File size: 9,232 Bytes
815415a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
from .configuration_mdlm import MDLMConfig

import math 

import torch 
import torch.nn as nn
import torch.nn.functional as F

import transformers
from transformers.modeling_attn_mask_utils import _prepare_4d_attention_mask


## ROPE 
class Rotary(nn.Module):
    def __init__(self, head_dim: int, base: int = 10_000):
        super().__init__()
        inv_freq = 1.0 / (base ** (torch.arange(0, head_dim, 2).float() / head_dim))
        # persistent=True for checkpoint compatibility (modern convention is False,
        # but the published MDLM state-dict includes this buffer).
        self.register_buffer("inv_freq", inv_freq, persistent=True)
        self._seq_len_cached = 0
        self._cos_cached = None
        self._sin_cached = None

    def _build_cache(self, seq_len: int, device, dtype):
        t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype)
        freqs = torch.outer(t, self.inv_freq)               # (T, Dh/2)
        emb = torch.cat((freqs, freqs), dim=-1)             # (T, Dh)
        self._cos_cached = emb.cos().to(dtype)
        self._sin_cached = emb.sin().to(dtype)
        self._seq_len_cached = seq_len

    def forward(self, seq_len: int, device, dtype):
        if (self._cos_cached is None
            or seq_len > self._seq_len_cached
            or self._cos_cached.device != device
            or self._cos_cached.dtype  != dtype):
            self._build_cache(seq_len, device, dtype)
        return self._cos_cached[:seq_len], self._sin_cached[:seq_len]


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


def apply_rope(q, k, cos, sin):
    # q, k: (B, H, T, Dh); cos, sin: (T, Dh) → broadcast to (1, 1, T, Dh)
    return (q * cos + rotate_half(q) * sin, k * cos + rotate_half(k) * sin)


## EMBED 
class TimestepEmbedder(nn.Module):
  def __init__(self, cond_dim: int, freq_dim: int = 256):
    super().__init__()
    self.mlp = nn.Sequential(
      nn.Linear(freq_dim, cond_dim, bias=True),
      nn.SiLU(),
      nn.Linear(cond_dim, cond_dim, bias=True))
    self.freq_dim = freq_dim

  def _fourier_features(self, t, max_period: int = 10_000):
    half = self.freq_dim // 2
    freqs = torch.exp(
        -math.log(max_period)
        * torch.arange(half, dtype=torch.float32, device=t.device)
        / half
    )
    args = t[:, None].float() * freqs[None]
    emb = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
    if self.freq_dim % 2:
      emb = torch.cat([emb, torch.zeros_like(emb[:, :1])], dim=-1)
    return emb

  def forward(self, t):
    return self.mlp(self._fourier_features(t))
  

# LEGACY EMBEDDING
class EmbeddingLayer(nn.Module):
  def __init__(self, hidden_dim, vocab_size):
    super().__init__()
    self.embedding = nn.Parameter(torch.empty((vocab_size, hidden_dim)))
    torch.nn.init.kaiming_uniform_(self.embedding, a=math.sqrt(5))

  def forward(self, x):
    return self.embedding[x]
  

## LM HEAD
class DDitFinalLayer(nn.Module):
  def __init__(self, hidden_dim: int, vocab_size: int, cond_dim: int):
    super().__init__()
    self.norm_final   = nn.LayerNorm(hidden_dim, bias=False)
    self.linear       = nn.Linear(hidden_dim, vocab_size)
    self.linear.weight.data.zero_()
    self.linear.bias.data.zero_()

    self.adaLN_modulation = nn.Linear(cond_dim, 2 * hidden_dim, bias=True)
    self.adaLN_modulation.weight.data.zero_()
    self.adaLN_modulation.bias.data.zero_()

  def forward(self, x, c):
    shift, scale = self.adaLN_modulation(c)[:, None].chunk(2, dim=2)
    return self.linear(modulate(self.norm_final(x), shift, scale))  
  

## TF BLOCK 
def modulate(x, shift, scale): return x * (1 + scale) + shift

class DDiTBlock(nn.Module):
    def __init__(self, hidden_dim, n_heads, cond_dim, mlp_ratio: int = 4, dropout: float = 0.1):
        super().__init__()
        self.n_heads  = n_heads
        self.head_dim = hidden_dim // n_heads
        self.dropout  = dropout
        self.mlp_ratio = mlp_ratio        

        self.norm1 = nn.LayerNorm(hidden_dim, bias=False)   # PyTorch ≥ 2.1 supports `bias=False`
        self.norm2 = nn.LayerNorm(hidden_dim, bias=False)   

        self.mlp = nn.Sequential(
          nn.Linear(hidden_dim, mlp_ratio * hidden_dim, bias=True),
          nn.GELU(approximate='tanh'),
          nn.Linear(mlp_ratio * hidden_dim, hidden_dim, bias=True))        
        
        self.attn_qkv = nn.Linear(hidden_dim, 3 * hidden_dim, bias=False) 
        self.attn_out = nn.Linear(hidden_dim, hidden_dim, bias=False)     ### ATT OUT

        self.adaLN_modulation = nn.Linear(cond_dim, 6 * hidden_dim, bias=True)         
        self.adaLN_modulation.weight.data.zero_()           
        self.adaLN_modulation.bias.data.zero_()             


    def forward(self, x, c, rotary_cos_sin, attention_mask=None):
        B, T, D = x.shape[0], x.shape[1], x.shape[2]

        (shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp) = self.adaLN_modulation(c)[:,None].chunk(6, dim=2)       

        x_skip  = x        
        x       = modulate(self.norm1(x), shift_msa, scale_msa)
        qkv     = self.attn_qkv(x).reshape(B, T, 3, self.n_heads, self.head_dim).permute(2, 0, 3, 1, 4)
        q, k, v = qkv.unbind(0)

        cos, sin = rotary_cos_sin
        q, k     = apply_rope(q, k, cos, sin)        #  new per-tensor RoPE
        att = F.scaled_dot_product_attention(q, k, v, is_causal=False, attn_mask=attention_mask).transpose(1, 2).contiguous().reshape(B, T, D)

        # ---- attention sub-block ----
        x = x_skip + gate_msa * F.dropout(
            self.attn_out(att), p=self.dropout, training=self.training)

        # ---- MLP sub-block ----
        x = x + gate_mlp * F.dropout(
            self.mlp(modulate(self.norm2(x), shift_mlp, scale_mlp)),
            p=self.dropout, training=self.training)
        return x


## LM 
class DITBackbone(nn.Module):
  def __init__(self, config):
    super().__init__()
    self.config = config
    self.vocab_embed = EmbeddingLayer(config.hidden_dim, config.vocab_size)
    self.sigma_map   = TimestepEmbedder(config.cond_dim)
    self.rotary_emb  = Rotary(config.hidden_dim // config.n_heads)    

    self.blocks = nn.ModuleList([
    DDiTBlock(config.hidden_dim,
              config.n_heads,
              config.cond_dim,
              dropout=config.dropout)
    for _ in range(config.n_blocks)
    ])
    self.output_layer = DDitFinalLayer(config.hidden_dim, config.vocab_size, config.cond_dim)

  def forward(self, input_ids, sigma, attention_mask=None, output_hidden_states=False):  

    if not self.config.time_conditioning:
      sigma = torch.zeros_like(sigma)    

    all_hidden_states = []
    x = self.vocab_embed(input_ids)
    if output_hidden_states: all_hidden_states.append(x)

    c = F.silu(self.sigma_map(sigma))     
    rotary_cos_sin = self.rotary_emb(x.shape[1], x.device, x.dtype)
    # --- prepare attention mask once (bidirectional, padding-only) ---------
    # SDPA expects either None, a bool/float (B,*,T,T) bias, or to be told
    # is_causal=True. A (B,T) padding mask must be expanded to an additive
    # (B,1,1,T) bias with -inf on pad keys.
    if attention_mask is not None and attention_mask.dim() == 2:
        attention_mask = _prepare_4d_attention_mask(
            attention_mask, dtype=x.dtype
        )

    for i in range(len(self.blocks)):
        x = self.blocks[i](x, c ,rotary_cos_sin, attention_mask=attention_mask)
        if output_hidden_states: all_hidden_states.append(x)

    logits = self.output_layer(x, c)
    return logits, all_hidden_states     



class MDLM(transformers.PreTrainedModel):
    config_class = MDLMConfig
    base_model_prefix = "mdlm"
    _tied_weights_keys = []  # Explicitly declare no tied weights

    def __init__(self, config: MDLMConfig):
        super().__init__(config)
        self.backbone = DITBackbone(config)
        # post_init() is called automatically by PreTrainedModel.from_pretrained()
        self.post_init()
        
    def forward(self, input_ids=None, timesteps=None, attention_mask=None, output_hidden_states=None, return_dict=None, labels=None,**kwargs):
        # Use config defaults only if not provided
        output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # Default timesteps if not provided
        if timesteps is None:
            timesteps = torch.zeros(input_ids.shape[0], device=input_ids.device, dtype=torch.float32)

        # Forward pass
        logits, all_hidden_states = self.backbone(
            input_ids=input_ids,
            sigma=timesteps,
            attention_mask=attention_mask,
            output_hidden_states=output_hidden_states
        )

        # Return based on return_dict flag
        if return_dict:
            return transformers.modeling_outputs.MaskedLMOutput(
                logits=logits,
                hidden_states=all_hidden_states if output_hidden_states else None,
                loss=None
            )
        
        # Non-dict return
        return (logits, all_hidden_states) if output_hidden_states else logits