refactor(linear_attn): extract GDN into shared LinearAttBackend (#1390)

Co-authored-by: wangzaijun <wzjhelloworld@qq.com>

Author: zhangts20

lightllm/common/basemodel/attention/base_att.py

@@ -70,6 +70,11 @@ class AttControl:
     nsa_prefill_dict: Dict = None
     nsa_decode: bool = False
     nsa_decode_dict: Dict = None
+    # linear attention 专用传参项
+    linear_att_prefill: bool = False
+    linear_att_prefill_dict: Dict = None
+    linear_att_decode: bool = False
+    linear_att_decode_dict: Dict = None
 
 
 @dataclass

lightllm/common/basemodel/attention/linear/__init__.py

@@ -0,0 +1,11 @@
+from .gdn import (
+    LinearAttBackend,
+    LinearAttPrefillAttState,
+    LinearAttDecodeAttState,
+)
+
+__all__ = [
+    "LinearAttBackend",
+    "LinearAttPrefillAttState",
+    "LinearAttDecodeAttState",
+]

lightllm/common/basemodel/attention/linear/gdn.py

@@ -0,0 +1,18 @@
+"""Linear-attention / GDN triton kernels shared across hybrid models."""
+
+from .causal_conv1d import causal_conv1d_fn
+from .causal_conv1d_spec import causal_conv1d_update
+from .fused_gdn_gating import fused_gdn_gating
+from .gdn_decode_pack import conv_pack_gdn_decode_inputs
+from .mtp_fused_recurrent import mtp_fused_recurrent_gated_delta_rule
+from .fla.ops import chunk_gated_delta_rule, fused_recurrent_gated_delta_rule
+
+__all__ = [
+    "causal_conv1d_fn",
+    "causal_conv1d_update",
+    "fused_gdn_gating",
+    "conv_pack_gdn_decode_inputs",
+    "mtp_fused_recurrent_gated_delta_rule",
+    "chunk_gated_delta_rule",
+    "fused_recurrent_gated_delta_rule",
+]