diff --git a/python/sglang/kernels/ops/attention/__init__.py b/python/sglang/kernels/ops/attention/__init__.py
index 7e46e7803..5d10a0662 100644
--- a/python/sglang/kernels/ops/attention/__init__.py
+++ b/python/sglang/kernels/ops/attention/__init__.py
@@ -43,6 +43,26 @@ del _mod, _fn
 __all__ = []
 
 
+# Generic attention kernels migrated in Phase 2.5 (RFC #29630).
+for _mod, _fn in [
+    ("utils", "mla_quantize_and_rope_for_fp8"),
+    ("utils", "launch_reshape_and_cache_flash"),
+    ("utils", "launch_reshape_and_cache_shuffle_5d"),
+    ("flash_mla_sm120", "flash_mla_with_kvcache_sm120"),
+    ("dcp_kernels", "create_dcp_kv_indices"),
+    ("dcp_kernels", "correct_attn_out"),
+    ("pa_page_table", "_build_pa_page_table"),
+    ("nsa_triton_decode", "triton_sparse_attn_decode"),
+]:
+    register_kernel(
+        KernelSpec(
+            op=f"attention.{_fn.lstrip('_')}",
+            backend=KernelBackend.TRITON,
+            target=f"sglang.kernels.ops.attention.{_mod}:{_fn}",
+        )
+    )
+del _mod, _fn
+
 # RoPE / QK-norm fusion kernels migrated from srt/layers top-level strays
 # (RFC #29630, Phase 2.5); registered for inventory.
 for _mod, _fn in [
diff --git a/python/sglang/srt/layers/dcp/kernels.py b/python/sglang/kernels/ops/attention/dcp_kernels.py
similarity index 100%
rename from python/sglang/srt/layers/dcp/kernels.py
rename to python/sglang/kernels/ops/attention/dcp_kernels.py
diff --git a/python/sglang/srt/layers/attention/flash_mla_sm120.py b/python/sglang/kernels/ops/attention/flash_mla_sm120.py
similarity index 99%
rename from python/sglang/srt/layers/attention/flash_mla_sm120.py
rename to python/sglang/kernels/ops/attention/flash_mla_sm120.py
index 12c74f105..5bca1372a 100644
--- a/python/sglang/srt/layers/attention/flash_mla_sm120.py
+++ b/python/sglang/kernels/ops/attention/flash_mla_sm120.py
@@ -235,7 +235,7 @@ def flash_mla_with_kvcache_sm120(**kwargs):
         )
 
     if _sm120_default_backend == "triton":
-        from sglang.srt.layers.attention.flash_mla_sm120_triton import (
+        from sglang.kernels.ops.attention.flash_mla_sm120_triton import (
             flash_mla_sparse_decode_triton,
         )
 
diff --git a/python/sglang/srt/layers/attention/flash_mla_sm120_triton.py b/python/sglang/kernels/ops/attention/flash_mla_sm120_triton.py
similarity index 100%
rename from python/sglang/srt/layers/attention/flash_mla_sm120_triton.py
rename to python/sglang/kernels/ops/attention/flash_mla_sm120_triton.py
diff --git a/python/sglang/srt/layers/attention/nsa/triton_decode/__init__.py b/python/sglang/kernels/ops/attention/nsa_triton_decode/__init__.py
similarity index 96%
rename from python/sglang/srt/layers/attention/nsa/triton_decode/__init__.py
rename to python/sglang/kernels/ops/attention/nsa_triton_decode/__init__.py
index 7762b8bd2..f3e8bfc39 100644
--- a/python/sglang/srt/layers/attention/nsa/triton_decode/__init__.py
+++ b/python/sglang/kernels/ops/attention/nsa_triton_decode/__init__.py
@@ -9,7 +9,7 @@ from typing import Optional, Tuple
 
 import torch
 
-from sglang.srt.layers.attention.nsa.triton_decode.triton_mla_kernels_decode_optimized import (
+from sglang.kernels.ops.attention.nsa_triton_decode.triton_mla_kernels_decode_optimized import (
     triton_sparse_attn_decode,
 )
 
diff --git a/python/sglang/srt/layers/attention/nsa/triton_decode/triton_mla_kernels_decode_fused.py b/python/sglang/kernels/ops/attention/nsa_triton_decode/triton_mla_kernels_decode_fused.py
similarity index 100%
rename from python/sglang/srt/layers/attention/nsa/triton_decode/triton_mla_kernels_decode_fused.py
rename to python/sglang/kernels/ops/attention/nsa_triton_decode/triton_mla_kernels_decode_fused.py
diff --git a/python/sglang/srt/layers/attention/nsa/triton_decode/triton_mla_kernels_decode_optimized.py b/python/sglang/kernels/ops/attention/nsa_triton_decode/triton_mla_kernels_decode_optimized.py
similarity index 100%
rename from python/sglang/srt/layers/attention/nsa/triton_decode/triton_mla_kernels_decode_optimized.py
rename to python/sglang/kernels/ops/attention/nsa_triton_decode/triton_mla_kernels_decode_optimized.py
diff --git a/python/sglang/kernels/ops/attention/pa_page_table.py b/python/sglang/kernels/ops/attention/pa_page_table.py
new file mode 100644
index 000000000..cb6f00968
--- /dev/null
+++ b/python/sglang/kernels/ops/attention/pa_page_table.py
@@ -0,0 +1,98 @@
+"""Paged-attention page-table builder, migrated from
+``sglang.srt.layers.attention.flashattention_backend`` (RFC #29630, Phase 2.5).
+"""
+
+from typing import Optional
+
+import torch
+import triton
+import triton.language as tl
+
+
+@triton.jit
+def _build_pa_page_table_kernel(
+    req_to_token_ptr,
+    req_pool_indices_ptr,
+    seq_lens_ptr,
+    prefill_lens_ptr,
+    dst_page_table_ptr,
+    kv_lens_ptr,
+    window_size: tl.constexpr,
+    req_to_token_stride,
+    dst_stride,
+    BLOCK_SIZE: tl.constexpr,
+):
+    """Build PA-SWA page_table directly from req_to_token.
+
+    For each request, dst row = [0..prefill_len) ∪ [decode_start..seq_len).
+    decode_start = max(prefill_len, seq_len - window_size)
+
+    prefill_lens_ptr is the full pool-sized buffer, prefill_len is loaded
+    via indirect indexing using req_idx.
+    """
+    bid = tl.program_id(0)
+    req_idx = tl.load(req_pool_indices_ptr + bid)
+    sl = tl.load(seq_lens_ptr + bid).to(tl.int32)
+    pf = tl.load(prefill_lens_ptr + req_idx).to(tl.int32)
+
+    decode_start = tl.maximum(pf, sl - window_size)
+    gap = tl.where(decode_start > pf, decode_start - pf, 0)
+    kv_len = sl - gap
+
+    tl.store(kv_lens_ptr + bid, kv_len)
+
+    src_base = req_idx * req_to_token_stride
+    dst_base = bid * dst_stride
+
+    for start in tl.range(0, kv_len, BLOCK_SIZE):
+        offs = start + tl.arange(0, BLOCK_SIZE)
+        mask = offs < kv_len
+        pos = tl.where(offs < pf, offs, offs + gap)
+        kv_loc = tl.load(
+            req_to_token_ptr + src_base + pos,
+            mask=mask,
+            other=0,
+        )
+        tl.store(dst_page_table_ptr + dst_base + offs, kv_loc.to(tl.int32), mask=mask)
+
+
+def _build_pa_page_table(
+    req_to_token: torch.Tensor,
+    req_pool_indices: torch.Tensor,
+    seq_lens: torch.Tensor,
+    prefill_lens: torch.Tensor,
+    window_size: int,
+    bs: int,
+    pa_max_len: int,
+    device: torch.device,
+    dst_page_table: Optional[torch.Tensor] = None,
+    dst_kv_lens: Optional[torch.Tensor] = None,
+):
+    """Build prefill-aware page_table from req_to_token.
+
+    When dst_page_table/dst_kv_lens are None, allocates new tensors (non-CUDA-graph).
+    When provided, writes in-place into existing buffers (CUDA-graph replay).
+
+    prefill_lens is the full pool-sized buffer; the kernel indexes it via
+    req_pool_indices values (indirect indexing, avoids external gather).
+
+    Returns (page_table, kv_lens).
+    """
+    if dst_page_table is None:
+        dst_page_table = torch.zeros(bs, pa_max_len, dtype=torch.int32, device=device)
+    if dst_kv_lens is None:
+        dst_kv_lens = torch.empty(bs, dtype=torch.int32, device=device)
+    if bs > 0 and pa_max_len > 0:
+        _build_pa_page_table_kernel[(bs,)](
+            req_to_token,
+            req_pool_indices.contiguous(),
+            seq_lens.to(torch.int32),
+            prefill_lens,
+            dst_page_table,
+            dst_kv_lens,
+            window_size,
+            req_to_token.stride(0),
+            dst_page_table.stride(0),
+            BLOCK_SIZE=256,
+        )
+    return dst_page_table, dst_kv_lens
diff --git a/python/sglang/srt/layers/attention/utils.py b/python/sglang/kernels/ops/attention/utils.py
similarity index 100%
rename from python/sglang/srt/layers/attention/utils.py
rename to python/sglang/kernels/ops/attention/utils.py
diff --git a/python/sglang/srt/layers/attention/aiter_backend.py b/python/sglang/srt/layers/attention/aiter_backend.py
index 047e1849c..60ef93923 100755
--- a/python/sglang/srt/layers/attention/aiter_backend.py
+++ b/python/sglang/srt/layers/attention/aiter_backend.py
@@ -14,17 +14,17 @@ from typing import TYPE_CHECKING, Optional
 import torch
 import triton
 
-from sglang.kernels.ops.kvcache.aiter_unified_attention import (
-    scatter_ragged_to_page_table_kernel,
-    scatter_req_to_token_to_page_table_kernel,
-)
-from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
-from sglang.srt.layers.attention.utils import (
+from sglang.kernels.ops.attention.utils import (
     assert_buffer_fits,
     create_flashinfer_kv_indices_triton,
     create_flashmla_kv_indices_triton,
     get_num_kv_index_blocks_flashmla,
 )
+from sglang.kernels.ops.kvcache.aiter_unified_attention import (
+    scatter_ragged_to_page_table_kernel,
+    scatter_req_to_token_to_page_table_kernel,
+)
+from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
 from sglang.srt.layers.dp_attention import (
     is_dp_attention_enabled,
 )
@@ -60,16 +60,16 @@ except ImportError:
         "aiter is AMD specific kernel library. Please make sure aiter is installed on your AMD device."
     )
 
+from sglang.kernels.ops.attention.utils import (
+    launch_reshape_and_cache_flash,
+    pad_sequence_with_mask,
+)
 from sglang.kernels.ops.quantization.fp8_kernel import fp8_dtype
 from sglang.srt.configs.model_config import AttentionArch
 from sglang.srt.layers.attention.aiter_utils import (
     forward_decode_vectorized_5d,
     forward_extend_vectorized_5d,
 )
-from sglang.srt.layers.attention.utils import (
-    launch_reshape_and_cache_flash,
-    pad_sequence_with_mask,
-)
 from sglang.srt.mem_cache.memory_pool import KVWriteLoc
 from sglang.srt.mem_cache.swa_memory_pool import SWAKVPool
 from sglang.srt.utils import get_bool_env_var
diff --git a/python/sglang/srt/layers/attention/aiter_utils.py b/python/sglang/srt/layers/attention/aiter_utils.py
index b0af9fa94..9d57584cc 100644
--- a/python/sglang/srt/layers/attention/aiter_utils.py
+++ b/python/sglang/srt/layers/attention/aiter_utils.py
@@ -33,8 +33,8 @@ except ImportError:  # pragma: no cover - import-time guard mirrors aiter_backen
     pa_decode_gluon = None
     get_recommended_splits = None
 
+from sglang.kernels.ops.attention.utils import launch_gather_shuffle_5d_to_linear
 from sglang.kernels.ops.quantization.fp8_kernel import fp8_dtype
-from sglang.srt.layers.attention.utils import launch_gather_shuffle_5d_to_linear
 
 if TYPE_CHECKING:
     from sglang.srt.layers.attention.aiter_backend import AiterAttnBackend
diff --git a/python/sglang/srt/layers/attention/cutlass_mla_backend.py b/python/sglang/srt/layers/attention/cutlass_mla_backend.py
index ca156d0e7..0bd908a8c 100644
--- a/python/sglang/srt/layers/attention/cutlass_mla_backend.py
+++ b/python/sglang/srt/layers/attention/cutlass_mla_backend.py
@@ -13,11 +13,11 @@ from typing import TYPE_CHECKING, Optional, Union
 import torch
 import triton
 
-from sglang.srt.layers.attention.flashinfer_mla_backend import FlashInferMLAAttnBackend
-from sglang.srt.layers.attention.utils import (
+from sglang.kernels.ops.attention.utils import (
     create_flashmla_kv_indices_triton,
     get_num_kv_index_blocks_flashmla,
 )
+from sglang.srt.layers.attention.flashinfer_mla_backend import FlashInferMLAAttnBackend
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.utils import is_cuda
 
diff --git a/python/sglang/srt/layers/attention/deepseek_v4_backend.py b/python/sglang/srt/layers/attention/deepseek_v4_backend.py
index 316f42685..6fc16e4b1 100644
--- a/python/sglang/srt/layers/attention/deepseek_v4_backend.py
+++ b/python/sglang/srt/layers/attention/deepseek_v4_backend.py
@@ -1671,7 +1671,7 @@ class DeepseekV4AttnBackend(
                 )
 
             if _is_sm120:
-                from sglang.srt.layers.attention.flash_mla_sm120 import (
+                from sglang.kernels.ops.attention.flash_mla_sm120 import (
                     flash_mla_with_kvcache_sm120,
                 )
 
diff --git a/python/sglang/srt/layers/attention/dsa/dsa_backend_mtp_precompute.py b/python/sglang/srt/layers/attention/dsa/dsa_backend_mtp_precompute.py
index ce26ff518..e4b90b7ef 100644
--- a/python/sglang/srt/layers/attention/dsa/dsa_backend_mtp_precompute.py
+++ b/python/sglang/srt/layers/attention/dsa/dsa_backend_mtp_precompute.py
@@ -11,8 +11,8 @@ from typing import TYPE_CHECKING, Optional
 
 import torch
 
+from sglang.kernels.ops.attention.utils import seqlens_expand_triton
 from sglang.srt.layers.attention.dsa.utils import compute_dsa_seqlens
-from sglang.srt.layers.attention.utils import seqlens_expand_triton
 from sglang.srt.utils import is_cuda, is_hip
 
 if TYPE_CHECKING:
diff --git a/python/sglang/srt/layers/attention/dsa_backend.py b/python/sglang/srt/layers/attention/dsa_backend.py
index 7416bb4f2..dae22035d 100644
--- a/python/sglang/srt/layers/attention/dsa_backend.py
+++ b/python/sglang/srt/layers/attention/dsa_backend.py
@@ -18,6 +18,11 @@ from sglang.srt.configs.model_config import get_dsa_index_topk, is_deepseek_dsa
 from sglang.srt.runtime_context import get_parallel
 
 logger = logging.getLogger(__name__)
+from sglang.kernels.ops.attention.utils import (
+    concat_mla_absorb_q_general,
+    mla_quantize_and_rope_for_fp8,
+    seqlens_expand_triton,
+)
 from sglang.srt.environ import envs
 from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
 from sglang.srt.layers.attention.dsa.dequant_k_cache import dequantize_k_cache_paged
@@ -47,11 +52,6 @@ from sglang.srt.layers.attention.dsa.utils import (
     pad_dsa_cache_seqlens,
     should_use_dsa_fused_topk,
 )
-from sglang.srt.layers.attention.utils import (
-    concat_mla_absorb_q_general,
-    mla_quantize_and_rope_for_fp8,
-    seqlens_expand_triton,
-)
 from sglang.srt.layers.utils.cp_utils import (
     cp_all_gather_rerange_output,
     cp_split_and_rebuild_position,
diff --git a/python/sglang/srt/layers/attention/flashattention_backend.py b/python/sglang/srt/layers/attention/flashattention_backend.py
index 038e515ef..ec88e1ac5 100644
--- a/python/sglang/srt/layers/attention/flashattention_backend.py
+++ b/python/sglang/srt/layers/attention/flashattention_backend.py
@@ -5,19 +5,18 @@ from typing import TYPE_CHECKING, Optional
 
 import numpy as np
 import torch
-import triton
-import triton.language as tl
 
 from sglang.kernels.ops.attention.metadata import (
     normal_decode_set_metadata,
     prepare_swa_spec_page_table_triton,
 )
+from sglang.kernels.ops.attention.pa_page_table import _build_pa_page_table
+from sglang.kernels.ops.attention.utils import assert_buffer_fits
 from sglang.kernels.ops.kvcache.trtllm_mha_page_table import (
     build_trtllm_mha_page_table,
 )
 from sglang.srt.configs.model_config import AttentionArch
 from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
-from sglang.srt.layers.attention.utils import assert_buffer_fits
 from sglang.srt.layers.cp.base import CPAttentionBackendKind, get_cp_strategy
 from sglang.srt.layers.cp.utils import is_cp_v2_active
 from sglang.srt.layers.radix_attention import AttentionType
@@ -50,95 +49,6 @@ def _should_disable_scheduler_metadata_precompute(server_args) -> bool:
     return bool(server_args.enable_prefill_cp or server_args.enable_dp_attention)
 
 
-@triton.jit
-def _build_pa_page_table_kernel(
-    req_to_token_ptr,
-    req_pool_indices_ptr,
-    seq_lens_ptr,
-    prefill_lens_ptr,
-    dst_page_table_ptr,
-    kv_lens_ptr,
-    window_size: tl.constexpr,
-    req_to_token_stride,
-    dst_stride,
-    BLOCK_SIZE: tl.constexpr,
-):
-    """Build PA-SWA page_table directly from req_to_token.
-
-    For each request, dst row = [0..prefill_len) ∪ [decode_start..seq_len).
-    decode_start = max(prefill_len, seq_len - window_size)
-
-    prefill_lens_ptr is the full pool-sized buffer, prefill_len is loaded
-    via indirect indexing using req_idx.
-    """
-    bid = tl.program_id(0)
-    req_idx = tl.load(req_pool_indices_ptr + bid)
-    sl = tl.load(seq_lens_ptr + bid).to(tl.int32)
-    pf = tl.load(prefill_lens_ptr + req_idx).to(tl.int32)
-
-    decode_start = tl.maximum(pf, sl - window_size)
-    gap = tl.where(decode_start > pf, decode_start - pf, 0)
-    kv_len = sl - gap
-
-    tl.store(kv_lens_ptr + bid, kv_len)
-
-    src_base = req_idx * req_to_token_stride
-    dst_base = bid * dst_stride
-
-    for start in tl.range(0, kv_len, BLOCK_SIZE):
-        offs = start + tl.arange(0, BLOCK_SIZE)
-        mask = offs < kv_len
-        pos = tl.where(offs < pf, offs, offs + gap)
-        kv_loc = tl.load(
-            req_to_token_ptr + src_base + pos,
-            mask=mask,
-            other=0,
-        )
-        tl.store(dst_page_table_ptr + dst_base + offs, kv_loc.to(tl.int32), mask=mask)
-
-
-def _build_pa_page_table(
-    req_to_token: torch.Tensor,
-    req_pool_indices: torch.Tensor,
-    seq_lens: torch.Tensor,
-    prefill_lens: torch.Tensor,
-    window_size: int,
-    bs: int,
-    pa_max_len: int,
-    device: torch.device,
-    dst_page_table: Optional[torch.Tensor] = None,
-    dst_kv_lens: Optional[torch.Tensor] = None,
-):
-    """Build prefill-aware page_table from req_to_token.
-
-    When dst_page_table/dst_kv_lens are None, allocates new tensors (non-CUDA-graph).
-    When provided, writes in-place into existing buffers (CUDA-graph replay).
-
-    prefill_lens is the full pool-sized buffer; the kernel indexes it via
-    req_pool_indices values (indirect indexing, avoids external gather).
-
-    Returns (page_table, kv_lens).
-    """
-    if dst_page_table is None:
-        dst_page_table = torch.zeros(bs, pa_max_len, dtype=torch.int32, device=device)
-    if dst_kv_lens is None:
-        dst_kv_lens = torch.empty(bs, dtype=torch.int32, device=device)
-    if bs > 0 and pa_max_len > 0:
-        _build_pa_page_table_kernel[(bs,)](
-            req_to_token,
-            req_pool_indices.contiguous(),
-            seq_lens.to(torch.int32),
-            prefill_lens,
-            dst_page_table,
-            dst_kv_lens,
-            window_size,
-            req_to_token.stride(0),
-            dst_page_table.stride(0),
-            BLOCK_SIZE=256,
-        )
-    return dst_page_table, dst_kv_lens
-
-
 @dataclass
 class FlashAttentionMetadata:
     """Metadata to be init once in the model forward pass,
diff --git a/python/sglang/srt/layers/attention/flashinfer_backend.py b/python/sglang/srt/layers/attention/flashinfer_backend.py
index a55f7f5e7..97b288ac3 100644
--- a/python/sglang/srt/layers/attention/flashinfer_backend.py
+++ b/python/sglang/srt/layers/attention/flashinfer_backend.py
@@ -19,13 +19,13 @@ from typing import TYPE_CHECKING, Callable, List, Optional, Union
 import torch
 
 from sglang.kernel_api_logging import debug_kernel_api
-from sglang.srt.dllm.config import DllmConfig
-from sglang.srt.environ import envs
-from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
-from sglang.srt.layers.attention.utils import (
+from sglang.kernels.ops.attention.utils import (
     assert_buffer_fits,
     create_flashinfer_kv_indices_triton,
 )
+from sglang.srt.dllm.config import DllmConfig
+from sglang.srt.environ import envs
+from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
 from sglang.srt.layers.radix_attention import AttentionType
 from sglang.srt.mem_cache.base_swa_memory_pool import BaseSWAKVPool
 from sglang.srt.mem_cache.memory_pool import KVWriteLoc
diff --git a/python/sglang/srt/layers/attention/flashinfer_mla_backend.py b/python/sglang/srt/layers/attention/flashinfer_mla_backend.py
index ce57f3567..0d68b303c 100644
--- a/python/sglang/srt/layers/attention/flashinfer_mla_backend.py
+++ b/python/sglang/srt/layers/attention/flashinfer_mla_backend.py
@@ -17,12 +17,12 @@ from typing import TYPE_CHECKING, Callable, Optional, Union
 
 import torch
 
+from sglang.kernels.ops.attention.utils import assert_buffer_fits
 from sglang.srt.environ import envs
 from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
 from sglang.srt.layers.attention.flashinfer_backend import (
     create_flashinfer_kv_indices_triton,
 )
-from sglang.srt.layers.attention.utils import assert_buffer_fits
 from sglang.srt.layers.dcp import (
     DecodeContextParallelMetadata,
     update_local_kv_lens_for_dcp,
diff --git a/python/sglang/srt/layers/attention/flashmla_backend.py b/python/sglang/srt/layers/attention/flashmla_backend.py
index 484f06a25..19e420d2a 100644
--- a/python/sglang/srt/layers/attention/flashmla_backend.py
+++ b/python/sglang/srt/layers/attention/flashmla_backend.py
@@ -12,12 +12,12 @@ import torch
 import triton
 from sgl_kernel.flash_mla import flash_mla_with_kvcache, get_mla_metadata
 
-from sglang.kernels.ops.quantization.fp8_kernel import scaled_fp8_quant
-from sglang.srt.layers.attention.flashinfer_mla_backend import FlashInferMLAAttnBackend
-from sglang.srt.layers.attention.utils import (
+from sglang.kernels.ops.attention.utils import (
     create_flashmla_kv_indices_triton,
     get_num_kv_index_blocks_flashmla,
 )
+from sglang.kernels.ops.quantization.fp8_kernel import scaled_fp8_quant
+from sglang.srt.layers.attention.flashinfer_mla_backend import FlashInferMLAAttnBackend
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
 from sglang.srt.runtime_context import get_parallel
 
diff --git a/python/sglang/srt/layers/attention/hip_flash_mla.py b/python/sglang/srt/layers/attention/hip_flash_mla.py
index d771926cc..7643d190e 100644
--- a/python/sglang/srt/layers/attention/hip_flash_mla.py
+++ b/python/sglang/srt/layers/attention/hip_flash_mla.py
@@ -38,7 +38,7 @@ def flash_mla_with_kvcache_entrypoint(backend: str, **kwargs):
         return dpsk_v4_fp8_attention_fwd(**kwargs)
 
     if backend == "triton":
-        from sglang.srt.layers.attention.nsa.triton_decode import (
+        from sglang.kernels.ops.attention.nsa_triton_decode import (
             triton_fp8_attention_fwd,
         )
 
diff --git a/python/sglang/srt/layers/attention/trtllm_mha_backend.py b/python/sglang/srt/layers/attention/trtllm_mha_backend.py
index b08a748f4..c3e3346b1 100644
--- a/python/sglang/srt/layers/attention/trtllm_mha_backend.py
+++ b/python/sglang/srt/layers/attention/trtllm_mha_backend.py
@@ -11,6 +11,7 @@ from typing import TYPE_CHECKING, Optional
 
 import torch
 
+from sglang.kernels.ops.attention.utils import canonicalize_stride
 from sglang.kernels.ops.kvcache.trtllm_fp8_kv_kernel import (
     fused_fp8_set_kv_buffer,
 )
@@ -27,7 +28,6 @@ from sglang.srt.layers.attention.flashinfer_backend import (
     FlashInferAttnBackend,
     FlashInferMultiStepDraftBackend,
 )
-from sglang.srt.layers.attention.utils import canonicalize_stride
 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
diff --git a/python/sglang/srt/layers/attention/trtllm_mla_backend.py b/python/sglang/srt/layers/attention/trtllm_mla_backend.py
index 49f072591..aeec6d1e4 100755
--- a/python/sglang/srt/layers/attention/trtllm_mla_backend.py
+++ b/python/sglang/srt/layers/attention/trtllm_mla_backend.py
@@ -19,6 +19,10 @@ from sglang.kernels.ops.attention.pad import (
 from sglang.kernels.ops.attention.pad import (
     unpad_draft_extend_output as unpad_draft_extend_output_triton,
 )
+from sglang.kernels.ops.attention.utils import (
+    concat_mla_absorb_q_general,
+    mla_quantize_and_rope_for_fp8,
+)
 from sglang.kernels.ops.kvcache.kv_indices import (
     create_flashmla_kv_indices_triton,
     get_num_kv_index_blocks_flashmla,
@@ -30,10 +34,6 @@ from sglang.srt.layers.attention.flashinfer_mla_backend import (
     FlashInferMLAAttnBackend,
     FlashInferMLAMultiStepDraftBackend,
 )
-from sglang.srt.layers.attention.utils import (
-    concat_mla_absorb_q_general,
-    mla_quantize_and_rope_for_fp8,
-)
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
 from sglang.srt.model_executor.runner_backend_utils.tc_piecewise_cuda_graph import (
     is_in_tc_piecewise_cuda_graph,
diff --git a/python/sglang/srt/layers/dcp/__init__.py b/python/sglang/srt/layers/dcp/__init__.py
index 631026d32..eb6aeea65 100644
--- a/python/sglang/srt/layers/dcp/__init__.py
+++ b/python/sglang/srt/layers/dcp/__init__.py
@@ -29,6 +29,9 @@ Package-internal helpers (the @triton.jit kernels, ``CPTritonContext``,
 out-of-tree callers; in-tree code should use ``get_parallel().dcp_enabled`` and
 ``get_parallel().attn_dcp_*``."""
 
+from sglang.kernels.ops.attention.dcp_kernels import (
+    create_triton_kv_indices_for_dcp_triton,
+)
 from sglang.srt.layers.dcp.comm import (
     all_gather_kv_cache_for_dcp,
     all_gather_kv_cache_for_mha_chunk_extend,
@@ -41,7 +44,6 @@ from sglang.srt.layers.dcp.comm import (
     get_attention_dcp_rank,
     get_attention_dcp_world_size,
 )
-from sglang.srt.layers.dcp.kernels import create_triton_kv_indices_for_dcp_triton
 from sglang.srt.layers.dcp.layout import (
     filter_dcp_local_kv_indices,
     get_dcp_lens,
diff --git a/python/sglang/srt/layers/dcp/comm.py b/python/sglang/srt/layers/dcp/comm.py
index 7ba6e813b..c540eeb20 100644
--- a/python/sglang/srt/layers/dcp/comm.py
+++ b/python/sglang/srt/layers/dcp/comm.py
@@ -25,11 +25,11 @@ from typing import Optional
 
 import torch
 
+from sglang.kernels.ops.attention.dcp_kernels import CPTritonContext, correct_attn_out
 from sglang.srt.distributed.device_communicators.pynccl_allocator import (
     use_symmetric_memory,
 )
 from sglang.srt.distributed.parallel_state import GroupCoordinator
-from sglang.srt.layers.dcp.kernels import CPTritonContext, correct_attn_out
 from sglang.srt.runtime_context import get_parallel
 
 
diff --git a/python/sglang/srt/layers/dcp/planner.py b/python/sglang/srt/layers/dcp/planner.py
index c7ee3748e..1a9caba76 100644
--- a/python/sglang/srt/layers/dcp/planner.py
+++ b/python/sglang/srt/layers/dcp/planner.py
@@ -20,7 +20,7 @@ from typing import Optional
 
 import torch
 
-from sglang.srt.layers.dcp.kernels import (
+from sglang.kernels.ops.attention.dcp_kernels import (
     create_dcp_kv_indices,
     update_kv_lens_and_indices,
 )
diff --git a/python/sglang/srt/layers/rotary_embedding/base.py b/python/sglang/srt/layers/rotary_embedding/base.py
index 95e5b2e7e..72e693b66 100644
--- a/python/sglang/srt/layers/rotary_embedding/base.py
+++ b/python/sglang/srt/layers/rotary_embedding/base.py
@@ -65,7 +65,7 @@ if _is_npu:
         )
 
 if _is_hip:
-    from sglang.srt.layers.attention.utils import (
+    from sglang.kernels.ops.attention.utils import (
         fused_qk_rope_reshape_and_cache,
     )
 
diff --git a/python/sglang/srt/mem_cache/memory_pool.py b/python/sglang/srt/mem_cache/memory_pool.py
index e863151c3..260376c9a 100644
--- a/python/sglang/srt/mem_cache/memory_pool.py
+++ b/python/sglang/srt/mem_cache/memory_pool.py
@@ -1837,7 +1837,7 @@ class MHATokenToKVPool(KVCache):
         # and viewed as ``store_dtype`` by ``set_kv_buffer``.
         if self.kv_cache_layout == "vectorized_5d":
             # Late-import to keep the NHD path import-clean.
-            from sglang.srt.layers.attention.utils import (
+            from sglang.kernels.ops.attention.utils import (
                 launch_reshape_and_cache_shuffle_5d,
             )
 
diff --git a/python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mha.py b/python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mha.py
index 0673ea5be..037e868fc 100644
--- a/python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mha.py
+++ b/python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mha.py
@@ -4,10 +4,10 @@ from typing import TYPE_CHECKING
 
 import torch
 
+from sglang.kernels.ops.attention.utils import concat_and_cast_mha_k_triton
 from sglang.srt.environ import envs
 from sglang.srt.layers.attention.dsa.dequant_k_cache import dequantize_k_cache_paged
 from sglang.srt.layers.attention.tbo_backend import TboAttnBackend
-from sglang.srt.layers.attention.utils import concat_and_cast_mha_k_triton
 from sglang.srt.layers.communicator import get_attn_tp_context
 from sglang.srt.layers.dcp import (
     all_gather_kv_cache_for_mha_chunk_extend,
diff --git a/python/sglang/srt/models/sarvam_moe.py b/python/sglang/srt/models/sarvam_moe.py
index 4a343a13f..4d184e11b 100644
--- a/python/sglang/srt/models/sarvam_moe.py
+++ b/python/sglang/srt/models/sarvam_moe.py
@@ -12,6 +12,7 @@ import torch.nn.functional as F
 from torch import nn
 from transformers import PretrainedConfig
 
+from sglang.kernels.ops.attention.utils import concat_and_cast_mha_k_triton
 from sglang.srt.distributed import (
     get_pp_group,
     tensor_model_parallel_all_reduce,
@@ -19,7 +20,6 @@ from sglang.srt.distributed import (
 from sglang.srt.eplb.expert_distribution import get_global_expert_distribution_recorder
 from sglang.srt.eplb.expert_location import ModelConfigForExpertLocation
 from sglang.srt.layers.activation import SiluAndMul
-from sglang.srt.layers.attention.utils import concat_and_cast_mha_k_triton
 from sglang.srt.layers.communicator import (
     LayerCommunicator,
     LayerScatterModes,
diff --git a/python/sglang/srt/speculative/dflash_info.py b/python/sglang/srt/speculative/dflash_info.py
index ef2b8c136..762e5d164 100644
--- a/python/sglang/srt/speculative/dflash_info.py
+++ b/python/sglang/srt/speculative/dflash_info.py
@@ -5,7 +5,7 @@ from typing import TYPE_CHECKING, Optional, Tuple
 
 import torch
 
-from sglang.srt.layers.attention.utils import create_flashinfer_kv_indices_triton
+from sglang.kernels.ops.attention.utils import create_flashinfer_kv_indices_triton
 from sglang.srt.managers.schedule_batch import ScheduleBatch
 from sglang.srt.model_executor.forward_batch_info import (
     CaptureHiddenMode,
diff --git a/python/sglang/srt/speculative/eagle_info.py b/python/sglang/srt/speculative/eagle_info.py
index ca4bd35f0..f6ff20d05 100644
--- a/python/sglang/srt/speculative/eagle_info.py
+++ b/python/sglang/srt/speculative/eagle_info.py
@@ -4,9 +4,9 @@ from typing import List, Optional, Tuple
 
 import torch
 
+from sglang.kernels.ops.attention.utils import create_flashinfer_kv_indices_triton
 from sglang.srt.constrained.base_grammar_backend import BaseGrammarObject
 from sglang.srt.environ import envs
-from sglang.srt.layers.attention.utils import create_flashinfer_kv_indices_triton
 from sglang.srt.model_executor.forward_batch_info import CaptureHiddenMode
 from sglang.srt.runtime_context import get_server_args
 from sglang.srt.speculative.spec_info import SpecInput, SpecInputType
diff --git a/python/sglang/srt/speculative/ngram_info.py b/python/sglang/srt/speculative/ngram_info.py
index 958b2cfe7..5c96872e9 100644
--- a/python/sglang/srt/speculative/ngram_info.py
+++ b/python/sglang/srt/speculative/ngram_info.py
@@ -4,8 +4,8 @@ from typing import Optional, Tuple
 
 import torch
 
+from sglang.kernels.ops.attention.utils import create_flashinfer_kv_indices_triton
 from sglang.srt.constrained.base_grammar_backend import BaseGrammarObject
-from sglang.srt.layers.attention.utils import create_flashinfer_kv_indices_triton
 from sglang.srt.speculative.spec_info import SpecInput, SpecInputType
 
 
diff --git a/test/manual/attention/test_trtllm_mla_backend.py b/test/manual/attention/test_trtllm_mla_backend.py
index f5aec11f7..344e5bcce 100755
--- a/test/manual/attention/test_trtllm_mla_backend.py
+++ b/test/manual/attention/test_trtllm_mla_backend.py
@@ -10,13 +10,13 @@ from sglang.srt.runtime_context import get_parallel, get_server_args
 _parallel_override = get_parallel().override(attn_tp_size=1)
 _parallel_override.__enter__()
 
+from sglang.kernels.ops.attention.utils import get_num_page_per_block_flashmla
 from sglang.srt.configs.model_config import AttentionArch
 from sglang.srt.layers.attention.flashinfer_mla_backend import FlashInferMLAAttnBackend
 from sglang.srt.layers.attention.trtllm_mla_backend import (
     TRTLLMMLABackend,
     TRTLLMMLADecodeMetadata,
 )
-from sglang.srt.layers.attention.utils import get_num_page_per_block_flashmla
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.mem_cache.memory_pool import MLATokenToKVPool
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
diff --git a/test/registered/attention/test_create_kvindices.py b/test/registered/attention/test_create_kvindices.py
index 898e84e80..907a12f86 100644
--- a/test/registered/attention/test_create_kvindices.py
+++ b/test/registered/attention/test_create_kvindices.py
@@ -3,7 +3,7 @@ import unittest
 import numpy as np
 import torch
 
-from sglang.srt.layers.attention.utils import create_flashinfer_kv_indices_triton
+from sglang.kernels.ops.attention.utils import create_flashinfer_kv_indices_triton
 from sglang.srt.utils import get_device
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
diff --git a/test/registered/kernels/test_flash_mla_backends.py b/test/registered/kernels/test_flash_mla_backends.py
index 06ab30e85..98dee7bf4 100644
--- a/test/registered/kernels/test_flash_mla_backends.py
+++ b/test/registered/kernels/test_flash_mla_backends.py
@@ -26,8 +26,7 @@ from unittest import mock
 
 import torch
 
-from sglang.srt.layers.attention import flash_mla_sm120 as fmod
-from sglang.srt.layers.attention.flash_mla_sm120 import (
+from sglang.kernels.ops.attention.flash_mla_sm120 import (
     _D,
     _NOPE_DIM,
     _NOPE_ROPE_STRIDE,
@@ -39,11 +38,12 @@ from sglang.srt.layers.attention.flash_mla_sm120 import (
     _sm120_sparse_decode_fwd,
     flash_mla_with_kvcache_sm120,
 )
-from sglang.srt.layers.attention.flash_mla_sm120_triton import (
+from sglang.kernels.ops.attention.flash_mla_sm120_triton import (
     _apply_attn_sink,
     _merge_partial_attn,
     flash_mla_sparse_decode_triton,
 )
+from sglang.srt.layers.attention import flash_mla_sm120 as fmod
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import CustomTestCase