Update model_loader deps and qqq quantization deps (#2220) (#2318)

Co-authored-by: HandH1998 <1335248067@qq.com>

python/sglang/srt/utils.py

@@ -430,16 +430,12 @@ def suppress_other_loggers():
     from vllm.logger import logger as vllm_default_logger
 
     vllm_default_logger.setLevel(logging.WARN)
-    logging.getLogger("vllm.config").setLevel(logging.ERROR)
     logging.getLogger("vllm.distributed.device_communicators.pynccl").setLevel(
         logging.WARN
     )
     logging.getLogger("vllm.distributed.device_communicators.shm_broadcast").setLevel(
         logging.WARN
     )
-    logging.getLogger("vllm.selector").setLevel(logging.WARN)
-    logging.getLogger("vllm.utils").setLevel(logging.ERROR)
-    logging.getLogger("vllm.model_executor.model_loader.loader").setLevel(logging.ERROR)
 
     warnings.filterwarnings(
         "ignore", category=UserWarning, message="The given NumPy array is not writable"
@@ -492,27 +488,6 @@ def kill_process_tree(parent_pid, include_parent: bool = True, skip_pid: int = N
             pass
 
 
-def monkey_patch_vllm_model_config():
-    from vllm.config import ModelConfig
-
-    if not hasattr(ModelConfig, "_resolve_task"):
-        return
-
-    def _resolve_task(
-        self,
-        task_option,
-        hf_config,
-    ):
-        supported_tasks = {
-            "generate": True,
-            "embedding": False,
-        }
-        selected_task = "generate"
-        return supported_tasks, selected_task
-
-    setattr(ModelConfig, "_resolve_task", _resolve_task)
-
-
 def monkey_patch_vllm_p2p_access_check(gpu_id: int):
     """
     Monkey patch the slow p2p access check in vllm.
@@ -1041,6 +1016,11 @@ def crash_on_warnings():
     return get_bool_env_var("SGLANG_IS_IN_CI")
 
 
+def print_warning_once(msg: str) -> None:
+    # Set the stacklevel to 2 to print the caller's line info
+    logger.warning(msg, stacklevel=2)
+
+
 def get_device_name(device_id: int = 0) -> str:
     if hasattr(torch, "cuda") and torch.cuda.is_available():
         return torch.cuda.get_device_name(device_id)
@@ -1055,6 +1035,33 @@ def get_device_name(device_id: int = 0) -> str:
         return torch.hpu.get_device_name(device_id)
 
 
+def get_device_capability(device_id: int = 0) -> Tuple[int, int]:
+    major, minor = None, None
+    if hasattr(torch, "cuda") and torch.cuda.is_available():
+        major, minor = torch.cuda.get_device_capability(device_id)
+
+    if hasattr(torch, "hip") and torch.hip.is_available():
+        major, minor = torch.cuda.get_device_capability(device_id)
+
+    if hasattr(torch, "xpu") and torch.xpu.is_available():
+        major, minor, *_ = torch.xpu.get_device_capability(device_id)["version"].split(
+            "."
+        )
+        major, minor = int(major), int(minor)
+
+    # TODO(HandH1998): `get_device_capability` is not supported by `torch.hpu` for now.
+    # Update this once the support is available.
+    if hasattr(torch, "hpu") and torch.hpu.is_available():
+        try:
+            major, minor = torch.hpu.get_device_capability(device_id)
+        except Exception as e:
+            raise RuntimeError(
+                f"An error occurred while getting device capability of hpu: {e}."
+            ) from e
+
+    return major, minor
+
+
 sglang_lib = Library("sglang", "FRAGMENT")  # noqa