[AMD] skip deterministic inference for MLA FP8 test (#23382)

Co-authored-by: YC Yen-Ching Tseng <yctseng@amd.com>

test/registered/mla/test_mla_fp8.py

@@ -8,6 +8,7 @@ from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
     CustomTestCase,
+    is_in_amd_ci,
     popen_launch_server,
 )
 
@@ -23,10 +24,6 @@ class TestMLA(CustomTestCase, MGSMEnMixin):
     def setUpClass(cls):
         cls.model = DEFAULT_MLA_FP8_MODEL_NAME_FOR_TEST
         cls.base_url = DEFAULT_URL_FOR_TEST
-        cls.process = popen_launch_server(
-            cls.model,
-            cls.base_url,
-            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
         other_args = [
             "--trust-remote-code",
             "--kv-cache-dtype",
@@ -38,8 +35,15 @@ class TestMLA(CustomTestCase, MGSMEnMixin):
             # straddle the 0.8 threshold. With deterministic inference,
             # the score becomes a fixed function of (model, weights, CUDA
             # stack), so threshold-edge flakes stop being random noise.
-                "--enable-deterministic-inference",
-            ],
+        ]
+        if not is_in_amd_ci():
+            # On AMD, the default attention backend (aiter) is not in the deterministic-inference allowlist, so the server fails to start, disable it.
+            other_args.append("--enable-deterministic-inference")
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=other_args,
         )
 
     @classmethod