RuntimeError when using batch size > 1

I get this error when batch size if larger than 1

`RuntimeError: The size of tensor a (2) must match the size of tensor b (4) at non-singleton dimension 0`

My sparse embedding size is [2,0,156] (empty with batch size 2)
My dense embedding size is [2,256,64,64] (batch size 2)

```
output_tokens = output_tokens.unsqueeze(0).expand(sparse_prompt_embeddings.size(0), -1, -1)
tokens = torch.cat((output_tokens, sparse_prompt_embeddings), dim=1)
src = torch.repeat_interleave(image_embeddings, tokens.shape[0], dim=0)
src = src + dense_prompt_embeddings
```

So the repeat_interleave extend the image_embeddings to 4 which is actually larger than the batch size

Am i missing something? or the repeat interleave is redundant?

Start at the code path containing the shown prompt-embedding and image-embedding operations, then reproduce the issue with batch size 2 and inspect the tensor shapes before concatenation and addition. Done means batched sparse and dense embeddings combine without a dimension mismatch and the existing batch-size-one behavior still works.

Written by the indexing model from the issue text.