app.py

import torch
from transformers import LlamaForCausalLM, LlamaTokenizer, LoraModel, LoraConfig, PeftModel
import gradio as gr

# Load tokenizer
tokenizer = LlamaTokenizer.from_pretrained("VanguardAI/BhashiniLLaMa3-8B_LoRA_Adapters")

# Load base model
base_model = LlamaForCausalLM.from_pretrained("unsloth/llama-3-8b-Instruct-bnb-4bit")

# Apply LoRA adapters
lora_config = LoraConfig(
    r=16,
    lora_alpha=16,
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj",],
    lora_dropout=0,
    bias="none",
    task_type="CAUSAL_LM"
)
model = PeftModel.from_pretrained(base_model, "VanguardAI/BhashiniLLaMa3-8B_LoRA_Adapters", config=lora_config)

condition= '''
ALWAYS provide output in a JSON format.
'''
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:
{}"""

@spaces.GPU(duration=300)
def chunk_it(inventory_list, user_input_text):
    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.
                report generation: Generate various inventory reports.
                delete item: Delete an existing Item.
        
                Required Parameters:
                Each task requires specific parameters to execute correctly:
        
                transaction:
                    ItemName (string)
                    ItemQt (quantity - integer)
                    Type (string: "sale" or "purchase" or "return")
                    ReorderPoint (integer)
                last n days transactions:
                    ItemName (string)
                    Duration (integer: number of days, if user input is in weeks, months or years then convert to days)
                view risk inventory:
                    RiskType (string: "overstock", "understock", or "Null" for all risk types)
                view inventory:
                    ItemName (string)
                new items:
                    ItemName (string)
                    SellingPrice (number)
                    CostPrice (number)
                report generation:
                    ItemName (string)
                    Duration (integer: number of days, if user input is in weeks, months or years then convert to days)
                    ReportType (string: "profit", "revenue", "inventory", or "Null" for all reports)
        
                The ItemName must always be matched from the below list of names, EXCEPT for when the Function is "new items".
                ''' + inventory_list +
                '''
                ALWAYS provide output in a JSON format.
                ''',  # instruction
                user_input_text,  # input
                "",  # output - leave this blank for generation!
            )
        ], return_tensors="pt").to("cuda")
    outputs = model.generate(**inputs, max_new_tokens=216, use_cache=True)
    content = tokenizer.batch_decode(outputs, skip_special_tokens=True)
    return content[0]


iface=gr.Interface(fn=chunk_it,
                  inputs="text",
                  outputs="text",
                  title="Bhashini_LLaMa_LoRA",
                  )
iface = gr.Interface(
    fn=chunk_it,
    inputs=[
        gr.Textbox(label="user_input_text", lines=3),
        gr.Textbox(label="inventory_list", lines=5)
    ],
    outputs="text",
    title="Formatter Pro",
)
iface.launch(inline=False)