[ONNX] Sam fix (#23110)

* [WIP] Fix for the ONNX export

* Apply changes

* Remove commented code

* Resolve todo

* empty -> zeros

* fix slow tests

---------

Co-authored-by: younesbelkada <younesbelkada@gmail.com>

src/transformers/models/sam/modeling_sam.py

@@ -223,9 +223,7 @@ class SamAttention(nn.Module):
     def _recombine_heads(self, hidden_states: Tensor, point_batch_size: int) -> Tensor:
         batch, n_heads, n_tokens, c_per_head = hidden_states.shape
         hidden_states = hidden_states.transpose(1, 2)
-        return hidden_states.reshape(
+        return hidden_states.reshape(batch // point_batch_size, point_batch_size, n_tokens, n_heads * c_per_head)
-            batch // max(1, point_batch_size), point_batch_size, n_tokens, n_heads * c_per_head
-        )
 
     def forward(self, query: Tensor, key: Tensor, value: Tensor) -> Tensor:
         # Input projections
@@ -482,7 +480,7 @@ class SamMaskDecoder(nn.Module):
                 Whether or not to return the attentions tensors of all attention layers.
         """
         batch_size, num_channels, height, width = image_embeddings.shape
-        point_batch_size = max(1, sparse_prompt_embeddings.shape[1])
+        point_batch_size = sparse_prompt_embeddings.shape[1]
         # Concatenate output tokens
         output_tokens = torch.cat([self.iou_token.weight, self.mask_tokens.weight], dim=0)
         output_tokens = output_tokens.repeat(batch_size, point_batch_size, 1, 1)
@@ -634,8 +632,18 @@ class SamPromptEncoder(nn.Module):
             torch.tensor(0.0, dtype=point_embedding.dtype, device=point_embedding.device),
         )
 
-        point_embedding[labels == 0] += self.point_embed[0].weight
+        point_embedding = torch.where(
-        point_embedding[labels == 1] += self.point_embed[1].weight
+            (labels == 0)[:, :, :, None],
+            point_embedding + self.point_embed[0].weight[None, None, :, :],
+            point_embedding,
+        )
+
+        point_embedding = torch.where(
+            (labels == 1)[:, :, :, None],
+            point_embedding + self.point_embed[1].weight[None, None, :, :],
+            point_embedding,
+        )
+
         return point_embedding
 
     def _embed_boxes(self, boxes: torch.Tensor) -> torch.Tensor:
@@ -675,8 +683,7 @@ class SamPromptEncoder(nn.Module):
             if input_labels is None:
                 raise ValueError("If points are provided, labels must also be provided.")
             point_embeddings = self._embed_points(input_points, input_labels, pad=(input_boxes is None))
-            sparse_embeddings = torch.empty((batch_size, point_batch_size, 0, self.hidden_size), device=target_device)
+            sparse_embeddings = point_embeddings
-            sparse_embeddings = torch.cat([sparse_embeddings, point_embeddings], dim=2)
         if input_boxes is not None:
             batch_size = input_boxes.shape[0]
             box_embeddings = self._embed_boxes(input_boxes)
@@ -692,7 +699,7 @@ class SamPromptEncoder(nn.Module):
             )
 
         if sparse_embeddings is None:
-            sparse_embeddings = torch.empty((batch_size, 0, 1, self.hidden_size), device=target_device)
+            sparse_embeddings = torch.zeros((batch_size, 1, 1, self.hidden_size), device=target_device)
 
         return sparse_embeddings, dense_embeddings
 
@@ -742,8 +749,6 @@ class SamVisionAttention(nn.Module):
             Extracted positional embeddings according to relative positions.
         """
         max_rel_dist = int(2 * max(q_size, k_size) - 1)
-        # Interpolate rel pos if needed.
-        if rel_pos.shape[0] != max_rel_dist:
         # Interpolate rel pos.
         rel_pos_resized = F.interpolate(
             rel_pos.reshape(1, rel_pos.shape[0], -1).permute(0, 2, 1),
@@ -751,8 +756,6 @@ class SamVisionAttention(nn.Module):
             mode="linear",
         )
         rel_pos_resized = rel_pos_resized.reshape(-1, max_rel_dist).permute(1, 0)
-        else:
-            rel_pos_resized = rel_pos
 
         # Scale the coords with short length if shapes for q and k are different.
         q_coords = torch.arange(q_size)[:, None] * max(k_size / q_size, 1.0)
@@ -865,7 +868,6 @@ class SamVisionLayer(nn.Module):
 
         pad_h = (window_size - height % window_size) % window_size
         pad_w = (window_size - width % window_size) % window_size
-        if pad_h > 0 or pad_w > 0:
         hidden_states = F.pad(hidden_states, (0, 0, 0, pad_w, 0, pad_h))
         pad_height, pad_width = height + pad_h, width + pad_w
 
@@ -902,7 +904,6 @@ class SamVisionLayer(nn.Module):
             hidden_states.permute(0, 1, 3, 2, 4, 5).contiguous().reshape(batch_size, pad_height, pad_width, -1)
         )
 
-        if pad_height > height or pad_width > width:
         hidden_states = hidden_states[:, :height, :width, :].contiguous()
         return hidden_states