bloom.py 12.1 KB
Newer Older
Woosuk Kwon's avatar
Woosuk Kwon committed
1
# coding=utf-8
2
3
# Adapted from
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/bloom/modeling_bloom.py
Woosuk Kwon's avatar
Woosuk Kwon committed
4
5
6
7
8
9
10
11
12
13
14
15
16
17
# Copyright 2023 The CacheFlow team.
# Copyright 2022 HuggingFace Inc. team and BigScience workshop.
#
# 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
18
"""Inference-only BLOOM model compatible with HuggingFace weights."""
Woosuk Kwon's avatar
Woosuk Kwon committed
19
import math
20
from typing import List, Optional, Tuple
Woosuk Kwon's avatar
Woosuk Kwon committed
21
22
23
24
25
26
27

import torch
from torch import nn
from transformers import BloomConfig

from vllm.model_executor.input_metadata import InputMetadata
from vllm.model_executor.layers.activation import get_act_fn
Woosuk Kwon's avatar
Woosuk Kwon committed
28
from vllm.model_executor.layers.attention import PagedAttention
29
30
31
32
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                               LinearMethodBase,
                                               QKVParallelLinear,
                                               RowParallelLinear)
Woosuk Kwon's avatar
Woosuk Kwon committed
33
from vllm.model_executor.layers.sampler import Sampler
34
35
from vllm.model_executor.layers.vocab_parallel_embedding import (
    VocabParallelEmbedding)
Woosuk Kwon's avatar
Woosuk Kwon committed
36
37
from vllm.model_executor.parallel_utils.parallel_state import (
    get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size)
38
from vllm.model_executor.sampling_metadata import SamplingMetadata
39
40
from vllm.model_executor.weight_utils import (default_weight_loader,
                                              hf_model_weights_iterator)
41
from vllm.sequence import SamplerOutput
Woosuk Kwon's avatar
Woosuk Kwon committed
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

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


def _get_alibi_slopes(total_num_heads: int) -> torch.Tensor:
    closest_power_of_2 = 2**math.floor(math.log2(total_num_heads))
    base = torch.tensor(
        2**(-(2**-(math.log2(closest_power_of_2) - 3))),
        dtype=torch.float32,
    )
    powers = torch.arange(1, 1 + closest_power_of_2, dtype=torch.int32)
    slopes = torch.pow(base, powers)

    if closest_power_of_2 != total_num_heads:
        extra_base = torch.tensor(
            2**(-(2**-(math.log2(2 * closest_power_of_2) - 3))),
            dtype=torch.float32,
        )
        num_remaining_heads = min(closest_power_of_2,
                                  total_num_heads - closest_power_of_2)
        extra_powers = torch.arange(start=1,
                                    end=1 + 2 * num_remaining_heads,
                                    step=2,
                                    dtype=torch.int32)
        slopes = torch.cat(
            [slopes, torch.pow(extra_base, extra_powers)], dim=0)
    return slopes


class BloomAttention(nn.Module):

73
74
75
76
77
    def __init__(
        self,
        config: BloomConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ):
Woosuk Kwon's avatar
Woosuk Kwon committed
78
79
80
81
82
83
84
85
86
87
        super().__init__()
        self.hidden_size = config.hidden_size
        self.total_num_heads = config.n_head
        self.head_dim = self.hidden_size // self.total_num_heads
        assert self.head_dim * self.total_num_heads == self.hidden_size

        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

88
        self.query_key_value = QKVParallelLinear(
Woosuk Kwon's avatar
Woosuk Kwon committed
89
            self.hidden_size,
90
91
            self.head_dim,
            self.total_num_heads,
Woosuk Kwon's avatar
Woosuk Kwon committed
92
            bias=True,
93
            linear_method=linear_method,
Woosuk Kwon's avatar
Woosuk Kwon committed
94
95
96
97
98
        )
        self.dense = RowParallelLinear(
            self.hidden_size,
            self.hidden_size,
            bias=True,
99
            linear_method=linear_method,
Woosuk Kwon's avatar
Woosuk Kwon committed
100
101
102
103
104
105
106
107
108
109
        )

        # Create the alibi slopes and slice them.
        tp_rank = get_tensor_model_parallel_rank()
        head_start = tp_rank * self.num_heads
        head_end = (tp_rank + 1) * self.num_heads
        alibi_slopes = _get_alibi_slopes(self.total_num_heads)
        alibi_slopes = alibi_slopes[head_start:head_end].tolist()

        scaling = self.head_dim**-0.5
Woosuk Kwon's avatar
Woosuk Kwon committed
110
111
112
113
        self.attn = PagedAttention(self.num_heads,
                                   self.head_dim,
                                   scaling,
                                   alibi_slopes=alibi_slopes)
Woosuk Kwon's avatar
Woosuk Kwon committed
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134

    def forward(
        self,
        position_ids: torch.Tensor,
        hidden_states: torch.Tensor,
        kv_cache: KVCache,
        input_metadata: InputMetadata,
        cache_event: Optional[torch.cuda.Event],
    ) -> torch.Tensor:
        del position_ids  # Unused.
        qkv, _ = self.query_key_value(hidden_states)
        q, k, v = qkv.chunk(chunks=3, dim=-1)
        k_cache, v_cache = kv_cache
        attn_output = self.attn(q, k, v, k_cache, v_cache, input_metadata,
                                cache_event)
        output, _ = self.dense(attn_output)
        return output


class BloomMLP(nn.Module):

135
136
137
138
139
    def __init__(
        self,
        config: BloomConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ):
Woosuk Kwon's avatar
Woosuk Kwon committed
140
141
        super().__init__()
        hidden_size = config.hidden_size
142
143
144
        self.dense_h_to_4h = ColumnParallelLinear(
            hidden_size,
            4 * hidden_size,
145
            linear_method=linear_method,
146
        )
147
148
        quant_config = getattr(linear_method, "quant_config", None)
        self.gelu_impl = get_act_fn("gelu", quant_config, 4 * hidden_size)
149
150
151
        self.dense_4h_to_h = RowParallelLinear(
            4 * hidden_size,
            hidden_size,
152
            linear_method=linear_method,
153
        )
Woosuk Kwon's avatar
Woosuk Kwon committed
154
155
156

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x, _ = self.dense_h_to_4h(x)
157
        x = self.gelu_impl(x)
Woosuk Kwon's avatar
Woosuk Kwon committed
158
159
160
161
162
163
        x, _ = self.dense_4h_to_h(x)
        return x


class BloomBlock(nn.Module):

164
165
166
167
168
    def __init__(
        self,
        config: BloomConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ):
Woosuk Kwon's avatar
Woosuk Kwon committed
169
170
171
172
173
        super().__init__()
        hidden_size = config.hidden_size

        self.input_layernorm = nn.LayerNorm(hidden_size,
                                            eps=config.layer_norm_epsilon)
174
        self.self_attention = BloomAttention(config, linear_method)
Woosuk Kwon's avatar
Woosuk Kwon committed
175
176
        self.post_attention_layernorm = nn.LayerNorm(
            hidden_size, eps=config.layer_norm_epsilon)
177
        self.mlp = BloomMLP(config, linear_method)
Woosuk Kwon's avatar
Woosuk Kwon committed
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
        self.apply_residual_connection_post_layernorm = (
            config.apply_residual_connection_post_layernorm)

    def forward(
        self,
        position_ids: torch.Tensor,
        hidden_states: torch.Tensor,
        kv_cache: KVCache,
        input_metadata: InputMetadata,
        cache_event: Optional[torch.cuda.Event],
    ) -> torch.Tensor:
        # Layer norm at the beginning of the transformer layer.
        layernorm_output = self.input_layernorm(hidden_states)

        # Layer norm post the self attention.
        if self.apply_residual_connection_post_layernorm:
            residual = layernorm_output
        else:
            residual = hidden_states

        # Self attention.
        attention_output = self.self_attention(
            position_ids=position_ids,
            hidden_states=layernorm_output,
            kv_cache=kv_cache,
            input_metadata=input_metadata,
            cache_event=cache_event,
        )
        attention_output = attention_output + residual
        layernorm_output = self.post_attention_layernorm(attention_output)

        # Get residual
        if self.apply_residual_connection_post_layernorm:
            residual = layernorm_output
        else:
            residual = attention_output

        # MLP.
        output = self.mlp(layernorm_output) + residual
        return output


class BloomModel(nn.Module):

222
223
224
225
226
    def __init__(
        self,
        config: BloomConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ):
Woosuk Kwon's avatar
Woosuk Kwon committed
227
228
229
230
231
        super().__init__()
        self.embed_dim = config.hidden_size

        # Embedding + LN Embedding
        self.word_embeddings = VocabParallelEmbedding(
232
233
234
            config.vocab_size,
            self.embed_dim,
        )
Woosuk Kwon's avatar
Woosuk Kwon committed
235
236
237
238
        self.word_embeddings_layernorm = nn.LayerNorm(
            self.embed_dim, eps=config.layer_norm_epsilon)

        # Transformer blocks
239
240
241
242
        self.h = nn.ModuleList([
            BloomBlock(config, linear_method)
            for _ in range(config.num_hidden_layers)
        ])
Woosuk Kwon's avatar
Woosuk Kwon committed
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257

        # Final Layer Norm
        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)

    def forward(
        self,
        input_ids: torch.Tensor,
        position_ids: torch.Tensor,
        kv_caches: List[KVCache],
        input_metadata: InputMetadata,
        cache_events: Optional[List[torch.cuda.Event]],
    ) -> torch.Tensor:
        hidden_states = self.word_embeddings(input_ids)
        hidden_states = self.word_embeddings_layernorm(hidden_states)
        for i in range(len(self.h)):
258
            cache_event = None if cache_events is None else cache_events[i]
Woosuk Kwon's avatar
Woosuk Kwon committed
259
260
261
262
263
264
265
266
267
268
269
270
271
272
            layer = self.h[i]
            hidden_states = layer(
                position_ids,
                hidden_states,
                kv_caches[i],
                input_metadata,
                cache_event,
            )
        hidden_states = self.ln_f(hidden_states)
        return hidden_states


class BloomForCausalLM(nn.Module):

273
274
275
276
277
    def __init__(
        self,
        config: BloomConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ):
Woosuk Kwon's avatar
Woosuk Kwon committed
278
279
        super().__init__()
        self.config = config
280
281
        self.linear_method = linear_method
        self.transformer = BloomModel(config, linear_method)
Woosuk Kwon's avatar
Woosuk Kwon committed
282
283
284
285
286
287
288
289
290
291
        self.lm_head_weight = self.transformer.word_embeddings.weight
        self.sampler = Sampler(config.vocab_size)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        kv_caches: List[KVCache],
        input_metadata: InputMetadata,
        cache_events: Optional[List[torch.cuda.Event]],
292
    ) -> torch.Tensor:
Woosuk Kwon's avatar
Woosuk Kwon committed
293
294
        hidden_states = self.transformer(input_ids, positions, kv_caches,
                                         input_metadata, cache_events)
295
296
297
298
299
300
301
        return hidden_states

    def sample(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> SamplerOutput:
Woosuk Kwon's avatar
Woosuk Kwon committed
302
        next_tokens = self.sampler(self.lm_head_weight, hidden_states,
303
                                   sampling_metadata)
Woosuk Kwon's avatar
Woosuk Kwon committed
304
305
306
307
308
        return next_tokens

    def load_weights(self,
                     model_name_or_path: str,
                     cache_dir: Optional[str] = None,
Jasmond L's avatar
Jasmond L committed
309
310
                     load_format: str = "auto",
                     revision: Optional[str] = None):
311
        params_dict = dict(self.named_parameters(remove_duplicate=False))
Woosuk Kwon's avatar
Woosuk Kwon committed
312
        for name, loaded_weight in hf_model_weights_iterator(
Jasmond L's avatar
Jasmond L committed
313
                model_name_or_path, cache_dir, load_format, revision):
314
            if name == "lm_head.weight":
315
316
317
318
                continue
            if not name.startswith("transformer."):
                name = "transformer." + name
            param = params_dict[name]
Woosuk Kwon's avatar
Woosuk Kwon committed
319
320

            if "query_key_value" in name:
321
322
323
                # NOTE: BLOOM's fused QKV's output_dim has the shape of
                # (num_heads * 3 * head_size), while the
                # required shape is (3 * num_heads * head_size).
Woosuk Kwon's avatar
Woosuk Kwon committed
324
                # Thus, we need weight conversion.
325
                output_dim = getattr(param, "output_dim", None)
Woosuk Kwon's avatar
Woosuk Kwon committed
326
                num_heads = self.config.num_attention_heads
327
328
329
330
331
332
333
334
335
336
337
338
                if output_dim is not None:
                    loaded_weight_shape = loaded_weight.shape
                    loaded_weight = loaded_weight.view(
                        loaded_weight_shape[:output_dim] + (num_heads, 3, -1) +
                        loaded_weight_shape[output_dim + 1:])
                    loaded_weight = loaded_weight.transpose(
                        output_dim, output_dim + 1)
                    loaded_weight = loaded_weight.reshape(loaded_weight_shape)

            weight_loader = getattr(param, "weight_loader",
                                    default_weight_loader)
            weight_loader(param, loaded_weight)