feat(embeddings): flush denormal numbers

handler.py

@@ -1,4 +1,3 @@
-import os
 import platform
 from typing import Union, Sequence, Sized
 
@@ -48,6 +47,10 @@ class SentenceTransformerHandler(Handler):
         self._allocate_model()
 
     def _allocate_model(self):
+        # Denormal number is used to store extremely small numbers that are close to 0.
+        # Computations with denormal numbers are remarkably slower than normalized number.
+        torch.set_flush_denormal(True)
+
         dtype = torch.bfloat16 if has_bf16_support() else torch.float32
         model = SentenceTransformer(self._config.model_id, device="cpu", model_kwargs={"torch_dtype": dtype})