Question about positional embeddings

[luzhoulz]

Genie-Envisioner/models/ltx_models/transformer_ltx_multiview.py

Lines 100 to 110 in 4c874d9

	if image_rotary_emb is not None: # for self attn, extend the sequence length according to the cross_view_attn param
	query = apply_rotary_emb(query, image_rotary_emb)
	key = apply_rotary_emb(key, image_rotary_emb)
	if cross_view_attn:
	query = rearrange(query, '(b v) l c -> b (v l) c', v=n_view)
	key = rearrange(key, '(b v) l c -> b (v l) c', v=n_view)
	value = rearrange(value, '(b v) l c -> b (v l) c', v=n_view)
	else: # for cross attn, extend the sequence length
	query = rearrange(query, '(b v) l c -> b (v l) c', v=n_view)

I noticed that you used positional embeddings in self-attention, but not in cross-attention. Is this a mistake, or did you do it intentionally?