migtissera

/

HelixNet

like 106

# Because ctransformers' AutoTokenizer.from_pretrain(:model) for GGML/GGUF is not implemented yet, 
# I use it from HF transformers.
import torch, json
from transformers import AutoTokenizer

from ctransformers import AutoModelForCausalLM

# Model paths must have everything from original files except original model weight *.bin files
# becase those files are necessary to load tokenizers from HF transformers.
model_path_actor = "/replace_it_with_your_model_path/migtissera_HelixNet/actor"
model_path_critic = "/replace_it_with_your_model_path/migtissera_HelixNet/critic"
model_path_regenerator = "/replace_it_with_your_model_path/migtissera_HelixNet/regenerator"

# Quantized GGUF files must be in each model path.
model_file_actor = "migtissera_HelixNet_actor-Q4_K_M.gguf"
model_file_critic = "migtissera_HelixNet_critic-Q4_K_M.gguf"
model_file_regenerator = "migtissera_HelixNet_regenerator-Q4_K_M.gguf"


def load_model_gguf(model_path, file_path):
    model = AutoModelForCausalLM.from_pretrained(
        model_path,
        model_file=file_path,
        model_type="llama",
        local_files_only=True,
        context_length=1024,
        # gpu_layers is for Metal acceleration for Apple Silicon.
        # Comment out or remove if you do not use Metal acceleration.
        gpu_layers=1,
        hf=True
    )
    return model

def load_tokenizer(model):
    tokenizer = AutoTokenizer.from_pretrained(model, local_files_only=True)
    return tokenizer

model_actor = load_model_gguf(model_path_actor, model_file_actor)
model_critic = load_model_gguf(model_path_critic, model_file_critic)
model_regenerator = load_model_gguf(model_path_regenerator, model_file_regenerator)

tokenizer_actor = load_tokenizer(model_path_actor)
tokenizer_critic = load_tokenizer(model_path_critic)
tokenizer_regenerator = load_tokenizer(model_path_regenerator)

def generate_text(instruction, model, tokenizer):
    tokens = tokenizer.encode(instruction)
    tokens = torch.LongTensor(tokens).unsqueeze(0)

    instance = {
        "input_ids": tokens,
        "top_p": 1.0,
        "temperature": 0.75,
        "generate_len": 1024,
        "top_k": 50,
    }

    length = len(tokens[0])
    with torch.no_grad():
        rest = model.generate(
            input_ids=tokens,
            max_length=length + instance["generate_len"],
            use_cache=True,
            do_sample=True,
            top_p=instance["top_p"],
            temperature=instance["temperature"],
            top_k=instance["top_k"],
            num_return_sequences=1,
        )
    output = rest[0][length:]
    string = tokenizer.decode(output, skip_special_tokens=True)
    return f"{string}"

system_prompt = "You are HelixNet. Elaborate on the topic using a Tree of Thoughts and backtrack when necessary to construct a clear, cohesive Chain of Thought reasoning. Always answer without hesitation."
  
conversation = f"SYSTEM:{system_prompt}"

while True:
    user_input = input("You: ")

    prompt_actor = f"{conversation} \nUSER: {user_input} \nASSISTANT: "
    actor_response = generate_text(prompt_actor, model_actor, tokenizer_actor)
    print("Generated ACTOR RESPONSE")

    prompt_critic = f"SYSTEM: {system_prompt} \nUSER: {user_input} \nRESPONSE: {actor_response} \nCRITIQUE:"
    critic_response = generate_text(prompt_critic, model_critic, tokenizer_critic)
    print("Generated CRITIQUE")

    prompt_regenerator = f"SYSTEM: {system_prompt} \nUSER: {user_input} \nRESPONSE: {actor_response} \nCRITIQUE: {critic_response} \nREGENERATOR: REGENERATED ANSWER:"
    regenerator_response = generate_text(prompt_regenerator, model_regenerator, tokenizer_regenerator)
    print("Generated REGENERATION")

    conversation = f"{conversation} \nUSER: {user_input} \nASSISTANT: {regenerator_response}"
    print(conversation)