【NPU】【bugfix】accuracy fix when enable both nsa cp and prefixcache (#23268)

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

@@ -359,7 +359,9 @@ def forward_dsa_prepare_npu(
             if q_event is not None:
                 torch.npu.current_stream().wait_event(q_event)
         else:
-            if fused_qkv_a_proj_out.shape[0] < 65535:
+            if fused_qkv_a_proj_out.shape[0] < 65535 and not nsa_use_prefill_cp(
+                forward_batch
+            ):
                 q_lora, k_nope, k_pe = fused_split_qk_norm(
                     fused_qkv_a_proj_out,
                     m.q_a_layernorm,

python/sglang/srt/layers/attention/nsa/nsa_indexer.py

@@ -1463,6 +1463,20 @@ class Indexer(MultiPlatformOp):
                     forward_batch.attn_cp_metadata.actual_seq_q_prev_tensor,
                     forward_batch.attn_cp_metadata.actual_seq_q_next_tensor,
                 )
+                if sum(forward_batch.extend_prefix_lens_cpu) > 0:
+                    total_kv_len_prev_tensor = (
+                        forward_batch.attn_cp_metadata.kv_len_prev_tensor
+                        + forward_batch.extend_prefix_lens.squeeze()
+                    )
+                    total_kv_len_next_tensor = (
+                        forward_batch.attn_cp_metadata.kv_len_next_tensor
+                        + forward_batch.extend_prefix_lens.squeeze()
+                    )
+                    forward_batch.attn_backend.forward_metadata.actual_seq_lengths_kv = (
+                        total_kv_len_prev_tensor,
+                        total_kv_len_next_tensor,
+                    )
+                else:
                     forward_batch.attn_backend.forward_metadata.actual_seq_lengths_kv = (
                         forward_batch.attn_cp_metadata.kv_len_prev_tensor,
                         forward_batch.attn_cp_metadata.kv_len_next_tensor,