[wenet/squeezeformer] compat with websocket server (#1544)

* [update] compat with websocket server

* compat with decoder chunk size -1

* fix format issues

wenet/squeezeformer/encoder.py

@@ -30,7 +30,6 @@
 from wenet.squeezeformer.convolution import ConvolutionModule
 from wenet.utils.mask import make_pad_mask, add_optional_chunk_mask
 from wenet.utils.common import get_activation
-import math
 
 
 class SqueezeformerEncoder(nn.Module):
@@ -381,16 +380,16 @@ def forward_chunk(
 
             xs, _, new_att_cache, new_cnn_cache = layer(
                 xs, att_mask, pos_emb,
-                att_cache=att_cache[i:i + 1][:, :, ::factor, :] if
-                elayers > 0 and att_cache.size(2) != 0 else
-                att_cache[:, :, ::factor, :],
+                att_cache=att_cache[i:i + 1][:, :, ::factor, :]
+                [:, :, :pos_emb.size(1) - xs.size(1), :] if
+                elayers > 0 else att_cache[:, :, ::factor, :],
                 cnn_cache=cnn_cache[i] if cnn_cache.size(0) > 0 else cnn_cache
             )
             # NOTE(xcsong): After layer.forward
             #   shape(new_att_cache) is (1, head, attention_key_size, d_k * 2),
             #   shape(new_cnn_cache) is (b=1, hidden-dim, cache_t2)
             cached_att \
-                = new_att_cache[:, :, math.ceil(next_cache_start / factor):, :]
+                = new_att_cache[:, :, next_cache_start // factor:, :]
             cached_cnn = new_cnn_cache.unsqueeze(0)
             cached_att = cached_att.repeat_interleave(repeats=factor, dim=2)
             if i == 0: