diff --git a/python/sglang/kernels/ops/speculative/multi_layer_eagle.py b/python/sglang/kernels/ops/speculative/multi_layer_eagle.py
index a41c9f1d7..ebdc10dc4 100644
--- a/python/sglang/kernels/ops/speculative/multi_layer_eagle.py
+++ b/python/sglang/kernels/ops/speculative/multi_layer_eagle.py
@@ -15,9 +15,10 @@
 import triton
 import triton.language as tl
 
-from sglang.srt.utils import is_cpu
+from sglang.srt.utils import is_cpu, is_npu
 
 _is_cpu = is_cpu()
+_is_npu = is_npu()
 
 if _is_cpu:
     from sgl_kernel import rotate_input_ids_cpu
@@ -74,6 +75,13 @@ def rotate_input_ids(
         return input_ids
 
     batch_size = extend_seq_lens.shape[0]
+
+    # rotate_input_ids_triton skipped: batch_size=0 (empty extend_seq_lens).
+    # This is expected when a DP rank has no requests.
+    # TODO: @iforgetmyname Remove NPU-specific guard after triton-ascend fixes zero-sized grid kernel launch abort
+    if batch_size == 0 and _is_npu:
+        return input_ids
+
     BLOCK_SIZE = 4096 if select_index is not None else 8
     grid = (batch_size,)
 
diff --git a/python/sglang/srt/hardware_backend/npu/memory_pool_npu.py b/python/sglang/srt/hardware_backend/npu/memory_pool_npu.py
index 718b46220..a685406f2 100644
--- a/python/sglang/srt/hardware_backend/npu/memory_pool_npu.py
+++ b/python/sglang/srt/hardware_backend/npu/memory_pool_npu.py
@@ -63,6 +63,7 @@ class NPUMHATokenToKVPool(MHATokenToKVPool):
         end_layer: Optional[int] = None,
         enable_alt_stream: bool = True,
         enable_kv_cache_copy: bool = False,
+        **kwargs,
     ):
         self.use_fia = get_bool_env_var("ASCEND_USE_FIA", "False")
         super().__init__(
@@ -82,6 +83,7 @@ class NPUMHATokenToKVPool(MHATokenToKVPool):
             end_layer=end_layer,
             enable_alt_stream=enable_alt_stream,
             enable_kv_cache_copy=enable_kv_cache_copy,
+            **kwargs,
         )
 
     def _create_buffers(self):