File size: 11,109 Bytes
c34ff1f | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 | """
Unit tests for the model module in the Speculators library.
"""
import tempfile
from typing import Literal
import pytest
import torch
from torch import nn
from speculators import (
SpeculatorModel,
SpeculatorModelConfig,
SpeculatorsConfig,
VerifierConfig,
reload_schemas,
)
from speculators.models.dflash.core import DFlashDraftModel
from speculators.models.eagle3.core import Eagle3DraftModel
from speculators.models.mtp.core import MTPDraftModel
from speculators.models.peagle.core import PEagleDraftModel
from speculators.proposals import GreedyTokenProposalConfig
# ===== Test Helper Classes =====
@SpeculatorModelConfig.register("test_speculator_model")
class SpeculatorModelTestConfig(SpeculatorModelConfig):
speculators_model_type: Literal["test_speculator_model"] = "test_speculator_model"
test_param: int = 123
@SpeculatorModel.register("test_speculator")
class SpeculatorTestModel(SpeculatorModel):
config_class = SpeculatorModelTestConfig # type: ignore[misc]
def __init__(self, config: SpeculatorModelTestConfig, **kwargs):
super().__init__(config, **kwargs)
self.test_module = nn.Linear(10, 10)
self.post_init() # type: ignore[attr-defined]
def forward(self, *args, **kwargs):
# Simple implementation for testing
return {"logits": torch.randn(1, 10, 1000)}
@classmethod
def from_training_args(cls, verifier_config, **kwargs):
"""Create model from training arguments."""
config = SpeculatorModelTestConfig(
speculators_config=SpeculatorsConfig(
algorithm="test_speculator",
proposal_methods=[GreedyTokenProposalConfig()],
default_proposal_method="greedy",
verifier=VerifierConfig.from_config(
verifier_config, name_or_path=kwargs.get("verifier_name_or_path")
),
)
)
return cls(config=config)
@staticmethod
def get_trainer_kwargs(**kwargs):
"""Get training and validation kwargs."""
return {}, {}
# Reload registries to include test classes
reload_schemas()
@pytest.fixture
def speculator_model_test_config():
return SpeculatorModelTestConfig(
test_param=456,
speculators_config=SpeculatorsConfig(
algorithm="test_algorithm",
proposal_methods=[GreedyTokenProposalConfig()],
default_proposal_method="greedy",
verifier=VerifierConfig(
name_or_path=None,
architectures=["TestModel"],
),
),
)
# ===== SpeculatorModel Class Attributes Tests =====
@pytest.mark.smoke
def test_speculator_model_class_attributes():
assert SpeculatorModel.auto_package == "speculators.models"
assert SpeculatorModel.registry_auto_discovery is True
assert SpeculatorModel.config_class == SpeculatorModelConfig
assert SpeculatorModel.base_model_prefix == "model"
assert SpeculatorModel.main_input_name == "input_ids"
# ===== SpeculatorModel Registry Tests =====
@pytest.mark.smoke
def test_speculator_model_registry_contains_test_model():
assert SpeculatorModel.registry is not None
assert "test_speculator" in SpeculatorModel.registry
assert SpeculatorModel.registry["test_speculator"] == SpeculatorTestModel
@pytest.mark.smoke
def test_speculator_model_registered_model_class_from_config(
speculator_model_test_config,
):
model_class = SpeculatorModel.registered_model_class_from_config(
speculator_model_test_config
)
assert model_class == SpeculatorTestModel
@pytest.mark.sanity
def test_speculator_model_registered_model_class_from_config_invalid():
with pytest.raises(
TypeError, match="Expected config to be an instance of SpeculatorModelConfig"
):
SpeculatorModel.registered_model_class_from_config("invalid_config") # type: ignore[arg-type]
config = SpeculatorModelConfig(
speculators_model_type="test_speculator_model",
test_param=456,
speculators_config=SpeculatorsConfig(
algorithm="test_algorithm",
proposal_methods=[GreedyTokenProposalConfig()],
default_proposal_method="greedy",
verifier=VerifierConfig(
name_or_path="test/verifier",
architectures=["TestModel"],
),
),
)
with pytest.raises(
TypeError,
match="Received a SpeculatorModelConfig instance but expected a subclass",
):
SpeculatorModel.registered_model_class_from_config(config)
class UnregisteredConfig(SpeculatorModelConfig):
speculators_model_type: Literal["unregistered"] = "unregistered"
config = UnregisteredConfig(
speculators_config=SpeculatorsConfig(
algorithm="test_algorithm",
proposal_methods=[GreedyTokenProposalConfig()],
default_proposal_method="greedy",
verifier=VerifierConfig(
name_or_path="test/verifier",
architectures=["TestModel"],
),
)
)
with pytest.raises(
ValueError, match="No registered model class found for config type"
):
SpeculatorModel.registered_model_class_from_config(config)
# # ===== SpeculatorModel Initialization Tests =====
@pytest.mark.smoke
def test_speculator_model_initialization(speculator_model_test_config):
model = SpeculatorTestModel(speculator_model_test_config)
assert model.config == speculator_model_test_config
@pytest.mark.sanity
def test_speculator_model_initialization_invalid():
# No config
with pytest.raises(
ValueError, match="Config must be provided to initialize a SpeculatorModel"
):
SpeculatorModel(config=None) # type: ignore[abstract, arg-type]
# Invalid config type
with pytest.raises(
TypeError, match="Expected config to be an instance of SpeculatorModelConfig"
):
SpeculatorModel( # type: ignore[abstract]
config="invalid_config", # type: ignore[arg-type]
)
# ===== SpeculatorModel from_pretrained Tests =====
@pytest.mark.smoke
def test_speculator_model_from_pretrained_config(speculator_model_test_config):
state_dict = SpeculatorTestModel(speculator_model_test_config).state_dict() # type: ignore[attr-defined]
model = SpeculatorModel.from_pretrained(
None, config=speculator_model_test_config, state_dict=state_dict
)
assert isinstance(model, SpeculatorTestModel)
assert model.test_module is not None
assert model.test_module.weight.abs().sum() > 0 # Ensure weights are initialized
assert isinstance(model.config, SpeculatorModelTestConfig)
assert model.config.speculators_model_type == "test_speculator_model"
assert model.config.test_param == 456
@pytest.mark.smoke
def test_speculator_model_from_pretrained_local_marshalling(
speculator_model_test_config,
):
original_model = SpeculatorTestModel(speculator_model_test_config)
with tempfile.TemporaryDirectory() as tmpdir:
# Save the model to a local directory
original_model.save_pretrained(tmpdir) # type: ignore[attr-defined]
# Load the model from the local directory
loaded_model = SpeculatorModel.from_pretrained(tmpdir)
assert isinstance(loaded_model, SpeculatorTestModel)
assert loaded_model.test_module is not None
assert (
pytest.approx(
(loaded_model.test_module.weight - original_model.test_module.weight)
.detach()
.abs()
.sum()
)
== 0
)
assert isinstance(loaded_model.config, SpeculatorModelTestConfig)
assert loaded_model.config.speculators_model_type == "test_speculator_model"
assert loaded_model.config.test_param == 456
@pytest.mark.smoke
def test_speculator_model_from_pretrained(
speculator_model_test_config,
):
state_dict = SpeculatorTestModel(speculator_model_test_config).state_dict() # type: ignore[attr-defined]
model = SpeculatorModel.from_pretrained(
None, config=speculator_model_test_config, state_dict=state_dict
)
assert isinstance(model, SpeculatorTestModel)
assert isinstance(model.config, SpeculatorModelTestConfig)
assert model.config.speculators_model_type == "test_speculator_model"
assert model.config.test_param == 456
@pytest.mark.sanity
def test_speculator_model_from_pretrained_invalid(speculator_model_test_config):
with pytest.raises(
ValueError,
match="Either `config` or `pretrained_model_name_or_path` must be provided",
):
SpeculatorModel.from_pretrained(None)
with pytest.raises(
ValueError,
match="Either `pretrained_model_name_or_path` or `state_dict` must be provided",
):
SpeculatorModel.from_pretrained(None, config=speculator_model_test_config)
with pytest.raises(
TypeError, match="Expected config to be an instance of SpeculatorModelConfig"
):
SpeculatorModel.from_pretrained("test/path", config="invalid_config")
with pytest.raises(OSError, match="'path/does/not/exist'."):
SpeculatorModel.from_pretrained(
"path/does/not/exist", config=speculator_model_test_config
)
# # ===== SpeculatorModel Forward Method Tests =====
@pytest.mark.smoke
def test_speculator_model_forward_concrete(
speculator_model_test_config,
):
model = SpeculatorTestModel(speculator_model_test_config, verifier=None)
result = model.forward()
assert "logits" in result
assert result["logits"].shape == (1, 10, 1000)
@pytest.mark.smoke
def test_speculator_model_forward_abstract(speculator_model_test_config):
model = SpeculatorModel( # type: ignore[abstract]
speculator_model_test_config, verifier=None, verifier_attachment_mode=None
)
with pytest.raises(
NotImplementedError, match="The forward method is only supported on concrete"
):
model.forward()
@pytest.mark.smoke
@pytest.mark.parametrize(
"model_class",
[Eagle3DraftModel, DFlashDraftModel, PEagleDraftModel, MTPDraftModel],
)
def test_save_ignore_keys_are_ignored_on_load_missing(model_class):
"""Weights excluded from saved checkpoints (e.g. verifier_lm_head, which is
reloaded from the verifier via load_verifier_weights) must also be ignored when
missing on load. Otherwise loading an initialized/trained checkpoint flags the
absent key as missing.
"""
save_ignore = set(getattr(model_class, "_keys_to_ignore_on_save", None) or [])
load_missing_ignore = set(
getattr(model_class, "_keys_to_ignore_on_load_missing", None) or []
)
not_ignored_on_load = save_ignore - load_missing_ignore
assert not not_ignored_on_load, (
f"{model_class.__name__} excludes {sorted(not_ignored_on_load)} from saved "
"checkpoints but does not list them in _keys_to_ignore_on_load_missing; "
"loading a checkpoint will raise on the absent key(s)."
)
|