llama.py 25.3 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

3
4
# Adapted from
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
Woosuk Kwon's avatar
Woosuk Kwon committed
5
# Copyright 2023 The vLLM team.
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
# 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 minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team 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.
Woosuk Kwon's avatar
Woosuk Kwon committed
24
"""Inference-only LLaMA model compatible with HuggingFace weights."""
25
26
from collections.abc import Iterable
from typing import Any, Optional, Union
Woosuk Kwon's avatar
Woosuk Kwon committed
27
28
29
30
31

import torch
from torch import nn
from transformers import LlamaConfig

32
from vllm.attention import Attention, AttentionType
33
from vllm.compilation.decorators import support_torch_compile
34
from vllm.config import CacheConfig, VllmConfig
35
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
Woosuk Kwon's avatar
Woosuk Kwon committed
36
from vllm.model_executor.layers.activation import SiluAndMul
37
from vllm.model_executor.layers.layernorm import RMSNorm
38
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
39
40
                                               QKVParallelLinear,
                                               RowParallelLinear)
41
from vllm.model_executor.layers.logits_processor import LogitsProcessor
42
from vllm.model_executor.layers.quantization import QuantizationConfig
43
from vllm.model_executor.layers.rotary_embedding import get_rope
44
from vllm.model_executor.layers.vocab_parallel_embedding import (
45
    DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
46
from vllm.model_executor.model_loader.weight_utils import (
47
    default_weight_loader, maybe_remap_kv_scale_name)
48
from vllm.model_executor.sampling_metadata import SamplingMetadata
49
from vllm.sequence import IntermediateTensors
Woosuk Kwon's avatar
Woosuk Kwon committed
50

51
from .interfaces import SupportsLoRA, SupportsPP
52
53
from .utils import (AutoWeightsLoader, PPMissingLayer, extract_layer_index,
                    is_pp_missing_parameter,
54
55
                    make_empty_intermediate_tensors_factory, make_layers,
                    maybe_prefix)
56

Woosuk Kwon's avatar
Woosuk Kwon committed
57
58

class LlamaMLP(nn.Module):
59

Woosuk Kwon's avatar
Woosuk Kwon committed
60
61
    def __init__(
        self,
62
63
64
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
65
        quant_config: Optional[QuantizationConfig] = None,
66
        bias: bool = False,
67
        prefix: str = "",
68
        reduce_results: bool = True,
69
    ) -> None:
Woosuk Kwon's avatar
Woosuk Kwon committed
70
        super().__init__()
71
        self.gate_up_proj = MergedColumnParallelLinear(
72
73
            input_size=hidden_size,
            output_sizes=[intermediate_size] * 2,
74
            bias=bias,
75
            quant_config=quant_config,
76
77
78
79
80
81
82
            prefix=f"{prefix}.gate_up_proj",
        )
        self.down_proj = RowParallelLinear(
            input_size=intermediate_size,
            output_size=hidden_size,
            bias=bias,
            quant_config=quant_config,
83
            reduce_results=reduce_results,
84
85
            prefix=f"{prefix}.down_proj",
        )
86
87
88
        if hidden_act != "silu":
            raise ValueError(f"Unsupported activation: {hidden_act}. "
                             "Only silu is supported for now.")
Woosuk Kwon's avatar
Woosuk Kwon committed
89
        self.act_fn = SiluAndMul()
Woosuk Kwon's avatar
Woosuk Kwon committed
90
91

    def forward(self, x):
92
93
        x, _ = self.gate_up_proj(x)
        x = self.act_fn(x)
Woosuk Kwon's avatar
Woosuk Kwon committed
94
95
96
97
98
99
        x, _ = self.down_proj(x)
        return x


class LlamaAttention(nn.Module):

100
101
102
103
104
105
106
    def __init__(
        self,
        config: LlamaConfig,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        rope_theta: float = 10000,
107
        rope_scaling: Optional[dict[str, Any]] = None,
108
109
110
111
112
113
114
115
        max_position_embeddings: int = 8192,
        quant_config: Optional[QuantizationConfig] = None,
        bias: bool = False,
        bias_o_proj: bool = False,
        cache_config: Optional[CacheConfig] = None,
        prefix: str = "",
        attn_type: str = AttentionType.DECODER,
    ) -> None:
Woosuk Kwon's avatar
Woosuk Kwon committed
116
        super().__init__()
117
        layer_idx = extract_layer_index(prefix)
118
        self.hidden_size = hidden_size
Zhuohan Li's avatar
Zhuohan Li committed
119
        tp_size = get_tensor_model_parallel_world_size()
120
        self.total_num_heads = num_heads
Zhuohan Li's avatar
Zhuohan Li committed
121
122
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
123
        self.total_num_kv_heads = num_kv_heads
124
125
126
127
128
129
130
131
132
        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)
133
        # MistralConfig has an optional head_dim introduced by Mistral-Nemo
134
135
136
137
        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
Amit Garg's avatar
Amit Garg committed
138
        # Phi models introduced a partial_rotary_factor parameter in the config
139
140
        self.partial_rotary_factor = getattr(config, "partial_rotary_factor",
                                             1)
Zhuohan Li's avatar
Zhuohan Li committed
141
142
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim
143
        self.scaling = self.head_dim**-0.5
144
145
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings
Woosuk Kwon's avatar
Woosuk Kwon committed
146

147
        self.qkv_proj = QKVParallelLinear(
148
149
150
151
            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,
152
            bias=bias,
153
            quant_config=quant_config,
154
            prefix=f"{prefix}.qkv_proj",
Woosuk Kwon's avatar
Woosuk Kwon committed
155
        )
156

157
        self.o_proj = RowParallelLinear(
158
159
            input_size=self.total_num_heads * self.head_dim,
            output_size=hidden_size,
160
            bias=bias_o_proj,
161
            quant_config=quant_config,
162
            prefix=f"{prefix}.o_proj",
Woosuk Kwon's avatar
Woosuk Kwon committed
163
        )
Woosuk Kwon's avatar
Woosuk Kwon committed
164

165
166
167
        self._init_rotary_emb(config,
                              rope_scaling=rope_scaling,
                              quant_config=quant_config)
168
169

        if hasattr(config, "interleaved_sliding_window"):
170
171
172
173
174
175
            interleaved_sliding_window = config.interleaved_sliding_window
            if isinstance(interleaved_sliding_window, int):
                sliding_window = interleaved_sliding_window
            elif isinstance(interleaved_sliding_window, list):
                sw_idx = layer_idx % len(interleaved_sliding_window)
                sliding_window = interleaved_sliding_window[sw_idx]
176
            else:
177
178
                raise ValueError(
                    f"{type(interleaved_sliding_window)} is not supported.")
179
180
181
        else:
            sliding_window = None

182
183
184
185
186
187
188
        self.attn = Attention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            cache_config=cache_config,
            quant_config=quant_config,
189
            per_layer_sliding_window=sliding_window,
190
            attn_type=attn_type,
191
            prefix=f"{prefix}.attn",
192
        )
Woosuk Kwon's avatar
Woosuk Kwon committed
193
194
195

    def forward(
        self,
196
        positions: torch.Tensor,
Woosuk Kwon's avatar
Woosuk Kwon committed
197
198
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
199
        qkv, _ = self.qkv_proj(hidden_states)
Zhuohan Li's avatar
Zhuohan Li committed
200
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
201
        q, k = self.rotary_emb(positions, q, k)
202
        attn_output = self.attn(q, k, v)
Woosuk Kwon's avatar
Woosuk Kwon committed
203
204
205
        output, _ = self.o_proj(attn_output)
        return output

206
207
208
209
210
    def _init_rotary_emb(self, config: LlamaConfig,
                         rope_scaling: Optional[dict[str, Any]],
                         quant_config: Optional[QuantizationConfig]) -> None:
        is_neox_style = True
        is_gguf = quant_config and quant_config.get_name() == "gguf"
211
        if is_gguf and config.model_type == "llama":
212
213
214
215
216
217
218
219
220
221
222
223
            is_neox_style = False

        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=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,
        )

Woosuk Kwon's avatar
Woosuk Kwon committed
224
225
226

class LlamaDecoderLayer(nn.Module):

227
228
229
    def __init__(
        self,
        config: LlamaConfig,
230
        cache_config: Optional[CacheConfig] = None,
231
        quant_config: Optional[QuantizationConfig] = None,
232
        prefix: str = "",
233
    ) -> None:
Woosuk Kwon's avatar
Woosuk Kwon committed
234
235
        super().__init__()
        self.hidden_size = config.hidden_size
236
237
        rope_theta = getattr(config, "rope_theta", 10000)
        rope_scaling = getattr(config, "rope_scaling", None)
238
239
240
241
        if rope_scaling is not None and getattr(
                config, "original_max_position_embeddings", None):
            rope_scaling["original_max_position_embeddings"] = (
                config.original_max_position_embeddings)
242
243
        max_position_embeddings = getattr(config, "max_position_embeddings",
                                          8192)
244
245
246
247
        # Support abacusai/Smaug-72B-v0.1 with attention_bias
        # Support internlm/internlm-7b with bias
        attention_bias = getattr(config, "attention_bias", False) or getattr(
            config, "bias", False)
248
249
250
251
252
        bias_o_proj = attention_bias
        # support internlm/internlm3-8b with qkv_bias
        if hasattr(config, 'qkv_bias'):
            attention_bias = config.qkv_bias

253
254
255
256
257
258
259
260
261
        # By default, Llama uses 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

Woosuk Kwon's avatar
Woosuk Kwon committed
262
        self.self_attn = LlamaAttention(
263
            config=config,
264
265
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
266
267
            num_kv_heads=getattr(config, "num_key_value_heads",
                                 config.num_attention_heads),
268
269
270
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            max_position_embeddings=max_position_embeddings,
271
            quant_config=quant_config,
272
            bias=attention_bias,
273
            bias_o_proj=bias_o_proj,
274
            cache_config=cache_config,
275
            prefix=f"{prefix}.self_attn",
276
            attn_type=attn_type,
Woosuk Kwon's avatar
Woosuk Kwon committed
277
278
        )
        self.mlp = LlamaMLP(
279
280
281
            hidden_size=self.hidden_size,
            intermediate_size=config.intermediate_size,
            hidden_act=config.hidden_act,
282
            quant_config=quant_config,
283
            bias=getattr(config, "mlp_bias", False),
284
            prefix=f"{prefix}.mlp",
Woosuk Kwon's avatar
Woosuk Kwon committed
285
        )
286
287
288
289
        self.input_layernorm = RMSNorm(config.hidden_size,
                                       eps=config.rms_norm_eps)
        self.post_attention_layernorm = RMSNorm(config.hidden_size,
                                                eps=config.rms_norm_eps)
Woosuk Kwon's avatar
Woosuk Kwon committed
290
291
292

    def forward(
        self,
293
        positions: torch.Tensor,
Woosuk Kwon's avatar
Woosuk Kwon committed
294
        hidden_states: torch.Tensor,
295
        residual: Optional[torch.Tensor],
296
    ) -> tuple[torch.Tensor, torch.Tensor]:
Woosuk Kwon's avatar
Woosuk Kwon committed
297
        # Self Attention
298
299
300
301
302
303
        if residual is None:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
        else:
            hidden_states, residual = self.input_layernorm(
                hidden_states, residual)
304
        hidden_states = self.self_attn(positions=positions,
305
                                       hidden_states=hidden_states)
Woosuk Kwon's avatar
Woosuk Kwon committed
306
307

        # Fully Connected
308
309
        hidden_states, residual = self.post_attention_layernorm(
            hidden_states, residual)
Woosuk Kwon's avatar
Woosuk Kwon committed
310
        hidden_states = self.mlp(hidden_states)
311
        return hidden_states, residual
Woosuk Kwon's avatar
Woosuk Kwon committed
312
313


314
@support_torch_compile
Woosuk Kwon's avatar
Woosuk Kwon committed
315
316
class LlamaModel(nn.Module):

317
318
319
320
    def __init__(self,
                 *,
                 vllm_config: VllmConfig,
                 prefix: str = "",
321
                 layer_type: type[nn.Module] = LlamaDecoderLayer):
Woosuk Kwon's avatar
Woosuk Kwon committed
322
        super().__init__()
323
324
325
326
327
328

        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

Woosuk Kwon's avatar
Woosuk Kwon committed
329
        self.config = config
330
        self.quant_config = quant_config
331
332
333
334
        lora_vocab = (lora_config.lora_extra_vocab_size *
                      (lora_config.max_loras or 1)) if lora_config else 0
        self.vocab_size = config.vocab_size + lora_vocab
        self.org_vocab_size = config.vocab_size
335
336
337
338
339
340
        if get_pp_group().is_first_rank or (config.tie_word_embeddings
                                            and get_pp_group().is_last_rank):
            self.embed_tokens = VocabParallelEmbedding(
                self.vocab_size,
                config.hidden_size,
                org_num_embeddings=config.vocab_size,
341
                quant_config=quant_config,
342
343
344
            )
        else:
            self.embed_tokens = PPMissingLayer()
345
        self.start_layer, self.end_layer, self.layers = make_layers(
346
            config.num_hidden_layers,
347
348
349
350
            lambda prefix: layer_type(config=config,
                                      cache_config=cache_config,
                                      quant_config=quant_config,
                                      prefix=prefix),
351
352
            prefix=f"{prefix}.layers",
        )
353
354
355
356
        if get_pp_group().is_last_rank:
            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        else:
            self.norm = PPMissingLayer()
Woosuk Kwon's avatar
Woosuk Kwon committed
357

358
359
        self.aux_hidden_state_layers: tuple[int] = tuple()

360
361
362
363
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(
                ["hidden_states", "residual"], config.hidden_size))

364
365
366
    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.embed_tokens(input_ids)

Woosuk Kwon's avatar
Woosuk Kwon committed
367
368
    def forward(
        self,
369
        input_ids: Optional[torch.Tensor],
370
        positions: torch.Tensor,
371
        intermediate_tensors: Optional[IntermediateTensors],
372
        inputs_embeds: Optional[torch.Tensor] = None,
373
374
    ) -> Union[torch.Tensor, IntermediateTensors, tuple[torch.Tensor,
                                                        list[torch.Tensor]]]:
375
376
377
378
379
380
        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
381
        else:
382
383
384
385
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

386
387
388
389
390
        aux_hidden_states = []
        for idx, layer in enumerate(
                self.layers[self.start_layer:self.end_layer]):
            if idx in self.aux_hidden_state_layers:
                aux_hidden_states.append(hidden_states + residual)
391
            hidden_states, residual = layer(positions, hidden_states, residual)
392
393
394
395
396
397
398

        if not get_pp_group().is_last_rank:
            return IntermediateTensors({
                "hidden_states": hidden_states,
                "residual": residual
            })

399
        hidden_states, _ = self.norm(hidden_states, residual)
400
401
402

        if len(aux_hidden_states) > 0:
            return hidden_states, aux_hidden_states
Woosuk Kwon's avatar
Woosuk Kwon committed
403
404
        return hidden_states

405
406
    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
407
408
409
410
411
412
413
414
415
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            (".qkv_proj", ".q_proj", "q"),
            (".qkv_proj", ".k_proj", "k"),
            (".qkv_proj", ".v_proj", "v"),
            (".gate_up_proj", ".gate_proj", 0),
            (".gate_up_proj", ".up_proj", 1),
        ]
        params_dict = dict(self.named_parameters())
416
        loaded_params: set[str] = set()
417
418
419
420
421
422
423
424
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue
            if ("rotary_emb.cos_cached" in name
                    or "rotary_emb.sin_cached" in name):
                # Models trained using ColossalAI may include these tensors in
                # the checkpoint. Skip them.
                continue
425
426
            if (self.quant_config is not None and
                (scale_name := self.quant_config.get_cache_scale(name))):
427
                # Loading kv cache quantization scales
428
429
430
                param = params_dict[scale_name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
431
432
                loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else
                                 loaded_weight[0])
433
                weight_loader(param, loaded_weight)
434
                loaded_params.add(scale_name)
435
                continue
436
437
438
439
440
            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
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
            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]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                weight_loader(param, loaded_weight)
468
469
            loaded_params.add(name)
        return loaded_params
470

Woosuk Kwon's avatar
Woosuk Kwon committed
471

472
class LlamaForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
Terry's avatar
Terry committed
473
    packed_modules_mapping = {
474
475
        "qkv_proj": ["q_proj", "k_proj", "v_proj"],
        "gate_up_proj": ["gate_proj", "up_proj"]
Terry's avatar
Terry committed
476
477
478
479
480
    }

    # LoRA specific attributes
    embedding_modules = {
        "embed_tokens": "input_embeddings",
481
        "lm_head": "output_embeddings"
Terry's avatar
Terry committed
482
483
    }
    embedding_padding_modules = ["lm_head"]
484

485
486
487
488
489
    # Mistral/Llama models can also be loaded with --load-format mistral
    # from consolidated.safetensors checkpoints
    mistral_mapping = {
        "layers": "model.layers",
        "attention": "self_attn",
490
491
492
        "qscale_act": "input_scale",
        "qscale_weight": "weight_scale",
        "kv_fake_quantizer.qscale_act": "kv_scale",
493
494
495
496
497
498
499
500
501
502
503
504
        "wq": "q_proj",
        "wk": "k_proj",
        "wv": "v_proj",
        "wo": "o_proj",
        "attention_norm": "input_layernorm",
        "feed_forward": "mlp",
        "w1": "gate_proj",
        "w2": "down_proj",
        "w3": "up_proj",
        "ffn_norm": "post_attention_layernorm",
        "tok_embeddings": "model.embed_tokens",
        "output": "lm_head",
505
        "norm": "model.norm",
506
    }
507

508
509
510
511
512
    def __init__(self,
                 *,
                 vllm_config: VllmConfig,
                 prefix: str = "",
                 layer_type: type[nn.Module] = LlamaDecoderLayer):
Woosuk Kwon's avatar
Woosuk Kwon committed
513
        super().__init__()
514
515
516
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config
Woosuk Kwon's avatar
Woosuk Kwon committed
517
        self.config = config
518
519
        self.lora_config = lora_config

520
        self.model = self._init_model(vllm_config=vllm_config,
521
522
                                      prefix=maybe_prefix(prefix, "model"),
                                      layer_type=layer_type)
523

524
525
526
527
528
529
530
531
        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,
532
533
534
535
536
537
                padding_size=(
                    DEFAULT_VOCAB_PADDING_SIZE
                    # We need bigger padding if using lora for kernel
                    # compatibility
                    if not lora_config else
                    lora_config.lora_vocab_padding_size),
538
                quant_config=quant_config,
539
                prefix=maybe_prefix(prefix, "lm_head"),
540
541
            )
            if config.tie_word_embeddings:
542
543
                self.lm_head = self.lm_head.tie_weights(
                    self.model.embed_tokens)
544
545
546
547
548
549
550

            logit_scale = getattr(config, "logit_scale", 1.0)
            self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                                    config.vocab_size,
                                                    logit_scale)
        else:
            self.lm_head = PPMissingLayer()
551

552
553
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)
Woosuk Kwon's avatar
Woosuk Kwon committed
554

555
556
557
558
559
560
561
    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)

562
563
564
565
566
567
568
    def _init_model(self,
                    vllm_config: VllmConfig,
                    prefix: str = "",
                    layer_type: type[nn.Module] = LlamaDecoderLayer):
        return LlamaModel(vllm_config=vllm_config,
                          prefix=prefix,
                          layer_type=layer_type)
569

570
571
572
    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.get_input_embeddings(input_ids)

Woosuk Kwon's avatar
Woosuk Kwon committed
573
574
    def forward(
        self,
575
576
        input_ids: torch.Tensor,
        positions: torch.Tensor,
577
        intermediate_tensors: Optional[IntermediateTensors] = None,
578
        inputs_embeds: Optional[torch.Tensor] = None,
579
    ) -> Union[torch.Tensor, IntermediateTensors]:
580
        model_output = self.model(input_ids, positions, intermediate_tensors,
581
                                  inputs_embeds)
582
        return model_output
583

584
585
586
587
588
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
589
        logits = self.logits_processor(self.lm_head, hidden_states,
590
591
592
                                       sampling_metadata)
        return logits

593
594
    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
595
596
597
598
599
        loader = AutoWeightsLoader(
            self,
            skip_prefixes=(["lm_head."]
                           if self.config.tie_word_embeddings else None),
        )
600
        return loader.load_weights(
601
            self.maybe_remap_mistral(name, loaded_weight)
602
            for name, loaded_weight in weights)
603

604
605
606
    # This function is used to remap the mistral format as
    # used by Mistral and Llama <=2
    def maybe_remap_mistral(
607
608
609
        self,
        name: str,
        loaded_weight: torch.Tensor,
610
    ) -> tuple[str, torch.Tensor]:
611

612
        def permute(w: torch.Tensor, n_heads: int):
613
614
615
616
617
618
619
620
621
622
            attn_in = self.config.head_dim * n_heads
            attn_out = self.config.hidden_size

            return w.view(n_heads, attn_in // n_heads // 2, 2,
                          attn_out).transpose(1, 2).reshape(attn_in, attn_out)

        mapping = self.mistral_mapping
        modules = name.split(".")

        # rotary embeds should be sliced
623
        if "wk" in modules and modules[-1] == "weight":
624
625
            loaded_weight = permute(loaded_weight,
                                    self.config.num_key_value_heads)
626
        elif "wq" in modules and modules[-1] == "weight":
627
628
629
            loaded_weight = permute(loaded_weight,
                                    self.config.num_attention_heads)

630
631
632
633
634
635
636
637
638
639
640
        num_modules = len(modules)
        for i in range(num_modules):
            item = modules[i]
            next_item = modules[i + 1] if i < num_modules - 1 else None

            combined_item = (f"{item}.{next_item}"
                             if next_item is not None else None)

            if combined_item in mapping:
                name = name.replace(combined_item, mapping[combined_item])
            elif item in mapping and mapping[item] not in name:
641
642
643
                name = name.replace(item, mapping[item])

        return name, loaded_weight