[Feature]Supports SWA based on appendattn (#6547)

fastdeploy/model_executor/layers/attention/ops/append_attention.py

@@ -83,11 +83,73 @@ def append_attention(
     causal: bool = True,
     speculate_decoder: bool = False,
     sliding_window: int = 0,
+    sink_size: int = 0,
+    head_wise_full_hidden: int = 0,
 ) -> paddle.Tensor:
     """
     append_attention
     """
     if current_platform.is_cuda():
+        if sliding_window > 0 and head_wise_full_hidden > 0:
+            out_swa = append_attention_gpu(
+                qkv.clone(),
+                key_cache,
+                value_cache,
+                seq_lens_encoder,
+                seq_lens_decoder,
+                seq_lens_this_time,
+                batch_id_per_token,
+                cu_seqlens_q,
+                block_tables,
+                encoder_batch_ids,
+                encoder_tile_ids_per_batch,
+                encoder_num_blocks,
+                kv_batch_ids,
+                kv_tile_ids_per_batch,
+                kv_num_blocks,
+                decoder_batch_ids,
+                decoder_tile_ids_per_batch,
+                decoder_num_blocks,
+                set_max_lengths,
+                rotary_embs,
+                attn_mask,
+                qkv_bias,
+                qkv_scale,
+                k_quant_scale,
+                v_quant_scale,
+                k_dequant_scale,
+                v_dequant_scale,
+                cache_k_zp,
+                cache_v_zp,
+                linear_shift,
+                linear_smooth,
+                mask_offset,
+                kv_signal_data,
+                q_norm_weight,
+                k_norm_weight,
+                sinks,
+                rms_norm_eps,
+                compute_type,
+                cache_quant_type,
+                use_neox_rotary_style,
+                rope_3d,
+                max_input_length,
+                quant_max_bound,
+                quant_min_bound,
+                out_linear_in_scale,
+                encoder_block_shape_q,
+                decoder_block_shape_q,
+                max_partition_size,
+                encoder_max_partition_size,
+                speculate_max_draft_token_num,
+                causal,
+                speculate_decoder,
+                sliding_window,
+                sink_size,
+            )
+            sliding_window = 0
+            sink_size = 0
+
         out = append_attention_gpu(
             qkv,
             key_cache,
@@ -142,7 +204,12 @@ def append_attention(
             causal,
             speculate_decoder,
             sliding_window,
+            sink_size,
         )
+        if head_wise_full_hidden > 0:
+            out_swa[:, :head_wise_full_hidden] = out[:, :head_wise_full_hidden]
+            return out_swa
+
         return out
     else:
         raise NotImplementedError
@@ -266,6 +333,7 @@ def append_attention_with_output(
             causal,
             speculate_decoder,
             sliding_window,
+            0,
         )
     else:
         raise NotImplementedError