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
n_heads × d_head -> d_head × d_head in DeltaNet (#903)
Clarified the explanation of the memory size calculation for `KV_cache_DeltaNet` and updated the quadratic term from `n_heads × d_head` to `d_head × d_head`.
Note that the`KV_cache_DeltaNet` memory size doesn't have a context length (`n_tokens`) dependency. Also, we have only the memory state S that we store instead of separate keys and values, hence`2 × bytes` becomes just`bytes`. However, note that we now have a quadratic`n_heads × d_head` in here. This comes from the state :
334
+
Note that the`KV_cache_DeltaNet` memory size doesn't have a context length (`n_tokens`) dependency. Also, we have only the memory state S that we store instead of separate keys and values, hence`2 × bytes` becomes just`bytes`. However, note that we now have a quadratic`d_head × d_head` in here. This comes from the state :
335
335
336
336
```
337
337
S = x.new_zeros(b, self.num_heads, self.head_dim, self.head_dim)