[Optimization] fake scaling for fp16

funasr/export/models/encoder/sanm_encoder.py

@@ -72,6 +72,7 @@ class SANMEncoder(nn.Module):
         else:
             xs_pad = self.embed(speech)
 
+        # xs_pad = xs_pad / scale
         encoder_outs = self.model.encoders0(xs_pad, mask)
         xs_pad, masks = encoder_outs[0], encoder_outs[1]
 

funasr/export/models/modules/encoder_layer.py

@@ -24,6 +24,7 @@ class EncoderLayerSANM(nn.Module):
         residual = x
         x = self.norm1(x)
         x = self.self_attn(x, mask)
+        # x = x / scale
         if self.in_size == self.size:
             x = x + residual
         residual = x