diff --git a/python/sglang/srt/arg_groups/speculative_hook.py b/python/sglang/srt/arg_groups/speculative_hook.py
index 0c24ae32e..427acdd38 100644
--- a/python/sglang/srt/arg_groups/speculative_hook.py
+++ b/python/sglang/srt/arg_groups/speculative_hook.py
@@ -147,8 +147,10 @@ def handle_speculative_decoding(server_args: ServerArgs) -> None:
 def _handle_dflash(server_args: ServerArgs) -> None:
     from sglang.srt.arg_groups.overrides import resolved_view
 
-    if not server_args.device.startswith("cuda"):
-        raise ValueError("DFLASH speculative decoding only supports CUDA device.")
+    if not (server_args.device.startswith("cuda") or server_args.device == "npu"):
+        raise ValueError(
+            "DFLASH speculative decoding only supports CUDA and NPU devices."
+        )
 
     if resolved_view(server_args).enable_dp_attention:
         raise ValueError(
diff --git a/python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py b/python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py
index 6b9e3e50c..30761f228 100644
--- a/python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py
+++ b/python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py
@@ -27,7 +27,7 @@ from sglang.srt.mem_cache.memory_pool import KVWriteLoc
 from sglang.srt.mem_cache.swa_memory_pool import SWAKVPool
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
 from sglang.srt.runtime_context import get_flags
-from sglang.srt.speculative.spec_info import SpecInput
+from sglang.srt.speculative.spec_info import SpecInput, SpecInputType
 from sglang.srt.utils import get_bool_env_var, get_current_device_stream_fast
 
 if TYPE_CHECKING:
@@ -44,6 +44,13 @@ logger = logging.getLogger(__name__)
 FULL_ATTENTION_WINDOW = 2147483647
 
 
+def _is_dflash_verify(spec_info: Optional[SpecInput]) -> bool:
+    return (
+        spec_info is not None
+        and spec_info.spec_input_type == SpecInputType.DFLASH_VERIFY
+    )
+
+
 def _expand_dsa_sparse_indices(topk_indices: torch.Tensor) -> torch.Tensor:
     """Expand [T, K] to [T, 1, K] for NPU sparse attention."""
     if topk_indices.dim() == 2:
@@ -480,7 +487,9 @@ class AscendAttnBackend(AttentionBackend):
             seq_lens_list_cumsum = np.cumsum(forward_batch.extend_seq_lens_cpu)
             self.forward_metadata.seq_lens_list_cumsum = seq_lens_list_cumsum
 
-        if forward_batch.forward_mode.is_target_verify():
+        if forward_batch.forward_mode.is_target_verify() and not _is_dflash_verify(
+            forward_batch.spec_info
+        ):
             self.forward_metadata.seq_lens_cpu_int += self.speculative_num_draft_tokens
         elif (
             forward_batch.forward_mode.is_decode_or_idle()
@@ -667,7 +676,7 @@ class AscendAttnBackend(AttentionBackend):
                 self.token_to_kv_pool.translate_loc_from_full_to_swa(out_cache_loc)
             )
         max_len = seq_lens_cpu[:bs].max().item()
-        if forward_mode.is_target_verify():
+        if forward_mode.is_target_verify() and not _is_dflash_verify(spec_info):
             max_len += self.speculative_num_draft_tokens
         elif forward_mode.is_decode_or_idle() and spec_info is not None:
             max_len += self.speculative_step_id + 1