[NPU] perf: precompute mamba conv-state track indices once per batch (#29105)

python/sglang/srt/hardware_backend/npu/attention/ascend_gdn_backend.py

@@ -47,6 +47,15 @@ class AscendGDNAttnBackend(AscendMambaAttnBackendBase):
         prefill_backend = get_linear_attn_prefill_backend()
         self.kernel_dispatcher = GDNKernelDispatcher(decode_backend, prefill_backend)
 
+    def _prepare_mamba_track_metadata(self, forward_batch: ForwardBatch):
+        if self.forward_metadata.has_mamba_track_mask:
+            mamba_track_mask_indices = forward_batch.mamba_track_mask.nonzero(
+                as_tuple=True
+            )[0]
+            self.forward_metadata.conv_states_mask_indices = (
+                forward_batch.mamba_track_indices[mamba_track_mask_indices]
+            )
+
     def prepare_gdn_inputs(
         self,
         bs: int,
@@ -85,6 +94,7 @@ class AscendGDNAttnBackend(AscendMambaAttnBackendBase):
             forward_batch.forward_mode,
             forward_batch.spec_info,
         )
+        self._prepare_mamba_track_metadata(forward_batch)
         self.graph_mode = True
 
     def init_forward_metadata(self, forward_batch: ForwardBatch):
@@ -96,6 +106,7 @@ class AscendGDNAttnBackend(AscendMambaAttnBackendBase):
             forward_batch.forward_mode,
             forward_batch.spec_info,
         )
+        self._prepare_mamba_track_metadata(forward_batch)
         self.graph_mode = False
 
     def forward_decode(
@@ -221,16 +232,13 @@ class AscendGDNAttnBackend(AscendMambaAttnBackendBase):
             ).view(seq_len, -1)
         else:
             mixed_qkv = mixed_qkv.transpose(0, 1)
-            if (
-                forward_batch.mamba_track_mask is not None
-                and forward_batch.mamba_track_mask.any()
-            ):
-                conv_dst = forward_batch.mamba_track_indices
+            if forward_metadata.has_mamba_track_mask:
                 mixed_qkv_to_track = mixed_qkv[
                     :, forward_metadata.track_conv_indices
                 ].transpose(0, 1)
-                mask_indices = forward_batch.mamba_track_mask.nonzero(as_tuple=True)[0]
-                conv_states.transpose(1, 2)[conv_dst[mask_indices]] = mixed_qkv_to_track
+                conv_states.transpose(1, 2)[
+                    forward_metadata.conv_states_mask_indices
+                ] = mixed_qkv_to_track
             kernel_size = layer.conv_weights.shape[-1]
             conv_states_for_prefill = conv_states[:, -(kernel_size - 1) :, :]
             conv_states_tmp = conv_states_for_prefill.transpose(1, 2).contiguous()