zzjrabbit3.py

from typing import Optional, Union

import torch
import torch.nn as nn
from tokenizers import Tokenizer, decoders, pre_tokenizers
from tokenizers.models import BPE
from transformers import (
    GenerationMixin,
    PreTrainedConfig,
    PreTrainedModel,
    TokenizersBackend,
)
from transformers.modeling_outputs import BaseModelOutput, CausalLMOutput


class ZZJRabbit3Config(PreTrainedConfig):
    model_type = "zzjrabbit3"

    def __init__(
        self,
        vocab_size: int = 100000,
        hidden_size: int = 1024,
        num_hidden_layers: int = 12,
        num_attention_heads: int = 8,
        attention_dropout: float | int = 0.0,
        pad_token_id: int | None = None,
        eos_token_id: int | list[int] | None = 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.attention_dropout = attention_dropout
        self.pad_token_id = pad_token_id
        self.eos_token_id = eos_token_id
        super().__init__(**kwargs)


class ZZJRabbit3RotaryEmbedding(nn.Module):
    def __init__(self, dim, max_position_embeddings=2048, base=10000):
        """
        Rotary Embedding 模块

        Args:
            dim: 每个 token embedding 的维度
            max_position_embeddings: 最大位置数
            base: rotary embedding 的频率基底
        """
        super().__init__()
        self.dim = dim
        self.base = base

        # 生成频率向量
        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
        self.register_buffer("inv_freq", inv_freq)

        # 可选：预先计算 cos/sin
        t = torch.arange(max_position_embeddings, dtype=torch.float32)
        freqs = torch.einsum("i,j->ij", t, inv_freq)
        self.register_buffer("cos_cached", freqs.cos())
        self.register_buffer("sin_cached", freqs.sin())

    def forward(self, position_ids):
        """
        position_ids: (batch_size, seq_len)
        返回：
            cos: (batch_size, seq_len, dim)
            sin: (batch_size, seq_len, dim)
        """
        # 从缓存中选取对应位置
        cos = self.cos_cached[position_ids]  # shape (batch, seq_len, dim/2)
        sin = self.sin_cached[position_ids]

        # 将维度对齐为 (dim)
        # cos/sin 当前 shape 为 (..., dim/2)，重复到 dim
        cos = torch.stack([cos, cos], dim=-1).flatten(-2)
        sin = torch.stack([sin, sin], dim=-1).flatten(-2)
        return cos, sin


def rotate_half(x):
    """[-x2, x1]"""
    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, sin, cos):
    cos = cos.unsqueeze(1)
    sin = sin.unsqueeze(1)

    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)

    return q_embed, k_embed


class ZZJRabbit3Attention(nn.Module):
    def __init__(self, config: ZZJRabbit3Config):
        super().__init__()
        self.config = config
        self.head_dim = config.hidden_size // config.num_attention_heads
        self.q_proj = nn.Linear(config.hidden_size, config.hidden_size)
        self.k_proj = nn.Linear(config.hidden_size, config.hidden_size)
        self.v_proj = nn.Linear(config.hidden_size, config.hidden_size)
        self.out_proj = nn.Linear(config.hidden_size, config.hidden_size)
        self.dropout = nn.Dropout(0.1)

    def forward(
        self,
        x: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        key_padding_mask: Optional[torch.BoolTensor] = None,
        attn_mask: Optional[torch.BoolTensor] = None,
    ) -> torch.Tensor:
        batch_size = x.size(0)
        Q = (
            self.q_proj(x)
            .view(batch_size, -1, self.config.num_attention_heads, self.head_dim)
            .transpose(1, 2)
        )
        K = (
            self.k_proj(x)
            .view(batch_size, -1, self.config.num_attention_heads, self.head_dim)
            .transpose(1, 2)
        )
        V = (
            self.v_proj(x)
            .view(batch_size, -1, self.config.num_attention_heads, self.head_dim)
            .transpose(1, 2)
        )
        cos, sin = position_embeddings
        Q, K = apply_rotary_pos_emb(Q, K, sin.to(Q.dtype), cos.to(Q.dtype))
        scores = torch.matmul(Q, K.transpose(-2, -1)) * (self.head_dim**-0.5)
        if key_padding_mask is not None:
            scores = scores.masked_fill(
                key_padding_mask.view(batch_size, 1, 1, -1), float("-inf")
            )
        if attn_mask is not None:
            scores = scores.masked_fill(attn_mask, float("-inf"))
        attn_weights = nn.functional.softmax(scores, dim=-1)
        attn_weights = self.dropout(attn_weights)
        context = torch.matmul(attn_weights, V)
        context = context.transpose(1, 2).contiguous()
        context = context.view(batch_size, -1, self.config.hidden_size)
        return self.out_proj(context)


class ZZJRabbit3Layer(nn.Module):
    def __init__(self, config: ZZJRabbit3Config):
        super().__init__()
        self.attn = ZZJRabbit3Attention(config)
        self.l1 = nn.Linear(config.hidden_size, config.hidden_size)
        self.l2 = nn.Linear(config.hidden_size, config.hidden_size)
        self.activate = nn.ReLU()
        self.norm = nn.RMSNorm(config.hidden_size)

    def forward(
        self,
        x: torch.Tensor,
        postition_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        key_padding_mask = None
        attn_mask = torch.gt(
            torch.triu(torch.ones(x.size(-2), x.size(-2), device=x.device), 1), 0
        )
        if attention_mask is not None:
            key_padding_mask = torch.lt(attention_mask, 1)
        attn = self.attn(
            x,
            postition_embeddings,
            key_padding_mask=key_padding_mask,
            attn_mask=attn_mask,
        )
        x = self.norm(x + attn)
        o = self.l1(x)
        o = self.activate(o)
        o = self.l2(o)
        return self.norm(x + o)


class ZZJRabbit3Model(PreTrainedModel):
    config_class = ZZJRabbit3Config

    def __init__(self, config: ZZJRabbit3Config, **kwargs):
        super().__init__(config, **kwargs)
        self.config = config
        self.embedding = nn.Embedding(config.vocab_size, config.hidden_size)
        self.rotary_emb = ZZJRabbit3RotaryEmbedding(
            config.hidden_size // config.num_attention_heads
        )
        self.layers = nn.ModuleList(
            [ZZJRabbit3Layer(config) for _ in range(config.num_hidden_layers)]
        )
        self.post_init()

    def forward(
        self,
        input_ids: torch.Tensor,
        return_dict: Optional[bool] = None,
        attention_mask: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> tuple | BaseModelOutput:
        res = self.embedding(input_ids)
        batch_size, seq_len = input_ids.shape
        position_ids = (
            torch.arange(seq_len, device=input_ids.device)
            .unsqueeze(0)
            .expand(batch_size, -1)
        )
        position_embeddings = self.rotary_emb(position_ids)
        for layer in self.layers:
            res = layer(res, position_embeddings, attention_mask)
        if not return_dict:
            return (res,)
        else:
            return BaseModelOutput(res)


class ZZJRabbit3ForCausalLM(PreTrainedModel, GenerationMixin):
    config_class = ZZJRabbit3Config

    def __init__(self, config: ZZJRabbit3Config, **kwargs):
        super().__init__(config, **kwargs)
        self.model = ZZJRabbit3Model(config, **kwargs)
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size)
        self.post_init()

    def forward(
        self,
        input_ids: torch.Tensor,
        return_dict: Optional[bool] = None,
        labels: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        logits_to_keep: Union[int, torch.Tensor] = 0,
        **kwargs,
    ) -> tuple | CausalLMOutput:
        hidden = self.model(input_ids=input_ids, attention_mask=attention_mask)[0]
        logits = self.lm_head(
            hidden[
                :,
                slice(-logits_to_keep, None)
                if isinstance(logits_to_keep, int)
                else logits_to_keep,
                :,
            ]
        )
        if labels is not None:
            loss = self.loss_function(
                logits=logits,
                labels=labels,
                vocab_size=self.config.vocab_size,
                **kwargs,
            )
        if not return_dict:
            return (loss, logits) if labels is not None else (logits,)
        else:
            return (
                CausalLMOutput(logits=logits, loss=loss)
                if labels is not None
                else CausalLMOutput(logits=logits)
            )

    @classmethod
    def can_generate(cls):
        return True

    def prepare_inputs_for_generation(self, input_ids, **kwargs):
        return {"input_ids": input_ids}


class ZZJRabbit3Tokenizer(TokenizersBackend):
    model = BPE

    def __init__(
        self,
        vocab=None,
        merges=None,
        unk_token="<eos>",
        eos_token="<eos>",
        pad_token="<eos>",
        **kwargs,
    ):
        self._vocab = vocab or {
            "<eos>": 0,
        }
        self._merges = merges or []

        self._tokenizer = Tokenizer(
            BPE(vocab=self._vocab, merges=self._merges, fuse_unk=True)
        )
        self._tokenizer.pre_tokenizer = pre_tokenizers.ByteLevel(add_prefix_space=False)
        self._tokenizer.decoder = decoders.ByteLevel()

        super().__init__(
            unk_token=unk_token, eos_token=eos_token, pad_token=pad_token, **kwargs
        )