In the paper，it mentioned that the work of the bidirectional language modeling pre-train has been done. Are you planning on releasing some pre-trained weights for the model?

this is more general(from my perspective)
q,k,v shape : (b,s,d)

q = q.contiguous()
q = rearrange(q, 'b n (h d) -> (b h) n d', h = self.num_heads)

(N * h, S, d)

k = k.contiguous()
k = rearrange(k, 'b n (h d) -> (b h) n d', h = self.num_heads)

(N * h, S, d)

v = v.contiguous()
v = rearrange(v, 'b n (h d) -> (b h) n d', h = self.num_heads)

Thanks for the awesome project.
I wonder when the code will be released. Thank you ;)

Thanks very much for your interesting work! I have a question about the O(N) space complexity mentioned in your paper. I am wondering whether you can help me to figure it out.

In Eq. (11) of your paper, you compute QK^T in the denominator, which may lead to O(N^2*d) space complexity?

bests

Compared with left_product function, attention mask is not used in forward() function.
How to use the attention mask in the forward method?

When implementing cosformer on MultiHeadAttention in Transformer-XL and running without extra long-range memory, the ReLU performance is worse than eLU. I think it is because the Attention and FF Net are different since XL-like transformer has different layer norm and residual connection. Why this ReLU(Q)ReLU(K).T softmax replacement is not robust on different transformer architectures?

Hello!

In the Figure 1, how do you get memory consumption of cosFormer on the LRA benchmark? Could you please open source script for computing it?

diff += torch.norm(left_res - right_res)
What does diff stand for

We are examining non-NLP applications of the cosformer self-attention, and would need to use attention masking for the padded tokens in the batch.
Is there a way to incorporate this ?
Because the code does not explicitly compute the attention weights on which masking is traditionally applied.

Thank you for your great work! I am currently working on a seq2seq task and I found the causal attention code only works the src_len and the tgt_len are the same. Also, I suggest that you could adopt EPFL's causal linear attention CUDA code to improve the speed of causal attention.