File size: 16,837 Bytes
f206855
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b4b0382
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional, Tuple, List, Union
from transformers import PreTrainedModel, PretrainedConfig, GenerationMixin
from transformers.modeling_outputs import CausalLMOutputWithPast
from transformers.cache_utils import Cache, DynamicCache


# ── Config ────────────────────────────────────────────────────────────────────

class DotLMConfig(PretrainedConfig):
    model_type = "dotlm"

    def __init__(
        self,
        vocab_size=16384,
        d_model=768,
        hidden_dim=2048,
        num_hidden_layers=24,
        n_heads=6,
        n_kv_heads=2,
        context_len=4096,
        theta_base=10000.0,
        norm_eps=1e-6,
        initializer_range=0.02,
        tie_word_embeddings=True,
        **kwargs
    ):
        super().__init__(**kwargs)
        self.vocab_size = vocab_size
        self.d_model = d_model
        self.hidden_dim = hidden_dim
        self.num_hidden_layers = num_hidden_layers
        self.n_heads = n_heads
        self.n_kv_heads = n_kv_heads
        self.context_len = context_len
        self.theta_base = theta_base
        self.norm_eps = norm_eps
        self.initializer_range = initializer_range
        self.tie_word_embeddings = tie_word_embeddings
        self.use_cache = kwargs.get("use_cache", True)
        self.pad_token_id = kwargs.get("pad_token_id", 0)
        self.bos_token_id = kwargs.get("bos_token_id", None)
        self.eos_token_id = kwargs.get("eos_token_id", 3)


# ── Architecture Components ───────────────────────────────────────────────────

def precompute_freqs_cis(dim, context_len, theta_base=10000.0):
    theta = 1.0 / (theta_base ** (torch.arange(0, dim, 2) / dim))
    seq_ids = torch.arange(context_len, dtype=torch.float32)
    m_theta = torch.outer(seq_ids, theta)
    m_theta = torch.cat([m_theta, m_theta], dim=-1)
    return torch.cos(m_theta), torch.sin(m_theta)


class SwiGLU(nn.Module):
    def __init__(self, d_model, hidden_dim):
        super().__init__()
        self.W = nn.Linear(d_model, hidden_dim, bias=False)
        self.V = nn.Linear(d_model, hidden_dim, bias=False)
        self.W2 = nn.Linear(hidden_dim, d_model, bias=False)
        self.silu = nn.SiLU()

    def forward(self, x):
        return self.W2(self.silu(self.W(x)) * self.V(x))


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

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


class RoPE(nn.Module):
    def forward(self, x, cos, sin):
        batch_size, num_heads, seq_len, head_dim = x.shape
        x1, x2 = x[..., : head_dim // 2], x[..., head_dim // 2 :]
        x_rotated = torch.cat([-x2, x1], dim=-1)
        return x * cos + x_rotated * sin


class GroupedQueryAttention(nn.Module):
    def __init__(self, d_model, n_heads, head_dim, n_kv_groups):
        super().__init__()
        self.n_heads = n_heads
        self.head_dim = head_dim
        self.n_kv_groups = n_kv_groups
        self.group_size = n_heads // n_kv_groups
        self.output_dim = n_heads * head_dim

        self.Wq = nn.Linear(d_model, self.output_dim, bias=False)
        self.Wk = nn.Linear(d_model, n_kv_groups * head_dim, bias=False)
        self.Wv = nn.Linear(d_model, n_kv_groups * head_dim, bias=False)
        self.Wo = nn.Linear(self.output_dim, d_model, bias=False)
        self.q_norm = RMSNorm(head_dim)
        self.k_norm = RMSNorm(head_dim)
        self.rope = RoPE()

    def forward(self, x, cos, sin, mask=None, past_key_value=None, use_cache=False):
        B, S, _ = x.shape

        q = self.Wq(x).view(B, S, self.n_heads, self.head_dim).transpose(1, 2)
        k = self.Wk(x).view(B, S, self.n_kv_groups, self.head_dim).transpose(1, 2)
        v = self.Wv(x).view(B, S, self.n_kv_groups, self.head_dim).transpose(1, 2)

        q, k = self.q_norm(q), self.k_norm(k)
        q, k = self.rope(q, cos, sin), self.rope(k, cos, sin)

        next_past = None
        if past_key_value is not None:
            if isinstance(past_key_value, Cache):
                # HF DynamicCache: update in-place and get concatenated K/V back.
                k, v = past_key_value.update(k, v, self.layer_idx)
                next_past = past_key_value
            else:
                # Legacy cache format: (k, v) per layer. Some generation paths
                # may pass placeholders like (None, None) on the first step.
                pk, pv = past_key_value
                if pk is not None:
                    k = torch.cat([pk, k], dim=2)
                    v = torch.cat([pv, v], dim=2)
                next_past = (k, v) if use_cache else None

        # Cache stores grouped K/V (n_kv_groups heads). We only expand for SDPA.
        kv_k, kv_v = k, v

        B, G, S_kv, D = kv_k.shape
        k = kv_k.unsqueeze(2).expand(B, G, self.group_size, S_kv, D).reshape(B, self.n_heads, S_kv, D)
        v = kv_v.unsqueeze(2).expand(B, G, self.group_size, S_kv, D).reshape(B, self.n_heads, S_kv, D)

        # Causal logic for SDPA: if mask is None, we assume causality if prefill
        # But for robustness, we always pass a mask if S > 1
        is_causal = (mask is None and S > 1 and past_key_value is None)
        
        out = F.scaled_dot_product_attention(
            q, k, v,
            attn_mask=None if (mask is None or is_causal) else ~mask,
            dropout_p=0.0,
            is_causal=is_causal,
        )
        out = out.transpose(1, 2).reshape(B, S, self.output_dim)
        if use_cache and past_key_value is None:
            # If we're not given a cache, return legacy K/V by default.
            next_past = (kv_k, kv_v)
        return self.Wo(out), next_past


class DotLMBlock(nn.Module):
    def __init__(self, d_model, n_heads, n_kv_heads, hidden_dim, norm_eps=1e-6, layer_idx=None):
        super().__init__()
        head_dim = d_model // n_heads
        self.attention = GroupedQueryAttention(d_model, n_heads, head_dim, n_kv_heads)
        self.attention.layer_idx = layer_idx
        self.feed_forward = SwiGLU(d_model, hidden_dim)
        self.norm1 = RMSNorm(d_model, norm_eps)
        self.norm2 = RMSNorm(d_model, norm_eps)

    def forward(self, x, cos, sin, mask=None, past_key_value=None, use_cache=False):
        residual = x
        x = self.norm1(x)
        attn_out, next_past = self.attention(x, cos, sin, mask, past_key_value, use_cache)
        x = residual + attn_out

        residual = x
        x = self.norm2(x)
        x = residual + self.feed_forward(x)
        return x, next_past


# ── Flat HF Wrapper ───────────────────────────────────────────────────────────

class DotLMForCausalLM(PreTrainedModel, GenerationMixin):
    config_class = DotLMConfig
    # Let HF know output head is tied to embeddings when enabled.
    _tied_weights_keys = {"head.weight": "embeddor.weight"}

    def __init__(self, config):
        super().__init__(config)
        self.config = config

        self.embeddor = nn.Embedding(config.vocab_size, config.d_model)
        self.blocks = nn.ModuleList([
            DotLMBlock(
                config.d_model, config.n_heads, config.n_kv_heads,
                config.hidden_dim, config.norm_eps, layer_idx=i
            )
            for i in range(config.num_hidden_layers)
        ])
        self.norm = RMSNorm(config.d_model, config.norm_eps)
        self.head = nn.Linear(config.d_model, config.vocab_size, bias=False)

        # Precompute RoPE
        head_dim = config.d_model // config.n_heads
        cos, sin = precompute_freqs_cis(head_dim, config.context_len, config.theta_base)
        self.register_buffer("cos_cache", cos, persistent=False)
        self.register_buffer("sin_cache", sin, persistent=False)

        # Causal mask placeholder
        mask = torch.triu(torch.ones(config.context_len, config.context_len, dtype=torch.bool), diagonal=1)
        self.register_buffer("causal_mask", mask, persistent=False)

        self.post_init()

    def _ensure_rope_and_mask(self):
        """
        `from_pretrained(..., low_cpu_mem_usage=True)` may build the module under
        meta tensors. In that case, our non-persistent buffers can end up as
        meta/zero tensors even though they are deterministic. Recompute them on
        demand.
        """
        need_rope = (
            self.cos_cache.device.type == "meta"
            or self.sin_cache.device.type == "meta"
            or self.cos_cache.numel() == 0
            or self.sin_cache.numel() == 0
            or (self.cos_cache.numel() > 0 and float(self.cos_cache.flatten()[0]) == 0.0)
        )
        need_mask = (
            self.causal_mask.device.type == "meta"
            or self.causal_mask.numel() == 0
            # causal_mask[0, 1] should be True for an upper-triangular mask.
            or (self.causal_mask.numel() > 1 and bool(self.causal_mask[0, 1]) is False)
        )
        if not (need_rope or need_mask):
            return

        head_dim = self.config.d_model // self.config.n_heads
        cos, sin = precompute_freqs_cis(head_dim, self.config.context_len, self.config.theta_base)
        self._buffers["cos_cache"] = cos
        self._buffers["sin_cache"] = sin

        mask = torch.triu(
            torch.ones(self.config.context_len, self.config.context_len, dtype=torch.bool), diagonal=1
        )
        self._buffers["causal_mask"] = mask

    def _init_weights(self, module):
        std = self.config.initializer_range
        if isinstance(module, nn.Linear):
            nn.init.normal_(module.weight, mean=0.0, std=std)
            if module.bias is not None:
                nn.init.zeros_(module.bias)
        elif isinstance(module, nn.Embedding):
            nn.init.normal_(module.weight, mean=0.0, std=std)

    def tie_weights(self, **kwargs):
        if self.config.tie_word_embeddings:
            self.head.weight = self.embeddor.weight

    def get_input_embeddings(self):
        return self.embeddor

    def set_input_embeddings(self, value):
        self.embeddor = value
        self.tie_weights()

    def get_output_embeddings(self):
        return self.head

    def set_output_embeddings(self, new_embeddings):
        self.head = new_embeddings
        self.tie_weights()

    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        token_type_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        **kwargs
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        B, S = input_ids.shape

        self._ensure_rope_and_mask()

        # Support both HF Cache (v5+) and legacy tuple-of-layer-caches.
        if use_cache and past_key_values is None:
            past_key_values = DynamicCache()

        # Positional tracking
        start_pos = 0
        if past_key_values is not None:
            if isinstance(past_key_values, Cache):
                start_pos = past_key_values.get_seq_length()
            else:
                layer0 = past_key_values[0]
                if layer0 is not None and layer0[0] is not None:
                    start_pos = layer0[0].shape[2]

        # Embeddings
        x = self.embeddor(input_ids)

        # RoPE slicing
        cos = self.cos_cache[start_pos : start_pos + S].to(device=x.device, dtype=x.dtype).unsqueeze(0).unsqueeze(0)
        sin = self.sin_cache[start_pos : start_pos + S].to(device=x.device, dtype=x.dtype).unsqueeze(0).unsqueeze(0)
        
        # Masking
        mask = None
        if S > 1:
            mask = self.causal_mask[start_pos : start_pos + S, : start_pos + S].to(device=x.device)

        next_past_key_values = [] if (use_cache and not isinstance(past_key_values, Cache)) else None

        # Blocks
        for i, block in enumerate(self.blocks):
            layer_past = None
            if past_key_values is not None:
                if isinstance(past_key_values, Cache):
                    layer_past = past_key_values
                else:
                    layer_past = past_key_values[i]
            x, new_layer_past = block(
                x, cos, sin, mask=mask, past_key_value=layer_past, use_cache=use_cache
            )
            if next_past_key_values is not None:
                next_past_key_values.append(new_layer_past)

        # Final head
        logits = self.head(self.norm(x))
        if not self.training:
            # Stability clip
            logits = torch.nan_to_num(logits, nan=0.0, posinf=1e4, neginf=-1e4)

        loss = None
        if labels is not None:
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            loss = F.cross_entropy(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))

        if not return_dict:
            return (logits, past_key_values) if use_cache else (logits,)

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=past_key_values if isinstance(past_key_values, Cache) else (tuple(next_past_key_values) if use_cache else None)
        )

    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, **kwargs):
        past_len = 0
        if past_key_values is not None:
            if isinstance(past_key_values, Cache):
                past_len = past_key_values.get_seq_length()
            else:
                layer0 = past_key_values[0] if len(past_key_values) > 0 else None
                if layer0 is not None and layer0[0] is not None:
                    past_len = layer0[0].shape[2]

        # Only slice for incremental decoding once we truly have cached history.
        if past_len > 0:
            input_ids = input_ids[:, -1:]
        return {
            "input_ids": input_ids,
            "past_key_values": past_key_values,
            "attention_mask": kwargs.get("attention_mask", None),
            "token_type_ids": kwargs.get("token_type_ids", None),
            "use_cache": True,
        }

    def _reorder_cache(self, past_key_values, beam_idx):
        if past_key_values is None:
            return past_key_values
        if isinstance(past_key_values, Cache):
            past_key_values.reorder_cache(beam_idx)
            return past_key_values
        return tuple(
            (k.index_select(0, beam_idx), v.index_select(0, beam_idx))
            for (k, v) in past_key_values
        )

    @torch.no_grad()
    def generate(self, input_ids=None, max_new_tokens=256, temperature=1.0,
                 top_k=None, do_sample=True, eos_token_id=None, **kwargs):
        """Custom autoregressive generate that bypasses GenerationMixin internals."""
        self._ensure_rope_and_mask()
        kv_cache = None
        curr_ids = input_ids

        for _ in range(max_new_tokens):
            if curr_ids.size(1) > self.config.context_len:
                curr_ids = curr_ids[:, -self.config.context_len:]

            model_input = curr_ids if kv_cache is None else curr_ids[:, -1:]
            out = self.forward(model_input, past_key_values=kv_cache, use_cache=True, return_dict=True)
            kv_cache = out.past_key_values

            logits = out.logits[:, -1, :]
            if do_sample:
                logits = logits / max(temperature, 1e-8)
                if top_k is not None:
                    v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
                    logits[logits < v[:, [-1]]] = -float("Inf")
                probs = F.softmax(logits, dim=-1)
                next_token = torch.multinomial(probs, num_samples=1)
            else:
                next_token = logits.argmax(dim=-1, keepdim=True)

            curr_ids = torch.cat([curr_ids, next_token], dim=1)
            if eos_token_id is not None and (next_token == eos_token_id).all():
                break

        return curr_ids