app.py

import spaces
import torch
import torch.nn.functional as F
from torch import Tensor
from transformers import AutoTokenizer, AutoModel
import threading
import queue
import gradio as gr
import os
import json
import numpy as np


title = """
# 👋🏻Welcome to 🙋🏻‍♂️Tonic's 📽️Nvidia 🛌🏻Embed V-1 !"""

description = """
You can use this Space to test out the current model [nvidia/NV-Embed-v1](https://huggingface.co/nvidia/NV-Embed-v1). 🐣a generalist embedding model that ranks No. 1 on the Massive Text Embedding Benchmark (MTEB benchmark)(as of May 24, 2024), with 56 tasks, encompassing retrieval, reranking, classification, clustering, and semantic textual similarity tasks.
You can also use 📽️Nvidia 🛌🏻Embed V-1 by cloning this space. 🧬🔬🔍 Simply click here: <a style="display:inline-block" href="https://huggingface.co/spaces/Tonic/NV-Embed?duplicate=true"><img src="https://img.shields.io/badge/-Duplicate%20Space-blue?labelColor=white&style=flat&logo=data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABAAAAAQCAYAAAAf8/9hAAAAAXNSR0IArs4c6QAAAP5JREFUOE+lk7FqAkEURY+ltunEgFXS2sZGIbXfEPdLlnxJyDdYB62sbbUKpLbVNhyYFzbrrA74YJlh9r079973psed0cvUD4A+4HoCjsA85X0Dfn/RBLBgBDxnQPfAEJgBY+A9gALA4tcbamSzS4xq4FOQAJgCDwV2CPKV8tZAJcAjMMkUe1vX+U+SMhfAJEHasQIWmXNN3abzDwHUrgcRGmYcgKe0bxrblHEB4E/pndMazNpSZGcsZdBlYJcEL9Afo75molJyM2FxmPgmgPqlWNLGfwZGG6UiyEvLzHYDmoPkDDiNm9JR9uboiONcBXrpY1qmgs21x1QwyZcpvxt9NS09PlsPAAAAAElFTkSuQmCC&logoWidth=14" alt="Duplicate Space"></a></h3> 
Join us : 🌟TeamTonic🌟 is always making cool demos! Join our active builder's🛠️community 👻  [![Join us on Discord](https://img.shields.io/discord/1109943800132010065?label=Discord&logo=discord&style=flat-square)](https://discord.gg/GWpVpekp) On 🤗Huggingface: [TeamTonic](https://huggingface.co/TeamTonic) & [MultiTransformer](https://huggingface.co/MultiTransformer) On 🌐Github: [Tonic-AI](https://github.com/tonic-ai) & contribute to 🌟 [MultiTonic](https://github.com/MultiTonic) 🤗Big thanks to Yuvi Sharma and all the folks at huggingface for the community grant 🤗
"""

tasks = {
        'ArguAna': 'Given a claim, find documents that refute the claim',
        'ClimateFEVER': 'Given a claim about climate change, retrieve documents that support or refute the claim',
        'DBPedia': 'Given a query, retrieve relevant entity descriptions from DBPedia',
        'FEVER': 'Given a claim, retrieve documents that support or refute the claim',
        'FiQA2018': 'Given a financial question, retrieve user replies that best answer the question',
        'HotpotQA': 'Given a multi-hop question, retrieve documents that can help answer the question',
        'MSMARCO': 'Given a web search query, retrieve relevant passages that answer the query',
        'NFCorpus': 'Given a question, retrieve relevant documents that best answer the question',
        'NQ': 'Given a question, retrieve Wikipedia passages that answer the question',
        'QuoraRetrieval': 'Given a question, retrieve questions that are semantically equivalent to the given question',
        'SCIDOCS': 'Given a scientific paper title, retrieve paper abstracts that are cited by the given paper',
        'SciFact': 'Given a scientific claim, retrieve documents that support or refute the claim',
        'Touche2020': 'Given a question, retrieve detailed and persuasive arguments that answer the question',
        'Natural Language Inference' : 'Retrieve semantically similar text',
        'Natural Language Inference' : 'Given a premise, retrieve a hypothesis that is entailed by the premise 20k',
        'PAQ, MSMARCO' : 'Given a web search query, retrieve relevant passages that answer the query',
        'PAQ, MSMARCO' : 'Given a question, retrieve passages that answer the question',
        'SQUAD' : 'Given a question, retrieve Wikipedia passages that answer the question' ,
        'StackExchange' : 'Given a question paragraph at StackExchange, retrieve a question duplicated paragraph',
        'Natural Question' : 'Given a question, retrieve Wikipedia passages that answer the question',
        'BioASQ' : 'Given a question, retrieve detailed question descriptions that are duplicates to the given question',
        'STS12, STS22, STSBenchmark' : 'Retrieve semantically similar text.',
        'AmazonCounterfactualClassification' : 'Classify a given Amazon customer review text as either counterfactual or not-counterfactual' , 
        'AmazonReviewsClassification' : 'Classify the given Amazon review into its appropriate rating category' , 
        'Banking77Classification' : 'Given a online banking query, find the corresponding intents',
        'EmotionClassification' : 'Classify the emotion expressed in the given Twitter message into one of the six emotions:anger, fear, joy, love, sadness, and surprise',
        'ImdbClassification': 'Classify the sentiment expressed in the given movie review text from the IMDB dataset',
        'MTOPIntentClassification' : 'Classify the intent of the given utterance in task-oriented conversation',
        'ToxicConversationsClassification' : 'Classify the given comments as either toxic or not toxic',
        'TweetSentimentExtractionClassification' : 'Classify the sentiment of a given tweet as either positive, negative, or neutral',
        'ArxivClusteringP2P' : 'Identify the main and secondary category of Arxiv papers based on the titles and abstracts',
        'ArxivClusteringS2S' : 'Identify the main and secondary category of Arxiv papers based on the titles',
        'BiorxivClusteringP2P' : 'Identify the main category of Biorxiv papers based on the titles and abstracts' ,
        'BiorxivClusteringS2S' : 'Identify the main category of Biorxiv papers based on the titles',
        'MedrxivClusteringP2P' : 'Identify the main category of Medrxiv papers based on the titles and abstracts',
        'MedrxivClusteringS2S' : 'Identify the main category of Medrxiv papers based on the titles',
        'TwentyNewsgroupsClustering' : 'Identify the topic or theme of the given news articles'
}

os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'max_split_size_mb:30'
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Define the model and tokenizer globally
tokenizer = AutoTokenizer.from_pretrained('nvidia/NV-Embed-v1', trust_remote_code=True)
model = AutoModel.from_pretrained('nvidia/NV-Embed-v1', trust_remote_code=True).to(device)

# Embedding requests and response queues
embedding_request_queue = queue.Queue()
embedding_response_queue = queue.Queue()

def clear_cuda_cache():
    torch.cuda.empty_cache()

def last_token_pool(last_hidden_states: Tensor, attention_mask: Tensor) -> Tensor:
    left_padding = (attention_mask[:, -1].sum() == attention_mask.shape[0])
    if left_padding:
        return last_hidden_states[:, -1]
    else:
        sequence_lengths = attention_mask.sum(dim=1) - 1
        batch_size = last_hidden_states.shape[0]
        return last_hidden_states[torch.arange(batch_size, device=last_hidden_states.device), sequence_lengths]

def format_response(embeddings):
    return {
        "data": [
            {
                "embedding": embeddings,
                "index": 0,
                "object": "embedding"
            }
        ],
        "model": "e5-mistral",
        "object": "list",
        "usage": {
            "prompt_tokens": 17,
            "total_tokens": 17
        }
    }

def embedding_worker():
    while True:
        # Wait for an item in the queue
        item = embedding_request_queue.get()
        if item is None:
            break
        selected_task, input_text = item
        embeddings = compute_embeddings(selected_task, input_text)
        formatted_response = format_response(embeddings)

        embedding_response_queue.put(formatted_response)
        embedding_request_queue.task_done()
        clear_cuda_cache()

@spaces.GPU
def compute_embeddings(selected_task, input_text):
    try:
        task_description = tasks[selected_task]
    except KeyError:
        print(f"Selected task not found: {selected_task}")
        return f"Error: Task '{selected_task}' not found. Please select a valid task."
    
    query_prefix = f"Instruct: {task_description}\nQuery: "
    queries = [input_text]
    
    # Get the embeddings
    query_embeddings = model.encode(queries, instruction=query_prefix, max_length=4096)
    # Normalize embeddings
    query_embeddings = F.normalize(query_embeddings, p=2, dim=1)
    embeddings_list = query_embeddings.detach().cpu().numpy().tolist()
    clear_cuda_cache()
    return embeddings_list

def compute_similarity(selected_task, sentence1, sentence2, extra_sentence1, extra_sentence2):
    try:
        task_description = tasks[selected_task]
    except KeyError:
        print(f"Selected task not found: {selected_task}")
        return f"Error: Task '{selected_task}' not found. Please select a valid task."
    
    # Compute embeddings for each sentence
    embeddings1 = compute_embeddings(selected_task, sentence1)
    embeddings2 = compute_embeddings(selected_task, sentence2)
    embeddings3 = compute_embeddings(selected_task, extra_sentence1)
    embeddings4 = compute_embeddings(selected_task, extra_sentence2)
    
    similarity1 = compute_cosine_similarity(embeddings1, embeddings2)
    similarity2 = compute_cosine_similarity(embeddings1, embeddings3)
    similarity3 = compute_cosine_similarity(embeddings1, embeddings4)

    similarity_scores = {"Similarity 1-2": similarity1, "Similarity 1-3": similarity2, "Similarity 1-4": similarity3}
    clear_cuda_cache()
    return similarity_scores

@spaces.GPU
def compute_cosine_similarity(emb1, emb2):
    tensor1 = torch.tensor(emb1).to(device).half()
    tensor2 = torch.tensor(emb2).to(device).half()
    similarity = F.cosine_similarity(tensor1, tensor2).item()
    clear_cuda_cache()
    return similarity

def app_interface():
    with gr.Blocks() as demo:
        gr.Markdown(title)
        gr.Markdown(description)

        with gr.Row():
            task_dropdown = gr.Dropdown(list(tasks.keys()), label="Select a Task", value=list(tasks.keys())[0])

        with gr.Tab("Sentence Similarity"):
            sentence1_box = gr.Textbox(label="'Focus Sentence' - The 'Subject'")
            sentence2_box = gr.Textbox(label="'Input Sentence' - 1")
            extra_sentence1_box = gr.Textbox(label="'Input Sentence' - 2")
            extra_sentence2_box = gr.Textbox(label="'Input Sentence' - 3")
            similarity_button = gr.Button("Compute Similarity")
            similarity_output = gr.Textbox(label="📽️Nvidia 🛌🏻Embed Similarity Scores")

            similarity_button.click(
                fn=compute_similarity,
                inputs=[task_dropdown, sentence1_box, sentence2_box, extra_sentence1_box, extra_sentence2_box],
                outputs=similarity_output
            )

    return demo

embedding_worker_thread = threading.Thread(target=embedding_worker, daemon=True)
embedding_worker_thread.start()

app_interface().queue()
app_interface().launch(share=True)