models/model_v1_earlyexit.py

"""
BitSkip v1 with Early Exit Loss and Quadratic Dropout
- BitLinear quantization (8-bit)
- Quadratic layer dropout (0 to 0.5 progression, sum=1 constraint)
- Early exit loss from all layers
- HuggingFace compatible
"""

import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from transformers import PreTrainedModel, PretrainedConfig, GenerationMixin
from transformers.modeling_outputs import CausalLMOutputWithPast
from typing import Optional, Tuple

from .bitlinear import BitLinear


class BitSkipV1EarlyExitConfig(PretrainedConfig):
    model_type = "bitskip_v1_earlyexit"
    
    def __init__(
        self,
        vocab_size=50257,
        hidden_size=2048,
        num_hidden_layers=24,
        num_attention_heads=32,
        num_key_value_heads=8,
        intermediate_size=4096,
        max_position_embeddings=2048,
        rms_norm_eps=1e-5,
        rope_theta=10000.0,
        # Early exit parameters
        early_exit_loss_weight=0.3,
        # Quadratic dropout parameters
        max_dropout_prob=0.5,
        # Inference
        inference_exit_layer=None,
        **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.max_position_embeddings = max_position_embeddings
        self.rms_norm_eps = rms_norm_eps
        self.rope_theta = rope_theta
        self.early_exit_loss_weight = early_exit_loss_weight
        self.max_dropout_prob = max_dropout_prob
        self.inference_exit_layer = inference_exit_layer
        super().__init__(**kwargs)


class QuadraticLayerDropout(nn.Module):
    """
    Quadratic layer dropout: p_l = p_max * (l/L)^2
    Normalized so sum of probabilities = 1
    """
    
    def __init__(self, num_layers, max_dropout_prob=0.5):
        super().__init__()
        self.num_layers = num_layers
        self.max_dropout_prob = max_dropout_prob
        
        # Compute quadratic dropout probabilities
        dropout_probs = []
        for i in range(num_layers):
            # Quadratic: p_l = p_max * (l/L)^2
            prob = max_dropout_prob * ((i / max(num_layers - 1, 1)) ** 2)
            dropout_probs.append(prob)
        
        # Normalize so sum = 1 (as per requirement)
        total_prob = sum(dropout_probs)
        if total_prob > 0:
            dropout_probs = [p / total_prob for p in dropout_probs]
        
        self.dropout_probs = dropout_probs
    
    def should_drop_layer(self, layer_idx):
        """Returns True if layer should be dropped during training."""
        if not self.training or layer_idx >= self.num_layers - 1:  # Never drop last layer
            return False
        return torch.rand(1).item() < self.dropout_probs[layer_idx]


class RMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        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):
    def __init__(self, dim, max_position_embeddings=2048, base=10000):
        super().__init__()
        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
        self.register_buffer("inv_freq", inv_freq)

    def forward(self, x, position_ids):
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
        position_ids_expanded = position_ids[:, None, :].float()
        freqs = (inv_freq_expanded @ position_ids_expanded).transpose(1, 2)
        emb = torch.cat((freqs, freqs), dim=-1)
        return emb.cos().to(x.dtype), emb.sin().to(x.dtype)


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


def apply_rotary_pos_emb(q, k, cos, sin):
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


class BitSkipV1Attention(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.hidden_size // self.num_heads
        self.num_key_value_heads = config.num_key_value_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        
        self.q_proj = BitLinear(self.hidden_size, self.num_heads * self.head_dim)
        self.k_proj = BitLinear(self.hidden_size, self.num_key_value_heads * self.head_dim)
        self.v_proj = BitLinear(self.hidden_size, self.num_key_value_heads * self.head_dim)
        self.o_proj = BitLinear(self.hidden_size, self.hidden_size)
        
        self.rotary_emb = RotaryEmbedding(self.head_dim, config.max_position_embeddings, config.rope_theta)

    def forward(self, hidden_states, attention_mask=None, position_ids=None, past_key_value=None, use_cache=False):
        bsz, q_len, _ = hidden_states.size()
        
        query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        cos, sin = self.rotary_emb(value_states, position_ids)
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if past_key_value is not None:
            key_states = torch.cat([past_key_value[0], key_states], dim=2)
            value_states = torch.cat([past_key_value[1], value_states], dim=2)

        past_key_value = (key_states, value_states) if use_cache else None

        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)

        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
        if attention_mask is not None:
            attn_weights = attn_weights + attention_mask
        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
        attn_output = torch.matmul(attn_weights, value_states)
        attn_output = attn_output.transpose(1, 2).contiguous().reshape(bsz, q_len, self.hidden_size)
        attn_output = self.o_proj(attn_output)

        return attn_output, None, past_key_value


class BitSkipV1MLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.gate_proj = BitLinear(config.hidden_size, config.intermediate_size)
        self.up_proj = BitLinear(config.hidden_size, config.intermediate_size)
        self.down_proj = BitLinear(config.intermediate_size, config.hidden_size)

    def forward(self, x):
        return self.down_proj(nn.functional.silu(self.gate_proj(x)) * self.up_proj(x))


class BitSkipV1DecoderLayer(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.self_attn = BitSkipV1Attention(config)
        self.mlp = BitSkipV1MLP(config)
        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)

    def forward(self, hidden_states, attention_mask=None, position_ids=None, past_key_value=None, use_cache=False):
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        hidden_states, _, present_key_value = self.self_attn(
            hidden_states, attention_mask, position_ids, past_key_value, use_cache
        )
        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,) + ((present_key_value,) if use_cache else ())


class BitSkipV1PreTrainedModel(PreTrainedModel):
    config_class = BitSkipV1EarlyExitConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True

    def _init_weights(self, module):
        if isinstance(module, (nn.Linear, BitLinear)):
            if hasattr(module, 'weight'):
                module.weight.data.normal_(mean=0.0, std=0.02)
            if hasattr(module, 'bias') and module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=0.02)


class BitSkipV1Model(BitSkipV1PreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
        self.layers = nn.ModuleList([BitSkipV1DecoderLayer(config) for _ in range(config.num_hidden_layers)])
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.gradient_checkpointing = False
        
        # Quadratic dropout module
        self.layer_dropout = QuadraticLayerDropout(config.num_hidden_layers, config.max_dropout_prob)
        
        self.post_init()

    def forward(self, input_ids, attention_mask=None, position_ids=None, past_key_values=None, use_cache=False, output_hidden_states=False, return_all_layer_outputs=False):
        hidden_states = self.embed_tokens(input_ids)
        
        if position_ids is None:
            position_ids = torch.arange(input_ids.shape[1], dtype=torch.long, device=input_ids.device)
            position_ids = position_ids.unsqueeze(0)
        
        next_decoder_cache = () if use_cache else None
        all_layer_hidden_states = []  # Store ALL layer outputs for early exit loss
        
        # Determine layers to run
        num_layers_to_run = self.config.inference_exit_layer if self.config.inference_exit_layer else len(self.layers)
        num_layers_to_run = min(num_layers_to_run, len(self.layers))
        
        for idx in range(num_layers_to_run):
            layer = self.layers[idx]
            past_key_value = past_key_values[idx] if past_key_values else None
            
            # Apply quadratic dropout during training
            if self.training and self.layer_dropout.should_drop_layer(idx):
                # Skip this layer - hidden_states pass through unchanged
                all_layer_hidden_states.append(hidden_states)
                continue
            
            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    layer.__call__,
                    hidden_states,
                    attention_mask,
                    position_ids,
                    past_key_value,
                    use_cache,
                )
            else:
                layer_outputs = layer(hidden_states, attention_mask, position_ids, past_key_value, use_cache)
            
            hidden_states = layer_outputs[0]
            all_layer_hidden_states.append(hidden_states)
            
            if use_cache:
                next_decoder_cache += (layer_outputs[1],)
        
        hidden_states = self.norm(hidden_states)
        all_layer_hidden_states.append(hidden_states)  # Add final normed output
        
        if return_all_layer_outputs:
            return hidden_states, next_decoder_cache, all_layer_hidden_states
        else:
            return hidden_states, next_decoder_cache, None


class BitSkipV1ForCausalLMWithEarlyExit(BitSkipV1PreTrainedModel, GenerationMixin):
    """BitSkip v1 with early exit loss."""
    
    _tied_weights_keys = ["lm_head.weight"]
    
    def __init__(self, config):
        super().__init__(config)
        self.model = BitSkipV1Model(config)
        self.vocab_size = config.vocab_size
        
        # Shared LM head for all layers (LayerSkip approach)
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        
        self.post_init()

    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 compute_early_exit_loss(self, all_layer_hidden_states, labels):
        """
        Compute early exit loss from all layers.
        Uses layer-proportional weighting: w_i = i/N
        """
        num_layers = len(all_layer_hidden_states)
        
        # Compute weights: layer-proportional (deeper layers get more weight)
        weights = [(i + 1) / num_layers for i in range(num_layers)]
        weight_sum = sum(weights)
        weights = [w / weight_sum for w in weights]  # Normalize
        
        total_exit_loss = 0.0
        
        for i, hidden_states in enumerate(all_layer_hidden_states):
            # Get logits for this layer
            logits = self.lm_head(hidden_states)
            
            # Compute cross-entropy loss
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            
            loss_fct = nn.CrossEntropyLoss()
            layer_loss = loss_fct(shift_logits.view(-1, self.vocab_size), shift_labels.view(-1))
            
            # Weight and accumulate
            total_exit_loss += weights[i] * layer_loss
        
        return total_exit_loss

    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        position_ids=None,
        past_key_values=None,
        inputs_embeds=None,
        labels=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        
        # Get all layer outputs if training with labels (for early exit loss)
        return_all = self.training and labels is not None
        
        hidden_states, past_key_values_output, all_layer_hidden_states = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            use_cache=use_cache,
            output_hidden_states=output_hidden_states,
            return_all_layer_outputs=return_all,
        )

        logits = self.lm_head(hidden_states)
        logits = logits.float()

        loss = None
        if labels is not None:
            # Main loss (final layer)
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            loss_fct = nn.CrossEntropyLoss()
            main_loss = loss_fct(shift_logits.view(-1, self.vocab_size), shift_labels.view(-1))
            
            # Early exit loss (all intermediate layers)
            if all_layer_hidden_states is not None and len(all_layer_hidden_states) > 0:
                early_exit_loss = self.compute_early_exit_loss(all_layer_hidden_states[:-1], labels)  # Exclude final layer
                
                # Combine: main + weighted early exit
                loss = main_loss + self.config.early_exit_loss_weight * early_exit_loss
            else:
                loss = main_loss

        if not return_dict:
            output = (logits,) + (past_key_values_output,)
            return (loss,) + output if loss is not None else output

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=past_key_values_output,
            hidden_states=None,
            attentions=None,
        )

    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs):
        if past_key_values is not None:
            past_length = past_key_values[0][0].shape[2]
            if input_ids.shape[1] > past_length:
                remove_prefix_length = past_length
            else:
                remove_prefix_length = input_ids.shape[1] - 1
            input_ids = input_ids[:, remove_prefix_length:]

        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[:, -input_ids.shape[1] :]

        if inputs_embeds is not None and past_key_values is None:
            model_inputs = {"inputs_embeds": inputs_embeds}
        else:
            model_inputs = {"input_ids": input_ids}

        model_inputs.update({
            "position_ids": position_ids,
            "past_key_values": past_key_values,
            "use_cache": kwargs.get("use_cache"),
            "attention_mask": attention_mask,
        })
        return model_inputs

    @staticmethod
    def _reorder_cache(past_key_values, beam_idx):
        reordered_past = ()
        for layer_past in past_key_values:
            reordered_past += (
                tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
            )
        return reordered_past
    
    def set_exit_layer(self, exit_layer):
        """Set early exit layer for inference."""
        self.config.inference_exit_layer = exit_layer
        self.model.config.inference_exit_layer = exit_layer


BitSkipV1EarlyExitConfig.register_for_auto_class()
BitSkipV1ForCausalLMWithEarlyExit.register_for_auto_class("AutoModelForCausalLM")