File size: 14,265 Bytes
807adfb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
# modeling_llada.py
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.checkpoint import checkpoint

from transformers import PreTrainedModel
from transformers.modeling_outputs import MaskedLMOutput
from ko_mini_llada.models.configuration_mini_llada import MiniLLaDAConfig

# RoPE (Rotary Positional Embedding)
def precompute_freqs_cis(dim: int, max_len: int, theta: float = 10000.0):
    """
    Precomputes the frequencies for the rotary positional embeddings.

    Args:
        dim (int): The dimension of the embeddings.
        max_len (int): The maximum sequence length.
        theta (float, optional): The theta parameter for frequency calculation. Defaults to 10000.0.

    Returns:
        torch.Tensor: Precomputed complex frequencies.
    """
    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
    t = torch.arange(max_len, dtype=torch.float)
    freqs = torch.outer(t, freqs)
    freqs_cis = torch.polar(torch.ones_like(freqs), freqs)
    return freqs_cis

def apply_rotary_emb(xq: torch.Tensor, xk: torch.Tensor, freqs_cis: torch.Tensor):
    """
    Applies rotary positional embeddings to input tensors.

    Args:
        xq (torch.Tensor): Query tensor.
        xk (torch.Tensor): Key tensor.
        freqs_cis (torch.Tensor): Precomputed complex frequencies.

    Returns:
        Tuple[torch.Tensor, torch.Tensor]: Tensors with rotary embeddings applied to query and key.
    """
    xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
    xk_ = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
    # Reshape freqs_cis for broadcasting. Assumes xq is (batch, n_heads, seq_len, head_dim)
    freqs_cis = freqs_cis[:xq_.shape[2]].view(1, 1, xq_.shape[2], xq_.shape[3])
    xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3)
    xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3)
    return xq_out.type_as(xq), xk_out.type_as(xk)

class Attention(nn.Module):
    """
    Multi-Head Attention module with Rotary Positional Embeddings.

    This module implements a multi-head attention mechanism, incorporating rotary
    positional embeddings (RoPE) for query and key tensors. It uses bidirectional
    attention as required by the LLaDA architecture.

    Args:
        dim (int): The input and output dimension of the module.
        heads (int): The number of attention heads.
    """
    def __init__(self, dim, heads):
        super().__init__()
        self.n_heads = heads
        self.head_dim = dim // heads
        self.q_proj = nn.Linear(dim, dim, bias=False)
        self.k_proj = nn.Linear(dim, dim, bias=False)
        self.v_proj = nn.Linear(dim, dim, bias=False)
        self.o_proj = nn.Linear(dim, dim, bias=False)

    def forward(self, x, freqs_cis, mask=None):
        """
        Forward pass for the Attention module.

        Args:
            x (torch.Tensor): Input tensor of shape (B, L, D).
            freqs_cis (torch.Tensor): Precomputed rotary frequencies.
            mask (torch.Tensor, optional): Attention mask. Defaults to None.

        Returns:
            torch.Tensor: Output tensor of shape (B, L, D).
        """
        B, L, D = x.shape
        q = self.q_proj(x).view(B, L, self.n_heads, self.head_dim).transpose(1, 2)
        k = self.k_proj(x).view(B, L, self.n_heads, self.head_dim).transpose(1, 2)
        v = self.v_proj(x).view(B, L, self.n_heads, self.head_dim).transpose(1, 2)

        freqs_cis = freqs_cis.to(x.device)
        q, k = apply_rotary_emb(q, k, freqs_cis[:L])

        # LLaDA: is_causal=False (Bidirectional Attention)
        attn_output = F.scaled_dot_product_attention(q, k, v, attn_mask=mask, is_causal=False)

        x = attn_output.transpose(1, 2).reshape(B, L, D)
        x = self.o_proj(x)
        return x

class Block(nn.Module):
    """
    Transformer block consisting of multi-head attention and a feed-forward network.

    This block applies pre-normalization (RMSNorm), followed by multi-head attention,
    a residual connection, another pre-normalization, a feed-forward network with
    SwiGLU activation, and a final residual connection.

    Args:
        dim (int): The dimension of the input and output.
        heads (int): The number of attention heads.
        intermediate_size (int): The intermediate size of the feed-forward network.
    """
    def __init__(self, dim, heads, intermediate_size):
        super().__init__()
        self.mha = Attention(dim, heads) # Multi-Head Attention
        self.norm1 = nn.RMSNorm(dim)
        self.norm2 = nn.RMSNorm(dim)

        # Feed-Forward Network with Gated Linear Unit
        self.gate_up_proj = nn.Linear(dim, intermediate_size * 2, bias=False)
        self.down_proj = nn.Linear(intermediate_size, dim, bias=False)
        self.gradient_checkpointing = False

    def run_block(self, x, freqs_cis, mask):
        """
        Actual forward pass logic.
        """
        # Multi-Head Attention
        residual = x

        x = self.norm1(x)
        x = self.mha(x, freqs_cis, mask)
        x = residual + x
        
        # Feed-Forward Network
        residual = x
        x = self.norm2(x)
        gate, up = self.gate_up_proj(x).chunk(2, dim=-1)
        x = F.silu(gate) * up
        x = self.down_proj(x)
        x = residual + x

        return x

    def forward(self, x, freqs_cis, mask=None):
        """
        Forward pass for the Transformer block.

        Args:
            x (torch.Tensor): Input tensor of shape (B, L, D).
            freqs_cis (torch.Tensor): Precomputed rotary frequencies.
            mask (torch.Tensor, optional): Attention mask. Defaults to None.

        Returns:
            torch.Tensor: Output tensor of shape (B, L, D).
        """
        if self.gradient_checkpointing and self.training:
            return checkpoint(self.run_block, x, freqs_cis, mask, use_reentrant=False)
        else:
            return self.run_block(x, freqs_cis, mask)

class Transformer(nn.Module):
    """
    The core Transformer model for Mini-LLaDA.

    This class stacks multiple Transformer blocks to form the main network.
    It includes token embeddings, a series of Transformer blocks, a final
    normalization layer, and a linear head to project to the vocabulary size.

    Args:
        vocab_size (int): The size of the vocabulary.
        dim (int): The embedding dimension.
        depth (int): The number of Transformer blocks.
        heads (int): The number of attention heads.
        intermediate_size (int): The intermediate size of the feed-forward networks.
        max_seq_len (int, optional): The maximum sequence length. Defaults to 2048.
    """
    def __init__(self, vocab_size, dim, depth, heads, intermediate_size, max_seq_len=2048):
        super().__init__()
        self.embed = nn.Embedding(vocab_size, dim) # Token Embedding
        self.layers = nn.ModuleList([
            Block(dim, heads, intermediate_size) for _ in range(depth)
        ])
        self.norm = nn.RMSNorm(dim)
        self.head = nn.Linear(dim, vocab_size, bias=False)

        freqs_cis = precompute_freqs_cis(dim // heads, max_seq_len)
        self.register_buffer("freqs_cis", freqs_cis, persistent=False)

    def forward(self, input_ids, attention_mask=None):
        """
        Forward pass for the Transformer model.

        Args:
            input_ids (torch.Tensor): Input token IDs of shape (B, L).
            attention_mask (torch.Tensor, optional): Mask to avoid attending to padding tokens.
                Shape (B, L). Defaults to None.

        Returns:
            torch.Tensor: Output logits of shape (B, L, vocab_size).
        """
        x = self.embed(input_ids)

        if attention_mask is not None:
            attention_mask = (attention_mask == 1).view(x.shape[0], 1, 1, x.shape[1])

        attention_mask = None

        for layer in self.layers:
            x = layer(x, self.freqs_cis, attention_mask)
        
        x = self.norm(x)
        x = self.head(x)
        return x

class MiniLLaDA(PreTrainedModel):
    """
    The Mini-LLaDA model, a Transformer-based model for masked language modeling
    inspired by diffusion models.

    This model is designed for pre-training using a diffusion-like noising process
    where a variable number of tokens are masked and the model learns to predict
    the original tokens. It can be used for both training (with labels) and
    inference (without labels).

    Args:
        config (MiniLLaDAConfig): The configuration object for the model.
    """
    config_class = MiniLLaDAConfig
    _supports_gradient_checkpointing = True

    def __init__(self, config: MiniLLaDAConfig):
        super().__init__(config)
        
        # 1. load backbone model
        self.network = Transformer(
            vocab_size=config.vocab_size,
            dim=config.dim,
            depth=config.depth,
            heads=config.head,
            intermediate_size=config.intermediate_size,
            max_seq_len=config.max_seq_len
        )
        self.mask_token_id = config.mask_token_id
        self.network.apply(self._init_weights)

    @property
    def supports_gradient_checkpointing(self):
        return True

    def forward(self, input_ids, attention_mask=None, labels=None, **kwargs):
        """
        Forward pass for the Mini-LLaDA model.

        If `labels` are provided, the model operates in training/evaluation mode,
        performing the diffusion forward process, running the noised input through
        the network, and computing the loss.

        If `labels` are not provided, the model operates in inference mode, simply
        passing the `input_ids` through the network to get logits.

        Args:
            input_ids (torch.Tensor): Input token IDs of shape (B, L).
            attention_mask (torch.Tensor, optional): Mask to avoid attending to padding tokens.
                Shape (B, L). Defaults to None.
            labels (torch.Tensor, optional): Labels for computing the loss. In SFT, this is used
                to identify which parts of the sequence to mask (assistant's response).
                Shape (B, L). Defaults to None.
            **kwargs: Additional keyword arguments.

        Returns:
            transformers.modeling_outputs.MaskedLMOutput: An output object containing the
                loss (if labels are provided) and logits.
        """
        # 1. Training and Evaluation Mode
        if labels is not None:
            # Diffusion Forward Process
            t, noisy_x, mask_indices = self.forward_process(input_ids, labels)
            
            # Reverse Process
            # network outputs: MaskedLMOutput (logits, hidden_states, etc.)
            outputs = self.network(input_ids=noisy_x, attention_mask=attention_mask)

            # Compute Loss
            loss = self.compute_diffusion_loss(outputs, input_ids, mask_indices, attention_mask)
            
            return MaskedLMOutput(loss=loss, logits=outputs)

        # 2. Inference Mode
        else:
            outputs = self.network(input_ids=input_ids, attention_mask=attention_mask)
            return MaskedLMOutput(logits=outputs)

    def forward_process(self, input_ids, labels=None):
        """
        Simulates the diffusion forward process by noising the input sequence.

        A random timestep `t` is sampled for each sequence in the batch, which
        determines the probability of a token being masked. Tokens are replaced
        with a `mask_token_id`. During supervised fine-tuning (SFT), masking is
        restricted to the assistant's response part of the sequence, identified
        by `labels != -100`.

        Args:
            input_ids (torch.Tensor): The original input token IDs of shape (B, L).
            labels (torch.Tensor, optional): Labels used to restrict masking during SFT.
                Defaults to None.

        Returns:
            Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: A tuple containing:
                - t (torch.Tensor): The sampled timesteps for each sequence (shape B).
                - noisy_x (torch.Tensor): The noised input IDs with masks (shape B, L).
                - mask_indices (torch.Tensor): A boolean tensor indicating which tokens
                  were masked (shape B, L).
        """
        B, L = input_ids.shape
        device = input_ids.device

        t = torch.rand(B, device=device)
        mask_probs = t.unsqueeze(1).expand(B, L)

        if labels is not None:
            train_mask = (labels != -100).float() 
            mask_probs = mask_probs * train_mask  # Make the probabilities zero where labels == -100, which is user message parts.

        random_matrix = torch.rand(B, L, device=device)
        mask_indices = (random_matrix < mask_probs)

        noisy_x = torch.where(mask_indices, self.mask_token_id, input_ids)
        return t, noisy_x, mask_indices

    def compute_diffusion_loss(self, logits, input_ids, mask_indices, attention_mask):
        """
        Computes the cross-entropy loss for the masked language modeling task.

        The loss is calculated only for the positions that were masked during the
        forward process. It also respects the attention mask to avoid computing
        loss on padding tokens.

        Args:
            logits (torch.Tensor): The model's output logits of shape (B, L, V).
            input_ids (torch.Tensor): The original input token IDs of shape (B, L).
            mask_indices (torch.Tensor): A boolean tensor indicating masked positions (shape B, L).
            attention_mask (torch.Tensor): The attention mask for the input (shape B, L).

        Returns:
            torch.Tensor: The computed cross-entropy loss.
        """
        B, L, V = logits.shape
        logits_flat = logits.view(-1, V)
        target_flat = input_ids.view(-1)
        
        # We compute loss only on tokens that were masked AND are not padding.
        final_loss_mask = mask_indices.view(-1)
        if attention_mask is not None:
            final_loss_mask = final_loss_mask & attention_mask.view(-1).bool()
        
        target_flat = torch.where(final_loss_mask, target_flat, -100)

        return F.cross_entropy(logits_flat, target_flat, ignore_index=-100)