[CI] Migrate JIT tests to runner config registration (#29066)

Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

test/README.md

@@ -73,8 +73,8 @@ Parameters: `est_time` (seconds), `stage` + `runner_config` (target stage and ru
 Keep `est_time`, `stage`, `runner_config` as **literal values** — `run_suite.py` collects them by AST parsing.
 
 JIT kernel correctness tests and benchmarks live under `test/registered/jit/`, same as other registered tests (their helpers stay alongside the kernel source under `python/sglang/jit_kernel/` and are imported by absolute path):
-- Correctness tests: `test/registered/jit/test_*.py` → `base-b-kernel-unit-1-gpu-large`
-- Benchmarks: `test/registered/jit/benchmark/bench_*.py` → `base-b-kernel-benchmark-1-gpu-large`
+- Correctness tests: `test/registered/jit/test_*.py` → `base-b-kernel-unit-test-1-gpu-large`
+- Benchmarks: `test/registered/jit/benchmark/bench_*.py` → `base-b-kernel-benchmark-test-1-gpu-large`
 
 ## Choosing a Suite
 
@@ -85,8 +85,8 @@ Use the lightest suite that meets your test's needs. Full suite tables are in th
 | No GPU required | `base-a-test-cpu` |
 | Small GPU (fits 5090, 32GB) | `base-b-test-1-gpu-small` (most tests go here) |
 | Large GPU memory or Hopper features | `base-b-test-1-gpu-large` |
-| JIT kernel correctness | `base-b-kernel-unit-1-gpu-large` |
-| JIT kernel benchmarks | `base-b-kernel-benchmark-1-gpu-large` |
+| JIT kernel correctness | `base-b-kernel-unit-test-1-gpu-large` |
+| JIT kernel benchmarks | `base-b-kernel-benchmark-test-1-gpu-large` |
 | Multi-GPU (2/4/8) | `base-b-test-2-gpu-large`, `base-c-test-*` |
 | Long-running or experimental | `nightly-*` suites |
 

test/registered/attention/test_gdn_prefill_cutedsl.py

@@ -15,7 +15,7 @@ from sglang.test.ci.ci_register import register_cuda_ci
 
 # CuteDSL prefill kernel only exists on Blackwell. Single-GPU kernel-unit
 # suite is the right slot (matches existing jit_kernel test_*.py pattern).
-register_cuda_ci(est_time=60, suite="base-b-kernel-unit-1-gpu-b200")
+register_cuda_ci(est_time=60, stage="base-b-kernel-unit", runner_config="4-gpu-b200")
 
 if not (torch.cuda.is_available() and torch.cuda.get_device_capability()[0] >= 10):
     pytest.skip(

test/registered/attention/test_kda_prefill_cutedsl.py

@@ -17,7 +17,7 @@ from sglang.test.ci.ci_register import register_cuda_ci
 
 # CuteDSL prefill kernel only exists on Blackwell. Single-GPU kernel-unit suite,
 # same slot as the GDN prefill test.
-register_cuda_ci(est_time=60, suite="base-b-kernel-unit-1-gpu-b200")
+register_cuda_ci(est_time=60, stage="base-b-kernel-unit", runner_config="4-gpu-b200")
 
 if not (torch.cuda.is_available() and torch.cuda.get_device_capability()[0] >= 10):
     pytest.skip(

test/registered/jit/benchmark/bench_activation.py

@@ -12,7 +12,9 @@ from sglang.jit_kernel.benchmark import marker
 from sglang.jit_kernel.benchmark.utils import create_random
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=30, suite="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=30, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 
 @torch.compile

test/registered/jit/benchmark/bench_add_constant.py

@@ -10,7 +10,9 @@ from sglang.jit_kernel.benchmark.utils import (
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=15, suite="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=15, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 CONSTANT = 7
 SIZE_LIST = get_benchmark_range(

test/registered/jit/benchmark/bench_awq_dequantize.py

@@ -9,7 +9,9 @@ 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="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=5, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 try:
     from sgl_kernel import awq_dequantize as aot_awq_dequantize

test/registered/jit/benchmark/bench_clamp_position.py

@@ -13,8 +13,10 @@ 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_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=13, suite="base-b-kernel-benchmark-1-gpu-large")
-register_amd_ci(est_time=16, suite="jit-kernel-unit-test-amd")
+register_cuda_ci(
+    est_time=13, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
+register_amd_ci(est_time=16, stage="jit-kernel-unit", runner_config="amd")
 
 SIZE_LIST = get_benchmark_range(
     full_range=[2**n for n in range(4, 16)],

test/registered/jit/benchmark/bench_concat_mla.py

@@ -12,7 +12,9 @@ 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="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=6, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 IS_CI = is_in_ci()
 

test/registered/jit/benchmark/bench_custom_all_reduce.py

@@ -33,7 +33,8 @@ from sglang.test.ci.ci_register import register_cuda_ci
 
 register_cuda_ci(
     est_time=120,
-    suite="base-b-kernel-benchmark-1-gpu-large",
+    stage="base-b-kernel-benchmark",
+    runner_config="1-gpu-large",
     disabled="requires multi-GPU, self-skips in CI",
 )
 

test/registered/jit/benchmark/bench_dsv4_fp4_indexer.py

@@ -10,7 +10,9 @@ from sglang.jit_kernel.benchmark.utils import get_benchmark_range
 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="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=5, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 try:
     import deep_gemm

test/registered/jit/benchmark/bench_fused_qknorm_rope.py

@@ -20,7 +20,9 @@ from sglang.jit_kernel.fused_qknorm_rope import (
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=6, suite="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=6, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 try:
     from sgl_kernel import fused_qk_norm_rope as fused_qk_norm_rope_aot

test/registered/jit/benchmark/bench_hadamard.py

@@ -16,7 +16,9 @@ from sglang.jit_kernel.benchmark.utils import (
 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="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=5, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 # AOT kernel: might not be available in all environments.
 # This is used for performance baseline comparison.

test/registered/jit/benchmark/bench_hicache.py

@@ -33,7 +33,9 @@ from sglang.jit_kernel.hicache import (
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=29, suite="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=29, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 DISABLE_TORCH = os.environ.get("DISABLE_TORCH", "0") == "1"
 PAGE_SIZE = 1

test/registered/jit/benchmark/bench_hisparse.py

@@ -9,7 +9,9 @@ from sglang.jit_kernel.benchmark.utils import DEFAULT_DEVICE, DEFAULT_DTYPE
 from sglang.jit_kernel.hisparse import load_cache_to_device_buffer_mla
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=12, suite="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=12, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 DEVICE = DEFAULT_DEVICE
 DTYPE = DEFAULT_DTYPE

test/registered/jit/benchmark/bench_mla_kv_pack_quantize_fp8.py

@@ -20,7 +20,9 @@ from sglang.jit_kernel.mla_kv_pack_quantize_fp8 import (
 from sglang.jit_kernel.utils import is_arch_support_pdl
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=15, suite="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=15, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 
 @triton.jit

test/registered/jit/benchmark/bench_mxfp8_moe.py

@@ -13,7 +13,9 @@ from sglang.jit_kernel.mxfp8 import (
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=5, suite="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=5, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 
 def is_sm100_supported(device=None) -> bool:

test/registered/jit/benchmark/bench_ngram_compute_decode.py

@@ -13,7 +13,9 @@ from sglang.jit_kernel.ngram_embedding import (
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=15, suite="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=15, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 NE_N = 8
 NE_K = 2

test/registered/jit/benchmark/bench_ngram_update_token_table.py

@@ -13,7 +13,9 @@ from sglang.jit_kernel.ngram_embedding import (
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=15, suite="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=15, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 MAX_CONTEXT_LEN = 4096
 BATCH_SIZE_LIST = get_benchmark_range(

test/registered/jit/benchmark/bench_norm.py

@@ -11,7 +11,9 @@ 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
 
-register_cuda_ci(est_time=30, suite="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=30, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 
 DTYPE = torch.bfloat16

test/registered/jit/benchmark/bench_nvfp4_blockwise_moe.py

@@ -15,7 +15,9 @@ from sglang.jit_kernel.nvfp4 import (
 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="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=5, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 FLOAT4_E2M1_MAX = 6.0
 FLOAT8_E4M3_MAX = torch.finfo(torch.float8_e4m3fn).max

test/registered/jit/benchmark/bench_nvfp4_quant.py

@@ -10,7 +10,9 @@ 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="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=5, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 FLOAT4_E2M1_MAX = 6.0
 FLOAT8_E4M3_MAX = torch.finfo(torch.float8_e4m3fn).max

test/registered/jit/benchmark/bench_nvfp4_scaled_mm.py

@@ -10,7 +10,9 @@ from sglang.jit_kernel.nvfp4 import cutlass_scaled_fp4_mm, scaled_fp4_quant
 from sglang.srt.utils import is_sm100_supported, is_sm120_supported
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=5, suite="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=5, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 FLOAT4_E2M1_MAX = 6.0
 FLOAT8_E4M3_MAX = torch.finfo(torch.float8_e4m3fn).max

test/registered/jit/benchmark/bench_online_c128_mtp.py

@@ -18,7 +18,9 @@ from sglang.jit_kernel.benchmark.utils import (
 from sglang.jit_kernel.dsv4.online_c128_mtp import _jit_online_c128_mtp_module
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=10, suite="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=10, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 HEAD_DIM = 512
 STATE_DIM = HEAD_DIM * 3

test/registered/jit/benchmark/bench_per_tensor_quant_fp8.py

@@ -8,7 +8,9 @@ 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="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=5, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 try:
     from vllm import _custom_ops as ops

test/registered/jit/benchmark/bench_per_token_group_quant_8bit.py

@@ -20,7 +20,9 @@ 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="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=13, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 IS_CI = is_in_ci()
 

test/registered/jit/benchmark/bench_per_token_group_quant_8bit_v2.py

@@ -14,7 +14,9 @@ from sglang.srt.layers.quantization.fp8_kernel import (
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=6, suite="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=6, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 G = 128
 HIDDEN = 4096

test/registered/jit/benchmark/bench_qknorm.py

@@ -6,7 +6,9 @@ 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="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=10, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 alt_stream = torch.cuda.Stream()
 

test/registered/jit/benchmark/bench_qknorm_across_heads.py

@@ -12,7 +12,9 @@ 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="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=12, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 IS_CI = is_in_ci()
 

test/registered/jit/benchmark/bench_renorm.py

@@ -9,7 +9,9 @@ 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="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=5, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 
 def torch_top_k_renorm_probs(probs, top_k):

test/registered/jit/benchmark/bench_resolve_future_token_ids.py

@@ -13,8 +13,10 @@ from sglang.jit_kernel.resolve_future_token_ids import resolve_future_token_ids_
 from sglang.srt.utils import get_compiler_backend
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=10, suite="base-b-kernel-benchmark-1-gpu-large")
-register_amd_ci(est_time=10, suite="jit-kernel-unit-test-amd")
+register_cuda_ci(
+    est_time=10, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
+register_amd_ci(est_time=10, stage="jit-kernel-unit", runner_config="amd")
 
 SIZE_LIST = get_benchmark_range(
     full_range=[2**n for n in range(4, 16)],  # 16 … 32K elements

test/registered/jit/benchmark/bench_rope.py

@@ -12,7 +12,9 @@ from sglang.jit_kernel.benchmark.utils import (
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=6, suite="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=6, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 MAX_SEQ_LEN = 131072
 ROPE_BASE = 10000.0

test/registered/jit/benchmark/bench_set_mla_kv_buffer.py

@@ -26,7 +26,9 @@ from sglang.srt.mem_cache.utils import set_mla_kv_buffer_kernel as sglang_triton
 from sglang.srt.mem_cache.utils import set_mla_kv_buffer_triton as sglang_wrapper
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=9, suite="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=9, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 
 def _triton_baseline(kv_buffer, loc, cache_k_nope, cache_k_rope):

test/registered/jit/benchmark/bench_store_cache.py

@@ -9,7 +9,9 @@ from sglang.jit_kernel.benchmark.utils import (
 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="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=9, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 
 @torch.compile()

test/registered/jit/benchmark/bench_tp_qknorm.py

@@ -41,7 +41,8 @@ from sglang.test.ci.ci_register import register_cuda_ci
 
 register_cuda_ci(
     est_time=120,
-    suite="base-b-kernel-benchmark-1-gpu-large",
+    stage="base-b-kernel-benchmark",
+    runner_config="1-gpu-large",
     disabled="requires multi-GPU, self-skips in CI",
 )
 

test/registered/jit/benchmark/diffusion/bench_diffusion_nvfp4_scaled_mm.py

@@ -18,7 +18,8 @@ from sglang.utils import is_in_ci
 
 register_cuda_ci(
     est_time=120,
-    suite="base-b-kernel-benchmark-1-gpu-large",
+    stage="base-b-kernel-benchmark",
+    runner_config="1-gpu-large",
     disabled="standalone diffusion NVFP4 benchmark",
 )
 

test/registered/jit/benchmark/diffusion/bench_fused_norm_scale_shift.py

@@ -20,7 +20,8 @@ from sglang.utils import is_in_ci
 
 register_cuda_ci(
     est_time=17,
-    suite="base-b-kernel-benchmark-1-gpu-large",
+    stage="base-b-kernel-benchmark",
+    runner_config="1-gpu-large",
     disabled="Temporarily skipped to unblock flashinfer upgrade. Ref: https://github.com/sgl-project/sglang/actions/runs/23735552939/job/69139238979?pr=21422",
 )
 

test/registered/jit/benchmark/diffusion/bench_group_norm_silu.py

@@ -16,7 +16,8 @@ from sglang.utils import is_in_ci
 
 register_cuda_ci(
     est_time=45,
-    suite="base-b-kernel-benchmark-1-gpu-large",
+    stage="base-b-kernel-benchmark",
+    runner_config="1-gpu-large",
     disabled="standalone benchmark",
 )
 

test/registered/jit/benchmark/diffusion/bench_norm_impls.py

@@ -24,7 +24,8 @@ from sglang.utils import is_in_ci
 
 register_cuda_ci(
     est_time=120,
-    suite="base-b-kernel-benchmark-1-gpu-large",
+    stage="base-b-kernel-benchmark",
+    runner_config="1-gpu-large",
     disabled="self-skips in CI, standalone tool",
 )
 

test/registered/jit/benchmark/diffusion/bench_qknorm_rope.py

@@ -13,7 +13,9 @@ from sglang.jit_kernel.benchmark.utils import (
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=13, suite="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=13, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 MAX_SEQ_LEN = 131072
 ROPE_BASE = 10000.0

test/registered/jit/benchmark/diffusion/bench_qwen_image_modulation.py

@@ -12,7 +12,9 @@ from sglang.jit_kernel.diffusion.triton.scale_shift import (
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.utils import is_in_ci
 
-register_cuda_ci(est_time=13, suite="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=13, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 if is_in_ci():
     B_RANGE, S_RANGE, D_RANGE = [1], [128], [3072]

test/registered/jit/benchmark/minimax/bench_minimax_decode_topk.py

@@ -17,7 +17,9 @@ from sglang.srt.layers.attention.minimax_sparse_ops.decode.flash_with_topk_idx i
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=8, suite="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=8, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 BLOCK_SIZE = 128
 TOPK = 16

test/registered/jit/benchmark/minimax/bench_minimax_qknorm_rope.py

@@ -12,7 +12,9 @@ from sglang.jit_kernel.minimax_qknorm_rope import (
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=6, suite="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=6, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 HEAD_DIM, ROTARY_DIM, BASE, EPS, MAXPOS = 128, 64, 5_000_000, 1e-6, 131072
 NQ, NK = 64, 4

test/registered/jit/benchmark/minimax/bench_minimax_store_kv_index.py

@@ -7,7 +7,9 @@ from sglang.jit_kernel.benchmark import marker
 from sglang.jit_kernel.minimax_store_kv_index import store_kv_index
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=6, suite="base-b-kernel-benchmark-1-gpu-large")
+register_cuda_ci(
+    est_time=6, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
+)
 
 HEAD_DIM = 128
 NUM_KV_HEADS = 1

test/registered/jit/deepseek_v4/test_c128_v2.py

@@ -18,7 +18,7 @@ from sglang.jit_kernel.tests.deepseek_v4.common import (
 )
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=30, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=30, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=30, suite="nightly-kernel-1-gpu", nightly=True)
 register_amd_ci(est_time=30, suite="nightly-amd-kernel-1-gpu", nightly=True)
 

test/registered/jit/deepseek_v4/test_c4_v2.py

@@ -18,7 +18,7 @@ from sglang.jit_kernel.tests.deepseek_v4.common import (
 )
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=30, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=30, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=30, suite="nightly-kernel-1-gpu", nightly=True)
 register_amd_ci(est_time=30, suite="nightly-amd-kernel-1-gpu", nightly=True)
 

test/registered/jit/deepseek_v4/test_fp4_indexer.py

@@ -21,7 +21,7 @@ from sglang.srt.layers.deepseek_v4_rope import (
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=60, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=60, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=60, suite="nightly-kernel-1-gpu", nightly=True)
 
 HEAD_DIM = 128

test/registered/jit/diffusion/test_diffusion_modelopt_fp8_scaled_mm.py

@@ -14,7 +14,7 @@ from sglang.srt.layers.quantization.fp8_utils import (
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=20, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=20, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=80, suite="nightly-kernel-1-gpu", nightly=True)
 
 DEVICE = "cuda"

test/registered/jit/diffusion/test_diffusion_nvfp4_scaled_mm.py

@@ -17,7 +17,7 @@ from sglang.srt.layers.quantization.modelopt_quant import pad_nvfp4_weight
 from sglang.test.ci.ci_register import register_cuda_ci
 
 # B200-only correctness coverage for diffusion NVFP4 scaled mm.
-register_cuda_ci(est_time=15, suite="base-b-kernel-unit-1-gpu-b200")
+register_cuda_ci(est_time=15, stage="base-b-kernel-unit", runner_config="4-gpu-b200")
 
 DEVICE = "cuda"
 DTYPE = torch.bfloat16

test/registered/jit/diffusion/test_flydsl_fused_norm.py

@@ -6,7 +6,7 @@ import torch.nn.functional as F
 
 from sglang.test.ci.ci_register import register_amd_ci
 
-register_amd_ci(est_time=30, suite="jit-kernel-unit-test-amd")
+register_amd_ci(est_time=30, stage="jit-kernel-unit", runner_config="amd")
 
 DEVICE = "cuda"
 D = 5120

test/registered/jit/diffusion/test_fused_norm_scale_shift.py

@@ -13,7 +13,7 @@ from sglang.jit_kernel.diffusion.cutedsl.scale_residual_norm_scale_shift import
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=28, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=28, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 DEVICE = "cuda"

test/registered/jit/diffusion/test_group_norm_silu.py

@@ -9,7 +9,7 @@ from sglang.jit_kernel.diffusion.group_norm_silu import apply_group_norm_silu
 from sglang.jit_kernel.diffusion.triton.group_norm_silu import triton_group_norm_silu
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=8, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=8, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 register_amd_ci(est_time=15, suite="nightly-amd-kernel-1-gpu", nightly=True)
 

test/registered/jit/diffusion/test_qknorm_rope.py

@@ -8,7 +8,7 @@ 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=44, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=44, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=176, suite="nightly-kernel-1-gpu", nightly=True)
 
 DEVICE = "cuda"

test/registered/jit/diffusion/test_qwen_image_modulation.py

@@ -12,7 +12,7 @@ from sglang.jit_kernel.diffusion.triton.scale_shift import (
 from sglang.jit_kernel.utils import get_ci_test_range
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=15, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=15, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 register_amd_ci(est_time=30, suite="nightly-amd-kernel-1-gpu", nightly=True)
 

test/registered/jit/diffusion/test_varlen_pack_pad.py

@@ -15,7 +15,7 @@ from sglang.jit_kernel.diffusion.triton.varlen_pack_pad import (
 from sglang.jit_kernel.utils import get_ci_test_range
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=10, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=10, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=60, suite="nightly-kernel-1-gpu", nightly=True)
 register_amd_ci(est_time=15, suite="nightly-amd-kernel-1-gpu", nightly=True)
 

test/registered/jit/diffusion/test_varlen_uspattn_equivalence.py

@@ -29,7 +29,7 @@ from sglang.multimodal_gen.runtime.layers.attention.layer import (
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=15, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=15, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=60, suite="nightly-kernel-1-gpu", nightly=True)
 
 DEVICE = "cuda"

test/registered/jit/kv_canary/test_const_sync.py

@@ -7,8 +7,8 @@ import sglang.jit_kernel
 from sglang.jit_kernel.kv_canary import consts
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=5, suite="base-b-kernel-unit-1-gpu-large")
-register_amd_ci(est_time=5, suite="jit-kernel-unit-test-amd")
+register_cuda_ci(est_time=5, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+register_amd_ci(est_time=5, stage="jit-kernel-unit", runner_config="amd")
 
 
 # Resolve the kernel source against the installed jit_kernel package rather

test/registered/jit/kv_canary/test_kernel_config.py

@@ -35,8 +35,8 @@ from sglang.jit_kernel.tests.kv_canary._fixtures import (
 )
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=60, suite="base-b-kernel-unit-1-gpu-large")
-register_amd_ci(est_time=60, suite="jit-kernel-unit-test-amd")
+register_cuda_ci(est_time=60, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+register_amd_ci(est_time=60, stage="jit-kernel-unit", runner_config="amd")
 
 _DEVICE = torch.device("cuda")
 

test/registered/jit/kv_canary/test_pipeline_e2e.py

@@ -40,8 +40,8 @@ from sglang.jit_kernel.tests.kv_canary._fixtures import (
 )
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=30, suite="base-b-kernel-unit-1-gpu-large")
-register_amd_ci(est_time=30, suite="jit-kernel-unit-test-amd")
+register_cuda_ci(est_time=30, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+register_amd_ci(est_time=30, stage="jit-kernel-unit", runner_config="amd")
 
 
 _DEVICE = torch.device("cuda")

test/registered/jit/kv_canary/test_plan_fuzz.py

@@ -22,8 +22,8 @@ from sglang.jit_kernel.tests.kv_canary._fuzz_driver import (
 from sglang.jit_kernel.tests.kv_canary._invariants import PlanInvariants
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=30, suite="base-b-kernel-unit-1-gpu-large")
-register_amd_ci(est_time=30, suite="jit-kernel-unit-test-amd")
+register_cuda_ci(est_time=30, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+register_amd_ci(est_time=30, stage="jit-kernel-unit", runner_config="amd")
 
 
 _DEVICE = torch.device("cuda")

test/registered/jit/kv_canary/test_plan_hand.py

@@ -22,8 +22,8 @@ from sglang.jit_kernel.tests.kv_canary._fixtures import (
 from sglang.jit_kernel.tests.kv_canary._invariants import PlanInvariants
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=30, suite="base-b-kernel-unit-1-gpu-large")
-register_amd_ci(est_time=30, suite="jit-kernel-unit-test-amd")
+register_cuda_ci(est_time=30, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+register_amd_ci(est_time=30, stage="jit-kernel-unit", runner_config="amd")
 
 
 _DEVICE = torch.device("cuda")

test/registered/jit/kv_canary/test_scatter_req_token_ids.py

@@ -13,8 +13,8 @@ from sglang.jit_kernel.kv_canary.scatter_req_token_ids import (
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cuda_ci(est_time=10, suite="base-b-kernel-unit-1-gpu-large")
-register_amd_ci(est_time=10, suite="jit-kernel-unit-test-amd")
+register_cuda_ci(est_time=10, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+register_amd_ci(est_time=10, stage="jit-kernel-unit", runner_config="amd")
 
 
 _DEVICE = torch.device("cuda")

test/registered/jit/kv_canary/test_utils.py

@@ -8,7 +8,7 @@ from sglang.jit_kernel.benchmark.kv_canary.utils import (
 )
 from sglang.test.ci.ci_register import register_cpu_ci
 
-register_cpu_ci(est_time=3, suite="base-a-test-cpu")
+register_cpu_ci(est_time=3, stage="base-a", runner_config="cpu")
 
 
 def test_fast_matrix_cases_include_e2e_decode_and_chunked_prefill_scenarios() -> None:

test/registered/jit/kv_canary/test_verify_fuzz.py

@@ -31,8 +31,8 @@ from sglang.jit_kernel.tests.kv_canary._fuzz_driver import (
 from sglang.jit_kernel.tests.kv_canary._invariants import VerifyInvariants
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=30, suite="base-b-kernel-unit-1-gpu-large")
-register_amd_ci(est_time=30, suite="jit-kernel-unit-test-amd")
+register_cuda_ci(est_time=30, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+register_amd_ci(est_time=30, stage="jit-kernel-unit", runner_config="amd")
 
 
 _DEVICE = torch.device("cuda")

test/registered/jit/kv_canary/test_verify_hand.py

@@ -53,8 +53,8 @@ from sglang.jit_kernel.tests.kv_canary._hand_oracle import (
 )
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=30, suite="base-b-kernel-unit-1-gpu-large")
-register_amd_ci(est_time=30, suite="jit-kernel-unit-test-amd")
+register_cuda_ci(est_time=30, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+register_amd_ci(est_time=30, stage="jit-kernel-unit", runner_config="amd")
 
 
 _DEVICE = torch.device("cuda")

test/registered/jit/kv_canary/test_write_fuzz.py

@@ -31,8 +31,8 @@ from sglang.jit_kernel.tests.kv_canary._fuzz_driver import (
 from sglang.jit_kernel.tests.kv_canary._invariants import WriteInvariants
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=30, suite="base-b-kernel-unit-1-gpu-large")
-register_amd_ci(est_time=30, suite="jit-kernel-unit-test-amd")
+register_cuda_ci(est_time=30, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+register_amd_ci(est_time=30, stage="jit-kernel-unit", runner_config="amd")
 
 
 _DEVICE = torch.device("cuda")

test/registered/jit/kv_canary/test_write_hand.py

@@ -51,8 +51,8 @@ from sglang.jit_kernel.tests.kv_canary._hand_oracle import (
 )
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=30, suite="base-b-kernel-unit-1-gpu-large")
-register_amd_ci(est_time=30, suite="jit-kernel-unit-test-amd")
+register_cuda_ci(est_time=30, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+register_amd_ci(est_time=30, stage="jit-kernel-unit", runner_config="amd")
 
 
 _DEVICE = torch.device("cuda")

test/registered/jit/minimax/test_minimax_decode_topk.py

@@ -12,8 +12,8 @@ import torch
 from sglang.jit_kernel.minimax_decode_topk import minimax_decode_topk
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=40, suite="base-b-kernel-unit-1-gpu-large")
-register_cuda_ci(est_time=40, suite="base-b-kernel-unit-1-gpu-b200")
+register_cuda_ci(est_time=40, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+register_cuda_ci(est_time=40, stage="base-b-kernel-unit", runner_config="4-gpu-b200")
 register_amd_ci(est_time=15, suite="nightly-amd-kernel-1-gpu", nightly=True)
 
 

test/registered/jit/minimax/test_minimax_decode_topk_page_table.py

@@ -25,7 +25,7 @@ from sglang.srt.layers.attention.minimax_sparse_ops.decode.topk_sparse import (
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=25, suite="base-b-kernel-unit-1-gpu-b200")
+register_cuda_ci(est_time=25, stage="base-b-kernel-unit", runner_config="4-gpu-b200")
 
 dev = "cuda"
 

test/registered/jit/minimax/test_minimax_qknorm_rope.py

@@ -14,8 +14,8 @@ from sglang.jit_kernel.minimax_qknorm_rope import (
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=30, suite="base-b-kernel-unit-1-gpu-large")
-register_cuda_ci(est_time=30, suite="base-b-kernel-unit-1-gpu-b200")
+register_cuda_ci(est_time=30, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+register_cuda_ci(est_time=30, stage="base-b-kernel-unit", runner_config="4-gpu-b200")
 
 dev = "cuda"
 HEAD_DIM, ROTARY_DIM, BASE, EPS = 128, 64, 5_000_000, 1e-6

test/registered/jit/minimax/test_minimax_store_kv_index.py

@@ -11,8 +11,8 @@ import torch
 from sglang.jit_kernel.minimax_store_kv_index import store_kv_index
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=30, suite="base-b-kernel-unit-1-gpu-large")
-register_cuda_ci(est_time=30, suite="base-b-kernel-unit-1-gpu-b200")
+register_cuda_ci(est_time=30, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+register_cuda_ci(est_time=30, stage="base-b-kernel-unit", runner_config="4-gpu-b200")
 register_amd_ci(est_time=10, suite="nightly-amd-kernel-1-gpu", nightly=True)
 
 dev = "cuda"

test/registered/jit/test_activation.py

@@ -12,9 +12,9 @@ from sglang.jit_kernel.activation import (
 from sglang.jit_kernel.utils import get_ci_test_range
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=20, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=20, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=30, suite="nightly-kernel-1-gpu", nightly=True)
-register_amd_ci(est_time=20, suite="jit-kernel-unit-test-amd")
+register_amd_ci(est_time=20, stage="jit-kernel-unit", runner_config="amd")
 
 
 OPS = SUPPORTED_ACTIVATIONS

test/registered/jit/test_add_constant.py

@@ -6,7 +6,7 @@ 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="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=45, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=180, suite="nightly-kernel-1-gpu", nightly=True)
 
 

test/registered/jit/test_amd_deterministic_custom_allreduce.py

@@ -25,7 +25,7 @@ import torch.distributed as dist
 from sglang.srt.environ import envs
 from sglang.test.ci.ci_register import register_amd_ci
 
-register_amd_ci(est_time=120, suite="sgl-kernel-unit-test-2-gpu-amd")
+register_amd_ci(est_time=120, stage="sgl-kernel-unit", runner_config="2-gpu-amd")
 
 
 def get_open_port():

test/registered/jit/test_amd_nccl_allreduce_determinism.py

@@ -22,7 +22,7 @@ import torch.distributed as dist
 
 from sglang.test.ci.ci_register import register_amd_ci
 
-register_amd_ci(est_time=120, suite="sgl-kernel-unit-test-2-gpu-amd")
+register_amd_ci(est_time=120, stage="sgl-kernel-unit", runner_config="2-gpu-amd")
 
 
 def get_open_port():

test/registered/jit/test_awq_dequantize.py

@@ -7,7 +7,7 @@ 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="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=9, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 try:

test/registered/jit/test_awq_marlin_moe_repack.py

@@ -11,7 +11,7 @@ from sglang.jit_kernel.awq_marlin_repack import (
 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="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=10, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 

test/registered/jit/test_awq_marlin_repack.py

@@ -12,7 +12,7 @@ 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="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=10, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 

test/registered/jit/test_clamp_position.py

@@ -6,9 +6,9 @@ import torch
 from sglang.jit_kernel.clamp_position import clamp_position_cuda
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=12, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=12, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
-register_amd_ci(est_time=12, suite="jit-kernel-unit-test-amd")
+register_amd_ci(est_time=12, stage="jit-kernel-unit", runner_config="amd")
 
 
 def _reference_clamp_position(seq_lens):

test/registered/jit/test_concat_mla.py

@@ -7,7 +7,7 @@ import triton
 
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=17, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=17, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 

test/registered/jit/test_custom_all_reduce.py

@@ -46,7 +46,8 @@ from sglang.test.ci.ci_register import register_cuda_ci
 
 register_cuda_ci(
     est_time=300,
-    suite="base-b-kernel-unit-8-gpu-h200",
+    stage="base-b-kernel-unit",
+    runner_config="8-gpu-h200",
 )
 register_cuda_ci(
     est_time=300,

test/registered/jit/test_cutedsl_gdn.py

@@ -29,7 +29,7 @@ try:
 except ImportError:
     TRITON_AVAILABLE = False
 
-register_cuda_ci(est_time=5, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=5, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 

test/registered/jit/test_deepseek_v4_compress_state_runtime_shapes.py

@@ -101,7 +101,7 @@ from sglang.jit_kernel.tests.deepseek_v4.common import (
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.utils import is_in_ci
 
-register_cuda_ci(est_time=30, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=30, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_amd_ci(est_time=25, suite="nightly-amd-kernel-1-gpu", nightly=True)
 
 Mode = Literal["decode", "prefill"]

test/registered/jit/test_flash_attention_3_only_qv.py

@@ -15,7 +15,7 @@ from sglang.jit_kernel.flash_attention import flash_attn_with_kvcache
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_cuda_ci(est_time=30, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=30, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 
 # FA3 only_qv path is SM90 (Hopper) only — skip on pre-Hopper and on
 # Blackwell+ (sm100+) where FA3 is not built.

test/registered/jit/test_flash_attention_4.py

@@ -14,8 +14,8 @@ from einops import rearrange, repeat
 from sglang.jit_kernel.flash_attention import flash_attn_varlen_func
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=120, suite="base-b-kernel-unit-1-gpu-large")
-register_cuda_ci(est_time=120, suite="base-b-kernel-unit-1-gpu-b200")
+register_cuda_ci(est_time=120, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+register_cuda_ci(est_time=120, stage="base-b-kernel-unit", runner_config="4-gpu-b200")
 register_cuda_ci(est_time=900, suite="nightly-kernel-1-gpu", nightly=True)
 
 # Skip this test on Hopper machine

test/registered/jit/test_fused_add_rmsnorm.py

@@ -7,7 +7,7 @@ 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=10, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=10, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 

test/registered/jit/test_fused_metadata_copy.py

@@ -16,7 +16,7 @@ import torch
 
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=100, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=100, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=400, suite="nightly-kernel-1-gpu", nightly=True)
 
 # =============================================================================

test/registered/jit/test_fused_qk_gemma_rmsnorm_gate.py

@@ -7,7 +7,7 @@ import torch
 from sglang.srt.models.utils import fused_qk_gemma_rmsnorm_with_gate
 from sglang.test.ci.ci_register import register_amd_ci
 
-register_amd_ci(est_time=20, suite="jit-kernel-unit-test-amd")
+register_amd_ci(est_time=20, stage="jit-kernel-unit", runner_config="amd")
 
 
 def reference_qk_gemma_rmsnorm_with_gate(

test/registered/jit/test_fused_store_index_cache.py

@@ -48,7 +48,7 @@ try:
 except ImportError:
     _is_fp8_fnuz = False
 
-register_cuda_ci(est_time=24, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=24, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 register_amd_ci(est_time=24, suite="nightly-amd-kernel-1-gpu", nightly=True)
 

test/registered/jit/test_fused_verify_triton_gdn.py

@@ -26,7 +26,7 @@ try:
 except ImportError:
     KERNELS_AVAILABLE = False
 
-register_cuda_ci(est_time=6, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=6, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 register_amd_ci(est_time=10, suite="nightly-amd-kernel-1-gpu", nightly=True)
 

test/registered/jit/test_gptq_marlin.py

@@ -23,7 +23,7 @@ from sglang.test.test_marlin_utils import (
     marlin_quantize,
 )
 
-register_cuda_ci(est_time=13, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=13, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 MNK_FACTORS = [

test/registered/jit/test_gptq_marlin_repack.py

@@ -13,7 +13,7 @@ from sglang.srt.layers.quantization.utils import (
 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="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=16, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 MARLIN_K_CHUNKS = [128]

test/registered/jit/test_grouped_topk.py

@@ -9,7 +9,7 @@ from sglang.jit_kernel.utils import get_ci_test_range
 from sglang.srt.layers.moe.topk import biased_grouped_topk_impl
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=30, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=30, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 

test/registered/jit/test_hadamard_jit.py

@@ -16,7 +16,7 @@ from sglang.jit_kernel.hadamard import (
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=128, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=128, stage="base-b-kernel-unit", runner_config="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

test/registered/jit/test_hicache.py

@@ -16,7 +16,7 @@ from sglang.srt.mem_cache.pool_host.common import (
 from sglang.srt.utils import is_cuda, is_hip, is_npu, is_xpu
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=10, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=10, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 pytestmark = pytest.mark.skipif(

test/registered/jit/test_hisparse.py

@@ -11,8 +11,8 @@ from sglang.jit_kernel.hisparse import (
 from sglang.srt.utils import is_cuda, is_hip, is_npu, is_xpu
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_amd_ci(est_time=30, suite="stage-b-test-1-gpu-small-amd")
-register_cuda_ci(est_time=10, suite="base-b-kernel-unit-1-gpu-large")
+register_amd_ci(est_time=30, stage="stage-b", runner_config="1-gpu-small-amd")
+register_cuda_ci(est_time=10, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)
 
 pytestmark = pytest.mark.skipif(

test/registered/jit/test_kpool_topk_transform.py

@@ -17,7 +17,7 @@ import torch
 from sglang.jit_kernel.kpool_topk_transform import fast_kpool_topk_transform_fused
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=60, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=60, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 
 
 def _ref_torch_kpool_transform_impl(

test/registered/jit/test_kvcacheio_asymmetric.py

@@ -7,7 +7,7 @@ import torch
 from sglang.srt.mem_cache.memory_pool_host import AsymmetricMHATokenToKVPoolHost
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=10, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=10, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_amd_ci(est_time=10, suite="nightly-amd-kernel-1-gpu", nightly=True)
 
 # These tests use AsymmetricMHATokenToKVPoolHost methods and let that class call

test/registered/jit/test_minimax_m3_qk_norm_rope.py

@@ -22,7 +22,7 @@ from sglang.jit_kernel.minimax_m3.qk_norm_rope import (  # noqa: E402
 from sglang.test.ci.ci_register import register_amd_ci  # noqa: E402
 
 # ROCm-only fused kernel; runs in the AMD jit-kernel unit suite.
-register_amd_ci(est_time=30, suite="jit-kernel-unit-test-amd")
+register_amd_ci(est_time=30, stage="jit-kernel-unit", runner_config="amd")
 
 DEVICE = "cuda"
 EPS = 1e-6

test/registered/jit/test_mla_kv_pack_quantize_fp8.py

@@ -7,7 +7,7 @@ from sglang.jit_kernel.mla_kv_pack_quantize_fp8 import mla_kv_pack_quantize_fp8
 from sglang.jit_kernel.utils import get_ci_test_range
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=60, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=60, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_amd_ci(est_time=60, suite="nightly-amd-kernel-1-gpu", nightly=True)
 
 DEVICE = "cuda"

test/registered/jit/test_moe_align_block_size.py

@@ -9,7 +9,7 @@ import triton.language as tl
 from sglang.jit_kernel.moe_align import moe_align_block_size
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=28, suite="base-b-kernel-unit-1-gpu-large")
+register_cuda_ci(est_time=28, stage="base-b-kernel-unit", runner_config="1-gpu-large")
 register_cuda_ci(est_time=120, suite="nightly-kernel-1-gpu", nightly=True)