diff --git a/python/sglang/srt/hardware_backend/npu/graph_runner/npu_graph_runner.py b/python/sglang/srt/hardware_backend/npu/graph_runner/npu_graph_runner.py
index 46daed9cc..2869c4bd1 100644
--- a/python/sglang/srt/hardware_backend/npu/graph_runner/npu_graph_runner.py
+++ b/python/sglang/srt/hardware_backend/npu/graph_runner/npu_graph_runner.py
@@ -241,8 +241,8 @@ class NPUGraphRunner(DecodeCudaGraphRunner):
             output = self.backend.replay_with_input_update(
                 graph_key,
                 seq_lens=seq_lens,
-                attr_name=self.attr_name[AttentionArch.MLA],
-                attr_type=self.attr_type[AttentionArch.MLA],
+                attr_name=self._get_update_attr_name(),
+                attr_type=self._get_update_attr_type(),
             )
         else:
             output = self.backend.replay(graph_key, forward_batch)
diff --git a/python/sglang/srt/mem_cache/allocator/swa.py b/python/sglang/srt/mem_cache/allocator/swa.py
index 4b44b39f1..66e49aadb 100644
--- a/python/sglang/srt/mem_cache/allocator/swa.py
+++ b/python/sglang/srt/mem_cache/allocator/swa.py
@@ -10,6 +10,8 @@ from sglang.srt.utils.common import get_num_new_pages
 _is_npu = is_npu()
 
 if _is_npu:
+    import torch_npu
+
     from sglang.srt.hardware_backend.npu.allocator_npu import (
         NPUPagedTokenToKVPoolAllocator,
     )
@@ -297,8 +299,11 @@ class SWATokenToKVPoolAllocator(BaseTokenToKVPoolAllocator):
             return None
 
         if _is_npu:
-            self.full_to_swa_index_mapping[alloc_full_indices.to(torch.int64)] = (
-                alloc_swa_indices.to(torch.int64)
+            indices_2d = alloc_full_indices.to(torch.int64).unsqueeze(-1)
+            torch_npu.npu_scatter_nd_update_(
+                self.full_to_swa_index_mapping,
+                indices_2d,
+                alloc_swa_indices.to(torch.int64),
             )
         else:
             self.full_to_swa_index_mapping[alloc_full_indices] = alloc_swa_indices
diff --git a/python/sglang/srt/speculative/eagle_info_v2.py b/python/sglang/srt/speculative/eagle_info_v2.py
index e54abb5a4..7b8a402d6 100644
--- a/python/sglang/srt/speculative/eagle_info_v2.py
+++ b/python/sglang/srt/speculative/eagle_info_v2.py
@@ -370,7 +370,11 @@ class EagleDraftExtendInputV2Mixin:
             # (the `_batch_size == batch_size` assertion, see #27091); the
             # marked pre-pad metadata is used as-is, matching the proven
             # skip_attn_backend_init=True behavior.
-            forward_batch.mark_forward_metadata_ready()
+            # On NPU with --disable-cuda-graph, block_table shape won't match
+            # after prepare_mlp_sync_batch padding; defer re-init to
+            # forward_extend (post-pad) instead.
+            if not _is_npu or can_cuda_graph:
+                forward_batch.mark_forward_metadata_ready()
         return forward_batch
 
 
diff --git a/python/sglang/srt/speculative/multi_layer_eagle_worker_v2.py b/python/sglang/srt/speculative/multi_layer_eagle_worker_v2.py
index 1f0e6d925..1f77b4269 100644
--- a/python/sglang/srt/speculative/multi_layer_eagle_worker_v2.py
+++ b/python/sglang/srt/speculative/multi_layer_eagle_worker_v2.py
@@ -562,7 +562,12 @@ class MultiLayerEagleDraftWorker(BaseDraftWorker):
             # pre-plan (see warning above). Mark the batch so the forward path
             # keeps skipping metadata init — preserves the pre-existing
             # behavior; the latent issue is tracked by the warning.
-            forward_batch.mark_forward_metadata_ready()
+            # On NPU with --disable-cuda-graph, leave each draft runner to init
+            # its own metadata in forward_extend (post-pad), otherwise
+            # per-runner attn_backend.forward_metadata is never initialized for
+            # draft_runner_list[1+].
+            if not _is_npu or can_cuda_graph:
+                forward_batch.mark_forward_metadata_ready()
 
         for step in range(self.speculative_num_steps):
             # log_info_on_rank0(logger, f"step: {step}, forward_batch.input_ids: {forward_batch.input_ids}")
@@ -821,7 +826,10 @@ class MultiLayerEagleWorkerV2(BaseSpecWorker):
         # prepare_for_v2_verify only plans when cuda-graph replay_prepare ran.
         # eagle_worker_v2 re-inits the non-graph path instead (post-pad); this
         # worker has not adopted that fix, so preserve its behavior verbatim.
-        verify_forward_batch.mark_forward_metadata_ready()
+        # On NPU with --disable-cuda-graph, non-graph verify needs metadata init
+        # in forward_extend (post-pad); only mark ready for the cuda-graph path.
+        if not _is_npu or can_run_cuda_graph:
+            verify_forward_batch.mark_forward_metadata_ready()
         # Run target verify batch in the main compute stream
         forward_batch_output = self.target_worker.forward_batch_generation(
             batch=None,