diff --git a/server/text_generation_server/utils/logits_process.py b/server/text_generation_server/utils/logits_process.py index d8c7c2c6..bc99819f 100644 --- a/server/text_generation_server/utils/logits_process.py +++ b/server/text_generation_server/utils/logits_process.py @@ -20,8 +20,6 @@ from transformers import ( TypicalLogitsWarper, ) -from transformers.generation.logits_process import _calc_banned_ngram_tokens - mempool = torch.cuda.graph_pool_handle() if torch.cuda.is_available() else None