NickyNicky/OpenHermes-2.5-Mistral-7B-nbits-4-group_size-128-HQQ

!python -m pip install --upgrade pip -q
!pip install -q accelerate safetensors deepspeed
!pip install -q bitsandbytes sentencepiece
!pip install -q scipy ninja -U
!pip install git+https://github.com/mobiusml/hqq/ transformers -U -q

import transformers
print(transformers.__version__)
### output: 4.39.2

model_id = 'NickyNicky/OpenHermes-2.5-Mistral-7B-nbits-4-group_size-128-HQQ'

from hqq.engine.hf import HQQModelForCausalLM, AutoTokenizer



tokenizer = AutoTokenizer.from_pretrained(model_id)
model     = HQQModelForCausalLM.from_quantized(model_id,device="cuda:0")

model.config.use_cache  = True
model.eval();

system=""""""

contenido="""escribe solo tres palabaras que contengan la letra 'T'"""

messages = [{"role": "system", "content": system+"eres un modelo de AI que responde adecuadamente a las tareas exactas que te pide el usuario, el idioma a la cual debes de responder es español."},
            {"role": "user", "content": contenido},

            ]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

input= tokenizer(prompt,
                 return_tensors="pt",
                 add_special_tokens=False).to(model.device)

generate_params = dict(
        max_new_tokens = 1900,
        do_sample = True,
        top_p = 0.90,
        top_k = 50,
        temperature =  0.6,
        repetition_penalty = 1.,
        pad_token_id = tokenizer.eos_token_id,
        eos_token_id = tokenizer.eos_token_id,
    )


output= model.generate(**input,**generate_params)
print(tokenizer.decode(output[0], skip_special_tokens=True))

use gpu.

colab.

https://colab.research.google.com/drive/1oEoH0qScGzkLV4WLGrMEMgl4qnEsZhTs?usp=sharing