-
Notifications
You must be signed in to change notification settings - Fork 78
/
generator.py
273 lines (207 loc) · 8.04 KB
/
generator.py
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
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
from __future__ import annotations
from abc import ABC, abstractmethod
from collections import OrderedDict
from dataclasses import dataclass
from typing import Protocol, final
from torch import Tensor
from torch.utils.hooks import RemovableHandle
from typing_extensions import override
from fairseq2.models.decoder import DecoderModel
from fairseq2.models.encoder_decoder import EncoderDecoderModel
from fairseq2.nn.padding import PaddingMask
class SequenceGenerator(ABC):
"""Represents a sequence generator."""
@abstractmethod
def __call__(
self, prompt_seqs: Tensor, prompt_padding_mask: PaddingMask | None
) -> SequenceGeneratorOutput:
"""
:param prompt_seqs:
The prompt sequences. *Shape:* :math:`(N,S)`, where :math:`N` is the
batch size and :math:`S` is the sequence length.
:param prompt_padding_mask:
The padding mask of ``prompt_seqs``. *Shape:* Same as ``prompt_seqs``.
"""
@abstractmethod
def register_step_hook(self, hook: StepHook) -> RemovableHandle:
"""Register a step hook on the sequence generator.
The hook will be called after every generation step.
:param hook:
The hook to register.
:returns:
A handle that can be used to remove the added hook by calling
``handle.remove()``.
"""
@property
@abstractmethod
def model(self) -> DecoderModel:
"""The associated decoder model."""
class AbstractSequenceGenerator(SequenceGenerator):
"""Provides a skeletal implementation of :class:`SequenceGenerator`."""
_model: DecoderModel
_step_hooks: dict[int, StepHook]
def __init__(self, model: DecoderModel) -> None:
"""
:param model:
The decoder model to use for generation.
"""
if model.vocab_info.eos_idx is None:
raise ValueError(
"`model.vocab_info` must have `eos_idx` set for sequence generation."
)
model.eval()
self._model = model
self._step_hooks = OrderedDict()
@final
@override
def register_step_hook(self, hook: StepHook) -> RemovableHandle:
handle = RemovableHandle(self._step_hooks)
self._step_hooks[handle.id] = hook
return handle
@final
@property
@override
def model(self) -> DecoderModel:
return self._model
@final
@dataclass
class SequenceGeneratorOutput:
"""Holds the output of a sequence generator."""
hypotheses: list[list[Hypothesis]]
"""The list of hypothesis generated per prompt, ordered by score."""
counters: GenerationCounters
"""The performance counters of the call."""
class Seq2SeqGenerator(ABC):
"""Represents a sequence-to-sequence generator."""
@abstractmethod
def __call__(
self,
source_seqs: Tensor,
source_padding_mask: PaddingMask | None,
prompt_seqs: Tensor,
prompt_padding_mask: PaddingMask | None,
) -> Seq2SeqGeneratorOutput:
"""
:param source_seqs:
The source sequences. *Shape:* :math:`(N,S,*)`, where :math:`N` is
the batch size, :math:`S` is the sequence length, and :math:`*` is
any number of sequence-specific dimensions including none.
:param source_padding_mask:
The padding mask of ``source_seqs``. *Shape:* :math:`(N,S)`, where
:math:`N` is the batch size and :math:`S` is the sequence length.
:param prompt_seqs:
The prompt sequences. *Shape:* :math:`(N,S_{prm})`, where :math:`N`
is the batch size and :math:`S_{prm}` is the prompt sequence length.
:param prompt_padding_mask:
The padding mask of ``prompt_seqs``. *Shape:* Same as ``prompt_seqs``.
"""
@abstractmethod
def register_step_hook(self, hook: StepHook) -> RemovableHandle:
"""Register a step hook on the sequence-to-sequence generator.
The hook will be called after every generation step.
:param hook:
The hook to register.
:returns:
A handle that can be used to remove the added hook by calling
``handle.remove()``.
"""
@property
@abstractmethod
def model(self) -> EncoderDecoderModel:
"""The associated encoder-decoder model."""
class AbstractSeq2SeqGenerator(Seq2SeqGenerator):
"""Provides a skeletal implementation of :class:`Seq2SeqGenerator`."""
_model: EncoderDecoderModel
_step_hooks: dict[int, StepHook]
def __init__(self, model: EncoderDecoderModel) -> None:
"""
:param model:
The encoder-decoder model to use for generation.
"""
if model.target_vocab_info.eos_idx is None:
raise ValueError(
"`model.vocab_info` must have `eos_idx` set for sequence generation."
)
model.eval()
self._model = model
self._step_hooks = OrderedDict()
@final
@override
def register_step_hook(self, hook: StepHook) -> RemovableHandle:
handle = RemovableHandle(self._step_hooks)
self._step_hooks[handle.id] = hook
return handle
@final
@property
@override
def model(self) -> EncoderDecoderModel:
return self._model
@final
@dataclass
class Seq2SeqGeneratorOutput:
hypotheses: list[list[Hypothesis]]
"""The list of hypothesis generated per prompt, ordered by score."""
encoder_output: Tensor
"""The encoder output used in encoder-decoder attention. *Shape:*
:math:`(N,S_{enc},M)`, where :math:`N` is the batch size, :math:`S_{enc}` is
the encoder output sequence length, and :math:`M` is the dimensionality of
the model."""
encoder_padding_mask: PaddingMask | None
"""The padding mask of :attr:`encoder_output`. *Shape:* :math:`(N,S_{enc})`,
where :math:`N` is the batch size and :math:`S_{enc}` is the encoder output
sequence length."""
counters: GenerationCounters
"""The performance counters of the call."""
@final
@dataclass
class Hypothesis:
"""Represents a hypothesis produced by a sequence generator."""
seq: Tensor
"""The generated sequence. *Shape:* :math:`(S)`, where :math:`S` is the
sequence length."""
score: Tensor | None
"""The score of the hypothesis. *Shape:* Scalar."""
step_scores: Tensor | None
"""The score of each sequence step. *Shape:* Same as ``seq``."""
@final
@dataclass
class GenerationCounters:
"""Holds the performance counters of a generator call."""
prefill_size: int = 0
"""The number of elements processed during the prefill step."""
num_generated_elements: int = 0
"""The number of generated elements."""
generation_time: float = 0
"""The generation time excluding prefill."""
cache_size: int = 0
"""The final size of the incremental cache in bytes."""
cache_capacity: int = 0
"""The final reserved capacity of the incremental cache in bytes."""
class StepHook(Protocol):
"""Represents a hook to pass to :meth:`~SequenceGenerator.register_step_hook`
or :meth:`~Seq2SeqGenerator.register_step_hook`."""
def __call__(
self,
prompt_indices: Tensor,
seqs: Tensor,
step_scores: Tensor | None,
prefill: bool,
) -> None:
"""
:param prompt_indices:
The indices of the input prompts corresponding to each sequence in
``seqs``. *Shape:* :math:`(N)`, where :math:`N` is the batch size.
:param seqs:
The sequences that are in process of being generated. *Shape:*
:math:`(N,S)`, where :math:`N` is the batch size and :math:`S` is
the sequence length generated so far.
:param step_scores:
The score of each step in ``seqs``. *Shape:* Same as ``seqs``.
:param prefill:
If ``True``, the hook is called as part of prompt prefill.
"""