling_deformable_detr.multi_scale_deformable_attention
 def multi_scale_deformable_attention(
-    value: Tensor, value_spatial_shapes: Tensor, sampling_locations: Tensor, attention_weights: Tensor
+    value: Tensor,
+    value_spatial_shapes: Union[Tensor, List[Tuple]],
+    sampling_locations: Tensor,
+    attention_weights: Tensor,
 ) -> Tensor:
     batch_size, _, num_heads, hidden_dim = value.shape
     _, num_queries, num_heads, num_levels, num_points, _ = sampling_locations.shape
-    # Ignore copy
     value_list = value.split([height * width for height, width in value_spatial_shapes], dim=1)
-
     sampling_grids = 2 * sampling_locations - 1
     sampling_value_list = []
     for level_id, (height, width) in enumerate(value_spatial_shapes):
@@ -861,8 +862,6 @@ def forward(
 
         batch_size, num_queries, _ = hidden_states.shape
         batch_size, sequence_length, _ = encoder_hidden_states.shape
-
-        # Ignore copy
         total_elements = sum(shape[0] * shape[1] for shape in spatial_shapes_list)
         if total_elements != sequence_length:
             raise ValueError(
@@ -899,7 +898,6 @@ def forward(
         else:
             raise ValueError(f"Last dim of reference_points must be 2 or 4, but got {reference_points.shape[-1]}")
 
-        # Ignore copy
         if self.disable_custom_kernels or MultiScaleDeformableAttention is None:
             # PyTorch implementation
             output = multi_scale_deformable_attention(