You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Both MultiheadAttention and TransformerLayer work fine without the rotary_pos_emb in the forward pass. However, when passing rotary_pos_emb, I get the error above.
It seems using RoPE (i.e., passing rotary embeddings in the forward of MHA) is causing query_layer and key_layer's memory layouts to change. Is this expected and is there a workaround?
The text was updated successfully, but these errors were encountered:
@cyanguwa Thank you for the quick response. We tested the changes, along with the newly implemented RotaryPositionEmbedding. They worked perfectly from our side so far. Also doing some further tests, will keep you posted on the result. Thanks again.
Both MultiheadAttention and TransformerLayer work fine without the rotary_pos_emb in the forward pass. However, when passing rotary_pos_emb, I get the error above.
It seems using RoPE (i.e., passing rotary embeddings in the forward of MHA) is causing query_layer and key_layer's memory layouts to change. Is this expected and is there a workaround?
The text was updated successfully, but these errors were encountered: