[NPU] enable index Cache for npu (#21502)

python/sglang/srt/hardware_backend/npu/modules/deepseek_v2_attention_mla_npu.py

@@ -302,6 +302,7 @@ def forward_dsa_prepare_npu(
     forward_batch: "ForwardBatch",
     zero_allocator: "BumpAllocator",
     layer_scatter_modes,
+    prev_topk_indices: torch.Tensor = None,
 ):
     dynamic_scale = None
     if is_mla_preprocess_enabled() and forward_batch.forward_mode.is_decode():
@@ -401,15 +402,18 @@ def forward_dsa_prepare_npu(
                 latent_cache, forward_batch, k_nope, k_pe
             )
 
-    topk_indices = m.indexer(
+    if m.skip_topk:
-        hidden_states,
+        topk_indices = prev_topk_indices
-        q_lora,
+    else:
-        positions,
+        topk_indices = m.indexer(
-        forward_batch,
+            hidden_states,
-        m.layer_id,
+            q_lora,
-        layer_scatter_modes,
+            positions,
-        dynamic_scale,
+            forward_batch,
-    )
+            m.layer_id,
+            layer_scatter_modes,
+            dynamic_scale,
+        )
 
     return (
         q_pe,
@@ -472,7 +476,10 @@ def forward_dsa_core_npu(
     attn_bmm_output = attn_bmm_output.reshape(-1, m.num_local_heads * m.v_head_dim)
 
     output, _ = m.o_proj(attn_bmm_output)
-    return output
+    if not m.next_skip_topk:
+        return output, None
+    else:
+        return output, topk_indices
 
 
 def npu_mla_preprocess(

python/sglang/srt/models/deepseek_v2.py

@@ -1467,6 +1467,7 @@ class DeepseekV2AttentionMLA(
                 forward_batch,
                 zero_allocator,
                 layer_scatter_modes,
+                prev_topk_indices,
             )
         else:
             raise NotImplementedError