update flash_attn.py

vllm/attention/backends/flash_attn.py

@@ -977,7 +977,6 @@ class FlashAttentionImpl(AttentionImpl):
                         v_descale=layer._v_scale.expand(descale_shape),
                     )
                 else:
-                    decode_output = decode_output.unsqueeze(1)
                     decode_output = flash_attn_with_kvcache(
                         q=decode_query.unsqueeze(1),
                         k_cache=key_cache,