timdettmers

/

guanaco-33b-merged

like 164

thanks for your work. I compared a lot llms and find this one is has good performance with low cost. but when i try to generate 512 tokens or 1024 tokens, both 33b and 65b versions will failed and my linux system will freeze.
i am currently using two 4090 with newest driver 525. any comments will be appreciated.

following is my inference config

nf4_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_use_double_quant=True,
bnb_4bit_compute_dtype=torch.float16
)

config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_path,
quantization_config=nf4_config,
trust_remote_code=True,
torch_dtype=torch.float16, # additional option to lower RAM consumtion
device_map={"": 1} #used for 33b version
#device_map='auto' #used for 65b version
)

tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False)

input_text = 'system: you are AI model. help user and responde carefully with long generation.\n
user:what is y-90 liver treatment. responde with 1000 words. \n
AI:'
input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to('cuda')

outputs = model.generate(input_ids, max_length=1024, top_k=10, temperature=0.5)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))