app.py

import os, json, itertools, bisect, gc

from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig
import transformers
import torch
from accelerate import Accelerator
import accelerate
import time
from huggingface_hub import login



# 🟢 (Option 2 - recommended for Render/Colab)
# Save your token as an environment variable called HUGGINGFACE_HUB_TOKEN
# Then this will automatically pick it up:
hf_token = os.getenv("HUGGINGFACE_HUB_TOKEN")

model = None
tokenizer = None
generator = None

def load_model(model_name, eight_bit=0, device_map="auto"):
    global model, tokenizer, generator

    print("Loading "+model_name+"...")

    if device_map == "zero":
        device_map = "balanced_low_0"

    # config
    gpu_count = torch.cuda.device_count()
    print('gpu_count', gpu_count)

    tokenizer = transformers.LlamaTokenizer.from_pretrained(model_name, token=hf_token)
    model = transformers.LlamaForCausalLM.from_pretrained(
        model_name,
        #device_map=device_map,
        #device_map="auto",
        torch_dtype=torch.float16,
        #max_memory = {0: "14GB", 1: "14GB", 2: "14GB", 3: "14GB",4: "14GB",5: "14GB",6: "14GB",7: "14GB"},
        #load_in_8bit=eight_bit,
        low_cpu_mem_usage=True,
        load_in_8bit=False,
        cache_dir="cache",
        token=hf_token 
    ).cuda()

    generator = model.generate

load_model("Muhammadidrees/JayConverstionalModel")

history = []

def go():
    invitation = "Assistant: "
    human_invitation = "Human: "

    # input
    msg = input(human_invitation)
    print("")

    history.append(human_invitation + msg)

    fulltext = "\n\n".join(history) + "\n\n" + invitation

#    print('SENDING==========')
#    print(fulltext)
#    print('==========')

    generated_text = ""
    gen_in = tokenizer(fulltext, return_tensors="pt").input_ids.cuda()
    in_tokens = len(gen_in)
    with torch.no_grad():
            generated_ids = generator(
                gen_in,
                max_new_tokens=200,
                use_cache=True,
                pad_token_id=tokenizer.eos_token_id,
                num_return_sequences=1,
                do_sample=True,
                repetition_penalty=1.1, # 1.0 means 'off'. unfortunately if we penalize it it will not output Sphynx:
                temperature=0.5, # default: 1.0
                top_k = 50, # default: 50
                top_p = 1.0, # default: 1.0
                early_stopping=True,
            )
            generated_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0] # for some reason, batch_decode returns an array of one element?

            text_without_prompt = generated_text[len(fulltext):]

    response = text_without_prompt

    response = response.split(human_invitation)[0]

    response.strip()

    print(invitation + response)

    print("")

    history.append(invitation + response)

while True:
    go()