diff --git a/.claude/skills/add-sgl-kernel/SKILL.md b/.claude/skills/add-sgl-kernel/SKILL.md
index 4c5dc7a52..662df875b 100644
--- a/.claude/skills/add-sgl-kernel/SKILL.md
+++ b/.claude/skills/add-sgl-kernel/SKILL.md
@@ -244,18 +244,15 @@ Create `sgl-kernel/benchmark/bench_scale.py`:
 
 ```python
 import itertools
-import os
 
 import torch
 import triton
 import triton.testing
 
 import sgl_kernel
+from sglang.utils import is_in_ci
 
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
+IS_CI = is_in_ci()
 
 dtypes  = [torch.float16] if IS_CI else [torch.float16, torch.bfloat16, torch.float32]
 sizes   = [4096] if IS_CI else [2**n for n in range(10, 20)]  # 1K … 512K
diff --git a/benchmark/kernels/all_reduce/benchmark_torch_symm_mem.py b/benchmark/kernels/all_reduce/benchmark_torch_symm_mem.py
index 030fd5bb2..5bdb7f5d6 100644
--- a/benchmark/kernels/all_reduce/benchmark_torch_symm_mem.py
+++ b/benchmark/kernels/all_reduce/benchmark_torch_symm_mem.py
@@ -44,12 +44,9 @@ from sglang.srt.distributed.parallel_state import (
     initialize_model_parallel,
     set_torch_symm_mem_all_reduce,
 )
+from sglang.utils import is_in_ci
 
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
+IS_CI = is_in_ci()
 
 
 def torch_allreduce(torch_input: torch.Tensor, group: ProcessGroup) -> torch.Tensor:
diff --git a/python/sglang/jit_kernel/benchmark/bench_awq_dequantize.py b/python/sglang/jit_kernel/benchmark/bench_awq_dequantize.py
index a39027ea2..6a3fcbc93 100644
--- a/python/sglang/jit_kernel/benchmark/bench_awq_dequantize.py
+++ b/python/sglang/jit_kernel/benchmark/bench_awq_dequantize.py
@@ -5,7 +5,8 @@ import triton
 import triton.testing
 
 from sglang.jit_kernel.awq_dequantize import awq_dequantize as jit_awq_dequantize
-from sglang.jit_kernel.benchmark.utils import is_in_ci, run_benchmark
+from sglang.jit_kernel.benchmark.utils import run_benchmark
+from sglang.utils import is_in_ci
 
 try:
     from sgl_kernel import awq_dequantize as aot_awq_dequantize
diff --git a/python/sglang/jit_kernel/benchmark/bench_awq_marlin_moe_repack.py b/python/sglang/jit_kernel/benchmark/bench_awq_marlin_moe_repack.py
index 7ba7a1f2c..ac9faa517 100644
--- a/python/sglang/jit_kernel/benchmark/bench_awq_marlin_moe_repack.py
+++ b/python/sglang/jit_kernel/benchmark/bench_awq_marlin_moe_repack.py
@@ -7,8 +7,9 @@ from sgl_kernel.scalar_type import scalar_types
 from sglang.jit_kernel.awq_marlin_repack import (
     awq_marlin_moe_repack as jit_awq_marlin_moe_repack,
 )
-from sglang.jit_kernel.benchmark.utils import is_in_ci, run_benchmark
+from sglang.jit_kernel.benchmark.utils import run_benchmark
 from sglang.srt.layers.quantization.utils import pack_cols, quantize_weights
+from sglang.utils import is_in_ci
 
 AOT_AVAILABLE = hasattr(torch.ops.sgl_kernel, "awq_marlin_moe_repack") and hasattr(
     torch.ops.sgl_kernel.awq_marlin_moe_repack, "default"
diff --git a/python/sglang/jit_kernel/benchmark/bench_awq_marlin_repack.py b/python/sglang/jit_kernel/benchmark/bench_awq_marlin_repack.py
index e774e381c..8cf748fc0 100644
--- a/python/sglang/jit_kernel/benchmark/bench_awq_marlin_repack.py
+++ b/python/sglang/jit_kernel/benchmark/bench_awq_marlin_repack.py
@@ -7,8 +7,9 @@ from sgl_kernel.scalar_type import scalar_types
 from sglang.jit_kernel.awq_marlin_repack import (
     awq_marlin_repack as jit_awq_marlin_repack,
 )
-from sglang.jit_kernel.benchmark.utils import is_in_ci, run_benchmark
+from sglang.jit_kernel.benchmark.utils import run_benchmark
 from sglang.srt.layers.quantization.utils import pack_cols, quantize_weights
+from sglang.utils import is_in_ci
 
 AOT_AVAILABLE = hasattr(torch.ops.sgl_kernel, "awq_marlin_repack") and hasattr(
     torch.ops.sgl_kernel.awq_marlin_repack, "default"
diff --git a/python/sglang/jit_kernel/benchmark/bench_concat_mla.py b/python/sglang/jit_kernel/benchmark/bench_concat_mla.py
index 7e5c5c43a..7f2b542a0 100644
--- a/python/sglang/jit_kernel/benchmark/bench_concat_mla.py
+++ b/python/sglang/jit_kernel/benchmark/bench_concat_mla.py
@@ -6,9 +6,10 @@ import triton.testing
 from sgl_kernel import concat_mla_absorb_q as aot_absorb_q
 from sgl_kernel import concat_mla_k as aot_k
 
-from sglang.jit_kernel.benchmark.utils import is_in_ci, run_benchmark
+from sglang.jit_kernel.benchmark.utils import run_benchmark
 from sglang.jit_kernel.concat_mla import concat_mla_absorb_q as jit_absorb_q
 from sglang.jit_kernel.concat_mla import concat_mla_k as jit_k
+from sglang.utils import is_in_ci
 
 IS_CI = is_in_ci()
 
diff --git a/python/sglang/jit_kernel/benchmark/bench_fused_add_rmsnorm.py b/python/sglang/jit_kernel/benchmark/bench_fused_add_rmsnorm.py
index ddce16b92..0d7b1ec23 100644
--- a/python/sglang/jit_kernel/benchmark/bench_fused_add_rmsnorm.py
+++ b/python/sglang/jit_kernel/benchmark/bench_fused_add_rmsnorm.py
@@ -5,8 +5,9 @@ import triton
 import triton.testing
 from flashinfer import fused_add_rmsnorm as fi_fused_add_rmsnorm
 
-from sglang.jit_kernel.benchmark.utils import is_in_ci, run_benchmark
+from sglang.jit_kernel.benchmark.utils import run_benchmark
 from sglang.jit_kernel.norm import fused_add_rmsnorm as jit_fused_add_rmsnorm
+from sglang.utils import is_in_ci
 
 IS_CI = is_in_ci()
 
diff --git a/python/sglang/jit_kernel/benchmark/bench_fused_norm_scale_shift.py b/python/sglang/jit_kernel/benchmark/bench_fused_norm_scale_shift.py
index f6bf7f57f..24971231c 100644
--- a/python/sglang/jit_kernel/benchmark/bench_fused_norm_scale_shift.py
+++ b/python/sglang/jit_kernel/benchmark/bench_fused_norm_scale_shift.py
@@ -8,13 +8,14 @@ import torch
 import triton
 import triton.testing
 
-from sglang.jit_kernel.benchmark.utils import is_in_ci, run_benchmark_no_cudagraph
+from sglang.jit_kernel.benchmark.utils import run_benchmark_no_cudagraph
 from sglang.multimodal_gen.runtime.layers.layernorm import (
     LayerNormScaleShift,
     RMSNormScaleShift,
     ScaleResidualLayerNormScaleShift,
     ScaleResidualRMSNormScaleShift,
 )
+from sglang.utils import is_in_ci
 
 if is_in_ci():
     B_RANGE, S_RANGE, D_RANGE = [1], [128], [1024]
diff --git a/python/sglang/jit_kernel/benchmark/bench_gptq_marlin.py b/python/sglang/jit_kernel/benchmark/bench_gptq_marlin.py
index a6b4d0514..31881d785 100644
--- a/python/sglang/jit_kernel/benchmark/bench_gptq_marlin.py
+++ b/python/sglang/jit_kernel/benchmark/bench_gptq_marlin.py
@@ -3,10 +3,11 @@ import triton
 import triton.testing
 from sgl_kernel.scalar_type import scalar_types
 
-from sglang.jit_kernel.benchmark.utils import is_in_ci, run_benchmark
+from sglang.jit_kernel.benchmark.utils import run_benchmark
 from sglang.jit_kernel.gptq_marlin import gptq_marlin_gemm as jit_gptq_marlin_gemm
 from sglang.srt.layers.quantization.marlin_utils import marlin_make_workspace
 from sglang.test.test_marlin_utils import marlin_quantize
+from sglang.utils import is_in_ci
 
 AOT_AVAILABLE = hasattr(torch.ops.sgl_kernel, "gptq_marlin_gemm") and hasattr(
     torch.ops.sgl_kernel.gptq_marlin_gemm, "default"
diff --git a/python/sglang/jit_kernel/benchmark/bench_gptq_marlin_repack.py b/python/sglang/jit_kernel/benchmark/bench_gptq_marlin_repack.py
index 027575b6c..3fe86a0e6 100644
--- a/python/sglang/jit_kernel/benchmark/bench_gptq_marlin_repack.py
+++ b/python/sglang/jit_kernel/benchmark/bench_gptq_marlin_repack.py
@@ -3,9 +3,10 @@ import triton
 import triton.testing
 from sgl_kernel.scalar_type import scalar_types
 
-from sglang.jit_kernel.benchmark.utils import is_in_ci, run_benchmark
+from sglang.jit_kernel.benchmark.utils import run_benchmark
 from sglang.jit_kernel.gptq_marlin_repack import gptq_marlin_repack as jit_fn
 from sglang.srt.layers.quantization.utils import gptq_quantize_weights, pack_rows
+from sglang.utils import is_in_ci
 
 AOT_AVAILABLE = hasattr(torch.ops.sgl_kernel, "gptq_marlin_repack") and hasattr(
     torch.ops.sgl_kernel.gptq_marlin_repack, "default"
diff --git a/python/sglang/jit_kernel/benchmark/bench_moe_wna16_marlin.py b/python/sglang/jit_kernel/benchmark/bench_moe_wna16_marlin.py
index 063ae908e..fa6e9b36d 100644
--- a/python/sglang/jit_kernel/benchmark/bench_moe_wna16_marlin.py
+++ b/python/sglang/jit_kernel/benchmark/bench_moe_wna16_marlin.py
@@ -3,10 +3,11 @@ import triton
 import triton.testing
 from sgl_kernel.scalar_type import scalar_types
 
-from sglang.jit_kernel.benchmark.utils import is_in_ci, run_benchmark
+from sglang.jit_kernel.benchmark.utils import run_benchmark
 from sglang.jit_kernel.moe_wna16_marlin import moe_wna16_marlin_gemm as jit_fn
 from sglang.srt.layers.moe.fused_moe_triton import moe_align_block_size
 from sglang.test.test_marlin_utils import marlin_quantize
+from sglang.utils import is_in_ci
 
 AOT_AVAILABLE = hasattr(torch.ops.sgl_kernel, "moe_wna16_marlin_gemm") and hasattr(
     torch.ops.sgl_kernel.moe_wna16_marlin_gemm, "default"
diff --git a/python/sglang/jit_kernel/benchmark/bench_norm.py b/python/sglang/jit_kernel/benchmark/bench_norm.py
index 2bccd672d..e58d9f5f9 100644
--- a/python/sglang/jit_kernel/benchmark/bench_norm.py
+++ b/python/sglang/jit_kernel/benchmark/bench_norm.py
@@ -6,9 +6,10 @@ import triton.testing
 from flashinfer.norm import fused_add_rmsnorm as fi_fused_add_rmsnorm
 from flashinfer.norm import rmsnorm as fi_rmsnorm
 
-from sglang.jit_kernel.benchmark.utils import is_in_ci, run_benchmark
+from sglang.jit_kernel.benchmark.utils import run_benchmark
 from sglang.jit_kernel.norm import fused_add_rmsnorm as jit_fused_add_rmsnorm
 from sglang.jit_kernel.norm import rmsnorm as jit_rmsnorm
+from sglang.utils import is_in_ci
 
 IS_CI = is_in_ci()
 
diff --git a/python/sglang/jit_kernel/benchmark/bench_norm_impls.py b/python/sglang/jit_kernel/benchmark/bench_norm_impls.py
index 5f3f74cd9..94b029597 100644
--- a/python/sglang/jit_kernel/benchmark/bench_norm_impls.py
+++ b/python/sglang/jit_kernel/benchmark/bench_norm_impls.py
@@ -15,12 +15,13 @@ from typing import Callable
 import torch
 import torch.nn.functional as F
 
-from sglang.jit_kernel.benchmark.utils import DEFAULT_DEVICE, is_in_ci
+from sglang.jit_kernel.benchmark.utils import DEFAULT_DEVICE
 from sglang.jit_kernel.diffusion.triton.norm import norm_infer, rms_norm_fn
 from sglang.jit_kernel.diffusion.triton.rmsnorm_onepass import triton_one_pass_rms_norm
 from sglang.jit_kernel.norm import fused_add_rmsnorm as jit_fused_add_rmsnorm
 from sglang.jit_kernel.norm import rmsnorm as jit_rmsnorm
 from sglang.jit_kernel.utils import KERNEL_PATH
+from sglang.utils import is_in_ci
 
 os.environ.setdefault("FLASHINFER_DISABLE_VERSION_CHECK", "1")
 
diff --git a/python/sglang/jit_kernel/benchmark/bench_per_token_group_quant_8bit.py b/python/sglang/jit_kernel/benchmark/bench_per_token_group_quant_8bit.py
index f0aadf145..b4515df0a 100644
--- a/python/sglang/jit_kernel/benchmark/bench_per_token_group_quant_8bit.py
+++ b/python/sglang/jit_kernel/benchmark/bench_per_token_group_quant_8bit.py
@@ -5,10 +5,7 @@ import torch
 import triton
 from sgl_kernel.test_utils import create_per_token_group_quant_test_data
 
-from sglang.jit_kernel.benchmark.utils import (
-    get_benchmark_range,
-    is_in_ci,
-)
+from sglang.jit_kernel.benchmark.utils import get_benchmark_range
 from sglang.jit_kernel.per_token_group_quant_8bit import (
     per_token_group_quant_8bit as sglang_per_token_group_quant_8bit,
 )
@@ -20,6 +17,7 @@ from sglang.srt.layers.quantization.fp8_kernel import (
 )
 from sglang.srt.utils import is_hip
 from sglang.srt.utils.bench_utils import bench_kineto
+from sglang.utils import is_in_ci
 
 IS_CI = is_in_ci()
 
diff --git a/python/sglang/jit_kernel/benchmark/bench_qknorm_across_heads.py b/python/sglang/jit_kernel/benchmark/bench_qknorm_across_heads.py
index ab660f489..cbcff1031 100644
--- a/python/sglang/jit_kernel/benchmark/bench_qknorm_across_heads.py
+++ b/python/sglang/jit_kernel/benchmark/bench_qknorm_across_heads.py
@@ -6,9 +6,10 @@ import triton
 import triton.testing
 from sgl_kernel import rmsnorm
 
-from sglang.jit_kernel.benchmark.utils import is_in_ci, run_benchmark
+from sglang.jit_kernel.benchmark.utils import run_benchmark
 from sglang.jit_kernel.norm import fused_inplace_qknorm_across_heads
 from sglang.srt.utils import get_current_device_stream_fast
+from sglang.utils import is_in_ci
 
 IS_CI = is_in_ci()
 
diff --git a/python/sglang/jit_kernel/benchmark/bench_qwen_image_modulation.py b/python/sglang/jit_kernel/benchmark/bench_qwen_image_modulation.py
index d16f9eda7..e0bc7dee0 100644
--- a/python/sglang/jit_kernel/benchmark/bench_qwen_image_modulation.py
+++ b/python/sglang/jit_kernel/benchmark/bench_qwen_image_modulation.py
@@ -3,13 +3,14 @@ from typing import Tuple
 import torch
 import triton.testing
 
-from sglang.jit_kernel.benchmark.utils import is_in_ci, run_benchmark_no_cudagraph
+from sglang.jit_kernel.benchmark.utils import run_benchmark_no_cudagraph
 from sglang.jit_kernel.diffusion.triton.norm import norm_infer
 from sglang.jit_kernel.diffusion.triton.scale_shift import (
     fuse_layernorm_scale_shift_gate_select01_kernel,
     fuse_residual_layernorm_scale_shift_gate_select01_kernel,
     fuse_scale_shift_gate_select01_kernel,
 )
+from sglang.utils import is_in_ci
 
 if is_in_ci():
     B_RANGE, S_RANGE, D_RANGE = [1], [128], [3072]
diff --git a/python/sglang/jit_kernel/benchmark/bench_renorm.py b/python/sglang/jit_kernel/benchmark/bench_renorm.py
index 9c654ab7f..20cc9ae71 100644
--- a/python/sglang/jit_kernel/benchmark/bench_renorm.py
+++ b/python/sglang/jit_kernel/benchmark/bench_renorm.py
@@ -5,7 +5,8 @@ import torch
 import triton
 import triton.testing
 
-from sglang.jit_kernel.benchmark.utils import is_in_ci, run_benchmark_no_cudagraph
+from sglang.jit_kernel.benchmark.utils import run_benchmark_no_cudagraph
+from sglang.utils import is_in_ci
 
 
 def torch_top_k_renorm_probs(probs, top_k):
diff --git a/python/sglang/jit_kernel/benchmark/utils.py b/python/sglang/jit_kernel/benchmark/utils.py
index 079a4489b..c17ef4f9a 100644
--- a/python/sglang/jit_kernel/benchmark/utils.py
+++ b/python/sglang/jit_kernel/benchmark/utils.py
@@ -5,7 +5,7 @@ from typing import Callable, List, Tuple
 import torch
 import triton.testing
 
-from sglang.jit_kernel.utils import is_in_ci as jit_kernel_is_in_ci
+from sglang.utils import is_in_ci
 
 # Common constants
 DEFAULT_DTYPE = torch.bfloat16
@@ -13,11 +13,6 @@ DEFAULT_DEVICE = "cuda"
 DEFAULT_QUANTILES = [0.5, 0.2, 0.8]
 
 
-def is_in_ci() -> bool:
-    """Check if running in CI environment."""
-    return jit_kernel_is_in_ci()
-
-
 def get_benchmark_range(full_range: List, ci_range: List) -> List:
     """Return appropriate benchmark range based on CI environment."""
     return ci_range if is_in_ci() else full_range
diff --git a/python/sglang/jit_kernel/utils.py b/python/sglang/jit_kernel/utils.py
index eb4748d71..fdcbc8e8a 100644
--- a/python/sglang/jit_kernel/utils.py
+++ b/python/sglang/jit_kernel/utils.py
@@ -7,6 +7,8 @@ from typing import TYPE_CHECKING, Any, Callable, List, Tuple, TypeAlias, TypeVar
 
 import torch
 
+from sglang.utils import is_in_ci
+
 if TYPE_CHECKING:
     from tvm_ffi import Module
 
@@ -14,11 +16,6 @@ F = TypeVar("F", bound=Callable[..., Any])
 _FULL_TEST_ENV_VAR = "SGLANG_JIT_KERNEL_RUN_FULL_TESTS"
 
 
-def is_in_ci() -> bool:
-    ci_env_vars = ("SGLANG_IS_IN_CI", "CI", "GITHUB_ACTIONS")
-    return any(os.getenv(env_var, "false").lower() == "true" for env_var in ci_env_vars)
-
-
 def should_run_full_tests() -> bool:
     return os.getenv(_FULL_TEST_ENV_VAR, "false").lower() == "true"
 
diff --git a/sgl-kernel/benchmark/bench_activation.py b/sgl-kernel/benchmark/bench_activation.py
index 3caa5b936..43ec8d14d 100644
--- a/sgl-kernel/benchmark/bench_activation.py
+++ b/sgl-kernel/benchmark/bench_activation.py
@@ -13,6 +13,8 @@ import triton
 import triton.testing
 from sgl_kernel import gelu_and_mul, gelu_tanh_and_mul, silu_and_mul
 
+from sglang.utils import is_in_ci
+
 # Optional vLLM import
 try:
     from vllm import _custom_ops as vllm_ops
@@ -22,11 +24,7 @@ except ImportError:
     vllm_ops = None
     VLLM_AVAILABLE = False
 
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
+IS_CI = is_in_ci()
 
 # gelu_quick is only available on HIP/ROCm platforms
 try:
diff --git a/sgl-kernel/benchmark/bench_awq_dequant.py b/sgl-kernel/benchmark/bench_awq_dequant.py
index 6bd03ab8a..cb22ba07f 100644
--- a/sgl-kernel/benchmark/bench_awq_dequant.py
+++ b/sgl-kernel/benchmark/bench_awq_dequant.py
@@ -7,6 +7,8 @@ import triton
 import triton.testing
 from sgl_kernel import awq_dequantize
 
+from sglang.utils import is_in_ci
+
 # Optional vLLM import
 try:
     from vllm import _custom_ops as ops
@@ -16,11 +18,7 @@ except ImportError:
     ops = None
     VLLM_AVAILABLE = False
 
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
+IS_CI = is_in_ci()
 
 
 def vllm_awq_dequantize(
diff --git a/sgl-kernel/benchmark/bench_cutlass_mla.py b/sgl-kernel/benchmark/bench_cutlass_mla.py
index 6947f309d..461d8862b 100644
--- a/sgl-kernel/benchmark/bench_cutlass_mla.py
+++ b/sgl-kernel/benchmark/bench_cutlass_mla.py
@@ -8,12 +8,9 @@ import triton
 from sgl_kernel import cutlass_mla_decode, cutlass_mla_get_workspace_size
 
 from sglang.srt.utils import get_device_capability
+from sglang.utils import is_in_ci
 
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
+IS_CI = is_in_ci()
 
 # CI environment uses simplified parameters
 if IS_CI:
diff --git a/sgl-kernel/benchmark/bench_dsv3_fused_a_gemm.py b/sgl-kernel/benchmark/bench_dsv3_fused_a_gemm.py
index bdf7f85de..43f0961f7 100644
--- a/sgl-kernel/benchmark/bench_dsv3_fused_a_gemm.py
+++ b/sgl-kernel/benchmark/bench_dsv3_fused_a_gemm.py
@@ -1,11 +1,4 @@
 import argparse
-import os
-
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
 
 import torch
 import torch.nn.functional as F
@@ -13,6 +6,10 @@ import triton
 import triton.testing
 from sgl_kernel import dsv3_fused_a_gemm
 
+from sglang.utils import is_in_ci
+
+IS_CI = is_in_ci()
+
 # CI environment uses simplified parameters
 if IS_CI:
     num_tokens_vals = [1]  # Only test 1 value in CI
diff --git a/sgl-kernel/benchmark/bench_dsv3_router_gemm.py b/sgl-kernel/benchmark/bench_dsv3_router_gemm.py
index 2daee279f..85ad88529 100644
--- a/sgl-kernel/benchmark/bench_dsv3_router_gemm.py
+++ b/sgl-kernel/benchmark/bench_dsv3_router_gemm.py
@@ -1,11 +1,4 @@
 import argparse
-import os
-
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
 
 import torch
 import torch.nn.functional as F
@@ -13,6 +6,10 @@ import triton
 import triton.testing
 from sgl_kernel import dsv3_router_gemm
 
+from sglang.utils import is_in_ci
+
+IS_CI = is_in_ci()
+
 # CI environment uses simplified parameters
 if IS_CI:
     num_tokens_vals = [1]  # Only test 1 value in CI
diff --git a/sgl-kernel/benchmark/bench_fp4_gemm.py b/sgl-kernel/benchmark/bench_fp4_gemm.py
index 278b965c2..d5d358611 100755
--- a/sgl-kernel/benchmark/bench_fp4_gemm.py
+++ b/sgl-kernel/benchmark/bench_fp4_gemm.py
@@ -10,12 +10,9 @@ from flashinfer.testing import bench_gpu_time_with_cupti
 from sgl_kernel import cutlass_scaled_fp4_mm, scaled_fp4_quant
 
 from sglang.srt.utils import get_device_capability, is_sm100_supported
+from sglang.utils import is_in_ci
 
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
+IS_CI = is_in_ci()
 
 FLOAT4_E2M1_MAX = 6.0
 FLOAT8_E4M3_MAX = torch.finfo(torch.float8_e4m3fn).max
diff --git a/sgl-kernel/benchmark/bench_fp8_blockwise_gemm.py b/sgl-kernel/benchmark/bench_fp8_blockwise_gemm.py
index 70766df94..29bd0a80a 100644
--- a/sgl-kernel/benchmark/bench_fp8_blockwise_gemm.py
+++ b/sgl-kernel/benchmark/bench_fp8_blockwise_gemm.py
@@ -9,6 +9,8 @@ import triton
 from deep_gemm.utils.layout import get_mn_major_tma_aligned_tensor
 from sgl_kernel import fp8_blockwise_scaled_mm
 
+from sglang.utils import is_in_ci
+
 # Optional vLLM import
 try:
     from vllm._custom_ops import cutlass_scaled_mm as vllm_scaled_mm
@@ -22,11 +24,7 @@ from sglang.srt.layers.quantization.fp8_kernel import (
     w8a8_block_fp8_matmul_triton as w8a8_block_fp8_matmul,
 )
 
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
+IS_CI = is_in_ci()
 
 
 def get_weight_shapes(args):
diff --git a/sgl-kernel/benchmark/bench_fp8_blockwise_group_gemm.py b/sgl-kernel/benchmark/bench_fp8_blockwise_group_gemm.py
index 19e425b52..063aca527 100644
--- a/sgl-kernel/benchmark/bench_fp8_blockwise_group_gemm.py
+++ b/sgl-kernel/benchmark/bench_fp8_blockwise_group_gemm.py
@@ -1,11 +1,4 @@
 import argparse
-import os
-
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
 import random
 from dataclasses import dataclass
 from typing import List, Tuple
@@ -14,6 +7,10 @@ import deep_gemm
 import torch
 from sgl_kernel import fp8_blockwise_scaled_grouped_mm
 
+from sglang.utils import is_in_ci
+
+IS_CI = is_in_ci()
+
 
 def get_m_alignment_for_contiguous_layout():
     return 128
diff --git a/sgl-kernel/benchmark/bench_fp8_gemm.py b/sgl-kernel/benchmark/bench_fp8_gemm.py
index 1dbedb66a..e37653d51 100644
--- a/sgl-kernel/benchmark/bench_fp8_gemm.py
+++ b/sgl-kernel/benchmark/bench_fp8_gemm.py
@@ -9,6 +9,7 @@ import triton
 from sgl_kernel import fp8_scaled_mm as sgl_scaled_mm
 
 from sglang.jit_kernel.per_tensor_quant_fp8 import per_tensor_quant_fp8
+from sglang.utils import is_in_ci
 
 # Optional vLLM import
 try:
@@ -21,11 +22,7 @@ except ImportError:
     vllm_scaled_fp8_quant = None
     VLLM_AVAILABLE = False
 
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
+IS_CI = is_in_ci()
 
 # Weight Shapes are in the format
 # ([K, N], TP_SPLIT_DIM)
diff --git a/sgl-kernel/benchmark/bench_int8_gemm.py b/sgl-kernel/benchmark/bench_int8_gemm.py
index 95f0f3bb8..722d89d7c 100644
--- a/sgl-kernel/benchmark/bench_int8_gemm.py
+++ b/sgl-kernel/benchmark/bench_int8_gemm.py
@@ -7,6 +7,8 @@ import torch
 import triton
 from sgl_kernel import int8_scaled_mm
 
+from sglang.utils import is_in_ci
+
 # Optional vLLM import
 try:
     from vllm._custom_ops import cutlass_scaled_mm as vllm_scaled_mm
@@ -16,11 +18,7 @@ except ImportError:
     vllm_scaled_mm = None
     VLLM_AVAILABLE = False
 
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
+IS_CI = is_in_ci()
 
 
 def to_int8(tensor: torch.Tensor) -> torch.Tensor:
diff --git a/sgl-kernel/benchmark/bench_kimi_k2_moe_fused_gate.py b/sgl-kernel/benchmark/bench_kimi_k2_moe_fused_gate.py
index 78b75231c..975a67a50 100644
--- a/sgl-kernel/benchmark/bench_kimi_k2_moe_fused_gate.py
+++ b/sgl-kernel/benchmark/bench_kimi_k2_moe_fused_gate.py
@@ -8,12 +8,9 @@ import triton.language as tl
 from sgl_kernel import kimi_k2_moe_fused_gate
 
 from sglang.srt.layers.moe.topk import kimi_k2_biased_topk_impl
+from sglang.utils import is_in_ci
 
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
+IS_CI = is_in_ci()
 
 
 def kimi_k2_biased_topk_torch_compile(scores, bias, topk, routed_scaling_factor):
diff --git a/sgl-kernel/benchmark/bench_moe_align_block_size.py b/sgl-kernel/benchmark/bench_moe_align_block_size.py
index 5dc4a0e4d..3debde2f4 100644
--- a/sgl-kernel/benchmark/bench_moe_align_block_size.py
+++ b/sgl-kernel/benchmark/bench_moe_align_block_size.py
@@ -7,6 +7,8 @@ import triton
 import triton.language as tl
 from sgl_kernel import moe_align_block_size as sgl_moe_align_block_size
 
+from sglang.utils import is_in_ci
+
 try:
     from vllm import _custom_ops as ops
 
@@ -15,11 +17,7 @@ except ImportError:
     ops = None
     VLLM_AVAILABLE = False
 
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
+IS_CI = is_in_ci()
 
 USE_RANDOM_PERM = False
 
diff --git a/sgl-kernel/benchmark/bench_moe_ep_post_reorder.py b/sgl-kernel/benchmark/bench_moe_ep_post_reorder.py
index 2a617d72d..c3d9415a4 100644
--- a/sgl-kernel/benchmark/bench_moe_ep_post_reorder.py
+++ b/sgl-kernel/benchmark/bench_moe_ep_post_reorder.py
@@ -1,15 +1,10 @@
-import os
-
 import torch
-
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
 import triton
 
 from sglang.srt.layers.moe.ep_moe.kernels import post_reorder_triton_kernel
+from sglang.utils import is_in_ci
+
+IS_CI = is_in_ci()
 
 # CI environment uses simplified parameters
 if IS_CI:
diff --git a/sgl-kernel/benchmark/bench_moe_fused_gate.py b/sgl-kernel/benchmark/bench_moe_fused_gate.py
index cb5ac1760..508dcf1a9 100644
--- a/sgl-kernel/benchmark/bench_moe_fused_gate.py
+++ b/sgl-kernel/benchmark/bench_moe_fused_gate.py
@@ -8,12 +8,9 @@ import triton.language as tl
 from sgl_kernel import moe_fused_gate
 
 from sglang.srt.layers.moe.topk import biased_grouped_topk
+from sglang.utils import is_in_ci
 
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
+IS_CI = is_in_ci()
 
 
 def biased_grouped_topk_org(scores, bias, num_expert_group, topk_group, topk):
diff --git a/sgl-kernel/benchmark/bench_moe_topk_sigmoid.py b/sgl-kernel/benchmark/bench_moe_topk_sigmoid.py
index d34e68b98..cbc80a607 100644
--- a/sgl-kernel/benchmark/bench_moe_topk_sigmoid.py
+++ b/sgl-kernel/benchmark/bench_moe_topk_sigmoid.py
@@ -6,11 +6,9 @@ import torch
 import triton
 from sgl_kernel import topk_sigmoid
 
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
+from sglang.utils import is_in_ci
+
+IS_CI = is_in_ci()
 
 
 def torch_topk_sigmoid_native(
diff --git a/sgl-kernel/benchmark/bench_moe_topk_softmax.py b/sgl-kernel/benchmark/bench_moe_topk_softmax.py
index e065981b8..4b2222405 100644
--- a/sgl-kernel/benchmark/bench_moe_topk_softmax.py
+++ b/sgl-kernel/benchmark/bench_moe_topk_softmax.py
@@ -6,6 +6,8 @@ import torch
 import triton
 from sgl_kernel import topk_softmax
 
+from sglang.utils import is_in_ci
+
 # Optional vLLM import
 try:
     from vllm import _custom_ops as vllm_custom_ops
@@ -15,11 +17,7 @@ except ImportError:
     vllm_custom_ops = None
     VLLM_AVAILABLE = False
 
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
+IS_CI = is_in_ci()
 
 
 def vllm_topk_softmax(gating_output, topk):
diff --git a/sgl-kernel/benchmark/bench_per_tensor_quant_fp8.py b/sgl-kernel/benchmark/bench_per_tensor_quant_fp8.py
index 43c2caddc..107123465 100644
--- a/sgl-kernel/benchmark/bench_per_tensor_quant_fp8.py
+++ b/sgl-kernel/benchmark/bench_per_tensor_quant_fp8.py
@@ -9,6 +9,7 @@ import triton
 import triton.testing
 
 from sglang.jit_kernel.per_tensor_quant_fp8 import per_tensor_quant_fp8
+from sglang.utils import is_in_ci
 
 # Optional imports
 try:
@@ -23,11 +24,7 @@ from sglang.srt.utils import is_hip
 
 _is_hip = is_hip()
 
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
+IS_CI = is_in_ci()
 
 fp8_type_ = torch.float8_e4m3fnuz if _is_hip else torch.float8_e4m3fn
 
diff --git a/sgl-kernel/benchmark/bench_per_token_group_quant_8bit.py b/sgl-kernel/benchmark/bench_per_token_group_quant_8bit.py
index 5ad6f9fc2..e4a8e7bbd 100644
--- a/sgl-kernel/benchmark/bench_per_token_group_quant_8bit.py
+++ b/sgl-kernel/benchmark/bench_per_token_group_quant_8bit.py
@@ -19,12 +19,9 @@ from sglang.srt.layers.quantization.fp8_kernel import (
 )
 from sglang.srt.utils import is_hip
 from sglang.srt.utils.bench_utils import bench_kineto
+from sglang.utils import is_in_ci
 
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
+IS_CI = is_in_ci()
 
 _is_hip = is_hip()
 fp8_type_ = torch.float8_e4m3fnuz if _is_hip else torch.float8_e4m3fn
diff --git a/sgl-kernel/benchmark/bench_per_token_quant_fp8.py b/sgl-kernel/benchmark/bench_per_token_quant_fp8.py
index 8db1869d1..dca014676 100644
--- a/sgl-kernel/benchmark/bench_per_token_quant_fp8.py
+++ b/sgl-kernel/benchmark/bench_per_token_quant_fp8.py
@@ -7,6 +7,8 @@ import triton
 import triton.testing
 from sgl_kernel import sgl_per_token_quant_fp8
 
+from sglang.utils import is_in_ci
+
 # Optional vLLM import
 try:
     from vllm import _custom_ops as ops
@@ -20,11 +22,7 @@ from sglang.srt.utils import is_hip
 
 _is_hip = is_hip()
 
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
+IS_CI = is_in_ci()
 
 fp8_type_ = torch.float8_e4m3fnuz if _is_hip else torch.float8_e4m3fn
 
diff --git a/sgl-kernel/benchmark/bench_qserve_w4a8_gemm.py b/sgl-kernel/benchmark/bench_qserve_w4a8_gemm.py
index 5827fa993..acef80f45 100644
--- a/sgl-kernel/benchmark/bench_qserve_w4a8_gemm.py
+++ b/sgl-kernel/benchmark/bench_qserve_w4a8_gemm.py
@@ -11,11 +11,9 @@ from sgl_kernel import (
     qserve_w4a8_per_group_gemm,
 )
 
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
+from sglang.utils import is_in_ci
+
+IS_CI = is_in_ci()
 
 
 def to_int8(tensor: torch.Tensor) -> torch.Tensor:
diff --git a/sgl-kernel/benchmark/bench_rmsnorm.py b/sgl-kernel/benchmark/bench_rmsnorm.py
index d521ab05f..953b79792 100644
--- a/sgl-kernel/benchmark/bench_rmsnorm.py
+++ b/sgl-kernel/benchmark/bench_rmsnorm.py
@@ -13,6 +13,8 @@ import triton
 import triton.testing
 from sgl_kernel.utils import is_arch_support_pdl
 
+from sglang.utils import is_in_ci
+
 # Optional imports
 try:
     from flashinfer.norm import fused_add_rmsnorm, rmsnorm
@@ -31,11 +33,7 @@ except ImportError:
     vllm_ops = None
     VLLM_AVAILABLE = False
 
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
+IS_CI = is_in_ci()
 
 
 def str2int_list(arg: str) -> List[int]:
diff --git a/sgl-kernel/benchmark/bench_rotary_embedding.py b/sgl-kernel/benchmark/bench_rotary_embedding.py
index 30614355f..bd3a900c7 100644
--- a/sgl-kernel/benchmark/bench_rotary_embedding.py
+++ b/sgl-kernel/benchmark/bench_rotary_embedding.py
@@ -12,12 +12,9 @@ from sgl_kernel.testing.rotary_embedding import (
 )
 
 from sglang.srt.utils.bench_utils import bench_kineto
+from sglang.utils import is_in_ci
 
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
+IS_CI = is_in_ci()
 
 # CI environment uses simplified parameters
 if IS_CI:
diff --git a/sgl-kernel/benchmark/bench_sum_scale.py b/sgl-kernel/benchmark/bench_sum_scale.py
index ad9621ee1..1c25ac151 100644
--- a/sgl-kernel/benchmark/bench_sum_scale.py
+++ b/sgl-kernel/benchmark/bench_sum_scale.py
@@ -6,11 +6,9 @@ import triton.language as tl
 from sgl_kernel import moe_sum_reduce as moe_sum_reduce_cuda
 from triton.testing import do_bench
 
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
+from sglang.utils import is_in_ci
+
+IS_CI = is_in_ci()
 
 
 @triton.jit
diff --git a/sgl-kernel/benchmark/bench_top_k_top_p_sampling.py b/sgl-kernel/benchmark/bench_top_k_top_p_sampling.py
index d84a34884..78a950362 100644
--- a/sgl-kernel/benchmark/bench_top_k_top_p_sampling.py
+++ b/sgl-kernel/benchmark/bench_top_k_top_p_sampling.py
@@ -7,11 +7,9 @@ import torch
 import triton
 import triton.testing
 
-# CI environment detection
-IS_CI = (
-    os.getenv("CI", "false").lower() == "true"
-    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
-)
+from sglang.utils import is_in_ci
+
+IS_CI = is_in_ci()
 
 
 def torch_top_k_top_p_joint_sampling_from_probs(
diff --git a/test/registered/test_srt_backend.py b/test/registered/language/test_srt_backend.py
similarity index 100%
rename from test/registered/test_srt_backend.py
rename to test/registered/language/test_srt_backend.py
diff --git a/test/registered/test_hybrid_dp_ep_tp_mtp.py b/test/registered/moe/test_hybrid_dp_ep_tp_mtp.py
similarity index 100%
rename from test/registered/test_hybrid_dp_ep_tp_mtp.py
rename to test/registered/moe/test_hybrid_dp_ep_tp_mtp.py