lab/poor_output.py

import gradio as gr
import torch
import time
import json
import uuid
import os
import pytz
from datetime import datetime
from transformers import AutoTokenizer
from unsloth import FastLanguageModel
from pathlib import Path
from huggingface_hub import CommitScheduler

# Load HF token from the environment
token = os.environ["HF_TOKEN"]

# Model Setup
max_seq_length = 2048
load_in_4bit = True
name = "large-traversaal/Phi-4-Hindi"
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=name,
    max_seq_length=max_seq_length,
    load_in_4bit=load_in_4bit,
)

model = FastLanguageModel.get_peft_model(
    model,
    r=16,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
    lora_alpha=16,
    lora_dropout=0,
    bias="none",
    use_gradient_checkpointing="unsloth",
    random_state=3407,
    use_rslora=False,
    loftq_config=None,
)
FastLanguageModel.for_inference(model)

# Task-Specific Prompt Mapping
option_mapping = {
    "translation": "### TRANSLATION ###",
    "mcq": "### MCQ ###",
    "nli": "### NLI ###",
    "summarization": "### SUMMARIZATION ###",
    "long response": "### LONG RESPONSE ###",
    "direct response": "### DIRECT RESPONSE ###",
    "paraphrase": "### PARAPHRASE ###",
    "code": "### CODE ###",
}

# Set up logging folder and CommitScheduler
log_folder = Path("logs")
log_folder.mkdir(parents=True, exist_ok=True)
log_file = log_folder / f"chat_log_{uuid.uuid4()}.json"

scheduler = CommitScheduler(
    repo_id="DrishtiSharma/phi-4-unsloth-logs",  
    repo_type="dataset",
    folder_path=log_folder,
    path_in_repo="data",
    every=10,
    token=token
)

# Fixed timezone
timezone = pytz.timezone("UTC")

def generate_response(message, temperature, max_new_tokens, top_p, task):
    append_text = option_mapping.get(task, "")
    prompt = f"### INPUT : {message} {append_text} RESPONSE : "
    print(f"Prompt: {prompt}")
    start_time = time.time()
    inputs = tokenizer.encode(prompt, return_tensors="pt").to(model.device)
    
    outputs = model.generate(
        input_ids=inputs,
        max_new_tokens=max_new_tokens,
        use_cache=True,
        temperature=temperature,
        top_p=top_p,
        pad_token_id=tokenizer.eos_token_id,
    )
    
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    processed_response = response.split("### RESPONSE :")[-1].strip()
    end_time = time.time()
    
    response_time = round(end_time - start_time, 2)
    timestamp = datetime.now(timezone).strftime("%Y-%m-%d %H:%M:%S %Z")
    log_data = {
        "timestamp": timestamp,
        "input": message,
        "output": processed_response,
        "response_time": response_time,
        "temperature": temperature,
        "max_tokens": max_new_tokens,
        "top_p": top_p
    }
    
    with scheduler.lock:
        with log_file.open("a") as f:
            f.write(json.dumps(log_data) + "\n")
    
    return processed_response

# Gradio UI
with gr.Blocks() as demo:
    gr.Markdown("## Chat with Phi-4-Hindi")
    
    task_dropdown = gr.Dropdown(
        choices=list(option_mapping.keys()),
        value="long response",
        label="Select Task"
    )
    message_input = gr.Textbox(label="Enter your message")
    
    with gr.Row():
        temperature_slider = gr.Slider(0.1, 1.0, value=0.7, step=0.1, label="Temperature")
        top_p_slider = gr.Slider(0.1, 1.0, value=0.9, step=0.1, label="Top P")
        max_tokens_slider = gr.Slider(50, 800, value=200, step=50, label="Max Tokens")
    
    output_box = gr.Textbox(label="Generated Response")
    generate_btn = gr.Button("Generate")
    
    generate_btn.click(
        generate_response,
        inputs=[message_input, temperature_slider, max_tokens_slider, top_p_slider, task_dropdown],
        outputs=output_box
    )

demo.launch()