aiter attention-backend (default enabled on AMD/ROCm) (#6381)

test/srt/models/test_dummy_grok_models.py

@@ -4,6 +4,7 @@ from sglang.test.test_utils import CustomTestCase, is_in_ci, run_bench_one_batch
 
 
 class TestDummyGrok1(CustomTestCase):
+
     def test_dummy_grok_1(self):
         output_throughput = run_bench_one_batch(
             None,

test/srt/test_eval_accuracy_large.py

@@ -3,6 +3,8 @@ Usage:
 python -m unittest test_eval_accuracy_large.TestEvalAccuracyLarge.test_mmlu
 """
 
+import os
+import time
 import unittest
 from types import SimpleNamespace
 
@@ -35,6 +37,11 @@ class TestEvalAccuracyLarge(CustomTestCase):
     def tearDownClass(cls):
         kill_process_tree(cls.process.pid)
 
+    def tearDown(self):
+        # Delay between tests to allow GPU memory cleanup
+        if os.getenv("SGLANG_AMD_CI") == "1":
+            time.sleep(180)
+
     def test_mmlu(self):
         args = SimpleNamespace(
             base_url=self.base_url,