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

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,
         )