[CI] Fold duplicate-server suites and prune the retract matrix on 1-gpu-5090 (#33745)

test/registered/constrained_decoding/test_constrained_decoding.py

@@ -1,20 +1,24 @@
 import unittest
 
+import openai
+
 from sglang.srt.utils import kill_process_tree
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.kits.ebnf_constrained_kit import EBNFConstrainedMixin
 from sglang.test.kits.json_constrained_kit import JSONConstrainedMixin
+from sglang.test.kits.json_mode_kit import JSONModeMixin
 from sglang.test.kits.regex_constrained_kit import RegexConstrainedMixin
 from sglang.test.test_utils import (
     DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
     CustomTestCase,
+    is_in_amd_ci,
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=120, stage="base-b", runner_config="1-gpu-small")
-register_amd_ci(est_time=179, suite="stage-b-test-1-gpu-small-amd")
+register_cuda_ci(est_time=135, stage="base-b", runner_config="1-gpu-small")
+register_amd_ci(est_time=220, suite="stage-b-test-1-gpu-small-amd")
 
 
 class ServerWithGrammar(CustomTestCase):
@@ -35,12 +39,16 @@ class ServerWithGrammar(CustomTestCase):
         if cls.disable_overlap:
             launch_args += ["--disable-overlap-schedule"]
 
+        if is_in_amd_ci():
+            launch_args.append("--constrained-json-disable-any-whitespace")
+
         cls.process = popen_launch_server(
             cls.model,
             cls.base_url,
             timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
             other_args=launch_args,
         )
+        cls.client = openai.Client(api_key="EMPTY", base_url=f"{cls.base_url}/v1")
 
     @classmethod
     def tearDownClass(cls):
@@ -50,19 +58,21 @@ class ServerWithGrammar(CustomTestCase):
 class TestXGrammarBackend(
     ServerWithGrammar,
     JSONConstrainedMixin,
+    JSONModeMixin,
     EBNFConstrainedMixin,
     RegexConstrainedMixin,
 ):
     backend = "xgrammar"
 
 
-class TestOutlinesBackend(ServerWithGrammar, JSONConstrainedMixin):
+class TestOutlinesBackend(ServerWithGrammar, JSONConstrainedMixin, JSONModeMixin):
     backend = "outlines"
 
 
 class TestLLGuidanceBackend(
     ServerWithGrammar,
     JSONConstrainedMixin,
+    JSONModeMixin,
     EBNFConstrainedMixin,
     RegexConstrainedMixin,
 ):