gpt_j.py 11.9 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
# Adapted from
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/gptj/modeling_gptj.py
# Copyright 2023 The vLLM team.
# Copyright 2021 The EleutherAI and HuggingFace Teams. 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.
Woosuk Kwon's avatar
Woosuk Kwon committed
17
"""Inference-only GPT-J model compatible with HuggingFace weights."""
18
from typing import Iterable, List, Optional, Tuple, Union
19
20
21
22
23

import torch
from torch import nn
from transformers import GPTJConfig

24
from vllm.attention import Attention, AttentionMetadata
25
from vllm.compilation.decorators import support_torch_compile
26
from vllm.config import CacheConfig, VllmConfig
27
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
28
from vllm.model_executor.layers.activation import get_act_fn
29
30
31
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                               QKVParallelLinear,
                                               RowParallelLinear)
32
from vllm.model_executor.layers.logits_processor import LogitsProcessor
33
from vllm.model_executor.layers.quantization import QuantizationConfig
34
from vllm.model_executor.layers.rotary_embedding import get_rope
Joe Runde's avatar
Joe Runde committed
35
from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
36
from vllm.model_executor.layers.vocab_parallel_embedding import (
37
    ParallelLMHead, VocabParallelEmbedding)
38
39
from vllm.model_executor.model_loader.weight_utils import (
    default_weight_loader, maybe_remap_kv_scale_name)
40
from vllm.model_executor.sampling_metadata import SamplingMetadata
41
from vllm.sequence import IntermediateTensors
42

43
44
45
46
from .interfaces import SupportsPP
from .utils import (is_pp_missing_parameter,
                    make_empty_intermediate_tensors_factory, make_layers)

47
48
49

class GPTJAttention(nn.Module):

50
51
52
    def __init__(
        self,
        config: GPTJConfig,
53
        cache_config: Optional[CacheConfig] = None,
54
        quant_config: Optional[QuantizationConfig] = None,
55
    ):
56
57
58
59
60
        super().__init__()
        self.total_num_heads = config.num_attention_heads
        self.hidden_size = config.hidden_size
        self.head_size = self.hidden_size // self.total_num_heads

61
        self.qkv_proj = QKVParallelLinear(
62
            config.hidden_size,
63
64
            self.head_size,
            self.total_num_heads,
65
            bias=False,
66
            quant_config=quant_config,
67
68
69
70
71
        )
        self.out_proj = RowParallelLinear(
            config.hidden_size,
            config.hidden_size,
            bias=False,
72
            quant_config=quant_config,
73
        )
74
75
76
77
78
79

        tp_world_size = get_tensor_model_parallel_world_size()
        assert self.total_num_heads % tp_world_size == 0
        self.num_heads = self.total_num_heads // tp_world_size

        scaling = self.head_size**-0.5
80
        assert getattr(config, "rotary", True)
81
        assert config.rotary_dim % 2 == 0
82
83
84
        rope_theta = getattr(config, "rope_theta", 10000)
        max_position_embeddings = getattr(config, "max_position_embeddings",
                                          8192)
Woosuk Kwon's avatar
Woosuk Kwon committed
85
        self.rotary_emb = get_rope(
86
            self.head_size,
Woosuk Kwon's avatar
Woosuk Kwon committed
87
            rotary_dim=config.rotary_dim,
88
            max_position=max_position_embeddings,
Woosuk Kwon's avatar
Woosuk Kwon committed
89
90
91
            base=rope_theta,
            is_neox_style=False,
        )
92
93
94
        self.attn = Attention(self.num_heads,
                              self.head_size,
                              scaling,
95
96
                              cache_config=cache_config,
                              quant_config=quant_config)
97
98
99
100
101

    def forward(
        self,
        position_ids: torch.Tensor,
        hidden_states: torch.Tensor,
102
103
        kv_cache: torch.Tensor,
        attn_metadata: AttentionMetadata,
104
105
106
    ) -> torch.Tensor:
        qkv, _ = self.qkv_proj(hidden_states)
        q, k, v = qkv.chunk(chunks=3, dim=-1)
Woosuk Kwon's avatar
Woosuk Kwon committed
107
        q, k = self.rotary_emb(position_ids, q, k)
108
        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
109
110
111
112
113
114
        attn_output, _ = self.out_proj(attn_output)
        return attn_output


class GPTJMLP(nn.Module):

115
116
117
118
    def __init__(
        self,
        intermediate_size: int,
        config: GPTJConfig,
119
        quant_config: Optional[QuantizationConfig] = None,
120
    ):
121
122
        super().__init__()
        hidden_size = config.n_embd
123
124
125
        self.fc_in = ColumnParallelLinear(
            hidden_size,
            intermediate_size,
126
            quant_config=quant_config,
127
128
129
130
        )
        self.fc_out = RowParallelLinear(
            intermediate_size,
            hidden_size,
131
            quant_config=quant_config,
132
        )
133
        self.act = get_act_fn(config.activation_function)
134
135
136
137
138
139
140
141
142
143

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states, _ = self.fc_in(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states, _ = self.fc_out(hidden_states)
        return hidden_states


class GPTJBlock(nn.Module):

144
145
146
    def __init__(
        self,
        config: GPTJConfig,
147
        cache_config: Optional[CacheConfig] = None,
148
        quant_config: Optional[QuantizationConfig] = None,
149
    ):
150
        super().__init__()
151
152
        inner_dim = (4 * config.n_embd
                     if config.n_inner is None else config.n_inner)
153
        self.ln_1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
154
        self.attn = GPTJAttention(config, cache_config, quant_config)
155
        self.mlp = GPTJMLP(inner_dim, config, quant_config)
156
157
158
159
160

    def forward(
        self,
        position_ids: torch.Tensor,
        hidden_states: torch.Tensor,
161
162
        kv_cache: torch.Tensor,
        attn_metadata: AttentionMetadata,
163
164
165
166
167
168
169
    ) -> torch.Tensor:
        residual = hidden_states
        hidden_states = self.ln_1(hidden_states)
        attn_output = self.attn(
            position_ids=position_ids,
            hidden_states=hidden_states,
            kv_cache=kv_cache,
170
            attn_metadata=attn_metadata,
171
172
173
174
175
176
        )
        mlp_output = self.mlp(hidden_states)
        hidden_states = attn_output + mlp_output + residual
        return hidden_states


177
@support_torch_compile
178
179
class GPTJModel(nn.Module):

180
181
182
    def __init__(
        self,
        config: GPTJConfig,
183
        cache_config: Optional[CacheConfig] = None,
184
        quant_config: Optional[QuantizationConfig] = None,
185
        prefix: str = "",
186
    ):
187
188
189
        super().__init__()
        self.config = config
        self.embed_dim = config.n_embd
190
191
192
193
        self.wte = VocabParallelEmbedding(
            config.vocab_size,
            self.embed_dim,
        )
194
195
196
197
198
        self.start_layer, self.end_layer, self.h = make_layers(
            config.n_layer,
            lambda prefix: GPTJBlock(config, cache_config, quant_config),
            prefix=f"{prefix}.h",
        )
199
        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
200
201
202
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(["hidden_states"],
                                                    config.n_embd))
203
204
205
206
207

    def forward(
        self,
        input_ids: torch.Tensor,
        position_ids: torch.Tensor,
208
209
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
210
211
212
213
214
215
216
        intermediate_tensors: Optional[IntermediateTensors],
    ) -> Union[torch.Tensor, IntermediateTensors]:
        if get_pp_group().is_first_rank:
            hidden_states = self.wte(input_ids)
        else:
            hidden_states = intermediate_tensors["hidden_states"]
        for i in range(self.start_layer, self.end_layer):
217
218
219
220
            layer = self.h[i]
            hidden_states = layer(
                position_ids,
                hidden_states,
221
                kv_caches[i - self.start_layer],
222
                attn_metadata,
223
            )
224
225
        if not get_pp_group().is_last_rank:
            return IntermediateTensors({"hidden_states": hidden_states})
226
227
228
229
        hidden_states = self.ln_f(hidden_states)
        return hidden_states


230
class GPTJForCausalLM(nn.Module, SupportsPP):
231

232
233
    def __init__(
        self,
234
235
        vllm_config: VllmConfig,
        prefix: str = "",
236
    ):
237
        super().__init__()
238
239
240
        config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
241
        self.config = config
242
        self.quant_config = quant_config
243
        assert not config.tie_word_embeddings
244
        self.transformer = GPTJModel(config, cache_config, quant_config)
245
        self.lm_head = ParallelLMHead(
246
            config.vocab_size,
247
248
            config.n_embd,
            bias=True,
249
            quant_config=quant_config,
250
        )
251
        self.logits_processor = LogitsProcessor(config.vocab_size)
Joe Runde's avatar
Joe Runde committed
252
        self.sampler = get_sampler()
253
254
        self.make_empty_intermediate_tensors = (
            self.transformer.make_empty_intermediate_tensors)
255
256
257
258
259

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
260
261
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
262
        intermediate_tensors: Optional[IntermediateTensors] = None,
263
    ) -> Union[torch.Tensor, IntermediateTensors]:
264
        hidden_states = self.transformer(input_ids, positions, kv_caches,
265
                                         attn_metadata, intermediate_tensors)
266
267
        return hidden_states

268
269
270
271
272
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
273
        logits = self.logits_processor(self.lm_head, hidden_states,
274
275
276
                                       sampling_metadata, self.lm_head.bias)
        return logits

277
278
    def sample(
        self,
279
        logits: torch.Tensor,
280
        sampling_metadata: SamplingMetadata,
281
    ) -> Optional[SamplerOutput]:
282
        next_tokens = self.sampler(logits, sampling_metadata)
283
284
        return next_tokens

285
    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
286
287
288
289
290
291
292
293
294
        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())
295
        for name, loaded_weight in weights:
296
297
            if "attn.bias" in name or "attn.masked_bias" in name:
                continue
298
299
            for (param_name, weight_name, shard_id) in stacked_params_mapping:
                if weight_name not in name:
300
                    continue
CHU Tianxiang's avatar
CHU Tianxiang committed
301
302
303
304
                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
305
306
                if is_pp_missing_parameter(name, self):
                    continue
CHU Tianxiang's avatar
CHU Tianxiang committed
307
                param = params_dict[name]
308
309
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
310
                break
311
            else:
312
313
314
                name = maybe_remap_kv_scale_name(name, params_dict)
                if name is None:
                    continue
CHU Tianxiang's avatar
CHU Tianxiang committed
315
316
317
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
318
319
                if is_pp_missing_parameter(name, self):
                    continue
320
321
322
323
                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                weight_loader(param, loaded_weight)