File size: 18,400 Bytes
d65cc40
29b3bed
 
d65cc40
 
29b3bed
 
 
 
d65cc40
29b3bed
 
 
d65cc40
fc5319a
 
 
 
29b3bed
 
 
 
 
 
d65cc40
29b3bed
fc5319a
29b3bed
 
 
 
 
 
 
 
 
 
d65cc40
29b3bed
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d65cc40
29b3bed
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d65cc40
 
29b3bed
 
 
 
d65cc40
 
29b3bed
d65cc40
29b3bed
 
 
 
 
 
 
d65cc40
29b3bed
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d65cc40
29b3bed
 
 
d65cc40
29b3bed
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fc5319a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
29b3bed
 
 
 
d65cc40
fc5319a
 
29b3bed
 
 
 
fc5319a
29b3bed
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d65cc40
29b3bed
 
 
 
 
 
fc5319a
 
 
 
 
29b3bed
fc5319a
 
29b3bed
fc5319a
 
 
 
 
29b3bed
 
 
 
 
 
 
 
 
 
 
fc5319a
 
29b3bed
fc5319a
29b3bed
 
 
fc5319a
 
29b3bed
d65cc40
29b3bed
d65cc40
 
fc5319a
 
 
 
29b3bed
 
 
 
d65cc40
 
29b3bed
d65cc40
fc5319a
 
 
 
 
 
 
 
 
 
29b3bed
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fc5319a
 
 
 
 
29b3bed
 
fc5319a
d65cc40
29b3bed
 
 
fc5319a
 
29b3bed
 
 
 
 
fc5319a
 
29b3bed
 
fc5319a
 
 
 
 
 
29b3bed
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fc5319a
 
 
 
 
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
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
"""
GSLM Unit Language Model - HuggingFace Compatible Implementation
Based on fairseq's transformer_lm_big architecture
"""

import torch
import torch.nn as nn
import torch.nn.functional as F
import math
import os
import json
from typing import Optional, Tuple, Dict, Union, List
from dataclasses import dataclass

from transformers.modeling_utils import PreTrainedModel
from transformers.modeling_outputs import BaseModelOutput, CausalLMOutputWithCrossAttentions
from transformers import AutoConfig, AutoModelForCausalLM

# Import config - handle both local and remote imports
try:
    from .config import GSLMConfig
except ImportError:
    # Fallback for when file is accessed directly
    from config import GSLMConfig


# For backward compatibility with the API
@dataclass
class CausalLMOutput:
    loss: Optional[torch.FloatTensor] = None
    logits: Union[torch.FloatTensor, List[torch.FloatTensor]] = None
    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[torch.FloatTensor]] = None


class PositionalEncoding(nn.Module):
    """Sinusoidal positional encoding for transformer models."""
    
    def __init__(self, d_model: int, max_len: int = 5000):
        super().__init__()
        pe = torch.zeros(max_len, d_model)
        position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
        div_term = torch.exp(torch.arange(0, d_model, 2).float() * 
                           (-math.log(10000.0) / d_model))
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        self.register_buffer('pe', pe.unsqueeze(0))
        
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """Add positional encoding to input tensor."""
        return x + self.pe[:, :x.size(1)]


class MultiheadAttention(nn.Module):
    """Multi-head attention mechanism."""
    
    def __init__(self, embed_dim: int, num_heads: int, dropout: float = 0.0):
        super().__init__()
        assert embed_dim % num_heads == 0
        self.embed_dim = embed_dim
        self.num_heads = num_heads
        self.head_dim = embed_dim // num_heads
        self.scaling = self.head_dim ** -0.5
        
        self.k_proj = nn.Linear(embed_dim, embed_dim)
        self.v_proj = nn.Linear(embed_dim, embed_dim)
        self.q_proj = nn.Linear(embed_dim, embed_dim)
        self.out_proj = nn.Linear(embed_dim, embed_dim)
        self.attn_dropout = nn.Dropout(dropout)
        
    def forward(
        self, 
        query: torch.Tensor, 
        key: Optional[torch.Tensor] = None, 
        value: Optional[torch.Tensor] = None,
        attn_mask: Optional[torch.Tensor] = None,
        key_padding_mask: Optional[torch.Tensor] = None
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """
        Args:
            query: [batch_size, tgt_len, embed_dim]
            key: [batch_size, src_len, embed_dim]
            value: [batch_size, src_len, embed_dim]
            attn_mask: [tgt_len, src_len] or [batch_size * num_heads, tgt_len, src_len]
            key_padding_mask: [batch_size, src_len]
        """
        if key is None:
            key = query
        if value is None:
            value = query
            
        batch_size, tgt_len, embed_dim = query.size()
        src_len = key.size(1)
        
        # Project and reshape
        q = self.q_proj(query) * self.scaling
        k = self.k_proj(key)
        v = self.v_proj(value)
        
        q = q.view(batch_size, tgt_len, self.num_heads, self.head_dim).transpose(1, 2)
        k = k.view(batch_size, src_len, self.num_heads, self.head_dim).transpose(1, 2)
        v = v.view(batch_size, src_len, self.num_heads, self.head_dim).transpose(1, 2)
        
        # Compute attention scores
        attn_weights = torch.matmul(q, k.transpose(-2, -1))
        
        # Apply masks
        if attn_mask is not None:
            if attn_mask.dim() == 2:
                attn_mask = attn_mask.unsqueeze(0).unsqueeze(0)
            attn_weights = attn_weights + attn_mask
            
        if key_padding_mask is not None:
            attn_weights = attn_weights.masked_fill(
                key_padding_mask.unsqueeze(1).unsqueeze(2),
                float('-inf')
            )
            
        # Softmax
        attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).type_as(attn_weights)
        attn_weights = self.attn_dropout(attn_weights)
        
        # Apply attention to values
        attn_output = torch.matmul(attn_weights, v)
        attn_output = attn_output.transpose(1, 2).contiguous().view(
            batch_size, tgt_len, embed_dim
        )
        attn_output = self.out_proj(attn_output)
        
        return attn_output, attn_weights


class TransformerDecoderLayer(nn.Module):
    """Transformer decoder layer."""
    
    def __init__(
        self, 
        d_model: int, 
        nhead: int, 
        dim_feedforward: int = 2048, 
        dropout: float = 0.1,
        attention_dropout: float = 0.1,
        activation: str = "relu"
    ):
        super().__init__()
        self.self_attn = MultiheadAttention(d_model, nhead, dropout=attention_dropout)
        
        # Feedforward network
        self.linear1 = nn.Linear(d_model, dim_feedforward)
        self.dropout = nn.Dropout(dropout)
        self.linear2 = nn.Linear(dim_feedforward, d_model)
        
        # Layer normalization
        self.norm1 = nn.LayerNorm(d_model)
        self.norm2 = nn.LayerNorm(d_model)
        
        # Dropout modules
        self.dropout1 = nn.Dropout(dropout)
        self.dropout2 = nn.Dropout(dropout)
        
        # Activation
        self.activation = F.relu if activation == "relu" else F.gelu
        
    def forward(
        self, 
        x: torch.Tensor,
        self_attn_mask: Optional[torch.Tensor] = None,
        self_attn_padding_mask: Optional[torch.Tensor] = None
    ) -> torch.Tensor:
        """
        Args:
            x: [batch_size, seq_len, d_model]
            self_attn_mask: [seq_len, seq_len]
            self_attn_padding_mask: [batch_size, seq_len]
        """
        # Self-attention block
        residual = x
        x = self.norm1(x)
        x, _ = self.self_attn(x, x, x, self_attn_mask, self_attn_padding_mask)
        x = self.dropout1(x)
        x = residual + x
        
        # Feedforward block
        residual = x
        x = self.norm2(x)
        x = self.linear2(self.dropout(self.activation(self.linear1(x))))
        x = self.dropout2(x)
        x = residual + x
        
        return x


class GSLMPreTrainedModel(PreTrainedModel):
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models.
    """
    config_class = GSLMConfig
    base_model_prefix = "gslm"
    supports_gradient_checkpointing = True
    _no_split_modules = ["TransformerDecoderLayer"]

    def _init_weights(self, module):
        """Initialize the weights"""
        if isinstance(module, nn.Linear):
            module.weight.data.normal_(mean=0.0, std=self.config.d_model ** -0.5)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=self.config.d_model ** -0.5)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, nn.LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)


class GSLMForCausalLM(GSLMPreTrainedModel):
    """
    GSLM Unit Language Model - Transformer LM Big Architecture
    HuggingFace compatible version with modified forward API
    """
    
    def __init__(self, config: GSLMConfig):
        super().__init__(config)
        self.config = config
        
        self.d_model = config.d_model
        self.vocab_size = config.vocab_size
        self.pad_idx = config.pad_idx
        self.max_seq_length = config.max_seq_length
        
        # Create transformer module container for compatibility
        self.transformer = nn.Module()
        
        # Token embeddings (wte for compatibility)
        self.transformer.wte = nn.Embedding(config.vocab_size, config.d_model, padding_idx=self.pad_idx)
        self.embed_scale = math.sqrt(config.d_model)
        
        # Positional encoding
        self.pos_encoder = PositionalEncoding(config.d_model, config.max_seq_length)
        
        # Transformer decoder layers (h for compatibility)
        self.transformer.h = nn.ModuleList([
            TransformerDecoderLayer(
                config.d_model, 
                config.nhead, 
                config.dim_feedforward, 
                config.dropout, 
                config.attention_dropout
            ) for _ in range(config.num_layers)
        ])
        
        # Final layer norm (ln_f for compatibility)
        self.transformer.ln_f = nn.LayerNorm(config.d_model)
        
        # Output projection (coch_head for compatibility)
        if config.share_input_output_embed:
            self.coch_head = lambda x: F.linear(x, self.transformer.wte.weight)
        else:
            self.coch_head = nn.Linear(config.d_model, config.vocab_size, bias=False)
            
        # Dropout
        self.transformer.drop = nn.Dropout(config.dropout)
        
        # Future heads not supported in GSLM
        self.future_heads = None
        
        # Initialize weights
        self.post_init()
        
    def get_input_embeddings(self):
        return self.transformer.wte
        
    def set_input_embeddings(self, value):
        self.transformer.wte = value
        
    def get_output_embeddings(self):
        if self.config.share_input_output_embed:
            return self.transformer.wte
        else:
            return self.coch_head
            
    def _create_causal_mask(self, seq_len: int, device) -> torch.Tensor:
        """Create causal attention mask."""
        mask = torch.triu(
            torch.full((seq_len, seq_len), float('-inf'), device=device), 
            diagonal=1
        )
        return mask
        
    def forward(
        self, 
        seq=None,
        input_ids=None,
        tgt=None,
        labels=None,
        output_logits=False,
        output_hidden_states=False,
        return_dict=False,
        up_until_layer=None,
        **kwargs
    ):
        """
        Compatible forward method with the specified API.
        
        Args:
            seq: torch.Tensor of shape (b, t) - input token IDs (legacy)
            input_ids: torch.Tensor of shape (b, t) - input token IDs (HF standard)
            tgt: torch.Tensor of shape (b, t) or None - target token IDs (legacy)
            labels: torch.Tensor of shape (b, t) or None - target token IDs (HF standard)
            output_logits: bool - whether to output logits
            output_hidden_states: bool - whether to output all hidden states
            return_dict: bool - whether to return dictionary output
            up_until_layer: int or None - stop at specific layer
            
        Returns:
            Depending on return_dict and other flags
        """
        # Handle both seq and input_ids for compatibility
        if seq is None and input_ids is not None:
            seq = input_ids
        elif seq is None and input_ids is None:
            raise ValueError("Either 'seq' or 'input_ids' must be provided")
            
        # Handle both tgt and labels for compatibility
        if tgt is None and labels is not None:
            tgt = labels
            
        batch_size, seq_len = seq.shape
        device = seq.device
        
        # Create causal mask
        causal_mask = self._create_causal_mask(seq_len, device)
        
        # Create padding mask
        padding_mask = seq.eq(self.pad_idx)
        
        # Token embeddings
        tok_emb = self.transformer.wte(seq) * self.embed_scale
        
        # Add positional encoding (sinusoidal, not learned)
        x = self.pos_encoder(tok_emb)
        x = self.transformer.drop(x)
        
        all_hidden_states = []
        
        # Pass through transformer layers
        for block_idx, block in enumerate(self.transformer.h):
            # Save hidden state before block
            if output_hidden_states:
                all_hidden_states.append(x)
                
            # Check if we should stop early
            if up_until_layer is not None and block_idx == up_until_layer:
                break
                
            # Forward the block
            x = block(x, causal_mask, padding_mask)
        
        # Append the last hidden state if we didn't exit early
        if output_hidden_states and (up_until_layer is None or block_idx == len(self.transformer.h) - 1):
            all_hidden_states.append(x)
        
        # If only hidden states requested
        if output_hidden_states and not output_logits and tgt is None:
            model_output = BaseModelOutput(
                last_hidden_state=x,
                hidden_states=tuple(all_hidden_states) if all_hidden_states else None,
            )
            return model_output
        
        # Final layer norm
        x = self.transformer.ln_f(x)
        
        # Compute logits
        logits = self.coch_head(x)
        
        # Compute loss if targets provided
        if tgt is not None:
            # Shift so that tokens < n predict n
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = tgt[..., 1:].contiguous()
            
            loss = F.cross_entropy(
                shift_logits.reshape(-1, self.config.vocab_size),
                shift_labels.reshape(-1),
                ignore_index=self.pad_idx
            )
            
            if return_dict:
                if output_logits:
                    # For compatibility, wrap single logits in list
                    all_logits = [logits]
                    
                if output_hidden_states:
                    model_output = CausalLMOutput(
                        loss=loss,
                        logits=all_logits if output_logits else logits,
                        hidden_states=tuple(all_hidden_states) if all_hidden_states else None,
                    )
                else:
                    model_output = CausalLMOutput(
                        loss=loss,
                        logits=all_logits if output_logits else logits,
                    )
                return model_output
            
            return logits, loss
        
        # No targets provided
        if return_dict:
            return CausalLMOutputWithCrossAttentions(
                logits=logits,
                hidden_states=tuple(all_hidden_states) if output_hidden_states else None,
            )
        
        return logits, None
    
    @torch.no_grad()
    def generate(
        self,
        input_ids: torch.Tensor = None,
        seq: torch.Tensor = None,
        max_length: int = 100,
        temperature: float = 1.0,
        top_k: Optional[int] = None,
        top_p: Optional[float] = None,
        pad_token_id: Optional[int] = None,
        eos_token_id: Optional[int] = None,
        **kwargs
    ) -> torch.Tensor:
        """Generate sequences using the language model."""
        # Handle both input_ids and seq
        if input_ids is None and seq is not None:
            input_ids = seq
        elif input_ids is None:
            raise ValueError("Either 'input_ids' or 'seq' must be provided")
            
        if pad_token_id is None:
            pad_token_id = self.pad_idx
            
        batch_size = input_ids.shape[0]
        device = input_ids.device
        
        # Keep track of which sequences are done
        unfinished_sequences = torch.ones(batch_size, dtype=torch.long, device=device)
        
        while input_ids.shape[1] < max_length:
            # Forward pass
            logits, _ = self.forward(input_ids)
            next_token_logits = logits[:, -1, :]
            
            # Apply temperature
            if temperature != 1.0:
                next_token_logits = next_token_logits / temperature
                
            # Apply top-k sampling
            if top_k is not None:
                indices_to_remove = next_token_logits < torch.topk(next_token_logits, top_k)[0][..., -1, None]
                next_token_logits[indices_to_remove] = -float('inf')
                
            # Apply top-p (nucleus) sampling  
            if top_p is not None:
                sorted_logits, sorted_indices = torch.sort(next_token_logits, descending=True)
                cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
                
                # Remove tokens with cumulative probability above the threshold
                sorted_indices_to_remove = cumulative_probs > top_p
                sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
                sorted_indices_to_remove[..., 0] = 0
                
                indices_to_remove = sorted_indices_to_remove.scatter(
                    dim=-1, index=sorted_indices, src=sorted_indices_to_remove
                )
                next_token_logits[indices_to_remove] = -float('inf')
                
            # Sample from the distribution
            probs = F.softmax(next_token_logits, dim=-1)
            next_tokens = torch.multinomial(probs, num_samples=1).squeeze(-1)
            
            # Update unfinished sequences
            if eos_token_id is not None:
                tokens_to_add = next_tokens * unfinished_sequences + pad_token_id * (1 - unfinished_sequences)
                unfinished_sequences = unfinished_sequences * (next_tokens != eos_token_id).long()
            else:
                tokens_to_add = next_tokens
                
            # Concatenate tokens
            input_ids = torch.cat([input_ids, tokens_to_add.unsqueeze(-1)], dim=-1)
            
            # Stop if all sequences are finished
            if eos_token_id is not None and unfinished_sequences.sum() == 0:
                break
                
        return input_ids


# Register the model with AutoModel
AutoConfig.register("gslm", GSLMConfig)
AutoModelForCausalLM.register(GSLMConfig, GSLMForCausalLM)