File size: 13,446 Bytes
ca1ecab
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# coding=utf-8
# Adapted from
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/opt/modeling_opt.py
# Copyright 2023 The vLLM team.
# Copyright 2022 The Fairseq Authors and The HuggingFace Inc. team. All rights
# reserved.
#
# 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 OPT model compatible with HuggingFace weights."""
from typing import List, Optional, Tuple

import torch
from torch import nn
from transformers import OPTConfig

from vllm.model_executor.input_metadata import InputMetadata
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.attention import PagedAttention
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                               LinearMethodBase,
                                               QKVParallelLinear,
                                               ReplicatedLinear,
                                               RowParallelLinear)
from vllm.model_executor.layers.sampler import Sampler
from vllm.model_executor.layers.vocab_parallel_embedding import (
    VocabParallelEmbedding)
from vllm.model_executor.parallel_utils.parallel_state import (
    get_tensor_model_parallel_world_size)
from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.model_executor.weight_utils import (default_weight_loader,
                                              hf_model_weights_iterator)
from vllm.sequence import SamplerOutput

KVCache = Tuple[torch.Tensor, torch.Tensor]


class OPTLearnedPositionalEmbedding(nn.Embedding):

    def __init__(self, num_embeddings: int, embedding_dim: int):
        # OPT is set up so that if padding_idx is specified then offset the
        # embedding ids by 2 and adjust num_embeddings appropriately. Other
        # models don't have this hack
        self.offset = 2
        super().__init__(num_embeddings + self.offset, embedding_dim)

    def forward(self, positions: torch.Tensor):
        return super().forward(positions + self.offset)


class OPTAttention(nn.Module):

    def __init__(
        self,
        embed_dim: int,
        num_heads: int,
        bias: bool = True,
        linear_method: Optional[LinearMethodBase] = None,
    ) -> None:
        super().__init__()
        self.embed_dim = embed_dim
        tensor_model_parallel_world_size = (
            get_tensor_model_parallel_world_size())
        total_num_heads = num_heads
        assert num_heads % tensor_model_parallel_world_size == 0
        self.num_heads = total_num_heads // tensor_model_parallel_world_size
        self.head_dim = embed_dim // total_num_heads
        self.scaling = self.head_dim**-0.5

        self.qkv_proj = QKVParallelLinear(
            embed_dim,
            self.head_dim,
            total_num_heads,
            bias=bias,
            linear_method=linear_method,
        )
        self.out_proj = RowParallelLinear(
            embed_dim,
            embed_dim,
            bias=bias,
            linear_method=linear_method,
        )
        self.attn = PagedAttention(self.num_heads,
                                   self.head_dim,
                                   scale=self.scaling)

    def forward(
        self,
        hidden_states: torch.Tensor,
        kv_cache: KVCache,
        input_metadata: InputMetadata,
    ) -> torch.Tensor:
        qkv, _ = self.qkv_proj(hidden_states)
        q, k, v = qkv.chunk(chunks=3, dim=-1)
        key_cache, value_cache = kv_cache
        attn_output = self.attn(q, k, v, key_cache, value_cache,
                                input_metadata)
        output, _ = self.out_proj(attn_output)
        return output


class OPTDecoderLayer(nn.Module):

    def __init__(
        self,
        config: OPTConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ):
        super().__init__()
        self.config = config
        self.embed_dim = config.hidden_size
        self.self_attn = OPTAttention(
            embed_dim=self.embed_dim,
            num_heads=config.num_attention_heads,
            bias=config.enable_bias,
            linear_method=linear_method,
        )
        self.do_layer_norm_before = config.do_layer_norm_before

        self.self_attn_layer_norm = nn.LayerNorm(
            self.embed_dim,
            elementwise_affine=config.layer_norm_elementwise_affine)
        self.fc1 = ColumnParallelLinear(
            self.embed_dim,
            config.ffn_dim,
            bias=config.enable_bias,
            linear_method=linear_method,
        )
        quant_config = getattr(linear_method, "quant_config", None)
        self.activation_fn = get_act_fn(config.activation_function,
                                        quant_config, config.ffn_dim)
        self.fc2 = RowParallelLinear(
            config.ffn_dim,
            self.embed_dim,
            bias=config.enable_bias,
            linear_method=linear_method,
        )
        self.final_layer_norm = nn.LayerNorm(
            self.embed_dim,
            elementwise_affine=config.layer_norm_elementwise_affine)

    def forward(
        self,
        hidden_states: torch.Tensor,
        kv_cache: KVCache,
        input_metadata: InputMetadata,
    ) -> torch.Tensor:
        # Self Attention
        residual = hidden_states
        # 125m, 1.7B, ..., 175B applies layer norm BEFORE attention
        if self.do_layer_norm_before:
            hidden_states = self.self_attn_layer_norm(hidden_states)
        hidden_states = self.self_attn(hidden_states=hidden_states,
                                       kv_cache=kv_cache,
                                       input_metadata=input_metadata)
        hidden_states = residual + hidden_states
        # 350m applies layer norm AFTER attention
        if not self.do_layer_norm_before:
            hidden_states = self.self_attn_layer_norm(hidden_states)

        # Fully Connected
        residual = hidden_states
        # 125m, 1.7B, ..., 175B applies layer norm BEFORE attention
        if self.do_layer_norm_before:
            hidden_states = self.final_layer_norm(hidden_states)
        hidden_states, _ = self.fc1(hidden_states)
        hidden_states = self.activation_fn(hidden_states)
        hidden_states, _ = self.fc2(hidden_states)
        hidden_states = residual + hidden_states
        # 350m applies layer norm AFTER attention
        if not self.do_layer_norm_before:
            hidden_states = self.final_layer_norm(hidden_states)
        return hidden_states


class OPTDecoder(nn.Module):

    def __init__(
        self,
        config: OPTConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ):
        super().__init__()
        self.config = config
        self.padding_idx = config.pad_token_id
        self.max_target_positions = config.max_position_embeddings
        self.vocab_size = config.vocab_size

        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.word_embed_proj_dim,
        )
        # Positional embeddings are replicated (not sharded).
        self.embed_positions = OPTLearnedPositionalEmbedding(
            config.max_position_embeddings, config.hidden_size)

        # Project out & in will be replicated if they exist.
        if config.word_embed_proj_dim != config.hidden_size:
            self.project_out = ReplicatedLinear(config.hidden_size,
                                                config.word_embed_proj_dim,
                                                bias=False,
                                                linear_method=linear_method)
        else:
            self.project_out = None

        if config.word_embed_proj_dim != config.hidden_size:
            self.project_in = ReplicatedLinear(config.word_embed_proj_dim,
                                               config.hidden_size,
                                               bias=False,
                                               linear_method=linear_method)
        else:
            self.project_in = None

        # Note that the only purpose of `config._remove_final_layer_norm` is to
        # keep backward compatibility with checkpoints that have been fine-tuned
        # before transformers v4.20.1
        # see https://github.com/facebookresearch/metaseq/pull/164
        if config.do_layer_norm_before and not config._remove_final_layer_norm:
            self.final_layer_norm = nn.LayerNorm(
                config.hidden_size,
                elementwise_affine=config.layer_norm_elementwise_affine)
        else:
            self.final_layer_norm = None

        self.layers = nn.ModuleList([
            OPTDecoderLayer(config, linear_method)
            for _ in range(config.num_hidden_layers)
        ])

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        kv_caches: List[KVCache],
        input_metadata: InputMetadata,
    ) -> torch.Tensor:
        inputs_embeds = self.embed_tokens(input_ids)
        pos_embeds = self.embed_positions(positions)
        if self.project_in is not None:
            inputs_embeds, _ = self.project_in(inputs_embeds)
        hidden_states = inputs_embeds + pos_embeds

        for i in range(len(self.layers)):
            layer = self.layers[i]
            hidden_states = layer(hidden_states, kv_caches[i], input_metadata)

        if self.final_layer_norm is not None:
            hidden_states = self.final_layer_norm(hidden_states)
        if self.project_out is not None:
            hidden_states, _ = self.project_out(hidden_states)
        return hidden_states


class OPTModel(nn.Module):

    def __init__(
        self,
        config: OPTConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ):
        super().__init__()
        self.decoder = OPTDecoder(config, linear_method)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        kv_caches: List[KVCache],
        input_metadata: InputMetadata,
    ) -> torch.Tensor:
        return self.decoder(input_ids, positions, kv_caches, input_metadata)


class OPTForCausalLM(nn.Module):

    def __init__(
        self,
        config,
        linear_method: Optional[LinearMethodBase] = None,
    ):
        super().__init__()
        self.config = config
        self.linear_method = linear_method
        self.model = OPTModel(config, linear_method)
        self.lm_head_weight = self.model.decoder.embed_tokens.weight
        self.sampler = Sampler(config.vocab_size)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        kv_caches: List[KVCache],
        input_metadata: InputMetadata,
    ) -> torch.Tensor:
        hidden_states = self.model(input_ids, positions, kv_caches,
                                   input_metadata)
        return hidden_states

    def sample(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[SamplerOutput]:
        next_tokens = self.sampler(self.lm_head_weight, hidden_states,
                                   sampling_metadata)
        return next_tokens

    def load_weights(self,
                     model_name_or_path: str,
                     cache_dir: Optional[str] = None,
                     load_format: str = "auto",
                     revision: Optional[str] = None):
        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(remove_duplicate=False))
        for name, loaded_weight in hf_model_weights_iterator(
                model_name_or_path, cache_dir, load_format, revision):
            if "lm_head.weight" in name:
                continue
            if name.startswith("decoder."):
                name = "model." + name

            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
                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
                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                weight_loader(param, loaded_weight)