checkpoint-3386/modeling_llama.py

from typing import Callable, List, Optional, Tuple, Union
import math
from transformers.cache_utils import Cache


import torch
import torch.nn as nn
from transformers.models.llama.modeling_llama import LlamaForCausalLM, LlamaConfig
from transformers.utils import logging

from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
logger = logging.get_logger(__name__)

def lambda_init_fn(layer_idx, num_layers=32):
    """Initialize lambda value based on layer index."""
    # Scale lambda_init from 0.1 to 0.9 based on layer depth
    # return 0.1 + 0.8 * (layer_idx / (num_layers - 1))
    return 0.8 - 0.6 * math.exp(-0.3 * layer_idx)

def lambda_init_fn_const(layer_idx, num_layers=32):
    """Initialize lambda value based on layer index."""
    # Scale lambda_init from 0.1 to 0.9 based on layer depth
    return 0.5

class LlamaForCausalLMj24(LlamaForCausalLM):
    
    def __init__(self, config: LlamaConfig):
        config._attn_implementation = "sdpa"
        super().__init__(config)

        for _ in range(10):
            print(f"** LLAMA Adapt j24 **")

        for layer_i, layer in enumerate(self.model.layers):

            hidden_dim = config.hidden_size
            n_heads = config.num_attention_heads
            depth = layer_i  # or pass 0

            layer.self_attn = LlamaAttention(
                config, layer_idx=layer_i
            )


    def freeze_original_unfreeze_adapters(self):
        """
        Freeze original weights and unfreeze newly initialized values 
        containing "lambda" and "lora" in their parameter names.
        """

        params_to_unfreeze = [
            "lambda",
            "lora",
        ]

        # First freeze all parameters
        for param in self.parameters():
            param.requires_grad = False
        
        # Then unfreeze parameters containing "lambda" or "lora"
        for name, param in self.named_parameters():
            for p in params_to_unfreeze:
                if p in name:
                    param.requires_grad = True
                    print(f"Unfreezing parameter: {name}")
        
        # Count trainable parameters
        # trainable_params = sum(p.numel() for p in self.parameters() if p.requires_grad)
        # total_params = sum(p.numel() for p in self.parameters())
        # print(f"Trainable parameters: {trainable_params:,} ({trainable_params/total_params:.2%} of total)")


class LlamaAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: LlamaConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
        self.scaling = self.head_dim**-0.5
        self.attention_dropout = config.attention_dropout
        self.is_causal = True
        config.lambda_std_dev = 0.1 # just use default value

        self.q_proj = nn.Linear(
            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
        )
        self.k_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.v_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.d_lora_proj = nn.Linear(
            config.num_attention_heads * self.head_dim // 2, config.hidden_size // 2, bias=config.attention_bias
        )
        self.o_proj = nn.Linear(
            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
        )

        self.lambda_init = lambda_init_fn(layer_idx)
        print(f"* Initializing lambda: {self.lambda_init}")
        self.i_step = 0 
        self.lambda_schedule = torch.linspace(0, self.lambda_init, 3386) # FIXME: hard-coded n_steps
        padding = torch.ones(5000) * self.lambda_init
        padding = padding.to(self.lambda_schedule.device)
        self.lambda_schedule = torch.cat((self.lambda_schedule, padding))
        self.lambda_q1 = nn.Parameter(torch.normal(0, config.lambda_std_dev, size=(self.head_dim,)))
        self.lambda_k1 = nn.Parameter(torch.normal(0, config.lambda_std_dev, size=(self.head_dim,)))
        self.lambda_q2 = nn.Parameter(torch.normal(0, config.lambda_std_dev, size=(self.head_dim,)))
        self.lambda_k2 = nn.Parameter(torch.normal(0, config.lambda_std_dev, size=(self.head_dim,)))
        self._init_lora()

        path = "/root/pegasus-llama-factory/scripts/analysis/head_importance_llama_1B/head_importance_indices.txt"
        with open(path, "r") as f:
            self.head_importance_indices = [line.strip() for line in f]
            self.head_importance_indices = self.head_importance_indices[self.layer_idx].split(",")
            self.head_importance_indices = [int(idx) for idx in self.head_importance_indices]

    def _init_lora(self):
        """Initialize LoRA for query, key, and value projection layers."""
        # Define LoRA rank and scaling factor
        self.lora_rank = 8
        self.lora_alpha = 16
        self.lora_dropout = 0.0
        self.scaling_factor = self.lora_alpha / self.lora_rank

        self.k_lora_full = nn.Parameter(torch.zeros(self.config.hidden_size, self.config.num_key_value_heads * self.head_dim))
        self.k_lora_A = nn.Parameter(torch.zeros(self.config.hidden_size, self.lora_rank))
        self.k_lora_B = nn.Parameter(torch.zeros(self.lora_rank, self.config.num_key_value_heads * self.head_dim))
        self.k_lora_scaling = nn.Parameter(torch.ones(self.config.num_key_value_heads * self.head_dim))
        nn.init.kaiming_uniform_(self.k_lora_A, a=math.sqrt(5))
        nn.init.zeros_(self.k_lora_B)
        nn.init.ones_(self.k_lora_scaling)  # Initialize scaling vector with ones

        self.v_lora_full = nn.Parameter(torch.zeros(self.config.hidden_size, self.config.num_key_value_heads * self.head_dim))
        self.v_lora_A = nn.Parameter(torch.zeros(self.config.hidden_size, self.lora_rank))
        self.v_lora_B = nn.Parameter(torch.zeros(self.lora_rank, self.config.num_key_value_heads * self.head_dim))
        self.v_lora_scaling = nn.Parameter(torch.ones(self.config.num_key_value_heads * self.head_dim))
        nn.init.kaiming_uniform_(self.v_lora_A, a=math.sqrt(5))
        nn.init.zeros_(self.v_lora_B)
        nn.init.ones_(self.v_lora_scaling)  # Initialize scaling vector with ones
        
        # Initialize LoRA weights for output projection
        self.o_lora_A = nn.Parameter(torch.zeros(self.config.hidden_size, self.lora_rank))
        self.o_lora_B = nn.Parameter(torch.zeros(self.lora_rank, self.config.hidden_size))
        self.o_lora_scaling = nn.Parameter(torch.ones(self.config.hidden_size))
        nn.init.kaiming_uniform_(self.o_lora_A, a=math.sqrt(5))
        nn.init.zeros_(self.o_lora_B)
        nn.init.ones_(self.o_lora_scaling)  # Initialize scaling vector with ones

        # LoRA dropout layer
        self.lora_dropout = nn.Dropout(self.lora_dropout)
        
        self.k_proj_forward_original = self.k_proj.forward
        self.v_proj_forward_original = self.v_proj.forward
        self.o_proj_forward_original = self.o_proj.forward

        self.k_proj.forward = self._k_proj_forward
        self.v_proj.forward = self._v_proj_forward
        self.o_proj.forward = self._o_proj_forward

    def _k_proj_forward(self, x):
        base_output = self.k_proj_forward_original(x)
        full_output = x @ self.k_lora_full
        lora_output = self.lora_dropout(x) @ self.k_lora_A @ self.k_lora_B * self.scaling_factor
        # offset = torch.cat((lora_output, full_output), dim=-1)
        offset = lora_output + full_output
        return (base_output + offset) * self.k_lora_scaling

    def _v_proj_forward(self, x):
        """LoRA-enabled forward for value projection."""
        base_output = self.v_proj_forward_original(x)
        full_output = x @ self.v_lora_full
        lora_output = self.lora_dropout(x) @ self.v_lora_A @ self.v_lora_B * self.scaling_factor
        # offset = torch.cat((lora_output, full_output), dim=-1)
        offset = lora_output + full_output
        return (base_output + offset ) * self.v_lora_scaling
    
    def _o_proj_forward(self, x):
        """LoRA-enabled forward for output projection."""
        base_output = self.o_proj_forward_original(x)
        lora_output = self.lora_dropout(x) @ self.o_lora_A @ self.o_lora_B
        return (base_output + lora_output * self.scaling_factor) * self.o_lora_scaling

    def __post_init__(self):
        """Initialize LoRA after the module is fully initialized."""
        self._init_lora()

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
        attention_mask: Optional[torch.Tensor],
        past_key_value: Optional[Cache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)

        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(hidden_shape).transpose(1, 2)
        key_states = key_states.view(hidden_shape).transpose(1, 2)
        value_states = value_states.view(hidden_shape).transpose(1, 2)

        cos, sin = position_embeddings
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if past_key_value is not None:
            # sin and cos are specific to RoPE models; cache_position needed for the static cache
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

        attention_interface: Callable = eager_attention_forward
        if self.config._attn_implementation != "eager":
            if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
                logger.warning_once(
                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
                )
            else:
                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

        attn_output, attn_weights = attention_interface(
                self,
                query_states,
                key_states,
                value_states,
                attention_mask,
                dropout=0.0 if not self.training else self.attention_dropout,
                scaling=self.scaling,
                **kwargs,
            ) # b, n, h, d

        lambda_1 = torch.exp(torch.sum(self.lambda_q1 * self.lambda_k1, dim=-1, dtype=torch.float32).to(query_states.dtype))
        lambda_2 = torch.exp(torch.sum(self.lambda_q2 * self.lambda_k2, dim=-1, dtype=torch.float32).to(query_states.dtype))

        alpha = (self.i_step / 3386)
        alpha = min(1, alpha)
        lambda_full = self.lambda_schedule[self.i_step] * (1 - alpha) + (lambda_1 - lambda_2 + self.lambda_init) * alpha

        self.i_step += 1

        modify_heads = self.head_importance_indices[:len(self.head_importance_indices)//2]
        main_heads = self.head_importance_indices[len(self.head_importance_indices)//2:]

        modify_attn_output = attn_output[:, :, modify_heads]
        main_attn_output = attn_output[:, :, main_heads]

        # attn_output = attn_output.reshape(*input_shape, -1).contiguous() # b, n, h * d
        modify_attn_output = modify_attn_output.reshape(*input_shape, -1).contiguous() # b, n, h // 2 * d
        main_attn_output = main_attn_output.reshape(*input_shape, -1).contiguous() # b, n, h // 2 * d

        modify_attn_output = modify_attn_output - lambda_full * self.d_lora_proj(modify_attn_output)
        # Restore the original head order while preserving the computation graph
        batch_size, seq_len, dim = modify_attn_output.shape
        restored_attn_output = torch.zeros(batch_size, seq_len, 2 * dim,
                                          device=attn_output.device, dtype=attn_output.dtype)
        
        for i, idx in enumerate(main_heads):
            restored_attn_output[:, :, idx*self.head_dim:(idx+1)*self.head_dim] = main_attn_output[:, :, i*self.head_dim:(i+1)*self.head_dim]
        for i, idx in enumerate(modify_heads):
            restored_attn_output[:, :, idx*self.head_dim:(idx+1)*self.head_dim] = modify_attn_output[:, :, i*self.head_dim:(i+1)*self.head_dim]
            
        attn_output = restored_attn_output

        attn_output = self.o_proj(attn_output)

        attn_weights = None  # No attention weights returned when using SDPA
        
        return attn_output, attn_weights

def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    """
    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
    """
    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
    if n_rep == 1:
        return hidden_states
    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)


def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: Optional[torch.Tensor],
    scaling: float,
    dropout: float = 0.0,
    **kwargs,
):
    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)

    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
    if attention_mask is not None:
        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
        attn_weights = attn_weights + causal_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
    attn_output = torch.matmul(attn_weights, value_states)
    attn_output = attn_output.transpose(1, 2).contiguous()

    return attn_output, attn_weights

class LlamaRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        LlamaRMSNorm is equivalent to T5LayerNorm
        """
        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)

    def extra_repr(self):
        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"



class LlamaRotaryEmbedding(nn.Module):
    def __init__(self, config: LlamaConfig, device=None):
        super().__init__()
        # BC: "rope_type" was originally "type"
        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
        else:
            self.rope_type = "default"
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

        self.config = config
        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]

        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.original_inv_freq = self.inv_freq

    def _dynamic_frequency_update(self, position_ids, device):
        """
        dynamic RoPE layers should recompute `inv_freq` in the following situations:
        1 - growing beyond the cached sequence length (allow scaling)
        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
        """
        seq_len = torch.max(position_ids) + 1
        if seq_len > self.max_seq_len_cached:  # growth
            inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, seq_len=seq_len)
            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
            self.max_seq_len_cached = seq_len

        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
            # This .to() is needed if the model has been moved to a device after being initialized (because
            # the buffer is automatically moved, but not the original copy)
            self.original_inv_freq = self.original_inv_freq.to(device)
            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
            self.max_seq_len_cached = self.original_max_seq_len

    @torch.no_grad()
    def forward(self, x, position_ids):
        if "dynamic" in self.rope_type:
            self._dynamic_frequency_update(position_ids, device=x.device)

        # Core RoPE block
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
        position_ids_expanded = position_ids[:, None, :].float()
        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
        device_type = x.device.type
        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos()
            sin = emb.sin()

        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
        cos = cos * self.attention_scaling
        sin = sin * self.attention_scaling

        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)


def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    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, k, cos, sin, position_ids=None, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.

    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        position_ids (`torch.Tensor`, *optional*):
            Deprecated and unused.
        unsqueeze_dim (`int`, *optional*, defaults to 1):
            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
    Returns:
        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
    """
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


class LlamaMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias)
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, x):
        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
        return down_proj