Fix: fallback to torch API when NVML memory query is not supported (#23426)

Co-authored-by: ronnie_zheng <zl19940307@163.com>

python/sglang/multimodal_gen/runtime/platforms/cuda.py

@@ -514,7 +514,10 @@ class NvmlCudaPlatform(CudaPlatformBase):
     def get_device_total_memory(cls, device_id: int = 0) -> int:
         physical_device_id = device_id_to_physical_device_id(device_id)
         handle = pynvml.nvmlDeviceGetHandleByIndex(physical_device_id)
-        return int(pynvml.nvmlDeviceGetMemoryInfo(handle).total)
+        try:
+            return int(pynvml.nvmlDeviceGetMemoryInfo(handle).total)
+        except pynvml.NVMLError_NotSupported:
+            return int(torch.cuda.get_device_properties(device_id).total_memory)
 
     @classmethod
     @with_nvml_context