modeling_latex_decoder.py

# update v2

import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional

from transformers import PreTrainedModel
from transformers.modeling_outputs import CausalLMOutput

from .configuration_latex_decoder import LaTeXDecoderConfig


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

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        rms = x.pow(2).mean(-1, keepdim=True).add(self.eps).sqrt()
        return x / rms * self.weight


def _build_rope_cache(seq_len, head_dim, theta=10000.0, device=None, dtype=torch.float32):
    half = head_dim // 2
    inv_freq = 1.0 / (theta ** (torch.arange(0, half, device=device, dtype=torch.float32) / half))
    pos = torch.arange(seq_len, device=device, dtype=torch.float32)
    freqs = torch.outer(pos, inv_freq)
    emb = torch.cat([freqs, freqs], dim=-1)
    return emb.cos().to(dtype), emb.sin().to(dtype)


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


def apply_rope(q, k, cos, sin):
    cos = cos.unsqueeze(0).unsqueeze(0)
    sin = sin.unsqueeze(0).unsqueeze(0)
    return q * cos + _rotate_half(q) * sin, k * cos + _rotate_half(k) * sin


class CausalSelfAttention(nn.Module):
    def __init__(self, cfg: LaTeXDecoderConfig):
        super().__init__()
        self.n_heads = cfg.n_heads
        self.head_dim = cfg.head_dim
        self.d_model = cfg.d_model
        self.dropout_p = cfg.dropout
        self.rope_theta = cfg.rope_theta

        self.qkv_proj = nn.Linear(cfg.d_model, 3 * cfg.d_model, bias=False)
        self.out_proj = nn.Linear(cfg.d_model, cfg.d_model, bias=False)
        self._rope_cache: dict = {}

    def _get_rope(self, seq_len, device, dtype):
        key = (seq_len, str(device), dtype)
        if key not in self._rope_cache:
            self._rope_cache[key] = _build_rope_cache(seq_len, self.head_dim, self.rope_theta, device, dtype)
        return self._rope_cache[key]

    def forward(self, x: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
        B, T, C = x.shape
        q, k, v = self.qkv_proj(x).chunk(3, dim=-1)

        q = q.view(B, T, self.n_heads, self.head_dim).transpose(1, 2)
        k = k.view(B, T, self.n_heads, self.head_dim).transpose(1, 2)
        v = v.view(B, T, self.n_heads, self.head_dim).transpose(1, 2)

        cos, sin = self._get_rope(T, x.device, q.dtype)
        q, k = apply_rope(q, k, cos, sin)

        dropout_p = self.dropout_p if self.training else 0.0

        if attention_mask is not None:
            causal = torch.triu(torch.full((T, T), float("-inf"), device=x.device, dtype=q.dtype), diagonal=1)
            pad = (~attention_mask).unsqueeze(1).unsqueeze(2)
            attn_bias = causal.unsqueeze(0).unsqueeze(0).expand(B, 1, T, T).clone()
            attn_bias = attn_bias.masked_fill(pad, float("-inf"))
            out = F.scaled_dot_product_attention(q, k, v, attn_mask=attn_bias, dropout_p=dropout_p, is_causal=False)
        else:
            out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p, is_causal=True)

        return self.out_proj(out.transpose(1, 2).contiguous().view(B, T, C))


class SwiGLUFFN(nn.Module):
    def __init__(self, cfg: LaTeXDecoderConfig):
        super().__init__()
        self.gate_proj = nn.Linear(cfg.d_model, cfg.d_ff, bias=False)
        self.up_proj = nn.Linear(cfg.d_model, cfg.d_ff, bias=False)
        self.down_proj = nn.Linear(cfg.d_ff, cfg.d_model, bias=False)
        self.dropout = nn.Dropout(cfg.dropout)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.dropout(self.down_proj(F.silu(self.gate_proj(x)) * self.up_proj(x)))


class TransformerBlock(nn.Module):
    def __init__(self, cfg: LaTeXDecoderConfig):
        super().__init__()
        self.norm1 = RMSNorm(cfg.d_model)
        self.attn = CausalSelfAttention(cfg)
        self.norm2 = RMSNorm(cfg.d_model)
        self.ffn = SwiGLUFFN(cfg)
        self.drop = nn.Dropout(cfg.dropout)

    def forward(self, x: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
        x = x + self.drop(self.attn(self.norm1(x), attention_mask))
        x = x + self.drop(self.ffn(self.norm2(x)))
        return x


class LaTeXDecoderForCausalLM(PreTrainedModel):
    config_class = LaTeXDecoderConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = False

    def __init__(self, config: LaTeXDecoderConfig):
        super().__init__(config)

        self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model, padding_idx=config.pad_id)
        self.embed_drop = nn.Dropout(config.dropout)
        self.layers = nn.ModuleList([TransformerBlock(config) for _ in range(config.n_layers)])
        self.norm_final = RMSNorm(config.d_model)
        self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False)

        if config.tie_weights:
            self.lm_head.weight = self.embed_tokens.weight

        self.post_init()

    def _init_weights(self, module: nn.Module):
        if isinstance(module, nn.Linear):
            nn.init.normal_(module.weight, mean=0.0, std=0.02)
            if module.bias is not None:
                nn.init.zeros_(module.bias)
        elif isinstance(module, nn.Embedding):
            nn.init.normal_(module.weight, mean=0.0, std=0.02)

    def forward(
        self,
        input_ids: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        labels: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> CausalLMOutput:
        x = self.embed_drop(self.embed_tokens(input_ids))
        for layer in self.layers:
            x = layer(x, attention_mask)
        logits = self.lm_head(self.norm_final(x))

        loss = None
        if labels is not None:
            shift_logits = logits[:, :-1, :].contiguous()
            shift_labels = labels[:, 1:].contiguous()
            shift_labels = shift_labels.masked_fill(shift_labels == self.config.pad_id, -100)
            loss = F.cross_entropy(
                shift_logits.view(-1, self.config.vocab_size),
                shift_labels.view(-1),
                ignore_index=-100,
            )

        return CausalLMOutput(loss=loss, logits=logits)

    @torch.inference_mode()
    def generate(
        self,
        prompt_ids: torch.Tensor,
        max_new_tokens: int = 200,
        temperature: float = 1.0,
        top_p: float = 0.9,
        eos_id: Optional[int] = None,
    ) -> torch.Tensor:
        eos = eos_id if eos_id is not None else self.config.eos_id
        generated = prompt_ids.clone()

        for _ in range(max_new_tokens):
            ctx = generated[:, -self.config.max_seq_len:]
            logits = self.forward(ctx).logits[:, -1, :]

            if temperature == 0.0:
                next_id = logits.argmax(dim=-1, keepdim=True)
            else:
                probs = F.softmax(logits / temperature, dim=-1)
                sorted_probs, sorted_idx = probs.sort(dim=-1, descending=True)
                cumsum = sorted_probs.cumsum(dim=-1)
                sorted_probs[cumsum - sorted_probs > top_p] = 0.0
                sorted_probs /= sorted_probs.sum(dim=-1, keepdim=True)
                next_id = sorted_idx.gather(-1, torch.multinomial(sorted_probs, 1))

            generated = torch.cat([generated, next_id], dim=-1)
            if next_id.item() == eos:
                break

        return generated