jordiclive/instruction-tuned-gpt-neox-20b

Experimental 20B instruction tuned model based on gpt-neox-20b.

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

name = "jordiclive/instruction-tuned-gpt-neox-20b"
model = AutoModelForCausalLM.from_pretrained(name, device_map=chip_map, torch_dtype=torch.float16)# load_in_8bit=True )
tokenizer = AutoTokenizer.from_pretrained(name)

def generate_from_model(model, tokenizer):
    encoded_input = tokenizer(text, return_tensors='pt')
    output_sequences = model.generate(
                                    input_ids=encoded_input['input_ids'].cuda(0),
                                    do_sample=True,
                                    max_new_tokens=35,
                                    num_return_sequences=1,
                                    top_p=0.95,
                                    temperature=0.5,
                                    penalty_alpha=0.6,
                                    top_k=4,
                                    output_scores=True,
                                    return_dict_in_generate=True,
                                    repetition_penalty=1.03,
                                    eos_token_id=0,
                                    use_cache=True
                                  )
    gen_sequences = output_sequences.sequences[:, encoded_input['input_ids'].shape[-1]:]
    for sequence in gen_sequences:
        new_line=tokenizer.decode(sequence, skip_special_tokens=True)
        print(new_line)

text = "User: Will big tech A.I be adulterated with advertisement?\n\nOA:"
generate_from_model(model,tokenizer)