Repository navigation
Expand file tree
/
Copy pathpooling_params.py
More file actions
226 lines (187 loc) · 8.08 KB
/
Copy pathpooling_params.py
File metadata and controls
226 lines (187 loc) · 8.08 K
8000
B
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
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from copy import deepcopy
from typing import Any
import msgspec
from vllm.config import ModelConfig, PoolerConfig
from vllm.logger import init_logger
from vllm.sampling_params import RequestOutputKind
from vllm.tasks import PoolingTask
logger = init_logger(__name__)
class LateInteractionParams(
msgspec.Struct,
omit_defaults=True, # type: ignore[call-arg]
array_like=True,
): # type: ignore[call-arg]
"""Metadata for worker-side late-interaction scoring.
Attributes:
mode:
- "cache_query": cache query token embeddings
- "score_doc": score a document against a cached query.
query_key: stable key used for both DP routing and worker cache lookup.
query_uses: expected number of document requests
"""
mode: str
query_key: str
query_uses: int | None = None
class PoolingParams(
msgspec.Struct,
omit_defaults=True, # type: ignore[call-arg]
array_like=True,
): # type: ignore[call-arg]
"""API parameters for pooling models.
Attributes:
use_activation: Whether to apply activation function to the pooler outputs.
`None` uses the pooler's default, which is `True` in most cases.
dimensions: Reduce the dimensions of embeddings
if model support matryoshka representation.
"""
# --8<-- [start:common-pooling-params]
use_activation: bool | None = None
# --8<-- [end:common-pooling-params]
## for embeddings models
# --8<-- [start:embed-pooling-params]
dimensions: int | None = None
# --8<-- [end:embed-pooling-params]
## for step pooling models
step_tag_id: int | None = None
returned_token_ids: list[int] | None = None
## Internal use only
task: PoolingTask | None = None
requires_token_ids: bool = False
skip_reading_prefix_cache: bool | None = None
late_interaction_params: LateInteractionParams | None = None
extra_kwargs: dict[str, Any] | None = None
output_kind: RequestOutputKind = RequestOutputKind.FINAL_ONLY
@property
def all_parameters(self) -> list[str]:
return ["dimensions", "use_activation"]
@property
def valid_parameters(self):
return {
"embed": ["dimensions", "use_activation"],
"classify": ["use_activation"],
"token_embed": ["dimensions", "use_activation"],
"token_classify": ["use_activation"],
}
def clone(self) -> "PoolingParams":
"""Returns a deep copy of the PoolingParams instance."""
return deepcopy(self)
def verify(self, model_config: ModelConfig) -> None:
# plugin task uses io_processor.parse_request to verify inputs,
# skipping PoolingParams verify
if self.task == "plugin":
if self.skip_reading_prefix_cache is None:
self.skip_reading_prefix_cache = True
return
# skipping verify, let plugins configure and validate pooling params
if self.task not in self.valid_parameters:
return
# NOTE: Task validation needs to done against the model instance,
# which is not available in model config. So, it's not included
# in this method
self._merge_default_parameters(model_config)
self._set_default_parameters(model_config)
self._verify_valid_parameters()
def _merge_default_parameters(self, model_config: ModelConfig) -> None:
pooler_config = model_config.pooler_config
if pooler_config is None:
return
assert self.task is not None, "task must be set"
valid_parameters = self.valid_parameters[self.task]
for k in valid_parameters:
if getattr(pooler_config, k, None) is None:
continue
if getattr(self, k, None) is None:
setattr(self, k, getattr(pooler_config, k))
if self.skip_reading_prefix_cache is None:
# If prefix caching is enabled,
# the output of all pooling may less than n_prompt_tokens,
# we need to skip reading cache at this request.
if self.task in ["token_embed", "token_classify"]:
self.skip_reading_prefix_cache = True
else:
self.skip_reading_prefix_cache = False
self._verify_step_pooling(pooler_config, valid_parameters)
def _verify_step_pooling(
self,
pooler_config: PoolerConfig,
valid_parameters: list[str],
):
step_pooling_parameters = ["step_tag_id", "returned_token_ids"]
if pooler_config.tok_pooling_type != "STEP":
invalid_parameters = []
for k in step_pooling_parameters:
if getattr(self, k, None) is not None:
invalid_parameters.append(k)
if invalid_parameters:
raise ValueError(
f"Task {self.task} only supports {valid_parameters} "
f"parameters, does not support "
f"{invalid_parameters} parameters"
)
else:
for k in step_pooling_parameters:
if getattr(pooler_config, k, None) is None:
continue
if getattr(self, k, None) is None:
setattr(self, k, getattr(pooler_config, k))
def _set_default_parameters(self, model_config: ModelConfig):
if self.task in ["embed", "token_embed"]:
if self.use_activation is None:
self.use_activation = True
if self.dimensions is not None:
if not model_config.is_matryoshka:
raise ValueError(
f'Model "{model_config.served_model_name}" does not '
f"support matryoshka representation, "
f"changing output dimensions will lead to poor results."
)
mds = model_config.matryoshka_dimensions
if mds is not None:
if self.dimensions not in mds:
raise ValueError(
f"Model {model_config.served_model_name!r} "
f"only supports {str(mds)} matryoshka dimensions, "
f"use other output dimensions will "
f"lead to poor results."
)
elif self.dimensions < 1:
raise ValueError("Dimensions must be greater than 0")
elif self.task in ["classify", "token_classify"]:
if self.use_activation is None:
self.use_activation = True
else:
raise ValueError(f"Unknown pooling task: {self.task!r}")
def _verify_valid_parameters(self):
assert self.task is not None, "task must be set"
valid_parameters = self.valid_parameters[self.task]
invalid_parameters = []
for k in self.all_parameters:
if k in valid_parameters:
continue
if getattr(self, k, None) is not None:
invalid_parameters.append(k)
if invalid_parameters:
raise ValueError(
f"Task {self.task!r} only supports {valid_parameters} "
f"parameters, does not support "
f"{invalid_parameters} parameters"
)
def __repr__(self) -> str:
return (
f"PoolingParams("
f"task={self.task}, "
f"dimensions={self.dimensions}, "
f"use_activation={self.use_activation}, "
f"step_tag_id={self.step_tag_id}, "
f"returned_token_ids={self.returned_token_ids}, "
f"requires_token_ids={self.requires_token_ids}, "
f"skip_reading_prefix_cache={self.skip_reading_prefix_cache}, "
f"late_interaction_params={self.late_interaction_params}, "
f"extra_kwargs={self.extra_kwargs})"
)
def __post_init__(self) -> None:
assert self.output_kind == RequestOutputKind.FINAL_ONLY, (
"For pooling output_kind has to be FINAL_ONLY"
)