diff --git a/python/sglang/srt/layers/attention/aiter_backend.py b/python/sglang/srt/layers/attention/aiter_backend.py
index ccf0c796a..273b71a83 100755
--- a/python/sglang/srt/layers/attention/aiter_backend.py
+++ b/python/sglang/srt/layers/attention/aiter_backend.py
@@ -27,6 +27,7 @@ from sglang.srt.layers.dp_attention import (
     is_dp_attention_enabled,
 )
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
+from sglang.srt.speculative.spec_utils import generate_draft_decode_kv_indices
 from sglang.srt.utils import is_gfx95_supported
 
 if TYPE_CHECKING:
@@ -2776,8 +2777,6 @@ class AiterMultiStepDraftBackend:
         topk: int,
         speculative_num_steps: int,
     ):
-        from sglang.srt.speculative.spec_utils import generate_draft_decode_kv_indices
-
         self.topk = topk
         self.speculative_num_steps = speculative_num_steps
         self.generate_draft_decode_kv_indices = generate_draft_decode_kv_indices
diff --git a/python/sglang/srt/layers/attention/flashinfer_backend.py b/python/sglang/srt/layers/attention/flashinfer_backend.py
index 70fd29d65..629cf4c25 100644
--- a/python/sglang/srt/layers/attention/flashinfer_backend.py
+++ b/python/sglang/srt/layers/attention/flashinfer_backend.py
@@ -27,6 +27,7 @@ from sglang.srt.layers.radix_attention import AttentionType
 from sglang.srt.mem_cache.allocator.swa import SWATokenToKVPoolAllocator
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
 from sglang.srt.speculative.spec_info import SpecInput
+from sglang.srt.speculative.spec_utils import generate_draft_decode_kv_indices
 from sglang.srt.utils import (
     get_int_env_var,
     is_flashinfer_available,
@@ -1538,8 +1539,6 @@ class FlashInferMultiStepDraftBackend:
         topk: int,
         speculative_num_steps: int,
     ):
-        from sglang.srt.speculative.spec_utils import generate_draft_decode_kv_indices
-
         self.topk = topk
         self.speculative_num_steps = speculative_num_steps
         self.generate_draft_decode_kv_indices = generate_draft_decode_kv_indices
diff --git a/python/sglang/srt/layers/attention/flashinfer_mla_backend.py b/python/sglang/srt/layers/attention/flashinfer_mla_backend.py
index da4881e88..716c947c7 100644
--- a/python/sglang/srt/layers/attention/flashinfer_mla_backend.py
+++ b/python/sglang/srt/layers/attention/flashinfer_mla_backend.py
@@ -25,6 +25,7 @@ from sglang.srt.layers.dp_attention import get_attention_tp_size
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
 from sglang.srt.server_args import get_global_server_args
 from sglang.srt.speculative.spec_info import SpecInput
+from sglang.srt.speculative.spec_utils import generate_draft_decode_kv_indices
 from sglang.srt.utils import (
     is_flashinfer_available,
     is_sm100_supported,
@@ -892,8 +893,6 @@ class FlashInferMLAMultiStepDraftBackend:
         topk: int,
         speculative_num_steps: int,
     ):
-        from sglang.srt.speculative.spec_utils import generate_draft_decode_kv_indices
-
         if topk > 1:
             raise ValueError(
                 "Currently Flashinfer MLA only supports topk=1 for speculative decoding"
diff --git a/python/sglang/srt/speculative/spec_utils.py b/python/sglang/srt/speculative/spec_utils.py
index 8a7b370d8..6938a5918 100644
--- a/python/sglang/srt/speculative/spec_utils.py
+++ b/python/sglang/srt/speculative/spec_utils.py
@@ -9,15 +9,13 @@ from typing import TYPE_CHECKING, List, Optional
 import torch
 from huggingface_hub import snapshot_download
 
-from sglang.srt.constrained.base_grammar_backend import BaseGrammarObject
 from sglang.srt.distributed.parallel_state import (
     GroupCoordinator,
     patch_tensor_parallel_group,
 )
 from sglang.srt.environ import envs
-from sglang.srt.managers.schedule_batch import Req
 from sglang.srt.mem_cache.common import get_last_loc
-from sglang.srt.server_args import ServerArgs, get_global_server_args
+from sglang.srt.server_args import get_global_server_args
 from sglang.srt.speculative.triton_ops.cache_locs import (
     align_evict_mask_to_page_size as align_evict_mask_to_page_size,
 )
@@ -53,6 +51,9 @@ _is_npu = is_npu()
 _is_musa = is_musa()
 
 if TYPE_CHECKING:
+    from sglang.srt.constrained.base_grammar_backend import BaseGrammarObject
+    from sglang.srt.managers.schedule_batch import Req
+    from sglang.srt.server_args import ServerArgs
     from sglang.srt.speculative.eagle_info import EagleVerifyInput