app.py

import gradio as gr
import spaces
import torch
import os
from transformers import AutoTokenizer, AutoModel
import plotly.graph_objects as go

TOKEN = os.getenv("HF_TOKEN")

default_model_name = "mistralai/Mistral-7B-Instruct-v0.1"
tokenizer = None
model = None

@spaces.GPU(duration=300)
def get_embedding(text, model_repo):
    global tokenizer, model
    
    if tokenizer is None or model is None or model.name_or_path != model_repo:
        try:
            tokenizer = AutoTokenizer.from_pretrained(model_repo)
            model = AutoModel.from_pretrained(model_repo, torch_dtype=torch.float16).cuda()
            
            # Set pad token to eos token if not defined
            if tokenizer.pad_token is None:
                tokenizer.pad_token = tokenizer.eos_token
            
            model.resize_token_embeddings(len(tokenizer))
        except Exception as e:
            return f"Error loading model: {str(e)}"
    
    inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True, max_length=512).to('cuda')
    with torch.no_grad():
        outputs = model(**inputs)
    return outputs.last_hidden_state.mean(dim=1).squeeze().cpu().numpy()

def reduce_to_3d(embedding):
    return embedding[:3]

@spaces.GPU
def compare_embeddings(model_repo, *texts):
    if not model_repo:
        model_repo = default_model_name
    
    embeddings = []
    for text in texts:
        if text.strip():
            emb = get_embedding(text, model_repo)
            if isinstance(emb, str):  # Error message
                return emb
            embeddings.append(emb)
    
    embeddings_3d = [reduce_to_3d(emb) for emb in embeddings]
    
    fig = go.Figure()
    
    for i, emb in enumerate(embeddings_3d):
        fig.add_trace(go.Scatter3d(x=[0, emb[0]], y=[0, emb[1]], z=[0, emb[2]], 
                                   mode='lines+markers', name=f'Text {i+1}'))
    
    fig.update_layout(scene=dict(xaxis_title='X', yaxis_title='Y', zaxis_title='Z'))
    
    return fig

def generate_text_boxes(n):
    return [gr.Textbox(label=f"Text {i+1}", visible=(i < n)) for i in range(10)]

with gr.Blocks() as iface:
    gr.Markdown("# 3D Embedding Comparison")
    gr.Markdown("Compare the embeddings of multiple strings visualized in 3D space using a custom model.")
    
    model_repo_input = gr.Textbox(label="Model Repository", value=default_model_name, placeholder="Enter the model repository (e.g., mistralai/Mistral-7B-Instruct-v0.3)")
    num_texts = gr.Slider(minimum=2, maximum=10, step=1, value=2, label="Number of texts to compare")
    
    with gr.Column() as input_column:
        text_boxes = generate_text_boxes(2)
    
    output = gr.Plot()
    
    compare_button = gr.Button("Compare Embeddings")
    
    def update_interface(n):
        return [gr.update(visible=(i < n)) for i in range(10)]

    num_texts.change(
        update_interface,
        inputs=[num_texts],
        outputs=text_boxes
    )
    
    compare_button.click(
        compare_embeddings,
        inputs=[model_repo_input] + text_boxes,
        outputs=output
    )

iface.launch()