Reduce GPU memory usage in the runtime.

After adding 'with torch.no_grad():', memory can be reduced form 11.77G to 3.19G when batch=4, token=1024.

README.md

@@ -157,7 +157,8 @@ batch_dict = tokenizer(
     return_tensors="pt",
 )
 batch_dict.to(model.device)
-outputs = model(**batch_dict)
+with torch.no_grad():
+    outputs = model(**batch_dict)
 embeddings = last_token_pool(outputs.last_hidden_state, batch_dict['attention_mask'])
 
 # normalize embeddings