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

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)