app.py

import torch
from torch import nn
from tokenizers import Tokenizer
import torch.nn.functional as F
import gradio as gr

# Enhanced Model with Custom Attention
class RachanaLLM(nn.Module):
    def __init__(self, vocab_size=50000, embed_dim=284, num_heads=4, num_layers=4, dropout=0.2, max_len=256):
        super(RachanaLLM, self).__init__()
        self.embed_dim = embed_dim
        self.embedding = nn.Embedding(vocab_size, embed_dim)
        self.positional_encoding = nn.Parameter(torch.zeros(1, max_len, embed_dim))
        encoder_layer = nn.TransformerEncoderLayer(d_model=embed_dim, nhead=num_heads, dropout=dropout, batch_first=True)
        self.transformer_encoder = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)
        self.output_layer = nn.Linear(embed_dim, vocab_size)

    def forward(self, src, src_mask=None):
        src = self.embedding(src) * torch.sqrt(torch.tensor(self.embed_dim, dtype=torch.float32))
        src = src + self.positional_encoding[:, :src.size(1)]
        output = self.transformer_encoder(src, src_key_padding_mask=src_mask)
        return self.output_layer(output)

# Load model and tokenizer
checkpoint = torch.load("best_model.pth")  # Use your actual path
model = RachanaLLM()
model.load_state_dict(checkpoint['model_state'] if 'model_state' in checkpoint else checkpoint)
model.eval()
tokenizer = Tokenizer.from_file("telugu_tokenizer_50k.json")  # Use your actual path

# Advanced Beam Search with Temperature, Top-K Sampling, and Dynamic Repetition Penalty
def beam_search_decoder(model, tokenizer, input_ids, beam_width=3, max_length=20, temperature=1.0, top_k=5, repetition_penalty=1.2):
    model.eval()
    with torch.no_grad():
        sequences = [[input_ids, 0.0]]
        
        for _ in range(max_length):
            all_candidates = []
            for seq, score in sequences:
                outputs = model(seq)
                logits = outputs[:, -1] / temperature

                # If a token has been used already, penalize its logit
                for token_id in set(seq[0].tolist()):
                    if token_id in seq[0]:
                        logits[0][token_id] /= repetition_penalty
                
                top_logits, top_indices = torch.topk(logits, top_k, dim=-1)
                softmax_scores = F.log_softmax(top_logits, dim=-1)

                for i in range(top_k):
                    next_token_id = top_indices[0][i].item()
                    if seq[0].tolist().count(next_token_id) < 2: # This ensures a token is penalized only if it appears more than once.
                        candidate = [torch.cat([seq, top_indices[:, i:i+1]], dim=1), score + softmax_scores[0][i].item()]
                        all_candidates.append(candidate)
                    
            # Sort candidates by score
            ordered = sorted(all_candidates, key=lambda tup: tup[1], reverse=True)
            sequences = ordered[:beam_width]
        
        best_seq = sequences[0][0]
        result_text = tokenizer.decode(best_seq[0].tolist())
        return result_text



# Predict function for Gradio Interface
def predict(input_sentence):
    input_ids = torch.tensor(tokenizer.encode(input_sentence).ids).unsqueeze(0)
    generated_text = beam_search_decoder(model, tokenizer, input_ids, temperature=0.8, top_k=10)
    return generated_text

# Gradio Interface
iface = gr.Interface(
    fn=predict,
    inputs=gr.Textbox(lines=2, placeholder="Type a sentence..."),  # Updated line
    outputs="text",
    examples=[
        "ఈ రోజు వాతావరణం చాలా బాగుంది.",
        "సినిమా బాగుందా లేదా చెప్పు!",
        "ఆలయ అధికారులు దర్శన ఏర్పాట్లు చేశారు."
    ],
    title="Rachana LLM"
)

iface.launch(share=True)