Garbled characters from Vicuna 7b-v1.5

#10
by ChocoWu - opened

I try to leverage Vicuna v1.5 for inference. However, when I input the prompt hi, how are you?, I got the wrong answers like this:
['hi, how are you? &=\\autoritéЉanesrices Villacci sobivent分нишalotsautoritéeusal kwietnehmardadrualal al'].
Here is the basic information:

vicuna=7b-v1.5
fschat==0.2.21
transformers=4.35.2

The inference code is as follows:

import torch
import transformers
from transformers import LlamaForCausalLM, LlamaTokenizer

model_name = "../pretrained_ckpt/vicuna/7b-v1.5/"

tokenizer = LlamaTokenizer.from_pretrained(model_name, torch_dtype=torch.float16)
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "right"

model = LlamaForCausalLM.from_pretrained(model_name).cuda()
prompt = 'hi, how are you?'
input_ids = tokenizer(prompt, return_tensors="pt").input_ids
print(input_ids)

generated_ids = model.generate(input_ids.to(model.device), max_length=30)
print(generated_ids)
Large Model Systems Organization org
edited Dec 3, 2023

Please use fastchat to apply the correct chat temploate. you can try this CLI command:

python3 -m fastchat.serve.cli --model-path lmsys/vicuna-7b-v1.5	--debug
weichiang changed discussion status to closed

Sign up or log in to comment