From e3f0f85b7099dae78dbd7d91dc8a5403eb66b7b7 Mon Sep 17 00:00:00 2001 From: Sun Choi Date: Mon, 12 Aug 2024 15:00:41 -0700 Subject: [PATCH] Pad token handling for Llama3.1 (#199) --- server/text_generation_server/models/causal_lm.py | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-) diff --git a/server/text_generation_server/models/causal_lm.py b/server/text_generation_server/models/causal_lm.py index 5eddd5b7..6437acf9 100644 --- a/server/text_generation_server/models/causal_lm.py +++ b/server/text_generation_server/models/causal_lm.py @@ -674,7 +674,14 @@ class CausalLM(Model): if model.config.pad_token_id is not None: tokenizer.pad_token_id = model.config.pad_token_id elif model.config.eos_token_id is not None: - tokenizer.pad_token_id = model.config.eos_token_id + if isinstance(model.config.eos_token_id, int): + tokenizer.pad_token_id = model.config.eos_token_id + elif isinstance(model.config.eos_token_id, list): + tokenizer.pad_token_id = model.config.eos_token_id[0] + else: + raise ValueError( + f"{type(model.config.eos_token_id)} type of eos_token_id in the model's config is not supported for tokenizer.pad_token_id" + ) elif tokenizer.eos_token_id is not None: tokenizer.pad_token_id = tokenizer.eos_token_id else: