[misc] clean up kernel API (#21325)

python/sglang/jit_kernel/all_reduce.py

@@ -12,6 +12,7 @@ from sglang.jit_kernel.utils import (
     load_jit,
     make_cpp_args,
 )
+from sglang.kernel_api_logging import debug_kernel_api
 
 
 class ConfigResult(NamedTuple):
@@ -158,6 +159,7 @@ def get_custom_all_reduce_cls() -> type[CustomAllReduceObj]:
         def world_size(self) -> int:
             return self._world_size
 
+        @debug_kernel_api
         def all_reduce(
             self,
             input: torch.Tensor,

python/sglang/jit_kernel/awq_marlin_repack.py

@@ -4,8 +4,8 @@ from typing import TYPE_CHECKING
 
 import torch
 
-from sglang.jit_kernel.debug_utils import maybe_wrap_jit_kernel_debug
 from sglang.jit_kernel.utils import cache_once, load_jit
+from sglang.kernel_api_logging import debug_kernel_api
 
 if TYPE_CHECKING:
     from tvm_ffi.module import Module
@@ -20,7 +20,7 @@ def _jit_awq_marlin_repack_module() -> Module:
     )
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def awq_marlin_repack(
     b_q_weight: torch.Tensor,
     size_k: int,
@@ -39,7 +39,7 @@ def awq_marlin_repack(
     return out
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def awq_marlin_moe_repack(
     b_q_weight: torch.Tensor,
     perm: torch.Tensor,

python/sglang/jit_kernel/debug_utils.py

@@ -1,45 +0,0 @@
-import os
-from typing import Any, Callable, TypeVar, cast, overload
-
-F = TypeVar("F", bound=Callable[..., Any])
-
-
-def _wrap_jit_kernel_debug(func: F, op_name: str | None = None) -> F:
-    try:
-        if int(os.environ.get("SGLANG_KERNEL_API_LOGLEVEL", "0")) == 0:
-            return func
-    except Exception:
-        return func
-
-    try:
-        from sglang.kernel_api_logging import debug_kernel_api
-    except Exception:
-        return func
-
-    if getattr(func, "_debug_kernel_wrapped", False):
-        return func
-
-    wrapped = debug_kernel_api(func, op_name=op_name)
-    setattr(wrapped, "_debug_kernel_wrapped", True)
-    return cast(F, wrapped)
-
-
-@overload
-def maybe_wrap_jit_kernel_debug(func: F) -> F: ...
-
-
-@overload
-def maybe_wrap_jit_kernel_debug(func: F, op_name: str) -> F: ...
-
-
-@overload
-def maybe_wrap_jit_kernel_debug(*, op_name: str | None = None) -> Callable[[F], F]: ...
-
-
-def maybe_wrap_jit_kernel_debug(
-    func: F | None = None, op_name: str | None = None
-) -> F | Callable[[F], F]:
-    if func is None:
-        return lambda wrapped_func: _wrap_jit_kernel_debug(wrapped_func, op_name)
-
-    return _wrap_jit_kernel_debug(func, op_name)

python/sglang/jit_kernel/diffusion/triton/rmsnorm_onepass.py

@@ -2,7 +2,7 @@ import torch
 import triton  # type: ignore
 import triton.language as tl  # type: ignore
 
-from sglang.jit_kernel.debug_utils import maybe_wrap_jit_kernel_debug
+from sglang.kernel_api_logging import debug_kernel_api
 from sglang.multimodal_gen.runtime.platforms import current_platform
 from sglang.srt.utils.custom_op import register_custom_op
 
@@ -37,7 +37,6 @@ def _rms_norm_tiled_onepass(
     tl.store(y_blk, x * rstd * w, mask=mask)
 
 
-@maybe_wrap_jit_kernel_debug
 @register_custom_op(op_name="triton_one_pass_rms_norm_cuda", out_shape="x")
 def _triton_one_pass_rms_norm_cuda(
     x: torch.Tensor, w: torch.Tensor, eps: float = 1e-6
@@ -73,6 +72,6 @@ def triton_one_pass_rms_norm(x: torch.Tensor, w: torch.Tensor, eps: float = 1e-6
 if current_platform.is_mps():
     from .mps_fallback import triton_one_pass_rms_norm_native
 
-    @maybe_wrap_jit_kernel_debug
+    @debug_kernel_api
     def triton_one_pass_rms_norm(x: torch.Tensor, w: torch.Tensor, eps: float = 1e-6):
         return triton_one_pass_rms_norm_native(x, w, eps)

python/sglang/jit_kernel/flash_attention_v4.py

@@ -4,7 +4,7 @@ from typing import Callable, Optional, Tuple, Union
 
 import torch
 
-from sglang.jit_kernel.debug_utils import maybe_wrap_jit_kernel_debug
+from sglang.kernel_api_logging import debug_kernel_api
 
 try:
     from flash_attn.cute import flash_attn_varlen_func as _flash_attn_varlen_func
@@ -19,7 +19,7 @@ def _maybe_contiguous(x: Optional[torch.Tensor]) -> Optional[torch.Tensor]:
     return x.contiguous() if x is not None and x.stride(-1) != 1 else x
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def flash_attn_varlen_func(
     q: torch.Tensor,
     k: torch.Tensor,
@@ -92,7 +92,7 @@ def flash_attn_varlen_func(
     return result
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def flash_attn_with_kvcache(
     q: torch.Tensor,
     k_cache: torch.Tensor,

python/sglang/jit_kernel/fused_store_index_cache.py

@@ -13,13 +13,13 @@ from typing import TYPE_CHECKING
 
 import torch
 
-from sglang.jit_kernel.debug_utils import maybe_wrap_jit_kernel_debug
 from sglang.jit_kernel.utils import (
     cache_once,
     is_arch_support_pdl,
     load_jit,
     make_cpp_args,
 )
+from sglang.kernel_api_logging import debug_kernel_api
 
 if TYPE_CHECKING:
     from tvm_ffi.module import Module
@@ -65,7 +65,7 @@ def can_use_nsa_fused_store(
         return False
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def fused_store_index_k_cache(
     key: torch.Tensor,
     index_k_with_scale: torch.Tensor,

python/sglang/jit_kernel/gptq_marlin.py

@@ -4,8 +4,8 @@ from typing import TYPE_CHECKING, Optional
 
 import torch
 
-from sglang.jit_kernel.debug_utils import maybe_wrap_jit_kernel_debug
 from sglang.jit_kernel.utils import cache_once, load_jit, make_cpp_args
+from sglang.kernel_api_logging import debug_kernel_api
 
 if TYPE_CHECKING:
     from sgl_kernel.scalar_type import ScalarType
@@ -32,7 +32,7 @@ def _or_empty(
     return t if t is not None else torch.empty(0, device=device, dtype=dtype)
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def gptq_marlin_gemm(
     a: torch.Tensor,
     c: Optional[torch.Tensor],

python/sglang/jit_kernel/gptq_marlin_repack.py

@@ -4,8 +4,8 @@ from typing import TYPE_CHECKING
 
 import torch
 
-from sglang.jit_kernel.debug_utils import maybe_wrap_jit_kernel_debug
 from sglang.jit_kernel.utils import cache_once, load_jit
+from sglang.kernel_api_logging import debug_kernel_api
 
 if TYPE_CHECKING:
     from tvm_ffi.module import Module
@@ -23,7 +23,7 @@ def _jit_gptq_marlin_repack_module() -> Module:
     )
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def gptq_marlin_repack(
     b_q_weight: torch.Tensor,
     perm: torch.Tensor,

python/sglang/jit_kernel/hicache.py

@@ -3,8 +3,8 @@ from __future__ import annotations
 import logging
 from typing import TYPE_CHECKING
 
-from sglang.jit_kernel.debug_utils import maybe_wrap_jit_kernel_debug
 from sglang.jit_kernel.utils import cache_once, load_jit, make_cpp_args
+from sglang.kernel_api_logging import debug_kernel_api
 
 if TYPE_CHECKING:
     import torch
@@ -67,7 +67,7 @@ def _default_unroll(element_size: int) -> int:
     return 1
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def transfer_hicache_one_layer(
     k_cache_dst: torch.Tensor,
     v_cache_dst: torch.Tensor,
@@ -103,7 +103,7 @@ def transfer_hicache_one_layer(
     )
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def transfer_hicache_all_layer(
     k_ptr_dst: torch.Tensor,
     v_ptr_dst: torch.Tensor,

python/sglang/jit_kernel/kvcache.py

@@ -11,6 +11,7 @@ from sglang.jit_kernel.utils import (
     load_jit,
     make_cpp_args,
 )
+from sglang.srt.utils.custom_op import register_custom_op
 
 if TYPE_CHECKING:
     from tvm_ffi.module import Module
@@ -46,6 +47,7 @@ def can_use_store_cache(size: int) -> bool:
         return False
 
 
+@register_custom_op(mutates_args=["k_cache", "v_cache"])
 def store_cache(
     k: torch.Tensor,
     v: torch.Tensor,

python/sglang/jit_kernel/moe_wna16_marlin.py

@@ -4,8 +4,8 @@ from typing import TYPE_CHECKING, Optional
 
 import torch
 
-from sglang.jit_kernel.debug_utils import maybe_wrap_jit_kernel_debug
 from sglang.jit_kernel.utils import cache_once, load_jit, make_cpp_args
+from sglang.kernel_api_logging import debug_kernel_api
 
 if TYPE_CHECKING:
     from sgl_kernel.scalar_type import ScalarType
@@ -37,7 +37,7 @@ def _or_empty(
     return t if t is not None else torch.empty(0, device=device, dtype=dtype)
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def moe_wna16_marlin_gemm(
     a: torch.Tensor,
     c_or_none: Optional[torch.Tensor],

python/sglang/jit_kernel/ngram_embedding.py

@@ -2,8 +2,8 @@ from __future__ import annotations
 
 from typing import TYPE_CHECKING
 
-from sglang.jit_kernel.debug_utils import maybe_wrap_jit_kernel_debug
 from sglang.jit_kernel.utils import cache_once, load_jit
+from sglang.kernel_api_logging import debug_kernel_api
 
 if TYPE_CHECKING:
     import torch
@@ -22,7 +22,7 @@ def _jit_ngram_embedding_module() -> Module:
     )
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def compute_n_gram_ids(
     ne_n: int,
     ne_k: int,
@@ -68,7 +68,7 @@ def compute_n_gram_ids(
     )
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def update_token_table(
     tokens: torch.Tensor,
     ne_token_table: torch.Tensor,

python/sglang/jit_kernel/norm.py

@@ -5,13 +5,13 @@ from typing import TYPE_CHECKING, Optional
 
 import torch
 
-from sglang.jit_kernel.debug_utils import maybe_wrap_jit_kernel_debug
 from sglang.jit_kernel.utils import (
     cache_once,
     is_arch_support_pdl,
     load_jit,
     make_cpp_args,
 )
+from sglang.kernel_api_logging import debug_kernel_api
 
 if TYPE_CHECKING:
     from tvm_ffi.module import Module
@@ -99,7 +99,7 @@ def can_use_fused_inplace_qknorm(head_dim: int, dtype: torch.dtype) -> bool:
         return False
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def fused_inplace_qknorm(
     q: torch.Tensor,
     k: torch.Tensor,
@@ -114,7 +114,7 @@ def fused_inplace_qknorm(
     module.qknorm(q, k, q_weight, k_weight, eps)
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def rmsnorm(
     input: torch.Tensor,
     weight: torch.Tensor,
@@ -133,7 +133,7 @@ def rmsnorm(
     module.rmsnorm(input, weight, output, eps)
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def fused_add_rmsnorm(
     input: torch.Tensor,
     residual: torch.Tensor,
@@ -144,7 +144,7 @@ def fused_add_rmsnorm(
     module.fused_add_rmsnorm(input, residual, weight, eps)
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def fused_inplace_qknorm_across_heads(
     q: torch.Tensor,
     k: torch.Tensor,

python/sglang/jit_kernel/nvfp4.py

@@ -8,8 +8,8 @@ from typing import TYPE_CHECKING, Optional, Tuple
 
 import torch
 
-from sglang.jit_kernel.debug_utils import maybe_wrap_jit_kernel_debug
 from sglang.jit_kernel.utils import cache_once, load_jit
+from sglang.kernel_api_logging import debug_kernel_api
 from sglang.srt.utils.custom_op import register_custom_op
 
 if TYPE_CHECKING:
@@ -196,7 +196,7 @@ def _jit_nvfp4_blockwise_moe_module() -> Module:
         )
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def cutlass_scaled_fp4_mm(
     a: torch.Tensor,
     b: torch.Tensor,
@@ -213,7 +213,7 @@ def cutlass_scaled_fp4_mm(
     return out
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def cutlass_fp4_group_mm(
     a_fp4: torch.Tensor,
     b_fp4: torch.Tensor,
@@ -293,7 +293,7 @@ def _scaled_fp4_quant_custom_op(
     module.scaled_fp4_quant(output, input, output_scale, input_global_scale)
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def scaled_fp4_quant(
     input: torch.Tensor, input_global_scale: torch.Tensor
 ) -> Tuple[torch.Tensor, torch.Tensor]:
@@ -363,7 +363,7 @@ def _scaled_fp4_experts_quant_custom_op(
     )
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def scaled_fp4_experts_quant(
     input_tensor: torch.Tensor,
     input_global_scale: torch.Tensor,
@@ -448,7 +448,7 @@ def _scaled_fp4_grouped_quant_custom_op(
     )
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def scaled_fp4_grouped_quant(
     input_tensor: torch.Tensor,
     input_global_scale: torch.Tensor,
@@ -509,7 +509,7 @@ def _silu_and_mul_scaled_fp4_grouped_quant_custom_op(
     )
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def silu_and_mul_scaled_fp4_grouped_quant(
     input_tensor: torch.Tensor,
     input_global_scale: torch.Tensor,

python/sglang/jit_kernel/per_tensor_quant_fp8.py

@@ -4,7 +4,6 @@ from typing import TYPE_CHECKING
 
 import torch
 
-from sglang.jit_kernel.debug_utils import maybe_wrap_jit_kernel_debug
 from sglang.jit_kernel.utils import cache_once, load_jit, make_cpp_args
 from sglang.srt.utils.custom_op import register_custom_op
 
@@ -23,7 +22,6 @@ def _jit_per_tensor_quant_fp8_module(is_static: bool, dtype: torch.dtype) -> Mod
     )
 
 
-@maybe_wrap_jit_kernel_debug
 @register_custom_op(
     op_name="per_tensor_quant_fp8",
     mutates_args=["output_q", "output_s"],

python/sglang/jit_kernel/per_token_group_quant_8bit.py

@@ -4,8 +4,8 @@ from typing import TYPE_CHECKING
 
 import torch
 
-from sglang.jit_kernel.debug_utils import maybe_wrap_jit_kernel_debug
 from sglang.jit_kernel.utils import cache_once, load_jit, make_cpp_args
+from sglang.kernel_api_logging import debug_kernel_api
 from sglang.srt.utils.custom_op import register_custom_op
 
 if TYPE_CHECKING:
@@ -73,7 +73,7 @@ def _per_token_group_quant_8bit_custom_op(
     return None
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def per_token_group_quant_8bit(
     input: torch.Tensor,
     output_q: torch.Tensor,

python/sglang/jit_kernel/rope.py

@@ -5,7 +5,6 @@ from typing import TYPE_CHECKING, Optional
 
 import torch
 
-from sglang.jit_kernel.debug_utils import maybe_wrap_jit_kernel_debug
 from sglang.jit_kernel.utils import (
     cache_once,
     is_arch_support_pdl,
@@ -177,7 +176,6 @@ def apply_rope_inplace_with_kvcache(
 
 
 # NOTE: this name is intentionally set as the old kernel in `sgl_kernel`
-@maybe_wrap_jit_kernel_debug
 def apply_rope_with_cos_sin_cache_inplace(
     q: torch.Tensor,
     k: torch.Tensor,

python/sglang/jit_kernel/timestep_embedding.py

@@ -4,8 +4,8 @@ from typing import TYPE_CHECKING
 
 import torch
 
-from sglang.jit_kernel.debug_utils import maybe_wrap_jit_kernel_debug
 from sglang.jit_kernel.utils import cache_once, load_jit, make_cpp_args
+from sglang.kernel_api_logging import debug_kernel_api
 
 if TYPE_CHECKING:
     from tvm_ffi.module import Module
@@ -22,7 +22,7 @@ def _jit_timestep_embedding_module(dtype: torch.dtype) -> Module:
     )
 
 
-@maybe_wrap_jit_kernel_debug
+@debug_kernel_api
 def timestep_embedding(
     t: torch.Tensor,
     dim: int,

python/sglang/kernel_api_logging.py

@@ -15,38 +15,51 @@ import os
 import sys
 from datetime import datetime
 from pathlib import Path
-from typing import Any, Callable
+from typing import Any, Callable, TypeVar, overload
 
 import torch
 
+_logger = logging.getLogger("sglang.kernel_api")
 
-def _substitute_process_id(path: str) -> str:
+_T = TypeVar("_T")
+_F = TypeVar("_F", bound=Callable[..., Any])
+
+
+def _str_with_pid(path: str) -> str:
     if "%i" in path:
         return path.replace("%i", str(os.getpid()))
     return path
 
 
-_KERNEL_API_LOG_LEVEL = int(os.environ.get("SGLANG_KERNEL_API_LOGLEVEL", "0"))
+def _get_env(key: str, type: Callable[..., _T], default: _T) -> _T:
-_KERNEL_API_LOG_DEST = _substitute_process_id(
+    value_str = os.environ.get(key, None)
-    os.environ.get("SGLANG_KERNEL_API_LOGDEST", "stdout")
+    if value_str is None:
+        return default
+    try:
+        return type(value_str)
+    except Exception:
+        _logger.warning(
+            "Failed to parse environment variable %s=%r as %s, using default %r",
+            key,
+            value_str,
+            type.__name__,
+            default,
         )
+        return default
+
+
+def _parse_pattern(value: str) -> list[str]:
+    return [p.strip() for p in value.split(",") if p.strip()]
+
+
+_KERNEL_API_LOG_LEVEL = _get_env("SGLANG_KERNEL_API_LOGLEVEL", int, 0)
+_KERNEL_API_LOG_DEST = _get_env("SGLANG_KERNEL_API_LOGDEST", _str_with_pid, "stdout")
 _DUMP_DIR = Path(
-    _substitute_process_id(
+    _get_env("SGLANG_KERNEL_API_DUMP_DIR", _str_with_pid, "sglang_kernel_api_dumps")
-        os.environ.get("SGLANG_KERNEL_API_DUMP_DIR", "sglang_kernel_api_dumps")
 )
-)
+_DUMP_INCLUDE_PATTERNS = _get_env("SGLANG_KERNEL_API_DUMP_INCLUDE", _parse_pattern, [])
-_DUMP_INCLUDE_PATTERNS = [
+_DUMP_EXCLUDE_PATTERNS = _get_env("SGLANG_KERNEL_API_DUMP_EXCLUDE", _parse_pattern, [])
-    p.strip()
-    for p in os.environ.get("SGLANG_KERNEL_API_DUMP_INCLUDE", "").split(",")
-    if p.strip()
-]
-_DUMP_EXCLUDE_PATTERNS = [
-    p.strip()
-    for p in os.environ.get("SGLANG_KERNEL_API_DUMP_EXCLUDE", "").split(",")
-    if p.strip()
-]
 
-_logger = logging.getLogger("sglang.kernel_api")
 _dump_call_counter: dict[str, int] = {}
 
 
@@ -371,17 +384,36 @@ def _infer_func_name(func: Callable) -> str:
     return qualname
 
 
+@overload
+def debug_kernel_api(
+    func: _F,
+    *,
+    op_name: str | None = None,
+) -> _F: ...
+
+
+@overload
+def debug_kernel_api(
+    *,
+    op_name: str | None = None,
+) -> Callable[[_F], _F]: ...
+
+
 def debug_kernel_api(
     func: Callable | None = None,
     *,
     op_name: str | None = None,
 ) -> Callable:
+    # NOTE: avoid any overhead in the hot path when logging is disabled
     if _KERNEL_API_LOG_LEVEL == 0:
         if func is None:
             return lambda f: f
         return func
 
     def decorator(f: Callable) -> Callable:
+        if hasattr(f, "_debug_kernel_wrapped"):
+            return f
+
         @functools.wraps(f)
         def wrapper(*args: Any, **kwargs: Any) -> Any:
             if _is_compiling():
@@ -434,18 +466,29 @@ def debug_kernel_api(
                 _log_section("Output:", {"return": result})
             return result
 
+        setattr(wrapper, "_debug_kernel_wrapped", True)
         return wrapper
 
-    if func is None:
+    return decorator if func is None else decorator(func)
-        return decorator
-    return decorator(func)
 
 
-def debug_torch_op(op_name: str, *, namespace: str = "sglang") -> Callable:
+def debug_torch_op(
-    def call(*args: Any, **kwargs: Any) -> Any:
+    op_func: Callable,
-        return getattr(getattr(torch.ops, namespace), op_name)(*args, **kwargs)
+    op_name: str,
-
+    *,
-    return debug_kernel_api(call, op_name=f"{namespace}.custom_op.{op_name}")
+    namespace: str = "sglang",
+) -> Callable:
+    """NOTE: For internal use. Prefer `debug_kernel_api` for general use cases."""
+    # NOTE: avoid any overhead in the hot path when logging is disabled
+    impl = getattr(getattr(torch.ops, namespace), op_name)
+    if _KERNEL_API_LOG_LEVEL == 0:
+        return impl
+    # NOTE: propagate the marker to avoid double-wrapping
+    if hasattr(op_func, "_debug_kernel_wrapped"):
+        setattr(impl, "_debug_kernel_wrapped", True)
+        return impl
+    # NOTE: redirect the function name
+    return debug_kernel_api(impl, op_name=_infer_func_name(op_func))
 
 
 def wrap_method_with_debug_kernel_once(
@@ -455,6 +498,10 @@ def wrap_method_with_debug_kernel_once(
     op_name: str,
     marker_attr: str | None = None,
 ) -> Any:
+    # NOTE: avoid any overhead in the hot path when logging is disabled
+    if _KERNEL_API_LOG_LEVEL == 0:
+        return obj
+
     if marker_attr is None:
         marker_attr = f"_debug_kernel_{method_name}_wrapped"
 

python/sglang/multimodal_gen/runtime/layers/utils.py

@@ -156,7 +156,7 @@ class CustomOpWrapper:
                     mutates_args=self.mutates_args,
                     fake_impl=self.fake_impl,
                 )
-            self._impl = debug_torch_op(self.op_name)
+            self._impl = debug_torch_op(self.op_func, self.op_name)
             assert self._impl is not None
         return self._impl
 

python/sglang/srt/layers/moe/moe_runner/flashinfer_trtllm.py

@@ -8,13 +8,16 @@ from torch.nn import Module
 from torch.nn.parameter import Parameter
 
 # Import to register custom ops for torch.compile compatibility
-import sglang.srt.layers.moe.flashinfer_trtllm_moe  # noqa: F401
-from sglang.kernel_api_logging import debug_torch_op
 from sglang.srt.distributed import get_tp_group
 from sglang.srt.distributed.device_communicators.pynccl_allocator import (
     use_symmetric_memory,
 )
 from sglang.srt.layers.dp_attention import is_allocation_symmetric
+from sglang.srt.layers.moe.flashinfer_trtllm_moe import (
+    trtllm_fp8_block_scale_moe_wrapper,
+    trtllm_fp8_block_scale_routed_moe_wrapper,
+    trtllm_fp8_per_tensor_scale_moe_wrapper,
+)
 from sglang.srt.layers.moe.moe_runner.base import (
     MoeQuantInfo,
     MoeRunnerConfig,
@@ -45,16 +48,6 @@ elif is_cuda_alike():
 else:
     fp4_quantize = None
 
-_trtllm_fp8_block_scale_routed_moe_wrapper = debug_torch_op(
-    "trtllm_fp8_block_scale_routed_moe_wrapper"
-)
-_trtllm_fp8_block_scale_moe_wrapper = debug_torch_op(
-    "trtllm_fp8_block_scale_moe_wrapper"
-)
-_trtllm_fp8_per_tensor_scale_moe = debug_torch_op(
-    "trtllm_fp8_per_tensor_scale_moe_wrapper"
-)
-
 
 def align_fp8_moe_weights_for_flashinfer_trtllm(
     layer: Module, swap_w13_halves: bool = False
@@ -386,7 +379,7 @@ def fused_experts_none_to_flashinfer_trtllm_fp8(
                 topk_weights=topk_output.topk_weights,
             )
 
-            output = _trtllm_fp8_block_scale_routed_moe_wrapper(
+            output = trtllm_fp8_block_scale_routed_moe_wrapper(
                 topk_ids=packed_topk_ids,
                 routing_bias=None,
                 hidden_states=a_q,
@@ -419,7 +412,7 @@ def fused_experts_none_to_flashinfer_trtllm_fp8(
         else:
             assert TopKOutputChecker.format_is_bypassed(topk_output)
 
-            output = _trtllm_fp8_block_scale_moe_wrapper(
+            output = trtllm_fp8_block_scale_moe_wrapper(
                 routing_logits=(
                     router_logits.to(torch.float32)
                     if routing_method_type == RoutingMethodType.DeepSeekV3
@@ -476,7 +469,7 @@ def fused_experts_none_to_flashinfer_trtllm_fp8(
         # Move kernel call outside context manager to avoid graph breaks
         # during torch.compile for piecewise cuda graph.
         # Use custom op wrapper for torch.compile compatibility.
-        output = _trtllm_fp8_per_tensor_scale_moe(
+        output = trtllm_fp8_per_tensor_scale_moe_wrapper(
             routing_logits=router_logits.to(torch.bfloat16),
             routing_bias=routing_bias_cast,
             hidden_states=a_q,

python/sglang/srt/layers/quantization/bitsandbytes.py

@@ -7,7 +7,6 @@ from typing import TYPE_CHECKING, Any, Optional
 import torch
 from packaging import version
 
-from sglang.kernel_api_logging import debug_torch_op
 from sglang.srt.layers.linear import LinearBase
 from sglang.srt.layers.quantization.base_config import (
     FusedMoEMethodBase,
@@ -16,7 +15,8 @@ from sglang.srt.layers.quantization.base_config import (
     QuantizeMethodBase,
 )
 from sglang.srt.layers.quantization.unquant import UnquantizedLinearMethod
-from sglang.srt.utils import direct_register_custom_op, set_weight_attrs
+from sglang.srt.utils import set_weight_attrs
+from sglang.srt.utils.custom_op import register_custom_op
 
 if TYPE_CHECKING:
     from sglang.srt.layers.moe.token_dispatcher import (
@@ -393,7 +393,8 @@ class BitsAndBytesLinearMethod(LinearMethodBase):
         return out
 
 
-def _apply_bnb_4bit(
+@register_custom_op(mutates_args=["out"])
+def apply_bnb_4bit(
     x: torch.Tensor,
     weight: torch.Tensor,
     offsets: torch.Tensor,
@@ -416,28 +417,6 @@ def _apply_bnb_4bit(
         current_index += output_size
 
 
-def _apply_bnb_4bit_fake(
-    x: torch.Tensor,
-    weight: torch.Tensor,
-    offsets: torch.Tensor,
-    out: torch.Tensor,
-) -> None:
-    return
-
-
-try:
-    direct_register_custom_op(
-        op_name="apply_bnb_4bit",
-        op_func=_apply_bnb_4bit,
-        mutates_args=["out"],
-        fake_impl=_apply_bnb_4bit_fake,
-    )
-    apply_bnb_4bit = debug_torch_op("apply_bnb_4bit")
-
-except AttributeError as error:
-    raise error
-
-
 class BitsAndBytesMoEMethod(FusedMoEMethodBase):
     """MoE method for BitsAndBytes.
 

python/sglang/srt/layers/quantization/fp8.py

@@ -10,7 +10,6 @@ import torch.nn.functional as F
 from torch.nn import Module
 from torch.nn.parameter import Parameter
 
-from sglang.kernel_api_logging import debug_torch_op
 from sglang.srt.distributed import get_tensor_model_parallel_world_size, get_tp_group
 from sglang.srt.distributed.device_communicators.pynccl_allocator import (
     use_symmetric_memory,
@@ -58,7 +57,10 @@ from sglang.srt.layers.quantization.fp8_utils import (
     requant_weight_ue8m0_inplace,
 )
 from sglang.srt.layers.quantization.kv_cache import BaseKVCacheMethod
-from sglang.srt.layers.quantization.marlin_utils_fp8 import prepare_fp8_layer_for_marlin
+from sglang.srt.layers.quantization.marlin_utils_fp8 import (
+    apply_fp8_marlin_linear,
+    prepare_fp8_layer_for_marlin,
+)
 from sglang.srt.layers.quantization.unquant import (
     UnquantizedFusedMoEMethod,
     UnquantizedLinearMethod,
@@ -111,8 +113,6 @@ ACTIVATION_SCHEMES = ["static", "dynamic"]
 
 logger = logging.getLogger(__name__)
 
-_apply_fp8_marlin_linear = debug_torch_op("apply_fp8_marlin_linear")
-
 
 class Fp8Config(QuantizationConfig):
     """Config class for FP8."""
@@ -646,7 +646,7 @@ class Fp8LinearMethod(LinearMethodBase):
         bias: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
         if self.use_marlin:
-            return _apply_fp8_marlin_linear(
+            return apply_fp8_marlin_linear(
                 input=x,
                 weight=layer.weight,
                 weight_scale=layer.weight_scale,

python/sglang/srt/layers/quantization/marlin_utils_fp8.py

@@ -13,7 +13,8 @@ from sglang.srt.layers.quantization.marlin_utils import (
     should_use_atomic_add_reduce,
 )
 from sglang.srt.layers.quantization.utils import get_scalar_types
-from sglang.srt.utils import direct_register_custom_op, is_cuda
+from sglang.srt.utils import is_cuda
+from sglang.srt.utils.custom_op import register_custom_op
 
 _is_cuda = is_cuda()
 if _is_cuda:
@@ -39,6 +40,22 @@ def fp8_fused_exponent_bias_into_scales(scales):
     return scales * s
 
 
+def fake_apply_fp8_marlin_linear(
+    input: torch.Tensor,
+    weight: torch.Tensor,
+    weight_scale: torch.Tensor,
+    workspace: torch.Tensor,
+    size_n: int,
+    size_k: int,
+    bias: Optional[torch.Tensor],
+    use_fp32_reduce: bool = USE_FP32_REDUCE_DEFAULT,
+) -> torch.Tensor:
+    out_shape = input.shape[:-1] + (size_n,)
+    fake_output = torch.empty(out_shape, dtype=input.dtype, device=input.device)
+    return fake_output
+
+
+@register_custom_op(fake_impl=fake_apply_fp8_marlin_linear)
 def apply_fp8_marlin_linear(
     input: torch.Tensor,
     weight: torch.Tensor,
@@ -83,30 +100,6 @@ def apply_fp8_marlin_linear(
     return output.reshape(out_shape)
 
 
-def fake_apply_fp8_marlin_linear(
-    input: torch.Tensor,
-    weight: torch.Tensor,
-    weight_scale: torch.Tensor,
-    workspace: torch.Tensor,
-    size_n: int,
-    size_k: int,
-    bias: Optional[torch.Tensor],
-    use_fp32_reduce: bool = USE_FP32_REDUCE_DEFAULT,
-) -> torch.Tensor:
-
-    out_shape = input.shape[:-1] + (size_n,)
-    fake_output = torch.empty(out_shape, dtype=input.dtype, device=input.device)
-    return fake_output
-
-
-direct_register_custom_op(
-    op_name="apply_fp8_marlin_linear",
-    op_func=apply_fp8_marlin_linear,
-    mutates_args=[],
-    fake_impl=fake_apply_fp8_marlin_linear,
-)
-
-
 def prepare_fp8_layer_for_marlin(
     layer: torch.nn.Module, size_k_first: bool = True
 ) -> None:

python/sglang/srt/mem_cache/memory_pool.py

@@ -37,7 +37,6 @@ import triton
 import triton.language as tl
 
 from sglang.jit_kernel.kvcache import can_use_store_cache, store_cache
-from sglang.kernel_api_logging import debug_kernel_api
 from sglang.srt.configs.mamba_utils import BaseLinearStateParams
 from sglang.srt.constants import GPU_MEMORY_TYPE_KV_CACHE
 from sglang.srt.environ import envs
@@ -61,14 +60,8 @@ from sglang.srt.utils import (
     is_npu,
     next_power_of_2,
 )
-from sglang.srt.utils.custom_op import register_custom_op
 from sglang.srt.utils.torch_memory_saver_adapter import TorchMemorySaverAdapter
 
-store_cache = register_custom_op(
-    debug_kernel_api(store_cache, op_name="jit_kernel.kvcache.store_cache"),
-    mutates_args=["k_cache", "v_cache"],
-)
-
 if TYPE_CHECKING:
     from sglang.srt.managers.cache_controller import LayerDoneCounter
     from sglang.srt.managers.schedule_batch import Req

python/sglang/srt/utils/common.py

@@ -1899,8 +1899,11 @@ def direct_register_custom_op(
     mutates_args: List[str],
     fake_impl: Optional[Callable] = None,
     target_lib: Optional[Library] = None,
-):
+) -> None:
     """
+    NOTE: Please try to use `register_custom_op` instead of this function.
+    See `python/sglang/srt/utils/custom_op.py` for details.
+
     `torch.library.custom_op` can have significant overhead because it
     needs to consider complicated dispatching logic. This function
     directly registers a custom op and dispatches it to the CUDA backend.

python/sglang/srt/utils/custom_op.py

@@ -161,7 +161,7 @@ class CustomOpWrapper:
                     mutates_args=self.mutates_args,
                     fake_impl=self.fake_impl,
                 )
-            self._impl = debug_torch_op(self.op_name)
+            self._impl = debug_torch_op(self.op_func, self.op_name)
             assert self._impl is not None
         return self._impl
 
@@ -290,7 +290,7 @@ def register_custom_op_from_extern(
         wrapper.__name__ = fn.__name__
         wrapper.__qualname__ = fn.__qualname__
         wrapper.__module__ = fn.__module__
-        wrapper.__signature__ = new_sig
+        wrapper.__signature__ = new_sig  # type: ignore[attr-defined]
         # Build annotations without computed args, preserving return type
         wrapper.__annotations__ = {
             k: v
@@ -334,4 +334,4 @@ def register_custom_op_from_extern(
         fake_impl=fake_impl,
     )
 
-    return debug_torch_op(name)
+    return debug_torch_op(fn, name)