app.py

import re
import spaces
import gradio as gr
import torch
# from peft import PeftModel, PeftConfig


# tokenizer = AutoTokenizer.from_pretrained("FlawedLLM/Bhashini")
# quantization_config = BitsAndBytesConfig(
#         load_in_4bit=True,
#         bnb_4bit_use_double_quant=True,
#         bnb_4bit_quant_type="nf4",  
#         bnb_4bit_compute_dtype=torch.float16)
# model = AutoModelForCausalLM.from_pretrained("FlawedLLM/BhashiniLLM",
#                                              device_map="auto",
#                                              quantization_config=quantization_config, 
#                                              torch_dtype =torch.float16, 
#                                              low_cpu_mem_usage=True, 
#                                              use_safetensors=True,
#                                             )

# # Assuming you have your HF repository in this format: "your_username/your_model_name"
# model_id = "FlawedLLM/BhashiniLLM"

# # Load the base model (the one you fine-tuned with LoRA)
# base_model = AutoModelForCausalLM.from_pretrained(model_id, device_map='auto')  # Load in 8-bit for efficiency
# for param in base_model.parameters():
#     param.data = param.data.to(torch.float16)  # or torch.float32

# # Load the LoRA adapter weights
# model = PeftModel.from_pretrained(base_model, model_id)
# tokenizer = AutoTokenizer.from_pretrained(model_id)



# model = AutoModel.from_pretrained("FlawedLLM/Bhashini", load_in_4bit=True, device_map='auto')
    # I highly do NOT suggest - use Unsloth if possible
# from peft import AutoPeftModelForCausalLM
# from transformers import AutoTokenizer
# model = AutoPeftModelForCausalLM.from_pretrained(
#         "FlawedLLM/Bhashini", # YOUR MODEL YOU USED FOR TRAINING
#         load_in_4bit = True,
#     )
# tokenizer = AutoTokenizer.from_pretrained("FlawedLLM/Bhashini")
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig

tokenizer = AutoTokenizer.from_pretrained("FlawedLLM/Bhashini_9")
quantization_config = BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_use_double_quant=True,
        bnb_4bit_quant_type="nf4",  
        bnb_4bit_compute_dtype=torch.float16)
torch_dtype =torch.float16
model = AutoModelForCausalLM.from_pretrained("FlawedLLM/Bhashini_9",quantization_config=quantization_config ,)

@spaces.GPU(duration=300)
def chunk_it(input_command):
    alpaca_prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.
    
    ### Instruction:
    {}
    
    ### Input:
    {}
    
    ### Response:
    {}"""
    inputs = tokenizer(
    [
        alpaca_prompt.format(
            '''
            You will receive  text input that you need to analyze to perform the following tasks:
    
            transaction: Record the details of an item transaction.
            last n days transactions: Retrieve transaction records for a specified time period.
            view risk inventory: View inventory items based on a risk category.
            view inventory: View inventory details.
            new items: Add new items to the inventory.
            old items: View old items in inventory.
            report generation: Generate various inventory reports.
            Required Parameters:
    
            Each task requires specific parameters to execute correctly:
    
            transaction:
                ItemName (string)
                ItemQt (quantity - integer)
                Flow (string: "in" or "out")
                ShelfNo (string or integer)
            last n days transactions:
                ItemName (string)
                Duration (integer: number of days, default: 30)
            view risk inventory:
                RiskType (string: "overstock", "understock", or Null for all risk types)
            view inventory:
                ItemName (string)
                ShelfNo (string or integer)
            new items:
                ItemName (string)
                SellingPrice (number)
                CostPrice (number)
            old items:
                ShelfNo (string or integer)
            report generation:
                ItemName (string)
                Duration (integer: number of days, default: 6)
                ReportType (string: "profit", "revenue", "inventory", or Null for all reports)
    
            ALWAYS provide output in a JSON format.''', # instruction
            input_command, # input
            "", # output - leave this blank for generation!
        )
    ], return_tensors = "pt").to("cuda")
    
    outputs = model.generate(**inputs, max_new_tokens = 216, use_cache = True)
    tokenizer.batch_decode(outputs)
    
    reply=tokenizer.batch_decode(outputs)
    # Regular expression pattern to match content between "### Response:" and "<|end_of_text|>"
    pattern = r"### Response:\n(.*?)<\|end_of_text\|>"
    # Search for the pattern in the text
    match = re.search(pattern, reply[0], re.DOTALL)  # re.DOTALL allows '.' to match newlines
    reply = match.group(1).strip()  # Extract and remove extra whitespace

    return reply


iface=gr.Interface(fn=chunk_it,
                  inputs="text",
                  outputs="text",
                  title="Formatter_Pro",
                  )
iface.launch(inline=False)