[AMD] rocm 7.2 image release, PR test, Nightly Test (#17799)

Co-authored-by: Alan Kao <akao@amd.com>
Co-authored-by: bingxche <Bingxu.Chen@amd.com>
Co-authored-by: Michael <13900043+michaelzhang-ai@users.noreply.github.com>

python/sglang/test/gpt_oss_common.py

@@ -41,7 +41,8 @@ class BaseTestGptOss(CustomTestCase):
 
         if model_variant == "20b":
             other_args += ["--cuda-graph-max-bs", "600"]
-        if _is_hip:
+        # Respect SGLANG_USE_AITER if already set, otherwise default to "0" for HIP
+        if _is_hip and "SGLANG_USE_AITER" not in os.environ:
             os.environ["SGLANG_USE_AITER"] = "0"
         self._run_test_raw(
             model=model,

python/sglang/test/nightly_utils.py

@@ -228,6 +228,7 @@ class NightlyBenchmarkRunner:
         variant: str = "",
         extra_bench_args: Optional[List[str]] = None,
         enable_profile: bool = True,
+        timeout: Optional[int] = None,
     ) -> Tuple[List[BenchmarkResult], bool, Optional[float]]:
         """Run a complete benchmark for a single model with server management.
 
@@ -247,6 +248,7 @@ class NightlyBenchmarkRunner:
             variant: Optional variant suffix (e.g., "basic", "mtp")
             extra_bench_args: Extra arguments for the benchmark command
             enable_profile: Whether to enable profiling (default True for NVIDIA)
+            timeout: Optional timeout for server launch (defaults to DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH)
 
         Returns:
             Tuple of (list of BenchmarkResult objects, success_bool, avg_spec_accept_length or None)
@@ -260,7 +262,9 @@ class NightlyBenchmarkRunner:
             model=model_path,
             base_url=self.base_url,
             other_args=other_args or [],
-            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            timeout=(
+                timeout if timeout is not None else DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
+            ),
         )
 
         try: