File size: 6,885 Bytes
df43f42
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4f9392a
 
df43f42
4f9392a
 
 
 
 
df43f42
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4f9392a
 
df43f42
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
"""
yk_diffusion: a from-scratch hybrid language model.

A single Transformer runs in two modes, selected by a learned mode embedding:
  - AR mode      (mode=0): causal attention -> standard next-token autoregressive LM ("normal")
  - DIFF mode    (mode=1): bidirectional attention -> masked discrete diffusion denoising

A time embedding conditions the diffusion mask ratio (MDLM / LLaDA-style absorbing-state training).
The same weights serve both behaviours; a <MODE> signal picks which one at inference time.
"""

import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.checkpoint import checkpoint


# ----------------------------------------------------------------------------
# Building blocks
# ----------------------------------------------------------------------------

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

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


class RotaryEmbedding(nn.Module):
    def __init__(self, head_dim, max_len=8192, base=10000.0,
                 rope_scale=1.0):
        super().__init__()
        # NTK-aware scaling: stretch the base so the trained (short) context
        # extends to longer inference windows without retraining. With
        # rope_scale = context_ratio (e.g. 128k/8k = 16) the model trained at
        # 8k still positions tokens correctly at 128k.
        base = base * (rope_scale ** (head_dim / (head_dim - 2)))
        inv = 1.0 / (base ** (torch.arange(0, head_dim, 2, dtype=torch.float32) / head_dim))
        self.register_buffer("inv_freq", inv)

    def forward(self, seq_len, device):
        t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype)
        freqs = torch.outer(t, self.inv_freq)            # [L, head_dim/2]
        emb = torch.cat([freqs, freqs], dim=-1)          # [L, head_dim]
        return emb.cos(), emb.sin()


def rotate_half(x):
    x1, x2 = x.chunk(2, dim=-1)
    return torch.cat((-x2, x1), dim=-1)


def apply_rope(x, cos, sin):
    # x: [B, h, L, hd];  cos/sin: [1, 1, L, hd]
    return x * cos + rotate_half(x) * sin


class Attention(nn.Module):
    def __init__(self, d, n_heads):
        super().__init__()
        assert d % n_heads == 0
        self.d = d
        self.n_heads = n_heads
        self.hd = d // n_heads
        self.scale = self.hd ** -0.5
        self.qkv = nn.Linear(d, 3 * d, bias=False)
        self.proj = nn.Linear(d, d, bias=False)

    def forward(self, x, cos, sin, attn_mask=None):
        B, L, _ = x.shape
        qkv = self.qkv(x).reshape(B, L, 3, self.n_heads, self.hd)
        qkv = qkv.permute(2, 0, 3, 1, 4)             # [3, B, h, L, hd]
        q, k, v = qkv[0], qkv[1], qkv[2]
        q = apply_rope(q, cos, sin)
        k = apply_rope(k, cos, sin)
        scores = (q @ k.transpose(-2, -1)) * self.scale
        if attn_mask is not None:
            scores = scores + attn_mask
        attn = scores.softmax(dim=-1)
        out = (attn @ v).transpose(1, 2).reshape(B, L, self.d)
        return self.proj(out)


class MLP(nn.Module):
    def __init__(self, d, d_ff):
        super().__init__()
        self.w1 = nn.Linear(d, d_ff, bias=False)
        self.w3 = nn.Linear(d, d_ff, bias=False)
        self.w2 = nn.Linear(d_ff, d, bias=False)

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


class Block(nn.Module):
    def __init__(self, d, n_heads, d_ff):
        super().__init__()
        self.ln1 = RMSNorm(d)
        self.attn = Attention(d, n_heads)
        self.ln2 = RMSNorm(d)
        self.mlp = MLP(d, d_ff)

    def forward(self, x, cos, sin, attn_mask):
        x = x + self.attn(self.ln1(x), cos, sin, attn_mask)
        x = x + self.mlp(self.ln2(x))
        return x


# ----------------------------------------------------------------------------
# The model
# ----------------------------------------------------------------------------

class YKDiff(nn.Module):
    def __init__(self, cfg):
        super().__init__()
        self.cfg = cfg
        d = cfg["d_model"]
        self.vocab = cfg["vocab_size"]
        self.max_len = cfg["max_len"]

        self.tok_emb = nn.Embedding(self.vocab, d)
        self.mode_emb = nn.Embedding(2, d)          # 0 = AR, 1 = DIFF
        self.time_emb = nn.Linear(1, d, bias=False)  # diffusion mask ratio conditioning
        self.rope = RotaryEmbedding(d // cfg["n_heads"], max_len=self.max_len,
                                     rope_scale=cfg.get("rope_scale", 1.0))

        self.blocks = nn.ModuleList([
            Block(d, cfg["n_heads"], cfg["d_ff"]) for _ in range(cfg["n_layers"])
        ])
        self.norm = RMSNorm(d)
        self.lm_head = nn.Linear(d, self.vocab, bias=False)
        with torch.no_grad():
            self.lm_head.weight.copy_(self.tok_emb.weight)   # weight tying

        self._causal = None

    @property
    def device(self):
        return next(self.parameters()).device

    def _causal_mask(self, L, device):
        if self._causal is None or self._causal.shape[-1] < L:
            m = torch.full((L, L), float("-inf"), device=device)
            m = torch.triu(m, diagonal=1)
            self._causal = m
        return self._causal[:L, :L]

    def forward(self, idx, mode, t=None, attn_mask=None):
        """
        idx:        [B, L] long   token ids
        mode:       [B] long      (0 AR, 1 DIFF)
        t:          [B] float|None  diffusion mask ratio (None -> 0)
        attn_mask:  [L, L]|None    explicit mask; if None, AR uses causal, DIFF uses none
        """
        B, L = idx.shape
        x = self.tok_emb(idx)
        x = x + self.mode_emb(mode).unsqueeze(1)
        if t is None:
            t = torch.zeros(B, device=idx.device)
        x = x + self.time_emb(t.unsqueeze(-1)).unsqueeze(1)

        cos, sin = self.rope(L, idx.device)
        cos = cos.unsqueeze(0).unsqueeze(0)        # [1,1,L,hd]
        sin = sin.unsqueeze(0).unsqueeze(0)

        if attn_mask is None:
            # AR default causal; DIFF default bidirectional (None)
            attn_mask = self._causal_mask(L, idx.device) if mode[0].item() == 0 else None
        else:
            attn_mask = attn_mask.to(idx.device)

        # gradient checkpointing: trade ~20% compute for ~4x less
        # activation memory, so a big batch fits on 16 GB.
        training = self.training and torch.is_grad_enabled()
        for blk in self.blocks:
            if training:
                x = checkpoint(blk, x, cos, sin, attn_mask,
                               use_reentrant=False)
            else:
                x = blk(x, cos, sin, attn_mask)
        x = self.norm(x)
        return self.lm_head(x)                     # [B, L, vocab]