[2/N] CI refactor: sperate some backend-independent CPU tasks. (#13447)

.github/workflows/pr-test.yml

@@ -394,7 +394,9 @@ jobs:
         timeout-minutes: 10
         run: |
           cd test/
-          python3 run_suite.py
+          python3 run_suite.py --hw cuda --suite stage-a-test-1
+          # temporarily put backend-independent cpu tests here
+          python3 run_suite.py --hw cpu --suite default
 
   unit-test-backend-1-gpu:
     needs: [check-changes, stage-a-test-1, sgl-kernel-build-wheels]

python/sglang/test/ci/ci_register.py

@@ -6,6 +6,7 @@ from typing import List
 
 
 class HWBackend(Enum):
+    CPU = auto()
     CUDA = auto()
     AMD = auto()
 
@@ -14,19 +15,24 @@ class HWBackend(Enum):
 class CIRegistry:
     backend: HWBackend
     filename: str
-    estimation_time: float
+    est_time: float
-    stage: str
+    suite: str
 
 
-def register_cuda_ci(estimation_time: float, ci_stage: str):
+def register_cpu_ci(est_time: float, suite: str):
     pass
 
 
-def register_amd_ci(estimation_time: float, ci_stage: str):
+def register_cuda_ci(est_time: float, suite: str):
+    pass
+
+
+def register_amd_ci(est_time: float, suite: str):
     pass
 
 
 REGISTER_MAPPING = {
+    "register_cpu_ci": HWBackend.CPU,
     "register_cuda_ci": HWBackend.CUDA,
     "register_amd_ci": HWBackend.AMD,
 }
@@ -45,28 +51,27 @@ class RegistryVisitor(ast.NodeVisitor):
             return None
 
         hw = REGISTER_MAPPING[func_call.func.id]
-        est_time = None
+        est_time, suite = None, None
-        ci_stage = None
         for kw in func_call.keywords:
-            if kw.arg == "estimation_time":
+            if kw.arg == "est_time":
                 if isinstance(kw.value, ast.Constant):
                     est_time = kw.value.value
-            elif kw.arg == "ci_stage":
+            elif kw.arg == "suite":
                 if isinstance(kw.value, ast.Constant):
-                    ci_stage = kw.value.value
+                    suite = kw.value.value
 
         for i, arg in enumerate(func_call.args):
             if isinstance(arg, ast.Constant):
                 if i == 0:
                     est_time = arg.value
                 elif i == 1:
-                    ci_stage = arg.value
+                    suite = arg.value
         assert (
             est_time is not None
         ), "esimation_time is required and should be a constant"
-        assert ci_stage is not None, "ci_stage is required and should be a constant"
+        assert suite is not None, "suite is required and should be a constant"
         return CIRegistry(
-            backend=hw, filename=self.filename, estimation_time=est_time, stage=ci_stage
+            backend=hw, filename=self.filename, est_time=est_time, suite=suite
         )
 
     def visit_Module(self, node):

test/per_commit/function_call/test_function_call_parser.py

@@ -12,6 +12,9 @@ from sglang.srt.function_call.llama32_detector import Llama32Detector
 from sglang.srt.function_call.mistral_detector import MistralDetector
 from sglang.srt.function_call.pythonic_detector import PythonicDetector
 from sglang.srt.function_call.qwen3_coder_detector import Qwen3CoderDetector
+from sglang.test.ci.ci_register import register_cpu_ci
+
+register_cpu_ci(1.0, "default")
 
 
 class TestPythonicDetector(unittest.TestCase):

test/per_commit/function_call/test_json_schema_constraint.py

@@ -16,6 +16,9 @@ from sglang.srt.function_call.utils import (
     _get_tool_schema_defs,
     get_json_schema_constraint,
 )
+from sglang.test.ci.ci_register import register_cpu_ci
+
+register_cpu_ci(1.0, "default")
 
 
 class TestJsonSchemaConstraint(unittest.TestCase):

test/per_commit/function_call/test_unknown_tool_name.py

@@ -1,10 +1,15 @@
 import json
 import logging
 
+import pytest
+
 from sglang.srt.entrypoints.openai.protocol import Function, Tool
 from sglang.srt.environ import envs
 from sglang.srt.function_call.base_format_detector import BaseFormatDetector
 from sglang.srt.function_call.core_types import StreamingParseResult
+from sglang.test.ci.ci_register import register_cpu_ci
+
+register_cpu_ci(1.0, "default")
 
 
 class DummyDetector(BaseFormatDetector):
@@ -17,6 +22,9 @@ class DummyDetector(BaseFormatDetector):
             normal_text="", calls=self.parse_base_json(action, tools)
         )
 
+    def structure_info(self):
+        pass
+
 
 def test_unknown_tool_name_dropped_default(caplog):
     """Test that unknown tools are dropped by default (legacy behavior)."""
@@ -67,3 +75,7 @@ def test_unknown_tool_name_forwarded(caplog):
         assert result.calls[0].name == "unknown_tool"
         assert result.calls[0].tool_index == -1
         assert json.loads(result.calls[0].parameters)["city"] == "Paris"
+
+
+if __name__ == "__main__":
+    pytest.main([__file__])

test/per_commit/test_srt_backend.py

@@ -19,7 +19,7 @@ from sglang.test.test_programs import (
 )
 from sglang.test.test_utils import DEFAULT_MODEL_NAME_FOR_TEST, CustomTestCase
 
-register_cuda_ci(estimation_time=80, ci_stage="stage-a-test-1")
+register_cuda_ci(est_time=80, suite="stage-a-test-1")
 
 
 class TestSRTBackend(CustomTestCase):

test/run_suite.py

@@ -1,10 +1,19 @@
+import argparse
 import glob
 from typing import List
 
 from sglang.test.ci.ci_register import CIRegistry, HWBackend, collect_tests
 from sglang.test.ci.ci_utils import TestFile, run_unittest_files
 
-LABEL_MAPPING = {HWBackend.CUDA: ["stage-a-test-1"]}
+HW_MAPPING = {
+    "cpu": HWBackend.CPU,
+    "cuda": HWBackend.CUDA,
+}
+
+LABEL_MAPPING = {
+    HWBackend.CUDA: ["stage-a-test-1"],
+    HWBackend.CPU: ["default"],
+}
 
 
 def _filter_tests(
@@ -13,8 +22,8 @@ def _filter_tests(
     ci_tests = [t for t in ci_tests if t.backend == hw]
     ret = []
     for t in ci_tests:
-        assert t.stage in LABEL_MAPPING[hw], f"Unknown stage {t.stage} for backend {hw}"
+        assert t.suite in LABEL_MAPPING[hw], f"Unknown stage {t.suite} for backend {hw}"
-        if t.stage == suite:
+        if t.suite == suite:
             ret.append(t)
     return ret
 
@@ -22,7 +31,7 @@ def _filter_tests(
 def run_per_commit(hw: HWBackend, suite: str):
     files = glob.glob("per_commit/**/*.py", recursive=True)
     ci_tests = _filter_tests(collect_tests(files), hw, suite)
-    test_files = [TestFile(t.filename, t.estimation_time) for t in ci_tests]
+    test_files = [TestFile(t.filename, t.est_time) for t in ci_tests]
 
     run_unittest_files(
         test_files,
@@ -32,7 +41,23 @@ def run_per_commit(hw: HWBackend, suite: str):
 
 
 def main():
-    run_per_commit(HWBackend.CUDA, "stage-a-test-1")
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--hw",
+        type=str,
+        choices=["cpu", "cuda"],
+        required=True,
+        help="Hardware backend to run tests on.",
+    )
+    parser.add_argument(
+        "--suite",
+        type=str,
+        required=True,
+        help="Test suite to run.",
+    )
+    args = parser.parse_args()
+    hw = HW_MAPPING[args.hw]
+    run_per_commit(hw, args.suite)
 
 
 if __name__ == "__main__":

test/srt/run_suite.py

@@ -8,7 +8,6 @@ from sglang.test.ci.ci_utils import TestFile, run_unittest_files
 suites = {
     "per-commit-1-gpu": [
         TestFile("debug_utils/test_tensor_dump_forward_hook.py", 15),
-        TestFile("function_call/test_json_schema_constraint.py", 1),
         TestFile("hicache/test_hicache_storage.py", 127),
         TestFile("hicache/test_hicache_variants.py", 393),
         TestFile("layers/attention/mamba/test_causal_conv1d.py", 25),
@@ -70,7 +69,6 @@ suites = {
         TestFile("test_flashmla.py", 230),
         TestFile("test_fp8_utils.py", 5),
         TestFile("rotary_embedding/test_mrope.py", 10),
-        TestFile("test_function_call_parser.py", 10),
         TestFile("test_fused_moe.py", 80),
         TestFile("test_gpt_oss_1gpu.py", 750),
         TestFile("test_harmony_parser.py", 20),
@@ -239,7 +237,6 @@ suites = {
         TestFile("ep/test_hybrid_dp_ep_tp_mtp.py"),
         TestFile("ep/test_moe_deepep.py"),
         TestFile("ep/test_moe_deepep_eval_accuracy_large.py"),
-        TestFile("function_call/test_unknown_tool_name.py"),
         TestFile("hicache/test_disaggregation_hicache.py"),
         TestFile("hicache/test_hicache_storage_benchmark.py"),
         TestFile("hicache/test_hicache_storage_e2e.py"),
@@ -381,7 +378,6 @@ suites = {
 # NOTE: please sort the test cases alphabetically by the test file name
 suite_amd = {
     "per-commit-amd": [
-        TestFile("function_call/test_json_schema_constraint.py", 1),
         # TestFile("hicache/test_hicache.py", 116), # Disabled temporarily, see https://github.com/sgl-project/sglang/issues/12575
         # TestFile("hicache/test_hicache_mla.py", 127), # Disabled temporarily,  # Temporarily disabled, see https://github.com/sgl-project/sglang/issues/12574
         # TestFile("hicache/test_hicache_storage.py", 127), # Disabled temporarily, see https://github.com/sgl-project/sglang/issues/12575
@@ -421,7 +417,6 @@ suite_amd = {
         TestFile("test_chunked_prefill.py", 410),
         TestFile("test_create_kvindices.py", 2),
         TestFile("test_eval_fp8_accuracy.py", 303),
-        TestFile("test_function_call_parser.py", 10),
         TestFile("test_fused_moe.py", 30),
         TestFile("test_harmony_parser.py", 20),
         TestFile("test_input_embeddings.py", 38),