[AMD] CI - enable deepseekv3.2 on MI325-8gpu and merge perf/accuracy test suites into stage-b suites (#17633)

Co-authored-by: Bingxu Chen <Bingxu.Chen@amd.com>

test/registered/amd/test_deepseek_v32_basic.py

@@ -15,11 +15,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_amd_ci(
-    est_time=3600,
-    suite="stage-c-test-large-8-gpu-amd-mi35x",
-    disabled="move to nightly for saving time",
-)
+register_amd_ci(est_time=1800, suite="stage-c-test-large-8-gpu-amd")
 
 DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
 
@@ -159,7 +155,7 @@ class TestDeepseekV32TP(CustomTestCase):
                 f"### test_bs_1_speed (deepseek-v32)\n" f"{speed=:.2f} token/s\n"
             )
             if is_in_amd_ci():
-                self.assertGreater(speed, 20)
+                self.assertGreater(speed, 15)
             else:
                 self.assertGreater(speed, 70)
 

test/registered/amd/test_deepseek_v32_mtp.py

@@ -17,7 +17,12 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_amd_ci(est_time=3600, suite="stage-c-test-large-8-gpu-amd-mi35x")
+register_amd_ci(
+    est_time=3600,
+    suite="stage-c-test-large-8-gpu-amd-mi35x",
+    disabled="move to nightly for saving time",
+)
+
 FULL_DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
 
 

test/registered/amd/test_kimi_k2_instruct.py

@@ -11,12 +11,13 @@ from sglang.test.send_one import BenchArgs, send_one_prompt
 from sglang.test.test_utils import (
     DEFAULT_URL_FOR_TEST,
     CustomTestCase,
+    is_in_amd_ci,
     is_in_ci,
     popen_launch_server,
     write_github_step_summary,
 )
 
-register_amd_ci(est_time=3600, suite="stage-c-test-large-8-gpu-amd-mi35x")
+register_amd_ci(est_time=3600, suite="stage-c-test-large-8-gpu-amd")
 
 KIMI_K2_MODEL_PATH = "moonshotai/Kimi-K2-Instruct-0905"
 SERVER_LAUNCH_TIMEOUT = 3600
@@ -88,7 +89,10 @@ class TestKimiK2Instruct0905(CustomTestCase):
                 f"### test_bs_1_speed (Kimi-K2-Instruct-0905)\n"
                 f"{speed=:.2f} token/s\n"
             )
-            self.assertGreater(speed, 45)
+            if is_in_amd_ci():
+                self.assertGreater(speed, 30)
+            else:
+                self.assertGreater(speed, 45)
 
 
 if __name__ == "__main__":

test/registered/attention/test_triton_sliding_window.py

@@ -10,6 +10,7 @@ from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
     CustomTestCase,
+    is_in_amd_ci,
     is_in_ci,
     popen_launch_server,
 )
@@ -62,7 +63,10 @@ class TestSlidingWindowAttentionTriton(CustomTestCase):
         metrics = run_eval(args)
         print(f"MMLU metrics with sliding window: {metrics}")
 
-        self.assertGreaterEqual(metrics["score"], 0.60)
+        if is_in_amd_ci():
+            self.assertGreaterEqual(metrics["score"], 0.55)
+        else:
+            self.assertGreaterEqual(metrics["score"], 0.60)
 
     def _test_short_context_generation(self):
         response = requests.post(

test/registered/eval/test_eval_accuracy_large.py

@@ -14,6 +14,7 @@ from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
     CustomTestCase,
+    is_in_amd_ci,
     is_in_ci,
     popen_launch_server,
     write_github_step_summary,
@@ -71,7 +72,10 @@ class TestEvalAccuracyLarge(CustomTestCase):
                 f"### test_human_eval\n" f'{metrics["score"]=:.4f}\n'
             )
 
-        self.assertGreater(metrics["score"], 0.64)
+        if is_in_amd_ci():
+            self.assertGreater(metrics["score"], 0.60)
+        else:
+            self.assertGreater(metrics["score"], 0.64)
 
     def test_mgsm_en(self):
         args = SimpleNamespace(

test/registered/eval/test_moe_eval_accuracy_large.py

@@ -3,6 +3,7 @@ Usage:
 python -m unittest test_moe_eval_accuracy_large.TestMoEEvalAccuracyLarge.test_mmlu
 """
 
+import os
 import unittest
 from types import SimpleNamespace
 
@@ -14,6 +15,7 @@ from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
     CustomTestCase,
+    is_in_amd_ci,
     is_in_ci,
     popen_launch_server,
     write_github_step_summary,
@@ -28,6 +30,15 @@ class TestMoEEvalAccuracyLarge(CustomTestCase):
     def setUpClass(cls):
         cls.model = DEFAULT_MOE_MODEL_NAME_FOR_TEST
         cls.base_url = DEFAULT_URL_FOR_TEST
+
+        # Disable AITER for AMD CI to ensure consistent results
+        env = None
+        if is_in_amd_ci():
+            env = os.environ.copy()
+            env["SGLANG_USE_AITER"] = "0"
+            env["SGLANG_USE_AITER_AR"] = "0"
+            env["HF_HUB_ENABLE_HF_TRANSFER"] = "0"
+
         cls.process = popen_launch_server(
             cls.model,
             cls.base_url,
@@ -38,6 +49,7 @@ class TestMoEEvalAccuracyLarge(CustomTestCase):
                 "--tp",
                 "2",
             ],
+            env=env,
         )
 
     @classmethod

test/run_suite.py

@@ -24,11 +24,6 @@ PER_COMMIT_SUITES = {
         "stage-b-test-small-1-gpu-amd-mi35x",
         "stage-b-test-large-1-gpu-amd",
         "stage-b-test-large-2-gpu-amd",
-        "stage-b-test-small-1-gpu-performance-amd",
-        "stage-b-test-large-1-gpu-performance-amd",
-        "stage-b-test-large-2-gpu-performance-amd",
-        "stage-b-test-small-1-gpu-accuracy-amd",
-        "stage-b-test-large-2-gpu-accuracy-amd",
         "stage-c-test-large-8-gpu-amd-mi35x",
     ],
     HWBackend.CUDA: [