fix: increasing H200 test timeout (#15600)

python/sglang/srt/environ.py

@@ -154,6 +154,7 @@ class Envs:
     SGLANG_TEST_STUCK_DP_CONTROLLER = EnvFloat(0)
     SGLANG_TEST_STUCK_TOKENIZER = EnvFloat(0)
     IS_BLACKWELL = EnvBool(False)
+    IS_H200 = EnvBool(False)
     SGLANG_SET_CPU_AFFINITY = EnvBool(False)
     SGLANG_PROFILE_WITH_STACK = EnvBool(True)
     SGLANG_PROFILE_RECORD_SHAPES = EnvBool(True)

python/sglang/test/run_combined_tests.py

@@ -1,3 +1,4 @@
+import time
 from typing import List, Optional
 
 from sglang.test.accuracy_test_runner import (
@@ -115,6 +116,10 @@ def run_combined_tests(
                 all_passed = False
                 model_result["errors"].append(perf_result.error)
 
+            # Wait for GPU memory and port cleanup
+            print("\nWaiting 20 seconds for resource cleanup...")
+            time.sleep(20)
+
         # Run accuracy test
         if run_accuracy:
             acc_result: AccuracyTestResult = run_accuracy_test(
@@ -127,6 +132,10 @@ def run_combined_tests(
                 all_passed = False
                 model_result["errors"].append(acc_result.error)
 
+            # Wait for GPU memory and port cleanup
+            print("\nWaiting 20 seconds for resource cleanup...")
+            time.sleep(20)
+
         all_results.append(model_result)
 
     # Write performance report if we ran perf tests

python/sglang/test/test_utils.py

@@ -168,6 +168,11 @@ def is_blackwell_system():
     return envs.IS_BLACKWELL.get()
 
 
+def is_h200_system():
+    """Return whether it is running on an H200 system."""
+    return envs.IS_H200.get()
+
+
 def _use_cached_default_models(model_repo: str):
     cache_dir = os.getenv("DEFAULT_MODEL_CACHE_DIR")
     if cache_dir and model_repo:
@@ -193,6 +198,9 @@ if is_in_amd_ci():
 if is_blackwell_system():
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH = 3000
 
+if is_h200_system():
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH = 3600
+
 
 def call_generate_lightllm(prompt, temperature, max_tokens, stop=None, url=None):
     assert url is not None