import streamlit as st
from transformers import AutoTokenizer, AutoModelForCausalLM
import time
import asyncio
import aiohttp
import json
import torch
import re

repo_name = "BeardedMonster/SabiYarn-125M"
device = "cuda" if torch.cuda.is_available() else "cpu"

@st.cache_resource(show_spinner=False)
def load_model():
    tokenizer = AutoTokenizer.from_pretrained(repo_name, trust_remote_code=True)
    model = AutoModelForCausalLM.from_pretrained(repo_name, trust_remote_code=True).to(device)
    return tokenizer, model

tokenizer, model = load_model()

# Add sidebar with instructions
st.sidebar.title("Instructions: How to use")
# st.sidebar.write("""
# 1. Write something in the text area (a prompt or random text) or use the dropdown menu to select predefined sample text.
# 2. Select a task from the **task dropdown menu** below only if you are providing your own text. **This is very important as it ensures the model responds accordingly.**
# 3. If you are providing your own text, please do not select any predefined sample text from the dropdown menu.
# 3. If a dropdown menu pops up for a nigerian language, **select the nigerian language (base language for diacritization and text cleaning tasks, target language for translation task).**
# 4. Then, click the Generate button.\n
# 5. For Translation tasks, setting english as the target language yields the best result (english as base language performs the worst).
# **Note: Model's overall performance vary (hallucinates) due to model size and training data distribution (majorly from articles and the bible). Performance may worsen with other task outside text generation and translation.
# For other tasks, we suggest you try them several times due to the generator's sampling method.**\n
# 6. Lastly, you can play with some of the generation parameters below to improve performance.
# """)

st.sidebar.write("""
1. **Write Text or Select Sample:**
   - Enter text in the text area or use the dropdown to choose a sample.

2. **Select a Task:**
   - Choose a task from the **task dropdown** if using your own text.
   - **Important:** This ensures correct model response.

3. **Avoid Conflicts:**
   - Don't select a sample text if using your own text.

4. **Select Nigerian Language:**
   - If prompted, choose the Nigerian language (base for diacritization/cleaning, target for translation).

5. **Generate Output:**
   - Click the Generate button.

6. **Translation Tips:**
   - English as the target language gives the best results. English as the base language performs poorly.

7. **Performance Note:**
   - The model's performance varies due to its size and training data. It performs best on text generation and translation.
   - For other tasks, try multiple times due to sampling.

8. **Adjust Parameters:**
   - Experiment with the generation parameters to improve performance.
""")

max_length = 100 
max_new_tokens = 50
num_beams = 5
temperature = 0.99 
top_k = 50
top_p = 0.95 
repetition_penalty = 4.0 
length_penalty = 3.0

# Create sliders in the sidebar
max_length = st.sidebar.slider("Max. output length", min_value=10, max_value=500, value=max_length)
max_new_tokens = st.sidebar.slider("Max. generated tokens", min_value=30, max_value=768, value=max_new_tokens)
num_beams = st.sidebar.slider("Number of Beams: Improves coherence of the model output.", min_value=1, max_value=10, value=num_beams)
temperature = st.sidebar.slider("Temperature: Controls the creativity of the model", min_value=0.1, max_value=2.0, value=temperature)
top_k = st.sidebar.slider("Top-K: Controls model's sampling space.", min_value=1, max_value=100, value=top_k)
top_p = st.sidebar.slider("Top-P", min_value=0.1, max_value=1.0, value=top_p)
repetition_penalty = st.sidebar.slider("Repetition Penalty: Discourages token repitition during generation.", min_value=1.0, max_value=10.0, value=repetition_penalty)
length_penalty = st.sidebar.slider("Length Penalty: Discourages poor output as token length grows.", min_value=0.1, max_value=10.0, value=length_penalty)


generation_config = {
    "max_length": max_length,
    "num_beams": num_beams,
    "do_sample": True,
    "temperature": temperature,
    "top_k": top_k,
    "top_p": top_p,
    "repetition_penalty": repetition_penalty,
    "length_penalty": length_penalty,
    "early_stopping": True
}

# Streamlit app
st.title("SabiYarn-125M : Generates text in multiple Nigerian languages.")

st.write("**Supported Languages: English, Yoruba, Igbo, Hausa, Pidgin, Efik, Urhobo, Fulfulde, Fulah. \nResults might not be coherent for less represented languages (i.e Efik, \
Urhobo, Fulfulde, Fulah).**")
st.write("**It might take a while (~25s) to return an output on the first 'generate' click.**")
st.write("**For convenience, you can use chatgpt to copy text and evaluate model output.**")
st.write("-" * 50)

# async def generate_from_api(user_input, generation_config):
#     url = "https://pauljeffrey--sabiyarn-fastapi-app.modal.run/predict"
    
#     payload = {
#         "prompt": user_input,
#         "config": generation_config
#     }
    
#     headers = {
#         'Content-Type': 'application/json'
#     }
    
#     async with aiohttp.ClientSession() as session:
#         async with session.post(url, headers=headers, json=payload) as response:
#             return await response.text()

async def generate_from_api(user_input, generation_config):
    urls = [
        "https://pauljeffrey--sabiyarn-fastapi-app.modal.run/predict",
        "https://daveokpare--sabiyarn-fastapi-app.modal.run/predict",
        "https://damilojohn--sabiyarn-fastapi-app.modal.run/predict"
    ]

    payload = {
        "prompt": user_input,
        "config": generation_config
    }
    
    headers = {
        'Content-Type': 'application/json'
    }

    async with aiohttp.ClientSession() as session:
        for url in urls:
            try:
                async with session.post(url, headers=headers, json=payload) as response:
                    if response.status == 200:
                        return await response.text()
                    else:
                        print(f"Failed to fetch from {url} with status code {response.status}")
            except Exception as e:
                print(f"Error fetching from {url}: {e}")

    return "FAILED"


# Sample texts
sample_texts = {
    "select":"",
    "Hausa: Afirka tana da al'adu...": "Afirka tana da al'adu da harsuna masu yawa. Tana da albarkatu da wuraren yawon shakatawa masu ban mamaki.",
    "Yoruba: Ìmọ̀ sáyẹ́nsì àti...": "Ìmọ̀ sáyẹ́nsì àti tẹ̀knọ́lójì ń ṣe émi lóore tó níye lori ní Áfíríkà. Ó ń fún àwọn ènìyàn ní ànfààní láti dá irọyin àti kí wọ́n lè ṣe àwọn nǹkan tuntun.",
    "Efik: Oma Ede, Mi ji ogede...": "Oma Ede, Mi ji ogede mi a foroma orhorho edha meji ri eka. ",
    "Igbo: N'ala Igbo ...": "N'ala Igbo, ọtụtụ ndị mmadụ kwenyere na e nwere mmiri ara na elu-ilu",
    "urhobo: Eshare nana ri...":"Eshare nana ri vwo ẹguọnọ rẹ iyono rẹ Aristotle vẹ Plato na",
    "Efik: Ke eyo ...":"Ke eyo Jesus ye mme mbet esie, etop emi ama ada ifụre ọsọk mme Jew oro esịt okobụn̄ọde ke ntak idiọkido ke Israel, oro ẹkenyụn̄ ẹdude ke mfụhọ ke itie-ufụn mme nsunsu ido edinam Ido Ukpono Mme Jew eke akpa isua ikie.",
    "Tell me a story in pidgin": "Tell me a story Pidgin",
    "who are you?": "who are you?",
    "Speak Yoruba": "Speak Yoruba",
    "Translate 'Often, all Yoruba children...' to Yoruba": "Often, all Yoruba children take pride in speaking the Yoruba language.",
    "Classify the sentiment": "Anyi na-echefu oke ike.",
    "what is the topic of this text": "Africa Free Trade Zone: Kò sí ìdènà láti kó ọjà láti orílẹ̀èdè kan sí òmíràn",
    "diacritize this text: ": "E sun, Alaga, fun ise amalayi ti e n se ni Naijiria. E maa ba a lo, egbon!", 
    "clean this text": "Abin mamaki ne aikin da shugabaZn HNajeriya ybake yi. kCiF 39gaba Tda haRkGa sir!",
    "headline of this text": '** Sylvain Itté French ambassador don comot Niger Republic **. Sylvain Itté, di French ambassador for Niger don comot Niamey and currently e dey for flight from Ndjamena to Paris. Sylvain Itté, di French ambassador for Niger don comot Niamey very early dis morning and currently e dey for flight from Ndjamena to Paris.\n\nDi military detain Bazoum and im family for di presidential palace. Niger na former French colony, and France still get 1,500 sojas for di African country.\n\n"France don decide to withdraw dia ambassador. In di next hours our ambassador and several diplomats go return to France," Oga Macron tok.\n\nE add say di military co-operation dey "over" and French troops go leave in "di months to come".\n\n"Dis Sunday we celebrate one new step towards di sovereignty of Niger," di junta tok, for one statement wey AFP news agency quote.\n\nDi decision by Paris dey come afta months of hostility and protest against di presence of French for di kontri, wit regular demonstrations for di capital Niamey.\n\nDi move don scata France operations against Islamist militants for di wider Sahel region and Paris influence for there. But oga Macron tok say "putschists no go hold France hostage,"'
}

instruction_wrap = {
    "Translate 'Often, all Yoruba children...' to Yoruba":"<translate> Often, all Yoruba children take pride in speaking the Yoruba language. <yor>",
    "Tell me a story in pidgin": "<prompt> Tell me a story in pidgin <response>:",
    "Translate 'how are you?' to Yoruba": "<prompt> Translate 'how are you?' to Yoruba <response>:",
    "who are you?": "<prompt> who are you? <response>:",
    "Speak Yoruba": "<prompt> Speak Yoruba <response>:",
    "Classify the sentiment" : "<classify> Anyi na-echefu oke ike. <sentiment>",
    "clean this text": "<clean> Abin mamaki ne aikin da shugabaZn HNajeriya ybake yi. kCiF 39gaba Tda haRkGa sir! <pcm>",
    "diacritize this text: ": "<diacritize> E sun, Alaga, fun ise amalayi ti e n se ni Naijiria. E maa ba a lo, egbon!  <yor>",
    "what is the topic of this text": "<classify> Africa Free Trade Zone: Kò sí ìdènà láti kó ọjà láti orílẹ̀èdè kan sí òmíràn <topic>",
    'headline of this text': '<title> ** Sylvain Itté French ambassador don comot Niger Republic **. Sylvain Itté, di French ambassador for Niger don comot Niamey and currently e dey for flight from Ndjamena to Paris. Sylvain Itté, di French ambassador for Niger don comot Niamey very early dis morning and currently e dey for flight from Ndjamena to Paris.\n\nDi military detain Bazoum and im family for di presidential palace. Niger na former French colony, and France still get 1,500 sojas for di African country.\n\n"France don decide to withdraw dia ambassador. In di next hours our ambassador and several diplomats go return to France," Oga Macron tok.\n\nE add say di military co-operation dey "over" and French troops go leave in "di months to come".\n\n"Dis Sunday we celebrate one new step towards di sovereignty of Niger," di junta tok, for one statement wey AFP news agency quote.\n\nDi decision by Paris dey come afta months of hostility and protest against di presence of French for di kontri, wit regular demonstrations for di capital Niamey.\n\nDi move don scata France operations against Islamist militants for di wider Sahel region and Paris influence for there. But oga Macron tok say "putschists no go hold France hostage,"  <headline>',
}

# Task options
task_options = {
    "select": "{}",
    "Text Generation": "{}",
    "Translation": "<translate> {} ",
    "Sentiment Classification": "<classify> {} <sentiment>:",
    "Topic Classification": "<classify> {} <topic>",
    "Instruction Following" : "<prompt> {} <response>:",
    "Headline Generation": "<title> {} <headline>",
    "Text Diacritization": "<diacritize> {} ",
    "Text Cleaning": "<clean> {} "
}

# Language options for diacritize, translation and clean tasks
language_options = {
    "select": "",
    "Yoruba": "<yor>",
    "Hausa": "<hau>",
    "Ibo": "<ibo>",
    "Pidgin": "<pcm>",
    "Efik": "<efi>",
    "Urhobo": "<urh>",
    "Fulah": "<ful>"
}

        
# Dropdown for sample text
sample_text = st.selectbox("Select a sample text to test the model:", list(sample_texts.keys()))

# Dropdown for tasks
task = st.selectbox("Select a task for the model:", list(task_options.keys()))

# Conditionally show language options dropdown for diacritize and clean tasks
if task in ["Text Diacritization", "Text Cleaning", "Translation"]:
    language = st.selectbox("Select a Nigerian language:", list(language_options.keys()))
    task_value = f"{task_options[task]} {language_options[language]}"
else:
    task_value = task_options[task]

def wrap_text(text, task_value):
    tasks = ["<classify>", "<prompt>", "<clean>", "<title>", "<diacritize>", "<translate>"]
    if any(task in text for task in tasks):
        return text
    return task_value.format(text)


# Text input 
user_input = st.text_area("Enter text below **(PLEASE, FIRST READ ALL INSTRUCTIONS IN THE SIDEBAR CAREFULLY FOR THE BEST EXPERIENCE)**: ", sample_texts[sample_text])
user_input = instruction_wrap.get(sample_texts.get(user_input, user_input), user_input)
print("Final user input: ", user_input)

if st.button("Generate"):
    if user_input:
        st.spinner("Please wait...")
        # try:
        st.write("**Generated Text Below:**")
        wrapped_input = wrap_text(user_input, task_value)
        print("wrapped_input: ", wrapped_input)
        generation_config["max_new_tokens"]= min(max_new_tokens, 1024 - len(tokenizer.tokenize(wrapped_input)))
        start_time = time.time()
        # try:
            # Attempt the asynchronous API call
        generation_config["max_new_tokens"] = min(max_new_tokens, 1024 - len(tokenizer.tokenize(wrapped_input)))
        # generated_text = asyncio.run(generate_from_api(wrapped_input, generation_config))
        
        loop = asyncio.new_event_loop()
        asyncio.set_event_loop(loop)
        generated_text = loop.run_until_complete(generate_from_api(wrapped_input, generation_config))
        # except Exception as e:
            # print(f"API call failed: {e}. Using local model for text generation.")                
            # Use the locally loaded model for text generation
            # input_ids = tokenizer(wrapped_input, return_tensors="pt")["input_ids"].to(device)
            # output = model.generate(input_ids, **generation_config)
            # generated_text = tokenizer.decode(output[0], skip_special_tokens=True)
            
        if generated_text == "FAILED":
            input_ids = tokenizer(wrapped_input, return_tensors="pt")["input_ids"].to(device)
            output = model.generate(input_ids, **generation_config)
            generated_text = tokenizer.decode(output[0], skip_special_tokens=True)

        generated_text  = re.sub(r"\|(end_f_text|end_of_text|end_ofext|end_oftext)|:|`", " ", generated_text.split("|end_of_text|")[0])
        
        if task == "Sentiment Classification":
            if "negative" in generated_text.lower():
                generated_text = "Negative"
            elif "positive" in generated_text.lower():
                generated_text = "Positive"
            elif "neutral" in generated_text.lower():
                generated_text = "Neutral"

        elif task == "Topic Classification":
            generated_text  = generated_text.split(" ")[0][:20]
        elif task == "Translation":
            n_sentences = len(user_input)
            generated_text = ".".join(generated_text.split(".")[: n_sentences])
                
        full_output = st.empty()
       
        
        output = ""
        for next_token in tokenizer.tokenize(generated_text):
            output += tokenizer.convert_tokens_to_string([next_token])
            full_output.markdown(f"<div style='word-wrap: break-word;'>{output}</div>", unsafe_allow_html=True)
            # full_output.text(output)
            time.sleep(0.1)
        end_time = time.time()
        time_diff = end_time - start_time
        st.write("Time taken: ", time_diff , "seconds.")
        # except Exception as e:
        #     st.error(f"Error during text generation: {e}")
    else:
        st.write("Please enter some text to generate.")