File size: 8,417 Bytes
6a22ec9 |
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 |
# Copyright (c) Microsoft Corporation
# Licensed under the MIT License.
# pylint: disable=import-outside-toplevel
from __future__ import annotations
from typing import Any, Sequence
import torch
import onnxscript.tools.transformers_models
def has_phi3() -> bool:
"""Tells if package *transformers* contains the phi3 model."""
try:
from transformers import Phi3Config
assert Phi3Config
except ImportError:
return False
return True
def _prepare_config_and_inputs(
batch_size: int,
seq_length: int,
vocab_size: int,
type_sequence_label_size: int = 2,
type_vocab_size: int = 16,
num_labels: int = 3,
num_choices: int = 4,
use_input_mask: bool = False,
use_token_type_ids: bool = False,
use_labels: bool = False,
) -> tuple[Any, ...]:
input_ids = onnxscript.tools.transformers_models.ids_tensor(
[batch_size, seq_length], vocab_size
)
input_mask = None
if use_input_mask:
input_mask = torch.tril(torch.ones(batch_size, seq_length))
token_type_ids = None
if use_token_type_ids:
assert type_vocab_size > 0, "type_vocab_size is null"
token_type_ids = onnxscript.tools.transformers_models.ids_tensor(
[batch_size, seq_length], type_vocab_size
)
sequence_labels = None
token_labels = None
choice_labels = None
if use_labels:
assert type_sequence_label_size > 0, "type_sequence_label_size is null"
assert num_labels > 0, "num_labels is null"
assert num_choices > 0, "num_choices is null"
sequence_labels = onnxscript.tools.transformers_models.ids_tensor(
[batch_size], type_sequence_label_size
)
token_labels = onnxscript.tools.transformers_models.ids_tensor(
[batch_size, seq_length], num_labels
)
choice_labels = onnxscript.tools.transformers_models.ids_tensor(
[batch_size], num_choices
)
return (
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
)
def get_phi3_model(
input_dims: Sequence[tuple[int, int]] = ((13, 7), (14, 7), (15, 8)),
hidden_size: int = 32,
num_hidden_layers: int = 2,
vocab_size: int = 99,
intermediate_size: int = 16,
max_position_embeddings: int = 512,
num_attention_heads: int = 4,
num_key_value_heads: int = 2,
_attn_implementation: str = "eager", # needed value to remove graph breaks
with_mask: bool = True,
) -> tuple[Any, list[tuple[torch.Tensor, ...]], dict]:
"""
Returns a model.
See `PhiConfig
<https://huggingface.co/docs/transformers/main/en/model_doc/phi#transformers.PhiConfig>`_.
The parameters are chosen for a unit test configuration from `test_modeling_phi.py
<https://github.com/huggingface/transformers/blob/main/tests/models/phi/test_modeling_phi.py>`_.
"""
from transformers import Phi3Config, Phi3Model
dynamic_shapes = {0: {0: "batch", 1: "length"}}
if with_mask:
dynamic_shapes.update({1: {0: "batch", 1: "length"}})
config = Phi3Config(
hidden_size=hidden_size,
num_hidden_layers=num_hidden_layers,
vocab_size=vocab_size,
intermediate_size=intermediate_size,
max_position_embeddings=max_position_embeddings,
num_attention_heads=num_attention_heads,
num_key_value_heads=num_key_value_heads,
pad_token_id=min(32000, vocab_size - 1),
)
if _attn_implementation:
config._attn_implementation = _attn_implementation # pylint: disable=protected-access
if with_mask:
class Phi3ModelWrapperNoMask(torch.nn.Module):
def __init__(self, config):
super().__init__()
self.model = Phi3Model(config)
def forward(self, input_ids, attention_mask):
model_output = self.model(
input_ids, attention_mask=attention_mask, use_cache=False
)
return model_output.to_tuple()
def generate_example_inputs_no_mask(batch: int, seq: int, vocab_size: int):
(
input_ids,
_, # token_type_ids,
input_mask,
_, # sequence_labels,
_, # token_labels,
_, # choice_labels,
) = _prepare_config_and_inputs(
batch_size=batch,
seq_length=seq,
vocab_size=vocab_size,
use_input_mask=True,
)
return input_ids, input_mask
example_args_collection = []
for b, s in input_dims:
example_args_collection.append(generate_example_inputs_no_mask(b, s, vocab_size))
return Phi3ModelWrapperNoMask(config), example_args_collection, dynamic_shapes
# no mask
class Phi3ModelWrapper(torch.nn.Module):
def __init__(self, config):
super().__init__()
self.model = Phi3Model(config)
def forward(self, input_ids):
model_output = self.model(input_ids, use_cache=False)
return model_output.to_tuple()
def generate_example_inputs(batch: int, seq: int, vocab_size: int):
(
input_ids,
*_,
# token_type_ids,
# input_mask,
# sequence_labels,
# token_labels,
# choice_labels,
) = _prepare_config_and_inputs(
batch_size=batch,
seq_length=seq,
vocab_size=vocab_size,
use_input_mask=True,
)
return (input_ids,)
example_args_collection = []
for b, s in input_dims:
example_args_collection.append(generate_example_inputs(b, s, vocab_size))
return Phi3ModelWrapper(config), example_args_collection, dynamic_shapes
def get_phi3_model_from_config(
warmup: int = 5,
repeat: int = 10,
config: str = "small",
num_hidden_layers: int = 1,
implementation: str = "eager",
dynamic_shapes: bool = False,
with_mask: bool = True,
) -> tuple[Any, list[tuple[torch.Tensor, ...]], dict]:
"""
Returns a model Phi to test or benchmark.
Args:
warmup: Number of inputs to generate.
repeat: Number of inputs to generate for repeat.
config: small, medium or large
num_hidden_layers: number of hidden layers
implementation: eager or sdpa
with_mask: One or two inputs.
dynamic_shapes: dynamic shapes or not
Returns:
Model and list of inputs.
"""
if config == "small":
conf_dict = dict(
input_dims=onnxscript.tools.transformers_models.get_input_dims_for_llm(
dynamic_shapes, warmup, repeat
),
hidden_size=32,
num_hidden_layers=num_hidden_layers,
vocab_size=99,
intermediate_size=16,
max_position_embeddings=512,
num_attention_heads=4,
num_key_value_heads=2,
_attn_implementation=implementation,
with_mask=with_mask,
)
elif config == "medium":
conf_dict = dict(
input_dims=onnxscript.tools.transformers_models.get_input_dims_for_llm(
dynamic_shapes, warmup, repeat
),
hidden_size=1024,
num_hidden_layers=num_hidden_layers,
vocab_size=1024,
intermediate_size=1024,
num_attention_heads=4,
num_key_value_heads=4,
max_position_embeddings=1024,
_attn_implementation=implementation,
with_mask=with_mask,
)
elif config in ("large", "default"):
conf_dict = dict(
input_dims=onnxscript.tools.transformers_models.get_input_dims_for_llm(
dynamic_shapes, warmup, repeat
),
hidden_size=2048,
num_hidden_layers=num_hidden_layers,
vocab_size=51200,
intermediate_size=8192,
num_attention_heads=32,
num_key_value_heads=None,
max_position_embeddings=2048,
_attn_implementation=implementation,
with_mask=with_mask,
)
else:
raise ValueError(f"Unexpected configuration {config!r}.")
return get_phi3_model(**conf_dict) # type: ignore[arg-type]
|