Question about TemporalSelfAttention

[usergxx]

TemporalSelfAttention does not address the issue of extracting indices from value based on query. There is still flaw here.

During training mode, in the TemporalSelfAttention module, the generation of the first frame's query takes the previous frame's value. However, this value is the original query without encoding. Therefore, the historical queries of the first frame actually lack positional encoding. Shouldn't they be added?