ci: decouple stage and runner for cuda registry

python/sglang/test/ci/ci_register.py

@@ -16,7 +16,12 @@
     "ut_parse_one_file",
 ]
 
+# `suite` stays in positional slot 2 for backward compat with existing
+# `register_cpu_ci(5, "stage-a-test-cpu")` style positional calls. New fields
+# (`stage`, `runner_config`) are kwarg-only.
 _PARAM_ORDER = ("est_time", "suite", "nightly", "disabled")
+_KWARG_ONLY = ("stage", "runner_config")
+_ALL_PARAMS = _PARAM_ORDER + _KWARG_ONLY
 _UNSET = object()
 
 
@@ -31,45 +36,69 @@ class HWBackend(Enum):
 class CIRegistry:
     backend: HWBackend
     filename: str
-    # Estimated time to run the test in seconds.
     est_time: float
-    # The suite this test is registered in.
-    suite: str
-    # Whether the test is a nightly test.
+    stage: Optional[str] = None
+    runner_config: Optional[str] = None
+    # Legacy single-string suite; kept for nightly/stress/weekly + AMD/CPU/NPU
+    # suites whose names don't follow `{stage}-test-{runner_config}` shape.
+    suite: Optional[str] = None
     nightly: bool = False
-    # Reason for disabling the test. None = enabled, string = disabled with reason.
     disabled: Optional[str] = None
 
+    @property
+    def effective_suite(self) -> Optional[str]:
+        if self.stage is not None and self.runner_config is not None:
+            return f"{self.stage}-test-{self.runner_config}"
+        return self.suite
+
 
 def register_cpu_ci(
-    est_time: float, suite: str, nightly: bool = False, disabled: Optional[str] = None
+    est_time: float,
+    suite: Optional[str] = None,
+    nightly: bool = False,
+    disabled: Optional[str] = None,
+    *,
+    stage: Optional[str] = None,
+    runner_config: Optional[str] = None,
 ):
     """Marker for CPU CI registration (parsed via AST; runtime no-op)."""
     return None
 
 
 def register_cuda_ci(
-    est_time: float, suite: str, nightly: bool = False, disabled: Optional[str] = None
+    est_time: float,
+    suite: Optional[str] = None,
+    nightly: bool = False,
+    disabled: Optional[str] = None,
+    *,
+    stage: Optional[str] = None,
+    runner_config: Optional[str] = None,
 ):
     """Marker for CUDA CI registration (parsed via AST; runtime no-op)."""
     return None
 
 
 def register_amd_ci(
     est_time: float,
-    suite: str,
+    suite: Optional[str] = None,
     nightly: bool = False,
     disabled: Optional[str] = None,
+    *,
+    stage: Optional[str] = None,
+    runner_config: Optional[str] = None,
 ):
     """Marker for AMD CI registration (parsed via AST; runtime no-op)."""
     return None
 
 
 def register_npu_ci(
     est_time: float,
-    suite: str,
+    suite: Optional[str] = None,
     nightly: bool = False,
     disabled: Optional[str] = None,
+    *,
+    stage: Optional[str] = None,
+    runner_config: Optional[str] = None,
 ):
     """Marker for NPU CI registration (parsed via AST; runtime no-op)."""
     return None
@@ -94,10 +123,8 @@ def _constant_value(self, node: ast.AST) -> object:
             return node.value
         return _UNSET
 
-    def _parse_call_args(
-        self, func_call: ast.Call
-    ) -> tuple[float, str, bool, Optional[str]]:
-        args = {name: _UNSET for name in _PARAM_ORDER}
+    def _parse_call_args(self, func_call: ast.Call) -> dict:
+        args = {name: _UNSET for name in _ALL_PARAMS}
         seen = set()
 
         if any(isinstance(arg, ast.Starred) for arg in func_call.args):
@@ -129,23 +156,49 @@ def _parse_call_args(
             seen.add(kw.arg)
             args[kw.arg] = self._constant_value(kw.value)
 
-        if args["est_time"] is _UNSET or args["suite"] is _UNSET:
+        if args["est_time"] is _UNSET:
             raise ValueError(
-                f"{self.filename}: est_time and suite are required constants in {func_call.func.id}()"
+                f"{self.filename}: est_time is a required constant in {func_call.func.id}()"
             )
 
-        est_time, suite = args["est_time"], args["suite"]
-        nightly_value = args["nightly"]
+        # The only valid (stage, runner_config, suite) shapes are:
+        #   (set,   set,   unset)  -> new-style pair
+        #   (unset, unset, set)    -> legacy single-string
+        # Any other combination is rejected with the actual triple in the error.
+        stage_set = args["stage"] is not _UNSET
+        runner_set = args["runner_config"] is not _UNSET
+        suite_set = args["suite"] is not _UNSET
+        valid_shape = (stage_set and runner_set and not suite_set) or (
+            not stage_set and not runner_set and suite_set
+        )
+        if not valid_shape:
+            raise ValueError(
+                f"{self.filename}: {func_call.func.id}() must specify exactly one of "
+                f"(stage, runner_config) pair or suite; got stage={stage_set}, "
+                f"runner_config={runner_set}, suite={suite_set}"
+            )
 
+        est_time = args["est_time"]
         if not isinstance(est_time, (int, float)):
             raise ValueError(
                 f"{self.filename}: est_time must be a number in {func_call.func.id}()"
             )
-        if not isinstance(suite, str):
+
+        suite = args["suite"] if suite_set else None
+        if suite is not None and not isinstance(suite, str):
             raise ValueError(
                 f"{self.filename}: suite must be a string in {func_call.func.id}()"
             )
 
+        stage = args["stage"] if stage_set else None
+        runner_config = args["runner_config"] if runner_set else None
+        for name, value in (("stage", stage), ("runner_config", runner_config)):
+            if value is not None and not isinstance(value, str):
+                raise ValueError(
+                    f"{self.filename}: {name} must be a string in {func_call.func.id}()"
+                )
+
+        nightly_value = args["nightly"]
         if nightly_value is _UNSET:
             nightly = False
         elif isinstance(nightly_value, bool):
@@ -161,7 +214,14 @@ def _parse_call_args(
                 f"{self.filename}: disabled must be a string in {func_call.func.id}()"
             )
 
-        return float(est_time), suite, nightly, disabled
+        return {
+            "est_time": float(est_time),
+            "stage": stage,
+            "runner_config": runner_config,
+            "suite": suite,
+            "nightly": nightly,
+            "disabled": disabled,
+        }
 
     def _collect_ci_registry(self, func_call: ast.Call):
         if not isinstance(func_call.func, ast.Name):
@@ -171,14 +231,11 @@ def _collect_ci_registry(self, func_call: ast.Call):
         if backend is None:
             return None
 
-        est_time, suite, nightly, disabled = self._parse_call_args(func_call)
+        parsed = self._parse_call_args(func_call)
         return CIRegistry(
             backend=backend,
             filename=self.filename,
-            est_time=est_time,
-            suite=suite,
-            nightly=nightly,
-            disabled=disabled,
+            **parsed,
         )
 
     @staticmethod

scripts/ci/utils/ci_coverage_report.py

@@ -155,7 +155,9 @@ def generate_summary_section(data: dict) -> str:
         for t in sorted(disabled_tests, key=lambda x: (x.backend.name, x.filename)):
             test_name = get_test_basename(t.filename)
             reason = t.disabled[:50] + "..." if len(t.disabled) > 50 else t.disabled
-            lines.append(f"| `{test_name}` | {t.backend.name} | {t.suite} | {reason} |")
+            lines.append(
+                f"| `{test_name}` | {t.backend.name} | {t.effective_suite} | {reason} |"
+            )
         lines.append("\n</details>\n")
 
     return "\n".join(lines)
@@ -197,7 +199,7 @@ def generate_by_folder_section(data: dict) -> str:
                     else ("Nightly" if t.nightly else "Per-Commit")
                 )
                 lines.append(
-                    f"| `{test_name}` | {t.suite} | {t.est_time:.0f}s | {status} |"
+                    f"| `{test_name}` | {t.effective_suite} | {t.est_time:.0f}s | {status} |"
                 )
 
             lines.append("")
@@ -231,7 +233,7 @@ def generate_by_suite_section(data: dict) -> str:
         # Group by suite within backend
         backend_suites = defaultdict(list)
         for t in backend_tests:
-            backend_suites[t.suite].append(t)
+            backend_suites[t.effective_suite].append(t)
 
         for suite in sorted(backend_suites.keys()):
             suite_tests = backend_suites[suite]
@@ -336,7 +338,7 @@ def generate_json_report(tests: list[CIRegistry]) -> str:
                 data["tests_by_folder"][folder]["backends"][backend] = [
                     {
                         "filename": get_test_basename(t.filename),
-                        "suite": t.suite,
+                        "suite": t.effective_suite,
                         "est_time": t.est_time,
                         "status": (
                             "disabled"
@@ -355,7 +357,7 @@ def generate_json_report(tests: list[CIRegistry]) -> str:
 
         backend_suites = defaultdict(list)
         for t in backend_tests:
-            backend_suites[t.suite].append(t)
+            backend_suites[t.effective_suite].append(t)
 
         data["tests_by_suite"][backend] = {
             "total": len(backend_tests),
@@ -423,7 +425,7 @@ def generate_json_report(tests: list[CIRegistry]) -> str:
             {
                 "filename": get_test_basename(t.filename),
                 "backend": t.backend.name,
-                "suite": t.suite,
+                "suite": t.effective_suite,
                 "reason": t.disabled,
             }
         )

scripts/ci/utils/compute_partitions.py

@@ -88,7 +88,7 @@ def compute_partitions(tests, full_parallel=False):
             continue
         if t.nightly or t.disabled is not None:
             continue
-        suite_tests[t.suite].append(t)
+        suite_tests[t.effective_suite].append(t)
 
     result = {}
     for suite, group in suite_tests.items():

test/registered/4-gpu-models/test_deepseek_v3_cutedsl_4gpu.py

@@ -14,7 +14,7 @@
     try_cached_model,
 )
 
-register_cuda_ci(est_time=1800, suite="stage-c-test-4-gpu-gb200")
+register_cuda_ci(est_time=1800, stage="stage-c", runner_config="4-gpu-gb200")
 
 
 class TestDeepseekR1Nvfp4CuteDSLDeepEP(CustomTestCase):

test/registered/4-gpu-models/test_gpt_oss_4gpu.py

@@ -3,8 +3,8 @@
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.gpt_oss_common import BaseTestGptOss
 
-register_cuda_ci(est_time=392, suite="stage-c-test-4-gpu-h100")
-register_cuda_ci(est_time=740, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=392, stage="stage-c", runner_config="4-gpu-h100")
+register_cuda_ci(est_time=740, stage="stage-c", runner_config="4-gpu-b200")
 
 
 class TestGptOss4Gpu(BaseTestGptOss):

test/registered/4-gpu-models/test_nvidia_nemotron_3_super_nvfp4.py

@@ -11,7 +11,7 @@
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=710, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=710, stage="stage-c", runner_config="4-gpu-b200")
 
 NEMOTRON_3_SUPER_NVFP4_MODEL = "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4"
 

test/registered/4-gpu-models/test_qwen35_fp4_mtp_v2.py

@@ -15,7 +15,7 @@
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=540, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=540, stage="stage-c", runner_config="4-gpu-b200")
 
 QWEN35_FP4_MODEL = "nvidia/Qwen3.5-397B-A17B-NVFP4"
 ACC_THRESHOLDS = {QWEN35_FP4_MODEL: {"gsm8k": 0.95}}

test/registered/4-gpu-models/test_qwen35_hicache.py

@@ -22,7 +22,7 @@
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=540, suite="stage-c-test-4-gpu-h100")
+register_cuda_ci(est_time=540, stage="stage-c", runner_config="4-gpu-h100")
 
 QWEN35_27B_MODEL = "Qwen/Qwen3.5-27B"
 ACC_THRESHOLDS = {QWEN35_27B_MODEL: {"gsm8k": 0.8}}

test/registered/4-gpu-models/test_qwen35_models.py

@@ -17,7 +17,7 @@
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=260, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(est_time=260, stage="stage-c", runner_config="4-gpu-b200")
 
 QWEN35_FP4_MODEL = "nvidia/Qwen3.5-397B-A17B-NVFP4"
 ACC_THRESHOLDS = {QWEN35_FP4_MODEL: {"gsm8k": 0.95}}

test/registered/4-gpu-models/test_qwen3_30b.py

@@ -11,7 +11,7 @@
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=261, suite="stage-c-test-4-gpu-h100")
+register_cuda_ci(est_time=261, stage="stage-c", runner_config="4-gpu-h100")
 
 QWEN3_30B_MODEL_PATH = "Qwen/Qwen3-30B-A3B-FP8"
 

test/registered/4-gpu-models/test_qwen3_next_models_mtp.py

@@ -6,7 +6,7 @@
 from sglang.test.kits.prefix_cache_branching_kit import PrefixCacheBranchingMixin
 from sglang.test.server_fixtures.default_fixture import DefaultServerBase
 
-register_cuda_ci(est_time=290, suite="stage-c-test-4-gpu-h100")
+register_cuda_ci(est_time=290, stage="stage-c", runner_config="4-gpu-h100")
 
 QWEN3_NEXT_MODEL = "Qwen/Qwen3-Next-80B-A3B-Instruct"
 

test/registered/8-gpu-models/test_deepseek_v32_indexcache.py

@@ -13,7 +13,7 @@
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=492, suite="stage-c-test-8-gpu-h200")
+register_cuda_ci(est_time=492, stage="stage-c", runner_config="8-gpu-h200")
 
 DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
 

test/registered/8-gpu-models/test_deepseek_v3_mtp.py

@@ -17,7 +17,7 @@
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=309, suite="stage-c-test-8-gpu-h200")
+register_cuda_ci(est_time=309, stage="stage-c", runner_config="8-gpu-h200")
 
 FULL_DEEPSEEK_V3_MODEL_PATH = "deepseek-ai/DeepSeek-V3-0324"
 

test/registered/8-gpu-models/test_dsa_models_hisparse.py

@@ -12,7 +12,9 @@
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=720, suite="stage-c-test-8-gpu-h200", nightly=True)
+register_cuda_ci(
+    est_time=720, stage="stage-c", runner_config="8-gpu-h200", nightly=True
+)
 
 GLM5_MODEL_PATH = "zai-org/GLM-5-FP8"
 

test/registered/8-gpu-models/test_dsa_models_mtp.py

@@ -18,7 +18,8 @@
 
 register_cuda_ci(
     est_time=1048,
-    suite="stage-c-test-8-gpu-h200",
+    stage="stage-c",
+    runner_config="8-gpu-h200",
 )
 
 FULL_DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"

test/registered/8-gpu-models/test_mimo_models.py

@@ -6,7 +6,7 @@
 from sglang.test.server_fixtures.default_fixture import DefaultServerBase
 from sglang.test.server_fixtures.mmmu_fixture import MMMUServerBase
 
-register_cuda_ci(est_time=610, suite="stage-c-test-8-gpu-h200")
+register_cuda_ci(est_time=610, stage="stage-c", runner_config="8-gpu-h200")
 
 
 class TestMiMoV2Flash(GSM8KMixin, SpecDecodingMixin, DefaultServerBase):

test/registered/8-gpu-models/test_minimax_m25_basic.py

@@ -14,7 +14,7 @@
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=307, suite="stage-c-test-8-gpu-h200")
+register_cuda_ci(est_time=307, stage="stage-c", runner_config="8-gpu-h200")
 
 MINIMAX_M25_MODEL_PATH = "MiniMaxAI/MiniMax-M2.5"
 

test/registered/8-gpu-models/test_nvidia_nemotron_3_super_bf16.py

@@ -11,7 +11,7 @@
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=376, suite="stage-c-test-8-gpu-h200")
+register_cuda_ci(est_time=376, stage="stage-c", runner_config="8-gpu-h200")
 
 NEMOTRON_3_SUPER_BF16_MODEL = "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16"