fix(server): t5 cannot run in f16 (#356)

Fix #349

server/text_generation_server/models/t5.py

@@ -40,7 +40,7 @@ class T5Sharded(Seq2SeqLM):
         self.process_group, rank, world_size = initialize_torch_distributed()
         if torch.cuda.is_available():
             device = torch.device(f"cuda:{rank}")
-            dtype = torch.float16
+            dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float32
         else:
             device = torch.device("cpu")
             dtype = torch.float32