Hindi-CausalLM

A Hindi language generation model with the following specifications:

Model Architecture

• Type: Causal Language Model with Transformer architecture
• Hidden size: 768
• Layers: 12
• Attention heads: 16
• Key-value heads: 4 (using grouped-query attention)
• Position encoding: Rotary Position Embeddings (RoPE)
• Vocabulary size: 16000
• Parameters: ~100M
• Context window: 512 tokens
• Trained on: Large corpus of Hindi text

Training

The model was trained on a large corpus of Hindi text using a cosine learning rate schedule with warmup. Training utilized mixed-precision and distributed data parallel across multiple GPUs.

Usage

You can use this model with the following code:

import torch
import math
import os
from hindi_embeddings import SentencePieceTokenizerWrapper
from safetensors.torch import load_file
from torch import nn
from transformers import PreTrainedModel, PretrainedConfig


class ConvaiCausalLMConfig(PretrainedConfig):
    model_type = "convaicausallm"
    
    def __init__(
        self,
        vocab_size=16000,
        hidden_size=768,
        num_hidden_layers=12,
        num_attention_heads=16,
        num_key_value_heads=4,
        intermediate_size=3072,
        hidden_act="silu",
        max_position_embeddings=512,
        rope_theta=10000.0,  # Base parameter for RoPE
        **kwargs
    ):
        super().__init__(**kwargs)
        self.vocab_size = vocab_size
        self.hidden_size = hidden_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.num_key_value_heads = num_key_value_heads
        self.intermediate_size = intermediate_size
        self.hidden_act = hidden_act
        self.max_position_embeddings = max_position_embeddings
        self.rope_theta = rope_theta


def precompute_freqs_cis(dim, end, theta=10000.0):
    """Precompute the frequency tensor for complex exponentials (cos, sin)"""
    # Ensure dim is even for complex numbers
    assert dim % 2 == 0, "Dimension must be even"
    
    # Create position indices for caching
    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2).float() / dim))
    t = torch.arange(end).float()
    freqs = torch.outer(t, freqs)  # [end, dim/2]
    
    # Create complex exponentials (cos, sin pairs)
    cos, sin = torch.cos(freqs), torch.sin(freqs)
    return cos, sin


def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None):
    """Apply rotary position embeddings to q and k tensors"""
    # Extract shapes
    batch, seq_len, n_heads, head_dim = q.shape
    _, kv_seq_len, n_kv_heads, _ = k.shape
    
    # Handle position IDs or use sequential positions
    if position_ids is None:
        # Default: Just use sequential positions
        position_ids = torch.arange(seq_len, device=q.device)
        position_ids = position_ids.unsqueeze(0).expand(batch, -1)
        
    # Get the cosine and sine for the positions we're using
    cos = cos[position_ids].unsqueeze(-2)  # [batch, seq, 1, dim/2]
    sin = sin[position_ids].unsqueeze(-2)  # [batch, seq, 1, dim/2]
    
    # q and k must be arranged in pairs for rotation
    q_embed_dim = q.shape[-1]
    q_half_dim = q_embed_dim // 2
    
    # Split the embedding dimensions into pairs
    q_half1, q_half2 = q[..., :q_half_dim], q[..., q_half_dim:]
    k_half1, k_half2 = k[..., :q_half_dim], k[..., q_half_dim:]
    
    # Apply rotary embeddings to each pair of dimensions
    # For each pair (a, b), we compute (a*cos - b*sin, a*sin + b*cos)
    q_out_half1 = q_half1 * cos - q_half2 * sin
    q_out_half2 = q_half1 * sin + q_half2 * cos
    k_out_half1 = k_half1 * cos - k_half2 * sin
    k_out_half2 = k_half1 * sin + k_half2 * cos
    
    # Concatenate back to original shape
    q_out = torch.cat([q_out_half1, q_out_half2], dim=-1)
    k_out = torch.cat([k_out_half1, k_out_half2], dim=-1)
    
    return q_out, k_out


class GroupedQueryAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.num_kv_heads = config.num_key_value_heads
        self.head_dim = config.hidden_size // config.num_attention_heads
        
        # For MQA/GQA support
        self.num_key_value_groups = self.num_heads // self.num_kv_heads
        
        self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.head_dim)
        self.k_proj = nn.Linear(config.hidden_size, self.num_kv_heads * self.head_dim)
        self.v_proj = nn.Linear(config.hidden_size, self.num_kv_heads * self.head_dim)
        self.o_proj = nn.Linear(config.hidden_size, config.hidden_size)
        
        # Precompute rotary position encoding frequencies
        max_seq_len = config.max_position_embeddings
        self.max_seq_len = max_seq_len
        
        # Register frequencies as buffers
        cos, sin = precompute_freqs_cis(self.head_dim, max_seq_len, config.rope_theta)
        self.register_buffer("cos", cos)  # [max_seq_len, dim/2]
        self.register_buffer("sin", sin)  # [max_seq_len, dim/2]
        
        # Create causal mask for attention
        self.register_buffer(
            "causal_mask", 
            torch.triu(torch.ones(max_seq_len, max_seq_len) * -1e9, diagonal=1)
        )

    def forward(self, hidden_states, attention_mask=None):
        batch_size, seq_len, _ = hidden_states.size()
        
        # Project queries, keys, values
        q = self.q_proj(hidden_states)
        k = self.k_proj(hidden_states)
        v = self.v_proj(hidden_states)
        
        # Reshape for attention computation
        q = q.view(batch_size, seq_len, self.num_heads, self.head_dim)
        k = k.view(batch_size, seq_len, self.num_kv_heads, self.head_dim)
        v = v.view(batch_size, seq_len, self.num_kv_heads, self.head_dim)
        
        # Apply rotary position embeddings
        q_rotary, k_rotary = apply_rotary_pos_emb(q, k, self.cos, self.sin)
        
        # Reshape for attention computation
        q_rotary = q_rotary.transpose(1, 2)  # [batch, heads, seq, dim]
        k_rotary = k_rotary.transpose(1, 2)  # [batch, kv_heads, seq, dim]
        v = v.transpose(1, 2)  # [batch, kv_heads, seq, dim]
        
        # Handle Multi-Query Attention / Grouped-Query Attention
        if self.num_key_value_groups > 1:
            # Repeat k, v for each query in the group
            k_rotary = k_rotary.repeat_interleave(self.num_key_value_groups, dim=1)
            v = v.repeat_interleave(self.num_key_value_groups, dim=1)
        
        # Compute attention scores
        attn_scores = torch.matmul(q_rotary, k_rotary.transpose(-1, -2)) / (self.head_dim ** 0.5)
        
        # Apply causal mask - only attend to previous tokens
        causal_mask = self.causal_mask[:seq_len, :seq_len]
        attn_scores = attn_scores + causal_mask
        
        # Apply attention mask if provided
        if attention_mask is not None:
            attn_scores = attn_scores + attention_mask
            
        # Normalize the attention scores to probabilities
        attn_probs = torch.softmax(attn_scores, dim=-1)
        
        # Apply attention to values
        context = torch.matmul(attn_probs, v)  # [b, n_heads, seq, head_dim]
        
        # Reshape back to [batch_size, seq_length, hidden_size]
        context = context.transpose(1, 2).contiguous()
        context = context.view(batch_size, seq_len, -1)
        
        # Final projection
        output = self.o_proj(context)
        
        return output


class ConvaiCausalLM(PreTrainedModel):
    config_class = ConvaiCausalLMConfig
    
    def __init__(self, config):
        super().__init__(config)
        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
        self.layers = nn.ModuleList([
            nn.ModuleDict({
                "self_attn": GroupedQueryAttention(config),
                "mlp": nn.Sequential(
                    nn.Linear(config.hidden_size, config.intermediate_size),
                    nn.SiLU(),
                    nn.Linear(config.intermediate_size, config.hidden_size)
                ),
                "input_layernorm": nn.LayerNorm(config.hidden_size),
                "post_attention_layernorm": nn.LayerNorm(config.hidden_size)
            }) for _ in range(config.num_hidden_layers)
        ])
        self.norm = nn.LayerNorm(config.hidden_size)
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        
        # Initialize weights
        self.apply(self._init_weights)
        
    def _init_weights(self, module):
        if isinstance(module, nn.Linear):
            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
            if module.bias is not None:
                torch.nn.init.zeros_(module.bias)
        elif isinstance(module, nn.Embedding):
            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
    
    def _prepare_attention_mask(self, attention_mask, input_shape, device):
        # Prepare masks for attention
        if attention_mask is None:
            attention_mask = torch.ones(input_shape, device=device)
            
        # Make broadcastable shape: [batch, 1, 1, seq_len]
        extended_mask = attention_mask.unsqueeze(1).unsqueeze(2)
        
        # Convert to additive mask (0 for valid, -10000 for masked)
        extended_mask = (1.0 - extended_mask) * -10000.0
        
        return extended_mask
    
    def forward(self, input_ids, attention_mask=None):
        batch_size, seq_len = input_ids.shape
        device = input_ids.device
        
        # Prepare attention mask
        if attention_mask is not None:
            attention_mask = self._prepare_attention_mask(
                attention_mask, (batch_size, seq_len), device
            )
        
        # Get embeddings
        hidden_states = self.embed_tokens(input_ids)
        
        # Apply each layer
        for layer in self.layers:
            residual = hidden_states
            
            # First norm and attention
            hidden_states = layer["input_layernorm"](hidden_states)
            hidden_states = layer["self_attn"](hidden_states, attention_mask)
            hidden_states = residual + hidden_states
            
            # Second norm and MLP
            residual = hidden_states
            hidden_states = layer["post_attention_layernorm"](hidden_states)
            hidden_states = layer["mlp"](hidden_states)
            hidden_states = residual + hidden_states
            
        # Final norm
        hidden_states = self.norm(hidden_states)
        
        # Compute logits
        logits = self.lm_head(hidden_states)
        
        return logits


class HindiLLMGenerator:
    def __init__(self, model_path, device=None):
        # Set device
        if device is None:
            self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        else:
            self.device = torch.device(device)
            
        print(f"Using device: {self.device}")
        
        # Load tokenizer
        tokenizer_path = os.path.join(model_path, "tokenizer.model")
        self.tokenizer = SentencePieceTokenizerWrapper(tokenizer_path)
        
        # Load model config
        config_path = os.path.join(model_path, "config.json")
        import json
        with open(config_path, 'r') as f:
            config_dict = json.load(f)
            
        self.config = ConvaiCausalLMConfig(**config_dict)
        
        # Load model - try safetensors first, fall back to PyTorch bin if needed
        safetensors_path = os.path.join(model_path, "model.safetensors")
        pytorch_path = os.path.join(model_path, "pytorch_model.bin")
        
        self.model = ConvaiCausalLM(self.config)
        
        # Check which format is available and load accordingly
        if os.path.exists(safetensors_path):
            print(f"Loading model from SafeTensors")
            state_dict = load_file(safetensors_path, device="cpu")
            self.model.load_state_dict(state_dict)
        elif os.path.exists(pytorch_path):
            print(f"Loading model from PyTorch bin")
            self.model.load_state_dict(torch.load(pytorch_path, map_location="cpu"))
        
        # Move model to device and set to evaluation mode
        self.model.to(self.device)
        self.model.eval()
    
    def generate(self, prompt, max_length=100, temperature=0.8, top_k=50, top_p=0.9, 
                 repetition_penalty=1.1, do_sample=True):
        # Tokenize the prompt
        input_ids = self.tokenizer.sp_model.EncodeAsIds(prompt)
        input_tensor = torch.tensor([input_ids], dtype=torch.long).to(self.device)
        
        # Start with the input tensor
        output_sequence = input_tensor.clone()
        
        # Generate tokens one by one
        for _ in range(max_length - len(input_ids)):
            with torch.no_grad():
                # Get the model's output for the current sequence
                outputs = self.model(output_sequence)
                next_token_logits = outputs[0, -1, :]
                
                # Apply temperature
                if temperature > 0:
                    next_token_logits = next_token_logits / temperature
                
                # Apply repetition penalty
                if repetition_penalty > 1.0:
                    for token_id in output_sequence[0].tolist():
                        next_token_logits[token_id] /= repetition_penalty
                
                # Filter with top-k sampling
                if top_k > 0:
                    top_k_values, top_k_indices = torch.topk(next_token_logits, top_k)
                    next_token_logits = torch.full_like(next_token_logits, float('-inf'))
                    next_token_logits.scatter_(0, top_k_indices, top_k_values)
                
                # Filter with top-p/nucleus sampling
                if top_p < 1.0 and do_sample:
                    sorted_logits, sorted_indices = torch.sort(next_token_logits, descending=True)
                    cumulative_probs = torch.cumsum(torch.softmax(sorted_logits, dim=-1), dim=-1)
                    
                    # Remove tokens with cumulative probability above the threshold
                    sorted_indices_to_remove = cumulative_probs > top_p
                    # Shift the indices to the right to keep the first token above the threshold
                    sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
                    sorted_indices_to_remove[..., 0] = 0
                    
                    indices_to_remove = sorted_indices[sorted_indices_to_remove]
                    next_token_logits[indices_to_remove] = float('-inf')
                
                # Sample or choose the next token
                if do_sample:
                    probs = torch.softmax(next_token_logits, dim=-1)
                    next_token = torch.multinomial(probs, num_samples=1)
                else:
                    next_token = torch.argmax(next_token_logits, dim=-1).unsqueeze(0)
                
                # Add the next token to the sequence
                output_sequence = torch.cat([output_sequence, next_token.unsqueeze(0)], dim=1)
                
                # Check if we've generated an end token
                if next_token.item() == self.tokenizer.eos_token_id:
                    break
        
        # Decode the generated sequence
        generated_ids = output_sequence[0].tolist()
        generated_text = self.tokenizer.sp_model.DecodeIds(generated_ids)
        
        return generated_text

# Example usage
if __name__ == "__main__":
    generator = HindiLLMGenerator("path/to/model")
    result = generator.generate("भारत एक विशाल देश है")
    print(result)

Example Prompts

Try the model with these example prompts:

भारत एक विशाल देश है
मुझे हिंदी में एक कहानी सुनाओ
आज का मौसम बहुत अच्छा है
हिंदी साहित्य की प्रमुख विशेषताएं

Capabilities

This model can:

• Generate coherent Hindi text
• Continue text from a given prompt
• Create stories, explanations, and other content in Hindi

Limitations

• Performance varies based on the similarity of the input to the training data
• May occasionally generate repetitive content for longer texts
• May produce grammatically incorrect Hindi in some contexts
• Has no knowledge of events beyond its training corpus

Intended Use

This model is intended for Hindi language generation tasks, creative writing assistance, and as a foundation for fine-tuning on specific tasks.

Ethical Considerations

Users should be aware that like all language models, this model may reproduce biases or generate problematic content in certain contexts.