modeling_llada.py

# modeling_llada.py

import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers.modeling_utils import PreTrainedModel
from transformers import PretrainedConfig

# --- 1) Config Sınıfı --------------------------------------------------------
class LLaDAConfig(PretrainedConfig):
    model_type = "llada"

    def __init__(
        self,
        vocab_size=50000,
        max_seq_len=512,
        d_model=128,
        n_layers=16,
        n_heads=8,
        dropout=0.1,
        **kwargs
    ):
        super().__init__(**kwargs)
        self.vocab_size   = vocab_size
        self.max_seq_len  = max_seq_len
        self.d_model      = d_model
        self.n_layers     = n_layers
        self.n_heads      = n_heads
        self.d_head       = d_model // n_heads
        self.d_ffn        = 4 * d_model
        self.dropout      = dropout


# --- 2) PreTrainedModel Tabanı ------------------------------------------------
class LLaDAPreTrainedModel(PreTrainedModel):
    config_class = LLaDAConfig
    base_model_prefix = "llada"


# --- 3) Alt Modüller ---------------------------------------------------------
class RMSNorm(nn.Module):
    def __init__(self, dim: int, eps: float = 1e-6):
        super().__init__()
        self.eps    = eps
        self.weight = nn.Parameter(torch.ones(dim))

    def _norm(self, x):
        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)

    def forward(self, x):
        return self._norm(x.float()).type_as(x) * self.weight


class RotaryPositionalEmbedding(nn.Module):
    def __init__(self, dim, max_seq_len=512):
        super().__init__()
        inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2).float() / dim))
        t = torch.arange(max_seq_len, device=inv_freq.device).type_as(inv_freq)
        freqs = torch.einsum("i,j->ij", t, inv_freq)
        emb = torch.cat((freqs, freqs), dim=-1)
        self.register_buffer("cos", emb.cos())
        self.register_buffer("sin", emb.sin())

    def forward(self, x):
        return self.cos[: x.shape[2], :], self.sin[: x.shape[2], :]


def apply_rotary(q, k, cos, sin):
    q2 = (q * cos) + (torch.cat([-q[..., 1::2], q[..., ::2]], -1) * sin)
    k2 = (k * cos) + (torch.cat([-k[..., 1::2], k[..., ::2]], -1) * sin)
    return q2, k2


class Attention(nn.Module):
    def __init__(self, config: LLaDAConfig):
        super().__init__()
        self.n_heads = config.n_heads
        self.d_head  = config.d_head

        self.wq = nn.Linear(config.d_model, config.n_heads * config.d_head, bias=False)
        self.wk = nn.Linear(config.d_model, config.n_heads * config.d_head, bias=False)
        self.wv = nn.Linear(config.d_model, config.n_heads * config.d_head, bias=False)
        self.wo = nn.Linear(config.n_heads * config.d_head, config.d_model, bias=False)

        self.rotary = RotaryPositionalEmbedding(config.d_head, config.max_seq_len)

    def forward(self, x):
        b, seq, _ = x.size()
        q = self.wq(x).view(b, seq, self.n_heads, self.d_head).transpose(1, 2)
        k = self.wk(x).view(b, seq, self.n_heads, self.d_head).transpose(1, 2)
        v = self.wv(x).view(b, seq, self.n_heads, self.d_head).transpose(1, 2)
        cos, sin = self.rotary(q)
        q, k     = apply_rotary(q, k, cos, sin)
        out      = F.scaled_dot_product_attention(q, k, v, is_causal=False)
        out      = out.transpose(1, 2).reshape(b, seq, -1)
        return self.wo(out)


class FeedForward(nn.Module):
    def __init__(self, config: LLaDAConfig):
        super().__init__()
        self.w1 = nn.Linear(config.d_model, config.d_ffn, bias=False)
        self.w2 = nn.Linear(config.d_ffn, config.d_model, bias=False)
        self.w3 = nn.Linear(config.d_model, config.d_ffn, bias=False)

    def forward(self, x):
        return self.w2(F.silu(self.w1(x)) * self.w3(x))


class TransformerBlock(nn.Module):
    def __init__(self, config: LLaDAConfig):
        super().__init__()
        self.attn      = Attention(config)
        self.ff        = FeedForward(config)
        self.norm1     = RMSNorm(config.d_model)
        self.norm2     = RMSNorm(config.d_model)

    def forward(self, x):
        h = x + self.attn(self.norm1(x))
        return h + self.ff(self.norm2(h))


# --- 4) Ana Model Sınıfı ------------------------------------------------------
class LLaDA_Model(nn.Module):
    def __init__(self, config: LLaDAConfig):
        super().__init__()
        self.embed  = nn.Embedding(config.vocab_size, config.d_model)
        self.layers = nn.ModuleList([TransformerBlock(config) for _ in range(config.n_layers)])
        self.norm   = RMSNorm(config.d_model)

    def forward(self, input_ids):
        x = self.embed(input_ids)
        for layer in self.layers:
            x = layer(x)
        return self.norm(x)


# --- 5) LM Head ile CausalLM --------------------------------------------------
class LLaDA_ModelForCausalLM(LLaDAPreTrainedModel):
    def __init__(self, config: LLaDAConfig):
        super().__init__(config)
        self.llada   = LLaDA_Model(config)
        self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False)

    def forward(self, input_ids, **kwargs):
        hidden = self.llada(input_ids)
        logits = self.lm_head(hidden)
        return logits