File size: 28,453 Bytes
c525dff
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
from typing import Callable, List, Optional, Tuple, Union

import numpy as np

import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import CrossEntropyLoss

from transformers.activations import ACT2FN
from transformers.generation.configuration_utils import GenerationConfig
from transformers.generation.logits_process import LogitsProcessorList
from transformers.generation.stopping_criteria import StoppingCriteriaList
from transformers.generation.streamers import BaseStreamer
from transformers.generation.utils import GenerateOutput
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
from transformers.modeling_utils import PreTrainedModel
from transformers.utils import logging
from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
from .configuration_progen import ProGenConfig

logger = logging.get_logger(__name__)

from .structure import StructureTransformer


def fixed_pos_embedding(x, seq_dim=1, seq_len=None):
    dim = x.shape[-1]
    if seq_len is None:
        seq_len = x.shape[seq_dim]
    inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2) / dim))
    sinusoid_inp = torch.einsum("i , j -> i j", torch.arange(seq_len), inv_freq).to(x.device).float()
    return torch.sin(sinusoid_inp), torch.cos(sinusoid_inp)


def rotate_every_two(x):
    x1 = x[:, :, :, ::2]
    x2 = x[:, :, :, 1::2]
    x = torch.stack((-x2, x1), axis=-1)
    return x.flatten(-2)  # in einsum notation: rearrange(x, '... d j -> ... (d j)')


def apply_rotary_pos_emb(x, sincos, offset=0):
    sin, cos = map(lambda t: t[None, offset : x.shape[1] + offset, None, :].repeat_interleave(2, 3), sincos)
    # einsum notation for lambda t: repeat(t[offset:x.shape[1]+offset,:], "n d -> () n () (d j)", j=2)
    return (x * cos) + (rotate_every_two(x) * sin)


class ProGenAttention(nn.Module):
    def __init__(self, config):
        super().__init__()

        max_positions = config.max_position_embeddings
        self.register_buffer(
            "bias",
            torch.tril(torch.ones((max_positions, max_positions), dtype=torch.bool)).view(
                1, 1, max_positions, max_positions
            ),
        )
        self.register_buffer("masked_bias", torch.tensor(-1e9))

        self.attn_dropout = nn.Dropout(config.attn_pdrop)
        self.resid_dropout = nn.Dropout(config.resid_pdrop)

        self.embed_dim = config.hidden_size
        self.num_attention_heads = config.num_attention_heads
        self.head_dim = self.embed_dim // self.num_attention_heads
        if self.head_dim * self.num_attention_heads != self.embed_dim:
            raise ValueError(
                f"embed_dim must be divisible by num_attention_heads (got `embed_dim`: {self.embed_dim} and `num_attention_heads`: {self.num_attention_heads})."
            )
        self.scale_attn = torch.sqrt(torch.tensor(self.head_dim, dtype=torch.float32)).to(torch.get_default_dtype())
        self.qkv_proj = nn.Linear(self.embed_dim, self.embed_dim * 3, bias=False)

        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)
        self.rotary_dim = None
        if config.rotary_dim is not None:
            self.rotary_dim = config.rotary_dim

    def _split_heads(self, x, n_head, dim_head, mp_num):
        reshaped = x.reshape(x.shape[:-1] + (n_head//mp_num, dim_head))
        reshaped = reshaped.reshape(x.shape[:-2] + (-1, ) + reshaped.shape[-1:])
        return reshaped

    def _merge_heads(self, tensor, num_attention_heads, attn_head_size):
        """
        Merges attn_head_size dim and num_attn_heads dim into n_ctx
        """
        if len(tensor.shape) == 5:
            tensor = tensor.permute(0, 1, 3, 2, 4).contiguous()
        elif len(tensor.shape) == 4:
            tensor = tensor.permute(0, 2, 1, 3).contiguous()
        else:
            raise ValueError(f"Input tensor rank should be one of [4, 5], but is: {len(tensor.shape)}")
        new_shape = tensor.size()[:-2] + (num_attention_heads * attn_head_size,)
        return tensor.view(new_shape)

    def _attn(
        self,
        query,
        key,
        value,
        attention_mask=None,
        head_mask=None,
    ):

        # compute causal mask from causal mask buffer
        query_length, key_length = query.size(-2), key.size(-2)
        causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length]

        # Keep the attention weights computation in fp32 to avoid overflow issues
        query = query.to(torch.float32)
        key = key.to(torch.float32)

        attn_weights = torch.matmul(query, key.transpose(-1, -2))

        attn_weights = attn_weights / self.scale_attn
        attn_weights = torch.where(causal_mask, attn_weights, self.masked_bias.to(attn_weights.dtype))

        if attention_mask is not None:
            # Apply the attention mask
            attn_weights = attn_weights + attention_mask

        attn_weights = nn.Softmax(dim=-1)(attn_weights)
        attn_weights = attn_weights.to(value.dtype)
        attn_weights = self.attn_dropout(attn_weights)

        # Mask heads if we want to
        if head_mask is not None:
            attn_weights = attn_weights * head_mask

        attn_output = torch.matmul(attn_weights, value)

        return attn_output, attn_weights

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        layer_past=None,
        head_mask=None,
        use_cache=False,
        output_attentions=False,
    ):

        qkv = self.qkv_proj(hidden_states)
        # TODO(enijkamp): factor out number of logical TPU-v3/v4 cores or make forward pass agnostic
        # mp_num = 4
        mp_num = 8
        qkv_split = qkv.reshape(qkv.shape[:-1] + (mp_num, -1))

        local_dim = self.head_dim * self.num_attention_heads // mp_num
        query, value, key = torch.split(qkv_split, local_dim, dim=-1)
        query = self._split_heads(query, self.num_attention_heads, self.head_dim, mp_num=mp_num)
        key = self._split_heads(key, self.num_attention_heads, self.head_dim, mp_num=mp_num)

        value = self._split_heads(value, self.num_attention_heads, self.head_dim, mp_num=mp_num)
        value = value.permute(0, 2, 1, 3)

        seq_len = key.shape[1]
        offset = 0

        if layer_past is not None:
            offset = layer_past[0].shape[-2]
            seq_len += offset

        if self.rotary_dim is not None:
            k_rot = key[:, :, :, : self.rotary_dim]
            k_pass = key[:, :, :, self.rotary_dim :]

            q_rot = query[:, :, :, : self.rotary_dim]
            q_pass = query[:, :, :, self.rotary_dim :]

            sincos = fixed_pos_embedding(k_rot, 1, seq_len=seq_len)
            k_rot = apply_rotary_pos_emb(k_rot, sincos, offset=offset)
            q_rot = apply_rotary_pos_emb(q_rot, sincos, offset=offset)

            key = torch.cat([k_rot, k_pass], dim=-1)
            query = torch.cat([q_rot, q_pass], dim=-1)
        else:
            sincos = fixed_pos_embedding(key, 1, seq_len=seq_len)
            key = apply_rotary_pos_emb(key, sincos, offset=offset)
            query = apply_rotary_pos_emb(query, sincos, offset=offset)

        key = key.permute(0, 2, 1, 3)
        query = query.permute(0, 2, 1, 3)

        if layer_past is not None:
            past_key = layer_past[0]
            past_value = layer_past[1]
            key = torch.cat((past_key, key), dim=-2)
            value = torch.cat((past_value, value), dim=-2)

        if use_cache is True:
            present = (key, value)
        else:
            present = None

        # compute self-attention: V x Softmax(QK^T)
        attn_output, attn_weights = self._attn(query, key, value, attention_mask, head_mask)

        attn_output = self._merge_heads(attn_output, self.num_attention_heads, self.head_dim)

        attn_output = self.out_proj(attn_output)
        attn_output = self.resid_dropout(attn_output)

        outputs = (attn_output, present)
        if output_attentions:
            outputs += (attn_weights,)

        return outputs  # a, present, (attentions)


class ProGenMLP(nn.Module):
    def __init__(self, intermediate_size, config):  # in MLP: intermediate_size= 4 * embed_dim
        super().__init__()
        embed_dim = config.n_embd

        self.fc_in = nn.Linear(embed_dim, intermediate_size)
        self.fc_out = nn.Linear(intermediate_size, embed_dim)

        self.act = ACT2FN[config.activation_function]
        self.dropout = nn.Dropout(config.resid_pdrop)

    def forward(self, hidden_states):
        hidden_states = self.fc_in(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states = self.fc_out(hidden_states)
        hidden_states = self.dropout(hidden_states)
        return hidden_states


class ProGenBlock(nn.Module):
    def __init__(self, config):
        super().__init__()
        inner_dim = config.n_inner if config.n_inner is not None else 4 * config.n_embd
        self.ln_1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
        self.attn = ProGenAttention(config)
        self.mlp = ProGenMLP(inner_dim, config)

    def forward(
        self,
        hidden_states,
        layer_past=None,
        attention_mask=None,
        head_mask=None,
        use_cache=False,
        output_attentions=False,
    ):
        residual = hidden_states
        hidden_states = self.ln_1(hidden_states)
        attn_outputs = self.attn(
            hidden_states,
            layer_past=layer_past,
            attention_mask=attention_mask,
            head_mask=head_mask,
            use_cache=use_cache,
            output_attentions=output_attentions,
        )
        attn_output = attn_outputs[0]  # output_attn: a, present, (attentions)
        outputs = attn_outputs[1:]

        feed_forward_hidden_states = self.mlp(hidden_states)
        hidden_states = attn_output + feed_forward_hidden_states + residual

        if use_cache:
            outputs = (hidden_states,) + outputs
        else:
            outputs = (hidden_states,) + outputs[1:]

        return outputs  # hidden_states, present, (attentions)


class ProGenPreTrainedModel(PreTrainedModel):
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    """

    config_class = ProGenConfig
    base_model_prefix = "transformer"
    supports_gradient_checkpointing = True
    is_parallelizable = True

    def __init__(self, *inputs, **kwargs):
        super().__init__(*inputs, **kwargs)

    def _init_weights(self, module):
        """Initialize the weights."""
        if isinstance(module, (nn.Linear,)):
            # Slightly different from Mesh Transformer JAX which uses truncated_normal for initialization
            # cf https://github.com/pytorch/pytorch/pull/5617
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            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.initializer_range)
            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)

    def _set_gradient_checkpointing(self, module, value=False):
        if isinstance(module, ProGenModel):
            module.gradient_checkpointing = value

class ProGenModel(ProGenPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.embed_dim = config.n_embd
        self.vocab_size = config.vocab_size
        self.wte = nn.Embedding(config.vocab_size, self.embed_dim)
        self.drop = nn.Dropout(config.embd_pdrop)
        self.h = nn.ModuleList([ProGenBlock(config) for _ in range(config.n_layer)])
        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
        self.rotary_dim = min(config.rotary_dim, config.n_ctx // config.num_attention_heads)

        self.gradient_checkpointing = False
        self.structure = StructureTransformer(**config.structure)
        
        self.init_weights()

        # Model parallel
        self.model_parallel = False
        self.device_map = None


    def parallelize(self, device_map=None):
        # Check validity of device_map
        self.device_map = (
            get_device_map(len(self.h), range(torch.cuda.device_count())) if device_map is None else device_map
        )
        assert_device_map(self.device_map, len(self.h))
        self.model_parallel = True
        self.first_device = "cpu" if "cpu" in self.device_map.keys() else "cuda:" + str(min(self.device_map.keys()))
        self.last_device = "cuda:" + str(max(self.device_map.keys()))
        self.wte = self.wte.to(self.first_device)
        # Load onto devices
        for k, v in self.device_map.items():
            for block in v:
                cuda_device = "cuda:" + str(k)
                self.h[block] = self.h[block].to(cuda_device)
        # ln_f to last
        self.ln_f = self.ln_f.to(self.last_device)


    def deparallelize(self):
        self.model_parallel = False
        self.device_map = None
        self.first_device = "cpu"
        self.last_device = "cpu"
        self.wte = self.wte.to("cpu")
        for index in range(len(self.h)):
            self.h[index] = self.h[index].to("cpu")
        self.ln_f = self.ln_f.to("cpu")
        torch.cuda.empty_cache()

    def get_input_embeddings(self):
        return self.wte

    def set_input_embeddings(self, new_embeddings):
        self.wte = new_embeddings

    def forward(
        self,
        input_ids=None,
        past_key_values=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        query_embeds=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        if past_key_values is None:
            # structure encode will check if input_ids contains valid 
            structure_embs = self.structure.encode(input_ids)
            if structure_embs is not None:
                input_ids = input_ids[:, self.structure.n_queries:]
        else:
            structure_embs = None
        
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
        elif input_ids is not None:
            input_shape = input_ids.size()
            input_ids = input_ids.view(-1, input_shape[-1])
            batch_size = input_ids.shape[0]
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
            batch_size = inputs_embeds.shape[0]
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        device = input_ids.device if input_ids is not None else inputs_embeds.device

        # if token_type_ids is not None:
        #     token_type_ids = token_type_ids.view(-1, input_shape[-1])

        if position_ids is not None:
            position_ids = position_ids.view(-1, input_shape[-1])

        if past_key_values is None:
            past_length = 0
            past_key_values = tuple([None] * len(self.h))
        else:
            past_length = past_key_values[0][0].size(-2)

        if position_ids is None:
            position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
            position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])

        # Attention mask.
        if attention_mask is not None:
            assert batch_size > 0, "batch_size has to be defined and > 0"
            attention_mask = attention_mask.view(batch_size, -1)
            # We create a 3D attention mask from a 2D tensor mask.
            # Sizes are [batch_size, 1, 1, to_seq_length]
            # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
            # this attention mask is more simple than the triangular masking of causal attention
            # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
            attention_mask = attention_mask[:, None, None, :]

            # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
            # masked positions, this operation will create a tensor which is 0.0 for
            # positions we want to attend and -10000.0 for masked positions.
            # Since we are adding it to the raw scores before the softmax, this is
            # effectively the same as removing these entirely.
            attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility
            attention_mask = (1.0 - attention_mask) * -10000.0

        # Prepare head mask if needed
        # 1.0 in head_mask indicate we keep the head
        # attention_probs has shape bsz x num_attention_heads x N x N
        # head_mask has shape n_layer x batch x num_attention_heads x N x N
        head_mask = self.get_head_mask(head_mask, self.config.n_layer)

        if inputs_embeds is None:
            inputs_embeds = self.wte(input_ids)
        
        if query_embeds is not None:
            inputs_embeds = torch.cat([query_embeds, inputs_embeds], dim=1)
            input_shape = inputs_embeds.size()[:-1]
        
        if structure_embs is not None:
            inputs_embeds = torch.cat([structure_embs, inputs_embeds], dim=1)
            input_shape = inputs_embeds.size()[:-1]

        hidden_states = inputs_embeds

        # disable token_type_ids
        # if token_type_ids is not None:
        #     token_type_embeds = self.wte(token_type_ids)
        #     hidden_states = hidden_states + token_type_embeds

        hidden_states = self.drop(hidden_states)

        output_shape = input_shape + (hidden_states.size(-1),)

        presents = () if use_cache else None
        all_self_attentions = () if output_attentions else None
        all_hidden_states = () if output_hidden_states else None
        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):

            # Model parallel
            if self.model_parallel:
                torch.cuda.set_device(hidden_states.device)
                # Ensure layer_past is on same device as hidden_states (might not be correct)
                if layer_past is not None:
                    layer_past = tuple(past_state.to(hidden_states.device) for past_state in layer_past)
                # Ensure that attention_mask is always on the same device as hidden_states
                if attention_mask is not None:
                    attention_mask = attention_mask.to(hidden_states.device)
                if isinstance(head_mask, torch.Tensor):
                    head_mask = head_mask.to(hidden_states.device)
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            if self.gradient_checkpointing and self.training:

                if use_cache:
                    # logger.warning(
                    #     "`use_cache=True` is incompatible with `config.gradient_checkpointing=True`. Setting "
                    #     "`use_cache=False`..."
                    # )
                    use_cache = False

                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        # None for past_key_value
                        return module(*inputs, use_cache, output_attentions)

                    return custom_forward

                outputs = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(block),
                    hidden_states,
                    None,
                    attention_mask,
                    head_mask[i],
                )
            else:
                outputs = block(
                    hidden_states,
                    layer_past=layer_past,
                    attention_mask=attention_mask,
                    head_mask=head_mask[i],
                    use_cache=use_cache,
                    output_attentions=output_attentions,
                )

            hidden_states = outputs[0]
            if use_cache is True:
                presents = presents + (outputs[1],)

            if output_attentions:
                all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)

            # Model Parallel: If it's the last layer for that device, put things on the next device
            if self.model_parallel:
                for k, v in self.device_map.items():
                    if i == v[-1] and "cuda:" + str(k) != self.last_device:
                        hidden_states = hidden_states.to("cuda:" + str(k + 1))

        hidden_states = self.ln_f(hidden_states)

        hidden_states = hidden_states.view(*output_shape)
        # Add last hidden state
        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        if not return_dict:
            return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)

        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=presents,
            hidden_states=all_hidden_states,
            attentions=all_self_attentions,
        )


class ProGenForCausalLM(ProGenPreTrainedModel):
    _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.masked_bias", r"h\.\d+\.attn\.bias", r"lm_head\.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.transformer = ProGenModel(config)
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size)
        self.init_weights()

        # Model parallel
        self.model_parallel = False
        self.device_map = None

    def parallelize(self, device_map=None):
        self.device_map = (
            get_device_map(len(self.transformer.h), range(torch.cuda.device_count()))
            if device_map is None
            else device_map
        )
        assert_device_map(self.device_map, len(self.transformer.h))
        self.transformer.parallelize(self.device_map)
        self.lm_head = self.lm_head.to(self.transformer.first_device)
        self.model_parallel = True

    def deparallelize(self):
        self.transformer.deparallelize()
        self.transformer = self.transformer.to("cpu")
        self.lm_head = self.lm_head.to("cpu")
        self.model_parallel = False
        torch.cuda.empty_cache()

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def prepare_inputs_for_generation(
        self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
    ):
        if past_key_values:
            input_ids = input_ids[:, -1:]

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and past_key_values is None:
            model_inputs = {"inputs_embeds": inputs_embeds}
        else:
            model_inputs = {"input_ids": input_ids}

        model_inputs.update(
            {
                "past_key_values": past_key_values,
                "use_cache": kwargs.get("use_cache"),
                "attention_mask": attention_mask,
            }
        )
        return model_inputs

    def forward(
        self,
        input_ids=None,
        past_key_values=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        labels=None,
        use_cache=None,
        query_embeds = None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        r"""
        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
            ``labels = input_ids`` Indices are selected in ``[-100, 0, ..., config.vocab_size]`` All labels set to
            ``-100`` are ignored (masked), the loss is only computed for labels in ``[0, ..., config.vocab_size]``
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        transformer_outputs = self.transformer(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            query_embeds=query_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        hidden_states = transformer_outputs[0]

        # Set device for model parallelism
        if self.model_parallel:
            torch.cuda.set_device(self.transformer.first_device)
            hidden_states = hidden_states.to(self.lm_head.weight.device)

        # make sure sampling in fp16 works correctly and
        # compute loss in fp32 to match with mesh-tf version
        # https://github.com/EleutherAI/gpt-neo/blob/89ce74164da2fb16179106f54e2269b5da8db333/models/gpt2/gpt2.py#L179
        lm_logits = self.lm_head(hidden_states).to(torch.float32)

        loss = None
        if labels is not None:
            # Shift so that tokens < n predict n
            shift_logits = lm_logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))

            loss = loss.to(hidden_states.dtype)

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

        return CausalLMOutputWithPast(
            loss=loss,
            logits=lm_logits,
            past_key_values=transformer_outputs.past_key_values,
            hidden_states=transformer_outputs.hidden_states,
            attentions=transformer_outputs.attentions,
        )

    @staticmethod
    def _reorder_cache(past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor) -> Tuple[Tuple[torch.Tensor]]:
        """
        This function is used to re-order the :obj:`past_key_values` cache if
        :meth:`~transformers.PretrainedModel.beam_search` or :meth:`~transformers.PretrainedModel.beam_sample` is
        called. This is required to match :obj:`past_key_values` with the correct beam_idx at every generation step.
        """
        return tuple(
            tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past)
            for layer_past in past
        )

    # def generate(self, inputs: Tensor | None = None, generation_config: GenerationConfig | None = None, logits_processor: LogitsProcessorList | None = None, stopping_criteria: StoppingCriteriaList | None = None, prefix_allowed_tokens_fn: Callable[[int, Tensor], List[int]] | None = None, synced_gpus: bool | None = None, assistant_model: PreTrainedModel | None = None, streamer: BaseStreamer | None = None, negative_prompt_ids: Tensor | None = None, negative_prompt_attention_mask: Tensor | None = None, **kwargs) -> GenerateOutput | LongTensor:
    #     return super().generate(inputs, generation_config, logits_processor, stopping_criteria, prefix_allowed_tokens_fn, synced_gpus, assistant_model, streamer, negative_prompt_ids, negative_prompt_attention_mask, **kwargs)