Refactor JIT kernel CI to use run_suite.py registration system (#21239)

python/sglang/jit_kernel/benchmark/bench_awq_dequantize.py

@@ -6,8 +6,11 @@ import triton.testing
 
 from sglang.jit_kernel.awq_dequantize import awq_dequantize as jit_awq_dequantize
 from sglang.jit_kernel.benchmark.utils import run_benchmark
+from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.utils import is_in_ci
 
+register_cuda_ci(est_time=5, suite="stage-b-kernel-benchmark-1-gpu-large")
+
 try:
     from sgl_kernel import awq_dequantize as aot_awq_dequantize
 

python/sglang/jit_kernel/benchmark/bench_clamp_position.py

@@ -11,6 +11,9 @@ from sglang.jit_kernel.benchmark.utils import (
 )
 from sglang.jit_kernel.clamp_position import clamp_position_cuda
 from sglang.srt.utils import get_compiler_backend
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=13, suite="stage-b-kernel-benchmark-1-gpu-large")
 
 SIZE_LIST = get_benchmark_range(
     full_range=[2**n for n in range(4, 16)],

python/sglang/jit_kernel/benchmark/bench_concat_mla.py

@@ -9,8 +9,11 @@ from sgl_kernel import concat_mla_k as aot_k
 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.test.ci.ci_register import register_cuda_ci
 from sglang.utils import is_in_ci
 
+register_cuda_ci(est_time=6, suite="stage-b-kernel-benchmark-1-gpu-large")
+
 IS_CI = is_in_ci()
 
 NUM_LOCAL_HEADS = 128

python/sglang/jit_kernel/benchmark/bench_custom_all_reduce.py

@@ -22,6 +22,13 @@ import torch
 import torch.distributed as dist
 
 from sglang.jit_kernel.benchmark.utils import is_in_ci
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(
+    est_time=120,
+    suite="stage-b-kernel-benchmark-1-gpu-large",
+    disabled="requires multi-GPU, self-skips in CI",
+)
 
 DTYPE_MAP = {
     "float16": torch.float16,

python/sglang/jit_kernel/benchmark/bench_fused_add_rmsnorm.py

@@ -7,8 +7,11 @@ from flashinfer import fused_add_rmsnorm as fi_fused_add_rmsnorm
 
 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.test.ci.ci_register import register_cuda_ci
 from sglang.utils import is_in_ci
 
+register_cuda_ci(est_time=6, suite="stage-b-kernel-benchmark-1-gpu-large")
+
 IS_CI = is_in_ci()
 
 

python/sglang/jit_kernel/benchmark/bench_fused_norm_scale_shift.py

@@ -15,8 +15,11 @@ from sglang.multimodal_gen.runtime.layers.layernorm import (
     ScaleResidualLayerNormScaleShift,
     ScaleResidualRMSNormScaleShift,
 )
+from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.utils import is_in_ci
 
+register_cuda_ci(est_time=17, suite="stage-b-kernel-benchmark-1-gpu-large")
+
 if is_in_ci():
     B_RANGE, S_RANGE, D_RANGE = [1], [128], [1024]
 else:

python/sglang/jit_kernel/benchmark/bench_hadamard.py

@@ -14,6 +14,9 @@ from sglang.jit_kernel.benchmark.utils import (
     run_benchmark,
 )
 from sglang.jit_kernel.hadamard import hadamard_transform
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=5, suite="stage-b-kernel-benchmark-1-gpu-large")
 
 # AOT kernel: might not be available in all environments.
 # This is used for performance baseline comparison.

python/sglang/jit_kernel/benchmark/bench_hicache.py

@@ -31,6 +31,9 @@ from sglang.jit_kernel.hicache import (
     transfer_hicache_all_layer,
     transfer_hicache_one_layer,
 )
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=29, suite="stage-b-kernel-benchmark-1-gpu-large")
 
 DISABLE_TORCH = os.environ.get("DISABLE_TORCH", "0") == "1"
 PAGE_SIZE = 1

python/sglang/jit_kernel/benchmark/bench_norm.py

@@ -9,8 +9,11 @@ from flashinfer.norm import rmsnorm as fi_rmsnorm
 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.test.ci.ci_register import register_cuda_ci
 from sglang.utils import is_in_ci
 
+register_cuda_ci(est_time=5, suite="stage-b-kernel-benchmark-1-gpu-large")
+
 IS_CI = is_in_ci()
 
 DTYPE = torch.bfloat16

python/sglang/jit_kernel/benchmark/bench_norm_impls.py

@@ -21,8 +21,15 @@ from sglang.jit_kernel.diffusion.triton.rmsnorm_onepass import triton_one_pass_r
 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.test.ci.ci_register import register_cuda_ci
 from sglang.utils import is_in_ci
 
+register_cuda_ci(
+    est_time=120,
+    suite="stage-b-kernel-benchmark-1-gpu-large",
+    disabled="self-skips in CI, standalone tool",
+)
+
 os.environ.setdefault("FLASHINFER_DISABLE_VERSION_CHECK", "1")
 
 REPO_ROOT = KERNEL_PATH.parents[2]

python/sglang/jit_kernel/benchmark/bench_nvfp4_blockwise_moe.py

@@ -13,6 +13,9 @@ from sglang.jit_kernel.nvfp4 import (
     scaled_fp4_quant,
 )
 from sglang.srt.utils import is_sm100_supported
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=5, suite="stage-b-kernel-benchmark-1-gpu-large")
 
 FLOAT4_E2M1_MAX = 6.0
 FLOAT8_E4M3_MAX = torch.finfo(torch.float8_e4m3fn).max

python/sglang/jit_kernel/benchmark/bench_nvfp4_quant.py

@@ -8,6 +8,9 @@ import triton
 from sglang.jit_kernel.benchmark.utils import get_benchmark_range, run_benchmark
 from sglang.jit_kernel.nvfp4 import scaled_fp4_quant
 from sglang.srt.utils import is_sm100_supported
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=5, suite="stage-b-kernel-benchmark-1-gpu-large")
 
 FLOAT4_E2M1_MAX = 6.0
 FLOAT8_E4M3_MAX = torch.finfo(torch.float8_e4m3fn).max

python/sglang/jit_kernel/benchmark/bench_nvfp4_scaled_mm.py

@@ -8,6 +8,9 @@ import triton
 from sglang.jit_kernel.benchmark.utils import get_benchmark_range, run_benchmark
 from sglang.jit_kernel.nvfp4 import cutlass_scaled_fp4_mm, scaled_fp4_quant
 from sglang.srt.utils import is_sm100_supported
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=5, suite="stage-b-kernel-benchmark-1-gpu-large")
 
 FLOAT4_E2M1_MAX = 6.0
 FLOAT8_E4M3_MAX = torch.finfo(torch.float8_e4m3fn).max

python/sglang/jit_kernel/benchmark/bench_per_tensor_quant_fp8.py

@@ -6,6 +6,9 @@ import triton.testing
 
 from sglang.jit_kernel.benchmark.utils import get_benchmark_range, run_benchmark
 from sglang.jit_kernel.per_tensor_quant_fp8 import per_tensor_quant_fp8
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=5, suite="stage-b-kernel-benchmark-1-gpu-large")
 
 try:
     from vllm import _custom_ops as ops

python/sglang/jit_kernel/benchmark/bench_per_token_group_quant_8bit.py

@@ -17,8 +17,11 @@ 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.test.ci.ci_register import register_cuda_ci
 from sglang.utils import is_in_ci
 
+register_cuda_ci(est_time=13, suite="stage-b-kernel-benchmark-1-gpu-large")
+
 IS_CI = is_in_ci()
 
 _is_hip = is_hip()

python/sglang/jit_kernel/benchmark/bench_qknorm.py

@@ -13,6 +13,9 @@ from sglang.jit_kernel.benchmark.utils import (
 )
 from sglang.jit_kernel.norm import fused_inplace_qknorm
 from sglang.srt.utils import get_current_device_stream_fast
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=10, suite="stage-b-kernel-benchmark-1-gpu-large")
 
 alt_stream = torch.cuda.Stream()
 

python/sglang/jit_kernel/benchmark/bench_qknorm_across_heads.py

@@ -9,8 +9,11 @@ from sgl_kernel import rmsnorm
 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.test.ci.ci_register import register_cuda_ci
 from sglang.utils import is_in_ci
 
+register_cuda_ci(est_time=12, suite="stage-b-kernel-benchmark-1-gpu-large")
+
 IS_CI = is_in_ci()
 
 alt_stream = torch.cuda.Stream()

python/sglang/jit_kernel/benchmark/bench_qwen_image_modulation.py

@@ -9,8 +9,11 @@ from sglang.jit_kernel.diffusion.triton.scale_shift import (
     fuse_layernorm_scale_shift_gate_select01_kernel,
     fuse_residual_layernorm_scale_shift_gate_select01_kernel,
 )
+from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.utils import is_in_ci
 
+register_cuda_ci(est_time=13, suite="stage-b-kernel-benchmark-1-gpu-large")
+
 if is_in_ci():
     B_RANGE, S_RANGE, D_RANGE = [1], [128], [3072]
 else:

python/sglang/jit_kernel/benchmark/bench_renorm.py

@@ -6,8 +6,11 @@ import triton
 import triton.testing
 
 from sglang.jit_kernel.benchmark.utils import run_benchmark_no_cudagraph
+from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.utils import is_in_ci
 
+register_cuda_ci(est_time=5, suite="stage-b-kernel-benchmark-1-gpu-large")
+
 
 def torch_top_k_renorm_probs(probs, top_k):
     """Vectorized PyTorch implementation of top-k renormalization."""

python/sglang/jit_kernel/benchmark/bench_resolve_future_token_ids.py

@@ -11,6 +11,9 @@ from sglang.jit_kernel.benchmark.utils import (
 )
 from sglang.jit_kernel.resolve_future_token_ids import resolve_future_token_ids_cuda
 from sglang.srt.utils import get_compiler_backend
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=10, suite="stage-b-kernel-benchmark-1-gpu-large")
 
 SIZE_LIST = get_benchmark_range(
     full_range=[2**n for n in range(4, 16)],  # 16 … 32K elements

python/sglang/jit_kernel/benchmark/bench_rmsnorm.py

@@ -13,6 +13,9 @@ from sglang.jit_kernel.benchmark.utils import (
     run_benchmark,
 )
 from sglang.jit_kernel.norm import rmsnorm as jit_rmsnorm
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=21, suite="stage-b-kernel-benchmark-1-gpu-large")
 
 
 def sglang_aot_rmsnorm(

python/sglang/jit_kernel/benchmark/bench_rope.py

@@ -10,6 +10,9 @@ from sglang.jit_kernel.benchmark.utils import (
     get_benchmark_range,
     run_benchmark,
 )
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=6, suite="stage-b-kernel-benchmark-1-gpu-large")
 
 MAX_SEQ_LEN = 131072
 ROPE_BASE = 10000.0

python/sglang/jit_kernel/benchmark/bench_store_cache.py

@@ -12,6 +12,9 @@ from sglang.jit_kernel.benchmark.utils import (
     get_benchmark_range,
 )
 from sglang.jit_kernel.kvcache import store_cache
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=9, suite="stage-b-kernel-benchmark-1-gpu-large")
 
 
 def sglang_jit_store_cache(

python/sglang/jit_kernel/norm.py

@@ -6,9 +6,6 @@ from typing import TYPE_CHECKING, Optional
 import torch
 
 from sglang.jit_kernel.debug_utils import maybe_wrap_jit_kernel_debug
-
-logger = logging.getLogger(__name__)
-
 from sglang.jit_kernel.utils import (
     cache_once,
     is_arch_support_pdl,
@@ -20,6 +17,9 @@ if TYPE_CHECKING:
     from tvm_ffi.module import Module
 
 
+logger = logging.getLogger(__name__)
+
+
 @cache_once
 def _jit_qknorm_module(head_dim: int, dtype: torch.dtype) -> Module:
     args = make_cpp_args(head_dim, is_arch_support_pdl(), dtype)

python/sglang/jit_kernel/tests/test_add_constant.py

@@ -4,6 +4,10 @@ import pytest
 import torch
 
 from sglang.jit_kernel.add_constant import add_constant
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=45, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=180, suite="nightly-kernel-1-gpu", nightly=True)
 
 
 @pytest.mark.parametrize("size", [1, 2, 127, 128, 1024, 1025])

python/sglang/jit_kernel/tests/test_awq_dequantize.py

@@ -5,6 +5,10 @@ import pytest
 import torch
 
 from sglang.jit_kernel.awq_dequantize import awq_dequantize as jit_awq_dequantize
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=9, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 try:
     from sgl_kernel import awq_dequantize as aot_awq_dequantize

python/sglang/jit_kernel/tests/test_awq_marlin_moe_repack.py

@@ -9,6 +9,10 @@ from sglang.jit_kernel.awq_marlin_repack import (
     awq_marlin_moe_repack as jit_awq_marlin_moe_repack,
 )
 from sglang.srt.layers.quantization.utils import pack_cols, quantize_weights
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=10, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 
 def _has_aot_awq_marlin_moe_repack() -> bool:

python/sglang/jit_kernel/tests/test_awq_marlin_repack.py

@@ -9,8 +9,12 @@ from sglang.jit_kernel.awq_marlin_repack import (
     awq_marlin_repack as jit_awq_marlin_repack,
 )
 from sglang.srt.layers.quantization.utils import pack_cols, quantize_weights
+from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_marlin_utils import get_weight_perm, marlin_weights
 
+register_cuda_ci(est_time=10, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
+
 
 def _has_aot_awq_marlin_repack() -> bool:
     return hasattr(torch.ops.sgl_kernel, "awq_marlin_repack") and hasattr(

python/sglang/jit_kernel/tests/test_clamp_position.py

@@ -4,6 +4,10 @@ import pytest
 import torch
 
 from sglang.jit_kernel.clamp_position import clamp_position_cuda
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=12, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 
 def _reference_clamp_position(seq_lens):

python/sglang/jit_kernel/tests/test_concat_mla.py

@@ -5,6 +5,11 @@ import pytest
 import torch
 import triton
 
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=17, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
+
 
 def torch_concat_mla_k(
     k: torch.Tensor, k_nope: torch.Tensor, k_rope: torch.Tensor

python/sglang/jit_kernel/tests/test_custom_all_reduce.py

@@ -18,6 +18,7 @@ import itertools
 import logging
 import os
 import subprocess
+import sys
 from typing import Optional
 
 import pytest
@@ -30,6 +31,19 @@ from sglang.jit_kernel.all_reduce import AllReduceAlgo
 from sglang.srt.distributed.device_communicators.custom_all_reduce_v2 import (
     CustomAllReduceV2,
 )
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(
+    est_time=120,
+    suite="stage-b-kernel-unit-1-gpu-large",
+    disabled="requires multi-GPU distributed setup",
+)
+register_cuda_ci(
+    est_time=120,
+    suite="nightly-kernel-1-gpu",
+    nightly=True,
+    disabled="requires multi-GPU distributed setup",
+)
 
 # ---------------------------------------------------------------------------
 # Test parameters (shared between test class and worker)
@@ -224,4 +238,7 @@ def worker_main() -> None:
 
 
 if __name__ == "__main__":
-    worker_main()
+    if "LOCAL_RANK" in os.environ:
+        worker_main()
+    else:
+        sys.exit(pytest.main([__file__, "-v", "-s"]))

python/sglang/jit_kernel/tests/test_cutedsl_gdn.py

@@ -6,6 +6,8 @@ import numpy as np
 import pytest
 import torch
 
+from sglang.test.ci.ci_register import register_cuda_ci
+
 try:
     import cuda.bindings.driver as cuda_driver
     import cutlass  # noqa: F401
@@ -27,6 +29,9 @@ try:
 except ImportError:
     TRITON_AVAILABLE = False
 
+register_cuda_ci(est_time=5, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
+
 
 def run_triton_kernel(A_log, dt_bias, q, k, v, a, b, initial_state, indices, scale):
     return fused_sigmoid_gating_delta_rule_update(

python/sglang/jit_kernel/tests/test_flash_attention_4.py

@@ -12,6 +12,10 @@ import torch.nn.functional as F
 from einops import rearrange, repeat
 
 from sglang.jit_kernel.flash_attention_v4 import flash_attn_varlen_func
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=120, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=900, suite="nightly-kernel-1-gpu", nightly=True)
 
 # Skip this test on Hopper machine
 skip_condition = torch.cuda.get_device_capability() < (10, 0)

python/sglang/jit_kernel/tests/test_fused_add_rmsnorm.py

@@ -5,6 +5,10 @@ import pytest
 import torch
 
 from sglang.jit_kernel.utils import get_ci_test_range
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=5, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 
 def sglang_jit_fused_add_rmsnorm(

python/sglang/jit_kernel/tests/test_fused_metadata_copy.py

@@ -14,6 +14,11 @@ import time
 import pytest
 import torch
 
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=100, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=400, suite="nightly-kernel-1-gpu", nightly=True)
+
 # =============================================================================
 # Helper Functions
 # =============================================================================

python/sglang/jit_kernel/tests/test_fused_norm_scale_shift.py

@@ -10,6 +10,10 @@ from sglang.jit_kernel.diffusion.cutedsl.scale_residual_norm_scale_shift import
     fused_norm_scale_shift,
     fused_scale_residual_norm_scale_shift,
 )
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=28, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 DEVICE = "cuda"
 SHAPE_MAP = {

python/sglang/jit_kernel/tests/test_fused_store_index_cache.py

@@ -22,6 +22,8 @@ from typing import Optional, Tuple
 import pytest
 import torch
 
+from sglang.test.ci.ci_register import register_cuda_ci
+
 try:
     from sglang.jit_kernel.fused_store_index_cache import (
         can_use_nsa_fused_store,
@@ -46,6 +48,9 @@ try:
 except ImportError:
     _is_fp8_fnuz = False
 
+register_cuda_ci(est_time=24, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
+
 PAGE_SIZE = 64
 HEAD_DIM = 128
 FP8_E4M3_MAX = 448.0

python/sglang/jit_kernel/tests/test_fused_verify_triton_gdn.py

@@ -11,6 +11,8 @@ import sys
 import pytest
 import torch
 
+from sglang.test.ci.ci_register import register_cuda_ci
+
 try:
     from sglang.srt.layers.attention.fla.fused_gdn_gating import fused_gdn_gating
     from sglang.srt.layers.attention.fla.fused_recurrent import (
@@ -24,6 +26,9 @@ try:
 except ImportError:
     KERNELS_AVAILABLE = False
 
+register_cuda_ci(est_time=6, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
+
 
 def _make_tensors(N, T, H, HV, K, V, device="cuda", seed=2025):
     """Create input tensors for GDN target_verify."""

python/sglang/jit_kernel/tests/test_gptq_marlin.py

@@ -6,8 +6,12 @@ from sgl_kernel.scalar_type import scalar_types
 
 from sglang.jit_kernel.gptq_marlin import gptq_marlin_gemm
 from sglang.srt.layers.quantization.marlin_utils import marlin_make_workspace
+from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_marlin_utils import awq_marlin_quantize, marlin_quantize
 
+register_cuda_ci(est_time=13, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
+
 MNK_FACTORS = [
     (1, 1, 1),
     (1, 4, 8),

python/sglang/jit_kernel/tests/test_gptq_marlin_repack.py

@@ -10,8 +10,12 @@ from sglang.srt.layers.quantization.utils import (
     pack_rows,
     sort_weights,
 )
+from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_marlin_utils import get_weight_perm, marlin_weights
 
+register_cuda_ci(est_time=16, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
+
 MARLIN_K_CHUNKS = [128]
 MARLIN_N_CHUNKS = [64, 256]
 

python/sglang/jit_kernel/tests/test_hadamard_jit.py

@@ -14,6 +14,10 @@ from sglang.jit_kernel.hadamard import (
     hadamard_transform_28n,
     hadamard_transform_40n,
 )
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=128, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=512, suite="nightly-kernel-1-gpu", nightly=True)
 
 # Exact M×N Hadamard matrices (±1 entries) copied from
 # python/sglang/jit_kernel/csrc/fast-hadamard-transform/code_gen.py.

python/sglang/jit_kernel/tests/test_moe_lora_align_block_size.py

@@ -11,7 +11,8 @@ import torch
 from sglang.jit_kernel.moe_lora_align import moe_lora_align_block_size
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=80, suite="stage-b-test-1-gpu-large")
+register_cuda_ci(est_time=28, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 
 def round_up(x, base):

python/sglang/jit_kernel/tests/test_moe_wna16_marlin.py

@@ -7,8 +7,12 @@ from sgl_kernel.scalar_type import scalar_types
 
 from sglang.jit_kernel.moe_wna16_marlin import moe_wna16_marlin_gemm
 from sglang.srt.layers.moe.fused_moe_triton import moe_align_block_size
+from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_marlin_utils import awq_marlin_quantize, marlin_quantize
 
+register_cuda_ci(est_time=10, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
+
 
 def _has_aot_moe_wna16_marlin_gemm() -> bool:
     return hasattr(torch.ops.sgl_kernel, "moe_wna16_marlin_gemm") and hasattr(

python/sglang/jit_kernel/tests/test_norm_jit.py

@@ -5,6 +5,11 @@ import sys
 import pytest
 import torch
 
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=125, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=500, suite="nightly-kernel-1-gpu", nightly=True)
+
 # JIT rmsnorm: fp16/bf16 only
 # - Warp norm path (one warp per token):  hidden_size in {64, 128, 256}
 # - CTA norm path (multi-warp per token): hidden_size is a multiple of 256, > 256, and <=8192

python/sglang/jit_kernel/tests/test_nvfp4_blockwise_moe.py

@@ -1,3 +1,5 @@
+import sys
+
 import pytest
 import torch
 
@@ -6,6 +8,10 @@ from sglang.jit_kernel.nvfp4 import (
     scaled_fp4_experts_quant,
     scaled_fp4_quant,
 )
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=5, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 FLOAT4_E2M1_MAX = 6.0
 FLOAT8_E4M3_MAX = torch.finfo(torch.float8_e4m3fn).max
@@ -125,3 +131,7 @@ def test_nvfp4_blockwise_moe_grouped_mm(dtype: torch.dtype) -> None:
         ref[start:end] = torch.matmul(a[start:end], b[i].t())
 
     torch.testing.assert_close(out, ref, atol=1e-1, rtol=1e-1)
+
+
+if __name__ == "__main__":
+    sys.exit(pytest.main([__file__, "-v", "-s"]))

python/sglang/jit_kernel/tests/test_nvfp4_gemm.py

@@ -1,7 +1,13 @@
+import sys
+
 import pytest
 import torch
 
 from sglang.jit_kernel.nvfp4 import cutlass_scaled_fp4_mm, scaled_fp4_quant
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=5, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 
 def _nvfp4_supported() -> bool:
@@ -140,3 +146,7 @@ def test_nvfp4_gemm(dtype: torch.dtype, shape: tuple[int, int, int]) -> None:
     )
 
     torch.testing.assert_close(out, expected_out.to(dtype=dtype), atol=1e-1, rtol=1e-1)
+
+
+if __name__ == "__main__":
+    sys.exit(pytest.main([__file__, "-v", "-s"]))

python/sglang/jit_kernel/tests/test_nvfp4_quant.py

@@ -1,3 +1,5 @@
+import sys
+
 import pytest
 import torch
 
@@ -12,6 +14,11 @@ try:
 except Exception:
     _sgl_silu_and_mul = None
 
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=5, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
+
 
 def _nvfp4_supported() -> bool:
     return torch.cuda.is_available() and torch.cuda.get_device_capability() >= (10, 0)
@@ -212,3 +219,7 @@ def test_silu_and_mul_quantize_to_fp4_grouped(shape: tuple[int, int, int]) -> No
         scale_ref = recover_swizzled_scales(ref_output_scales[i], m, k)
         scale_ans = recover_swizzled_scales(output_scales[i], m, k)
         torch.testing.assert_close(scale_ref[: mask[i]], scale_ans[: mask[i]])
+
+
+if __name__ == "__main__":
+    sys.exit(pytest.main([__file__, "-v", "-s"]))

python/sglang/jit_kernel/tests/test_per_tensor_quant_fp8.py

@@ -6,6 +6,10 @@ import pytest
 import torch
 
 from sglang.jit_kernel.per_tensor_quant_fp8 import per_tensor_quant_fp8
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=16, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 try:
     from sglang.srt.utils import is_hip

python/sglang/jit_kernel/tests/test_per_token_group_quant_8bit.py

@@ -23,6 +23,10 @@ from sglang.srt.layers.quantization.fp8_kernel import (
 from sglang.srt.layers.quantization.fp8_kernel import (
     per_token_group_quant_8bit as triton_per_token_group_quant_8bit,
 )
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=16, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 configs = list(
     itertools.product(

python/sglang/jit_kernel/tests/test_pos_enc.py

@@ -8,6 +8,10 @@ import triton
 import triton.language as tl
 
 from sglang.jit_kernel.rope import rotary_embedding
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=18, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 
 @triton.jit

python/sglang/jit_kernel/tests/test_qknorm.py

@@ -6,6 +6,10 @@ import torch
 import triton
 
 from sglang.jit_kernel.utils import get_ci_test_range
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=37, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=148, suite="nightly-kernel-1-gpu", nightly=True)
 
 
 def sglang_aot_qknorm(

python/sglang/jit_kernel/tests/test_qknorm_across_heads.py

@@ -6,6 +6,10 @@ import torch
 import triton
 
 from sglang.jit_kernel.utils import get_ci_test_range
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=15, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 
 def sglang_jit_qknorm_across_heads(

python/sglang/jit_kernel/tests/test_qwen_image_modulation.py

@@ -10,6 +10,10 @@ from sglang.jit_kernel.diffusion.triton.scale_shift import (
     fuse_residual_layernorm_scale_shift_gate_select01_kernel,
 )
 from sglang.jit_kernel.utils import get_ci_test_range
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=15, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 DEVICE = "cuda"
 DTYPES = get_ci_test_range(

python/sglang/jit_kernel/tests/test_renorm.py

@@ -7,6 +7,11 @@ import pytest
 import sgl_kernel
 import torch
 
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=6, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
+
 
 @pytest.mark.parametrize("batch_size", [1, 99, 989])
 @pytest.mark.parametrize("vocab_size", [111, 32000, 128256])

python/sglang/jit_kernel/tests/test_resolve_future_token_ids.py

@@ -4,6 +4,10 @@ import pytest
 import torch
 
 from sglang.jit_kernel.resolve_future_token_ids import resolve_future_token_ids_cuda
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=9, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 
 def _reference_resolve(input_ids, future_map):

python/sglang/jit_kernel/tests/test_rmsnorm.py

@@ -6,6 +6,10 @@ import torch
 import triton
 
 from sglang.jit_kernel.utils import get_ci_test_range
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=18, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 
 def sglang_jit_rmsnorm(input: torch.Tensor, weight: torch.Tensor) -> None:

python/sglang/jit_kernel/tests/test_rope.py

@@ -5,6 +5,10 @@ import torch
 import triton
 
 from sglang.jit_kernel.utils import get_ci_test_range
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=64, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=256, suite="nightly-kernel-1-gpu", nightly=True)
 
 DEVICE = "cuda"
 DTYPE = torch.bfloat16

python/sglang/jit_kernel/tests/test_store_cache.py

@@ -6,6 +6,10 @@ import torch
 
 from sglang.jit_kernel.kvcache import can_use_store_cache, store_cache
 from sglang.jit_kernel.utils import get_ci_test_range
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=28, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 BS_LIST = [2**n for n in range(0, 15)]
 BS_LIST += [x + 1 + i for i, x in enumerate(BS_LIST)]

python/sglang/jit_kernel/tests/test_timestep_embedding.py

@@ -14,6 +14,10 @@ from sglang.jit_kernel.timestep_embedding import (
     timestep_embedding as timestep_embedding_cuda,
 )
 from sglang.jit_kernel.utils import get_ci_test_range
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=16, suite="stage-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 CORRECTNESS_BATCH_SIZES = get_ci_test_range(
     [1, 2, 8, 128, 256, 512, 1536, 2048, 4096, 11008, 16384],