Invoke the out_balancer of attention_squeeze

This commit is contained in:
Daniel Povey 2022-12-12 23:31:22 +08:00
parent 0522425ea8
commit b5e0676f14

View File

@ -1445,6 +1445,7 @@ attn_weights: a Tensor of shape (num_heads, batch_size, seq_len, seq_len)
x = x * scales
x = self.activation(x) # Identity only. For diagnostics.
x = self.out_proj(x)
x = self.out_balancer(x)
return x