tiiuae

/

falcon-7b-instruct

like 917

Follow

Technology Innovation Institute 919

I built a tool that generates inferences in a loop with Falcon 7B in 16/8/4 bit. Weirdly, Falcon generates the first 3 inferences rapidly and then it blocks and never returns the 4th inference when I use it in 8 or 4 bit. Plus, the problem persists when I change the set of instructions I use to generate inferences.

Here is the code I use to load the model and generate inferences:

def load_falcon(weight_encoding):
model_path = "/falcon-7b-instruct"
eight_bit = weight_encoding == "8-bit"
four_bit = weight_encoding == "4-bit"

#creating a model 
fmodel = AutoModelForCausalLM.from_pretrained(
    model_path,
    load_in_8bit = eight_bit,
    load_in_4bit = four_bit,
    trust_remote_code = True,
    torch_dtype= torch.float16,
    device_map = "auto")

fmodel.eval() 
# fmodel.to(device)
tokenizer = AutoTokenizer.from_pretrained(model_path)

gen_text = transformers.pipeline(
model=fmodel, 
tokenizer=tokenizer, 
task='text-generation', 
return_full_text=False,     
max_length=5000, 
temperature=0.1, 
top_p=0.75, #select from top tokens whose probability adds up to 15%
top_k=40, #selecting from top 0 tokens 
repetition_penalty=1.9, #without a penalty, output starts to repeat 
do_sample=True, 
num_return_sequences=1,
eos_token_id=tokenizer.eos_token_id

)
return gen_text, tokenizer

def falcon_inference(instruction, gen_text):
response = gen_text(instruction)
texts = [ seq['generated_text'] for seq in response]
text = '\n'.join(texts)
return text

Configuration:
CUDA Version: 12.1
NVIDIA GeForce RTX 4090 (24 GO VRAM)
I have the latest versions of bitsandbytes (0.39.0), transformers (4.31.0.dev0) and accelerate (0.21.0.dev0).

Thank you for your help !