[NPU] support dsv32 radixcache on ascend (#17964)

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

@@ -628,7 +628,7 @@ class AscendAttnBackend(AttentionBackend):
             if self.forward_metadata.actual_seq_lengths_q is not None:
                 actual_seq_qlen = self.forward_metadata.actual_seq_lengths_q
             else:
-                actual_seq_qlen = torch.cumsum(forward_batch.seq_lens, dim=0)
+                actual_seq_qlen = torch.cumsum(forward_batch.extend_seq_lens, dim=0)
         else:
             if self.forward_metadata.actual_seq_lengths_q is None:
                 if (

python/sglang/srt/hardware_backend/npu/memory_pool_npu.py

@@ -221,6 +221,7 @@ class NPUMLATokenToKVPool(MLATokenToKVPool):
                 dtype=self.store_dtype,
                 device=self.device,
             )
+            self.index_k_buffer = None
             if self.index_head_dim is not None:
                 self.index_k_buffer = torch.zeros(
                     (