model.py

# =============================================================================
# Gemma-Style Model Architecture (110M Parameters)
# =============================================================================
"""
This module implements a Gemma-style transformer architecture with:
- RMSNorm (Root Mean Square Layer Normalization)
- RoPE (Rotary Position Embeddings)
- GeGLU (Gated Linear Unit with GELU activation)
- GQA (Grouped Query Attention)

Architecture Specifications:
- hidden_size: 768
- num_layers: 12
- intermediate_size: 3072
- num_attention_heads: 12
- num_key_value_heads: 4 (GQA ratio of 3:1)
- vocab_size: 50257 (GPT-2 tokenizer)
"""

import math
import logging
from typing import Optional, Tuple, Dict, Any, Union, List

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch import Tensor

from transformers import PretrainedConfig, PreTrainedModel
from transformers.modeling_outputs import CausalLMOutputWithPast

logger = logging.getLogger(__name__)


class GemmaConfig(PretrainedConfig):
    """
    Configuration class for the Gemma-style model.
    
    Inherits from HuggingFace PretrainedConfig for compatibility with
    AutoConfig and the transformers ecosystem.
    """
    model_type = "gemma_custom"
    
    def __init__(
        self,
        hidden_size: int = 768,
        num_layers: int = 12,
        intermediate_size: int = 3072,
        num_attention_heads: int = 12,
        num_key_value_heads: int = 4,
        max_position_embeddings: int = 1024,
        vocab_size: int = 50257,
        rope_theta: float = 10000.0,
        rms_norm_eps: float = 1e-6,
        hidden_act: str = "gelu_pytorch_tanh",
        attention_dropout: float = 0.0,
        hidden_dropout: float = 0.0,
        tie_word_embeddings: bool = True,
        initializer_range: float = 0.02,
        **kwargs
    ):
        self.hidden_size = hidden_size
        self.num_layers = num_layers
        self.intermediate_size = intermediate_size
        self.num_attention_heads = num_attention_heads
        self.num_key_value_heads = num_key_value_heads
        self.max_position_embeddings = max_position_embeddings
        self.vocab_size = vocab_size
        self.rope_theta = rope_theta
        self.rms_norm_eps = rms_norm_eps
        self.hidden_act = hidden_act
        self.attention_dropout = attention_dropout
        self.hidden_dropout = hidden_dropout
        self.initializer_range = initializer_range
        
        # Derived attributes
        self.head_dim = self.hidden_size // self.num_attention_heads
        self.num_key_value_groups = self.num_attention_heads // self.num_key_value_heads
        
        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)


class RMSNorm(nn.Module):
    """
    Root Mean Square Layer Normalization.
    
    RMSNorm(x) = x * rsqrt(mean(x^2) + eps) * weight
    """
    
    def __init__(self, hidden_size: int, eps: float = 1e-6):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states: Tensor) -> Tensor:
        input_dtype = hidden_states.dtype
        hidden_states = hidden_states.to(torch.float32)
        variance = hidden_states.pow(2).mean(-1, keepdim=True)
        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
        return self.weight * hidden_states.to(input_dtype)


class RotaryEmbedding(nn.Module):
    """Rotary Position Embedding (RoPE)."""
    
    def __init__(self, dim: int, max_position_embeddings: int = 1024, theta: float = 10000.0):
        super().__init__()
        self.dim = dim
        self.max_position_embeddings = max_position_embeddings
        self.theta = theta
        
        inv_freq = 1.0 / (self.theta ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self._set_cos_sin_cache(max_position_embeddings)

    def _set_cos_sin_cache(self, seq_len: int):
        t = torch.arange(seq_len, dtype=torch.float32)
        freqs = torch.einsum("i,j->ij", t, self.inv_freq)
        emb = torch.cat((freqs, freqs), dim=-1)
        self.register_buffer("cos_cached", emb.cos(), persistent=False)
        self.register_buffer("sin_cached", emb.sin(), persistent=False)

    def forward(self, x: Tensor, position_ids: Tensor) -> Tuple[Tensor, Tensor]:
        seq_len = position_ids.max() + 1
        if seq_len > self.cos_cached.shape[0]:
            self._set_cos_sin_cache(seq_len)
            self.cos_cached = self.cos_cached.to(x.device)
            self.sin_cached = self.sin_cached.to(x.device)
        
        cos = self.cos_cached[position_ids].unsqueeze(2)
        sin = self.sin_cached[position_ids].unsqueeze(2)
        return cos.to(x.dtype), sin.to(x.dtype)


def rotate_half(x: Tensor) -> Tensor:
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)


def apply_rotary_pos_emb(q: Tensor, k: Tensor, cos: Tensor, sin: Tensor) -> Tuple[Tensor, Tensor]:
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


class GemmaMLP(nn.Module):
    """Gemma-style MLP with GeGLU activation."""
    
    def __init__(self, config: GemmaConfig):
        super().__init__()
        self.gate_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
        self.up_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
        self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
        self.act_fn = nn.GELU(approximate="tanh")

    def forward(self, x: Tensor) -> Tensor:
        gate = self.act_fn(self.gate_proj(x))
        up = self.up_proj(x)
        return self.down_proj(gate * up)


class GemmaAttention(nn.Module):
    """Grouped Query Attention (GQA) module."""
    
    def __init__(self, config: GemmaConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = config.head_dim
        self.num_key_value_heads = config.num_key_value_heads
        self.num_key_value_groups = config.num_key_value_groups
        self.attention_dropout = config.attention_dropout
        
        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
        
        self.rotary_emb = RotaryEmbedding(
            self.head_dim,
            max_position_embeddings=config.max_position_embeddings,
            theta=config.rope_theta
        )

    def forward(
        self,
        hidden_states: Tensor,
        attention_mask: Optional[Tensor] = None,
        position_ids: Optional[Tensor] = None,
    ) -> Tensor:
        batch_size, seq_len, _ = hidden_states.shape
        
        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)
        
        query_states = query_states.view(batch_size, seq_len, self.num_heads, self.head_dim)
        key_states = key_states.view(batch_size, seq_len, self.num_key_value_heads, self.head_dim)
        value_states = value_states.view(batch_size, seq_len, self.num_key_value_heads, self.head_dim)
        
        cos, sin = self.rotary_emb(query_states, position_ids)
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
        
        query_states = query_states.transpose(1, 2)
        key_states = key_states.transpose(1, 2)
        value_states = value_states.transpose(1, 2)
        
        if self.num_key_value_groups > 1:
            key_states = key_states.repeat_interleave(self.num_key_value_groups, dim=1)
            value_states = value_states.repeat_interleave(self.num_key_value_groups, dim=1)
        
        scale = 1.0 / math.sqrt(self.head_dim)
        attn_weights = torch.matmul(query_states, key_states.transpose(-2, -1)) * scale
        
        if attention_mask is not None:
            attn_weights = attn_weights + attention_mask
        
        attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
        attn_weights = F.dropout(attn_weights, p=self.attention_dropout, training=self.training)
        
        attn_output = torch.matmul(attn_weights, value_states)
        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.reshape(batch_size, seq_len, self.hidden_size)
        
        return self.o_proj(attn_output)


class GemmaDecoderLayer(nn.Module):
    """Single transformer decoder layer with pre-norm architecture."""
    
    def __init__(self, config: GemmaConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.self_attn = GemmaAttention(config, layer_idx)
        self.mlp = GemmaMLP(config)

    def forward(
        self,
        hidden_states: Tensor,
        attention_mask: Optional[Tensor] = None,
        position_ids: Optional[Tensor] = None,
    ) -> Tensor:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        hidden_states = self.self_attn(hidden_states, attention_mask, position_ids)
        hidden_states = residual + hidden_states
        
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states
        
        return hidden_states


class GemmaModel(nn.Module):
    """Core Gemma transformer model (without LM head)."""
    
    def __init__(self, config: GemmaConfig):
        super().__init__()
        self.config = config
        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
        self.layers = nn.ModuleList([
            GemmaDecoderLayer(config, layer_idx) 
            for layer_idx in range(config.num_layers)
        ])
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.gradient_checkpointing = False

    def forward(
        self,
        input_ids: Tensor,
        attention_mask: Optional[Tensor] = None,
        position_ids: Optional[Tensor] = None,
    ) -> Tensor:
        batch_size, seq_len = input_ids.shape
        hidden_states = self.embed_tokens(input_ids)
        
        if position_ids is None:
            position_ids = torch.arange(seq_len, device=input_ids.device).unsqueeze(0).expand(batch_size, -1)
        
        causal_mask = self._create_causal_mask(seq_len, hidden_states.device, hidden_states.dtype)
        
        for layer in self.layers:
            if self.gradient_checkpointing and self.training:
                hidden_states = torch.utils.checkpoint.checkpoint(
                    layer, hidden_states, causal_mask, position_ids, use_reentrant=False
                )
            else:
                hidden_states = layer(hidden_states, causal_mask, position_ids)
        
        return self.norm(hidden_states)

    def _create_causal_mask(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> Tensor:
        mask = torch.full((seq_len, seq_len), float("-inf"), device=device, dtype=dtype)
        mask = torch.triu(mask, diagonal=1)
        return mask.unsqueeze(0).unsqueeze(0)


class GemmaForCausalLM(PreTrainedModel):
    """
    Gemma model with language modeling head for causal text generation.
    
    Inherits from HuggingFace PreTrainedModel for full compatibility with
    AutoModelForCausalLM and the transformers ecosystem.
    """
    config_class = GemmaConfig
    supports_gradient_checkpointing = True
    _no_split_modules = ["GemmaDecoderLayer"]
    _supports_param_buffer_assignment = False  # Fix for accelerate weight loading
    
    def __init__(self, config: GemmaConfig):
        super().__init__(config)
        self.config = config
        
        self.model = GemmaModel(config)
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        
        if config.tie_word_embeddings:
            self.lm_head.weight = self.model.embed_tokens.weight
        
        self.post_init()

    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path, *args, **kwargs):
        """
        Custom from_pretrained that properly loads weights for this custom model.
        This overrides the default behavior to ensure weights are loaded correctly.
        """
        import os
        from huggingface_hub import hf_hub_download
        
        # Get config
        trust_remote_code = kwargs.pop("trust_remote_code", True)
        torch_dtype = kwargs.pop("torch_dtype", None)
        device_map = kwargs.pop("device_map", None)
        
        # Load config
        config = cls.config_class.from_pretrained(
            pretrained_model_name_or_path, 
            trust_remote_code=trust_remote_code,
            **kwargs
        )
        
        # Create model
        model = cls(config)
        
        # Find weight file
        if os.path.isdir(pretrained_model_name_or_path):
            weight_file = os.path.join(pretrained_model_name_or_path, "pytorch_model.bin")
        else:
            # Download from hub
            weight_file = hf_hub_download(
                repo_id=pretrained_model_name_or_path,
                filename="pytorch_model.bin"
            )
        
        # Load weights
        state_dict = torch.load(weight_file, map_location="cpu")
        model.load_state_dict(state_dict, strict=False)
        
        # Handle dtype and device
        if torch_dtype is not None:
            model = model.to(torch_dtype)
        if device_map == "auto":
            if torch.cuda.is_available():
                model = model.to("cuda")
        elif device_map is not None:
            model = model.to(device_map)
        
        return model

    def get_input_embeddings(self):
        return self.model.embed_tokens

    def set_input_embeddings(self, value):
        self.model.embed_tokens = value

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def _init_weights(self, module: nn.Module):
        std = self.config.initializer_range
        if isinstance(module, nn.Linear):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=std)

    def num_parameters(self, only_trainable: bool = False) -> int:
        return sum(p.numel() for p in self.parameters() if not only_trainable or p.requires_grad)

    def forward(
        self,
        input_ids: Tensor,
        attention_mask: Optional[Tensor] = None,
        position_ids: Optional[Tensor] = None,
        labels: Optional[Tensor] = None,
        return_dict: Optional[bool] = None,
        **kwargs
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        
        hidden_states = self.model(input_ids, attention_mask, position_ids)
        logits = self.lm_head(hidden_states)
        
        loss = None
        if labels is not None:
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            loss_fct = nn.CrossEntropyLoss()
            loss = loss_fct(
                shift_logits.view(-1, self.config.vocab_size),
                shift_labels.view(-1)
            )
        
        if not return_dict:
            output = (logits,)
            return ((loss,) + output) if loss is not None else output
        
        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=None,
            hidden_states=None,
            attentions=None,
        )

    def prepare_inputs_for_generation(
        self, input_ids, past_key_values=None, attention_mask=None, **kwargs
    ):
        if past_key_values is not None:
            input_ids = input_ids[:, -1:]
        
        position_ids = kwargs.get("position_ids", None)
        if attention_mask is not None and position_ids is None:
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 1)
            if past_key_values:
                position_ids = position_ids[:, -1].unsqueeze(-1)
        
        return {
            "input_ids": input_ids,
            "attention_mask": attention_mask,
            "position_ids": position_ids,
            "past_key_values": past_key_values,
        }

    def generate(
        self,
        input_ids: Tensor,
        max_new_tokens: int = 50,
        temperature: float = 1.0,
        top_k: int = 50,
        top_p: float = 0.95,
        do_sample: bool = True,
        eos_token_id: Optional[int] = None,
        **kwargs
    ) -> Tensor:
        """Custom generate method for simple autoregressive generation."""
        self.eval()
        
        for _ in range(max_new_tokens):
            with torch.no_grad():
                outputs = self.forward(input_ids)
                next_token_logits = outputs.logits[:, -1, :] / temperature
            
            if top_k > 0:
                indices_to_remove = next_token_logits < torch.topk(next_token_logits, top_k)[0][..., -1, None]
                next_token_logits[indices_to_remove] = float("-inf")
            
            if top_p < 1.0:
                sorted_logits, sorted_indices = torch.sort(next_token_logits, descending=True)
                cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
                sorted_indices_to_remove = cumulative_probs > top_p
                sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
                sorted_indices_to_remove[..., 0] = 0
                indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
                next_token_logits[indices_to_remove] = float("-inf")
            
            if do_sample:
                probs = F.softmax(next_token_logits, dim=-1)
                next_token = torch.multinomial(probs, num_samples=1)
            else:
                next_token = torch.argmax(next_token_logits, dim=-1, keepdim=True)
            
            input_ids = torch.cat([input_ids, next_token], dim=-1)
            
            if eos_token_id is not None and (next_token == eos_token_id).all():
                break
        
        return input_ids


def create_model_from_config(config_dict: Optional[Dict[str, Any]] = None) -> GemmaForCausalLM:
    """Factory function to create a model from a configuration dictionary."""
    if config_dict is None:
        config = GemmaConfig()
    else:
        config = GemmaConfig(**{k: v for k, v in config_dict.items() if hasattr(GemmaConfig, k) or k in ['hidden_size', 'num_layers', 'intermediate_size', 'num_attention_heads', 'num_key_value_heads', 'max_position_embeddings', 'vocab_size', 'rope_theta', 'rms_norm_eps', 'hidden_act', 'attention_dropout', 'hidden_dropout', 'tie_word_embeddings', 'initializer_range']})
    
    model = GemmaForCausalLM(config)
    logger.info(f"Created model with {model.num_parameters():,} parameters")
    return model