Fix 2-gpu CI test and suppress some warnings (#4930)

python/sglang/srt/models/deepseek_v2.py

@@ -30,9 +30,6 @@ from sglang.srt.distributed import (
     tensor_model_parallel_all_reduce,
 )
 from sglang.srt.layers.activation import SiluAndMul
-from sglang.srt.layers.attention.triton_ops.rocm_mla_decode_rope import (
-    decode_attention_fwd_grouped_rope,
-)
 from sglang.srt.layers.dp_attention import (
     dp_gather_partial,
     dp_scatter,
@@ -83,6 +80,11 @@ if _is_cuda:
 else:
     from vllm import _custom_ops as ops
 
+if _is_hip:
+    from sglang.srt.layers.attention.triton_ops.rocm_mla_decode_rope import (
+        decode_attention_fwd_grouped_rope,
+    )
+
 expert_distribution_recorder = ExpertDistributionRecorder()
 
 

python/sglang/srt/utils.py

@@ -564,6 +564,10 @@ def load_image(image_file: Union[str, bytes]) -> tuple[Image, tuple[int, int]]:
 
 
 def suppress_other_loggers():
+    warnings.filterwarnings(
+        "ignore", category=UserWarning, message="The given NumPy array is not writable"
+    )
+
     try:
         from vllm.logger import logger as vllm_default_logger
     except ImportError:
@@ -578,10 +582,6 @@ def suppress_other_loggers():
     )
     logging.getLogger("vllm.config").setLevel(logging.ERROR)
 
-    warnings.filterwarnings(
-        "ignore", category=UserWarning, message="The given NumPy array is not writable"
-    )
-
 
 def assert_pkg_version(pkg: str, min_version: str, message: str):
     try: