File size: 1,886 Bytes
e118aba
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""V4 decoder config.

Three variants for the ablation:
- A: full LASER2 per-layer cross-attention
- B: no LASER2 (pure decoder baseline)
- C: LASER2 input-only (first layer only)
"""

from dataclasses import dataclass, field


@dataclass
class V4Config:
    # Vocab — LASER2 SPM 50K
    vocab_size: int = 50004  # LASER2 fairseq dict: bos/pad/eos/unk + 50000 SPM
    pad_token_id: int = 1
    bos_token_id: int = 0
    eos_token_id: int = 2

    # Model shape
    n_layer: int = 28
    n_embd: int = 1024
    n_head: int = 16         # query heads
    n_kv_head: int = 2       # GQA 8:1
    head_dim: int = 64

    # FFN
    ffn_mult: float = 5.4    # SwiGLU 5.4× = 5504 hidden
    ffn_hidden: int = 5504   # explicit

    # Context
    max_seq_len: int = 2048

    # Positional
    rope_theta: float = 10000.0

    # Cross-attention to LASER2
    cross_attention_mode: str = "per_layer"  # "per_layer" | "input_only" | "none"
    laser_dim: int = 1024    # LASER2 BiLSTM output is 512*2 = 1024

    # Training
    dropout: float = 0.0
    tied_embeddings: bool = True

    # Init
    init_std: float = 0.02


def config_variant_a() -> V4Config:
    """Variant A: full LASER2 per-layer cross-attention."""
    return V4Config(cross_attention_mode="per_layer")


def config_variant_b() -> V4Config:
    """Variant B: pure decoder baseline (no LASER2)."""
    return V4Config(cross_attention_mode="none")


def config_variant_c() -> V4Config:
    """Variant C: LASER2 input-only (first layer)."""
    return V4Config(cross_attention_mode="input_only")


def config_test() -> V4Config:
    """Small config for unit tests."""
    return V4Config(
        vocab_size=256,
        n_layer=2,
        n_embd=64,
        n_head=4,
        n_kv_head=2,
        head_dim=16,
        ffn_mult=4,
        ffn_hidden=256,
        max_seq_len=128,
        cross_attention_mode="per_layer",
    )