File size: 16,238 Bytes
03d9aff
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""UTR-LM ported to Hugging Face PreTrainedModel."""

import math
from typing import Optional, Tuple, Union

import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import PreTrainedModel
from transformers.modeling_outputs import BaseModelOutput, MaskedLMOutput

from .configuration_utrlm import UtrLmConfig


# ---------------------------------------------------------------------------
# Rotary embeddings
# ---------------------------------------------------------------------------

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


def _apply_rotary_pos_emb(x, cos, sin):
    cos = cos[:, : x.shape[-2], :].to(x.dtype)
    sin = sin[:, : x.shape[-2], :].to(x.dtype)
    return (x * cos) + (_rotate_half(x) * sin)


class RotaryEmbedding(nn.Module):
    def __init__(self, dim: int):
        super().__init__()
        inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2).float() / dim))
        self.register_buffer("inv_freq", inv_freq)
        self._seq_len_cached: Optional[int] = None
        self._cos_cached: Optional[torch.Tensor] = None
        self._sin_cached: Optional[torch.Tensor] = None

    def _update_cos_sin_tables(self, x: torch.Tensor, seq_dimension: int = 1):
        seq_len = x.shape[seq_dimension]
        if seq_len != self._seq_len_cached or self._cos_cached.device != x.device:
            self._seq_len_cached = seq_len
            t = torch.arange(x.shape[seq_dimension], device=x.device).type_as(self.inv_freq)
            freqs = torch.einsum("i,j->ij", t, self.inv_freq)
            emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
            self._cos_cached = emb.cos()[None, :, :]
            self._sin_cached = emb.sin()[None, :, :]
        return self._cos_cached, self._sin_cached

    def forward(self, q, k):
        self._cos_cached, self._sin_cached = self._update_cos_sin_tables(k, seq_dimension=-2)
        return (
            _apply_rotary_pos_emb(q, self._cos_cached, self._sin_cached),
            _apply_rotary_pos_emb(k, self._cos_cached, self._sin_cached),
        )


# ---------------------------------------------------------------------------
# Attention variants
# ---------------------------------------------------------------------------

class UtrLmAttention(nn.Module):
    """Eager (standard) attention."""

    def __init__(self, embed_dim: int, num_heads: int):
        super().__init__()
        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.rot_emb = RotaryEmbedding(dim=self.head_dim)

    def _project(self, x):
        """Project and reshape x (T, B, E) -> q/k/v in (B*H, T, head_dim)."""
        tgt_len, bsz, _ = x.size()
        q = (self.q_proj(x) * self.scaling).contiguous().view(tgt_len, bsz * self.num_heads, self.head_dim).transpose(0, 1)
        k = self.k_proj(x).contiguous().view(tgt_len, bsz * self.num_heads, self.head_dim).transpose(0, 1)
        v = self.v_proj(x).contiguous().view(tgt_len, bsz * self.num_heads, self.head_dim).transpose(0, 1)
        q, k = self.rot_emb(q, k)
        return q, k, v

    def forward(self, x, key_padding_mask, output_attentions: bool = False):
        tgt_len, bsz, _ = x.size()
        q, k, v = self._project(x)

        attn_weights = torch.bmm(q, k.transpose(1, 2))
        attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, tgt_len)
        if key_padding_mask is not None:
            attn_weights = attn_weights.masked_fill(
                key_padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool), float("-inf")
            )
        attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, tgt_len)

        attn_probs = F.softmax(attn_weights, dim=-1, dtype=torch.float32).type_as(attn_weights)
        attn = torch.bmm(attn_probs, v)
        attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, self.embed_dim)
        out = self.out_proj(attn)

        if output_attentions:
            return out, attn_probs.view(bsz, self.num_heads, tgt_len, tgt_len)
        return out, None


class UtrLmSdpaAttention(UtrLmAttention):
    """SDPA attention via torch.nn.functional.scaled_dot_product_attention."""

    def forward(self, x, key_padding_mask, output_attentions: bool = False):
        if output_attentions:
            # SDPA doesn't expose attention weights; fall back to eager.
            return super().forward(x, key_padding_mask, output_attentions=True)

        tgt_len, bsz, _ = x.size()
        q, k, v = self._project(x)  # (B*H, T, head_dim)

        # Reshape to (B, H, T, head_dim) for SDPA
        q = q.view(bsz, self.num_heads, tgt_len, self.head_dim)
        k = k.view(bsz, self.num_heads, tgt_len, self.head_dim)
        v = v.view(bsz, self.num_heads, tgt_len, self.head_dim)

        # Convert bool padding mask -> additive float mask (B, 1, 1, T)
        attn_mask = None
        if key_padding_mask is not None:
            attn_mask = torch.zeros(bsz, 1, 1, tgt_len, dtype=q.dtype, device=q.device)
            attn_mask = attn_mask.masked_fill(key_padding_mask[:, None, None, :], float("-inf"))

        # scale=1.0 because q is already pre-scaled by self.scaling
        out = F.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask, scale=1.0)
        out = out.permute(2, 0, 1, 3).contiguous().view(tgt_len, bsz, self.embed_dim)
        return self.out_proj(out), None


class UtrLmFlashAttention2(UtrLmAttention):
    """Flash Attention 2 via flash_attn (must be installed separately)."""

    def forward(self, x, key_padding_mask, output_attentions: bool = False):
        if output_attentions:
            # Flash attention doesn't expose attention weights; fall back to eager.
            return super().forward(x, key_padding_mask, output_attentions=True)

        try:
            from flash_attn import flash_attn_func
            from flash_attn.bert_padding import pad_input, unpad_input
        except ImportError as e:
            raise ImportError("flash_attn is required for attn_implementation='flash_attention_2'. "
                              "Install with: pip install flash-attn --no-build-isolation") from e

        tgt_len, bsz, _ = x.size()
        q, k, v = self._project(x)  # (B*H, T, head_dim)

        # Reshape to (B, T, H, head_dim) - flash_attn's expected layout
        q = q.view(bsz, self.num_heads, tgt_len, self.head_dim).permute(0, 2, 1, 3)
        k = k.view(bsz, self.num_heads, tgt_len, self.head_dim).permute(0, 2, 1, 3)
        v = v.view(bsz, self.num_heads, tgt_len, self.head_dim).permute(0, 2, 1, 3)

        # Flash attention requires fp16 or bf16
        orig_dtype = q.dtype
        if orig_dtype not in (torch.float16, torch.bfloat16):
            q, k, v = q.to(torch.bfloat16), k.to(torch.bfloat16), v.to(torch.bfloat16)

        if key_padding_mask is not None:
            # Unpad, run varlen flash attention, repad
            from flash_attn import flash_attn_varlen_func
            attention_mask = ~key_padding_mask  # True = valid token
            q_unpad, indices, cu_seqlens, max_seqlen, _ = unpad_input(q, attention_mask)
            k_unpad, _, _, _, _ = unpad_input(k, attention_mask)
            v_unpad, _, _, _, _ = unpad_input(v, attention_mask)

            out_unpad = flash_attn_varlen_func(
                q_unpad, k_unpad, v_unpad,
                cu_seqlens_q=cu_seqlens,
                cu_seqlens_k=cu_seqlens,
                max_seqlen_q=max_seqlen,
                max_seqlen_k=max_seqlen,
                softmax_scale=1.0,  # q already pre-scaled
                causal=False,
            )
            out = pad_input(out_unpad, indices, bsz, tgt_len)
        else:
            out = flash_attn_func(q, k, v, softmax_scale=1.0, causal=False)

        out = out.to(orig_dtype).permute(1, 0, 2, 3).contiguous().view(tgt_len, bsz, self.embed_dim)
        return self.out_proj(out), None


UTRLM_ATTENTION_CLASSES = {
    "eager": UtrLmAttention,
    "sdpa": UtrLmSdpaAttention,
    "flash_attention_2": UtrLmFlashAttention2,
}


# ---------------------------------------------------------------------------
# Transformer layer (pre-LN)
# ---------------------------------------------------------------------------

def _gelu(x):
    return x * 0.5 * (1.0 + torch.erf(x / math.sqrt(2.0)))


class UtrLmLayer(nn.Module):
    def __init__(self, embed_dim: int, attention_heads: int, config: UtrLmConfig):
        super().__init__()
        attn_cls = UTRLM_ATTENTION_CLASSES[getattr(config, "_attn_implementation", "eager")]
        self.self_attn = attn_cls(embed_dim, attention_heads)
        self.self_attn_layer_norm = nn.LayerNorm(embed_dim)
        self.fc1 = nn.Linear(embed_dim, 4 * embed_dim)
        self.fc2 = nn.Linear(4 * embed_dim, embed_dim)
        self.final_layer_norm = nn.LayerNorm(embed_dim)

    def forward(self, x, padding_mask, output_attentions: bool = False):
        residual = x
        x = self.self_attn_layer_norm(x)
        x, attn_weights = self.self_attn(x, key_padding_mask=padding_mask, output_attentions=output_attentions)
        x = residual + x

        residual = x
        x = self.final_layer_norm(x)
        x = _gelu(self.fc1(x))
        x = self.fc2(x)
        return residual + x, attn_weights


# ---------------------------------------------------------------------------
# Backbone
# ---------------------------------------------------------------------------

class UtrLmModel(PreTrainedModel):
    """
    UTR-LM encoder backbone. Returns last_hidden_state (B, T, E).
    The [CLS] token sits at position 0 (prepend_bos=True by default).
    """

    config_class = UtrLmConfig
    base_model_prefix = "utrlm"
    _supports_sdpa = True
    _supports_flash_attn_2 = True

    def __init__(self, config: UtrLmConfig):
        super().__init__(config)
        self.embed_scale = 1
        self.embed_tokens = nn.Embedding(
            config.alphabet_size, config.embed_dim, padding_idx=config.padding_idx
        )
        self.layers = nn.ModuleList(
            [UtrLmLayer(config.embed_dim, config.attention_heads, config) for _ in range(config.num_layers)]
        )
        self.emb_layer_norm_after = nn.LayerNorm(config.embed_dim)
        self.post_init()

    def get_input_embeddings(self):
        return self.embed_tokens

    def set_input_embeddings(self, value):
        self.embed_tokens = value

    def forward(
        self,
        input_ids: torch.LongTensor,
        attention_mask: Optional[torch.BoolTensor] = None,
        output_hidden_states: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, BaseModelOutput]:
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None
            else self.config.output_hidden_states
        )
        output_attentions = (
            output_attentions if output_attentions is not None else self.config.output_attentions
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        cfg = self.config
        # HF convention: attention_mask is 1=attend, 0=pad.
        # Convert to bool padding_mask (True = ignore) or derive from input_ids.
        if attention_mask is not None:
            padding_mask = attention_mask.eq(0)
        else:
            padding_mask = input_ids.eq(cfg.padding_idx)

        x = self.embed_scale * self.embed_tokens(input_ids)

        if cfg.token_dropout:
            x.masked_fill_((input_ids == cfg.mask_idx).unsqueeze(-1), 0.0)
            mask_ratio_train = 0.15 * 0.8
            src_lengths = (~padding_mask).sum(-1)
            mask_ratio_observed = (input_ids == cfg.mask_idx).sum(-1).to(x.dtype) / src_lengths.to(x.dtype)
            x = x * (1 - mask_ratio_train) / (1 - mask_ratio_observed)[:, None, None]

        if padding_mask is not None:
            x = x * (1 - padding_mask.unsqueeze(-1).type_as(x))

        all_hidden_states = () if output_hidden_states else None
        all_attentions = () if output_attentions else None
        if output_hidden_states:
            all_hidden_states += (x,)

        x = x.transpose(0, 1)  # (B, T, E) -> (T, B, E)
        effective_padding = padding_mask if padding_mask.any() else None

        for layer in self.layers:
            x, attn_weights = layer(x, padding_mask=effective_padding, output_attentions=output_attentions)
            if output_hidden_states:
                all_hidden_states += (x.transpose(0, 1),)
            if output_attentions:
                all_attentions += (attn_weights,)

        x = self.emb_layer_norm_after(x)
        x = x.transpose(0, 1)  # (T, B, E) -> (B, T, E)

        if output_hidden_states:
            all_hidden_states = all_hidden_states[:-1] + (x,)

        if not return_dict:
            return tuple(v for v in [x, all_hidden_states, all_attentions] if v is not None)

        return BaseModelOutput(
            last_hidden_state=x,
            hidden_states=all_hidden_states,
            attentions=all_attentions,
        )


# ---------------------------------------------------------------------------
# MLM head
# ---------------------------------------------------------------------------

class UtrLmForMaskedLM(PreTrainedModel):
    """
    UTR-LM with a masked-language-modelling head.
    Returns MaskedLMOutput with logits (B, T, vocab_size).
    """

    config_class = UtrLmConfig
    base_model_prefix = "utrlm"
    _supports_sdpa = True
    _supports_flash_attn_2 = True

    def __init__(self, config: UtrLmConfig):
        super().__init__(config)
        self.utrlm = UtrLmModel(config)

        embed_dim = config.embed_dim
        vocab_size = config.alphabet_size
        self.lm_head = nn.ModuleDict({
            "dense": nn.Linear(embed_dim, embed_dim),
            "layer_norm": nn.LayerNorm(embed_dim),
        })
        self.lm_head_bias = nn.Parameter(torch.zeros(vocab_size))

        self.post_init()

    def get_input_embeddings(self):
        return self.utrlm.embed_tokens

    def set_input_embeddings(self, value):
        self.utrlm.embed_tokens = value

    def get_output_embeddings(self):
        return self.utrlm.embed_tokens

    def set_output_embeddings(self, new_embeddings):
        self.utrlm.embed_tokens = new_embeddings

    def _lm_head_forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.lm_head["dense"](x)
        x = _gelu(x)
        x = self.lm_head["layer_norm"](x)
        return F.linear(x, self.utrlm.embed_tokens.weight) + self.lm_head_bias

    def forward(
        self,
        input_ids: torch.LongTensor,
        attention_mask: Optional[torch.BoolTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        output_hidden_states: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, MaskedLMOutput]:
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = self.utrlm(
            input_ids,
            attention_mask=attention_mask,
            output_hidden_states=output_hidden_states,
            output_attentions=output_attentions,
            return_dict=True,
        )
        logits = self._lm_head_forward(outputs.last_hidden_state)

        loss = None
        if labels is not None:
            loss = F.cross_entropy(
                logits.view(-1, self.config.alphabet_size),
                labels.view(-1),
                ignore_index=self.config.padding_idx,
            )

        if not return_dict:
            output = (logits,) + outputs[1:]
            return (loss,) + output if loss is not None else output

        return MaskedLMOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )