gemma2.py 17.1 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

Woosuk Kwon's avatar
Woosuk Kwon committed
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
# Copyright 2024 The vLLM team.
# Copyright 2024 Google Inc. 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.
18
19
from collections.abc import Iterable
from typing import Optional, Union
Woosuk Kwon's avatar
Woosuk Kwon committed
20
21
22
23
24

import torch
from torch import nn
from transformers import Gemma2Config

25
from vllm.attention import Attention
26
from vllm.compilation.decorators import support_torch_compile
27
from vllm.config import CacheConfig, VllmConfig
28
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
29
from vllm.logger import init_logger
Woosuk Kwon's avatar
Woosuk Kwon committed
30
31
32
33
34
35
from vllm.model_executor.layers.activation import GeluAndMul
from vllm.model_executor.layers.layernorm import GemmaRMSNorm
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
                                               QKVParallelLinear,
                                               RowParallelLinear)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
36
from vllm.model_executor.layers.quantization import QuantizationConfig
Woosuk Kwon's avatar
Woosuk Kwon committed
37
from vllm.model_executor.layers.rotary_embedding import get_rope
Woosuk Kwon's avatar
Woosuk Kwon committed
38
39
from vllm.model_executor.layers.vocab_parallel_embedding import (
    VocabParallelEmbedding)
40
41
from vllm.model_executor.model_loader.weight_utils import (
    default_weight_loader, maybe_remap_kv_scale_name)
Woosuk Kwon's avatar
Woosuk Kwon committed
42
from vllm.model_executor.sampling_metadata import SamplingMetadata
43
from vllm.sequence import IntermediateTensors
Woosuk Kwon's avatar
Woosuk Kwon committed
44

45
from .interfaces import SupportsLoRA, SupportsPP
46
from .utils import (AutoWeightsLoader, extract_layer_index,
47
                    is_pp_missing_parameter,
48
49
                    make_empty_intermediate_tensors_factory, make_layers,
                    maybe_prefix)
Woosuk Kwon's avatar
Woosuk Kwon committed
50

51
52
logger = init_logger(__name__)

Woosuk Kwon's avatar
Woosuk Kwon committed
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

class Gemma2MLP(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        hidden_activation: str,
        quant_config: Optional[QuantizationConfig] = None,
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size, [intermediate_size] * 2,
            bias=False,
            quant_config=quant_config)
        self.down_proj = RowParallelLinear(intermediate_size,
                                           hidden_size,
                                           bias=False,
                                           quant_config=quant_config)
        if not (hidden_act == hidden_activation == "gelu_pytorch_tanh"):
            raise ValueError(
                "Gemma2 uses `gelu_pytorch_tanh` as the hidden activation "
                "function. Please set `hidden_act` and `hidden_activation` to "
                "`gelu_pytorch_tanh`.")
        self.act_fn = GeluAndMul(approximate="tanh")

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        gate_up, _ = self.gate_up_proj(x)
        x = self.act_fn(gate_up)
        x, _ = self.down_proj(x)
        return x


class Gemma2Attention(nn.Module):

    def __init__(self,
                 config: Gemma2Config,
                 hidden_size: int,
                 num_heads: int,
                 num_kv_heads: int,
                 head_dim: int,
                 max_position_embeddings: int,
                 rope_theta: float,
                 cache_config: Optional[CacheConfig] = None,
98
                 quant_config: Optional[QuantizationConfig] = None,
99
100
                 attn_logits_soft_cap: Optional[float] = None,
                 prefix: str = "") -> None:
Woosuk Kwon's avatar
Woosuk Kwon committed
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
        super().__init__()
        self.config = config
        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)
        self.head_dim = head_dim
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim
        self.scaling = config.query_pre_attn_scalar**-0.5
        self.rope_theta = rope_theta

        self.qkv_proj = QKVParallelLinear(
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=config.attention_bias,
            quant_config=quant_config,
        )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=config.attention_bias,
            quant_config=quant_config,
        )
Woosuk Kwon's avatar
Woosuk Kwon committed
138
        self.rotary_emb = get_rope(
Woosuk Kwon's avatar
Woosuk Kwon committed
139
            self.head_dim,
Woosuk Kwon's avatar
Woosuk Kwon committed
140
141
            rotary_dim=self.head_dim,
            max_position=max_position_embeddings,
Woosuk Kwon's avatar
Woosuk Kwon committed
142
143
144
145
            base=self.rope_theta,
            is_neox_style=True,
        )

146
147
        # reference:
        # https://github.com/huggingface/transformers/blob/54be2d7ae87e873482b984cc956e165ca4dc0ba3/src/transformers/models/gemma2/modeling_gemma2.py#L312 # noqa
148
        layer_idx = extract_layer_index(prefix)
149
150
        use_sliding_window = (layer_idx % 2 == 0 and getattr(
            config, "interleaved_sliding_window", None) is not None)
151
152
        sliding_window = config.interleaved_sliding_window if \
            use_sliding_window else None
Woosuk Kwon's avatar
Woosuk Kwon committed
153
154
155
156
157
        self.attn = Attention(self.num_heads,
                              self.head_dim,
                              self.scaling,
                              num_kv_heads=self.num_kv_heads,
                              cache_config=cache_config,
158
                              quant_config=quant_config,
159
                              logits_soft_cap=attn_logits_soft_cap,
160
                              per_layer_sliding_window=sliding_window,
161
                              prefix=f"{prefix}.attn")
Woosuk Kwon's avatar
Woosuk Kwon committed
162
163
164
165
166
167
168
169
170

    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, k = self.rotary_emb(positions, q, k)
171
        attn_output = self.attn(q, k, v)
Woosuk Kwon's avatar
Woosuk Kwon committed
172
173
174
175
176
177
178
179
180
181
182
        output, _ = self.o_proj(attn_output)
        return output


class Gemma2DecoderLayer(nn.Module):

    def __init__(
        self,
        config: Gemma2Config,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
183
        prefix: str = "",
Woosuk Kwon's avatar
Woosuk Kwon committed
184
185
186
187
188
189
190
191
192
193
194
195
196
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = Gemma2Attention(
            config=config,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            num_kv_heads=config.num_key_value_heads,
            head_dim=config.head_dim,
            max_position_embeddings=config.max_position_embeddings,
            rope_theta=config.rope_theta,
            cache_config=cache_config,
            quant_config=quant_config,
197
            attn_logits_soft_cap=config.attn_logit_softcapping,
198
            prefix=f"{prefix}.self_attn",
Woosuk Kwon's avatar
Woosuk Kwon committed
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
        )
        self.hidden_size = config.hidden_size
        self.mlp = Gemma2MLP(
            hidden_size=self.hidden_size,
            intermediate_size=config.intermediate_size,
            hidden_act=config.hidden_act,
            hidden_activation=config.hidden_activation,
            quant_config=quant_config,
        )
        self.input_layernorm = GemmaRMSNorm(config.hidden_size,
                                            eps=config.rms_norm_eps)
        self.post_attention_layernorm = GemmaRMSNorm(config.hidden_size,
                                                     eps=config.rms_norm_eps)
        self.pre_feedforward_layernorm = GemmaRMSNorm(config.hidden_size,
                                                      eps=config.rms_norm_eps)
        self.post_feedforward_layernorm = GemmaRMSNorm(config.hidden_size,
                                                       eps=config.rms_norm_eps)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
222
    ) -> tuple[torch.Tensor, torch.Tensor]:
Woosuk Kwon's avatar
Woosuk Kwon committed
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
        if residual is None:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
        else:
            hidden_states, residual = self.input_layernorm(
                hidden_states, residual)
        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
        )
        hidden_states = self.post_attention_layernorm(hidden_states)

        hidden_states, residual = self.pre_feedforward_layernorm(
            hidden_states, residual)
        hidden_states = self.mlp(hidden_states)
        hidden_states = self.post_feedforward_layernorm(hidden_states)
        return hidden_states, residual


242
@support_torch_compile
Woosuk Kwon's avatar
Woosuk Kwon committed
243
244
class Gemma2Model(nn.Module):

245
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
Woosuk Kwon's avatar
Woosuk Kwon committed
246
        super().__init__()
247
248
249
        config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
Woosuk Kwon's avatar
Woosuk Kwon committed
250
        self.config = config
251
        self.quant_config = quant_config
Woosuk Kwon's avatar
Woosuk Kwon committed
252
253
254
255
256

        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
        )
257
258
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
259
260
            lambda prefix: Gemma2DecoderLayer(
                config, cache_config, quant_config, prefix=prefix),
261
            prefix=f"{prefix}.layers")
Woosuk Kwon's avatar
Woosuk Kwon committed
262
263
264
265
266
267
268
269
        self.norm = GemmaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

        # Normalize the embedding by sqrt(hidden_size)
        # The normalizer's data type should be downcasted to the model's
        # data type such as bfloat16, not float32.
        # See https://github.com/huggingface/transformers/pull/29402
        normalizer = self.config.hidden_size**0.5
        self.register_buffer("normalizer", torch.tensor(normalizer))
270
271
272
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(
                ["hidden_states", "residual"], config.hidden_size))
Woosuk Kwon's avatar
Woosuk Kwon committed
273

274
275
276
    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.embed_tokens(input_ids)

Woosuk Kwon's avatar
Woosuk Kwon committed
277
278
    def forward(
        self,
279
        input_ids: Optional[torch.Tensor],
Woosuk Kwon's avatar
Woosuk Kwon committed
280
        positions: torch.Tensor,
281
        intermediate_tensors: Optional[IntermediateTensors],
282
        inputs_embeds: Optional[torch.Tensor] = None,
283
284
    ) -> Union[torch.Tensor, IntermediateTensors]:
        if get_pp_group().is_first_rank:
285
286
287
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
288
                hidden_states = self.get_input_embeddings(input_ids)
289
290
291
292
293
294
            hidden_states *= self.normalizer
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]
295
        for layer in self.layers[self.start_layer:self.end_layer]:
Woosuk Kwon's avatar
Woosuk Kwon committed
296
297
298
299
300
            hidden_states, residual = layer(
                positions,
                hidden_states,
                residual,
            )
301
302
303
304
305
        if not get_pp_group().is_last_rank:
            return IntermediateTensors({
                "hidden_states": hidden_states,
                "residual": residual
            })
Woosuk Kwon's avatar
Woosuk Kwon committed
306
307
308
        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states

309
310
    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
311
312
313
314
315
316
317
318
319
        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())
320
        loaded_params: set[str] = set()
321
        for name, loaded_weight in weights:
322
323
            if (self.quant_config is not None and
                (scale_name := self.quant_config.get_cache_scale(name))):
324
325
326
327
328
329
330
331
                # Loading kv cache scales for compressed-tensors quantization
                param = params_dict[scale_name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                loaded_weight = loaded_weight[0]
                weight_loader(param, loaded_weight)
                loaded_params.add(scale_name)
                continue
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
            for (param_name, shard_name, shard_id) in stacked_params_mapping:
                if shard_name not in name:
                    continue
                name = name.replace(shard_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
349
350
351
352
                # Remapping the name of FP8 kv-scale.
                name = maybe_remap_kv_scale_name(name, params_dict)
                if name is None:
                    continue
353
354
355
356
357
358
359
360
                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)
            loaded_params.add(name)

361
        return loaded_params
362

Woosuk Kwon's avatar
Woosuk Kwon committed
363

364
class Gemma2ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
Woosuk Kwon's avatar
Woosuk Kwon committed
365
366
367
368
369
370
371
372
373
374
375
376
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }

377
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
378
379
380
        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
381
382
383
        del lora_config  # Unused.
        super().__init__()
        self.config = config
384
385
        # currently all existing Gemma models have `tie_word_embeddings` enabled
        assert config.tie_word_embeddings
Woosuk Kwon's avatar
Woosuk Kwon committed
386
        self.quant_config = quant_config
387
388
        self.model = Gemma2Model(vllm_config=vllm_config,
                                 prefix=maybe_prefix(prefix, "model"))
Woosuk Kwon's avatar
Woosuk Kwon committed
389
390
        self.logits_processor = LogitsProcessor(
            config.vocab_size, soft_cap=config.final_logit_softcapping)
391
392
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)
Woosuk Kwon's avatar
Woosuk Kwon committed
393

394
395
396
    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
397
398
399
400
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
401
        intermediate_tensors: Optional[IntermediateTensors] = None,
402
        inputs_embeds: Optional[torch.Tensor] = None,
403
    ) -> Union[torch.Tensor, IntermediateTensors]:
404
        hidden_states = self.model(input_ids, positions, intermediate_tensors,
405
                                   inputs_embeds)
Woosuk Kwon's avatar
Woosuk Kwon committed
406
407
        return hidden_states

408
409
410
411
412
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
413
414
        logits = self.logits_processor(self.model.embed_tokens, hidden_states,
                                       sampling_metadata)
Woosuk Kwon's avatar
Woosuk Kwon committed
415
416
        return logits

417
418
    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
419
420
421
422
423
        loader = AutoWeightsLoader(
            self,
            skip_prefixes=(["lm_head."]
                           if self.config.tie_word_embeddings else None),
        )
424
        return loader.load_weights(weights)