|
from transformers import AutoTokenizer, AutoModelForCausalLM |
|
import torch |
|
|
|
model_id = "scb10x/llama-3-typhoon-v1.5-8b-instruct" |
|
|
|
tokenizer = AutoTokenizer.from_pretrained(model_id) |
|
model = AutoModelForCausalLM.from_pretrained( |
|
model_id, |
|
torch_dtype=torch.bfloat16, |
|
device_map="auto", |
|
) |
|
|
|
messages = [ |
|
{"role": "system", "content": "You are a helpful assistant who're always speak Thai."}, |
|
{"role": "user", "content": "ขอสูตรไก่ย่าง"}, |
|
] |
|
|
|
input_ids = tokenizer.apply_chat_template( |
|
messages, |
|
add_generation_prompt=True, |
|
return_tensors="pt" |
|
).to(model.device) |
|
|
|
terminators = [ |
|
tokenizer.eos_token_id, |
|
tokenizer.convert_tokens_to_ids("<|eot_id|>") |
|
] |
|
|
|
outputs = model.generate( |
|
input_ids, |
|
max_new_tokens=512, |
|
eos_token_id=terminators, |
|
do_sample=True, |
|
temperature=0.4, |
|
top_p=0.9, |
|
) |
|
response = outputs[0][input_ids.shape[-1]:] |
|
print(tokenizer.decode(response, skip_special_tokens=True)) |
|
|