[Hardware] Initial TPU integration (#5292)

vllm/utils.py

@@ -146,6 +146,15 @@ def is_neuron() -> bool:
     return transformers_neuronx is not None
 
 
+@lru_cache(maxsize=None)
+def is_tpu() -> bool:
+    try:
+        import libtpu
+    except ImportError:
+        libtpu = None
+    return libtpu is not None
+
+
 @lru_cache(maxsize=None)
 def get_max_shared_memory_bytes(gpu: int = 0) -> int:
     """Returns the maximum shared memory per thread block in bytes."""
@@ -546,6 +555,11 @@ def maybe_expand_dim(tensor: torch.Tensor,
     return tensor
 
 
+def get_dtype_size(dtype: torch.dtype) -> int:
+    """Get the size of the data type in bytes."""
+    return torch.tensor([], dtype=dtype).element_size()
+
+
 def merge_dicts(dict1: Dict[Any, List[Any]],
                 dict2: Dict[Any, List[Any]]) -> Dict[Any, List[Any]]:
     """Merge 2 dicts that have key -> List of items.