File size: 4,878 Bytes
6efaeab
 
 
 
 
 
 
ac59af7
6efaeab
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
08ab6b6
 
6efaeab
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
"""Token embeddings, task token embeddings, and RoPE for Ogma."""

from __future__ import annotations

import torch
import torch.nn as nn

from .config import OgmaConfig

__all__ = ["TokenEmbedding", "RotaryPositionalEncoding"]


class TokenEmbedding(nn.Module):
    """Token embedding with optional linear projection.

    Loads a vocab_size x d_embed embedding table and projects to d_model.
    Includes 3 learnable task token embeddings ([QRY], [DOC], [SYM]).
    """

    def __init__(self, config: OgmaConfig) -> None:
        super().__init__()
        self.config = config
        self.embed = nn.Embedding(
            config.vocab_size + config.n_special_tokens,
            config.d_embed,
            padding_idx=config.pad_id,
        )
        if config.d_embed != config.d_model:
            self.proj = nn.Linear(config.d_embed, config.d_model)
        else:
            self.proj = nn.Identity()  # type: ignore[assignment]

        # Task token embeddings are learned separately at d_model
        self.task_tokens = nn.Embedding(3, config.d_model)

    def forward(
        self,
        token_ids: torch.Tensor,
        task_token_ids: torch.Tensor,
    ) -> torch.Tensor:
        """Embed tokens and prepend task token.

        Args:
            token_ids: (B, S) token IDs.
            task_token_ids: (B,) task token IDs (4=QRY, 5=DOC, 6=SYM).

        Returns:
            (B, S+1, d_model) embeddings with task token prepended.
        """
        # Embed and project regular tokens
        x = self.embed(token_ids)  # (B, S, d_embed)
        x = self.proj(x)  # (B, S, d_model)

        # Get task token embeddings (map 4,5,6 -> 0,1,2)
        task_idx = task_token_ids - self.config.qry_id  # (B,)
        task_emb = self.task_tokens(task_idx)  # (B, d_model)
        task_emb = task_emb.unsqueeze(1)  # (B, 1, d_model)

        # Prepend task token
        return torch.cat([task_emb, x], dim=1)  # (B, S+1, d_model)

    def load_pretrained_embeddings(
        self, embeddings: torch.Tensor
    ) -> None:
        """Load pre-computed token embeddings (e.g., from teacher PCA).

        Args:
            embeddings: (vocab_size, d_embed) tensor.
        """
        with torch.no_grad():
            n = min(embeddings.shape[0], self.config.vocab_size)
            start = self.config.n_special_tokens
            self.embed.weight[start : n + start] = embeddings[:n]


class RotaryPositionalEncoding(nn.Module):
    """Rotary Position Embedding (RoPE). Zero trainable parameters."""

    def __init__(self, dim: int, max_seq_len: int = 512) -> None:
        super().__init__()
        inv_freq = 1.0 / (
            10000.0 ** (torch.arange(0, dim, 2).float() / dim)
        )
        self.register_buffer("inv_freq", inv_freq)
        self._build_cache(max_seq_len)

    def _build_cache(self, seq_len: int) -> None:
        inv_freq: torch.Tensor = self.inv_freq  # type: ignore[assignment]
        t = torch.arange(seq_len, dtype=inv_freq.dtype)
        freqs = torch.outer(t, inv_freq)
        cos_cached = freqs.cos()
        sin_cached = freqs.sin()
        self.register_buffer("cos_cached", cos_cached, persistent=False)
        self.register_buffer("sin_cached", sin_cached, persistent=False)

    def forward(self, x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        """Return cos and sin for sequence length of x.

        Args:
            x: (B, S, ...) tensor to determine sequence length.

        Returns:
            Tuple of (cos, sin) each of shape (S, d_head//2).
        """
        seq_len = x.shape[1]
        cos: torch.Tensor = self.cos_cached  # type: ignore[assignment]
        sin: torch.Tensor = self.sin_cached  # type: ignore[assignment]
        if seq_len > cos.shape[0] or not torch.isfinite(cos[:seq_len]).all():
            self._build_cache(max(seq_len, cos.shape[0]))
            cos = self.cos_cached  # type: ignore[assignment]
            sin = self.sin_cached  # type: ignore[assignment]
        return cos[:seq_len], sin[:seq_len]


def apply_rope(
    q: torch.Tensor,
    k: torch.Tensor,
    cos: torch.Tensor,
    sin: torch.Tensor,
) -> tuple[torch.Tensor, torch.Tensor]:
    """Apply rotary embeddings to query and key tensors.

    Args:
        q: (B, n_heads, S, d_head) query tensor.
        k: (B, n_heads, S, d_head) key tensor.
        cos: (S, d_head//2) cosine cache.
        sin: (S, d_head//2) sine cache.

    Returns:
        Rotated (q, k) tensors.
    """

    def _rotate(x: torch.Tensor) -> torch.Tensor:
        x1 = x[..., : x.shape[-1] // 2]
        x2 = x[..., x.shape[-1] // 2 :]
        cos_exp = cos.unsqueeze(0).unsqueeze(0)  # (1, 1, S, d_head//2)
        sin_exp = sin.unsqueeze(0).unsqueeze(0)
        return torch.cat(
            [x1 * cos_exp - x2 * sin_exp, x2 * cos_exp + x1 * sin_exp],
            dim=-1,
        )

    return _rotate(q), _rotate(k)