apertus.py 22.2 KB
Newer Older
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
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

# Adapted from
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
# Copyright 2025 The Swiss AI Initiative.
# Copyright 2023 The vLLM team.
# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate the architectural differences made by
# the Swiss AI Initiative that trained the model.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Inference-only Apertus model compatible with HuggingFace weights."""
27

28
from collections.abc import Iterable
29
from itertools import islice
30
from typing import Any
31
32
33
34
35
36
37
38
39
40
41
42

import torch
from torch import nn
from transformers import ApertusConfig

from vllm.attention import Attention, AttentionType
from vllm.attention.layers.encoder_only_attention import EncoderOnlyAttention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import XIELU
from vllm.model_executor.layers.layernorm import RMSNorm
43
44
45
46
47
from vllm.model_executor.layers.linear import (
    ColumnParallelLinear,
    QKVParallelLinear,
    RowParallelLinear,
)
48
49
50
51
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
52
53
54
55
    DEFAULT_VOCAB_PADDING_SIZE,
    ParallelLMHead,
    VocabParallelEmbedding,
)
56
from vllm.model_executor.model_loader.weight_utils import (
57
58
59
    default_weight_loader,
    maybe_remap_kv_scale_name,
)
60
61
62
from vllm.sequence import IntermediateTensors

from .interfaces import SupportsLoRA, SupportsPP
63
64
65
66
67
68
69
70
71
from .utils import (
    AutoWeightsLoader,
    PPMissingLayer,
    extract_layer_index,
    is_pp_missing_parameter,
    make_empty_intermediate_tensors_factory,
    make_layers,
    maybe_prefix,
)
72
73
74
75
76
77
78
79


class ApertusMLP(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
80
        quant_config: QuantizationConfig | None = None,
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
        bias: bool = False,
        prefix: str = "",
        reduce_results: bool = True,
    ) -> None:
        super().__init__()
        self.up_proj = ColumnParallelLinear(
            input_size=hidden_size,
            output_size=intermediate_size,
            bias=bias,
            quant_config=quant_config,
            prefix=f"{prefix}.up_proj",
        )
        self.down_proj = RowParallelLinear(
            input_size=intermediate_size,
            output_size=hidden_size,
            bias=bias,
            quant_config=quant_config,
            reduce_results=reduce_results,
            prefix=f"{prefix}.down_proj",
        )
        if hidden_act != "xielu":
102
103
104
105
            raise ValueError(
                f"Unsupported activation: {hidden_act}. "
                "Only xIELU is supported for now."
            )
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
        self.act_fn = XIELU()

    def forward(self, x):
        x, _ = self.up_proj(x)
        x = self.act_fn(x)
        x, _ = self.down_proj(x)
        return x


class ApertusAttention(nn.Module):
    def __init__(
        self,
        config: ApertusConfig,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        rope_theta: float = 10000,
123
        rope_scaling: dict[str, Any] | None = None,
124
        max_position_embeddings: int = 8192,
125
        quant_config: QuantizationConfig | None = None,
126
127
        bias: bool = False,
        bias_o_proj: bool = False,
128
        cache_config: CacheConfig | None = None,
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
        prefix: str = "",
        attn_type: str = AttentionType.DECODER,
    ) -> None:
        super().__init__()
        layer_idx = extract_layer_index(prefix)
        self.hidden_size = hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = num_heads
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
        self.total_num_kv_heads = num_kv_heads
        if self.total_num_kv_heads >= tp_size:
            # Number of KV heads is greater than TP size, so we partition
            # the KV heads across multiple tensor parallel GPUs.
            assert self.total_num_kv_heads % tp_size == 0
        else:
            # Number of KV heads is less than TP size, so we replicate
            # the KV heads across multiple tensor parallel GPUs.
            assert tp_size % self.total_num_kv_heads == 0
        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
        # MistralConfig has an optional head_dim introduced by Mistral-Nemo
        head_dim = getattr(config, "head_dim", None)
        if head_dim is None:
            head_dim = self.hidden_size // self.total_num_heads
        self.head_dim = head_dim
        # Phi models introduced a partial_rotary_factor parameter in the config
155
        self.partial_rotary_factor = getattr(config, "partial_rotary_factor", 1)
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim
        self.scaling = self.head_dim**-0.5
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings

        self.qkv_proj = QKVParallelLinear(
            hidden_size=hidden_size,
            head_size=self.head_dim,
            total_num_heads=self.total_num_heads,
            total_num_kv_heads=self.total_num_kv_heads,
            bias=bias,
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
        )

        self.o_proj = RowParallelLinear(
            input_size=self.total_num_heads * self.head_dim,
            output_size=hidden_size,
            bias=bias_o_proj,
            quant_config=quant_config,
            prefix=f"{prefix}.o_proj",
        )

180
181
182
        self._init_rotary_emb(
            config, rope_scaling=rope_scaling, quant_config=quant_config
        )
183
184
185
186
187
188
189

        sliding_window = None
        if layer_types := getattr(config, "layer_types", None):
            is_sliding = layer_types[layer_idx] == "sliding_attention"
            if is_sliding:
                sliding_window = config.sliding_window

190
191
192
193
194
        attn_cls = (
            EncoderOnlyAttention
            if attn_type == AttentionType.ENCODER_ONLY
            else Attention
        )
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

        self.attn = attn_cls(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            cache_config=cache_config,
            quant_config=quant_config,
            per_layer_sliding_window=sliding_window,
            attn_type=attn_type,
            prefix=f"{prefix}.attn",
        )

        self.q_norm = RMSNorm(self.head_dim, eps=config.rms_norm_eps)
        self.k_norm = RMSNorm(self.head_dim, eps=config.rms_norm_eps)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        qkv, _ = self.qkv_proj(hidden_states)
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        q = self.q_norm(q.contiguous().view(-1, self.head_dim)).view_as(q)
        k = self.k_norm(k.contiguous().view(-1, self.head_dim)).view_as(k)
        q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(q, k, v)
        output, _ = self.o_proj(attn_output)
        return output

225
226
227
    def _init_rotary_emb(
        self,
        config: ApertusConfig,
228
229
        rope_scaling: dict[str, Any] | None,
        quant_config: QuantizationConfig | None,
230
    ) -> None:
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
        is_neox_style = True
        is_gguf = quant_config and quant_config.get_name() == "gguf"
        if is_gguf and config.model_type == "apertus":
            is_neox_style = False

        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=int(self.partial_rotary_factor * self.head_dim),
            max_position=self.max_position_embeddings,
            base=self.rope_theta,
            rope_scaling=rope_scaling,
            is_neox_style=is_neox_style,
            partial_rotary_factor=self.partial_rotary_factor,
        )


class ApertusDecoderLayer(nn.Module):
    def __init__(
        self,
        config: ApertusConfig,
251
252
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
253
254
255
256
257
258
259
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        rope_theta = getattr(config, "rope_theta", 10000)
        rope_scaling = getattr(config, "rope_scaling", None)
        if rope_scaling is not None and getattr(
260
261
            config, "original_max_position_embeddings", None
        ):
262
            rope_scaling["original_max_position_embeddings"] = (
263
264
265
                config.original_max_position_embeddings
            )
        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
266
267
268
        # Support abacusai/Smaug-72B-v0.1 with attention_bias
        # Support internlm/internlm-7b with bias
        attention_bias = getattr(config, "attention_bias", False) or getattr(
269
270
            config, "bias", False
        )
271
272
        bias_o_proj = attention_bias
        # support internlm/internlm3-8b with qkv_bias
273
        if hasattr(config, "qkv_bias"):
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
            attention_bias = config.qkv_bias

        # Apertus defaults to causal attention as it is a decoder-only model.
        # You can override the HF config with `is_causal=False` to enable
        # bidirectional attention, which is used in some embedding models
        # (e.g. parasail-ai/GritLM-7B-vllm)
        if getattr(config, "is_causal", True):
            attn_type = AttentionType.DECODER
        else:
            attn_type = AttentionType.ENCODER_ONLY

        self.self_attn = ApertusAttention(
            config=config,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
289
290
291
            num_kv_heads=getattr(
                config, "num_key_value_heads", config.num_attention_heads
            ),
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            max_position_embeddings=max_position_embeddings,
            quant_config=quant_config,
            bias=attention_bias,
            bias_o_proj=bias_o_proj,
            cache_config=cache_config,
            prefix=f"{prefix}.self_attn",
            attn_type=attn_type,
        )
        self.mlp = ApertusMLP(
            hidden_size=self.hidden_size,
            intermediate_size=config.intermediate_size,
            hidden_act=config.hidden_act,
            quant_config=quant_config,
            bias=getattr(config, "mlp_bias", False),
            prefix=f"{prefix}.mlp",
        )
310
311
312
313
        self.attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.feedforward_layernorm = RMSNorm(
            config.hidden_size, eps=config.rms_norm_eps
        )
314
315
316
317
318

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
319
        residual: torch.Tensor | None,
320
321
322
323
324
325
    ) -> tuple[torch.Tensor, torch.Tensor]:
        # Self Attention
        if residual is None:
            residual = hidden_states
            hidden_states = self.attention_layernorm(hidden_states)
        else:
326
327
            hidden_states, residual = self.attention_layernorm(hidden_states, residual)
        hidden_states = self.self_attn(positions=positions, hidden_states=hidden_states)
328
329

        # Fully Connected
330
        hidden_states, residual = self.feedforward_layernorm(hidden_states, residual)
331
332
333
334
335
336
        hidden_states = self.mlp(hidden_states)
        return hidden_states, residual


@support_torch_compile
class ApertusModel(nn.Module):
337
338
339
340
341
342
343
    def __init__(
        self,
        *,
        vllm_config: VllmConfig,
        prefix: str = "",
        layer_type: type[nn.Module] = ApertusDecoderLayer,
    ):
344
345
346
347
348
349
350
351
352
        super().__init__()

        config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config

        self.config = config
        self.quant_config = quant_config
353
354
355
356
357
        lora_vocab = (
            (lora_config.lora_extra_vocab_size * (lora_config.max_loras or 1))
            if lora_config
            else 0
        )
358
359
        self.vocab_size = config.vocab_size + lora_vocab
        self.org_vocab_size = config.vocab_size
360
361
362
        if get_pp_group().is_first_rank or (
            config.tie_word_embeddings and get_pp_group().is_last_rank
        ):
363
364
365
366
367
368
369
370
371
372
            self.embed_tokens = VocabParallelEmbedding(
                self.vocab_size,
                config.hidden_size,
                org_num_embeddings=config.vocab_size,
                quant_config=quant_config,
            )
        else:
            self.embed_tokens = PPMissingLayer()
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
373
374
375
376
377
378
            lambda prefix: layer_type(
                config=config,
                cache_config=cache_config,
                quant_config=quant_config,
                prefix=prefix,
            ),
379
380
381
382
383
384
385
386
387
            prefix=f"{prefix}.layers",
        )
        if get_pp_group().is_last_rank:
            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        else:
            self.norm = PPMissingLayer()

        self.aux_hidden_state_layers = tuple[int, ...]()

388
389
390
        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], config.hidden_size
        )
391
392
393
394
395
396

    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.embed_tokens(input_ids)

    def forward(
        self,
397
        input_ids: torch.Tensor | None,
398
        positions: torch.Tensor,
399
400
401
        intermediate_tensors: IntermediateTensors | None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors | tuple[torch.Tensor, list[torch.Tensor]]:
402
403
404
405
406
407
408
409
410
411
412
413
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.get_input_embeddings(input_ids)
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

        aux_hidden_states = []
414
415
416
        for idx, layer in enumerate(
            islice(self.layers, self.start_layer, self.end_layer)
        ):
417
418
419
420
421
            if idx in self.aux_hidden_state_layers:
                aux_hidden_states.append(hidden_states + residual)
            hidden_states, residual = layer(positions, hidden_states, residual)

        if not get_pp_group().is_last_rank:
422
423
424
            return IntermediateTensors(
                {"hidden_states": hidden_states, "residual": residual}
            )
425
426
427
428
429
430
431

        hidden_states, _ = self.norm(hidden_states, residual)

        if len(aux_hidden_states) > 0:
            return hidden_states, aux_hidden_states
        return hidden_states

432
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
433
434
435
436
437
438
439
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            (".qkv_proj", ".q_proj", "q"),
            (".qkv_proj", ".k_proj", "k"),
            (".qkv_proj", ".v_proj", "v"),
        ]
        params_dict = dict(self.named_parameters())
440
441
442
443
444
445

        # we need to load the buffers for beta and eps (XIELU)
        for name, buffer in self.named_buffers():
            if name.endswith(".beta") or name.endswith(".eps"):
                params_dict[name] = buffer

446
447
448
449
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue
450
            if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
451
452
453
                # Models trained using ColossalAI may include these tensors in
                # the checkpoint. Skip them.
                continue
454
455
456
            if self.quant_config is not None and (
                scale_name := self.quant_config.get_cache_scale(name)
            ):
457
458
                # Loading kv cache quantization scales
                param = params_dict[scale_name]
459
460
461
462
                weight_loader = getattr(param, "weight_loader", default_weight_loader)
                loaded_weight = (
                    loaded_weight if loaded_weight.dim() == 0 else loaded_weight[0]
                )
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
                weight_loader(param, loaded_weight)
                loaded_params.add(scale_name)
                continue
            if "scale" in name:
                # Remapping the name of FP8 kv-scale.
                name = maybe_remap_kv_scale_name(name, params_dict)
                if name is None:
                    continue
            for param_name, weight_name, shard_id in stacked_params_mapping:
                if weight_name not in name:
                    continue
                name = name.replace(weight_name, param_name)
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue

                if is_pp_missing_parameter(name, self):
                    continue

                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue

                if is_pp_missing_parameter(name, self):
                    continue

                param = params_dict[name]
495
                weight_loader = getattr(param, "weight_loader", default_weight_loader)
496
497
498
499
500
501
502
503
504
505
506
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params


class ApertusForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
    packed_modules_mapping = {"qkv_proj": ["q_proj", "k_proj", "v_proj"]}

    # LoRA specific attributes
    embedding_modules = {
        "embed_tokens": "input_embeddings",
507
        "lm_head": "output_embeddings",
508
509
510
    }
    embedding_padding_modules = ["lm_head"]

511
512
513
514
515
516
517
    def __init__(
        self,
        *,
        vllm_config: VllmConfig,
        prefix: str = "",
        layer_type: type[nn.Module] = ApertusDecoderLayer,
    ):
518
519
520
521
522
523
524
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config
        self.config = config
        self.lora_config = lora_config

525
526
527
528
529
        self.model = self._init_model(
            vllm_config=vllm_config,
            prefix=maybe_prefix(prefix, "model"),
            layer_type=layer_type,
        )
530
531
532
533
534
535
536
537
538
539
540
541
542

        if get_pp_group().is_last_rank:
            self.unpadded_vocab_size = config.vocab_size
            if lora_config:
                self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
            self.lm_head = ParallelLMHead(
                self.unpadded_vocab_size,
                config.hidden_size,
                org_num_embeddings=config.vocab_size,
                padding_size=(
                    DEFAULT_VOCAB_PADDING_SIZE
                    # We need bigger padding if using lora for kernel
                    # compatibility
543
544
545
                    if not lora_config
                    else lora_config.lora_vocab_padding_size
                ),
546
547
548
549
                quant_config=quant_config,
                prefix=maybe_prefix(prefix, "lm_head"),
            )
            if config.tie_word_embeddings:
550
                self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens)
551
552

            logit_scale = getattr(config, "logit_scale", 1.0)
553
554
555
            self.logits_processor = LogitsProcessor(
                self.unpadded_vocab_size, config.vocab_size, logit_scale
            )
556
557
558
559
        else:
            self.lm_head = PPMissingLayer()

        self.make_empty_intermediate_tensors = (
560
561
            self.model.make_empty_intermediate_tensors
        )
562
563
564
565
566
567
568
569

    def set_aux_hidden_state_layers(self, layers: tuple[int, ...]) -> None:
        self.model.aux_hidden_state_layers = layers

    def get_eagle3_aux_hidden_state_layers(self) -> tuple[int, ...]:
        num_layers = len(self.model.layers)
        return (2, num_layers // 2, num_layers - 3)

570
571
572
573
574
575
576
577
578
    def _init_model(
        self,
        vllm_config: VllmConfig,
        prefix: str = "",
        layer_type: type[nn.Module] = ApertusDecoderLayer,
    ):
        return ApertusModel(
            vllm_config=vllm_config, prefix=prefix, layer_type=layer_type
        )
579
580
581
582
583
584
585
586

    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.get_input_embeddings(input_ids)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
587
588
589
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
590
591
592
        model_output = self.model(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
593
594
595
596
597
        return model_output

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
598
    ) -> torch.Tensor | None:
599
        logits = self.logits_processor(self.lm_head, hidden_states)
600
601
        return logits

602
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
603
604
        loader = AutoWeightsLoader(
            self,
605
            skip_prefixes=(["lm_head."] if self.config.tie_word_embeddings else None),
606
607
        )
        return loader.load_weights(weights)