[CI] Share VLM engines and prune launch matrices on the per-commit H100/H200 suites (#33944)

python/sglang/test/server_fixtures/streaming_session_fixture.py

@@ -51,6 +51,15 @@ LEAK_FILLER = (
     "We promptly judged antique ivory buckles for the next prize. "
 ) * 20
 
+SWA_MODEL = "openai/gpt-oss-20b"
+
+# Common gpt-oss-20b launch args. Matches TestSessionLatency/TestSWARadixCacheKL.
+SWA_COMMON_ARGS = [
+    "--mem-fraction-static",
+    "0.70",
+    "--cuda-graph-backend-prefill=disabled",
+]
+
 ABORT_REPRO_CONTEXT_LEN = 512
 ABORT_REPRO_PAGE_SIZE = 256
 ABORT_REPRO_GEN_LEN = 4

test/registered/cuda_graph/breakable/test_breakable_cuda_graph.py

@@ -24,7 +24,7 @@ from sglang.test.test_utils import (
 )
 
 # CI Registration — large suite to fit the integration test's server startup.
-register_cuda_ci(est_time=79, stage="base-b", runner_config="1-gpu-large")
+register_cuda_ci(est_time=60, stage="base-b", runner_config="1-gpu-large")
 register_amd_ci(est_time=200, suite="stage-c-test-large-8-gpu-amd-mi35x")
 
 
@@ -388,7 +388,7 @@ class TestBreakableCudaGraph(CustomTestCase):
             base_url=self.base_url,
             model=self.model,
             eval_name="mgsm_en",
-            num_examples=1319,
+            num_examples=200,
             num_threads=1024,
         )
 

test/registered/lora/test_lora_overlap_loading.py

@@ -30,7 +30,7 @@ from sglang.test.lora_utils import (
 )
 from sglang.test.test_utils import CustomTestCase
 
-register_cuda_ci(est_time=48, stage="base-b", runner_config="1-gpu-large")
+register_cuda_ci(est_time=380, stage="base-b", runner_config="1-gpu-large")
 register_amd_ci(est_time=75, suite="stage-b-test-1-gpu-small-amd")
 
 

test/registered/mem_cache/test_post_capture_kv_sizing.py

@@ -28,7 +28,7 @@ from sglang.test.test_utils import (
 )
 
 # CI Registration
-register_cuda_ci(est_time=240, stage="base-b", runner_config="1-gpu-large")
+register_cuda_ci(est_time=100, stage="base-b", runner_config="1-gpu-large")
 
 STDOUT_FILENAME = "post_capture_kv_sizing_stdout.log"
 STDERR_FILENAME = "post_capture_kv_sizing_stderr.log"
@@ -88,7 +88,7 @@ class TestPostCaptureKVSizing(CustomTestCase):
             base_url=self.base_url,
             model=self.model,
             eval_name="gsm8k",
-            num_examples=500,
+            num_examples=200,
             num_threads=1024,
         )
         metrics = run_eval(args)

test/registered/models_e2e/test_inkling.py

@@ -31,7 +31,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=600, stage="base-b", runner_config="1-gpu-large")
+register_cuda_ci(est_time=250, stage="base-b", runner_config="1-gpu-large")
 
 # Defaults to the HF `test` revision; override MODEL/REVISION to point at a
 # local checkpoint. Empty REVISION drops the flag (for local paths).

test/registered/models_e2e/test_mimo_v2_flash.py

@@ -7,13 +7,11 @@ from sglang.test.kits.eval_accuracy_kit import GSM8KMixin
 from sglang.test.kits.spec_decoding_kit import SpecDecodingMixin
 from sglang.test.server_fixtures.default_fixture import DefaultServerBase
 
-register_cuda_ci(est_time=350, stage="base-c", runner_config="8-gpu-h200")
+register_cuda_ci(est_time=200, stage="base-c", runner_config="8-gpu-h200")
 
 
 class TestMiMoV2Flash(GSM8KMixin, SpecDecodingMixin, DefaultServerBase):
     gsm8k_accuracy_thres = 0.75
-    gsm8k_num_questions = 1319
-    gsm8k_num_threads = 1319
     model = "XiaomiMiMo/MiMo-V2-Flash"
 
     other_args = [

test/registered/models_e2e/test_minimax_m25_basic.py

@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=250, stage="base-c", runner_config="8-gpu-h200")
+register_cuda_ci(est_time=160, stage="base-c", runner_config="8-gpu-h200")
 
 MINIMAX_M25_MODEL_PATH = "MiniMaxAI/MiniMax-M2.5"
 
@@ -53,8 +53,8 @@ class TestMiniMaxM25Basic(CustomTestCase):
         args = SimpleNamespace(
             num_shots=20,
             data_path=None,
-            num_questions=1400,
-            parallel=1400,
+            num_questions=200,
+            parallel=200,
             max_new_tokens=512,
             host="http://127.0.0.1",
             port=int(self.base_url.split(":")[-1]),

test/registered/models_e2e/test_qwen3_next_models_mtp.py

@@ -7,7 +7,7 @@ from sglang.test.kits.kl_divergence_kit import KLDivergenceMixin
 from sglang.test.kits.prefix_cache_branching_kit import PrefixCacheBranchingMixin
 from sglang.test.server_fixtures.default_fixture import DefaultServerBase
 
-register_cuda_ci(est_time=430, stage="base-c", runner_config="4-gpu-h100")
+register_cuda_ci(est_time=290, stage="base-c", runner_config="4-gpu-h100")
 
 QWEN3_NEXT_MODEL = "Qwen/Qwen3-Next-80B-A3B-Instruct"
 
@@ -69,6 +69,11 @@ class TestQwen3NextMTPTopk(
     ]
 
 
+@unittest.skip(
+    "Manual-only: topk == 1 is covered by TestQwen3NextMTPLazyV2 and extra_buffer "
+    "by TestQwen3NextMTPTopk. Kept runnable locally for the plain (topk=1, "
+    "extra_buffer) baseline."
+)
 class TestQwen3NextMTPV2(GSM8KMixin, KLDivergenceMixin, DefaultServerBase):
     model = QWEN3_NEXT_MODEL
     gsm8k_accuracy_thres = 0.93

test/registered/sessions/test_streaming_session_swa.py

@@ -1,3 +1,9 @@
+"""Per-commit streaming-session tests on a hybrid-SWA model.
+
+Baseline + large-page retract + abort-leak repro stay per-commit; the
+mixed-chunk retract variant lives in test_streaming_session_swa_extra.py.
+"""
+
 import unittest
 
 from sglang.test.ci.ci_register import register_cuda_ci
@@ -9,20 +15,12 @@ from sglang.test.server_fixtures.streaming_session_fixture import (
     ABORT_REPRO_CHUNKED_PREFILL_SIZE,
     ABORT_REPRO_CONTEXT_LEN,
     ABORT_REPRO_PAGE_SIZE,
+    SWA_COMMON_ARGS,
+    SWA_MODEL,
     StreamingSessionServerBase,
 )
 
-register_cuda_ci(est_time=519, stage="base-b", runner_config="1-gpu-large")
-
-
-SWA_MODEL = "openai/gpt-oss-20b"
-
-# Common gpt-oss-20b launch args. Matches TestSessionLatency/TestSWARadixCacheKL.
-SWA_COMMON_ARGS = [
-    "--mem-fraction-static",
-    "0.70",
-    "--cuda-graph-backend-prefill=disabled",
-]
+register_cuda_ci(est_time=390, stage="base-b", runner_config="1-gpu-large")
 
 
 class TestStreamingSessionSWA(StreamingSessionServerBase, StreamingSessionKitMixin):
@@ -48,21 +46,6 @@ class TestStreamingSessionSWARetractLargePage(
     env_overrides = [("SGLANG_TEST_RETRACT", True)]
 
 
-class TestStreamingSessionSWARetractMixedChunk(
-    StreamingSessionServerBase, StreamingSessionKitMixin
-):
-    """SWA under retract decode with --enable-mixed-chunk."""
-
-    model = SWA_MODEL
-    extra_args = [
-        "--chunked-prefill-size",
-        "128",
-        "--enable-mixed-chunk",
-        *SWA_COMMON_ARGS,
-    ]
-    env_overrides = [("SGLANG_TEST_RETRACT", True)]
-
-
 class TestStreamingSessionSWAAbortLeakRepro(
     StreamingSessionServerBase, AbortLeakReproKitMixin
 ):

test/registered/sessions/test_streaming_session_swa_extra.py

@@ -0,0 +1,35 @@
+"""Label-gated SWA streaming-session variants.
+
+CUDA-only: gpt-oss-20b is not part of the AMD streaming-session coverage.
+"""
+
+import unittest
+
+from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.kits.streaming_session_kit import StreamingSessionKitMixin
+from sglang.test.server_fixtures.streaming_session_fixture import (
+    SWA_COMMON_ARGS,
+    SWA_MODEL,
+    StreamingSessionServerBase,
+)
+
+register_cuda_ci(est_time=130, stage="extra-a", runner_config="1-gpu-large")
+
+
+class TestStreamingSessionSWARetractMixedChunk(
+    StreamingSessionServerBase, StreamingSessionKitMixin
+):
+    """SWA under retract decode with --enable-mixed-chunk."""
+
+    model = SWA_MODEL
+    extra_args = [
+        "--chunked-prefill-size",
+        "128",
+        "--enable-mixed-chunk",
+        *SWA_COMMON_ARGS,
+    ]
+    env_overrides = [("SGLANG_TEST_RETRACT", True)]
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/vlm/test_vision_openai_server_a.py

@@ -20,7 +20,7 @@ from sglang.test.vlm_utils import (
     terminate_and_kill_process_tree,
 )
 
-register_cuda_ci(est_time=780, stage="base-b", runner_config="1-gpu-large")
+register_cuda_ci(est_time=560, stage="base-b", runner_config="1-gpu-large")
 
 
 class TestLlavaServer(ImageOpenAITestMixin):
@@ -43,17 +43,10 @@ class TestQwen3VLServer(ImageOpenAITestMixin, VideoOpenAITestMixin):
     extra_args = ["--cuda-graph-max-bs-decode=4"]
 
 
-class TestQwen3OmniServer(OmniOpenAITestMixin):
-    model = "Qwen/Qwen3-Omni-30B-A3B-Instruct"
-    extra_args = [  # workaround to fit into H100
-        "--mem-fraction-static=0.90",
-        "--disable-cuda-graph",
-        "--disable-fast-image-processor",
-        "--grammar-backend=none",
-    ]
-
-
 class TestQwen2VLContextLengthServer(CustomTestCase):
+    # --context-length 300 is calibrated to this model's mm-token expansion:
+    # it must sit above the warmup image's expanded length but below the test
+    # image's. A cheaper VLM needs the bound recalibrated, not just swapped.
     @classmethod
     def setUpClass(cls):
         cls.model = "Qwen/Qwen2-VL-7B-Instruct"

test/registered/vlm/test_vision_openai_server_extra.py

@@ -0,0 +1,27 @@
+"""Label-gated vision/omni server launches too expensive for the per-commit
+budget; the per-commit set lives in test_vision_openai_server_a.py."""
+
+import unittest
+
+from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.vlm_utils import OmniOpenAITestMixin
+
+register_cuda_ci(est_time=180, stage="extra-a", runner_config="1-gpu-large")
+
+
+class TestQwen3OmniServer(OmniOpenAITestMixin):
+    model = "Qwen/Qwen3-Omni-30B-A3B-Instruct"
+    extra_args = [  # workaround to fit into H100
+        "--mem-fraction-static=0.90",
+        "--disable-cuda-graph",
+        "--disable-fast-image-processor",
+        "--grammar-backend=none",
+    ]
+
+
+# Delete the mixin so it is not collected as a test case in its own right.
+del OmniOpenAITestMixin
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/vlm/test_vlm_input_format.py

@@ -37,8 +37,9 @@ from sglang.srt.entrypoints.openai.protocol import ChatCompletionRequest
 from sglang.srt.parser.conversation import generate_chat_conv
 from sglang.srt.utils.common import is_cuda, is_xpu
 from sglang.srt.utils.hf_transformers_utils import _fix_added_tokens_encoding
+from sglang.test.test_utils import CustomTestCase
 
-register_cuda_ci(est_time=747, stage="base-b", runner_config="1-gpu-large")
+register_cuda_ci(est_time=300, stage="base-b", runner_config="1-gpu-large")
 
 IMAGE_MAN_IRONING_URL = "https://raw.githubusercontent.com/sgl-project/sgl-test-files/refs/heads/main/images/man_ironing_on_back_of_suv.png"
 IMAGE_SGL_LOGO_URL = "https://raw.githubusercontent.com/sgl-project/sgl-test-files/refs/heads/main/images/sgl_logo.png"
@@ -52,6 +53,7 @@ class VLMInputTestBase:
     chat_template = None
     processor = None
     visual = None  # Should be a callable for precomputed embeddings
+    engine = None
 
     @classmethod
     def setUpClass(cls):
@@ -76,25 +78,34 @@ class VLMInputTestBase:
         )
         _fix_added_tokens_encoding(cls.processor.tokenizer)
         cls._init_visual()
+        cls._start_engine()
 
     @classmethod
     def _init_visual(cls):
         """Override in subclass to set up cls.visual as a callable for precomputed embeddings."""
         raise NotImplementedError
 
-    def setUp(self):
-        self.engine = Engine(
-            model_path=self.model_path,
-            chat_template=self.chat_template,
-            device=self.device.type,
+    @classmethod
+    def _start_engine(cls):
+        # One engine per class: every test only reads through it. The tests are
+        # sync rather than async because the tokenizer manager pins handle_loop
+        # to the loop of the first request, so a per-test loop would strand it.
+        cls.engine = Engine(
+            model_path=cls.model_path,
+            chat_template=cls.chat_template,
+            device=cls.device.type,
             mem_fraction_static=0.8,
             enable_multimodal=True,
             disable_cuda_graph=True,
             trust_remote_code=True,
         )
 
-    def tearDown(self):
-        self.engine.shutdown()
+    @classmethod
+    def tearDownClass(cls):
+        # CustomTestCase runs tearDownClass even when setUpClass raised.
+        if cls.engine is not None:
+            cls.engine.shutdown()
+            cls.engine = None
 
     def verify_response(self, output):
         # The goal is to check that the model roughly understands:
@@ -157,25 +168,25 @@ class VLMInputTestBase:
 
         return inputs, text
 
-    async def test_accepts_image(self):
+    def test_accepts_image(self):
         req = self.get_completion_request()
         conv = generate_chat_conv(req, template_name=self.chat_template)
         text = conv.get_prompt()
-        output = await self.engine.async_generate(
+        output = self.engine.generate(
             prompt=text,
             image_data=self.main_image,
             sampling_params=dict(temperature=0.0, max_new_tokens=512),
         )
         self.verify_response(output)
 
-    async def test_accepts_precomputed_embeddings(self):
+    def test_accepts_precomputed_embeddings(self):
         req = self.get_completion_request()
         processor_output, _ = self.get_processor_output(req=req)
 
         with torch.inference_mode():
             precomputed_embeddings = self.__class__.visual(processor_output)
 
-        output = await self.engine.async_generate(
+        output = self.engine.generate(
             input_ids=processor_output["input_ids"][0].detach().cpu().tolist(),
             image_data=[
                 self._precomputed_image_data(processor_output, precomputed_embeddings)
@@ -184,10 +195,10 @@ class VLMInputTestBase:
         )
         self.verify_response(output)
 
-    async def test_accepts_processor_output(self):
+    def test_accepts_processor_output(self):
         req = self.get_completion_request()
         processor_output, prompt = self.get_processor_output(req=req)
-        output = await self.engine.async_generate(
+        output = self.engine.generate(
             input_ids=processor_output["input_ids"][0].detach().cpu().tolist(),
             image_data=[self._processor_output_image_data(processor_output)],
             sampling_params=dict(temperature=0.0, max_new_tokens=512),
@@ -207,7 +218,7 @@ class VLMInputTestBase:
         raise NotImplementedError
 
 
-class TestQwenVLUnderstandsImage(VLMInputTestBase, unittest.IsolatedAsyncioTestCase):
+class TestQwenVLUnderstandsImage(VLMInputTestBase, CustomTestCase):
     model_path = "Qwen/Qwen2.5-VL-3B-Instruct"
     chat_template = "qwen2-vl"
 
@@ -234,7 +245,7 @@ class TestQwenVLUnderstandsImage(VLMInputTestBase, unittest.IsolatedAsyncioTestC
         return dict(processor_output, format="processor_output")
 
 
-class TestGemmaUnderstandsImage(VLMInputTestBase, unittest.IsolatedAsyncioTestCase):
+class TestGemmaUnderstandsImage(VLMInputTestBase, CustomTestCase):
     model_path = "google/gemma-3-4b-it"
     chat_template = "gemma-it"
 
@@ -263,9 +274,7 @@ class TestGemmaUnderstandsImage(VLMInputTestBase, unittest.IsolatedAsyncioTestCa
 
 
 # Updated Kimi-VL test to use the new input format.
-class TestKimiVLImageUnderstandsImage(
-    VLMInputTestBase, unittest.IsolatedAsyncioTestCase
-):
+class TestKimiVLImageUnderstandsImage(VLMInputTestBase, CustomTestCase):
     model_path = "moonshotai/Kimi-VL-A3B-Instruct"
     chat_template = "kimi-vl"
 
@@ -322,7 +331,7 @@ class TestKimiVLImageUnderstandsImage(
 
 # not for CI: too large
 # class TestLlama4ImageUnderstandsImage(
-#     VLMInputTestBase, unittest.IsolatedAsyncioTestCase
+#     VLMInputTestBase, CustomTestCase
 # ):
 #     # Allow overriding via env for local/offline runs.
 #     model_path = "meta-llama/Llama-4-Scout-17B-16E-Instruct"
@@ -364,7 +373,7 @@ class TestKimiVLImageUnderstandsImage(
 #         return dict(processor_output, format="processor_output")
 
 
-# class TestLlavaUnderstandsImage(VLMInputTestBase, unittest.IsolatedAsyncioTestCase):
+# class TestLlavaUnderstandsImage(VLMInputTestBase, CustomTestCase):
 #     model_path = "llava-hf/llava-1.5-7b-hf"
 #     chat_template = "vicuna_v1.1"
 
@@ -403,7 +412,7 @@ class TestKimiVLImageUnderstandsImage(
 #         return dict(processor_output, format="processor_output")
 
 
-class TestInternVLUnderstandsImage(VLMInputTestBase, unittest.IsolatedAsyncioTestCase):
+class TestInternVLUnderstandsImage(VLMInputTestBase, CustomTestCase):
     model_path = "OpenGVLab/InternVL2-2B"
     chat_template = "internvl-2-5"
 
@@ -427,6 +436,7 @@ class TestInternVLUnderstandsImage(VLMInputTestBase, unittest.IsolatedAsyncioTes
             cls.model_path, trust_remote_code=True
         )
         cls._init_visual()
+        cls._start_engine()
 
     @classmethod
     def _init_visual(cls):
@@ -591,7 +601,7 @@ class TestInternVLUnderstandsImage(VLMInputTestBase, unittest.IsolatedAsyncioTes
 
 
 @unittest.skip("temporarily disabled: NaN in next_token_logits")
-class TestMiniCPMVUnderstandsImage(VLMInputTestBase, unittest.IsolatedAsyncioTestCase):
+class TestMiniCPMVUnderstandsImage(VLMInputTestBase, CustomTestCase):
     model_path = "openbmb/MiniCPM-V-4"
     chat_template = "minicpmv"
 
@@ -618,6 +628,7 @@ class TestMiniCPMVUnderstandsImage(VLMInputTestBase, unittest.IsolatedAsyncioTes
         cls.processor.tokenizer = get_tokenizer(cls.model_path, trust_remote_code=True)
         _fix_added_tokens_encoding(cls.processor.tokenizer)
         cls._init_visual()
+        cls._start_engine()
 
     @classmethod
     def _init_visual(cls):