'eos_token_id' for llama model.generate is not working

[devymex]

System Info

• transformers version: 4.30.2
• Platform: Linux-5.4.0-137-generic-x86_64-with-glibc2.31
• Python version: 3.10.0
• Huggingface_hub version: 0.15.1
• Safetensors version: 0.3.1
• PyTorch version (GPU?): 2.0.1+cu117 (True)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using GPU in script?:
• Using distributed or parallel set-up in script?:

Who can help?

No response

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

import transformers, torch

weights_dir = "weights/recovered"
question = 'Hello, there!'

model = transformers.AutoModelForCausalLM.from_pretrained(weights_dir)
model = model.cuda()
print(model.config)
# LlamaConfig {
#   "_name_or_path": "weights/recovered",
#   "architectures": [
#     "LlamaForCausalLM"
#   ],
#   "bos_token_id": 1,
#   "eos_token_id": 2,
#   "hidden_act": "silu",
#   "hidden_size": 4096,
#   "initializer_range": 0.02,
#   "intermediate_size": 11008,
#   "max_position_embeddings": 2048,
#   "model_type": "llama",
#   "num_attention_heads": 32,
#   "num_hidden_layers": 32,
#   "pad_token_id": 0,
#   "rms_norm_eps": 1e-06,
#   "tie_word_embeddings": false,
#   "torch_dtype": "float32",
#   "transformers_version": "4.30.2",
#   "use_cache": true,
#   "vocab_size": 32001
# }

tokenizer = transformers.AutoTokenizer.from_pretrained(weights_dir)
question_ids = tokenizer.encode(question + tokenizer.eos_token, return_tensors='pt')
question_ids = question_ids.cuda()

print(tokenizer.eos_token_id, tokenizer.bos_token_id, tokenizer.pad_token_id)
# 2, 1, 32000

print(question_ids)
# tensor([[    1, 15043, 29892,   727, 29991,   829, 29879, 29958]],
       device='cuda:0')

print(tokenizer.decode(question_ids[0]))
# <s> Hello, there!</s>

outputs = model.generate(
        question_ids,
        eos_token_id=2,
        max_new_tokens=200,
        num_beams=4,
        num_return_sequences=2,
        early_stopping=True
    )
answer = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(answer)
# Hello, there!</s>
# Hello, there!</s>
# <s>Hello, there!</s>

No matter how I changing the parameters of model.generate, it always ignores the </s> as the ending token (id:2).

In addition, the skip_special_tokens of tokenizer is not working too.

Where am I doing wrong? Please help, many thanks!

Expected behavior

The model.generate stop at the first time of </s>

[amyeroberts]

cc @ArthurZucker

[ArthurZucker]

Hey! A few things to note:

• LlamaTokenizerFast (which you are using through the AutoTokenizer API) has been fixed here [Lllama] Update tokenization code to ensure parsing of the special tokens [core] #24042, addressing the issue with special tokens being encode.
• You are not sharing any repo, so we can't reproduce potential bugs.
• it always ignores the </s> as the ending token what does that mean? Does the generation not stop? Then have a look here LLaMA FastTokenizer does not add eos_token_id at the end. #22794.
• skip_special_tokens will work if you have the correct version of LlamaTokenizer.
• If you wish to add the ending token in your prompt, set add_eos_token to True. It will be done automatically

Here is a working snippet:

from transformers import LlamaTokenizer, AutoModelForCausalLM, AutoTokenizer
weights_dir = "huggyllama/llama-7b"
question = 'Hello, there!'

# if you want to add eos, set `add_eos_token=True`
tokenizer = LlamaTokenizer.from_pretrained(weights_dir, add_eos_token=True)
question_ids = tokenizer.encode(question, return_tensors='pt')
print(question_ids)
# tensor([[    1, 15043, 29892,   727, 29991,     2]])
print( tokenizer.decode(question_ids[0], skip_special_tokens = True))
# 'Hello, there!'


# if you are not using the correct version of tokenizer, special tokens are wrong
tokenizer = AutoTokenizer.from_pretrained(weights_dir, add_eos_token=True)
print(tokenizer.is_fast)
# True
question_ids = tokenizer.encode('Hello, there!</s>', return_tensors='pt')
print(question_ids)
# tensor([[    1, 15043, 29892,   727, 29991,   829, 29879, 29958,     2]])
question_ids = tokenizer.encode('Hello, there! </s>', return_tensors='pt')
# tensor([[    1, 15043, 29892,   727, 29991,     2,     2]])
print(question_ids)

[devymex]

@ArthurZucker Many thanks! add_eos_token=True did the trick!