Use BF16 on HPU by default (HabanaAI#361)

We don't *officially* support FP16, and for the most part, we use BF16
wherever we can. This removes the need of specifying `--dtype bfloat16`
- when `dtype` is not provided (is `auto`), and model default data type
is `float16`, we cast it to `bfloat16` for HPU.

vllm/config.py

@@ -1635,6 +1635,13 @@ def _get_and_verify_dtype(
                     torch_dtype = torch.float16
             else:
                 torch_dtype = config_dtype
+
+            if current_platform.is_hpu() and config_dtype == torch.float16:
+                logger.info(
+                    "For HPU, we cast models to bfloat16 instead of"
+                    "using float16 by default. Please specify `dtype` if you "
+                    "want to use float16.")
+                torch_dtype = torch.bfloat16
         else:
             if dtype not in _STR_DTYPE_TO_TORCH_DTYPE:
                 raise ValueError(f"Unknown dtype: {dtype}")